Theoretical and Mathematical Physics

Syed Twareque Ali

Jean-Pierre Antoine
Jean-Pierre Gazeau

Coherent States,
Wavelets,
and Their
Generalizations
Second Edition
Coherent States, Wavelets, and Their Generalizations
Theoretical and Mathematical Physics

The series founded in 1975 and formerly (until 2005) entitled Texts and Monographs
in Physics (TMP) publishes high-level monographs in theoretical and mathematical
physics. The change of title to Theoretical and Mathematical Physics (TMP) signals that
the series is a suitable publication platform for both the mathematical and the theoretical
physicist. The wider scope of the series is reflected by the composition of the editorial
board, comprising both physicists and mathematicians.
The books, written in a didactic style and containing a certain amount of elementary
background material, bridge the gap between advanced textbooks and research mono-
graphs. They can thus serve as basis for advanced studies, not only for lectures and
seminars at graduate level, but also for scientists entering a field of research.

Editorial Board
W. Beiglbck, Institute of Applied Mathematics, University of Heidelberg, Heidelberg,
Germany
P. Chrusciel, Gravitational Physics, University of Vienna, Vienna, Austria
J.-P. Eckmann, Dpartement de Physique Thorique, Universit de Genve, Geneva,
Switzerland
H. Grosse, Institute of Theoretical Physics, University of Vienna, Vienna, Austria
A. Kupiainen, Department of Mathematics, University of Helsinki, Helsinki, Finland
H. Lwen, Institute of Theoretical Physics, Heinrich-Heine-University of Dsseldorf,
Dsseldorf, Germany
M. Loss, School of Mathematics, Georgia Institute of Technology, Atlanta, USA
N.A. Nekrasov, IHS, Bures-sur-Yvette, France
M. Ohya, Tokyo University of Science, Noda, Japan
M. Salmhofer, Institute of Theoretical Physics, University of Heidelberg, Heidelberg,
Germany
S. Smirnov, Mathematics Section, University of Geneva, Geneva, Switzerland
L. Takhtajan, Department of Mathematics, Stony Brook University, Stony Brook, USA
J. Yngvason, Institute of Theoretical Physics, University of Vienna, Vienna, Austria

For further volumes:

http://www.springer.com/series/720
Syed Twareque Ali Jean-Pierre Antoine
Jean-Pierre Gazeau

Coherent States, Wavelets,

and Their Generalizations
Second Edition

123
Syed Twareque Ali Jean-Pierre Antoine
Department of Mathematics and Statistics Institut de Recherche en Mathmatique et
Concordia University Physique (IRMP)
Montral, Qubec Universit Catholique de Louvain
Canada Louvain-la-Neuve, Belgium

Jean-Pierre Gazeau
Astroparticules et Cosmologie
(APC, UMR 7164)
Universit Paris Diderot, Sorbonne Paris Cit
Paris, France

Centro Brasileiro de Pesquisas

Fisicas (CBPF)
Rio de Janeiro, Brasil

ISSN 1864-5879 ISSN 1864-5887 (electronic)

ISBN 978-1-4614-8534-6 ISBN 978-1-4614-8535-3 (eBook)
DOI 10.1007/978-1-4614-8535-3
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2013947161

Springer Science+Business Media New York 1999, 2014

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publishers location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

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

Preface to the Second Edition

More than 12 years have passed since the publication of the first edition of this
book and the topic of coherent states (CS) is more flourishing than ever. The sheer
abundance of new developments in the field made the appearance of a new edition
both desirable and indeed necessary. The range of physical applications of CS has
increased enormously in the past decade, from the traditional aspects of quantum
physics (nuclear, condensed matter, optics) all the way to quantum gravity and
quantum information theory. A good illustration of this situation is given by the
recent special issue [38] of the Journal of Physics A, devoted to CS, which, in 37
articles, covers much of the wide panorama of the field.
On the other hand, CS have turned into a genuine domain of mathematics, reach-
ing into applied group theory and harmonic analysis to the theory of quantization
and its probabilistic formulation. The topic of quantization, using coherent states,
was not dealt with in the first edition. We decided that the subject now fully merits
due attention in a book such as ours. The other domain of focus in this book, that
of wavelets, has continued its explosive growth. As before, our approach is to look
upon CS and wavelets as mathematically related aspects of one theory. Wavelets
on manifolds (spheres and other conics, for instance) and various generalizations
(ridgelets, curvelets, shearlets) have been developed, prompted by the demands
of signal and image processing. All these new aspects are covered, sometimes
in considerable detail, in the present edition. As a consequence, our bibliography
has grown considerably. Certain sections of the old book have been eliminated
or rewritten, while others have been merged in an attempt to clarify, update,
and streamline the presentation. On the other hand, there are other mathematical
developments, such as CS on Hilbert C - and W -modules and CS on quaternionic
Hilbert spaces, which are mentioned here but have not been discussed in any detail,
for otherwise the book would have grown to ponderous lengths.
Additionally, we should also mention two new textbooks on CS, namely, those of
Gazeau [Gaz09] (with emphasis on physical applications) and of Combescure and
Robert [Com12] (rather mathematically oriented).
In the course of the years past, all three of us have continued to teach and lecture
about the various aspects of CS and wavelets. In doing so, we have had the pleasure

v
vi Preface to the Second Edition

of discussing these topics with uncountably many colleagues and, of course, we

benefited enormously from these discussions. We wish to thank them all for their
fruitful suggestions. Particular thanks are due to Gitta Kutyniok and Daniela Rosca,
who read and commented on Chaps. 14 and 15, respectively, and to Herv Bergeron,
who influenced heavily the contents of Chap. 11.
Altogether, the subject of CS continues to be a fascinating domain, both in
physics and mathematics. As before we wanted to share this fascination with our
readers. We made an effort to make the book useful to beginning students as well as
to seasoned researchers.

Montral, QC, Canada Syed Twareque Ali

Louvain-la-Neuve, Belgium Jean-Pierre Antoine
Paris, France and Rio de Janeiro, Brazil Jean-Pierre Gazeau
Preface to the First Edition

Nitya kaaler utshab taba

Bishyer-i-dipaalika
Aami shudhu tar-i-mateer pradeep
Jaalao tahaar shikhaa1
Tagore

Should authors feel compelled to justify the writing of yet another book? In an
overpopulated world, must parents feel compelled to justify the bringing forth of yet
another child? Perhaps not! But an act of creation is also an act of love and a love
story can always be happily shared.
In writing this book, it has been the authors feeling that in all the wealth of
material on coherent states and wavelets, in which the literature on the subject
abounds, there is lack of a discernable unifying mathematical perspective. The
use of wavelets in research and technology has witnessed explosive growth in
recent years, while the use of coherent states in numerous areas of theoretical and
experimental physics has been an established trend since decades. Yet it is not at
all uncommon to find practitioners in either one of the two disciplines, who are
hardly aware of its links to the other. Currently, there are many books in the market
which treat the subject of wavelets from a wide range of perspectives and with
windows on one or several areas of a large spectrum of possible applications. On
the theory of coherent states, likewise, there exist several excellent monographs,
edited collections of papers, synthetic reviews or specialized articles. The emphasis
in most of these works is usually on physical applications. In the more mathematical
works, the focus is usually on specific propertiesarising either from group theory,
holomorphic function theory and, more recently, differential geometry. The point of
view put forward in this book is that both the theory of wavelets and the theory of

1 Thine is an eternal celebration . . . A cosmic Festival of Lights! . . . Therein I am a mere flicker

of a wicker lamp . . . O kindle its flame, (my Master!).

vii
viii Preface to the First Edition

coherent states can be subsumed into certain broad functional analytic structures,
namely, positive operator valued measures on a Hilbert space and reproducing
kernel Hilbert spaces. The specific context in which these structures arise, to
generate particular families of coherent states or wavelet transforms, could of course
be very diverse, but typically they emanate either from the property of square
integrability of certain unitary group representations on Hilbert spaces or from
holomorphic structures associated to certain differential manifolds. In talking about
square integrable representations, a broad generalization of the concept has been
introduced here, moving from the well-known notion of square integrability with
respect to the whole group to one based upon some of its homogeneous spaces. This
generalization, while often implicit in the past, in physical discussions of coherent
states (notably, in the works of Klauder, Barut and Girardello, or even in the case of
the time honoured canonical coherent states, discovered by Schrdinger), had not
been readily recognized in the mathematical literature until the work of Gilmore
and Perelomov. In this book, this generalization is taken even further, with the result
that the classes of coherent states that can be constructed and usefully employed
extend to a vast array of physically pertinent groups. Similarly, it is generally
known and recognized that wavelets are coherent states arising from the affine
group of the real line. But using coherent states of other groups to generate higher
dimensional wavelets, or alternative wavelet-like transforms, is not such a common
preoccupation among practitioners of the trade (an exception being the recent book
by Torrsani). About a third of the present book is devoted to looking at wavelets
from precisely this point of view, displaying thereby the richness of possibilities that
exists in this domain. Considerable attention has also been paid in the book to the
discretization problem, in particular with the discussion of -wavelets. The interplay
between discrete and continuous wavelets is a rich aspect of the theory, which does
not seem to have been exploited sufficiently in the past.
In presenting this unifying backdrop, for the understanding of a wide sweep of
mathematical and physical structures, it is the authors hope that the relationship
between the two disciplinesof wavelets and coherent stateswill have been
made more transparent, aiding thereby the process of cross-fertilization as well.
For graduate students or research workers, approaching the disciplines for the first
time, such an overall perspective should also make the subject matter easier to
assimilate with the book acting as a dovetailed introduction to both subjects
unfortunately, a frustrating incoherence blurs the existing literature on coherence!
Besides being a primer for instruction, the book, of course, is also meant to be
a source material for a wide range of very recent results, both in the theory of
wavelets and of coherent states. The emphasis is decidedly on the mathematical
aspect of the theory, although enough physical examples have been introduced, from
time to time, to illustrate the material. While the book is aimed mainly at graduate
students and entering research workers in physics and mathematical physics, it
is nevertheless hoped that professional physicists and mathematicians would also
find it interesting reading, being an area of mathematical physics in which the
intermingling of theory and practice is most thoroughgoing. Prerequisites for an
understanding of most of the material in the book is a familiarity with standard
Preface to the First Edition ix

Hilbert space operator techniques and group representation theory, such as every
physicist would acknowledge from the days of graduate quantum mechanics and
angular momentum. For the more specialized topics, an attempt has been made to
make the treatment self-contained and indeed, a large part of the book is devoted to
the development of the mathematical formalism.
If a book such as this can make any claims to originality at all, it can mainly
be in the manner of its presentation. Beyond that, the authors believe that there is
also a body of material presented here (for example in the use of POV measures or
in dealing with the discretization problem), which has not appeared in book form
before. An attempt has been made throughout to cite as many references to the
original literature as were known to the authorsomissions should therefore be
attributed to their collective ignorance and the authors would like to extend their
unconditional apologies for any resulting oversight.
This book has grown out of many years of shared research interest and indeed,
camaraderie, between the three authors. Almost all of the material presented here
has been touched upon in courses, lectures and seminars, given to students and
among colleagues at various institutions in Europe, America, Asia and Africa
notably in graduate courses and research workshops, given at different times by all
three authors, in Louvain-la-Neuve, Montral, Paris, Porto-Novo, Biaystok, Dhaka,
Fukuoka, Havana and Prague. One is tempted to say that the geographical diversity
here rivals the mathematical menagerie!
To all their colleagues and students who have participated in these discussions,
the authors would like to extend their heartfelt thanks. In particular, a few colleagues
graciously volunteered to critically read parts of the manuscript and to offer numer-
ous suggestions for improvement and clarity. Among them, one ought to specially
mention J. Hilgert, G.G. Emch, S. De Bivre and J. Renaud. In addition, the
figures would not exist without the programming skills of A. Coron, L. Jacques and
P. Vandergheynst (Louvain-la-Neuve), and we thank them all for their gracious help.
During the writing of the book, the authors made numerous reciprocal visits to each
others institutions. To Concordia University, Montral, the Universit Catholique
de Louvain, Louvain-la-Neuve and the Universit Paris 7Denis Diderot, Paris, the
authors would like to express their appreciation for hospitality and collegiality. The
editor from Springer-Verlag, Thomas von Foerster, deserves a special vote of thanks
for his cooperation and for the exemplary patience he displayed, even as the event
horizon for the completion of the manuscript kept receding further and further! It
goes without saying, however, that all responsibility for errors, imperfections and
residual or outright mistakes, is shared jointly by all three authors.

Montral, QC, Canada Syed Twareque Ali

Louvain-la-Neuve, Belgium Jean-Pierre Antoine
Paris, France and Rio de Janeiro, Brazil Jean-Pierre Gazeau
About the Authors

Dr. S.T. Ali is a full professor in the Department of Mathematics and Statistics,
Concordia University, Montreal. He has held teaching and research positions at
the Abdus Salam International Centre for Theoretical Physics (Trieste, Italy);
the Department of Mathematics (University of Toronto, Canada); the Institut fr
Theoretische Physike (Technische Universitt Clausthal, Germany). He has written
about 150 research publications, including three books.
J-P. Antoine has a full career as a professor of mathematical physics at Universit
Catholique de Louvain, Belgium, after postdocs at Princeton University, University
of Pittsburgh, and Universit de Genve. He was also invited professor at many
foreign universities (Universit Paris 7, Universit du Burundi, Fukuoka University
(Japan), Concordia University, Universit Nationale du Benin). He is now professor
emeritus, but continues his research activities. At the Universit Catholique de
Louvain, he was head of the Institute of Theoretical and Mathematical Physics
(FYMA), Chairman of the Department of Physics. His research interests cover a
wide range. From mathematical physics, in particular, the formulation of quantum
mechanics beyond Hilbert space, he moved to several innovative topics in pure
mathematics, such as partial inner product spaces and partial algebras of unbounded
operators. On the other hand, he also pursued several research directions in
mathematical physics, namely, classical gauge field theories, coherent states and
wavelets, including the applications of the latter in signal and image processing. He
is the author or coauthor of more than 130 research papers, four monographs, and
plenty of book chapters or conference papers.
J-P. Gazeau is a full professor of physics at the Universit Paris Diderot
(Sorbonne Paris Cit), France, a member of the Astroparticles and Cosmology
Laboratory (CNRS, UMR 7164), and currently Chairman of the Standing Com-
mittee of the International Colloquium on Group Theoretical Methods in Physics.
Having obtained his academic degrees from Sorbonne University and Pierre-and
Marie Curie Universit (Paris 6), he spent most of his academic career in Paris and
as invited professor and researcher in many other places, such as UCLA, Louvain,
Montreal, Prague, Newcastle, Krakow, Rio de Janeiro, and Sao Paulo. Professor
Gazeau has authored more than 180 scientific publications, including three books, in

xi
xii About the Authors

theoretical and mathematical physics, mostly devoted to group theoretical methods

in physics, coherent states, quantization methods, and number theory for aperiodic
systems.
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Canonical Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Minimal Uncertainty States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 The Group Theoretical Backdrop. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Some Functional Analytic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 A Complex Analytic Viewpoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 An Alternative Representation and Squeezed States . . . . . . . . . . . . . . . 31
2.6 Some Geometrical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.7 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Positive Operator-Valued Measures and Frames . . . . . . . . . . . . . . . . . . . . . . . 37
3.1 Definitions and Main Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.1 Examples of POV-Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 The Case of a Tight Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Frames and Semi-frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3.1 Frames Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3.2 Upper Semi-frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.3 Lower Semi-frames, Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 Discrete Frames and Semi-frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.1 Discrete Frames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.2 Weighted and Controlled Frames . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.3 Fusion Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.4 Discrete Semi-frames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4.5 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4 Some Group Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant Measures . . 61
4.1.1 A Simple Example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1.2 An Example Using the Affine Group . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Induced Representations and Systems of Covariance . . . . . . . . . . . . . . 69
4.2.1 Vector Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

xiii
xiv Contents

4.2.2 Discrete Series Representations of SU(1,1) . . . . . . . . . . . . . . . 76

4.2.3 The Regular Representations of a Group . . . . . . . . . . . . . . . . . . 83
4.3 An Extended Schurs Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4 Harmonic Analysis on Locally Compact Abelian Groups . . . . . . . . . 86
4.4.1 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4.2 Lattices in LCA Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.3 Sampling in LCA Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Lie Groups and Lie Algebras: A Reminder . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5.1 Lie Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5.2 Lie Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5.3 Extensions of Lie Algebras and Lie Groups. . . . . . . . . . . . . . . 98
4.5.4 Contraction of Lie Algebras and Lie Groups. . . . . . . . . . . . . . 101
5 Hilbert Spaces with Reproducing Kernels and Coherent States . . . . . . 105
5.1 A First Look at Reproducing Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.2 Some Illustrative Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.1 The Canonical CS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.2 An Example from a Hardy Space . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.3 An Example of VCS from a Matrix Domain . . . . . . . . . . . . . . 112
5.3 A Second Look at Reproducing Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3.1 A Motivating Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3.2 Measurable Fields and Direct Integrals . . . . . . . . . . . . . . . . . . . . 115
5.3.3 Example Using a POV Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4 Reproducing Kernel Hilbert Spaces: General Construction . . . . . . . . 117
5.4.1 Positive-Definite Kernels and Evaluation Maps. . . . . . . . . . . 118
5.4.2 Coherent States and POV Functions . . . . . . . . . . . . . . . . . . . . . . . 122
5.4.3 Some Isomorphisms, Bases, and -Selections . . . . . . . . . . . . 125
5.4.4 A Reconstruction Problem: Example of a
Holomorphic Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.5 Some Properties of Reproducing Kernel Hilbert Spaces . . . . . . . . . . . 130
6 Square Integrable and Holomorphic Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.1 Square Integrable Kernels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.2 Holomorphic Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.3 Some Examples of Coherent States from Square
Integrable Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.3.1 Standard Versus Circle Coherent States . . . . . . . . . . . . . . . . . . . 141
6.3.2 Coherent States for Motion on the Circle . . . . . . . . . . . . . . . . . 141
6.3.3 A General Holomorphic Construction . . . . . . . . . . . . . . . . . . . . . 143
6.3.4 Nonlinear Coherent States and Orthogonal Polynomials . 146
6.4 GazeauKlauder CS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.4.1 Coherent States for Discrete Spectrum Dynamics . . . . . . . . 151
6.4.2 Statistical and Semi-Classical Aspects . . . . . . . . . . . . . . . . . . . . 153
6.4.3 Imposing the Hamiltonian Lower Symbol . . . . . . . . . . . . . . . . 155
6.4.4 Action-Angle Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.4.5 Two Examples of Action-Angle Coherent States . . . . . . . . . 157
Contents xv

6.4.6 Coherent States for Continuum Dynamics . . . . . . . . . . . . . . . . 160

6.4.7 Coherent States for Discrete and Continuum Dynamics . 161
6.4.8 A General Expression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.5 CS on Quaternionic Hilbert Spaces and Hilbert Modules . . . . . . . . . . 163
7 Covariant Coherent States. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.1 Square-Integrable Covariant Coherent States . . . . . . . . . . . . . . . . . . . . . . . 166
7.1.1 A General Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.1.2 The GilmorePerelomov CS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.1.3 Vector and Matrix CS: A Geometrical Setting . . . . . . . . . . . . 170
7.2 Example: The Classical Theory of Coherent States . . . . . . . . . . . . . . . . 174
7.2.1 CS of Compact Semisimple Lie Groups. . . . . . . . . . . . . . . . . . . 174
7.2.2 CS of Noncompact Semisimple Lie Groups . . . . . . . . . . . . . . 177
7.2.3 CS of Non-Semisimple Lie Groups. . . . . . . . . . . . . . . . . . . . . . . . 179
7.3 Covariant CS: The General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.3.1 Generalized Perelomov CS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.3.2 Continuous Semi-Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7.3.3 Some Further Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.4 CS on Spheres Through Heat Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
7.5 CS in Loop Quantum Gravity and Quantum Cosmology . . . . . . . . . . 193
7.6 CS on Conformal Classical Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.6.1 Beyond Square Integrability and Singular Orbit . . . . . . . . . . 201
8 Coherent States from Square Integrable Representations . . . . . . . . . . . . . 203
8.1 Square Integrable Group Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.1.1 Example: The Connected Affine Group . . . . . . . . . . . . . . . . . . . 209
8.2 Orthogonality Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
8.3 A Class of Semidirect Product Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
8.3.1 Three Concrete Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
8.3.2 A Broader Setting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
8.4 A Generalization: - and V -Admissibility . . . . . . . . . . . . . . . . . . . . . . . . . . 226
8.4.1 Example of the Galilei Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
8.4.2 CS of the Isochronous Galilei Group . . . . . . . . . . . . . . . . . . . . . 236
8.4.3 Atomic Coherent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
9 CS of General Semidirect Product Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.1 Squeezed States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
9.2 Geometry of Semidirect Product Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
9.2.1 A Special Class of Orbits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
9.2.2 The Coadjoint Orbit Structure of . . . . . . . . . . . . . . . . . . . . . . . . 252
9.2.3 Measures on . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
9.2.4 Induced Representations of Semidirect Products . . . . . . . . . 258
9.3 CS of Semidirect Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
9.3.1 Admissible Affine Sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
xvi Contents

10 CS of the Relativity Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) . . . . . . . . . . . . . . . . . . . . . . 271
10.1.1 The Poincar Group P+ (1, 3) in 1 + 3 Dimensions . . . . . . 271
10.1.2 The Poincar Group P+ (1, 1) in 1 + 1 Dimensions . . . . . . 282
10.1.3 Poincar CS: The Massless Case . . . . . . . . . . . . . . . . . . . . . . . . . . 287
10.2 The Galilei Groups G(1, 1) and G G(3, 1) . . . . . . . . . . . . . . . . . . . . . . . 290
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2)
and Its Contraction(s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
11 Integral Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
11.1 What is Really Quantization? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
11.2 Sea Star Algebra. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
11.2.1 Exploring the Plane with a Five-Fold Set of Arms. . . . . . . . 307
11.2.2 What About the Continuous Frame?. . . . . . . . . . . . . . . . . . . . . . . 311
11.2.3 Probabilistic Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
11.2.4 Beyond Coherent States: Integral Quantization . . . . . . . . . . . 312
11.3 Integral Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
11.3.1 Covariant Integral Quantization. . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
11.4 Exemple: Weyl-Heisenberg Covariant Integral Quantization(s) . . . 316
11.4.1 Quantization with a Generic Weight Function . . . . . . . . . . . . 316
11.4.2 Regular and Isometric Quantizations . . . . . . . . . . . . . . . . . . . . . . 318
11.4.3 Elliptic Regular Quantizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
11.4.4 Elliptic Regular Quantizations that Are Isometric . . . . . . . . 321
11.4.5 Hyperbolic Regular Quantizations . . . . . . . . . . . . . . . . . . . . . . . . . 321
11.4.6 Hyperbolic Regular Quantizations that Are Isometric . . . . 321
11.4.7 Quantum Harmonic Oscillator According to . . . . . . . . . . . 321
11.4.8 Variations on the Wigner Function . . . . . . . . . . . . . . . . . . . . . . . . 322
11.5 Weyl-Heisenberg Integral Quantizations of Functions
and Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
11.5.1 Acceptable Probes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
11.5.2 Quantizable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
11.5.3 Quantizable Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
11.6 Quantization with Coherent States or Frame: General . . . . . . . . . . . . . 332
11.6.1 A First Example: Frame Quantization of Finite Sets . . . . . 333
11.6.2 A Second Example: CS Quantization of
Motion on the Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
11.6.3 Quantization With Action-Angle CS for
Bounded Motions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
11.7 Application to Various Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
12 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
12.1 A Word of Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
12.2 Derivation and Properties of the 1-D Continuous
Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
12.3 A Mathematical Aside: Extension to Distributions . . . . . . . . . . . . . . . . . 357
Contents xvii

12.4 Interpretation of the Continuous Wavelet Transform . . . . . . . . . . . . . . . 362

12.4.1 The CWT As Phase Space Representation . . . . . . . . . . . . . . . . 362
12.4.2 Localization Properties and Physical
Interpretation of the CWT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
12.5 Discretization of the Continuous WT: Discrete Frames . . . . . . . . . . . . 365
12.6 Ridges and Skeletons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
12.7 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
12.7.1 Application to NMR Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . 370
13 Discrete Wavelet Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
13.1 The Discrete WT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
13.1.1 Multiresolution Analysis, Orthonormal
Wavelet Bases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
13.1.2 Connection with Filters and the Subband
Coding Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
13.1.3 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
13.1.4 Applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
13.2 Towards a Fast CWT: Continuous Wavelet Packets . . . . . . . . . . . . . . . . 385
13.3 Algebraic Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
13.4 A Group-Theoretical Approach to Discrete Wavelet Transforms . . 390
13.4.1 Wavelets on the Finite Field Z p . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
13.4.2 Wavelets on Z p : Pseudodilations and Group Structure . . . 392
13.5 Wavelets on a Discrete Abelian Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
13.5.1 Compatible Filters: The General Case . . . . . . . . . . . . . . . . . . . . . 395
13.5.2 Compatible Filters in the Case G = R, A Z+ . . . . . . . . . . 400
13.5.3 Cohomological Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
13.5.4 Compatible Filters and Discretized Wavelet Transform . . 404
14 Multidimensional Wavelets and Generalizations . . . . . . . . . . . . . . . . . . . . . . . 411
14.1 Going to Higher Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
14.2 Mathematical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
14.2.1 The CWT in n Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
14.2.2 The CWT of Radial Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
14.3 The Two-Dimensional CWT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
14.3.1 Minimality Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
14.3.2 Interpretation, Visualization Problems, Calibration. . . . . . . 426
14.3.3 Practical Applications of the CWT in Two Dimensions . . 429
14.4 The Discrete WT in Higher Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
14.4.1 Applications of the 2-D DWT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
14.4.2 The DWT of Radial Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
14.5 Generalizations of 2-D Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
14.5.1 Continuous Wavelet Packets in Two Dimensions . . . . . . . . . 439
14.5.2 General Directional Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
14.5.3 Multiselective Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
14.5.4 Ridgelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
xviii Contents

14.5.5 Curvelets and Contourlets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447

14.5.6 Shearlets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
14.5.7 Geometrical Wavelets: Dictionaries, Molecules . . . . . . . . 455
15 Wavelets on Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
15.1 Wavelets on the Two-Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
15.1.1 Stereographic Wavelets on the Two-Sphere . . . . . . . . . . . . . . 458
15.1.2 Poisson Wavelets on the Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
15.1.3 Discrete Spherical Wavelets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
15.2 Wavelets on Other Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
15.2.1 Wavelets on Conic Sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
15.2.2 Wavelets on the n-Sphere and the Two-Torus . . . . . . . . . . . . . 479
15.2.3 Wavelets on Surfaces of Revolution . . . . . . . . . . . . . . . . . . . . . . . 482
15.2.4 Local Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
15.3 Wavelets on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
16 Wavelets Related to Affine Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
16.1 The Affine WeylHeisenberg Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
16.2 The Affine or Similitude Groups of Spacetime . . . . . . . . . . . . . . . . . . . . . 499
16.2.1 Kinematical Wavelets, Motion Analysis . . . . . . . . . . . . . . . . . . 499
16.2.2 The Affine Galilei Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
16.2.3 The (Restricted) GalileiSchrdinger Group . . . . . . . . . . . . . . 506
16.2.4 The Affine Poincar Group and the Conformal Group . . . 509
16.3 Some Generalizations: Wavelets on Riemannian
Symmetric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
17 The Discretization Problem: Frames, Sampling, and All That . . . . . . . . 515
17.1 The WeylHeisenberg Group or Canonical CS . . . . . . . . . . . . . . . . . . . . . 516
17.2 Wavelet Frames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
17.2.1 Alternative Approaches to the ax + b Group. . . . . . . . . . . . 520
17.3 Frames for Affine Semidirect Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
17.3.1 The Affine WeylHeisenberg Group. . . . . . . . . . . . . . . . . . . . . . . 521
17.3.2 The Affine Poincar Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
17.3.3 Discrete Frames for General Semidirect Products . . . . . . . . 524
17.4 Groups Without Dilations: The Poincar Groups . . . . . . . . . . . . . . . . . . 526
17.4.1 The Poincar Group P+ (1, 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
17.4.2 The Poincar Group P+ (1, 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
17.5 Generalities on Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
18 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
18.1 Present Status of CS and Wavelet Research. . . . . . . . . . . . . . . . . . . . . . . . . 537
18.2 What Is the Future from Our Point of View? . . . . . . . . . . . . . . . . . . . . . . . 539
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
Chapter 1
Introduction

Abstract We start with a description of the canonical coherent states (CS) and
some historical remarks on the evolution of the concept and its applications. Then
we present in detail the organization of the book.

The notion of coherent states (CS1 ) is rooted in quantum physics and its relationship
to classical physics. The term coherent itself originates in the current language of
quantum optics (for instance, coherent radiation, sources emitting coherently, etc.).
It was introduced in the 1960s by Glauber [333,334],2 one of the founding fathers of
the theory of CS, together with Klauder [409,410], Klauder and Sudarshan [Kla68],
and Sudarshan [570], in the context of a quantum optical description of coherent
light beams emitted by lasers. Since then, coherent states have pervaded nearly
all branches of quantum physicsquantum optics, of course, but also nuclear,
atomic, and solid-state physics, quantum electrodynamics (the infrared problem),
quantization and dequantization problems, path integrals, quantum gravity, quantum
information theory, just to mention a few. It has been said, even convincingly [412],
that coherent states are the natural language of quantum theory!
As early as 1926, at the very beginning of quantum mechanics, Schrdinger [552]
was interested in studying quantum states, which restore the classical behavior of
the position operator of a quantum system (in the Heisenberg picture):
i i
Q(t) = e h Ht Q e h Ht . (1.1)

1 The acronym CS will be used throughout this book to mean coherent state or coherent states,

depending on the context.

2 For convenience, we have split the bibliography into two separate lists, books and theses (denoted

by letters) and articles (denoted by numbers).

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 1
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__1,
Springer Science+Business Media New York 2014
2 1 Introduction

In this relation, H = P2 /2m + V (Q) is the quantum Hamiltonian of the system.

Schrdinger understood classical behavior to mean that the average or expected
value q(t) of the position operator Q(t), in the desired state, would obey the classical
equation of motion:

V
mq(t) + = 0. (1.2)
q
The first example of CS discovered by Schrdinger were of course those pertaining
to the harmonic oscillator, V (q) = 12 m2 2 q2 , known universally to physicists,
and the prototype of every integrable model. These states |z, parametrized by
the complex number z, are defined in such a way that one recovers the familiar
sinusoidal solution

z | Q(t) | z = 2Qo |z| cos( t ), (1.3)

where z = |z|ei and Qo = (h/2m )1/2 is a fundamental quantum length built from
the universal constant h and the constants m and characterizing the quantum
harmonic oscillator under consideration.
In this way, CS mediate a smooth transition from classical mechanics to
quantum mechanics. One should not be misled, however: CS are rigorously quantum
states (witness the constant h appearing in the definition of Qo ), yet they allow for a
classical reading in a host of quantum situations. This unique qualification results
from a set of properties satisfied by these SchrdingerKlauderGlauberSudarshan
CS, also called canonical CS. The most important among them, and indeed the
guiding spirit in the organization of the present volume, are as follows:
P1. The states |z saturate the Heisenberg inequality:

1
 Qz  Pz = h, (1.4)
2

where  Qz := [z|Q2 |z z|Q|z2 ]1/2 .

P2. The states |z are eigenvectors of the annihilation operator, with eigenvalue z:

a|z = z|z, z C, (1.5)

where a = (2mh )1/2 (m Q + iP).

P3. The states |z are obtained from the ground state |0 of the harmonic oscillator
by a unitary action of the WeylHeisenberg group. This is a key Lie group in
quantum mechanics, whose Lie algebra is generated by {Q, P, I}, where I is the
identity operator, with [Q, P] = i hI (which implies [a, a ] = I):

|z = e(za za) |0,

(1.6)

where e(za za) := D(z) is the so-called displacement operator.

1 Introduction 3

P4. The CS {|z} constitute an overcomplete family of vectors in the Hilbert space
of the states of the harmonic oscillator. This property is encoded in the following
resolution of the identity:

1
I= |zz| d Re z d Im z . (1.7)
C

These four properties are, to various extents, the basis of the many generalizations
of the canonical notion of CS, illustrated by the family {|z}.
It is important to note here that the properties P1P4, satisfied by the canonical
CS, set them apart radically from ordinary wave packets, familiar from elementary
quantum mechanics. Indeed, these latter states are often characterized by good
localization at a given instant, either in space or in momentum (but not in both), but
generically they spread in the course of their time evolution. Moreover, one could
not build a continuous family of such wave packets and expect them to generate a
resolution of the identity. In this sense, the canonical CS constitute a unique family
of wave packets that maintain the same support, both in space and in momentum, as
exemplified by P1 (coherent states remain coherent!), and also resolve the identity
in the sense of P4.
We emphasize, however, that for the various generalizations of CS presented
in this book (or elsewhere) there is a priori no reason to expect any localization
property, either in space or in some other observable, and, even if they do have
such a property, they have no reason to preserve it in time. For instance, localization
requires the existence of a well-defined position operator, and this is not always
available (e.g., in a relativistic quantum mechanics). For a pedagogical review of
some of these questions and references to original articles, we refer the reader
to [153].
Let us now give some examples of such generalizations, with no pretention
of being exhaustive. (Most of these examples are borrowed from the well-known
reprint volume edited by Klauder and Skagerstam [Kla85], in which original
references may also be found.)
Concerning Property P1, one might mention the approach of Nieto and Simmons
[485, 486], who use it as a defining criterion for more general CS adapted to a local
potential that has at least one region of confinement.
Property P2 was the takeoff point of Barut and Girardello [126] for constructing
CS associated with representations of the group SL(2, R) SU(1, 1) SOo (1, 2),
which belong to the so-called discrete series. These are representations on Hilbert
spaces that can be constructed with the help of a raising operator a and its adjoint a,
the commutator of the two being equal to the generator Xo of the compact subgroup
SO(2) U(1):

[a, a ] = Xo . (1.8)

Thus, the BarutGirardello CS are again states |z, which satisfy the equation

a|z = z|z, z C. (1.9)

4 1 Introduction

Actually, the same method was used by Atkins and Dobson [105] for building
angular momentum or spin CS, associated to SU(2). Here the angular momentum
operators are obtained, following Schwinger, by combining two sets of harmonic
oscillator generators, as described in the work of Carruthers and Nieto on phase
and angle variables [195]. The same approach has been used by various authors for
generalizing CS to supersymmetric quantum mechanics. In that context, one may
quote also the paper of BergeronValance [140], who obtain oscillator-like CS for
a one dimensional Hamiltonian with an arbitrary potential, using a supersymmetric
formalism.
Property P3, which has a clear group-theoretical flavor, is at the core of the
approach developed independently by Gilmore [330, 331] and Perelomov [502]
(and, in spirit, already anticipated by Klauder [409, 410]). The basic ingredients
here are a group G, a subgroup H of G, a unitary representation U of G in a Hilbert
space H, and a probe H, chosen because of some particular good properties
that it possesses. A family of CS {x }, indexed by the points x of the coset space
X = G/H, is constructed as follows:

x = U( (x)) , x X, (1.10)

where : G/H G is a suitable section or lifting : X Gthat is, a map

associating to each x X a certain element (x) G, such that the coset (x)H
is precisely x (for Lie groups, this is obviously a fiber bundle construction). In a
picturesque way, one may say that the probe is transported by the group G
modulo H.
Finally, Property P4 is in fact, both historically and conceptually, the one
that survives (or should survive, presumably after some technical modifications)
once one has been forced (however reluctantly!) to abandon some of the other
propertiesperhaps even all of them. As far as physical applications are concerned,
this property has gradually emerged as the one most fundamental for the analysis, or
decomposition, of states in the Hilbert space of the problem, or of operators acting
on this space. Thus, it will come as no surprise that Property P4 is the raison dtre
of the present volume. We shall present a comprehensive survey of this point of
view in the following pages. Very schematically, given a measure space (X, ) and
a Hilbert space H, a family of CS {x | x X} must satisfy an operator identity of
the form

|x x | d (x) = A, (1.11)
X

where A is a positive, invertible operator on H. Note that (1.11) must be understood

mathematically with a certain bit of caution (for instance, it must be interpreted in a
weak sense, that is, in terms of expectation values in states, possibly even resorting
to a triplet of spaces H1 H H1 , in the sense of distribution theory). In the ultimate
analysis, what is desired is to make the family {x } operational through the identity
(1.11). This means being able to use it as a frame through which one reads the
1 Introduction 5

information contained in an arbitrary state in H, in an operator on H, or in a setup

involving both operators and states, such as an evolution equation on H. At this
point, it is legitimate to borrow from the language of quantum physics and say that
(1.11) realizes a quantization of the classical space (X, ) and the measurable
functions on it. The resulting CS or integral quantization will be studied in depth in
Chap. 11, which is new in this second edition.
One could, however, get rid of this quantum mechanical dressing as well,
which goes back to Schrdingers original work, and instead switch squarely to
the terminology of signal processing. In that setting, H is a Hilbert space of finite
energy signals and (X, ) is a space of parameters, suitably chosen for emphasizing
certain aspects of the signal that may interest us in particular situations. Every signal
contains noise, but the nature and the amount of noise is different for different
signals. In this context, choosing (X, , {x }) amounts to selecting a part of the
signal that we wish to isolate and interpret, while eliminating a noise that has (once
and for all) been deemed useless. Here too, we have in effect chosen a frame. In our
opinion, this formal analogy between the CS quantization of (X, ), namely,

x  |x x |, (1.12)

on the one hand, and signal processing, on the other, merits a deeper investigation.
It is another goal of the present volume to induce the reader to undertake such a
reflection.
A perfect illustration of the deep analogy between quantum mechanics and signal
processing is provided by two widely used techniques, namely, Gabor analysis and
wavelet analysis. While the former goes back to the work of Gabor in the 1940s
[294], the latter is a relatively recent (about 25 years old) approach in signal and
image analysis that has undergone an explosive development, with applications to
a large number of topics in engineering, applied mathematics, and physics. The
crucial point is that, while Gabor analysis yields a time-frequency representation
of the signal, wavelets provide a time-scale representation. The built-in scaling
operation makes it a very efficient tool for analyzing singularities in a signal, a
function, an image, and so on; that is, the portion of the signal that contains the
most significant information. In addition, wavelets can be adapted to signals living
on various types of manifolds. For instance, wavelets on the two-sphere have found
application in the analysis of the cosmic background radiation (CMB).
Now, not surprisingly, wavelets are CS, namely, those associated with the
affine group of the appropriate dimension, consisting of translations, dilations,
and rotations (in dimensions higher than one). In fact, it was Aslaksen and
Klauder [102, 103] who first studied the one-dimensional affine group, for the
purely quantum mechanical purpose of generalizing Property P1 above to dilations-
translations [411], and it was yet another mathematical physicist, Grossmann and
Morlet [349] and Grossmann et al. [351, 352], who discovered the crucial link
between the representations of the affine group and an intriguing technique in
signal analysis developed by the geophysicist Jean Morlet. Once again, we see
here CS bridging the gap between classical physics (this time microseismology in
6 1 Introduction

oil prospecting!) and quantum mechanical techniques. This explains why wavelets
naturally find their place in a book on coherent states: They constitute one of the
most promising developments in the venerable theory of coherent states.
Let us come back for a while to the original motivation for introducing CS,
namely, to study quantum reality through a framework formally similar to classical
reality. As a matter of fact, a formalism based on Hilbert space or distribution
theory (such as a complete set of commuting observables) is often heavy,
cumbersome, replete with countless indices, and perhaps even turns out to be
intractable. On the contrary, the use of CS restores to quantum states the appearance
of classical points [see (1.11)], and to observables the practicality of a functional
formulation in terms of symbols. The latter come in two varieties [135,136,333,334],
namely, to each observable O, one associates the covariant symbol O  and the

contravariant symbol O:

  , x) = x |O|x 
O(x and O=

O(x) |x x | d (x). (1.13)
x |x  X

It was precisely this symbolic formulation that enabled Glauber and others [333,
334, 371, 372, 445] to treat a quantized boson or fermion field like a classical field,
particularly, for computing correlation functions or other quantities of statistical
physics, such as partition functions and derived quantities. Indeed, these quantities
are often given as the trace of a product of operators (e.g., the density operator of the
system), in which the use of the resolution of the identity (1.11) (with x |x  = 1)
is particularly efficient. For instance:

Tr [O1 O2 . . . ON ] = x |A1/2 O1 O2 . . . ON A1/2 |x  d (x)
X
N 
= xi |A1/2 Oi A1/2 |xi+1  d (xi )
i=1 X
N 
= (A1/2 Oi A1/2 )(xi , xi+1 ) d (xi ),
i=1 X

where xN+1 x1 x. Such expressions are convenient for deriving estimates when
considering various limits, such as a thermodynamical limit or a path integral
[Kla11]. In particular, one can follow the dynamical evolution of a system in
a classical way, elegantly going back to the study of classical trajectories
in the space X. It is a striking fact that the latter, which is often a genuine
phase space or a space of classical states, is being used as a parameter space
for quantum events. We encounter here, in a purely operational fashion, a debate
whose formulation goes back to the beginnings of quantum theory, namely, the
preeminence given by the physicistor the signal engineerto aspects of reality
that appear most immediate on the phenomenological level and are then called
classical or quantal, depending on ones practice or epistemological preference.
1 Introduction 7

So much for the physical background of the book. Indeed, the reader will
readily sense a strong mathematical bias throughout the work. The reason for
this is twofold. On the one hand, there are several review papers or books on
the market that cover the physical developments of the theory of CS. Besides
the KlauderSkagerstam reprint volume already quoted [Kla85], we also ought to
mention the review article by Zhang et al. [617] and the proceedings of the 1993
Oak Ridge Conference, entitled Coherent States: Past, Present and Future [Fen94],
celebrating the 50th anniversary of the GlauberKlauder CS. More recently, there
appeared review papers by Dodonov [256], Vourdas [597] and Robert [521], and
the textbooks of DodonovManko, mostly devoted to quantum optics [Dod03],
BengtssonZyczkowski [Ben07], which contains a lot of pictorial information
about CS, Gazeau [Gaz09], more oriented towards the physical, probabilistic
and quantization aspects, and CombescureRobert [Com12], more mathematically
oriented. In addition, there appeared also a special issue of the Journal of Physics
A, devoted to CS and covering in 37 articles most aspects of the field [38]. This
collection illustrates perfectly the extreme versatility of the CS concept. Indeed,
coherent states constitute nowadays a flourishing research topic, with applications
to a wide spectrum of domains. They are everywhere in physics: condensed matter
physics, atomic physics, nuclear and particle physics, quantum optics, dynamics
both quantum and classical potentials, quantum gravity, quantization and quantum
information theory.
On the other hand, the mathematical developments around the concept of CS
have been equally spectacular, if less well-known among physicists. In fact, the only
comprehensive book (until the first edition of the present work, in 2000) on the topic
is the monograph of Perelomov [Per86], going back to 1986. As we shall see, far-
reaching generalizations of the standard CS have appeared in the meantime, which
substantially change the perspective. An early survey of these developments is given
in our 1995 review article [31]. Since then, CS have grown into a full-fledged
domain in mathematics, incorporating many tools such as group representations,
POV measures, frames, holomorphic functions, orthogonal polynomials and so on.
A recent review paper is that of Robert [521], but a more comprehensive panorama
is clearly needed, hence the present second edition.
A third field related to CS is signal processing. Indeed both Gabor analysis and
wavelet analysis derive in the first place from CS theory, namely, CS associated to
the WeylHeisenberg and the ax + b group, respectively. Here too, a tremendous
development has taken place in the last years, another testimony to the richness
of the notion of CS. Throughout the book, we will strive to illustrate many of
these topics, starting from the basic definitions and providing references to the most
significant papers on applications of CS.
Coming back to the title of the Oak Ridge volume, it is not our intention (nor is
it within our competence) to write here a detailed history of the past. Concerning
the present, a good testimony to the immediate pertinence of our topic is the
status of CS for the hydrogen atom. As early as his 1926 work, Schrdinger [552]
had tried to construct CS appropriate to the hydrogen atom problem, but without
success. His search was motivated by the quasi-classical character of the canonical
8 1 Introduction

CS, which made them very desirable for studying the quantization of a classical
dynamical system, a point that is still pertinent today. What Schrdinger sought, but
could not obtain, was a set of well-localized wave packets, concentrated around the
classical Kepler orbits, and thus permitting an easy transition from the classical
to the quantum description of the hydrogen atom. Several systems of CS have
been proposed since Schrdingers original work, derived mainly from the groups
SU(1,1) and SO(4,2), although in a regularized sense (the latter is the full dynamical
group of the H-atom; the former is the subgroup describing its radial motion only),
but none of them is fully convincing (see [414] for a list of references and also
Sect. 7.6).
In this context, we may note that the two most celebrated models of classical
and quantum physics, namely the r2 harmonic oscillator potential (stability around
an equilibrium position) and the 1/r Kepler potential (Gausss theorem) are at
the same time mathematically very close (they may even be transformed into
each othersee below), and yet profoundly different from the point of view of
physical interpretation. The confining property of the former potential leads to a
purely discrete, semibounded spectrum, whereas the latter simultaneously exhibits
(at least, for the radial part) a bounded discrete spectrum and a positive continuum.
It is a highly nontrivial task to construct CS for the H-atom from a superposition
of its (generalized) eigenstates, so as to satisfy Properties P1 and P4 or, if that
is too stringent a requirement, to obtain states with a satisfactory semiclassical
behavior, i.e., their temporal evolution remains always well localized around the
corresponding classical phase space trajectory, as it is the case for canonical CS.
Of course, an alternative approach is possible, exploiting the mathematical
equivalence between the harmonic oscillator and the H-atom, regularized at the
origin and restricted to negative energies. Indeed, the so-called OC-transformation
maps one problem into the other, interchanging energy and coupling constant
[304, Kem37]. Thus there appears a natural basis in the state space of the H-atom,
the Sturmian basis {i, ( x)}, where i (n, l, m) is the standard triple of quantum
numbers [300303, 540]. This basis is orthonormal with respect to the measure
r1 dx and the basis vectors depend on the dilation parameter as follows:

i, ( x) = [U( )hi ](x) = 3/2 hi ( x).

Choosing = 1/n for i = (n, l, m) yields precisely the incomplete set of H-atom
bound states. The dynamical symmetries SO(4,1) and SO(4,2) of the Keplerian
potential are in fact the ones that describe best the Sturmian system, in the sense
that they reflect the symmetries of the harmonic oscillator. The function hi is the
eigenfunction of the operator r(1 ) corresponding to the eigenvalue n. Thus it is
natural to construct Sturmian CS related to SO(4,1) or SO(4,2) (see for instance
[618]), for this construction is fully in the spirit of the canonical CS, or at least
of the BarutGirardello CS [126]. Within the same circle of ideas, we may also
mention the CS built by Onofri [494], again based on SO(4,2) and close to the spirit
of the present book. We will review this construction in Sect. 7.6. In fact, we feel
that the Sturmian approach has not been totally exploited in the construction of
CS. For instance, Sturmian functions could be used as probes for building orbits
1 Introduction 9

coherent under the action of a relativity group, such as the Galilei or the affine
Galilei group. CS for these groups will be discussed at length in Sects. 10.2 and
16.2.2, respectively.
However, no matter how elegant they are, all these CS give only a very imperfect
answer to the physical problems raised by some spectacular experiments in atomic
physics. Indeed, the Kepler wave packets, that Schrdinger looked for, have by
now been seen experimentally, and they exhibit the fascinating phenomenon of
revival: A state, initially well localized, spreads all over its Kepler orbit, but,
after a large number of revolutions, reconcentrates itself where it began, passing
in between through various multilocalized stages (see [455, 481, 611, 612] and
the popular account in [482]). Moreover, the story underwent a new development
when Klauder introduced a new set of CS for the H-atom [414], abandoning the
group-theoretical underpinning, but imposing time stability (a CS must remain a CS
throughout its time evolution), as well as continuity in the labels, that is, continuity
of the map x  x (this condition was in fact always assumed, at least implicitly).
While these new states do not yet fully answer the original question (they do not
follow the classical Kepler orbits with little or no dispersion [130]), they open the
door to a whole host of new CS, adapted to systems with either a discrete or a
continuous spectrum, or both, and are specified by insisting on classical action
angle variables [309]. We will come back to these CS (often called GazeauKlauder
CS) in Sect. 6.4. So the present stage is still very lively!
As for the future, we will not take the risk of making definite predictions. Only
one thing is certain: CS will continue to occupy a central position in quantum
theories, and new developments may be expected both in the mathematical aspects
and for specific physical applications.
A word now about the organization of the book. First, let us emphasize that we
will skip a number of topics that the reader might expect, because of their current
fashionable character. To mention a few, we will not discuss superCS (that is, CS
associated with supergroups or supersymmetries), q-deformed CS, nor the use of CS
for defining path integrals. Some of these topics have been described, for instance,
in our review paper [31], in the KlauderSkagerstam volume [Kla85], in the book
by Gazeau [Gaz09], or in the latest book of Klauder [Kla11].
Let us now go into the details. Quite naturally, we start, in Chap. 2, with a
comprehensive survey of the canonical CS, discussing in detail properties P1P4
above. In particular, P3 means that the canonical coherent states can be obtained by
acting on the oscillator ground state with the operators of a unitary representation
U of the WeylHeisenberg group GWH . It turns out that the representation U has
the property of square integrability modulo the center Z of the group GWH , which
is isomorphic to the real line R the matrix elements of U are square integrable
when the integration is performed with respect to the measure on the coset space
GWH /Z R2 . Furthermore, the physical states (or rays) associated with the CS are
not indexed by elements of GWH itself, but by points in the coset space X = GWH /Z.
All of these aspects will lead us to far-reaching generalizations. In addition, the
canonical CS have remarkable functional analytic and meromorphic properties. For
example, using them, one can derive the FockBargmann representation of quantum
mechanics on Hilbert spaces of analytic functions.
10 1 Introduction

The stage being thus set, we begin our study by four mostly mathematical
chapters (Chaps. 36). As we said above, the resolution of the operator (1.11) is
the central object of the theory. Mathematically, it is equivalent to the existence of
a reproducing kernel K(x, y) = x |A1 y  (for the moment, we assume A1 to be
bounded), which means that the Hilbert space H is realized as a space of functions
f (x) := x | f  and one has

f (x) = K(x, y) f (y) d (y), f H. (1.14)
X

Thus, the basic framework of our theory is a reproducing kernel Hilbert space.
Hilbert spaces of this type have many attractive properties, and reproducing kernels
are well-studied objects in mathematics [99,Mes62], which may be defined on quite
general measure spaces. Besides, they arise naturally in large classes of complex
domains [145], leading to preferred families of vectors in the Hilbert space that
generate resolutions of the identity.
As a first step (Chap. 3), we review a number of mathematical topics needed
in the sequel, centered on the notion of positive operator-valued (POV) measures
and frames. Then, in Chap. 4, we give a quick reminder of a number of concepts
and results from group theory, such as (quasi-) invariant measures on homogeneous
spaces, induced group representations, localization properties, harmonic analysis on
abelian groups, Lie groups and Lie algebras, their extensions and contractions, and
so on. In Chap. 5, we investigate systematically reproducing kernel Hilbert spaces.
Particular attention is given, in Chap. 6, to the two classes of square integrable and
holomorphic reproducing kernels. In Sect. 6.4, in particular, we discuss in detail the
GazeauKlauder CS.
Thus, equipped with the necessary mathematical machinery, we are in a position
to state our definition of a coherent state (Chap. 7). It is general enough to cover
such widely different situations as GilmorePerelomov CS, abstract CS, or discrete
frames in Hilbert space. In the same chapter, which occupies a pivotal position
in the book, we also describe briefly what may be called the classical theory of
coherent states, largely due to Gilmore and Perelomov. It basically deals with the
CS associated with representations of simple groups, as examplified by SU(2) and
SU(1,1) in the compact and noncompact cases, respectively. Our treatment is rather
short here, except for a typical physical application in each case, and we refer to
the monograph of Perelomov [502] for further information. Instead, we discuss a
number of more or less recent developments: several generalizations of frames; CS
on spheres obtained via heat kernels; application of CS in quantum gravity (here we
give only a brief sketch, for lack of space); and CS on conformal classical domains.
All these topics are new to the present edition.
From here on, we concentrate on CS associated with group representations.
Although the theory described so far is more general, this is the class that is readily
applicable in specific physical situations. Chapter 8 covers in detail the class of
CS most studied in the literature, namely, the CS associated with square integrable
group representations. Given a locally compact group or a Lie group G, and a square
1 Introduction 11

integrable representation U of G (if one such exists) in some Hilbert space H, fix a
vector H (possibly satisfying an additional admissibility condition). Then, a CS
system, associated to U, is defined to be the following set of vectors in H:

S := {g = U(g) : g G}. (1.15)

In other words, coherent states are the elements of the orbit of under the (square
integrable) representation U(G). They possess a number of remarkable properties,
in particular, their matrix elements satisfy very useful orthogonality relations,
allowing one to derive a generalized Wigner map (briefly, this amounts to lifting the
analysis from the Hilbert space H to the Hilbert space B2 (H) of all HilbertSchmidt
operators on H). For this topic and its application in statistical mechanics, we refer
to the first edition of this work. While this is the basic construction, the approach of
Gilmore [330,331] and Perelomov [502] introduces an additional degree of freedom
in that it indexes the CS by points, not of G itself, but of X = G/H , where H is the
subgroup of elements of G leaving invariant up to a phase. The resulting CS enjoy
all of the desirable properties of the canonical CS. Actually, a slight generalization
of the notion of admissibility permits one to cover more general situations, such as
the Galilei group or the so-called vector CS.
The following chapters are basically devoted to applications of the preceding
theory and its generalizations to various groups of physical interest. In Chaps. 9
and 10, we turn to a class of great physical interest, namely Lie groups of the
semidirect product type, G = V  S, where V is a vector space (typically, Rn ) and
S is a group of automorphisms of V . It is possible to give a unified treatment of all
such groups, since their representations may be constructed by Mackeys method
of induced representations [Mac68, Mac76]. In addition, one can decide when such
a representation is square integrable, thanks to work by KleppnerLipsman [419],
Aniello et al. [41,42], BernierTaylor [146], Fhr [292], and Fhr and Mayer [293].
One may also refer to the monograph of Fhr [Fue05]. The physical interest of
this class is that it covers the so-called relativity groups, that is, the kinematical
groups associated to the Galilean, Newtonian, and Einsteinian relativities [108],
which are fully treated in Chap. 10. The problem is that these groups have no
square integrable representations in the usual sense; hence the preceding theory
is not applicable as it stands. There exist, however, families of vectors in certain
representation spaces of these groups, which strongly resemble coherent states, and
appear in an essential way in localization and quantization problems, but are not
obtainable by the GilmorePerelomov method [2225, 511, 512]. A possible way to
incorporate all of these families of states within a unified framework of generalized
coherent states was suggested in [13] and later developed in a series of papers
[15, 17, 2630]. The point of departure here is the observation that, while, in the
GilmorePerelomov framework, CS are indexed by points of X = G/H , in many
instances (particularly for the relativity groups), more general classes of CS may
be constructed using a suitable homogeneous space X = G/H, where H is a closed
subgroup of G that does not necessarily coincide with H . Then, one needs to find a
section : X G, and one ends with the general definition of CS given in Chap. 7;
12 1 Introduction

that is, the vectors (x) = U( (x)) , x X. As a prototype of the general theory,
we treat in detail the Poincar group, probably the most important group of quantum
physics. Comparable information about the other relativity groups may be found in
the literature, in particular, our earlier papers [15, 17, 2630]. We also discuss the
relationship between the various cases, the precise link being the notion of group
contraction and its extension to group representations.
A word of warning may be in order here. Whereas Chap. 10 is fairly concrete,
with explicit calculations, Chap. 9 is rather abstract, with a strong geometrical flavor.
We feel indeed that a detailed and mathematically precise treatment of the general
situation is worth giving, since it would be difficult to find it in the literature in a
form directly applicable to the relativity groups of Chap. 10.
Chapter 11, which is totally new, is a general reflection on the notion of
quantization, that encompasses those of WeylHeisenberg quantization and of CS
quantization. The emphasis here is on the probabilistic aspects. Since most formulas
are based on integrals, we have coined the term integral quantization for our general
formalism.
Chapters 1216 present a survey of wavelets, first in one dimension, then in
higher dimensions (mostly two, for the purposes of image analysis). As said before,
wavelets are the CS associated with the affine groups: the affine group or the ax + b
group of the line and the similitude group of Rn for n > 1, consisting of translations,
dilations, and rotations. Note that, in both cases, the group is a semidirect product:

G+ = R  R+
, SIM(n) = Rn  (R+
SO(n)),

where Rn corresponds to translations, R+ corresponds to (positive) dilations and

SO(n) corresponds to rotations. While stressing the group-theoretical origin of
the (continuous) wavelet transform (WT), we also describe in Chap. 12 various
concrete aspects, such as the explicit construction of wavelets, the explanation of
the efficiency of the WT in signal and image analysis, a brief survey of physical
and technological applications, and so on. In addition, we discuss a number
of mathematical properties, such as the minimal uncertainty property (i.e., the
generalization of P1 above), or the extension of the WT to distributions (largely
following [378]).
We also give, in Chap. 13, a brief account of the so-called discrete WT, based on
the notion of multiresolution analysis, which is based on a totally different philos-
ophy and more widely known in applications. A natural generalization consists of
wavelet analysis on abstract abelian groups, first the finite fields Z p , then general
ones. In this way, one is approaching the goal of giving a general, group-related
formulation of the discrete wavelet transform and the multiresolution formalism.
Although this development is quite abstract and mathematically demanding, we
felt that it was worth discussing it in detail, because it opens the door to far
reaching generalizationsbesides having in itself a great mathematical interest.
A slightly different development, which we also touch briefly in this chapter, is
the construction of wavelets based on irrational numbers, such as the golden mean
and, more generally, on the so-called Pisot numbers.
1 Introduction 13

Next we survey in Chap. 14 the case of multidimensional wavelets and a number

of generalizations, all issued from signal/image analysis, such as ridgelets, curvelets
and shearlets. Then we turn to generalized wavelets or CS associated with other
groups, which are of two kinds. On one hand, wavelets on manifolds, such as
spheres, surfaces of revolution, or discrete graphs, are discussed in Chap. 15.
On the other hand, CS associated with other affine groups, obtained by adjoining
dilations to the corresponding kinematical groups (Galilei, Poincar) or to the Weyl
Heisenberg group, are described in Chap. 16.
Finally, Chap. 17 touches on the crucial aspect of the discretization of the
wavelet or the CS transform. The application of any kind of CS to a concrete
problem requires numerical computations, and this means that all formulas have to
be discretized: Signals and images are discretely sampled, integrals are replaced
by sums, and the WT is thus also a digital signal. This transition raises deep
mathematical questions, some of which we discuss as well as briefly describing
the results obtained so far in the literature.
In retrospect, we might say that Chaps. 1216 almost constitute a book within
the book. The development of wavelet theory has been so spectacular in the past
few years that we were compelled to cover a huge number of different issues and
new results. The outcome illustrates the richness and versatility of the CS concept.
Indeed, wavelets are after all only CS of a special type, namely, those for which a
scale variable plays a central rle. This ultimately justifies the central point of view
of this book, namely, that CS provide the unifying notion.
A final word of warning before proceeding. This book is about mathematical
physics; hence, its audience consists of both mathematicians and physicists. Of
course, it is impossible to please everyone at the same time. So, we have on some
occasions dwelled on mathematics longer than expected, following the internal
dynamics of the topic at hand. Conversely, we have, at times, indulged in a thorough
discussion of certain physical examples or applications. We apologize to both
communities for these various wanderings, our sole excuse being our desire to
illustrate the general theory and make the text livelier.
This in fact characterizes the spirit of the book. We have conceived it more as a
research monograph, designed for self-study, rather than a textbook suitable for the
classroom. We have therefore not included any standard end-of-chapter exercises.
Instead, one will find dispersed throughout the book a large number of worked out
examples, which could be thought of as being research level problems.
Chapter 2
Canonical Coherent States

Abstract This chapter is devoted to a detailed examination of the canonical

coherent states (CS), originally introduced by Schrdinger in 1926 and rediscovered
in the 1960s for the description of coherent light (lasers). We discuss successively
the minimal uncertainty problem, the group-theoretical background of CS, their
functional analytic properties and the geometrical context, both in the real and in
the complex formulation. We conclude with some simple examples.

This chapter is devoted to a fairly detailed examination of the quintessential example

of coherent statesthe canonical coherent states. It is fair to say that the entire
subject of coherent states developed by analogy from this example. As mentioned
in the Introduction, this set of states, or rays in the Hilbert space of a quantum
mechanical system, was originally discovered by Schrdinger [552] in 1926, as
a convenient set of quantum states for studying the transition from quantum to
classical mechanics. They are endowed with a remarkable array of interesting
properties, some of which we shall survey in this chapter. Apart from initiating the
discussion, this will also help us in motivating the various mathematical directions
in which one can try to generalize the notion of a CS.
Throughout the book, all Hilbert spaces will be taken to be separable and defined
over the complexes. Thus, if H is a Hilbert space, its dimension, denoted dim H, is
(countably) infinite. Also, for any two vectors , in H, their scalar product  | 
will be taken to be antilinear in the first variable and linear in the second variable
(the standard physicists convention). Unless otherwise stated, we shall use the
natural system of units, in which c = h = 1.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 15
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__2,
Springer Science+Business Media New York 2014
16 2 Canonical Coherent States

2.1 Minimal Uncertainty States

Recall that the quantum kinematics of a free n-particle system is based upon the
existence of an irreducible representation of the canonical commutation relations
(CCR),

[Qi , Pj ] = iI i j , i, j = 1, 2, . . . , n, (2.1)

on a Hilbert space H. (Here I denotes the identity operator on H). If n is finite, then
according to the well-known uniqueness theorem of von Neumann [vNe55], up to
unitary equivalence, there exists only one irreducible representation of (2.1) by self-
adjoint operators, on a (separable, complex) Hilbert space (see Sect. 2.2 for a more
precise statement). Furthermore, the CCR imply that for any state vector in H
(note,   = 1), the Heisenberg uncertainty relations hold:

1
 Qi   Pi  , i = 1, 2, . . . , n, (2.2)
2
where, for an arbitrary operator A on H,
1
 A = [ |A2  | A |2 ] 2 (2.3)

is its standard deviation in the state (to be precise, the vector must belong to
the domain of the operator A2 ). As already pointed out by Schrdinger (see also
[vNe55, Bie81]), there exists an entire family of states, s in the Hilbert space,
labelled by a vector parameter s = (s1 , s2 , . . . , sn ) Rn , each one of which saturates
the uncertainty relations (2.2):

1
 Qi  s  Pi  s = , i = 1, 2, . . . , n. (2.4)
2
We call these vectors minimal uncertainty states (MUSTs). In the configuration
space, or Schrdinger representation of the CCR, in which

H = L2 (Rn , dx), x = (x1 , x2 , . . . , xn ),

(Qi )(x) = xi (x), (Pi )(x) = i (x), (2.5)
xi

the MUSTs, s , are just the Gaussian wave packets

n
xi2
s (x) = ( s2i ) 4 exp[
1
]. (2.6)
i=1 2s2i
2.1 Minimal Uncertainty States 17

Not surprisingly, quantum systems in these states display behaviour very close to
classical systems. More generally, there exists a larger family of states, namely gaus-
sons or gaussian pure states [557], which exhibits the minimal uncertainty property.
These latter states q,p U,V
are parametrized by two vectors, q = (q1 , q2 , . . . , qn ), p =
(p1 , p2 , . . . , pn ) Rn and two real n n matrices U and V , of which U is positive
definite. In the Schrdinger representation,
n 1 q
q,p
U,V
(x) = 4 [detU] 4 exp [i(x ) p]
2
1
exp [ (x q) (U + iV )(x q)]. (2.7)
2

In the optical literature, states of the type q,p

U,V
, for which U is a diagonal matrix
but not the identity matrix, are called squeezed states [150, 598]. Note that when
q = p = 0 and U is diagonal, with eigenvalues 1/s2i , i = 1, 2, . . . , n, the gaussons
(2.7) are exactly the MUSTs (2.6). Moreover, if T denotes the orthogonal matrix
which diagonalizes U, i.e., TUT 1 = D, where D is the matrix of eigenvalues of U,
then defining the vectors x = T x, q = T q, p = T p, and the matrix V  = TV T 1 ,
we may rewrite (2.7) as


 n 1  q
qD,V
 ,p (x ) = 4 [det D] 4 exp [i(x ) p ]
2
1
exp [ (x q ) (D + iV  )(x q )]. (2.8)
2

It is clear from this relation that, if Qi , Pi , i = 1, 2, . . . , n, are the components of the

rotated vector operators, Q = T 1 Q, P = T 1 P, where Q = (Q1 , Q2 , . . . , Qn ), P =
(P1 , P2 , . . . , Pn ) are the vector operators of position and momentum, respectively,
[see (2.5)] then

1
 Qi  U,V  Pi  U,V = , i = 1, 2, . . . , n. (2.9)
q,p q,p 2
To examine some properties of the MUSTs (2.6), take n = 1, for simplicity (of
notation), and define the creation and annihilation operators,1

1 1
a = (s1 Q isP), a = (s1 Q + isP),
2 2
[a, a ] = 1. (2.10)

1 For historical reasons, we use here the physicists notation a for the creation operator, but in the
rest of the book, we denote by A the adjoint of the operator A.
18 2 Canonical Coherent States

Using these operators and the MUST s , for a fixed s R, we can generate a
very interesting class of other MUSTs (which of course is a subclass of (2.7), and
which was already noticed by von Neumann [vNe55]). To do so, define the complex
variable
1
z = x + iy = (s1 q + isp), (q, p) R2 (2.11)
2

and write

s = |0. (2.12)

(Note that a|0 = 0). Also let {|n}

n=0 be the normalized eigenstates of the number
operator N = a a:
n
(a )
N|n = n|n, |n = |0, m|n = mn . (2.13)
n!

Then the set of states in H,

|z|2
|z = exp ( + za ) |0
2

1 zn
= exp ( |z|2 ) |n, (2.14)
2 n=0 n!

for all z C, have the eigenvalue property

a|z = z |z. (2.15)

It is straightforward to verify that each one of these states |z is again a MUST
satisfying (2.4).
Suppose now that we have a quantized electromagnetic field (in a box), and let
ak , ak , k = 0, 1, 2, . . ., be the creation and annihilation operators for the various
Fourier modes k. Then in the states

|{zk } = |zk ,
k

the electromagnetic field behaves classically. More precisely [333, 334], the
correlation functions for the field factorize in these states. Thus, let x = (x,t) be
a space-time point and E+ (x) the positive frequency part of the quantized electric
field (Note: E (x) = E+ (x) is the negative frequency part of the field). Then,

E+ (x) |{zk } = E (x) |{zk },

2.1 Minimal Uncertainty States 19

where E is a 3-vector valued function of x, giving the observed field strength at the
point x. Let be the density matrix,

= |{zk }{zk }|,

(n)
and G1 ,2 ,...,2n the correlation functions,

(n)
G1 ,2 ,...,2n (x1 , x2 , . . . , x2n ) =
= Tr[ E1 (x1 ) . . . En (xn )E+n+1 (xn+1 ) . . . E+2n (x2n )], (2.16)

where Ek denotes the k -th component of E . It is then easily verified that

n 2n
G1 ,2 ,...,2n (x1 , x2 , . . . , x2n ) = E k (xk )
(n)
E (x ). (2.17)
k=1 =n+1

It is because of this factorizability property that the states |{zk } or the MUSTs |z
were called coherent states. However, in the current mathematical literature (though
not always in the optical literature), the term coherent state is used to designate an
entire array of other mathematically related states, which do not necessarily display
either the factorizability property (2.17) or the minimal uncertainty property (2.4).
We shall reserve the term canonical coherent states for the MUSTs (2.14).
In order to bring out some additional properties of the canonical CS |z, let us
use (2.11) to write

|z = s (q,p) , (2.18)

where z and q, p are related by (2.11). The significance of the in this notation will
become clear in a while. A short computation shows that

s (q,p) |Q|s (q,p)  = q,

s (q,p) |P|s (q,p)  = p. (2.19)

In other words, the MUST s (q,p) is a translated Gaussian wave packet, centered
at the point q in position and p in momentum space. Explicitly, as a vector in
L2 (R, dx),

(xq)2
ipx 2s2
qp
s (q,p) (x) = ( s2 )1/4 ei 2 e e , (2.20)

which should be compared to (2.7).

20 2 Canonical Coherent States

2.2 The Group Theoretical Backdrop

A group theoretical property of |z emerges if we use the BakerCampbell

Hausdorff identity,
1
eA+B = e 2 [A,B] eA eB , (2.21)

for two operators A, B, the commutator [A, B] of which commutes with both A and
B, and the fact that an |0 = 0, n 1, to write |z in (2.14) as

1
|z = exp( |z|2 ) eza eza |0 = eza za |0,

(2.22)
2

where z denotes the complex conjugate of z. In terms of (q, p) this is

s (q,p) = ei(pQqP) s U(q, p) s , (2.23)

where, (q, p) R2 , the operators U(q, p) = ei(pQqP) are, of course, unitary.

Moreover, we have the integral relation,

1
|s (q,p) s (q,p) | dq dp = I. (2.24)
2 R2

The convergence of the above integral is in the weak sense (see Sect. 3.1), i.e., for
any two vectors , in the Hilbert space H,

1
 |s (q,p) s (q,p) |  dq dp =  | . (2.25)
2 R2

To check the validity of this relation, we use (2.20) to obtain


1 ipq (x q)2
 |s (q,p)  = ( s2 ) 4 e 2 (x) exp [ipx] exp [ ] dx. (2.26)
R 2s2

Hence the left-hand side of (2.25) becomes

 
1
dq dp dx dx (x) exp [ip(x x )]
2 s R2 R2

(x q)2 (x q)2
exp [ ] (x ). (2.27)
2s2 2s2
Using the representation

1 
eip(xx ) dp = (x x ), (2.28)
2 R
2.2 The Group Theoretical Backdrop 21

for the -distribution, and performing the integration over x , the above integral
becomes

1 (x q)2
(x) exp [ ] (x) dq dx
s R2 s2
 
1
= exp [q2 ] dq (x) (x) dx
R R

=  | . (2.29)

The relation (2.24) is called the resolution of the identity generated by the
canonical CS.
The operators U(q, p) in (2.23) arise from a unitary, irreducible representation
(UIR) of the WeylHeisenberg group, GWH , which is a central extension of the group
of translations of the two-dimensional Euclidean plane. The UIR in question is the
unitary representation of GWH which integrates the CCR (2.1). An arbitrary element
g of GWH is of the form

g = ( , q, p), R, (q, p) R2 ,

with multiplication law,

g1 g2 = (1 + 2 + ((q1 , p1 ); (q2 , p2 )), q1 + q2 , p1 + p2 ), (2.30)

where is the multiplier function

1
((q1 , p1 ); (q2 , p2 )) = (p1 q2 p2 q1 ). (2.31)
2
Any infinite-dimensional UIR, U , of GWH is characterized by a real number =
0 (in addition, there are also degenerate, one-dimensional, UIRs corresponding to
= 0, but they are irrelevant here [Per86]) and may be realized on the same Hilbert
space H, as the one carrying an irreducible representation of the CCR:
pq
U ( , q, p) = ei U (q, p) := ei ( 2 ) ei pQ ei qP . (2.32)

If H = L2 (R, dx), these operators are defined by the action

q
(U ( , q, p) )(x) = ei ei p(x 2 ) (x q), L2 (R, dx). (2.33)

Thus, the three operators, I, Q, P, appear now as the infinitesimal generators of this
representation and are realized as:

i i
(Q )(x) = x (x), (P )(x) = (x), [Q, P] = I. (2.34)
x

For our purposes, we take for the specific value, = 1

h = 1, and simply write U
for the corresponding representation.
22 2 Canonical Coherent States

Denoting the phase subgroup of GWH (the subgroup of elements g = ( , 0, 0),

R), by , it is easily seen that the left coset space GWH / can be identified
with R2 and a general element in it parametrized by (q, p). In terms of this
parametrization, GWH / carries the invariant measure

dq dp
d (q, p) = . (2.35)
2
The function

: GWH / GWH , (q, p) = (0, q, p), (2.36)

then defines a section in the group GWH , now viewed as a fibre bundle, over the base
space GWH /, having fibres isomorphic to . Thus, the family of canonical CS is
the set,

S = {s (q,p) = U( (q, p)) s : (q, p) GWH /}, (2.37)

and the operator integral in (2.24) becomes


|s (q,p) s (q,p) | d (q, p) = I. (2.38)
GWH /

In other words, the CS s (q,p) are labelled by the points (q, p) in the homogeneous
space GWH / of the WeylHeisenberg group, and they are obtained by the action
of the unitary operators U( (q, p)), of a UIR of GWH , on a fixed vector s H. The
resolution of the identity (2.38) is then a statement of the square-integrability of the
UIR, U, with respect to the homogeneous space GWH /. This way of looking at
coherent states turns out to be extremely fruitful. Indeed, one could ask if it might
not be possible to use this idea to generalize the notion of a CS and to build families
of such states, using UIRs of groups other than the WeylHeisenberg group, making
sure in the process that basic relations of the type (2.36)(2.38) are still fulfilled.
We shall see in Chaps. 7 and 8 that this is indeed possible, and that such an approach
yields a powerful generalization of the notion of a coherent state.
Two remarks are in order before proceeding. First and not surprisingly, the same
canonical CS may be obtained from the oscillator group H(4), which is the group
with the Lie algebra generated by {a, a , N = a a, I}. Secondly, it is interesting that
the canonical CS are widely used in signal processing, where they generate the
so-called windowed Fourier transform or Gabor transform, since it was introduced
in that context by Dennis Gabor in his 1946 landmark paper on communication
theory [294]. This is a hint that CS will have an important role in classical physics
as well as in quantum physics, and as a matter of fact they may be viewed as a
natural bridge between the two. We shall discuss these matters in detail in Chaps. 11
and 12.
2.3 Some Functional Analytic Properties 23

2.3 Some Functional Analytic Properties

The resolution of the identity given by the operator integral in (2.38) leads to some
interesting functional analytic properties of the CS, s (q,p) . These properties can be
studied in their abstract forms and be used to obtain a generalization of the notion
of a CS, but now independently of any group theoretical implications.
Let H = L2 (GWH /, d ) be the Hilbert space of all complex valued functions
on GWH / which are square integrable with respect to d . Then (2.38) implies that
functions : GWH / C of the type

(q, p) = s (q,p) | , (2.39)

 and moreover, writing W : H H

for H, define elements in H,  for the linear
map which associates, via (2.39), an element in H to an element in H  (i.e.,
W = ), we see that W is an isometry or a norm-preserving linear map:

W 2 =  2 = | (q, p)|2 d (q, p) =  2 . (2.40)
GWH /

The range of this isometry, which we denote by HK ,


HK = W H H, (2.41)
 and furthermore, it is a reproducing kernel Hilbert space.
is a closed subspace of H
To understand the meaning of this, consider the function K(q, p; q , p ) defined on
GWH / GWH /:

K(q, p; q , p ) = s (q,p) |s (q ,p ) 

i s2 1
= exp [ (qp q p)] exp [ (p p )2 ] exp [ 2 (q q )2 ]
2 4 4s
1 1
= exp [zz |z|2 |z |2 ]
2 2
= z|z  = K(z, z ), (2.42)

the third and fourth equalities following from (2.18) and (2.20). The function K is
a reproducing kernel, a name that reflects the reproducing property satisfied by any
vector HK :

(q, p) = K(q, p; q , p ) (q , p ) d (q , p ). (2.43)
GWH /

The function K enjoys the properties:

1. Hermiticity,

K(q, p; q , p ) = K(q , p ; q, p). (2.44)

24 2 Canonical Coherent States

2. Positivity,

K(q, p; q, p) > 0. (2.45)

3. Idempotence,

K(q, p; q , p )K(q , p ; q , p ) d (q , p ) = K(q, p; q , p ). (2.46)
GWH /

The above relations hold for all (q, p), (q , p ) GWH /. Condition (2.46) is a
consequence of (2.38) and is also called the square integrability property of K.
All three relations are the transcription of the fact that the orthogonal projection
operator PK of H onto HK is an integral operator, with kernel K(q, p; q , p ).
It is easy to see that the kernel K actually determines the Hilbert space HK .
Comparing (2.39) and (2.42) we see that

(W s (q ,p ) )(q, p) = K(q, p; q , p ); (2.47)

in other words, for fixed (q , p ), the function (q, p)  K(q, p; q , p ) is simply the
image in HK of the CS s (q ,p ) under the isometry (2.39). Additionally, if is
an element of the Hilbert space HK , it is necessarily of the form (2.39). Hence,
multiplying both sides of that equation by s (q,p) and integrating, we get, upon
using (2.38),

= (q, p)s (q,p) d (q, p). (2.48)
GWH /

This shows that the set of vectors s (q,p) , (q, p) GWH /, is overcomplete in H
and hence, since W is an isometry, the set of vectors

(q,p) = W s (q,p) , (q,p) (q , p ) = K(q , p ; q, p), (2.49)

(for all (q, p) GWH /), is overcomplete in HK . (Note that the vectors (q,p)
are the same CS as the s (q,p) , but now written as vectors in the Hilbert space of
functions HK ). The term overcompleteness is to be understood in the following way:
Since HK is a separable Hilbert space, it is always possible to choose a countable
basis {i } i=1 in it, and to express any vector H as a linear combination of the
vectors in this basis. By contrast, the family of CS, S in (2.37) is labelled by a pair
of continuous parameters (q, p), and (2.48), or equivalently (2.38) is the statement
of the fact that any vector can be expressed in terms of the vectors in this family.
Clearly, it should be possible to choose a countable set of vectors {s (qi ,pi ) }i=1
from S and still obtain a basis for H. This is in fact possible and many different
discretizations exist. The most familiar situation is that where the set of points
{qi , pi } is a lattice, such that the area of the unit cell is smaller than a critical
value (to be sure, the resulting set of CS is then overcomplete). The determination
of adequate subsets {qi , pi } leads to very interesting mathematical problems, for
2.3 Some Functional Analytic Properties 25

instance in number theory and in the theory of analytic functions [Per86, Sect. 1.4].
These considerations are part of the general problem of CS discretization, that we
shall tackle in the last chapter of the book (Chap. 17).
Equation (2.39) also implies a boundedness property for the functions in the
reproducing kernel Hilbert space HK . Indeed, using the unitarity of U( (q, p)),

| (q, p)|    , (q, p) GWH /, (2.50)

implying that the vectors in HK are all bounded functions. More importantly, this
also shows that the linear map

EK (q, p) : HK C, EK (q, p) = (q, p), (2.51)

which simply evaluates each function HK at the point (q, p), and hence called
an evaluation map, is continuous. As we shall see later, this can in fact be taken to
be the defining property of a reproducing kernel Hilbert space and used to arrive at
coherent states via a relation of the type (2.47).
The CS s (q,p) , along with the resolution of the identity relation (2.38), can be
used to obtain a useful family of localization operators on the phase space :=
GWH /. Indeed, relations such as (2.19) tend to indicate that the CS s (q,p) do in
some sense describe the localization properties of the quantum system in the phase
space . To pursue this point a little further, denote by an arbitrary Borel set in
, considered as a measure space, and let B( ) denote the -algebra of all Borel
sets of . Define the positive, bounded operator

a( ) = |s (q,p) s (q,p) | d (q, p). (2.52)

This family of operators, as runs through B( ), enjoys certain measure theoreti-

cal properties:
1. If J is a countable index set and i , i J, are mutually disjoint elements of B( ),
i.e., i j = 0,
/ for i = j (0/ denoting the empty set), then

a(iJ i ) = a(i ), (2.53)

iJ

the sum being understood to converge weakly.

2. Normalization:
a( ) = I, / = 0.
also a(0) (2.54)

Such a family of operators a( ) is said to constitute a normalized, positive

operator-valued (POV) measure on H. Using the isometry W in (2.39) and the CS
(q,p) in (2.49), we obtain the normalized POV-measure aK ( ) on HK :

aK ( ) = | (q,p)  (q,p) | d (q, p)

= Wa( )W . (2.55)
26 2 Canonical Coherent States

Note that

aK ( ) = | (q,p)  (q,p) | d (q, p)

= PK , (2.56)

 which projects onto the reproducing

where PK is the projection operator, HK = PK H,
kernel subspace HK of H.
If HK is an arbitrary state vector, and = W , H, then by (2.55) and
(2.39)

 |aK ( )  =  |a( )  = | (q, p)|2 d (q, p). (2.57)

This means that if (q, p) is considered as being the phase space wave function
of the system, then aK ( ) is the operator of localization in the region of phase
space. Of course, in order to interpret | (q, p)|2 as a phase space probability density,
an appropriate concept of joint measurement of position and momentum has to be
developed. This can in fact be done (see, for example, [13, Bus91, Pru86]). Here, let
us just indicate, without proof, an interesting fact which reinforces the interpretation
of the aK ( ) as localization operators. On HK define the two unbounded operators
QK and PK

 |QK  = (q, p) q (q, p) d (q, p),


 |PK  = (q, p) p (q, p) d (q, p), (2.58)

on vectors , chosen from appropriate dense sets in HK . Then it can be shown

(see, e.g., [13]) that

[QK , PK ] = iIK , IK = identity operator on HK . (2.59)

Thus, multiplication by q and p on , followed by the projection PK , yields the

position and momentum operators on HK .
Mathematically, the virtue of this functional analytic description of the coherent
states s (q,p) is that it points up another possibility of generalization: We would
like to associate CS to arbitrary reproducing kernel Hilbert spaces. This will be
dealt with in Chap. 5, and it will also take us into the theory of frames, in Chaps. 7
and 17.
2.4 A Complex Analytic Viewpoint 27

2.4 A Complex Analytic Viewpoint

To bring out some complex analytic properties of the canonical CS, let H be an
arbitrary vector. Computing its scalar product with the CS |z using (2.14), we get

1 n| 
z|  = exp ( |z|2 ) zn
2 n=0 n!
1
= exp ( |z|2 ) f (z). (2.60)
2
Here f is an analytic function of the complex variable z. Note that in (2.60) we
took the scalar product of |  with |z and not with |z, in order to come out with
an analytic function f (z), rather than an anti-analytic function. This was made
necessary by our convention that scalar products are linear in the right hand and
antilinear in the left hand term. In terms of z, z we may write the measure (2.35) as

dz dz
d (q, p) = , (2.61)
2 i
and let us define the new measure
dz dz
d (z, z) = exp (|z|2 ) . (2.62)
2 i
In this notation, denotes the exterior product of the two differentials dz and dz
(considered as one-forms on the complex manifold C). In measure theoretic terms,
the quantity i dz dz/2 simply represents the Lebesgue measure dxdy, z = x + iy,
on C. Comparing (2.60) and (2.61) with (2.39) and (2.40), we see that HK can be
identified with the Hilbert space of all analytic functions in z which are square-
integrable with respect to d . Let Hhol denote this Hilbert space. Then, the linear
map

1
Whol : H Hhol , (Whol )(z) = exp ( |z|2 ) z| , (2.63)
2
is an isometry. Using (2.18) and (2.20) we can compute the vectors

1
f (q,p) = Whol s (q,p) = Whol |z = exp ( |z|2 ) z , (2.64)
2

which are the images of the s (q,p) in Hhol . The vectors z Hhol represent the
analytic functions [see (2.42)]:

 1
z (z ) = ez z = exp [ (|z|2 + |z |2 )] K(z , z). (2.65)
2
28 2 Canonical Coherent States

From this it is clear that the function Khol : C C C,


Khol (z , z) = z |z Hhol = ez z , (2.66)

is a reproducing kernel for Hhol . (Here |Hhol denotes the scalar product in Hhol ).
Indeed, for any f Hhol and z C,

Khol (z, z ) f (z ) d (z , z ) = f (z) = z | f Hhol . (2.67)
C

The vectors z satisfy the resolution of the identity relation on Hhol :


|z z | d (z, z) = IHhol . (2.68)
C

The MUST, s = |0, is represented in Hhol as the constant vector

Whol s = u0 = Khol ( , 0), u0 (z) = 1, z C. (2.69)

Since Khol (z, z) = z 2 , (2.67) implies that

1
| f (z)| [Khol (z, z)] 2  f . (2.70)

Hence, in agreement with (2.51), the evaluation map

Ehol (z) : Hhol C, Ehol (z) f = f (z), (2.71)

is continuous. Actually, the CS z could have been obtained by using this fact alone,
i.e., by defining it to be the vector which for arbitrary f Hhol gives f (z) = z | f .
Such a construction would be independent of any group theoretical considerations,
and is intrinsic to complex manifolds admitting Khler structures (see Sect. 2.6
below).
The representation exp (|z|2 /2) z of the canonical CS, with the z being vectors
in the space of holomorphic functions Hhol is known among physicists as the
FockBargmann representation, and the Hilbert space Hhol as the Bargmann or
BargmannSegal space, often denoted by F B [123,124]. As noted earlier, it consists
of entire analytic functions which are square integrable with respect to d (z, z),

f (z) = f n zn , the sum converging absolutely for all z C
n=0

and

 f 2 := | f (z)|2 d (z, z) < . (2.72)
C
2.4 A Complex Analytic Viewpoint 29

The inner product is


 f | g =
C
f (z) g(z) d (z, z) = n! fn gn . (2.73)
n=0

Going back to the position representation, where the vectors |n read as x | n =
n (x), we see that the operator Whol is an integral operator,

f (z) = (Whol )(z) = K (z, x) (x) dx, H, (2.74)
R

with kernel
1

K (z, x) = 1/4 e 2 (z
2 +x2 )+ 2zx

zn
= n (x) n! (2.75)
n=0

In particular, the basis vectors |n H, [see (2.13)] are mapped by Whol to the vectors

zn
Whol |n = |un , un (z) = . (2.76)
n!

The inverse transformation is simply


(x) = K (z, x) f (z) d (z, z), f Hhol . (2.77)
C

In the FockBargmann representation, the operators a, a are given by

f
(a f )(z) = (z), (a f )(z) = z f (z), f Hhol . (2.78)
z

Equation (2.66) then implies:

Khol (z , z) = un (z )un (z). (2.79)
n=0

At this point, it is worthwhile to reinterpret the group-theoretical considerations

of Sect. 2.2 in the present complex analytic formulation. The starting point is the
unitary displacement operator

D(z) = e(za za) (z C),

D(z) = (D(z))1 = D(z) . (2.80)
30 2 Canonical Coherent States

introduced in Chap. 1, property P3 [see, also (2.22)]. Using the BakerCampbell

Hausdorff formula (2.21), we have
1 1
D(z) = eza eza e 2 |z| = eza eza e 2 |z| ,
2 2
(2.81)

from which follow the formulas

   
1 1
D(z) = a z D(z) = D(z) a + z ,
z 2 2
   
1 1
D(z) = a z D(z) = D(z) a + z .
z 2 2
These operators satisfy the addition formula:
1 
D(z)D(z ) = e 2 zz D(z + z ) ,

where z z is the symplectic product z z = zz zz = 2i Im zz . This in turn entails

the covariance formula on a global level:

D(z)D(z )D(z) = ezz D(z ),

and on a Lie algebra level

D(z) a D(z) = a z, D(z) a D(z) = a z.

The matrix elements of the operator D(z) involve the associated Laguerre
polynomials:

n! |z|2 /2 mn (mn) 2
m|D(z)|n = Dmn (z) = e z Ln (|z| ) , for m n , (2.82)
m!
(mn) (nm)
with Ln (t) = m!
n! (t)
nm L
m (t) for n m.
Coming back to the WeylHeisenberg group GWH = {( , z) , R, z C}, we
see that the group law (2.30) becomes

(1 , z1 )(2 , z2 ) = (1 + 2 + Im z1 z2 , z1 + z2 ) , ( , z)1 = ( , z) ,

so that the unitary representation (2.32) reads in Hhol as

( , z)  ei D(z) , (2.83)

(1 , z1 )(2 , z2 )  ei 1 D(z1 )ei 2 D(z2 ) = ei(1 +2 +Im z1 z2 ) D(z1 + z2 ) . (2.84)

To complete the picture we need to define also a discrete symmetry, the parity P,
acting on H as a linear operator through

P = ei a a .

P|n = (1)n |n , or (2.85)
2.5 An Alternative Representation and Squeezed States 31

It satisfies the following identities:

P2 = 1,
P a P = a ; P a P = a , (2.86)
PD(z)P = D(z) .

The displacement operators D(z) obey some interesting integral formulas. First, by
an explicit calculation using associated Laguerre polynomials, (2.82), one has,

dz dz
m | D(z)| n = mn 2(1)m , (2.87)
C i
which in turn implies

dz dz
D(z) = 2P. (2.88)
C i
Moreover, using the orthogonality of the associated Laguerre polynomials, one
obtains from (2.82) the ground state projector P0 := |00| as the Gaussian average
of D(z):

1 dz dz
e 2 |z| D(z)
2
= |00| . (2.89)
C i
More generally, for Re(s) < 1,

s dz dz 2 s+1
e 2 |z| D(z)
2
= exp(ln a a) , (2.90)
C i 1s s1
where the convergence holds in norm for Re(s) < 0 and weakly for 0 Re(s) < 1.
Finally, combining (2.88) with the third relation from (2.86), one gets a resolution
of the identity

dz dz
D(z) 2P D(z) = 1. (2.91)
C i

2.5 An Alternative Representation and Squeezed States

There is an interesting alternative representation of the canonical CS (2.14) in terms

of vectors in the Bargmann space. Indeed, denoting them again by |z, in terms of
the Bargmann space vectors z , we have |z = exp [|z|2 /2] |z . Thus we may write,
in view of our earlier discussion,

zn
|z = exp [|z|2 /2] un (z) |un  = exp [|z|2 /2] |un  . (2.92)
n=0 n=0 n!
32 2 Canonical Coherent States

A second representation can be obtained obtained using the squeezed basis,

which is the basis associated to the squeezed states of quantum optics (see, e.g.,
[577, 598, 610, 613]). The point of departure here is the expression (2.79) for
the reproducing kernel. As will be shown in Chap. 5, the functional form of the
reproducing kernel is independent of the orthonormal basis used to express it. Thus,
in (2.79) we may replace the basis {un }n=0 , also called the FockBargmann basis,
by any other orthonormal basis of Hhol . To this end we choose the squeezed basis

un = S( )un , n = 0, 1, 2, . . . , (2.93)

where S( ) is the (unitary) squeeze operator

S( ) = e K+ K , C, (2.94)

with

1 1 1 1
K+ = z2 , K = z2 , K0 = z z + (2.95)
2 2 2 2

being the well-known generators of the Lie algebra of the SU(1,1) group on Hhol .

The vectors un have the explicit forms

n 

1 2 41 2 2 z2 1 | |2  12
tanh(| |)
un (z) = (1 | | ) e Hn z , , =
n! 2 2 | |
(2.96)
where the Hn are the complex Hermite polynomials (i.e., the usual Hermite
polynomials, written in terms of the complex variable z),

[n/2]
(1)m (2z)n2m
Hn (z) = n! . (2.97)
m=0 m! (n 2m)!

They satisfy the orthogonality relations [39, 40, 215, 269, 318],
 
x2 y2 1 2 2 n
Hm (z)Hn (z) e2 [ 1+ + 1 ] dx dy = n! mn , z = x + iy, (2.98)
C 2

for any nonzero < 1. Going back to (2.79) we see that we may also write


Khol (z , z) = ez z = un (z )un (z), (2.99)
n=0

form which it follows that

2.5 An Alternative Representation and Squeezed States 33

|z|2
|z = exp [ ] un (z) un
2 n=0

 n  1
|z|2 (1 | |2 ) 4
1
2 z2 1 | |2 2
= exp [ ] e 2 Hn z un ,
2 n=0 n! 2 2

(2.100)

and this holds for any complex . This expression should be compared to (2.92).
More generally, since any two orthonormal bases of a Hilbert can be mapped
unitarily into one another, taking an arbitrary orthonormal basis {vn }
n=0 of Hhol we
may write

|z = exp [|z|2 /2] vn (z) vn . (2.101)
n=0

Finally, let us note that the squeezed states of quantum optics, which we already
encountered in Sect. 2.1, are themselves defined to be the vectors,
|z|2 zn
|z,  := S( )|z = exp [ ] un , (2.102)
2 n=0 n!
in the squeezed basis, so that alternatively, we may also write them as
|z|2
|z,  = exp [ ] un (z) un , (2.103)
2 n=0

in the FockBargmann basis.

The unitary squeeze operator S( ) in (2.94) can be written as

S( ) = |un un | , (2.104)
n=0

and thus on the Bargmann space of analytic functions Hhol it has the integral kernel,

S (z, w) = un (z)un (w)
n=0
 n
 1
| |
2 2 2
1 2 1 wn
= (1 | |2 ) 4 e 2 z Hn z .
1

n=0 n! 2 2 n!

After some manipulation this can be brought into the form,

S (z, w) = (1 | |2 ) 4 e 2 ( z
2 w2 )
e 1| | zw .
1 1 2
(2.105)
34 2 Canonical Coherent States

From (2.102) it follows that the function,

|w|2 |w|2

S (z, w) = (1 | |2 ) 4 e e 2 ( z
2 w2 )
e 1| | zw .
1 1
z  e
2
2 2 (2.106)

is just the Bargmann space function representing the squeezed coherent state vector
|w, . Furthermore, taking the limit 0 = | | 0, in the above equation,
yields the Bargmann space function for the canonical coherent state |w [see (2.99)],
as it should.

2.6 Some Geometrical Considerations

As already pointed out, the existence of the CS z can be traced back to certain
intrinsic geometrical properties of C, considered as a one-dimensional, complex
Khler manifold. While we do not intend, at this point, to discuss this notion in any
depth, it is still possible to get a general idea of what is involved. To begin with, C
may be thought of as being either a one-dimensional complex manifold or a two-
dimensional real manifold R2 , equipped with a complex structure. In the first case,
one works with the holomorphic coordinate z (or the antiholomorphic coordinate z).
In the second case, one uses the real coordinates q, p. Considered as a real manifold,
R2 is symplectic, i.e., it comes equipped with a closed, non-degenerate two-form
[compare with (2.35) and (2.61)]

1
= dq dp = dz dz, (2.107)
i

while considered as a complex manifold, C admits the Khler potential function:

(z , z) = z z, (2.108)

from which the two-form emerges upon differentiation:

1 2 (z, z)
= dz dz. (2.109)
i z z

Similarly, the Khler potential also determines the reproducing kernel:

Khol (z , z) = exp [ (z , z)], (2.110)

while the measure d [see (2.62)], defining the Hilbert space Hhol of holomorphic
functions, is given in terms of it by

dz dz
d (z, z) = exp [ (z, z)] . (2.111)
2 i
2.7 Outlook 35

Continuing, if we define the complex one-form

= iz (z, z) = iz dz, (2.112)

we get

= z , (2.113)

where z , z denote (exterior) differentiation with respect to z and z, respectively.

It appears therefore, that it is the Khler structure of C, (or the fact that it comes
equipped with the Khler potential ) which leads to the existence of the Hilbert
space Hhol of holomorphic functions and consequently, the CS z [the appearance
of these latter being a consequence of the continuity of the evaluation map (2.71)].
Once again, this situation is generic to all Khler manifolds.
Let P(z) be the one dimensional projection operator onto the vector subspace
of Hhol generated by the vector z , and denote this subspace by Hhol (z). The
collection of all these one-dimensional subspaces, as z ranges over C, defines a
(holomorphic) line bundle over the manifold C a structure which is intimately
related to the existence of a geometric prequantization of C. We hasten to add,
however, that while a complex Khler structure is in some sense ideally suited to
the existence of a geometric prequantization, a family of CS may define a geometric
prequantization even in the absence of such a structure and, in fact, it provides a
complete quantization, as we shall see in Chap. 11.

2.7 Outlook

We have quickly gleaned through a number of illustrative properties of the canonical

coherent states. As mentioned earlier, each one of these properties can be taken as
the starting point for a generalization of the notion of a CS. From a purely physical
point of view, for example, it is useful to look for generalizations which preserve the
minimal uncertainty property. In doing so, it is useful to exploit some of the group
theoretical properties as well. Mathematical generalizations could be based on group
theoretical, analytic or related geometrical properties. We shall attempt to describe
a bit of all of these various possibilities and along the way, we shall be naturally led
to some powerful applications of the theory of CS to wavelets and signal analysis.
But before digging further into the mathematics, we shall exhibit some examples
of physical applications illustrating the use of the canonical coherent states. For
instance, one could mention
1. Quantization theory: this is one of the most obvious applications of canonical
CS which goes back to the early days of quantum mechanics. The idea is
to establish a correspondence between classical observables, that is, real valued
functions on phase space, and quantum observables, represented by self-adjoint
36 2 Canonical Coherent States

operators on a Hilbert space, in such a way that Poisson brackets correspond

to commutators. Note, however, that such a goal is impossible to reach in most
cases. Details on this quantization technique may be found in earlier works [17,
33,135,136,488,511]. More important, we will devote to the topic of quantization
the full Chap. 11.
2. Atomic physics: as another illustration of the efficiency of CS methods, one
may consider the example of a system of N two-level atoms interacting with
a radiation field, for instance, the Dicke model [251, Scu97]. An extensive
description of such applications of CS to atomic physics may be found in the
textbook [Gaz09].
3. Quantum measurements and quantum information theory: we have here the most
recent application of canonical CS, in the domain which has witnessed in the last
years a spectacular renewal, thanks to the prospect of using the phenomenon of
entanglement for designing a quantum computer. For details we refer to the books
by Helstrom [Hel76], Holevo [H0l01] or Busch et al. [Bus91]. Another direction
to quote is the rle of CS in the description of the phenomenon of decoherence
[619, 620].
Chapter 3
Positive Operator-Valued Measures and Frames

Abstract Chapter 3 is a mathematical analysis of two related notions, namely,

positive operator-valued (POV) measures and frames, and some generalizations.
We start with continuous frames and introduce two generalizations, called upper,
resp. lower, semi-frames, for which only the upper, resp. lower, frame bound is
satisfied. Then we turn to the more familiar discrete frames and their generalizations.

This chapter, and the three succeeding it, constitute a mathematical interlude,
preparing the ground for the formal definition of a coherent state in Chap. 7 and
the subsequent development of the general theory. As should be clear already, from
a look at the last chapter, in order to define CS mathematically and to obtain a
synthetic overview of the different contexts in which they appear, it is necessary to
understand a bit about positive operator-valued (POV) measures on Hilbert spaces
and their close connection with certain types of group representations. In Chap. 2,
we have also encountered examples of reproducing kernels and reproducing kernel
Hilbert spaces, which in turn are intimately connected with the notion of POV
measures and hence coherent states. In this chapter, we gather together the relevant
mathematical concepts and results about POV measures. In the next chapter we
will do the same for the theory of groups and group representations. Chapters 5
and 6 will then be devoted to a study of reproducing kernel Hilbert spaces. The
treatment is necessarily condensed, but we give ample reference to more exhaustive
literature. Although the mathematically initiated reader may wish to skip these four
mathematical chapters, the discussion of many of the topics here is sufficiently
different from their treatment in standard texts to warrant at least a cursory glance
at it.
As in the previous chapter, let H be a separable, complex Hilbert space. The
set of all bounded, linear operators on H will be denoted by L (H), and its subset
of positive elements by L (H)+ . Note that a bounded operator A is an element of
L (H)+ if and only if A is self-adjoint and  |A  0, for all vectors in H.
In particular, A = P is an (orthogonal) projection operator if P = P = P2 . We shall

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 37
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__3,
Springer Science+Business Media New York 2014
38 3 Positive Operator-Valued Measures and Frames

not have much occasion to deal with unbounded operators. When we do have one,
call it T , we will denote by D(T ) its (usually dense) domain and by Ran(T ) its range.
Throughout, and as far as possible, we have attempted to illustrate new concepts
with simple examples.

3.1 Definitions and Main Properties

The following is a brief recapitulation of some aspects of the theory of POV

measures. We only mention those concepts which will be needed in the sequel.
A more detailed treatment may be found in, for example, [Ber66]. Note that
similar considerations on the relationship between POV measures1 and frames have
appeared in a recent paper by Moran et al. [472].
Let X be a metrizable, locally compact space. (All the group spaces and
parameter spaces for defining CS will be of this type. Metrizability is a technical
assumption, which entails that Baire subsets and Borel subsets of X coincide). The
Borel sets B(X) of X consist of elements of the -algebra formed by its open
sets. A positive Borel measure on B(X) is a map : B(X) R+ = R+ {}
satisfying:
1. For the empty set, 0,
/

(0)
/ = 0. (3.1)

2. If J is a countable index set and i , i J, are mutually disjoint elements of B(X)

(i.e., i j = 0,
/ for i = j), then

(iJ i ) = (i ). (3.2)
iJ

If in addition, satisfies

( ) = sup (C), (3.3)

C , C = compact

then it is called a regular Borel measure [Par05]. Unless the contrary is stated, all
Borel measures will be assumed to be regular. The measure is said to be bounded
or finite if (X) < .
While referring to a measure , we shall use either of the two notations, or d .
However, the measure of a set will be written as ( ). The most common example
of a Borel measure is the Lebesgue measure, d = dx, on the real line R. The
Borel sets in this case are the elements of the -algebra formed by open intervals

1 These are often referred to as POVMs.

3.1 Definitions and Main Properties 39

= (a, b) in R, with ( ) = b a. Furthermore, if (x), x R, is a positive

density function, then d (x) ( dx) = (x) dx defines another Borel measure
on R. Of course, the Lebesgue measure is not finite, but if the density has a finite
integral over R, then d is a finite measure.
Definition 3.1.1 (POV-measure). Let H be a Hilbert space and X a locally com-
pact space. A POV measure is a map a : B(X) L (H)+ , satisfying:
/ = 0.
1. a(0)
2. For a countable index set J and mutually disjoint elements i , (i J), in B(X),

a(iJ i ) = a(i ). (3.4)

iJ

The above sum is understood to converge weakly, which means that, if and
are arbitrary elements of H, then the complex sum iJ  |a(i )  converges. This
is the same as saying that the POV-measure a is equivalent to the entire collection
of positive Borel measures, ( ) =  |a( ) , for all H.
The POV-measure a is said to be regular if, for each H, the measure is
regular [Par05]. It is said to be bounded if

a(X) = A L (H)+ . (3.5)

This just means that each one of the positive Borel measures above is bounded.
In particular, the POV-measure is said to be normalized if

a(X) = I (= identity operator on H). (3.6)

Let F : X L (H)+ be a weakly measurable function, i.e., for each x X, F(x)

is a bounded, positive operator, and for arbitrary , H, the complex function
x   |F(x)  is measurable in B(X) (that is, the inverse image of any Borel
subset of C belongs to B(X)). Using F and the regular Borel measure , we may
define the regular POV-measure a, such that for all B(X),

a( ) = F(x) d (x). (3.7)

The operator function F is then called a density for the POV-measure a. We shall
assume in all such cases that the support of the measure is all of X. (This is not
a severe restriction on the class of measures used, since one can always delete the
complement of the support of in X, and work with a smaller space X  which is
just the support of ). Furthermore, we shall mainly be concerned with bounded
POV-measures with densities, such that the operator

A= F(x) d (x) (3.8)
X
40 3 Positive Operator-Valued Measures and Frames

has an inverse which is densely defined on H. (This technical requirement will

ensure, in situations that we shall encounter, that using F it is possible to define
a family of vectors which span H). The operator A is sometimes referred to as the
resolution operator, most often as the frame operator. We will not, in general,
assume A1 to be bounded. However, the following special case will be of particular
interest to us. Assume that for each x X, the operators F(x) have the same
finite rank n. Then there exists, for each x, a set of linearly independent vectors,
xi , i = 1, 2, . . . , n, in H for which the map x  xi is measurable and
n
F(x) = |xi xi |, (3.9)
i=1

so that (3.8) can be rewritten as

n 
A= |xi xi | d (x). (3.10)
i=1 X

Definition 3.1.2 (Frame). A set of vectors xi , i = 1, 2, . . . , n, x X, constitutes a

rank-n frame, denoted F {xi , A, n}, if for each x X, the vectors xi , i = 1, 2, . . . , n
are linearly independent and if they satisfy (3.10), with both A and A1 being
bounded positive operators on H. If A = I, for some > 0, the frame is said to be
tight.
Alternatively, the frame condition may be written in the following more familiar
form:
n 
m(A) f 2 |xi | f |2 d (x) M(A) f 2 , f H, (3.11)
i=1 X

where m(A) and M(A) denote, respectively, the infimum and the supremum of the
spectrum of A. These numbers, which satisfy m(A) > 0, M(A) < , are called the
frame bounds. [Note: It is customary, in the literature, to denote frame bounds by
the letters A and B, but this would conflict with our use of the letter A for the
frame operator (3.8)]. In case the frame is tight, m(A) = M(A). When n = 1, we
occasionally write, for simplicity, F {x , A} F {x , A, 1}.
All this takes a more familiar shape if the space X is discrete, with a counting
measure. Indeed the relation (3.10) then reads:
n
A= |xi xi |, (3.12)
i=1 xX

and the frame is discretethe type usually encountered in the literature. We shall
take up the notion of a discrete frame in detail in Sect. 3.4.1 below.
A POV-measure a is said to be commutative if [a( ), a(  )] = 0, for all , 
B(X), i.e., if all the operators a( ), B(X), mutually commute. A POV-
measure a = P is called a projection-valued (PV) measure if P( ) is a projection
operator for each B(X). A PV-measure is necessarily commutative.
3.1 Definitions and Main Properties 41

Let a be a POV-measure and a positive Borel measure on X. We say that a

is smooth with respect to if they have the same sets of measure zero, i.e., for
any B(X), a( ) = 0 implies ( ) = 0 and vice versa. Given a bounded POV-
measure a, it is always possible to find a Borel measure with respect to which it
is smooth. Indeed, let {n }Ni=1 be an orthonormal basis for H. Then

N
1
( ) = 2n n |a( )n , (3.13)
n=1

defines a positive, bounded Borel measure ( (X) a(X)). Clearly, a( ) = 0

implies ( ) = 0. Next, since each term in the sum on the right hand side of the
above equation is positive, ( ) = 0 implies n |a( )n  = 0 for all n. On the other
hand, since a( ) is a positive operator, there exists b L (H) for which b2 = a( ).
Thus bn 2 = 0, for all n, implying that b = 0, whence a( ) = 0, proving that a
is smooth with respect to . Note that if a is a normalized POV-measure, i.e., if
a(X) = I, then (X) = 1, so that becomes a probability measure.

3.1.1 Examples of POV-Measures

We have already seen examples of POV-measures in Chap. 2. In fact (2.52) defines

a POV-measure, with density F(q, p) = |s (q,p) s (q,p) |, on the Borel sets of the
phase space . This is an example of a non-commutative POV-measure, and indeed,
we shall mostly be concerned with POV-measures of this particular type in this
book. Examples of commutative POV-measures are also easily obtained. Let H =
L2 (R, dx) and denote by the characteristic function of the set B(R):

1, if x ,
(x) =
0, otherwise.

Then, the operators P( ), such that

(P( ) )(x) = (x) (x), L2 (R, dx), (3.14)

define a normalized PV-measure (the above relation is assumed to hold for almost
all x R, with respect to the Lebesgue measure). More generally, let f : R R+
be an integrable function satisfying

f (x) dx = 1.
R

For arbitrary B(R), define the operator a( ) on L2 (R, dx) by

(a( ) )(x) = ( f )(x) (x), L2 (R, dx). (3.15)

42 3 Positive Operator-Valued Measures and Frames

Here, denotes convolution:

 
( f )(x) = (x y) f (y) dy = f (x y) dy.
R

Then, it is easy to verify that a is a normalized, commutative POV-measure.

For general normalized, commutative POV-measures there is a representation
theorem [11,12,36] which says that every such measure can in a sense be expressed
in the above manner. To see this in the present context, let us rewrite (3.15)
somewhat differently: for each y R, define the PV-measure Py by

(Py ( ) )(x) = (x y) (x),

and let be the Borel measure

d (x) = f (x) dx, (R) = 1.

Then, it is not hard to see that (3.15) may be reexpressed as

(a( ) )(x) = ( ) (x),

and, as a weak integral, a( ) has the representation


a( ) = Py ( ) d (y). (3.16)
R

This shows that the commutative POV measure a can be expressed as a probability
average (since is a probability measure) over PV-measures. It is this sort of
a representation, as a probability average, which can be generalized to arbitrary
regular, normalized, commutative POV-measures.
For arbitrary POV-measures, commutative or not, there is a general theorem, due
to Namark [479], which states that every such measure can be embedded in a PV-
measure on an enlarged Hilbert space. Applications of this theorem will show up in
various contexts in later chapters, warranting a full statement of it here.
Theorem 3.1.3 (Namarks extension theorem). Let a be an arbitrary normal-
 an isometric
ized POV-measure on B(X). Then there exist a Hilbert space H,
   such
embedding W : H H and a PV-measure P on B(X), with values in L (H),
that

 )P,
Wa( )W 1 = PP( (3.17)

for all B(X), where P is the projection operator

 = W H.
PH (3.18)
3.2 The Case of a Tight Frame 43

 of {a, H} can be chosen to be minimal in the sense that the set

 H}
The extension {P,
of vectors

 ) | B(X), W H}
S = {P( (3.19)

 This minimal extension is unique up to unitary equivalence.

be dense in H.

(Note that in (3.17), the inverse map W 1 is defined on the range of W in H).
For illustrative purposes, we work out the Namark extension for the cases where
a defines a frame, with A = I [see (3.9) and (3.10)].

3.2 The Case of a Tight Frame

 = Cn L2 (X, d ).
For the tight frame F {xi , I, n}, consider the Hilbert space H
This space consists of all -measurable functions : X C such that
n

n 
 2 = | i (x)|2 d (x) < ,
i=1 X

 be the linear map,

i (x) denoting the i-th component of (x) in Cn . Let W : H H

(W )i (x) = i (x) = xi | , i = 1, 2, . . . , n, x X. (3.20)

Then W is an isometry. Indeed,

n  n 
| i (x)|2 d (x) =  |xi H xi | H d (x),
i=1 X i=1 X

which, in view of the weak convergence of the integral (3.10), implies that

 2H
 =  H .
2
(3.21)

(The notation  K means that the norm is taken in the Hilbert space K. Similarly
|K denotes the scalar product in K). If is an arbitrary vector in the range, W H,
 it is easily verified that the inverse map W 1 acts on it as
of H in H,
n 
W 1 = i (x)xi d (x) H, (3.22)
i=1 X

giving a reconstruction formula for = W 1 , when the coefficients i (x) =

 define the projection operators
xi | , i = 1, 2, . . . , n, x X are known. On H

P( ), B(X), by

 ) )(x) = (x),
(P( 
H, (3.23)
44 3 Positive Operator-Valued Measures and Frames

where is the characteristic function of the set . Clearly P is a PV-measure on

 We then have the result:
B(X) with values in L (H).
Theorem 3.2.1. The PV-measure P extends the POV-measure,
n 
a( ) = |xi xi | d (x), B(X), (3.24)
i=1

minimally in the sense of Namark.

 be arbitrary. Then there exist ,  H such that
Proof. Let ,  H

= W 1 P ,  = W 1 P  .

Hence, using (3.23) and the definition of the scalar product on H,
n 
 )P    =
 |PP( H (x)(W )i (x)(W  )i (x) d (x)
i=1 X
n 
=  |xi H xi |  H , by (3.20)
i=1

=  |a( )  H , by (3.7) and (3.9)

= W 1 P |a( )W 1 P  H .

Since W is an isometry from H to W H H,  and ,  are arbitrary, the last equality

is immediately seen to imply (3.17). Thus, {P,  is an extension of {a, H}. To see
 H}
that this extension is minimal, it is enough to prove that if H  is a vector for
which   |P(
 )   = 0, for all B(X) and all W H [see (3.19)], then  = 0.
H
Now
n 
  )   =
 |P(  i (x) i (x) d (x),

H
i=1

and hence the vanishing of the left hand side of this equation, for all B(X),
implies that
n n
 i (x) i (x) =  i (x)xi |  = 0
i=1 i=1

for almost all x X (with respect to the measure ), where = W 1 . Thus, since
H is arbitrary, this means that
n
 i (x)xi = 0
i=1

for almost all x X. The linear independence of the xi , i = 1, 2, . . . , n, then shows

 i (x) = 0, for all i and almost all x X. Hence = 0.
that 

3.3 Frames and Semi-frames 45

3.3 Frames and Semi-frames

In the whole analysis so far, we have assumed that the operator A has a bounded
inverse A1 , but this is in fact not necessary. The important point is that A be
invertible. If A1 is unbounded (it is necessarily densely defined, as the inverse
of a self-adjoint operator), the whole development goes through, with some minor
modifications. In order to appreciate the difference between the two situations, let us
return for a while to frames. For simplicity, we restrict the discussion to the rank-1
case, but the generalization to rank-n is straightforward. We mostly follow [52, 53],
where a complete analysis may be found.

3.3.1 Frames Revisited

Thus we start with a frame F {x , A}, following Definition 3.1.2. First, F {x , A}

is a total set in H. Next define the analysis operator by the (coherent state) map
W : H L2 (X, d ) introduced in (3.20), given as

(W f )(x) = x | f , f H,

and denote its range by RW . Its adjoint W : L2 (X, d ) H, called the synthesis
operator, reads (the integral being understood in the weak sense, as usual [283])

W F = F(x) x d (x), for F L2 (X, d ). (3.25)
X

Then WW = A and W f 2L2 (X) = A1/2 f 2H =  f | A f . Furthermore, W is

injective, since A > 0, so that W1 : RW H is well-defined. (In the signal
processing literature [52, 53], the analysis operator W is often denoted by C and
the synthesis operator W by D = C ).
Next, the lower frame bound implies that RW is a closed subspace of L2 (X, d ).
The corresponding orthogonal projection is P : L2 (X, d ) RW defined by

P := W A1W = W W+ ,

where W+ = A1W is the pseudo-inverse of W . The projection P is an integral

operator with (reproducing) kernel K(x, y) = x | A1 y , thus RW is a reproducing
kernel Hilbert space.
In addition, the subspace RW is also complete in the norm   , associated to the
inner product

F | F   := F |W A1 W1 F  L2 (X) , for F, F  RW . (3.26)

46 3 Positive Operator-Valued Measures and Frames

Hence (RW ,   ) is a Hilbert space, denoted by H , and the map W : H H is

unitary, since it is both isometric and onto. One has indeed, for every F, F  RW ,

F | F   = W f |W f   = W f |W A1 W1W f  L2 (X, d )

= W f |W A1 f  L2 (X, d )

=  f |WW A1 f  H
=  f | f  H .
( )
Therefore, it can be inverted on its range by the adjoint operator W : H H,
which is precisely the pseudo-inverse W+ = A1W . Thus one gets, for every f H,
a reconstruction formula, with a weakly convergent integral:

( )
f = W F= F(x) A1 x d (x), for F = W f H . (3.27)
X

Before going further, it is worth mentioning that continuous or generalized

frames have been discussed by a number of authors, for instance [101, 283, 295,
516, 616, Kai94].

3.3.2 Upper Semi-frames

As a matter of fact, there are situations where the notion of frame is too restrictive,
in the sense that one cannot satisfy both frame bounds simultaneously. Thus there
is room for two natural generalizations, called semi-frames. More precisely, we say
that a family F {x , A} is an upper (resp. lower) semi-frame, if it satisfies the upper
(resp. lower) frame inequality in (3.11). Thus, in the upper case, A is bounded and
A1 is unbounded, whereas, in the lower case, A is unbounded and A1 is bounded.
In the sequel of this section, we discuss the case of upper semi-frames. We will come
back to lower semi-frames and the duality relation between the two in Sect. 3.3.3.
Let now F {x , A} be an upper semi-frame, which is not a frame, that is, there
exists M < such that

0< |x | f |2 d (x) M  f 2 , f H, f = 0. (3.28)
X

In this case, F {x , A} is a total set in H, the operators W and A are bounded, A is

injective and self-adjoint. Therefore, the range RA of A is dense in H and A1 is also
self-adjoint. A1 is unbounded, with dense domain D(A1 ) = RA .
Define the Hilbert space H as the completion of W (RA ) with respect to the
norm  defined in (3.26). Then, the same calculation as above shows that the map
W , restricted to the dense domain D(A1 ) = RA , is an isometry onto W (RA ) H :

W f |W f   =  f | f  H , f , f  RA .
3.3 Frames and Semi-frames 47

Thus W extends by continuity to a unitary map between the respective completions,

namely, from H onto H . Therefore, H and RW coincide as sets, so that H is a
vector subspace (though not necessarily closed) of L2 (X, d ).
Consider now the operators

GA := W AW1 : RW W (RA ),

G1 1 1
A := W A W : W (RA ) RW ,

both acting in the Hilbert space RW , the closure of RW in L2 (X, d ). Then one shows
[27, Sect. 3] that GA is a bounded, positive and symmetric operator, while G1 A is
positive and essentially self-adjoint. These two operators are bijective and inverse
to each other. Thus one gets the following commutative diagram:

W
- =RW RW L2(X,d)
6 6

W
A A1 GA GA1

? ?
- W (R )L2(X,d)
?

D(A1)=RA A
W
(3.29)

Next let G = GA and let G1 be the self-adjoint extension of G1A . Both operators
are self-adjoint and positive, G is bounded and G1 is densely defined in RW .
Furthermore, they are inverse of each other on the appropriate domains. Moreover,
since the spectrum of G1 is bounded away from zero, the norm  is equivalent
 1/2
to the graph norm of G1/2 = G1 , so that

Ran(G1/2 ) = D(G1/2 ) = H = RW RW L2 (X, d ).

For further details, we refer to [27, 52, 53], where the analysis is made in full
generality.
At this point, we make a distinction. We will say that the upper semi-frame
F {x , A} is regular if all the vectors x , x X, belong to D(A1 ). This will simplify
some statements below.
Indeed, let us first assume that F {x , A} is regular. Then the discussion
proceeds exactly as in the bounded case. In particular, the reproducing kernel
K(x, y) = x | A1 y  is a bona fide function on X X. One obtains the same weak
reconstruction formula, but restricted to the subspace RA = D(A1 ):

( )
f = W F= F(x) A1 x d (x), f RA , F = W f H . (3.30)
X
48 3 Positive Operator-Valued Measures and Frames

On the other hand, if F {x , A} is not regular, one has to treat the kernel K(x, y)
as a bounded sesquilinear form over H and use the language of distributions,
for instance, with a Gelfand triplet or rigged Hilbert space [Gel64]. This kind of
approach has become familiar, both in the theory of representations of noncompact
groups [170, 446, 478] and in quantum mechanics [48].
The construction, originating from [29, Sect. 4], proceeds as follows [52, 53].
If F {x , A} is regular, one has indeed

F(x)K(x, y)F  (y) d (x) d (y) = W1 F | AW1 F  H , F, F  H . (3.31)
XX

Since W is an isometry and A is bounded, the relation (3.31) defines a bounded

sesquilinear form over H , namely

K (F, F  ) = W1 F | AW1 F  H , (3.32)

and this remains true even if F {x , A} is not regular. Denote by H

the Hilbert
space obtained by completing H in the norm given by this sesquilinear form. Now,
(3.31) and (3.32) imply that

K (F, F  ) = W1 F | AW1 F  H = F |W AW1 F   = F | F  L2 .

Therefore, one obtains the following Gelfand triplet, with continuous and dense
range embeddings,

H H0 H
, (3.33)

where
H = RW , which is a Hilbert space for the norm  =  | G1 L2 .
1/2

H0 = H = RW is the closure of H in L2 (X, d ).

H
1/2
is the completion of H0 (or H ) in the norm  :=  | GL2 , as well as
the conjugate dual of H .
If F {x , A} is regular, all three spaces H , H0 , H are reproducing kernel Hilbert
space s, with the same kernel K(x, y) = x | A y . Of course, if A1 is bounded,
1

that is, in the case of a frame, the three Hilbert spaces of (3.33) coincide as sets, with
equivalent norms, since then both A and A1 belong to GL(H), thus G and G1 are
both bounded on L2 (X, d ).
Actually, one can go further. As we have seen, H = RW is the form domain of
G1 , that is, the domain D(G1/2 ), with its graph norm. On the side of H, i.e., on
the l.h.s. of the diagram (3.29), this corresponds to the form domain of A1 , with
norm  =  | A1 H .
1/2
3.3 Frames and Semi-frames 49

We can instead consider the smaller space RA = D(A1 ), with norm A =
A1 | A1 H , which is complete, hence a Hilbert space.
1/2

Taking the (conjugate) dual again, we obtain a new Gelfand triple. Mapping
everything into L2 by W , we obtain the following scale of Hilbert spaces:

HA H H0 = RW H
HA . (3.34)

In the quintet (3.34), HA := W (RA ), with norm A =  | G3 L2 .

1/2

Iterating, we see that the multiplet (3.34) is the central part of the Hilbert scale
built on the powers of the positive self-adjoint operator G1/2 , namely, Hm :=
D(Gm/2 ), m Z

H H1 , W (RW ) H2 , HA H3 , H
H1 , ...,

where RW denotes the range of W . Similar considerations apply on the side

of H, that is, we obtain the scale built on the powers of A1/2 , i.e., H  m :=
D(Am/2 ), m Z. Putting everything together, we obtain the commutative diagram
presented on page 50.
In the upper row of the diagram, the operator A1/2 is unitary from each space
onto the next one. The same is true for the operator G1/2 in the lower row. On the
left side of the central spaces H and RW , the operators W and W are defined as
usual, and to the right they are defined by duality. Clearly, using the above notation
we have W : Hm H  m+1 , W : H
 m Hm+1 and A : H m H  m+2 . Furthermore, Ak

being unitary for every k R, one has (A ) = A .
k k

In both cases, we obtain in this way a simple partial inner product space [Ant09].
A natural question then is to identify the end spaces,
 
H (G1/2 ) := Hm , H (G1/2 ) := Hm , (3.35)
mZ mZ

and similarly for the scale H  m := D(Am/2 ), m Z. In some simple examples, the
question can be answered explicitly. In this way, one has at ones disposal the full
machinery of partial inner product spaces [Ant09]. For instance, one can ask under
which conditions the space H is nuclear, and similar questions whose answer relies
on the structure of the full scale.

3.3.3 Lower Semi-frames, Duality

Given a frame := F {x , A}, one says [616] that a frame := F {x , B} is dual

to the frame if one has, in the weak sense,

f= x | f  x d (x), f H.
X
 50

    
2 1 0 1 2

    
A1/2 A1/2 A1/2 A1/2 A1/2 A 1/2
RA RW RA1/2 RA 1

W W W W W W W W W

G 1/2 G 1/2 G 1/2 G 1/2 G 1/2 G 1/2

W (RA) W(RW ) RW RW W(RA1 )

    

A 3 2 1 0 1
3 Positive Operator-Valued Measures and Frames
3.3 Frames and Semi-frames 51

Then the frame is dual to the frame . This applies, in particular, to a given
frame = F {x , A} and its canonical dual  := F { x , A1 }. This notion can be
extended to semi-frames. First, it is known [295] that an upper semi-frame is a
frame if and only if there exists another upper semi-frame which is dual to , in
the sense that

 f | f  = x | f  x | f   d (x), f , f  H.
X

Let first = {x } be an arbitrary total family in H. Then, as usual, we define the

analysis operator W : D(W ) L2 (X, d ) as W f (x) = x | f  on the domain

D(W ) := { f H : |x | f |2 d (x) < }.
X

Next, we define the synthesis operator V : D(V ) H as


V F = F(x) x d (x), F D(V ) L2 (X, d ), (3.36)
X

on the domain

D(V ) := {F L2 (X, d ) : F(x) x d (x) converges weakly in H }.
X

A priori they are both unbounded. Following [198, Lemma 3.1] and [113,
Lemma 3.1 and Proposition 3.3], we have, as in the discrete case,
Lemma 3.3.1. (i) Given any total family , the analysis operator W is closed.
Then satisfies the lower frame condition if and only if W has closed range
and is injective.
(ii) If the function x  x | f  is locally integrable for all f H, then the operator
V is densely defined and one has W = V .
A proof is given in [52, Lemmas 2.1 and 2.2]. The condition of local integrability
is satisfied for all f D(W ), but not necessarily for all f H, unless is an upper
semi-frame, since then D(W ) = H.
Finally, one defines the frame operator as A = V W , so that, in the weak sense,

A f = x | f x d (x), f D(A ),
X

where

D(A ) := { f H : x | f  x d (x) converges weakly in H }.
X
52 3 Positive Operator-Valued Measures and Frames

Notice that one has in general D(A )  D(W ). As in the discrete case [113,
Lemma 3.1], one has D(A ) = D(W ) if and only if Ran(W ) D(V ). This
happens, in particular, for an upper semi-frame , for which one has D(A ) =
D(W ) = H.
1
For an upper semi-frame, A : H H is a bounded injective operator and A is
unbounded. If = {x } satisfies the lower frame condition, then A : D(A ) H
1
is an injective operator, possibly unbounded, with a bounded inverse A . However,
if the upper frame inequality is not satisfied, A and W could have nondense
domains, in which case one cannot define a unique adjoint W and A may not be
self-adjoint. However, if y D(W ), y X, then W is densely defined, V W
and V is closable. Finally, V is closed if and only if V = W . Then A = W W
is self-adjoint [52, Lemmas 5.3 and 5.4].
Next, we say that a family = {x } is a lower semi-frame if it satisfies the lower
frame condition, that is, there exists a constant m > 0 such that

m  f 2 |x | f |2 d (x), f H. (3.37)
X

Clearly, (3.37) implies that the family is total in H. With these definitions, we
obtain a nice duality property between upper and lower semi-frames.
Proposition 3.3.2. (i) Let = F {x , A} be an upper semi-frame, with upper
frame bound M and let = {x } be a total family dual to . Then is a
lower semi-frame, with lower frame bound M1 .
(ii) Conversely, if = {x } is a lower semi-frame, there exists an upper semi-frame
= {x } dual to , that is, one has, in the weak sense,

f= x | f  x d (x), f D(W ).
X

A proof may be found in [52, Lemma 2.5 and Proposition 2.6].2 The same paper
(Sects. 2.6 and 2.7) presents concrete examples of a non-regular upper semi-frame
(from affine coherent states) and of a lower semi-frame (from wavelets on the two-
sphere).
The first example concerns the affine coherent states introduced by Paul [499,
Pau85] (see also [29]). These coherent states stem from the following unitary
irreducible representation of the connected affine group G+ or ax + b group (12.21):
 
Un (b, a) f (r) = an/2 eibr f (ar), a > 0, b R, f H(n) ,

where H(n) := L2 (R+ , rn1 dr), n = integer 1. The coherent states are indexed
by points of the quotient space G+ /H R, where H denotes the subgroup of
dilations. It turns out that the representation Un is square integrable mod(H, ), for

2 That paper contains an error, which was corrected in the Corrigendum. See also [53].
3.3 Frames and Semi-frames 53

a suitable section : R G+ . This concept, sketched in Sect. 2.2 for the case of the
WeylHeisenberg group, will be studied in detail in Sect. 7.1.1. As a consequence,
coherent states may be constructed by the general formalism developed there. They
take the form

x (r) = eixr (r), r R+ ,

where is admissible if it satisfies the two conditions

(i) sup a(r) = 1, where a(r) := rn1 | (r)|2 ,

rR+

(ii) | (r)|2 = 0, except perhaps at isolated points r R+ .

We emphasize that these wavelets living on the half-line R+ are different from the
usual wavelets developed in Chap. 12. In fact they resemble more those encountered
in Sect. 14.2.2 devoted to radial functions. The frame operator A and its inverse A1
are multiplication operators on H(n) , namely

(A1 f )(r) = [a(r)]1 f (r).

Since a(r) 1, the inverse A1 is indeed unbounded and no frame vector x belongs
to its domain. Thus we have indeed a non-regular upper semi-frame. Of course, we
have also, for every n Z,

(An f )(r) = [a(r)]n f (r).

 (n)
Thus the scale generated by A1/2 consists of the spaces H m = D(A
m/2 ), m Z,

with norm

1/2
 f m =  f | Am f 1/2 = | f (r)| [a(r)]
2 m n1
r dr . (3.38)
0

However, the end spaces of the scale, namely

 
 (A1/2 ) :=
H  (n)
H m ,
 (A1/2 ) :=
H  (n)
H m
mZ mZ

do not seem to have an easy interpretation.

As for the second example, we shall discuss it in Sect. 15.1.1.1.
54 3 Positive Operator-Valued Measures and Frames

3.4 Discrete Frames and Semi-frames

3.4.1 Discrete Frames

As mentioned in Sect. 3.1 above, an important case for applications is that of a

discrete frame, which corresponds to a discrete space X, equipped with a counting
measure . The theory of discrete frames was originally developed in the 50s in
the context of nonharmonic Fourier series [265], and regained prominence recently,
in the context of Gabor and wavelet analysis [234, 370, Chr03, Dau92, Gr01] (see
Chaps. 12 and 17). In view of their practical importance, it is fitting to discuss them
here at some length.
A countable family of vectors { j , j } in a separable Hilbert space H is called
a (discrete) frame if there are two positive constants m, M, with 0 < m < M < ,
such that

m  2 | j | |2 M  2 , H (3.39)
j

The index set may be finite or infinite. The two constants m, M are the frame
bounds. If m = M, the frame is said to be tight. If m = M = 1, one speaks of a
Parseval frame. If, in addition,  j  = 1, j, the set { j } is simply an orthonormal
basis. Note that finite frames will be discussed within the framework of quantization,
in Sect. 11.6.
It should be clear that the definition of (discrete) frame just given coincides with
that given in (3.12) above, namely

A= | j  j |, (3.40)
j

that is,

 | A  = | j | |2 . (3.41)
j

Thus the frame condition (3.11), which says that A and A1 are both bounded, coin-
cides with (3.39), with frame bounds m = m(A), M = M(A), denoting respectively
the infimum and the supremum of the spectrum of A. In the notation introduced
above, the frame would be denoted as F { j , A}.
The properties of a frame are best discussed in terms of the analysis operator
W : H 2 , defined, as in (3.20), by

W :  { j | }.

Its adjoint is the synthesis operator D = W : 2 H given by Dc = j c j j ,

where c = (c j ). Then the frame operator is A = W W and is given by

Af =  j | f  j .
j
3.4 Discrete Frames and Semi-frames 55

Thus the inequalities (3.39) may be written as

mI W W MI, (3.42)

where I is the identity operator. This in turn implies that W W is invertible and

M1 I (W W )1 m1 I. (3.43)

Define now, for each j N :

 j = A1 j = (W W )1 j ,
(3.44)

so that j = A
 j . Then the following is true:
Theorem 3.4.1. The vectors { j } constitute a frame, with frame bounds M1 , m1
 = W (W W )1 . In addition, the expansion
and analysis operator W

(x) =  j |   j (x), (3.45)

j

 W = I.
converges strongly in H, i.e., W
Proof. That {
 j } is the frame described in the statement results from the equalities

| j | |2 = |(W W )1 j | |2

j j

= | j |(W W )1 |2
j

= W (W W )1 2 =  |(W W )1 

and the inequalities (3.39). Furthermore

 W = (W W )1W W = I,
W
that is, (3.45) is an identity. 

In other words, the duality between the two frames may be written as  W
W =
W W = I or explicitly,

| j  j | = | j  j | = I. (3.46)
j j

The frame F {  j , A1 } is called the canonical dual of F { j , A}. In general,

however, a frame may have many different duals, and it has a unique dual if and
only if it is a Riesz basis [197, Corollary 6.65], [Chr03, Theorem 5.6.5]. This notion
is crucial for applications. In the case of wavelet expansions, it is the basis of the
so-called biorthogonal scheme [Dau92], briefly discussed in Sect. 13.1.3.
56 3 Positive Operator-Valued Measures and Frames

The important point here is that, for all practical purposes, a good frame is
almost as good as an orthonormal basis. By good frame, we mean that the
expansion (3.45) converges sufficiently fast. How would one estimate the speed of
this convergence? By (3.45), we need to compute  j = (W W )1 j . If M and m are
close to each other, A = W W is close to 2 (M + m)I, hence A1 is close to M+
1 2
mI
and thus  j is close to M+m j . Hence we may write
2

2
=
M+m  j |  j + R , (3.47)
j

where
2
R=I A. (3.48)
M+m
Hence
2
A1 = (I R)1
M+m

2
=
M+m Rk . (3.49)
k=0

The series converges in norm, since, by (3.48),

Mm Mm
I R I, (3.50)
M+m M+m
which implies
M m M/m 1
R = .
M + m M/m + 1
Therefore the expansion (3.45) converges as a power series in |M/m 1|. Thus the
frame is good if |M/m 1|  1, in particular if it is tight. To the first order, the
expansion (3.45) becomes
2
=
M+m  j |  j . (3.51)
j

The quantity

Mm
w(F ) = (3.52)
M+m

is called the width of the frame F . It measures the lack of tightness, since w(F ) = 0
if and only if the frame F is tight. Notice that a frame and its dual have the same
width. More details on frames may be found in [232, 370, Chr03, Dau92].
3.4 Discrete Frames and Semi-frames 57

3.4.2 Weighted and Controlled Frames

As we will see in Chap. 15, the classical notion of frame given in (3.39) is not always
sufficient, we have to generalize it. We introduce two variants to this classical notion,
both originally introduced in the problem of discretizing the spherical wavelet
transform [154, Jac04] and later analyzed in [112].
As compared to the classical definition (3.39), the family {k , k } is a
controlled frame in H if there is a positive bounded operator C, with bounded
inverse, such that

m  f 2 k | f   f |Ck  M  f 2 , for all f H. (3.53)

k

The family {k } is a weighted frame in H if there are positive weights v2k > 0
such that

m  f 2 v2k |k | f |2 M  f 2 , for all f H. (3.54)

k

These two notions are in fact mathematically equivalent to the classical notion of
frame [112], namely, a family of vectors {k } is a controlled frame, resp. a weighted
frame, if and only if it is a frame in the standard sense (with different frame bounds,
of course). However, this is not true numerically, the convergence properties of the
respective frame expansions may be quite different [112, Ant04]. And, indeed, the
new notions are used precisely for improving the convergence of the reconstruction
process, as we will see in Sect. 15.1.1.2.

3.4.3 Fusion Frames

Given a weighted frame := {k }, suppose now the weights are constant by blocks
of finite size n j , so that (3.54) takes the form

nj
m  f 2 v2j |i j | f |2 M  f 2 , for all f H.
jJ i=1

Then, for each j, the family {i j , i = 1, 2, . . . , n j } is a frame for its span, call it H j ,
which is at most n j -dimensional. Call H j the corresponding orthogonal projection.
Let m j , M j be the frame bounds,

 2 nj  2
   
m j H j f  |i j | f |2 M j H j f  ,
i=1
58 3 Positive Operator-Valued Measures and Frames

and assume that minf := inf j m j > 0 and Msup := sup j M j < . Then we get
 2
m   M
Msup
 f 2 v2j H j f 
minf
 f 2 , f H. (3.55)
jJ

In that case, the family {H j } jJ is a fusion frame3 with respect to the weights
{v j } jJ , a notion introduced by Casazza and Kutyniok [196, 199]. Actually, in
the general definition, the subspaces {H j } jJ are closed subspaces of H, of
arbitrary dimension. This structure nicely generalizes frames, in particular, it yields
associated analysis, synthesis and frame operators and a dual object.
Given the family {H j } jJ , one considers their direct sum
  2
H := H j = {{ f j } jJ : f j H j ,  f j  < }} ,
jJ jJ

and this is the ambient Hilbert space. In terms of H , one considers, following the
standard pattern,
1. The analysis operator W ,v : H H defined by W ,v f = {v j H j f } jJ .

2. The synthesis operator D ,v = WW,v : H H, given by

D ,v f = vj fj , for all f = { f j } H .
jJ

Note that the series on the r.h.s. converges unconditionally.

3. The frame operator A ,v : H H given, as usual, by A ,v = W ,v W ,v , so that

A ,v f = v2j H j f .
jJ

Most of the standard results about ordinary frames extend to fusion frames, for
instance, the duality relation and the reconstruction formula. We refer to [52, 53] or
the original papers for details.

3.4.4 Discrete Semi-frames

This, however, is not the end of the story. Indeed, in the discrete case also, the notion
of frame is sometimes too restrictive, and one has to resort to semi-frames, upper
or lower (an upper semi-frame is often called a (total) Bessel sequence in the signal

3 Initially called frame of subspaces in [196].

3.4 Discrete Frames and Semi-frames 59

processing literature). The analysis is the same as in the general case, as it should
[53]. But here we can be more concrete. First we consider a simple example.
Let (ek ), k N, be an orthonormal basis in H. Let k = 1k ek . Then = (k ) is
an upper semi-frame:

0< |k | f |2 |ek | f |2 =  f 2 .

k k

There is no lower frame bound, since for f = e p , one has |k | f |2 = 1
p2
.
k
Let k = k ek . The sequence = (k ) is dual to (k ), since it obviously satisfies
the relations

f = k | f  k = k | f  k . (3.56)
k k

In addition, we have

|ek | f |2 =  f 2 |k | f |2 ,

k k

and this is unbounded, since k |k | f |2 = p2 for f = e p . Hence, (k ) is a lower

semi-frame, dual to (k ).
In this case, in the basis (ek ), the frame operator associated to (k ) is A =
diag(1/n2 ). Thus A1 = diag(n2 ), which is clearly unbounded. It follows that (k )
is the canonical dual of (k ), since k = A1 k . The sequence used by Gabor in his
original IEE-paper [294], a Gabor system with a Gaussian window, a = 1 and b = 1,
is exactly such an upper semi-frame.
Similarly, G1 = diag(n2 ), acting in 2 , so that the inner products of the three
spaces in the Gelfand triplet analogous to (3.33) are, respectively:
For H : c | d = n n2 cn dn ;
For H0 : c | d0 = n cn dn ;
For H
: c | d
= n n2 cn dn .
1

Both W = (W k ) and W = (W k ) live in H , since {k }n = 1k kn and

{k }n = k kn . In addition, and for the same reason, the upper semi-frame is totally
regular, which means that, for every k Z, k H  (G1/2 ) = nZ H
 n.
Now, in this example, we can identify the end spaces in the scale {Hm :=
D(Gm/2 ), m Z}. We get
 
H (G1/2 ) = Hm = s, H (G1/2 ) = Hm = s , (3.57)
mZ mZ

the space of fast decreasing, resp. slowly increasing, sequences (the so-called
Hermite representation of tempered distributions [555]). And, indeed, s is a nuclear
space, the proof using precisely this representation.
60 3 Positive Operator-Valued Measures and Frames

The example ( 1k ek ), (kek ) can be generalized to weighted sequences (k ), (k ),

with k := wk ek , k := w1 ek , where w  has a subsequence converging to zero and
k
wk = 0, k. Hence the former is an upper semi-frame and not a frame, whereas the
latter satisfies the lower frame condition, but not the upper one. The frame operator
associated to the sequence (wk ek ) is still diagonal, namely, A = diag(w2n ). Thus
A1 = diag(w2 1
n ), which is clearly unbounded, and k = A k . The inner products
read as:
For H : c | d = n m2n cn dn
For H0 : c | d0 = n cn dn
For H
: c | d
= n mn cn dn
2

Here too, the upper semi-frame (k ) is totally regular, since {k }n = wk kn . The

considerations made above about the triplet (3.33) or the relations (3.57) apply in
the case of weighted sequences as well. For instance, if the sequence (1/wk ) grows

polynomially, one gets the same result, namely, the end spaces H (G1/2 ) = m Hm ,

resp. H (G1/2 ) = m Hm still coincide with s and s , respectively.
Finally we remark that all these considerations extend to fusion upper semi-
frames [53].

3.4.5 Discretization

An interesting question of discretization now arises, namely, what is the connection

between continuous frames and discrete ones? As we shall see in Sect. 12.5, the
practical implementation of the wavelet transform (a special case of coherent states)
in signal processing requires the selection of a discrete set of points in the transform
space. Indeed, all formulas must generally be evaluated numerically, and a computer
is an intrinsically discrete object (even finite!). But this operation must be performed
in such a way that no information is lost. This requirement then immediately leads
to the determination of a discrete frame. The same situation actually prevails for
any CS transform, of the type we shall develop at length in this book. Putting the
question in general terms, it reads now: Given a continuous frame, {x , Ac , x X},
based on a general space X, can one find a discrete set of points {x j X, j J} such
that {x j , Ad , j J} is a discrete frame, possibly with a different width? Moreover,
a priori the two operators Ac , Ad need not coincide.
Besides its obvious mathematical interest, a positive answer to this question is
crucial for practical applications of CS. In view of its importance, we shall devote
the last chapter of this book (Chap. 17) to the discretization problem, listing on the
way the explicit results that are known in particular examples.
Chapter 4
Some Group Theory

Abstract In this chapter, we provide the necessary group-theoretical background

that will be used in later chapters. We start with the notion of homogeneous space
of a locally compact group and its (quasi-) invariant measures. Then we turn to
induced representations and the attending notion of system of covariance. This is
illustrated by vector CS and the discrete series representations of SU(1,1). Next we
describe briefly some aspects of harmonic analysis on a locally compact Abelian
group. Finally we survey the basic facts concerning Lie groups and Lie algebras,
their extensions and their contractions.

In this chapter, we introduce a few concepts from the theory of groups, Lie algebras,
transformation spaces, and group representations, presenting them in a form and
with notations adapted to the aims of this book. (A good source for more detailed
information is, for example, [Bar77].)

4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant

Measures

Let G be a locally compact (metrizable) groupin fact, most of the time, we shall
take it to be a Lie group (additional information, specific to Lie groups and Lie
algebras, is given in Sect. 4.5). Suppose that X is a transformation space for G,
also called a G-space. This means that there is defined on X an action (we shall
only consider a left action) of the group: G X X, to be written (g, x)  gx and
assumed to be continuous in the topologies of G and X. If G is a Lie group, X
will be taken to be a manifold and the above action assumed smooth, i.e., infinitely
differentiable. We shall mostly be concerned with the case in which the action of
G on X is transitive, which means that given x, y X, it is always possible to find

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 61
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__4,
Springer Science+Business Media New York 2014
62 4 Some Group Theory

g G, which solves the equation y = gx. The space X will then be called a transitive
G-space or a homogeneous space. The stability subgroup of a point x X is the set

Hx := {g G : gx = x}. (4.1)

This is a closed subgroup of G. As an example of a G-space, let H be a closed

subgroup of G, and consider the left coset space G/H. Elements of G/H are of
the form gH, for g G, and G itself has the transitive left action gH  g gH (for
arbitrary g G) on G/H. If G is a Lie group and H is a Lie subgroup of G,
then G/H has a natural structure of a manifold and the above action is smooth.
In fact, such a transformation space is the generic example of a homogeneous
space. Indeed, let X be any homogeneous space of G, and let x0 X be fixed.
Denote by H the stability subgroup of x0 . Since G acts transitively on X, for any
x X, we can find g G, for which x = gx0 . The identification x  gH is then a
homeomorphism (diffeomorphism, if G is a Lie group) between X and G/H. Let X
be a transformation space for G and x G. The orbit of x under G is the set

Gx := {y = gx : g G} X. (4.2)

If X is a homogeneous space, it corresponds to a single orbit under G.

The group G carries a left and a right invariant Haar measure. We shall denote
the left Haar measure by and use it systematically. The right Haar measure, when
used, will be denoted by r . If = r , the group is called unimodular. In general,
and r are different but equivalent measures; i.e., they have the same null sets.
Thus, there exists a measurable function  : G R+ , such that

d (g) = (g) dr (g). (4.3)

This function, called the modular function of the group, is an R+ -valued character
satisfying, for -almost all g, g1 , g2 G,

(g) > 0,
(e) = 1, e = identity element of G, (4.4)

(g1 g2 ) = (g1 )(g2 ).

Furthermore, for -almost all g, g G, the following relations hold:


dr (g) = (g1 ) d (g) = d (g1 ),
(4.5)
d (gg ) = (g ) d (g).
While the group itself always carries a left (and a right) invariant Haar measure,
the homogeneous space X need not carry any measure invariant under the action
x  gx. Quasi-invariant measures, however, always exist on X. The measure on
X is said to be quasi-invariant if and g are equivalent measures, for all g G,
where g is defined to be the measure obtained by the natural action of g on :

g ( ) = (g ), B(X). (4.6)
4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant Measures 63

The RadonNikodym derivative of g with respect to ,

dg (x)
(g, x) = , (4.7)
d (x)
is then a cocycle, : G X R+ , with the properties

(g1 g2 , x) = (g1 , x) (g2 , g1
1 x), (4.8)
(e, x) = 1,
the above equations holding for -almost all g1 , g2 G, and -almost all x X.
(Note that all the measures g , g G, belong to the same measure class; i.e., they
all have the same measure-zero sets.) Given a cocycle such as above, however, it
is always possible to find a strict cocyle [i.e., one for which the conditions (4.8)
hold for all g1 , g2 G, and all x X], which is equal to the given cocyle almost
everywhere. We shall always assume that such a strict cocycle has been chosen.
Let H be a closed subgroup of G, and write X = G/H and : G X, the
canonical surjection (or projection map), (g) = gH. A section on X is a map
: X G, satisfying ( (x)) = x, for all x X. While it is always possible to find
sections that are Borel maps, this is not the case if one insists on smooth sections (in
the case in which G is Lie group). On the other hand, local smooth sections always
exist. This means that, given x X, it is always possible to find an open set U X,
with x U, and a smooth section U : U G. (It ought to be noted, however, that
U can be quite arbitrary.) More than that, if X is a connected manifold, one may
assume the open set U X to be dense in X. (Note that X \ U is then a set of
-measure zero.)

4.1.1 A Simple Example

Let us illustrate, with a simple example, the various measures appearing on a

group and on homogeneous spaces. Consider the group Gut of all 2 2 real, upper
triangular matrices with nonzero determinant:
 
xy
g= , x, y, z R, xz = 0. (4.9)
0z
The left and right invariant Haar measures for Gut can be easily computed as
dx dy dz dx dy dz
d (g) = , dr (g) = , (4.10)
|x2 z| |xz2 |
from which we find that
z
(g) = . (4.11)
x
64 4 Some Group Theory

Now, let H1 be the subgroup of Gut , which consists of all elements of the type:
 
x0
h= . (4.12)
0z

Since an arbitrary g Gut can be written as

   y  
xy 1 z x0
g= = ,
0z 01 0z

we see that the coset space Gut /H1 can be parametrized by a single variable u (= yz )
R. Alternatively, an element in Gut /H1 can be represented by a matrix
 
1u
u= Gut . (4.13)
01

To obtain the transformation properties of u under the action of g Gut , note that
 xu+y   
1 z x0
gu = .
0 1 0z

This gives the transformation rule,

xu + y
u  gu = .
z
Thus, the Lebesgue measure du is quasi-invariant on Gut /H1 R, with
z
d(g1 u) = (g, u) du = du. (4.14)
x
It is also easy to see that no invariant measure exists on Gut /H1 . Consider, however,
the subgroup H2 of Gut , which consists of all matrices of the type
 
1y
h= , y R. (4.15)
01

Then proceeding as before, it can be seen that the coset space Gut /H2 is identifiable
with all matrices of the type
 
x0
v= Gut .
0y

Since this is topologically isomorphic to R2 , with the origin taken out, we find that
Gut /H2 carries the invariant measure,

dx dy
d (x, y) = . (4.16)
|xy|
4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant Measures 65

Finally, note that, when we chose to represent an element of Gut /H1 by the matrix
u in (4.13), we actually made the association uH1  u; in other words, we chose a
section
 
1u
0 : Gut /H1 R Gut , 0 (u) = . (4.17)
01

An entire class of other sections may then be obtained from 0 by writing

   
f1 (u) 0 f1 (u) u
(u) = 0 (u) = , (4.18)
0 f2 (u) 0 f2 (u)

where f1 and f2 are real-valued, nowhere vanishing Borel functions.

Returning to the general discussion, it is possible, using a global Borel section
and a quasi-invariant measure on the homogeneous space X = G/H, to construct
another quasi-invariant measure , which in a sense is the standard quasi-invariant
measure for the chosen section. Indeed, using a strict cocycle , let us write

d (x) = ( (x), x) d (x). (4.19)

The measure inherits interesting properties from g . Indeed, since is a Borel

section, is a Borel function, and thus is a Borel measure. Also, let x0 X be
the point that is stable under H, so that

x = (x)x0 . (4.20)

Furthermore, if f : X C is an arbitrary Borel function, then the defining relation

(4.6) implies, for all g G,
  
f (x) dg (x) = f (gx) d (x)= f (x) (g, x) d (x)
X X X

= f (x) d (g1 x). (4.21)
X

Alternatively,

dg (x) = (g, x) d (x) = d (g1 x). (4.22)

From these relations, we may now derive the following properties of :

1. is a quasi-invariant measure. Indeed, this follows immediately from (4.6) and
the fact that (g, x) > 0, for all g, x. Also, from (4.6) and (4.22),

d (g1 x) = ( (g1 x), x) (g, x) d (x). (4.23)

66 4 Some Group Theory

2. If X admits an (left) invariant measure m, then is a scalar multiple of m.

Indeed, in this case one can find a function f , which can be taken to be positive
everywhere, such that the quasi-invariant measure satisfies the relation d (x) =
f (x) dm(x). Thus,

dg (x) = d (g1 x) = f (g1 x) dm(x),

implying that

dg (x) f (g1 x)
(g, x) = = .
d (x) f (x)

Setting g = (x), and using (4.20),

f (x0 )
( (x), x) = ,
f (x)

so that

d (x) = f (x0 ) dm(x). (4.24)

3. is independent of the quasi-invariant measure used to define it.

To see this, let  be another quasi-invariant measure and set

dg (x)
d = 
( (x), x) d(x), 
( (x), x) = .
d(x)

Since both and  are quasi-invariant measures, there is a Borel function f :

X R+ , f (x) > 0, for all x X, such that

d(x) = f (x) d (x),

and, hence,

dg (x) = d(g1 x) = f (g1 x) d (g1 x).

Thus,

 d(g1 x) f (g1 x)
(g, x) = = (g, x).
d(x) f (x)
(This relation has a natural meaning in the language of cohomology [Asc72,
Hil71]; namely, and  are related by a coboundary and, therefore, are
cohomologically equivalent.) This implies that


( (x), x) d(x) = f (x0 ) ( (x), x) d (x);
4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant Measures 67

that is,

 = f (x0 ) ,

and f (x0 ) is a constant.

4. The measure [ ] , formed by applying the prescription (4.19) to is a
constant multiple of . Since is a quasi-invariant measure, this assertion
is a consequence of the previous result. To prove it directly, let us write

d (g1 x)
= (g, x).
d (x)
Then,

d[ ] (x) = ( (x), x) d (x).

But,

d (g1 x) ( (g1 x, g1 x) d (g1 x)

(g, x) = = ,
d (x) ( (x), x) d (x)

implying that

( ( (x)1 x), (x)1 x)

d[ ] (x) = ( (x), x) d (x)
( (x), x)
= ( (x0 ), x0 ) d (x), by (4.20).

Since (x0 ) H is fixed, the result follows.

The above results show that, once the section is fixed, the natural quasi-
invariant measure to use on X is . In fact, this is the measure we shall use in
Sect. 7.1, to give a general definition of coherent states.

4.1.2 An Example Using the Affine Group

As an example of an explicit construction of the measure , consider the connected

affine group G+ , also called the ax + b group, consisting of transformations of R of
the type x  ax + b, x R, where a > 0, b R. Writing

g = (b, a) G+ , (4.25)

one has the multiplication law

g1 g2 = (b1 + a1 b2 , a1 a2 ). (4.26)
68 4 Some Group Theory

Consider the subgroup H of G+ ,

H := {g G+ : g = (0, a), a R+
}. (4.27)

Then, G+ /H R, since, for (b, a) G+ ,

(b, a) = (b, 1)(0, a), b R. (4.28)

Also, since, for x R,

(b, a)(x, 1) = (ax + b, 1)(0, a), (4.29)

the action of G+ on the coset space G+ /H, parametrized globally by points in R,

can be written as

gx = ax + b, g = (b, a) G+ . (4.30)

Thus, R is a homogeneous space for G+ ; however, there is no invariant measure

on R under the action (4.30). On the other hand, for any Borel function : R
R+ , (x) = 0 (for all x), the measure

d (x) = (x) dx (4.31)

is quasi-invariant. Indeed, for g = (a, b) G+ ,

!x b"1
dg (x) = d (g1 x) = dx, (4.32)
a a a
so that

dg (x) ( ax ba )
(g, x) = = . (4.33)
d (x) a (x)

A (global) section 0 : R G+ is now defined by

0 (x) = (x, 1), (4.34)

and, in view of (4.28), any other section : R G+ can be expressed in terms of

0 as

(x) = 0 (x)(0, f (x)) = (x, f (x)), (4.35)

where f : R R+
. Using such a general section, we compute the measure to be
[see (4.33)]

(0)
d (x) = ( (x), x) d (x) = dx. (4.36)
f (x)
4.2 Induced Representations and Systems of Covariance 69

Note that this also explicitly demonstrates the independence of of the quasi-
invariant measure in (4.31) with which we started. The measure does depend,
however, on , through f . In fact, choosing f (x) = (0)/ (x), we may actually
make = .

4.2 Induced Representations and Systems of Covariance

The inducing construction of Mackey [453] is a method for obtaining unitary

representations of locally compact groups starting from known representations of
subgroups. We present here a very brief account of the construction of such a
representation, mainly to set up the notation and terminology that we shall be
using, and later in Sect. 9.2.4, we shall give a more detailed account of the inducing
construction for semi-direct product type of groups. (We assume that the reader
has at least a basic idea of the representation theory of locally compact groups and
in particular the theory of induced representations. Detailed and highly readable
accounts of the latter may be found in, for example, [Kir76, Lip74], or [Var85].)
Suppose that the group G has a representation by unitary operators U(g), g G,
on a Hilbert space H. We shall consider strongly continuous, unitary representations
only. Strong continuity means that, for all H,

(U(g) I)  0 as g e.

The method of induced representations constructs such strongly continuous

representations of G, starting from strongly continuous unitary representations
of subgroups of G. The following is an outline of the procedure, adapted to our
special needs.
Let H be a closed subgroup of G and X = G/H, let be a quasi-invariant measure
on X, and (g, ) be the RadonNikodym derivative of the transformed measure
g , g G with respect to , as defined in (4.7). Fix a Borel section : X G. For
g G and x X, we may write

g (x) = (gx)h(g, x), where h(g, x) = (gx)1 g (x) H. (4.37)

Here, h : G X H is again a cocycle, with h (g, x) = [h(g1 , x)]1 satisfying

conditions similar to (4.8):

h (g1 g2 , x) = h (g1 , x)h (g2 , g1
1 x), (4.38)
h (e, x) = e,

for all g1 , g2 G and all x X (see the comment following (4.8) and Fig. 4.1).
Suppose that H has a strongly continuous unitary representation h  V (h), h H,
on a Hilbert space K. Denoting by U (K) the group of all unitary operators on K,
define the map B : G X U (K) by
70 4 Some Group Theory

H
E
E
E
H E s (gx)
E
J E
J E h(g, x)
J E
J E gs (x)
s (x) J
E
J E
J E
J
gx
J
J
J
x

Fig. 4.1 Origin of the cocycle h(g, x)

1
B(g, x) = [ (g, x)] 2 V (h(g1 , x))1 . (4.39)

It is not hard to see that B also satisfies cocycle conditions for all g1 , g2 G and
x X (again, after adjustment on null sets, if necessary):

B(g1 g2 , x) = B(g1 , x)B(g2 , g1
1 x), (4.40)
B(e, x) = IK (= identity operator on K).

Consider next the Hilbert space H = K L2 (X, d ), of functions : X K, which

are square integrable in the norm

 2H
 =  (x)2K d (x).
X


It is straightforward to verify that the operators U(g), g G, defined by

 )(x) = B(g, x) (g1 x),

(U(g) 
H, (4.41)

are unitary on H. Moreover, they define a strongly continuous representation of

G. The representation g  U(g) so constructed is called the representation of
G induced from the representation V of the subgroup H. In general, U  is not
irreducible, even when V is irreducible. The specific choice of the Borel section
and the quasi-invariant measure is inconsequential, since a different choice leads
to a unitarily equivalent representation.
4.2 Induced Representations and Systems of Covariance 71

There is an invariant formulation of the inducing construction, which avoids the

use of the section and, in fact, is the form in which the construction appears
most often in the mathematical literature. We briefly mention this construction,
since we shall be using it also in the sequel. Let us denote by [g], the element
x = gH in the coset space X = G/H, and consider the Hilbert space LK 2 (X, d ),

of all measurable functions f : G K, which satisfy the condition,

f(gh) = V (h)1 f(g), g G, h H, (4.42)

and are square integrable in the sense that


f(g)2K d ([g]) < . (4.43)
X

It is easily verified that, in view of (4.42), the integrand in (4.43) is indeed a function
over X, so that the integral is well defined, as is also the corresponding scalar
product. Now define a linear map, W : H  = K L2 (X, d ) L2 (X, d ), by
K

(W )(g) = f(g) = V (h)1 (x), x = [g] X, (4.44)

where h H is determined from g = ([g])h. (Note that, for given g G and a

fixed section , this decomposition is unique.) It is straightforward to verify that
this map is onto and, in fact, unitary. The image of the induced representation U, 
on LK (X, d ) under this unitary map, then has a simple covariant form. Indeed,
2


writing V U(g) = WU(g)W 1 , a straightforward computation, using (4.37), (4.39),

and (4.41), yields

1
(V U(g) f)(g ) = (g, x) 2 f(g1 g ), (4.45)

being the cocycle defined in (4.7).

Every induced representation has a canonically associated PV measure, which
in a sense determines the representation. To see this, consider the following PV
 defined on the Borel sets of the homogeneous space X:
measure P,

 ) )(x) = (x) (x).

(P( (4.46)

 ) transform covariantly under the action of the U(g):

The operators P( 

 P(
U(g)  = P(g
 )U(g)  ), (4.47)

where g is the transform of the set under the natural action of G on X.

 P}
A pair {U,  satisfying (4.46) is called a system of imprimitivity based on X.
Thus, associated to an induced representation, there is always a canonical system
72 4 Some Group Theory

of imprimitivity, and, as asserted in Theorem 4.2.2, this system of imprimitivity

actually determines the representation. Before coming to that, however, it will be
useful to also consider the following, more general concept:
Definition 4.2.1 (System of covariance). A system of covariance based on X is a
pair {U, a}, consisting of a unitary representation U of a locally compact group G,
and a normalized POV measure a, defined on the Borel sets of a G-space X, such
that the relation

U(g)a( )U(g) = a(g ) (4.48)

holds for all g G and B(X). In particular, if a = P is a PV measure, the

system of covariance is called a system of imprimitivity. The system {U, a} is called
a transitive system of covariance, if the action of G on X is transitive.
As mentioned earlier, a system of imprimitivity is determinative of an induced
representation. The precise sense in which this is to be understood is brought out by
the next theorem, due to Mackey [453].
Theorem 4.2.2 (Mackeys imprimitivity theorem). Let {U, P} be a transitive
system of imprimitivity, based on the homogeneous space X of the locally compact
group G. Then, there exist a closed subgroup H of G, a Hilbert space K, and
a continuous unitary representation V of H on K, such that the given system is
 P},
unitarily equivalent to the canonical system of imprimitivity {U,  arising from
 of G induced from V .
the representation U
Explicitly, the above theorem states that, if H is the Hilbert space of the
representation U, and H that of the induced representation U,
 then there exists a

unitary map W : H H, such that


WU(g)W 1 = U(g), g G,
1  ), (4.49)
W P( )W = P( B(X).

While a similar result does not hold for POV measures, as a consequence of
Namarks extension theorem (Theorem 3.1.3), it is possible to show that every tran-
sitive system of covariance can always be embedded into a system of imprimitivity.
Indeed, we have the following result [200, 483, 554]:
Theorem 4.2.3. If {U, a} is a transitive system of covariance, then U is a subrep-
resentation of an induced representation.
Proof. Let {U, a} act on the Hilbert space H. Extend a to P on the enlarged Hilbert
space H using Theorem 3.1.3, and let W and P be as in Namarks theorem [see

(3.17) and (3.18)]. To extend U(g) to a unitary operator U(g)  we proceed as
on H,
follows: Since the set of vectors

 ) : B(X), W H}
S = {P(
4.2 Induced Representations and Systems of Covariance 73

 it is enough to describe the action of U(g)

[see (3.19)] is dense H,  on this set. Define
 P(
U(g)  )WU(g)W 1 ,
 ) = P(g B(X), W H. (4.50)

Then,
 P(
U(g)  )WU(g)W 1 |P(g
 ) 2 = P(g  )WU(g)W 1  

H H

 )WU(g)W 1   ,
= WU(g)W 1 |P(g H
 ) is a projection operator
since P(g
 )WU(g)W 1  
= PWU(g)W 1 |P(g H

= WU(g)W 1 |Wa(g )U(g)W 1 H

, by (3.17)

=  |Wa( )W 1 H
, by (4.48)
 )   = P(
=  |P(  ) 2 .
H


Thus, U(g) is unitary.
Next, for ,  B(X),
 P(
U(g)  )U(g)   ) = U(g)
 P(  P(  )P(g
 1  )WU(g)W 1 , by (4.50)
 P(
= U(g)  g1  )WU(g)W 1
 g1  ))WU(g)W 1WU(g)W 1
= P(g(
  ) = P(g
= P(g  )P(
  ) .

Thus, the imprimitivity condition (4.47) is satisfied on the dense set S , and, hence,
 In other words, the pair {U,
it is satisfied on all of H.  P}
 is a system of imprimitivity
that minimally extends the system of covariance {U, a}. 

To illustrate the usefulness of this result, and to convince the reader of its
pertinence to the subject matter of this book, we work out a few examples in some
detail.

4.2.1 Vector Coherent States

Vector coherent states will be defined in their proper context in Sect. 5.1 and again,
in a geometrical setting, in Sect. 7.1.2. Here, we obtain a special class of such vectors
as an illustration of the relationship between systems of covariance and induced
representations. Let g  U(g) be a continuous unitary representation of the (locally
compact group) G on the Hilbert space H. Let K H be a subspace of dimension
n < . Let H G consist of all elements h G for which U(h)K K (i.e., K is
stable under H). If PK is the projection operator PK H = K, then

[U(h), PK ] = 0, h H. (4.51)
74 4 Some Group Theory

It is not hard to see that H is a closed subgroup of G. Let F L (K) be a strictly

positive operator, satisfying

[U(h), F] = 0, h H. (4.52)

Note that, if the restriction of the representation h  U(h) of H to K is irreducible,

then F = PK , for some > 0. In general, one can find an orthonormal basis ui
K, i = 1, 2, . . . , n, and numbers i > 0, such that
n n 1
F = i |ui ui | = | i  i |, i = i 2 ui , (4.53)
i=1 i=1

and
n
U(h)ui = V (h) ji u j , (4.54)
j=1

where the V (h) ji are the matrix elements of the n n unitary matrices V (h), h H,
realizing a unitary representation of H on Cn . Let X = G/H, and suppose that it
carries the (left) invariant measure . Let : X G be a global Borel section and
h : G X H be a cocycle defined as in (4.37). Define the set of vectors in H:

S = {i (x) = U( (x)) i : i = 1, 2, . . . , n; x X}. (4.55)

Then, by virtue of (4.52) and (4.37), the operator

n
F(x) = |i (x) i (x) | = U( (x))FU( (x)) (4.56)
i=1

has the property

U(g)F(x)U(g) = F(gx), x X, g G. (4.57)

The invariance of F under H, as follows from (4.52), implies that F(x) does not
depend on the section . Indeed, if  : X G is another section, then there is a
Borel function h : X H such that  (x) = (x)h(x), and, hence,
n
|i  (x) i  (x) | = U(  (x))FU(  (x)) = F(x).
i=1

Suppose now that the vectors S form a tight frame F {i (x) , I, n}, so that

n 
|i (x) i (x) | d (x) = I. (4.58)
i=1 X
4.2 Induced Representations and Systems of Covariance 75

[If the representation g  U(g) is irreducible, then just the convergence of the
integral on the left-hand side of the above equation to an operator A would imply
A = I, > 0. This is because (4.57) and the invariance of the measure together
imply U(g)AU(g) = A; and, hence, A commutes with every U(g), g H. Thus,
for an irreducible representation, the frame is necessarily tight.] The associated
normalized POV measure
n 
a( ) = |i (x) i (x) | d (x) (4.59)
i=1

is independent of , and, furthermore, for any g G,



U(g)a( )U(g) = F(gx) d (x), by (4.57)


= F(x) d (x), by the invariance of
g

= a(g ).

Thus, {U, a} is a transitive system of covariance. By Theorem 4.2.3, U must be a

subrepresentation of an induced representation. To work this out directly, we extend
a to the PV measure P,  on the Hilbert space H  = Cn L2 (X, d ), following the

construction of Sect. 3.2 [see (3.23)]. Thus, W : H H,

(W )i (x) = i (x) |  = i (x), i = 1, 2, . . . , n,

 Writing
is a unitary embedding of H onto a subspace H = W H of H.

U (g) = WU(g)W 1 , (4.60)

we find, for any H ,

(U (g) )i (x) = i (x) |U(g)W 1 H

= U(h(g1 , x))i (g1 x) |W 1 H , by (4.37)

n
= [V (h(g1 , x)) ]i j j (g1 x), (4.61)
j=1


by (4.54) and the unitarity of V (h), h H. Defining U(g)  by
on all of H

 )(x) = V (h(g1 , x)) (g1 x),

(U(g) (4.62)

and comparing with (4.41), we see that U  is unitary, and is in fact the representation
of G that is induced from the representation V of H on Cn . Furthermore, {U,  P},


with P as in (3.23), is a system of imprimitivity that minimally extends {U, a}.
76 4 Some Group Theory

The vectors S in (4.55) are called vector coherent states [547]. The above
analysis shows that such states, when they form a tight frame, arise from repre-
sentations U that are subrepresentations of induced representations.

4.2.2 Discrete Series Representations of SU(1,1)

Representations of the group SU(1,1), which belong to the discrete series, provide
illustrations of both the inducing construction and systems of covariance. The group
SU(1,1) consists of all 2 2 complex matrices g of the type
 

g= , , C, det g = | |2 | |2 = 1. (4.63)

The inverse element is

 

g1 = , (4.64)

while the elements k SU(1, 1) of the type

$
%
ei 2 0
k= , 0 < 4 , (4.65)
0 ei 2

constitute the maximal compact subgroup K of SU(1,1). The Cartan decomposition

of an arbitrary group element, g = pk, k K, is then obtained (see Sect. 4.5.2) as
   $
%
1z | | 0
g= = | | ,
z1 0 | |
1
z = 1 , | | = (1 |z|2 ) 2 , (4.66)

from which it is seen that the coset space SU(1, 1)/K is homeomorphic to the open
unit disc D = {z C : |z| < 1} of the complex plane. Furthermore, since
   $ 
%
1z 1 z
 |  | 0
g = | |   ,
z1 z 1 0 |  |

where

z +
 = z + , z = ,
z+
4.2 Induced Representations and Systems of Covariance 77

the action of SU(1,1) on D is captured by the fractional holomorphic transformation

z +
z  z = gz = . (4.67)
z+

The invariant measure on D, under this transformation, is now easily computed.

Indeed, we have
dz
dz = , (4.68)
( z + )2
and, hence,
dz dz
dz dz = . (4.69)
| z + |4
On the other hand, by (4.67),

1 |z|2
1 |z |2 = , (4.70)
| z + |2
so that
 2
1 1 |z |2
= . (4.71)
| z + |4 1 |z|2

Using this in (4.69), we get

dz dz dz dz

= . (4.72)
(1 |z | )
2 2 (1 |z|2 )2
Thus, the measure
1 dz dz
d0 (z, z) := (4.73)
2 i (1 |z|2 )2
on D is invariant under the action (4.67) of SU(1,1).
Note that, as a complex manifold, D admits a Khler structure determined by the
potential function

(z , z) = ln(1 z z), (4.74)

in terms of which we obtain the invariant two-form

1 2 (z, z) 1 dz dz
= dz dz = , (4.75)
i z z i (1 |z|2 )2

which is closed, i.e., d = 0, and determines the measure d0 .

78 4 Some Group Theory

Next, for each j = 1, 3/2, 2, 5/2, . . ., let us define a quasi-invariant measure on D,

d j (z, z) = exp [2 j (z, z)] d0 (z, z) = (1 |z|2 )2 j d0 (z, z)

dz dz
= (1 |z|2 )2 j2 , (4.76)
2 i
and consider the Hilbert space

H j = L2 (D, (2 j 1) d j ), (4.77)

of all complex measurable functions f on D, which are square integrable with

respect to (2 j 1) d j , i.e.,

 f 2j = (2 j 1) | f (z)|2 d j (z, z) < . (4.78)
D

(More properly, we should write f (z, z) rather than f (z) in the above, but we shall
ignore this technicality.) The factor (2 j 1) is a normalization constant that ensures
that

u0 2j = 1, where u0 (z) = 1, z D.

Indeed, using polar coordinates, in which

dz dz r dr d
= ,
2 i
we immediately find
 2  1
2j1
u0 2j = d (1 r2 )2 j2 r dr = 1.
0 0

An induced representation g  U j (g) of SU(1,1) can now be constructed, for

each j, following the procedure outlined in Sect. 4.2. Let V j be the (irreducible)
representation of the subgroup K, of dimension 1,

V j (k) = exp[i( j )], (4.79)

where the element k K is given by (4.65). We proceed to compute [see (4.39)]

1
B(g, z) = [ (g, z)] 2 V j (h(g1 , z))1 (4.80)

for an arbitrary group element g SU(1,1). Take the section : D SU(1,1),

 
1 1z
(z) = & , (4.81)
1 |z|2 z1
4.2 Induced Representations and Systems of Covariance 79

and note that, since

 
1 z + z
g1 (z) = & ,
1 |z|2 z z +

the coset decomposition (4.66) implies that

$ z+ %
1 | z+ | 0
h(g , z) = z+
. (4.82)
0 | z+ |

Hence,

( z + )2 j | z|2 j
V j (h(g1 , z)) = V j (h(g1 , z))1 = . (4.83)
| z + |2 j ( z)2 j

Next, for the quasi-invariant measure j , (4.22) and (4.76) together imply that
 2 j
d j (g1 z, g1 z) 1 |g1 z|2
(g, z) = =
d j 1 |z|2
1
= , (4.84)
| z|4 j

the last equality being obtained by taking z = g1 z in (4.71). Combining (4.80) with
(4.76) and (4.84) leads to

B(g, z) = ( z)2 j . (4.85)

Consequently, inserting (4.67), with g replaced by g1 , and (4.85) into (4.41), we

obtain the representation U j of SU(1,1) , on the Hilbert space H j , which is induced
from the representation V j of the subgroup K:
 
2 j z
(U (g) f )(z) = ( z)
j
f , f H j. (4.86)
z

Being an induced representation, U j is unitary. It is not irreducible, however, and

we shall presently construct an irreducible subrepresentation of it. First, let us note
a few general features.
The restriction of U j to the subgroup K is seen to be

(U j (k) f )(z) = ei j f (ei z) = V j (k) f (ei z), k K. (4.87)

In particular, if f (z) = zn , for some positive integer n, then

(U j (k) f )(z) = e( j+n) f (z), k K. (4.88)

80 4 Some Group Theory

Similarly, for arbitrary f H j ,

 
(1 |z|2 ) j z z
(U j ( (z)) f )(z ) = f . (4.89)
(1 z z)2 j 1 z z

Defining the projection operators P( ) on H j ,

(P( ) f )(z) = (z) f (z),

it is easily checked that

U j (g)P( )U j (g) = P(g ), g SU(1,1), (4.90)

giving a system of imprimitivity.

j
In order to isolate an irreducible subrepresentation of H j , let Hhol denote the set
of all elements f H , which are holomorphic in z. Theorem 6.2.2 then allows us to
j
j j
infer that Hhol is a Hilbert subspace of H j . Consider the function Khol : C C C,

j
Khol (z , z) = (1 z z)2 j = exp [2 j (z , z)], (4.91)

where (z , z) is the potential function defined in (4.74). Since |z | < 1, |z| < 1, the
middle term in this equation can be expanded in an infinite series to obtain

j
Khol (z , z) = un (z ) un (z), (4.92)
n=0

where the un are the monomials

 1
(2 j + n) 2
un (z) = zn , n = 0, 1, 2, . . . . (4.93)
(n + 1) (2 j)

j j
Clearly, un Hhol , for all n, and every function f Hhol can be written in terms of
them. Also, it is not hard to check that

(2 j 1) j
Khol (z , w)Khol
j
(w, z) d j (w, w) = Khol
j
(z , z), (4.94)
D

j
and we recognize here the condition for Khol to be a reproducing kernel (see
Sect. 2.3). The elements z Hhol , generated using Khol
j j j
,

zj (z ) = Khol
j
(z , z) = (1 z z)2 j , (4.95)

that is,

zj = un (z) un , (4.96)
n=0
4.2 Induced Representations and Systems of Covariance 81

enjoy the properties

'
 zj | zj  = Khol
j
(z , z)
, (4.97)
 zj | f  = f (z), f Hholj

as can easily be checked using (4.94). Furthermore, these properties also ensure that
the resolution of the identity relation

(2 j 1) | zj  zj | d j (z, z) = Ihol (4.98)
D

j
is satisfied (where Ihol is the identity operator on Hhol ), and thus the vectors zj , z
j
D, form an overcomplete set in Hhol . Using (4.89), let us define a second set of
vectors,
1 1
(z) = (2 j 1) 2 U j ( (z))u0 = (2 j 1) 2 (1 |z|2 ) j zj , (4.99)

and note that they also satisfy a resolution of the identity, but with respect to the
invariant measure 0 on D [see (4.73)]:

| (z)  (z) | d0 (z, z) = Ihol . (4.100)
D

The vectors (z) are generated by acting on the single vector u0 with the operators
U j ( (z)) of the group representation. On the other hand, being essentially the
vectors zj (i.e., up to a scale factor), the (z) also span Hhol j
. Thus, Hholj
is the
smallest closed subspace of H , containing u0 and invariant under U . Furthermore,
j j
j
as will be proved in Lemma 4.2.4, every closed subspace of Hhol that is stable under
U and different from the trivial subspace {0} must contain the vector u0 . Thus, the
j
j
representation U j restricted to Hhol is irreducible. (Actually, this is a special case of
j
a more general result [422].) Denote this restriction by Uhol . For j = 1, 3/2, 2, . . .,
these are representations in the so-called discrete series representations of SU(1,1) .
The vectors (z) , z D, are the coherent states of the discrete series representations
of SU(1,1) or CS associated to the unit disc [Per86].
For each Borel set of D, now construct the positive operator
 
a( ) = | (z)  (z) | d0 (z, z) = (2 j 1) | zj  zj | d j (z, z). (4.101)

Since, for arbitrary g SU(1,1), g (z) = (gz)h(g, z) [see (4.82)], it follows from
(4.89) and the invariance of the measure 0 that the covariance relation,
j
Uhol (g)a( )Uhol
j
(g) = a(g ), B(D), (4.102)
82 4 Some Group Theory

j
is satisfied. Thus, {Uhol , a} is a transitive system of covariance, and of course we
have here a subrepresentation of an induced representation, in conformity with
Theorem 4.2.3. Moreover, the system of imprimitivity {U j , P} in (4.90) is its
minimal Namark extension, as is evident from the discussion in Sects. 3.2 and 4.2.1.
j
Note, finally, that the projection operator Phol corresponding to the subspace Hhol
H j acts as

(Phol f )(z) = (2 j 1) j
Khol (z, w) f (w) d j (w, w) (4.103)
D

[this is a straightforward consequence of (4.94)], and, hence, a( ) and P( ) are

related through

Phol P( )Phol = a( ), B(D), (4.104)

j
as required by (3.17). (Note: in the context of that equation, Hhol  and
= H, H j = H,
j
W is just the inclusion map Hhol  H .)
j
j
We end this section by proving a technical result used to show that Uhol is
irreducible [Sug90].
j
Lemma 4.2.4. Let H0 Hhol be a nontrivial, closed subspace that is stable under
j
Uhol . Then, u0 H0 .
Proof. Let f H0 , and write f (z) = n
n=0 an z . By (4.88),

j
(Uhol (k) f )(z) = ei j an en zn , k K.
n=0

Thus,
 2
1
ei j (U j (k) f )(z) d = a0 = f (0), z D.
2 0

j j
Since H0 is stable under Uhol , it follows that Uhol (k) f H0 . Also,
 2
1
ei j Uhol
j
(k) f d H0 ,
2 0

since H0 is closed and the integral is the limit of Riemann sums of the type
$ m
%
m+1 m ei 2 0
e i j m j
Uhol (km ) f
2
, where km =
i
0 e 2
m ,
m

with each U j (km ) f H0 . Thus, the function f (0)u0 belongs to H0 , so that u0 H0 ,

provided f (0) = 0. We still need to show, however, that it is possible to find an
4.2 Induced Representations and Systems of Covariance 83

f H0 with a0 = f (0) = 0. Since H0 = {0}, one can find an element f H0 and

a point z0 D such that f(z0 ) = 0. Moreover, the action of SU(1,1) on D being
transitive, there is a g SU(1,1) for which g1 0 = z0 . Hence, by (4.86),
j
(Uhol (g) f)(0) = 2 j f(z0 ) = 0,

since = 0. Because U j (g) f is again an element in H0 , the desired result follows

j
upon writing f = Uhol (g) f. 


4.2.3 The Regular Representations of a Group

There are two representations of a locally compact group G, both induced represen-
tations, which are of great importance in harmonic analysis and will reappear later in
our development of the theory of CS. These are the so-called regular representations
of G. Let be the left Haar measure on G, and consider the trivial subgroup
H = {e}, consisting of just the identity element. The representation of G induced by
the trivial representation of H is carried by the Hilbert space L2 (G, d ). Denoting
this representation by U , we have, for all f L2 (G, d ),

(U (g) f )(g ) = f (g1 g ), g, g G. (4.105)

This representation is called the left regular representation of G. Similarly, using the
right Haar measure r and the Hilbert space L2 (G, dr ), we can construct another
unitary representation Ur as

(Ur (g) f )(g ) = f (g g), g, g G, (4.106)

for all f L2 (G, dr ). This representation is called the right regular representation
of G. In general, these representations are reducible. On the other hand, U and Ur
are unitarily equivalent representations. Indeed, the map

V : L2 (G, d ) L2 (G, dr ), (V f )(g) = f (g1 ), g G, (4.107)

is easily seen [using (4.5)] to be unitary, and

VU (g)V 1 = Ur (g), g G. (4.108)

The regular representation Ur can also be realized on the Hilbert space L2 (G, d ),
rather than on L2 (G, dr ), using the fact that and r are related by the modular
function  through (4.3). Thus, the map
1
W : L2 (G, dr ) L2 (G, d ), (W f )(g) = (g) 2 f (g) (4.109)
84 4 Some Group Theory

is unitary, and, for all f L2 (G, d ),

r (g) f )(g ) = (g) 2 f (g g),

(U
1
r (g) = WUr (g)W 1 ,
where U g G. (4.110)

From this, we see that the left and right regular representations commute:
r (g2 )] = 0
[U (g1 ), U g1 , g2 G. (4.111)

Clearly, the two representations U and Ur on L2 (G, d ) are also unitarily equiva-
lent. More interesting, however, is the map J : L2 (G, d ) L2 (G, d ),
1
(J f )(g) = f (g1 )(g) 2 , J2 = I
r (g),
JU (g)J = U g G, (4.112)

which is an antiunitary isomorphism.

4.3 An Extended Schurs Lemma

In harmonic analysis, the irreducibility of a unitary group representation is usually

determined by an application of Schurs lemma (see, for example, [Kir76]).
We discuss this lemma in some detail here, since we shall require an extended
version of it when dealing with square integrable representations in Chap. 8.
Lemma 4.3.1 (Classical Schurs lemma). Let U be a continuous unitary irre-
ducible representation of G on the Hilbert space H. If T L (H), and T commutes
with U(g), for all g G, then T = I, for some C.
A general proof of this theorem may, for example, be found in [Sug90]. Actually,
there is a more general version of this lemma, which we now state and prove
and then use to prove a further extended version of it. Let U1 and U2 be two
representations of G on the Hilbert spaces H1 and H2 , respectively. A bounded linear
map T : H1 H2 is said to intertwine U1 and U2 if

TU1 (g) = U2 (g)T, g G. (4.113)

Given two Hilbert spaces H1 and H2 , a linear map T : H1 H2 is said to be a

multiple of an isometry if there exists > 0 such that

T 2H2 =  2H1 , H1 . (4.114)

Lemma 4.3.2 (Generalized Schurs lemma). Let U1 be a unitary irreducible

representation of G on H1 and U2 be a unitary, but not necessarily irreducible,
representation of G on H2 . Let T : H1 H2 be a bounded linear map that
intertwines U1 and U2 . Then, T is either null or a multiple of an isometry.
4.3 An Extended Schurs Lemma 85

Proof. Consider the adjoint map T : H2 H1 , and take the adjoint of the relation
(4.113), getting

U1 (g)T = T U2 (g), g G, (4.115)

since U j (g) = U j (g1 ). From this and (4.113), we obtain

T TU1 (g) = T U2 (g)T = U1 (g)T T, g G. (4.116)

Now, U1 is unitary and irreducible, and T T : H1 H1 commutes with it. Hence,

by the classical Schurs lemma, T T = I, for some C; that is, T is either null
or a multiple of an isometry. 

The next extended version of Schurs lemma, which we now state and prove, has
been adapted from [351]. This is the version that we shall eventually require.
Lemma 4.3.3 (Extended Schurs lemma). Let U1 be a unitary irreducible
representation of G on H1 and U2 be a unitary, but not necessarily irreducible,
representation of G on H2 . Let T : H1 H2 be a closed linear map, the domain
D(T ) of which is dense in H1 and stable under U1 [i.e., U1 (g) D(T ), for all
g G and D(T )], and suppose that T intertwines U1 and U2 . Then, T is either
null or a multiple of an isometry.
Proof. (The idea of this proof is to convert D(T ) into a Hilbert space by equipping
it with the graph norm of T , and then to use the generalized Schurs lemma above.)
On D(T ), define the new scalar product

 | T =  | H1 + T |T H2 ,  2T =  2H1 + T 2H2 , (4.117)

for all , , D(T ). Then, since T is closed, D(T ) equipped with this scalar product
is a Hilbert space, which we denote by HT . Consider T as a linear map T : HT H2 .
Then,

T H2 T H2
T  = sup = sup .
HT  T HT [ 2 + T 2 ] 12
H1 H2

Thus, T  1, so that T : HT H2 is a bounded linear map. Next, it is seen that

the representation U1 is unitary on HT . Indeed, for all HT and g G,

U1 (g) 2T = U1 (g) 2H1 + TU1 (g) 2H2 =  2H1 + U2 (g)T 2H2

=  2H1 + T 2H2 =  2T .

Also, for all g G, the restriction of U1 (g) to HT is surjective (i.e., onto HT ), so

that, for all HT , U1 (g) = implies HT . This follows from the stability
of D(T ) under U1 (g) and the fact that U1 (g1 ) = U1 (g1 )U1 (g) = . Thus, by
86 4 Some Group Theory

Lemma 4.3.2, T : HT H2 is either a multiple of an isometry, in which case there

exists a > 0 such that, for all HT ,

T 2H2 =  2T =  2H1 + T 2H2 ,

or else T is null. In any case, = 1, and

T 2H2 =  2H1 , 0.
1

Thus, T : D(T ) H2 is either null and, hence, null on all of H1 , or else it is a

multiple of an isometry, in which case, by virtue of the closedness of T , it extends
to a multiple of the isometry from H1 to H2 . 

As a corollary, if H1 = H2 and U1 = U2 , then, as a consequence of the classical
Schurs lemma, T is a multiple of the identity.

4.4 Harmonic Analysis on Locally Compact Abelian Groups

4.4.1 Basic Notions

We give in this section a summary of some basic concepts and useful results on
locally compact abelian (LCA for short) groups. Further information may be found
in [Rud62] or any textbook on harmonic analysis (for instance, [Fol95]).
Given an LCA group G (whose group operation shall be written additively),
we denote by G the Haar measure on G. When G is discrete, G is naturally
the counting measure, and discrete versions of relations that we shall define on the
whole group G will then hold.
 be the unitary dual of G, that is, the set of all its unitary characters. We use
Let G
the notation

(x) =  , x, 
x G, G,

to express the duality pairing between G and G.  We again denote the group law on
G additively: For any ,  G,
 we write  +  , x =  , x  , x, x G. On the
dual group G, we take the Haar measure to be the dual of the given Haar measure
on G.
The Fourier transform is defined as usual: Given f L1 (G), its Fourier transform
is the function F f = f L (G) defined by

f( ) = f (x) , x d (x) . (4.118)
G
4.4 Harmonic Analysis on Locally Compact Abelian Groups 87

It is a standard result (see, e.g., [Rud62] for more details) that the Fourier transform
so defined extends to an isometry between L2 (G) and L2 (G).  More precisely, there

exists a measure G on G (the dual Haar measure) such that the following Plancherel
formula holds: For all f L2 (G), one has f L2 (G) and
 2

f( ) dG ( ) = | f (x)|2 d (x). (4.119)

G G

This Fourier transform has the usual properties with respect to the operations of
translation and modulation. Here, the translation of parameter b G is the unitary
operator Tb defined on L2 (G) by

(Tb f )(x) = f (x b),

and the modulation of parameter b G is the unitary operator Eb defined on

 by
L2 (G)

(Eb f)( ) =  , b f( ) .

T E
The maps b  Tb and b  Eb provide two unitarily equivalent representations of
 in fact, the regular representa-
G, with respective carrier spaces L2 (G) and L2 (G)),
tion of G. The associated intertwining operator is the Fourier transform. Explicitly,
we have

F Tb = Eb F for all b G.

If f and g are two functions over G, their convolution product is defined as


( f g)(x) = f (x y)g(y) d (y), (4.120)
G

if the integral on the right-hand side converges. The convolution product is

associative and, since G is abelian, it is also commutative; that is,

( f g) h = f (g h) , f g = g f.

In the Fourier space, the convolution is given by pointwise multiplication:

f(
g = f g.

The standard example, of course, is G = SO(2) S1 T, the unit circle, with Haar
measure d /2 . The unitary characters of G are the trigonometric functions

m ,  = eim , m Z,
88 4 Some Group Theory

 = Z and the harmonic analysis on SO(2) is the theory of Fourier series.

so that G
Note that part of the theory extends to nonabelian groups. If G is compact, G  is
discrete, and harmonic analysis leads to expansions in appropriate special functions.
For instance, in the case of SO(3), one gets the theory of spherical harmonics, which
indeed yields Fourier expansions on the two-sphere S2 .

4.4.2 Lattices in LCA Groups

 be its annihilator, defined as

Let G be any subgroup of G, and let G

 :  , x = 0 for all x }.
= { G (4.121)

is a closed subgroup of G.  We always have ( ) , and the reverse inclusion

is true if is itself a closed subgroup of G.
There is a remarkable duality between subgroups and quotient groups of a locally
compact abelian group, as stated in the following theorem, which extends the Pon-
trjagin duality theorem (see [Rud62, Theorem 2.1.2] or [Fol95, Theorem 4.39]).

Theorem 4.4.1 (Pontrjagin duality). Let be a closed subgroup of G, and define

(
: G/ and  
: G/

by ( ) = q, ( ) = | , where q : G G/ is the canonical

projection. Then, and are isomorphisms of topological groups.
Note that the quotient groups G/ and G/  are both locally compact abelian
groups. The main idea of Theorem 4.4.1 is that, on the one hand, the characters of
G/ can be identified with the subset of G  and, on the other hand,  can be

identified with the quotient G/ . Note also that the surjectivity of yields a sort
of HahnBanach theorem for locally compact abelian groups, in the sense that every
character on extends to a character of G.
Let us now suppose that is a lattice in G; that is, is a discrete subgroup of
G and the quotient G/ is compact. By a fundamental domain of in G, we mean
a -measurable set G, such that, for each x G, (x + ) consists of a
single point. Given a lattice G, it can be shown that a fundamental domain
of always exists in G. It is standard to take = G/ , but many other choices are
possible. The lattice size, which we denote by s( ), is defined as the measure of a
fundamental domain, that is, s( ) = ( ), and it is independent of the particular
choice of . The quantity s( )1 then serves as a measure of the density of .
Taking = G/ allows us to take the measure naturally inherited from G as the
measure of G/ ; we normalize it to be a probability measure such that the Weyl
formula (see, e.g., [Fol95] for details) holds:
4.4 Harmonic Analysis on Locally Compact Abelian Groups 89

 

G
f (x) d (x) =
G/
f (x + ) dG/ (x), f L1 (G). (4.122)

If is a lattice in G, then its annihilator is a lattice in G  and s( )s( ) = 1

(see, for instance, [343, Lemma 6.2.3]). If i ( i = 1, 2) are two lattices in G, such
that 2 1 , then in the dual space we have 1 2 ; furthermore, in all of the
practical situations we know, it is always possible to choose a fundamental domain
i of i , i = 1, 2, in such a way that, as subsets of G, they satisfy 1 2 .
The whole discussion on the existence and choices of fundamental domains may
be repeated in the dual space. In particular, we have the following version of the
Weyl formula:
 


G
f ( ) dG ( ) =

G/
f ( + ) dG/ ( ), 
f L1 (G). (4.123)

The canonical example is G = R. The unitary characters of R are given by the

pure oscillations

 ,t = ei t .

Each character of R can, therefore, be identified with a point R that is

interpreted as a frequency, and R is thus identified with R. We take the Lebesgue
measure dx as the Haar measure on R; if d denotes the Lebesgue measure on R, 
then the dual measure of dx is d /2 .
Let = Z; then = 2 Z. By Theorem 4.4.1, we have Z  R/Z
 ; the dual
of Z can, therefore, be identified with the unit circle T = R/2 Z. A fundamental
domain of Z in R  can be chosen to be = [ , ), and then s(Z ) = 1 = s(Z) .
The Haar measure on Z is the counting measure, and its dual measure on T is then
d /2 .

4.4.3 Sampling in LCA Groups

A natural map from functions defined over the whole group G to functions over the
lattice is given by the sampling operator (perfect sampling in the terminology
of [Hol95]) associated to , defined as

f = f | ,

where f is a function with suitable properties (i.e., for which point values make
sense).
A function f over G is said to be -periodic if, for all x G,

f (x + ) = f (x), for all .

90 4 Some Group Theory

Such a function can be identified with a function over G/ in the usual way. The
periodization operator P : L1 (G) L1 (G/ ) is defined as

(P f )(x) = f (x + ) . (4.124)

By the Weyl formula (4.122), we have

 | f L2 (G) =  | P f L2 (G/ ) , for any -periodic function . (4.125)

In other words, the periodization operator allows us to write the scalar product of a
-periodic function with an arbitrary function f over G as a scalar product over
G/ . In the particular case in which G = R and = Z, the Fourier transform of a
sequence is a Z -periodic function, i.e., a 2 -periodic function.
We close this section by giving the useful Poisson summation formula, which
links sampling to periodization via the Fourier transform. The proof can be found in
most treatises on harmonic analysis (see [343, Fol95], and also [Hol95]) .
Theorem 4.4.2 (Poisson summation formula). Let G be a lattice. Assume
 Then the following is true.
that f L1 (G) F 1 L1 (G).
1. The -periodized (x) = f (x + ) of f is in L1 (G/ ) (where x = x + )
(
and for G/ ,
1 
( ) = f ( ). (4.126)
s( )
2. We have
1
f (x + ) =
s( )
f( ) , x a.e. (4.127)

f( + ) = s( ) f (x) , x a.e. (4.128)

x

2
3. If, in addition, | f( ) | < , then L2 (G/ ) and the relations (4.127)
and (4.128) hold in the L2 - sense.
For the canonical example, G = R, = Z, one recovers, of course, the usual
Poisson summation formula (which will be used in Sect. 6.3 and in Sect. 7.4),
familiar in signal processing [Lyn82, Pap02] and in harmonic analysis [Kah98].

4.5 Lie Groups and Lie Algebras: A Reminder

The theory of Lie groups, Lie algebras, and their representations is widely known
and many excellent books cover it, for instance [Hel78, Jac62], or [Kna96].
Nevertheless, it is useful to recall here a few basic facts that will be used later in
this book, and at the same time fix our notation.
4.5 Lie Groups and Lie Algebras: A Reminder 91

4.5.1 Lie Algebras

Let g be a complex Lie algebra, that is, a complex vector space with an
antisymmetric bracket [., .] that satisfies the Jacobi identity

[[X,Y ], Z] + [[Y, Z], X] + [[Z, X],Y ] = 0, X,Y, Z g. (4.129)

For X,Y g, the relation (adX)(Y ) = [X,Y ] gives a linear map ad: g End g
(endomorphisms of g), called the adjoint representation of g. Next, if dim g < , it
makes sense to define

B(X,Y ) = Tr[(adX)(adY )], X,Y g. (4.130)

B is a symmetric bilinear form on g, called the Killing form of g. Alternatively,

one may choose a basis {X j , j = 1, . . . , n} in g, in terms of which the commutation
relations read
n
[Xi , X j ] = ckij Xk , i, j = 1, . . . , n, (4.131)
k=1

where ckij are called the structure constants and n = dim g. Then it is easy to see
that gi j = nk,m=1 cm
ik c jm = B(Xi , X j ) defines a metric on g, called the CartanKilling
k

metric.
The Lie algebra g is said to be simple (respectively, semisimple) if it does not
contain any nontrivial ideal (respectively, abelian ideal). A semisimple Lie algebra
may be decomposed into a direct sum of simple ones. Furthermore, g is semisimple
iff the Killing form is nondegenerate (Cartans criterion).
Let g be semisimple. Choose in g a Cartan subalgebra h, i.e., a maximal nilpotent
subalgebra (it is in fact maximal abelian and unique up to conjugation). The
dimension  of h is called the rank of g. A root of g with respect to h is a linear form
on h, h , for which there exists X = 0 in g such that (adH)X = (H)X, H h.
Then, according to the fundamental work of Cartan and Chevalley, one can find
a basis {Hi , E } of g, with the following properties. {H j , j = 1, . . . , } is a basis of
h, and each generator E is indexed by a nonzero root , in such a way that the
commutation relations (4.131) may be written in the following form:

[Hi , H j ] = 0, j = 1, . . . , ,
[Hi , E ] = (Hi )E , i = 1, . . . , , (4.132)

[E , E ] = H i Hi h,
i=1

[E , E ] = N E + ,
92 4 Some Group Theory

where N = 0 if + is not a root. Let denote the set of roots of g. Note that the
nonzero roots come in pairs , and no other nonzero multiple of a
root is a root. Accordingly, the set of nonzero roots may be split into a subset + of
positive roots and the corresponding subset of negative roots = { , + }.
The set + is contained in a simplex (convex pyramid) in h , the edges of which are
the so-called simple positive roots, i.e., positive roots that cannot be decomposed as
the sum of two other positive roots [61, Jac62, Kna96]. Of course, the same holds
for . The consideration of root systems is the basis of the Cartan classification
of simple Lie algebras into four infinite series A , B , C , D and five exceptional
algebras G2 , F4 , E6 , E7 , E8 [Jac62].
In addition to the Lie algebra g, one may also consider the universal enveloping
algebra U(g), which consists of all polynomials in the elements of g, modulo
the commutation relations (4.131). Of special interest are the so-called Casimir
elements, which generate the center of U(g), and in particular, the quadratic Casimir
element C2 = ni,k=1 gik X i X k , where {X i } is the basis of g dual to {Xi }, i,e,
B(X i , X j ) = ji and gik is the CartanKilling metric. The element C2 does not depend
on the choice of the basis {Xi }. In a CartanChevalley basis {H j , E }, one gets


C2 = (H j )2 + E E . (4.133)
j=1

In the same way, one defines a real Lie algebra as a real vector space with an
antisymmetric bracket [., .] that satisfies the Jacobi identity. The two concepts are
closely related. If g is a real Lie algebra, one can define its complexification gc by
complexifying it as a vector space and extending the Lie bracket by linearity. If dim
g = n, the complex dimension of gc is still n, but its real dimension is 2n. Conversely,
if g is a complex Lie algebra, and one restricts its parameter space to real numbers,
one obtains a real form gr , i.e., a real Lie algebra whose complexification is again g.
A given complex Lie algebra has in general several nonisomorphic real forms (also
classified by Cartan), among them a unique compact one, characterized by the fact
that the CartanKilling metric gi j is negative definite. For instance, the complex Lie
algebra A2 yields two real forms, su(3), the compact one, and su(2, 1).
For physical applications, it is not so much the Lie algebras themselves that
matter, but their representations in Hilbert spaces. The key ingredient for building
the latter, also due to Cartan, is the notion of a weight vector. Let T be a
representation of the Lie algebra g in the Hilbert space H, that is, a linear map
of g into the operators on H, such that

T ([X,Y ]) = [T (X), T (Y )], X,Y g. (4.134)

The representation T is called Hermitian if T (X ) = T (X) , for every X g.

By Schurs lemma, it follows that the Casimir elements of U(g) can be simulta-
neously diagonalized in any irreducible Hermitian representation of g, and in fact
their eigenvalues characterize the representation uniquely.
4.5 Lie Groups and Lie Algebras: A Reminder 93

Let {H j , E } be a CartanChevalley basis of the complexified Lie algebra gc of

g. Then a weight for T is a linear form h , for which a nonzero vector |  H
exists, such that [T (H) (H)]n |  = 0, for all H h and some n. Then |  is
called a weight vector if n = 1, that is, T (H)|  = (H)| , H h. In particular,
|  is called a highest (respectively lowest) weight vector if

T (E+ )|  = 0, + + , (resp. T (E )|  = 0, ), (4.135)

where + , respectively , is the set of the positive, resp. negative, roots. The
interest of this notion is the fundamental result of Cartan, which says that the
irreducible, finite-dimensional, Hermitian representations of simple Lie algebras
are in one-to-one correspondence with highest weight vectors (see [Hum72, Jac62],
and [61] for the geometrical construction of representations in those terms).
We illustrate this on the simplest example, namely su(2), with the familiar
angular momentum basis {Jo , J+ , J } and = {J }. (Here there is an implicit
identification between the Lie algebra and its dual.) In the unitary spin- j irreducible
representation D( j) of SU(2), of dimension 2 j + 1, with standard basis {| j, m | m =
j, . . . , j}, the highest (respectively, lowest) weight vector | j, j (respectively,
| j, j) satisfies the relation J+ | j, j = 0 (respectively, J | j, j = 0) and has an
isotropy subalgebra u(1) (by this, we mean the subalgebra that annihilates the given
vector). Notice that here C2 = Jo2 + 12 (J+ J + J J+ ) is indeed diagonal in D( j) , with
eigenvalue j( j + 1).

4.5.2 Lie Groups

A Lie group G may be defined in several ways [Hel78, Kna96], for instance, as a
smooth manifold with a group structure, such that the group operations (g1 , g2 ) 
g1 g2 , g  g1 are Ck , for some k 2. This actually implies that all group operations
are in fact (real) analytic. A Lie group is said to be simple, respectively, semisimple,
if it does not have any nontrivial invariant subgroup, respectively, abelian invariant
subgroup.
Let G be a Lie group. For g G, consider the map Lg : G G with Lg (g ) =
gg . The derivative of this map at g G, denoted Tg (Lg ), sets up an isomorphism


Tg (Lg ) : Tg G Tgg G between the tangent spaces at g and gg . For any vector
X Te G (the tangent space at the identity), let us define a vector field X on G by
Xg = Te (Lg )X. Such a vector field is said to be left invariant. It can be demonstrated
that the usual Lie bracket [X,  Y ] of two left invariant vector fields is again a left
invariant vector field. Using this fact, we see that, if Yg = Te (Lg )Y, Y Te G, the
 Y ] defines a Lie bracket on the tangent space, Te G, at the identity
relation [X,Y ] = [X,
of the Lie group G. Thus, equipped with this bracket relation, Te G becomes a Lie
algebra that we denote by g.
Next, it can be shown that, for any X g, there exists a unique analytic
homomorphism, t  X (t) of R into G such that
94 4 Some Group Theory

d
X (0) = (t)|t=0 = X,
dt X

and, for each X g, we then write exp X = X (1). The map X  exp X is called the
exponential map, and it defines a homeomorphism between an open neighborhood
N0 of the origin 0 g and an open neighborhood Ne of the identity element e of G.
Each X (t) defines a one-parameter subgroup of G, with infinitesimal generator X,

X (t) = exp(tX),

and every one-parameter subgroup is obtained in this way. If G is a matrix group,

the elements X of the Lie algebra are also matrices and the exponential map comes
out in terms of matrix exponentials.
Using this tool, the fundamental theorems of Lie may then be sketched as
follows.
1. Every Lie group G has a unique Lie algebra g, obtained as the vector space
of infinitesimal generators of all one parameter subgroups, in other words, the
tangent space Te G, at the identity element of G.
2. Given a Lie algebra g, one may associate to it in a unique way, by the exponential
map X  exp X, a connected and simply connected Lie group G, with Lie algebra
g (G is called the universal covering of G). Any other connected Lie group G with
the same Lie algebra g is of the form G = G/D, where D is an invariant discrete
subgroup of G.
Furthermore, a Lie group G is simple, respectively, semisimple, if and only if its Lie
algebra g is simple, respectively, semisimple.
Here again, one may start with a real Lie group G, build its complexification Gc ,
and find all real forms of Gc . One, and only one, of them is compact (it corresponds,
of course, to the compact real form of the Lie algebra). For instance, the complex Lie
group SL(2, C) has two real forms, SU(2) and SU(1, 1), the former being compact.
A Lie group has natural actions on its Lie algebra and its dual. These are the
adjoint and coadjoint actions, respectively, and may be understood in terms of the
exponential map. For g G, g  gg g1 defines a differentiable map from G to
itself. The derivative of this map at g = e is an invertible linear transformation,
Adg , of Te G (or, equivalently, of g) onto itself, giving the adjoint action. Thus, for
t ( , ), for some > 0, such that exp(tX) G and X g,

d
Y= [g exp(tX) g1 ]|t=0 := Adg (X) (4.136)
dt
is a tangent vector in Te G = g. If G is a matrix group, the adjoint action is simply

Adg (X) = gXg1 . (4.137)

4.5 Lie Groups and Lie Algebras: A Reminder 95

Now considering g  Adg as a function on G with values in End g, its derivative

at the identity, g = e, defines a linear map ad : g End g. Thus, if g = exp X, then
Adg = exp(adX), and it can be verified that (adX)(Y ) = [X,Y ].
The corresponding coadjoint actions are now obtained by dualization: The
coadjoint action Ad#g of g G on the dual g , of the Lie algebra, is given by

< Ad#g (X ) ; X > = < X ; Adg1 (X) >, X g , X g, (4.138)

where < ; > < ; >g ,g denotes the dual pairing between g and g. Once again,
the (negative of the) derivative of the map g  Ad#g at g = e is a linear transformation
ad# : g End g , such that, for any X g, (ad# )(X) is the map,

< (ad# )(X)(X ) ; Y > = < X ; (ad)(X)(Y ) >, X g , Y g. (4.139)

Clearly, Ad#g = exp(ad# X). If we introduce a basis in g and represent Adg by a

matrix in this basis, then, in terms of the dual basis in g , Ad#g is represented by the
transposed inverse of this matrix.
Under the coadjoint action, the vector space g splits into a union of disjoint
coadjoint orbits

OX = {Ad#g (X ) : g G}. (4.140)

According to the KirillovSouriauKostant theory (see [Gui84] or [Woo92]), each

coadjoint orbit carries a natural symplectic structure. This arises as follows: Let
Y be an arbitrary point on the coadjoint orbit OX and t  s(t), t ( , ), be a
smooth curve in OX passing through Y . We take s(t) to be of the form

s(t) = Ad#g(t)Y , g(t) = exp(tX), X g.

(Note that s(0) = Y .) Differentiating with respect to t,

s(0) = (ad# )(X)(Y ) := vX ,

is a tangent vector to the orbit at the point Y . Let

gY = {X g : (ad# X)(Y ) = 0}, (4.141)

be the annihilator of Y . Then, denoting elements in g/gY by X gY , X g, it is

clear that the map

iY : g/gY TY OX , iY (X gY ) = vX = (ad# X)(Y ),

is a vector space isomorphism. This enables us to identify the tangent space TY OX

to the orbit OX at the point Y , with g/gY , for all Y OX . For any Y OX ,
consider now the antisymmetric bilinear functional Y , defined on the tangent
space TY OX ,

Y (vX , vY ) = Y ; [X,Y ]. (4.142)

96 4 Some Group Theory

Since

(ad# )(X)(Y ) ; Y  = Y ; [X,Y ],

the expression in (4.142) is well defined and nondegenerate on TY OX . Thus, we

have defined an antisymmetric bilinear form on the coadjoint orbit OX , given
at any point Y OX by (4.142). This form is nondegenerate and closed, i.e., its
exterior derivative, d = 0, which follows from the Jacobi identity (4.129) satisfied
by the elements of the Lie algebra g. Hence, defines a symplectic structure on
the manifold OX . In addition, is G-invariant. This implies, in particular, that the
orbit is of even dimension and carries a natural G-invariant (Liouville) measure.
Therefore, a coadjoint orbit is a natural candidate for realizing the phase space of a
classical system and, hence, a starting point for a quantization procedure. Explicit
examples of coadjoint orbits for semidirect product groups will be worked out
in detail in Sect. 9.2, where they will turn out to be natural parameter spaces for
building coherent states.
Semisimple Lie groups have several interesting decompositions. In the sequel,
we are going to use three of them, the Cartan, the Iwasawa, and the Gauss
decompositions.
1. Cartan decomposition
This is the simplest case. Given a semisimple Lie group G, its real Lie algebra
g always possesses a Cartan involution, that is, an automorphism : g g, with
square equal to the identity:

[X,Y ] = [ (X), (Y )], X,Y g, 2 = I, (4.143)

and such that the symmetric bilinear form B (X,Y ) = B(X, Y ) is positive
definite, where B is the CartanKilling form. Then, the Cartan involution yields
an eigenspace decomposition

g = kp (4.144)

of g into +1 and 1 eigenspaces. It follows that

[k, k] k, [k, p] p, [p, p] k. (4.145)

Assume for simplicity that the center of G is finite, and let K denote the analytic
subgroup of G with Lie algebra k. Then [Kna96], (i) K is closed and maximal
compact; (ii) there exists a Lie group automorphism of G, with differential ,
such that 2 = I and the subgroup fixed by is K; and (iii) the mapping K p G
given by (k, X)  k exp X is a diffeomorphism onto. One may as well write the
diffeomorphism as (k, X)  exp X k p k, and in that form we recognize the Cartan
decomposition given explicitly in (4.66) for SU(1, 1).
4.5 Lie Groups and Lie Algebras: A Reminder 97

2. Iwasawa decomposition
Any connected semisimple Lie group G has an Iwasawa decomposition into three
closed subgroups, namely G = KAN, where K is a maximal compact subgroup, A
is abelian, and N is nilpotent, and the last two are simply connected [Kna96]. This
means that every element g G admits a unique factorization g = kan, k K, a
A, n N, and the multiplication map K A N G given by (k, a, n)  kan is a
diffeomorphism onto.
Assume that G has a finite center. Let M be the centralizer of A in K, that is,
M = {k K : ka = ak, a A} (if the center of G is not finite, the definition of
M is slightly more involved). Then, P = MAN is a closed subgroup of G, called
the minimal parabolic subgroup. The interest of this subgroup is that the quotient
manifold X = G/P K/M carries the unitary irreducible representations of the
principal series of G (which are induced representations), in the sense that these
representations are realized in the Hilbert space L2 (X, d ), where is the natural
G-invariant measure. To give a concrete example, take G = SOo (3, 1), the Lorentz
group. Then, the Iwasawa decomposition reads as

SOo (3, 1) = SO(3) A N, (4.146)

where A SOo (1, 1) R is the subgroup of Lorentz boosts in the z-direction and
N C has dimension two and is abelian. In this case, M = SO(2), the subgroup of
rotations around the z-axis, so that X = G/P SO(3)/SO(2) S2 , the two-sphere.
This example is the geometrical framework for the construction of wavelets on the
sphere S2 , which we shall describe in Sect. 15.1.
A closely related decomposition is the so-called KAK decomposition [Kna96].
Again, let G be a semisimple Lie group with a finite center, K be a maximal
compact subgroup, and G = KAN be the corresponding Iwasawa decomposition.
Then, every element in G has a decomposition as k1 ak2 with k1 , k2 K and a A.
This decomposition is in general not unique, but a is unique up to conjugation.
A familiar example of a KAK decomposition is the expression of a general rotation
SO(3) as the product of three rotations, parametrized by the Euler angles
= m( ) u( ) m( ), (4.147)

where m and u denote rotations around the z-axis and the y-axis, respectively.
3. Gauss decomposition
Again, let G be a semisimple Lie group and Gc = exp gc be the corresponding
complexified group. If b is a subalgebra of gc , we call it maximal (in the sense of
Perelomov [Per86]) if b b = gc , where b is the conjugate of b in gc . Let {H j , E }
be a CartanChevalley basis of the complexified Lie algebra gc .
The complex group Gc possesses remarkable subgroups, namely,
1. H c , the Cartan subgroup, which is generated by {H j }.
2. B , called the Borel subgroups, which are maximal connected, solvable, sub-
groups, corresponding to the subalgebras b , generated by {H j , E : }; if
b is maximal , then b+ = b and b = b generate Borel subgroups.
98 4 Some Group Theory

3. Z , the connected nilpotent subgroups generated by {E : }.

The interest of these subgroups is that almost all elements of Gc admit a Gauss
decomposition:

g = z+ hz = b+ z = z+ b , z Z , h H c, b B . (4.148)

It follows that the quotients X+ = Gc /B and X = B+ \Gc are compact complex

homogeneous manifolds, on which Gc acts by holomorphic transformations. This is
the crucial fact behind the geometry of some CS, of group-theoretical origin, as first
pointed out in [493].

4.5.3 Extensions of Lie Algebras and Lie Groups

Exactly as Mackey induction is a method for building representations of a group

from representations of a subgroup, it is useful to have a method for constructing a
group G from two smaller ones, one of them at least becoming a closed subgroup of
G. Several possibilities are available, of increasing sophistication.
1. Direct product
This is the most trivial solution, which amounts to glue the two groups together
without interaction. Given two (topological or Lie) groups G1 , G2 , their direct
product G = G1 G2 is simply their Cartesian product, endowed with the group law:

(g1 , g2 )(g1 , g2 ) = (g1 g1 , g2 g2 ), g1 , g1 G1 , g2 , g2 G2 . (4.149)

With the obvious identifications g1 (g1 , e2 ), g2 (e1 , g2 ), where e j denotes the

neutral element of G j , j = 1, 2, it is clear that both G1 and G2 are invariant subgroups
of G1 G2 . In the case of Lie groups, the notion of direct product corresponds to
that of direct sum of the corresponding Lie algebras, g = g1 g2 , and again both g1
and g2 are ideals of g.
2. Semidirect product
A more interesting construction arises when one of the groups, say, G2 , acts on
the other one, G1 , by automorphisms. More precisely, one has a homomorphism
from G2 into the group Aut G1 of automorphisms of G1 . Although the general
definition may be given as in the first case, we will consider in the sequel only the
case in which G1 V is abelian, in fact, a vector space (hence, group operations are
noted additively), and G2 S is a subgroup of AutV . Then, we define the semidirect
product G = V  S as the Cartesian product, endowed with the group law:

(v, s)(v , s ) = (v + s (v ), ss ), v, v V, s, s S. (4.150)

According to the law (4.150), the neutral element of G is (0, eS ) and the inverse of
(v, s) is (v, s)1 = (s1 (v), s1 ) = (s1 (v), s1 ). It is easy to check that V is an
invariant subgroup of G, while S is not in general. As a matter of fact, S is invariant
iff the automorphism is trivial, i.e., the product is direct. Indeed, one has readily
4.5 Lie Groups and Lie Algebras: A Reminder 99

(v, s)(v , eS )(v, s)1 = (s (v ), eS ) V

and

(v, s)(0, s )(v, s)1 = (v, ss )(s1 (v), s1 ) = (v ss s1 (v), ss s1 ).

In addition to the WeylHeisenberg group GWH = R  R2 that we discussed in

Chap. 2, examples of semidirect products of this type are the following groups,
which we will study in detail in the sequel:
1. The Euclidean group E(n) = Rn  SO(n);
2. The Poincar group P+ (1, 3) = Rn  SOo (3, 1), where the second factor is the
Lorentz group;
3. The similitude group SIM(n) = Rn  (R+ +
SO(n)), where R is the group of
dilations, whereas SO(n) denotes the rotations, as in the first example (since these
two operations commute, one gets here a direct product).
3. Group extension
An even more general concept is that of extension. We say that G is an extension of
G1 by G2 if G1 G/G2 . Thus, here G2 is an invariant subgroup of G, but G1 need
not even be a subgroup. The extension is said to be central if G2 is contained in the
center of G.
The most efficient way of studying group extensions is by using the language of
exact sequences [467,Bar77]. Let {Hk , k I N} be a sequence of groups, and, for
each k I, let ik : Hk1 Hk be a homomorphism:

i ik+1
. . . Hk1
k
Hk Hk+1 . . .

Then, the sequence {Hk , k I} is exact if, for each k I, Im ik = Ker ik+1 Hk . Let
f : G G be any homomorphism from G into another group G . Then, an obvious
example of an exact sequence is
f
1 Ker f G Im f 1,

where 1 denotes the trivial group with one element. Other ones are

1 Z(G) G Int G 1,

where Z(G) and Int G denote, respectively, the center and the group of inner
automorphisms (i.e., conjugations) of G, and

1 Int G Aut G Out G 1,

where Out G Aut G/Int G denotes the group of outer automorphisms of G.

Using this language, we say that G is an extension of G1 by G2 if the following
sequence is exact:
100 4 Some Group Theory

i
1 G2 G G1 1. (4.151)

i
Now, the exactness of the first triple, 1 G2 G, means that i is an injection; that

is, G2 is a subgroup of G. Similarly, the exactness of G G1 1 means that is a
surjection. Then, the exactness of the central triple means precisely that G2 Ker ,
i.e., G1 = Im G/G2 , as announced.
As a consequence of the definition, each element g G induces an automorphism
of G2 by conjugation, h  ghg1 , h G2 , and thus one gets a homomorphism from
G into Aut G2 . From this, one deduces several additional exact sequences, and one
gets:

1

Z(G2 )

1 G2 G G1 1 (4.152)

1 Int G2 Aut G2 Out G2 1.

1

This diagram simplifies if G2 is abelian, which is the only case we shall need. Then,
indeed, Z(G2 ) = G2 , so that Int G2 = 1 and Aut G2 = Out G2 . Thus, (4.152) reduces
to the scheme:
i
1 G2 G G1 1.
(4.153)
Aut G2

In these terms, the extension problem may be formulated as follows: Given a group
G1 , an abelian group G2 , and an homomorphism : G1 Aut G2 , to find all groups
G such that the conjugation by g G induces on G2 the automorphism (g).
Whereas the general extension problem (4.152) may have no solution for a given
homomorphism : G1 Out G2 , the restricted problem (4.153) always has one
(trivial) solution, namely, the semidirect product G = G2  G1 , which corresponds
to the double exact sequence

i
1 G2 G  G1  1, (4.154)

where the homomorphism : G1 G is an injection (so that G1 is now a subgroup

of G), such that = 1, the identity [in that case, one says that the exact sequence
(4.154) splits].
4.5 Lie Groups and Lie Algebras: A Reminder 101

The canonical example of central extension is, of course, the WeylHeisenberg

group GWH = R  R2 . Besides that one, the most familiar case is the construction
of the full or extended Galilei group G ( G) as a central extension of the ordinary
Galilei group G ( G1 ) by R ( G2 ). The elements of G are pairs g = ( , g), with
R and g G , and the group law reads

g g = ( , g)(  , g ) = ( +  + (g, g ), gg ),

where : G G R is again a multiplier, as in (2.31). This example will be treated

in great detail in Sect. 8.4.1. The problem of central extensions will also reappear in
a crucial fashion in the study of CS for the affine Galilei group, which is an extension
of a dilation group of dimension 2 by the Galilei group, which will be discussed in
Sect. 16.2.2.

4.5.4 Contraction of Lie Algebras and Lie Groups

We conclude this section by recalling some basic facts concerning the process of
contraction, both for Lie algebras and Lie groups and their representations. Let g1 =
(V, [., .]1 ) and g2 = (V, [., .]2 ) be two Lie algebras on the same vector space V . We
say that g2 is a contraction of g1 if there is a one-parameter family of invertible
linear mappings , (0, 1], from V to V such that

lim 1 [ X , Y ]1 = [X,Y ]2 , X,Y V. (4.155)

0

The limit (4.155) defines a new Lie algebra structure on V , which is not isomorphic
to the original one. In the case of an InnWigner contraction [391], a particular
subalgebra of g1 is conserved throughout the process. More precisely, suppose g1
decomposes into a subalgebra s in and a vector subspace vc , that is

g1 = s + vc , (4.156)

in such a way that

[ s , s ]2 s , [ v c , v c ]2 = 0 , [ s , v c ]2 v c . (4.157)

Using (4.156), we can decompose any X V as

X = Xs + Xc , Xs s, Xc vc ,

and define the contracting mappings

(X) = Xs + Xc .
102 4 Some Group Theory

Then, applying (4.155) does not affect the subalgebra s. We say in this case that we
have a contraction of g1 along s.
The contraction process may be lifted to the corresponding Lie groups
[260262]. Again, let g1 and g2 be two Lie algebras such that g1 is of the form
(4.156) and g2 is a contraction of g1 along s. Let G1 be the simply connected Lie
group with Lie algebra g1 . Let S be the subgroup of G1 whose Lie algebra is s in
the decomposition (4.156). Defining the semidirect product

G2 Vc  S , Vc = exp vc vc ,

one easily checks that g2 = (V, [. , .]2 ) is the Lie algebra of G2 . Consider now the
family of maps : G2 G1 given by
 
: (v , s)  expG1 v s . (4.158)

They play essentially the same role as the maps of (4.155) at the level of the
corresponding groups, namely,
! 1
" 2
lim 1 (g) (g ) = g g , g, g G2 , (4.159)
0

1 2
where , denote the product in G1 , G2 , respectively. Indeed, one checks readily
that Te = , where Te is the derivative of evaluated at the neutral element
of G1 . It is easy to see on (4.158) that the subgroup S is preserved during the
contraction.
Classical examples of group contraction are that of SO(3) into the 2-D Euclidean
group E(2) = R2  SO(2) and those of the relativity groups

SOo (4, 1)
"
0 P+ (3, 1) G(3, 1),
c
(4.160)
#
SOo (3, 2)

where SOo (4, 1) is the de Sitter group, SOo (3, 2) is the Anti-de Sitter group,
P+ (3, 1) is the Poincar group and G(3, 1) is the extended Galilei group, whereas
is the curvature of the de Sitter universe or the Anti-de Sitter universe and, of
course, c is the speed of light. We will study similar examples in Chapter 10. More
precisely, we will study there the contraction of SOo (1, 2) SU(1, 1) into P+ (1, 1).
The last step is to apply the contraction method to group representations.
Whereas contractions of Lie algebras and Lie groups are relatively ancient and well
known [391, 549], the extension of the procedure to group representations is rather
recent [468]. A rigorous version has been given by Dooley [260262], which we
follow. The additional difficulty here is that the representation space itself varies
during the contraction.
4.5 Lie Groups and Lie Algebras: A Reminder 103

Let G2 be a contraction of G1 , defined by the contraction map : G2 G1 ,

and let U be a representation of G2 in a Hilbert space H. Suppose that, for each
(0, 1], we have a representation {H , U } of G1 , a dense subspace D of H and
a linear injective map I : H D (this map is called a precontraction in [519]).
Then, one says that the representation U of G2 is a contraction of the family of
representations {U } of G1 if one can find a dense subspace D of H such that, for
all D and g G2 , for every small enough, D , U ( (g)) I1 I1 (D )
and

lim I U ( g) I1 U(g) H = 0. (4.161)

0

This definition is not yet fully satisfactory, however, in the sense that the vector
is fixed, independently of , and this is not always natural. An example will be
discussed in Sect. 10.3. There, in the contraction of SU(1, 1) towards the Poincar
group P+ (1, 1), the states that contract correctly (namely, the CS) do indeed depend
on the contraction parameter. Hence, we replace in (4.161) the fixed vector by I
and demand [519]:

lim I U ( g) U(g)I H = 0. (4.162)

0

Then, one has to find the appropriate domain of contraction, that is, the set of vectors
D , such that the limit lim 0 I exists in some suitable sense.
As we will see in Chap. 10, this (extended) notion of contraction of group
representations gives a precise mathematical meaning to the relationship between
the three relativity groups mentioned above, at the quantum level. As a consequence,
it yields also the convergence into one another of the corresponding CS systems.
Moreover, the same procedure of group contraction will allow us in Sect. 15.1.1.1
to give a clear mathematical formulation to the so-called Euclidean limit of wavelets
on the two-sphere. This means that, when the radius of the sphere goes to , the
spherical wavelet analysis converges into the plane one, on the 2-D tangent plane.
Chapter 5
Hilbert Spaces with Reproducing Kernels and
Coherent States

Abstract This chapter is again purely mathematical. The central theme is the
concept of reproducing kernel Hilbert spaces and the attending measure problems,
which are then used as a tool for constructing CS.

We have already seen in Sect. 2.3 the intimate connection between the canonical
coherent states and a certain reproducing kernel Hilbert space and again in
Sect. 4.2.2, the connection between the coherent states of the discrete series
representation of SU(1,1) and an associated reproducing kernel. It turns out that
coherent states can in fact be defined in great generality as certain preferred sets of
vectors in appropriate reproducing kernel Hilbert spaces. We proceed to explore this
connection in some detail in the present chapter.
Unfortunately, this chapter is somewhat technical in nature and we have to
deviate a bit from standard treatments of reproducing kernels and reproducing kernel
Hilbert spaces, as can be found, for example, in [99, 574, 575, Mes62]. We give here
a somewhat more general treatment of the subject in view of the fact that we need to
define kernels for vector valued function spaces as well as for scalar valued function
spaces, in order to be able to incorporate vector coherent states into the framework,
which usually is not done in the literature. Moreover, since we also wish to link
coherent states to measurable operator valued functions and POV measures, we need
to widen the setting to deal with measurable fields of Hilbert spaces as well. We start
out, in the next section, with a slight generalization of the standard treatment of
reproducing kernels, with the resulting theory being sufficient for most constructions
of coherent states found in the physical literature. In the following section we work
out some illustrative examples. For most readers it is not necessary to go beyond
these two sections. After that we go into the subject of measurable fields of Hilbert
spaces, direct integrals, their relation to POV measures and associated reproducing
kernels. The adopted level of generality assures immediate applicability of the
concept to the various geometric and functional analytic contexts in which they

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 105
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__5,
Springer Science+Business Media New York 2014
106 5 Hilbert Spaces with Reproducing Kernels and Coherent States

are often required. While the level of technicality in this part is rather high, it was
deemed necessary for the development of some of the non-traditional parts of the
theory treated in this book.

5.1 A First Look at Reproducing Kernels

Let X be a set (typically a subset of a locally compact space, which in most cases
of interest here would be Rm or Cm ) and K a finite dimensional Hilbert space of
dimension n. Often K will just be Cn , with the usual Euclidean scalar product. Let
fi : X K, i = 0, 1, 2, . . . , N, where N could be finite or infinite (the latter being most
often the case) be a set of (vector valued) functions which satisfy the conditions:
1. For each x X,

N
0 <  fi (x)2K < , (5.1)
i=1

where  K denotes the norm in K.

2. For any sequence of complex numbers, c0 , c1 , c2 , . . . cN , such that Ni=0 |ci |2 < ,

N
ci fi (x) = 0, for all x X, ci = 0, for all i. (5.2)
i=0

3. For each x X, the set of vectors fi (x), i = 0, 1, 2, . . . , N, spans K .

Using these functions we define a positive definite kernel, K : X X L (K),

N
K(x, y) = | fi (x) fi (y)|, (5.3)
i=0

the convergence of the sum being guaranteed by (5.1). The positive definiteness of
K is understood in the sense that:
1. If x1 , x2 , . . . , xm is any finite set of points in X and vi K, i = 1, 2, 3, . . . m, are
arbitrary vectors, then
m
vi | K(xi , x j )v j  0. (5.4)
i, j=1

2. For each x X, the operator K(x, x) L (K) is strictly positive-definite, i.e.,

K(x, x)v = 0 v = 0. (5.5)

5.1 A First Look at Reproducing Kernels 107

Note that K(x, y) = K(y, x) for all x, y X. If K = Cn then K(x, y) is an n n

matrix and K(x, x) has positive, non-zero eigenvalues. It should also be pointed out
here that, in the usual treatments of reproducing kernels and reproducing kernel
Hilbert spaces, this second condition (5.5) is not assumed. However, we impose it
here to avoid unnecessary technicalities later.
For each x X and v K define the function xv : X K by
N
xv (y) =  fi (x) | v fi (y) = K(y, x)v . (5.6)
i=0

In view of this definition, we shall also write

N
xv = K(, x)v =  fi (x) | v fi . (5.7)
i=0

Let H be the vector space formed by taking the (complex) linear span of these
functions. On it we define a scalar product  | K by the prescription

xu | yv K = u | K(x, y)vK . (5.8)

An easy extension of a standard argument [99,574,Mes62] shows that this defines a

non-degenerate sesquilinear form and hence a proper scalar product on H . We take
the completion of H with respect to the norm defined by this scalar product to get
a Hilbert space which we denote by HK .
Note that if the dimension of K is one, K(x, y) is a complex valued function
and the vectors (5.6) become simply x = K(, x), x X. The vectors xv are not
normalized. Indeed, from (5.8) we see that

xv K = K(x, x)1/2 vK K(x, x)1/2 vK , (5.9)

K(x, x)1/2 denoting the square root of the positive operator K(x, x) and K(x, x)
is the norm of the positive operator K(x, x) which is also its highest eigenvalue.
In particular, if u+ is the (normalized) eigenvector of K(x, x) corresponding to the
highest eigenvalue, then

xu+ K = K(x, x)1/2 . (5.10)

The kernel K(x, y) has a reproducing property. To see this, note that from (5.6)
and (5.8) we have

xu | vy K = u | yv (x)K ,

and since an arbitrary element HK can be written as a linear combination of the

vectors yv , we get the reproducing property of the kernel:

K(, x)u | K = xu | K = u | (x)K . (5.11)

108 5 Hilbert Spaces with Reproducing Kernels and Coherent States

In case K is one dimensional this relation assumes the more familiar form,

K(, x) | K = x | K = (x) .
i
Let vi , i = 1, 2, . . . n be an orthonormal basis of K and set xi = xv .
Definition 5.1.1 (RKHS and coherent states). The Hilbert space HK defined
above is called a reproducing kernel Hilbert space (RKHS), with reproducing kernel
K(x, y), and the vectors,

SK = {xi : x X, i = 1, 2, . . . n} (5.12)

its associated coherent states (CS).

It is not hard to see that for fixed x X, the vectors xi , i = 1, 2, . . . n, are linearly
independent.
For physical applications one usually works with the normalized versions of
these vectors, xi = xi /xi , and often a somewhat wider definition of coherent
states, as vectors in a more general Hilbert space (see (5.23) below), is adopted.
Moreover, for many physical applications a resolution of the identity (see (5.25)
is assumed for the coherent states. However, to the extent that coherent states are
special sets of vectors associated to a reproducing kernel Hilbert space, a resolution
of the identity is not required and indeed, there are also physical coherent states for
which no resolution of the identity is postulated.
In the above definition, if n > 1 we call the xi vector coherent states (VCS),
otherwise, just coherent states. Thus generally, vector coherent states come from an
operator or matrix kernel while coherent states arise from scalar kernels and it ought
to be stressed here that all coherent states arise from positive definite kernels.
Consider now the bounded operator,
n
FK (x) = | xi xi | . (5.13)
i=1

This is a positive, rank-n operator on HK and its range, which we denote by Kx

is an n-dimensional subspace of HK . Furthermore, x  FK (x) defines a positive
operator valued (POV) function on X, taking values in L (HK ). We say that this
POV-function is canonically associated to the reproducing kernel Hilbert space HK .
This function turns out to have many useful properties and indeed, as shown in a
later section, such a POV-function always determines a reproducing kernel Hilbert
space. Let , HK . Then it is easily checked that

 | FK (x) K =  (x) | (x)K , (5.14)

in view of which FK (x) may also be called the localization operator at the point x.
Next, at each point x X define the linear evaluation map, Ek (x) : HK K,

EK (x) = (x) . (5.15)

5.1 A First Look at Reproducing Kernels 109

There follows the very useful property of the reproducing kernel Hilbert space, that
this map is continuous. Indeed, from (5.9) and (5.11),

|u | (x)| = |xu | | xu    K(x, x)1/2 u   .

Hence

EK (x)  =  (x) = sup |u | (x)| K(x, x)1/2   . (5.16)

u1

On the other hand, if u+ is the eigenvector corresponding to the highest eigenvalue

of K(x, x), then taking = x + we get EK (x) = (x) = x + (x) = u+ | K(x, x)u+ 
u u

= K(x, x). And thus, from (5.10), for this choice of ,

EK (x)  = K(x, x)1/2   . (5.17)

Combining (5.16) and (5.17) we have proved the following theorem.

Theorem 5.1.2. For all x X, the evaluation map EK (x) is continuous and
EK (x) = K(x, x)1/2 . For all HK ,

 (x) K(x, x)1/2   . (5.18)

The adjoint EK (x) : K HK of the evaluation map is easily computed. Since, for
all v K and HK ,

 | EK (x)vK = EK (x) | vK =  (x) | vK =  | xv K ,

the last equality following from the reproducing property (5.11), we get

EK (x)v = xv , v K. (5.19)

Moreover, since for any v K,

EK (x)EK (y)v = EK (x)yv = yv (x) = K(x, y)v ,

by (5.6) we get the alternative expression for the reproducing kernel,

K(x, y) = EK (x)EK (y) , (5.20)

which is often useful. Similarly, from (5.14), we get for the POV function

FK (x) = EK (x) EK (x) . (5.21)

Going back to (5.8) and replacing the vectors xu , yv in the left hand side of the
equation by their expansions in (5.6), in terms of the fi , we see that

N
u | fi (x) f j (y) | v fi | f j K = u | K(x, y)v .
i, j=0
110 5 Hilbert Spaces with Reproducing Kernels and Coherent States

Comparing with (5.3) and using (5.2) this can be shown [574] to imply that

 f i | f j  K = i j , i, j = 0, 1, 2, . . . N . (5.22)

Thus the vectors fi , which were initially used to define the kernel K become an
orthonormal basis of the reproducing kernel Hilbert space. This means that the
reproducing kernel Hilbert space has dimension N + 1, if N is finite or infinity
if N is infinite. Moreover, in a standard manner it can be shown that any other
orthonormal basis of HK also satisfies (5.1) and has the spanning property, just like
the fi , and hence the kernel K(x, y) can also be written in terms of this basis. Thus,
the expression of the reproducing kernel in (5.3) is independent of the choice of the
orthonormal basis.
If H is an abstract Hilbert space, of the same dimension as HK , it is possible to
define coherent states in H by choosing an orthonormal basis {i }Ni=0 in H. With the
same fi as before we define the vectors in H
N
xv =  fi (x) | v i , x X, v K (5.23)
i=0

and similarly we define the coherent states xi , i = 1, 2, . . . n, x X, in H. These

vectors are just the unitary images in H of the corresponding vectors in HK , obtained
via the unitary map W : HK H, with W fi = i , i = 0, 1, 2, . . . , N. They define the
same reproducing kernel K(x, y), since

xu | xv H = u | K(x, y)vK . (5.24)

Indeed, most often the coherent states one uses in physics are of this type, i.e.,
they are constructed in a Hilbert space which is different from, but isomorphic to,
a reproducing kernel Hilbert space. The construction outlined above is the most
general technique for building coherent states, as we will see in Sect. 11.6: a family
of coherent states is always associated to an orthonormal basis in a reproducing
kernel Hilbert space.
As mentioned earlier, for most physical applications coherent states are also
required to satisfy a resolution of the identity. This means that the space X is
a measure space, with some convenient measure d and the following operator
identity holds:
n 
| xi xi | d (x) = IH , (5.25)
i=1 X

where IH is the identity operator on H. In this case the reproducing kernel can be
seen to possess the additional property of being square integrable, in the sense that

K(x, z)K(z, y) d (z) = K(x, y). (5.26)
X
5.2 Some Illustrative Examples 111

This also means that the family of coherent states xi forms a rank-n tight frame in
the sense of Chap. 3. We shall come back to this point again in Chap. 6.
On H we again have a POV function
n
F(x) = | xi xi | .
i=1

The range of the positive operator F(x) is an n-dimensional subspace, Kx , of H

and the union of these subspaces for all x X is total in H. If the resolution of
the identity (5.25) is also satisfied by this POV function, it becomes a measurable
function, in the sense that for all , H, the function x   | F(x)  is
measurable. In fact, starting from such a POV function on an abstract Hilbert space
H, we could define a very general reproducing kernel K(x, y) = F(x)1/2 F(y)1/2
and associated coherent states. This way of constructing reproducing kernels is
discussed in Sects. 5.3 and 5.4.

5.2 Some Illustrative Examples

To illustrate the above discussion, a few examples of reproducing kernel Hilbert

spaces and their associated coherent states are in order.

5.2.1 The Canonical CS

Going back to the representation of the canonical coherent states in the Fock
Bargmann space Hhol , discussed in Sect. 2.4, we see that the kernel Khol (z , z) =

z |z Hhol = ez z is a square integrable kernel for Hhol and the vectors z are the
associated coherent states. However, these are not normalized states, so that dividing
by Khol (z, z) gives the normalized coherent states e|z| z . Being a square integrable
2

kernel, there is also the resolution of the identity (2.68). The orthonormal basis
vectors spanning Hhol are the vectors un in (2.76).

5.2.2 An Example from a Hardy Space

Consider now the Hardy space H 2 (D) of all analytic functions f on the open unit
disc D, satisfying the condition,
  2
1 i
 f  := sup
2
| f (re )| d < ,
2
z = rei . (5.27)
0<r<1 2 0
112 5 Hilbert Spaces with Reproducing Kernels and Coherent States

This is a Hilbert space, with the above norm and associated scalar product
  2
1 i
 f | g = sup f (re ) g(re ) d
i (5.28)
0<r<1 2 0

and the vectors, fn (z) = zn , n = 0, 1, 2, . . . , , form an orthonormal basis of it. This

space has the reproducing kernel
1
K(z, z ) = , z, z D . (5.29)
1 zz
This kernel is not square integrable, since the scalar product (5.28) is not of L2 -type.
Coherent states can be defined for the above kernel as

z = fn (z) fn = zn fn , zD. (5.30)
n=0 n=0

These do, however, satisfy a formal resolution of the identity on H 2 (D),

  2
1
sup |z z | d = I , (5.31)
0<r<1 2 0

when the operator integral on the left hand side is appropriately interpreted.
We will meet similar examples in the part of Sect. 7.6 devoted to singular orbits
on conformal groups.

5.2.3 An Example of VCS from a Matrix Domain

The following is an interesting example of a construction of vector coherent states

over a matrix domain [19, Ali07].
Let X = CNN be the domain consisting of all complex N N matrices with
nonzero determinants (i.e., the group GL(N, C)). We denote matrices in X by Z,
which have the matrix elements Zi j = xi j + iyi j , i, j = 1, 2, . . . N and equip X with
the measure

N 2 Tr [Z Z] N
d (Z) = e dZ, dZ = dxi j dyi j , (5.32)
i, j=1

being a positive constant. The measure is normalized to be of total mass one.

It can be shown [332, 426] that the functions n : X X, n (Z) = Z n , n =
0, 1, 2, . . . , , satisfy the matrix orthogonality condition,

m (Z)n (Z) d (Z) = n cn IN mn , (5.33)
X

where IN is the N N identity matrix and the quantities ck > 0 have the explicit
forms
5.2 Some Illustrative Examples 113

 n+1 n+1
1
cn =
(n + 1)(n + 2) (N + j) (N j) . (5.34)
j=1 j=1

In particular, c0 = 1, c1 = N, c2 = N 2 + 1, etc. Let 1 , 2 , . . . , N be the standard

orthonormal basis of CN . Then the functions

Zn i
fni (Z) = n (Z) i = & , i = 1, 2, . . . , N, n = 0, 1, 2, . . . , , (5.35)
ck k

2 (X, d ) of all square integrable functions

are orthonormal in the Hilbert space LC N h
from X to C . Since any Z X has nonzero determinant, the vectors fni (Z), for all
N

i and N, span CN . Let H be the subspace spanned by these functions. Then the
function

N
Z n W n
K (Z,W ) = | fni (Z) fni (W )| = (5.36)
n=0 cn
n
n=0 i=1

converges for all Z,W X and hence is the reproducing kernel of H . This kernel
is square integrable,

K (Z,W )K (W, Z  ) d (W ) = K (Z, Z  ) .
X

Using the above reproducing kernel we now define the vector coherent states

Zi = K (, Z ) i , i = 1, 2, . . . , , Z X, (5.37)

i.e.,

W k Z k i
Zi (W ) = .
n=0 cn
n

These VCS also satisfy the resolution of the identity on H .

N 
|Zi Zi | d (Z) = I .
i=1 X

It ought to be noted that, in the above construction, X consisted of matrices with

non-zero determinants. We could also have taken X to be the set of all complex,
N N matrices. However, in that case the spanning property, that for fixed Z, the
vectors fni (Z), i = 1, 2, . . . N, n = 0, 1, 2, . . . , , span C, would not hold for all Z.
However, the set of matrices with zero determinant are a null set of the measure
d . Thus the spanning property would hold almost everywhere and we could still
build the same VCS, satisfying again the above resolution of the identity. A general
discussion of such constructions appears in the following sections.
114 5 Hilbert Spaces with Reproducing Kernels and Coherent States

5.3 A Second Look at Reproducing Kernels

As mentioned earlier, the material in this section and the next, up to Sect. 5.4.3,
is rather technical and most readers might wish to skip it. On the other hand, we
elaborate here a rather far reaching generalization of the notion of reproducing
kernel Hilbert spaces, in a direction not readily available in the present literature.
In particular, we introduce a construction of a reproducing kernel Hilbert space
using measurable fields of vectors and analyze in some detail the relation between
such Hilbert spaces and measurable positive operator valued functions.

5.3.1 A Motivating Example

In order to motivate the discussion, let us go back to the canonical coherent states,

s (q,p) = ei(pQqP) s ,

as defined in (2.23). Each one of these vectors, s (q,p) , defines a one-dimensional

subspace, K(q,p) = C[s (q,p) ], of the underlying Hilbert space H (which is also the
space spanned by these vectors). The collection of all of these spaces, {K(q,p) :
(q, p) R2 }, forms a measurable field of Hilbert spaces, in a sense to be made
precise below. (As a matter of fact, in this case, the set {K(q,p) : (q, p) R2 }
actually defines a line bundle, with base space R2 and fiber K(q,p) , above each point
(q, p) R2 .) Consider now the operator

 p; q p ) = | s
(q,p)  (q,p) | (q ,p )  (q ,p ) |
s s s
K(q, (5.38)

on H, defined for each choice of pairs, (q, p) and (q , p ), of points in R2 . Using the
positive operators (in this case, simply one-dimensional projection operators)

F(q, p) = | s (q,p) s (q,p) |, (q, p) R2 , (5.39)

(5.38) may be rewritten in the form,

 p; q , p ) = F(q, p)1/2 F(q , p )1/2 .

K(q, (5.40)

Indeed, F(q, p) being a positive operator, it has a square root, denoted F(q, p)1/2 ,
that satisfies the relation F(q, p)1/2 F(q, p)1/2 = F(q, p). Since, for any vector
 p; q p ) lies in K(q,p) , we may see K(q,
K(q ,p ) , the vector K(q,  p; q p ) as a linear
map from K(q ,p ) to K(q,p) , which satisfies

 p; q p ) = K(q
K(q,   , p ; q, p) ,
5.3 A Second Look at Reproducing Kernels 115

and, for (q, p) = (q , p ), is a strictly positive operator on K(q ,p ) and the mapping
defines a positive-definite kernel. Next, for each (q, p) R2 , let us define a vector-
valued function  (q,p) that associates to each (q , p ) R2 the vector

 (q,p) (q , p ) = K(q
  , p ; q, p) s
(q,p)

in K(q ,p ) , and equip the linear span of all of these vector functions  (q,p) , with the
scalar product

 (q,p) |  (q ,p ) K = (q,p)

s
| (q
s
 ,p ) H . (5.41)

The closure of this linear span, in the above scalar product, is then a Hilbert space,
which we denote by H   . Let us emphasize again that this is a space of vector-
K
valued functions. The set of vectors  (q,p) , (q, p) R2 , is clearly total in H
  , and
K
the association s  
(q,p) sets up a bijective isometry W between the spaces
(q,p)
H and H   . The Hilbert space H   is a reproducing kernel Hilbert space with the
K K
reproducing kernel given by (5.38).
In the rest of this chapter, we shall put all of these observations on a wider footing,
in obtaining a comprehensive generalization of the concept of a reproducing kernel
Hilbert space.

5.3.2 Measurable Fields and Direct Integrals

Let us begin with a locally compact space X equipped with a Borel measure .
For technical convenience, we shall assume throughout that the support of is all
of X. Suppose that to each x X we associate a Hilbert space Kx . Let  | x
and  . . . x denote the scalar product and norm, respectively, in Kx . (The Cartesian
product space xX Kx has a natural vector space structure that we assume for it.
The following discussion on measurable fields and direct integrals of Hilbert spaces
has been adapted from [Tak79].)
Definition 5.3.1. The family {Kx : x X} is called a measurable field of Hilbert
spaces if there exists a subspace M of the product space xX Kx such that
1. For each M, the positive, real-valued function x   (x)x on X is
-measurable;
2. If, for any xX Kx , the complex-valued functions x   (x)| (x)x , for
all M, are -measurable, then M; and
3. There exists a countable subset {n } n=1 of M such that for each x X the set of
vectors {n (x)}n=1 is total in K x .
Elements in M are called -measurable vector fields, and the set {n }
n=1 is
called a fundamental sequence of -measurable vector fields. Condition 3 also
116 5 Hilbert Spaces with Reproducing Kernels and Coherent States

implies that each Kx is a separable Hilbert space, and, indeed, we shall mostly
deal with the case in which they are all finite dimensional. Measurable fields of
Hilbert spaces are useful for building direct integrals of Hilbert spaces, which are
generalizations of direct sums. The next two lemmata give criteria for identifying
measurable fields of Hilbert spaces and the existence of a very convenient type of
fundamental sequences.

Lemma 5.3.2. Let {n }

n=1 xX Kx satisfy the following:

1. For each m and n, the function x  m (x) | n (x)x on X is -measurable; and

2. For each x X, the sequence of vectors {n (x)} n=1 is total in Kx .

Then the set

M = { Kx : x  n (x) | (x)x is -measurable for all n}

xX

satisfies conditions 1, 2 and 3 of Definition 5.3.1, and hence {Kx : x X} is a

measurable field of Hilbert spaces and M is a -measurable field of vectors.
Lemma 5.3.3. Let {Kx : x X} be a measurable field of Hilbert spaces on {X, }.
Then the function x  N(x) = dim Kx on X is measurable, and there exists a
fundamental sequence {n }
n=1 of measurable vector fields such that
N(x)
1. {n (x)}n=1 is an orthonormal basis for Kx , x X;
2. N(x)+k (x) = 0, k = 1, 2, . . ., if N(x) < .
Suppose now that we are given {X, }, as before, and a measurable field of
Hilbert spaces {Kx : x X} along with the set of -measurable vector fields M. Let
 M be the collection of all ( -equivalence classes of) vector fields satisfying
H

 2 :=  (x)2x d (x) < , (5.42)
X

and define on it the scalar product


 |  =  (x) | (x)x d (x) , .
, H (5.43)
X

(As usual, this scalar product is nondegenerate on equivalence classes.) It can be

shown that H is complete in the norm (5.42), and, hence, equipped with the scalar
product (5.43), it becomes a Hilbert space. We call H  the direct integral of the
measurable field of Hilbert spaces {Kx : x X} and write

=
H Kx d (x) . (5.44)
X
5.4 Reproducing Kernel Hilbert Spaces: General Construction 117

If Kx = K, some fixed Hilbert space, for all x X, the field {Kx : x X} is

called a constant field of Hilbert spaces and then it is easily seen that (as a unitary
equivalence)

Kx d (x)
= K L2 (X, d ) . (5.45)
X

5.3.3 Example Using a POV Function

The following example of the construction of a measurable field of Hilbert spaces

and their direct integral will be extremely valuable to us in the sequel. Let H be
an arbitrary Hilbert space and let {X, } be as above. Let F : X L (H)+ be a
weakly measurable, positive operator-valued function on X, i.e., a function of the
variable x, such that its values F(x) are positive operators on H and, for arbitrary
, H, the complex-valued function x   |F(x)  is measurable, and assume
that F(x) = 0, x X. (Note, we are also assuming that the support of is all of
X.) An example of such a function is the map (q, p)  F(q, p) = | s (q,p) s (q,p) |
defined in (5.39). We now construct a measurable field of Hilbert spaces and a direct
integral associated in a natural way to the operator-valued function F. For each
x X, consider the range of this square root operator:

Ran F(x)1/2 = {F(x)1/2 : H} H. (5.46)

In general, this is not a closed subspace of H; let Kx be its closure. Denote by  . . . x

and  | x the restrictions of the norm and scalar product, respectively, of H to Kx .
(Note that, in particular, if F(x) has finite rank, then Ran F(x)1/2 = Kx .) Let {n }
n=1
be a basis of H, and consider the sequence of vectors n xX Kx , n = 1, 2, . . .,
with

n (x) = F(x)1/2 n , n = 1, 2, . . . , x X. (5.47)

The measurability of the function x  F(x) now implies that {n }n=1 satisfies the
conditions of Lemma 5.3.2. Hence, we can construct the set of -measurable fields
M following that lemma and obtain a direct integral Hilbert space, as in (5.44).

5.4 Reproducing Kernel Hilbert Spaces: General

Construction

Using the foregoing notion of measurable fields of Hilbert spaces, we now proceed
to discuss reproducing kernel Hilbert spaces of equivalence classes of measurable
fields. The presentation given here borrows heavily from [201, 359, Mes62],
118 5 Hilbert Spaces with Reproducing Kernels and Coherent States

although we introduce a few convenient modifications which bring out the con-
nection between measurable POV functions and reproducing kernel Hilbert spaces
(see Theorem 5.4.8).

5.4.1 Positive-Definite Kernels and Evaluation Maps

The starting point is once more a locally compact space X equipped with a
regular Borel measure , the support of which is all of X, a measurable field of
Hilbert spaces {Kx : x X} and an associated set M of -measurable vector fields.
We assume that dim Kx is a fixed finite number n for all x X. (The assumption of
finiteness of n is made for technical convenience and can be relaxed if necessary.)
For x, y X, denote by L (Kx , Ky ) the space of all (bounded) linear maps from Kx
to Ky . If A L (Kx , Ky ) and A : Ky Kx is the adjoint map,

A vy | vx x = vy | Avx y , vx Kx , vy Ky ,

then A L (Ky , Kx ). [Note: L (Kx , Kx ) = L (Kx ).]

Definition 5.4.1 (Measurable positive-definite kernel). A measurable positive-
definite -kernel, on a measurable field of Hilbert spaces {Kx : x X}, is a map that
associates with each pair of points (x, y) X X an element K(x, y) L (Kx , Ky )
and has the following properties:
1. If x1 , x2 , . . . , xN is any finite set of points in X and vi Kxi , i = 1, 2, . . . , N, are
arbitrary vectors, then

N
vi | K(xi , x j )v j  0. (5.48)
i, j=1

2. For each x X, the operator K(x, x) L (Kx ) is strictly positive-definite, i.e.,

K(x, x)vx = 0 vx = 0. (5.49)

3. For fixed y X and vy Ky , the function x  K(x, y)vy , denoted K(, y)vy ,
is an element of the set M of the -measurable vector fields associated with
{Kx : x X}.
Note that conditions 1 and 2 imply that, for all x, y X,

K(x, y) = K(y, x) . (5.50)

For M, let [ ] denote its -equivalence class, i.e., if M, then [ ] if and

only if (x) = (x) for -almost all x X. In other words, any two elements of M,
differing only on a set of -measure zero, belong to the same -equivalence class.
The measurable, positive-definite -kernel K is said to be admissible if, for any
5.4 Reproducing Kernel Hilbert Spaces: General Construction 119

x, y X and vx Kx , vy Ky , the equality [K(, y)vy ] = [K(, x)vx ] holds if and only
if K(z, y)vy = K(z, x)vx for all z X. Let [M] denote the set of all -equivalence
classes of M, and similarly, if F is a subspace of M, we shall denote by [F] the
corresponding subspace of [M].
Definition 5.4.2. A -selection on a linear subspace [F] of [M] is a linear map
: [F] F, such that, for all [ ] [F], one has [ ([ ])] = [ ].
To understand the significance of the above definition, we note that, given [F], it is
always possible to choose an element F as representative of an equivalence class
[ ] [F]. It is not always possible, however, to make the correspondence [ ]
linear for the entire subspace [F]. The existence of a -selection on a particular
subspace [F] ensures this linearity. When a -selection exists on a subspace [F] of
[M], we shall simply write for [ ], with the understanding that ([ ]) = .
Theorem 5.4.3. Let K be an admissible, positive-definite -kernel on the mea-
surable field of Hilbert spaces {Kx : x X}. Then, there exists a unique Hilbert
space HK [M] of -equivalence classes of measurable vector fields satisfying the
following conditions:
1. The set

SK = {[K(, x)vx ] : x X, vx Kx } (5.51)

is total in HK .
2. There exists a -selection : HK M such that the linear map

EK (x) : HK Kx , EK (x)[ ] = ([ ])(x), (5.52)

is bounded (in the norms of HK and Kx ) for each x X.

3. If EK (x) : Kx HK is the dual map of EK (x), then

K(x, y) = EK (x)EK (y) . (5.53)

Proof. Consider finite sums of vectors in SK of the type

N M
[ N ] [K(, xi )vi ], [ M ] = [K(, y j )w j ], xi , y j X, vi Kxi , w j Ky j .
i=1 j=1

Using these, we may define a unique sesquilinear form,

N M
[ N ] | [ M ]K = vi | K(xi , y j )w j xi , (5.54)
i=1 j=1

or, equivalently,

[K(, x)vx ] | [K(, y)wy ]K = vx | K(x, y)wy x , (5.55)

120 5 Hilbert Spaces with Reproducing Kernels and Coherent States

on the linear span of SK . [This is obviously a reproducing condition, analogous to

(5.8) or (5.11).] Indeed, the admissibility of K shows first of all that  | K is well
defined; for, if [K(, x )vx ] = [K(, x)vx ] and [K(, y )wy ] = [K(, y)wy ], then

vx | K(x , y )wy x = vx | K(x , y)wy x = K(y, x )vx | wy y

= K(y, x)vx | wy y = vx | K(x, y)wy x ,

which is the consistency condition required by (5.55). Furthermore,  | K is

nondegenerate, since, for each x X,

N
 N (x)2x = K(x, xi )vi | K(x, x j )v j x
i, j=1

N
= [K(, x)K(x, xi )vi ] | [K(, x j )v j ]K
i, j=1

= [K(, x) N (x)] | [ N ]K [K(, x) N (x)]K [ N ]K ,

in virtue of the CauchySchwarz inequality (with an obvious meaning for  . . . K ).

But,

[K(, x) N (x)]2K = [K(, x) N (x)] | K(, x) N (x)]K

=  N (x) | K(x, x) N (x)K

= K(x, x)1/2 N (x)2x .

Thus,

 N (x)x K(x, x)1/2 [ N ]K , (5.56)

and, hence, if [ N ]K = 0, then N (x) = 0 for all x X. On the other hand,
N (x) = 0, x X, surely implies that [ N ]K = 0.
Completing the linear span of SK in the scalar product (5.54), we obtain a Hilbert
space HK . To see that HK [M], let [ N ] = jIN [K(, x j )v j ], N = 1, 2, 3, . . ., where
IN is a finite index set for each N, be a Cauchy sequence in the linear span of SK .
Then, (5.56) shows that, for each x X, N (x) is a Cauchy sequence in Kx . Since Kx
is a Hilbert space, there is, for each x X, a (x) Kx such that limN  N (x)
(x)x = 0. Moreover, standard measure theoretic arguments show that M.
Similarly, letting [ ] = limN [ N ], it is easily seen that [ ] HK and that
[ N ] [ ]K 0.
To prove the existence of a -selection on HK , define a pointwise map on
vectors [ N ] HK as

([ N ])(x) = N (x), x X. (5.57)

5.4 Reproducing Kernel Hilbert Spaces: General Construction 121

Then, by (5.56),

 ([ N ])(x)x K(x, x)1/2 [ N ]K ,

and therefore can be extended to all of HK . Thus, the norm estimate

 ([ ])(x)x K(x, x)1/2 [ ]K , [ ] HK , (5.58)

is generally valid, and, consequently, defines a -selection on HK . This also

proves the boundedness of the evaluation map EK (x) in (5.52) as well as of the
dual map EK (x) . Next, for arbitrary x, y X, vx Kx and wy Ky ,

vx | K(x, y)wy x = vx | EK (x)[K(, y)wy ]x = EK (x) vx | [K(, y)wy ]K .

From this, using (5.55) and the fact that SK is total in HK , we get

EK (x) vx = [K(, x)vx ]. (5.59)

Hence,

EK (x)EK (y) wy = EK (x)[K(, y)wy ] = K(x, y)wy ,

proving (5.53). 

From now on, we assume that n < . It is not hard to see that the following three
conditions are equivalent (recall that we are assuming that dim Kx = n < , for all
x X):

K(x, x) > 0 , (5.60a)

EK (x)HK = Kx , (5.60b)
EK (x) vx = 0 vx = 0 . (5.60c)

In view of Theorem 5.4.3, from now on, we shall write for elements [ ] HK , it
being understood that

(x) = EK (x) = ([ ])(x) Kx . (5.61)

With this convention, elements in SK will be written

[vx ] = EK (x) vx = K(, x)vx . (5.62)

Definition 5.4.4 (Reproducing kernel Hilbert space). The Hilbert space HK ,

constructed in the above manner, using a measurable field of Hilbert spaces, is called
a reproducing kernel Hilbert space, and the operator function K(x, y) is called a
reproducing kernel for HK . The map EK (x) : HK Kx is called the evaluation map
at x X.
122 5 Hilbert Spaces with Reproducing Kernels and Coherent States

It is readily proven that, given a Hilbert space of equivalence classes of measur-

able vector fields (taken from a measurable field of Hilbert spaces {Kx : x X}),
a reproducing kernel, when it exists, is unique. Furthermore, given such a Hilbert
space, the defining condition for a reproducing kernel is that it constitute a map
that associates with each pair (x, y) X X an operator K(x, y) L (Ky , Kx ), such
that, for fixed x and vx , the vector field K(, x)vx is measurable and the reproducing
condition (5.55) is satisfied. Alternatively, in view of (5.62), the reproducing
condition may also be written as

K(, x)vx | K = vx | (x)x , (5.63)

for all x X, vx Kx and HK .

5.4.2 Coherent States and POV Functions

Let us write for the range of EK (x)

HK (x) = EK (x) Kx HK . (5.64)

Then, (5.60c) shows that HK (x) is an n-dimensional subspace of HK and EK (x) is

a vector space isomorphism. Let PK (x) be the projection operator

PK (x)HK = HK (x). (5.65)

We claim that

PK (x) = EK (x) K(x, x)1 EK (x) , (5.66)

that is, for any HK ,

PK (x) = EK (x) K(x, x)1 (x) , (5.67)

and, in particular,

= EK (x) K(x, x)1 (x), HK (x) . (5.68)

Equation (5.66) indeed defines a projection operator. The relations PK (x) =

PK (x)2 = PK (x) are immediate to verify, noting the equality K(x, x) = EK (x)EK (x)
that follows from (5.53).
Define the map VK (x) : Kx HK (x), by

VK (x) = EK (x) K(x, x)1/2 , (5.69)

5.4 Reproducing Kernel Hilbert Spaces: General Construction 123

where K(x, x)1/2 denotes the inverse of the (bounded) positive square root operator
K(x, x)1/2 L (Kx ). Then VK (x) = K(x, x)1/2 EK (x).
The map VK (x) is unitary. Indeed, by inspection, VK (x)VK (x) = IKx (the unit
operator on Kx ), and, for all vx Kx , one has

VK (x)VK (x) EK (x) vx = EK (x) K(x, x)1 EK (x)EK (x) vx = EK (x) vx .

Defining the operator

FK (x) = VK (x)K(x, x)VK (x)1 (5.70)

as the image of K(x, x) on HK (x) under VK (x), we find that

FK (x)EK (x) vx = EK (x) K(x, x)vx . (5.71)

Extend FK (x) to all of HK by defining it to be zero on the orthogonal complement

of HK (x), and, for simplicity of notation, denote both this extended operator and its
restriction to HK (x) by the same symbol. Then, we obtain

FK (x) = EK (x) EK (x). (5.72)

Thus, FK (x) is a bounded positive operator on HK of norm

FK (x) = K(x, x) . (5.73)

The projection operator PK (x) is the support of FK (x), which implies that

PK (x)FK (x)PK (x) = FK (x) .

Furthermore, since the square root operator is FK (x)1/2 = VK (x)K(x, x)1/2VK (x)1 ,
the following is easily seen to hold:

FK (x)1/2 = VK (x) (x) HK (x), HK (x) , (5.74)

whence,

EK (x) vx = FK (x)1/2VK (x)vx (5.75)

and

FK (x)1/2 EK (x) vx = EK (x) K(x, x)1/2 vx . (5.76)

Since, for arbitrary , HK , (5.72) implies

 | FK (x) K =  (x) | (x)x , (5.77)

124 5 Hilbert Spaces with Reproducing Kernels and Coherent States

and , are elements of the measurable field of vectors M, the positive operator-
valued function x  FK (x) is measurable which, in view of the example worked
out in Sect. 5.3.2, also implies that {HK (x) : x X} is a measurable field of Hilbert
spaces. The associated set MK of measurable vector fields consists of all elements
in xX HK (x) of the type x  VK (x) (x), where M. Additionally, the rank
of the operator FK (x) is finite and equal to n, for each x X, and the set of vectors
{FK (x)1/2 : x X, HK } is total in HK .
Definition 5.4.5. The function FK is said to be the POV function canonically
associated with the reproducing kernel Hilbert space HK .
The operator FK (x), being a rank n positive operator, can be written in the form
n
FK (x) = | xi xi | , (5.78)
i=1

where the vectors xi HK (x), i = 1, 2, . . . , n, are linearly independent (and

nonzero). Moreover, if vix = VK (x)1 xi Kx , i = 1, 2, . . . , n, is the corresponding
(linearly independent) set of vectors in Kx , then
n
K(x, x) = | vix vix | , (5.79)
i=1

and we infer from (5.69) that

xi = K(, x)K(x, x)1/2 vix = EK (x) K(x, x)1/2 vix . (5.80)

Definition 5.4.6 (Coherent states). The set of vectors

SK = {xi : x X, i = 1, 2, . . . , n} SK (5.81)

is called a family of coherent states (CS), associated with the reproducing kernel K.
The above definition of CS is the most general one employed in this book. In later
chapters, we shall look at more specific families of CS, which satisfy additional
properties, usually arising from group symmetries. Clearly, the set of CS is total
in HK . Moreover, the measurability properties of K(x, y) and an application of
Lemma 5.3.3 demonstrate that, for each i, x  xi HK (x) can be chosen as a
measurable vector field. A point ought to be made here about the choice of the
linearly independent set of vectors {xi }ni=1 used to represent FK (x). If we were to
choose for these vectors the eigenvectors of FK (x), then (5.78) would simply be
its spectral representation and the xi its eigenvectors. On the other hand, starting
with this representation, if we define a second set of linearly independent vectors
xi = nj=1 U (x) ji xj , i = 1, 2, . . . , n, where the U (x) ji are the elements of an n n
unitary matrix, then once again FK (x) = ni=1 | xi xi |. The vectors xi , however, are
no longer necessarily mutually orthogonal.
5.4 Reproducing Kernel Hilbert Spaces: General Construction 125

5.4.3 Some Isomorphisms, Bases, and -Selections

Given the reproducing kernel Hilbert space, constructed above, it is immediate to

verify that the function
 y) = FK (x)1/2 FK (y)1/2
K(x, (5.82)

is an admissible, positive-definite -kernel on the measurable field of Hilbert spaces

{HK (x) : x X}. Thus, using this kernel, we could construct a second Hilbert space
 would be the reproducing kernel. Canonically associated with this
HK , for which K
new Hilbert space, one would have a POV function x  FK (x). As we show in the
next lemma, however, this new Hilbert space HK is not sensibly different from HK .
Lemma 5.4.7. The mapping

W : HK HK , (W )(x) = FK (x)1/2 , (5.83)

is unitary and maps FK (x) onto FK (x), for each x X; i.e.,

FK (x) = W FK (x)W 1 . (5.84)

Furthermore, the respective evaluation maps and reproducing kernels are related by

EK (x) = VK (x)EK (x)W 1 = FK (x)1/2W 1 (5.85)

 y) = VK (x)K(x, y)VK (y)1 = E  (x)E  (y) ,
K(x, (5.86)
K K

while the coherent states xi , i = 1, 2, . . . , n, satisfying

n
FK (x) = | xi xi | , (5.87)
i=1

are the vectors

xi = W xi = WVK (x)vix . (5.88)

Proof. Clearly W is a linear map, and since the vectors HK (x), as x ranges
through X, are total in HK , it is enough to prove the unitarity of W on such vectors.
Let FK (x)1/2 : HK (x) HK (x) be the inverse of the operator FK (x)1/2 restricted to
HK (x). For such a vector ,

 x)FK (x)1/2 ,
(W )(y) = FK (y)1/2 = K(y,

and, hence, for all HK (x),

 x)FK (x)1/2 .
W = K(, (5.89)
126 5 Hilbert Spaces with Reproducing Kernels and Coherent States

Thus, using the reproducing property (5.63) in HK , we obtain, for , HK (x),

 x)FK (x)1/2 |  
W | K = K(, K

= FK (x)1/2 | HK (x)

=  | FK (x)1/2 HK (x)

=  |W K ,

which yields W = FK (x)1/2 , for each HK (x). With this and (5.83), we
obtain immediately that WW = W W = IHK (x) , thus proving the unitarity of W .
Next, for HK , (5.83) and (5.74) imply

EK (x)(W ) = (W )(x) = FK (x)1/2 = VK (x) (x) = VK (x) EK (x) ,

from which (5.84), (5.85), and (5.86) follow. Finally, (5.88) results from (5.80),
upon noting that FK (x) = EK (x) EK (x). 

It emerges from the above discussion that a reproducing kernel Hilbert space is
always associated to a POV function. This is made precise in the following theorem,
as a direct consequence of the foregoing lemma.
Theorem 5.4.8. Let H be an abstract Hilbert space, x  F(x) a weakly measurable
POV function on {X, }, with values in L (H)+ , satisfying:
1. Rank F(x) = n < , n = 0, for all x X;
2. The set of vectors

S = {F(x)1/2 : H x X} (5.90)

is total in H.
Then, {Kx : x X}, where Kx H is the range of F(x), is a measurable field of
Hilbert spaces on which

K(x, y) = F(x)1/2 F(y)1/2 (5.91)

is an admissible -kernel. The associated reproducing kernel Hilbert space HK and

the canonical POV function x  FK (x) are related by the unitary map

WK : H HK , (WK )(x) = F(x)1/2 , (5.92)

WK F(x)WK1 = FK (x). (5.93)

If
n
F(x) = | xi xi | , (5.94)
i=1
5.4 Reproducing Kernel Hilbert Spaces: General Construction 127

where xi Kx , i = 1, 2, . . . , n, are linearly independent and x  xi are measurable

vector fields, then the coherent states

SK = {xi = WK xi : i = 1, 2, . . . , n, x X} (5.95)

form a total set in HK . In general, a vector Kx maps to

WK = EK (x) F(x)1/2 = K(, x)F(x)1/2 HK , (5.96)

where F(x)1/2 is the inverse of the square root operator F(x)1/2 restricted to Kx .
(Again, in the definition of the kernel in (5.91), the restriction of the operator F(x)1/2
to Kx is implied.)
Thus, we have shown that having a reproducing kernel Hilbert space is com-
pletely equivalent to having a POV function on a Hilbert space satisfying the
conditions of this theorem. Note that, in the statement of this theorem, we could have
replaced the hypothesis on the existence of a measurable POV function x  F(x)
and conditions 1 and 2 by the following equivalent hypothesis.
There exists a set of vectors

S = {xi : i = 1, 2, . . . , n, x X},

which is total in H and satisfies:

1. For each x X, the vectors xi i = 1, 2, . . . , n, are linearly independent;
2. For each H, the functions x  xi | , i = 1, 2, . . . , n, are measurable.
The above construction of a reproducing kernel Hilbert space can also be carried
out in terms of the set of basis vectors xi i = 1, 2, . . . , n, x X, F(x) =
ni=1 |xi xi |, and indeed, for explicit computations, it is more useful to do it this
way: The n n matrix-valued function K(x, y), with matrix elements

K(x, y)i j = xi | yj , i, j = 1, 2, . . . , n, (5.97)

defines a positive-definite kernel on the measurable field of Hilbert spaces

{Kx : x X}, with Kx = Cn for each x. The resulting reproducing kernel Hilbert
space consists of measurable maps : X Cn . In particular,

(EK (x) v)(y) = K(y, x)v, v Cn . (5.98)

The unitary map WK : H HK is now given, componentwise, by

(WK )i (x) = i (x) = xi | . (5.99)

The matrix K(x, y) is related to the operator K(x, y) in (5.91) by

n
K(x, y) = | uix K(x, y)i j uyj | , (5.100)
i, j=1
128 5 Hilbert Spaces with Reproducing Kernels and Coherent States

where uix is the vector F(x)1 xi (= K(x, x)1/2 vix ), considered as an element in
HK (x) [the restriction of F(x) to HK (x) being again implied here] and uyj | is the
dual of | uyj  as a vector in HK (y).
Suppose now that H  is a Hilbert space of (equivalence classes) of -measurable
vector fields from {Kx : x X} and HK is a reproducing kernel Hilbert space that is
a subspace of H.  Let PK be the projection operator

 = HK .
PK H (5.101)

Then, for arbitrary [ ] H, PK [ ] is the unique element in HK , such that

(PK )(x) Kx is defined by

vx | (PK )(x)x =  [vx ] | K = K(, x)vx | K , (5.102)

for all vx Kx [see (5.63)]. On the other hand, if again H is a Hilbert space of
(equivalence classes) of -measurable vector fields from {Kx : x X}, we can ask
under what conditions H itself would be a reproducing kernel Hilbert space. This is
settled by the following theorem (see [201] for proof).
Theorem 5.4.9. Let H be a Hilbert space of equivalence classes of -measurable
vector fields taken from the measurable field of Hilbert spaces {Kx : x X}. Then
the following are equivalent:
1. There exists an admissible, positive-definite -kernel K on {Kx : x X} such that
H = HK .
2. There exists a -selection on H, such that for each x X, the evaluation
map, EK (x) : H Kx , EK (x)[ ] = ([ ])(x), is continuous (in the norms of H
and Kx ).

5.4.4 A Reconstruction Problem: Example of a

Holomorphic Map

An immediate application of the theory of reproducing kernel Hilbert spaces is in

the reconstruction of certain types of functions from a knowledge of their values
at countable sets of points. Let HK be a reproducing kernel Hilbert space of
-measurable vector fields from {Kx : x X}. In general, X is not a discrete space
and the set of coherent states (5.81) contains more vectors than is required to form a
basis (in physicists terminology, they form an overcomplete set of vectors). Thus,
it should be possible to choose a countable subset of points {x j }Nj=1 , N dim HK ,
(note that the dimension of HK is in general infinite), such that the set of CS
xi j , i = 1, 2, . . . , n, j = 1, 2, . . . , N, is total in HK . Choosing such a set of CS, if,
for HK , the set of complex numbers
5.4 Reproducing Kernel Hilbert Spaces: General Construction 129

i (x j ) = xi j |  , i = 1, 2, . . . , n, j = 1, 2, . . . , N , (5.103)

are known, it is possible to determine itself. In other words, it is possible to

reconstruct HK from a knowledge of its values (x j ) Kx j at the points x j
X, j = 1, 2, . . . , N.
A simple but powerful example of the construction of reproducing kernel Hilbert
spaces and discretizations in the above sense is provided by a class of holomorphic
maps [489]. Once again, let D = {z C : |z| < 1} be the open unit disc in the
complex plane and H be an abstract Hilbert space. Let {cn } n=0 be an infinite
complex sequence satisfying

cn
cn = 0, lim = 1, n = 0, 1, 2, . . . , (5.104)
n cn+1

and {en }
n=0 be an orthonormal basis in H. It is then readily checked that the series

z = cn zn en (5.105)
n=0

is strongly convergent, so that z H. Furthermore, the convergence is uniform on

every compact subset of D. Thus, for each H, the function

(z) = z |  = cn en | zn (5.106)
n=0

is analytic in D. From this, it also follows that the set of vectors, z , z D, is total
in H. The function K : C C C [see (5.97)] given by

K(z, z ) = z | z 

is then a positive-definite kernel on the space HK (D) of all holomorphic functions

of the type (5.106). Thus, HK (D) is a reproducing kernel Hilbert space and the map

WK : H HK (D), (WK )(z) = (z) = z |  , (5.107)

is a unitary isomorphism.
If {z j }j=1 is a sequence of points in D converging to z0 D, then the values of
the analytic function in (5.106) at the points {z j }j=0 determine it completely on
all of D. This implies that the countable set of vectors

z j = WK z j , z j (z) = z | z , j = 0, 1, 2, . . . , ,

is total in HK (D), and, hence, any vector in it can be written in terms of these.
We ought to point out, however, that it is in general not possible to build frames
using the vectors z j , since the operator represented by j=0 | z j z j | is in general
130 5 Hilbert Spaces with Reproducing Kernels and Coherent States

unbounded. On the other hand, it can be verified that, if j , j = 0, 1, 2, . . . , , are

positive nonzero numbers, satisfying j=0 j < , then the operator defined by the
weak sum j=0 j | z j z j | is bounded with finite trace.
Actually, our reconstruction problem is rendered particularly simple if the vectors
xi j form a (discrete) frame in HK , in the sense of Sect. 3.4.1. Then, indeed,
n N
| xi j xi j | = A , (5.108)
i=1 j=1

where both A and A1 are bounded operators on HK . Let xi j , i = 1, 2, . . . , n, j =

1, 2, . . . , N, constitute the dual frame F {i , A1 , n}. Then, by (3.46), we see that
xj

n N
= i (x j ) xi j . (5.109)
i=1 j=1

The kth component of (x) Kx , in the basis vix , i = 1, 2, . . . , n, [see (5.79)] is,
therefore,
n N
k (x) = i (x j ) xk | A1 xi j  (5.110)
i=1 j=1

and
n
(x) = k (x)K(x, x)1 vkx . (5.111)
k=1

Clearly, the important quantity here, as always, is the frame kernel KA (x, y), with
matrix elements

KA (x, y)i j = xi | A1 yj  . (5.112)

5.5 Some Properties of Reproducing Kernel Hilbert Spaces

Reproducing kernel Hilbert spaces have an array of interesting and useful prop-
erties, some of which we now derive. The assumptions and setting are as for
Theorem 5.4.3. We start out with a minimum norm problem.
Theorem 5.5.1. If HK is a reproducing kernel Hilbert space and HK (x) HK ,
as in (5.64), the vectors HK (x) are determined by their values at x. Of all
HK , satisfying  (x)x = (fixed constant), the vector min HK (x), such
that min (x)x = and min (x) is an eigenvector of K(x, x) corresponding to the
highest eigenvalue max , has minimum norm in HK , and
5.5 Some Properties of Reproducing Kernel Hilbert Spaces 131

1
min = [min (x)/max ] = K(, x)min (x) . (5.113)
max

Proof. The first statement is contained in (5.68) and (5.62). For the proof of the
second statement, let us rewrite (5.58) as

 K K(x, x)1/2  (x)x , HK . (5.114)

On the other hand, [vx ](x) = K(x, x)vx , so that  [vx ](x)x = K(x, x)vx x . Suppose
that vx is an eigenvector of K(x, x) with eigenvalue [we have > 0, since K(x, x) is
a strictly positive operator]. Thus, K(x, x)1/2 vx = 1/2 vx which implies K(x, x)vx =
1/2 K(x, x)1/2 vx . Hence, for this vector

 [vx ](x)x = 1/2 K(x, x)1/2 vx x = [vx | K(x, x)vx x ]1/2 = 1/2  [vx ]K ,

by (5.62) and (5.55). Thus,

 [vx ]K = 1/2  [vx ](x)x .

If vx is the eigenvector for the highest eigenvalue max of K(x, x), then since
K(x, x) = max , we have for this vector

 [vx ]K = K(x, x)1/2  [vx ](x)x ,

which thus minimizes (5.114). Writing min = [vx ], with vx normalized so as to

ensure min  = , using (5.68), we get

min (x)
vx = ,
max

from which (5.113) results. 


The next theorem gives a kind of internal condition for ascertaining when a
Hilbert space of measurable vector fields is a reproducing kernel Hilbert space.
H
Theorem 5.5.2. If {ui }dim
i=1 is an orthonormal basis in HK , then

dim H
K(x, y) = | ui (x)ui (y)| . (5.115)
i=1

If H [M] is a Hilbert space of -equivalence classes of measurable vector fields in

{Kx : x X}, then H is a reproducing kernel Hilbert space if there exists a sequence
H
i=1 M, such that the sequence of -equivalence classes
of vector fields {ui }dim
dim H
{[ui ]}i=1 is an orthonormal basis of H and:
1. For each x X,
132 5 Hilbert Spaces with Reproducing Kernels and Coherent States

dim H
ui (x)x < . (5.116)
i=1

2. The set of vectors {ui (x) : i = 1, 2, . . . , dim H} is total in Kx , for each x X.

In this case, the reproducing kernel is constructed using (5.115), and
dim H
[vx ] = ui (x) | vx x ui . (5.117)
i=1

Furthermore, every orthonormal basis of vectors in a reproducing kernel Hilbert

space, with K satisfying the strict positivity condition (5.49), fulfils conditions 1 and
2 above.
The proof of this theorem is an adaptation of a standard proof, given for example,
in [Mes62], and we omit it. A number of useful norm estimates hold for reproducing
kernel Hilbert spaces, which can be derived using relations such as (5.58), (5.62),
(5.91) and (5.92).
Theorem 5.5.3. Let HK be a reproducing kernel Hilbert space (of -measurable
vector fields taken from a measurable field of Hilbert spaces {Kx : x X}), FK be
the canonically associated POV function and EK (x) : HK Kx be the evaluation
map at x X. Then,

K(x, y) = FK (x)1/2 FK (y)1/2  [FK (x) FK (y)]1/2 , x, y X, (5.118)

EK (x) = EK (x)  = FK (x)1/2 , x X, (5.119)

 (x)x K(x, x)1/2  K , x X, HK . (5.120)

The vectors [vx ], x X, vx Kx , in (5.62) satisfy

 [vx ]K = K(x, x)1/2 vx x and  [vx ](x)x = K(x, x)vx x . (5.121)
Chapter 6
Square Integrable and Holomorphic Kernels

Abstract This chapter is devoted to a systematic study of two special types of

reproducing kernel Hilbert spaces, namely, those corresponding to square integrable
kernels and to holomorphic kernels. The first case is illustrated by the construction
of CS on the circle. This example suggests introducing action-angle variables, which
are then used to extend the theory to a non-holomorphic set-up, namely, the so-called
GazeauKlauder CS. The latter in turn lead to probabilistic considerations, that will
be the focus of Chap. 11.

In this chapter, we study two special types of reproducing kernel Hilbert spaces,
which are probably the most widely occurring types in the physical literature. While
the very general reproducing kernel Hilbert spaces, constructed in the last chapter,
were spaces of vector-valued functions, they were not assumed to be Hilbert spaces
of square integrable functions, with respect to any measure. Most reproducing kernel
Hilbert spaces which arise in physics and in group representation theory do, on the
other hand, turn out to be spaces of square integrable functions. Another widely
occurring variety of reproducing kernel Hilbert spaces are spaces of holomorphic
or square integrable holomorphic functions. We look at these two situations more
closely in this chapter. Recall from the discussion in Chap. 2, that the family of
canonical CS arise from a reproducing kernel Hilbert space of square integrable
functions and indeed, they may also be associated to a space of analytic functions
(the Bargmann space). Square integrable kernels are illustrated by the construction
of CS on the circle. This example suggests introducing action-angle variables,
which are then used to extend the theory to a non-holomorphic set-up, namely, the
so-called GazeauKlauder CS [309, 310]. The latter in turn lead to probabilistic
considerations, that will be the focus of Chap. 11.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 133
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__6,
Springer Science+Business Media New York 2014
134 6 Square Integrable and Holomorphic Kernels

6.1 Square Integrable Kernels

The two examples of reproducing kernels, appearing in (2.66) and (4.91), respec-
tively, (the former coming from the Bargmann space for the representation of the
canonical CS and the latter from spaces of holomorphic functions on the unit
disc, used to describe representations of SU(1,1) from the discrete series) are both
holomorphic kernels, i.e., they arise in Hilbert spaces of holomorphic functions.
They are also square integrable kernels, in that they both satisfy an integral relation
of the type:

K(x, z)K(z, y) d (z) = K(x, y). (6.1)
X

Motivated by this observation, we introduce a somewhat more general notion of

square integrability for reproducing kernels.
Definition 6.1.1 (Square integrable kernel). Let {Kx : x X} be a measurable
field of Hilbert spaces (with respect to the Borel measure ), K an admissible,
positive-definite -kernel, and HK the corresponding reproducing kernel Hilbert
space. The kernel K is said to be square integrable if there exists a bounded operator
A on HK , with bounded inverse, such that for each vx Kx and vy Ky ,

vx |K(x, z)K(z, y)vy  d (z) =  [vx ]|A [vy ] (6.2)
X

[here we are using the notation [vx ] introduced in (5.62)]. An equivalent way of
expressing the above condition for square integrability is, according to (5.78),
 n 
FK (x) d (x) = |xi xi | d (x) = A, A, A1 L (HK ), (6.3)
X i=1 X

the integral being assumed to converge weakly. This, in turn, is equivalent to the
frame condition, already given in (3.11),

m(A) 2K  |FK (x) K d (x) M(A) 2K , HK , (6.4)
X

where m(A) and M(A) are positive real numbers, such that 0 m(A) M(A) < .
In the special case of A = IK (the identity operator on HK ), the condition (6.2) is
exactly the relation (6.1).
For square integrable kernels, a useful reconstruction formula for vectors HK
can be obtained, in terms of the values i (x) = xi |  of their components i
evaluated at the points x. Indeed, it is not hard to verify that for any HK ,
n 
= i (x)A1 xi d (x). (6.5)
i=1 X
6.1 Square Integrable Kernels 135

Let , HK . Then,

 | K =  |[ FK (x) d (x)](A1 )K ,
X

which, by virtue of (5.72) and the weak convergence of (6.3), implies that

 | K =  |EK (x) EK (x)(A1 )K d (x)
X

=  (x)|(A1 )(x)x d (x). (6.6)
X

Thus, the scalar product in HK is realized as an L2 -product with an operator density.

Again, when A = IK ,

 | K =  (x)| (x)x d (x), (6.7)
X

and comparing with (5.43), we see that in this case the reproducing kernel Hilbert
 in (5.44). Also,
space HK becomes a subspace of the direct integral Hilbert space H
now the operators

aK ( ) = FK (x) d (x), B(X), (6.8)

constitute a normalized POV-measure, and following the discussion in Sect. 3.2, the
PV-measure

 ) )(x) = (x) (x),

(P( 
H

 The projection operator PK : H

is its minimal Namark extension to H.  HK now
acts in the manner [see (5.101) and (5.102)]:

(PK )(x) = K(x, y) (y) d (y), 
H, (6.9)
X

while (5.63) appears in the integral form:


EK (x) = K(x, y) (y) d (y) = (x), HK . (6.10)
X

In general, however (i.e., when A is not necessarily equal to IK ), the relation (5.63)
assumes the form:

EK (x) = K(x, y)(A1 )(y) d (y) = (x), HK . (6.11)
X
136 6 Square Integrable and Holomorphic Kernels

Although in the above definition of square integrability we assumed both A and A1

to be bounded operators, a more general situation, where A is bounded but A1 is
only assumed to be densely defined, can also be envisaged. In this case, instead of a
frame, we will get an upper semiframe, as discussed in Sect. 3.3.2 (see also [27]).
Since we are assuming A1 to be bounded, quite generally, we may rewrite
(6.3) as

FK (A) (x) d (x) = IK , where, FK (A) (x) = A1/2 FK (x)A1/2 . (6.12)
X

Since FK (A) (x) is again a positive operator-valued function of x, satisfying all the
conditions of Theorem 5.4.8, there exists another reproducing kernel Hilbert space
HK (A) , with kernel

K (A) (x, y) = FK (A) (x)1/2 FK (A) (y)1/2 . (6.13)

This is an admissible, positive-definite -kernel, defined on the measurable field of

(A)
Hilbert spaces {Kx : x X}, where
(A)
Kx = {A1/2 : HK (x)}, (6.14)

[HK (x) being as in (5.64)], and satisfies the square integrability condition

K (A) (x, z)K (A) (z, y) d (z) = K (A) (x, y). (6.15)
X

Following (5.92) we could construct the unitary map,

WK (A) : HK HK (A) , (WK (A) )(x) = FK (A) (x)1/2 . (6.16)

However, although this map makes the two Hilbert spaces HK and HK (A) isomet-
rically isomorphic, it does not map the coherent states of these spaces into one
another, unless A = IK . Indeed, let xi , i = 1, 2, . . . , n, x X, be the coherent states
i
of HK [see (5.78)(5.81)] and x , i = 1, 2, . . . , n, x X, those of HK (A) . Then it is
easily verified that
i
WK (A) (A1/2 xi ) = x , i = 1, 2, . . . , n, x X (6.17)

6.2 Holomorphic Kernels

We proceed now to a consideration of a class of holomorphic kernels, which as will

turn out, are all square integrable. Let D C be a domain, i.e., an open connected
set,

dz dz 1
d (z, z) = = dy dx, z = x + iy, (6.18)
2 i
6.2 Holomorphic Kernels 137

the Lebesgue measure on D and

d (z, z) = (z, z) d (z, z) (6.19)

any other measure, equivalent to , where is a continuous, positive function, which

does not vanish anywhere on D. Let H = L2 (D, d ), and denote the norm in it by
 . . . hol . Suppose that there exists a non-empty subset of vectors in H, which can
2 (D, d ) H denote this subset.
be identified with functions analytic in z. Let Lhol
Note that if, for example, D = C and = , then there are no nonvanishing analytic
functions in H at all. On the other hand, with as in (2.62), Lhol 2 (D, d ) is the

Bargmann space of entire analytic functions discussed in Sect. 2.4. Similarly, when
D = D, the unit disc, we have the entire class of Hilbert spaces of holomorphic
j
functions Hhol discussed in Sect. 4.2.2. We now have the important result (proof
adapted from [Sug90]):
2 (D, d ) is a closed Hilbert subspace of H, on which the
Lemma 6.2.1. Lhol
evaluation map

Ehol (z) : Lhol

2
(D, d ) C, Ehol (z) f = f (z), (6.20)

is bounded and linear for all z D, and moreover, for any compact subset K D,
there exists a constant c(K) > 0, such that

| f (z)| c(K) f hol , (6.21)

2 (D, d ) and z C.
for all f Lhol
Proof. The linearity of Ehol (z) is obvious and its boundedness would follow directly
2 (D, d ) and
once (6.21) is proved. Let us therefore prove this relation. Let f Lhol
z D. Choose (0, 1) such that V (z) = {w : |w z| < } D. Taylor expanding
f around z in V (z), we may write

f (w) = ak (w z)k , ak C.
k=0

Setting fk (w) = (w z)k ,

   2
1
 fk | f  = fk (w) f (w) d (w, w) = r dr rk+l ei(k) d
V (z) 0 0

2 k++2
= k .
k++2
Thus,

2(k+1)
 f 2 = |ak |2  fk 2 = |ak |2 k+1
,
k=0 k=0
138 6 Square Integrable and Holomorphic Kernels

and since a0 = f (z), this implies


| f (w)|2 d (w, w) | f (z)|2 2 . (6.22)
V (z)

Now let < 1 be chosen so that the closed compact set

K = {w C : dist(K, w) }

is contained in K. Here dist(K, w) is the infimum of |z w|, over all z K. Then, for
any z K, V (z) K . Going back to the measure d (w, w) = (w, w) d (w, w), let

r(K) = inf  (w, w).

wK

Hence, for all z K,

 
 f hol = | f (w)|2 (w, w) d (w, w) | f (w)|2 (w, w) d (w, w)
D V (z)

2 r(K)| f (z)|2 ,

by (6.22); so that taking c(K) = [r(K)]1/2 1 we obtain (6.21).

It only remains to prove that Lhol2 (D, d ) is closed. Let { f }
m m=0 be a Cauchy
sequence in Lhol (D, d ). Since Lhol
2 2 (D, d ) H, there exists f H such that

limm  fm f hol = 0. By virtue of (6.22), it follows that the complex sequence

{ fm (z)}
m=0 converges to some function g(z), and that this convergence is uniform
on every compact subset K of D. Being the uniform limit of holomorphic functions,
g must then also be holomorphic and, as in the standard proof of the completeness
of L2 -spaces, we infer that f (z) = g(z) almost everywhere. Hence g Lhol 2 (D, )

which implies that Lhol (D, ) is closed.

2 

Combining the various results in this section with this lemma, we arrive at our
main theorem on holomorphic kernels, which in fact is a special case of a more
general result proved in [422]:
Theorem 6.2.2 (Holomorphic kernel-space theorem). The subspace Lhol2 (D, d )

of H = L (D, d ) is a reproducing kernel Hilbert space with square integrable

2

kernel KD : D D C,

KD (z, z ) = Ehol (z)Ehol (z ) , (6.23)

such that

KD (z, w)KD (w, z ) d (w, w) = KD (z, z ). (6.24)
D

For fixed w D, the kernel KD (z, w) is holomorphic in z.

6.2 Holomorphic Kernels 139

From the proof of Lemma 6.2.1 it is also clear that, if D is a bounded domain, then
2 (D, d ) (i.e., w.r.t. the Lebesgue measure) is always non-empty. (Indeed, the
Lhol
identity function I(z) = 1, z D, is always in Lhol 2 (D, d )). In this case, the

reproducing kernel KD is called the Bergman kernel of the domain D. As an

j
example, the kernel Khol in (4.91), with j = 1, is the Bergman kernel of D = D,
the unit disc. In general, the kernel Khol is called the -Bergmann kernel of D.
The above theorem admits several generalizations. First, D could be taken to be
a domain in Ck , so that we would be considering Hilbert spaces of holomorphic
functions of k complex variables, z1 , z2 , . . . , zk . Writing z = (z1 , z2 , . . . , zk ), the
measure in (6.18) would now be replaced by

k
1
d (z, z) =
(2 i)k i=1
dzi dzi . (6.25)

Furthermore, the density in the definition of ,

d (z, z) = (z, z) d (z, z), (6.26)

could in general be an admissible weight [497]. To understand the meaning of

this concept, let L2 (D, d ) be, as before, the Hilbert space of all complex-valued
functions on D Ck , which are square integrable with respect to (6.26), and
2 (D, d ) be the set of vectors in L2 (D, d ) which can be identified
again let Lhol
with holomorphic functions. We assume that this set is non-empty. Let Ehol (z) :
2 (D, d ) C, E (z) f = f (z), be the evaluation map at z D. Then is said
Lhol hol
to be an admissible weight if, (1) it is Lebesgue measurable, positive and non-zero
on all of D; and (2) for any z D, there exists a neighbourhood V (z) and a constant
k(z), such that for all w V (z), Ehol (w) k(z). It can be shown that if is an
admissible weight then Lhol 2 (D, d ) is a closed subspace of L2 (D, d ), admitting a

reproducing kernel Khol (z, w) = Ehol (z)Ehol (w) , which is holomorphic in z and is
square integrable in the sense of (6.24).
The second generalization is a little more sophisticated, in that it involves
holomorphic vector bundles [498]. Let E = (E, , M) be a holomorphic vector
bundle, with base space M (which is a complex analytic manifold), total space E and
canonical projection map : E M. Suppose that M has complex dimension k and
that the fibre 1 (x), over the point x M, is isomorphic to Cn . Let h be a Hermitian
structure on E, i.e., it associates a scalar product |x to each fibre 1 (x), x M,
such that if s and t are smooth sections of E, then hx (s,t) = s(x)|t(x)x , and the map
x  s(x)|t(x)x is smooth. Let be a volume form on M and consider the Hilbert
space L2 (E, h, d ). This is the completion of the space of all smooth, compactly
supported sections of E, with respect to the scalar product

s|t = s(x)|t(x)x d (x). (6.27)
M
140 6 Square Integrable and Holomorphic Kernels

2 (E, h, d ) be the set of all vectors in L2 (E, h, d ) which can be identified

Let Lhol
with holomorphic sections of E and assume that this set is non-empty. Let E denote
the dual bundle of E and  : E M the corresponding projection map, i.e., 1 (x)
is the vector space dual of 1 (x), for each x M. Then consider for any v E ,
the linear map

Ev : Lhol
2
(E, h, d ) C, Ev (s) = v ; s(  (v )), (6.28)

;  denoting the dual pairing between E and E . It can be shown that there exists a
neighbourhood V (v ) of v and a constant c(v ) > 0, such that for all w V (v ),

|Ew (s)| c(v )s . (6.29)

From this it follows that the evaluation map

Ehol (x) : Lhol

2
(E, h, d ) 1 (x), Ehol (x)s = s(x), (6.30)

2 (E, h, d ) is a closed subspace of L2 (E, h, d ). The maps

is continuous and that Lhol
Ev and Ehol (x) are related by

Ev (s) = v ; Ehol (x)s, (6.31)

where x  (v ). The Hilbert space Lhol

2 (E, h, d ) admits the reproducing kernel

Khol (x, y) = Ehol (x)Ehol (y) : 1 (y) 1 (x), (6.32)

which again satisfies the square integrability condition (6.1), and is holomorphic
in x. The earlier constructions of -Bergman kernels all become special cases of
this general result.

6.3 Some Examples of Coherent States from Square

Integrable Kernels

As we have seen in the previous chapter, given a reproducing kernel Hilbert space
HK of measurable vector fields on a locally compact space X, there is always an
overcomplete set of vectors xi , i = 1, 2, . . . , n, x X in HK , defined by the kernel
K [see (5.80)(5.81)], which constitutes its coherent states. In concrete situations,
such states may not be related to any group representation at all. In this section, we
shall discuss examples of such CS. The first one pertains to spaces of holomorphic
functions, the second one is a nonholomorphic extension of the first.
Another example is the set of CS on the unit circle, first obtained in [238].
These are derived from canonical CS on the line by a direct integral decomposition
6.3 Some Examples of Coherent States from Square Integrable Kernels 141

(of Bloch type) of L2 (R, dx) into infinitely many copies of L2 (S1 , d ). There exists a
simple and somewhat little known derivation of these states from standard coherent
states. It is noticeable that this construction was proposed earlier by Chang and Shi
[203] for the treatment of a generalized quantum Chirikov map under the rational
resonance condition 2 h = M/N (see also Zyczkowski [621]).

6.3.1 Standard Versus Circle Coherent States

We first express the standard (canonical) coherent states in the position representa-
tion, as in (2.20) (in simplified notations and with s = 1):
qp 1
q,p (x) = 1/4 eixp ei 2 e 2 (xq)
2
(6.33)

Then we restrict the variable x in (6.33) to the interval 0 x < 2 and change its
name to x = . Similarly, we restrict the range of the CS parameter q to the same
interval and change its name to q = . We keep the range R of p and also change
its name as to p = J. Then let us consistently extend the function (6.33) to the 2 -
periodic function

J, ( ) = 1/4 eiJ eiJ /2 e2 inJ e 2 (2 n+ ) .
1 2
(6.34)
n=

Using the Poisson summation formula

 +
(n) = 2  (2 k) ,  (k) := 1
(t) eikt dt , (6.35)
nZ kZ 2

allows one to write (6.34) in the alternate form:

1
J, ( ) = eiJ /2 ein( ) e 2 (Jn) .
1 2
(6.36)
2 3/2
n=

These states are the unnormalized coherent states for the circle in the Fourier
representation, as shown in the next section.

6.3.2 Coherent States for Motion on the Circle

The phase space of a particle moving on a circle (e.g., a simple pendulum in its
rotational regime) is viewed as the cylinder S1 R = {( , J), | 0 < 2 , J R},
equipped with the measure d (x) = 21 dJ d . With > 0 a regularization

parameter which is taken to be 1/ 2 in (6.36), a family of corresponding normal-
ized coherent states is defined as [239, 338, 423]
142 6 Square Integrable and Holomorphic Kernels

 1/4
1 1 1 2 (Jn)2 in
|J,  = &
N (J) 2 2 e 4 e |en  , (6.37)
nZ

where the states |en s form an orthonormal basis of some separable Hilbert space
H . For instance, they could be chosen to be the Fourier exponentials ein forming
the orthonormal basis of the Hilbert space L2 (S1 , d /2 )
= H . They would be the
spatial or angular modes in this representation. In this representation, the coherent
states read as the following Fourier series:
 1/4
1 1 1 (Jn)2 in( )
J, ( ) = &
N (J) 2 2 e 4 2 e , (6.38)
nZ

which, up to a factor, is the expression (6.36). The normalization factor is a periodic

sum of normalized Gaussians which can be written as the elliptic theta function 3
[Mag66], using the Poisson formula (6.35):

1 1 (Jn)2

= e2 inJ e2 n
2 2 2
N (J) = e 2 2
2 nZ
2
nZ

= 3 (J, 2 i 2 ) . (6.39)

Its asymptotic behavior at small and large values of the parameter is given by

lim N (J) =
0
(J n) (Dirac comb) , (6.40)
nZ

lim N (J) = 1 . (6.41)

We also note that lim 0 2 2 N (J) = 1 if J Z and = 0 otherwise.
By construction, the states (6.37) are normalized and resolve the identity in the
Hilbert space H :
 +  2
d
dJ N (J) |J, J, | = IH . (6.42)
0 2

Hence, these coherent states yield a tight frame, but they cannot be obtained as an
orbit of a fixed vector under some group action. They overlap as

1 2 (JJ  )2 
e 8 1 2 ( J+J
2 n) ein( )
2 
 
J, |J ,  = &
2 2 N (J) N (J  )
e 2 (6.43)
nZ

1 2 (JJ  )2 i J+J  (  )
e8 e 2 2  
e 2 ( 2 n) ei n(J+J ) ,
2
= & (6.44)
N (J) N (J  )
nZ

where the last equality results again from the Poisson formula (6.35).
6.3 Some Examples of Coherent States from Square Integrable Kernels 143

The above example was reexamined in [423], including the holomorphic aspects
discussed here, and a complete and rigorous treatment has finally appeared in [338].
Note the recent comprehensive study [590] of squeezed states on the circle (and
more generally on an interval with periodic or nonperiodic boundary conditions)
which result from a truncation of the Gaussian states (6.33).
It is interesting to note that the unitary transformation mapping the Hilbert space
L2 (R) of the canonical CS onto the one of the new CS, namely L2 (S1 R), where
R is the dual of S1 , is analogous to the Zak transform, familiar in solid state theory.
We shall meet this transform explicitly again in the discussion of Gabor frames in
Sect. 17.1. As an extreme example, of course, one finds the discrete frames discussed
in Sect. 3.4.1.

6.3.3 A General Holomorphic Construction

Although most of the subsequent chapters will deal with CS originating from
a group representation, CS are also often associated to the complex structures
possessed by certain types of spaces, unrelated to any group representation. The
Hilbert spaces of holomorphic functions Lhol 2 (D, d ) discussed in Sect. 6.1 admit

such CS. Similarly, for the class of holomorphic maps mentioned in the example
of Sect. 5.4.4, the vectors z in (5.105) or (6.45), or equivalently, the vectors z in
(5.107), are CS which are not necessarily related to any group representation.
In such cases, the reproducing kernel can often be realized as a square integrable
kernel in the sense of Definition 6.1.1, i.e., it may be possible to find a Borel measure
for which (6.2) holds. In the case of the holomorphic maps of the unit disc D C
of Sect. 5.4.4, it is particularly interesting to examine this point since it can be related
[489] to the classical
moment problem [556,Akh65,Sho50]. To begin with, note that
if we take cn = n + 1 and en = un , n = 0, 1, . . . , in (5.105) (or (6.45) below),
then the vectors z are precisely the vectors zj appearing in (4.95)(4.97), for j = 1,
in the holomorphic discrete series representation of SU(1,1). Thus the Hilbert space
HK in (5.107) is now the L2 -space of (4.77) and the kernel K is the same as that in
(4.91)(4.92) for j = 1. Of course this kernel is square integrable since it satisfies
(4.98). Generally, going back to the vectors

z = cn zn en (6.45)
n=0

defined in (5.105), assume, for simplicity, that the coefficients cn are all real, and
consider the formal operator:

A= |z z | d (z, z), (6.46)
D
144 6 Square Integrable and Holomorphic Kernels

where the measure is defined as [see (6.18) and (6.26)]:

1 1 dz dz
d (z, z) = d (z, z) = ,
K(z, z) K(z, z) 2 i
K(z, z ) = z | z , K(z, z) = z 2 , z, z D. (6.47)

For arbitrary , H,
 
 |z z |  d (z, z)     d (z, z) =    .
D D

Thus, as a weak integral, the right hand side of (6.46) defines A as a bounded positive
operator, with A 1. Since the vectors z are overcomplete in H,  |A  = 0
implies = 0, so that A1 exists, although it may not be bounded. The spectrum of
A is, however, easily computed. Indeed,

em |Aen  = cm cn zm zn d (z, z), (6.48)
D

as is readily inferred using (5.105). Introducing polar coordinates (r, ) and

noting that

K(z, z) = c2k |z|2n = c2k r2n
k=0 k=0

depends on r2 only, the integration with respect to may be completed to obtain

 1
em |Aen  = c2n r2n+1 (r2 ) dr mn , (6.49)
0

where we have written

 1

(r ) =
2
c2k r2k . (6.50)
k=0

Thus,
 1  1
c2n
A= n |en en |, n = c2n
0
r2n+1 (r2 ) dr =
2 0
rn (r) dr. (6.51)
n=0

The boundedness of A ensures that, for finite n, n < . Thus A has a pure point
spectrum. However, since n could go to zero as n , A1 could be unbounded.
In any case,

n1 |en en |, n
1/2
A1 = and A1/2 = |en en |. (6.52)
n=0 n=0
6.3 Some Examples of Coherent States from Square Integrable Kernels 145

If we wanted A to be the identity operator, (6.49) would impose the constraints

 1
2
rn (r) dr = , n = 0, 1, 2, . . . , . (6.53)
0 c2n

For a given set of constants cn , these cannot in general be satisfied. However,

replacing (r2 ) dr in the above equation by an arbitrary measure dF(r), (6.53)
becomes just the moment problem for determining this measure, given the set of
constants cn [Sho50]. A solution to this problem would then yield a measure

1
d
 (z, z) = r dF(r) d , (6.54)

for which

|z z | d
 (z, z) = I. (6.55)
D

Thus, with respect to this measure, the kernel K(z, z ) = z |z  would become
square integrable in the sense of Definition 6.1.1, i.e.:

K(z, z )K(z , z ) d
 (z , z ) = K(z, z ),
D

and HK = Lhol2 (D, d  ). Recall that HK is the Hilbert space consisting of all
the holomorphic functions (z) = z | , H, and these now become square
integrable with respect to d  . While this is an attractive procedure, the measure
dF(r) may turn out to be highly singular and generally d  would not be absolutely
continuous with respect to the Lebesgue measure d of D.
Considering the operator A1/2 , which has the eigenvectors en and eigenvalues
1/2
n , n = 0, 1, 2, . . . , we see that the weighted CS

  1 1/2
z = A 1/2
z = bn z en ,
n
bn = 1/2
0
r (r) dr
n
, (6.56)
n=0

are well-defined vectors in H. Indeed, it is not hard to see that bn+1 /bn 1 as n ,
and hence the coefficients bn , n = 0, 1, 2, . . . , , also define a holomorphic map on
the unit disc, in the sense that each one of the functions (z) = z | , H, is
holomorphic on D. Furthermore,

|z z | d (z, z) = I, (6.57)
D

and the corresponding kernel K (z, z ) = z |z  is square integrable with respect

to the original measure d . By Theorem 6.2.2, the functions (z) constitute
2 (D, d ), of holomorphic functions square integrable with
the Hilbert space Lhol
146 6 Square Integrable and Holomorphic Kernels

respect to d , and K is the reproducing kernel for this space. However, the
norm of Lhol2 (D, d ) is different from the norm of the original reproducing kernel

Hilbert space HK , having the kernel K(z, z ) = z |z  = z |Az . Since z |  =

z |A1/2 , for any H, we see that with (z) = z | ,

 2L2 (D, d )
= | (z)|2 d (z, z) =  |A H ,
hol D

while  2K =  2H , where  K denotes the norm in HK [see (5.107)]. Thus,
2 (D, d ), and this inclusion is continuous, if A1/2 is unbounded
while HK  Lhol
HK only forms an open dense set in Lhol 2 (D, d ). As sets, the two Hilbert spaces H
K
and Lhol (D, d ) become equal if A
2 1/2 is bounded.

6.3.4 Nonlinear Coherent States and Orthogonal Polynomials

In this Section we illustrate the holomorphic construction mentioned above for the
case of the so-called deformed coherent states, also known as nonlinear coherent
states in the quantum optical literature [4, 558]. In view of the importance of the
topic, particularly the appearance of the subject of orthogonal polynomials, we make
the discussion self-contained, so that it may be read independently.
Going back to the canonical coherent states, we rewrite them in the form, using
again the ket notation,

|z|2 zn
|z = e 2
1 |en , (6.58)
n=0 [n!] 2

where the en , n = 0, 1, 2, . . . , , are an orthonormal basis in an arbitrary (complex,

separable, infinite dimensional) Hilbert space H. The related deformed or nonlinear
coherent states are then the generalized versions:

zn

1
|z = N (|z|2 ) 2 1 |en , (6.59)
n=0 [xn !] 2

where the x1 , x2 , x3 , . . . , is an appropriate sequence of positive numbers and xn ! =

x1 x2 . . . xn (by convention, x0 ! = 1). The normalization factor N (|z|2 ) is chosen
so that the vectors |z are normalized to one. As usual, these coherent states are
required to form an overcomplete set of vectors in the Hilbert space H, satisfying
also a resolution of the identity of the type,

|zz| N (|z|2 ) d (z, z) = I , (6.60)
D

for some measure d and where D is some appropriate open disk in the complex
plane (often the open unit disc, but which could also be the entire plane).
6.3 Some Examples of Coherent States from Square Integrable Kernels 147

The measure d is of the type d d (r) (for z = rei ), where d is related to

the xn ! through a moment condition [556, Akh65] :
 L  L
xn ! 1
= r2n d (r) , = d (r) , (6.61)
2 0 2 0

zn
L being the radius of convergence of the series n=0 . This means that once the
xn !
quantities (n) = xn ! are specified, the measure d is to be determined by solving,
exactly or approximately, the moment problem (6.61), some representative samples
of which may be found in [35, 309, 312, 418, 461, 489].
More generally, it is possible to start with a function f (z), holomorphic in the
open disc DL = {z : |z| < L}, with a Taylor expansion of the type,

z2n
f (z) = , n > 0, n , 0 = 1 , (6.62)
n=0 n

where the sequence {n }

n=0 satisfies,
n+1
lim = L2 > 0 . (6.63)
n n
n
Then, setting xn = , for n 1, and x0 = 0, the vectors
n1

zn
|z = f (|z|)1/2 [xn !]1/2 |en  , (6.64)
n=0

define a set of deformed or nonlinear coherent states for all z DL . The moment
problem (6.61) is used to determine the measure d and then one has the resolution
of the identity,
 2  L
d d (r) f (|z|) |zz| = I , (6.65)
0 0

and normalization z | z = 1.
It is also known [162, 163, 490] that if the sum 1
n=0 xn diverges then the above
family of coherent states is naturally associated to a set of polynomials {pn (x)} n=0 ,
orthogonal with respect to some measure dw(x) on the real line, which may then
be used to replace the |en  in the definition (6.64) of the CS. To see this, define the
generalized annihilation operator a f by its action on the vectors |z,

a f |z = z|z , (6.66)

and its adjoint af . Their actions on the basis vectors are easily seen to be:

a f |en  = xn |en1  , af |en  = xn+1 |en+1  . (6.67)
148 6 Square Integrable and Holomorphic Kernels

Note that the numbers xn are the eigenvalues of the self-adjoint operator af a f , with
eigenvectors en . We now define the operators,

1 1
Q f = [a f + af ] , Pf = [a f af ] , (6.68)
2 i 2

which are the deformed analogues of the standard position and momentum opera-
tors. The operator Q f has the following action on the basis vectors:
 
xn xn+1
Q f |en  = |en1  + |en+1  . (6.69)
2 2

1
If now the sum n=0 diverges, the operator Q f is essentially self-adjoint and
xn
hence has a unique self-adjoint extension, which we again denote by Q f . Let Ex ,
x R, be the spectral family of Q f , so that,

Qf = x dEx .

Thus there is a measure dw(x) = de0 |Ex e0  on R such that the vectors en can be
realized as elements pn (x) in L2 (R, dw). Furthermore, on this Hilbert space, Q f is
just the operator of multiplication by x and consequently, the relation (6.69) assumes
the form

xn
xpn (x) = cn pn1 (x) + cn+1 pn+1 (x) , cn = . (6.70)
2

This is a two-term recursion relation, familiar from the theory of orthogonal

polynomials and thus the functions pn (x) are polynomials, that may also be obtained
by orthonormalizing the sequence of monomials 1, x, x2 , x3 , . . . , with respect to the
measure dw(x) (using a GramSchmidt procedure):

en |e  = pn (x)p (x) dw(x) = n . (6.71)
R

Then, for any w-measurable set R,


en |E( )e  = pn (x)p (x) dw(x) . (6.72)

Also setting z = |z, the nonlinear coherent states (6.59), as functions of x,

become

zk
z (x) = f (|z|)1/2 [xk !]1/2 pk (x) . (6.73)
k=0
6.3 Some Examples of Coherent States from Square Integrable Kernels 149

The polynomials pn are not monic polynomials, i.e., the coefficient of xn in pn is

not one. However, the renormalized polynomials

qn (x) = cn ! pn (x) , (6.74)

are seen to satisfy the recursion relation

qn+1 (x) = xqn (x) c2n qn1 (x) , (6.75)

from which it is clear that these polynomials are indeed monic.

It is immediately seen that for the canonical coherent states, with xn = n,
this procedure would give us the well-known Hermite polynomials. Similarly, the
coherent states (4.93)(4.99) arising from the discrete series representations of the
SU(1, 1) group, give rise to the Pollaczek polynomials [20].
Additionally, the function

zn 2n/2 zn
G(z, x) = xn ! pn (x) = qn (x) , (6.76)
n=0 n=0 xn !

is the generating function for the polynomials qn (or pn ), in the sense that

1 xn ! n G(z, x)
qn (x) = . (6.77)
2n/2 n! zn z=0

It is interesting to note that, for fixed z, the generating function G(z, x) is the
functional form of the (non-normalized) nonlinear coherent state, when expressed
as a vector in L2 (R, dw), the Hilbert space spanned by the orthogonal polynomials
pn . In other words, for fixed z, G(z, x) gives a sort of coordinate space representation
of the nonlinear coherent state (analogous to writing the canonical coherent states
in the Hermite polynomial basis of L2 (R, ex dx)).
2

There is a simple way to compute the monic polynomials qn . To see this, note
first that by virtue of (6.69) and (6.70), the operator Q f can be represented in the en
basis as the infinite tridiagonal matrix,

0 c1 0 0 0 ...
c . . .
1 0 c2 0 0

0 c2 0 c3 0 . . .
Qf =
0 0 c3 0 c4

. . . . (6.78)

0 0 0 c4 0 . . .

.. .. .. .. .. ..
. . . . . .

Let Qn be the truncated matrix consisting of the first n rows and columns of Q f and
In the n n identity matrix. Then,
150 6 Square Integrable and Holomorphic Kernels

x c1 0 0 0 ... 0 0 0
c1 x c2 0 0 ... 0 0 0

0 c2 c3 ...
x 0 0 0 0

0 0 c3 x c4 ... 0 0 0

In Qn = 0
.
0 0 c4 x ... 0 0 0 .
(6.79)
. .. .. .. .. .. .. .. ..
. . . . . . . . .

0 0 0 0 0 . . . x cn2 0

0 0 0 0 0 . . . cn2 x cn1
0 0 0 0 0 . . . 0 cn1 x

It now follows that qn is just the characteristic polynomial of Qn :

qn (x) = det[x In Qn ] . (6.80)

Indeed, expanding the determinant with respect to the last row, starting at the lower
right corner, we easily get

det[x In Qn ] = x det[x In1 Qn1 ] c2n1 det[x In2 Qn2 ] , (6.81)

which is precisely the recursion relation (6.75). Consequently the roots of the
polynomial pn (or qn ) are the eigenvalues of Qn .
There is a second set of orthogonal polynomials related to the family of nonlinear
coherent states (6.59), which can be obtained by symmetrizing the measure d in
(6.61), giving the resolution of the identity, to define it over the whole of R and
looking at its moments. It is then interesting to study the relationship between this
family of orthogonal polynomials and the pn above [20].
Note that the reproducing kernel which arises from the coherent states (6.64) is

(zz )n
K(z, z ) = , (6.82)
n=o xn !

and the corresponding reproducing kernel Hilbert space HK is a subspace, consisting

of functions which are holomorphic in the domain DL , of the larger Hilbert space
L2 (DL , d (z, z)). In addition, introducing the vectors [see (6.64)]

zn
| z  = [xn !]1/2 |en  (6.83)
n=0

in H, we easily see [in view of the resolution of the identity (6.65)] that the map

W : H HK , (W )(z) = z | , H, (6.84)
6.4 GazeauKlauder CS 151

is unitary, embedding H into a holomorphic subspace of L2 (DL , d (z, z)). Under

this unitary map, the basis vectors |en  H transform as
zn
(W |en )(z) := un (z) = , n = 0, 1, 2, . . . . (6.85)
[xn !]1/2
Finally, a word about quantization on the domain DL using nonlinear coherent
states. The coherent state or BerezinToeplitz procedure, that we will discuss in
detail in Chap. 11 is particularly efficient in the holomorphic context. We briefly
describe here a somewhat different, but entirely equivalent procedure for carrying
out such a quantization using the Hilbert space HK of holomorphic functions
introduced above.
Let PK be the projection operator mapping the Hilbert space L2 (DL , d (z, z)) to
its holomorphic subspace HK , which as we know, is a reproducing kernel subspace.
Let f be a sufficiently well behaved, real valued function on DL (a classical
observable on the phase space DL ). Let fmult denote the operator of multiplication
by the function f on L2 (DL , d (z, z)). Then, the restricted operator

O f = PK fmult PK (6.86)

is the corresponding quantized operator, acting on the Hilbert space HK , now

identified as the Hilbert space of quantum states. It can then be shown that

Of = f (z, z) |z z | d (z, z) , (6.87)
D

where the z = W z are the images in HK of the vectors (6.83), under the unitary
map W . The action of O f on a vector HK is easily computed to be

(O f )(z) = K(z, z ) f (z , z ) (z ) d (z , z ) , (6.88)
D

with K as in (6.82).
Several examples of sequences {xn } leading to such frames may be found in
[35, 312]. In some cases, the moment problem can be solved explicitly, in other
ones, only approximately.

6.4 GazeauKlauder CS

6.4.1 Coherent States for Discrete Spectrum Dynamics

Suppose that we deal with a quantum system described by a positive Hamiltonian

H with purely discrete nondegenerate spectrum,

H|en  = xn |en , n 0, (6.89)

152 6 Square Integrable and Holomorphic Kernels

where the strictly increasing sequence 0 = x0 < x1 < x2 < < xn < consists of
pure numbers and can be finite or infinite. Let us denote by H the Hilbertian span of
the orthonormal set {|en }n0 . We then consider the unit norm vector in H:

1 J n/2
|J,  = & eixn |en ,
N (J) n0 n
(6.90)

where (n ) is, like in the previous section, a positive sequence with 0 = 1, J 0

and < < . It is clear that we have another sort of generalization of the

standard coherent states |z,where z = Jein , n = n!, and N (J) = e|z| . This
2

generalization is different of the one described in the previous section since we now
abandon any requirement of holomorphy. Let us examine if some of the standard
CS properties listed in Chaps. 1 and 2 are still valid in the present case.
1. Eigenvector property: Property (1.5) formally holds with the lowering operator
defined by
&
a |en  = n /n1 |en1 , eiH / a eiH / |J,  = J|J,  , (6.91)

for n 1, and a |0  = 0.
2. Temporal stability: We easily check that the standard CS family is temporally
stable: eiHho t |z  = |ei t z , where Hho is the shifted harmonic oscillator
Hamiltonian with eigenvalues n. Here we have,

eiHt |J,  =| J, + t . (6.92)

3. Normalization: This property imposes the power series

Jn
N (J) = n (6.93)
n0

to be non trivial, which means that the convergence radius R = limn n n >
J 0 is nonzero.
4. Continuity: The continuity of the map [0, R) R % (J, )  |J,  H holds
true.
5. Resolution of the identity: This is the crucial point, because we have to solve
a classical moment problem. Indeed, n > 0 is assumed to be a probability
distribution moment
 R
n = J n w (J) dJ, w (J) 0. (6.94)
0

Then the resolution of the identity holds in the following sense:


I= |J,  J, | d (J, )
  R (6.95)
1
= lim d N (J)w (J) dJ |J,  J, |.
2 0
6.4 GazeauKlauder CS 153

It often happens that the determination of the distribution w , with support in [0, R],
is quite intractable, while another one, say w  , with the same support, and giving
rise to the moments
 R
n = J n w  (J) dJ, w  (J) 0, (6.96)
0

may be also relevant to our purposes. Of course n should be equivalent to n in the

sense that both series N (J) and N  (J) have same convergence radius. So let us
see what happens when we replace w by w  in the d of (6.95). Suppose that
both the positive sequences (n := n /n ) and (n1 ) are bounded. We now get the
resolution of a positive diagonal operator (N), |en  = n |en ,

= |J,  J, | d  (J, )
 (6.97)
I = ( )1/2 |J,  J, |( )1/2 d  (J, ),

where d  (J, ) = N (J) w  (J) dJ d . This operator should be bounded with

bounded inverse so that N (J) and N  (J) do have same convergence radius.
Hence, the triple ( , {|J,  }, d  ) is a frame.

6.4.2 Statistical and Semi-Classical Aspects

At this point we may notice the remarkable interplay between two statistical aspects
of the states (6.90):
1. Probability distribution w (J):
We could say this w (J) pertains to the classical (statistical) side of the states
|J,  if we give the measure d (J, ) in (6.95) a classical phase space inter-
pretation, which would be perfectly allowable in the standard case. As a matter
of fact, let us associate to a quantum observable A its (non-unique) components
 ) as seen from the continuous coherent frame {|J,  J, |}:
A(J,

A=  )|J,  J, | d (J, ).
A(J, (6.98)

 ), which exists for a large class of operators A, is also called

The function A(J,
upper (see the contribution of Lieb in [Fen94]) or contravariant (Berezin in
[135, 136]) symbol. In particular, for an observable f (N) diagonal with respect
to {|en }, we have

f (N) = f(J)|J,  J, | d (J, ),
 R
1
which implies f (n) = f(J)J n w (J) dJ. (6.99)
n 0
154 6 Square Integrable and Holomorphic Kernels

2. Wave-packet weighting probability :

Alternatively, the moment n is part of the purely quantal weighting probability

1 Jn
|n |J,  |2 |cn |2 = . (6.100)
N (J) n

Let us now associate to the observable A its lower (or covariant) symbol which
is nothing but its mean value in the state |J,  :

A(J, ) = J, |A|J,  . (6.101)

In particular, we have for the diagonal observable f (N):

f(J) = |cn |2 f (n)  f (n). (6.102)

n0

Therefore, the weighting probability determines lower symbols of diagonal observ-

ables. It also determines the semi-classical revival structure of the coherent state
|J,  . Let us briefly recall what we mean by this temporal concept of revival and
refer to the comprehensive review [153] for more details. Suppose the expansion
(6.90) is strongly weighted around a mean value N = n for the number operator
N. For instance, in the case of the harmonic oscillator, the distribution |cn |2 =
e|z| |z|2 /n! is Poissonian and is weighted around N = |z|2 with spread |z| =
2

N. In the general case (6.90), let us denote by n N the closest integer to N for a
given J. With the assumption that the spread N is small compared with N n,
what we should expect in general for large J (semi-classical behavior), we expand
the energy En = xn in a Taylor series in n around the centrally excited value n:
1 1
En En + En (n n) + En (n n)2 + En (n n)3 + , (6.103)
2 6
where each prime on En denotes a derivative. These derivatives define distinct
time scales: the classical period Tcl = 2 /|En |, the revival time trev = 2 / 12 |En |,
the super-revival time tsr = 2 / 16 |En |, and so on. Inserting this expansion in the
evolution factor ei En / of (6.90) allows us to understand the possible occurrence
of a quasiperiodic revival structure of these coherent states at high J. The latter
would be well illustrated by the behavior as a function of the evolution parameter
of the absolute square of the so-called autocorrelation function of the coherent
states:

| J, 0|J,  |2 = | J, 0 | eiH / | J, 0 |2 . (6.104)

It is clear from (6.103) that the revival features will be more or less apparent
according to the value of the deviation (n n) relatively to n. A quantitative estimate
is provided by the so-called Mandel parameter Q(J) defined as follows.
6.4 GazeauKlauder CS 155

( N)2 d d
Q(J) = 1 = J log log N (J). (6.105)
N dJ dJ

This parameter allows one to compare the weighting distribution |cn |2 with the
Poisson distribution. The latter is
Poissonian if Q = 0,
super-Poissonian if Q > 0,
sub-Poissonian if Q < 0.
We are now aware of the basic importance of the distribution w (J) in regard to
these statistical considerations.

6.4.3 Imposing the Hamiltonian Lower Symbol

So far the only physical quantity we have at our disposal is the Hamiltonian.
Suppose there exists an explicit functional form

H(J) = J, |H|J,  = |cn |2 xn = xn , (6.106)

n0

for the lower symbol of the Hamiltonian. We could for instance consider as a
prerequisite the classical energy-action dependence for systems with one degree
of freedom,
+
1
J= p(E = H) dq, (6.107)
2
and this way deserves to be carefully explored in order to study the (semi-) classical
behavior of the corresponding quantum system. Suppose that H(J) has a series
expansion in the neighborhood of J = 0:
H(J) = J n J n . (6.108)
n0

Then, by simple identification with the expansion

xn
N (J)H(J) = n J n , (6.109)
n0

we obtain
n
xn 1
= 0, 0 = 0, = nm , 0 = 1. (6.110)
n m=0 m

The solution n to (6.110) can be written in terms of the factorial xn ! = x1 x2 xn ,

x0 ! = 1, as follows:
n
xn ! xn1 !
n =
pn
, with pn = nm , p0 = 1 . (6.111)
m=0 xm !
156 6 Square Integrable and Holomorphic Kernels

For instance,
x1 x1 x2 x1 x2 x3
1 = , 2 = , 3 =
1 x1 2 + 12 x1 2 + (x1 + x2 )1 2 + 13

Let us give a few explicit examples.

The linear case H(J) = J: It is the simplest one, of course, and gives rise to
the so-called action-angle coherent states, introduced in [309]. Here the n s are
exactly the factorials:

n = xn ! . (6.112)

The single power case H(J) = J q , q = 2, 3, : Then the corresponding CS do

not generate the whole Hilbert space H:

1 J nq/2
|J,  = &
N (J) n0 xq x2q xnq
eixnq |enq . (6.113)

The geometric case H(J) = /(1 J): Then the denominator pn in (4.7) is
given by

n1
pn = xi1 xi2 xik . (6.114)
k=1 1i1 <i2 <<ik n1

Let us end this section by noting that the converse to the above approach is also
of great interest, at least from a more experimental point of view. Suppose the
coherent states |J,  are prepared in such a way that the weighting distribution
n
|cn |2 = N1(J) J n ( e.g. a Gaussian-like distribution) is determined by the experimen-
tal conditions. Then the coefficients of the lower symbol series (6.108) will be given
by inverting (6.110): 0 = 0, n = nm=0 cm (1 , , n )xm .

6.4.4 Action-Angle Coherent States

Let us now concentrate on the linear case H(J) = J, which amounts to require that
the mean value (or lower symbol) of the Hamiltonian mimics the classical relation
energy-action in the case of the harmonic oscillator: J, |H|J,  = J. We shall
adopt here the simplified notations:

1 J n/2
|J,  = &
N (J)
xn ! eixn |en , (6.115)
n0
6.4 GazeauKlauder CS 157

where N (J) is the generalized exponential

Jn
N (J) = xn ! . (6.116)
n0

The algebraic structure exemplified by (6.91) may now display nontrivial interesting
features. Indeed, it is clear from the definition of lowering and raising operators

a|en  = xn |en1 , a |en  = xn+1 |en+1 , (6.117)

that the Hamiltonian factorizes as

H = a a x(N), x(N)|en  = xn |en . (6.118)

Now, given any lowering operator a, a|0  = 0, of the above type, and any diagonal
(real) operator f (N), f (N)|en  = fn |en , the following commutation rules hold true:

a f (N) = f (N + 1)a, f (N)a = a f (N + 1),

[a, f (N)] = f  (N)a = a f  (N 1), [a , f (N)] = a f  (N) = f  (N 1)a ,

where f  (N) := f (N + 1) f (N) is the finite difference derivative of f (N). Of

course, the sense of the above any should be compounded by operator domain
considerations. In the present case, we also have

[a, a ] = x (N). (6.119)

So we can think about the existence of an interesting dynamical or spectrum

generating Lie algebra generated by {a, a , x (N), . . .}, where the dots mean further
generators coming from iterating the above commutators. This algebra is generically
infinite. However, nontrivial structures, such as a quadratic algebra, can occur when
the operator function x(N) is a solution of some required finite difference equation.
This approach also leads to the many types of nonlinear CS generated by Lie
algebras obtained by deforming the Heisenberg Lie algebra {a, a , I}. Examples of
such structures are, for instance, discussed in [217, 226, 227, 364].

6.4.5 Two Examples of Action-Angle Coherent States

6.4.5.1 Coulomb-Like Spectrum

Our first example concerns a Coulomb-like discrete spectrum:

 
1 1 n+2
xn = 1 , for which n = xn ! = . (6.120)
(n + 1)2 2 n+1
158 6 Square Integrable and Holomorphic Kernels

Thus the convergence radius of the generalized exponential is R = 1. This series can
be given the closed form:

2! 1 ln(1 J) "
E(J) = + . (6.121)
J 1J J

The distribution probability w(J) is explicitly given by:

 1
1
n = J n w(J) dJ where w(J) = (1 + (J 1 )). (6.122)
0 J

On the other hand, the weighting distribution is given by

n+1 (1 J)J n+2

|cn |2 = . (6.123)
n + 2 J + (1 J) ln(1 J)

This distribution is super-Poissonian. This is easily inferred from the estimate on

the parameter [560]:

xn+1 /xn
< 1 Q < 0. (6.124)
(n + 1)/n

6.4.5.2 Infinite Square Well and PschlTeller Potentials

The article [77] is devoted to the examples of the infinite square well and Pschl
Teller potentials. So we just give here a brief account of this nice illustration of
many features of the action-angle coherent states. We consider the Hamiltonian

P2 h2
H= +V (Q) ( + )2 , (6.125)
2m 8ma2
for a particle on the line submitted to the potential

0, 0 < x < a,
V (x) = (6.126)
, x 0, x a,

for = 1 = (infinite square well),

2  
h ( 1) ( 1)
+ , 0 < x < a,
V (x) = 8ma2 cos2 x/2a sin2 x/2a (6.127)

, x 0, x a,

for > 1, > 1 (PschlTeller potential). Note that we have reintroduced

the Planck constant into the physical quantities expressions. In a certain sense,
the PschlTeller potential [509] interpolates between the harmonic oscillator and
6.4 GazeauKlauder CS 159

the infinite square well. We shall here treat them both on the same footing, since the
infinite well is obtained in the limit = 1 = . Eigenvalues and eigenvectors are
given by:

h
H|en  = h xn |en , xn = n(n + + ), = ,
2ma2
 x |en  = kn (cos x/2a) (sin x/2a) 2 F1 (n, n + + ; + 1/2; sin2 x/2a).
(6.128)

Then the salient points of [77] may be summarized as follows.

1. Algebraic side:
We put x(N) = N(N + + ). The Lie algebra generated by {a, a , x (N), } is
finite and is su(1, 1), as in fact for any algebra built from an x(N) quadratic in N.
Indeed,

[a, a ] = x (N) = 2N + 1 + + , [a, x (N)] = 2a, [a , x (N)] = 2a . (6.129)

Note that x (N) = 2, x (N) = 0. Therefore, we can identify the Hilbertian span
of the |en s via the values assumed by the su(1, 1) Casimir operator:

+ +1 3
C = 2x (N)(2x (N)1)+x(N) ( 1)I = . (6.130)
2 2

This range of values falls into the discrete series representation interval 1/2 <
< for su(1, 1) or, equivalently, for the universal covering of SU(1,1). The
value = 3/2 corresponds to the infinite well.
2. Coherent states:
They read in the present case:

1 J n/2
|J,  = & eixn |en ,
N (J) n0 xn !
(6.131)

n! (n + + 1)
n = xn ! = , + 2, (6.132)
( + 1)

I (2 J)
E(J) = ( + 1) /2 , R = , (6.133)
J
where I is a modified Bessel function. Note that the = 2 case corresponds to
the infinite well. At this point, we should remark that, for = 0 and for complex
values of J, these coherent states are nothing else but the BarutGirardello
states [126]. However, the construction of the latter was motivated by purely
mathematical considerations, more precisely by the aim to extend to the algebra
su(1, 1) the eigenvector property.
160 6 Square Integrable and Holomorphic Kernels

3. Resolution of the identity:

 
1
I = lim d E(J)w(J) |J, J, | dJ ,
2 0

2J /2
with n = J n w(J) dJ, w(J) = K (2 J), (6.134)
0 ( + 1)

where K is a modified Bessel function.

4. Infinite square well:
Let us end this section by some considerations on the specific case of the infinite
well [77], since it is the simplest one and should not qualitatively differ from the
PschlTeller cases. The weighting probability |cn |2 is sub-Poissonian and reads

1 J n+1
|cn |2 = . (6.135)
n!(n + 2)! I2 (2 J)

The revival structure [107] is encoded by the expansion

h2 h2
En = 2
n(n + 2) = [n(n + 2) + 2(n + 1)(n n) + (n n)2 ], (6.136)
2ma 2ma2

thus the two characteristic times are the classical period Tcl = 2 ma2 /h(n + 1)
and the revival time trev = 4 ma2 /h = 2(n + 1)Tcl . There is no super revival
time here. We refer to [77] for a systematic study of the temporal behavior of
the probability density, |x|J, |2 with = t, of the square modulus of the
autocorrelation, |J, 0|J, |2 , of the mean position and mean momentum and
their respective dispersions, and so on. On the same topic, we may also quote the
recent paper [396], which contains a large number of references and the papers
[141,142], where the study of these potentials is made with different types of CS.

6.4.6 Coherent States for Continuum Dynamics

We now turn to the continuum case. Suppose the quantum system has a Hamiltonian
with positive non degenerate purely continuous spectrum,

H|E = E|E, 0 E < E, (6.137)

where |E denotes some eigendistribution. Mimicking the discrete case, we define
the normalized coherent states as the continuous superposition

1 E J E/2 i E
|J,  = & & e |E dE. (6.138)
N (J) 0 (E)
6.4 GazeauKlauder CS 161

The normalization condition reads

 E JE
N (J) = dE, (6.139)
0 (E)

which has to be finite for, say, 0 J < R. Equation (6.139) gives the ratio
J E /(N (J) (E)) the meaning of a weighting distribution.
These coherent states clearly have continuity and temporal stability properties.
For having the resolution of the identity, we first suppose the existence of a
nonnegative density w (J) such that
 R
J E w (J) dJ = (E). (6.140)
0

Then we have,
   R
I= |J,  J, | d (J, ) d N (J)w (J) |J,  J, | dJ. (6.141)
0

Note the simplification in the integration provided by the continuous spectrum.

Like in the discrete case, there is an interplay between the functional dependence of
the lower symbol and the weighting distribution:

1 E EJ E
H(J) = J, |H|J,  = dE = J log N (J) (6.142)
N (J) 0 (E) J
 J dx 
i.e., N (J) = N (J0 ) exp H(x) . (6.143)
J0 x

An explicit example of a situation analogous to the one described in this section

may be found in [131, 132].

6.4.7 Coherent States for Discrete and Continuum Dynamics

Finally, we briefly present the more general case of a Hamiltonian having both
discrete and continuous spectra:

H = HD HC , 0 HD , < HC : (6.144)

We introduce for this situation the following superposition of respective coherent

states:

| J, ; K, ; ,  = f (K, )|J,  + ei g(J) |K,  , (6.145)

162 6 Square Integrable and Holomorphic Kernels

where |J,  (resp. |K,  ) corresponds to the discrete (resp. continuous) spec-
trum. Continuity and temporal stability follow from this expression:

eiHt | J, ; K, ; ,  =| J, + t; K, + t; + t, + t . (6.146)

Deriving the resolution of the identity,


I= | J, ; K, ; ,  J, ; K, ; , | d , (6.147)

where

d (J, ; K, ; , ) := dD (J, ) dC (K, ) h( ) d d /2 ,

is more elaborate and explains the presence in (6.145) of the functions f , g, and the
parameters and . The ranges of values assumed by the latter are respectively
(, ) and [0, 2 ). The conditions the three functions f , g, h, have to satisfy are
the following:

| f (K, )|2 |J,  J, | d = ID , (6.148)

|g(J)|2 K, | K, | d = IC , (6.149)
 2 
d ei g(J) f (K, )|J,  K, | d = 0, (6.150)
0

  
with h( ) d = 1, | f (K, )|2 dC = 1, |g(J)|2 dD = 1.

Again here, a predefinite functional form assigned to the lower symbol

H(J, K, , , ) := | f (K, )|2 HD + |g(J)|2 HC

implies determining constraints for linear combination coefficients and normaliza-

tion factors, i.e., for the weighting distributions present in the construction.

6.4.8 A General Expression

Through the successive generalizations of coherent states presented in this section,

we have been becoming gradually aware of the existence of some unifying for-
malism which underlies all various constructions. So let us conclude by proposing
the following general form. Let a quantum system be with one degree of freedom
and with Hamiltonian H. We suppose that the spectrum of H, Sp(H), is bounded
from below (as it should be!), and we choose its lower bound equal to zero. Let
6.5 CS on Quaternionic Hilbert Spaces and Hilbert Modules 163

be a complex measure with support contained in Sp(H). We then consider the linear
superposition of Hamiltonian eigenstates or eigendistributions | , H |  = | ,
in a possibly abusive notation:

|  := |  d ( ), (6.151)

with the following requirements on :

1. Temporal stability:
The complex nature of the measure should ensure the temporal stability:

eiHt |  = |ei t . (6.152)

2. Normalization:

d 2
1 =  |  = d , (6.153)
d

with again a possibly abusive notation for the density d / d of the measure .
3. Operator resolution:

A= |  | d ( ), (6.154)

where A is some positive bounded operator, e.g. A = P P where P is a projection

and defines a frame.
4. Lower symbol identity:

 |H|  = H( ). (6.155)

The extension of the formalism presented here to systems with several degrees of
freedom is possible, and rather straightforward for completely integrable systems.
As a matter of fact, this approach has opened the door to a vast literature,
extending the theory to different potentials, like power-law potentials, q-deformed
CS, supersymmetric CS, and so on. For recent developments, see for instance
[385, 392] and references therein.

6.5 CS on Quaternionic Hilbert Spaces and Hilbert Modules

Finally, we should mention two non-conventional constructions of coherent states

and square-integrable type of kernels on spaces other than complex Hilbert spaces,
that have recently appeared in the literature. The first is the problem of quaternionic
coherent states. Coherent states on a quaternionic Hilbert space (i.e., a Hilbert space
which is a vector space over the field of quaternions), as analogues of the canonical
164 6 Square Integrable and Holomorphic Kernels

CS, were first constructed in [3, Adl95] and later also in [584]. This work is within
the context of building quantum mechanics and field theory on quaternionic Hilbert
spaces. Exploiting some recent work [327, 328, Gen13] on extending the concept
of holomorphy to functions of a quaternionic variable, CS of the holomorphic and
nonlinear types, on quaternionic Hilbert spaces, obeying resolutions of the identity
and possessing reproducing kernels were obtained in [583]. Quaternionic CS is
a direction of research which promises to have implications to wavelet analysis
as well.
The second area of generalization involves building coherent states on Hilbert
C -modules. These are vector spaces, similar to Hilbert spaces, but in which the
inner product takes values in C -algebras. Coherent states can be constructed on
such spaces [37], which obey resolutions of the identity as well as possessing
reproducing kernels. As such they are a far reaching generalization of the standard
notion of coherent states. It is surmised that such CS could have applications in the
theory of representations of quantum groups and quantum stochastic processes.
Chapter 7
Covariant Coherent States

Abstract This chapter returns to a group-theoretical context, namely, a systematic

study of CS associated to group representations. After a general definition of these
covariant CS, we describe the well-known GilmorePerelomov CS, as well as vector
and matrix CS. Some generalizations are mentioned, including continuous semi-
frames. We conclude the chapter by a thorough description of two interesting cases.
First we treat CS on spheres constructed via heat kernels (such CS are not of the
GilmorePerelomov type). Next we turn to CS on conformal classical domains, i.e.,
classical domains associated to the conformal group SO(n, 2).

By now we have encountered several examples of coherent statesthe canonical

CS, associated to the WeylHeisenberg group, discussed in detail in Chap. 2, vector
CS, introduced in Sect. 4.2.1, CS associated to the discrete series representations
of SU(1,1), in Sect. 4.2.2, general CS associated to any reproducing kernel Hilbert
space in Sect. 5.4.2, holomorphic nonlinear CS in Sect. 6.3, action-angle CS in
Sect. 6.4. At this point, certain common features are already seen to emerge:
1. Coherent states form an overcomplete family of vectors in a Hilbert space, and
define a reproducing kernel on it;
2. The existence of the reproducing kernel enables one to map the Hilbert space
unitarily to a space of (vector-valued) functionsin quantum mechanical lan-
guage, the CS become wave functions on some appropriate space (usually, the
phase space of the underlying classical system);
3. Group representations often facilitate the construction of such families, the
individual CS being now labelled by the points of some homogeneous space of
the groupthereby inheriting covariance properties under the action of the
group;
4. Square integrability of the reproducing kernel, when it is present, reflects special
properties of the functions constituting the reproducing kernel Hilbert space and
of the group representation.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 165
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__7,
Springer Science+Business Media New York 2014
166 7 Covariant Coherent States

In this and the next few chapters, we embark on a systematic study of CS

associated to group representations. We begin by giving a general definition of such
CS and indicate how it is broad enough to encompass all examples of group-related
CS found in the literature. Later, in Chap. 8, we pass on to a detailed treatment of
square integrable representations of groupsor representations emanating from the
discrete seriesand CS associated to them. In Chap. 9 we specialize to a class of
groups of the semidirect product type, admitting square integrable representations,
and for which many of the quantities appearing in the general theory can be
explicitly calculated.

7.1 Square-Integrable Covariant Coherent States

At the expense of being pedantic, we wish to reiterate that generally, coherent

states are simply a preferred overcomplete family of vectors in a reproducing kernel
Hilbert space. There is no assumption of group covariance or square-integrability in
the general setting. However, most coherent states arising in physics are associated
to symmetry groups of the system. We introduce, therefore, a general definition to
encompass all such group related coherent states.

7.1.1 A General Definition

As before, let G be a locally compact group, the left Haar measure on it, H a
closed subgroup of G and X = G/H the corresponding left coset space, with a quasi-
invariant measure . For a global Borel section : X G of the group, let be
the unique quasi-invariant measure defined in (4.19):

d (x) = ( (x), x) d (x), where (g, x) d (x) = d (g1 x), (g G).

Definition 7.1.1 (Covariant coherent states).

(i) Let U be a unitary representation of G in a Hilbert space H and i , i =
1, 2, . . . , n, a set of linearly independent vectors in H. Then, if the set of vectors,

S = {i (x) = U( (x)) i : i = 1, 2, . . . , n, x X}, (7.1)

is total in H, we call it a family of covariant coherent states for U.

(ii) Let F = ni=1 | i  i |, F (x) = U( (x))FU( (x)) , and suppose that

F (x) d (x) = A , A , A1
L (H), (7.2)
X
7.1 Square-Integrable Covariant Coherent States 167

the integral converging weakly. Then the coherent states S are called square
integrable and the representation U is said to be square integrable mod(H, ).
In this case, the vectors i , i = 1, 2, . . . , n, are called admissible mod(H, ) and
the operator F a resolution generator for the representation.
Unless otherwise stated, the number n of vectors i (which are often called
fiducial vectors [409, 410, Kla85]) will be assumed to be finite. Given a family of
covariant CS, square integrable or not, there is an associated reproducing kernel

K (x, y) = F (x)1/2 F (y)1/2 , (7.3)

and a reproducing kernel Hilbert space HK , to which H is unitarily equivalent as a

consequence of Theorem 5.4.8 (see, in particular (5.92)(5.96); see also [428]).
Consider now the sections g : X G, g G, which are covariant translates of
under g:

g (x) = g (g1 x) = (x)h(g, g1 x), (7.4)

where h is the cocycle defined in (4.37). Let g be the measure dg (x) =

(g (x), x) d , again constructed using (4.19), and Fg (x) = U(g (x))FU(g (x)) .
If U is square integrable mod(H, ), there is a general covariance property enjoyed
by the positive operators defined by the weak integrals,

ag ( ) = Fg (x) dg (x), g G, B(X). (7.5)

Indeed, it is easily verified [using (4.8)] that the covariance condition

U(g)a ( )U(g) = ag (g ), U(g)A U(g) = Ag (7.6)

holds. In the case where X admits a left invariant measure m and one takes = m,
and

U(h)FU(h) = F, h H, (7.7)

it is an immediate consequence of (7.4) that U(g)A U(g) = A , for all g G. If

the representation U is assumed to be irreducible, Schurs lemma (see Sect. 4.3)
then implies that A = I, where is a positive real number. Moreover, in this
case, the operators a ( ) do not depend on a situation which has already been
encountered for vector CS in Sect. 4.2.1and thus, the pair {U, a} forms a system
of covariance in the sense of Definition 4.2.1. We now call the representation U
simply square integrable mod (H) and the vectors i , i = 1, 2, . . . , n, admissible
mod (H).
168 7 Covariant Coherent States

7.1.2 The GilmorePerelomov CS

A ready example of a family of covariant coherent states is obtained for any unitary
representation U of the group G, which admits a cyclic vector (a fortiori if it
is irreducible). In this case, the set of vectors g = U(g) , g G, is dense in H
and hence satisfies condition (i) of the Definition 7.1.1, with H = {e}, the trivial
subgroup consisting of the identity element of G, (so that X = G), the trivial section
(g) = g and n = 1. Thus the orbit of any cyclic vector is a family of coherent states,
which do not in general have to satisfy condition (ii), i.e., they are not necessarily
square integrable. Furthermore, square integrability in this case would imply that

| g g | d (g) = I, (7.8)
G

for some real number (and a similar relation would hold if the right invariant
measure were to be substituted for d ).
On the other hand, consider the closed subgroup H of G which stabilizes the
cyclic vector up to a multiplicative factor, i.e.,

U(h) = exp [i (h)] , h H, (7.9)

where is a real-valued function on H (the fact that H is a subgroup is trivial to

verify.) It follows that (h) = exp [i (h)] is a character of H, since clearly

(h1 h2 ) = (h1 ) (h2 ), h1 , h2 H, and (e) = 1. (7.10)

Letting X = G/H and taking any Borel section : X G, we may define another
family of covariant CS

(x) = U( (x)) , x X, (7.11)

since this set of vectors is also total in H. Such CS will be referred to as the Gilmore
Perelomov coherent states [330,331,502,Per86]. Note that if now the representation
is also square integrable in the sense of (7.8) and the measure is actually
invariant under the action of G on X, then H is necessarily a compact subgroup
of G, although just square integrability mod(H, ) will not ensure this. From a
mathematical point of view, it is natural to consider the notion of square integrability
of a representation of a group G modulo its center Z(G). While this is commonly
done in the mathematical literature [Bor72], it is however unduly restrictive. The
more general setting mentioned above was introduced, independently, by Gilmore
[330, 331] and Perelomov [502]. It was anticipated, to some extent, by Klauder
[409, 410], using the argument that coherent states arise in quantum theories, and
there, as is well-known, states are represented not by individual vectors in a Hilbert
space H, but rather by rays (one-dimensional subspaces of H). Upon normalization,
a quantum state is thus represented by a family { = exp [i ]o : [0, 2 )}.
7.1 Square-Integrable Covariant Coherent States 169

Take now a coherent state g = U(g) , g G, of the GilmorePerelomov type;

the corresponding quantum state does not depend on g, but rather on its equivalence
class modulo H. Although a section is not explicitly used in either [330, 331] or
[502], it appears implicitly in the form of a particular factorization of the elements
of G. Indeed take g G and let x = gH be the coset to which it belongs (or,
equivalently, its canonical projection in the principal bundle G X = G/H). Then
any Borel section : X G defines a factorization g = (x)h(x), with h(x) H,
and conversely, there always exists a factorization g = go (g)h(g), with h(g) H,
defining a Borel section (x) = go (g).
Before passing to some concrete examples, we might point out that the formu-
lations of Gilmore [330, 331] and Perelomov [502, Per86] differ in the assumptions
they make concerning G,U and . Gilmore considers a general symmetry group
G, but restricts U to be a square integrable representation and a highest (or
lowest) weight vector of U, while Perelomov requires G to be a Lie group, which
allows for the use of the Lie algebra, e.g. in the choice of the basic vector (see
Sect. 4.5 below). Coherent states are then defined through (7.11), initially without
any square integrability assumptions, thus allowing U to be any unitary irreducible
representation (including those of the principal or the complementary series), and
any normalized vector in the representation space.
Both the canonical coherent states (2.37) and the CS arising from the discrete
series representations of SU(1,1) [see (4.99)] are examples of GilmorePerelomov
coherent states. Indeed, the form of s in (2.6) and of the representation
U ( , q, p) = ei U (q, p) of the WeylHeisenberg group in (2.33) tell us that
with H = , the phase subgroup,

U ( ) s = ei s , . (7.12)

Thus, using the section (q, p) = (0, q, p) in (2.36), we identify the canonical CS,
s (q,p) = U (q, p) s in (2.37), as being precisely of the GilmorePerelomov type.
If, as noted in Sect. 2.2, we derive the canonical CS from a unitary representation of
the oscillator group H(4), then, choosing again for the ground state |0, one finds
that H = U(1) U(1), and the construction proceeds as above, with X C [617].
In the case of the discrete series representation U j of SU(1,1) given in (4.86), if
we take H = K U(1), the maximal compact subgroup [see (4.65)], and the vector

= (2 j 1)1/2 u0 , with u0 (z) = 1, z D , (7.13)

then once again,

U j (k) = ei j , k = diag(ei 2 , ei 2 ) K , (7.14)

by (4.88). Hence the CS, (z) = U j ( (z)) in (4.99), generated by the vector ,
are again of the GilmorePerelomov type.
170 7 Covariant Coherent States

CS associated to square integrable representations of semisimple Lie groups,

which are of the GilmorePerelomov type, will be discussed more systematically
in Sect. 7.2.

7.1.3 Vector and Matrix CS: A Geometrical Setting

For vector coherent states (VCS), which we defined in Sect. 5.1 and of which we
examined a square integrable version in Sect. 4.2.1, there is an obvious generaliza-
tion of the GilmorePerelomov CS. Going back to the definition of covariant CS, let
K be the n-dimensional subspace of H generated by the vectors i , i = 1, 2, . . . , n.
Assume that K is stable under the action of the operators U(h), h H, i.e.,
U(h) K, for all h H and K. Then the functions [see (5.99)], : X Cn ,
with components i (x) = i (x) | , i = 1, 2, . . . , n, for all H, constitute a
reproducing kernel Hilbert space with matrix-kernel [see (5.97)] K(x, y), having
matrix elements

K(x, y)i j = i (x) | j (y) , i, j = 1, 2, . . . , n . (7.15)

The mapping,

W : H HK , (W )i (x) = i (x) = i (x) | , i = 1, 2, . . . , n, x X, (7.16)

is a Hilbert space isometry. Denoting the image of U(g) under W (as an operator
on HK ) by U (g), we find as in (4.60)(4.62),

(U (g) )(x) = V (h(g1 , x)) (g1 x), HK , x X, (7.17)

where, as defined in (4.54), V (h), h H, constitutes an n-dimensional unitary

representation of H on Cn . Since in Sect. 4.2.1 we had assumed [see (4.58)] square
integrability of the kernel K(x, y), and consequently there was the associated system
of covariance (4.59), it was possible to extend the representation U to an induced
representation U  on the Hilbert space Cn L2 (X, d ), of which HK then became a
subspace. The conditions for ensuring square integrability will be taken up again
in Sect. 8.4 below. In general however one does not have square integrability,
although the form of the representation U in (7.17) resembles that of an induced
representation. The vectors i (x) , i = 1, 2, . . . , n, x X, are called VCS and the
representation U a VCS-representation [241244, 513, 514, 543, 544, 546, 547]. In
[546] the subspace K is not assumed to be finite-dimensional, nor is one necessarily
restricted to representations on Hilbert spaces.
Let us emphasize again that the GilmorePerelomov CS are a special case of the
VCS when the dimension of the subspace K is unity. Moreover, while the VCS arise
7.1 Square-Integrable Covariant Coherent States 171

from representations of the subgroup H on the subspace K, for general covariant

CS, Definition 7.1.1 given above does not require K to carry a representation of H.
It is possible to give an invariant geometrical description of VCS [14], using
vector bundles, which generalizes a similar construction proposed for the Gilmore
Perelomov type CS, employing complex line bundles, in [518]. In order to present
this construction, it is necessary to use some elementary notions from the theory of
principal bundles and their associated vector bundles. (An easy introduction to the
subject may, for instance, be found in [vWe86]).
The point of this construction is to obtain the VCS as sections in a vector bundle
associated to the principal bundle : G X, i.e., the bundle giving the natural
fibration of G over the homogeneous space G/H. (Here denotes the projection
map (g) = gH.) For the purposes of this discussion, we shall use the symbol
[g] to denote the left coset gH = x X = G/H and assume that U is a unitary
representation of G on H, admitting a finite dimensional cyclic subspace K, which is
stable under U when the latter is restricted to H, i.e., the set {U(g) : K, g G}
is total in H and,

U(h) K , for all K, h H. (7.18)

Let PK be the projection operator for the subspace K and write

U(h)PK = V (h), h H. (7.19)

We also assume that the representation h  V (h) of the subgroup H acts irreducibly
on K.
Looking upon G as a principal bundle, fibred over X and with structure group H,
we consider the G-homogeneous associated bundle B = G V (H) K. This bundle, is
also fibred over X, and is obtained by identifying elements (g, ), (g ,  ) G K,
whenever g = gh and  = V (h)1 for some h H. Thus, elements in B can be
put into one to one correspondence with vectors g = U(g) H, for all g G
and K. In the terminology of [518], these vectors may be called a family
of coherent vectors. In view of the cyclicity of the subspace K with respect to
the representation U, it should be possible to extract a suitable subset of vectors,
from among this set, which would be the CS of this representation, satisfying the
conditions of Definition 7.1.1, and we shall do so momentarily. Identifying the
bundle B = G V (H) K with this family of coherent vectors, the canonical projection
B : B X in it is, very simply, B (g ) = [g]. Sections in this bundle can be put into
one to one correspondence with functions f : G K, which satisfy

f(gh) = V (h)1 f(g), h H. (7.20)

Consider next the functions : G K, defined by

(g) = PKU(g)1 , H. (7.21)

172 7 Covariant Coherent States

Clearly these functions satisfy (7.20) and hence also describe sections of B.
We denote this class of sections (for all H) by HK , and then it is clear that
HK has the natural structure of a reproducing kernel Hilbert space, with reproducing
kernel K : G G L (K) (= bounded operators on K),

K(g, g ) = PKU(g)1U(g )PK . (7.22)

Indeed, since K(g, g ) = K(g , g) and as an operator on K, K(g, g) is strictly positive,

K is a strictly positive definite kernel and therefore, following Theorem 5.4.3, the
relationship

K( , g) | K( , g )  K =  | K(g, g )  H , (7.23)

can be used to define a scalar product [see (5.55)],  | K , on the linear span of
sections of the type K( , g) , for all g G and K, and which will then extend
to all sections of the type (7.21). Note that the section K( , g) assumes the value

K(g , g) = PKU(g) g K , (7.24)

at the point g G. Note also, that the measurable field of Hilbert spaces, postulated
by Theorem 5.4.3, is in this case {Kg : g G}, with Kg = K for all g G. It then
follows that HK is complete in the above scalar product, hence a reproducing kernel
Hilbert space, and the map W : H HK , such that

(WK )(g) = (g) = PKU(g)1 , g G, H , (7.25)

is unitary. Clearly,

WK g = K( , g) , g G, (7.26)

meaning that the coherent vectors are mapped to sections in the associated bundle
B and the linear span of this family of sections defines a reproducing kernel
Hilbert space HK , which is unitarily equivalent to the original space H of the
representation U. The reproducing property of the kernel HK is reflected in the
evaluation condition for functions in HK at points g G:

K( , g) | K =  | (g)H , H, g G. (7.27)

Let { i }ni=1 (n = dimension of K) be an orthonormal basis for K and : X G

a (global) Borel section of G. Then, the set of vectors

S = {i ([g]) : i = 1, 2, . . . n, [g] X}, (7.28)

is total in H, and forms a family of VCS for the representation U. This, then, is
the differential geometric interpretation of vector CS. They arise as (in general
measurable) sections in a vector bundle, naturally associated with the group
representation in question.
7.1 Square-Integrable Covariant Coherent States 173

We shall also be interested in conditions for the square integrability mod(H) of

these VCS, i.e., conditions to ensure a resolution of the identity of the type:
n  
| i ([g]) i ([g]) | d ([g]) = PK ([g]) d ([g]) = I < , > 0,
i=1 X X
(7.29)
where we have written, with obvious justification,

PK ([g]) = U(g)PKU(g) . (7.30)

In this case, however, it will be necessary to impose the condition that the measure
be invariant on X. We shall return to this point in Sect. 8.4, but let us note here that
from our discussion on induced representations in Sect. 4.2, we know that the set of
sections f in (7.20), which are square integrable with respect to the invariant measure
, forms a Hilbert space LK
2 (X, d ), on which the representation g  V U(g), g G,

(V U(g)f)(g ) = f(g1 g ), g G, (7.31)

is unitary (by virtue of the invariance of ). Furthermore, this is just the standard
induced representation (4.45) of G (induced from the UIR V of the subgroup H).
Suppose now, that the reproducing kernel Hilbert space HK forms a subspace
of LK2 (X, d ). Then, the unitary map W in (7.25) will clearly intertwine the
K
representation U (of G on H) with the induced representation V U, and thus, U
would be a subrepresentation of an induced representation. Additionally, if all the
sections of the type (7.21) are elements of LK 2 (X, d ), then for any two sections

,  HK ,

 (g) |  (g)H d ([g]) < , (7.32)
X

and in particular, using sections of the type (7.26),

 
K(g , g) | K(g , g )  H d ([g ]) =  |U(g1 )PK ([g ])U(g)   d ([g ])
X X

=  | PK ([g ])  ; d ([g ]) < , (7.33)
X

for arbitrary ,  K, the last equality following from the invariance of the
measure . Thus, we arrive at the condition

I( )  | PK ([g])  d ([g]) < , (7.34)
X

for all K. It will be shown in Theorem 8.4.5, that this condition is both necessary
and sufficient for the square integrability mod(H) of the representation U of G.
174 7 Covariant Coherent States

7.2 Example: The Classical Theory of Coherent States

The classical theory of coherent states, largely due to Gilmore [330, 331] and
Perelomov [502], applies primarily to semisimple and nilpotent Lie groups. We will
briefly survey it in this section, referring the reader to the monograph of Perelomov
[Per86] for a more detailed presentation and concrete examples, together with
original references.

7.2.1 CS of Compact Semisimple Lie Groups

Let G be a compact semisimple Lie group, with Lie algebra g. Then any unitary
irreducible representation of G is square integrable and finite dimensional. So the
only question, for constructing CS is: What is the best choice for the basic vector
in a given physical situation?
An obvious requirement is to try and find a vector that yields CS which are
closest to the classical states, in other words, minimal uncertainty states. A crite-
rion for that was found independently by Perelomov [502, Per86] and by Delbourgo
and Fox [245, 246], and generalized further by Spera [565]. If is a highest (or
lowest) weight vector in the chosen representation, then the corresponding coherent
states have minimal uncertainty. This means that they minimize the dispersion C2
of the quadratic Casimir operator C2 = ni,k=1 gik X i X k , that is, the quantity:

n
C2 = C2  gik X i  X k , (7.35)
i,k=1

where . denotes the mean value in the state , namely A =  |A| .
We illustrate this again on the simplest example, namely SU(2), already men-
tioned in Sect. 4.5.1. In the unitary spin- j irreducible representation D( j) , the vector
| j, j is a lowest weight vector, i.e. J | j, j = 0) and has an isotropy subgroup
U(1). Then the corresponding CS, D( j) (g)| j, j, have minimal uncertainty and are
indexed by points of SU(2)/U(1) S2 , the Bloch sphere of quantum optics. These
are the spin CS or Bloch states, widely used in atomic physics, quantum optics
[87, 515, Scu97] and quantum information [420, 421, Ben07].
In fact this choice of a highest weight vector is a particular case of a general
criterion introduced by Perelomov [502, Per86], which points towards some deep
geometric properties of CS. Namely, if the isotropy subalgebra b gc of is
maximal in the sense of Perelomov (here gc denotes the complexified Lie algebra of
g, see Sect. 4.5.1), that is, b b = gc , then generates minimal uncertainty CS.
Coming back to SU(2), it is easy to see that b = {J0 , J }, Gc = SL(2, C),
H , Z are the one parameter subgroups generated by J0 , J , respectively, and X+ =
c

Gc /B X = B+ \Gc C S2 (the last homeomorphism being the stereographic

projection from the Riemann sphere onto the complex plane). Then the Gauss
7.2 Example: The Classical Theory of Coherent States 175

decomposition (4.148) of elements of Gc , g = z+ hz , z Z , h H c , yields a

corresponding factorization for the representation operators D( j) (g):

D( j) (g) = exp(zJ+ ) exp(xJ0 ) exp(z J ). (7.36)

Choosing the lowest weight vector | j, j, with isotropy subalgebra b = {Jo , J },
one gets for the corresponding CS:

|z = D( j) (g)| j, j = c j exp(zJ+ )| j, j, zC, (7.37)

with c j a suitable constant. The same technique is widely used in nuclear physics,
for the unitary groups SU(n), in particular SU(3) and U(3) (which, while not
semisimple, can be treated in the same way). We refer to [Kla85] for original
references or [355] for a more recent one.
On the other hand, as mentioned in Chap. 1, (spin) CS associated to SU(2) have
been constructed in [105], by the BarutGirardello method.

7.2.1.1 An Example: Spin CS and the HeppLieb Model

As mentioned in Sect. 2.7, a system of N two-level atoms interacting with a radiation

field exhibits a phase transition, radiance superradiance, at a certain critical
temperature Tc , when the coupling is strong enough. This result was obtained in
the simplest approximation by Wang and Hioe [599] with the help of canonical CS,
and derived rigorously by Hepp and Lieb by standard methods [371]. In a further
paper [372], the last two authors obtained a far reaching generalization, this time
using Bloch atomic or spin CS [87, 515, Scu97]. The difference between the two
approaches may be caricatured by saying that canonical CS make the radiation field
classical, while the atomic CS make the atomic variables classical.
The mathematical problem, as usual, resides in the calculation of the partition
function. One starts with the Hamiltonian

H = H0 + HI , (7.38)

where H0 and HI are the free and interaction Hamiltonians respectively:

H0 = Hrad + Hat
1 N
= k ak ak + zj , (7.39)
k 2 j=1
 
N N
1
HI = ( k ak )( + j ) + ( k ak )( j ) .

(7.40)
2 L k j=1 k j=1
176 7 Covariant Coherent States

written for N two-level atoms interacting with M photon modes (k = 1, . . . , M).

Here, ak and ak are the creation and annihilation operators, respectively, for the
kth radiation mode with frequency k , is the two-level energy shift, and the
atom-field coupling strength. Two level atoms are described by Pauli matrices

j = j i j ,
x y
(7.41)

     
01 0 i 1 0
xj = , yj = , zj = .
10 j
i 0 j
0 1 j

The quantity to be estimated is Z = Tr e H , where the trace is over the tensor

product of the Fock space F of the photon modes and the Hilbert space H of the N
spin-1/2 atoms. The key point is that the total spin, J, is a constant of the motion,
so that
N/2
Z= Y (N, J) Z(J), Z(J) = Tr e H(J) , (7.42)
J=0

where H(J) is the Hamiltonian for a spin, S, of magnitude J, and the combinatorial
factor Y (N, J) = N!(2J + 1)[(J + 1 + 12 N)!( 12 N J)!]1 is the number of ways to
construct an angular momentum J from N spin-1/2 particles. At this point, one uses
the following bounds, obtained in [445] with atomic CS,

 (2J + 1)1 Z(J) Z(J

Z(J)  + 1), (7.43)

where


 = (4 )1 Tr
Z(J) F
e H(J, ) d ,

 ) is the Hamiltonian H(J) with S replaced by J

and H(J,  ( is a point on the unit
sphere). Calculating the trace explicitly, one then finds that, for large N and fixed

M, the free energy ln Z(J) coincides with that obtained for a single mode in the
rotating-wave approximation, but with an effective coupling constant
M
2 = max |2k cos |2 k1 ,
k=1

that is, the result of Wang and Hioe [599].

Next, atomic CS may be used to understand the thermodynamical stability of a
system with infinitely many modes. The result is that the system is stable, that is,
Z exp NA, for some constant A, iff the atoms have very repulsive cores. Otherwise,
a lower bound may be derived, namely,
2
Z > eCN , C > 0.
7.2 Example: The Classical Theory of Coherent States 177

This is true irrespective of whether the atomic coordinates are treated classically or
quantum-mechanically. A final result of [372] is to extend the treatment to a finite-
mode case, where the atoms have translational degrees of freedom. In each of these
various situations, the reasoning always rests on estimates obtained in [445] using
atomic CS.

7.2.2 CS of Noncompact Semisimple Lie Groups

If G is a noncompact semisimple Lie group, the square integrable representations are

exactly those of the discrete series. For such a representation, all that we have said
for compact groups goes through, except for the definition of minimal uncertainty
states. Indeed, the quadratic Casimir operator C2 is not positive definite for a
noncompact group, so that the minimization of C2 can no longer be interpreted
as optimal localization of the state, which is the crucial semiclassical feature of
CS [240]. However, in the case of SU(1,1), it turns out that the correct CS not only
minimize C2 , but they also saturate another inequality directly derived from the
commutation relations, of the general form

1
A B [A, B], (7.44)
2
and which fully justifies their quasi-classical characterization. As a matter of fact,
they share this property with the BarutGirardello CS [126], mentioned in Chap. 1.
We shall examine these CS of SU(1,1) in great detail in Sect. 10.3, in the study of the
contraction process from an Anti-deSitter to a Minkowskian space-time (SU(1,1)
SO(1,2) is the Anti-deSitter group in 1+1 space-time dimensions) [240, 519]. See
also [386].
Applications of CS, coming from noncompact semisimple Lie groups, are again
numerous. For instance, besides the example just mentioned, SU(1,1) CS have been
used for deriving a consistent path integral formalism [Ino92]; CS for Sp(3,R) and
other noncompact symplectic groups underlie the detailed structure of many nuclei
[543, 544, 546, 547] (see below). Another striking application is the long standing
open problem of the hydrogen atom CS. The question was originally asked by
Schrdinger whether one could find quantum states of the atom sufficiently well
localized on the Kepler orbits such that they would converge to their classical
counterparts in the semiclassical limit h 0. Schrdingers idea was to find
states analogous to the corresponding states of the harmonic oscillator, namely the
canonical CS, but he was not successful.
Now, it is well known [127, 494] that the conformal group SOo (4, 2) is the
dynamical group of the hydrogen atom, in a regularized sense (see Sect. 7.6 below).
By this, one means that the group is generated by the observables of the system,
and is such that a single UIR of it yields the entire energy spectrum. In the present
case, the H-atom, both classical and quantum, has an accidental degeneracy, due
178 7 Covariant Coherent States

to the existence of two conserved vector observables, the angular momentum (as
for any central potential) and the RungeLenz vector. As a consequence, the group
that explains the degeneracy of the spectrum is SO(4), for the discrete spectrum,
and SOo (1, 3) for the continuous spectrum [4345, 120]. More generally, the de
Sitter group SOo (4, 1) may be taken as the invariance group of the system, when
it is bounded, in the sense that a single UIR of the latter describes the whole
discrete spectrum. In the same way, the Anti-de Sitter group SOo (3, 2) is the
invariance group of the unbounded system and, similarly, P+ (1, 3) = SOo (3, 1) 
R4 , the Poincar group, for the zero energy case.
On top of that, if one adds ladder operators, describing electromagnetic dipole
transitions between different energy levels, one obtains the group SOo (4, 2). Then
the relevant UIR of the latter remains irreducible when restricted to SOo (4, 1). On
that basis, various CS associated to SOo (4, 2) or its subgroup SU(1,1) have been
proposed as candidates for the states sought by Schrdinger [128], but no really
satisfactory solution has been found. It was even suggested that the problem has no
solution, because the SOo (4, 2) CS cannot be well localized [618]. Thus the problem
remains open. It is worth mentioning here that a different solution to this problem,
using CS not related to any group, was presented in [309].

7.2.2.1 An Example: CS for the Nuclear Collective Model

The microscopic theory of nuclear collective structure, a refinement of the tradi-

tional BohrMottelson collective model of nuclei, has a long history and a huge
literature surrounds it. A comprehensive survey and the original references may be
found in the review papers [544, 545]. Group theory plays a central rle in these
models.
The starting point is to realize that the relevant observables for a nucleus contain
both positions and momenta, which obey the CCR (2.1). Therefore, the symmetry
group necessarily contains canonical transformations in R3 , that is, the group
Sp(3,R) (sometimes called Sp(6,R)). More precisely, one starts with an independent
particle Hamiltonian

1 A1 3 2 1 A1 3
H= si 2
2m s=1
P + m 2
Q2si +V (Q), (7.45)
i=1 s=1 i=1

where A is the number of nucleons (the center of mass motion has been removed).
The first two terms in (7.45) simply constitute a harmonic oscillator Hamiltonian
and V (Q) is a collective potential. One then notes that the model is defined in terms
of the following observables: the six quadrupole moments Qi j , the generators of
deformations and rotations Si j , Li j and the six components of the quadrupole flow
tensor Ki j , where
7.2 Example: The Classical Theory of Coherent States 179

Qi j = Qsi Qs j ,
s

Si j = (Qsi Ps j + Psi Qs j ),
s
(7.46)
Li j = (Qsi Ps j Qs j Psi ),
s

Ki j = Psi Ps j .
s

Using again the CCR, one finds that the operators (7.46) generate the Lie algebra
sp(3, R), and the Hamiltonian (7.45) is a polynomial in these generators.
The next step is to build an irreducible representation of this Lie algebra. The
standard technique, pioneered by Rowe (see also [241244]), is to use the vector
coherent state method, that we have sketched in Sect. 4.2.1. Using this approach,
one is then able to build an orthonormal basis in the representation Hilbert space,
adapted to the subgroup chain

Sp(3, R) U(3) U(1) SU(3) SO(3) (7.47)

(SU(3) is the invariance group of the Elliott model, a precursor of the present
collective model). Basis vectors of this kind can be chosen as eigenstates of the
Hamiltonian and with definite angular momentum. In that way, CS of Sp(3,R) are
indeed the building blocks of a host of different collective models of the nuclei.
For further details, we refer to [544] or to [545], which contains a host of recent
references.

7.2.3 CS of Non-Semisimple Lie Groups

For non-semisimple groups, the emphasis has been put on the existence of CS
systems based on a coset space X = G/H (see [Per86, Sect. 10.2] for a review).
For G a nilpotent Lie group, rather complete results have been obtained in [473],
making extensive use of Kirillovs method of orbits [Kir76]. Some of these results
remain valid when G is a solvable or a reductive Lie group [474].
Let G be a Lie group of one of these types. Under the coadjoint action, the space
g* is foliated into orbits, as usual, but, in addition, each orbit is associated to a
single unitary irreducible representation of G. In addition, every coadjoint orbit is a
symplectic manifold, making it a natural candidate for a classical phase space and
hence a starting point for a quantization procedure. This is the basis of the method
of geometric quantization [Woo92], within the framework of which much work on
CS has been done [447, 487, 488, 518, 565]. Coadjoint orbits have been used also
for the construction of CS for inhomogeneous groups such as the Euclidean, Galilei
and Poincar groups [18, 57, 237], where in addition to quantization, they have also
180 7 Covariant Coherent States

been studied in connection with notions of quantum localizability and covariance

[554]. Actually each of these examples is the semidirect product of a translation
group Rn by the corresponding homogeneous group. Accordingly, we will discuss
them in detail, albeit mostly by the method of induced representations, in Chap. 9.
An exception is the affine WeylHeisenberg group, studied in [400] by the method
of coadjoint orbits. We will describe this example briefly in Sect. 16.1.

7.3 Covariant CS: The General Case

It is true that the GilmorePerelomov CS cover a large number of interesting

situations, of which Perelomovs monograph [Per86] offers a vast choice of
examples. Yet this class is not general enough for all physical purposes. In practice,
the index space X = G/H is often specified a priori, on physical grounds, and it
may be that H is not the stability subgroup of any vector in the representation
space. For instance, in (geometric) quantization [Woo92], one takes for X a phase
space, which could be a coadjoint orbit of G or, more generally, a symplectic
G-space [237]. This is the situation for the kinematical or relativity groups
(Euclidean, Galilei or Poincar), that we will analyze in detail in Chap. 9. Inter-
estingly enough, there are phase spaces that do lead to GilmorePerelomov CS. A
notable example is the theory of wavelets, that we shall study closely in Chaps. 12
and 14.
In the more general situations, one has to exploit the theory of covariant CS
in almost its full generality, starting from Definition 7.1.1. The only restriction is
that the unitary representation U is usually assumed to be irreducible and square
integrable mod (H, ), for some Borel section . In principle, all the relevant results
may be read off Chaps. 5 and 6, in particular Theorem 5.4.8 and Sect. 6.1. However,
it may be useful to summarize these results in the form in which we will use them in
Chaps. 9 and 1416. This has the additional advantage of facilitating the comparison
with the existing literature.
To make things slightly simpler, we shall assume here that the measure on X is
invariant. The general case of a quasi-invariant measure is then easily deduced from
our previous analysis.
Theorem 7.3.1. Let G be a locally compact group, H a closed subgroup of G, and
X = G/H, equipped with an invariant measure . Let : X G be a global Borel
section. Let U be a unitary irreducible representation of G in a Hilbert space H.
Assume that the representation U is square integrable mod(H, ), that is, there
exist n linearly independent vectors i , i = 1, 2, . . . , n in H such that the integral
n 
U( (x))| i  i |U( (x)) d (x) (7.48)
i=1 X
7.3 Covariant CS: The General Case 181

converges weakly to a bounded positive self-adjoint operator A , with bounded

inverse A1
. Then the family of CS

S = {i (x) = U( (x)) i : i = 1, 2, . . . , n, x X} (7.49)

has the following properties:

(1) The set S is total in H.
(2) The range of the linear map WK : H Cn L2 (X, d ), given componentwise by

(WK )i (x) = i (x) | , H, (7.50)

is a (closed) subspace HK of Cn L2 (X, d ). It is complete with respect to the

new scalar product

 | K :=  |WK A1 1
WK Cn L2 (X, d ) , (7.51)

and then WK : H HK is unitary.

(3) The projection PK = WK WK from Cn L2 (X, d ) onto HK is an integral
operator, and HK is a reproducing kernel Hilbert space of functions:

(x) = K (x, y) (y) d (y), HK . (7.52)
X

The kernel of PK is given explicitly by the matrix K (x, y)i j = i (x) | A1

(y),
j

and it is square integrable:

n 
K (x, y)i j K (y, z) jk d (y) = K (x, z)ik , (7.53)
j=1 X

(4) Being unitary, the map WK may be inverted on its range by the adjoint operator,
WK1 = WK on HK , to obtain the reconstruction formula [see (6.5)]:

n 
= WK1 = i (x) A1
(x) d (x), HK .
i
(7.54)
i=1 X

For practical purposes, it may be useful to rewrite the square integrability condition
in the form of the familiar frame condition
n 
m(A ) 2 |i (x) | |2 d (x) M(A ) 2 , H , (7.55)
i=1 X

where the frame bounds have been denoted m(A ), M(A ), because they are in fact
the infimum and the supremum, respectively, of the self-adjoint operator A .
182 7 Covariant Coherent States

Suppose that the Hilbert space H corresponds to the description of a quantum

system, which admits G as a symmetry group. Then, in more familiar terms,
properties (1) and (2) mean that the CS family S yields a new representation of
the system under consideration, unitarily equivalent to the original one. By property
(3), the Hilbert space of this CS representation is a space of (usually well behaved)
functions. Finally, property (4) means that the vector = WK1 may be expanded
in (or reconstructed from) the CS i (x) . This is the interpretation in terms of which
the reconstruction formula is most often used, for instance in signal processing
(see Chap. 12). Notice that part of this construction is standard in noncommutative
harmonic analysis, where the CS map (7.50) is sometimes called the Poisson
transform [470].

7.3.1 Generalized Perelomov CS

Besides the usual GilmorePerelomov CS discussed in Sect. 7.1.2, Perelomov also

considers CS associated to non-square integrable unitary representations, namely,
the so-called class-I representations [Per86, Sect. 13.1] or [471, 503]. This means,
unitary irreducible representations for which there exists a vector 0 invariant under
the maximal compact subgroup K G.
If U is a class-I representation and g1 , g2 are two elements of G in the same coset
modulo K, i.e., g1 = g2 k, for some k K, clearly one has

U(g1 )0 = U(g2 )U(k)0 = U(g2 )0 ,

so that the family of CS {U(g)0 , g G} is indexed by the points of the quotient

space X = G/K. Here again, one may consider a Borel section : X G and write
the CS as U( (x))0 = U(g)0 , but the result does not depend on the choice of
section.
In particular, Perelomov studies class-I representations of the principal series,
which can be realized as induced representations, acting in the Hilbert space L2 (X ).
Here X = G/P = K/M, where G = KAN is the Iwasawa decomposition of G and
P = MAN its minimal parabolic subgroup (see Sect. 4.5.2). These representations
are obviously not square-integrable.
According to Cartan [Hel78], such quotients exhaust the irreducible Hermitian
symmetric spaces, that is, those that have a complex structure. In the case of rank-1
spaces, there are three series and one exceptional case, namely,
1. The real n-dimensional hyperbolic space XnI = SO(n, 1)/SO(n);
2. The complex n-dimensional hyperbolic space XnII = SU(n, 1)/SU(n) U(1)
(real dimension = 2n);
3. The quaternion n-dimensional hyperbolic space XnIII = Sp(n, 1)/Sp(n) Sp(1)
(real dimension = 4n);
7.3 Covariant CS: The General Case 183

4. A two-dimensional hyperbolic space over the octonions X IV = FR

4 /SO(9) (real
dimension = 16)
In addition, all three series 13 can be realized as bounded homogeneous domains,
that can be treated in a uniform way simply by the substitutions R C H (the
quaternions). Similar results holds in the case of symmetric spaces of higher rank
[Per86, Sect. 13.6] or [471,503] (the rank of the symmetric space G/K is the number
of independent invariants determined by any pair of its points, and it equals the
dimension of the subgroup A in the Iwasawa decomposition of G). This offers an
interesting point of view on all these CS. We will come back to some of them in
Sect. 16.2.4.

7.3.2 Continuous Semi-Frames

In the case of a representation that is square integrable mod(H, ), we had to

introduce a section : G/K G and the set of corresponding CS will in general
depend on the choice of the section . For the generalized Perelomov CS, on the
contrary, the representation is not square integrable and we also need a section, but
the result does not depend on the choice of the section.
The next step is to combine both difficulties. If the representation U is not
square integrable mod(H, ), and if by analogy with (7.2), we formally write (for
simplicity, we take here n = 1 and assume that the measure is invariant)

| (x)  (x) | d (x) = A , (7.56)
X

then A is not a bounded operator perhaps not even a bona fide operator on
the Hilbert space H. However, it may be possible to give a meaning to A as a
sesquilinear form on K K, where K is some dense domain in H. An interesting
possibility is to reformulate this statement in the language of rigged Hilbert spaces
(RHS) [Gel64], as we did already in Sect. 3.3.2. This means that one would consider
the triplet of spaces

K H K , (7.57)

where K denotes the topological conjugate dual of K, in a suitable locally convex

topology. Then a sesquilinear form over K K may be identified with a continuous
map from K into K .
An example of this approach may be found in [496] (actually, this paper
treats vector CS, not only plain CS). Namely, the authors generalize the Gilmore
Perelomov CS by allowing the stability subgroup H of the (fiducial) vector to
be noncompact. As for Perelomov, he gives only examples where H is compact,
in particular in the case where the indexing manifold is G/K, with K the maximal
184 7 Covariant Coherent States

compact subgroup of G, that is, when G/K is a bounded homogeneous domain, as

explained in Sect. 7.3.1 above.
In fact, Perelomov [Per86, Chaps. 67] mentions explicitly the RHS formulation
in the case of covariant CS associated to class-I representations of the principal
series of SOo (3, 1), the Lorentz group, or SOo (n, 1), but does not develop it.
Let us now come back to the symbolic relation (7.56). There are three cases.
1. The operator A is bounded, self-adjoint and invertible, but its inverse A1 is
not bounded (it is necessarily densely defined, as the inverse of a self-adjoint
operator). In that case, A defines an upper semi-frame, as we have discussed
in Sect. 3.3.2 (this notion was introduced briefly in the first edition of this book
(Sect. 7.3.1) under the name unbounded frame). Assuming this semi-frame to be
regular, that is, all the vectors i (x) , i = 1, 2, . . . , n, x X, belong to the domain
D(A1
), we obtain the following triplet of Hilbert space s:

HK HK HK . (7.58)

In this relation, HK denotes, as before, the completion of WK [D(A1 )] with

respect to the norm  K , which is the graph norm of the self-adjoint operator
1/2
WK A WK1 acting in Cn L2 (X, d ). Then HK is the closure of HK in Cn
L2 (X, d ), and HK is the conjugate dual of HK , that is, the completion of HK in
the norm  |WK A WK1 Cn L2 (X, d ) . If i (x) D(A1
), i = 1, 2, . . . , n, x X,
then all three spaces HK , HK , HK  are reproducing kernel Hilbert spaces, with the
same reproducing kernel K (x, y). A concrete example (Th.Pauls wavelets on
the half-line) was given in Sect. 3.3.2.
2. The operator A is self-adjoint and invertible, but unbounded (of course, densely
defined), and its inverse A1
is bounded. Then we obtain a lower semi-frame,
as described in Sect. 3.3.2. An explicit example (wavelets on the two-sphere) is
given in Sect. 15.1.1.1.
3. A is not densely defined or not even a bona fide operator on the Hilbert space.
In that case, we have nothing to say, except that we dont know any concrete
example of this bad situation.

7.3.3 Some Further Generalizations

Before concluding this section, it may be worthwhile to make contact with several
notions related to CS that have cropped up in the literature.
(1) Weighted CS
In the discussion at the end of Sect. 6.1, we have seen that it is always possible to
transform a non-tight frame, with AK = IK , into a tight one, by mapping the Hilbert
space HK unitarily onto another one HK with help of the well-defined operator
7.3 Covariant CS: The General Case 185

1/2
AK . This corresponds to the replacement of (6.3) by (6.12). However the abstract
CS on both sides do not map into each other. When applied to the present situation,
this move corresponds to the definition of weighted CS introduced in [28], namely
the vectors

xi = AK1/2 i (x) .

(7.59)

While these vectors do constitute a tight frame, they are not covariant CS, since they
1/2
have been weighted along the orbit {U( (x)) : x X} by the operator AK . This
procedure becomes more suggestive and useful if, and probably only if, the operator
AK is diagonal in Hk , that is, a multiplication operator by a function AK (x). This
indeed happens for a number of physically interesting cases, such as the Poincar
group or the affine WeylHeisenberg group, that we shall analyze in subsequent
chapters.
At this point, we ought to mention that weighted frames, i.e., frames (n )
with complex weights (n ) such that the sequence (k k ) is again a frame,
are well-known in the discrete case. They were introduced in the PhD thesis of
Jacques [Jac04] and then taken over in [154], in order to get a numerically more
efficient approximation algorithm for frame expansions into spherical wavelets (see
Sect. 15.1.1). A systematic analysis was then performed in [112], in particular
concerning the choice of optimal weights. Closely related is the notion of frame
multipliers, which are the operators defined as

Mm, , f = mk  f | k k (7.60)
k

for the frames = (n ) and = (n ) and the weight sequence m = (mk ) (see [154]
for further references).
The interesting point here is that both notions, weighted CS as well as frame
multipliers, can be extended to the general case. Weighted CS are obvious. Let be
a measurable function on X. Then the vectors i (x) form a weighted CS system, in
particular, a weighed frame, if the following vectors constitute a CS system, resp. a
CS frame
xi := (x)i (x) .
(7.61)

Actually the same definitions apply when {i (x) } is only an upper semi-frame
(the map x  i (x) is sometimes called a continuous Bessel mapping). Using this
language, given two frames or upper semi-frames x , x and a measurable function
, the operator defined weakly as

M , , = (x) | x x | d (x) (7.62)
X

is called a continuous frame (or Bessel) multiplier. A systematic analysis has been
made in [114], including finding conditions under which the operator M , , is
compact, trace class, HilbertSchmidt, etc.
186 7 Covariant Coherent States

(2) Quasi-coherent states

It may happen that the square integrability condition (7.2) is not satisfied as
it stands, but that the integral does converge when the translated measure is
replaced by some other quasi-invariant measure. Then the formalism goes through,
as developed in Chap. 5, but one loses covariance properties with respect to
translation of the section, such as (7.6) [31].
(3) Quasi-sections
Finally, it may happen that no section is admissible, but that the square integra-
bility condition is satisfied when the base space X is suitably reparameterized, i.e.
when is replaced by the composition : X G, where : X X is
a homeomorphism (reparametrization) of the base space. This map, called a quasi-
section of the principal bundle : G G/H = X, satisfies the condition = .
Here again the whole machinery of Chap. 5 applies and CS may be defined. A case
where this more general approach is needed is that of the massless representations
of the Poincar group P+ (1, 1) in one space and one time dimensions (see [57] and
Sect. 10.1).
Of course, all these variants are particular cases of the general theory developed
here, and of unequal importance per se. But, from a historical point of view, they
have contributed to expand the scope of the theory, thus paving the way to its
present status. Note, however, that weighted CS are fundamental if we insist on a
probabilistic interpretation of the POV measure (e.g. in quantum mechanics). Also
the lack of covariance, which leads to quasi-coherent states, may sometimes be very
meaningful from a physical point of view (e.g. for Poincar CS, that we will discuss
in Sect. 10.1).

7.4 CS on Spheres Through Heat Kernels

The classical SegalBargmann transform is a unitary transform between the Schr-

dinger and Fock representations of quantum mechanics on Euclidean space. As was
explained in Sect. 2.4 the standard coherent states are the kernel of this integral
transform, and this kernel appears as a heat kernel. Indeed, we know that a solution
of the heat equation on Rn ,

u
(u = 0 , (7.63)
t
can be analytically continued to an entire function on Cn through the classical
SegalBargmann transform

1
f (y) e(xy)
2 /4t
L2 (Rn , dx) % f (x)  (et( f )(x) = dy
(4 t)n/2 Rn
7.4 CS on Spheres Through Heat Kernels 187

 ! e|Im z| /2t dz "

2
1
f (y) e(zy)
2 /4t
dy O(Cn ) L2 Cn , , (7.64)
(4 t)n/2 Rn (2 t)n/2

where O(Cn ) denotes the holomorphic functions on Cn . This is an isometric

isomorphism of Hilbert spaces, identifying L2 (Rn ) with a reproducing kernel space
of entire functions on Cn , known as the FockBargmannSegal Hilbert space.
In [361] Hall has generalized the SegalBargmann transform by replacing Rn
with a compact Lie group G equipped with a bi-invariant metric (see [492] for
further developments). In this context, he has presented a family of coherent states
for a system whose classical configuration space is the group manifold of G. These
states are labeled by points in the associated phase space, namely the cotangent
bundle T (G), and are defined in terms of the heat kernel on G. One may identify
T (G) with the complexified group Gc . For instance, SL(2, C) is the complexified
of SU(2). Such coherent states have been applied to quantum gravity in [100]
with various successive generalizations essentially due to Thiemann [579582]. An
important objective of these works was to determine the classical limit of the so-
called loop quantum gravity (LQG) theory (see for instance the review [495]).
These coherent states for the group G are not of Perelomov type, but instead are
realized as eigenvectors of certain non-self-adjoint annihilation operators. In this
sense they are of the BarutGirardello type. In this section, we give an account of
the Hall method (see also Stenzel in [569]) with the case of rotation groups SO(d)
and spheres viewed as symmetric spaces Sd = SO(d + 1)/SO(d). In [361,362], Hall
describes these states using the complexifier point of view of Thiemann (for the
general case) and generalizing an independent approach followed by Kowalski and
Rembielinski for S2 [424, 425].
The phase space for the motion of a particle on the sphere Sd , of radius r, is
viewed as the co-tangent bundle

T (Sd ) = {(x, p) Rd+1 Rd+1 : x2 = r2 , x p = 0} . (7.65)

Let us show how the complexifier method allows to identify this symplectic
manifold with the complex sphere in Cd+1 . In this section, devoted to an example
of mechanical system, we (re)introduce physical units as a guideline for specifying
the relations between classical and quantum versions of the system.
On the classical level, a complexifier is the classical observable

kinetic energy J2
= ,
2m r2

where J 2 is built from the components Jkl of the angular momentum in Rd+1 :

J 2 = (Jkl )2 , Jkl = pk xl pl xk . (7.66)

k<l
188 7 Covariant Coherent States

In terms of the (d + 1) (d + 1)-matrix J = (Jkl ), we have the relations Jx = r2 p,

Jp = p2 x . The constants m (mass), (frequency), r (length) provide the studied
system with physical units or standards. Complex coordinates of the manifold
T (Sd ) are obtained through the following exponentiated symplectic action of the
kinetic energy:
/ +  n
J2 0 i 1
ak = exp i ,
2m 2 r2
xk := 2m r2 n! { {{xk , J 2 }, J 2 }, . . . , J 2 } ,
n=0
(7.67)
where {, } is the Poisson bracket. The computation has to be done with the
following constraints on phase space:
xk xl
{xk , pl } = kl . (7.68)
r2
It is proved in [362] that the series on the r.h.s. of (7.67) converges. The outcome is
the following parametrization of the phase space:
   
J r2 J
a = cosh x + i sinh p. (7.69)
m r2 J m r2

Thus we obtain the complex sphere realization of T (Sd ) :

T (Sd ) Sdc = {a = (a1 , a2 , . . . , ad+1 )T Cd+1 : a2 = a2k = r2 } , (7.70)

k

The quantization of this model is worked out in accordance with the underlying
Euclidean group SO(d + 1)  Rd+1 symmetry exemplified, on the classical level,
by the Poisson commutation rules issued from the unconstrained canonical rules
{xk , pl } = kl in the phase space Rd+1 Rd+1 :

{Jkl , Jmn } = kn Jlm + lm Jkn km Jln ln Jkm , (7.71)

{xk , Jlm } = kl xm km xl , {xk , xl } = 0 . (7.72)

Thus, introducing the quantum counterparts Jkl , xk of the classical Jkl , xk , these
operators are requested to be self-adjoint generators of a spinless unitary irreducible
representation of SO(d + 1) Rd+1 , with commutation rules consistent with (7.71)
(7.72):
1  
[Jkl , Jmn ] = kn Jlm + lm Jkn km Jln ln Jkm , (7.73)
ih
1
xk , Jlm ] = kl xm km xl , [
[ xk , xl ] = 0 . (7.74)
ih
The quantum complexifier is then defined as

kinetic energy operator J2

= . (7.75)
2m r2
7.4 CS on Spheres Through Heat Kernels 189

The annihilation operators are the counterparts of the phase space complex
coordinates (7.67). They are obtained from position operators through the expo-
nentiated adjoint action of the enveloping Lie algebra:
$  %
1 J2
ak = exp , xk
h 2m r2
+
1 1
:= (2m r2 )n n! [ [[xk , J2 ], J2 ], . . . , J2 ] (7.76)
n=0

ak , al ] = 0 and a21 + a22 + + a2d+1 = r2 I. They correspond to the

They satisfy [
spinless UIR of SO(d + 1)  Rd+1 for which x2 Jkl = Jkm xm xl Jlm xm xk .
For convenience, let us introduce the dimensionless parameter

h
= . (7.77)
m r2
As a ratio of the quantum action to the classical action characteristic of the system, it
encodes precisely the semi-classical character of its quantum model: semi-classical
aspects become manifest for  1. We have two possible expressions of these
operators. The first one involves the heat operator on the sphere which transports
the Cartesian position operators for Sd like an imaginary time evolution ruled by the
LaplaceBeltrami operator J2 :
$ % $ %
J2 J2
a = exp 2
 
x exp . (7.78)
2h 2h2

The second one corresponds to a polar coordinate factorization of the operator 

a up
to a corrective factor of quantum nature:

a = ed /2 ei J/h 
 x. (7.79)

We introduce now the (nonsymmetric) momentum operator:

1
=
p x,
J (7.80)
r2
and we check that p = p2
J x + ih(d 1)
p. Combining these two formulas, we may
express the operator 
a in terms of x and 
p,
 
h r2
a = e /2
 cosh( Js /h) + (d 1) sinh( Js /h) x + i sinh( Js /h)
p ,
2Js Js
1 (7.81)
 
where Js = J + h (d 1) /4. This equation describes the quantized complex
2 2 2

sphere and has to be compared with (7.69).

190 7 Covariant Coherent States

Given x Sd , a (Dirac) simultaneous eigenstate of the position operators for a

localization in x is denoted | x ,

xk | x  = xk | x .

In a first step, we define the coherent states | x  as the action of the heat operator
on | x , for x Sd ,
$ %
J2
| x  = exp 2 | x  . (7.82)
h

Performing the analytic continuation Sd % x  a Sdc , we get eventually

the coherent states | a . The relation (7.78) implies that they are simultaneous
eigenvectors of the annihilation operators: ak | a  = ak | a . They are normalizable,
a | a  < . Because of (7.78), they are covariant with respect to all complex
rotations R SO(d + 1, C) (by analytical continuation):

Uc (R)| a  = | Ra  , (7.83)

where Uc is the analytical continuation of the concerned UIR of the Euclidean group
SO(d + 1)  Rd+1 .
They resolve the identity

|a a | dh (a) = ICS , (7.84)
Sdc

where ICS denotes the identity operator in the closed subspace of L2 (Sdc , dh (a))
generated by the CS |a . For the sake of conciseness we choose physical units in
which r = 1 and m = 1, so that position x and momentum p are now dimensionless,
and h and become equal. The measure h on the complex sphere is given in terms
of real coordinates x and p by:
 
sinh 2p d1 d
dh (a) = (2 , 2p) 2 (x2 r2 ) (x p) dx dp, (7.85)
2p
&
where p = p2 and is the heat kernel for the d-dimensional hyperbolic space
[Dav90], that is, the fundamental solution of the heat equation. More precisely, it is
the solution to the differential equation,

(s, R) 1 2
= + (d 1) coth R , (7.86)
s 2 R2 R

subject to the initial condition


lim cd f (R) (s, R) (sinh R)d1 dR = f (0)
s0 0
7.4 CS on Spheres Through Heat Kernels 191

for all continuous functions f on [0, ) with at most exponential growth at infinity.
Here cd is the volume of the unit sphere in Rd . The heat kernel is given through the
recursion formula

eds/2 d (s, R)
d+2 (s, R) = .
2 sinh R R
The lowest-dimensional expressions read as:

1 (s, R) =(2 s)1/2 eR

2 /2s
,

e /2s
2
1
2 (s, R) =(2 s)1 es/8 & d ,
s R cosh cosh R
R R2 /2s
3 (s, R) =(2 s)3/2 es/2 e .
sinh R
The proof of the resolution of the identity is based on the fact that the measure
2 3d1 d
(2p)1 sinh 2p 2 dx dp is invariant under the action of the complexified
rotation group SO(d + 1, C) and on the properties of the coherent states with regard
to the action of the self-adjoint LaplaceBeltrami operators Ja2 and Ja2 on the
complex sphere.
 
2 2 1
Ja 2 := al ak , Ja |a  = 0 , |a  = Ja2 |a  .
k<l ak al 2

Expressed in terms of p, the LaplaceBeltrami operator reads

 
2
Ja2 (2p) = + (d 1) coth R (2p).
R2 R R=2p

Differentiating (7.84) with respect to under the integral sign, using the self-
adjointness and the above equation allows to conclude that the integral (7.84) does
not depend on . Hence, it is enough to prove that
 
lim |x,p x,p | dh (a) = |x x | dx = ICS .
0 Sdc Sd

The coherent states |a  are conveniently realized as elements of the Hilbert

space L2 (Sd , dy) of square integrable functions on the sphere equipped with its usual
SO(d + 1)-invariant measure dy,

(a, y) = y |a  , a Sdc , y Sd . (7.87)

192 7 Covariant Coherent States

(a, y) depends only on the complex angle = (a, y) defined through analytic
continuation by
   
ay J xy 1 J
cos = 2 = cosh + i sinh py. (7.88)
r h r2 J h
In the lowest dimensions, (a, y) is given by
+
1 (a, y) = (2 )1/2 e 2 (2 n) ,
1 2
(7.89)


e /8 d +
(1)n ( 2 n)e 2 ( 2 n) ,
1
2 (a, y) = (2 )1
2

cos cos n=
(7.90)
e /2 +
( 2 n)e 2 (2 n) ,
1
3 (a, y) = (2 )3/2
2
(7.91)
sin n=

For higher dimensions, there exists the inductive formula,

1 d d
d+2 (a, y) = ed /2 (a, y) . (7.92)
2 sin d
We observe the fundamental importance of the elliptic function 3 [Mag66],
namely,
+
ei (z+n)
2 /
3 (z, ) := (i )1/2 ,
n=

in all these expressions. The Poisson summation formula allows to give them an
alternative form to them, more appropriate to the harmonic analysis on the sphere.
+

+ n2 in
e 21 (2 n)2
= e 2 e .
2 n=
(7.93)
n=

In particular for d = 1 we recover the familiar form (up to a normalization factor)

of coherent states for the circle in Fourier representation, described in Sect. 6.3.
More generally, as elements of H := L2 (Sd , dy) and from their analytical
dependence on the complex angle which is the continuation to the complex
plane of the real angle between the CS parameter x and the vector y, their
Hilbertian expansion in terms of the orthonormal basis of H made of (hyper-)
spherical harmonics Y ( y), 
y := y/r, reads as

(a, y) = c (J) Y (a) Y (

y) . (7.94)

Here N stands for a set of d indices labeling the orthonormal basis, e.g.
encoding the successive reductions SO(d + 1) SO(d) . . . SO(2), and Y is
the holomorphic continuation of the Y ( x) from Sd to Sdc . The resolution of the
unity (7.84) implies that the set of functions
7.5 CS in Loop Quantum Gravity and Quantum Cosmology 193

&
O = { c (J) Y (a) , N } (7.95)

is an orthonormal basis in the CS subspace and an orthonormal system (but not a

basis!) in the Hilbert space L2 (Sdc , dh (a)).

7.5 CS in Loop Quantum Gravity and Quantum Cosmology

The most interesting part of the development of CS in LQG is the construction of

SU(2)-invariant CS in tensor products of representations. These states then obtain
a geometrical interpretation as polyhedrons. Actually this development proceeded
along two directions, namely, that of Thiemann [579582] and that of Freidel et al.
[291]. The latter started from a somewhat intuitive paper by Rovelli and Speziale
[542], in which the authors postulated the existence of a state over the tensor
product of four irreducible representations, supported on fixed values of the angles
between generators of the different representations (which are SU(2)-invariant).
That construction was then put on a sounder mathematical basis by Livine and
Speziale [450], then culminating in the geometrical construction of Freidel and
Speziale [291]. All that was developing in a direction orthogonal to the approach
of Thiemann, until Bianchi et al. [151] realized that the construction was in fact a
particular case of that of Thiemann!
In a recent paper, Freidel and Livine [289] went one step further. Noting that
the tensor product of N representations of SU(2) has a natural U(N) invariance,
they introduced a new set of coherent states covariant under U(N) transformations.
These CS are indexed by points of the Grassmannian of complex two-planes in
CN and possess a direct geometrical interpretation in terms of framed polyhedrons.
Further references along these lines may be found in [289].
These CS in LQG underlie all calculations, be it the asymptotic behavior towards
the classical limit, the derivation of n-point functions, the applications in cosmology
or, more recently, the entropy of black holes. As a matter of fact, applications of CS
in LQG has become a field itself, rather technical, with a large body of literature, and
it clearly would lead us far beyond the present volume. Thus we will mostly refer
the reader to the very recent review paper by Oriti et al. [495] (which belongs to the
special issue [38]). The starting point is that of Halls CS [361,362], suitably adapted
to the cotangent bundle of SU(2), T SU(2), discussed in Sect. 7.4. Then, instead
of the analytic continuation method of Hall, the authors consider noncommutative
translations on the Lie algebra of SU(2). The outcome is a new class of CS, better
adapted to LQG, in particular concerning their peakedness and the fluctuations
around the expectation values of certain (flux) observables, the latter taking exactly
the corresponding classical values. For details, we refer to the original paper [495].
A very recent application of CS, more precisely 1-D wavelets, is in the realm
of quantum cosmology [144]. The authors of that paper explore the possibility
of dealing with gravitational singularities on a quantum level by using CS (or
wavelet) quantization instead of the canonical quantization. The gravitational
194 7 Covariant Coherent States

collapse model studied is a FriedmanLematre universe with a RobertsonWalker

matrix coupled to a perfect fluid. The outcome is that the CS quantization method
offers an additional flexibility, that leads to the occurrence of a quantum repulsive
potential which regularizes the singularity. We shall discuss the various CS related
quantization methods at length in Chap. 11.

7.6 CS on Conformal Classical Domains

There is a conformal group SO(n, 2) construction of coherent states, of a different

nature, labelled by points of classical Khlerian domains and of complex spheres.
It was developed in particular by Onofri [493, 494] with the aim to obtain a
quantization of the so-called Kepler manifold through a limit of manifolds which
are quantizable la KostantSouriau. Onofri had in view the regularized Kepler
Coulomb problem which corresponds to the free motion on the sphere Sn1 , but
with the restriction that the phase space is the Moser or Kepler (Souriau) manifold
T0 (Sn1 ), that is, the cotangent bundle T (Sn1 ) with the zero-section deleted (i.e.,
the points of zero momentum removed). This restriction amounts to put r = 0
in (7.70), i.e., to deal with the complex sphere of radius 0. Indeed, the Moser
manifold can be identified with the null quadric in Cn :
 
T0 Sn1 {z = (z1 , z2 , . . . , zn )T Cn : z2 = zT z = 0}, (7.96)

Due to its projective nature, the latter is a singular orbit for the coadjoint action of
SO(n, 2), actually the most singular one with real dimension 2(n 1). It is related
to the limit case l = 0 of a continuous family of orbits Ol (in Onofris notations).
The symplectic manifold Ol is defined as the orbit of the point l = lS , l > 0,
under the action of the coadjoint representation of SO(n, 2). Here S stands for the
basis element dual to the element S so(n, 2) which generates the subalgebra so(2).
To be more explicit about notations, one denotes the basis elements of so(n, 2) by

X12 = S , X +2 +2 = M , X1 +2 = V , X2 +2 = W , (7.97)

where Greek indices run from 1 to n. They are tensorial with respect to the subgroup
SO(n) whose generators are the M . Accordingly, a generic point of the dual
space so(n, 2) is denoted by
 
= sS + m M

+ v V + w W .
<

In terms of these coordinates, the orbit Ol , l > 0, is given by

4  
Ol = = s, m , v , w :
7.6 CS on Conformal Classical Domains 195

  5
s2 + m2 v2 + w2 = l 2 ; sm = v w v w .
<

The manifold Ol , l > 0, is homeomorphic to the homogeneous bounded domain of

type IV (Cartan domain) [Hua63]. Such a domain, which generalizes the unit disk
in the complex plane (see Sect. 6.3) is Khlerian and homeomorphic to the coset
SO(n, 2)/SO(n) SO(2). It is described in terms of vectors in Cn by

D (n) = { Cn : < 1 , 1 2 + | T |2 > 0} , (7.98)

where is the hermitian conjugate of the column vector . The one-to-one

differentiable map l : Ol D (n) is given by:
 
i zT z
= z+ z , z := v iw , (7.99)
2s 2s(s + l) z z
and yields the relations
2l(l + s)
= 1 & (7.100)
2ls + 4l 2 s2 + |zT z|2
4l 2
1 2 + | T |2 = & (7.101)
2ls + 4l 2 s2 + |zT z|2

The inverse map l1 is given by

T 1 | T |2
z = 2il , s=l , m = 2il , (7.102)
( , ) ( , ) ( , )

where ( , ) := 1 2 + | T |2 . Homogeneity of the orbit Ol is lost at the

limit l = 0; it splits into three orbits, the trivial one and two manifolds of respective
real dimension 2n and 2(n 1):

60 := lim O = {z = 0} {zT z = 0} {zT z = 0} := {0} O O .

O l 0
l0

Because of (7.100)(7.101) taken at l = 0, the map (7.99) sends O to the so-called

Lie sphere G(n) :
/ 0
G(n) := = ei 
x:
x Rn , x = 1 , [0, ] = [0, ] Sn1 .
x (7.103)

It is the characteristic manifold (or BergmanShilov (or Sylow) boundary) of D (n) ,

i.e., the manifold on the boundary of D (n) having the following properties:
1. The modulus of every analytic function regular in D (n) assumes its maximum
on G(n) .
196 7 Covariant Coherent States

2. For every point a on G(n) there exists a function f ( ), regular on D (n) , such
that the modulus of f ( ) assumes its maximum at = a.
The third manifold O0 is precisely the cotangent bundle (7.96). It is the
orbit of the point 0 := S + V1 + W2 + M12 , and is homeomorphic to the

coset R SO(n)/SO(n 2). Its identification with the phase space of Kepler
motion
 is obtained
 by introducing a local chart of canonical coordinates (x, p)
T Rn1 \ {0} . With the CoulombKepler Hamiltonian given by
p2
H= , p2 = p p , r2 = x x ,
2m r

and with p0 := 2mH (Fock parameter), the following transformations are locally
canonical:
m
s= , (7.104)
p0
 
m x/r (x p) p
z = rp + i exp(ip0 x p/m ) , (7.105)
p0
 
rp2 m
zn = + ix p exp(ip0 x p/m ) , (7.106)
p0
m = x p x p , , = 1, 2, . . . , n 1 , (7.107)
p2 x (x p) p m x/r
a := (m1n , . . . , mn1n ) = . (7.108)
p0
Physically, a is the well-known RungeLenz vector [4345] and the components of
z = v iw Cn are the BacryGyrgyi parameters [360]. In the formulas (7.98)
(7.102) there appears the weight function ( , ) which derives from the more
general function
( , ) := 1 2 + T T . (7.109)
This object plays a key role in the geometric and analytic structure of the domain
D (n) . Indeed, suppose that we have a sequence of analytic functions { }, =
0, 1, . . . , in D (n) such that any analytic function f ( ) in D (n) can be developed in a
series

f ( ) = a ( ) ,
=0

which is convergent in D (n) . The basis { } can be chosen to be orthonormal, such

that:

( ) ( ) d = , (7.110)
G(n)

( ) ( ) d = . (7.111)
D (n)
7.6 CS on Conformal Classical Domains 197

Here, d = d d x is the SO(n) invariant measure on Sn1 , and

x, where d

d d
d = .
2i

This basis allows to define the Bergman kernel K( , ) through the expansion

1 1
K( , ) :=   [ ( , )]n = ( ) ( ) , (7.112)
Vol D (n)
=0

where the volume of the domain is given by:

! " n
Vol D (n) = .
2n1 n!

The numbers n are pseudoconformal invariant: they do not depend on the choice
(n) (n)
of the basis and are preserved under analytic mappings D (n)  D1 , G(n)  G1 .
The Bergman kernel is reproducing in the following sense. For any function f ( ),
analytic in D (n) , we have

f ( ) = K( , ) f ( ) d .
D (n)

It can be used as the Khlerian potential from which derives the Riemannian metric
on the space D (n) :
n
2
d d ln K( , ) = ln K( , ) di d j .
i, j i j

An orthonormal system on the Lie sphere G(n) simply derives from the spherical
harmonics Y ( x) introduced in Sect. 7.4 with d = n 1. We know that they are
restriction on Sn1 of homogeneous harmonic polynomials on Rn ,
x
Ys, (x) (xT x)s/2 Ys, (
x) , 
x= ,
xT x

where = (k1 , k2 , . . . , kn2 ), with s k1 k2 . . . |kn2 |, stands for the set of

remaining indices characterizing spherical harmonics [Vil69]. The integer s is the
degree of homogeneity of the corresponding harmonic polynomials Y . Now, any
homogeneous polynomial of degree f on Rn decomposes in a unique way in a sum
of polynomials of the form
7 8
f
(xT x)k0 Y f 2k0 , (x) , 0 k0 , f N.
2
198 7 Covariant Coherent States

By the extension x  ei x and restriction to the Lie sphere, we find the orthonormal
system

1
( ) := f 2k0 , ( ) = e2ik0 Y f 2k0 , (
x) . (7.113)

This set can be extended to an orthonormal basis of L2 (G(n) , d ) by replacing k0 

k0 in the phase factor. By analytic continuation of (7.113) to the domain D (n) , we
find the orthonormal basis
1
( ) := f 2k0 , ( ) = ( T )k0 Y f 2k0 , ( ) , (7.114)

of the Hilbert space (of the FockBargmannSegal type) H(n) of holomorphic

functions ( ), D (n) , which are square integrable with respect to the scalar
product

1
 1 | 2  =   1 ( ) 2 ( ) d . (7.115)
Vol D (n) D (n)

With this definition, the basis element 0 ( ) := 1 has norm 1. The constant has
been given by [Hua63, Eq. (7.5.2)]:
   
k0 ! (n) n2 + 1 f + n2 k0
= n   .
2 2 k0 + n2 + 1 ( f + n k0 )
Besides the Bergman kernel there exist two other kernels which can be expanded in
terms of these basis functions and which play a central role in the harmonic analysis
on the domain D (n) and its BergmanShilov boundary G(n) .

(i) The Cauchy kernel

It is given by

H( , ) = ( ) ( )

1 1
= 
(n)
2 3n/2 , = ei x , xT x = 1 ,
Vol G (x ei ) (x ei )
T

and is reproducing in the sense that for any analytic function f ( ) with a series
development f ( ) = a ( ) on G(n)

f ( ) = H( , ) f ( ) d .
G(n)
7.6 CS on Conformal Classical Domains 199

Note the value of the volume of the BergmanShilov boundary:

! " n
2 2 +1
(n)
Vol G = n .
2

(ii) The Poisson kernel

It is given by
! T
"n
H( , )H( , ) 1 1 2 + | T |2 /2
P( , ) = =   ,
H( , ) Vol G(n) |( )T ( )|n

and is reproducing in the sense that to any continuous function u( ) on G(n) the
function defined by

u( ) = P( , ) u( ) d .
G(n)

is harmonic on D (n) .
The conformal group action on the homogeneous space D (n) derives from
the Cartan factorization (see Sect. 4.5.2) SO(n, 2) = PK, K SO(n) SO(2).
An element g SO(n, 2), as a block ([2 + n)] [2 + n] matrix, obeys
 
I22 02n
gJ gT = J := .
0n2 Inn

An element k in the maximal compact subgroup K is characterized by k1 = kT ,

whereas an element in P is symmetric. Since the map D (n) %  p( ) P is one-
to-one, the action g :  g of SO(n, 2) on D (n) derives from the left action on
P followed by a Cartan factorization: gp( ) = p(g )kg . The explicit form of this
 
ab
symplectic action of g = SO(n, 2) on D (n) is given by:
cd

(1, i)cT + 2 d T + (1, i)cT T

g :  g =  . (7.116)
1
[(1, i)aT + 2 bT + (1, i)aT T ]
i

Note that the map l1 defined in (7.102) is the so-called moment of the above
action.
Let us now build la GilmorePerelomov a family of conformally covariant
coherent states labelled by the points of the domain D (n) or equivalently by
200 7 Covariant Coherent States

(n)
the points of the orbit Ol . For l > n/2, they live in the Hilbert space Hl of
holomorphic functions ( ), D (n) , which are square integrable with respect
to the scalar product

 1 | 2  l = 1 ( ) 2 ( ) d l ( ) , (7.117)
D (n)

where the measure, given by

 n
2 (l 1) (l + n/2 1)
d l ( ) = ( ( , ))ln/21 d ,
(l n/2)
! "
reduces to d /Vol D (n) for l = n/2 + 1 and is such that the unity function
0 ( ) = 1 has still norm 1. The corresponding UIR g  Ul (g) of SO(n, 2) belongs
(n)
to the holomorphic discrete series and is defined on Hl by

Ul (g) : ( )  (Ul (g) ) ( ) = l (g, ) (g1 ) . (7.118)

The multiplier l is given by

   ln/2+1
1 1
l (g, ) = ((1, i)a 2 c + (1, i)a )
T
,
2 i
 
and is related to the jacobian jg g1 = det( g / ) =g1 of the map
 g by

2  3 l+n/21
l (g, ) = jg g1 n
.

We notice that the minimal weight (e.g. ground  vector 0 transform as

 state)
a 0
Ul (k)0 = ei(l+n/21) 0 under the action of k = SO(n)SO(2), with a =
0d
 
cos sin
. Hence, following the group theoretical construction la Gilmore
sin cos
Perelomov, normalized coherent states are defined in this FockBargmannSegal
representation as:
! "
D (n) %  | ; l := Ul p( ) |0  . (7.119)

By construction these coherent states are covariant with respect to Ul up to a phase:

Ul (g)| ; l = Ul (gp( ))|0 

= Ul (p(g )kg )|0  = ei(l+n/21)g |g ; l . (7.120)

7.6 CS on Conformal Classical Domains 201

The resolution of the identity follows from Schurs lemma and square integrability
of Ul :
 
| ; l ; l| ( ( , ))ln/2+1 dl ( ) := | ; l ; l| dl ( ) = I , (7.121)
D (n) D (n)

where the measure dl includes the Bergman kernel as follows

 n
2 (l 1) (l + n/2 1)
dl ( ) = ( ( , ))n d
(l n/2)
(2l 2) (l + n/2 1)
= K( , ) d . (7.122)
n! (l n/2)
The overlap between two coherent states is given by:
ln/2+1
( , )
 ; l| ; l = 1 , (7.123)
( , ) ( , )
(n)
and we can define from it an element of the Hilbert space Hl

;l ( ) := ( ( , ) ( , ))(l+n/21)/2  ; l| ; l ,

= (1 2 T + T T )ln/2+1 . (7.124)
(n)
In virtue of (7.121), this is a reproducing kernel for Hl whose expansion in terms
of an orthonormal basis is

;l ( ) := Rl ( , ) = ;l ( ) ;l ( ) . (7.125)

Equivalently,

| ; l = ( ( , ))(l+n/21)/2 ;l ( )| ;l  . (7.126)

Note that the Bergman kernel corresponds to the value l = n/2 + 1, i.e. the lowest
UIR in the stricto sensu holomorphic discrete series.

7.6.1 Beyond Square Integrability and Singular Orbit

It is possible to consider the continuation of the representations Ul of SO(n, 2) for

l n/2 and the limit l 0 corresponding to the quantization of the Kepler manifold
(n)
O0 . According to Onofri [494], the Hilbert space Hl may be equivalently viewed
as the space of holomorphic functions on D (n) linearly spanned by the vectors in the
orbit of the minimal weight vector 0 ,

{Ul (g)0 = l (g, ) : g SO(n, 2)} , (7.127)

202 7 Covariant Coherent States

and completed according to the topology of uniform convergence. The norm is

implicitly defined by requiring that the kernel Rl ( , ) be reproducing, i.e.,

(n)
Rl ( , ) = ;l ( ) Hl ,  ;l | l = ( ) .

(n)
This definition of Hl makes sense also for the range l n/2, where the square-
integrability of Ul does not hold anymore. At the limit l = 0, we have the expansion

R0 ( , ) = (1 2 T + T T )n/2+1 = ;0 ( ) ;0 ( ) , (7.128)

(n)
where the holomorphic functions ;0 , which form an orthonormal basis of H0 , are
uniquely characterized by their values on the BergmanShilov boundary G(n) . They
are given in terms of the harmonic polynomials in (7.114) with k0 = 0, f = s, as
 1/2
2 n/2
;0 ( ) := s, ( ) = Ys, ( ) , (7.129)
(n/2 1)(s + n/2 1)

Thanks to the constraint k0 = 0, the orthonormality of the basis may be understood

(n)
in the realization of H0 as the Hilbert space of square integrable functions on the
unit sphere S n1 with respect to the inner product

1 [ + (n/2 1)2 ]1/2
 1 | 2  0 = 1 (
x) 2 (
x) d
x (7.130)
Vol (Sn1 ) Sn1 n/2 1
  n/2
where Vol Sn1 = 2(n/2) and is the LaplaceBeltrami operator on the sphere.
A function (x)) = s, as, Ys, ( x) on the unit sphere can be extended to a
holomorphic function on D (n) as ( ) = s, as, Ys, ( ), and the reproducing
property of R0 ( , ) reads as

1 [ + (n/2 1)2 ]1/2
(
x)) = (
y) d
y. (7.131)
Vol (Sn1 ) Sn1 x
(n/2 1) yn2
This is an SO(n)-invariant equation, which reduces in the case of spherical
harmonics to the integral eigenequation:

1 s + n/2 1 Ys, (
y)
Ys, (
x)) = y.
d (7.132)
Vol (S ) n/2 1
n1
Sn1 x
 yn2

For n = 4, we recover the Fock formulation on S3 of the stationary Schrdinger

equation for the CoulombKepler problem [282]. The Balmer quantization of the
2
energy follows from (7.104): m/p0 = s + 1 = 1, 2, . . . , i.e., E = 2m(s+1) 2.
Chapter 8
Coherent States from Square Integrable
Representations

Abstract This chapter is devoted to a detailed development of the theory of square

integrable group representations, including the resulting orthogonality relations.
Then we study a particular class of semidirect product groups, namely, groups of
the form G = Rn  H, where H is an n-dimensional closed subgroup of GL(n, R).
Several concrete examples are presented. Finally we generalize the theory to
representations that are only square integrable on a homogeneous space. This allows
one to treat CS of the Gilmore-Perelomov type and, in particular, CS of the Galilei
group.

This chapter is devoted to a fairly detailed development of the theory of square

integrable group representations, particularly in Sect. 4.2.2 and Sect. 7.6. We
have already seen examples of such representations. Indeed, the discrete series
representations, U j of SU(1,1) discussed in Sect. 4.2.2, are square integrable. To
make this clear, observe that the discussion in that section, leading up to (4.73), tells
us that the invariant Haar measure on SU(1,1) can be written as

1
d (g) = d0 (z, z) d , (8.1)
4

where d0 (z, z) is the invariant measure (4.73) on the open unit disc. This follows
from the fact that an arbitrary group element, g SU(1,1), can be written as g =
(z)k [see (4.65) and (4.66)]. Moreover, the measure d , so defined, is both left and
right invariant, the group being unimodular. On the other hand, from the definition
of the CS (z) in (4.99) and the resolution of the identity in (4.100), we find that the
vectors g = U j (g) , with = (2 j 1)1/2 u0 [see (4.93)], satisfy the resolution of
the identity (on the Hilbert space Hhol of analytic functions on the open unit disc),

|g g | d (g) = Ihol . (8.2)
SU(1,1)

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 203
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__8,
Springer Science+Business Media New York 2014
204 8 Coherent States from Square Integrable Representations

A resolution of the identity of the above type, with respect to the entire group, forms
the basis for the square integrability of the representationa notion that we now
make precise.

8.1 Square Integrable Group Representations

Group representations U that are square integrable mod({e}, ), where (g) = g,

for all g G and n = 1, in Definition 7.1.1 of covariant CS, are simply called
square integrable. In this section and the next, we give a detailed account of
the properties of such representations, following mainly [193, 351, 352, Gaa73],
and [505]. An algebraic treatment of the subject may be found in [266]; the case in
which the group G is unimodular is treated in detail in [Dix64]. Square integrable
representations of a group G are also said to belong to the discrete series of its
representations. We emphasize again that square integrability in this setting implies
the existence of a vector H

|g g | d (g) = I, g = U(g) , (8.3)
G

( = left Haar measure). For a compact group, the above relation would hold for
any (suitably normalized) vector in the representation space H. Hence, every
representation of a compact group is square integrable. For noncompact groups,
however, there do exist representations that are not square integrable, and we have
already encountered this situation. For example, the representation U of the Weyl
Heisenberg group, given in (2.33), is not square integrable, for the Hilbert space
does not contain any nontrivial vector that would satisfy (8.3). The reason for this
is that the invariant Haar measure is now d dq dp, and the integration with respect
to would clearly diverge.
CS of semisimple Lie groups, compact and noncompact, have been discussed in
Sects. 7.2.1 and 7.2.2, respectively, the simplest cases being SU(2) and SU(1,1).
For the remainder of this section, and in the next section, U will be taken to be a
unitary, irreducible representation of G on the Hilbert space H.
Definition 8.1.1 (Admissible vector). A vector H is said to be admissible if

I( ) = |U(g) | |2 d (g) < . (8.4)
G

Note that since dr (g) = d (g1 ), [see (4.5)] and since U(g) is unitary,
 
I( ) = |U(g1 ) | |2 dr (g) = | |U(g) |2 dr (g).
G G

Hence,

I( ) = |U(g) | |2 dr (g), (8.5)
G
8.1 Square Integrable Group Representations 205

so that it is immaterial whether the left or the right invariant Haar measure is used in
the definition of admissibility. Note also that, if = 0, then I( ) = 0. Indeed, since
g  U(g) |  is a continuous function, and the measure d is invariant under left
translations, I( ) = 0 implies U(g) |  = 0, for all g G. Since U(g) , g G,
is a dense set of vectors in H, this implies that = 0.
Lemma 8.1.2. If H is an admissible vector, then so also is g = U(g) , for all
g G.
Proof. Indeed,
 
I(g ) = |U(g )g |g |2 d (g ) = |U(g1 g g) | |2 d (g ),
G G

= |U(g g) | |2 d (g ), by the left invariance of d ,
G

= |U(g ) | |2 (g1 ) d (g ), by (4.5),
G

1
= |U(g ) | |2 d (g ), by (4.4).
(g) G

Thus,
1
I(g ) = I( ) < . (8.6)
(g)


Let A denote the set of all admissible vectors. Then, as a consequence of this
lemma, A is stable under U(g), g G. Since U is irreducible, either A = {0}, i.e.,
it consists of the zero vector only, or A is total in H. Furthermore, it turns out (see
the proof of Theorem 8.1.3) that

A iff |U(g) | |2 d (g) < , H, (8.7)
G

and this in turn implies that 1 + 2 is admissible if 1 , 2 are, i.e., A is a vector

subspace of H. Therefore, either A = {0} or A is dense in H. For A , = 0,
we shall write
I( )
c( ) = . (8.8)
 2
The condition for the admissibility mod(H, ), of a set of vectors i , i =
1, 2, . . . , n, was given in Definition 7.1.1 (see also Theorem 7.3.1). The next theorem
shows that in the present context that definition is equivalent to the one given
above; the theorem also captures practically all of the important properties of
square integrable representations. Recall the definitions of the left and right regular
representations, U and Ur , of G, given in Chap. 3, Sect. 4.2.3.
206 8 Coherent States from Square Integrable Representations

Theorem 8.1.3. The UIR g  U(g) of the locally compact group G is square
integrable if and only if A = {0}. In this case, for any A , the mapping

W : H L2 (G, d ), (W )(g) = [c( )]1/2 g | , H, g G, (8.9)

is a linear isometry onto a (closed) subspace H of L2 (G, d ). On H, the resolution

of the identity

1
|g g | d (g) = I (8.10)
c( ) G

holds. The subspace H = W H L2 (G, d ) is a reproducing kernel Hilbert space,

so that the corresponding projection operator

P = W W , P L2 (G, d ) = H , (8.11)

has the reproducing kernel K ,


 )(g) =
(P  (g ) d (g ),
K (g, g )  L2 (G, d ),

G
1
K (g, g ) = g |g , (8.12)
c( )

as its integral kernel. Furthermore, W intertwines U and the left regular represen-
tation U ,

W U(g) = U (g)W , g G. (8.13)

Of course, since here n = 1, we have identified the operator K (g, g ) and its matrix
kernel K (g, g ).
Before proving this theorem, we observe that an entirely analogous result holds
with the right regular representation Ur . Thus, for each A , there exists a linear
isometry,

Wr : H L2 (G, dr ), (Wr )(g) = [c( )]1/2 g1 | , H, g G. (8.14)

The corresponding reproducing kernel is

1
Kr (g, g ) =  1 |g1  = K (g1 , g1 ). (8.15)
c( ) g

Proof. The domain D(W ) of W is the set of all vectors H such that

1
|g | H |2 d (g) < .
c( ) G
8.1 Square Integrable Group Representations 207

But, for any D(W ) and g G, we have

 
1 1
|g |U(g ) H |2 d (g) = |g1 g | H |2 d (g)
c( ) G c( ) G

1
= |g | H |2 d (g),
c( ) G

by the invariance of . Thus, D(W ) is stable under U, and, hence, dense in H,

since U is irreducible. Moreover, on D(W ), the intertwining property (8.13) follows
easily from (8.9) and the definition of the left regular representation in (4.105).
We prove next that, as a linear map, W is closed. Let {n } n=1 D(W ) be a
sequence converging to H, and let the corresponding sequence {W n } n=1
L2 (G, d ) converge to L2 (G, d ). Then, by the continuity of the scalar product
in H,

lim W n (g) = lim [c( )]1/2 g |n  = [c( )]1/2 g | . (8.16)

n n

Thus, since W n in L2 (G, d ) and W n (g) [c( )]1/2 g |  pointwise,

[c( )]1/2 g |  = (g),

almost everywhere (with respect to ), whence,


|g | |2 d (g) < ,
G

implying that D(W ) and W = ; i.e., W is closed.

Using Lemma 4.3.3 (the extended Schurs lemma), we establish the boundedness
of W : D(W ) L2 (G, d ). Hence, D(W ) = H, and, furthermore, W is a multiple
of an isometry,

W 2L2 (G, d ) =  2H , H, (8.17)

where R+ . To fix , take = . Then,

W 2L2 (G, d ) I( )
= = = 1, (8.18)
 2 c( ) 2

by (8.4), (8.8), and (8.9).

Thus, W is an isometry; i.e., WW = I, which implies that (8.10) holds.
Therefore, the range of W is a closed subspace of L2 (G, d ), and the projection
on it is P = W W . Then, (8.12) and (8.13) follow from a direct computation. 

An immediate consequence of this theorem is the following important result.
208 8 Coherent States from Square Integrable Representations

Corollary 8.1.4. Every square integrable representation of a locally compact

group G is unitarily equivalent to a subrepresentation of its left regular representa-
tion (and, hence, of its right regular representation as well).
The proof of this corollary consists simply in showing that the projection P on the
range of W commutes with the left regular representation. Indeed:

P U (g)=W WU (g) = W (U (g1 )W ) = W (W U(g1 ))

=W U(g)W = U (g)W W = U (g)P .

Since W is an isometry, its inverse is equal to its adjoint on its range; i.e., W1 =
W on H . Then, applying both sides of (8.10) to an arbitrary vector H, we
obtain the reconstruction formula [see (6.5)]

1
= W = (g)g d (g). (8.19)
[c( )]1/2 G

In (8.52), we shall obtain a generalized version of this reconstruction formula using

two different admissible vectors.
The outcome of Theorem 8.1.3 is a total set of CS indexed by the points of
the group G itself. While this does happen in simple examples, it is rather rare.
A case in point are the 1-D wavelets, which we shall discuss in great detail in
Chap. 12. Another, rather exotic, example has been given in [519], for the group
SU(1, 1). In general, however, the CS systems of physical interest are supported
by a quotient manifold X = G/H, where H may be the stability subgroup of the
generating vector (the GilmorePerelomov situation), but need not be so. Yet,
CS on a group have a number of interesting mathematical properties that warrant a
thorough analysis, which is the purpose of the present chapter.
We note that a vector W H = H = P L2 (G, d ), if and only if there
exists a vector H such that (g) = [c( )]1/2 g |  for almost all g G (with
respect to the measure ). This also means, in view of the strong continuity of the
representation g  U(g), that (g) can be identified with the bounded continuous
function of G,

g  [c( )]1/2 g |  = U(g) | ,

(8.20)
sup|g | | sup U(g)    =     .
gG gG

Hence, the reproducing kernel subspace H can be identified with a space of

bounded, continuous functions on the group G.
In addition, the reproducing kernel K (g, g ) is in the present case a convolution
kernel on G : K (g, g ) =  |U(g1 g ) . This implies that K has a regularizing
effect. For instance, if G is a Lie group, and is appropriately chosen, the
elements of H can be made to be infinitely differentiable functions, which extend to
8.1 Square Integrable Group Representations 209

holomorphic functions on the complexified group Gc [493]. This gives rise to some
of the attractive holomorphic properties of CS, and their geometrical implications.
Another consequence of the convolution character of K is that the kernel, and
hence the elements of H , have interpolation properties that prove useful in practical
computations [351].
Finally, it should be emphasized that the reproducing kernel K is the main tool
for computing the efficiency or resolving power of the transform W , in wavelet
analysis (see Sect. 14.3.2). Notice that each admissible vector determines its own
reproducing kernel K and reproducing kernel subspace H .

8.1.1 Example: The Connected Affine Group

We have encountered the connected affine group of the line, G+ , in Sect. 4.1. As one
of the simplest examples of an application of Theorem 8.1.3, let us explicitly display
the square integrability properties of a UIR of this group. Recall that G+ , which
is also called the ax + b group, consists of all transformations of R of the form
x  ax + b, x R, with a > 0, b R and a group element is given by a pair
(b, a) R R+ . The invariant measures on G+ are easily computed to be

db da db da
d (b, a) = , and dr (b, a) = , (8.21)
a2 a
so that the modular function is
1
(b, a) = . (8.22)
a
There are only two nontrivial, unitary irreducible representations of G+ . In order
to examine these, consider the Hilbert space L2 (R, dx) and on it the representation
U(b, a),
!xb"
(U(b, a) f )(x) = a1/2 f , f L2 (R, dx), g = (b, a) G+ . (8.23)
a
This representation is unitary, but not irreducible. In order to isolate its irreducible
components, let us go over to the Fourier-transformed Hilbert space, which we
 d ). On this space, the above representation transforms to U(b,
denote by L2 (R,  a),
with

 a)
(U(b, h)( ) = a1/2 
h(a )eib ,   d ).
h L2 (R, (8.24)

It is now clear that the two subspaces,

 = L 2 (R
H  , d ), (8.25)
210 8 Coherent States from Square Integrable Representations

are stable under the action of the representation U(b, a), and in fact constitute
irreducible subspaces under this action. We denote the restrictions of U  to these
two subspaces by U  + and U  , respectively. The two subrepresentations are then
inequivalent, but both are square integrable, and we only look at U  + . Let
H +
be an admissible vector, in the sense of Definition 8.1.1. An easy computation then
yields
 
db da  ( )|2
|
I( ) =  + (b, a)
|U  |
 |2  2
= 2  d < . (8.26)
+
RR a2 0

 ) in (8.8) now assumes the form,

The constant c(

 ( )|2
|
 ) = 2
c( d < , (8.27)
0

 + is admissible if and only if it satisfies the

H
which means that the vector
condition

2
 ( )|2 d < .
| (8.28)
0

 has to lie in the domain of the positive,

In other words, for admissibility,
unbounded operator C:
 1/2
2
 )( ) =
(C  ( ),
0. (8.29)

As we shall see in Theorem 8.2.1, this is in fact the generic situation, in the sense
that admissible vectors are always characterized by being in the domain of a certain
positive operator, determined uniquely by the group representation itself. Coherent
states for this representation are then the families of vectors,

 + (b, a)
ba = U , (b, a) G+ , (8.30)

 being any admissible vector. The CS satisfy the resolution of the identity (on
with
 +
H ):

1 db da
ba 
| ba ba| = I. (8.31)
)
c( RR+
a2

The corresponding reproducing kernel is


1 
K(b, a; b , a ) = ei (bb ) [aa ]1/2  (a ) d .
 (a ) (8.32)
)
c( 0
8.1 Square Integrable Group Representations 211

This example is of great interest for applications, since it is the group-theoretical

backbone of the theory of wavelets, which we shall discuss at length in Sect. 12.2
[the vectors (8.30) are exactly the 1-D wavelets]. It is also fundamental for a new
approach to quantum cosmology singularities [144]. Before leaving it, let us note
that, in the case of the SU(1,1) CS, mentioned at the beginning of this chapter, the
operator C is (2 j + 1)1/2 Ihol , or just a multiple of the identity.
In our discussion of square integrable representations so far, the representation U
has been assumed to be irreducible. This requirement may be weakened in several
ways. A first possibility is to take a direct sum of square integrable representations
(actually every square integrable representation is of this type [Gaa73]). In this case
one may prove the following result [281].
Theorem 8.1.5. Let G be a locally compact group, with left Haar measure . Let
U be a strongly continuous unitary representation of G into a Hilbert space H, and
assume that U is a direct sum of disjoint square integrable representations Ui :
 
U= Ui , in H= Hi . (8.33)
i i

Let be an admissible vector, in the sense of Definition 8.1.1. Then,


|U(g) | |2 d (g) = ci Pi 2 , H, (8.34)
G i

where Pi is the projection on Hi and


ci = Pi 2 |Ui (g)Pi |Pi |2 d (g). (8.35)
G

If, in addition, all constants ci are equal, then the map W :  U(g) |  is an
isometry (up to a constant) from H into L2 (G, d ).
Thus, when the conditions of this theorem are satisfied, CS may be built in the
usual way. By similar arguments, the same is true if some of the components Ui are
mutually unitarily equivalent.
Another generalization is to take for U a cyclic representation, with a
cyclic vector [194]. In this case, assuming the admissibility condition (8.4), all of
Theorem 8.1.3 can be recovered. The relationship between cyclicity and square
integrability has been studied systematically, for the case of semidirect product
groups, in [293].
A more radical situation has been considered recently, in the case of the
Euclidean [393] and Poincar [416] groups. The idea is to take for U a direct integral
of irreducible, Wigner-type representations U (m) over the mass parameter m, on an
interval [m1 , m2 ], and for such a reducible representation, CS may be built. Taking
all positive values of m, however, results in an irreducible representation of the
corresponding affine or similitude group, to which we shall return in Sect. 16.2. The
same idea was exploited to a certain extent in [146].
212 8 Coherent States from Square Integrable Representations

Finally, although the framework used here is that of square integrable represen-
tations, a parallel theory may be derived for integrable representations, for which
many of the above results extend to Banach spaces [266, 278, 279].

8.2 Orthogonality Relations

If G is a compact group and U a unitary irreducible representation of G, then,

according to the PeterWeyl theorem [Gaa73], the matrix elements U(g) |  of
U satisfy certain orthogonality relations, and one may construct an orthonormal
basis of L2 (G, d ) consisting of such matrix elements. When G is only locally
compact, square integrable representations have the same property. Thus, among
all UIRs, the square integrable representations are the direct generalizations of the
irreducible representations of compact groups. These orthogonality relations, well
known when G is unimodular [336, Dix64], extend to non-unimodular groups as
well [193, 266, 351].
Theorem 8.2.1 (Orthogonality relations). Let G be a locally compact group and
U be a square integrable representation of G on the Hilbert space H. Then there
exists a unique positive, self-adjoint, invertible operator C in H, the domain D(C)
of which is dense in H and is equal to A , the set of all admissible vectors; if and
 are any two admissible vectors and ,  are arbitrary vectors in H, then

g |  g |  d (g) = C |C     | . (8.36)
G

Furthermore, C = I, > 0, if and only if G is unimodular.

Proof. Let ,  A , and consider the corresponding isometries W ,W  , defined
as in (8.9). With W : L2 (G, d ) H denoting, as before, the adjoint of the linear
map W : H L2 (G, d ), the operator W W is bounded on H. Next, for all g G,

W W U(g) = W U (g)W , by (8.13),

= [U (g1 )W  ]W = [W  U(g1 )]W

= U(g)W W .

By Schurs lemma, W W is therefore a multiple of the identity on H:

W W = ( ,  )I, ( ,  ) C (8.37)

[ ( ,  ) is antilinear in and linear in  ]. Applying Theorem 8.1.3, we find, for

= ,

( , ) = 1, A. (8.38)
8.2 Orthogonality Relations 213

Set

q( ,  ) = [c( )c(  )]1/2 ( ,  ), (8.39)

with c( ) as in (8.8). Using (8.9), we obtain

 
g |  g |  d (g) = [c( )c(  )]1/2 (W   )(g)(W )(g) d (g)
G G

= [c( )c(  )]1/2 W   |W L2 (G, d )

= [c( )c(  )]1/2   |W W  H ,

for all ,  A , and ,  H. Inserting (8.37) and (8.39) into this, we get

g |  g |  d (g) = q( ,  )  | H . (8.40)
G

From (8.37) and (8.40), we also find that


1
W W = |g g | d (g). (8.41)
[c( )c(  )]1/2 G

Now, (8.40) and (8.37) together imply that q : A A C is a positive, symmetric,

sesquilinear form on the dense domain A . Moreover, since q is independent of ,  ,
taking =  = 0, we obtain

1
q( ,  ) = U(g)  | U(g) |  d (g). (8.42)
 2 G

We next prove that the sesquilinear form q is closed on its form domain A . Indeed,
consider on A the scalar product and associated norm:

 |  q =  |  H + q( ,  ),  2q =  2H + q( , ), ,  A . (8.43)

Let {k }
k=1 A be a Cauchy sequence in the  . . . q -norm. Clearly, {k }k=1 is
also a Cauchy sequence in the norm of H, implying that there exists a vector H
such that limk k H = 0. Also, since the sequence is Cauchy in the  . . . q -
norm, q( j k , j k ) 0 for j, k . From (8.42), we infer that the sequence
of functions,

k }
{ k=1 L (G, d ),
2 k (g) = U(g)k | H ,

 L2 (G, d )
is a Cauchy sequence in L2 (G, d ). Thus there exists a vector
satisfying

k
lim   2
L (G, d ) = 0, (8.44)
k
214 8 Coherent States from Square Integrable Representations

and, therefore, the sequence { k } also converges to

 weakly, with the sequence
k=1

of norms {k L2 (G, d ) }k=1 remaining bounded. Moreover, for any g G,

lim U(g)k | H = U(g) | H k (g)

lim |  (g)| = 0.
k k

Thus, by (8.44),  (g) = U(g) | H , for all g G and all H, so that g 

U(g) | H defines a vector in L2 (G, d ). Taking = , we see that this implies
A . Next,

lim k 2q = lim k 2H + lim q(k , k )

k k k
1 k
 2 2
= 0 + lim  L (G, d ) , by (8.42),
k  2
= 0, by (8.44).

Consequently, A is complete in the  . . . q -norm, so that q is closed (see, for

example, [Ree80]). Since q is a closed, symmetric, positive form, the well-known
second representation theorem [Kat76] implies that the existence of a unique
positive self-adjoint operator C, with domain A , such that

q( ,  ) = C |C  H . (8.45)

Next, if = 0, then, by (8.37), (8.39), and (8.45),

I( )
C 2 = c( ) = = 0 (8.46)
 2

[see (8.8)]. So, C is injective, and, consequently, it is invertible. Moreover, its inverse
C1 is densely defined, as the inverse of an invertible self-adjoint operator [indeed,
it is easily seen that Ran(C) (the range of C) is dense in H].
It remains to prove the last statement. By (8.42), we have, for all g G,

1
q(U(g) ,U(g)  ) = U(g g)  | U(g g) |  d (g ),
 2 G

(g1 )
= U(g )  | U(g ) |  d (g ),
 2 G

= (g1 ) q( ,  ).

Comparing with (8.45), we find that, for all ,  A ,

1
CU(g) |CU(g)  H = C |C  H . (8.47)
(g)
8.2 Orthogonality Relations 215

Now, C2 is positive and densely defined in H. In addition, its domain is invariant

under U. Indeed, let  D(C2 ), which implies that  D(C), C  D(C), and
g D(C). Then, (8.47) may be written as
1
Cg |Cg H =  |C2  H ,
(g)

which shows that Cg D(C) as well, i.e., g D(C2 ). Thus, (8.47) implies that,
on the dense invariant domain D(C2 ):
1
C2U(g) = U(g)C2 . (8.48)
(g)
This then shows, using the extended Schurs Lemma 4.3.3, with U1 = U2 , that
(g) = 1, for all g G; that is, G is unimodular if and only if C = I, > 0. 

The operator C is known in the mathematical literature as the Duflo-Moore
operator [146, 266, 505], often denoted C = K 1/2 . Actually, it can be shown (see,
for example, [Sug90]), that, if G is compact, then

C = [dim H]1/2 I. (8.49)

(Note that, with G compact and U irreducible, dim H is finite.) If G is not compact,
but just unimodular, then [see (8.46)] with   = 1,

C = [c( )]1/2 I, (8.50)

so that the value of c( ) does not depend of A . In that case, we call

dU c( )1 the formal dimension of the representation U. In this terminology,
when G is a non-unimodular group, the formal dimension of a square integrable
representation U is the positive self-adjoint (possibly unbounded) operator C2 .
To conclude this section, we derive a generalized version of the resolution of the
identity (8.10).
Corollary 8.2.2. Let U be a square integrable representation of the locally compact
group G. If and  are any two nonzero admissible vectors, then, provided
C |C   = 0,

1
|g g | d (g) = I. (8.51)
C |C   G

Proof. This is mere restatement of the orthogonality relation (8.36), since the
vectors and  are arbitrary. 

From here, we get the reconstruction formula

[c(  )]1/2
= (g)g d (g), H, (g) = (W )(g), (8.52)
C |C   G
216 8 Coherent States from Square Integrable Representations

provided C |C   = 0, which generalizes (8.19). Here, is the analyzing vector

and  the vector used for reconstruction. This reconstruction formula can be further
generalized to situations in which only one of the vectors ,  is admissible, pro-
vided the other one satisfies a more stringent condition and the mutual compatibility
condition

1
0 < |c  | < , where c  =   |g g |  d (g) (8.53)
C |C   G

is satisfied. If # denotes the admissible vector, then one requires that W # #

L1 (G, d ). While the extended formula (8.52) has been known for a long time
in wavelet analysis [232], where it is widely used in practicefor example, when
 is taken to be the -function, (8.52) yields the so-called Morlet reconstruction
formula, its validity for a general group was proved in [379]. Another example,
in two dimensions, is obtained by taking a singular analyzing waveleta delta
function on a lineand this yields the inverse Radon transform [377].
An important consequence of (8.52) is that there are many kernels associated to
a given , namely, all the functions

1
K  (g, g ) = g |g  , (8.54)
C |C  

each one of which defines the evaluation map on H L2 (G, d ):


K  (g, g ) (g ) d (g ) = (g), H = W (H). (8.55)
G

It ought to be noted, however, that, if =  , K  is not a positive-definite kernel,

and, hence, not a reproducing kernel, although, as an integral operator on L2 (G, d ),
it is idempotent, that is, square integrable:

K  (g, g )K  (g , g ) d (g ) = K  (g, g ). (8.56)
G

8.3 A Class of Semidirect Product Groups

We have seen explicit examples of square integrable representations in the last

chapter, related to the SU(1,1) and the connected affine groups. Here we work out a
few more examples, in order to both illustrate the general theory of square integrable
representations better and to get a deeper understanding of the nature of the Duflo-
Moore operator C, appearing in the orthogonality relations. We then move on to
deriving a generalization of the notion of square integrability, in order to accomodate
CS of the Gilmore-Perelomov type and vector CS.
8.3 A Class of Semidirect Product Groups 217

The particular type of square integrable representations which we examine here,

following [146], arises from a class of semidirect product groups. The regular rep-
resentation of every group in this class consists of square integrable representations
only. This class of groups includes the affine group Gaff , on which the theory of
wavelets is based. We have already looked at the connected component of this group
in Sect. 8.1 and we shall take it up again, in detail, Sect. 12.2 [see also (4.25)(4.26)
in Sect. 4.1].
Consider Rn as an abelian group and denote by R  n its dual group. We shall

actually identify R with R and express duality through the usual scalar product,
n n

k ; x = k x, k R n , x Rn . The group GL(n, R) of all real n n non-singular

matrices has a natural action on Rn : (x, M)  Mx, where x Rn , M GL(n, R),
and Mx is just the matrix M acting on the vector x. The dual action on R  n is
T 
(k, M)  M k, k R and M denoting, as before, the transpose of the matrix
n T

M. Let H be an n-dimensional closed subgroup of GL(n, R), and assume that H has
an open free orbit in R n . This means that there exists a k0 R  n such that its orbit
under H,

O = H T k0 = {M T k0 : M H}, (8.57)
 n and furthermore,
is an open set in R

MT k = k M = In , (8.58)
 where In is the n n identity matrix. In other words, the stabilizer of
for any k O,

any k O is trivial. Note that in (8.57) k0 can be chosen to be any point in the orbit.
Also, for fixed k O, the map M  M T k, M H, is a homeomorphism between
H and O [335].
Let G = Rn  H be the semidirect product group, with elements g = (a, M), a
Rn , M H and multiplication law:

(a1 , M1 )(a2 , M2 ) = (a1 + M1 a2 , M1 M2 ). (8.59)

It is this class of semidirect products with which we shall be working i.e.,

semidirect products of Rn with closed n-dimensional subgroups of GL(n, R) having
open free orbits in R n . The affine group Gaff , mentioned above, is of this type, with
n = 1 and H = R\{0}. Let dG and dH denote the left Haar measures on G and
H, respectively. The Lebesgue measure da on Rn , when the latter is considered as
an abelian group, is the invariant measure on it. If det M is the determinant of M,
then it is not hard to see that
1
dG (a, M) = da dH (M). (8.60)
| det M|
Similarly, if G and H are the modular functions of G and H, respectively
[see (4.3)(4.5)], then

H (M)
G (a, M) = , (a, M) Rn  H. (8.61)
| det M|
218 8 Coherent States from Square Integrable Representations

Recall that for any integrable function f on Rn ,

 
1
f (Mx) dx = f (x) dx
Rn | det M| Rn

( dx being the Lebesgue measure on Rn ).

Consider next the Hilbert space H = L2 (Rn , dx) and its Fourier transformed
space H = L 2 (R  being given by
 n , dk), with the Fourier isometry F : H H

1
(F f )(k) = f(k) = exp [ik x] f (x) dx.
(2 )n/2 Rn

On H define the unitary representation U of G = Rn  H:

1
(U(a, M) f )(x) = f (M 1 (x a)), (a, M) G, f H. (8.62)
| det M|1/2
Since Rn G/H, this representation resembles a regular representation, but is
defined on a coset space, instead of on the group itself. Such a representation is
also called a quasi-regular representation. The unitarily equivalent representation
 M) = FU(a, M)F 1 , on the Fourier transformed space H,
U(a,  has the form:

 M) f)(k) = | det M|1/2 exp [ik a] f(M T k),

(U(a,  n , dk).
f L2 (R (8.63)

Referring to the discussion of induced representations in Sect. 4.2, we realize that U

is the representation of G induced from the trivial representation of the subgroup H.
In general, this representation is reducible. Let O R  n be an open free orbit and let
  
HO = L (O, dk) be the Hilbert subspace of L (R , dk) consisting of elements with
2 2 n

supports contained in O.  Denote by H  the subspace of L2 (Rn , dx) which is the

O
inverse Fourier transform of H  , i.e., if f H , then its Fourier transform f H
 .
O O O
  is an invariant subspace for the representation U.
From (8.63) it is then clear that H 
O
In fact, we shall see in Theorem 8.3.2 below that U  restricted to this subspace (or
equivalently, U restricted to HO) is irreducible.
Using the homeomorphism M  M T k (for some fixed k) let us transfer the
measure H from H to O,  and denote the resulting measure by . Then, for any
Borel set E H,

H (E) = (E  ), where E  = E T k = {M T k O : M E}. (8.64)

A simple argument using the left invariance of H then shows that is independent
of the point k chosen in the definition of the homeomorphism M  M T k. Further-
more, the Lebesgue measure dk of R  can be shown to be
 n , when restricted to O,
equivalent to d . Thus, there exists a positive, Lebesgue measurable function C on
O such that

d (k) = C (k) dk, 0 < C (k) < , (8.65)


the above relations holding for almost all k O.
8.3 A Class of Semidirect Product Groups 219

Lemma 8.3.1. The function C : O R+ satisfies the relations:

1. For any k O and integrable function f : O C,
  
f(M T k) dH (M) = f(k ) d (k ) = f(k )C (k ) dk ; (8.66)
H O O


2. For any M H and ( -almost) all k O,
H (M)
C (M T k) = C (k). (8.67)
|det M|

Proof. The two equalities in (8.66) follow from the definition of in (8.64) and the
relation between d and dk in (8.65). To prove (8.67) it is enough to note that for
any integrable f,
 
1
f(k)C (M T k) dk = f((M T )1 k)C (k) dk
O | det M| O

1
= f((M T )1 k) d (k),
| det M| O

 N H, and using (8.66) we

whence, writing k = N T k0 , for some fixed k0 O,
arrive at
 
1
f(k)C (M T k) dk = f((NM 1 )T k0 ) dH (N)
O | det M| H

(M)
= f(N T k0 ) dH (N),
| det M| H

(M)
= f(k) d (k).
| det M| O

where the second equality results from using the property (4.5) of the modular
function. 

Combining (8.61) and (8.67) we see that C is essentially the image of the
modular function on O,  and furthermore, if G is unimodular, then C (k) is a
constant function. In general, however, C could be an unbounded function. Define
 dk) as
  = L2 (O,
an operator C on H O

(C f)(k) = (2 )n/2 C (k)1/2 f(k), (8.68)

and let A denote its domain, i.e., f A if and only if


C (k)| f(k)|2 dk < . (8.69)
O

Clearly, A is dense in H  . It will turn out that C is precisely the Duflo-Moore

O
operator appearing in the orthogonality relations (8.36).
220 8 Coherent States from Square Integrable Representations

Theorem 8.3.2. The representation U  of the group G = Rn  H, restricted to the

Hilbert space H  dk), is irreducible and square integrable and every
  = L2 (O,
O
vector f A is admissible.
Proof. Let f1 , f2 A and 
h1 ,   , and assume these functions to be sufficiently
h2 HO
smooth so that integrations with respect to delta-measures and interchanges under
integrals may be performed. Then, writing

I=  M) f1 |
U(a,  M) f2 |
h1 U(a, h2  dG (a, M),
G

we find that
 
I= exp [ik a] f1 (M T k)
h1 (k) dk
G O

exp [ik a] f2 (M T k )

h2 (k ) dk da dH (M)
O
   
= h1 (k)
exp [i(k k ) a] da  h2 (k )
H O O Rn

f1 (M T k) f2 (M T k ) dk dk dH (M).

Next using the relation,


1
exp [i(k k ) a] da = (k k ),
(2 )n Rn

which holds in the sense of distributions (i.e., when integrated with respect to
smooth functions), and performing the integration in k , we obtain
 
I = (2 ) n 
h1 (k) f1 (M k) f2 (M k) dH (M) 
 T  T h2 (k) dk.
O H

Using (8.66) and (8.68) we get finally,


 M) f1 |
U(a,  M) f2 |
h1 U(a, h2  dG (a, M) = 
h1 |
h2  C f2 |C f1 . (8.70)
G

A simple continuity argument now enables us to extend this relation to all f1 , f2 A
and 
h1 ,   . We have recovered in this way, in the present context, exactly the
h2 H O
orthogonality relations appearing in (8.36) for a square integrable representation.
Taking f1 = f2 = f we obtain the admissibility condition (8.4) for f. To prove
 when restricted to H
square integrability of the representation U,  , it only remains
O
to establish its irreducibility, and then A would be the set of admissible vectors.
8.3 A Class of Semidirect Product Groups 221

Let f A be arbitrary but nonzero, and suppose that  hH   is a vector which

O
 M) f for every (a, M) G. Then, U(a,
is orthogonal to U(a,  M) f| h = 0, for all
(a, M) G. Hence, by (8.70),  h2 C f2 = 0, implying  h = 0, and since A is
dense in H  , this means that every vector in H   is cyclic. Thus, the representation
O O
U restricted to H  is irreducible and hence square integrable. 

O
In order to facilitate the comparison of this result with the general formalism
developed in Chap. 7, we may point out that the measure on O is invariant
under H, whereas the Lebesgue measure is only quasi-invariant. Hence we get a
unitarily equivalent picture if we use as representation space L2 (O,  d ) and the
corresponding representation UO, which is defined exactly as in (8.62)(8.63), but
without the factor | det M|1/2 . As stated in (8.65), this factor is the Radon-Nikodym
derivative, and equals C (k), upon using the homeomorphism M  M T ko , where M
is the element of H that maps the fixed base point ko to k.
Given a group of the type Rn  H, with H an n-dimensional subgroup of
GL(n, R), it is not always the case that open free orbits of H exist in R  n . However,
if there does exist one such orbit, then R  is in fact a disjoint union of such orbits.
n

Indeed, the left regular representation is in this case a direct sum of irreducible
subrepresentations, i.e., of square integrable representations [110]. In addition, the
restriction to free orbits may be dropped, as we shall see in Sect. 8.3.2 below.

8.3.1 Three Concrete Examples

8.3.1.1 The Similitude Group in Two Dimensions

As our first example, we take G = R2  H, where H = R+ SO(2) consists of

dilations and rotations of the plane. The group G is called the similitude group of
R2 and denoted SIM(2). This is the group that underlies the 2-D wavelet transform
that we shall study in Chap. 14.
For > 0, [0, 2 ), the element M = ( , ) H acts on R2 as:
    
x1 cos sin x1
x=  r( )x . (8.71)
x2 sin cos x2

The dual action on R 2 is k  r( )k. Hence there are two orbits, namely Oo =
{0} and O1 = R
 2 \ {0}. The latter is open and free, as follows from inspection.
Therefore the corresponding unitary representation

1
(U(a, M( , )) f )(x) = f ( 1 r( )(x a)), f L2 (R2 , dx), (8.72)

or, equivalently,
222 8 Coherent States from Square Integrable Representations

 M( , )) f)(k) = eika f( r( )k),

(U(a,  2 , dk),
f L 2 (R (8.73)

is irreducible and square integrable. The group H = R+ SO(2) is unimodular

(being abelian) and its Haar measure is dH ( , ) = 1 d d . The homeomor-
phism M  M T ko between H and the open orbit O1 may be expressed by choosing
the fixed vector ko = (1, 0) R2 and mapping M = ( , ) onto k = M T ko = 0,
where = |k|, = arg k. Then the measure on O1 , transferred from dH , is

d d|k| 1
d (k) = d = d = 2 dk.
|k| |k|

Thus the Duflo-Moore operator is

(C f)(k) = 2 |k|1 (k). (8.74)

8.3.1.2 The Affine Poincar Group in 1+1 Dimensions

Let G = R2  H, where H = R+ L+ (1, 1) is the Lorentz group in 1-time and 1-

space dimensions together with dilations. Elements in R+ L+ (1, 1) are 2 2 real
matrices M of the form
 
cosh sinh
M = M( , ) = , M = M T , > 0, < < . (8.75)
sinh cosh

Thus G is the affine Poincar group in 1-space and 1-time dimensions, i.e., the
Poincar group along with dilations. This group may also be called SIM(1,1), since
it is the similitude group of the 1+1 dimensional Minkowski space. Although this
group has square integrable representations, the Poincar group itself, P+ (1, 1) =
R2  L+ (1, 1) (i.e., without dilations), has no representations in the discrete series.
(We shall return to this point later in Sect. 10.1, and demonstrate that P+ (1, 1) has
nevertheless a representation which is square integrable mod(T, ), for a certain
class of sections , T being the time-translation subgroup).
Following the physicists convention, we denote vectors in R2 and R  2 in this
case by
   
x0 k0  2,
x= R ,2
k= R (8.76)
x k

and use the Minkowski inner product

x k = x0 k0 xk (8.77)
8.3 A Class of Semidirect Product Groups 223

to define the dual pairing k ; x. Also, instead of the usual Fourier transform we
shall use, as is customary in a relativistic theory, the relativistic Fourier transform:

1
f(k) = eikx f (x) dx, f L2 (R2 , dx),
2 R2

1  2 , dk),
f (x) = eikx f(k) dk, f L2 (R (8.78)
2 R 2

(where of course, dx = dx0 dx and dk = dk0 dk). The representation (8.62) now
assumes the form [note that M( , )1 = M( 1 , )]:
1
(U(a, M( , )) f )(x) = f (M( 1 , )(x a)),

a = (a0 , a) R2 , f L2 (R2 , dx). (8.79)

Using the identity,

   
1 1 0 1 0
M( ,) M( , ) = , (8.80)
0 1 0 1
the relativistic Fourier transformed representation is seen to be
 M( , )) f)(k) = eika f(M( , )k),
(U(a,  2 , dk),
f L2 (R (8.81)

which of course is unitarily equivalent to (8.79). The dual action of an element

M( , ) R+ 2
L+ (1, 1) on R is found using k (M( , )x) = (M( , )k) x,
from which orbits under this action can be identified. Indeed, for the four vectors,
       
1 1 0 0
k+ = , k = , kr = , k = , (8.82)
0 0 1 1
we obtain the four orbits

Oj = {M( , )k j : j = +, , r, ; > 0, < < }. (8.83)

The orbit O+ , for example, consists of all vectors k R

 2,
   
k0 cosh
k= = , > 0, < < ,
k sinh

and is therefore the open cone in the (k0 k)-plane, lying above the lines k0 = k,
k0 > 0. Similarly, O is the open cone lying below the lines k0 = k, k0 < 0, while
Or and O are the open right and left cones, i.e., lying to the right and left of the
lines k = k0 , k > 0 and k = k0 , k < 0, respectively. Each one of these cones is
 2:
an open free orbit and their union is dense in R

O+ O Or O = R

 2 \{(k0 , k) : k0 = k}.
224 8 Coherent States from Square Integrable Representations

Besides these orbits, there are the two null cones and the origin, corresponding to
     
1 1 0
k= , k= , k= ,
1 1 0

 = L 2 (R
respectively, which are not open and free. Thus H  2 , dk) decomposes as the
orthogonal direct sum

 = L2 (O+ , dk) L2 (O , dk) L2 (Or , dk) L2 (O , dk),

H

and restricted to each one of these component Hilbert spaces, the representation
 M) is irreducible and the four subrepresentations are mutually inequivalent.
U(a,
It is not hard to calculate the invariant measure on these orbits. Indeed, on any one
of them,
dk0 dk
d (k) = , (8.84)
|k02 k2 |
and consequently one gets for the Duflo-Moore operator C in (8.68),

(C
h)(k) = 2 |k02 k2 |1/2
h(k), (8.85)

on a dense set of vectors in the appropriate Hilbert space. (The square integrability
of the affine Poincar group in higher space-time dimensions has been studied in
[160]. We shall look at these higher dimensional analogues in Chap. 14).
As a final remark, notice that the square integrable representations considered
here are the restrictions of certain square integrable representations of the (semisim-
ple) conformal group SO(2,2), of which the affine Poincar group P+ (1, 1) is a
subgroup.

8.3.1.3 Dilations and Translations in R2

For our third example we take the group G = R2  H, with H consisting of 2 2

matrices of the type
 
a1 0
M= , a1 a2 = 0. (8.86)
0 a2
Then G = Gaff Gaff , where Gaff is the full (disconnected) affine group [see
Sect. 12.2 and also (4.25)(4.26)]. For this group, there is only one open free orbit
 2:
in R
 
1
O = {k = M : M H}. (8.87)
1
8.3 A Class of Semidirect Product Groups 225

Clearly, O = R
 2 \ {R
 R}
 {k R  2 : k1 k2 = 0}, which is dense in R
 2 , and the
invariant measure on it is easily computed to be
 
dk k1
d (k) = , k= . (8.88)
|k1 k2 | k2
The representation
 M)
(U(a, h)(k) = exp [ik a]
h(M 1 k), (8.89)
 dk) is now irreducible and square integrable, and the operator C
of G on L2 (O,
assumes the form

(C
h)(k) = 2 |k1 k2 |1/2
h(k), (8.90)
 dk).
for a dense set of vectors in L2 (O,

8.3.2 A Broader Setting

The first example in the preceding section extends in a straightforward way to higher
dimensions, that is, to the similitude group of Rn ,

SIM(n) = Rn  (R+
SO(n)).

However, for n > 2, the theory outlined above no longer applies. Indeed, here too,
there are only two orbits, namely Oo = {0} and O1 = R  n \ {0}, and the unitary
representation associated to the latter is irreducible. The orbit O1 is open in R  n , but
it is not free. For instance, the point ko = (0, 0, . . . , 1) is stabilized by SO(n 1) and
so does any other point of the orbit. Nevertheless, this representation is still square
integrable, as a consequence of a more general result, proved in [292], which we
now sketch.
With the same notation as in the preceding section, the following holds:
Theorem 8.3.3. Let U  be the unitary (quasi-regular) representation of G = Rn  H
 , dk). Let O be any measurable subset of R
in L (R
2 n  n , invariant under the action of
  be the restriction of U
H and let U  dk). Then:
 to H  = L2 (O,
O O

(i)   is irreducible iff the action of H on O is ergodic;

UO
(ii)   is square integrable iff the stabilizer of any point of O is compact.
UO
Here ergodicity of the action of H means that the closure of any H-invariant subset
of O either is of measure zero or coincides with O. In other words, O consists of the
closure of a unique dense orbit under H, plus possibly some sets of measure zero
(closures of orbits of smaller dimension). Note that this condition essentially rules
out discrete sets of dilations. The proof of this theorem is almost the same as that of
Theorem 8.3.2.
226 8 Coherent States from Square Integrable Representations

The Duflo-Moore operator (8.68), in this general situation, is again the operator
of multiplication by the function C(k)1/2 , where C(k) is a Radon-Nikodym
derivative, just as in (8.65). Identifying as before a point k O1 with an element
h = h(k) H by k = h(k)ko , for some fixed base point ko O1 , one has

C(k) = H (h) | det h|1 , k = h(k)ko .

The above theorem is tailor-made to fit the SIM(n) group. Here H = R+ SO(n)
is unimodular, and, for k O1 , one may choose ko = (0, 0, . . . , 1), so that h(k) =
(|k|, R(ko k)), where R(ko k) denotes some rotation that maps ko to k. Thus
det h(k) = |k|n and the Duflo-Moore operator is the operator of multiplication by
(2 )n/2 |k|n/2 .
An identical analysis can be carried out for the positive mass representations of
the affine Poincar group in 1+3 dimensions. The corresponding orbit is the interior
of the future lightcone,

O+ = {k : k0 > 0, k02 k2 > 0},

(generalizing the orbit O+ of the 1+1 dimensional case discussed in the previous
section), and the stabilizer of any of its points is isomorphic to SO(3). The same is
true for the past lightcone O , but not for the space-like region k2 < 0. For instance,
the point k(o) = (0, . . . , 0, 1), and thus every point of that orbit, has a stabilizer
isomorphic to SOo (1,2). We shall come back to these higher dimensional examples
in Sect. 16.2.
Actually, both Theorems 8.3.2 and 8.3.3 follow from a more general result [419]
(see also [41, 42]), that completely settles the question of the square integrability of
all induced representations of semidirect products of the type Rn  H, where H is a
locally compact topological group, acting continuously on Rn . This result is stated
in Theorem 9.3.1 where the induced representations of such groups are discussed in
detail (see Sect. 9.2.4). Simply stated, according to this theorem, the quasi-regular
representation U  is square integrable, if and only if the orbit O has positive Lebesgue

measure in R . n

8.4 A Generalization: - and V -Admissibility

It is possible to generalize the notion of square integrability of a group representation

as envisaged in Sect. 8.1, to include CS of the Gilmore-Perelomov type. As a first
example of such a generalization, we sketch a number of results, following [369],
which extend Theorems 8.1.3 and 8.2.1 in an interesting way. Later, we go further
and discuss the square integrability of vector CS as well in this context.
Let H G be a closed subgroup and suppose that : H C is a unitary character
of H, i.e., for all h, h1 , h2 H,

(h) C, | (h)| = 1 and (h1 ) (h2 ) = (h1 h2 ). (8.91)

8.4 A Generalization: - and V -Admissibility 227

Suppose that we are given a UIR g  U(g) of G on H and assume that there are
nonzero vectors H such that [see (7.9)]

U(h) = (h) , h H. (8.92)

Also assume that X = G/H carries an invariant measure and let : X G be a

Borel section.
Definition 8.4.1. A nonzero vector H is said to be -admissible for the UIR
g  U(g) if it satisfies (8.92) and

I ( ) = |U( (x)) | |2 d (x) < . (8.93)
X

Obviously, this definition of -admissibility is independent of the choice of the

particular section . Denote by A the set of all -admissible vectors in H and by
H the closure of this set, which clearly is a subspace of H. If 1 and 2 are two
different unitary characters of H and the vectors 1 , 2 H are respectively 1 - and
2 -admissible, then the fact that 1 and 2 are eigenvectors of the unitary operator
U(h) corresponding to different eigenvalues 1 (h) and 2 (h), implies that 1 and
2 are mutually orthogonal. From this it follows that H1 and H2 are mutually
orthogonal subspaces of H.
Next let h : G X H be the cocycle h(g, x) = (gx)1 g (x) [see (4.37)].
Then since defines a one-dimensional unitary representation of H, the operators
U(g), g G, acting on the Hilbert space L2 (X, d ),

( U(g) )(x) = (h(g1 , x))1 (g1 x), L2 (X, d ), (8.94)

realize the representation of G which is induced from , as is immediately verified

by comparing with (4.39)(4.41). If it can now be proved that (with an appropriate
normalization N) the operators

1
a( ) = | (x)  (x) | d (x), (x) = U( (x)) ,
N

where runs through all Borel sets of X and is an -admissible vector,

define a normalized POV-measure, then from our discussion of vector CS in
Sects. 4.2.1, 7.1.2 and 7.1.3 it would follow that the representation U is equivalent
to a subrepresentation of the induced representation U. This indeed is the case,
and following the lines of the proof of Theorem 8.1.3, one arrives at the following
generalization of it:
Theorem 8.4.2. Let g  U(g) be a UIR of the locally compact group G and
suppose that is a unitary character of the closed subgroup H. Let X = G/H
and let : X G be any Borel section. Assume that A = {0} and A . Then
the representation U is square integrable mod (H, ) and the mapping W : H
L2 (X, d ) defined by
1
(W )(x) = [c ( )] 2  (x) | , H, (8.95)
228 8 Coherent States from Square Integrable Representations

where,

I ( )
c ( ) = and (x) = U( (x)) , (8.96)
 2

is a linear isometry onto a (closed) subspace H of L2 (X, d ). On H the resolution

of the identity,

1
| (x)  (x) | d (x) = I, (8.97)
c ( ) X

holds. The subspace H = W H L2 (X, d ) is a reproducing kernel Hilbert space

with reproducing kernel

1
K (x, y) =  | . (8.98)
c ( ) (x) (y)

Furthermore, W intertwines U and the induced representation U:

W U(g) = U(g)W , g G. (8.99)

The above theorem settles the question of square integrability of representations

admitting Gilmore-Perelomov type CS, for if in the definition of these states
in (7.9)(7.11) the vector is -admissible with (h) = exp [i (h)], then U is
square integrable mod(H). Note that if (X), the total measure of the homogeneous
space X = G/H, is finite, then any H, satisfying (8.92) is -admissible and
consequently the representation U is square integrable.
Obvious examples of applications of the above theorem are the representa-
tions (2.32) of the Weyl-Heisenberg group GWH giving rise to the canonical
CS (2.37) and the discrete series representations (4.86) of SU(1,1), leading to
the CS (4.99). In the case of the Weyl-Heisenberg group, the character is the
representation,  exp(i ), of the phase subgroup and every vector in
the representation space is -admissible. For the SU(1,1) representations U j , the
representation k  exp(i j ) of the maximal compact subgroup K plays the role
of , with (2 j + 1)1/2 u0 [see (4.99)] being an -admissible vector.
In general, A is not dense in H, i.e., H = H. However, the following version
of the orthogonality relations can be proved, generalizing Theorem 8.2.1 [369]:
Theorem 8.4.3. Let G be a locally compact group, H a closed subgroup admitting
the unitary character , X = G/H and : X G any Borel section. Let U be a
UIR of G on the Hilbert space H, A = {0} and H H the closure of A . Then
there exists a unique positive, invertible, self-adjoint operator C on H such that
for arbitrary ,  A and ,  H,

 (x) |  (x) |   d (x) = C  |C  |  . (8.100)
X
8.4 A Generalization: - and V -Admissibility 229

If U  is another UIR of G, on the Hilbert space H , which is not unitarily equivalent

to U,  : X G any Borel section and any -admissible vector for the
representation U  , then for arbitrary H and  H ,

 (x) | H   (x) |  H d (x) = 0. (8.101)
X

Again, the obvious examples, illustrating this theorem are the representations
U of the Weyl-Heisenberg group [see (2.32)]. Taking = 1, writing U for
the corresponding representation and U( , q, p) = ei U(q, p), the orthogonality
relations now assume the form:

1
U(q, p)1 |1 U(q, p)2 |2  dq dp = 1 |1 2 |2 , (8.102)
2 R2

for arbitrary
vectors 1 , 2 , 1 and 2 in the Hilbert space, and the operator
C = 2 I.
An elegant generalization of Theorem 8.4.2, to include cases where the unitary
character is replaced by a representation of the subgroup H on a finite dimensional
Hilbert spaceleading thereby to a square integrability condition for vector coherent
statescan be derived based on the geometrical considerations of Sect. 7.1.3.
We begin by generalizing the concept of -admissibility. As before, we assume
that the coset space X = G/H carries an invariant measure, d .
Definition 8.4.4. Let U be a UIR of G on the Hilbert space H and H a closed
subgroup of G. Let V be a finite dimensional UIR of H. A subspace K H is said
to be V -admissible if U(H) restricted to K is unitarily equivalent to V and if there
exists a non-zero vector K such that the condition (7.34) is satisfied, namely,

I( ) :=  |PK ([g])  d ([g]) < ,
X

where PK is the projection operator for the subspace K H and PK ([g]) =

U(g)PKU(g) , is satisfied. If such a subspace K exists, the representation U is said
to be square integrable mod(H).
With the conditions of the above definition being satisfied, let us set

I( )
cV ( ) = , (8.103)
 2

and denote by D the dense set in H spanned by the vectors U(g) , K, g G.

2 (X, d ),
Then, in view of the second equality in (7.33), the map W : D LK

(W )(g) = [cV ( )]1/2 PKU(g1 ) , (8.104)

230 8 Coherent States from Square Integrable Representations

is well defined. The set D is stable under U and W (being a multiple of the map
WK in (7.25)) intertwines U and the induced representation V U. Retracing the lines
of the proof of Theorem 8.1.3, it is easily established that W is closed, while an
application of Lemma 4.3.3 then shows that W is an isometry. Thus we have proved
the following result:
Theorem 8.4.5. Let G be a locally compact group, H a closed subgroup of G, U
a UIR of G on the Hilbert space H and V a finite dimensional UIR of H. Let K be
a V -admissible subspace of H. Then, the mapping (8.104) extends to all of H as a
2 (X, d ). On H, the
linear isometry, so that its range, H , is a closed subspace of LK
resolution of the identity

1
PK ([g]) d ([g]) = I, (8.105)
cV ( ) X

holds. The subspace H = W (H) LK 2 (X, d ) is a reproducing kernel Hilbert

space; the corresponding projection operator

P = W W , P LK
2
(X, d ) = H , (8.106)

has the reproducing kernel K :


(P )(g) = K (g, g ) f(g ) d ([g ]), f LK
2
(X, d ),
X
1
K (g, g ) = PKU(g1 )U(g )PK , (8.107)
c( )

as its integral kernel. Furthermore, W intertwines U and the induced representa-

tion V U,

W U(g) = V U(g)W , g G. (8.108)

A few remarks are in order here:

As a consequence of this theorem, the quantity cV ( ) cV (K) is independent of
the particular K chosen to evaluate it.
The condition (8.107) translates into the property

K (g, g ) K (g , g ) d ([g ]) = K (g, g ). (8.109)
X

The set of positive, bounded operators


a( ) = PK ([g]) d ([g]),

8.4 A Generalization: - and V -Admissibility 231

for arbitrary Borel sets X, form a normalized positive operator-valued

measure and {U, a} is a system of covariance:

U(g)a( )U(g) = a(g ), (8.110)

which of course, is just a restatement of the intertwining property (8.108).

If we choose H = {e}, the subgroup consisting of only the identity element,
and take for V (or ) the trivial representation, then the V -admissibility (or -
admissibility) of a vector H, for the representation U, is the same as
admissibility in the sense of Definition 8.1.1. Thus if such a vector exists, U
is square integrable.
In order to define coherent states it is necessary to choose a section : G/H G
and an orthonormal basis, i , i = 1, 2, . . . , n, in K (n being its dimension). Then,
PK = ni=1 | i  i |, and

S = {[cV (K)]1/2 i (x) = [cV (K)]1/2 U( (x)) i : i = 1, 2, . . . , n, x X = G/H},

is a family of (in general vector) coherent states for the representation U. The
resolution of the identity (8.105), in terms of these CS, now has the familiar form:
n 
1
[cV (K)] |i (x) i (x) | d (x) = I,
i=1 X

which again is independent of the section chosen to define the CS.

8.4.1 Example of the Galilei Group

Let us illustrate the results of Theorems 8.4.2 and 8.4.5 by examples drawn from
certain representations of the Galilei group. The kinematics of a free nonrelativistic
physical particle is governed by its invariance under the action of the Galilei
group [442444, In54], which is a ten parameter group G0 of transformations of
Newtonian space-time. An element g G0 is of the form

g = (b, a, v, R), b R, a, v R3 , R SO(3), (8.111)

where b is a time and a a spatial translation, v a velocity boost and R a spatial

rotation. The action of g on a space-time point (x,t) is given by

g (x,t) = (x ,t  )
x = Rx + vt + a, t  = t + b. (8.112)
232 8 Coherent States from Square Integrable Representations

Actually, in quantum mechanics one needs to work with a central extension

of G0 , and also, in order to accomodate particles with half-integral spins, it is
necessary to replace the rotation group SO(3) by its universal covering group
SU(2). Denoting this extended Galilei group by G, its elements will be written as
g = ( , b, a, v, ), R, SU(2). A general element of SU(2) is a 22 complex
matrix,
 

= , , C, | |2 + | |2 = 1, (8.113)

and the elements of the corresponding 3 3 rotation matrix R( ) are then

R( )i j = Tr[ j 1 i ], i, j = 1, 2, 3, (8.114)

the j s being the three Pauli matrices:

     
01 0 i 1 0
1 = , 2 = , 3 = . (8.115)
10 i 0 0 1

The group multiplication in G is

g1 g2 = (1 + 2 + (g1 , g2 ), b1 + b2 , a1 + R(1 )a2 + v1 b2 , v1 + R(1 )v2 , 1 2 ),

gi = (i , bi , ai , vi , i ), i = 1, 2, (8.116)

where : G G R is a multiplier, which up to equivalence (in the sense of

cocycles), can be taken to be

1
(g1 , g2 ) = m v21 b2 + v1 R(1 )a2 (8.117)
2

Thus there is a one parameter family of multipliers (that is, of central extensions),
indexed by the parameter m, which is interpreted as mass [442, 443].
The multiplier has the properties

(e, e) = 0,
(g1 , g2 ) + (g1 g2 , g3 ) = (g1 , g2 g3 ) + (g2 , g3 ), g1 , g2 , g3 G. (8.118)

The identity element of G is e = (0, 0, 0, 0, I2 ) (where I2 = 2 2 identity matrix),

while the inverse of g = ( , b, a, v, ) is the element
  
1
g1 = m v2 b v a , b, R( 1 )(vb a), R( 1 )v, 1 . (8.119)
2

A number of subgroups of G are of particular interest to us here for the construction

of CS. The subroup G having elements ( , 0, a, v, ), i.e., with the time part b
8.4 A Generalization: - and V -Admissibility 233

set equal to zero, is called the isochronous extended Galilei group. Another useful
subgroup of Gone which is also a subgroup of G consists of elements of the
type g = ( , 0, a, v, I2 ). We denote this group by GWH and note that it is isomorphic
to the Weyl-Heisenberg group for three degrees of freedom. The subgroup S of
space translations with elements g = (0, 0, a, 0, I2 ), and of pure Galilean translations
(or velocity boosts) V , with g = (0, 0, 0, v, I2 ), will also play important roles in
the construction to follow. To these we add the subgroup of time translations T
with g = (0, b, 0, 0, I2 ) and the phase subgroup having elements g = ( , 0, 0, 0, I2 ).
Finally, we note that G can be looked upon as the semidirect product,

G = ( T S )  L, L = V  SU(2), (8.120)

and observe that T S is an abelian subgroup of G.

Following [443], we proceed now to explicitly compute a class of UIRs of G,
before actually computing their CS. The reasons for this extended exercise are
twofold: First, it will illustrate how the general technique of induced representations
may be used to generate unitary irreducible representations of certain semidirect
product groups [453, Mac68, Mac76], and secondly, in Chap. 9 we shall encounter
representations of other semidirect product groups which are similarly obtained. The
particular representations of G which concern us here are the ones which describe
free nonrelativistic quantum particles of mass m > 0 and spin j = 0, 1/2, 1, 3/2, . . ..
The construction of these representations exploits the semidirect product nature of G
displayed in (8.120). The momentum space realizations of these representations are
carried by the Hilbert spaces K j L2 (R  3 , dk), where K j is a (2 j + 1)-dimensional
spinor space (which we identify with C2 j+1 ), R  3 the dual of R3 , (which we again
identify with R ), and dk the Lebesgue measure on R
3  3 . In Sect. 8.4.2, we shall
discuss the representations of the isochronous Galilei group G related to these
representations of G and then construct the corresponding CS. In Chap. 10, finally,
we shall obtain the CS of the full group G.
The spaces K j carry the unitary irreducible representations of SU(2), which we
denote by D j ( ), SU(2). The abelian subgroup T S of G has unitary
characters ,E,p of the type,

,E,p ( , b, a) = exp [i( + Eb + p a)], , E R, p R3 . (8.121)

We determine the dual action of the subgroup L on such a character. Since

 
21 2 3
(0, 0, 0, v, )( , b, a, I2 ) = + m v b + v R( )a , b, R( )a + vb, v, ,
2
writing  ,E  ,p = (v, ) [ ,E,p ] for the transformed character, under the dual action
of (v, ) L, we find
 
21 2 3
 ,E  ,p ( , b, a) = ,E,p + m v b + v R( )a , b, R( )a + vb, v,
2

m 2 
= exp i{ + (E + v + p v)b + R( 1 )(m v + p) a} ,
2
234 8 Coherent States from Square Integrable Representations

and thus,

 = ,
m v2
E = E + + p v, (8.122)
2
p = R( 1 )(m v + p).

Next we observe that

p 2 p2
E =E = E0 = const. (8.123)
2m 2m

Thus, the orbit of a character under the dual action of the subgroup L of G is a
paraboloid characterized by two constants, E0 , R and, without loss of generality,
we choose = 1. Denoting an orbit by OE0 , the constant E0 may be identified with
the internal energy of the system. The character 1,E0 ,0 OE0 is then a natural choice
of a representative for the orbit. From (8.122), the stability subgroup of this character
is seen to be SU(2), so that

OE0 L/SU(2) R
 3. (8.124)

The invariant measure on this orbit may therefore be simply taken to be the
Lebesgue measure dk on R  3 . According to the general theory, the irreducible
representations of G can now be obtained by inducing from the UIRs VE0 , j of
the subgroup ( T S ) SU(2), where

VE0 , j ( , b, a, ) = 1,E0 ,0 ( , b, a) D j ( ) = exp [i( + E0 b)] D j ( ). (8.125)

 3 G:
Define the section : L/SU(2) R
! k "
(k) = 0, 0, 0, , I2 . (8.126)
m

Computing g1 (k), for g = ( , b, a, v, ) G, explicitly,

g1 (k) =
! 1
k  "
= m[ v2 b v a], b, R( 1 )[a vb], R( 1 ) v , 1
2 m
!
k  "
= 0, 0, 0, R( 1 ) v , I2
m
! k2 b
k  "
+ k a, b, R( 1 ) a b , 0, 1 ,
2m m
8.4 A Generalization: - and V -Admissibility 235

which yields the cocycle h : GR

 3 L, h(g, k)= (gk)1 g (k) [see (4.37)(4.38)],
! k2 b
k  "
h(g1 , k) = + k a, b, R( 1 ) a b , 0, 1 ,
2m m
! 2
k b k "
h(g1 , k)1 = + k a, b, a b, 0, . (8.127)
2m m
Thus finally,

! k2 " 
VE0 , j (h(g1 , k))1 = exp i{ + E0 + b k a} D j ( ), (8.128)
2m
and inserting this into (4.41), we obtain the representation U  E0 , j of G which
is induced from the UIR VE0 , j of the subgroup ( T S ) SU(2). This
 3 , dk) in the manner:
representation acts on the Hilbert space K j L2 (R

! 2" 
 E0 , j ( , b, a, v, ))(k) = exp i{ + E0 + k b k a}
(U
2m
D ( ) (R( 1 )(k mv)),
j
(8.129)
 3 , dk). This is the so-called momentum space representation
for all  K j L2 (R
for a non-relativistic system of mass m, internal energy E0 and spin j. When dealing
with a single system, the internal energy may be set equal to zero, simply by
adjusting the rest state. We shall do this in what follows and write the representations
as U j . The representation on configuration space is obtained by going over to the
Fourier transformed space K j L2 (R3 , dx) of vectors

(x) = (2 ) 3/2
eikx (k) dk.
3
R

The transformed representation U j assumes the form:

P2 
(U j ( , b, a, v, ) )(x) = exp i{ + b + mv (x a)}
2m
D j ( ) (R( 1 )(x a)), (8.130)

where P = i, or, in terms of the time-translated functions,

P2 
(x,t) = (exp i t )(x),
2m
the representation appears in the more familiar form:

mv2 
(U j ( , b, a, v, ) )(x,t) = exp i{ (t b) + mv (x a)}
2
 
D j ( ) R( 1 )(x a v(t b),t b . (8.131)
236 8 Coherent States from Square Integrable Representations

8.4.2 CS of the Isochronous Galilei Group

Unitary irreducible representations of the isochronous Galilei group G can now be

read off from (8.129) after setting b = 0. These representations are also carried by
 3 , dk), and using the same notation as before (also taking E0 = 0),
K j L 2 (R

 j ( , a, v, ))(k) = exp [i( k a)] D j ( ) (R( 1 )(k mv)),

(U

( , a, v, ) G . (8.132)

We proceed to demonstrate the square integrability mod( SU(2)) of these

representations in the sense of Theorems 8.4.2 and 8.4.5, following [2325, 369].
Denote by e j , = j, j + 1, . . . , j 1, j, the canonical basis of K j (i.e., e j is
the vector with components j+ +1, ,  = 1, 2, . . . , 2 j + 1), and let D j ( )  be the
matrix elements of D j ( ) in this basis. Define a function : R+ C, normalized
in the manner

| (r)|2 r2 dr = 1. (8.133)
0

(This is an element of the radial Hilbert space L2 (R+ , r2 dr)). Consider first the
representations U  3,
 j for integral values of the spin ( j = 0, 1, 2, . . .). For any k R

denote by Y j (k), = j, j + 1, . . . , j 1, j, the j-th spherical harmonics in the
angles of the 3-vector k,1 and let d (k) be the area element on the unit sphere S2 .
The following properties of the spherical harmonics are well known:


Y j (k)Y j (k) d (k) = j j  , (8.134)
S2
j

Y j (R1 k) =

D j (R)  Y j (k),  3,
kR R SO(3), (8.135)
= j

4 j

Pj (cos ) = Y (k)Y j (k ),
2 j + 1 = j j

k k  3;
where cos = , k, k R (8.136)
k k 

and P j is the Legendre polynomial of order j.

1 This is an unconventional notation for spherical harmonics, and it differs from the one used in

Chap. 7.
8.4 A Generalization: - and V -Admissibility 237

 3 , dk),
 K j L 2 (R
Consider finally the vector
j
1
 (k) =

(2 j + 1)1/2 = j
e j Y j (k) (k),  3,
kR (8.137)

and with as in (8.133). Let be the unitary character

( , ) = ei . (8.138)

of the subgroup H = SU(2) of G . An application of (8.135) then yields

 j ( , 0, 0, )
U  = ( , )
. (8.139)

The quotient space = G /H is isomorphic to R6 , by virtue of the

decomposition

( , a, v, ) = (0, a, v, I2 )( , 0, 0, ). (8.140)

We parametrize by (q, p) R6 , with q = a and p = mv, in terms of which the

invariant measure on it is just the Lebesgue measure dq dp. A natural choice of a
section : G is then
! p "
(q, p) = 0, q, , I2 , (8.141)
m
and by a straightforward computation, of the type used in the proof of
Theorem 8.3.2, and using relations such as (8.134), we find that

) =
I (  (q,p) |
|  |2 dq dp = (2 )3 . (8.142)

We have thus shown that the vector  3 , dk) is -admissible, and

 K j L 2 (R

therefore U is square integrable mod( SU(2)). Since 
j  2 = 1, c = (2 )3
2 
[see (8.96)], and hence the map W : K L (R , dk) L ( , dq dp),
j 3 2

1 1 9 ! " :
(W )(q, p) =   (q,p) |  =
U j 0, q, p , I2   (8.143)
(2 )3/2 (2 )3/2 m
is an isometry. Furthermore, we have the resolution of the identity,

1
 (q,p) 
|  (q,p) | dq dp = I, (8.144)
(2 )3

for the Galilei coherent states, which we now define to be the vectors
1
 (q,p) ,
(q, p) R6 ,
(2 )3/2
 3 , dk). The image of K j L2 (R
in K j L2 (R  3 , dk) in L2 ( , dq dp) under the
isometry (8.143) is a reproducing kernel Hilbert space H , the elements of which
are the continuous functions:
238 8 Coherent States from Square Integrable Representations

j 
1
(q, p) = 3 1/2
[(2 j + 1)(2 ) ] m= j 3
R
eikq Y jm (k p) (k p) m (k) dk,

(8.145)
 3 , dk).
where (k) = ( j (k),  j+1 (k), . . .  j1 (k),  j (k)) K j L2 (R
The reproducing kernel defining H is

1
K (q, p; q , p ) =  (q,p) |
  (q ,p ) 
(2 )3
  
1  (k p) (k p )
= ei[k (q q )] Pj
4 (2 )3 3
R k p k p 

(k p) (k p ) dk, (8.146)

which can be obtained using (8.136). The coset decomposition (8.140) implies the
following action of g = ( , a, v, ) G on :

g1 (q, p) = (R( 1 )(q a), R( 1 )(p mv)). (8.147)

Similarly, the cocycle h : G SU(2) is found to be

h(g1 , (q, p)) = ( mv (q a), 1 ). (8.148)

Using these we obtain the form of the representation

 j (g)W1 ,
U (g) = W U g G ,

 j under the isometry (8.143):

on L2 ( , dq dp), arising as the image of U

(U ( , a, v, ) )(q, p) = exp [i{ + mv (q a)}]

(R( 1 )(q a), R( 1 )(p mv)), (8.149)

( H ), which extends to the whole of L2 ( , dq dp) as the representation of G

which is induced from .
To obtain a version of the orthogonality relations (recall that we are still
considering representations for j = 0, 1, 2, . . .), first note that any -admissible
vector must be of the form (8.137) with L2 (R+ , r2 dr) (but not necessarily
normalized). Then if   are -admissible vectors, a similar computation as that
,
leading to (8.142) results in the orthogonality relation:

 (q,p) |
  (q,p) |  dq dp = (2 )3    | ,
  | (8.150)

8.4 A Generalization: - and V -Admissibility 239

 3 , dk), illustrating Theorem 8.4.3. This relation also

for arbitrary ,  K j L2 (R
shows that in this case the Duflo-Moore operator C , appearing in Theorem 8.4.3,
is a multiple of the identity: C = (2 )3 I.
In order to take into account the representations for j = 1/2, 3/2, 5/2, . . .,
it is necessary to adopt a more general approach. In fact, using the notion of
V -admissibility, we shall now derive a wide class of square integrable CS for all
values of j, integral or half-odd integral. Let us first decompose the representation
space K j L2 (R  3 , dk) into rotationally invariant subspaces. For any  K j
2 
L (R , dk), and SU(2),
3

 j (0, 0, 0, ) = D j ( )(R( 1 )k).

(U (8.151)

 3 , dk) can be written as the tensor product

The space L2 (R

L2 (S2 , d (k)) L2 (R+ , r2 dr),

where S2 is the unit sphere and r = k. Let {n }

n=1 be an orthonormal basis of
L2 (R+ , r2 dr):

n (r)n (r) r2 dr = nn . (8.152)
R+

For each  = 0, 1, 2, . . ., the spherical harmonics Y (k), = ,  + 1, . . . ,  1, ,
span a subspace of L2 (S2 , d (k)) which is stable under rotations. Furthermore,


fn (R1 k) = D  (R)  fn (k), (8.153)


where

fn (k) = Y (k) n (k). (8.154)

 3 , dk) spanned by the unit vectors

For fixed , n, let Kn be the subspace of L2 (R

fn , = ,  + 1, . . . ,  1, . Then,


 3 , dk) =
L 2 (R Kn , (8.155)
n=1 =0

and each subspace Kn carries a representation of SU(2) unitarily equivalent to the
UIR D  .
Next, in view of the well-known decomposition (e.g., for angular momenta)
j+

D j ( ) D  ( ) D J ( ), (8.156)
J=| j|
240 8 Coherent States from Square Integrable Representations

 3 , dk) decomposes as
( implying unitary equivalence), the space K j L2 (R


 j+

 3 , dk) =
K j L 2 (R KnJ , (8.157)
n=1 =1 J=| j|

in which each KnJ is a (2J + 1)-dimensional space, carrying a UIR of SU(2),

unitarily equivalent to D J . The vectors

 nJM (k) = C( j, ; , M | JM) ej YM (k) (k),

(8.158)

for M = J, J + 1, . . . , J 1, J, and where the C( j, ; , M | JM) are SU(2)

Clebsch-Gordan coefficients, form an orthonormal basis in KnJ and satisfy

 j (0, 0, 0, )
U  nJM = D J ( )
 nJM , M = J, J + 1, . . . , J 1, J. (8.159)

The sum over in (8.158) runs through a subset of | j |, | j | + 1, . . . , j + ,

determined by the constraints j j and  M . Let V J denote the
UIR of the subgroup H = SU(2) of G :

V J ( , ) = ei D J ( ). (8.160)

 KnJ
Then (8.159) implies that for any vector

 j ( , 0, 0, )
U  = V J ( , )
. (8.161)

This also means that the projection operator,

J
PnJ =  nJM 
|  nJM |, (8.162)
M=J

for the subspace KnJ is a minimal projector satisfying

U  j (0, 0, 0, ) = PnJ ,
 j (0, 0, 0, )PnJU (8.163)

for all SU(2) [minimality implying that there is no projector P PnJ also
satisfying (8.163)]. Furthermore, a straightforward computation using properties of
spherical harmonics shows that

J 
 
 nJM ) =
cV J ( nJM
|  nJM 2
(q,p) | (q,p) | dq dp
M=J
8.4 A Generalization: - and V -Admissibility 241

J
= (2 )3 |C( j, ; , M | JM)|2
M=J

|C( j, ; , M  | JM  )|2 . (8.164)

Thus, the subspace KnJ of K j L2 (R  3 , dk) is V J -admissible, and again the

representation is square integrable mod( SU(2)) , since the above integral is
clearly independent of the section . This also implies an interesting relation
among Clebsch-Gordan coefficients, namely, that the sum on the RHS of (8.164)
is independent of M  , and indeed, using the fact that the vectors  nJM , M =
J, J + 1, . . . , J 1, J, form an orthonormal basis for K , it is not hard to show
nJ

[by computing the integral in (8.164) directly] that


 nJM ) = cV J (
cV J (  ) = (2 )3 (2J + 1), (8.165)

 KnJ .
for any vector
Consequently, for each n = 1, 2, . . ., and each J = | j |, | j | + 1, . . . , j + , the
 3 , dk) KJ L2 ( , dq dp):
map WnJ : K j L2 (R

(WnJ )M (q, p) = [(2 )3 (2J + 1)]1/2 

nJM 
(q,p) | ,

M = J, J + 1, . . . , J 1, J, (8.166)

 3 , dk) the resolution of the identity,

is an isometry. On K j L2 (R

J 
1

(2 )3 (2J + 1) M=J
nJM
|  nJM
(q,p)  (q,p) | dq dp = I, (8.167)

is valid for the class of Galilei coherent states,

[(2 )3 (2J + 1)]1/2

nJM
(q,p) , (q, p) R6 , M = J, J + 1, . . . , J 1, J.
(8.168)
Explicitly, the CS are the functions

1
nJM
(q,p) (k) =
[(2 )3 (2J + 1)]1/2
1 ! "
= U j (0, q, p , I2  nJM )(k)
[(2 )3 (2J + 1)]1/2 m
1
= eikq
 nJM (k p)
[(2 )3 (2J + 1)]1/2
242 8 Coherent States from Square Integrable Representations

1
= eikq C( j, ; , M | JM) e j
[(2 )3 (2J + 1)]1/2

M
Y (k p) n (k p), (8.169)

(M = J, J + 1, . . . , J 1, J), which in general are vector coherent states. The

 3 , dk) in KJ L2 ( , dq dp), under the isometry WnJ , is a
image of K j L2 (R
reproducing kernel Hilbert space HnJ . Its elements are vector-valued functions
with components,

1
M (q, p) =
[(2 )3 (2J + 1)]1/2
C( j, ; , M | JM)


M
eikq Y (k p) n (k p) (k) dk,
3
R
M = J, J + 1, . . . , J 1, J. (8.170)

The reproducing kernel for HnJ is matrix-valued and has the components

[KnJ (q, p; q p )]MM =

1
=
(2 )3 (2J + 1) C( j, ; , M | JM)C( j, ; , M | JM )


M M 
exp [ik (q q )] Y (k p)Y (k p )
R3

n (k p) n (k p ) dk, (8.171)

 j in HnJ KJ L2 ( , dq dp)
M, M  = J, J + 1, . . . , J. Finally, the image of U
under WnJ is
 j (g)W 1 ,
UnJ (g) = WnJU nJ g G , (8.172)

(UnJ ( , a, v, ) )(q, p) = exp [i{ + mv (q a)}]

D J ( ) (R( 1 )(q a), R( 1 )(p mv)), (8.173)

which again extends to the whole of KJ L2 ( , dq dp) as the representation of G

induced from V J .
In conclusion observe that since the isochronous Galilei group G is essentially
the Weyl-Heisenberg group for three degrees of freedom together with rotations, the
coherent states (8.168) are also the VCS of this latter group.
8.4 A Generalization: - and V -Admissibility 243

8.4.3 Atomic Coherent States

We remark that the coherent states, (8.168)(8.169) obtained above, are all possible
such states arising from the (extended) Galilei group, which are labelled by phase
space points, which satisfy a resolution of the identity and incorporate spin degrees
of freedom. Let denote the Fourier transformed function corresponding to any
one of the vectors nJM
(q,p) . Then is a quantum mechanical wave function in
configuration space. Its time evolved version,

(x,t) = (eiHt )(x), (8.174)

for some quantum mechanical Hamiltonian H, obeys the Schrdinger equation.

The fiducial vectors  nJM , M = J, J 1, . . . , J, from which the coherent
states (8.168)(8.169) are built, can be looked upon as quantum mechanical wave
functions for a system with spin j, orbital angular momentum  and total angular
momentum J. Thus the associated coherent states (8.168)(8.169) can be called
atomic (or molecular) coherent states.
Actually, the resolution of the identity (8.167) is not so surprising and it is easy
to check it by direct computation. In fact, a much more general expression can be
obtained as follows: Let  i , i = 1, 2, . . . , n, be any set of orthonormal vectors in K j
L (R
2  , dk), spanning any n-dimensional subspace K having projection operator PK .
3

Then by direct computation it can be seen that


1  j ( (q, p)) ,
 j ( (q, p)) PK U
PK (q, p) dq dp = I, PK (q, p) = U
(2 )3 n
(8.175)
for any one of the following sections [obtained from the basic Sect. (8.141)]:
! p "
(q, p) = 0, q, , I2 ( (q, p), 0, 0, ). (8.176)
m

Here SU(2) is fixed and (q, p) is an arbitrary, real valued, measurable

function of (q, p). Taking the i to be the n eigenstates of some observable (e.g.,
the Hamiltonian of some atomic system in an external potential), the general
coherent states [(2 )3 n]1/2
i (q,p) , i = 1, 2, . . . , n, (q, p) , can be used in
atomic computations involving such n-level systems. However, these CS would
not, in general, display the covariance property (8.110). Even more generally, if
F is any trace-class operator on K j L2 (R  3 , dk), then with as in (8.176) and
c(F) = (2 ) [Tr F],
3


1  j ( (q, p))F U
 j ( (q, p)) . (8.177)
F(q, p) dq dp = I, F(q, p) = U
c(F)
Chapter 9
CS of General Semidirect Product Groups

Abstract In this chapter we examine general semidirect product groups, with

special emphasis on their geometrical structure, and the construction of their CS.
Examples include squeezed states, the Euclidean groups and affine sections in the
general case.

In Sect. 8.3, we studied a class of semidirect product groups, the regular representa-
tions of which consisted entirely of a discrete sum of irreducible subrepresentations,
all square integrable. These groups were of the general form G = Rn  H, where
H was an n-dimensional subgroup of GL(n, R) and its action on the dual space
 n gave rise to open free orbits. In this chapter, we generalize this setting and
R
consider semidirect products of the type G = V  S, where V is an n-dimensional
real vector space (n is assumed to be finite) and S is usually a subgroup of GL(V )
(the group of all nonsingular linear transformations of V ). We shall again examine
the action of S on the dual vector space V , but now without restriction on the
orbits, in particular, the dimensions of these orbits could be lower than n, the
dimension of the vector space V on which S acts. The analysis, however, will have
features similar to those encountered in Sect. 8.3, although we shall no longer look
for the sort of simple square integrability with respect to the entire group as was
done there. An interesting interplay between the geometry of the orbits and the
existence of CS, square integrable over the associated homogeneous spaces, will
be uncovered, which will turn out to be intimately related to the theory of group
representations emanating from coadjoint orbits [Kir76]. Our exposition follows
[16, 2325, 28, 237], and the main result on square integrability is formulated in
Theorem 9.3.4.
In the spirit of Sect. 8.3, however, we shall also look at a somewhat more
general result for the square integrability of a group representation, which does not
necessarily require the presence of open free orbits. This result is stated (without
proof) in Theorem 9.3.1, leading to the following hierarchy of generalizations:
Theorem 8.3.2 proves the square integrability of the quasi-regular representation

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 245
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__9,
Springer Science+Business Media New York 2014
246 9 CS of General Semidirect Product Groups

of G, when open free orbits are present; Theorem 8.3.3 generalizes this result
to the case in which the orbit is open but not necessarily free; however, the
stabilizing subgroup is compact; Theorem 9.3.1 generalizes further to the case of
an arbitrary induced representation of G, under the restriction that the orbit have
positive Lebesgue measure in V and that the representation of the subgroup of S0
of S, from which the representation of G is induced, be square integrable; finally,
Theorem 9.3.4 obtains conditions for the square integrability of an arbitrary induced
representation of G, assuming only that the representation of the subgroup of S0 be
finite-dimensional.
We begin with a discussion of squeezed states as an illustration of the use of
semidirect product groups for constructing CS.

9.1 Squeezed States

Squeezed states were introduced in (2.7) of Chap. 2 (see the discussion following
that equation). We study here the origin of these states in a certain representation
of the metaplectic group. The discussion is based on [557]. Recall that the general
gaussons, or Gaussian pure states were defined in Eq. (2.7), as

q 
1 
q,p
U,V
(x) = n/4 [detU]1/4 exp i(x ) p exp (x q) (U + iV )(x q) ,
2 2
(9.1)

where q, p Rn , U is a (strictly) positive definite n n real, symmetric matrix and

V is an arbitrary n n real, symmetric matrix. In the case in which U is the identity
matrix and V = 0, these states reduce to the canonical coherent states (for n degrees
of freedom), while if V = 0 and U is not the identity matrix, we get squeezed states.
We begin by introducing the notation,

X = (q, p) R2n , Z = V iU Mn (C) Cnn , (9.2)

to rewrite the q,p

U,V
as (X,Z ) , in which the will presently be identified with a
section of the metaplectic group. Note that, for each fixed Z , we have a resolution
of the identity on L2 (Rn , dx):

1
| (X,Z )  (X ,Z ) | dX = I, dX = dq dp. (9.3)
(2 )n R6

The metaplectic group is the semidirect product, Mp(2n, R) = GWH (n) 

Sp(2n, R), of the WeylHeisenberg group GWH (n) for n degrees of freedom and the
real symplectic group Sp(2n, R) of R2n . Elements g Mp(2n, R) are of the form

g = ( , X , M), where R, ( , X ) GWH (n), M Sp(2n, R). (9.4)

9.1 Squeezed States 247

Recall that the product rule in GWH is

(1 , X 1 )(2 , X 2 ) = (1 + 2 + (X 1 , X 2 ), X 1 + X 2 ), (9.5)

where, of course, X i = (qi , pi ), i = 1, 2, and

1
(X 1 , X 2 ) = (p1 q2 p2 q1 ). (9.6)
2
The elements M Sp(2n, R) are 2n 2n real matrices satisfying
 
0 In
M M = , det M = 1, where =
T
(9.7)
In 0

(In = n n identity matrix). The product rule of Mp(2n, R) is, thus,

(1 , X 1 , M1 )(2 , X 2 , M2 ) = (1 + 2 + (X 1 , M1 X 2 ), X 1 + M1 X 2 , M1 M2 ). (9.8)

The n n complex matrices Z = V iU, where V = V T and U = U T , U > 0, are

n n real matrices, define a certain coset space of Sp(2n, R), as we shall now see.
Any element M Sp(2n, R) can be written in block form as
 
AB
M= , with ADT BCT = In , ABT = BAT , CDT = DCT , (9.9)
CD
where A, B,C, and D are n n real matrices. The maximal compact subgroup of
Sp(2n, R) is Sp(2n, R) SO(2n), a group that is isomorphic to U(n) (the group of
all n n complex unitary matrices of unit determinant). This means that a general
M Sp(2n, R) has the polar decomposition:

M = p(M)1/2 k, p(M) = MM T , k Sp(2n, R) SO(2n) U(n). (9.10)

The matrix p(M)1/2 is the unique positive square root of the positive-definite matrix
p(M). Moreover, as shown in [557], p(M) can be conveniently written in the form

p(M) = M(U,V )M(U,V )T ,

   1/2 
In 0 U 0
where M(U,V ) = , (9.11)
V In 0 U 1/2
in which, once again, U is a (strictly) positive-definite, n n real symmetric matrix
and V is an n n real symmetric matrix. Using the complex notation introduced
above, we see that the coset space Sp(2n, R)/[Sp(2n, R) SO(2n)] can be identified
with the complex tubular domain T of matrices Z = V iU (with V a real
symmetric and U a positive-definite, symmetric, n n matrix). In view of (9.11),
we shall write p(M) p(Z ). Then, for arbitrary Z T and M Sp(2n, R), if we
write

p(Z  ) = M p(Z ) M T , (9.12)

248 9 CS of General Semidirect Product Groups

a straightforward computation yields the transformation law

Z  = M[Z ] = (DZ C)(A BZ )1 , (9.13)

for the action of M Sp(2n, R) on T . Furthermore, it is clear that

Z = M[Z0 ], where Z0 = iIn , M = p(Z )1/2 k, (9.14)

meaning that T is the orbit of Z0 = iIn under the action (9.13) of Sp(2n, R). It
also means that the elements of the complex matrices Z provide a global coordi-
natization for the homogeneous space Sp(2n, R)/[Sp(2n, R) SO(2n)], which, as
a geometrical object, is therefore a complex (Khler) manifold, identified with the
unbounded version of a classical domain of type RII in the Cartan classification
[Hua63].
The Lie algebra sp(2n, R) of Sp(2n, R) consists of all 2n 2n real matrices X
with the property that X is a symmetric matrix. Thus, any X sp(2n, R) has the
block form,
 

X= , = T, = T, (9.15)
T
in terms of three n n real matrices, , , and . If = T and = , in the
above, the corresponding X lies in the Lie algebra of the maximal compact subgroup
Sp(2n, R) SO(2n). Clearly, is such an element of the Lie algebra. Consider now
the adjoint action of M Sp(2n, R) on , as defined in (4.137):

AdM ( ) = M M 1 . (9.16)

If the pairing between sp(2n, R) and its dual sp(2n, R) is given via the trace
operation, may itself be identified with an element of the dual space sp(2n, R) ,
and the coadjoint action of Sp(2n, R) on is then identifiable with:
1
Ad#M ( ) = M T MT. (9.17)

It is easily verified, using the decomposition (9.10), that the stability subgroup of
under both the coadjoint and adjoint actions is precisely the maximal compact
subgroup Sp(2n, R) SO(2n). Thus, the complex domain T is in fact the orbit of
under the coadjoint (or adjoint) action.
In order to obtain the squeezed states q,p U,V
in (9.1), it will be necessary to
construct an appropriate representation of the metaplectic group Mp(2n, R). We do
this by starting with a representation of the WeylHeisenberg group GWH (n),
obtained by generalizing to n degrees of freedom the representation introduced
in (2.33) for one degree of freedom. This extended UIR, denoted again by U ,
R, = 0, acts on the Hilbert space L2 (Rn , dx) in the manner
q
(U ( , X ) )(x) = ei ei p(x 2 ) (xq), X =(q, p), L2 (Rn , dx).
(9.18)
9.1 Squeezed States 249

While, for different values of , the representations U are unitarily inequivalent,

for our purposes, it will be enough to choose = 1. We shall do so in the following
and write the representation simply as U. Next, we note that the representation U
can actually be extended to a UIR of the entire metaplectic group. Indeed, let us
define the 2n-component vector operator X on L2 (Rn , dx), in terms of the position
and momentum operators Qi , Pi ,

X = (Q1 , Q2 , . . . Qn , P1 , P2 , . . . , Pn ),

(Qi )(x) = xi (x), (Pi )(x) = i (x). (9.19)
xi

In terms of the operator X, the canonical commutation relations [Qk , P ] = iI k are

conveniently rewritten as

[Xk , X ] = iI( )k , k,  = 1, 2, . . . , 2n. (9.20)

If M is a 2n 2n real matrix and X = MX, then the components of X satisfy

the CCR (9.20) if and only if M Sp(2n, R), as expected from a symplectic
action. Moreover, for each M Sp(2n, R), there exists a unitary operator U (M)
on L2 (Rn , dx), such that M  U (M) is a representation of Sp(2n, R), which acts
on the vector X in the manner,

U (M)X U (M) = M 1 X. (9.21)

Consequently, in this representation of Sp(2n, R), the elements X of the Lie algebra
sp(2n, R) [see (9.15)] are realized as the operators

X = 1/2X XX, (9.22)

on L2 (Rn , dx), which means that, if M(t) = exp[tX] is the one-parameter subgroup
of Sp(2n, R) generated by X, then U (M(t)) = exp[iXt].  Furthermore, for all
( , X ) GWH (n) and M Sp(2n, R), the relations

U( , X )XU( , X ) = X X , U (M)U( , X ) U (M) = U( , M X ), (9.23)

 one also has the equivalent

are easily seen to hold. At the level of the generators X,
relations

 = iXX,
[X, X] [X, 
 X ] = i[X, X  ]. (9.24)

The relations (9.20)(9.23) together [or, alternatively, the relations (9.24)] imply
that the operators

U( , X , M) = U( , X ) U (M), ( , X , M) Mp(2n, R), (9.25)

250 9 CS of General Semidirect Product Groups

define a UIR of Mp(2n, R) on L2 (Rn , dx). It is difficult to write explicitly the

action of a general representation operator U( , X , M) on an arbitrary vector

L2 (Rn , dx). Identifying the coset Mp(2n, R)/[ (Sp(2n, R) SO(2n))], however,
( being the phase subgroup of GWH ) with R2n T , let us define the section

: R2n T Mp(2n, R), (X , Z ) = (0, X , M(U,V )),

X = (q, p), Z = V iU, (9.26)

M(U,V ) being as in (9.11). It is then straightforward to verify that

(X ,Z ) = XU,V
,Z = U( (X , Z )) , where (x) = n/4 exp [1/2x x].
(9.27)

Thus, finally, the general squeezed states or gaussons are all obtained from the
ground state wave function of a system of n harmonic oscillators using a section
in the metaplectic group. Moreover, writing

P = |  |, and P(X , Z ) = | (X ,Z )  (X ,Z ) |, (9.28)

we find that
P(X , Z ) = U( , X , M) P U( , X , M) (9.29)
[assuming the decomposition (9.10) and (9.11) of M Sp(2n, R)], showing that the
corresponding projection operators form an orbit of the metaplectic group.

9.2 Geometry of Semidirect Product Groups

For a general analysis of CS of semidirect product groups, it will now be necessary

to examine some detailed features of the group G = V  S and of the associated
geometry of orbits. Much of the discussion in the next three subsections follows
[Gui84] and [237]. Unfortunately the discussion is a bit technical and a proliferation
of symbols and notation will occur in the course of it. In order to make the reading
easier, we have drawn up a table of the more important quantities at the end of
Sect. 9.2.3 (Table 9.1).

9.2.1 A Special Class of Orbits

A general group element in G = V  S will be written as g = (x, s), with x V and

s S; the group multiplication being given by (x1 , s1 )(x2 , s2 ) = (x1 + s1 x2 , s1 s2 ), in
which the action of S on V is indicated by x  sx. If k V , the dual of V , and < ; >
denotes the dual pairing between V and V , we shall indicate by k  sk the action
of S on V , defined in the manner
9.2 Geometry of Semidirect Product Groups 251

< sk ; x > = < k ; s1 x >, (9.30)

which we shall later identify with the coadjoint action.

Let k0 V be a fixed vector and O be its orbit under S. Recall that O then
consists of all elements k V that are of the type k = sk0 , s S. Let S0 be the
stability subgroup of k0 ; i.e., s S0 iff sk0 = k0 . Then, S/S0 O via the map
gS0  gk0 . Moreover, as a geometrical object, O is a smooth manifold, and, at
any point k O , we may consider the tangent space Tk O . If we assume that the
dimension of O , as a real manifold, is m ( n, the dimension of V ), then Tk O is a
real vector space, also of dimension m. Moreover, Tk O can be identified in a natural
way with a subspace of V . Indeed, let {e i }ni=1 be a basis of V and suppose that
k(t) = ni=1 ki (t)e i is a smooth curve in O , passing through k = k(0) and defined
for values of t lying in some open set ( , ) R. The components ki (t) are some
smooth functions of t. Then, differentiating with respect to t,
n
k(0) = ki (0)e i V
i=1

is a vector tangent to O at k. We shall always consider Tk O as being embedded in


this manner: Tk O V . The space T O = kO Tk O of all tangent vectors at all
points k is again a smooth manifold, of dimension 2m, called the tangent bundle of
O . The cotangent space Tk O at the point k O is the dual space of Tk O , which
we therefore identify with a subspace of V , i.e., Tk O V . Correspondingly we

have the cotangent bundle T O = kO Tk O . As a manifold, T O has certain
properties that make it resemble a classical phase space. It is a symplectic manifold,
which means that it comes equipped with a nondegenerate, closed two-forma
point we shall return to a little later.
Along with the two vector bundles T O and T O , there is a second pair of
vector bundles that appear here simultaneously and that we also need to consider.
For the tangent space Tk O , let Nk be its annihilator in V , i.e.,

Nk = {x V : < p ; x > = 0, p Tk O }. (9.31)

Then,

Nk Tk O = V, (9.32)

and N = kO Nk is a manifold, referred to as the normal bundle. Analogously we
define

Nk = {p V : < p ; x > = 0, x Tk O }, (9.33)

so that, again,

Nk Tk O = V , (9.34)

with Nk being identifiable with the dual of Nk and N =
kO Nk with the dual
bundle of N.
252 9 CS of General Semidirect Product Groups

Consider again the coadjoint action [see (9.30)] of S on V , under which k 

k = sk. The derivative of this map at the point k V , which we denote by Dk (s),
is a linear map between the two tangent spaces TkV and Tk V . The restriction
of the map k  sk to the set of points k O , which we denote by s|O , leaves O
invariant, and thus the restriction of the derivative Dk (s) to Tk O is exactly Dk (s|O ),
the derivative of the restricted map s|O at k, and this is a linear map between Tk O
and Tk O . In particular, since, for s S0 , sk0 = k0 , its derivative Dk0 (s) at k0 maps
Tk0 V to itself and, consequently, Dk0 (s|O ), for s S0 , maps Tk0 O to itself. Since
all tangent spaces TkV can be naturally identified with V , we identify the action
of Dk (s) on TkV with that of s itself, and it then follows that, for s S0 and p
Tk0 O , sp Tk0 O . (One can think of s as a matrix acting on the vector p. Then
the derivative of this map, with respect to p, is simply s itself.) By duality, we then
immediately obtain that S0 leaves Nk0 invariant. Thus, N0  S0 is a subgroup of G
(for simplicity, we write N0 = Nk0 ). Let denote the resulting left coset space,

= G/H0 , H0 = N0  S0 . (9.35)

Also, set

V0 = Tk0 O , V0 = Tk0 O , (9.36)

so that, by (9.32) and (9.34), N0 V0 = V and N0 V0 = V . Since any x V can be

written uniquely as x = n + v, with n N0 and v V0 , it would seem plausible that
should be isomorphic, as a Borel space, to V0 O . In fact, as we explain below,
is isomorphic to T O as a symplectic manifold and to V0 O as a Borel space.

9.2.2 The Coadjoint Orbit Structure of

We now present a reasonably self-contained and not-too-technical description of the

coset space in (9.35) as a coadjoint orbit of G and as a symplectic manifold, in
a sense that we make precise. While no rigorous differential geometric proofs are
given, the discussion can in fact be developed into exact proofs without too much
difficulty. A more rigorous treatment may, for instance, be found in [Gui84]. Let us
rewrite the elements g = (x, s) of G in matrix form,
 
sx
g= , (9.37)
01
with group multiplication being replicated by matrix multiplication. Denoting by
g, v, and s the Lie algebras of G,V , and S, respectively, an element X = (a, J) g,
with a v and J s, then appears as the matrix
 
Ja
X= . (9.38)
00
9.2 Geometry of Semidirect Product Groups 253

Note that J can be identified with a linear transformation of V , and v with V itself
in an obvious fashion. The adjoint action of G on g is then defined by
 
sJs1 sa sJs1 x
Adg (X) = gXg1 = . (9.39)
0 0

The resultant action on the dual space g of the Lie algebra g is now easily obtained.
Indeed, let v , s be the dual spaces of v and s, respectively, and let X = (a , J )
g , with a v and J s . Also, the dual pairing in (g , g) has the form

< X ; X >g ,g = < a ; a >v ,v + < J ; J >s ,s . (9.40)

We now define the coadjoint action of G on g :

< Ad#g (X ) ; X >g ,g = < X ; Adg1 (X) >g ,g . (9.41)

  
Writing Ad#g (X ) = X = (a , J ), using (9.39) and noting that v can be identified
with V , we easily derive,

a = sa

(9.42)
J = Ad#S
s (J ) + (sa ) ) x,

where the first equation reproduces the coadjoint action defined in (9.30),

< sa ; a >v ,v = < a ; s1 a >v ,v , a v,

while
1
s (J ) ; J >s ,s = < J ; s Js >s ,s ,
< Ad#S J s, (9.43)

and (sa ) ) x denotes the element in s for which

< (sa ) ) x ; J >s ,s = < sa ; Jx >v ,v , J s. (9.44)

Let us examine the orbit of X0 = (k0 , 0) g under the coadjoint action.

From (9.42)(9.44),

Ad#g (k0 , 0) = (sk0 , sk0 ) x) = (k, k ) x), g = (x, s). (9.45)

Let d be the dimension of the Lie group S, considered as a real manifold. Since
dim O = m, the dimension of the subgroup S0 is d m. Choose a basis {Ji }di=1
for the Lie algebra s, in such a way that Jm+1 , Jm+2 , . . . , Jd is a basis for s0 , the
Lie algebra of S0 . Recall that each Ji can be identified with a linear map from V to
254 9 CS of General Semidirect Product Groups

itself. Denote by Ji the dual map on V (now identified with v ): < Ji k ; x > =
< k ; Ji x >, for all x V, k V . Then, since S0 is the stability subgroup of k0 ,

Ji k0 = 0, i = 1, 2, . . . , m,
(9.46)
Ji k0 = 0, i = m + 1, m + 2, . . . , d.

Moreover, Ji k0 , i = 1, 2, . . . , m, is a basis for the tangent space Tk0 O = V0 . Since

sk0 = k implies that s(Tk0 O ) = Tk O , it follows that sJi s1 k, i = 1, 2, . . . , m, is a
basis for Tk O .
Consider now the element k0 ) x s . By (9.44),

< k0 ) x ; J >s ,s = < k0 ; Jx >v ,v ,

and, in particular, for all Ji , i = m + 1, m + 2, . . . , M,

< k0 ) x ; Ji >s ,s = < Ji k0 ; x >v ,v = 0, x V.

On the other hand, for Ji = 1, 2, . . . , m, and x N0 (the annihilator of Tk0 O ),

< k0 ) x ; Ji >s ,s = < Ji k0 ; x >v ,v = 0.

Thus, writing x = nk0 + vk0 , with nk0 N0 = Nk0 and vk0 V0 = Tk0 O , we get

k0 ) x = k0 ) vk0 ,

which means that each k0 ) x s can be identified with an element vk0 Tk0 O .
Conversely, given any vk0 Tk0 O , we can define the element k0 ) vk0 s
corresponding to it. Generally, using the decomposition

x = nk + vk , x V, nk Nk , vk Tk O , (9.47)

we establish the correspondence

(k, k ) x)  (k, vk ), (9.48)

between elements in the orbit O(k0 ,0) of (k0 , 0) g , under the coadjoint action of G,
and elements in the cotangent bundle T O . Moreover, it is clear from (9.45) that
H0 = N0  S0 is the stability subgroup of (k0 , 0) g . Hence, finally,

= G/H0 {Ad#g (k0 , 0) : g G} = O(k0 ,0) T O . (9.49)

The action of G on T O is readily computed. On V , corresponding to the

decomposition V = Nk Tk O [see (9.32)], let us introduce the two projection
operators, Pnk and Pvk , such that, for x V and x = nk + vk [see (9.47)],

Pnk x = nk , Pvk x = vk . (9.50)

9.2 Geometry of Semidirect Product Groups 255

Next, writing

Ad#(x,s) (k, k ) vk ) = (k , k ) wk ), (9.51)

and applying (9.42), we get

k = sk, wk = Pvk x + svk Tsk O . (9.52)

(Note that Pvk svk = svk .) Thus, under the action of g = (x, s) G, an element (k, vk )
T O transforms to

g(k, vk ) = (sk, Pvsk x + svk ). (9.53)

To make the correspondence between T O and G/H0 more explicit, recall first
that, as manifolds, O S/S0 . Let : O S be a global Borel section, such that

(k0 ) = e = identity element of S,

(k)k0 = k, k O , (9.54)

and which is a smooth map on some open dense set in O . Then, any element s S
can be uniquely written as

s = (k)s0 , k O , s0 S0 . (9.55)

This section will be fixed once and for all. Although the specific families of coherent
states we shall obtain using it will depend on the particular choice of this section, the
results on square integrability, such as Theorems 9.3.4 and 9.3.6, will not, however.
If vk Tk O , then (k)1 vk Tk0 O = V0 . For any (x, s) G, let x = nk + vk be the
decomposition (9.47) and s = (k)s0 . Then, the coset (x, s)H0 can conveniently be
represented by the element (vk , (k)) G, following the coset decomposition,

(x, s) = (vk , (k))(n0 , s0 ), n0 = (k)1 nk N0 . (9.56)

The transformation property of an arbitrary (v p , (p)) under the action of g = (x, s)

is then obtained as usual by computing (x, s)(v p , (p)) and decomposing as above.
We get

(x, s) : (v p , (p))  (Pvsp x + sv p , (sp)), (9.57)

which is the same transformation rule as for (p, v p ) Tp O obtained in (9.53). Thus,
an element (p, v p ) Tp O corresponds to the coset (v p , (p))H0 G/H0 and vice
versa.
The following section : G/H0 G will be useful for the construction of CS
later:

((v p , (p))H0 ) = (v p , (p)). (9.58)

256 9 CS of General Semidirect Product Groups

Since any v p Tp O can be written as v p = (p)q, for some q V0 , we shall

actually write the above section as the mapping,

: V0 O G, (q, p) = ( (p)q, (p)). (9.59)

We shall refer to as the principal section.

9.2.3 Measures on

Since the space T O is a cotangent bundle, it is a symplectic manifold [Gui84] and

consequently comes equipped with a nondegenerate two-form, which is invariant
under the action of G and gives rise to an invariant measure on it. Alternatively, since
T O is a coadjoint orbit, the discussion in Sect. 4.5.2 [see, in particular (4.142)],
also tells us that it carries an invariant measure. By virtue of the isomorphisms
displayed in (9.49), these properties are also inherited by the coset space . Locally,
the two-form on T O may be constructed as follows. Let p1 , p2 , . . . , pm be a set of
local coordinates, defined on an open set U O . Then, there is a smooth map
: U Rm , invertible on its range, with smooth inverse, and such that pi = (p)i
is the ith component of (p) in Rm , for p U. The corresponding tangent vectors,
{ pi }m
i=1 , span the tangent space Tp O . Let { dpi }i=1 be the dual basis in Tp O , for
m

each p U. Then, a general cotangent vector v p Tp O can be written as

m
v p = vip dpi , vip R.
i=1

The canonical two-form , expressed locally on U in terms of these coordinates,

is now
m
= dvip dpi , (9.60)
i=1

which is easily checked to be invariant, closed, and nondegenerate. The associated

left-invariant measure d on T O is then, locally,

d = dv1p dv2p . . . dvmp dp1 dp2 . . . dpm . (9.61)

Again, the invariance of this measure is easily established using the fact that the dvip
and dpi transform contragrediently under S. We recognize dv1p dv2p . . . dvmp
to be simply a version of the Lebesgue measure dv p on Tp O , expressed in these
coordinates.
For our purposes, it will be useful to find a convenient Borel isomorphism
between T O and V0 O [V0 = Tk0 O , see (9.36)], and, eventually, to work with
9.2 Geometry of Semidirect Product Groups 257

V0 O as the underlying space for labeling coherent states, rather than with T O .
Recall that, as a manifold, O is isomorphic to the coset space S/S0 , and hence
the map

c : T O V0 O , c(v p , p) = ( (p)1 v p , p) := (q, p), (9.62)

where p O and v p = (p)q Tp O is a Borel isomorphism. Denoting the

Lebesgue measure on V0 by dq, we may then write,

dv p = f (p) dq, (9.63)

where

r(p)
f (p) = (9.64)
|det[ (p)1 |Tp O ]|

(p)1 |Tp O denoting the restriction of (p)1 to Tp O and r(p) being a measur-
able function on O , which is positive and nonzero on U, and can in fact be chosen
to be smooth on it.
At this point we make the further assumption that O carries an invariant
measure d (under the action of S), which on U can be written in the form

d (p) = m(p) dp1 dp2 . . . dpm , (9.65)

where again m(p) is a measurable function (which also can be chosen to be smooth)
on O , which is positive and nonzero on the open set U, used to define the local
coordinates pi . It follows then that under the Borel isomorphism c, the invariant
measure d on T O transforms locally on U to the measure
f (p)
d (q, p) = dq d (p). (9.66)
m(p)
Hence, globally on V0 O , we may write

d (q, p) = (p) dq d (p), (9.67)

where now (p) is a measurable function that is positive and nonzero almost
everywhere on O (with respect to d ). Henceforth, we shall identify the coset
space = G/H0 , H0 = N0  S0 , or the cotangent bundle T O , as Borel spaces,
with V0 O , equipped with the invariant measure d in (9.67). The invariance of
d is with respect to the group transformation (9.52), transported to V0 O via the
isomorphism (9.62). Thus, under the action of the group element g = (x, s), (q, p) 
(q , p ) = g(q, p), with

q = (sp)1 Pvsp x + h0 (s; p)q, p = sp, (9.68)

258 9 CS of General Semidirect Product Groups

as follows from (9.57) and (9.62), where h0 (s; p) = (sp)1 s (p) S0 . Note also,
that in view of this transformation property, the section introduced in (9.59)
transforms as:

g(q, p) = (g(q, p)) ( (sp)1 Pnsp x, h0 (s, p)). (9.69)

The following set of local coordinates for O will later turn out to be useful for
computational purposes. Let {ei }ni=1 be a basis of V , such that e1 , e2 , . . . , em is a basis
for V0 . Let {e i }ni=1 be the dual basis for V , i.e.,

e i ; e j  = i j , i, j = 1, 2, . . . , n,

and e 1 , e 2 , . . . , e m span V0 . Then, since V0 is the tangent space to O at the point

k0 , it is always possible to find an open set O(k0 ) O , containing k0 , such that the
map, : O(k0 ) Rm , with components

(k)i := ki = k k0 ; ei , i = 1, 2, . . . , m, (9.70)

is a diffeomorphism. In this basis, an arbitrary k O(k0 ) has the expansion

m
k = (ki + i )e i + n , i = k0 ; ei , i = 1, 2, . . . , m, (9.71)
i=1

where n N0 and is completely determined by the ki . If q V0 and q = m i

i=1 q ei ,
then
m
k ; q = (ki + i )qi , (9.72)
i=1

a relation we shall need later.

As an aid to further reading, we have collected together in Table 9.1 some of the
special symbols, defined quantities, and notations introduced in the last few pages.

9.2.4 Induced Representations of Semidirect Products

For semidirect product groups of the type we are considering here, all irreducible
representations arise as induced representations and correspond to orbits O in V
[Mac68]. Let us work out these representations in some detail, for we shall have
to rely on their specific features to derive a condition for the existence of square
integrable coherent states. It will turn out that the CS will be labeled by points in
= G/H0 [see (9.35)], or equivalently, by points in V0 O (recall that V0 = Tk0 O ).
9.2 Geometry of Semidirect Product Groups 259

Table 9.1 Main notions and notations used in this chapter

Symbol Definition Relation
V Vector space, abelian group of dimension n
V Dual space of V
ei (e i ) Basis of V (V )
S Subgroup of GL(V )
S0 Stabilizer of k0 V S0 S
G Semidirect product group G =V S
O Orbit of k0 in V under S, of dimension m O S/S0
Global Borel section, : O S s = (k)s0 , s S, s0 S0
g, v, s Lie algebra of G, V , S X = (a, J) g, a v, J s
g , v , s Dual spaces of g, v, s X = (a , J ) g
O(k0 ) Open set around k0 , for local coordinates O(k0 ) O
Coordinate chart, : O(k0 ) Rm (k)i = ki = k k0 ; ei 
d (k) Invariant measure on O d (k) = m(k) m i=1 dki
Tk O Tangent space of O at k, V0 = Tk0 O Tk O V
Tk O Cotangent space of O at k, V0 = Tk0 O Tk O V
Nk Annihilator of Tk O in V, N0 = Nk0 V = Nk Tk O
Nk Annihilator of Tk O in V , N0 = Nk0 V = Nk Tk O
Pvk (Pnk ) Projection operator from V to Tk O (Nk )
H0 Stabilizer of (k0 , 0) g under coad-action H0 = N0  S0
T O Cotangent bundle of O T O G/H0
Borel
Parameter space for labeling CS = V0 O T O
d (q, p) Invariant measure on d (q, p) = (p) dq d (p)
Principal section, : G (q, p) = ( (p)q, (p))

These CS will be square integrable and the associated induced representations

square integrable mod(H0 , ) for appropriate sections .
Consider again the element k0 V , of which O is the orbit under S. The
associated unitary character of the abelian subgroup V ,

(x) = exp[i < k0 ; x >], x V, (9.73)

defines a one-dimensional representation of V . Let s  L(s) be a unitary irreducible

representation of S0 , the stability subgroup of k0 , and carried by some Hilbert
space K. Consider now the UIR, L, of V  S0 carried by K:

( L)(x, s) = exp[i < k0 ; x >] L(s). (9.74)

We need the representation of G = V  S, which is induced from L. Clearly,

G/(V  S0 ) O , (9.75)

and we shall need the section,

: O G, (k) = (0, (k)), (9.76)

260 9 CS of General Semidirect Product Groups

with as defined in (9.54), which corresponds to the coset decomposition

(x, s) = (0, (k)) ( (k)1 x, s0 ), (x, s) G

[see (9.54)]. Also, since for an arbitrary (x, s) G and p O ,

(x, s)(0, (p)) = (0, (sp)) ( (sp)1 x, (sp)1 s (p)), (9.77)

the action of G on O is also given by k  (x, s)k = sk [see (9.30)]. Recall that we are
assuming the measure d on O to be invariant under this action. (This will be the
case if, for example, S0 is a compact subgroup of S.) Denote by h : G O V  S0
and h0 : S O S0 the two cocycles appearing in (9.77):

h((x, s), p) = ( (sp)1 x, h0 (s, p)),

h0 (s, p) = (sp)1 s (p). (9.78)

Then a straightforward computation yields

( L)(h(x, s)1 , p)) = exp[i < k ; x >] L(h0 (s1 , p)). (9.79)

Following Sect. 4.2 [see, in particular, (4.41)], we write the representation of G

induced from L and carried by the Hilbert space L H = K L2 (O , d ). Denoting
this representation by LU, we obtain

( LU(x, s) )(k) = exp[i < k ; x >] L(h0 (s1 , k))1 (s1 k). (9.80)

This representation is irreducible.

9.3 CS of Semidirect Products

The foregoing discussion on semidirect product groups can be effectively used to

derive general square integrability conditions for induced representations of the type
in (9.80). We first state, without proof, a theorem on the square integrability of such
a representation [419], which generalizes Theorems 8.3.2 and 8.3.3. The setting is
the same as in the previous section.
Theorem 9.3.1. The induced representation LU of the semidirect product group
G = V  S is square integrable if and only if the orbit O has positive Lebesgue
measure in V and the representation L is square integrable.
As a corollary, we recover Theorem 8.3.3 for L = I:
Corollary 9.3.2. Let L = I, the trivial representation. Then the representation U
is square integrable if and only if the orbit O has positive Lebesgue measure in V
and the stabilizer of any of its points is compact.
9.3 CS of Semidirect Products 261

Note that, if the orbit O has positive Lebesgue measure in V , then its dimension
is necessarily n (i.e., the same as that of V ) and, hence, the cotangent bundle T O
has dimension 2n.
We now derive a condition for the square integrability mod(H0 , ) of the
representation LU, for the principal section , introduced in (9.59). We also
assume that the Hilbert space K is finite-dimensional. Let i , i = 1, 2, . . . , N < , be
vectors in L H = K L2 (O , d ) that are smooth functions on O and have supports
contained in the set O(k0 ), used to define the local coordinates ki in (9.70). Introduce
the vectors
L
i(q,p) = U( (p)q, (p)) i , i = 1, 2, . . . , N, (q, p) V0 O ,

and the positive bounded operator on L H:

N
F = | i  i |. (9.81)
i=1

Later, we shall assume that it satisfies the condition,

L L
U(0, s0 ) F U(0, s0 ) = F, forall s0 S0 . (9.82)

The operator F has a kernel,

N
F(k, k ) = | i (k)K K i (k )| L (K), k, k O (9.83)
i=1

(the subscripts in the above expression indicating that the vectors come from the
Hilbert space K), for which

 |F  =  (k)|F(k, k ) (k )K d (k) d (k ),
O O

and if F also satisfies (9.82), then, using the explicit form of the representation LU
in (9.80), this is seen to imply

L(s0 )F(k, k )L(s0 ) = F(s0 k, s0 k ), s0 S0 , (9.84)

almost everywhere.
For arbitrary , K L2 (O , d ), consider the formal integral,

N 
I , =
 | LU((q, p)) F L
U((q, p))  d (q, p)
i=1 V0 O
N 
=
 |i(q,p) i(q,p) |  d (q, p), (9.85)
i=1 V0 O
262 9 CS of General Semidirect Product Groups

where d is the measure obtained in (9.66). Using the explicit form of the
representation LU and the invariance of the measure d , the above integral can
be brought into the form,
N 
I , = O O V0 O
 ( (p)k) : L(h0 ( (p)1 , (p)k))1 i (k)K
i=1

 i (k ) : L(h0 ( (p)1 , (p)k)) ( (p)k )K


eikk ; q
d (k) d (k ) d (q, p).

At this point we introduce the coordinates (9.70) for the variables k and k , use the
relation (9.72), and noting that the i are smooth functions, with supports contained
in O(k0 ), we obtain a -measure type of integral with respect to q. Performing this
integration and a second one over k , we obtain
N  
I , =(2 )m
d (k)
O
d (p) m( (p)1 k) (p)
i=1 O

 (k)|L(h0 ( (p)1 , k))1 i ( (p)1 k)K

 i ( (p)1 k)| L(h0 ( (p)1 , k)) (k)K ,

m being the density function [see (9.65)] that appears when the special coordi-
nates (9.70) are introduced. Thus,
N  
I , = (2 )m
d (k)
O
d (p) m( (p)1 k) (p)
i=1 O (9.86)
L L
 (k)|( U(0, (p)) )(k)K ( U(0, (p)) )(k)| (k)K
i i

Set
N 
A (k) = (2 )m
m( (p)1 k) (p)
i=1 O (9.87)
L L
|( U(0, (p)) )(k)K i
K
( U(0, (p)) )(k)| d (p),
i

the convergence of the integral in K being in the weak sense. This is a measurable
function on O that defines a formal operator A on the Hilbert space K
L2 (O , d ):
(A )(k) = A (k) (k), (9.88)
for almost all k O . If we make the assumption that A (k) is a bounded operator
(for almost all k, with respect to the invariant measure d ) and, furthermore, that
the function k  A (k)K is (essentially) bounded, then A becomes a bounded
operator and we may write

I , =  |A , (9.89)
9.3 CS of Semidirect Products 263

and, moreover,

N 
A = |i(q,p) i(q,p) | d (q, p). (9.90)

i=1 V0 O

Suppose now that  is any other section that is related to the principal section
in the manner

 (q, p) = (q, p) (n(p), s0 (q, p)), (q, p) V0 O , (9.91)

where n(p) is a measurable function on O , with values in N0 = Tk0 O , and s0 (q, p)

is a measurable function on V0 O , with values in S0 . Define the vectors, i  (q,p) =
U(  (q, p)) i , i = 1, 2, . . . , N, and the operator

N 
A  = |i  (q,p) i  (q,p) | d (q, p). (9.92)

i=1 V0 O

Note that we have not yet imposed the invariance property (9.82) on F.
Lemma 9.3.3. The operator A is bounded if and only if there exists a constant
c > 0, such that
N 
(2 ) m

 i ( (p)1 k)2 m( (p)1 k) (p) d (p) < c,
K
(9.93)
i=1 O

for almost all k O . If, furthermore, F satisfies the invariance property (9.82),
then for all other sections  of the type (9.91), A  = A .
Proof. By virtue of the unitarity of the representation L, the condition in (9.93) is
seen to be the same as the condition:
N 
(2 )m
( LU(0, (p)) i )(k)2 m( (p)1 k) (p) d (p) < c.
K
i=1 O

Since A (k) in (9.87) is a positive operator on K, the integral on the left-hand side
of the expression above is exactly equal to A (k)K , and, since A is a (block)
multiplication operator, A  is equal to the (essential) supremum of the function
k  A (k)K , which proves the first part of the lemma. The rest of the lemma is
proved by first noting that, if F satisfies (9.82), then
L L
U(n(p), s0 (q, p)) F U(n(p), s0 (q, p)) = LU(n(p), e) F L
U(n(p), e) ,

and then repeating the same computations as were done to arrive at (9.87), but now
using the vectors i  (q,p) , instead of the i(q,p) . 

We can now prove the main result of this section.
264 9 CS of General Semidirect Product Groups

Theorem 9.3.4. Suppose there exist vectors, i L H = K L2 (O , d ), i =

1, 2, . . . . . . , N, which (1) are smooth as functions on O , having supports contained
in the set O(k0 ) used to define the local coordinates ki in (9.70), and (2) satisfy the
conditions (9.82) and (9.93).
Then, the induced representation LU of the group V  S on L H is square
integrable mod(H0 , ). The operator-valued function k  A (k) is a constant,

A (k) = c() IK , (9.94)

where c() > 0 and IK is the identity operator on K. The vectors

S = { [c()]1/2 i(q,p) = [c()]1/2 LU( (p)q, (p)) i :

i = 1, 2, . . . , N, (q, p) V0 O } (9.95)

form a family of square integrable, covariant CS for this representation, and, by

(9.94), the resolution of the identity

N 
1
c()
|i(q,p) i(q,p) | d (q, p) = I (9.96)
i=1 V0 O

holds on L H.
Proof. Going back to (9.90), by virtue of Lemma 9.3.3, the operator A is seen to
be bounded, once the vectors i satisfy conditions (1) and (2). For arbitrary g G,
consider the operator LU(g)A LU(g) . Using the coset decomposition (9.69),
we find
L
U(g) LU((q, p)) = LU((g(q, p))) LU(n(p), h0 (s, p)),

where n(p) V0 and depends on p only. Thus,

L
U(g) LU((q, p)) = L
U(  (g(q, p))),

where  is a section of the type (9.91). From (9.90), we then obtain

N 

L
U(g) A LU(g) = |i  (g(q,p)) i  (g((q,p)) | d (q, p).

i=1 V0 O

Invoking the invariance of the measure d and applying Lemma 9.3.3, we see that,
for all g G,
L
U(g) A LU(g) = A .
9.3 CS of Semidirect Products 265

The irreducibility of the representation LU(g) then implies that

A = c() I, (9.97)

for some positive, nonzero constant c(), and, hence, the rest of the theorem
follows. 

Conditions (1) and (2) of this theorem constitute the admissibility conditions
mod(H0 , ) for the vectors i . Note also that, written out explicitly, (9.94)
expresses the remarkable result,

N 
c() = (2 ) m

|( LU(0, (p)) i )(k)|uK |2 m( (p)1 k) (p) d (p),
i=1 O

(9.98)

for any unit vector u K and, in particular,

N 
c() = (2 )m ( LU(0, (p)) i )(k)2 m( (p)1 k) (p) d (p),
K
i=1 O

(9.99)

for almost all k O .

Theorem 9.3.4 is readily seen to include Theorem 8.3.2 as a special case. Indeed,
if O is an open free orbit in V , it is in fact an open subset of V . Also, S0 = {e}, and
Tp O can be identified with V , for all p O . In the coset decomposition (9.55),
(k) = s. The open set O(k0 ) is O itself. Moreover, comparing with (8.60) and
noting that we have used the Borel isomorphism (9.62), d (q, p) = dq d (p), with
dq being exactly the Lebesgue measure on V . Thus, f (p) = 1, for all p O .
Furthermore, m is defined on all of O and can be identified with the function C
appearing in (8.65). Thus, any vector in the representation Hilbert space satisfies
Condition (1) of Theorem 9.3.4 and (9.82), while the condition (9.93) simply
becomes

(2 )n | (s1 k)|2 C (s1 k) d (p) < c. (9.100)
O

Comparing with (8.66), this is equivalent to


(2 )n | (p)|2 C (p) d (p) < c, (9.101)
O

which is the admissibility condition postulated in (8.68), whence Theorem 9.3.4

follows.
A few remarks are in order here, as follows.
While Theorem 9.3.4 has been proved specifically for the section , and the
definition of the CS in (9.95) depends also on the specific choice of the section
266 9 CS of General Semidirect Product Groups

(p) [see (9.54)], the admissibility condition (9.93) itself is independent of the
choice of (p), as are the constant c() and the resolution of the identity (9.96),
as can be easily verified. The same holds true for the quantity

N
1
F (q, p) =
c() |i(q,p) i(q,p) | .
i=1

While Theorem 9.3.1 is not quite a special case of Theorem 9.3.4, the former can
in fact be proved using similar arguments and the square integrability of L.
If the assumption of invariance of F under the subgroup S0 in (9.82) is dropped,
but (9.93) is retained, the square integrability of the representation LU would
still hold, mod(H0 , ), but the resolution of the identity (9.96) would have to be
replaced by the more general resolution of the operator A given in (9.90).

An Example: The Euclidean Group E(n)

An easy example of an application of Theorem 9.3.4 is provided by the Euclidean

group E(2) = R2  SO(2) [237]. This group has elements g = (v, ), v = (v1 , v2 )
R2 , [0, 2 ). The orbits of S = SO(2) on V = R2 are circles, so one chooses
O = S1 , the unit circle. The corresponding induced representation of E(2) lives in
H = L2 (S1 , d ) and reads as

(U(v, ) )( ) = ei(v1 cos + v2 sin ) ( ). (9.102)

Writing x = (x1 , x2 ) = (cos , sin ) for points on the unit circle, the cotangent
bundle is

T S1 = {(v, x) R2 S1 : v1 x1 + v2 x2 = 0}. (9.103)

It carries coordinates (a, ) R S1 defined by v1 (a, ) = a sin , v2 (a, ) =

a cos , and the corresponding invariant form is da d . A vector L2 (S1 , d )
is admissible if it is a smooth function satisfying the following conditions:
1. the support of is contained in the half-circle ( 2 , 2 );
2. is even: ( ) = ( );
 /2
| ( )|2
3. d < .
/2 cos

The coherent states are a, = U(v(a, ), ) , with v(a, ) = (v1 (a, ), v2 (a, )),
a R, 0 < 2 , and with proper normalization of ,
  2
|a, a, | da d = I. (9.104)
0
9.3 CS of Semidirect Products 267

Finally, if is admissible, the isometric map W : H L2 ( , da d ) reads as



(W )(a, ) = eia sin( ) (  ) d  . (9.105)
S1

Again, the characteristic feature here is the necessity of imposing a restriction on the
support of to guarantee its admissibility, in addition to the symmetry Condition 2
and the growth Condition 3.
The same method applies to the Euclidean group in n dimensions E(n) =
Rn  SO(n). The orbits O are now spheres Sn1 , and the representation lives in
L2 (Sn1 ). Furthermore, the vector has to be chosen SO(n 1)-invariant, this
condition replacing the symmetry Condition 2. The method also applies to the
Galilei and Poincar groups. On the other hand, as stated in Sect. 8.1, CS of the
Euclidean group have also been obtained using a reducible representation [393].

9.3.1 Admissible Affine Sections

The principal section , introduced in (9.59) and used to derive Theorem 9.3.4, is
not the only section for which the representation LU could be square integrable.
Indeed, it is possible to find conditions for square integrabilityalthough not
necessarily with a resolution of identityfor a wide class of other sections, which
we now discuss. Explicit examples, using the Poincar and Galilei groups, will be
worked out in the next chapter.
Since V = N0 V0 and H0 = N0  S0 , any other section can be expressed in
terms of in the following manner:

(q, p) = (q, p) (n(q, p), s0 (q, p)), (9.106)

where n : V0 O N0 and s0 : V0 O S0 are Borel functions. In particular, we

shall isolate a class of sections, called affine sections. These are characterized by the
following forms for the functions n and s0 :
n(q, p) = (p)q + (p),
(9.107)
s0 (q, p) = s0 (p), (q, p) V0 O ,

where, for fixed p, (p) : V V is a linear map such that Ker( (p)) =
Ran( (p)) = N0 and all three functions : O L (V ), : O N0 and
s0 : O S0 depend on p only. Additionally, they are assumed to be smooth on
the open dense set on which (p) is smooth. Then,

(q, p) = (F(p)q + (p) (p), (p)s0 (p)), (9.108)

where F(p) : V V is the linear map

268 9 CS of General Semidirect Product Groups

F(p) = (p)(IV + (p)), (IV = identity operator on V ). (9.109)

Let (p) : V V be the adjoint map to (p) and IV be the identity operator
on V .
Definition 9.3.5. The section is called an admissible affine section if, for each
p O , IV + (p) maps the set O(k0 ) used to define the local coordinates ki
in (9.70) into itself and

det [J (p, k)] = 0, p O , k O(k0 ), (9.110)

where J (p, k) is the Jacobian of the map IV + (p) restricted to O(k0 ).

To understand the nature of the admissible affine section (q, p) better, let us
write

(q, p) = (
q, (p)s0 (p)), q = F(p)q + (p) (p) V. (9.111)

For fixed p O , the set of all vectors q span an m-dimensional affine subspace p
of V . Let {e i }ni=1 be a basis of V , such that the vectors e 1 , e 2 , . . . , e m span V0
and the vectors e (m+1) , e (m+2) , . . . , e n span N0 . For each p O , define the basis,

bi (p) = F(p)1 e i , i = 1, 2, . . . , n, (9.112)

in V [where, of course, F(p) : V V is the adjoint of the operator F(p)]. Then,

since q V0 if and only if < e i ; q > = 0, i = m + 1, m + 2, . . . , n, we obtain
from (9.111) the defining equations for the affine subspace p ,

< bi (p) ; q > = i (p), i = m + 1, m + 2, . . . , n, (9.113)

where the constant term i (p) (for fixed p), which could be zero, is given by

i (p) = < bi (p) ; (p) (p) > . (9.114)

Thus, q p if and only if it satisfies the system of linear equations (9.113), i.e., if
and only if q lies in the common intersection of the n m affine hyperplanes in V ,
having normals bi (p) and constants i (p). For appropriate choices of and F, it

may turn out that the set pO p has the structure of a vector bundle. In particular,
if i (p) = 0, i = m + 1, m + 2, . . . , n, and the vectors bi (p) , i = m + 1, m + 2, . . . , n,
span the subspace Np (the annihilator of the cotangent space Tp O at p) of V , it

would follow that V = Np p , for each p, and, in this case, pO p is called a
parallel bundle.
Note, finally, that once the section : O S and the basis vectors {e i }ni=1
V are fixed, the vectors bi (p) and the scalars i (p), i = m + 1, m + 2, . . . , n, are

uniquely determined as soon as and are known.

9.3 CS of Semidirect Products 269

The class of affine admissible sections can be shown to be stable under the group
G = H  S, i.e., if (q, p) is an admissible affine section, then so also is the section
 (q, p) = g (g1 (q, p)), for each g = (x, s) G.
Using arguments similar to those which led to the proof of Theorem 9.3.4, it is
now possible to prove the more general result on the square integrability of LU
mod(H0 , ):
Theorem 9.3.6. Let be any admissible affine section, and suppose that there
exist vectors, i L H = K L2 (O , d ), i = 1, 2, . . . , N, which (1) are smooth as
functions on O , having supports contained in the set O(k0 ) used to define the local
coordinates ki in (9.70), and (2) satisfy the invariance condition (9.82) under the
subgroup S0 .
Let
L
S = { i (q,p) = U( (p)q, (p)) i : i = 1, 2, . . . , N, (q, p) V0 O }.
(9.115)
Then, the representation LU is square integrable mod(H0 , ), with coherent states
S and the expression

N 
A = |i (q,p) i (q,p) | d (q, p), (9.116)

i=1 V0 O

defining a rank-N frame F {i (q,p) , A , N}, if and only if there exist two nonzero,
positive numbers, m and M, m < M, for which

N
m( (p)1 k)
m < (2 ) m

 i ( (p)1 k)2
K |det [J (p, k)]|
(p) d (p) < M. (9.117)
i=1 O

In this case, A is the operator on L H,

(A )(k) = A (k) (k), (9.118)

with
N 
A (k) = (2 )m
|( LU(0, (p)) i )(k)K
i=1 O

m( (p)1 k)
K( LU(0, (p)) i )(k)| (p) d (p). (9.119)
|det [J (p, k)]|

Generally, admissible affine sections may exist, for which the operator A can
never be a multiple of the identity. We shall encounter such cases, for the Poincar
group, in the next chapter. The principal section is in a sense the most obvious
choice of a section, in that it is the one naturally associated to the cotangent
bundle T O .
Chapter 10
CS of the Relativity Groups

Abstract In this chapter, we examine CS for some relativity groups, namely, the
Poincar group in 1 + 1 and 1 + 3 dimensions, the Galilei groups and the Anti-de
Sitter group.

In this chapter, we examine a few of the various relativity groups, which are of great
importance in physics. The discussion will also illustrate the use of the results on
semidirect products obtained in the last chapter.

10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1)

Among the relativity groups, the most important are the Poincar and the Galilei
groups in various spacetime dimensions. These groups are semidirect products of a
translation group Rn by the corresponding isometry group of spacetime. Thus they
fall entirely within the scope of the preceding discussion.

10.1.1 The Poincar Group P+ (1, 3) in 1 + 3 Dimensions

We begin with the most fundamental relativity group, namely the Poincar group
P+ (1, 3) in one time and three space dimensions. Accordingly we treat it in a rather
detailed fashion, except for some proofs which may be found in [32].
The full Poincar group P+ (1, 3) is the twofold covering group, P+ (1, 3) =
R1,3  SL(2, C), P+ (1, 3) where R1,3 is the group of spacetime translations, with
Minkowski metric (+, , , ) (accordingly, we use in this chapter the notation
P+ (1, 3) instead of P+ (3, 1)). Elements of P+ (1, 3) will be denoted by (a, A),

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 271
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__10,
Springer Science+Business Media New York 2014
272 10 CS of the Relativity Groups

a = (a0 , a) R1,3 , A SL(2, C). The multiplication law is (a, A)(a , A ) = (a +

a , AA ) where SOo (3, 1) is the (proper, orthochronous) Lorentz transforma-
tion corresponding to A:

1
= Tr[A A ], , = 0, 1, 2, 3, (10.1)
2

1 = 1 = x , 2 = 2 = y and 3 = 3 = z are the Pauli matrices

[see (8.115)], 0 = 0 = I2 , and the metric tensor is g00 = 1 = g11 = g22 = g33 .
For typographical simplicity, we will often write p p for the spatial part of a
4-vector p.
Let

Vm+ = {k = (k0 , k) R3,1 : k2 = k02 k2 = m2 , k0 > 0}, (10.2)

be the forward mass hyperboloid. Then

k = k k = A kA , (10.3)

with k = k = k0 I2 k , = (x , y , z ).
The group P+ (1, 3) is a semidirect product of the type described in Sect. 9.2,
hence its unitary irreducible representations may be obtained by the Mackey method
of induced representations, described in Sect. 9.2.4. In the Wigner realization, the
unitary, irreducible representation UW of P+ (1, 3) for a particle of mass m > 0 and
spin s = 0, 12 , 1, 32 , 2, . . ., is carried by the Hilbert space

s = C2s+1 L2 (V + , dk
HW m ) (10.4)
k0
of C2s+1 -valued functions on Vm+ which are square integrable:

dk
(k) (k) =  2 =  |  < . (10.5)
Vm+ k0
Explicitly, in agreement with (9.80), we have

(UWs (a, A) )(k) = eika D s (h(k)1 Ah( 1 k)) ( 1 k),

(10.6)
k a = k0 a0 k a,

where D s is the (2s + 1)-dimensional irreducible spinor representation of SU(2)

(carried by C2s+1 ) and

mI2 + k
h(k) = & , (k = (k0 , k)), (10.7)
2m(k0 + m)
is the image in SL(2, C) of the Lorentz boost k , which brings the four vector (m, 0)
to the 4-vector k in Vm+ ( is a 4-vector of Pauli matrices).
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 273

The matrix form of the Lorentz boost is

 
1 k0 k
k = = k , (10.8)
m k mVk

where Vk is the 3 3 symmetric matrix

k k
Vk = I3 + = Vk . (10.9)
m(k0 + m)

The Wigner representation (10.6) is not square integrable over P+ (1, 3), as
may be shown by a straightforward computation. Hence we have to look for an
appropriate quotient. The obvious choice is phase space, which, for a classical
(spinless), relativistic particle, can be identified with

= P+ (1, 3)/(T SU(2)), (10.10)

T denoting the subgroup of time translations (as a further hint that T is a subgroup to
quotient out, one may note that it is the integral over a0 that diverges when checking
for square integrability). For a particle with nonzero spin (treated as an additional
classical degree of freedom), a geometric quantization programme [Woo92] would
normally start with the phase space  = P+ (1, 3)/(T SO(2)). However, since
geometric quantization is not our objective here, we choose in (10.10) as the
phase space for a particle with arbitrary spin s. (In the terminology of geometric
quantization, this means working on a C2s+1 -bundle, rather than on a line bundle.)
Of course, T SU(2) cannot be the stability subgroup of any vector in HW s , since

it contains translations. Hence the CS we are going to construct will not be of the
GilmorePerelomov type.
For A SL(2, C), let

A = h(k)u(k), u(k) SU(2), (10.11)

be its Cartan decomposition (see Sect. 4.5.2), in which we write explicitly the
dependance on the parameter k. An arbitrary element (a, A) P+ (1, 3) has the
left coset decomposition,
!  ! 
a0 k " ma0 "
(a, A) = 0, a , h(k) , 0 , u(k) , (10.12)
k0 k0

according to (10.11). Thus, the elements in have the global coordinatization,

(q, p) R6 :

a0 k
q = a , p = k. (10.13)
k0
274 10 CS of the Relativity Groups

In terms of these variables, the action of P+ (1, 3) on is given by (q, p) 

(q , p ) = (a, A)(q, p),

1 !  "
q = p [a + (0, q)] p
[a0 + { (0, q)}0 ]
p0 0
(10.14)
&
p = p, p = ( m2 + p2 , p),

where L+ (3, 1) is related to A by (10.1), and p0 = ( p)0 . It can be shown

[13] that the measure dq dp is invariant under this action, and hence represents the
invariant measure on , in the variables (q, p).
Comparing with the general theory in Sect. 9.2, the coset space in this case
is isomorphic to the cotangent bundle T Vm+ and Vm+ is the orbit of the vector
(m, 0) R1,3 under L+ (1, 3). The cotangent space T(m,0) Vm+ is just R3 and since
Vm+ can be globally coordinatized using vectors in R3 [see (10.2)], the open set
U(k0 ) introduced in the last chapter [see (9.70)], for defining local coordinates is
now all of Vm+ .
Although it would be possible, for constructing CS, to work with the principal
section, introduced in (9.58) and (9.59), for reasons of physical transparency, it is
better to begin, in the present case, with a different section. Accordingly, in terms of
the variables (q, p), let us define the basic section, 0 : P+ (1, 3), by

0 (q, p) = ((0, q), h(p)), (10.15)

We call 0 the Galilean section. Later we shall also link it to the principal section.
Any other section : P+ (1, 3) is then related to 0 in the manner

(q, p) = 0 (q, p) (( f (q, p), 0), u(q, p)) (10.16)

where f : R6 R and u : R6 SU(2) are smooth functions. As explained in

Sect. 9.3.1, we work with the particular class of affine sections, which now assume
the form,

f (q, p) = (p) + q #(p), (10.17)

where : R3 R, # : R3 R3 are smooth functions of p alone, and

u(q, p) = u(p), a function of p alone. Moreover, we shall only deal here with
sections for which = 0. Writing

(q, p) = (
q, h(p)u(p)), q0 , 
q = ( q) R1,3 , (10.18)

we see that

q0 = (p) 
q, (10.19)
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 275

where is the 3-vector field

p0 (p) m(p)
(p) = , so that #(p) = . (10.20)
m + p (p) p0 p (p)

We also introduce the dual vector field ,

p mVp (p)
(p) = , (10.21)
p0 p (p)

where Vp is the matrix defined in (10.9). Note that

1
= and #(p) = [p mVp (p)]. (10.22)
m

The vector fields (p) and (p) have an interesting physical and geometrical
interpretation, but that does not concern us here (see [32] for details).
We now take an arbitrary affine section , and going back to the Hilbert space
HWs in (10.4) choose a set of vectors i , i = 1, 2, . . . , 2s + 1, in it to define the formal

operator [see (7.48) and (10.6)]:

2s+1 
A = R6
|i (q,p) i (q,p) | dq dp, i (q,p) = UWs ( (q, p)) i . (10.23)
i=1

From the general definition, in order for the set of vectors

S = {i (q,p) : (q, p) R6 , i = 1, 2, . . . , 2s + 1} HW
s , (10.24)

to constitute a family of coherent states for the representation UWs , the integral
in (10.23) must converge weakly, and define A as a bounded invertible operator.
In fact, it will be possible to choose vectors i such that for each affine section ,
both A and A1 are bounded, i.e., each family S of CS will define a rank-(2s + 1)
frame.
To study the convergence properties of the operator integral in (10.23), we have
to determine the convergence of the ordinary integral

2s+1 
I , = R6
 |i (q,p) i (q,p) |  dq dp (10.25)
i=1

for arbitrary , HW
s . In (10.18) set


A(p) = h(p)u(p),  (p) = p (p)
and (10.26)
276 10 CS of the Relativity Groups

where  (p) and (p) are the matrices in the Lorentz group L (3, 1) which
+

correspond to A(p) and u(p), respectively. Then,
1
i (q,p) (k) = exp{iX(k) q} D s (v(k, p)) i ( (p) k), (10.27)

where
k p
X(k) = k #(p), (10.28)
m
and

v(k, p) = h(k)1 A(p)h( (p)1 k) SU(2). (10.29)

Substituting into (10.25) yields

2s+1 
I , = Vm+ Vm+ R6
exp{i[X(k) X(k )] q} (k) D s (v(k, p))
i=1

1
 (p) k) i ( 1
 (p) k ) D s (v(k , p)) (k ) dk dk
i ( dq dp. (10.30)
k0 k0

In order to perform the k, k integrations in (10.30), we need to change variables:

k  X(k). Computing the Jacobian JX (k) of this transformation from (10.28), we
obtain for its determinant
1
det [JX (k)] = 1 + #(p) [k0 p kp0 ]. (10.31)
mk0

Since at k = p = 0, det [JX (k)] = 1, and since we need det [JX (k)] = 0 in order to
change variables, we must impose the condition that det [JX (k)] > 0, (k, p). This,
in turn, imposes restrictions on , and hence on the 4-vector q = (
q0 , 
q) in (10.18).
Rewriting (10.31) in terms of , we get
 
p0 (k p)0 (k p)
det [JX (k)] = 1+ R(k p) (p) , (10.32)
mk0 (k p)0

where we have introduced the rotation matrix R(k p) = p1k pk1 . Thus, the
positivity of det [JX (k)] would be ensured if the second term within the square
brackets in (10.32) does not exceed 1 in magnitude, i.e., if  (p) < 1, p. To
ensure this we need the result [32]:
Proposition 10.1.1. The following conditions are equivalent:
1. The 4-vector q = ( ) is space-like, i.e., |
q0 , q q0 |2 
q2 < 0.
2. For all p R3 , the 3-vector field obeys (p) < 1 .
3. For all p R3 , the 3-vector field obeys  (p) < 1.
Consequently, we may state:
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 277

Proposition 10.1.2. The condition det [JX (k)] > 0 holds for all k, p R3 if and
q0 , 
only if the 4-vector q = ( q) is space-like, i.e., if and only if any one of the
equivalent conditions in Proposition 10.1.1 is satisfied.
An affine section , for which the corresponding 3-vector satisfies any one of the
equivalent conditions of Proposition 10.1.1 will be called a space-like affine section.
This, in the present instance, is the equivalent of the admissibility condition on the
section , postulated in Sect. 9.3.1.
We return now to the integral (10.30). For simplicity, we assume that the vectors
i , defining the CS, satisfy the condition of rotational invariance:

2s+1 2s+1
D s (u)( | i  i |)D s (u) = | i  i |, u SU(2), (10.33)
i=1 i=1

which implies that we may take

i = ei , i = 1, 2, . . . , 2s + 1, (10.34)

where ei are the unit vectors (i j ), j = 1, 2, . . . , 2s+1, in C2s+1 and L2 (Vm+ , dk

k0 ).
We assume, furthermore, that the function | (k)| itself is also rotationally invariant,
2

i.e., | (Rk)|2 = | (k)|2 , R SO(3).

Going back to the computation of I , in (10.30), we note that the dq integration
yields a -measure in X, and hence, making the change of variables k  X,
integrating and rearranging [using (10.32)], we obtain

dp dk
I , = (k) A (k, p) (k) , (10.35)
Vm+ Vm+ p0 k0

where A (k, p) is the (2s + 1) (2s + 1)-matrix kernel

A (k, p) = a (k, p) | (p)|2 I2s+1 ,

(2 )3 m
a (k, p) = , (10.36)
p0 + p R(k k1 p) (k1 p)

(2 )3 m(k1 p)0
= , (10.37)
mk0 [k0 (k p) + k(k1 p)0 ] #(k1 p)
1

and R(k p) is the rotation matrix defined above. Assuming the integral (10.25) to
exist for all , HW
s , let us write


dp
A (k) = A (k, p) . (10.38)
Vm+ p0

Then the operator A in (10.23) is a matrix-valued multiplication operator:

278 10 CS of the Relativity Groups

(A )(k) = A (k) (k), HW

s . (10.39)

On HW
s define the operators (P , P),
0

(P )(k) = k (k). (10.40)

We shall also denote the analogous operators on L2 (Vm+ , dk/k0 ) by the same
symbols. P01 is a bounded operator with spectrum [0, m1 ]. With the above
simplifications (10.38) becomes

A (k) = a (k, P) I2s+1 , (10.41)

where  denotes the L2 (Vm+ , dk/k0 ) expectation value with respect to the vector
. Hence for the operator A [see (10.39)]

A  = sup |a (k, P) |, (10.42)

kVm+

provided this supremum exists. On the other hand, since we have  (k1 p) < 1
and R(k k1 p)  = 1, from (10.36) we get

1 1 1
(p0 p) < < (p0 + p) (10.43)
(2 )3 m a (k, p) (2 )3 m

Thus, we have the result:

Lemma 10.1.3. If (p) 1, p, then a (k, p) is a bounded function satisfying

(2 )3 (2 )3
(p0 p) a (k, p) (p0 + p). (10.44)
m m
1/2
Suppose now that lies in the domain of P0 , i.e.,

| (k)|2 dk < , (10.45)
Vm+

and set

dp
P0 P = (p0 P)| (p)|2 . (10.46)
Vm+ p0

Then (10.35), (10.36) and (10.44) together imply

Lemma 10.1.4. If the assumption of rotational invariance on i , i = 1, 2, . . ., 2s + 1,
1/2
is satisfied, and if D(P0 ), then for all such that (p) 1, p,
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 279

(2 )3 (2 )3
P0 P     |I , | P0 + P    . (10.47)
m m
As a consequence of this lemma we see that both the operator A in (10.23) and
its inverse, A1
, are bounded, with

(A1 1
)(k) = [a (k, P) ] (k), HW
s . (10.48)

Indeed, collecting all these results we obtain,

Proposition 10.1.5. Let i , i = 1, 2, . . . , 2s + 1, satisfy the condition of rotational
invariance. Then for every space-like affine section , the set of vectors S
in (10.24) is a family of spin-s coherent states, forming a rank-(2s + 1) frame
1/2
F {i (q,p) , A , 2s + 1}, if and only if D(P0 ). The operator A acts via
multiplication by a bounded invertible function A (k) given by (10.41) and A1
via multiplication by the function A1 (k). Moreover, the norm of A is bounded by
the estimate,
(2 )3 (2 )3
P0 P A  P0 + P , (10.49)
m m
whilst its spectrum satisfies
(2 )3
Spec(A ) [P0 P , P0 + P ]. (10.50)
m

It is interesting to note that the estimates (10.49) and (10.50) are universal, i.e., the
bounds are valid for any space-like affine section . Indeed:
(2 )3
P0 P = inf Spec(A ), ( ) (10.51)
m 1

(2 )3
P0 + P = sup Spec(A ), ( ). (10.52)
m 1

In addition, the class of affine sections is stable under the action of P+ (1, 3).
If : P+ (1, 3) is any section, then for arbitrary (a, A) P+ (1, 3), (a,A) is
again a section [see (7.4)], where

(a,A) (q, p) = (a, A) ((a, A)1 (q, p)) = (q, p) h((a, A), (a, A)1 (q, p)), (10.53)

(a, A)1 (q, p) being the translate of (q, p) by (a, A)1 under the action (10.14) and

h((a, A), (a, A)1 (q, p)) = (q, p)1 (a, A) ((a, A)1 (q, p)) T SU(2).
280 10 CS of the Relativity Groups

Moreover, if the section defines the frame F {i (q,p) , A , 2s + 1}, then (a,A)
defines the frame F {i , A(a,A) , 2s + 1}, where A(a,A) = U(a, A) A U(a, A).
(a,A) (q,p)
Let now A denote the class of all affine space-like sections, but with not
necessarily assumed to be zero. Then,
Proposition 10.1.6. If A, then (a,A) A, for all (a, A) P+ (1, 3).
In view of this result (see [32] for a proof), starting with any family of coherent
states S , we may generate an entire class of covariantly translated families S(a,A)
of other coherent states, using the natural action (10.53) of P+ (1, 3) on the space
of sections. If is characterized by and , then the functions  and  that
characterize (a,A) may be computed explicitly without difficulty.
Some special cases of the affine admissible sections are of particular interest:

10.1.1.1 The Galilean Section 0

This section is defined by

p p
(p) = 0 (p) = 0, #(p) = #0 (p) = 0 and 0 (p) = , #0 (p) = .
p0 m
(10.54)
Thus, here 0 (p) < 1, 0 (p) < 1, p. From (10.54) and (10.37),

(2 )3 k0 p0 k p
a (k, p) = a0 (k, p) = , (10.55)
m k0

and using the rotational invariance of | (k)|2 ,

(2 )3
A (k) = A0 (k) = P0  I2s+1 . (10.56)
m
Hence,

(2 )3
A = A0 = P0  I, (10.57)
m
so that the frame is tight.

10.1.1.2 The Lorentz Section 

This is the principal section introduced in the previous chapter. This time

(p) =  (p) = 0 (p), #(p) = # (p) = #0 (p), (10.58)

10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 281

in other words, the Galilean and Lorentz sections are duals to each other. Further-
more, for this section one again gets a tight frame:

A = A = (2 )3 mP01  I. (10.59)

Indeed, from (10.58) and (10.36), we see that a (k, p) does not depend on k:

(2 )3 m
a (k, p) = a (k, p) = . (10.60)
p0
In fact, assuming rotational invariance, one sees from (10.36) that this property holds
only if (p) 0, that is, for the Lorentz section.

10.1.1.3 The Symmetric Section s

This section is self-dual, being given by

p p
(p) = s (p) = s (p) = , #(p) = #s (p) = #s (p) = . (10.61)
m + p0 m + p0
Again, s (p) < 1, p. Proceeding in the same way as before, we find that the
operator A = As is now given by
; <
k0 P0 + m2
(As )(k) = As (k) (k) = (2 )3 (k), HW
s . (10.62)
m(k0 + P0 )

To determine the spectrum of As , note that the function

k0 p0 + m2
f (k0 ) as (k, p) = (2 )3 , (10.63)
m(k0 + p0 )

satisfies the estimate (2 )3 f (k0 ) (2 )3 p0 /m, k0 [m, ), This implies that

the spectrum of the operator A = As is the entire interval [(2 )3  2 , (2 )3
P0  /m], so that for this section the frame is never tight.

10.1.1.4 The Limiting Sections

In addition to these three space-like sections, it is worth mentioning two limiting

cases, given by
p mp
(p) = + (p) = , #(p) = #+ (p) =
p p(p0 p)
. (10.64)
(p) p mp
(p) = + = , (p) = #+ (p) =
p p(p0 + p)
282 10 CS of the Relativity Groups

In this limiting situation, + (p) =  (p) = 1, p. Thus these two sections are
light-like, and duals of each other. Here again, the frame is never tight. In addition,
each of these light-like sections saturates one of the bounds in (10.49), + the lower
bound and the upper bound.
One can also consider time-like sections, with + (p) > 1, but then admissibil-
ity [i.e., convergence of the integral (10.30)] will require restrictions on the support
of the function . This situation seems to be generic, we have met it in the case
of Euclidean CS in Sect. 9.3, and we will meet it again for the group in Sect. 10.2
below.
A final comment is in order here. The Galilean and the Lorentz sections are dual
to each other, yet completely dissimilar. More generally, the nature of the operator
A depends strongly on the choice of the section. Every space-like affine section
yields a multiplication operator, but the spectrum of the latter may be very different
for different sections. The geometrical analysis made in [32] shows that the Lorentz
section is in fact the only truly covariant section, as expected from the principal
section, and thus should yield simpler results. We will see shortly that the same
holds in one space dimension. In this sense, choosing a section is very similar to
choosing a gauge in electrodynamics. There too, some choices are more convenient
than other ones for a given problem.

10.1.2 The Poincar Group P+ (1, 1) in 1 + 1 Dimensions

The 1+1 dimensional case P+ (1, 1) is entirely parallel to the previous one, for spin
s = 0. Accordingly we will be very brief, referring the reader to the original papers
[2730] for further details.
The group elements (a, p ) of P+ (1, 1) = R2  SOo (1, 1) are parametrized by
a = (a0 , a) R2 and p SOo (1, 1), the latter being a Lorentz boost indexed by
p = (p0 , p), where p0 = (p2 + m2 )1/2 , m > 0.
The Wigner representation UW of mass m 0 reads:

(UW (a, p ) )(k) = eika ( p1 k), Hm = L2 (Vm+ , dk/k0 ). (10.65)

Here Vm+ denotes the forward mass hyperbola of mass m 0, Vm+ = {k = (k0 , k) :
k2 = k02 k2 = m2 , k0 > 0}, k a = k0 a0 k a, with invariant measure dk/k0 . Again
UW is not square integrable over P+ (1, 1) and we look for appropriate quotients.
The natural phase spaces, i.e., the coadjoint orbits of P+ (1, 1), are the following
[21, 57]:
For m > 0: = P+ (1, 1)/T , where T is the subgroup of time translations.
For m = 0: l(r) = P+ (1, 1)/Tl(r) , where Tl , Tr are subgroups of lightlike
translations and the indices l, r correspond to the left, resp. the right, branch of
the light cone (see Sect. 10.1.3)
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 283

Neither T , nor Tl(r) , is the stability subgroup of any vector in Hm , and hence we
again have to use the general construction for building covariant coherent states.
Let us consider first the massive case, m > 0. The phase space = P+ (1, 1)/T
carries canonical coordinates (q, p) and the P+ (1, 1)-invariant measure dq dp. In
the principal bundle P+ (1, 1) = P+ (1, 1)/T , a section : P+ (1, 1)
reads:

(q, p) = 0 (q, p) (( f (q, p), 0), I) =: (

q, p ), (10.66)

where 0 (q, p) = ((0, q), p ) is again called the Galilean section, and the expression
multiplying 0 (q, p) is an arbitrary element of T , indexed by the function f . As in
the 1+3 case, we consider only affine sections, corresponding to f (q, p) = q #(p).
Thus the section is characterized by the function #. The class of affine sections is
stable under the action  (a, p ) of P+ (1, 1), and it has the advantage, that for
such sections, all calculations may be done explicitly.
As before, the section may be parametrized in terms of a momentum-
dependent speed :

q0 = (p) 
q. (10.67)

Geometrically, (10.67) implies that for fixed p, the choice of determines a

particular reference frame in q-space [28]. In particular, we see that the vector q is
time-like, light-like or space-like, whenever || > 1, || = 1 or || < 1, respectively,
and this classification is also P+ (1, 1)-invariant (compare Proposition 10.1.1).
Sections for which || < 1 will thus be called space-like.
We introduce also the dual speed (compare (10.21))

p p0 (p) 1
(p) = = (p m#(p)) . (10.68)
p0 p (p) p0

One has, as before, = , which means that the sections come in dual pairs, and
moreover the duality preserves the space-like (respectively, time-like or
light-like) character of the vector q.
According to the general scheme, we now define the Poincar CS

(q,p) = UW ( (q, p)) , (q, p) , (10.69)

and the corresponding resolution operator


A = | (q,p)  (q,p) | dq dp. (10.70)

Then the results of [2730] may be summarized as follows.

284 10 CS of the Relativity Groups

Theorem 10.1.7. Let : = P+ (1, 1)/T P+ (1, 1) be any affine space-like

section. Then a vector is admissible mod(T, ) iff it is of finite energy,
1/2
D(P0 ), and every such vector generates a family of coherent states indexed by
points in the phase space , which constitute a rank one frame:

S = { (q,p) = UW ( (q, p)) , (q, p) }. (10.71)

More precisely, A is a positive bounded multiplication operator with bounded

inverse: (A )(k) = A (k) (k), where

dp
A (k) = A (k, p) | (p)|2 . (10.72)
Vm+ p0

The kernel A (k, p) is defined in terms of the function # or the speed defining the
section as follows:

2 k p
A (k, p) = (10.73)
m k0 p #( p1 k/m)
2 m
= . (10.74)
p0 + p ( p1 k/m)

The spectrum of the operator A obeys the following universal bounds:

2 2
P0 |P| A  P0 + |P| , (10.75)
m m

where   |  denotes a mean value in the state .

In the relations (10.73) and (10.74), z z denotes again the space component of a
2-vector z = (z0 , z).
Since the resolution operator A is a positive multiplication operator, so are its
inverse and its square root. The latter fact allows to compute explicitly weighted CS,
as indicated in (7.59), Sect. 7.3.
Proposition 10.1.8. For any space-like affine section and admissible vector
, the frame S = { (q,p) , (q, p) } may be turned into a tight frame of
weighted CS:

w(q,p) = W (q, p) (q,p) , (10.76)

where W (q, p) is a measurable family of weighting operators, given explicitly by

1/2
W (q, p) =  2C (p)| (q,p)   (q,p) |. (10.77)
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 285

1/2
In this expression, C (p) is the positive self-adjoint operator of multiplication by
the function A (k, k p)1/2 , where p = (p0 , p).
We conclude with some concrete examples, which are essentially identical to
those discussed in the 1+3 case. Among the class of admissible space-like sections,
we recover the three remarkable cases discussed in the previous section, with almost
identical properties. Once again, of course, the Lorentz section is the principal
section.
p
1. Galilean section 0 : #(p) = 0 (p) = 0 (p) = :
p0

A0 = 2 m1 [P0  I P P P0 1 ],

and thus the frame can be made tight if one requires that P = 0. This is the
one-dimensional remnant of the rotational invariance imposed in the 1+3 case.
p p
2. Lorentz section  : #(p) = (p) = (p) = 0 :
m p0
Al = 2 mP01  I, thus the frame is tight for any admissible vector , and
this is the only section with this property (here, and only here, the kernel A (k, p)
does not depend on k).
p p
3. Symmetric section s : #(p) = (p) = (p) = :
m + p0 m + p0
the frame is not tight for any .
Again the first two sections are dual of each other and the third one is self-dual.
Also, non-space-like sections can be considered, exactly as in the previous case.
As for admissible vectors, the following functions are convenient for explicit
computations:
Gaussian vector: G (k) ek0 /U
Binomial vector: (k) (1 + (k0 m)/U) /2 , > 1/2.
In both cases, U is a normalization constant, with the dimension of an energy.
We will see in Sect. 10.3 below that these two types of admissible vectors have
attractive properties when comparing the Poincar group P+ (1, 1) with the Anti-
deSitter group SOo (1, 2) and the Galilei group G(1, 1), via the technique of group
contraction.
Finally a word can be said about orthogonality relations, in the spirit of Sect. 8.2.
In order to derive these relations, we have to compute the following matrix element,
and in particular to find sufficient conditions for its convergence:

I1 2 (2 , 1 ) = 2 |A1 2 |1 , (10.78)

where

A1 2 = |1 (q,p)  2 (q,p) | dq dp. (10.79)

286 10 CS of the Relativity Groups

In the case of a tight frame, we will get A1 2 = (1 , 2 )I, where (, ) is a

sesquilinear form. Then one gets

I1 2 (2 , 1 ) = (1 , 2 ) 2 |1 , (10.80)

If the form is closed, we may write (1 , 2 ) = C1 |C2 , as in the proof of

Theorem 8.2.1, and ordinary orthogonality relations follow.
Coming back to our Poincar CS, and proceeding exactly as before, one finds
1/2
that the integral (10.79) converges if both 1 and 2 belong to the domain of P0
1 2
and then A is the operator of multiplication by the function

dp
A1 2 (k) = A (k, p) 2 (p) 1 (p) , (10.81)
Vm+ p0

with the kernel A (k, p) given in (10.74). Of course, putting 1 = 2 = , one

recovers the operator A A .
From (10.81), we can now explicitly compute the matrix element (10.78). Let
1/2
1 , 2 D(P0 ) and 1 , 2 H, and write j = | j  j |, j = 1, 2. Then one has

I1 2 (2 , 1 ) = U( (q, p))2 |2 U( (q, p))1 |1  dq dp


= Tr[U( (q, p)) 2 ] Tr[U( (q, p)) 1 ] dq dp

= 2 |A 1 B2 (H) , (10.82)

where A is a positive self-adjoint operator on B2 (H) :

(A |  |)(k, p) = A (k, p) (k) (p). (10.83)

Consider first the case of a tight frame. For the Galilean section 0 with P = 0,
we obtain
2
0 (1 , 2 ) = 1 |P0 2 , (10.84)
m
which is manifestly a closed sesquilinear form. Hence we have orthogonality
relations, with the DufloMoore operator

2 1/2 m 1
C0 = P or K0 = P . (10.85)
m 0 2 0

Similarly, for the Lorentz section  , we get

 (1 , 2 ) = 2 m 1 |P01 2 , (10.86)
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 287

and thus
1/2 1
C = 2 mP0 or K = P0 . (10.87)
2 m
We see again how much the result depends on the choice of the section, even
for tight frames. If the frame is not tight, however, there seems to be no way of
transforming (10.82) into a useful orthogonality relation.

10.1.3 Poincar CS: The Massless Case

We turn now to the massless case. In 1+3 dimensions, there is no significant

difference from the massive case, since the limit m 0 is known to be nonsingular
(but there is more freedom in choosing an appropriate phase space, as indicated
in [Sch96]). However, in 1+1 dimensions, the situation becomes more intricate.
Indeed, because of the infrared divergences inherent to any massless quantum field
theory in 2 dimensions, one is forced to work in an indefinite metric (Krein) space,
with a nonunitary, indecomposable representation of P+ (1, 1). Yet CS may be
obtained by a suitable adaptation of the general formalism used so far, namely the
introduction of quasi-sections, as done in [57], which we follow.
Let us first recall that a Krein space K is a vector space equipped with a
nondegenerate sesquilinear form , , such that K = K1 K2 , where the sesquilinear
form is positive definite (resp. negative definite) on K1 (resp. K2 ) and both K1 and K2
are complete (i.e., Hilbert spaces) with respect to the restriction of the sesquilinear
form. It follows that any vector f K may be decomposed as f = f1 + f2 ( f j K j ),
with

 f , f  =  f1 2  f2 2 ,  f1 2 =  f1 , f1  0,  f2 2 =  f2 , f2  0.

The Krein space K is called a Pontrjagin space if one of K1 , K2 is finite dimensional.

Finally, given a Krein space K with indefinite inner product , , a majorant Hilbert
topology on K is a Hilbert inner product (i.e., positive definite) (, ) such that

| f , g| ( f , f )1/2 (g, g)1/2 , f , g K.

A systematic study of Krein spaces and other indefinite inner product spaces may
be found in [Bog74].
Let us now come back to P+ (1, 1). Starting from the relevant Wightman two-
point function, which is not positive definite, one gets an indefinite sesquilinear
form ,  on the Schwartz space S (R2 ). In order to obtain an appropriate majorant
Hilbert topology, let S (R2 ) such that (0) = 1 . Given f S (R2 ), define
f0 (x) = f (x) f(0) . Then the desired inner product may be written as

( f , g) =  f0 , g0  +  f ,  , g + f(0)
g(0) (10.88)

It may be directly verified that (10.88) defines a Hilbert majorant topology.

288 10 CS of the Relativity Groups

Upon completion with respect to this Hilbert topology, one obtains the Krein
space (actually it is a Pontrjagin space with one negative dimension):
! dk "
K = L2 C+ , V X, (10.89)
|k|

where C+ is the positive light cone, and X and V are both one dimensional
subspaces. X is the one-dimensional subspace generated by the function , while
V is the one dimensional subspace generated by the vector v which is defined (by
the Riesz lemma) as the representative of the functional f   , f  = (v, f ).
The nonpositivity of , . comes from the fact that the metric operator J,
! " by (, ) = , J, is the identity on the first summand of K, but equals
defined
0 1
1 0
on the complementary subspace V X. Then the Wigner representation UW
may be defined on a dense subset of K. However it is now only J-unitary, i.e.
UW (a, ) f ,UW (a, )g =  f , g, and not unitary for the associated Hilbert space
inner product (, ). Furthermore, UW is neither irreducible nor completely reducible,
but indecomposable, that is, there is an invariant subspace whose orthogonal
complement is not invariant.
Since the positive light cone is disconnected,
! " consisting of two half-lines, C+ =
R+ R , the Hilbert subspace L2 C dk/|k| may in turn be decomposed into a
+
direct sum:
! dk "
L2 C+ , = Hl H r , (10.90)
|k|

where
! dk " ! dk "
Hl = L2 R , , Hr = L2 R+ , , (10.91)
|k| |k|

(left and right Hilbert spaces, respectively). Correspondingly, one may quotient the
representation UW and obtain finally two unitary irreducible representations Ul(r)
defined on Hl(r) . This amounts to considering the matrix elements

1 ,UW (a, )2 Hl(r) , 1 , 2 Hl(r) , (10.92)

and associating to the sesquilinear forms so defined the operators Ul(r) (a, ). The
final result is

(Ul(r) (a, ) )(k) = eika ( 1 k), Hl(r) . (10.93)

These are the representations of P+ (1, 1) that are used in the construction of
systems of massless coherent states (but the resolution of the identity (10.98) that
we will obtain in the end lives in the Krein space K).
10.1 The Poincar Groups P+ (1, 3) and P+ (1, 1) 289

We consider first the right coadjoint orbit r = P+ (1, 1)/Tr , corresponding to

the representation Ur of P+ (1, 1). A global parametrization of this orbit is given
by ( , p), R, p > 0 and it carries the P+ (1, 1)-invariant measure d ( , p) =
d dp/p. But it turns out that the natural (Galilean) section n ( , p) = ((0, ), p )
is not admissible, because the integral of the type (7.48),

cr ( , ) = |Ur (n ( , p)) | |2 d dp/p, (10.94)
r

giving the admissibility condition, is infrared divergent (that is, at p = 0), for any
, Hr . However, the following quasi-section is admissible:

p 1
r ( , p) = ((0, p1 ), (p) ), where (p) = . (10.95)
2 2p

A straightforward calculation then shows that a vector Hr is admissible for the

1/2
quasi-section (10.95) if and only if D(Hr ), where Hr is the usual momentum
operator on Hr (but now both Hr and Hr1 are unbounded).
Using this quasi-section, we define right coherent states as the states:
1 k
r ( ,p) (k) = (Ur (r ( , p)) )(k) = ei p (kp1 ), Hr . (10.96)

Then an explicit calculation leads to the following (weak) identity:


1
|Ur (r ( , p)) Ur (r ( , p)) | d ( , p) = I, (10.97)
cr ( ) r

where cr ( ) cr ( , ) is given in (10.94). Thus we get a genuine resolution of

the identity and a tight frame. An intriguing fact, albeit not totally unexpected (since
the phase spaces are the same in the two cases, namely a half-plane, although the
two groups are different), is that these right CS of P+ (1, 1) are identical with the
familiar wavelets associated to the ax + b group, that we shall discuss in Chap. 12
[see also Sect. 8.1, in particular (8.24)].
In the same way, we may construct a corresponding set of left coherent states.
Collecting together all these results, one obtains a resolution of the identity in the
Krein space (10.89):

1
I = PV + PX + |Ul (l ( , p)) Ul (l ( , p)) | dl ( , p)
cl ( ) l

1
+ |Ur (r ( , p)) Ur (r ( , p)) | dr ( , p), (10.98)
cr ( ) r
290 10 CS of the Relativity Groups

where Hl , Hr , both admissible, and PV , PX denote the rank 1 projection

operators on V, X, respectively. This result suggests that most of the CS formalism
might possibly be extended to (simple) indefinite metric spaces.

10.2 The Galilei Groups G(1, 1) and G G(3, 1)

In Sect. 8.4.1, we have discussed at great length the structure of the extended Galilei
group G G(3, 1), and we have constructed, by the Mackey method, a class of
induced unitary irreducible representations, given in (8.129) and (8.131). Coming
to CS, however, we have limited our analysis to the isochronous subgroup G ,
because the restriction of these representations to G are square integrable mod
( SU(2)), where is the phase subgroup coming from the central extension.
The corresponding parameter space is the phase space = G /( SU(2)) R6 ,
and the associated Galilei CS constitute a tight frame, as seen in (8.144).
In the present section, we are going to extend the analysis to the full extended
Galilei group G. However, exactly as for the Poincar case, the rotation subgroup
inherent to three space dimensions does not change things much except, of course,
for the possibility of having nonzero spin values. Indeed the essential features of
the construction are already present in one space dimension. Hence we shall mainly
discuss here the corresponding group G(1, 1). Moreover, we shall be rather brief,
since the procedure parallels that used above for the Poincar group P+ (1, 1).
Details may be found in [30].
Group elements of G(1, 1) are parametrized as g = ( , b, a, v), where is
the parameter corresponding to the central extension of mass m, (b, a) time and
space translations and v the boost parameter. In momentum space, the unitary
irreducible representation that we shall use reads (compare (8.129)), with j = 0
and E0 = 0):
! 2 "
 )(k) = exp i + k b k a (k mv),
(U(g) L2 (R, dk). (10.99)
2m

This representation is not square integrable over G(1, 1) (the integral over b
diverges), hence we have to take a quotient. Following the now standard procedure,
we choose the associated coadjoint orbit of G(1, 1), which is the quotient =
G(1, 1)/( T ), where T is the subgroup of G(1, 1) consisting of phase
changes and time translations. It carries global coordinates (q, p) := (a bv, mv)
R2 and the invariant measure is dq dp.
As in the Poincar case, we consider first the Galilean section 0 : G(1, 1),
which in this case is the principal section, and is given by
! p"
0 (q, p) = 0, 0, q, . (10.100)
m
10.2 The Galilei Groups G(1, 1) and G G(3, 1) 291

Then an arbitrary affine section is of the form (again an irrelevant term (p) has
been set to zero):

(q, p) = 0 (q, p) (0, q #(p), 0, 0)

! p2 p p"
= q #(p) , q #(p), q + q #(p), , (10.101)
2m m m
with # a real continuous function that indexes the section.
Next one looks for conditions to be imposed on admissible vectors to ensure
that the integral I1 2 (2 , 1 ), the analogue of (10.78), converges and is explicitly
calculable. As compared to the Poincar case, the new feature here is that (except for
the Galilean section) the Jacobian JX (k) corresponding to the change of variables
k  X(k) in the integral analogous to (10.30) is totally unwieldy, unless one
imposes a restriction on the support of the admissible vectors j . Accordingly, one
defines L2 (R, dk) to be admissible mod( T , ) if it satisfies the following
two conditions:
(i) k R, p #(k p) m implies (p) = 0;
(ii) The following integral converges:
 ! "1
p
A (k) = 2 dp 1 #(k p) | (p)|2 . (10.102)
p#(kp)<m m

For the Galilean section, corresponding to # 0, every vector L2 (R, dk) is

admissibleagain this is the natural or fully covariant section. For any other affine
section, the support of an admissible vector must be restricted, and, contrary to
the Poincar case, the admissibility of depends now explicitly on the particular
section . Notice that the integral in (10.102) defines a positive function, which is a
constant iff is a constant. We summarize the results in a proposition:
Proposition 10.2.1. Let : G(1, 1) be an affine section. Then any vector
L2 (R, dk) admissible mod ( T , ) generates a frame of CS indexed by points
in phase space, given by the vectors

 (q, p)) ](k)

(q,p) (k) = [U(
 
|k p|2
= exp i q #(p) k q (k p). (10.103)
2m

The corresponding resolution generator A is a positive, bounded, invertible

multiplication operator :

(A )(k) = A (k) (k), L2 (R, dk), (10.104)

with kernel A (k) given in (10.102).

292 10 CS of the Relativity Groups

Here too, as in the Poincar case (Proposition 10.1.8), each of the CS frames so
obtained may be turned into a tight frame of weighted CS, using weighting operators
similar to (10.77). We refer to [30] for details.
Among affine sections, we mention the special (rather trivial!) cases:
1. The Galilean section 0 , corresponding to #(p) = 0: every vector is admissible,
and the frame is tight.
2. Constant sections, with #(p) = #c , a positive constant: one gets a tight frame for
any vector with support in (, m#c 1 ) and such that (m P #c )1  < .
For these two sections, one may derive orthogonality relations, exactly as for
P+ (1, 1). The DufloMoore operator is

Cc = 2 m (m P c )1/2 (c 0). (10.105)

Other sections will in general require support restrictions on and give nontight
frames.
We conclude with two examples of admissible vectors, similar to those given in
the previous section for P+ (1, 1):
Gaussian vector: G (k) exp(k2 /2mU).
Binomial vector: (k) (1 + k2 /2mU) /2 , > 1/2.
Here too U is a normalization factor, which may be interpreted as a kind of internal
energy.
Exactly the same analysis may be made in the three dimensional case G =
G(3, 1), with very similar results. In the simplest situation of spin j = 0, one may
choose as parameter space the coadjoint orbit = G/( T SU(2)) R6 , with
coordinates (q, p) := (a bv, mv) and invariant measure dq dp.

10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2)

and Its Contraction(s)

Among many other applications in physics [305], the group SOo (1, 2) and its double
covering SU(1, 1) SL(2, R) play a central rle in the kinematical symmetries
of (1+1)-dimensional spacetime. Different kinematical (or relativity) Lie groups
are possible. Two of them have been considered in the previous sections, namely
the Galilei group G(1, 1) and the Poincar group P+ (1, 1). But one can in fact
distinguish (at least) seven different kinematics, by adapting to (1+1)-dimensional
spacetimes the classification of Bacry and Lvy-Leblond [108]. Restricting to 1+1
dimensions is just a matter of technical (but not conceptual!) simplification. These
seven kinematics or relativities are well encapsulated in the following diagram
concerning the relations between their respective Lie groups :
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 293

SOo (2, 1) R2  SOo (1, 1) =: N+

(de Sitter) (Newton)
"
P+ (1, 1) G(1, 1) R3
(Poincar) (Galilei) (Static)
#
SOo (1, 2) R2  SO(2) =: N
(Anti-de Sitter) (Newton)

Two of the relativities have maximal symmetry, i.e., their kinematical groups are
the de Sitterian pseudo-orthogonal groups SOo (1, 2) or SOo (2, 1). The difference
between them stems from the compact or noncompact nature of the time translation
subgroup. Here, no physical unit is necessary to standardize their three (pseudo-)
angular parameters. They are departure points for successive contractions (denoted
by arrows on the diagram) until the ultimate kinematics is reached, where nothing
moves (the static situation). At each step, some of the parameters acquire a physical
dimension. Hence, they may be interpreted as a length, time, or momentum.
Correspondingly, part of the original simple group structure breaks down into a
(semi)direct product structure. Note that, on a quantum level, central extensions
(trivial or not) of the respective groups have to be taken into account in order to get
rid of infinite phases resulting from the contraction limits.
The Anti-de Sitter group SOo (1, 2) is of special interest, because it is the
symmetry group for an elementary system which is simultaneously a deformation
of the relativistic free particle (rest energy mc2 ) and the harmonic oscillator (rest
energy 12 h , where := c and is the spacetime curvature) [311, 316]. Indeed,
writing the commutation rules for so(1, 2) su(1, 1):
1 1
[X0 , X1 ] = mc2 2 X2 , [X2 , X0 ] = X1 , [X2 , X1 ] = X0 , (10.106)
m m2 c
where X0 is the compact generator, one can check that the limit 0 gives the
Poincar Lie algebra whereas the limit 0, c , c = , gives the harmonic
oscillator Lie algebra, renamed as the Newton (N ) Lie algebra in the present
context. The contraction toward the harmonic oscillator has been studied in [316].
Note that, in this context, the GilmorePerelomov CS for the discrete series of
SU(1,1) contract smoothly toward the WeylHeisenberg canonical CS [315]. On
the other hand, the zero curvature limit toward P+ (1, 1) and its representations
was studied in [26], with the aim of analyzing the (singular) behavior of the
SU(1, 1) GilmorePerelomov CS as 0. The problem was considered again in
[206, 240, 307, 519], in order to give a more rigorous treatment to the contraction
procedure . The present section is strongly inspired by [519]. Henceforth, we put
m = 1, c = 1, h = 1 for convenience.
The definition of the contraction given in (4.155) at the level of the Lie algebras
g1 and g2 is adapted to the present problem by defining the deformation map
294 10 CS of the Relativity Groups

:= : g2 = p(1, 1) g1 = su(1, 1), R+

, (10.107)

in the following way. First, we identify the Poincar group P+ (1, 1) with the group
of (projective) 3 3 matrices

cosh sinh a0
(a, p ) sinh cosh a = M(a, ), (10.108)
0 0 1
with cosh = p0 , sinh = p. The Lie algebra basis resulting from this parametriza-
tion is given by

f0 = M(a, )|a=0, =0 (10.109)
a0

f1 = M(a, )|a=0, =0 (10.110)
a

f2 = M(a, )|a=0, =0 . (10.111)

We now define the invertible linear map : p(1, 1) su(1, 1) by

f0  e0
: f 1  e1 (10.112)
f2  e2

where {e0 , e1 , e2 } is the basis of su(1, 1):

     
1 i 0 1 01 1 0 i
e0 = , e1 = , e2 = , (10.113)
2 0 i 2 10 2 i 0
with the commutation rules [compare to (10.106)]

[e0 , e1 ] = e2 , [e2 , e0 ] = e1 , [e2 , e1 ] = e0 . (10.114)

Then one can see that

lim 1 [ fi , f j ]1 = [ fi , f j ]2 , i, j = 0, 1, 2. (10.115)
0

The corresponding map at the group level, : P+ (1, 1) SU(1, 1), then reads

: M(a, ) = ea0 f0 +a f1 e f2  e (a0 e0 +ae1 ) e e2 (10.116)

Note the factorization formula

 
e  
e (a0 e0 +ae1 ) = e 0 e e1 e e2 ,

10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 295

where the different parameters are defined or related by the following relations,
where 12 (a20 a2 )1/2 :

sin
= cos + ia0 C,
2
sin
= a R,
2
 
 sin 2
= arg cos 2 + ia0 ,
2
 
sin 2
 = sinh1 a
2

tanh 2 ei


tanh = i  .
2 tanh 2 ei 

Let us now recall the respective UIRs involved in the contraction scheme. For
SU(1, 1), they are the representations U j of the discrete series already described in
Sect. 4.2.2, and extended to the universal covering group of SU(1, 1). This means
that the parameter j = 1, 3/2, 2, 5/2, . . . in (4.76) runs now over the continuous range

1
< j = 1/ < . (10.117)
2
We recall also the relevant Hilbert spaces and unitary actions. For typographical
simplicity, we write U := U 1/ , H := H1/ , := 1/ :

H = L2 (D, (2/ 1) d ), (10.118)

dz dz
d (z, z) = (1 |z|2 )2/ 2 , (10.119)
2 i
 
z
(U (g) f )(z) = ( z)2/ f , (10.120)
z
 

for f H and g = SU(1, 1).

For the Poincar group, we shall choose another parametrization for the Wigner
representation (10.65), better suited to the contraction procedure. To that effect, we
introduce the carrier Hilbert space
 
H = L2 (1, 1); 2 dx/(1 x2 ) ,

obtained from L2 (Vm+ , dk/k0 ) by changing k into x = k/(k0 + 1), i.e.,

k = 2x/(1 x2 ), and putting m = 1. The (unit mass) Wigner representation
U(a, ) := UW (a, p ) is then given by
296 10 CS of the Relativity Groups

  $ %
1 + x2 2x x cosh 2 + sinh 2
(U(a, ) f ) (x) = exp i a0 + a f .
1 x2 1 x2 cosh 2 + x sinh 2
(10.121)

Let us consider an element f of the FockBargmann space H on the disk

D and its restriction, suitably weighted, to the purely imaginary diameter
{z = ix, 1 < x < 1}:

I ( f )(x) := (1 x2 )1/ f (ix). (10.122)

The map I from H to H is called a precontraction (see Sect. 4.5.4). It has the
crucial property :
Theorem 10.3.1. The precontraction I is a continuous injection from H into H.
Proof. Injectivity results from the holomorphy of f . Continuity results from the
following majoration, proved in [519]:
 1
(2/ + 1)22/
I ( f )2H = 2 | f (ix)|2 (1 x2 )2/ 1 dx  f 2H , (10.123)
1 2/ 1


Of course, the map I is not surjective. For instance, the function (1x2 )1/ e1/(1ix)
is in H for all < 2, but e1/(1z) is not in H for any < 2.
Let us put C = I (H ). It is clear that, for  < , we have H  H , and so
C  C . On the other hand, the functions of the form (1 x2 )p(x), where p is
a polynomial, form a dense subset of H and they are also elements of C1 . Since
C C1 for all 1, C is dense in H for all 1. (Actually the same holds true
for any < 2). From this discussion the following commutative diagram results:

. . . H H  . . . (  < )
" * (10.124)
H

where the arrows represent continuous injections. Hence the Wigner Hilbert space
H is the universal arena into which all the spaces H are conveniently injected for
the purpose of contraction. Accordingly, we propose the following definition.
Definition 10.3.2. A family of states H is said to contract to H if I ( )
tends to in some (possibly weak) sense when 0.
We are now in a position to see what happens to the GilmorePerelomov coherent
states (4.99),
 1/2
1 2/ 1
(z) (z ) = (1 |z|2 )1/ (1 z z)2/ , (10.125)
2 2
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 297

which are mapped to

 1/2  1/
1 2/ 1 (1 x2 )(1 |z|2 )
I ( (z) )(x ) = . (10.126)
2 2 (1 ix z)2

When considered as a distribution, the action of this vector on C0 (1, 1) reads

1
I ( (z) )|  =
2
    1/ (10.127)
1 /2 1/2 1 (1 x2 )(1 |z|2 )  dx
=2 (x ) .
1 (1 + ix z)2 1 x2

This expression has in general no limit for 0, except for z = ix, x (1, 1). In
this case, we compute that (10.127) indeed converges to 2 (x).
So, when the CS label z is concentrated on the purely imaginary diameter of
the disk, the limit of (10.126) is a Dirac distribution. This result could actually be
expected from the fact that the GilmorePerelomov CS (10.125) are themselves a
quantization of the open unit disk D considered as an SU(1, 1) coadjoint orbit or,
better, as a classical phase space D SU(1, 1)/U(1) for a (massive) free particle
in an Anti-de Sitter spacetime. This phase space interpretation is well captured by
the following (a, k) parametrization of D:

sinh a + ik cosh a &

D % z = z(a, k; ) = , k0 = 1 + k2 , a, k R. (10.128)
1 + k0 cosh a

The singular behavior of these CS is then obvious. They exactly reproduce the
singular behavior of the parametrization (10.128) at zero curvature,

ik
lim z(a, k; ) = = ix, (10.129)
0 k0 + 1

i.e., the disk D contracts into the segment (i, i) . However, although the original
complex structure is lost, this does not imply a singular behavior at the level of
the respective coadjoint orbits of SU(1, 1) and P+ (1, 1). The limit (10.129) is just
a (rather exotic) parametrization of the cylindric coadjoint orbit of the Poincar
group {(a, k0 , k) : k02 k2 = 1, k0 > 0}, which is interpreted as the phase space for
a massive free particle in Minkowski spacetime.
The singular behavior of the GilmorePerelomov coherent states originates in the
choice of the basis function u0 (z) = 1 in H , the image of which, namely (1x2 )1/ ,
goes to zero as 0. The same holds for any element un (z) of the orthonormal
basis (4.93). It is thus necessary to choose some admissible vector in H such
that the family

g = U (g) , g SU(1, 1), (10.130)

298 10 CS of the Relativity Groups

contracts to a Poincar counterpart

(a, ) = U(a, ) . (10.131)

Among them it may be possible to select a subfamily of coherent states through a

suitable choice of section (a, ) (q, p), as in (10.66). Let us try, for instance, to
recover in the contraction limit the Gaussian vector G (k) exp(k0 ) mentioned in
the previous section. The vector G has many interesting properties from a quantum
as well as a statistical point of view [306]. One of them is related to the relativistic
Heisenberg inequality

1
P1 P2 P0 , (10.132)
2
where the infinitesimal generator P0 corresponds to the basis element f0 in (10.111)
in the representation U(a, ). Explicitly we have (in the x-parametrization)

1 + x2
P0 f (x) = f (x) (energy) (10.133)
1 x2
2x
P1 f (x) = f (x) (momentum P1 := P) (10.134)
1 x2
1 x2 d
P2 f (x) = i f (x) (Lorentz boost P2 := K). (10.135)
2 dx
The inequality (10.132) is saturated precisely when computed in the state G . This
is related to the fact that the latter is annihilated by the operator P2 + iP1 :

(P2 + iP1 ) G = 0, (10.136)

which should be compared with the Galilean (or WeylHeisenberg) analogue (2.15).
The Anti-de Sitterian counterpart of Pi is Ki :

d
Ki f = i U (e ei )| =0 . (10.137)
d
So it is natural to make a parallel with (10.136) and look for solutions to equations
of the type

(AK2 + iK1 ) = , (10.138)

with A > 0, C and H . These solutions saturate the following Heisenberg

inequality [485]:

K1 K2 K0 , (10.139)
2
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 299

if A = K2 / K1 . The set of solutions of (10.138) splits into three classes, according

to the values taken by A.
(i) A > 1:

$  % 1 (1i / A2 1) $  % 1 (1+i / A2 1)
A+1 A+1
(z)=c z+i zi ,
A1 A1
(10.140)

where c is a normalization constant.

(ii) A = 1:

1 z
(z) = c e . (10.141)

These states, indexed by = 1 C, are the BarutGirardello (BG) CS

[126], in their FockBargmann representation on the disk, i.e., as elements of
H . The alternative expression is

n
&n! (n + 2/ ) |un ,
1/2
|  |b  = n (10.142)
n=0

where the normalization factor n is given in terms of the modified Bessel

function I by

n = | |12/ I2/ 1 (2| |). (10.143)

Note that the normalization factor c in (10.140) is given by

c = [n (2/ )]1/2 . (10.144)

The BG states yield the following resolution of the identity


I= |b  b | dBG ( ), (10.145)
C

with the measure given by

1
dBG ( ) = | | I2/ 1 (2| |) K2/ 1 (2| |) d(| |2 ) d(arg ). (10.146)

The occurrence in this expression of the modified Bessel function of the third
kind, K , stems from the classical moment problem

(n + 1) (n + 2/ ) = un (u) du, (10.147)
0
300 10 CS of the Relativity Groups

with

(u) = 2u1/ K2/ 1 (2 u). (10.148)

We emphasize that the BG states do not contract correctly for any C, since
the limit as 0 of the expression

I (x) = [n 1 (2/ )]1/2 (1 x2 )1/ ei x/ (10.149)

is not square integrable, although it may exist as a distribution supported on a

set of measure zero.
(iii) 0 < A < 1:

$  % 1(1 / A2 1) $  % 1 (1+ / A2 1)
1+A 1+A
(z)=c z + z .
1A 1A
(10.150)

It is interesting to note that similar distribution-valued CS appear in [496],

in a different context, and also in [589]. The states derived in the latter paper,
called generalized intelligent states, are obtained as solutions of an equation
similar to (10.138), and contain both GilmorePerelomov and BarutGirardello CS.
We also remark that SU(1,1) CS similar to, but different from the BarutGirardello
CS, have been constructed by Hongoh [386].
In view of the effect of the contraction, only the states with nonzero values
concentrated around the imaginary axis will be relevant in the present context. Now,
the classical observables ki corresponding to the Ki read [315]:
1 + zz zz z+z
k0 = , k1 = i , k2 = . (10.151)
1 zz 1 zz 1 zz
Hence, k2 must vanish in the limit 0 and this condition in turn entails K2 0
at the quantum level. In view of the equality

A2 ( K2 )2 + ( K1 )2 = A K0 , (10.152)

that follows from (10.138), it seems natural to put A = 1/ . Finally, by choosing the
value 0 for the unessential parameter , we are left with the binomial state
 
1 + 1/
b (z) = z +
2
. (10.153)
1
It is then easily checked that this state contracts to G (up to a factor):


b = I b , lim 
b (x) = e2/(1x ) b(x) G (k).
2
(10.154)
0
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 301

The square integrability of the representation then allows one to consider the family
of SU(1, 1)-transported vectors

b g = U (g)b , g SU(1, 1), (10.155)

as a (tight) frame, in the spirit of Theorem 8.1.3 (Sect. 8.1):


1
I= |b g  b g | d (g), (10.156)
c(b ) SU(1,1)

where
1
d (g) = (1 |z|2 )2 d2 z d
2 2

is the Haar measure on SU(1, 1) expressed in terms of the factorization (4.66),

g = (z) k( ), z D, [0, 2 ). (10.157)

The constant c(b ) is the one defined in (8.8).

The important point concerning the CS g is their behavior under contraction.
A precise answer is given by the following results [519].
Theorem 10.3.3. For fixed (a, ), one has

lim I U ( (a, ))b = U(a, )b,

0

in the sense of the norm topology on H:

lim I U ( (a, ))b U(a, )bH = 0, for g = (a, ) P+ (1, 1). (10.158)
0

Note that (10.158) is a significant generalization of the contraction la Dooley (see

Sect. 4.5.4), in the sense that it involves states b which depend on the deformation
parameter , as it should be in the generic case.
The proof of Theorem 10.3.3 partially rests upon the following limiting argu-
ment, which is interesting in itself.
Lemma 10.3.4. Let f be in H. Suppose that the family { f } satisfies the following
two conditions:
lim 0 f = f , almost everywhere on (1, 1),
f (x)/(1 x2 ) is bounded on (1, 1), uniformly in .
Then

lim  f f H = 0. (10.159)
0
302 10 CS of the Relativity Groups

Theorem 10.3.3 in particular guarantees that the subfamily of {b g , g SU(1, 1)}

obtained by restricting g SU(1, 1) to (Cartan) sections (4.66),

b (z) = U ( (z))b , g SU(1, 1), (10.160)

contracts in our sense to Poincar coherent states of the form (10.69), defined
precisely with the symmetric section (see Sect. 10.1.2)

(q, p)  s (q, p) = (( ), p),

q0 , q (10.161)

where
qp
q0 = p0 , 
q = p0 q, 
p = p. (10.162)
p0 + 1

More precisely, we have (see [26] for detailed calculations)

(q, p) = (z( , ) + O( )),

q, p (10.163)

with the notations (10.116) and (10.128).

Note that (10.162) implies the spacetime relation


pq
q0 = , (10.164)
p0 + 1

which has many interesting features from a kinematical point of view in

Minkowskian geometry [26, 306].
More general Poincar sections of the type (10.66) can be reached through
contraction, that is, lim 0 1 , starting from sections

(z) = (z ) k( ), (10.165)

where
 
z
s (z) = , = (1 |z|2 )1/2 , (10.166)
z
 i /2 
e 0
k( ) = U(1), (10.167)
0 ei /2

and : D R is a C function depending on in a smooth way. This section

leads to a new parametrization of the open unit disk D analogous to a gauge
transformation

z  z = z ( , ) := ei (z,z, ) z(
q, p , ),
q, p 
q = q. (10.168)
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 303

If one makes a more precise assumption for the asymptotic behavior of :

(z, z, ) := ( , ) = 0 (
q, p p) + 
q 1 (k) + o( ), (10.169)

then the resulting Poincar section is of affine type

), p ) = lim 1 (z) := (q, p)
q0 , q
((
0
(10.170)
= ((0, q), p )(( (p) + q #(p), 0), I),
where

q p
q0 = + p) + 0 (
q 1 ( p). (10.171)
p0 + 1
Hence, comparing with (10.17) and (10.19), and taking into account the fact that
is not equal to zero, we get

p
1 (
p) = (p) p + 1
0
(10.172)
0 (
p) = (p)[p0 p (p)].
A last point relative to the present discussion on sections in the two groups
SU(1, 1) and P+ (1, 1) is the following question. Let the coherent states (q,p)
be precisely those obtained by contraction from the CS b (z) given in (10.160).
In that case, how does the resolution of the identity (10.156) in SU(1, 1) behave
under precontraction and contraction, and similarly for the Poincar resolution
operator (10.70) under the action of I1 , when 0? This important question
deserves attention and no precise answer can be given at the moment.
Let us end this section by quoting another important result of [519] concerning
contractions of (essentially) self-adjoint operators (called observables in a quantum
context) on H . Let B be an observable on H such that b g D(B ) and let B be
its precontracted partner:

B = I B I1 . (10.173)

Then we say that B contracts to an observable B acting in H iff

lim 
b g |B |
b g H = bg |B|bg H , g, g P+ (1, 1). (10.174)
0

Let p = p(K  , K
 , K  , ) be a polynomial of the variables K
i . We have:
0 1 2

Theorem 10.3.5. For any g, g P+ (1, 1), the following limit holds true:

lim  0 , K
b g |p(K 2 , )|
1 , K b g  = bg |p(P0 , P1 , P2 , 0)|bg . (10.175)
0
304 10 CS of the Relativity Groups

As a consequence, we have
Proposition 10.3.6. Let f H be of the form
 a 1 +b+O( )
d( ) 1
f (x) = P(x) 1 + , (10.176)
1 x2 (1 x2 )no

where P is a polynomial, a > 0, no N and d is a nonnegative function defined

for > 0, such that 0 < lim 0 d( ) < . Then p(K  , K
 , K
 , ) f tends in the
0 1 2
H-norm to p(P0 , P1 , P2 , 0) f , where f is the limit of f .
In conclusion, Anti-de Sitterian quantum observables, possibly with the exception of
some pathological ones, contract to their Poincar counterparts, and this contraction
procedure involves covariant symbols on both sides. We should note, however, that
we make use in H of a continuous family {bg } bigger than the set of Poincar
coherent states {b (q,p) }.
Finally, similar contraction results can be found with quantum probes b
more general than (10.153), i.e., probes which do not necessarily saturate some
Heisenberg inequality.
Chapter 11
Integral Quantization

Abstract In this chapter, we present an original approach to quantization based

on operator-valued measures, under the generic name of integral quantization. The
so-called Berezin-Klauder-Toeplitz quantization, and more generally coherent states
quantization are particular (and mostly manageable) cases of this approach, and
we work out some illuminating examples of them. We particularly insist on the
probabilistic aspects appearing at each stage of our quantization procedure.

In this chapter, we present an original approach to quantization based on operator-

valued measures, under the generic name of integral quantization. The so-called
Berezin or Klauder or yet Toeplitz quantization, and more generally coherent states
quantization are particular (and mostly manageable) cases of this approach, and
we work out some illuminating examples of them. We particularly insist on the
probabilistic aspects appearing at each stage of our quantization procedure.
Motivated by the five-fold symmetry of sea stars (i.e., starfish), we develop in
a comprehensive way the quantization of functions on a set with five elements,
which yields in particular the notion of a quantum angle. We eventually extend
the method to the unit circle, introducing in this way the general presentation of
integral quantization, including in particular the one based on coherent states built
from group representation in the spirit of this book. Then we particularize this
material to the Weyl-Heisenberg group underlying the canonical commutation rule,
a paradigm of quantum physics. We show that there is a world of quantizations
that follow this rule. In addition, we enlarge the set of objects to be quantized in
order to include singular functions or distributions. The example of the quantum
phase obtained in this way is illuminating. Finally we present a general Bayesian
construction of action-angle coherent states in view of quantizing classical phase
spaces for bounded motions.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 305
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__11,
Springer Science+Business Media New York 2014
306 11 Integral Quantization

11.1 What is Really Quantization?

In digital signal processing, one generally understands by quantization the process

of mapping a large set of input values to a smaller set such as rounding values to
some unit of precision. A device or algorithmic function that performs quantization
is called a quantizer. In physics or mathematics, the term has a different meaning.
For instance, one can find the following definition in the Oxford Dictionary:
Quantize: form into quanta, in particular restrict the number of possible values of
(a quantity) or states of (a system) so that certain variables can assume only certain discrete
magnitudes,

or the following, less restricted ones, which in fact are more to the point for our
purposes here,
We understand (quantization) in the sense first sketched by Weyl [Wey28], where it describes
a harmonic analysis". It consists in Fourier analyzing a (fairly arbitrary) function on
phase space, and then replacing the elementary building blocks" (i.e. exponentials on
phase space) by appropriate operators [229].
Quantization can be any procedure that associates a quantum mechanical observable to a
given classical dynamical variable [408].

The standard (canonical) construction is well known. It is based on the replace-

ment, in the expression of classical observables f (q, p), of the conjugate variables
(q, p) (i.e., obeying {q, p} = 1) by their respective operator counterparts (Q, P).
The latter obey the canonical commutation rule [Q, P] = ihI. The substitution has
to be followed by a symmetrization. This last step is avoided by the integral
version of this method, namely the Weyl or phase-space quantization (see for
instance [Zac06] and references therein). The canonical procedure is universally
accepted in view of its numerous experimental validations, one of the most famous
and simplest one going back to the early period of Quantum Mechanics with
the quantitative prediction of the isotopic effect in vibrational spectra of diatomic
molecules [Her89]. These data validated the canonical quantization, contrary to the
Bohr-Sommerfeld Ansatz (which predicts no isotopic effect). Nevertheless this does
not prove that another method of quantization fails to yield the same prediction (see
for instance [143]). Moreover, canonical or Weyl quantization is difficult, if not
impossible, to implement when f is singular, or when barriers or other impassable
boundaries are involved.
To be more mathematically precise, yet remaining at an elementary or minimal
level, let us adopt the following.
Definition 11.1.1. Quantization is a linear map Q : C (X)  (H) from a vector
space C (X) of complex-valued functions f (x) on a set X to a vector space (H) of
linear operators Q( f ) A f in some complex Hilbert space H such that
(i) To the function f = 1 there corresponds the identity operator I on H,
(ii) To a real function f C (X) there corresponds a(n) (essentially) self-adjoint
operator A f in H.
11.2 Sea Star Algebra 307

Physics adds into the game further requirements, depending on various mathemat-
ical structures put on X and C (X) such as a measure, a topology, a manifold,
closure under algebraic operations, etc. together with an interpretation in terms of
measurements of the elements of C (X) (resp. (H)), to which are given the status
of classical (resp. quantum) physical quantities or observables. In particular, the
spectral properties of elements of (H) are at the heart of the quantum measurement
protocols. Moreover, one usually requires an unambiguous classical limit of the
quantum physical quantities, the limit operation being associated to a change of
scale. It is precisely at this point that both definitions of quantization might be
related.
A warning is in order at this point. According to the standard (von Neumanns)
interpretation, every self-adjoint operator represents an observable, but this is
untenable. On the mathematical side, observables are most often represented by
unbounded operators and there will be in general no dense domain common to all
of them. On the physical side, a self-adjoint operator may represent an observable
for one system, but not for another one. Think (following Wigner) to the meaning
of the operator p2 + q2 for the hydrogen atom . . .
This leads to the notion of labeled observable introduced by Roberts [522, 523].
In a nutshell, the idea is to characterize a quantum system, with Hilbert space H, by
a family of distinguished (labeled) observables, which are given both a physical
interpretation (how does one measure it?) and a mathematical definition (as a self-
adjoint operator in H). In addition, these are required to have a common dense
invariant domain. This is the basis of the Rigged Hilbert Space formulation of
quantum mechanics, developed in [4648, 159, 522, 523] (see also the related
approach of [138]).

11.2 Sea Star Algebra

11.2.1 Exploring the Plane with a Five-Fold Set of Arms

The purpose of this introductory model is to enhance our intuition that quantization
can be viewed as the analysis of a set from the point of view of a family of
coherent states, i.e., a total, not necessarily linearly independent, family of vectors
in a Hilbert space resolving the identity, i.e., forming an overcomplete family in
the language of physicists. Let us ask ourselves how the ocean floor is five-fold
orientationally explored by a starfish (or seastar) (Fig.11.1)1 Starfish possibly do

1 A marine echinoderm with five radiating arms. The undersides of the arms bear tube feet for

locomotion and, in predatory species, for opening the shells of mollusks. On the end of each arm
or ray there is a microscopic eye which allows the sea star to see, although it only allows it to
distinguish between light and darkness, which is useful for detecting movements. Very recently
seven-fold sea stars have been observed in Antarctic deep-sea hydrothermal vents [205].
308 11 Integral Quantization

Fig. 11.1 A sea star as a

five-fold frame for the plane

this in a noncommutative way through the pentagonal set of unit vectors (the arms)
forming a five-fold frame in the plane:
<  
2n 2n
=R |0 coherent state n = 0, 1, 2, 3, 4 mod 5 .
5 5

with
 
cos sin
R( ) := SO(2) .
sin cos

Denoting the orthogonal projection on unit vector in the plane with polar angle by
 
cos2 cos sin
|  | = , (11.1)
cos sin sin2

we get the resolution of the identity:

<;  
2 4 2 n 2 n 10
5
5 n=0 5
=
01
I (11.2)

Actually this property holds for any regular N-fold polygon in the plane.
<;  
2 N1 2 n 2 n 10
N
N n=0 N
=
01
11.2 Sea Star Algebra 309

This resolution of the identity remains valid in the continuum limit N , 0,

2/N / :
 2  
1 10
|   | d = (11.3)
0 01

Going back to the starfish, let us make more precise the set X explored by it, a
probabilistic construction of the frame by using the Hilbert space structure of the
plane, and the exact way of exploring X through functions on this finite set.
The starfish senses its possible orientations via the five angles 2 n/5, i.e.,
X = {0, 1, 2, 3, 4}, and equipped with the discrete measure:

2 4
X
f (x) d (x) = f (n) ,
5 n=0
(11.4)

Let us choose two orthonormal elements, 0 (n) = cos(2 n/5) and 1 (n) =
sin(2 n/5), in the Hilbert space L2 (X, ), and build the five unit vectors (the
starfish CS!) in the real two-dimensional= Hilbert, i.e., the Euclidean plane R2 H,
with the usual orthonormal basis |0 , 2 :
<
2 n :
X % n  |n = 0 (n)|0 + 1 (n) . (11.5)
5 2

There are two probabilistic aspects in this construction, interlacing each other
in a sort of Bayesian duality [34]. For each n, {0, 1} % k  |k (n)|2 represents
the probability of detection of the direction k /2 when the orientation is actually
2 n/5. On the other hand, for each k, n  |k (n)|2 is a distribution probability on
the set X equipped with the discrete measure with uniform weight 2/5.
The five unit vectors (11.5) resolve the identity in H. They form a finite unit
frame for analyzing the complex-valued functions n  f (n) on X through what we
call a coherent state (CS) quantization [214, 216, Gaz09]:
 <;
2 4 2n 2n
f (n)  |xx| f (x) d (x) = f (n)
X 5 n=0 5 5
 

:= A f = M(2, C) . (11.6)

The map (11.6) is surjective from C5 to C3 . As a consequence, some information is

lost, corresponding to the existence of the complex two-dimensional kernel
 
2 4 1 4
Ke = f C5 : f14 = f23 , f14+ = 2 f 23+ , f (0) = 2 f23+ ,
4 2 1 4 2 1
310 11 Integral Quantization

where f14 = 12 ( f (1) f (4)), f23 = 12 ( f (2) f (3)), and = 12 (1 + 5) =
2 cos /5 is the golden mean which satisfies 2 = +1. This map should be analyzed
more precisely through spectral values of the 2 2 symmetric matrix A f , and also
through CS mean values or lower symbols of A f :
; <
2n 2n 2 4 2(n n )
n  f(n) :=
5
Af
5
=
5 n =0
f (n ) cos2
5
. (11.7)

Note that f is the average of f with respect to the distribution |k (n)|2 on X

mentioned above. We expect that the new function f(n) will be more regular than
f , since its dependence on n is carried by cosine functions.
It should be added that if we had chosen instead as a finite frame the orthonormal
basis |0, | /2, in R2 , we would have obtained the trivial commutative quantiza-
tion:

( f (0), f (1))  A f = diag( f (0), f (1)) .

Let us apply this formalism to the classical angle function (n) = 2 n/5 mod 2 .
Its quantum version is given by
<; $ %
2 4 2n 2n 2n 3 
(n)  A = = , (11.8)
5 n=0 5 5 5 5  5

where  = 15 (3 )3/2 0 325. By construction, this operator is covariant with

respect to rotations R(2 n /5), n = 0, 1, 2, 3, 4 mod 5:

R(2 n /5)A R(2 n /5) = A+ 2n mod 2 ,

5

More generally, given f = (a0 , a1 , a2 , a3 , a4 ), i.e., f (n) = an and extended periodi-

cally mod 5, the operator A f is covariant in the following sense:

R(2 n /5)A f R(2 n /5) = AT (2 n /5) f (11.9)

where

T (2 n /5) f (n) = f (n n mod 5) .

The eigenvalues of this starfish quantum angle are given by

 1  1 01

= 4 1 + 2 2 95 .
5
5
11.2 Sea Star Algebra 311

They correspond to eigenvectors |9 /20 and |19 /20 respectively. Note here the
rationality of the factors of despite that eigenvalues are algebraic irrational factors
of . The lower symbol of the quantum angle is the following function on X:
3 4.31 3.39 2.46 4.14
(0) = , (1) = , (2) = , (3) = , (4) = ,
5 5 5 5 5
We observe that its values oscillate around 4 /5, which is the mean value of the
two eigenvalues and which is also the average of the original angle function in the
following classical sense:

1 4 1
 f class :=
5 n=0
f (n) = tr(A f ) .
2

It is proved in [214] that  f class is the limit which is reached after infinitely iterated
maps f  f  f  . . . .
We ignore considerations of whether this formalism has some meaning for the
biomechanics of the sea star or whether its sensory perceptions are quantum in
nature!

11.2.2 What About the Continuous Frame?

It is interesting to implement CS quantization by using the continuous set of unit

vectors |  in the plane which obeys (11.3). The quantization of a complex-valued
function on the circle yields the symmetric complex matrix:

1 2
f ( ) |  | f ( ) d
0
  (11.10)
 f class + C ( f ) S (f)
= := A f ,
S(f)  f class C ( f )
where
 2
1
 f class = f ( ) d ,
2 0
 2
1
C(f) = f ( ) cos 2 d ,
2 0
 2
1
S(f) = f ( ) sin 2 d .
2 0

The lower symbol is obtained from f through the integral transform:

 2
1
f( ) =  |A f |  = cos2 (  ) f (  ) d  . (11.11)
0
312 11 Integral Quantization

Thus, the quantum version of the classical angle function reads as

 
12
 A = . (11.12)
12

Its eigenvalues are = 1/2 corresponding to eigenvectors | /4, |3 /4

respectively. The corresponding lower symbol
sin 2
 |A |  = (11.13)
2
is a loose smooth regularization of the angle function which varies between the two
eigenvalues of A .

11.2.3 Probabilistic Aspects

Behind the resolution of the identity (11.2) lies an interpretation in terms of

geometrical probability, which was described in [Gaz09] and which could prove
interesting for understanding the quantum starfish model. Consider a subset X
(finite case) or a Borel set [0, 2 ) and the quantities
<; 
2 2 n 2 n 1
a( ) = or = |  | d . (11.14)
N n N N

One easily checks:

/ = 0,
a(0) a(X or [0, 2 )) = I
a(iJ i ) = a(i ) , if i j = 0/ for all i = j . (11.15)
iJ

The map  a( ) defines a normalized measure on the set of subsets of X or on

the -algebra of the Borel sets in the interval [0, 2 ), taking its values in the set of
positive linear operators on the Euclidean plane, namely, it is a POVM.

11.2.4 Beyond Coherent States: Integral Quantization

The above quantization procedure is based on coherent states |  R( )|0.

  P0 in P := |  | = R( )P0 R( ) by
Actually we can replace the projection
ab
a 2 2 symmetric matrix M = . By Schurs Lemma 4.3.1, we still have the
bd
resolution of the identity:
 2
1
R( )MR( ) d = I .
(a + d) 0
11.3 Integral Quantization 313

 
a b
In particular, with a density matrix M = := , 0 a 1 and det =
b 1a
&
a (a2 + b2 ) 0, i.e., |b| a(1 a), we obtain a POVM. The corresponding
quantum angle then reads as
 
b 12 a
A = 1 ,
2 a b
 
with eigenvalues b 12 a and corresponding eigenvectors | /4, |3 /4.
We exploit in the next section these elementary ideas and results.

11.3 Integral Quantization

The conditions (i) and (ii) given in Definition 11.1.1 and obeyed in the above
examples are fulfilled more generally if we use the resources offered by the pair
(measure, integration). Note that such an approach was followed in a more specific
context by authors like those of [228230, 348, 408].
Let (X, ) be a measure space. Let H be a complex Hilbert space and X %
x  M(x) L (H) an X-labelled family of bounded operators on H resolving the
identity:

M(x) d (x) = I , (11.16)
X

provided that the equality is valid in a weak sense, which implies -integrability for
the family M(x). If the operators M(x) are positive and have unit trace, they will be
preferentially denoted by M(x) = (x) in order to comply with the usual notation
for a density matrix in quantum mechanics.
The corresponding quantization of complex-valued functions f (x) on X is then
defined by the linear map:

f  A f = M(x) f (x) d (x) . (11.17)
X

This operator-valued integral is again understood in the weak sense, i.e., as the
sesquilinear form,

B f (1 , 2 ) = 1 |M(x)|2  f (x) d (x) . (11.18)
X

The form B f is assumed to be defined on a dense subspace of H. If f is real and at

least semi-bounded, the Friedrichs extension [Ree80, Thm. X.23] of B f univocally
defines a self-adjoint operator. However, if f is not semi-bounded, there is no natural
314 11 Integral Quantization

choice of a self-adjoint operator associated with B f . In this case, we can consider

directly the symmetric operator A f , enabling us to obtain a self-adjoint extension
(unique for particular operators). The question of what is the class of operators that
may be so represented is a subtle one (see for instance [141, 142]).
If M(x) = (x) and given a positive unit trace operator 0 , we obtain the exact
classical-like formula

tr(0 A f ) = f (x) (x) d (x) , (11.19)
X

where (x) = tr(0 (x)), because of (11.16), is a true classical probability density
on X. More generally, if we have at our disposal another (it could be the same)
X-labelled family of positive unit trace operators X % x  (x) L + (H), the latter
may be associated with the following map to the classical element in C (X)

A f  f(x) := tr((x) (x )) f (x ) d (x ) , (11.20)
X

provided that the integral is defined. Note the map f  f is a generalization of the
Segal-Bargmann transform (7.64). It is at this point that the stated classical limit can
be given its full meaning, given a scale parameter and a distance d( f , f):

d( f , f) 0 as 0. (11.21)

One of the interesting aspects of this integral quantization scheme is precisely to

analyze the quantization issues, particularly the spectral properties of operators A f ,
from functional properties of the lower symbol f.
Another appealing aspect is the possibility to quantize constraints. Suppose that
the measure set (X, ) is also a smooth manifold of dimension n, on which is defined
the space D  (X) of distributions as the topological dual of the (LF)-space cn (X) of
compactly supported n-forms on X [347]. Some of these distributions, e.g. (u(x)),
express geometrical constraints. Extending the map (11.17) to these objects yields
the quantum version A (u(x)) of these constraints.
A different starting point for quantizing constraints, more in Diracs spirit [Dir01]
and intensively employed in (Loop) Quantum Gravity and Quantum Cosmology
(see [161,250,401,541,606] and references therein) would consist in quantizing the
function u  Au and determining the kernel of the operator Au . The two methods
are obviously not equivalent, except for a few cases.

11.3.1 Covariant Integral Quantization

Lie group representation theory offers a wide range of possibilities for building
explicit integral quantizations. Let G be a Lie group with left Haar measure d (g),
11.3 Integral Quantization 315

and let g  U(g) be a unitary irreducible representation of G in a Hilbert space H.

Let M a bounded operator on H. Suppose that the operator

R := M(g) d (g) , M(g) := U(g) MU(g) , (11.22)
G

is defined in a weak sense. From the left invariance of d (g) we have


U(g0 ) RU(g0 ) = d (g) M(g0 g) = R
G

and so R commutes with all operators U(g), g G. Thus Schurs Lemma 4.3.1
guarantees that R is a multiple of the identity, R = cM I. The constant cM is given by

cM = tr (0 M(g)) d (g) , (11.23)
G

where the unit trace positive operator 0 is suitably chosen in order to make the
integral convergent. Incorporating this constant into the measure, we obtain the
resolution of the identity provided by the family of operators U(g) MU(g) :

M(g) d (g) = I , d (g) := c1
M d (g). (11.24)
G

For instance, in the case of a square-integrable UIR

>
U for which |  is an admissible
unit vector, with M = |  | and cM c( ) = G d (g) | |U(g) |2 , we recover
the resolution of the identity provided by the family of coherent states |g  =
U(g)|  defined in Sect. 8.1.
This construction provides
>
an integral quantization of complex-valued functions
on the group f  A f = G M(g) f (g) d (g). Moreover, this quantization is covariant
in the sense that

U(g)A f U(g) = AU (g) f , (11.25)

where (U (g) f )(g ) := f (g1 g ) is the left regular representation (4.105) if f
L2 (G, d (g)).
If M = , with the most immediate choice  = in (11.20), we get the
generalization of the Berezin or heat kernel transform on G,

f(g) := tr( (g) (g )) f (g ) d (g ).
G

Within the same group-theoretical framework, the definition, given in Sect. 7.1,
of square-integrable covariant coherent states with respect to a left coset manifold
X = G/H equipped with a quasi-invariant measure opens a world of possible
covariant quantizations. With the notations of that section, for a global Borel section
316 11 Integral Quantization

: X G of the group, let be the unique quasi-invariant measure defined

in (4.19). Let U be a UIR which is square integrable mod(H) and n vectors i which
are admissible mod(H). With F = ni=1 | i  i |, F (x) = U( (x))FU( (x)) , we
have the resolution of the identity

1
F (x) d (x) = I, (11.26)
cF X

the integral
>
converging weakly, if tr F = 1. The constant cF is given by the integral
cF = X tr (F F (x)) d (x). Consider now the sections g : X G, g G, which
are covariant translates of under g:

g (x) = g (g1 x) = (x)h(g, g1 x), (11.27)

where h is the cocycle defined in (4.37). Let g be the measure dg (x) :=

(g (x), x) d , again constructed using (4.19), and Fg (x) = U(g (x))FU(g (x)) .
If U is square integrable mod(H, ), there is a general covariance property enjoyed
not only by the positive operators
>
(7.5) but also, when they are properly defined,
by the operators A f = cF 1 X F (x) f (x) d (x), i.e., the quantized versions of
functions f (x) on X:

1
U(g)A f U(g) = AUg(g) f , A f g := Fg (x) f (x)d g (x).
l cF X

Of course, it is possible to establish similar results by replacing the operator F by

a more general bounded operator M provided integrability and weak convergence
hold in the above expressions.

11.4 Exemple: Weyl-Heisenberg Covariant Integral

Quantization(s)

11.4.1 Quantization with a Generic Weight Function

The formalism and the quantization procedure introduced above allow to enlarge
considerably, in the case of the standard quantum mechanics, the Berezin-Klauder-
Toeplitz quantization which has been sketched in the end of Sect. 6.3. With the
notations of Sect. 2.2, the group G is the Weyl-Heisenberg group GWH , H is the
phase subgroup and the left coset space X = GWH / is here identified with C, a
general element in it being denoted by z = (q + ip)/ 2. The invariant measure is
noted with the shorthand
dz dz dq dp
d2 z := = .
2i 2
11.4 Exemple: Weyl-Heisenberg Covariant Integral Quantization(s) 317

With the UIR U and the section defined in Eqs. (2.32) and (2.36) respectively, we
end with the unitary displacement operator D(z) = exp(za za) U( (q, p)) of
Eq. (2.80), where lowering and raising operators act in some separable Hilbert space
H with orthonormal basis {|en }, n N. A list of important properties, useful for
the sequel, of D(z) are given in Eqs. (2.81)(2.91).
Let (z) be a function on the complex plane obeying the condition

(0) = 1 , (11.28)

and defining a bounded operator M on H through the operator-valued integral


d2 z
M= (z)D(z) , (11.29)
C

Then, the family M(z) := D(z)MD(z) of operators displaced under the unitary
action D(z) resolves the identity

d2 z
M(z) = I. (11.30)
C
 
It is a direct consequence of the property D(z)D(z )D(z) = ezz zz D(z ), of the
formula

d2
ez z = 2 (z) , (11.31)
C

and of the condition (11.28) with D(0) = I. The resulting quantization map is
given by

d2 z
Af = M(z) f (z) . (11.32)
C

Using the symplectic Fourier transform defined according to Eq. (11.31) by


d2
f(z) = ez z f ( ) , (11.33)
C

we have the alternative expression


d2 z
Af = (z) D(z) f(z) , (11.34)
C

Formally, and in accordance with (11.31) and (11.33), the operator M is recovered
through the quantization of the Dirac distribution on C:

M = A (2) (z) . (11.35)

318 11 Integral Quantization

We will give this equality its full mathematical sense in Sect. 11.5.
The covariance property of the above quantization reads as

A f (zz0 ) = D(z0 )A f (z) D(z0 ) . (11.36)

Moreover we have successively

A f (z) = PA f (z) P, f (z) = (z), z , (11.37)

where P =
n=0 (1) |en en | is the parity operator defined in (2.85),
n

A f (z) = Af (z) , f (z) = (z), z , (11.38)

and

d2 z
tr(Af A f ) = | f (z)|2 , f | (z)| = 1, z . (11.39)
C

11.4.2 Regular and Isometric Quantizations

The quantization map (11.32) is said to be regular (in the sense that it yields the
canonical commutation rule [a, a ] = I), if the weight function verifies the two
conditions
(i) (z) = (z),
(ii) (z) = (z).
In that case we have

Az = a , A f (z) = Af (z) . (11.40)

Moreover, the quantization of the canonical variables is equivalent to the canonical

quantization:

a + a a a
Aq = := Q , Ap = := P . (11.41)
2 i 2

We say that the quantization map f  A f is isometric if | (z)| = 1. In that case


d2 z
tr(Af A f ) = | f (z)|2 , (11.42)
C

which means that the map f  A f is invertible (the inverse is given by a trace
formula).
11.4 Exemple: Weyl-Heisenberg Covariant Integral Quantization(s) 319

11.4.3 Elliptic Regular Quantizations

Clearly, the quantization map (11.32) includes the so-called normal (as a limit case),
anti-normal and Wigner-Weyl (i.e., canonical) quantizations [57,174,175,348] (see
also the interesting paper by Royer [548] and references therein). More precisely,
the quantization map f  A f is said to be elliptic regular if the weight function
satisfies (z) w(|z|2 ) with w : R  R (i.e., it leads to a regular quantization).
The normal, Wigner-Weyl and anti-normal (i.e., CS) quantizations correspond to a
specific choice of the weight function . Indeed let us choose
2 /2
s (z) = es|z| , Re s < 1 . (11.43)

Since this function is isotropic in the complex plane, the resulting operator M Ms
is diagonal. From the expression of the matrix elements of D(z) in (2.82) involving
associated Laguerre polynomials, and the integral [Mag66],

( + 1) ( + n + 1) 1
e x x Ln (x) dx = 1
2 F1 (n, + 1; + 1; ) ,
0 n! ( + 1)
(11.44)
we get the diagonal elements of Ms :
 n
2 s+1
en |Ms |en  = , (11.45)
1s s1

and so
  
d2 z 2 s+1
Ms = s (z)D(z) = exp ln a a . (11.46)
C i 1s s1

Then s = 1 corresponds to the CS (anti-normal) quantization, since

 
2 s+1
M1 = exp ln a a = |00| ,
1s s1 s=1

and so
 
d2 z d2 z
Af = D(z)MD(z) f (z) = |zz| f (z) . (11.47)
C C
The choice s = 0 corresponds to the Wigner-Weyl quantization since, from
Eq. (2.88), M = 2P, and so

d2 z
Af = D(z) 2P D(z) f (z) . (11.48)
C

The case s = 1 is the normal quantization in an asymptotic sense.

320 11 Integral Quantization

The parameter s was originally introduced by Cahill and Glauber in [175], where
they discuss the problem of expanding an arbitrary operator as an ordered power
series in a and a , a typical question encountered in quantum field theory, specially
in quantum optics. They associate with every complex number s a unique way of
ordering all products of these operators. We saw that normal ordering, antinormal
ordering (yielded by CS quantization), and symmetric ordering (yielded by Weyl
quantization) correspond to the values s = +1, s = 1, and s = 0 respectively.
Actually, Cahill and Glauber were not interested in the question of quantization
itself. They start from a symmetric operator-valued series A(a, a ) in terms of
powers of a and a , without considering its classical counterpart from which its
could be built from a given quantization procedure. They ask about mathematical
relations between the A(a, a ) As (a, a )s when a certain s-dependent order is
chosen. Nevertheless, their work allows one to give, to a certain extent, a unified
view of different quantizations of the functions on C.
Now, an important point favoring in particular the CS quantization is that
(from (11.45)) the operator Ms is positive and trace class for s 1 (and only
trace class if Re s < 0). In both cases its trace is equal to 1. This is an important
result since it proves that for any real s 1, Ms s is a density operator and its
associated quantization can be given a consistent probabilistic interpretation like in
(11.19). In that case the operator-valued measure

d2 z
C  D(z)Ms D(z) , (11.49)
B(C)

is a positive operator-valued measure. Furthermore, the form of (11.46) suggests

that for s 1 and given an elementary quantum energy, say h , the s-dependent
temperature introduced through

s+1 h h
ln = s = coth (11.50)
s1 kB T 2kB T

is intrinsically involved in the quantization procedure of the phase space C through

the Boltzmann-Planck like density operator
 h
 nh

s = 1 e kB T e kB T |en en | . (11.51)
n=0

Interestingly, the temperature-dependent operators s (z) = D(z) s D(z) defines

a Weyl-Heisenberg covariant family of POVMs on the phase space C, the null
temperature limit case being the POVM built from standard CS.
11.4 Exemple: Weyl-Heisenberg Covariant Integral Quantization(s) 321

11.4.4 Elliptic Regular Quantizations that Are Isometric

An elliptic regular quantization is isometric iff w(u) {1, +1}. A simple example
is given by the family :

(z) = 2 (1 |z|2 ) 1 (11.52)

where is the Heaviside function. The Wigner-Weyl quantization is a special case

( = 0).

11.4.5 Hyperbolic Regular Quantizations

We say that the quantization map f  A f is hyperbolic regular if the weight function
verifies (z) m(Im (z2 )) with m : R  R (i.e., it is a regular quantization).
In that case we have

A f (q) = f (Q) , A f (p) = f (P) . (11.53)

Hence, one notices that the Wigner-Weyl (i.e. canonical) quantization (m(u) = 1) is
not the only one providing these relations.

11.4.6 Hyperbolic Regular Quantizations that Are Isometric

An hyperbolic regular quantization is isometric iff m(u) {1, +1}. A simple

example is given by the family :

(z) = 2 (1 Im (z2 )) 1 (11.54)

The Wigner-Weyl quantization is a special case ( = 0). The above relations imply
that all the main properties of the Wigner-Weyl quantization scheme are verified in
that case.

11.4.7 Quantum Harmonic Oscillator According to

Given a general weight function , the quantization of the classical harmonic

oscillator energy (in suitable units) |z|2 = (p2 + q2 )/2 yields the operator

(0)
A|z|2 = (0)a a + z |z=0 a z |z=0 a + z z |z=0 . (11.55)
2
322 11 Integral Quantization

In the case of a regular quantization, we obtain the operator

1
A|z|2 = a a + z z |z=0 . (11.56)
2
Furthermore,

1! 2 "
Aq2 = Q2 z z |z=0 + + z2 (11.57)
2 z z=0 z=0

1! 2 "
A p2 = P2 z z |z=0 + z2 (11.58)
2 z z=0 z=0

One observes that the difference between the ground state energy of the quantum
harmonic oscillator, namely E0 = 1/2 z z |z=0 , and the minimum of the
quantum potential energy, namely Em = [min(Aq2 ) + min(A p2 )]/2 = z z |z=0 is
independent of the particular (regular) quantization chosen, namely E0 Em = 1/2.
In the exponential Cahill-Glauber case s (z) = es|z| /2 the above operators
2

reduce to
1s s s
A|z|2 = a a + , Aq2 = Q2 , A p2 = P2 . (11.59)
2 2 2
It has been proven in [143] that these constant shifts in energy are inaccessible to
measurement.

11.4.8 Variations on the Wigner Function

In quantum mechanics, there are several transforms of the wave function or

in general of the density operator which provide a phase space portrait of the
quantum state. The Wigner function (also called the Wigner function or the Wigner
Ville distribution [344, 602, 607, Wey28, Wey31, Zac06] is the one that has found
more applications, mainly in statistical mechanics, but also in quantum chemistry,
molecular dynamics, scattering theory, or quantum optics. It is a quasiprobability
distribution introduced by Eugene Wigner in 1932 to study quantum corrections to
classical statistical mechanics. The goal was to link the wavefunction that appears
in Schrdingers equation to a probability distribution in phase space. Precisely,
because it might be negative in some regions of the phase space, the Wigner function
fails to be a probability distribution.
The Wigner function is (up to a constant factor) the Weyl transform of the
quantum-mechanical density operator. For a particle in one dimension it takes the
form (in units h = 1)
 + 9
1 y y : ipy
W(q, p) = q |q + e dy . (11.60)
2 2 2
11.5 Weyl-Heisenberg Integral Quantizations of Functions and Distributions 323

Adapting this definition to the present context, and given an operator A, the
corresponding Wigner function is defined as

WA (z) = tr (D(z)2PD(z) A) , (11.61)

In the case of the quantization map f  A f based on a weight function , we have


d2
WA f (z) =  ( z) f ( )
, (11.62)
C

which becomes in the case of Weyl quantization

WA f = f (11.63)

(this one-to-one correspondence of the Weyl quantization is related to the isometry

property).
In the case of the anti-normal quantization, the above convolution corresponds to
the Husimi transform (when f is the Wigner transform of a quantum pure state).
If the quantization map f  A f is regular and isometric, the corresponding
inverse map A  WA is given by

  
d2 z
WA (z) = tr D(z)MD(z) A , where M = M = (z)D(z) . (11.64)
C

In general this map A  WA is only the dual of the quantization map f  A f in the
sense that

d2 z
WA (z) f (z) = tr(A A f ) . (11.65)
C

This dual map becomes the inverse of the quantization map only in the case of a
Hilbertian isometry f L2 (C, d2 z/ )  A f LHS
2 (H) (Hilbert-Schmidt operators).

11.5 Weyl-Heisenberg Integral Quantizations of Functions

and Distributions

11.5.1 Acceptable Probes

When it is well defined, the map (11.20) appears as a very convenient tool to
characterize the class of acceptable operator-valued functions M(x) and the class of
functions f quantizable with respect to the latter. For the Weyl-Heisenberg integral
quantization, we only consider in this section the positive unit trace operators
324 11 Integral Quantization

(or probe) M = and, in particular, the examples = s in the range < s 1,

including the most manageable CS case. Accordingly, the mean value or lower
symbol of A f is defined by
   d2 z
f(z) = tr (z) (z ) f (z ) . (11.66)
C

In particular, the resolution of the identity guarantees that

   d2 z
tr (z) (z ) = 1. (11.67)
C
 
Thus, for each z, tr (z) (z ) is a probability distribution on the phase space, and
so f stems from the corresponding kernel averaging of the original f . In the CS case
(z) = |zz|, (11.66) is the Gaussian convolution (Berezin or heat kernel transform)
of the function f (z):

d2 z
e|zz | f (z )
 2
f(z) = z|A f |z = . (11.68)
C

An inescapable aspect of any quantization procedure is its classical limit. We should

expect that the lower symbol f of A f approximates f better and better in this limit.
Actually, as can be already observed in the Gaussian case, the limit operation is not
so straightforward. We first have to give the complex plane a phase-space meaning
after introducing physical units through

q p
z := + i , (11.69)
 2 h 2

where h is the Planck constant and  an arbitrary length scale. Secondly, a general
operator includes in its definition a set of physical constants, such a temperature, a
time scale, etc. (as in (11.51)). We thus adopt the following classicality requirement
on the choice of density operators .
Definition 11.5.1. A density operator is acceptable from the classical point of
view if
(i) The following limit holds true,
 
tr (z) (z ) (z z ) as h 0 ,  0 , h/ 0 , (11.70)

which implies a suitable h and  dependence of all other parameters involved in

the expression of .
(ii) Given an orthonormal basis {en }, the matrix elements en | (z)|en  are rapidly
decreasing C functions in z. This condition will appear natural for the
quantization of distributions.
11.5 Weyl-Heisenberg Integral Quantizations of Functions and Distributions 325

11.5.2 Quantizable Functions

Equation (11.68) illustrates nicely the regularizing role of quantum mechanics

versus classical singularities. Note also that the Gaussian convolution allows one to
carry out the semi-classical limit by using a saddle point approximation. For regular
functions for which A f exists, the application of the saddle point approximation is
trivial and  f f L1 0 holds true. For singular functions, the semi-classical limit
is less obvious and has to be verified for each particular case. Inspired by the CS
case (11.68) in which, with mild constraints imposed to f , the transform f inherits
infinite differentiability from the Gaussian, let us adopt now a second acceptance
criterion imposed on the function f to be quantized.
Definition 11.5.2. Given an acceptable density operator , a function C % z 
f (z) C is -quantizable along the map f  A f defined by (11.32), if the map
C % z = 12 (q + ip) (q, p)  f(z) is a C function with respect to the (q, p)
coordinates of the complex plane.
This definition is reasonable insofar as differentiability properties of (11.66)
are those of the displacement operator D(z). We will extend this definition to
distributions T D  (R2 ) in the next subsection. In the CS case, the fact that
the Berezin transform f  f is a Gaussian convolution is of great importance. It
explains the robustness of CS quantization, since it is well defined for a very large
class of non smooth functions and even, as will be shown below, for a class of
distributions including the tempered ones.

11.5.2.1 An Example: The Quantum Phase

When they are expressed interms of the action-angle variables J = |z|2 0 and
= arg(z) [0, ), i.e., z = J ei , for the harmonic oscillator (see [Gol81, Lan81]
and below), the standard coherent states read as action-angle CS encountered in
Sect. 6.4.4:
&
|z |J,  = pn (J)ein |en  , (11.71)
n

where n  pn (J) = eJ J n /n! is the Poisson distribution. The action variable is

precisely the Poisson average of the discrete variable n, npoisson = J. Note that in
electromagnetism, the variables J and represent the field intensity and the phase,
respectively. The CS quantization procedure yields the quantum versions of classical
functions f (J, ) which are 2 -periodic in the angle variable:
 +  2
1
f (J, )  dJ f (J, ) |J, J, | d = A f . (11.72)
0 2 0
326 11 Integral Quantization

Defining the unitary representation  U( ) of the unit circle S1 on CS by

U( )|J,  = ei |J, +  where is arbitrary real, we easily verify from (11.72)
the covariance property (compare with (11.9))

U( )A f U( ) = AT ( ) f , T ( ) f (J, ) = f (J, ) . (11.73)

Since we do not retain in our quantization scheme the condition of smoothness on

the classical observables, we feel free to CS quantize a larger class of functions
on the plane, like the 2 -periodic angle function ( ) = for [0, 2 ). Since this
function is real and bounded, its quantum counterpart A f yielded by CS quantization
is a bounded self-adjoint operator. In the present basis, it is given by the infinite
matrix:
!  "
n+n
2 +1 1
A = 1H + i 
|en en | . (11.74)
n=n

n!n ! n n

This operator has spectral measure with support [0, 2 ]. The corresponding lower
symbol reads as the Fourier sine series:
!  "
n+n + 1 
zn zn
J, |A| J,  = + i e
J 2

n=n
n!n ! n n
sin q
= 2 dq ( J) , (11.75)
q=1 q

where the function

q 
+ m + 1 2m+q
dq (r) = e r2
2
m!(m + q)!
r
m=0

( q2 + 1) q
= er rq
2
1 F1 ( + 1; q + 1; r )
2
(q + 1) 2
balances the trigonometric Fourier coefficient 2/q of the angle function . It can be
shown that this positive function is bounded by 1 [312].
Let us evaluate the asymptotic behavior of the function J, |A| J,  for small
and large J respectively. For
small J, it oscillates around its classical average value
with amplitude equal to J (compare with (11.13)):

J, |A| J,  J sin .

For large J, we recover the Fourier series of the 2 -periodic angle function:

1
J, |A| J,  2 q sin q = ( ) for [0, 2 ) .
q=1
11.5 Weyl-Heisenberg Integral Quantizations of Functions and Distributions 327

5
4

4 3

6 2
3 1.0

4
2
0.5
2 r
1

1 2 3 4 5 6 0 0.0

11.2 Lower symbol of the angle operator for J = {0.5, 1, 5} and [0, 2 ) (left) and
Fig.
for J, [0, 1] [0, 2 ) (right) (courtesy of A. Youssef)

This behavior is understood in terms of the classical limit of these quantum objects.
Indeed, by re-injecting physical dimensions into our formula, we know that the
quantity |z|2 = J acquires the dimension of an action and should appear in the
formulas as divided by the Planck constant h or h. Hence, the limit J in our
previous expressions can also be considered as the classical limit h 0. On the
left panel of Fig. 11.2, we show the behavior of (11.75) as a function of for
different values of J. One can observe how close to the classical behavior it becomes
for the largest value of J.
The number operator N  = a a is, up to a constant shift, the quantization of the

classical action, AJ = N + 1: AJ = n (n + 1)|en en |. Let us ask to what extent the
commutator of the action and angle operators and its lower symbol are close to
the canonical value, namely i.
! "
n+n
+1

2
[A , AJ ] = i |en en | , (11.76)
n=n n!n !

J, |[A , AJ ]|J,  = 2i dq ( J) cos q =: i C (J, ) . (11.77)
q=1

For small J, the

function C (J, ) oscillates around 0 with an amplitude equal to
J: C (J, ) J cos . Applying the Poisson summation formula, we get for
J (or h 0) the expected canonical behavior for [0, 2 ):

J, |[A , AJ ]|J,  i + 2 i ( 2 n) . (11.78)

nZ

One can observe that, for h 0, the commutator symbol becomes canonical for
= 2 n, n Z. Dirac singularities are located at the discontinuity points of the 2
328 11 Integral Quantization

periodic function ( ). The fact that the action-angle commutator is not canonical
(see [195] for a comprehensive discussion on this point) should not frighten us since,
on a more general level, we know that there exist such classical canonical pairs for
which mathematics (e.g. the Pauli theorem and its correct forms [296]) prevents the
corresponding quantum commutator from being exactly canonical.

11.5.3 Quantizable Distributions

While proceeding with a quantization scheme, specially with canonical quantiza-

tion, one usually imposes too restrictive conditions on the original function f (z).
Since it is viewed as a classical observable on the phase space, it is forced to belong
to the space of infinitely differentiable functions on R2 , essentially because of the
prerequisite Lagrangian and Hamiltonian structures. These conditions prevent us
from reaching a large class of operators, including elementary ones like nn :=
|en en |. Nevertheless, the latter have also a CS diagonal representation. For that,
it is enough to extend the class of quantizable objects to distributions on R2 (for
canonical coordinates (q, p)) or possibly on R+ [0, 2 ) (for polar coordinates
(r, )).
When we examine the matrix elements of the operator A f yielded by CS
quantization, 2

d2 z
f  A f = |zz| f (z)
C

1  d2 z
 |en en | n!n ! e|z| zn zn f (z)
2
=
n,n =0 C (11.79)

 
:=  Af nn
|en en | ,
n,n =0

one immediately thinks of tempered distributions on the plane as acceptable


objects. Indeed functions like n,n (z) = e|z| zn zn / n!n ! are rapidly decreasing
2

C functions on the plane with respect to the canonical coordinates (q, p), or
equivalently with respect to the coordinates (z, z): they belong to the Schwartz space
S (R2 ).
Using complex coordinates is clearly more convenient and we will adopt the
following definitions and notations for tempered distributions. First, any func-
tion f (z) which is slowly increasing and locally integrable with respect to the
Lebesgue measure d2 z on the plane defines a regular tempered distribution T f ,

2 The equalities hold in a weak sense, with the difficulties for unbounded operators already

mentioned.
11.5 Weyl-Heisenberg Integral Quantizations of Functions and Distributions 329

i.e., a continuous linear form on the vector space S (R2 ) equipped with the usual
topology of uniform convergence at each order of partial derivatives multiplied by a
polynomial of arbitrary degree [Sch61]. This definition rests on the map,

S (R2 ) %  T f ,  := f (z) (z) d2 z . (11.80)
C

For any tempered distribution T S  (R2 ), we define the quantization map T 

AT as

1
n,n
T  AT := T, n,n |en en | , (11.81)
 =0

where the convergence is assumed to hold in a weak sense. In the sequel the integral
notation will be kept, with the usual abuse of notation, for all (tempered or not)
distributions T :

d2 z 1
n,n
|zz| T (z) := T, n,n |en en | . (11.82)
C  =0

In this way, returning to the general scheme (11.32) with M a positive unit trace
operator, we formally define the quantization of a distribution T as

d2 z 1
n,n
T  AT = (z) T (z) := T, n,n |en en | , (11.83)
C  =0

where the n,n (z) = en | (z)|en  are assumed to belong to S (R2 ). The resulting
lower symbol is

d2 z 1
T (z) = tr( (z) (z )) T (z ) =  T, tr( (z) ())  . (11.84)
C

Hence we adopt the following extended definition of quantizable objects.

Definition 11.5.3. Given an acceptable density operator , a distribution T
D  (R2 ) is -quantizable along the map T  AT defined by (11.83) if the map
C % z = 12 (q + ip) (q, p)  T (z) is a C function with respect to the (q, p)
coordinates of the complex plane.
In the case of CS quantization, we easily check that such definitions are
mathematically justified for all tempered distributions. The following result allows
one to extend the set of such acceptable observables.
Proposition 11.5.4. A distribution T D  (R2 ) is CS quantizable if there exists
< 1 such that the product e |z| T S  (R2 ), i.e., it is a tempered distribution.
2
330 11 Integral Quantization

Note that similar extensions to distributions have been considered in [228230,570],

and [157,158] for the Weyl quantization. In the general case, one expects to have a
similar result with suitably chosen weight functions (z). In the sequel, we suppose
that such a choice has been made.
Thus, according to Proposition 11.5.4, the definition 11.5.3 can be extended to
locally integrable functions f (z) increasing like e |z| p(z) for some < 1 and some
2

polynomial p, and it is easily understood in which way this extends to distributions.

Actually, the latter can be characterized as derivatives (in the distributional sense) of
such functions. We recall here that partial derivatives of distributions are given by
; < ; <
r s r s
T, = (1) r+s
T, r s . (11.85)
zr zs z z

We also recall that the multiplication of distributions T by smooth functions (z)

C (R2 ) is understood through:

C (R2 ) %   T,  := T,  . (11.86)

Of course, all compactly supported distributions like Diracs and its derivatives, are
tempered and so are expected to be -quantizable. For the Dirac distribution (z),
already encountered in Eq. (11.35), the -quantization yields:

1 1
(z) (z) d2 z = (0) , (11.87)
C
which, in the CS case, particularizes to

1 1
A = |zz| (z) d2 z = 00 . (11.88)
C
Hence, the ground state, identified with the corresponding projection operator, is
the quantized version of the Dirac distribution supported at the origin of the phase
space. Similarly, the quantization of the Dirac distribution z0 (z z0 ) at the
point z0 reads as

Az = D(z0 )|00|D(z0 ) = |z0 z0 | . (11.89)

0

The obtention of all possible projections nn = |en en | or even all possible simple
operators nn = |en en | is based on the quantization of partial derivatives of the
distribution.
From now on, let us pursue with CS quantization which has the advantage to
yield concrete results. The quantized derivatives of the Dirac distribution at any
order are found as

inf (n,n )
 (1) p 1
A n n = (1)n+n n! n !
z z
p!
&
(n p)! (n p)!
np,n p . (11.90)
p=0
11.5 Weyl-Heisenberg Integral Quantizations of Functions and Distributions 331

With this quantity A n n at hand, one can invert the formula (11.90) in order to
z z
get n,n as:

inf(n,n ) 
 1
n,n = (1)n+n n! n ! A np n p . (11.91)
p=0 p! (n p)! (n p)! z z

Therefore the upper symbol of n,n is given by the distribution Tn,n supported at
the origin:
  

inf(n,n )
n+n 1 np n p
Tn,n = (1) n! n ! . (11.92)
p=0 p!(n p)!(n p)! znp zn p

Let us note that this distribution, as is well known, can be approximated, in the
sense of the D  (R2 ) (or S  (R2 )) topology, by smooth functions, such as linear
combinations of derivatives of Gaussians. The projections n,n are then obtained
trivially from (11.91) to get
n  
1 n
n,n = Azp p . (11.93)
p=0 p! p z

Many of the ideas around this combination of coherent states with distributions
pertain to the domain of Quantum Optics. They are already present in the original
works by Sudarshan [570], Glauber [333,334], Klauder [Kla68], Cahill [173], Miller
[469] and others. In Quantum Optics, the basic idea is that replacing the nondiagonal
representation of quantum operators, say B, (usually, in this context, one focuses on
the density operators ) given by

B= |z1 z2 | z1 | B |z2  d2 z1 d2 z2
C2
>
by a diagonal one, also called the P-representation, B = C |zz| P(z) d2 z, can
simplify considerably some calculations. Although this can be considered as the CS
quantization of P(z), the spirit is quite different since their approach is the inverse
of the one presented here: given a B, the question is to find P(z). The main result
obtained in this direction is that one can formally write a P-representation for each
quantum operator B, which is given by [570]

n| B |m n!m! r2 +i(mn)( ) (m+n)
P(z) = e (r), z = rei , (11.94)
m,n=0 2 r(n + m)!

or, by [Kla68, Sec.8.4],

x2 +y2

 y) e 2 ,
P(z) = F 1 B(x, z = (q + ip)/ 2, (11.95)
332 11 Integral Quantization

 y) = F [z| B |z] . Here F is the Fourier transform from the (p, q)-space
where B(x,
to the (x, y)-space, and F 1 is its inverse. However the question of the validity
of such formulas is mathematically nontrivial: the convergence in the sense of
distributions of (11.94) is a difficult problem (which might even have no explicit
solution!), and for instance has been partially studied by Miller in [469]. Manifestly,
the work done in this direction was concentrated on the dequantization problem
(finding an associated classical function to each quantum operator) and this was
done in a quite pragmatic spirit in order to simplify computations. Let us note
that the existence of such a well-defined dequantization procedure is by no means
a physical requirement, since the quantum realm is by definition richer than the
classical one. A more physical requirement is that the semi-classical limit be well
behaved.

11.6 Quantization with Coherent States or Frame: General

We explained above how integral quantization, and more specifically CS quanti-

zation (we could use as well the term frame quantization), is a generic expression
for calling a certain point of view in analyzing a set (say a set X of parameters)
equipped with a measure, say . This approach matches what physicists understand
by quantization when the observed measure space X is equipped with a phase
space or symplectic structure, like the phase space {position q momentum p} for
the motion of a particle on the line. It matches also well established approaches
by signal analysts, like time-frequency (Gabor) or time-scale (wavelet) analysis.
The set X can be finite, countably infinite, or uncountably infinite. The approach is
generically simple, in essence Hilbert space based, and generically the same.
Discarding any group theoretical construction la Gilmore-Perelomov, one
follows the construction suggested by (11.5) and discussed in Sect. 5.1, by first
considering the Hilbert space L2 (X, ) of complex square integrable functions on
X with respect to the measure . One then chooses in it an orthonormal set O
of functions n (x) (set aside the question of evaluation map in their respective
equivalence classes), satisfying the finiteness and positiveness conditions 0 <
N (x) = n |n (x)|2 < (a.e.), and a companion Hilbert space H (the space
of quantum states) with orthonormal basis {|en } in one-to-one correspondence
{|en  n } with the elements of O. There results a family C of unit vectors |x
(the coherent states) in H, which are labelled by elements of X and which resolve
the identity operator in H:

1
X % x  |x = & n (x)|en  H .
N (x) n
(11.96)


x|x = 1 , |xx| N (x) d (x) = IH . (11.97)
X
11.6 Quantization with Coherent States or Frame: General 333

This certainly represents the most straightforward way to build total families of
states resolving the identity, and one can easily show that most of the various CS
families proposed during the last 60 years could have been obtained in this way.
Note that there are interesting alternate methods, which do not require the existence
of a measure for defining coherent states, but only the existence of a reproducing
kernel Hilbert space developed in Section 5.1 (see also [387]).
A direct and important consequence of the resolution of the identity (11.97) is
the existence of a positive operator-valued measure on the measure space (X, F ),
for a algebra F of subsets of X,

F %  |xx| N (x) d (x) L (H)+ . (11.98)

As we already stated in this chapter, Eq. (11.97) is the departure point for
analyzing the original set X and functions living on it from the point of view of
the frame (in its true sense) C through the CS quantization map

f (x)  A f = |xx| f (x) N (x) d (x) .
X

We end in general with a non-commutative algebra of operators in H, leaving aside

the usual questions of domains in the infinite dimensional case. In turn, considering
the properties of the lower symbol f(x) := x|A f |x as compared to the original
f (x) allows one to decide whether the procedure does or does not make sense
mathematically.
There is a kind of manifest universality in this approach. Changing the frame
family C produces another quantization, another point of view, possibly mathemat-
ically and/or physically equivalent to the previous one, possibly not.
There is also, underlying the construction, a deep Bayesian content [34],
based or not on experimental evidences or on selective information choice,
namely,> an interplay between the set of probability distributions x  |n (x)|2
(from X |n (x)|2 d (x) =1), labelled by n, on the classical measure space
(X, ), and the discrete set of probability distributions n  |n (x)|2 /N (x) (from
N (x) = n |n (x)|2 ).

11.6.1 A First Example: Frame Quantization of Finite Sets

As a first illustration of our general construction, we generalize to an arbitrary finite

set the seastar model of Sect. 11.2 [214, 216].
334 11 Integral Quantization

11.6.1.1 A Short Review of Finite Frames

Let K be the field R or C, and let H be a d-dimensional Hilbert space over K with
{|en }dn=1 a fixed orthonormal basis. We recall that a system of vectors {|wi }M
i=1 is
a finite frame [Chr03] for H if there are constants 0 < m M < such that

M
m||v||2 |wi |v|2 M||v||2 for all |v H.
i=1

i=1 |wi wi |v satisfies the relation

The frame operator S|v = M

M
v|mv = m||v||2 |wi |v|2 = v|Sv M||v||2 = v|Mv
i=1

that is,

m IH S M IH .

If m = M, the frame is called a tight frame and

S = M IH .

A frame {|wi }M

i=1 is called an equal norm frame if

||w1 || = ||w2 || = = ||wM || .

A tight frame with M = 1 is usually called a Parseval frame and in this case
M
|wi wi | = IH . (11.99)
i=1

If {|wi }M
i=1 is a tight frame with frame bound M, then { M |wi }i=1 is a Parseval
1 M

frame. In the sequel, we consider only Parseval frames which do not contain the null
vector and express their vectors in terms of some unit vectors. Let {|wi }M i=1 be a
Parseval frame. Denoting i = wi |wi  and |ui  = 1i |wi  the resolution of identity
becomes
M
i |ui ui | = IH .
i=1

Two relations are easily derived from these definitions.

M M
v|w = i v|ui  ui |w, ||v||2 = i |ui |v|2 (11.100)
i=1 i=1
11.6 Quantization with Coherent States or Frame: General 335

for any |v, |w H, and

d d M M d M
d= en |en  = i |ui |en |2 = i |ui |en |2 = i . (11.101)
n=1 n=1 i=1 i=1 n=1 i=1

Next, by a normalized Parseval frame in H, we mean any system of vectors {|ui }M
i=1
satisfying the following two conditions:
(i) The vectors |ui  are unit vectors, that is,

ui |ui  = 1, for any i {1, 2, . . . M};

(ii) There are M positive constants {i }M

i=1 such that

M
i |ui ui | = IH .
i=1

In the case 1 = 2 = . . . = M , we are in the uniform case, like for the seastar:

d M
|ui ui | = IH
M i=1

From now on, frame will mean normalized Parseval frame. Let X =
{a1 , . . . , aM } be a set of parameters (or indices), with M d, equipped with the
uniform measure:
 M
f (x) d (x) := f (ai ) .
X i=1

Consider an orthonormal system of d functions {n : X C}dn=1 :

M
m (ai ) n (ai ) = mn (11.102)
i=1

such that
d
i = |n (ai )|2 = 0 for every i {1, . . . , M} .
n=1

We then consider the map X % ai  |ai  H defined by

1 d
|ai  = n (ai ) |en  , i = 1, . . . , M .
i n=1
336 11 Integral Quantization

By construction, these vectors resolve the identity in H, i.e., they form a (tight finite
normalized) frame:
M
i |ai ai | = IH . (11.103)
i=1

We easily infer that i 1, and this inequality is strict when d < M. Their overlap
reads as:
d
1
ai |a j  =
i j n=1
n (ai ) n (a j ) .

11.6.1.2 Bayesian Duality

The tight frame {|ai }M i=1 defines two probability distributions which can be
interpreted in terms of a Bayesian duality:
(i) A prior distribution on the set of indices n {1, . . . , d}, with parameter an X ,

|n (ai )|2 d
n 
i
= |en |ai |2 , |en |ai |2 = 1 for i {1, . . . , M} .
n=1
(11.104)
(ii) A posterior distribution on the original set of parameters ai X , equipped
with the uniform (discrete) measure, and with parameter n {1, . . . , d},

M
ai  |n (ai )|2 , |n (ai )|2 = 1 . (11.105)
i=1

Therefore, the resolution of the identity satisfied by the states |ai  introduces a
preferred prior measure, precisely the uniform one, on X , which is the set of
parameters of the distribution n  i 1 |n (ai )|2 , with this distribution itself playing
the role of the likelihood function. The associated distributions ai  |n (ai )|2 ,
indexed by n, become the related conditional posterior distributions.

11.6.1.3 Distance of a Frame to Orthonormality

Let us introduce the real M M matrix U with matrix elements Ui j = |ai |a j |2 .
These elements obey Uii = 1 for 1 i M and we suppose that 0 < Ui j = U ji < 1
for every pair (i, j), with i = j.
By the Perron-Frobenius theorem [Mey00], the spectral radius r = r(U) > 0
and it is the dominant simple eigenvalue of U. There exists a unique vector vr ,
vr  = 1, which is strictly positive (all components are > 0 and can be interpreted
as probabilities) and Uvr = rvr . All other eigenvalues of U lie within the open disk
11.6 Quantization with Coherent States or Frame: General 337

of radius r : | | < r. Since trU = M, and since U has M eigenvalues, one should have
r > 1. The value r = 1 represents precisely the limit case in which all eigenvalues
are equal to 1, i.e., U = I and the frame is simply an orthonormal basis of CM . It is
then natural to view the number := r 1 as a kind of distance of the frame to
the orthonormality. An interesting question concerns the relation between the set
{1 , 2 , . . . , M } of weights defining the frame and . Projecting on each vector |ai 
from both sides the frame resolution of the identity leads to the M equations:

M
1 = ai |ai  = j |ai |a j |2 , i.e. U v = v , (11.106)
j=1

where vT := (1 2 . . . M ) and vT := (1 1 . . . 1) CM . In the uniform case for which

i = d/M for all i, i.e., in the case of a finite equal norm Parseval frame, which
means that v = (d/M) v , then r = M/d and vr = 1/ M v , i.e., = (M d)/d,
a relation which clearly exemplifies what we can expect at the limit d M.

11.6.1.4 Stochastic & Ergodic Aspects of a Frame

Let us examine the matrix P := U K, where

K := diag(1 , 2 , . . . , M ) .
 
It is (right) stochastic. The row vector := vT /d = d1 2
d ... M
d is a stationary
probability vector:

P= . (11.107)

As is well known, this vector obeys the ergodic property:

! " j
lim Pk = j = . (11.108)
k ij d

11.6.1.5 Frame Quantization

Following our CS quantization scheme, we associate to each function f : X C

the operator

M
A f : H H, A f = i f (ai ) |an an | .
i=1

The operator A f corresponding to a real function f is self-adjoint. The eigenvalues

of the matrix A f form the quantum spectrum of f (by opposition to its classical
spectrum, that is, the set of its values f (ai )). The matrix elements of A f with respect
338 11 Integral Quantization

to the orthonormal basis {|en } are

M
em |A f |en  = f (ai ) m (ai ) n (ai ) .
i=1

11.6.1.6 Lower Symbols

In the present context, and given a function f and its quantized version A f , its lower
or covariant symbol is the function f = A f : X R, such that
M
f(ai ) = ai |A f |ai  = j f (a j ) |ai |a j |2 , (11.109)
j=1

Note that
M M
tr A f = i f(ai ) = i f (ai ) .
i=1 i=1

Semi-classical aspects of the finite frame quantization of f can be analyzed through

this lower symbol. Indeed, with the stochastic matrix approach, the relation (11.109)
may be rewritten as

f = Pf , P = UK , (11.110)
 
with f T := ( f (a1 ) f (a2 ) . . . f (aM )) and f T := f(a1 ) f(a2 ) . . . f(aM ) . This formula
is interesting because it can be iterated:

f [k] = Pk f , f [k] = P f [k1] , f [1] f , (11.111)

and so we find from the stochastic nature of P that the ergodic limit (or long-
term average) of the iteration stabilizes to the classical average of the observable
f defined as:
M
i
f [] =  f cl v , where  f cl := f (ai ) , (11.112)
i=1 d

a feature already mentioned for the seastar model of Sect. 11.2. Now, we can
evaluate the distance between the lower symbol and its classical counterpart through
the inequality:

f f := max | f (ai ) f(ai )| I P f , (11.113)

1id

where the induced norm [Mey00] on the matrix A is A = max1iM Mj=1 |ai j |.
In the present case, because of the stochastic nature of P, we have
11.6 Quantization with Coherent States or Frame: General 339

 
I P = 2 1 min i . (11.114)
1iM

In the uniform case, i = d/M for all n, we thus have an estimate of how far from
each other the two functions f and f are: ff 2(M d)/M f . In the general
case, we can view the parameter

:= 1 min i (11.115)
1iM

as a distance of the quantum world to the classical one, of non-commutativity to

commutativity, or again of the frame to an orthonormal basis, like the distance
= r 1 introduced above. Another way to check that = 1 d/M 0 means,
in the uniform case n = d/M for all n, that we go back to the classical spectrum of
the observable f results from the following relations. We have

d d d
M |ai |a j |2 = ||ai ||2 M |ai |ai |2 = 1 M ,
j=i

and, from the relation

M
d d
f(ai ) =
M f (a j ) |ai |a j |2 =
M
f (ai ) + f (a j ) |ai |a j |2 ,
j=1 j=i

we get
   
d d d
1 min f (ai ) f (a j ) f (a j ) 1
max f (ai ) . (11.116)
M i M M i

11.6.2 A Second Example: CS Quantization of Motion

on the Circle

11.6.2.1 Construction and Implementation

Another basic illustration of the method concerns the motion of a particle on the
circle (see also a similar model in tomographic measurements [434]). Here, the
phase space in terms of action-angle variables is the cylinder X = {(J, ) , J R ,
[0, 2 )} equipped with its symplectic measure d (x) = (2 h)1 dJ d , where h is
the Planck constant. Our choice of an orthonormal set in L2 (X, dJ d /2 ) is
 ! "1/4 
 2 i n
2 (Jn)
O = n (x) = e e ,nZ , (11.117)

340 11 Integral Quantization

Fig. 11.3 Spectrum of the angle operator A obtained by CS quantization of the angle of rotation
on the circle for different values of the parameter (courtesy of R. Kanamoto)

where n = n or n = n2 . This expression involving Gaussian coefficients is directly

inspired from the first example (with n = n) given in Sect. 6.3 and by original works
about coherent states on the circle
[239,338,423]. We observe that a free parameter,
namely the half-width = 1/ 2 of the Gaussian, is present in the construction.
It has the meaning of a resolving power, like the fundamental length  introduced
in (11.69) for the standard CS. Let H be a separable (complex) Hilbert space with
orthonormal basis {|en  , n Z}. The coherent states read, with J= J/h,

1 ! "1/4

e 2 (Jn) ein |en  .
1 2
|x |J,  = & (11.118)
N (J) nZ

The normalization function N (J) N (J) is given by (6.39). The corresponding

CS quantization of the action J gives the diagonal operator AJ with eigenvalues
Jn = hn, i.e., the quantum angular momentum, and for the energy E J 2 , En n2 +
const., as it should be.
Concerning the alternative about the frequency function n  n , the choice n
n is the usual one, whereas the choice n n2 is appropriate for the quantization of
the energy since it yields temporal evolution stability.
As for the motion on the line, the quantization of the 2 -periodic angle function
( ) = for [0, 2 ) provides a bounded self-adjoint angle operator. We show
in Figs. 11.3 and 11.4 its spectrum and lower symbol, respectively. We observe
the extent to which the classical limit is reached at large . For more details, see
[308, Gaz09].
Instead of the Gaussian probability distribution in (11.117), we can actually
choose a non-negative, even, well localized and normalized integrable function
11.6 Quantization with Coherent States or Frame: General 341

Fig. 11.4 Lower symbol of the angle operator A obtained by CS quantization of the angle of
rotation on the circle for different values of the parameter (courtesy of R. Kanamoto)

 +
R % J  p (J) , p (J) = p (J) , dJp (J) = 1 , (11.119)

where > 0 is some width parameter, and obeying

0 < N (J) := pn (J) < , for all J R,

nZ

where pn (J) := p0 (J hn). The simplest choice is the uniform function p (J) =

2 [ , ] (J ). The functions n (x), for n Z, are now given by:
1

&
n (x) = pn (J) ein , n Z, (11.120)

with n = n or n2 . The corresponding family of coherent states on the circle reads

as:
1 &
|J,  = &
N (J) pn (J) ein |en  . (11.121)
nZ

They are normalized, resolve the identity and give the correct quantization for action
J and energy J 2 . For more details see [88].
342 11 Integral Quantization

11.6.3 Quantization With Action-Angle CS for Bounded

Motions

Let us now examine in which way the action-angle material presented above can be
extended to more generic mechanical systems.

11.6.3.1 Action & Angle in Classical Mechanics

We start with a quick review of the action-angle formalism in Classical Mechanics

[Gol81, Lan81]. Consider a conservative confined mechanical system with phase
space conjugate variables (q, p). For a given motion, its Hamiltonian function is
fixed to a certain value E of the energy:

H(q, p) = E p = p(q, E) (11.122)

The action variable is defined by:

+
J= p(q, E) dq = J(E) , (11.123)

where the loop integral is understood as performed over a complete period of

libration or rotation. The conjugate angle> variable is then derived from the Hamilton
characteristic function W = W (J, q) = p dq which generates contact transforma-
tion (q, p)  (J, ) at constant Hamiltonian, where = WJ . Its time evolution is
-periodic:
t J
= + 0 , = = (E) . (11.124)
(E) E
The two examples considered in the previous sections, namely the harmonic
oscillator and the free rotator represent extreme cases:
Free rotator: this means a particle of mass m, freely moving on the circle of
radius l. The canonical coordinates are
J
= 2 , p = ml 2 = ,
2
p2
and the energy is E = .
2ml 2
p2 1 2
Harmonic oscillator: from E = + kq ,
2m 2
2
J= E, = + 0 , q = qmax sin .
2
One of the most familiar prototype of intermediate system is the simple
pendulum.
11.6 Quantization with Coherent States or Frame: General 343

11.6.3.2 At the Origin of Quantum Mechanics: Bohr-Sommerfeld

Quantization

It was asserted by Bohr and Sommerfeld that the only permitted confined motions
are those which obey:
+
J= p(E) dq = n h , n N, (11.125)

where h is the Planck constant or quantum of action.

This old condition is correct for the quantization of the angular momentum for the
free rotator, p = J/(2 ) = nh and so the quantization of the free rotator quantum
energy E = n2 h2 /(2ml 2 ), whereas it gives E = h n for the harmonic oscillator
energy and so does not provide the observed ground state [Her89] one-half quantum,
measured from the minimum of the potential energy [143].
The motivation for the Bohr-Sommerfeld quantum condition was the correspon-
dence principle: the behavior of a quantum system reproduces classical physics in
the limit of large quantum numbers. This principle has to be complemented by the
physical observation that the quantity to be quantized must be an adiabatic invariant
(i.e., it remains constant when changes occur slowly).

11.6.3.3 A Bayesian Probabilistic Construction of Action-Angle Coherent

States and Related Quantizations

Suppose that energy measurements on a confined one-dimensional system yields

the sequence of values (up to a constant shift):

E0 < E1 < < En < . (11.126)

We define a corresponding sequence of probability distributions J  pn (J), i.e.,

>
 
R or R+ dJ pn (J) = 1, with J = J/h, supposing a (prior) uniform distribution on the
range of the action variable J, obeying the two conditions:

0 < N (J) := pn (J) < , En + cst =
R or R+
E(J) pn (J) dJ (11.127)
nZ or N

where R and Z (resp. R+ and N) stand for the rotation (resp. libration) type of
motion.
The finiteness condition allows to consider the map n  pn (J)/N (J) as a
probabilistic model referring to the discrete data, which might be viewed in the
present context as a prior distribution. Let H be a complex separable Hilbert space
with orthonormal basis {|en  n Z or N}. Let > 0 be a rescaled period of the
344 11 Integral Quantization

angle variable and X = {(J, ) , J R or R+ , 0 < }3 be the action-angle

phase space for a rotation (resp. libration) motion with measured energies given
by the discrete sequence E0 < E1 < < En < . Let (pn (J))nZ or N be the
sequence of probability distributions associated with these energies. We suppose
pn (J) = pn (J) in the rotation case.
One then constructs the family of states in H for the considered motion as the
following continuous map from X into H:

1 &
X % (J, )  |J,  = &
N (J)
pn (J) ein |en  H , (11.128)
n

where the choice of the real sequence n  n is left to us in order to comply with
some criteria imposed a priori. In both cases the coherent states |J, 
(i) Are unit vector s: J, |J,  = 1
(ii) Resolve the identity operator in H with respect to a measure in the Bohr sense
B ( dJ d ) on the phase space X :
   T
1
dJN (J) lim
2
N (J) |J, J, | dB (J , ) = |J, J, | d = IH .
X Z or N T T T2
(11.129)

(iii) Allow a quantization of classical observables f (J, ) which is energy compat-

ible with our construction of the posterior distribution J  pn (J):

f (J, )  A f = N (J) f (J, ) |J, J, | dB (J , ). (11.130)
X

(iv) Since it is trivially verified that in both cases the quantum Hamiltonian is
exactly what we expect:

AE(J) = (En + cst)|en en | . (11.131)

n

11.7 Application to Various Systems

We present here (in chronological order) a series of works that constitute

explorations of the integral (mainly CS) quantization method we have exposed
in this chapter: Gazeau et al. [297299, 314, 322], Ali et al. [35], Baldiotti et al.
[117, 118], Gazeaus textbook [Gaz09], Cotfas et al. [214, 215], ElBaz et al. [270],

3 Actually we keep the freedom of making vary from to + as we do for any angle variable.
11.7 Application to Various Systems 345

Bergeron et al. [141], Cotfas et al. [216], Gazeau-Szafraniec [318], Gazeau-del

Olmo [312], Aremua et al. [88], Gazeau-Kanamoto [308], Bagrov et al. [111], Ali
et al. [39], Bergeron-Gazeau [139].
Note that these approaches were already present, to a certain extent, in the first
works by Klauder at the beginning of the sixties (see for instance [413, Kla00] and
references therein), in the famous 1975 paper on quantization by Berezin [135, 136]
and in the works by others. We may quote, for instance, [Unt91] and [137]. The
latter, in particular, seeks a possible continuous path from the classical world to
the quantum one, in a way compatible with all axioms of quantum mechanics. Note
that the method fits as well the so-called fuzzyfication la Madore [Mad95](see
also [10] and references therein) of the sphere [299] or of the hyperboloid [323].
As a matter of fact, in the 2010 paper [270], the authors explored a new
and quite exotic set equipped with a measure la Berezin-Majid-Rodriguez-
Plaza [454]. There are interesting and unexpected issues, such as the one-to-one
correspondence between the paragrassmann or k-fermionic algebra and the standard
algebra of matrices, or the fact that q-commutation relations between paragrassmann
generators are mapped to the Biedenharn-Macfarlane commutation relations for
the q-oscillator. These results, recently revisited and extended by S. Sontz [561
564], yield a non-commutative classical description of matrix algebras viewed
as quantum operators. These works shed a new light on finite-dimensional Hilbert
spaces and their operators as quantum versions of finite sets endowed or not with
some more or less exotic algebraic structure. This opens new perspectives for
tackling some problems typically encountered in Quantum Information or Logic.
In [88, 308] are presented families of coherent states built from various prob-
ability distributions of the action variable for the motion on the circle, and more
generally for systems for which the phase space has a cylindric topology. The
resulting quantizations are particularly relevant to recent models in superconducting
circuit QED, for which is raised again the longstanding question of commutation
relation between phase ( angle) and number operator ( angular momentum),
more precisely excess Cooper pair number operator with spectrum Z and not just N.
They could also offer new perspectives in the study of time behavior of quantum-
chaotic phenomena.
Coherent states for unbounded motions have been built in [111]. The construction
extends to a continuous spectrum previous constructions of action-angle coherent
states in view of a consistent energy quantization. Integral quantization based on
the affine group have been explored recently. An interesting application in quantum
cosmology may be found in [144].
A variety of such situations involving CS quantization is once more the mark of
the universality and the easy implementation of this type of analysis, in comparison
with usual quantization methods (see the review [33]). Integral quantization allows
us to enlarge our vision on the still problematic relation between classical and
quantum worlds, in particular by tackling examples which can be considered as
toy examples in Physics and interesting objects to study in Mathematics.
Chapter 12
Wavelets

Abstract This is the first of four chapters devoted to a very successful type of CS,
namely, wavelets. In the present one, we treat the simplest case, the continuous
wavelet transform (CWT) in 1-D. Starting from the beginning, we rewrite the
general CS formalism for the case at hand, that is, the connected affine group of
the line. We discuss the basic properties, the interpretation of the CWT as a phase
space representation and some examples, with emphasis on a recent application to
NMR spectroscopy.

12.1 A Word of Motivation

Wavelet analysis is a particular time-scale or space-scale representation of signals

that has become popular in physics, mathematics and engineering in the last few
years. The genesis of the method is interesting for the present book, so we will
spend a paragraph outlining it. After the empirical discovery by Morlet (who
was analyzing microseismic data in the context of oil exploration [341]), it was
recognized from the very beginning by Grossmann, Morlet and Paul [349353]
that wavelets are simply coherent states associated to the affine group of the line
(dilations and translations). Thus immediately the stage was set for a far-reaching
generalization, using the formalism developed in Chap. 8 (it is revealing to note
that two out of those three authors are mathematical physicists). But then the wind
changed. Lemari and Meyer [440], Meyer [464, 465], and Mallat [457] made
the crucial discovery that orthonormal bases of regular wavelets could be built,
and even with compact support, as shown by Daubechies [231], by changing the
perspective (of course, the orthonormal basis of the Haar wavelets was known since
the beginning of the last century, but these are piecewise constant, discontinuous
functions). Group theory lost its priority, in favor of the so-called multiresolution
analysis (more about this in Sect. 13.1.1 below), which made contact with the world
of signal processing and engineering. Then the theory really caught the attention of

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 347
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__12,
Springer Science+Business Media New York 2014
348 12 Wavelets

practitioners and it started to grow explosively. For a thorough survey of the early
papers on wavelet theory, we refer to the reprint collection edited by Heil and Walnut
[Hei06].
In this chapter, we will nevertheless focus on the continuous wavelet transform
or CWT, based on group representations, in one space (or time) dimension. This
approach fits with the general trend of the book, but, more important, it is by far
the most natural way to extend wavelets to higher dimensions and more general
situations, that will be described in Chap. 14.
Let us begin with an intuitive discussion, in order to convey to the reader a feeling
for the theory, and also to understand its success. It is a fact that most real life signals
are nonstationary. They often contain transient components, sometimes physically
very significant, and mostly cover a wide range of frequencies. In addition, there is
frequently (but not always) a direct correlation between the characteristic frequency
of a given segment of the signal and the time duration of that segment. Low
frequency pieces tend to last a long interval, whereas high frequencies occur
in general for a short moment only. Human speech signals are typical in this
respect. Vowels have a relatively low mean frequency and last quite long, whereas
consonants contain a wide spectrum, especially in the attack, and are often very
short.
Clearly the usual Fourier transform is inadequate for treating such signals. It
gives a purely frequency domain representation, and loses all information on time
localization, which may be crucial. For this reason, signal analysts turn to time-
frequency representations. The idea is that one needs two parameters: one, called
a, refers to frequency, the other one, b, indicates the position in the signal. Thus,
assuming the transform to be linear on signals, one writes a general time-frequency
transform of a signal s as

s(t) S(b, a) = ba (t) s(t) dt, (12.1)
R

where ba is the analyzing function and ba its complex conjugate, a > 0, b R.

This concept of a time-frequency representation is in fact quite old and familiar:
the most obvious example is simply a musical score! Clearly it is important to know
when to play a given note, and not only to which frequency it corresponds. Eloquent
comments along this line, by J. Ville and by L. de Broglie, may be found in [Fla98,
p. 9].
Among the possible linear time-frequency transforms, two stand out as particu-
larly simple and efficient: the Windowed or short-time Fourier transform (STFT)1
and the wavelet transform (WT). Both are particular cases of CS, only the group
differs, and in fact they should be developed in parallel [Dau92]. The essential
difference between the two is in the way the frequency parameter a is introduced

1 Sometimes called the Gabor transform, although Gabor considered only a discretized version of

it [294].
12.1 A Word of Motivation 349

in the analyzing function. In both cases, b is simply a time translation. The kernels
of the two transforms can be written as follows :
1. For the Windowed Fourier Transform, one uses

ba (t) = eit/a (t b). (12.2)

Here is a window function and the a-dependence is a modulation (1/a

frequency); the window has constant width, but the lower a, the larger the number
of oscillations in the window. Comparing with Sect. 2.2, we see that (12.2) yields
exactly the standard CS associated to the WeylHeisenberg group (canonical CS
if is a Gaussian). In signal processing, the corresponding functions , which
are canonical CS, are sometimes called gaborettes.
2. For the wavelet transform, instead, one takes

1 !t b"
ba (t) = . (12.3)
a a

The action of a on the function is a dilation (a > 1) or a contraction (a < 1).

The shape of the function is unchanged, it is simply spread out or squeezed. In
particular, the effective support of ba varies as a function of a. As we will see
below, (12.3) yields the CS associated to the (connected) affine group of the real
line or ax + b group.
Thus, already at this stage, we may infer that the wavelet transform will be better
adapted to the analysis of the type of signals described above.
Notice that the requirement of linearity is nontrivial, for there exists a whole
class of quadratic, or more properly sesquilinear, time-frequency representations.
The prototype is the so-called WignerVille transform, introduced originally by
Wigner in quantum mechanics (in 1932!) [607] and extended by Ville to signal
analysis [595]:
 ! t" ! t"
W [s](b, a) = eit/a s b s b+ dt = 2Psb,a |sb,a , (12.4)
R 2 2

where sb,a sba is defined in (12.2) above and P is the space reflection operator
(Ps)(x) = s(x), introduced in (2.85) and in Sect. 11.4. Actually, the WignerVille
transform of s(t) is nothing but the so-called Wigner distribution W(q, p), for q =
b, p = 1/a (see Sect. 11.4). Further information may be found in [208, Fla98].
As it turns out, the examples discussed in this chapter and the following ones
are of a special character. On the mathematical side, they are distinguished by the
fact that the underlying group is a group of space or spacetime transformations
that contains dilations. Actually, we have already encountered them already in
Sects. 8.3.1 and 8.3.2, and we will see other ones in Chaps. 14 and 16. In addition,
the range of applications and the rle are different. Whereas the previous cases
were essentially concerned with quantum physics (atomic physics, optics, quantum
350 12 Wavelets

mechanics, elementary particles, even quantum gravity), here we deal mostly with
classical physics, even engineering: signal and image processing, fluid mechanics,
geophysics, astrophysics, . . . (notable exceptions are atomic and solid state physics
[339,Ber99]). In all these cases, wavelets are used essentially as a tool for analyzing
experimental data (e.g. time-frequency analysis of a signal). In most cases, no
explicit physical signification has been found for them. By contrast, the familiar
Fourier transform is a mathematical tool, but it has also a physical support. For
instance, the diffraction phenomenon in optics may be seen as the physical realiza-
tion of the Fourier transform. In quantum mechanics also, the latter plays a central
physical rle, encoding the transition from position to momentum representation,
and this is finally expressed in the form of Heisenberg uncertainty relations. But for
wavelets, it is only in the context of (quantum!) cosmology that they acquire a direct
physical meaning, as one can see for instance in [102, 103, 161, 273, 415] or [144].
In any case, they have reached a remarkably wide range of applications.
Let us say a word here on dilations. Physical applications will be described
briefly in the following sections. For every n 1, we consider global dilations of
Rn (zoom):

x  ax, a > 0, x Rn . (12.5)

These form a one-parameter, abelian group D(n) R+ R, hence unimodular. The

(left and right invariant) Haar measure is d (a) = a1 da. The group D(n) has a
natural unitary representation in the Hilbert space L2 (Rn , dx):

(U(a) f ) (x) = an/2 f (a1 x), a > 0. (12.6)

Most of the groups we will consider below, the so-called affine or similitude groups,
have the general structure

(n)
Gaff = G(n)  D(n)

of a semidirect product of D(n) with an appropriate transformation group G(n) of Rn ,

for instance, the Euclidean, the Galilei or the Poincar groupeach of which is in
fact the isometry group of a given geometry of space or spacetime.
Now each of these groups G(n) is itself a semidirect product G(n) = Rn  S(n) ,
where S(n) is the corresponding homogeneous group, and D(n) acts nontrivially
(n)
on Rn , according to (12.5). As a consequence, the affine group Gaff becomes
nonunimodular. If we denote by db the Lebesgue measure on Rn , and by d (n)
the Haar measure on S(n) (which is unimodular), then the left and right invariant
(n)
Haar measures on Gaff read, respectively:

db da da
d (b, S, a) = d (n) , dr (b, S, a) = db d (n) . (12.7)
an a a
12.2 Derivation and Properties of the 1-D Continuous Wavelet Transform 351

Clearly nonunimodularity does complicate the situation (the DufloMoore oper-

ator C of Theorem 8.2.1 is now nontrivial, and necessarily unbounded). However the
addition of dilations more than compensates, in that it restores square integrability.
As we will see in Chap. 14, most of the affine groups of interest have a square
integrable representation, possibly modulo the center. The only notable exception
is the affine WeylHeisenberg group (Sect. 16.1), for which we will have to the use
the general formalism developed in the previous chapters.
Remark. Our Fourier transform in Rn is defined as

f( ) = (2 )n/2 ei x f (x) dx, (12.8)
Rn

for , x Rn and x = 1 x1 + 2 x2 + + n xn . Note that, for better homogeneity,

we will use the variable x both for n = 1, instead of the more familiar t of signal
processing, and for n > 1, instead of x, as used in the preceding chapters.

12.2 Derivation and Properties of the 1-D Continuous

Wavelet Transform

Take first the full affine group of the line Gaff = {(b, a) : b R, a = 0, }, with the
natural action x  ax + b and group law

(b, a)(b , a ) = (b + ab , aa ). (12.9)

Thus Gaff is a semidirect product of the translation group R by the full dilation
group R : Gaff = R  R . The unit element is (0, 1) and the inverse of (b, a) is
(a1 b, a1 ).
The group Gaff is non-unimodular, the left Haar measure is d (b, a) = |a|2 da db
and the right Haar measure is dr (b, a) = |a|1 da db (see also Sect. 8.1).
The whole theory of wavelets rests on the following central result [102, 103]:
Theorem 12.2.1. Up to unitary equivalence, Gaff has a unique UIR, acting in
L2 (R, dx), namely
!xb"
(U(b, a) f ) (x) = |a|1/2 f =: fba (x) (a = 0, b R), (12.10)
a
or, on Fourier transforms:
 
 a) f ) = |a|1/2 f(a )eib
U(b, (a = 0, b R). (12.11)

The representation U is square integrable, and a vector L2 (R, dx) is admissible

if and only if it satisfies the condition
352 12 Wavelets

 +
d
c := 2 |
 ( )|2 < . (12.12)
| |

Proof. Unitarity and irreducibility of U are immediate.

Uniqueness follows from the fact that the representation U is obtained by
Mackeys standard method of induction (there is only one nontrivial orbit in -
space) [102, 103, Bar77] (see Chaps. 3 and 8).
Finally, the square integrability follows from a straightforward calculation:

 a) da db
|U(b,  |
 |2
Gaff a2

 da
= eib( )  (a  )
 (a )  (  ) d d 
 ( ) db
|a|

da
= 2 |
 (a )|2 |
 ( )|2 d
|a|
 +
d
= 2  2 |
 ( )|2 <
| |

(the exchange of integrals is justified by Fubinis theorem). 


In practice, the admissibility condition (12.12) (plus some regularity: L1 L2
suffices) is equivalent to a zero mean condition:
 +
()
admissible
 (0) = 0 (x) dx = 0. (12.13)

Indeed, if> L1 (R),  is continuous and (12.12) requires

>
 (0) = 0. Conversely
[232], if (x) dx = 0 and (1 + |x|) | (x)| dx < for some > 0 (a condition
slightly stronger than L1 ), then |  ( )| | | , = min{ , 1}, and (12.12)
holds.
From now on, an admissible function will be called a wavelet [note that some
authors [Woj97] use more restrictive definitions in the context of multiresolution
analysis (see Sect. 13.1.1 below)]. Thus a wavelet is by necessity an oscillating
function, real or complex-valued (see the examples below). However, contrary to a
popular belief, this is in fact not the origin of the term wavelet (indeed, they were
called wavelets of constant shape in the very first paper [349], because wavelet
was already in use in the geophysics communityof course, such a long name did
not survive this first paper, it was immediately abbreviated to wavelets). We refer
to [50] for a historical account.
Let be a wavelet and s L2 (R) a signal. Then the CWT of s with respect to
is the function S := T s given by the scalar product of s with the transformed
wavelet ba :
12.2 Derivation and Properties of the 1-D Continuous Wavelet Transform 353

S(b, a) = ba |s

 +
= |a|1/2 (a1 (x b)) s(x) dx (12.14)

 +
= |a|1/2  (a ) s( ) eib d .
(12.15)

Clearly the map T : s  S coincides, up to a constant, with the coherent state map
WK or W described in Sects. 7.1.1 and 8.1, respectively. Thus all the properties
listed below are immediate translations of the corresponding ones discussed there.
Notice that a different normalization is often used in the examples, namely replacing
|a|1/2 by |a|1 in (12.10) and (12.14), which has the effect of enhancing small
scales in transforms. This is called the L1 -normalization, because it ensures that the
L1 norm is conserved under dilation. Indeed, denoting ba (x) = |a|1 f ((x b)/a),
we have  ba 1 =  1 , b, a. However the L -normalization given above is the
2

only one that comes from a unitary representation of Gaff .

In practice one often imposes on the analyzing wavelet a number of additional
properties, for instance, restrictions on the support of and of  (see Sect. 12.4
below). Or may be required to have a certain number N 1 of vanishing moments
(by the admissibility condition (12.13), the moment of order 0 must always vanish):
 +
xn (x) dx = 0, n = 0, 1, . . . N. (12.16)

This property improves its efficiency at detecting singularities in the signal. Indeed,
the transform (12.14) is then blind to the smoothest part of the signal, that which
is polynomial of degree up to Nand less interesting, in general. Only the sharper
part remains, including all singularities (jumps in the signal or one of its derivatives).
For instance, if the first moment (n = 1) vanishes, the transform will erase any linear
trend in the signal.
Before proceeding to the general properties of the CWT, let us give two
examples, actually the most widely used wavelets, which are depicted in Fig. 12.1.
The Mexican hat or Marr wavelet:
This is simply the second derivative of a Gaussian:

H (x) = (1 x2 ) ex ?H ( ) = 2 e
2 /2 2 /2
, . (12.17)

It is a real wavelet, with two vanishing moments (n = 0, 1). Similar wavelets,

with more vanishing moments, are obtained by taking higher derivatives of the
Gaussian [353]:
 m
(m) 1 d ((m)
ex H ( ) = m e /2 .
2 /2 2
H (x) = , (12.18)
i dx
354 12 Wavelets

10 0 10 10 0 10

Fig. 12.1 Two usual wavelets; (left) The Mexican hat or Marr wavelet; (right) The real part of the
Morlet wavelet (o = 5.6)

The Morlet wavelet:

This is just a modulated Gaussian:
! "
M (x) = 1/4 eio x eo /2 ex /2
2 2


?M ( ) = 1/4 e( o ) /2 e /2 eo /2 .
2 2 2
(12.19)

In fact the first term alone does not satisfy the admissibility condition, hence
the necessity of a correction. However, for o large enough (typically o 5.5),
this correction term is numerically negligible ( 104 ). The Morlet wavelet
is complex, hence the corresponding transform S(b, a) is also complex. This
enables one to deal separately with the phase and the modulus of the transform,
and the phase turns out to be a crucial ingredient in most algorithms used in
applications such as feature detection.
Besides these standard, universally used wavelets, there are other ones, specif-
ically designed for a given type of signals. Such are, for instance, the 1-D
CauchyPaul wavelets introduced in [500, Pau85] for the purpose of quantum
mechanics, namely, for m > 0,

m! 1
m (x) = ,
2 (1 ix) m+1

@
m e , 0,

?m ( ) = (12.20)
0, <0

(see also Sect. 14.3.1). The wavelet m is progressive and extends to an analytic
function in the upper half-plane Im x > 0.
12.2 Derivation and Properties of the 1-D Continuous Wavelet Transform 355

Another example is provided by the so-called metabolite-based autocorrela-

tion wavelets, adapted to NMR spectroscopy, that we shall discuss in detail in
Sect. 12.7.1.
Consider now the connected subgroup of Gaff , called G+ or the ax + b group:

G+ = {(b, a) : b R, a > 0}. (12.21)

When restricted to G+ , the representation U splits into the direct sum of two
unitarily inequivalent, square integrable, UIRs U , acting in the two Hardy spaces
(see also Section 8.1):

H+ (R) = { f L2 (R) : f( ) = 0 for < 0},

H (R) = { f L2 (R) : f( ) = 0 for > 0}, (12.22)

and

L2 (R) = H+ (R) H (R). (12.23)

Elements of H+ (R), resp. H (R), extend to functions analytic in the upper, resp.
lower, complex half-plane, and accordingly they are called upper, resp. lower
analytic signals [Lyn82, Pap02].
From now on we shall mostly restrict ourselves to the representation U+ , acting in
H+ (R) (actually, Theorem 8.1.5 allows one to extend wavelet analysis from H+ (R)
to the full space L2 (R)). Thus a vector H+ (R) is admissible if it satisfies

d
c = 2 |
 ( )|2 < . (12.24)
0

If, in addition,
 is real, is then called a progressive wavelet (notice that neither
the Mexican hat, nor the Morlet wavelet are progressive).
The general properties of the CWT may then by summarized into the following
theorem, which is simply the particularization of Theorem 8.1.3.
Theorem 12.2.2. Let T : s  S be the wavelet transform with respect to the
wavelet . Then:
1/2
(1) The map W := c T is an isometry from the space H = L2 (R) or H+ (R)
of finite energy signals (in the sense of signal processing) onto the space H of
wavelet transforms, which is a closed subspace of L2 (G+ , a2 da db):
 
da db
|S(b, a)|2 = c |s(x)|2 dx. (12.25)
G+ a2 R
356 12 Wavelets

Equivalently, the wavelet generates a resolution of the identity (tight frame):


da db
c1
|ba ba | = I. (12.26)
G+ a2

(2) The projection P = W W from L2 (G+ , a2 da db) onto H is an integral

operator, with (reproducing) kernel

K(b , a ; b, a) = c1
b a |ba . (12.27)

Hence a function f L2 (G+ , a2 da db) is the wavelet transform of a signal if

and only if it satisfies the reproduction property

da db
f (b , a ) = K(b , a ; b, a) f (b, a) . (12.28)
G+ a2

(3) By (1), the map W may be inverted on its range H by its adjoint: W1 H =
W . As a consequence, the signal s may be recovered from its wavelet transform
S with help of the reconstruction (or synthesis) operator R = T :

da db
s(x) = c1 1
(R S)(x) = c ba (x) S(b, a) . (12.29)
G+ a2

(4) Finally, the CWT is covariant with respect to the group G+ :

s(x bo )  S(b bo , a), bo R, (12.30)

1/2
ao s(a1 1 1
o x)  S(ao b, ao a), ao > 0.

The unitary operator W : H+ (R) H is the particularization to the ax+b group

of the coherent state map WK or W described in Sects. 7.1.1 and 8.1, respectively,
and similarly in the case of the full affine group Gaff .
Coming back to statement (1) of the theorem, we recall that in signal analysis, as
in classical electromagnetism, the L2 norm is interpreted as the total energy of the
signal. Therefore, the relation (12.25) means that W conserves energy. In turn, this
suggests to interpret |S(b, a)|2 as the energy density in the wavelet parameter space.
We will make use of this interpretation in Sect. 12.4.2 for the study of localization
properties and the physical interpretation of the CWT.
Clearly the reconstruction formula (12.29) yields a decomposition of the signal
into a linear superposition of the wavelets ba with coefficients S(b, a). In the
same way, one may reconstruct the signal by using a wavelet different from the
analyzing wavelet [384, Hol95]:
12.3 A Mathematical Aside: Extension to Distributions 357


da db
s(x) = c1
ba (x) (T s)(b, a) , (12.31)
G+ a2

provided and satisfy the compatibility condition


d
0 < |c | < , where c =
 ( ) ( ) . (12.32)
| |

When (12.32) holds, we say that is a reconstruction wavelet for . Notice that
such a extended reconstruction formula also follows from the general theory of
square integrable representations discussed in Sect. 8.2, namely the relation (8.52).
Analogous reconstruction formulas may also be written for the WT based on the
full affine group Gaff [232, Dau92].
We emphasize the covariance property (12.30), especially that with respect to
translations. It is one of the reasons why the CWT is extremely useful for detecting
particular features in signals. Indeed translation covariance is lost in the discrete
time WT discussed briefly in Sect. 13.1 below, and this is one of the drawbacks of
the discrete wavelet schemes based on multiresolution.
Now since the wavelet transform (12.14) is a scalar product, it makes sense to
take for a singular function, provided the signal is correspondingly smooth (this
is the basic idea of weak (Sobolev) derivatives or rigged Hilbert spaces (RHSs)
[Gel64]). For instance, if the signal s is continuous, we may take for a Dirac
measure ( function), so that  = 1. Thus, if the analyzing wavelet satisfies the
condition resulting from (12.32),

d
|
 ( )| < , (12.33)
| |

then one gets the simplified reconstruction formula used by Grossmann and Morlet
[350] (and originally due to Caldern):

da
s(x) = S(a, x) . (12.34)
0 a

The extension of the CWT to distributions, in the context of RHSs, is discussed in

detail in the next section.

12.3 A Mathematical Aside: Extension to Distributions

Before going over to the practical problems of implementing the CWT and
applying it to physical problems, we shall devote a few comments to a remarkable
mathematical aspect, namely its extension to distributions.
Let us consider for a while the standard RHS [Gel64]:
358 12 Wavelets

S (R) L2 (R) S (R), (12.35)

where S (R) is the Schwartz space of fast decreasing C functions and S (R) the
space of tempered distributions, that is, continuous antilinear functionals on S (R).
Now the sesquilinear form < , > that puts S and S in duality may be identified,
up to a complex conjugation, with the inner product of L2 :

< f , s > = s| f L2 (R) , f , s S (R). (12.36)

Note that the convention we have adopted here is the one used in the RHS theory. It
allows both embeddings in (12.35) to be linear. Note that the opposite convention
is used in [Hol95], considering instead the space S  (R) of continuous linear
functionals on S (R), and then an antilinear embedding of S into S  .
Many operations from analysis, such as derivation, multiplication by a polyno-
mial, convolution, map S (R) continuously into itself, with respect to its natural
Frchet topology. It follows that all these operations extend by duality to S (R).
So does, in particular, the Fourier transform, which is an isomorphism of the triplet
(12.35) onto itself.
Coming back to the CWT, observe from (12.14) that it is an L2 inner product.
Thus the usual RHS reasoning applies: one of the two functions or s may be taken
as singular (a tempered distribution) provided the other one is correspondingly good.
This justifies the use of plane waves or -functions as signals, or that of a -function
as analyzing wavelet, as discussed in the previous section.
But we may go further. Indeed, the CWT in many ways provides a generalization
of the Fourier transform, so it is a natural question to ask whether it extends to
distributions, and maps the triplet (12.35) onto a similar one. The answer to both
questions is in fact yes, as shown by Holschneider [378, Hol95], that we follow
here. Of course the situation is somewhat more complicated here, since the wavelet
transform maps functions on R, even R+ in the progressive case, into functions on
the half plane R2+ .
The key tool is proving these properties is that of localization: Given the
asymptotic behavior of a signal s as x , what is that of its wavelet transform
S(b, a)? This aspect is particularly important in the present context. Indeed we are
talking here about localization in the half-plane G+ = {(b, a) : b R, a > 0}, and
the latter as a natural interpretation as phase space, as we will see in Sect. 12.4.1.
Thus what is involved here is localization in phase space, a key aspect of coherent
state theory.
In order to make precise the concept of localization, as introduced in [378,
Hol95], we need standard functions. For , 0, consider the functions
 
(1 + |x| )1 , x 0, 0, x > 0,
+ (x) = (x) = 1 (12.37)
0, x < 0, (1 + |x| ) , x 0,

and
12.3 A Mathematical Aside: Extension to Distributions 359

(x) = (x) + + (x), (x) := (x). (12.38)

A function s is called polynomially localized if |s(x)| c (x) for some .

Define also
 
1
( ) = + + ( ), ( ) := ( ). (12.39)

The function vanishes for < 0, and tends to zero as for 0 and as for
. A progressive function s (that is, s( ) = 0 for < 0) is called polynomially
band-limited or strip-localized if its Fourier transform is majorized by some :
s = O( ).
Then the basic localization result asserts that if the wavelet and the signal s are
both polynomially localized, resp. strip-localized, so is the wavelet transform T s.
Using this concept of localization, one may now build a RHS appropriate for the
CWT. Denote by S+ (R) the set of functions in S (R) whose Fourier transform is
arbitrarily well polynomially band-localized: for every > 0, there is a constant
c > 0 such that

|s(x)| c (x) and s( )| c ( ).

| (12.40)

It turns out that S+ (R) is a closed subspace of S (R), consisting exactly of those
functions whose Fourier transform vanishes for 0 (the progressive functions)
hence the notation is consistent. There is a natural topology on S+ (R), given by the
seminorms :
(+)
s = sup inf c  , (12.41)
 [0, ]

where the infimum is taken over all constants c  that verify (12.40). It is easy to
(+) (+) (+)
show that {  , > 0} is an ordered set of seminorms: s s if
. Thus S+ (R) becomes a locally convex topological vector space. A sequence
(+)
sn S+ (R) tends to zero if limn sn  = 0 for all 0 (or equivalently for a
directed set {m }).2
Now we turn to wavelet transforms. If both the wavelet and the signal s are
in S+ (R), the localization result quoted above tells us that the wavelet transform
S(b, a) is well localized both in scale and position; namely for every > 0, there is
a constant c > 0 such that
 
b
|S(b, a)| c (a) . (12.42)
1+a

2The set {m } is directed if, for any pair m , n , there is an element p such that m p and
m p .
360 12 Wavelets

Denote the space of such functions by S (G+ ). As above, this space becomes a
locally convex topological vector space when equipped with the seminorms:

(2)
S = sup inf c  . (12.43)
 [0, ]

The two spaces S+ (R) and S (G+ ) are for the CWT what Schwartzs space
S (R) is for the Fourier transform, as results from the following theorem.
Theorem 12.3.1. (i) Given a wavelet S+ (R), the wavelet transform T :
S+ (R) S (G+ ) is continuous with respect to the respective topologies and
the following estimate holds for > 0:

(2) (+) (+)

T s c  2( +1) s2( +1) , (12.44)

with some constant c > 0 depending on only.

(ii) Similarly the reconstruction operator R = T is continuous from S (G+ ) into
S+ (R), and one has the estimate, for > 1 and some constant c > 0,

(+) (+) (2)

R S c   S , (12.45)

Given Theorem 12.3.1, the extension to distributions is easy. First we define

the duals S+ (R), resp. S (G+ ), as the space of continuous antilinear maps on
S+ (R), resp S (G+ ). Continuity of F S+ (R) means, for instance, that there
exists c > 0 such that, for all s S+ (R), there exists 0 such that

(+)
| < F, s > | c s (12.46)

(here < , > is the sesquilinear form defining the duality). Clearly one has, exactly
as for the Schwartz triplet (12.35),

S+ (R) L2 (R+ , d ) S+ (R), (12.47)

and the duality form < , > may be seen as an extension/restriction of the inner
product of L2 (R+ ):

< h, s > = s | hL2 (G+ ) , h L2 (G+ ), s S+ (R). (12.48)

Notice that, as before, this relation provides a linear embedding of S+ into S+ .

The same triplet structure is present, of course, on the side of the transforms :

S (G+ ) L2 (G+ , a2 da db) S (G+ ), (12.49)

12.3 A Mathematical Aside: Extension to Distributions 361

We equip both duals with the weak*-topology. This means, for instance, that a
sequence Fn S+ (R) tends to zero in S+ (R) if

lim < Fn , s > = 0, s S+ (R). (12.50)

n

Now we are ready to extend the wavelet transform to distributions. Using

the reconstruction operator R = T , we may write, by definition, for any s
S+ (R), S (G+ ):

T s| L2 (G+ ) = s|R L2 (R+ ) . (12.51)

Thus, using the identification (12.48), we may define the wavelet transform of F
S+ (R) by duality as the distribution T F S (G+ ) given by

< T F, > = < F, R >, S (G+ ). (12.52)

Similarly, the reconstruction operator R is defined on S (G+ ) by the

relation:

< R , s > = < , T s >, s S+ (R). (12.53)

The outcome is twofold.

Theorem 12.3.2. (i) For any , S+ (R), the wavelet transform T and the
reconstruction operator R :

T : S+ (R) S (G+ ), R : S (G+ ) S+ (R)

are continuous for the respective weak*-topologies.

(ii) If is a reconstruction wavelet for the analyzing wavelet , then

R T = c I (12.54)
>
with c = R  ( ) ( ) | |1 d and I is the identity operator on S+ (R).

In addition, the localization results are extended to distributions. For a full analysis,
we refer to the original work of Holschneider [378, Hol95].
The conclusion is twofold. First, exactly as the Fourier transform, the continuous
wavelet transform on progressive functions extends by duality to the appropriate
spaces of distributions, including continuity. Second, since the continuous wavelet
transform is after all a convolution, it has a regularizing effect: the wavelet transform
of a tempered distribution is a C polynomially bounded function, exactly as in the
usual theory of tempered distributions on R or R2 . The only new fact is that here
the wavelet transform maps functions on R+ into functions on the half-plane G+ .
Otherwise all the results follow the standard pattern.
362 12 Wavelets

12.4 Interpretation of the Continuous Wavelet Transform

Besides the fact that the wavelet transform is a particular case of the CS machinery,
it possesses other aspects that lead to a fruitful interpretation, both from the
mathematical and the physical vantage points.

12.4.1 The CWT As Phase Space Representation

As we have discussed in Sect. 12.2, the continuous wavelet transform derives

from a UIR of the affine group Gaff or G+ , which is of the semidirect product
type. Hence the considerations developed in Sects. 8.1 and 8.3 and, in particular,
Sect. 9.2.2 apply, and the coadjoint orbits yield precious information concerning
the representations. More precisely, according to the theory of Kirillov [Kir76],
each coadjoint orbit corresponds to a unique (up to unitary equivalence) unitary
irreducible representation. On the other hand, each coadjoints orbit is a symplectic
manifold, precisely the kind of manifold that, in mechanics, is usually taken as the
phase space of a classical system, and starting point for the procedure of (geometric)
quantization. Thus we have to compute the coadjoint orbits of G+ [Gui84], an
exercise that will nicely illustrate the general treatment developed in Chaps. 8 and 9.
An easy calculation [57] shows that G+ has two kinds of coadjoint orbits. First an
infinite set of orbits of dimension 0, namely the single point sets {0, x}, x R. Then
two orbits of dimension 2, namely O = R R, with Lebesgue measure, which
correspond to the two representations U+ and U . Thus O+ is homeomorphic to
G+ . Writing = a1 , we see that the left Haar measure on G+ reduces to the latter:

dl (b, a) = a2 da db = d db. (12.55)

Thus (G+ , dl ) is a phase space associated to U+ , so that the CWT yields a phase
space realization of signals [58, 272]. This fact is physically significant, in that it
opens the way to the application of wavelets to quantum problems, in particular
in the quantization process [102, 103, 144, 161, 273, 415]. Similar considerations
hold true in higher dimensions, as we shall see in the next chapter. Note also that
the interpretation of a1 as a momentum variable is implicit in the identification
of the WignerVille transform W [s](b, a) with the Wigner distribution W(q, p), for
q = b, p = 1/a [see the comments after (12.4)].
Note that the Poincar half-plane we obtain here is a coadjoint orbit of
SL(2, R) SU(1, 1) too (see Sect. 4.2.2). This is related to the fact that the
representation U+ extends to a discrete series representation of SL(2, R), of which
G+ is a subgroup.
12.4 Interpretation of the Continuous Wavelet Transform 363

12.4.2 Localization Properties and Physical Interpretation

of the CWT

The efficiency of the CWT hinges on the interplay between two facts:
>
1. The admissibility condition for , reduced to (x) dx = 0;
2. The support properties of .
Consider the latter first. We have seen above that it is essential to require that be
square integrable, possibly integrable, and satisfy the admissibility condition. But
this is far from sufficient in practice.
On the contrary, one usually assumes and  to be as well localized as possible
(compatible with the FourierHeisenberg uncertainty principle, of course). More
specifically, assume that has an numerical support3 of width T , centered around
0, while  has an numerical support of width , centered around o . Then the
transformed wavelets ba and (ba have, respectively, a numerical support of width
aT around b and a numerical support of width /a around o /a. Notice that the
product of the two widths is constant (we know it has to be bounded below by a
fixed constant, by Fouriers theorem). Therefore:
If a , 1, ba is a wide window, whereas ( ba is very peaked around a small
frequency o /a : this transform will be mostly sensitive to low frequencies.
If a  1, ba is a narrow window and (ba is wide and centered around a high
frequency o /a : this wavelet has a good localization capability in the time
domain and is mostly sensitive to high frequencies.
Thus we have obtained a tool that reproduces the correlation between duration
and average frequency often encountered in real life signals, and announced in
Sect. 12.1: Low frequency portions of the signal tend to be long, whereas high
frequencies occur briefly in general.
Combining now these localization properties with the zero mean condition and
the fact that ba acts like a filter (convolution in x-space),
 +
S(b, a) = |a|1/2 (a1 (x b)) s(x) dx

 +
= |a|1/2  (a ) s( ) eib d ,

we see that the CWT performs a local filtering, both in position and in scale. The
wavelet transform S(b, a) is nonnegligible only when the wavelet ba matches the
signal, that is, it selects the part of the signal, if any, that is concentrated around the
time b and the scale a.

3 By this, we mean that the function is numerically negligible outside that region.
364 12 Wavelets

1
a 10 b 10 c

log2(a)

log2(a)
0.5 8 8

6 6
0
0 0.5 1 0 0.5 1 0 0.5 1
x x x

Fig. 12.2 Unraveling a fractal function: (a) The devils staircase; (b) Its wavelet transform (with
the first derivative of a Gaussian); (c) The corresponding skeleton (see Sect. 12.6)

In order to get a physical interpretation of this feature, we recall (see the

discussion after Theorem 12.2.2) that |S(b, a)|2 may be interpreted as the energy
density in the wavelet parameter space. Hence, if the wavelet is well localized,
the local filtering effect means that the energy density of the transform will be
concentrated on the significant parts of the signal. This is the key to all the
approximation schemes that make wavelets such an efficient tool.
Furthermore, combining the support properties of with the covariance of the
CWT under G+ , one sees that the CWT analysis has constant relative bandwidth:
/ = constant, contrary to the STFT analysis, which has a constant bandwidth,
= constant. This implies that it has a better resolution at high frequency, i.e.,
small scales: the CWT is a singularity detector.
In addition to its localization properties, the wavelet is often required to have a
certain number of vanishing moments, as we already mentioned in Sect. 12.2 above
[see (12.16)]. This condition determines the capacity of the wavelet transform to
detect and measure singularities. Indeed, if has all its moments vanishing up to
order N 1, then it is blind to polynomials of degree up to N. Equivalently, it detects
singularities down to the (N + 1)th derivative of the signal. This property is crucial
for a whole class of applications, namely the determination of local regularity of
functions or measures, more generally the characterization of singularities. Note
that the latter aspect often requires more specialized wavelets, such as chirps [466,
Jaf96, Tor95].
All taken together, the CWT may be called a mathematical microscope, with
optics , position b and global magnification 1/a. In addition, by its very definition,
the wavelet transform is an ideal tool for analyzing scale dependent features, in
particular fractals [92, Arn95]. A simple>
but striking example is that of the devils
staircase, that is, the function f (x) = 0x d , where is a uniform measure on the
triadic Cantor set. The wavelet transform of f exhibits in a transparent way the
fractal structure of the function, as shown in Fig. 12.2.
As we said above, the so-called wavelet spectrum |S(b, a)|2 may be interpreted as
the energy density of the signal. In fact a finer analysis may be performed with the
linear intermittency measure (LIM) obtained by normalizing the wavelet spectrum
to its translational average,
12.5 Discretization of the Continuous WT: Discrete Frames 365

|S(b, a)|2
LIM(b, a) :=  . (12.56)
|S(b, a)|2 db
R

Indeed, LIM(b, a) > 1 means that the signal is more active than average at (b, a).
This is the signature of the phenomenon of intermittency (see Sect. 12.7 for some
applications).
Finally, as we shall see in Chap. 14, all the considerations made in this section
extend to higher dimensions, where all together they contribute to make the CWT
into a remarkably efficient tool for image analysis (in 2 dimensions).

12.5 Discretization of the Continuous WT: Discrete Frames

The reproduction property (12.28) implies, as usual for CS, that the information
content of the wavelet transform S(b, a) is highly redundant. In fact the signal has
been unfolded from one to two dimensions, and this explains the practical efficiency
of the CWT for disentangling parts of the signal that live at the same time, but on
different scales. This redundancy may be eliminated this is the rationale behind
the discrete wavelet transform that we will discuss below. It may also be exploited,
in several ways. A first possibility is to obtain interesting interpolation properties,
which in fact rely on the Lie group structure. We will skip the point here and refer
the interested reader to [351]. What will concern us here is the observation that it
must be possible to obtain the full information about the signal from a small subset
of the values of the transform S(b, a), for instance, a discrete subset. In fact, the
actual (numerical) reconstruction of a signal from its wavelet transform requires the
discretization of the integral in (12.29). What is the minimal sampling grid ensuring
no loss of information?
The question may be answered at two levels. On the practical side, the minimal
grid may be determined from the reproducing kernel K. The latter is the autocorre-
lation function of and thus defines a correlation length in a and b. This approach
certainly yields a qualitative answer, which may be made quantitative by defining a
precise notion of resolving power [58, Chu92].
From a mathematical point of view, the answer lies in the theory of discrete
frames described in Sect. 3.4.1 [234, Dau92].
Let us explain how frames enter naturally in the picture (the reasoning is valid
for any linear transformation that yields a different representation of the signal, for
instance, the Windowed Fourier Transform). Let = {bk , a j , j, k Z} be a discrete
lattice in the (b, a)-half-plane G+ . We say that yields a good discretization if an
arbitrary signal s(x) may be represented as a discrete superposition

s(x) =  jk |s
 jk (x), (12.57)
j,k Z
366 12 Wavelets

where jk := bk a j and
 jk may be explicitly constructed from jk . We emphasize
that (12.57) must be an exact representation, i.e., there is no loss of information as
compared to the continuous reconstruction (12.29). Actually (12.57) means that the
signal s(x) may be replaced by the set { jk |s} of its wavelet coefficients. Thus we
consider the analysis operator

W : s  { jk |s}.

Since s L2 , it is natural to require that the sequence of coefficients be also square

integrable and that F be continuous from L2 (R) to 2 , i.e.,

| jk |s|2 Ms2 , M > 0. (12.58)

j,k Z

In addition, one wants the reconstruction of s(x) from its coefficients to be

numerically stable, that is, a small error in the coefficients implies a small error
in the reconstructed signal. In particular, if the l.h.s. of (12.58) is small, s2 should
be small also. Therefore, there must exist a constant m > 0 such that

m s2 | jk |s|2 M s2 (12.59)

j,k Z

(the lower bound indeed guarantees the numerical stability [Dau92]). In other words,
as indicated in Sect. 3.4.1, the set { jk } constitutes a discrete frame, with frame
bounds m and M.
Notice that it is precisely at this point that arises the basic difference between the
discretized CWT and the discrete wavelet transform (DWT), that we will discuss
briefly in Sect. 13.1. In the former case, the wavelet is chosen a priori (with very
few constraints, see above), and the question is whether one can find a lattice such
that { jk } is a frame with decent frame bounds m, M. In the other approach, one
usually imposes that the set { jk } be an orthonormal basis and tries to construct
a function to that effect. The construction is rather indirect and the resulting
function is usually very complicated (often it has a fractal behavior).
Of course the practical question is: How does one build a good frame? In view
of the discussion in Sect. 3.4.1, a good frame is a frame with a width as small
as possible, in order to ensure that truncation of the expansion (3.51) yields a good
approximation. Since this question is more general than the specific example of
wavelets, we will postpone the discussion to Chap. 17, where we will treat in parallel
various classes of CS, including wavelets.
12.6 Ridges and Skeletons 367

12.6 Ridges and Skeletons

Real signals are frequently very entangled and noisy, and their wavelet transform is
difficult to understand. However, a clever exploitation of the intrinsic redundancy of
the CWT is often able to bypass the difficulty. Now reduction to a discrete subset is
not the only way to exploit the redundancy. Another one, which in a sense is more
intrinsic, is the restriction of the transform to its ridges [248], that we are going to
describe now.
As a matter of fact, many signals are well approximated by a superposition of
spectral lines :

s(x) = Al (x)eil x , (12.60)

l

or, more generally, by a so-called asymptotic signal:

s(x) = Al (x)eil (x) , (12.61)

l

where the amplitude Al (x) varies slowly (in a logarithmic sense) with respect to the
phase l (x), namely,

1 dAl (x) dl (x)

 . (12.62)
Al (x) dx dx

Typical examples are spectra in NMR spectroscopy [248], that we shall discuss in
detail in Sect. 12.7.1.
For a signal of this kind, the wavelet transform (12.14) in the position domain is
a sum of rapidly oscillating integrals, and the essential contribution to each of them
is given by the stationary points of the phase of the integrand. Assume for simplicity
there is only one such point xs = xs (a), which is defined as a solution of the equation

dl o
(xs ) = , (12.63)
dx a

where o is a basic frequency of the analyzing wavelet (typically a Morlet wavelet

(12.19). In fact (12.63) means that the instantaneous frequency dl (x)/ dx coincides
with the scaled frequency o /a of the wavelet. Then the ridge of the wavelet
transform is defined as the set of points (b, a) for which xs (a) = b. These constitute
a curve in the (b, a)-half-plane and a detailed analysis shows that, on this curve,
the wavelet transform S(b, a) coincides, up to a small correction, with the analytic
signal Z(b) associated to s(x) [248, 354, Lyn82] (we recall that the analytic
signal associated to the signal s is obtained by subtracting the negative frequency
component of s [Lyn82, Pap02, Tor95]; this is done by taking the Hilbert transform
of s). In the general case, the wavelet transform of a signal has, of course, many
368 12 Wavelets

ridges, each of them being essentially a line of local maxima. The set of all ridges
is called the skeleton of the transform.
The conclusion is that the restriction of the wavelet transform S(b, a) to its
skeleton contains the whole information. In particular, the frequency modulation
law x1 arg{s(x)} of s(x) is easily recovered from the skeleton. Thus it is not
necessary to compute the whole wavelet transform, but only its skeleton. This is of
course much less costly computationally, because there are fast algorithms available.
More important, the ridge concept is quite robust in the presence of noise, so that
this technique will usually improve considerably the signal-to-noise ratio, thus the
efficiency of the wavelet analysis method. Spectacular examples may be found,
for instance, in the analysis of (multi)fractal curves [69, 92, Arn95], or that of the
fluctuations of the Earths magnetic field [9].

12.7 Applications

The CWT has found a wide variety of applications in various branches of physics
and/or signal processing. We will list here a representative selection; some of the
earlier ones may be found in the proceedings volumes [Com90, Mey91, Mey93],
with the original references. In addition, a survey of some physical applications is
given in the volume [Ber99]. More recent applications are scattered throughout the
specialized literature. In all cases, the CWT is primarily used for analyzing transient
phenomena, detecting abrupt changes in a signal or comparing it with a given
pattern. In certain cases, the nature or the geometry of the physical phenomenon
to be analyzed essentially imposes the use of wavelets, in which case one can speak
of intrinsic wavelet applications, a concept forged by Benedetto [133].
Sound and acoustics:
For historical reasons, the first applications of the CWT were in the field of
acoustics. A few examples are musical synthesis, speech analysis [Mae94] and
modeling of the sonar system of bats and dolphins. Other examples include
various problems in underwater acoustics, such as the disentangling of the
different components of an underwater refracted wave and the identification of an
obstacle (a submarine is a good example!). A beautiful example is the wavelet
auditory modelling (WAM) [133, 134]. The point here is that the structure of
the cochlea effectively dictates the use of wavelets for compression and noise
reduction algorithms, so that this wavelet application is indeed intrinsic.
Geophysics, meteorology:
This is the origin of the method, which was designed in an empirical fashion by
J. Morlet for analyzing the recordings of microseisms used in oil prospecting.
Later on, the CWT has been applied to the analysis of various long time series
of geophysical origin, e.g., in gravimetry (fluctuations of the local gravitational
field), in geomagnetism (fluctuation of the Earths magnetic field [9]) or in
astronomy (fluctuations of the length of the day, variations of solar activity,
12.7 Applications 369

measured by the sunspots, etc). Interesting results have been obtained in [220]
for the analysis of meteorological radar data, namely, the reconstruction of
continuous reflectivity densities and the analysis of discrete radar wind profiler
data. More recently, the CWT has been used systematically by Holschneider
and his collaborators for the polarization and dispersion analysis of two- and
three-component signals in geophysics and a dedicated software (Geophysical
Wavelet Library) has been designed (see the review paper [427], where refer-
ences to previous work can be found).
Fractals:
As mentioned above, the CWT is an ideal tool for studying fractals, or more
generally phenomena with particular properties under scale changes. Thus it is
quite natural that the CWT has found many applications in the analysis of (1-D
and 2-D) fractals, artificial (diffusion limited aggregates) or natural (arborescent
growth phenomena) [92,Arn95]. An interesting example of fractal or self-similar
behavior is that of telecommunications network traffic, and here too the WT
(although rather the DWT) has given interesting results [1].
Turbulence, solar physics:
Related to fractals is the use of the CWT in the analysis of 2-D developed
turbulence: identification of coherent structures, uncovering of hierarchical
structure, detection of intermittency [89, 274276, Abr97]. The latter is based
on the use of LIM, defined in (12.56). Exactly the same method has been used
recently in solar physics, namely, for discriminating between two scenarios for
hard X-rays in solar flares [252, 253].
Atomic physics:
When an atom is hit by a short intense laser pulse, it emits radiation that cover
a whole spectrum of harmonics of the laser frequency (experimentally, harmonic
order up to several hundreds have been observed). This is a fast and complex
physical process, which cannot be understood without a time-frequency analysis.
This has been done, both with STFT and wavelets, yielding for instance the time
profile of each individual harmonic [70, 71, 83] and the effect of the polarization
of the laser field on harmonic generation [84, 85]. This technique leads to the
controlled emission of ultrashort light pulses, in the 1018 s (attosecond) range,
a potentially very useful tool in many applications.
Spectroscopy:
This was one of the earliest and most successful applications, in particular for
NMR spectroscopy, where the method proved extremely efficient in subtracting
unwanted spectral lines or filtering out background noise, but also for a much
deeper analysis of spectra. We will discuss this example in more detail in
Sect. 12.7.1 below [54, 76, 78, 122, Suv11].
Analysis of local singularities:
The strong point of the CWT is to detect singularities in a signal, but it yields also
a fine characterization of their strengths, in particular in the case of oscillating
singularities [94, 97, 98].
370 12 Wavelets

Quantum cosmology:
A very recent application of the 1-D CWT concerns quantum cosmology [144].
As mentioned in Sect. 7.5, wavelet quantization regularizes the gravitational
singularity in a FriedmannLematre universe. Such quantization methods is the
subject matter of Chap. 11.
Edge detection and shape characterization:
A particular case of analysis of local singularities is the determination of the
shape of an object, a standard problem in image processing, for instance in
robotic vision. An original approach [69] consists in treating the contour of the
object as a complex curve in the plane and analyzing it with the 1-D CWT.
The method benefits from all the good properties of the wavelet transform, for
instance its robustness to noise, and has led to many applications (see [Cos01,
Chap. 7] for a general overview of multiscale shape characterization and [600]
for a review of wavelet-based edge detection methods).
Medical and biological applications:
The CWT has been used for analyzing or monitoring various electrical or
mechanical phenomena in the brain (EEG, VEP), the heart (ECG) or the
visual system [264, Ald96, Tho98]. It also yields good models for the auditory
mechanism [233]. A remarkable success is the statistical analysis of correlations
in DNA sequences, resolving a raging controversy among biologists [95, 96].
Engineering and applied science:
Here again the important aspect is monitoring, for instance in detecting anomalies
in the functioning of nuclear, electrical or mechanical installations, or testing the
behavior of materials under impact [593].

12.7.1 Application to NMR Spectroscopy

Before concluding this section, let us illustrate the use of the CWT by the example
of NMR spectroscopy.
The physical phenomenon may be described as follows. When a sample is placed
in a static magnetic field, nuclei with a magnetic moment align along this applied
field, resulting in a net magnetization. In its equilibrium state, the magnetization
is static and does not induce a signal in the receiver antenna. In order to obtain
information, one must first excite the nuclei with a radio frequency pulse. After such
a pulse, the magnetizations precess around the static field at angular frequencies
characteristic of their chemical environment and relax to their equilibrium state.
This precession induces a signal in the receiver antenna. The signal to be analyzed
is the Fourier transform (spectrum) of the damped response curve of the protons
(this is called an FID signal, for Free Induction Decay). It contains a large number
of narrow peaks, the spectral lines, but many among them are useless, coming for
instance from the protons of the solvent. These peaks, which may be quite big,
must be subtracted, and the position and amplitude of the relevant ones measured
12.7 Applications 371

x 105 x 105
10 10
a b

8 8

6 6
abs(FFT)

abs(FFT)
4 4

2 2

0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
frequency frequency

Fig. 12.3 Application of the CWT in NMR spectroscopy: Subtraction of an unwanted peak.
(a) The original spectrum; (b) The spectrum reconstructed after subtraction of the water peak

with precision. In addition, the spectra are often quite noisy, and must be cleaned
before any useful measurement can be performed.
The problem of suppression of the solvent peak is crucial in NMR spectroscopy,
in particular for the 1 H spectroscopy. As a consequence, various methods have
been designed for achieving it efficiently, along two different lines. One approach
is experimental, namely one submits the sample to a particular sequence of RF
pulses before the actual measurement (the so-called saturation recovery or selective
inversion recovery methods [Ala94, Gun94]). The other one consists in processing
the data after the measurement, by various techniques of signal analysis (for
instance, convolution in time-domain [429, 430, 463] or statistical methods linked
to singular value decomposition (SVD), see [167, 236, 507, 594] and references
therein). Compared with these standard methods, the wavelet technique (which
resorts to the second approach) is both highly efficient and simple to implement.
A typical analysis of NMR spectra with help of wavelets is given in Figs. 12.3
and 12.4 [122, 248, 354]. The first one is an example of peak subtraction. The
original spectrum (left) exhibits a huge parasite peak, due to the protons of the
solvent (water), that masks to a large extent the interesting structures. The analysis
consists in isolating this peak, on the CWT, subtracting it from the spectrum and
reconstructing the remaining part. The result (right) is a spectrum where all the
fine details are now clearly visible, and have not been perturbed by the removal of
the large peak. Indeed, the prominent structures appear at exactly the same place
on the frequency (horizontal) axis in both pictures. The reason for the remarkable
372 12 Wavelets

15000
a
abs(FFT)
10000

5000

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency

15000
b
abs(FFT)

10000

5000

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency

Fig. 12.4 Noise filtering in spectroscopy: (a) The original spectrum; (b) The spectrum recon-
structed after noise removal

efficiency of the method in this case is that the huge line and the rest of the spectrum
live at different scales, hence they are decoupled in the CWT (unfolding) and can
be readily separated with very little distortion [213].
Figure 12.4 is an example of noise filtering. The original signal (top) consists of a
number of damped sinusoids embedded in noise. The dominant peaks are localized
with help of the CWT ridge algorithm, the remnant of the spectrum is subtracted
and the filtered spectrum is reconstructed, using one of the reconstruction formulas.
The result (bottom) is quite spectacular, and compares well with standard methods
of noise suppression (signal filtering). This method has been applied successfully to
real life NMR spectra.
We may also remark that, in this problem as well as in the analysis of harmonic
generation in laser-atom interaction, both the wavelet and the STFT (Gabor)
analyses give very good, and in fact comparable results. The lesson of those studies
[54, 78, 122] is that it is more important to choose the adequate range of parameters
than to select a particular wavelet, or even a particular time-frequency method,
wavelets or STFT.
However, real life NMR spectroscopy demands much more than the rather
simple application just described. Indeed, proton magnetic resonance spectroscopic
(MRS) signals, typically acquired in the brain of a subject, either in vitro or in
vivo, have to be analyzed quantitatively. These signals contain many resonating
components whose frequencies are characteristic of the observed metabolites and
the chemical environment around them in a particular voxel, and amplitudes are
directly related to these concentrations of metabolites. Identifying and measuring
individual metabolite components is crucial for the evolving field of MRS for
12.7 Applications 373

NAA
Cr
Cho

Cr

Glx
mIns
Tau
Glx Lip CH2
+ Lac
Lip CH3

4.0 3.5 3.0 2.5 2.0 1.5 1.0

frequency (ppm)

Fig. 12.5 In vivo short-echo time MRS spectrum acquired at 4.7T in a mouse brain, superimposed
on a measured in vivo macromolecular spectrum (from [571])

clinical applications, since the relative abundance of certain metabolites is a clear

signature of certain brain disorders (tumors, epilepsy). To give an idea, we present
in Fig. 12.5 a typical MRS spectrum, measured in vivo, showing a large number of
overlapping peaks, with the symbols of the corresponding metabolites. The task is
now to evaluate the parameters of all these resonance peaks.
However, the analysis and quantification of these metabolite contributions is dif-
ficult due to the low signal to noise ratio, the number of overlapping frequencies and
the contamination of the signal of interest with water and a baseline originating from
macromolecules and lipids. The baseline is a major obstacle for MRS quantification
as its shape and intensity are generally not known a priori. In a first approach [571],
a CWT with the Morlet wavelet has been used for quantifying signals acquired in
vivo, that is, recovering parameters such as metabolite amplitudes, frequencies and
damping factors, while facing successively quantification challenges arising from
the non-Lorentzian lineshapes, overlapping frequencies, and noise or baseline.
Although the Morlet wavelet is quite efficient for obtaining such results, it is not
optimal for each kind of spectrum. It is capable of detecting single peaks in the
spectrum of a composite signal and providing parameter estimates to describe these
peaks, but it cannot separate different metabolite contributions. The problem is that
the Morlet wavelet is standard, but it is not directly related to the analyzed signal.
Already in [571], it was pointed out that other wavelets, with more resemblance
with the metabolites themselves, should improve the detection. To that effect,
the concept presented in [441] is to define wavelet functions directly from the
374 12 Wavelets

MRS data themselves, by the autocorrelation function of the metabolite profiles,

properly averaged to zero. In this way, one obtains wavelets, called metabolite-based
autocorrelation wavelets, with a spectrum that is very similar to the MRS spectrum
and one has a tailor-made tool for signal detection.
The technique runs as follows. Given an ergodic process time series s( ), its
autocorrelation function estimator Rss (t) is defined [Pap02] by
 
Rss (t) = s( )s( t) d = s( )s( + t) d . (12.64)

The autocorrelation function in the frequency domain Sss ( ), i.e., the Fourier
transform of Rss (t),4 can be evaluated by the WienerKhintchine relation:

Sss ( ) = F {Rss (t)} = 2 |
s( )|2 . (12.65)

This definition allows one to find the expression of the autocorrelation of a signal,
given its spectrum.
To start with, let us take as signal s(t) a truncated complex damped exponential
function, giving a single truncated Lorentzian spectral line (the phase is omitted
since it drops out in the autocorrelation function):

s(t) = AeDt ei1 t (t), A > 0, D > 0, (12.66)

where (t) is the Heaviside step function. This type of signals, which vanish for
t < 0, resembles a measured FID signal, i.e. the spectrum of the response curve of
the protons, where there are no available values before the start of the measurement.
Thus, computing the Fourier transform of (12.66), namely,

1 A
s( ) = , (12.67)
2 D + i( 1 )

we obtain the autocorrelation function of (12.66) by (12.65),

1 A2
Sss ( ) = . (12.68)
2 D + ( 1 )2
2

Note that the non-truncated Lorentzian, i.e., the signal (12.66) without the
function, cannot have an analytical autocorrelation function, since its Fourier
transform is a Dirac delta function.
Taking the inverse Fourier transform of (12.68) or using (12.64), we obtain the
autocorrelation function of the signal (12.66) given by

4 We return here to the (t, ) notations, more familiar in signal processing.

12.7 Applications 375

Cr 4 Cr
0.012 1.5 x 10

0.010

0.008 1
|Cr(ppm)|

|Cr(f)|
0.006

0.004 0.5

0.002

0 0
5 4 3 2 1 0 5 4 3 2 1 0
ppm ppm
7
6 x 10

2
Re{Cr(t)}

6
1000 500 0 500 1000
t [ms]

Fig. 12.6 Wavelet constructed from a Creatine (Cr) profile acquired in vitro: (Top left) Modulus
of the spectrum of the Cr profile; (Top right) Modulus of the Cr-based wavelet spectrum; (Bottom)
Real part of the Cr-based wavelet in the time domain

A2 D|t|+i1t
Rss (t) = e , < t < . (12.69)
2D
This function is localized and oscillating and can be used as a wavelet function
if it is admissible. In fact, it is almost admissible, since the correction term to be
added, as for the Morlet wavelet, is numerically negligible for realistic values of the
parameters and can be omitted. With this approximation, the wavelet associated to
the signal s(t) is defined as s (t) = Rss (t).
The CWT of the signal s(t) given in (12.66) with respect to the corresponding
wavelet s (t) is:

A21
S(b, a) = a s(b) .
D2 + [1 (a 1) + iaD]2

Clearly this CWT diverges as a 1 (the match between the wavelet and the signal
is too perfect!). By allowing different damping factors for the signal and the wavelet,
namely Ds and D, respectively, we get for a = 1:
376 12 Wavelets

CWT Naa: Max at scale=1 and =24 CWT Naa: Max at scale=0.81 and =16
0.8 0.8
scales

scales
1 1

1.2 1.2
0 100 200 300 400 500 0 100 200 300 400 500

horizontal ridges, thresh=0.4 horizontal ridges, thresh=0.4
0.8 0.8
scales

scales
1 1

1.2 1.2
0 100 200 300 400 500 0 100 200 300 400 500

Fig. 12.7 CWT using the in vitro NAA-wavelet on (left panels) the composite signal and (right
panels) the composite signal lacking the NAA contribution

|A21 |
|S(b, 1)| = eDs b . (12.70)
|D2 D2s |

and one can estimate the damping factor of the signal by

d
Ds = ln |S(b, 1)| . (12.71)
db
The technique can be extended to a sum of Lorentzian lineshapes, with some
additional complications and approximations, due to the interference terms between
different components [441].
The next step is to discretize the whole process and to derive the metabolite-
based wavelets directly from the available data, that is, metabolite profiles acquired
by MRS measurements or by simulation. These profiles are, by the very nature of
their acquisition, discrete sequences. As the resulting wavelets are discrete, too, they
require an extra step of up- and downsampling in order to perform a proper CWT,
namely, for allowing a dilation operation [441].
Since these wavelets are designed from a specific metabolite signal, the CWT
using one of them is supposed to be more sensitive to the metabolite from which
it was derived. In order to show this, we take as an example an in vitro Creatine
(Cr) profile. Figure 12.6 shows, in the top panels, the spectrum of the original
Creatine profile (after some preprocessing) and the spectrum of the derived Cr-
based wavelet, and in the bottom panel, the real part of the wavelet in the time
domain. Note that both spectra present a pair of very similar characteristic peaks.
The wavelet function so obtained is well localized and has zero mean, as required
for an admissible wavelet. The same operation is then performed for all the needed
metabolites and lipids. Because of the strong resemblance between a given wavelet
and its mother metabolite, that wavelet will be very efficient in detecting the
presence or absence of that metabolite in a mixture, by comparing the response at
the scale a = 1 with other scales, because it exploits a priori knowledge optimally.
12.7 Applications 377

To give a (spectacular) example, we show in Fig. 12.7 the analysis of two composite
signals. The left panels present, on top, the CWT of the composite signal using the
NAA-wavelet, and, below, the corresponding ridges (lines of maximum modulus).
In the same way, the right panels present the CWT and its ridges for the composite
signal, but the NAA contribution removed. Indeed, we see a strong ridge at scale
a = 1 when NAA is present (left panels), but that ridge is no longer present if the
NAA contribution is removed from the signal.
This technique exploits optimally the properties of wavelet analysis, whose
philosophy is always to exploit a maximum of a priori information. It could prove
very useful in practical implementations of MRs spectroscopy, in particular for
clinical applications. For a fully detailed exposition, we refer to the report [Suv11].
Chapter 13
Discrete Wavelet Transforms

Abstract This chapter is devoted to discrete wavelets. We start with the standard
version, related to multiresolution analysis, and some of its generalizations. Next we
extend the analysis to a group-theoretical approach to discrete wavelet transforms.
Starting from wavelets on the finite field Z p , we introduce pseudo-dilations and a
group structure. Then we generalize this approach to wavelets on a discrete abelian
group. Finally we discuss algebraic wavelets, by which we mean wavelets based
on different numbers, replacing, for instance, the dilation factor 2 by the golden
mean (we speak then of -wavelets) or arbitrary real numbers, which lead to Pisot
wavelets.

13.1 The Discrete WT

As we have seen in Sect. 12.5, the discretization of the CWT leads, among other
things, to the theory of frames. For many practical purposes of signal processing,
a tight frame is almost as good as an orthonormal basis. And if one stays with the
standard wavelets, as we have done so far, one cannot do better, these wavelets do
not generate any orthonormal basis (like the usual coherent states). However there
are cases where an orthonormal basis is really required. A typical example is data
compression, which is performed (in the simplest case) by removing all wavelet
expansion coefficients below a fixed threshold. In order not introduce any bias in
this operation, the coefficients have to be as decorrelated as possible, and of course
an orthonormal basis is ideal in this respect.
Fortunately, it turns out that there exists wavelets that do generate orthonormal
bases, and this leads to the discrete WT (DWT). The key step was the discovery
that almost all examples of orthonormal bases of wavelets can be associated to a
multiresolution analysis [457, Mey92], and furthermore that the whole construction
may be transcripted into the language of the so-called quadrature mirror filters
(QMF). We describe briefly these two developments in turn.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 379
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__13,
Springer Science+Business Media New York 2014
380 13 Discrete Wavelet Transforms

13.1.1 Multiresolution Analysis, Orthonormal Wavelet Bases

The key to the discrete WT is the notion of multiresolution analysis (MRA), so we

begin with a formal definition.
Definition 13.1.1. A MRA of L2 (R) is an increasing sequence of closed subspaces

. . . V2 V1 V0 V1 V2 . . . , (13.1)
 
with jZ V j dense in L2 (R) and jZ V j = {0}, and such that
1. f (x) V j f (2x) V j+1
2. There exists a function V0 , called a scaling function, such that {0,k , k Z}
is an orthonormal basis of V0 , where 0,k (x) = (x k).
Combining conditions 1 and 2, one gets an orthonormal basis of V j , { j,k (x)
2 j/2 (2 j x k), k Z}.
Remark. Actually, it is enough to require that {0,k , k Z} be a Riesz basis of V0 ,
that is, the span of {0,k , k Z} is dense in V0 and there exist constants m, M > 0
such that, for all c = (ck )kN 2 (N), one has
 2
 
 
m c22  ck 0,k  M c22 .
k=1 

Equivalently, 0,k = T fk , where { fk } is an orthonormal basis of V0 and T is a

bijective bounded operator from V0 onto itself.
Each V j can be interpreted as an approximation space: the approximation of f
L2 (R) at the resolution 2 j is defined by its projection onto V j . The additional details
needed for increasing the resolution from 2 j to 2 j+1 are given by the projection of
f onto the orthogonal complement W j of V j in V j+1 :

V j W j = V j+1 , (13.2)

and we have:
$ %


L (R) =
2
W j = V jo Wj , (13.3)
j Z j= jo

where jo is an arbitrary lowest resolution level. Then the theory asserts the existence
of a function , sometimes called the mother wavelet, explicitly computable from ,
such that, for each j Z, { j,k (x) := 2 j/2 (2 j x k), k Z} is an orthonormal basis
of the detail space W j , and thus { j,k (x), j, k Z} is an orthonormal basis of L2 (R):
these are the orthonormal wavelets. Like the { j,k } above, they are indexed by points
13.1 The Discrete WT 381

of the dyadic lattice {(k2 j , 2 j ), j, k Z}, adapted to the Lobachesky geometry of

the Poincar half-plane, following the pattern described in Sect. 12.5.
The construction of proceeds as follows. First, the inclusion V0 V1 yields the
relation (called the scaling or refining equation, or a two-scale relation):

(x) = 2 hn (2x n), hn = 1,n | , |hn |2 = 1 . (13.4)
nZ nZ

Taking Fourier transforms, this gives

1
(2 ) = m0 ( ) ( ), with m0 ( ) = hn ein . (13.5)
2 nZ

Thus m0 is a 2 -periodic function and it satisfies the relation

|m0 ( )|2 + |m0 ( + )|2 = 1, a.e.. (13.6)

By (13.5), this implies m0 (0) = 1 and thus m0 ( ) = 0. Iterating (13.5), one gets the
scaling function as the (convergent!) infinite product

( ) = (2 )1/2 m0 (2 j ), (13.7)
j=1

so that (0) = (2 )1/2 . By the same argument as in (13.4) and (13.5), a function
W0 V1 may be defined by the relation

 (2 ) = m1 ( ) ( ),
(13.8)

where m1 is another 2 -periodic function. As a consequence of the relation V1 =

V0 W0 and the orthonormality of the functions { j,k }, the functions m0 , m1 must
satisfy the identity

m1 ( ) m0 ( ) + m1 ( + ) m0 ( + ) = 0, a.e.. (13.9)

The simplest solution is to put

m1 ( ) = ei m0 ( + ) , (13.10)

which implies, in particular, |m0 ( )|2 + |m1 ( )|2 = 1, a.e.. Then one obtains

(x) = 2 (1)n1 hn1 (2x n) , (13.11)
nZ

or, equivalently, using the freedom in the choice of the solution to (13.9),
382 13 Discrete Wavelet Transforms

(x) = 2 (1)n hn+1 (2x n) . (13.12)
nZ

Then one proves that this function indeed generates an orthonormal basis with all
the required properties. Various additional conditions may then be imposed on the
function (hence on the basis wavelets): arbitrary regularity, several vanishing
moments (in any case, has always mean zero, as in the CWT), symmetry, fast
decrease at infinity, even compact support (see below).

13.1.2 Connection with Filters and the Subband Coding

Scheme

Actually, the discussion above means that we have translated the multiresolution
structure into the language of digital filters (by a filter, we mean a multiplication
operator in frequency space or a linear convolution in the time variable). For
instance, m0 ( ) is a filter, with Fourier coefficients hn , m1 ( ) is another one, and
{m0 , m1 } are called quadrature mirror filters or QMF whenever (13.9) is satisfied.
Then the various restrictions imposed on translate into suitable constraints on the
filter coefficients hn . For example, we will see below that has compact support if
only finitely many hn differ from zero (one then speaks of a finite impulse response
or FIR filter). The advantage of turning to that filter language is that efficient and
well understood algorithms are available, very similar to the so-called Laplacian
pyramid familiar in signal processing [Vet95].
Take indeed f V0 . Then, using the decomposition V0 = V1 W1 , one
may write:

f = c0,k 0,k
kZ

= c1,k 1,k + d1,k 1,k .

kZ kZ

From the orthonormality of the bases, one gets immediately:

c1,k = hn2k c0,n , d1,k = gn2k c0,n , (13.13)

nZ nZ

with hn = 0,n | 1,0  = 1,n | 0,0  (of course, 0,0 = ), gn = 0,n | 1,0 . (In
practice the signals are real-valued functions, so that the coefficients hn , gn are real
too.) Thus the sequences c1 = (c1,k ), d1 = (d1,k ) are uniquely determined from c0 by
application of the filters h and g, acting by convolution and decimation. The filter
h is a low-pass, g a high-pass filter, and they allow perfect reconstruction, namely
(simply by taking adjoints):
13.1 The Discrete WT 383

 
c0,n = hn2k c1,k + gn2k d1,k . (13.14)
kZ

In addition, it turns out that h and g are conjugate quadrature filters (CQF), a special
case of QMF, characterized by the relation gn = (1)n hn+1 , that is, h m0 and g
m1 , as defined in (13.5) and (13.10), respectively. The interpretation of the relations
(13.13), (13.14) is the following. If c0 describes the signal at a given resolution, then
c1 corresponds to its approximation at half the resolution and d1 to the additional
details needed for recovering the initial resolution.
Now the key point is that this operation may be iterated. The next approximation
c2 and the corresponding details d2 are obtained from the sequence c1 by the
same filters h and g, by relations analogous to (13.13). This procedure of splitting
repeatedly a given signal cn into its low frequency component cn+1 and its high
frequency component dn+1 is a standard technique in signal processing, known as
the subband coding scheme [Vet95]. In the case where the filter h (hence also g)
has finite length (finite number of nonzero coefficients hn ), one may visualize the
resemblance of this algorithm with the pyramidal one. Let L be the length of the
filter h. Then, at level j, one has:

c j,k = hn2k c j1,n , d j,k = gn2k c j1,n , (13.15)

nZ nZ

that is, every c2,k depends on L coefficients c1,k , each of which in turns depends on L
coefficients c0,k , and so on, and similarly for the detail coefficients d j,k . This indeed
leads graphically to a pyramid of nonvanishing coefficients. In other words, at
each resolution j, the wavelet coefficients may be obtained in terms of those at
lower resolution by adding finer and finer details, and this is obtained by repeated
application of the two filters h, corresponding to the scaling function (low-pass),
and g, corresponding to the wavelet (high-pass).
The construction made so far shows the validity of the implications: multires-
olution analysis orthonormal wavelet basis CQF. The crucial point is
the opposite implication: Under what conditions does a pair of CQF filters h, g
generate an orthonormal wavelet basis? This question was answered by Daubechies
[Dau92], in the form of a regularity condition (implying some vanishing moments).
In addition, she showed that, if one takes for h a filter with a finite number of nonzero
coefficients hn , then one generates an orthonormal basis of wavelets with compact
support.
However this is not sufficient to guarantee the efficiency of the method, because
the rapidity of the algorithm depends crucially on the length of the filters involved,
as for pyramidal schemes in general. This remark opens the way to various
improvements, to which we will come back in Sect. 13.1.3 below.
On the other hand, it is clear from the definition of a multiresolution analysis that
the general translation covariance is lost. The dyadic lattice is invariant only row by
row, and the row j = jo is invariant under discrete translations by 2 jo k, k Z. This
feature creates many difficulties in applications, in particular, for pattern recognition
or texture analysis (see the discussion in [Mal99, Sect. 5.4]).
384 13 Discrete Wavelet Transforms

13.1.3 Generalizations

As we just saw, appropriate (CQF) filters generate orthonormal wavelet bases.

However, this result turns out to be too rigid and various generalizations have been
proposed.
1. Biorthogonal wavelet bases
As we mentioned in Sect. 12.2, the wavelet used for reconstruction in the con-
tinuous wavelet transform need not be the same as that used for decomposition,
the two have only to satisfy a cross-compatibility condition. The same idea
in the discrete case leads to biorthogonal bases, i.e., one has two hierarchies
of approximation spaces, V j and Vj , with cross-orthogonality relations. This
gives a better control, for instance, on the regularity or decrease properties of
the wavelets [209].
2. Wavelet packets and best basis algorithm
The construction of orthonormal wavelet bases leads to a special subband
coding scheme, rather asymmetrical: each sequence c j gets further decom-
posed into c j+1 and d j+1 , whereas the detail sequence d j is left unmodified.
Thus more flexible subband schemes have been considered, called wavelet
packets where both subspaces V j1 and W j1 are decomposed at each step
[210, 211, Mey92, Wic94]. They provide rich libraries of orthonormal bases,
and also strategies for determining the optimal basis in a given situation.
3. The lifting scheme: Second generation wavelets
One can go further and abandon the regular dyadic scheme and the Fourier
transform altogether. Using the lifting scheme of Sweldens [572, 573],
one obtains the so-called second-generation wavelets, which are essentially
custom-designed for any given problem. This approach uses the fact that, in
the biorthogonal scheme, a given wavelet fixes its biorthogonal partner only
up to a rather arbitrary function. Then one starts from a very simple wavelet
and gradually obtains the needed one by choosing successive appropriate
biorthogonal partners.
4. Integer wavelet transforms
In their standard numerical implementation, the classical (discrete) WT
converts floating point numbers into floating point numbers. However, in
many applications (data transmission from satellites, multimedia), the input
data consists of integer values only and one cannot afford to lose information:
only lossless compression schemes are allowed. Subsequent developments
have produced new methods that allow one to perform all calculations in
integer arithmetic [176].
13.2 Towards a Fast CWT: Continuous Wavelet Packets 385

13.1.4 Applications

Finally, as far as applications are concerned, wavelet (bi)orthogonal bases and

wavelet packets may be used for most problems previously treated with the
continuous wavelet transform (for precise references to early applications, see
for instance the conference volumes [Mey91, Mey93]; more recent applications
are too scattered in the literature to be given here). Their main virtue is a
remarkable efficiency in data compression. For achieving useful rates, one has to
determine which information is really essential and which one may be discarded
with acceptable loss of signal quality. High compression rates have been achieved
with the discrete wavelet transform, especially when wavelets are combined with
vector coding. Significant results have been obtained also in speech analysis (signal
segmentation, analysissynthesis, recognition).
Another field where wavelet bases have brought spectacular progress is
numerical analysis. Very fast algorithms may be designed for matrix multiplication,
and this has opened many doors.
Similarly, wavelet bases have been applied successfully to a whole class of hard
problems in pure mathematics, such as the construction of universal unconditional
bases for many function spaces (L p , Sobolev, Besov, etc.), efficient resolution of
partial differential equations, analysis of singular integral operators (Caldern
Zygmund operators), etc.
Finally, expansion into orthogonal or biorthogonal wavelet bases has brought a
whole new perspective in ab initio structure calculations in atomic and in solid state
physics. For a survey of these developments, see [Ber99, Chap.8], [91,202,249,326]
and the review [339].

13.2 Towards a Fast CWT: Continuous Wavelet Packets

Besides the full discretization described in Sect. 12.5, and the discrete WT just
discussed, there is an intermediate procedure, introduced in [268], under the name
of infinitesimal multiresolution analysis. It consists in discretizing the scale variable
alone, on an arbitrary sequence of values (not necessarily powers of a fixed ratio).
This leads to fast algorithms that could put the continuous wavelet transform on the
same footing as the discrete wavelet transform in terms of speed and efficiency, by
extending the advantages of the latter to cases where no exact QMF is available. Let
us sketch the method. Further details may be found in [Tor95].
Instead of the standard L2 -normalization used so far, it is more convenient
 to
choose the L1 -normalization, namely to use b,a (x) = a1 a1 (x b) . Then,
given a wavelet , normalized to c = 1, one lumps together all the low frequency
components in a scaling function
 ! x " da 1  x 
 ( ) = da
(x) = = (s) ds,
 (a ) , (13.16)
1 a a2 x 0 1 a
386 13 Discrete Wavelet Transforms

and one introduces the integrated wavelet

 1 ! x " da 1  2x  1
 ( ) = da
(x) = = (s) ds,
 (a ) . (13.17)
1/2 a a2 x x 1/2 a

These functions satisfy two-scale relations:

(x) = 2 (2x) (x),  ( ) =

 ( /2)
 ( ). (13.18)

Next, one chooses a regular grid, as opposed to the dyadic one used in the discrete
case, namely:

 j = 2 j (2 j ( x)),
xj xj = 2 j (2 j ( x)). (13.19)
x,2

Although the resulting transform will be redundant, it has the great advantage over
the conventional DWT of maintaining (integer) translation covariance. Then, exactly
as in (13.3), one gets a discrete reconstruction formula:

s(x) = xjo | s + xj | s. (13.20)
j= jo

Then assume there exists two functions 0 , 1 satisfying the following relations,
analogous to (13.5), (13.8),

 ( ),
 (2 ) = 0 ( )  (2 ) = 1 ( )
 ( ), a.e. (13.21)

These functions are not necessarily 2 -periodic. However, since using the regular
grid means sampling (x) at unit rate, we have to assume that the function 
is essentially supported in [ , ]. Therefore, since the functions 0 , 1 always
appear in a product with  , according to the relations (13.21), it is reasonable to
approximate the functions 0 , 1 in a neighborhood of 0 by 2 -periodic functions
m0 , m1 . In fact it can be shown [Tor95] that there exists a unique pair m0 , m1 that
minimizes the quantities
 1/2
(i , mi ) =  ( )|2 d
|(i ( ) mi ( )) , i = 0, 1,
R

namely

 ( + 2k )
kZ  (2 + 4k )
m0 ( ) = , (13.22)
 ( + 2k )|2
kZ |

 (2 + 4k )
 ( + 2k )
kZ
m1 ( ) = . (13.23)
 ( + 2k )|2
kZ |
13.3 Algebraic Wavelets 387

These approximate filters m0 , m1 , which are called pseudo-QMF, satisfy the identity
m0 ( ) + m1 ( ) = 1, a.e..
More flexibility is obtained if one subdivides the scale interval [1/2, 1] into n
subbands, by ao = 1/2 < a1 < . . . < an = 1. In that case one ends up with one
scaling function (x) and n integrated wavelets i (x), i = 0, . . . n1, corresponding
to integration from ai1 to ai . An additional improvement consists in periodizing the
signal and computing filters m0 , m1 of the same length as the signal. The resulting
pyramidal algorithm has a complexity equal to one half of the traditional fast Fourier
transform value O(N log22 N). Thus one obtains a very fast implementation of the
continuous wavelet transform, truly competitive with the discrete wavelet transform,
both in 1-D and in 2-D [592, Van98].

13.3 Algebraic Wavelets

There is a different way of generalizing the discrete WT, namely by replacing the
usual natural numbers, and the dyadic numeration underlying the multiresolution
approach, by another system of numeration.
A simple though unusual example is
based on the golden mean = 12 (1 + 5). We shall give here some indications on
that construction, following [313] (see also the first edition of the present volume,
Sect. 13.4). Two interesting aspects emerge. One is a genuine generalization of the
standard multiresolution, which yields a corresponding orthonormal wavelet basis,
the so-called -Haar basis. The other is the occurrence of a quasiperiodic structure
at each multiresolution level, instead of the usual lattice structure. More general
cases can be envisaged also, namely wavelet systems based on the so-called Pisot
numbers, first studied in [317].
We begin by sketching the construction of the Haar basis on the line, obtained
by using as scaling factor the irrational number = 12 (1 + 5), instead of the usual
factor of 2 [313]. The algebraic nature of , based on the equation 2 = + 1, entails
the -adic property
1 1 1
= j+1 + j+2 , j Z. (13.24)
j
This equation provides a subdivision of the unit interval into two parts
1 1
A = [0, 1] = [0, ] [ , 1]. (13.25)

Equation (13.25) is the starting point of an iterative sequence of subdivisions of A
into -adic intervals of the type

b b+1
A j,b = , , j N, A0,0 = A, (13.26)
j j
388 13 Discrete Wavelet Transforms

A j,b = A j+1, b A j+2, 2 b+ , (13.27)

where b is called a -integer [313], a concept derived from the system of numeration
based on the irrational [148,172,520]. Explicitly, the set of nonnegative -integers
is the following (lexicographically ordered) strictly increasing sequence of real
numbers:

Z+
= {0, 1, , , + 1, , + 1, + , , . . .}.
2 2 3 3 3 4
(13.28)

The set Z of -integers is then defined by

Z = Z+ +
(Z ). (13.29)

The numbers in Z+ are the nodes of a quasiperiodic chain, called the Fibonacci
tiling of the positive real line with two types of tiles, a long one with length 1 and a
short one with length 1/ . Define the increasing sequence of finite sets

B j = {x Z+
: 0 x < }, j N.
j
(13.30)

Then one shows that the -adic interval A j,b = [b j , (b + 1) j ] appears at a

certain step of the subdivision process (13.27) starting with (13.25) if b satisfies
the condition b B j1 , which means that b Z+ and 0 b 1.
j

Another approach to this subdivision procedure consists in using affine semi-

group actions (a similar technique was used in [121] for the analysis of 1-D
quasicrystals). We denote again by (b, a) the generic element of the affine group Gaff
of the line described in Sect. 12.2, with action on R given by (b, a)x = ax + b. Then
the generic interval A j,b in (13.26) is the result of an affine action on the original
A = A00 :

A j,b = (b j , j ) A. (13.31)

We now come to the explicit construction of the -Haar basis in L2 [0, 1]. For that
purpose, we need a mother wavelet, called here -Haar wavelet.
Definition 13.3.1. The -Haar wavelet is the function on the real line R defined as
1/2
, for 0 x 1/ ,
h (x) = 1/2 , for 1/ < x 1, (13.32)

0, otherwise.

One can verify directly that

 1  
1 1
(h (x)) dx = 2 + 1
2
= 1, (13.33)
0
13.3 Algebraic Wavelets 389

and that
 1
1 1
h (x) dx = 1/2 1/2 2 = 0. (13.34)
0

Thus h is orthogonal to [0,1] A , the characteristic function of the unit interval.

Following the standard pattern in wavelet analysis, we can now assert
Theorem 13.3.2. (1) The system

{A (x), j/2 h ( j x b) hj,b (x); j N, b B j1 } (13.35)

is an orthonormal basis in L2 [0, 1].

(2) The system

{ j/2 h ( j x b), j/2 h ( j x + b + 1); j Z, b Z+

} (13.36)

is an orthonormal basis of L2 (R), called the -Haar basis of L2 (R).

The proof, both for orthogonality and completeness, is based on the properties of
the -adic intervals A j,b = [b j , (b + 1) j ].
It is remarkable that the concept of multiresolution may also be extended to this
environment. From this, one can define more general wavelets, called -wavelets,
which have all the expected properties, namely
Definition 13.3.3. A -wavelet is a function (x) L2 (R) such that the family of
functions

{ j,b (x) := j/2 ( j x b), j,b1 (x) := j/2 ( j x + b + 1)},

where j is an arbitrary integer and b is an even positive -integer, i.e., b Z+

, is
an orthonormal basis in the Hilbert space L2 (R).
The -Haar wavelets are an example of -wavelet, but few examples beyond Haar
are known at the moment [106]. Clearly a more tractable requirement is to replace
the orthonormal basis in Definition 13.3.3 by a Riesz basis, for instance, a basis
of spline wavelets. Spline wavelets associated to the Fibonacci sequence have been
constructed in [320, 321].
More general algebraic wavelets may be obtained by replacing by any real
number > 1. Indeed, for such a , there exists a numeration system based
on the so-called Rnyi -expansion of real numbers [520]. Examples of such
a construction have been given in [317] for the case of the so-called Pisot
Vijayaraghavan (for short, PV or Pisot) numbers. Then this whole formalism may
be extended to two or more dimensions, leading to orthonormal or Riesz bases of
algebraic wavelets in the plane [319] or in Rn [147, Ber98]. For details of these
constructions, we refer to the original papers or (partially) to the first edition of the
present volume.
390 13 Discrete Wavelet Transforms

Note that in a parent context related to aperiodic structures, the concept of

multiresolution analysis has been introduced in crystallography [271] for analyzing
and classifying 2-D pure-point diffraction spectra (i.e., sets of Bragg peaks) of
certain self-similar structures with scaling factor > 1, like quasicrystals. The
diffraction spectrum is viewed as a point set in the complex plane in which
each point is assigned a positive number, its Bragg intensity. Then, by using a
nested sequence of self-similar subsets called beta-lattices, it becomes possible
to implement a multiresolution analysis of the spectrum . This analysis yields
a partition of at once in geometry, in scale, and in intensity (fingerprint
of the spectrum, not of the diffracting structure itself!). The method has been
experimented through numerical explorations of pure-point diffraction spectra of
various mathematical structures and also on the diffraction pattern of a realistic
model of quasicrystal [271].

13.4 A Group-Theoretical Approach to Discrete Wavelet

Transforms

13.4.1 Wavelets on the Finite Field Z p

Although the construction of algebraic wavelets just described may be given a

group-theoretical interpretation, in terms of what we could call the -adic affine
group G = {(b j , j ), b a -adic integer, j Z, }, this is not the point of view
generally adopted. On the contrary, the construction of the discrete WT described in
Sect. 13.1 above, as well as its biorthogonal generalization [Dau92] are incompatible
with the group-theoretical structure. The set of points {(k2 j , 2 j ), j, k Z} G+
defining the above basis is not a subgroup of G+ (but it is a subgroup of the dyadic
affine subgroup {(k2 j , 2l ), k, j, l Z} G+ .) However, there exists a construction
that goes a long way towards a group-theoretical understanding of the DWT, namely
wavelets on the finite field Z p [281] (note that the field Z p is sometimes denoted
by F p in the mathematical literature [Lan93]). Since this is an immediate and
interesting application of the CS formalism, which, moreover, leads to far-reaching
generalizations, we sketch it here.
Instead of R, one considers the set Z p = Z/pZ of remainders modulo p. When
p is a prime number, Z p is a field, with addition and multiplication modulo p; the
elements of Z p are then simply {0, 1, . . . , p 1}. The affine transformations of Z p ,
viz. n  an + b, a Zp , b Z p , are well-defined and form a group called the affine
group G p , with the usual group law of Gaff (taken modulo p). The group G p has a
natural unitary representation on 2 (Z p ):
13.4 A Group-Theoretical Approach to Discrete Wavelet Transforms 391

!nb"
(U(b, a) f )(n) = fba (n) := f . (13.37)
a

Let E denote the closed subspace of 2 (Z p ) defined by

p1
E = { f 2 (Z p ) : f (n) = 0}, (13.38)
n=0

so that 2 (Z p ) = C E. This decomposition reduces the representation U, and the

restriction of U to E is irreducible (square integrability is automatic here, since G p
is finite). Accordingly, given E, a wavelet transform may be defined, as usual,
as the linear map T : 2 (Z p ) 2 (G p ) given by

p1
(T f )(b, a) = ba | f  = ba (n) f (n). (13.39)
n=0

Noting that ba (n) = (a f )(n b), where (a f )(n) = f (a1 n) is the discrete
dilation operator acting on 2 (Z p ), we may also write this wavelet transform as:

(T f )(b, a) = (a
 f )(b), (13.40)

where  (n) := (n).

The map T is an isometry from E to 2 (G p ), up to the constant c = p  22 (Z p ) ,
and thus it yields an inversion formula:

1
f (n) =
c (T f )(b, a)ba (n), f E. (13.41)
(b,a)G p

A similar result exists for the full space 2 (Z p ).

When applied to explicit signals, this discrete WT leads to efficient decom-
positions and reconstructions. More interestingand surprisingis the result of
[281], according to which a discrete signal over Z p may be interpreted as a
sampled version of a continuous one. This interpretation requires the introduction
of deformed dilations (pseudodilations), and these lead automatically to an
algorithmic structure very similar to the multiresolution structure (the so-called
algorithme trous [Com90]). Similar constructions have been performed for
wavelets on an arbitrary finite field F [397], and on Z as well [75] (see also [49]
for a review).
The key word here is sampling, which transform a continuous signal into a
discrete one. This brings us to the general discretization problem, that we shall
discuss in Chap. 17.
392 13 Discrete Wavelet Transforms

Interestingly, a group structure also emerges in the lifting scheme approach,

mentioned in Sect. 13.1.3 (iii). In that case, one gets matrix groups and, more
generally, an algebraic structure called a group lifting structure [169].

13.4.2 Wavelets on Z p : Pseudodilations and Group Structure

Because Z p is a finite field, the dilation by a mod p, a : f (n)  f (a1 n),

preserves the number of nonzero values of the wavelet: the set of nonzero values
{ f (a1 n), n Z p } is just a permutation of the set { f (n), n Z p }. Therefore the
discrete dilation operator a preserves the length of the support of f , unlike the
continuous dilation operator used in Chap. 12 (this remark is the starting point of
the algorithme trous introduced in [383]). Thus the discrete case described in
Sect. 13.4.1 cannot be the sampled version of the continuous case.
In order to correct this situation, one introduces pseudodilation operators [281]

Da = Ka a , (13.42)

where Ka is a bounded operator on 2 (Z p ). Then the crucial lemma reads as follows:

Lemma 13.4.1. Let Tb and a denote, respectively, the translation and dilation
operators on 2 (Z p ):

(Tb f )(n) = f (n b),

b, n Z p , a Zp ,
(a f )(n) = f (a1 n),

and define the Fourier transform as

p1
f(k) = e2 ink/p f (n).
n=0

Then the operators (b, a) = Tb Da , where Da is the pseudodilation (13.42), form a

representation of the affine group G p if and only if Ka is a convolution operator,

Da = Fa a , (13.43)

where the filter Fa satisfies the compatibility relations

Faa = Fa a Fa (13.44)

or, equivalently, in Fourier space,

Faa (k) = Fa (k) Fa (ak). (13.45)

13.4 A Group-Theoretical Approach to Discrete Wavelet Transforms 393

The corresponding representation of G p is unitary if and only if

|Fa (k)| = 1, for all a Zp , k Z p .

Functions Fa that satisfy (13.44) and (13.45) are called compatible filters, in analogy
with the notion of filter sketched in Sect. 13.1.2.
Given a compatible filter Fa , the corresponding wavelet transform is defined as
in (13.40):

(T f )(b, a) = (Da  (Fa f ))(b),

 f )(b) = (a (13.46)

where  (x) = (x).

As a result of Lemma 13.4.1, one gets for this new WT the same algorithmic
structure as in the usual multiresolution analysis described in Sect. 13.1.1:

 (Faa f )(b)
(T f )(b, aa ) = aa
 (a Fa (Fa f ))(b).
= aa (13.47)

Indeed, writing the same relation on 2 (Z) with a = 2 j1 , a = 2 (the structure is the
same), one gets:
! "
 2 j1 F2 (F2 j1 f ) (b),
(T f )(b, 2 j ) = 2 j (13.48)

in which one recognizes the standard pyramidal algorithm of the Discrete Wavelet
Transform (see Sect. 13.1.2). Indeed these relations mean that the computation of
the wavelet coefficients at scale 2 j can be performed through a pyramidal algorithm
involving only dilated copies of two filters, F2 (which stands for the low-pass filter)
and (band-pass filter). Since only the true dilation a is involved, all the filters
have constant length, and one has a fast algorithm.
In conclusion, every compatible filter Fa gives rise to an efficient wavelet
transform on Z p . It remains to characterize these compatible filters, and the relations
(13.44), (13.45) show that this is a group cohomology problem, namely Fa is a one-
cocycle over the multiplicative group Zp , with values in the p-dimensional torus
T p = {(c1 , c2 , . . . , c p ), |c j | = 1}, in the unitary case, in C p in general.
Actually unitary equivalent representations of G p yield the same transform, so
one needs only to characterize the unitary dual G  p or, correspondingly, equivalence
classes of cocycles Fa . In other words one has to identify the first cohomology group
H 1 (Zp , T p ), or H 1 (Zp , C p ) in the nonunitary case (see Sect. 13.5.3 for general
definitions).
Surprisingly, it turns out that the general solution of this problem may be obtained
explicitly [397], once one notes that the multiplicative group Zp = Z p \ {0} is
isomorphic to the cyclic group Z p1 of order p 1. Let us give a concrete example
for p = 5:
394 13 Discrete Wavelet Transforms

. Z5 = {0, 1, 2, 3, 4} with addition mod 5

. Z5 = {1, 2, 3, 4} with multiplication mod 5
. Z4 = {1, i, 1, i} = {1, , 2 , 3 } with = i = ei2 /4 .
Then the correspondence

1 1, 3 3
2 , 4 2

gives the required group isomorphism between Z5 and Z4 .

In these terms, the result of [397] is that the group H 1 (Z p1 , T p ), which labels
the nonequivalent unitary cocycles, is itself isomorphic to Zp Z p1 . Indeed the
most general cocycle may be described as follows. Let v be a function on Z p such
that |v(n)| = 1, i.e., v : Z p T . Then Zp acts on such functions as:

(a v)(n) = v(a1 n), a Zp , n Z p .

Lemma 13.4.2. Let be a generator of the cyclic group Z p1 Zp and let a =

( )
r Zp . Then the functions Fa defined as

( ( )
r1
Fa (k) = ( j v)(k), k Z p ,
j=0

are compatible filters giving rise to unitary pseudodilation representations of G p .

The elements j v are to be viewed as functions Z p T and v must satisfy the
relations

v(k) = 1 and v(0) p1 = 1.

kZ p

( )
Moreover, these functions Fa , Z p1 , are all unitary compatible filters.
Exactly the same result holds for an arbitrary finite field F, as shown in [397],
but we refer the reader to the original paper for the details.

13.5 Wavelets on a Discrete Abelian Group

The case of Z p (or more generally F) is still particular, in the sense that the dilations
a form a multiplicative abelian group Zp which acts on 2 (Z p ). Thus, as stated in
Sect. 13.4.1, the affine group G p is the semidirect product Z p  Zp , and the general
theory of group-related CS applies.
13.5 Wavelets on a Discrete Abelian Group 395

This is not the case in the usual multiresolution context: there is no dilation
group acting on 2 (Z) (the space of discrete signals), only an abelian multiplicative
semigroup A = {2 j , j = 0, 1, 2, . . .}. Thus a purely group-theoretical formulation of
the discrete WT is not available (except the group-theoretical structure underlying
the lifting scheme [169], but this is completely different). However, the results
of [75] go a long way towards this goal and effectively provide the missing link
between the DWT and the CWT. We will conclude this chapter with a rather detailed
survey of these results, because we feel they are both important and promising.
As often, it is more instructive to treat a general abstract case, the multiresolution
situation being just a (particularly simple) example. Thus we consider a locally
compact abelian (LCA) group G, a lattice in G and an abelian semigroup A with
unit (called a commutative monoid in the mathematical literature [Lan93]) acting on
(in the multiresolution case, G = R, = Z, and A = {2 j , j = 0, 1, 2, . . .}). We use
the notions and the terminology introduced in Sect. 4.4. Notice that transforms on
LCA groups have been considered before, in [342] for the Gabor case, and in
[380, Hol95] for the wavelet case. The setup of the latter is a hierarchy of three
LCA groups

J H G,

with H and J discrete, G/H compact (that is, H is a lattice) and H/J finite. A typical
relevant example is G = R, H = Z, J = 2Z. However the aim of [380, Hol95] is
to generalize the notion of QMF and particular attention is paid to the sampling
problem. In the paper discussed here, on the contrary, the goal is to provide an
abstract, group-theoretical setting for the pyramidal algorithms of the DWT, using
the cohomological language introduced in [281, 397].

13.5.1 Compatible Filters: The General Case

In order to make things precise, we start by defining the action of a semigroup A

on the lattice contained in the LCA group G.
Definition 13.5.1. (1) Let A be a semigroup with unit element 1 and a lattice
in G. An action of A on is a mapping (a, x) A  ax , such that:
(i) a(a ) = (aa ) , a, a A , ; and (ii) 1 = , . The action of
A is one-to-one if for any a A , the mapping  a is one-to-one.
(2) We say that the action of A on is compatible (with the group law) if

a(  ) = (a )(a  ), a A , ,  . (13.49)
396 13 Discrete Wavelet Transforms

(3) Assume the action of A on is compatible and one-to-one. We say that ao

divides if there exists a (unique)  such that = ao  . In such a case
we write ao | , and  = a1
o .

Note that a1  
o only denotes an element of such that ao = ; this does not
mean that ao is invertible in A .
From now on, we shall assume that the action of A on is compatible and
one-to-one. In addition, we suppose that for each a A , a = 1, a is a lattice of
G and the quotient Qa = /a is a nontrivial finite group of order d(a). The action
of A on induces an action on , the dual action, in the natural way: for any
, , and a A , a ,  =  , a . The dual action is compatible, but
not necessarily one-to-one. For convenience, we shall call sequences the functions
on .
As discussed in the previous section, there is no natural group of dilations acting
on 2 (Z), only semigroups A Z+ , and these have a compatible and one-to-one
action. As in the case of Z p , we first consider the natural dilation on 2 (Z).
Definition 13.5.2. The natural dilation by a A is the bounded operator a ,
defined on 2 ( ) as follows: for any sequence {u , } 2 ( ),
@
ua1 , if a| ,
(a u) = (13.50)
0 , otherwise.

 each a A , we thus obtain a map a which dilates an arbitrary sequence u =

For
u by inserting zeros.
The action of a in the Fourier domain is as follows: for all u 2 ( ),

(
a u( ) = u
(a ),  . (13.51)

These maps a (a A ) satisfy a number of obvious properties. For instance,

a a = aa , a, a A , (13.52)
a T = Ta a , where T is the translation by . (13.53)

Combining (13.52) and (13.53), we get:

(T a )(T  a ) = T +a  aa , (13.54)

which shows that ( , a) = T a is a representation of the composition law

( , a)(  , a ) = ( + a  , aa ),

characteristic of a semidirect product. Indeed we might say that the underlying

structure is the semidirect product GA =  A of the abelian group by the
multiplicative semigroup A (both discrete).
13.5 Wavelets on a Discrete Abelian Group 397

The natural dilation a is the analogue of the familiar dilation trous used
in the discrete WT [383], and it is also full of holes, namely zeros, a fact which
prevents a nice sampling interpretation. Consider for instance 2 , the natural dilation
by a factor 2 on 2 (Z). The main difference between 2 and the continuously defined
D2 lies in the fact that, by construction, half of the coefficients of a sequence dilated
using 2 vanish: (2 f )2k+1 = 0, k Z. Therefore, a sequence dilated with 2
can hardly be interpreted as a sampling of a continuously defined function, dilated
with D2 .
More generally, the fundamental difference between these Da and a is as
follows: given a finite sequence {u }, the measure of the support of a dilated
sequence a u, does not change, since the measure we use on is the counting
measure.
As in the case of Z p , the solution is to replace the natural dilations a by
pseudodilations, defined in the same way.
Definition 13.5.3. Let A be an abelian semigroup with unit, acting on a lattice
G, in such a way that the action is compatible and one-to-one. A pseudodilation
(or principal pseudodilation) on 2 ( ) is a bounded operator

Da = Ka a , (13.55)

where Ka is a bounded linear operator acting on 2 ( ), in such a way that

Da Da = Daa , a, a A . (13.56)
Da T = Ta Da , a A , (13.57)

Notice that (13.56), (13.57) are identical to (13.52), (13.53) above, and thus one may
suspect the occurrence of a representation of GA . Indeed, exactly as for Z p (and for
the same reason), the operator Ka must be a convolution operator, as follows from
the following lemma, the exact parallel of Lemma 13.4.1.
Lemma 13.5.4. The operator Da = Ka a is a pseudodilation, in the sense of
(a)
Definition 13.5.3, if and only if Ka is the convolution by some sequence {h , }
satisfying the compatibility condition
! "
(aa )  (a )
= h(a) a h(a ) = h a  h  , a, a A .
(a)
h (13.58)



Thus we have, for all u 2 ( ),

! "

(a)
(Da u) = h a  u  = h(a) (a u) , (13.59)


398 13 Discrete Wavelet Transforms

and we shall write:

Da = h(a) a , aA . (13.60)

Introduce now, as described in Sect. 4.4, the (group) Fourier transform Ha = h( (a)

of the sequences h(a) , which is a function on the dual  of . Under the Fourier
transform, the relations (13.59), (13.60) read

(
D u(a ), .
a u( ) = Ha ( ) (13.61)

Therefore the compatibility condition (13.58), which simply translates (13.56),

reads in Fourier space,

Haa ( ) = Ha ( )Ha (a ), . (13.62)

In other words, we have recovered for the filters Ha exactly the cocycle
equations (13.44) and (13.45).
In the usual multiresolution approach, one uses in fact two filters, a low-pass filter
h and a high-pass filter g, satisfying some identities, for instance the QMF relations
discussed in Sect. 13.1.2. Thus, in addition to the pseudodilation Da , which lead to
the filters Ha , we need another ingredient. As mentioned above, we assume that,
for all a = 1, Qa is finite and nontrivial, so that d(a) > 1. We now introduce the
following
Definition 13.5.5. Given a principal pseudodilation D, an associated pseudodila-
tion on 2 ( ) is a mapping D  which assigns to any pair a, a A , a = 1, the family
of linear operators D a,a ; , = 1, . . . , d(a ) 1, of the form

a,a ; = K
D a,a ; aa , (13.63)

a,a ; is a bounded linear operator acting on 2 ( ), such that

where K

a,a ; T = T(aa ) D
D a,a ; , a, a A , , (13.64)
a ,a ; = D
Da 0 D a a ,a ; , a0 , a1 , a A . (13.65)
1 0 1

 is completely determined by the

Notice that an associated pseudodilation D
operators

D 1,a; ,
a; := D a A , a = 1, = 1, . . . d(a) 1

and the relation (13.65).

Exactly as in the case of principal pseudodilations (Lemma 13.5.4), one then
shows that each K a,a ; must be the convolution operator by a sequence g(a,a ; ) , so
13.5 Wavelets on a Discrete Abelian Group 399

that the associated pseudodilation reads

a,a ; u = g(a,a ; ) (aa u) .

D (13.66)

(a,a ; )
In addition, the sequences {g , } must satisfy (and may actually be
generated by) the following compatibility equations:
 
g(a0 a1 ,a ; ) = h(a0 ) a0 g(a1 ,a ; ) . (13.67)

Let G(a,a ; ) denote the Fourier transform of the sequence g(a,a ; ) . In the Fourier
domain, the compatibility relations (13.67) read

Ga0 a1 ,a ; ( ) = Ha0 ( )Ga1 ,a ; (a0 ) . (13.68)

These arguments motivate the following definition, generalizing the corresponding

one for Z p given in Sect. 13.4.2:
Definition 13.5.6. A family {(Ha , Ga,a ; ), a, a A , a = 1 , = 1, . . . , d(a ) 1}
of bounded functions on  satisfying the conditions (13.62) and (13.68) is called a
family of compatible filters.
Therefore, we can summarize the results of this section as follows:
Theorem 13.5.7. Let D be a principal pseudodilation, and D  be an associated
pseudodilation. Then they are necessarily filters of the form (13.60) and (13.66),
and the corresponding functions on  {(Ha , Ga,a ; ), a, a A , = 1, . . . , d(a )1}
constitute a family of compatible filters.
Since h(a) can be recovered from Ha by inverse Fourier transform, the complete
description of the Da is obtained if one can completely classify the solutions
of (13.62).
To give a trivial example, consider the natural pseudodilations on , defined by
the sequences:

(a) (a; )
h = ,0 , g = , , = 1, . . . d(a) 1, (13.69)

and define the other sequences g(a,a ; ) using (13.67). It is readily verified that
the corresponding family of functions Ga,a ; , together with the functions Ha = 1
occurring in the case of natural dilations, yield a family of compatible filters. Other
explicit examples may be found in [75].
At this point, it is useful to introduce the adjoints of the principal and associated
pseudodilation operators Da , D a,a ; and define the synthesis operators:

d(a)1
Sa = Da Da + a; D
D  .
a; (13.70)
=1
400 13 Discrete Wavelet Transforms

Then one says that a family of compatible filters defines a subband coding scheme
if Sa = 1, for all a A . Indeed, this condition guarantees that the signal may be
reconstructed perfectly from its transform.
Hence, given such a subband coding scheme, one may decompose any s 2 ( )

as follows. Let s1 = s, and define the sequences sa , a A and d a,a ; inductively by

sa = Da s ,   sa
d a,a ; = D (13.71)
a ;

Then we have the wavelet-like decomposition (compare Sect. 13.1.2):

d(a)1
s = Da s a + a; d a,1;
D
=1
d(a)1 d(a)1
a,a; d a,a; + a; d 1,a;
2
= Da 2 s a + D D
=1 =1
= ...

As an example, let us take the natural pseudodilations. Then we easily verify that
with the sequences given in (13.69), we have

s , if a| ,
(Da Da s) =
0 , otherwise,

and

a; D
 s) = s , if a|( + ),
(D a;
0 , otherwise.

In such a case, it is clear that Sa = 1, so we have again a subband coding scheme.

We notice that this example is the abstract version of the scheme used in [380] to
generate families of perfect reconstruction QMF. Let us also note that, exactly as in
the continuous case (Sect. 13.1.3), this analysis may be extended to a biorthogonal
scheme, which consists in decoupling the decomposition and reconstruction steps,
by using different families of compatible filters in those two steps.

13.5.2 Compatible Filters in the Case G = R, A Z+

We specialize now to the case of interest in classical one-dimensional wavelet

theory, namely G = R and = Z. Let A be any subsemigroup of the multiplicative
semigroup Z+ . In order to make more explicit the fact that we are now dealing
with integers, we shall in this case denote the elements of A by n. The usual
13.5 Wavelets on a Discrete Abelian Group 401

multiresolution case corresponds to n = 2 j . Let {(Hn , Gn ), n A } be a system of

compatible filters. After identifying the dual T of Z with the interval [ , ), the
functions Hn and Gn on T are now considered as 2 -periodic functions.
We first notice that, given a 2 -periodic function , and setting, for all n A ,

(n )
Hn ( ) = , (13.72)
( )

we immediately obtain a solution of (13.62), provided that the quotients are well-
defined and define bounded 2 -periodic functions for all n A . We call such
families trivial, for reasons that will become clear soon. We notice that (13.72) is
reminiscent of the relationship (13.5), (2 ) = m0 ( ) ( ), between the Fourier
transform of the scaling function and the low-pass filter in classical multiresolution
theory. However, the identification is not perfect, for ( ) is 2 -periodic, unlike
the Fourier transform of the scaling function. However, any bounded function
L (R), such that the quotient (13.72) defines a bounded 2 -periodic function
for all n A , also yields a solution to the compatibility relation (13.62).
More generally, functions Hn ( ) of the form

(n )
Hn ( ) = n , R, (13.73)
( )

also satisfy the compatibility relations (13.62). This comes from the fact that
(nn ) = n n , in other words that n  n is a character of the semigroup A .
Another simple example of such a trivial family of compatible filters is
provided by the theory of spline functions. Let us again consider any subsemigroup
A Z+ , and define, for all n A ,

(n) n1 , if 0 k n 1,
hk = (13.74)
0, otherwise.

A direct calculation yields

1 1 ein
Hn ( ) = . (13.75)
n 1 ei

It is easily verified that such a function satisfies (13.62), since it is of the

form (13.73), with ( ) = 1 ei . Now, for each = 1, . . . , n 1, let
 
2
Gn; ( ) = Hn . (13.76)
n

2
=0 sin ( + /n)
Using the trigonometric identity n1 = n2 sin(n )2 , valid for

all n Z+ and R, we easily see that, for all n A ,
402 13 Discrete Wavelet Transforms

d(n)1
|Hn ( )|2 + |Gn; ( )|2 = 1. (13.77)
=1

This relation tells us that we have constructed a subband coding scheme, called
the Haar multiresolution analysis in the classical wavelet literature [Dau92, Mal99,
Vet95].
Actually, the Haar MRA is the simplest instance of the so-called spline mul-
tiresolution analyses [Chu92]. Spline filters of order N correspond to filters Hn of
the form
 N
1 1 ein
Hn ( ) = N , (13.78)
n 1 ei

and Ga; filters given in (13.76). Since the filters Hn again assume the form (13.73),
the compatibility condition (13.62) is verified by construction. However, the perfect
reconstruction condition (13.77) is not satisfied any more. This problem is usually
bypassed by going to a biorthogonal scheme, as alluded to above.

13.5.3 Cohomological Interpretation

Let us come back now to the general case. As for Z p , the relation (13.62)
satisfied by the filters defining principal pseudodilations may be given an interesting
cohomological interpretation. We refer to [75] for details and to [Asc72] or [Hil71]
for the terminology and mathematical background.
As before we assume that the semigroup A operates on  in such a way that the
Haar measure  on  is A -quasi-invariant. Let M denote the commutative ring of
all complex-valued functions on  such that ( ) = 0 almost everywhere on .
Then the mapping

(a, )  a , a A , M with (a )( ) = (a ) for all , (13.79)

defines a compatible action of A on M .

For each integer r 1, we define r-cochains as functions H r : A . . . A
M that map ordered n-tuples of elements of A into M . We denote by C r :=
C r (A , M ) the set of all n-cochains. For n = 0, we define

C 0 (A , M ) := M

i.e., the 0-cochains are the elements of M . The set C r is an abelian group under the
composition law (U rV r )(a1 , . . . , ar ) = U r (a1 , . . . , ar )V r (a1 , . . . , ar ). The semigroup
A acts on C r by
13.5 Wavelets on a Discrete Abelian Group 403

[a(H r )] (a1 , . . . , ar ) = a (H r (a1 , . . . , ar ))

and one can easily check that this action is compatible with the group law of C r .
Next one introduces, in the standard way, the so-called differentials

r : C r (A , M ) C r+1 (A , M ),

which are defined as follows:

r+1 2 3(1)r+i
r (H r ) = pir+1 (H r ) , (13.80)
i=0

where pir+1 (H r ) is given by

(a1 H r )(a2 , . . . , ar+1 ), if i = 0,
pr+1 (H )(a1 , . . . , ar+1 ) = H r (a1 , . . . , ai ai+1 , . . . , ar+1 ), if 0 < i < r + 1,
i r
r
H (a1 , . . . , ar ), if i = r + 1.
(13.81)
and the inverse of a cochain is simply the inverse of the function. It turns out that r
is a group homomorphism for each integer r 0, and

Im (r r1 ) = {1}.

We define

Z r (A , M ) = Ker r , B r (A , M ) = Im r1 .

Following the usual terminology, the elements of Z r (A , M ) are called r-cocycles

and those of B r (A , M ), r-coboundaries. For each r 1, the quotient group

H r (A , M ) = Z r (A , M )/B r (A , M )

is the r-th cohomology group of A in M , and it measures the deviation from

exactness of the sequence [compare Sect. 4.5.3 (3)]

r1
C 0
0
C 1
1
C 2 . . . C r1 C r
r
C r+1 . . .

In fact, for characterizing the compatible filters, we only need the first cohomology
group. Indeed, an explicit calculation gives that, for all M and a A ,

(a )
[0 ] (a)( ) = , for all , (13.82)
( )

and, for all H C 1 (A , M ) and a, a A ,

404 13 Discrete Wavelet Transforms

Haa ( )
[1 H] (a, a )( ) = , . (13.83)
Ha ( )Ha (a )

Comparing (13.83) with (13.62), it is clear that each family {Hn : n A } of

elements of M is uniquely identified with an element H C 1 (A , M ). Then,
according to (13.83), an element H C 1 (A , M ) defines a compatible family of
filters {Hn }, that is, a solution of the cocycle equation (13.62), if and only if
H Ker 1 , and a trivial solution if and only if H Im 0 , by (13.82). In addition, we
want the filters Hn to be bounded, which we write H L (A , ). Then we obtain
the following
Proposition 13.5.8. The set of all families of compatible filters indexed by A is
given by Ker 1 L (A , ).
Given an element H of C 1 which defines a family of compatible filters, it is
interesting to know how to define explicitly the functions Hn for a given a A .
The answer is known for the particular class of solutions of the form (13.82), that
we have called trivial. More generally, the complete classification of compatible
filters would require to compute H 1 (A , M ), which indexes classes of solutions of
the cocycle equation modulo trivial solutions. This problem is still open in abstract
wavelet analysis, even in the simple case G = R, A = {2 j , j = 0, 1, 2, . . .}. But it is
also interesting by itself from the point of view of cohomology theory. Conversely,
one may also hope that the tools of cohomology theory will help to classify the
wavelet filters.

13.5.4 Compatible Filters and Discretized Wavelet Transform

As a last point, we show how the pseudodilation operators and compatible filters
occur naturally in the context of discretization of the continuous wavelet transform.
We shall see that discretization appears as a sort of generalized intertwining operator
between the canonical action of the affine group on L2 (R) and the action of
an affine semigroup on 2 (Z). As a byproduct, this approach also provides a
geometrical interpretation of the pyramid algorithms which have been developed
for the computation of the wavelet coefficients, and a generalization for an arbitrary
integral scale.
Up to now, we have only worked at the level of sequences defined on a lattice .
To proceed further, our first step will be to transport the action of the filters on L2 (G).
In order to do that, we need additional assumptions. From now on, we will assume
that the semigroup A acts on G also, and that this latter action is compatible,
one-to-one and transitive, i.e., for all g G and a A , there is some g G
such that g = ag . Finally, we assume that the Haar measure = G on G is
A -quasi-invariant (see Sect. 4.1). Let us denote by a (g) = d (ag)/ d (g) the
corresponding RadonNikodym derivative (formally, in the notation of (4.7), we
13.5 Wavelets on a Discrete Abelian Group 405

could write a (g) (a1 , g), but this does not make sense, since A is only a
semigroup; see the remark after Definition 13.5.1). The function a satisfies the
usual cocycle condition (4.8): for all a, a A and g G,

aa (g) = a (g)a (ag) . (13.84)

Note that the translation invariance of the Haar measure implies that a (ag ) =
a (ag) for all a, a A , g G, .
Let now D and D  be a principal and associated pseudodilations, and let Ha and
Ga,a ; be the corresponding filters, assumed to be bounded. Let us introduce the


following operators, acting on L2 (G):

Ua f (g) = a (g) h f (ag ) ,

(a)
(13.85)

(a,a ; )
Va,a ; f (g) = aa (g) g f (aa g ) . (13.86)

These operators are the L2 (G) analogues of the discrete operators Da and D a; ,
and they enable us to transport the action of the pseudodilations onto L2 (G). The
following is a direct consequence of the compatibility relations and the cocycle
condition (13.84):
Corollary 13.5.9. The operators Ua and Va,a ; satisfy

Uaa = Ua Ua , Va,a ; = Ua V1,a ; . (13.87)

In the Fourier domain, the operators Ua and Va; V1,a; read

(
U 
a f (a ) = Ha ( ) f ( ), V 
a; f (a ) = Ga; ( ) f ( ),
.
G (13.88)

Finally, the counterpart of the perfect reconstruction property is given by the

following
Corollary 13.5.10. Assume that the filters Ha and Ga; generate a subband coding
scheme. Then we have the equality

d(a)1
Ua Ua + Va; Va; = 1.
=1

These properties motivate the following definition.

Definition 13.5.11. Let {Ha , a A } be a family of filters satisfying the compati-
bility condition (13.62). A scaling function associated to the family {Ha , a A } is
a square integrable function L2 (G) which is invariant under the operator Ua , for
all a A .
406 13 Discrete Wavelet Transforms

 one can then

Given a scaling function (g) and an associated pseudodilation D,
construct the associated generalized wavelets, defined by

a; (g) = Va; (g) , (13.89)

and develop an abstract theory of wavelets. Instead of doing so, we rather come back
to the usual situation G = R, = Z.
Let A Z+ be a multiplicative semigroup, whose elements we denote again
by n. The Lebesgue measure (which is also the Haar measure) on R is A -quasi-
invariant, and the corresponding RadonNikodym derivative is

n (t) = n , nA,t R.

Let {(Hn , Gn,n ; ) : n, n A , = 1, . . . d(n ) 1} be a system of compatible filters

labelled
>
by A . Let us assume there exists a scaling function L1 (R) L2 (R),
with (t) dt = 1, satisfying the following two requirements:
Stability: the functions (t k), k Z form a frame of their closed linear span V .
Refinability:

(t) = n hl (at l), n A .

(a)
(13.90)
l

Then it follows from (13.90) that for all [compare with (13.8)],

(n ) = Hn ( )( ) , n A , R, (13.91)

that is, is a scaling function in the sense of Definition 13.5.11.

Notice that formally, (13.91) and the normalization condition define the func-
tion . Indeed, for all n A , n = 1, we may write, like in (13.7),
     

( ) = Hn Hn 2 Hn 3 . . .
n n n

However, the convergence of such infinite products to nice L2 functions is often

problematic, although it has been proved in the standard multiresolution case, see
(13.7).
Let us now consider an associated pseudodilation D  with filters {g(n,n ; ) }.
According to (13.89), we introduce the family of (square integrable) functions
{ n; (t)}, defined by

(n; )
n; (t) = (Vn; ) (t) = n gl (at l) . (13.92)
l

In the Fourier domain, we have

13.5 Wavelets on a Discrete Abelian Group 407

! " ! "
 n; ( ) = Gn;
 , (13.93)
n n
where we have set
(n; ) ik
Gn; ( ) = gk e .
kZ

Using the scaling function and the wavelets n; , we shall now make the
connection between compatible filters and the discretized version of the continuous
wavelet transform.
As stated in (12.14), the continuous wavelet transform on L2 (R) is given by

1
(T f )(b, a) = Tb Da | f  = & (a1 (x b)) f (x) dx , (13.94)
|a| R

where Tb , Da denote the translation, resp. dilation operator in L2 (R). Instead of the
usual multiresolution discretization described in Sect. 12.5, we want to compute the
following transforms, for any discrete semigroup A Z+ :

1
(Tm; f ) (b, n) = m; (n1 (t b)) f (t) dt , (13.95)
n R

where b Z and n A is any multiple of m A , and = 1 . . . d(m) 1.

We are also interested in

1
(S f ) (b, n) = (n1 (t b)) f (t) dt , n A , b . (13.96)
n R

The key point is the following. Given any (finite energy) sequence s 2 (Z), the
stability
>
(or frame) property tells us that one can always find f L2 (R) such that
sk = f (t) (t k) dt. In other words, using the terminology of [380], the sequence
s = {sk , k Z} may be obtained by sampling imperfectly some f with respect to
the lattice Z and the function , that is,

s = Z ( f ) = Z f , (13.97)

where Z : L2 (R) 2 (Z) is the (perfect) sampling operator, (Z h)k = h(k), k Z,

and (t) = (t). Let us consider the Fourier transform of s, given by s( ) =
kZ sk eik . Then the Poisson summation formula (4.128) yields:

(
s( ) =  f ( + ) = ( + 2 l) f( + 2 l) . (13.98)
Z lZ
408 13 Discrete Wavelet Transforms

Let us now set

! "
snk = (Dn s)k , dkn; = D s , (13.99)
n;
k

and more generally, for n, m A , and = 1, . . . d(m) 1,

! " ! "
dkn,m; = D s = D
n,m;
 D s .
m; n (13.100)
k k

The action of the adjoint pseudodilation Dn in the Fourier domain reads

1 n1
(
D s( ) =
n Hn (n1 ( + 2 l)) s(n1 ( + 2 l)) .
n l=0
(13.101)

Then, using (13.98) and the 2 -periodicity of Hn , we get

 n1
1
snk =
2 n
eik Hn (n1 ( + 2 l)) s(n1 ( + 2 l)) d
l=0

1
= eikn f( )Hn ( ) ( ) d
2

1
= eikn f ( ) (n ) d , by (13.91).
2

In other words

1 1
snk = f (t) (n1t k) dt = (S f )(nk, n) . (13.102)
n R n

Similarly, we obtain, using (13.93),


1
dkn; = eink f( ) Gn; ( ) ( ) d
2 R

1 1
= f (t) n; (n1t k) dt = (Tn; f )(nk, n) . (13.103)
n R n

More generally, we obtain from (13.92)

1 ! "
dkn,m; = (Tm; f ) (nmk, nm) = D s .
n,m; (13.104)
nm k

In other words, the wavelet coefficients dkn,m; of a discrete signal s 2 (Z), obtained
in the discrete scheme based on (A , Z), may also be computed by sampling the
continuous WT corresponding to the scaling function and the wavelets a; .
Conversely, the sampled values of the continuous wavelet coefficients may be
13.5 Wavelets on a Discrete Abelian Group 409

obtained by imperfect sampling by the function followed by (the adjoint of) a

pseudodilation. This makes the announced connection between compatible filters
and continuously defined wavelets. We summarize this result as a theorem:
Theorem 13.5.12. Let {(Hn , Gn,m; ) : n, m A , m = 1, = 1, . . . d(m) 1} be a
family of compatible filters such that there exists a (scaling) function L1 (R)
L2 (R) such that

(n ) = Hn ( )( ) , n A , R. (13.105)

Let n; be the wavelets defined by

 n; (n ) = Gn; ( )( ) ,
R. (13.106)

Then the sampled values of the corresponding wavelet coefficients may be

obtained by imperfect sampling by the function followed by (the adjoint of) a
pseudodilation:
1 1 ! "
(Tm; f ) (nmk, nm) = Tnmk Dnm m; | f  = D s , k Z, n, m A ,
n,m;
nm nm k

(13.107)
where s = Z ( f ) = Z f . Similar statements hold for the projection on the
scaling function,
1 1
(S f )(nk, n) = Tnk Dn | f  = (Dn s)k . (13.108)
n n

The theorem may also be reformulated by stating that the following diagram is
commutative, where the top line lives in L2 (R) and the bottom one in 2 (Z):

L2(R): f - ( m;k f )(,nm)

f 1 X nmZ
XZ nm

? ?

2(Z): s - dn,m;k
~
n,m;k

The commutative diagram above suggests that the (normalized) sampling oper-
ators 1a aZ act as generalized intertwining operators between the representation
of the affine group on L2 (R) and the representation of the affine semigroup Z  A
on 2 (Z). This may offer a different point of view on the harmonic analysis of the
affine group.
410 13 Discrete Wavelet Transforms

In conclusion, this approach yields a general setting for multiresolutions

associated to the action of a semigroup A on a lattice , especially when A is
the semidirect product of the group of lattice translations by a discrete abelian
semigroup of dilations. The tool is the notion of compatible filters, built on
pseudodilations.
Transporting the action of the pseudodilations to the case of continuously
defined functions yields structures which are extremely close to multiresolution
analyses. In particular, the pseudodilations provide the algebraic setting for the
pyramid algorithms familiar in discrete wavelet analysis. The pyramid algorithm for
computing recursively scaling function coefficients is nothing but a recursive use of
a pseudodilation (more precisely the adjoint of a pseudodilation). The corresponding
wavelet coefficients are obtained via the (adjoint) associated pseudodilations.
The outcome is a reformulation, in a very general group-theoretical framework,
of discrete wavelet transforms based on multiresolution. Such algebraic structures
associated to multiresolution pave the way towards generalizations. The latter
may for example include (separable or non separable) wavelet bases associated
with nontrivial dilations in more than one dimension (see Sect. 14.4), or the
wavelet bases associated to aperiodic tilings developed for studying quasicrystals
in [147, 319, Ber98].
Chapter 14
Multidimensional Wavelets and Generalizations

Abstract The continuous wavelet transform can be extended to arbitrary

dimensions and this is the topic of this chapter. We begin with the general
mathematical analysis, with some emphasis on the distinction between isotropic
and directional wavelets. Next we particularize to 2-D, the most important case
for applications in image analysis, discussing its distinctive properties and some
applications. Finally we describe in some detail a number of generalizations, such
as multiselective wavelets, ridgelets, curvelets and shearlets.

14.1 Going to Higher Dimensions

Exactly as in one dimension, multidimensional wavelets may be derived from the

similitude group of Rn (n > 1), consisting of dilations, rotations and translations. Of
course, the most interesting case for applications is n = 2, where wavelets have
become a standard tool in image processing, including radar imaging [Mey91,
Mey93]. Also n = 3 may have a practical importance, since there are physical
phenomena that are intrinsically multiscale and three-dimensional. Typical exam-
ples may be found in fluid dynamics, for instance the appearance of coherent
structures in turbulent flows, or the disentangling of a wave train in acoustics. In such
cases, a 3-D wavelet analysis is likely to yield a deeper understanding [104]. The
same comment is valid for the applications of the WT in quantum physics (quantum
mechanics, atomic physics, solid state physics, etc). For the latter, we refer to the
survey volume [Ber99, Chap.8], [91, 202, 249, 326] and the review [339].
More important, the derivation of the CWT in dimensions larger than 2 requires
in certain cases (axisymmetric wavelets) the general CS construction developed at
length in the previous chapters, because the natural parameter space is no longer
the similitude group itself, but rather a coset space (although here we are still in the
GilmorePerelomov setting, the dividing subgroup is the isotropy subgroup of the
wavelet). This opens the door to a host of other generalizations, namely wavelets
on manifolds on which there is a group action defined. We will discuss below

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 411
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__14,
Springer Science+Business Media New York 2014
412 14 Multidimensional Wavelets and Generalizations

several instances, namely wavelets on spheres and various kinds of time-dependent

wavelets. The latter could prove to be useful tools for motion tracking, including in
relativistic situations (here the relevant group would be the Galilei or the Poincar
group augmented with dilations, that is, the corresponding affine groups; we will
cover these examples in Sect. 16.2). It should be noted that, in most of these cases,
including wavelets on spheres, one needs the full-fledged extension of the CS
method, beyond GilmorePerelomov.

14.2 Mathematical Analysis

14.2.1 The CWT in n Dimensions

An n-dimensional signal of finite energy is represented by a complex-valued

function s L2 (Rn , dx). The natural operations, usually applied to a signal s, are
obtained by combining three elementary transformations:

(i) Translation: Tb : s(x)  s(x b), b Rn , (14.1)

(ii) Dilation: Da : s(x)  an/2 s(a1 x), a > 0, (14.2)
1
(iii) Rotation: UR : s(x)  s(R x), R SO(n), (14.3)

where b Rn is the displacement parameter, a > 0 is the dilation parameter and R

is an n n rotation (orthogonal) matrix. In particular:
For n = 2, R R( ), a 22 rotation matrix, parametrized by an angle [0, 2 );
For n = 3, R R( , , ), a 3 3 rotation matrix, parametrized for instance by
the three Euler angles , , .
These three operations generate the n-dimensional Euclidean group with dila-
tions [476,Mur90], also known as the similitude group of Rn , SIM(n) = Rn (R+
SO(n)). This group SIM(n) has the following natural action on a signal s, in position
and Fourier space, respectively:

sb,a,R (x) = [U(b, a, R)s ](x) = an/2 s(a1 R1 (x b)) (14.4)

 s ](k) = an/2 s(a R1 k)eikb .
b,a,R (k) = [U(b, a, R)
s (14.5)

We recall that our Fourier transform in Rn is defined as [see (12.8)]


f(k) = (2 )n/2 eikx f (x) dx, (14.6)
Rn

for k, x Rn and k x = k1 x1 + k2 x2 + + kn xn .
14.2 Mathematical Analysis 413

Thus U(b, a, R) = Tb Da UR , where the order of the variables (b, a, R) is imposed

by the structure of the similitude group SIM(n) as a semidirect product: dilations
and rotations (a, R) act on translations (b). We note in passing that one could replace
the group SO(n) by O(n) in (14.3), that is, include also mirror symmetries (or more
general discrete operations), but this has never been considered.
The basic structure of the n-dimensional WT is the same as in the 1-D case, as
results from the following theorem.
Theorem 14.2.1. The operator family U() defined in (14.4) is a unitary irreducible
representation of SIM(n) in L2 (Rn , dx), and it is unique up to unitary equivalence.
This representation is square integrable: a vector L2 (Rn , dx) is admissible iff it
satisfies the condition

dk
c = (2 ) An1 n
|
 (k)|2 < , (14.7)
Rn |k|n

n1
2 k/2
where |k|2 = k k, and An1 = is the volume of SO(n 1).
k=2 (k/2)

Proof. The proof is elementary. The unitarity of U is obvious. Irreducibility follows

from the fact that, for every nonzero s L2 (Rn , dx), the linear span of the set

D = {sb,a,R = U(b, a, R)s, (b, a, R) SIM(n)} (14.8)

is dense in L2 (Rn , dx). Indeed, let f L2 (Rn , dx) be orthogonal to D, that is

sb,a,R | f  = 0, (b, a, R) SIM(n). This means

sb,a,R | f = an/2
 eikb s(a R1 k) f(k) dk = 0, (b, a, R) SIM(n),
Rn

which implies that s(a R1 k) f(k) = 0 a.e., for all (a, R) R+ SO(n). Since the

action of the latter on R is transitive, this implies f (k) = 0 a.e., that is, f = 0.
n

Uniqueness follows from the fact that the representation U is obtained by

Mackeys standard method of induction, since there is only one nontrivial orbit in
k-space [476, Bar77, Mur90] (see Sect. 9.2.4).
Finally, the square integrability follows again from the explicit calculation of the
L2 norm of the matrix element of U(b, a, R) with respect to the left Haar measure
dg = a(n+1) db da dR, where dR is the Haar measure on SO(n):

da
|U(b, a, R) | |2 db dR = c  2 , (14.9)
SIM(n) an+1

where c is the constant defined in (14.7). 


Actually the two results stated in this theorem, namely, uniqueness and square
integrability of the representation U, already follow from the analysis of semidirect
414 14 Multidimensional Wavelets and Generalizations

products made in Sect. 8.3. Nevertheless, we found it instructive to give the present
direct proof. It is interesting to note that the presence of dilations is crucial both for
the uniqueness and for the square integrability of U. Now the restriction of U to the
rotation group SO(n) is the quasiregular representation, which decomposes into the
direct sum of all irreducible vector representations. In other words, the inclusion of
dilations forbids the appearance of spinorial representations of SO(n).
A nonzero signal that satisfies the admissibility condition (14.7) is called an
n-dimensional wavelet. Again, if is regular enough ( L1 (Rn , dx) L2 (Rn , dx)
suffices), the admissibility condition (14.7) simply means that the wavelet has zero
mean:


 (0) = 0 (x) dx = 0, (14.10)
Rn

In addition, both and  are supposed to be well-localized. Clearly b,a,R satisfies

all these conditions whenever does.
Given a finite energy signal s L2 (Rn , dx), its continuous wavelet transform
(with respect to the fixed wavelet ) S is given by:

S (b, a, R) = b,a,R |s (14.11)


= an/2 (a1 R1 (x b))s(x) dx (14.12)
Rn

= an/2 eibk
 (aR1 k) s(k) dk. (14.13)
Rn

The most interesting case arises when the wavelet is axially symmetric, i.e.,
SO(n 1)-invariant. Then one can replace everywhere SO(n) by SO(n)/SO(n 1)
Sn1 , the unit sphere in Rn . The rotation R becomes R R( ), Sn1 , and the
CWT is S S (a, , b) L2 (X, d ), where the parameter space is

X = SIM(n)/SO(n 1) = Rn  (R+
SO(n)/SO(n 1))

= Rn  (R+
S
n1
) Rn Rn , (14.14)

with SIM(n)-invariant measure

da
d (x) = db d ( ), (14.15)
an+1

where d ( ) is the surface measure on Sn1 . Furthermore, while b Rn

corresponds to the position variables, the pair (a1 , ) may be interpreted as a
spatial frequency in spherical polar coordinates. To be sure, writing = a1 , the
volume element on X becomes

d (x) = n1 d d ( ) db dk db, (14.16)

14.2 Mathematical Analysis 415

the (Lebesgue) volume element of R2n . As for n = 1 (see Sect. 12.4.1), one may
easily compute the coadjoint orbits of SIM(n). It turns out that there are only two:
a trivial one, Oo = {0} Rn , associated to the identity representation, and one
of dimension 2n, OU = Rn Rn , with canonical variables k Rn , x Rn , which
corresponds to the representation U. Thus OU (X, dk db), so that here too, the
CWT is a phase space representation [272].
From now on, we will restrict ourselves to the case of an axially symmetric
wavelet . Of course, this covers also the fully isotropic case (SO(n)-invariant
wavelet), where S does not depend on at all.
The main properties of the wavelet transform may be summarized as follows
(compare with Theorem 12.2.2):
1/2
Theorem 14.2.2. Let the map W : s  c S be defined by

1/2
(W s)(b, a, ) = c b,a, |s, s L2 (Rn , dx). (14.17)

Then:
(1) W conserves the norm of the signal:
 
da
|S (b, a, )|2 db d ( ) = c |s(x)|2 dx, (14.18)
X an+1 Rn

i.e., it is an isometry from the space of signals into the space of transforms,
which is a closed subspace H of L2 (X, d ). Equivalently, the family of
wavelets {b,a, }, with a > 0, Sn1 , b Rn , generates a resolution of the
identity:

da
c1
|b,a,  b,a, | db d ( ) = I. (14.19)
X an+1

(2) The adjoint of W , restricted to H , yields the reconstruction formula:


da
s(x) = c1
b,a, (x) S (b, a, ) db d ( ) . (14.20)
X an+1

(3) The projection from L2 (X, d ) onto H is an integral operator, with

(reproducing) kernel

K(b , a ,  ; b, a, ) = c1
b ,a ,  | b,a, , (14.21)

that is, the autocorrelation function of .

(4) The CWT is covariant with respect to the group SIM(n):

s(x bo )  S (b bo , a, ), bo Rn ,
416 14 Multidimensional Wavelets and Generalizations

s(R1 1 1
o x)  S (Ro b, a, Ro ), Ro SO(n), (14.22)
n/2
ao s(a1 1 1
o x)  S (ao b, ao a, ), ao > 0.

No proof is needed, since this theorem is the exact adaptation to the case at hand of
Theorem 12.2.2.
As we can see, all the formulas are the exact analogues of those given in
Sect. 12.2 for n = 1. It follows that the interpretation of the n-dimensional CWT
is entirely similar to that given previously. The only new element is the presence of
the rotation degree of freedom, indexed by the variable R SO(n), or Sn1 in
the axisymmetric case. Here also, the covariance is crucial for applications and it is
lost in the common version of the 2-D discrete WT (see Sect. 14.4). In particular,
the presence of the rotation parameter makes the multidimensional CWT sensitive
to directions, provided one uses an oriented wavelet. As we will see in Sect. 14.3.3,
many applications are based on this property.
This remark brings us precisely to the question of the choice of the analyzing
wavelet . There are two possibilities, depending on the problem at hand.
1. If one wants to perform a pointwise analysis, that is, when no oriented features
are present or relevant in the signal, one may choose an analyzing wavelet
which is invariant under rotation (full SO(n) invariance).
2. When the aim is to detect directional features in a signal, for instance to perform
directional filtering, one has to use a wavelet which is not rotation invariant. The
best angular selectivity will be obtained if is directional, which means that
its (essential, i.e., numerical) support in spatial frequency space is contained in
a convex cone with apex at the origin. Since it may sound counterintuitive, this
definition requires a word of justification. According to (14.13), the wavelet acts
as a filter in k-space (multiplication by the function  ). Take, for instance, n = 2
and suppose the signal s(x) is strongly oriented, for instance a long segment
along the x-axis. Then its Fourier transform s(k) is essentially supported along
the ky -axis. In order to detect such a signal, with a good directional selectivity,
one needs a wavelet supported in a narrow cone in k-space. Then the WT
is negligible unless  (k) is essentially concentrated on the support of s(k):
directional selectivity demands to restrict the support of  , not .
Let us examine in more detail some examples of n-dimensional wavelets of each
kind. In most cases, they are the obvious generalizations of those given in Chap. 12
for n = 1.
1. Isotropic wavelets
The Mexican hat or Marr wavelet
This is a real, rotation invariant wavelet, given by the Laplacian of a Gaussian:
1
H (x) = exp( |x|2 ), = x21 + x22 + . . . + x2n
2
1
= (n |x|2 ) exp( |x|2 ). (14.23)
2
14.2 Mathematical Analysis 417

One uses also higher order Laplacians of the Gaussian,

(m)
H (x) = ( )m exp( 12 |x|2 ). (14.24)

For increasing m, these wavelets have more and more vanishing moments, and
are thus sensitive to increasingly sharper details. Note that m could be extended
to real numbers, giving numerically more freedom, but then ( )m becomes a
pseudodifferential operator.
An interesting technique [95, 96] is to analyze the same signal with several
(m)
wavelets H , for different m. The features common to all the transforms almost
surely belong to the signal, they are not artifacts of the analysis.
Difference wavelets
An interesting class consists of wavelets obtained as the difference between a
function h and a contracted version of the latter. If h is a smooth non-negative
function, integrable and square integrable, with all moments of order 1 vanishing,
then the function ( ) given by the relation:

( ) (x) = h(x) 2 h( 1 x) ( > 1) (14.25)

is easily seen to be a wavelet satisfying the admissibility condition (14.10). Such

difference wavelets have the additional advantage that they lead to interesting and
fast algorithms [268]. We will come back to this point.
A typical example is the Difference-of-Gaussians or DOG wavelet, obtained
by taking for h a Gaussian. The DOG filter is a good substitute for the Mexican hat
(for = 1.6, their shapes are extremely similar), frequently used in psychophysics
works [DeV88]. We will encounter a similar wavelet in the CWT on the 2-sphere,
in Sect. 15.1.1.
2. Directional wavelets
The Morlet wavelet
This one is the prototype of an oriented wavelet:

1
M (x) = exp(iko x) exp( |Ax|2 ) + corr. term. (14.26)
2
1/2 1/2
The parameter ko Rn is the wave vector, and A = diag[1 , 2 , . . .], j 1, is
an n n anisotropy matrix. The correction term enforces the admissibility condition

M (0) = 0, but it is numerically negligible for |ko | 5.6 and will usually be
dropped. The most useful choice is the axially symmetric Morlet wavelet, with wave
vector ko pointing in the xn direction and rotation invariant around that axis. For
n = 3, this yields the following wavelet:
418 14 Multidimensional Wavelets and Generalizations

20 20

15 15

10 10

5 5

0 0

5 5

10 10

15 15

20 20
20 15 10 5 0 5 10 15 20 20 15 10 5 0 5 10 15 20

Fig. 14.1 Two directional 2-D wavelets, in k space, seen in level curves: (left) The Morlet
(C10 )
wavelet M (ko = (0, 6), = 5); (right) The Cauchy wavelet
44 with supporting cone C10 =

C (10 , 10 ), rotated by 90 for the sake of comparison

1
ax (x) = exp(iko z) exp [ 1 (x2 + y2 ) + z2 ]. (14.27)
2
The 2-D Morlet wavelet is shown, in k-space, in Fig. 14.1 (left).
The conical wavelets
In order to achieve a genuinely oriented wavelet, it suffices to consider a smooth
function  (C ) (k) with support in a strictly convex cone C in spatial frequency
space and behaving inside C as P(k1 , . . . , kn )e k , with C and P(.) denotes
a polynomial in n variables. Alternatively one may replace the exponential by a
Gaussian, which gives a better localization in spatial frequency. Thus one obtains
the class of conical wavelets. Let us study in some detail the former case, called the
Cauchy wavelets.
Take first n = 2. Let C C ( , ) = {k R2 : arg k } be the convex
cone given by the directions and , with corresponding unit vectors e , e . The
dual cone, also convex, is

 , ) = {k R2 : k k > 0, k C ( , )},
C C ( (14.28)

where  = + /2,  = /2, and therefore e e = e e = 0, whereas

 , ), we define the 2-D
e  e = e  e = sin( ). Given a fixed vector C(
Cauchy wavelet in spatial frequency variables [58, 68, 72]:
@
(C ) (k e )l (k e )m ek , k C ( , );

lm (k) = (14.29)
0, otherwise.
14.2 Mathematical Analysis 419

(C )
The Cauchy wavelet lm (k) is strictly supported in the cone C ( , ) and the
parameters l, m N give the number of vanishing moments on the edges of the
cone. The crucial parameter is the opening angle of the cone = , since it
characterizes the angular resolving power (ARP) of the wavelet [58].
An explicit calculation yields the following result:

(C )
lm (x) = const. (z e )l1 (z e )m1 , (14.30)

where we have introduced the complex variable z = x + i R2 + i C (a tubular

(C )
domain in C2 ). We show in Fig. 14.1 (right) the wavelet 44 (k) for C10 =

C (10 , 10 ), with an ARP = 20 . This is manifestly a highly directional filter.
Actually the origin of the name Cauchy is the following example. For =
0, = /2, = e /4 and l = m = 1, one gets:

(C ) 1
11 (x) = (1 ix)2 (1 iy)2 , (14.31)
2
i.e., the product of two 1-D CauchyPaul wavelets [500, Hol95, Pau85], that is,
derivatives of the Cauchy kernel (z t)1 . Of course, this example is of little use
in practice. The main interest of Cauchy wavelets is their good angular selectivity,
(C )
which requires a narrow cone. For applications, it turns out that the wavelet 44 10 ,

with support in the cone C10 = C (10 , 10 ) has properties very similar to those
of the Morlet wavelet (14.26) with |ko | = 5.6, except that here the opening angle
of the cone is totally controllable (see Sect. 14.3.2 for the calibration problem).
For a Morlet wavelet, on the contrary, the cone gets narrower for increasing |ko |,
but then the amplitude decreases as exp(|ko |2 ). In that sense, Cauchy wavelets are
better adapted. This may be related to the fact that they have minimal uncertainty,
as discussed in the next section.
An alternative possibility is to replace in (14.30) the exponential by a Gaussian
centered on the axis = e 1 ( + ) of the cone, exp(|k ao |2 ) (ao > 0).
2
The resulting conical wavelet is very similar to the previous one, except that it is
more concentrated in spatial frequency space, since it is also localized in scale,
around the central scale ao . However, although the pure Gaussian is well peaked,
the addition of a large number of vanishing moments tends to spread it. Thus one
can achieve an even better scale localization by using an appropriate width for the
Gaussian. It was found empirically [Jac04] that the following function, strongly
peaked along the x-axis, has an almost ideal behavior:
@
(k e )l (k e )m e 2 (kx ko ) ,
1 2
(C ) k C ( , ),

lm (k) = (14.32)
0, otherwise.

with ko = (1) (l + m)1/2 . This shows how wavelets may be tailor-made for
specific applications. For instance, the conical wavelet (14.32), with supporting cone
420 14 Multidimensional Wavelets and Generalizations

C10 = C (10 , 10 ), has been used in the analysis of the Penrose tiling depicted
in Fig. 14.5. It has also proved to be very efficient for motion analysis, where it
replaces the Morlet wavelet for the spatial component of a separable space-time
wavelet [166]. See also Sect. 16.2.1.
Let us take now n = 3 and consider the convex simplicial (or pyramidal) cone
C ( , , ) defined by the three unit vectors e , e , e , the angle between any two of
them being smaller than . The dual cone is also simplicial, namely C = C (  , , ),
where e = e e is orthogonal to the - face, etc. With these notations, given a
vector C and l, m, n N , we define a 3-D Cauchy wavelet in spatial frequency
space as:
@
(C , ) (k e )l (k e )m (k e)n ek , k C ( , , ),

lmn (k) = (14.33)
0, otherwise.

Here too the expression for the 3-D wavelet in position space may be obtained
explicitly:

(C , ) il+m+n+3 (e e )l (e e )m (e e )n

lmn (x) = l! m! n! det A , (14.34)
2 (z e )l+1 (z e )m+1 (z e )n+1

where A is the matrix that transforms the unit vectors e1 , e2 , e3 into the triple
e , e , e and we have written z = x + i . From the expressions (14.33) and (14.34),
one may then obtain other 3-D Cauchy wavelets, for instance one supported in a
circular cone. Take a circular convex cone, aligned on the positive kz -axis, with total
opening angle 2o (0 < o < /2). In spherical polar coordinates k = (|k|, , ), the
interior of the cone is simply C (o ) = {k R3 : o }. Then an axisymmetric
3-D Cauchy wavelet supported in this cone may be defined, for instance, by

|k|l (tan2 o tan2 )m e|k| cos , 0 o ;
m(o ) (k) =
(14.35)
0, otherwise.

Again m N defines the number of vanishing moments on the surface of the cone,
that is, the regularity of the wavelet. For o very small, this wavelet lives inside a
narrow pencil: it clearly evokes the beam of a searchlighta vivid illustration of the
wavelet as a directional probe!
If we note that the expression on the r.h.s. of (14.35) may be written as

|k|l (tan2 o kz2 kx2 + ky2 )m ekz ,

we see that all these wavelets are built on the same model, namely F(k)m ekz , where
F(k) = 0 is the equation of the cone. Clearly the whole construction extends to any
number of dimensions n 2, and the last remark gives a hint on how to design
Cauchy or conical wavelets adapted to a general cone.
14.2 Mathematical Analysis 421

An interesting property of these conical wavelets, in any dimension, is their

(C )
analyticity. Indeed, since the function lm (k) has support in the convex cone C
and is of fast decrease at infinity, it follows from general theorems [Str64] that its
(C )
Fourier transform (C ) (x) is the boundary value of a function lm (z), holomorphic
in the tubular domain Rn + iC. This is seen on the expressions (14.29) and (14.34)
of the Cauchy wavelets, for which the inverse Fourier transform can be computed
explicitly.
Another interpretation is that we have here in fact a construction of an n-D
progressive wavelet. Indeed, the definitions (14.30) and (14.33), shows that the
Cauchy wavelet is obtained by taking the directional derivative of the exponential
exp(k ), and then taking an analytic signal associated to it, by putting to zero
the part that lives outside of the convex cone C . The same is true for more general
conical wavelets of the type (14.32). In that sense, the conical wavelets yield a
genuine multidimensional generalization of the 1-D Hardy functions obtained as
Fourier transforms of progressive wavelets [349], much more so than the so-called
2-D Hardy functions defined in [225].

14.2.2 The CWT of Radial Functions

A radial setting arises naturally when separating variables in polar coordinates, since
Rn R+ Sn1 , which leads to L2 (Rn ) L2 (R+ ) L2 (Sn1 ). For wavelets on the
sphere Sn1 , there exists a choice of methods that we shall describe in Sect. 15.1 (at
least for n = 3, the general case being treated in [62]). Thus it remains to see how to
design wavelets on the half-line R+ .
If the wavelet is isotropic, that is, invariant under SO(n), then the CWT (14.11)
of a signal s is a function S (b, a) that does not depend on R SO(n) any more. Let
now the signal s itself be isotropic or radial, i.e., there is a function s defined on R+
such that s(x) = s(|x|). Then its CWT is also radial. Indeed, for every R SO(n),
we have

S (R1 b, a) = TR1 b Da | s = UR TbUR1 Da | s = Tb DaUR1 |UR1 s

= Tb Da | s = S (b, a).

However, although the CWT preserves the essential symmetry property, the pres-
ence of the latter does not reduce the cost of computing the CWT if one uses (14.12)
or (14.13). Clearly, the culprit is the translation operator Tb , which does not respect
the symmetry [the same is true if one replaces the full rotation group by any closed
subgroup of SO(n)]. In order to alleviate this defect, one may introduce generalized
translation operators, following [517, Rau04].
Denote by (Lloc 1 ) (Rn ) the closed subspace of radial, locally integrable,
rad
functions in Lloc (Rn ) and by Prad the corresponding projection operator. Then
1

one shows that

422 14 Multidimensional Wavelets and Generalizations


1
Prad f (x) = f (|x| ) d ( ),
|Sn1 | Sn1

where |Sn1 | = 2 n/2 (n/2)1 is the area of Sn1 . Let g (Lloc1 ) (Rn ) be
rad
+
radial, with corresponding function g Lloc (R ). Then Ty g(x) = g(x y) is no
1

longer radial in x and y. Thus, we define a generalized translation operator t :

1 )(R+ ) (L1 )(R+ ),
(Lloc loc

t g(r) = Prad Ty g(x), for x, y Rn , with r = |x|, t = |y|.

Hence we obtain

1
t g(r) = g(|t r |) d ( ), r, t R+ , Sn1 , (14.36)
|Sn1 | Sn1

where the expression on the r.h.s. does not depend on the choice of Sn1 .
Now Lrad2 (Rn ) := P L2 (Rn ) is unitary equivalent to L2 (R+ , d ), where the
rad n
measure n on R+ is given by dn (r) := |Sn1 | rn1 dr. Therefore, the CWT of
the radial function s becomes

S (b, a) = b Da | s = b Da (r) s(r) dn (r), (14.37)
0

where is the function on R+ corresponding to . In conclusion, the presence of

the radial symmetry reduces the n-dimensional CWT to a single integral over R+ .
The relation (14.37) solves the problem of the CWT of a radial function.
However, this approach needs to be further refined for designing the corresponding
discrete WT, via a multiresolution. We shall discuss this problem in Sect. 14.4.2.
We note that b c = b+c , b, c R+ \ {0}, which indicates that we are facing
a different mathematical structure. First there is a good notion of convolution
on R+ . Indeed, for any two functions f, g on R+ , one defines (whenever the integral
converges)

f  g(t) := f(r)t g(r) dn (r). (14.38)
0

One sees easily that, for any two radial functions f , g on Rn , with corresponding
functions f, g on R+ , one has f g(x) := f  g(|x|), where denotes the ordinary
convolution on Rn . This implies that L1 (R+ , dn ) is a commutative Banach algebra
for the convolution (14.38).
Next one introduces a radial -function:

r ( f ) := Prad f (x), r = |x|.

14.3 The Two-Dimensional CWT 423

This object satisfies the following relations:

t g(r) = t r (g) = t g(x) ,

t g(r) = r g(t), t 1 = 1 and 0 g(r) = g(r).

Then the convolution (14.38) can be extended to the space Mb (R+ ) of regular
bounded Borel measures on R+ , so that the latter becomes a commutative Banach
*-algebra, with neutral element 0 and  as involution. This object is called
a hypergroup. More precisely, the pair (R+ , ) is called the BesselKingman
hypergroup of order n/2 1, denoted Hn/21 . This is the structure that will allow,
in Sect. 14.4.2, the transition from the CWT to the DWT on R+ . We refer to
[517, Rau04] for a short introduction to this topic.

14.3 The Two-Dimensional CWT

In this section, we examine some additional aspects of two-dimensional wavelets,

since this is the most important case for applications (image processing mainly, but
not only). One specificity of the two-dimensional analysis is the importance of the
geometry of the object. This is manifest already in the choice of the 2-D wavelet for
image analysis with the CWT, as discussed in the previous section, but even more in
the case of the DWT, as we will see in Sect. 14.4. Indeed, the geometrical inadequacy
of the 2D DWT is the main motivation for a number of generalizations that we will
describe in Sect. 14.5. This point of view has been emphasized in the recent review
paper [395], which also contains a large number of references to image processing.

14.3.1 Minimality Properties

As we have seen in Chap. 2, the canonical coherent states have the characteristic
property of minimal uncertainty, which means that they saturate the inequality in
the Heisenberg uncertainty relations (2.2), and this is interpreted by saying that
canonical CS are quantum states whose behavior is closest to classical. What about
wavelets, which are the coherent states associated to the similitude groups?
According to the standard discussion in quantum mechanics textbooks [Coh77,
Got66], two observables of a quantum system, represented by self-adjoint operators
A and B, obey the uncertainty relation
1
A. B |[A, B]|, (14.39)
2
where
1
A A = A2  A2 (14.40)
424 14 Multidimensional Wavelets and Generalizations

denotes the variance of A in the state and C =  |C  is the average of the
operator C in the state . The state is said to have minimal uncertainty if equality
holds in (14.39), which happens iff

(A A) = io (B B) , (14.41)

for some o > 0. (Note that the notion of minimal uncertainty has also been
emphasized, and the result (14.41) proven, by Gabor in his pioneering paper on
communication [294].)
In the case of canonical CS, the noncommuting operators are Q and P, that is,
the infinitesimal generators of translations in phase space. In order to apply this
concept to 2-D wavelets, we return to the discussion following Theorem 14.2.1,
which showed that the 2-D CWT is also a phase space representation. Accordingly,
we consider the infinitesimal generators of the transformation (14.4) or its equiva-
lent (14.5) in k-space, and denote them by P1 and P2 for translations, D for dilations
and J for rotations. Among these, there are four nonzero commutators, namely

[D, P1 ] = iP1 , [J, P2 ] = iP1 , [D, P2 ] = iP2 , [J, P1 ] = iP2 , (14.42)

but the first two transform into the last two under a rotation by /2. More generally,
defining P = P1 cos + P2 sin , we replace both pairs of commutators by the
relations:

[D, P ] = iP ; [J, P + /2 ] = iP , (14.43)

and we look for wavelets which are minimal with respect to this pair. Thus,
minimality has to be defined with respect to a fixed direction e , and it is impossible
to do it for two directions at the same time, for instance, for all four relations (14.42)
simultaneously [218].
For fixing ideas, we consider the uncertainty relations for the first pair in (14.42),
corresponding to = 0:

1 1
D. P1 |P1 |, J. P2 |P1 |. (14.44)
2 2

Then, according to (14.41), a vector

 saturates these inequalities iff it satisfies the
following system of equations

(D + i1 P1 )
 (k) = (D + i1 P1 )
 (k)
(1 , 2 > 0). (14.45)
(J + i2 P2 )
 (k) = (J + i2 P1 )
 (k)

This system of partial differential equations may be solved in polar coordinates,

k = (|k|, ), imposing successively five conditions: (i) Integrability of the system
requires 1 = 2 = > 0; (ii) 2 -periodicity in implies that P2  = 0 and J =
m Z; (iii) Square integrability,
 L2 , implies that the support of
 is restricted
14.3 The Two-Dimensional CWT 425

to a convex cone in the right half-plane; (iv) Admissibility of  implies P1  > 1;
(v) Finally, imposing the condition that
 (k) be real implies that J = D = 0.
The result is that a real wavelet  is minimal with respect to the first pair of
commutation relations (14.42) iff it vanishes outside some convex cone C in the
half-plane kx > 0 and is exponentially decreasing inside:

c |k| e ke1 ( > 0, > 0), k C,

 (k) = (14.46)
0, otherwise.

More generally, if one chooses the commutation relations in (14.43), one obtains a
similar result, rotated by , that is, a wavelet supported in a convex cone C with
axis in the direction e , and exponentially decreasing inside.
Proposition 14.3.1. A real wavelet  has minimal uncertainty iff it vanishes
outside some convex cone C with apex at the origin and opening angle < ,
and is exponentially decreasing inside:

c |k| e k ( > 0, > 0, C), kC

 (k) = (14.47)
0, otherwise.

In other words,
 must be of the form

 (k) = c C (k) |k| e k

( > 0, > 0), (14.48)

where C is the characteristic function of C , or a smoothened version thereof.

We may now impose some degree of regularity (vanishing moments) at the
boundary of the cone, by taking an appropriate linear superposition of such minimal
wavelets
 . Thus we obtain finally:

 C (k) = c C (k) F(k) e e1 k ,

( > 0) (14.49)

where F(k) is a polynomial in kx , ky , vanishing at the boundaries of the cone C ,

including the origin. Clearly a Cauchy wavelet with = e1 is of this type.
Other minimal wavelets may be obtained if one includes commutators with ele-
ments of the enveloping algebra, i.e., polynomials in the generators. For instance, if
one requires the wavelet to be rotation invariant, one may start from the commutator
between D and the Laplacian = P12 + P22 . Then one finds a whole family of
minimal isotropic wavelets, among them all powers of the Laplacian, n , acting
on a Gaussian [59]. For n = 2, this gives the 2-D isotropic Mexican hat [218]. By
comparison, the same problem in 1-D was posed and solved by Klauder [411], with
the result that the minimal wavelets are the equivalent of the CauchyPaul wavelets
described in (12.20), namely m ( ) = m e , for 0 and 0 for < 0. Indeed,
the positive half-line is a convex cone in 1-D!
426 14 Multidimensional Wavelets and Generalizations

We ought to emphasize at this point that the property of minimality for wavelets
is a mathematical one, and it is not clear whether it implies an operational meaning
in the same way as was discussed for gaborettes. Cauchy wavelets are linear
combinations of minimal wavelets, but they are not the most efficient conical
wavelets for directional analysis. This is not new: in 1-D too, the CauchyPaul
wavelet (12.20) is minimal, but many others are as least as useful in practice, for
instance the derivatives of the Gaussian or the Morlet wavelet.

14.3.2 Interpretation, Visualization Problems, Calibration

As we have said above, all the formulas for the 2-D continuous wavelet transform
are the exact analogues of the corresponding ones in 1-D, that we discussed in
Chap. 12. Hence the interpretation of the CWT as a singularity scanner is the same
also. In practice, indeed, one assumes the wavelet to be well localized both in
position space (x) and in spatial frequency space (k). From this one deduces again
that wavelet analysis operates at constant relative bandwidth, |k|/|k| = const.
Therefore, the analysis is most efficient at high frequencies or small scales, and so
it is particularly apt at detecting discontinuities in images, either point singularities
(contours, corners) or directional features (edges, segments).
In addition, the wavelet is often required to have a number of vanishing
moments (as for instance the mth order Mexican hat or the Cauchy wavelets), which
increases its capacity at detecting the singular part of images, in particular the
discontinuities (this is where the most significant information lies).
In conclusion, as in the 1-D case, the 2-D wavelet transform may be interpreted as
a mathematical, direction selective, microscope, with optics , magnification 1/a
and orientation tuning parameter [92]. Notice that the magnification is global,
independently of the direction, and there is the additional property of directivity,
given by the rotation angle .
Now comes a striking difference between 1-D and 2-D analyses. In 1-D, the
wavelet parameters (a, b) span a half-plane, so that a lot of information may
be obtained visually (one may say that the wavelet transform unfolds the signal
from 1-D to 2-D, thus decoupling parts that live at different scales). And more
sophisticated methods of feature detection exploit this, for instance the WTMM
method (wavelet transform modulus maxima) [92, 248], which is based on the
ridges of the transform. But in 2-D, one faces a problem of visualization, since the
transform S(b, a, ) is a function of four variables: two position variables b R2 ,
and the pair (a, ) R+ [0, 2 ) R .
2

Since it is impossible to compute and visualize the full transform, it is necessary

to fix certain of the variables a, , bx , by . There are six possible choices of two-
dimensional sections, but the phase space interpretation discussed above indicates
that two of them are more natural: either (a, ) or (bx , by ) are fixed, and the WT is
treated as a function of the two remaining variables. Thus we obtain the two standard
representations [58, 66, 68]:
14.3 The Two-Dimensional CWT 427

1. The position or aspect-angle representation: a and are fixed and the CWT
is considered as a function of position b alone. Alternatively, one may use
polar coordinates, in which case the variables are interpreted as range |b| and
perception angle , a familiar representation of images.
2. The scale-angle representation: for fixed b, the CWT is considered as a function
of scale a and anisotropy angle , i.e., of spatial frequency. In other words, one
looks at the full transform as through a keyhole located at b, and observes all
scales and all directions at once.
The position representation is the most familiar one, and it is useful for the
general purposes of image processing: detection of position, shape and contours
of objects; pattern recognition; image filtering by resynthesis after elimination
of unwanted features (for instance, noise). The scale-angle representation will
be particularly interesting whenever scaling behavior (as in fractals) or angular
selection is important, in particular when directional wavelets are used. In fact,
both representations are needed for a full understanding of the properties of the
continuous wavelet transform in all four variables.
In addition to these two familiar representations, there are four other
two-dimensional sections, obtained by fixing two of the four variables (a, , |b|, ),
and analyzing the CWT as a function of the remaining two. Among these, the
angleangle representation might be useful for applications [67]. Here one fixes
the range |b| and the scale a and considers the CWT at all perception angles and
all anisotropy angles . This case is particularly interesting, because the parameter
space is now compact (it is a torus).
The last point we want to address in this more practical section is that of cali-
bration. We have compared the continuous wavelet transform to a (mathematical)
microscope, so one should determine its resolving power. This is particularly true
in 2-D: how well does the wavelet tool separate scales or directions? Interestingly
enough (from our point of view), it turns out that a large part of the answer
resides in the reproducing kernel K. This was to be expected, since K is in fact
the autocorrelation function of the wavelet. Of course, K has to be studied in
all four variables (a, , bx , by ), that is, both in the position and in the scale-angle
representations. The latter, in particular, yields the angular informationand that is
where directional wavelets in fact originate from.
In addition, one also uses benchmark signals for testing definite properties.
In order to give the flavor of the technique evoked here, let us give an example
(C )
of wavelet calibration. For testing the angular selectivity of the wavelet 44 10 , one
uses as benchmark signal a semi-infinite rod, sitting along the positive x-axis, and
modeled with a delta function:

s(x) = (x) (y), (14.50)

where (x) is the step function. Then one computes the CWT of s as a function of
the position x along the rod, for the fixed scale ao = 1. Thus is the misorientation
of the wavelet with respect to the signal (the rod). The result, shown in Fig. 14.2,
428 14 Multidimensional Wavelets and Generalizations

| S(x,a0,) |
= 20
1

= 10
0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1 = 20
= 30
0
5 0 5 10
x
(C )
Fig. 14.2 Testing the angular selectivity of the Cauchy wavelet 44 10 with the semi-infinite rod
signal. The figure shows the modulus of the CWT as a function of the position x along the rod, for
various values of the misorientation angle and fixed scale ao = 1

(C )
is that the wavelet 44 10 , with an opening angle (ARP) = 20 , detects the
orientation of the rod with a precision of the order of 10 . Indeed, for = 0,
the WT is a wall, increasing smoothly from 0, for x 5, to its asymptotic value
(normalized to 1) for x 5. Then, for increasing misorientation , the wall gradually
collapses, and essentially disappears for > 10 . Only the tip of the rod remains
visible, and for large ( > 45 ), it gives a sharp peak. In the same way, one can
also test the angular selectivity as a function of the opening angle of the support
cone. We refer the reader to the original articles for more details [58, 66, 72].
The outcome of the calibration process is that a wavelet  (k) with support in
a narrow cone in spatial frequency space is highly selective in direction. In that
case, the CWT offers the possibility of directional filtering, and this is crucial in
many applications. In order to see what this means and to show the efficiency of
a directional wavelet for that purpose, we consider in Fig. 14.3(a) a pattern made
of rods in many different directions. Applying the CWT with a Cauchy wavelet,
oriented horizontally (that is,  (k) is concentrated along the ky -axis), selects all
the rods which have roughly the same direction (b), whereas the other ones, which
are misaligned, yield only a faint signal corresponding to their tips. Since this is in
fact noise, one performs a thresholding for removing it, thus getting an improved
14.3 The Two-Dimensional CWT 429

Fig. 14.3 Directional filtering with the CWT: (a) The pattern; (b) CWT with a Cauchy wavelet
supported in C (10 , 10 ) (c) The same after thresholding

picture (c). The same two operations are then repeated with various successive
orientations of the wavelet. In this way, one can count the number of objects
that lie in any particular direction (similar results may be obtained with a Morlet
wavelet [58]).

14.3.3 Practical Applications of the CWT in Two Dimensions

Among the multidimensional wavelets, the case n = 2 is certainly the most

interesting for applications, in particular in image processing. The 2-D continuous
wavelet transform has been used by a number of authors, in a wide variety of
physical or engineering problems [Com90,Mey91,Mey93]. In all cases, its main use
is for the analysis of images, that is, the detection of specific features, for instance
a hierarchical structure or particular discontinuities, such as edges, filaments,
contours, boundaries between areas of different luminosity, etc. Of course, the type
of wavelet chosen, oriented or not, depends on the precise aim.
A large number of applications are discussed in our textbook [Ant04]. Among
the most successful ones, we may quote:
Contour detection, character recognition:
The contour or the edges of an object are discontinuities in luminosity. Hence
the CWT will detect them with efficiency [66, Mur90]. Possible applications are
segmentation of images (e.g. medical images) or character recognition [67].
Symmetry detection:
The 2-D CWT with a highly directional wavelet (e.g. Cauchy) allows to
detect discrete rotational or combined dilation-rotation symmetries, even locally,
in objects like geometrical figures, Penrose tilings or diffraction patterns of
quasicrystals [72]. More details are given below.
430 14 Multidimensional Wavelets and Generalizations

Analysis of astronomical and astrophysical images:

The CWT has been used for several purposes in astronomy: noise filtering
(background sky), with a technique known as unsharp masking, unraveling
of the hierarchical structure of a galactic nebula, or that of the universe itself
(galaxy counts, detection of galaxy clusters or voids) [559]. Here one usually
couples an isotropic CWT analysis with statistical methods. It has been used also
in solar physics for the automatic detection and analysis of special objects from
the Solar corona: (i) suppressing cosmic ray hits on the CCD camera; [Jac04]
(ii) detecting and classifying bright points, [Jac04]; (iii) then for detecting and
analyzing Solar coronal magnetic loops [152, Bis10]. In the same way, the CWT
has been applied to a number of topics in astrophysics, for instance, the detection
of gamma sources in the Universe. In both cases, the CWT is applied to data
collected by satellites (SoHO, EGRET, GLAST) [79, 247, 578]. In addition, the
detection should be performed in an automatic way, on board of the satellite. This
justifies the use of wavelets, because of their simplicity.
Analysis of 2-D fractals:
The CWT is an ideal tool for analyzing 2-D (multi)fractals, either artificial
(numerical snowflakes, diffusion limited aggregates) or natural (electrodepo-
sition clusters, various arborescent phenomena), since the scaling behavior
is the crucial aspect. Particular applications include the measurement of the
fractal dimensions and the unraveling of universal laws (mean angle between
branches, azimuthal Cantor structures, etc.), with help of the WTMM method
[90,92,375]. In addition, a sophisticated statistical treatment is necessary (fractals
are never exact), which leads to the so-called thermodynamical formalism of
multifractal analysis [93]. Here too wavelet maxima are the appropriate tool for
characterizing the multifractals [388].
Analysis of 2-D turbulence in fluids:
Two-dimensional developed turbulence in fluids is a field where the CWT
gives new insights, in particular concerning the localization of small scales in the
distribution of energy or enstrophy (i.e., the integral of the square of the vorticity
of the given velocity field) and the evolution of coherent structures [274276].
Other applications in fluid dynamics include the visualization and measurement
of a velocity field with help of an oriented wavelet (see below).
Texture analysis:
The determination and classification of textures in images is an old and
difficult problem, with many potential applications (see [235] for a review and
further references). The preferred tool to that effect consists of Markov random
fields. However, since most textures are oriented, it is natural to try and use the
2-D CWT with directional wavelets for attacking the problem. Some progress on
the texture problem has been achieved along these lines in [337], in particular for
the so-called shape from texture problem [207,389], which consists in estimating
the shape of a 3-D object from the texture of an image. One of the key steps is the
generalization to 2-D of the algorithm for measuring the instantaneous frequency
of the signal (which becomes here the local wave vector) and the systematic use
of the ridge or skeleton of the CWT, both familiar in 1-D (see Sect. 12.6).
14.3 The Two-Dimensional CWT 431

Fig. 14.4 Measuring the velocity field in a turbulent fluid, with a Morlet wavelet. (Left) The dot-
bar signature of tracers in the fluid; (Right) A quasi-laminar flow (top), a turbulent flow around an
obstacle (bottom)

Medical physics and psychophysics:

Medical imaging, in particular 2-D NMR imaging and tomography, is an
important application, and much work is in progress in this field. Another
one is modelling of human vision, for instance the definition of local contrast
in images. This notion is based on a nonlinear extension of the CWT, which
leads to interesting mathematical developments, under the name of infinitesimal
multiresolution analysis (MRA) [65, 268, Duv91], but also to faster algorithms,
even competitive with the discrete WT.
As an illustration of the physical applications of 2-D wavelets, we give three
examples, all of them of a directional nature. The first two concern fluid mechanics,
and both rely on the possibility of directional filtering described in the previous
section.
The first example is a straightforward application of the method described above
[609,Wis93]. The aim is to measure the velocity field of a 2-D turbulent flow around
an obstacle. Velocity vectors are materialized by small segments (tiny plastic balls
are injected into the stream and photographed with a fast CCD camera). Then the
WT with a Morlet wavelet is computed twice: first the WT selects those vectors that
are closely aligned with the wavelet; then the analysis is repeated with a wavelet
oriented in the orthogonal direction, thus completely misoriented with respect to
the selected vectors: now the WT sees only the tips of the vectors and their length
may be easily measured. Using appropriate thresholdings, the complete velocity
field may thus be obtained, in a totally automated fashion, with an efficiency at least
comparable to that of more traditional methods. An example of such an analysis is
given in Fig. 14.4.
432 14 Multidimensional Wavelets and Generalizations

A second example concerns the disentangling of a wave train, represented

by a linear superposition of damped plane waves. The problem originates from
underwater acoustics: when a point source emits a sound wave above the surface
of water, the wave hitting the surface splits into several components of very
different characteristics (called respectively direct, lateral and transient), and
the goal is to measure the parameters of all components. This phenomenon has
been analyzed successfully with the WT both in 1-D [550] and in 2-D [58]. In the
latter case, the signal representing the underwater wave train is taken as a linear
superposition of damped plane waves:

N
f (x) = cn eikn x eln x , (14.51)
n=1

where, for each component, kn is the wave vector, ln is the damping vector, and
cn a complex amplitude. Then, using successively the scale-angle and the position
representations described above, one is able to measure all the 6N parameters of this
signal with surprising ease and precision. And the 3-D version is no more difficult,
except for the numerical cost.
Our third example shows the CWT as a symmetry scanner. Many objects are
invariant, at least locally, under specific rotations, sometimes combined with dilation
with a given factor (inflation). Such are, for instance, regular geometrical figures
(a square, a hexagon), Penrose tilings or the diffraction pattern of quasicrystals. For
such objects, the 2-D CWT with a highly directional wavelet (e.g. Cauchy) is an
efficient symmetry detector. The relevant tool is the so-called scale-angle measure
of the object. For a signal s(x), this is the function

S (a, ) = |S (b, a, )|2 db, (14.52)
R2

which may also be interpreted as the partial energy density in the scale-angle
variables. If only the rotational behavior is required, one may further integrate
S over a. When the object is invariant under rotation by 2 /n, this function is
2 /n-periodic in . If the object is invariant under dilation by a factor ao , the
same behavior is visible in the function S (a, ). Thus in the case of an inflation
invariance, S (a, ) is a doubly periodic function.
We give in Fig. 14.5 an application of this analysis on a Penrose tiling (top).
The figure on the bottom left shows the scale-angle measure S (a, ) itself, and
the one on the right its local maxima. On this picture, it is clear that the object is
invariant under dilation by , the golden mean, under rotation by /5, but also under
a combined rotation-dilation symmetry, namely a rotation by /10 together with a
dilation by a factor = 1.36. Note that in this particular case, in order to achieve
a better scale localization, we have used the modified Gaussian Cauchy wavelet
(14.32), with parameters m = n = 4, = 16.
A similar result may be obtained with a quasiregular lattice, for instance the
diffraction pattern of a quasicrystal. An example is given in Fig. 14.6. As can
14.3 The Two-Dimensional CWT 433

2.8 2.8

2.6 2.6

2.4 2.4

2.2 2.2
In(a)

In(a)

2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2

0 18 36 54 72 90 108 126 144 162 180

0 18 36 54 72 90 108 126 144 162 180

Fig. 14.5 Symmetry detection with the CWT: (top) A Penrose tiling; (bottom) The corresponding
scale-angle measure S (a, ), obtained with a (modified) Gaussian Cauchy wavelet supported in
C (10 , 10 ) (left); the same in contour levels (right); this pattern has a rotation symmetry by
/5, a dilation symmetry by and a mixed symmetry, consisting of a rotation by /10 combined
with a dilation by = 1.36. Homologous maxima are linked by a line segment

be read from the graph of the scale-angle measure, this pattern has, in addition
to the obvious invariance under rotation by /4, two distinct combined rotation-
symmetries, namely rotation by /8 together with a dilation by a factor
dilation
1 = 2 cos( /8), resp. 2 = 2 cos( /8) (the first one is only approximate). Other
examples include a tiling with fivefold symmetry built from the root diagram of a
certain infinite-dimensional Lie algebra [73]; or a pattern of symmetry 12 generated
434 14 Multidimensional Wavelets and Generalizations

a b 2

1.8

1.6

1.4

1.2

ln(a)
1

0.8

0.6

0.4

0.2

0
0 20 40 60 80 100 120 140 160 180

Fig. 14.6 Similar analysis of an octagonal pattern: (a) the pattern; (b) the local maxima of its
scale angle measure S (a, ); this pattern has a rotation symmetry by /4, and
two distinct mixed
symmetries, consisting of a rotation by /8 combined with a dilation by 1 = 2 cos( /8), resp.
2 = 2 cos( /8). Homologous maxima are linked by a line segment, continuous for 1 and dashed
for 2

by the Faraday instability in fluid dynamics [74]. In each case, the pattern possesses
combined rotationdilation symmetries, in addition to the obvious rotation symme-
try. It should be remarked that, in all cases analyzed so far, including the Penrose
tiling of Fig. 14.5, these symmetries, previously unknown, were discovered on the
graph of the scale-angle measure, not on the pattern itself. This illustrates the power
of the method. Furthermore, there exists a technique that guarantees that no other
hidden symmetry is present (see [72, 388] for further details).
An interesting feature of this type of situation is that we have here (approximate)
invariance under a semigroup (dilation + rotation). This opens the way to a whole
new class of problems in wavelet (or CS) theory: What happens if groups are
replaced by semigroups? The question is fascinating, but very little is known so
far in this direction, except in the discrete abelian case that we have discussed in
detail in Sect. 13.5.

14.4 The Discrete WT in Higher Dimensions

As we have seen in Sect. 13.1, a key step in the success of the 1-D discrete WT
was the discovery that almost all examples of orthonormal bases of wavelets may
be derived from a MRA, and furthermore that the whole construction may be
transcripted into the language of QMF filters. In the 2-D case, the situation is exactly
the same, as we shall sketch in this section. Further information may be found in
[232, Mey92].
14.4 The Discrete WT in Higher Dimensions 435

The simplest approach consists in building a 2-D MRA simply by taking the
direct (tensor) product of two such structures in 1-D, one for the x direction, one for
the y direction.
If {V j , j Z} is a MRA of L2 (R), then {V j = V j V j , j Z} is a MRA of
L (R2 ). Writing again V j W j = V j+1 , it is easy to see that this 2-D analysis
2

requires one scaling function: (x, y) = (x) (y), but three wavelets:

h (x, y) = (x) (y), v (x, y) = (x) (y), d (x, y) = (x) (y). (14.53)

As the notation suggests, h detects preferentially horizontal edges, that is,

discontinuities in the vertical direction, v and d , vertical and oblique edges,
respectively.
From these three wavelets, one gets an orthonormal basis of V j as { j,k (x, y) =
j,k1 (x) j,k2 (y), k = (k1 , k2 ) Z2 }, and one for W j in the same way, namely
{j,k (x, y), = h, v, d, and k = (k , k ) Z2 }. Here, as in Sect. 13.1.1, for any
1 2
F L (R ), the function Fj,k is defined as
2 2

Fj,k (x, y) = 2 j F(2 j x k1 , 2 j y k2 ). (14.54)

As in the 1-D case, the implementation of this construction rests on the pyramidal
algorithm of Mallat [456, 457]. The technique consists in translating the multireso-
lution structure into the language of QMF filters, and putting suitable constraints on
the filter coefficients. For instance, has compact support if only finitely many of
them differ from zero.
Clearly this construction enforces a Cartesian geometry, with the horizontal and
the vertical directions playing a preferential role. This is natural for certain types of
images, such as in television, but is poorly adapted for detecting edges in arbitrary
directions. Other solutions are possible, however. More isotropic wavelets may
be obtained, either [Mar82] by superposition of wavelets with specific orientation
tuning or by choosing a different way of dilating, using a nondiagonal 2-D dilation
matrix (which amounts to dilating by a noninteger factor) [232]. Consider, for
instance, the following dilation matrices:
     
20 1 1 1 1
D0 = , D1 = , D2 = . (14.55)
02 1 1 1 1

The matrix D0 corresponds to the usual dilation scheme by powers of 2, whereas

D1 and D2 lead to the so-called quincunx scheme [Fea90]. Actually only one
wavelet is needed in this scheme, instead of three. This is consistent with the
theorem according to which the number of independent wavelets needed in a given
multiresolution scheme equals | det D| 1, where D is the dilation matrix used.
However, the scheme based on orthonormal wavelet bases is too rigid for most
applications and various generalizations have been proposed, such as biorthogonal
wavelet bases, wavelet packets or second-generation wavelets, exactly as in 1-D.
436 14 Multidimensional Wavelets and Generalizations

In particular, the lack of translation covariance is a serious defect, for which various
remedies have been proposed, for instance the use of a Cartesian grid instead of the
dyadic onewhich amounts to make a step toward the continuous transform, as in
the construction of pseudo-QMFs discussed in Sect. 14.5.1 below.
Another generalization of the standard DWT is the extension to other numeration
systems, as we have discussed in one dimension in Sect. 13.3. Precisely, the case of
-wavelets in the plane has already been explored in [319]. The aim is to construct
wavelet bases which would be better adapted to the intrinsic geometry of certain
patterns, such as the various aperiodic tilings developed for studying quasicrystals
[147]. This new orientation reveals the need for a more general algebraic structure
than the one commonly used in multiresolution. Significant progress in this direction
has been achieved in [75], as we have discussed in Sects. 13.3 and 13.5.
It should be noted that both in 1-D and in 2-D, the -wavelets obtained are
of the Haar type, i.e., piecewise constant. The same result can be reached for
orthonormal wavelet bases on arbitrary plane domains. The usual technique is that
of triangulation followed by successive partitions. This means, one starts from a
given triangulation of the domain, then one subdivides each triangle into a number
of triangles of equal areas and iterates the refinement process. Typically, one speaks
of 13 splitting [530] or 14 splitting [508]. The outcome is an MRA that leads to
a basis of piecewise constant wavelets, orthogonal with respect to a weighted inner
product adapted to the original triangulation, common in computer graphics.

14.4.1 Applications of the 2-D DWT

As with other methods, wavelet bases may be applied to all the standard problems
of image processing. The main problem of course is data compression, and for
achieving useful rates one has to determine which information is really essential
and which one may be discarded with acceptable loss of image quality. Significant
results have been obtained in the following directions:
Representation of images in terms of wavelet maxima [460], as a substitute for
the familiar zero-crossing schemes [Mar82];
In particular, application of this maxima representation to the detection of edges,
and more generally detection and analysis of local singularities [458];
Image compression: a first approach consists in combining the previous wavelet
maxima method for contours and biorthogonal wavelet bases for texture descrip-
tion [290]; however, this is now superseded by the new standard JPEG-2000,
entirely based on biorthogonal wavelet bases [2].
Denoising, by combining the DWT with statistical techniques [257].
A systematic analysis of multiscale methods in image and texture analysis may
be found in the thesis of Peyr [Per05].
Some applications are less conventional. For instance, a technique based on
biorthogonal wavelet bases is nowadays used by the FBI for the identification
14.4 The Discrete WT in Higher Dimensions 437

of fingerprints [168]. The advantages over more conventional tools are the ease
of pattern identification and the superior compression rates, which allows to
store and transmit a much bigger amount of information in real time. Another
striking application is the deconvolution of noisy images from the Hubble Space
Telescope, by a technique combining the DWT with a statistical analysis of the data
[164, 524, Bou93]. The results compare favorably in quality with those obtained
by conventional methods, but the new method is much faster. Yet another field
of applications (although it was done before the wavelet techniques were born) is
constructive quantum field theory: various perturbation expansions (the so-called
cluster expansion) used in the analysis of field theory models [129] are in
fact discrete wavelet expansions (the summation over scales, indexed by j, was
originally motivated by renormalization group arguments). Finally, one should also
quote a large amount of work under development in the field of High Definition
Television, where wavelet techniques are being actively exploited; here again the
huge compression rates make them specially interesting.
Nevertheless, the standard 2-D DWT remains suboptimal in image processing,
because it does not take into account the geometry of the image. Typically, a line
or a curve in an image is a one-dimensional feature. Hence analyzing it with the
DWT, with square cells, is not natural and results in an explosion of the number of
relevant cells, thus of non-negligible wavelet coefficient. This is the so-called curse
of dimensionality. As a remedy, a number of generalizations, more geometrically
driven, have been introduced, such as ridgelets, contourlets, beamlets, or shearlets.
We will briefly describe some of these in Sect. 14.5 below, referring the reader to
[Ant04, Chap. 11], [395] or the original papers for more details.

14.4.2 The DWT of Radial Functions

In Sect. 14.2.2, we have formulated the CWT of radial functions on Rn and this has
led us to the notion of Bessel-Kingman hypergroup Hn/21 . Now we proceed to
discretize this formalism. In the context of radial functions, the Fourier transform is
replaced by that of Hankel transform, namely,

f( ) = jn/21 ( r)( f (r) dn (r), f L2 (Rn , n ), (14.56)
0

where jn/21 (z) is a spherical Bessel function and n is the measure on R+ given
by dn (r) := |Sn1 | rn1 dr. Then f  f is a self-inverse isometric isomorphism of
L2 (Rn , n ). One may note also the relations

jn/21 (|z|) = eiz |  d ( ) (14.57)
Sn1

? 
r f ( ) = jn/21 ( r) f ( ). (14.58)

These relations explain why spherical Bessel functions will play a central role in the
sequel.
438 14 Multidimensional Wavelets and Generalizations

For wavelets on R, the sampling lattice of the (b, a) half-plane underlying

multiresolution is the dyadic lattice {(k2 j , 2 j ), k, j Z}, as described in Sect. 13.1.1.
In this scheme, the discretization of the translation parameter corresponds to
the related group structure of R and, therefore, is not the appropriate for radial
functions. Instead, Rauhut-Rssler [517] propose a lattice where the discretization
of the translation parameter involves the positive zeros of spherical Bessel function
jn/21 . The trouble is that those zeros are, in general, not equidistant. Fortunately,
this objection does not apply in the physically interesting case of R3 , since then
j1/2 (r) = sinr r , with positive zeros k , k N. We write (k) = k .
Let us write L2 (H) = L2 (H1/2 , d3 ), where H1/2 is the Bessel-Kingman
hypergroup of order 1/2. The generalized translation on H may be written as

 r+u
1
r f (u) = f (t)t dt.
2ru |ru|

Then the normalized Fourier-Bessel functions Mk j1/2 (k r), with Mk := 21/4 5/4 k,
&
k N, form an orthonormal basis in L2 ([0, 1], 2/ r2 dr).
Using these ingredients, one may now introduce a radial multiresolution analysis
in L2 (H) exactly as in L2 (R), as given in Definition 13.1.1, except for the condition
2, which is replaced by
2 . There exists a (scaling) function L2 (H) such that B := {Mk (k) , k N}
is a Riesz basis of V0 .
Notice that, contrary to the classical case, the function itself does not belong to V0
and V0 is not shift-invariant. In fact, it can be shown that, if f V0 , then (k) f  V0
for all k.
With this definition, one can develop the whole machinery of the DWT, following
the standard pattern. We refer to the original paper [517] for the details.
A similar method was used recently in [439] for constructing wavelets supported
in the unit ball B3 , both continuous and discrete, with applications to denoising data.

14.5 Generalizations of 2-D Wavelets

The 2-D discrete wavelet transform that we have sketched in Sect. 14.4 is widely
used, but it has drawbacks. Although it is optimally economical, since it yields
(bi)orthogonal bases, it is too rigid. In fact redundancy is helpful, in that it
increases both flexibility and robustness to noise of the transform. Indeed, the
wavelet community has seen in the last few years a growing trend towards more
redundancy and the development of tools more efficient than wavelets, such as
ridgelets, curvelets, shearlets, etc. The key words here are sparsity and geometry: the
new transforms and approximation methods yield sparser representations of signals,
14.5 Generalizations of 2-D Wavelets 439

because they take much better into account their geometrical features. To give a
simple example, a smooth curve is in fact a 1-D object and it is a terrible waste (of
time or bits) to represent it by a 2-D transform designed for genuine 2-D images.
A review of these generalizations may be found in [Wel03], aptly entitled Beyond
Wavelets. Before turning to the new tools, we have a quick look at wavelet packets.

14.5.1 Continuous Wavelet Packets in Two Dimensions

Continuous wavelet packets may be defined in 2-D by a straightforward extension

of the 1-D construction described in Sect. 13.2. However, the full benefit of this
approach shows up if one works in polar coordinates, for then the directional
variable may be controlled. To make things clearer, we shall change notation in
this section. The 2-D spatial frequency vector is denoted henceforth byk, with polar
coordinates (k, ), k = |k|, = argk.
Then, starting from an isotropic wavelet , one gets an isotropic scaling function
 iso (k). But one can do better and
 (k) and a family of isotropic integrated wavelets
i
design directional pseudo-QMFs as follows. Let {l ( ), l = 1, . . . d} be a resolution
of the identity consisting of C , 2 -periodic, functions of compact support, so that
dl=1 l ( ) = 1. Then one obtains a family of directional integrated wavelets, in
polar coordinates, as

 iso (k) l ( ),
i,l (k, ) = (14.59)
i

and indeed one has

d
i,l (k, ) = iiso (k). (14.60)
l=1

The net result is a discrete, fast, implementation of the 2-D CWT, including the
directional degree of freedom. Here the technique of periodizing the signal and
adapting the length of the pseudo-QMFs to that of the signal becomes crucial
for obtaining a fast algorithm. This approach has found interesting applications,
for instance in directional filtering [592, Van98]. More information about this
construction may be found in [Ant04, Sect. 2.6].
Actually, there is another technique for designing 2-D wavelets that are sensitive
to directions, namely, the use of the so-called steerable filters. The idea is to
build a finite wavelet frame by taking a number of oriented wavelets, obtained by
successive rotations of a given one, in such a way that all directions are covered.
This idea, introduced by Freeman and Adelson [288], has been extremely fruitful in
applications (see [591] for an up-to-date review and further references). The concept
was later extended to directional analysis on the two-sphere in [604]. But in fact,
directional wavelet packets and steerable wavelets are two incompatible notions.
440 14 Multidimensional Wavelets and Generalizations

14.5.2 General Directional Wavelets

Actually, the most interesting idea in the previous construction is that of wavelets of
the type (14.59), separable in polar coordinates in Fourier space. Indeed this yields
an interesting class of 2-D directional wavelets of the general form

 (k) = (k) w( ),
(14.61)

where w is a normalized, positive angular distribution:


w( ) 0, w( ) d = 1.

If one takes a function w with support in [ , ], with < /4, one obtains a
conical wavelet supported in the cone C ( , ). For instance, the Gaussian conical
wavelet (14.32), for = 1, would correspond to
@
sinm ( ) sinm ( + ), if | | < ,
w( ) =
0, otherwise.

On the same model, one can obtain a wavelet supported in a double cone (thus as
directional as the previous one) by taking
2 3
 (k) = k1 (k) w( ) + (k) w( ) .
(14.62)

Another example is the so-called von Mises distribution, which is the angular
analogue of the Gaussian distribution:

w( ) = exp( cos )/{2 I0 ( )},

where is a dispersion parameter and I0 is a modified Bessel function. Coupling

1 2
this with (k) = kL e 2 k , one gets a wavelet with properties very similar to those
of the Gaussian conical wavelet (14.32).
The idea of using wavelets that are polar separable in frequency space will be
exploited systematically in the generalizations that we shall study in the following
sections.

14.5.3 Multiselective Wavelets

The first example is a family of wavelets with adaptive angular resolution, developed
in [394, Jac04], that we follow closely.
Let us particularize to 2-D the CWT given in (14.12) and (14.13):
14.5 Generalizations of 2-D Wavelets 441


(W s)(b, a, ) = a2 (a1 r1 (x b))s(x) dx (14.63)
R2


=  (ar1k) s(k) dk.
ei kb (14.64)
R2

Here Ws = a1 S , is an L1 -normalized wavelet, that we assume directional, as

defined in Sect. 14.2 and r is the usual 2 2 rotation matrix by an angle S1
[0, 2 ). In addition, we assume that
 is separable in polar coordinates, namely, of
the form (14.61) where w is a positive function in L2 (S1 , d ).
Then, the integral (14.64) becomes in polar coordinates

(W s)(b, a, ) = (ak) w ( ) s(k, ) ei kb cos( ) k dk d , (14.65)
R+ S1

with w ( ) = w( ), b = |b| and = argb. The last equation may be rewritten as

(W s)(b, a, ) = w | Rb,a S , (14.66)

1

>
where  f | g 1 = S1 d f ( ) g( ) is the scalar product on the circle S1 of the
S
2 -periodic functions f and g, and

Rb,a ( ) = (ak) s(k, ) ei kb cos( ) k dk. (14.67)
R+

In other words, the relation (14.66) means that the wavelet transform of the image s
can be interpreted as the projection of Rb,a on a kind of scaling function w localized
around S1 .
In addition, one can control the angular selectivity of the wavelet by applying a
dilation D on the circle, to get a family of functions w , = D w , which in turn
lead to wavelets  (k) = (k) w ( ) of angular selectivity controlled by .
Next we discretize the whole process. Indeed, let L1 (R2 ) L2 (R2 ) be a polar
separable wavelet of the form (14.61). Given a dyadic sampling of the scales a j =
a0 2 j ( j Z, a0 R+ ) and a regular sampling of the angles n = 2K n (0 n < K,
K N ), assume that the frame property is satisfied, that is, there are two positive
constants m > 0, M < such that

K1
m | (a j k)|2 |w( n )|2 M. (14.68)
jZ n=0

Then, there exists a dual wavelet

 , nonunique, of course, which yields the
following reconstruction formula:

K1
s(x) = W j,n 
a j ,n (x), s L2 (R2 ), (14.69)
jZ n=0
442 14 Multidimensional Wavelets and Generalizations

where W j,n (b) = (W s)(b, a j , n ) and  denotes the standard convolution of two
sequences defined by u  v[n] = mZ u[n m] v[m].
A particular case known as a LittlewoodPaley (LP) frame1 arises when

K1
(a j k) w( n ) = c, c R+ . (14.70)
jZ n=0

Then,  is simply a Dirac distribution (2) (x) and the reconstruction formula
(14.69) reduces to a LittlewoodPaley decomposition

K1
1
s(x) =
c W j,n (x). (14.71)
jZ n=0

Next we proceed to define new 2-D wavelets with multiple, but controlled, angular
selectivities, with the following properties [394, Jac04]:
They combine with each other in a pyramidal scheme to form less selective
directional wavelets until one obtains a totally isotropic one;
They define for each selectivity level a LittlewoodPaley (LP) directional frame.
The idea is to use a MRA on the circle, obtained by periodization of an MRA on the
line, as described in Sect. 13.1.1. We start from the relation (14.66) and discretize it
as indicated above. Then, given a biorthogonal multiresolution analysis of the circle,
with a scaling function and a wavelet , we project Rb, j := Rb,a onto the functions
    j
l,n ( ) = l,n 2 and l,n ( ) = l,n 2 . Thus we get the new coefficients

a 
W j,l,n (b) = l,n | Rb, j  1 ,
S
d 
W j,l,n (b) = l,n | Rb, j  1 ,
S

W j (b) = 0,0 | Rb, j 

i
=  1 | Rb, j  1 , (14.72)
S1 S

for l N and 0 n < 2l . These amount, respectively, to the projection of the image
s on translated and dilated copies of the functions l,n
a , d and i , respectively,
l,n
defined in frequency space by
a 

l,n (k) = (k) l,n ( ),
d 

l,n (k) = (k) l,n ( ),

 i (k) =
a 
0,0 (k) = (k). (14.73)

1 These frames are sometimes called linear, e.g. in [Tor95].

14.5 Generalizations of 2-D Wavelets 443

Here the exponent a stands for angular approximation, d for angular details, and
i for isotropic. The full parameterization of the L1 (R2 )-normalized wavelets thus
reads
!x b "
b,a j,l,n (x) = a2
j l,n
a
, (14.74)
aj

and similarly for d (x) and i .

b, j,l,n b, j
Besides the scale and translation parameters a j and b, the rotation of the
2-D wavelets is obviously given by the parameter n, 0 n < 2l , which precisely
translates a function on S1 by an angle 2 n2l . Since the aperture of the fre-
quency space cones containing these wavelets is proportional to 2l , the angular
selectivity of l,n
a and d is proportional to 2l . Thus we may call the parameter l
l,n
the (angular) selectivity level.
We have thus generated a new family of wavelets, {l,n a , d , i } ranging from
l,n
very directional ones to a totally isotropic one, depending on the value of l. The
key point is that, for each selectivity level l N, the family { j,l,n
a } generates a

Littlewood-Paley (LP) frame, as stated in the following proposition [394, Jac04].

Proposition 14.5.1. Let a j = a0 2 j , j Z, be a dyadic scale discretization and
suppose that jZ (a j k) = 1 a.e. for k R+ . Then, for any l N, the family
{ aj,l,n : j Z, 0 n < 2l } is a LP frame of L2 (R2 ) with constant c = 2l/2 .

This is a simple consequence of the fact that the function R , which leads to
by periodization, realizes a partition of the line, i.e., mZ R (t + m) = 1 for every
t R.
Like every wavelet, for any l N , the functions and verify simple scaling
rules analogous to (13.4), namely,

2l 1
l1,0 ( ) = hl [n] l,n ( ), (14.75)
n=0

2l 1
l1,0 ( ) = gl [n] l,n ( ). (14.76)
n=0

where hl [n], gl [n] are the filter sequences coming from the MRA. Thus, several
wavelets of level l merge into a single wavelet at level l 1, i.e., with half the
angular selectivity.
Next comes the choice of the basic wavelet. Given an initial selectivity level
L N, we define L,n a and d in frequency as
L,n

a 

L,n (k) = R (log2 k) L,n ( ), (14.77)
d 

L,n (k) = R (log2 k) L,n ( ), (14.78)
444 14 Multidimensional Wavelets and Generalizations

with 0 n < 2L , L,n ( ) = L,n ( 2 ), L,n ( ) = L,n ( 2 ), and and the

periodization of R and R , respectively. This yields K = 2L differently oriented
wavelets. Remark that, choosing (k) = N (log2 k), as in (14.77), implies that
jZ (a j k) = jZ R (log2 a0 k j) = 1, since mZ R (t m) = 1. Thus, by
Proposition 14.5.1, the wavelets { aj,l,n } constructed from (14.77) constitute a LP
frame for each l N. Therefore, starting from a selectivity level L N, we can
generate inductively L + 1 frames with l,n a (0 l L), characterized by an angular

selectivity 2 . In particular, for each level l [0, L], Proposition 14.5.1 shows that
l

the wavelets j,l,n

a generate a LP frame of constant c = 2l/2 which, according to
(14.71), may be used to reconstruct the original image. But there is more. Indeed,
as shown in the following proposition, we can mix different frames inside the same
reconstruction formula [394, Jac04].
Proposition 14.5.2. If satisfies the same condition as in Proposition 14.5.1, then,
given any function 
l : (x, j) R2 Z  
l(x, j) N, a function f L2 (R2 ) can be
decomposed as


2l 1

f (x) = 2l/2 W j,al,n (x), 
l 
l(x, j). (14.79)
jZ n=0

This property yields a new degree of freedom to describe images adaptively. Indeed,
at each point b R2 and each scale j Z, we may search the best frame, that is,
the selectivity level l(b, j) characterizing best the content of f . To that effect, one
possibility is to choose the frame that offers the best match between the image and
the wavelets, that is, at (b, j), to choose the selectivity level

|a | f |
(b, j) = arg max max
b, j,l,n
. (14.80)
l[0,L] n[0,2l ) a 
b, j,l,n

Of course, other selection rules may be chosen, for instance, a sparsity criterion
based on the 1 minimization of normalized coefficients.
The prime application of this technique is image denoising. Let s be a band-
limited image corrupted by a Gaussian white noise of vanishing mean and variance
2 , that is,

s (x) = s(x) + N (x), (14.81)

where s is the noisy image and N is the white noise term. The problem is to
estimate s from s . This is done by computing the wavelet coefficients of the noisy
image s , (softly) thresholding them and recover the rest using the reconstruction
formula that derives from (14.79). In the multiselective case, the thresholding is
adapted to the local selectivity 
l(x, j). Experiments have been performed in [394,
Jac04], using a standard CohenDaubechiesFeauveau, compactly supported, spline
14.5 Generalizations of 2-D Wavelets 445

biorthogonal wavelet basis and the selection criterion (b, j) defined in (14.80). The
result is that the multiselective method is substantially more efficient at denoising
the images, in terms of peak signal to noise ratio (PSNR). We refer to the original
source for quantitative details.

14.5.4 Ridgelets

A given n-dimensional signal may have different types of singularities:

0-dimensional (point singularities), 1-dimensional (line or curve with bad behavior),
(n 1)-dimensional (hypersurface with bad behavior). Among these, the DWT is
fully efficient only for dealing with 0-dimensional singularities. Thus, in the case
of images, one needs another tool for analyzing 1-dimensional singularities. In a
nutshell, the goal is to find the method that yields the sparsest representation of
signals [i.e., functions f (x, y))] that contain both smooth regions and edges.
The simplest case is that of line singularities, and the ridgelet transform is
precisely designed for dealing with them. Notice that ridgelets, in turn, are not well-
adapted to curves: these require a more sophisticated tool, naturally called curvelets,
that we shall discuss in Sect. 14.5.5 below.
We have seen in Sect. 14.3.2 that the 2-D CWT is also a very efficient tool to
this effect, since it is able to localize the singularity and its orientation. On the other
hand, if the data to be analyzed is a truly infinitely long ridge, there is no information
along its principal direction. Thus one could save an unnecessary degree of freedom
in the transform. These considerations led Cands [Can98] to introduce a slightly
modified version of the 2-D CWT, namely the continuous ridgelet transform (CRT).
Starting from the original definition (14.63) and (14.64), one requires that a
ridgelet behaves like a 1-D wavelet in a given direction, represented by the unit
vector u of orientation , and be constant along the orthogonal direction. Such an
object can then be shifted along its oscillating direction, rotated and scaled in order
to yield a family of continuous ridgelets:
!u x b "

a, ,b (x) = a1/2 . (14.82)
a
Equivalently, ridgelets can be seen as usual 2-D wavelets, except they are constant
along a preferred direction and will thus never be admissible. Note that, in (14.82),
is a function of one variable only, typically a 1-D wavelet. We manifest this by using
a different notation, namely, a, , b as subscript instead of the traditional b, a, .
Thus the parameter space of the CRT is = {(a, , b), a R+ , SO(2), b R},
with measure d (a, , b) = a3 da d db. Notice that 1,0,0 (x) = (x), with Fourier
2 3
transform 1,0,0 (k) = 2
 (kx ) (ky ) = k1 (k) ( ) + (k) ( ) , that is,
the ultimate directional wavelet as defined in (14.62).
446 14 Multidimensional Wavelets and Generalizations

The definition of the CRT follows that of the CWT and shares most of its
properties. First, define an admissible ridgelet, or simply ridgelet, as a function of
zero mean, i.e., a 1-D wavelet, satisfying the condition

1 |
 ( )|2
c = d < +.
R | |
2 2

As in the wavelet case, this condition is essentially equivalent to imposing at least

one vanishing moment on . Then the CRT of a function f L2 (R2 ) reads as
follows.
Proposition 14.5.3. Let be a (real-valued) ridgelet. Then the CRT of a function
f L2 (R2 ) is the function R f L2 ( , d (a, , b)) defined by
 !u x b "

(R f )(a, , b) = a, ,b | f  = a1/2 f (x) dx .
R2 a

As for the CWT, the CRT admits a reconstruction formula.

Theorem 14.5.4. Any f L1 (R2 ) such that f L1 (R2 ) admits the following
decomposition:

f = c1
a, ,b | f a, ,b d (a, , b).

The CRT is a stable representation, since it satisfies a Plancherel formula:

Theorem 14.5.5. For any f L1 (R2 ) L2 (R2 ) and an admissible ridgelet,
one has

 f 22 = c1
|a, ,b | f |2 d (a, , b).

The next step, of course, is discretizing the CRT for practical implementation. In
particular, one wonders about the existence of ridgelet frames and their properties.
Such a construction was proposed by Cands [180, 186, Can98], to which we refer
for details. As usual, the problem is to find a discretization (a j , j, , b j,k ) of the
parameter set such that one can sample the CRT in a stable manner, that is,
satisfying a norm equivalence (frame condition):

m f 2 |(R f )(a j , j, , b j,k )|2 M f 2 . (14.83)

j,k,

First, the scale and position variables are sampled exactly as in the case of wavelets
or LittlewoodPaley analysis, i.e., a j = a0 2 j , b j,k = k b0 a0 2 j . The difference
with respect to usual 2-D wavelet frames comes in the particular sampling of the
rotation parameter. In order to reach the equivalence (14.83), Cands found out that
14.5 Generalizations of 2-D Wavelets 447

Fig. 14.7 Comparison between the wavelet (left) and ridgelet (right) tilings of the frequency plane
(from [Ant04, Sect. 11.1.1])

the angular resolution should increase at finer scales, j, = 2  2 j , corresponding

to a j = a0 2 j . The difference between the ridgelet and wavelet discretization of the
frequency plane is illustrated in Fig. 14.7 (in which j increases outwards).
More information about the properties of the CRT and its application in image
processing can be found in the seminal reference [Can98] and in the papers [180,
185, 186]. Applications to approximation theory and characterization of function
spaces are discussed in [181]. A different and illuminating way of understanding
the CRT consists in seeing it as a combination of a Radon transform followed by
a classical 1-D wavelet transform, as already indicated by Holschneider [377]. For
more details on this, see [186] or [Ant04, Sect. 11.1.2].

14.5.5 Curvelets and Contourlets

As we said above, ridgelets are specifically designed and optimal for representing
line singularities. Now real images may also contain curved singularities, contours
or edges, in addition to smooth regions (the so-called cartoon images). In that case,
neither wavelets nor ridgelets will give a sparse representation, one has to resort to
a better adapted tool, namely, curvelets.
To make things precise, suppose an image contains an edge modelled by a C2
curve y(x), that we represent by its Taylor series:

x2 
y(x) y(0) + xy (0) + y (0) + O(x3 ).
2
Using the translation and rotation degrees of freedom and neglecting higher order
terms, we can always assume the reduced form:
c 2
y(x) x ,
2
448 14 Multidimensional Wavelets and Generalizations

where c = y (0) is the curvature at the analysis point. Such a curve has a natural
parabolic scaling: Dilating along the x axis by a factor a amounts to dilate along the
y axis by a factor a2 . Restricting our attention to signals smooth away from C2 edges,
the ideal representation would thus consist of localized atoms obeying a parabolic
scaling law and looking like anisotropic needles of aspect width length2 .
This is the key ingredient for the curvelet transform introduced by Cands and
Donoho. In a first approach [185, 187], they derived it from ridgelets. Later [189,
190], they introduced an improved version (second generation curvelets), that we
shall now sketch. The continuous curvelet transform (CCT) is based on curvelet
atoms, defined as

a,b, (x) = a (r1 (x b)),

where a (x) = a3/4 (a1 x, a1/2 y). In this relation, behaves as a wavelet in
the x-direction, with vanishing moments, and as a scaling, bump-like function in the
y-direction. Thus, for very small scales a  1, a,b, is a needle of effective length
a1/2 and effective width a, thus verifying the parabolic scaling law width length2 .
Using this idea, Cands and Donoho [190] define the CCT as follows. Take, in
frequency space, a radial window (k), supported in [1/2, 2] and an angular window
v(t), t = cos , both positive and real-valued, and satisfying the normalization
conditions
  1
dk
(k)2 = 1, v(t)2 dt = 1. (14.84)
0 k 1

Then curvelet atoms are obtained as usual by translation and rotation from a basic
function

a,b, (x) := a,0,0 (r (x b)),

where the latter is defined in frequency space as

a,0,0 (k, ) := a3/4 (ak)v(a1/2 ), a < 1.

Notice that  is again polar separable. In terms of these objects, the CCT of f
L2 (R2 ) is defined in [190] as the function (we consider only the high frequencies)

( f )(a,b, ) :=  f | a,b, , a < 1,b R2 , [0, 2 ). (14.85)

This transform admits a reconstruction formula and a Parseval relation.

Theorem 14.5.6. Let f L2 (R2 ) have a Fourier transform vanishing for k < 2.
Let the windows and v satisfy the conditions (14.84). Then one has the following
reconstruction formula
14.5 Generalizations of 2-D Wavelets 449


da 
f (x) = ( f )(a,b, ) a,b, (x) db d ,
a3

and a Parseval relation for high-frequency functions


da 
 f 2L2 = |( f )(a,b, )|2 db d .
a3

The main virtue of the CCT is to localize in a precise fashion singular curves (edges,
contours) in images, in the sense that the quantity  f | a,b,  is nonnegligible iff the
point b lies on the curve and the curvelet element is tangent to it. More generally,
the CCT resolves the wavefront set of a distribution, a concept used in microlocal
analysis. For technical definitions and more details about this, we refer to [190,
Sect. 5].
Of course, all that should be discretized for practical implementations. Fol-
lowing the same pattern as before, one defines [189, 190] a scale parameter j =
0, 1, 2, . . ., an orientation parameter  = 0, 1, . . . , 2 j , and a translation parameter
k = (k1 , k2 ), k1 , k2 Z. Then one introduces:
(1) The parabolic scaling matrix D j = diag(22 j , 2 j );
(2) The rotation angle J = 2 .2 j ., where J = ( j, );
(3) A scaled translation parameter k , where takes discrete values (on a lattice)
depending on j.
With this notation, curvelets are defined as the functions

(x) = 23 j/2 (D j rJx k ), = ( j, ,k ). (14.86)

Here again, the waveform is smooth and oscillatory in the x-direction (wavelet-
like) and bell-shaped (scaling function) in the y-direction. Thus the curvelet (14.86)
is well-localized in space and obeys approximately the relations

length 2 j , width 22 j ,

i.e., it follows the parabolic scaling law. In the frequency domain, one obtains a
tiling of the frequency plane with wedges pointing in the direction J , of length 2 j
and width 2 j/2 , and there are 2 j directions (see Fig. 2.2 in [189]). Thus, here too,
as for the ridgelets of Sect. 14.5.4 and for the multiselective analysis of Sect. 14.5.3,
the angular resolution increases with j. Then the final result reads as
Proposition 14.5.7. Let the curvelets be defined in (14.86). Then, for any F
L2 (R2 ), one has

2
 | F = (2 )2 F2L2 . (14.87)

450 14 Multidimensional Wavelets and Generalizations

By the Plancherel formula, one has also

2
 | f  = (2 )2  f 2L2 . (14.88)

Therefore, the family { } is a tight frame in L2 (R2 ).

The outcome is that curvelets provide an optimally sparse representation of
objects with singularities along piecewise C2 edges. In addition, the correspond-
ing algorithms are fast [192]. They have a wide spectrum of applications, in
image processing (image denoising [566], image reconstruction [191], astronomical
images [567]), statistical estimation [188], representation of Fourier integral opera-
tors [182] or wave propagators [184]. Actually, some of the earlier applications use
first generation curvelets [185].
A technique analogous to curvelets is the contourlet transform [254, 255]. This
construction is made directly in the spatial domain, suitably discretized, instead of
the continuous frequency space. Thus it is closer in spirit to the standard techniques
of signal processing, using a Laplacian pyramid and filter banks. As for the resulting
tiling of frequency space, it is again made of directional wedges, but the geometry
is pseudo-polar instead of polar, which means that the disk is replaced by a square
(see [254, Fig. 4.4] or [395]).

14.5.6 Shearlets

With shearlets, we are coming back to our original philosophy: there is a continuous
transform coming from a square integrable unitary representation of a Lie group
(the so-called shearlet group), then it is discretized for implementation purposes.
Originally introduced in [357,433], the shearlet transform was largely developed
by Kutyniok and her collaborators [221, 431]. For an up to date review, we refer to
[Kut12]. We will sketch the main features of this method, for which there is by now
a fairly large literature.
We start with the shearlet group. We consider triples (a, s,t) where a R+ , s R
and t R2 and define the parabolic scaling matrix Aa and the shear matrix Ss :
  
a 0 1s
Aa = , Ss = . (14.89)
0 a 01

Then one has

Lemma 14.5.8. The set R+ R R2 equipped with the multiplication given by

(a, s,t) (a , s ,t  ) = (aa , s + s a,t + Ss Aat  ) (14.90)

is a Lie group, called the (reduced) shearlet group and denoted Sh(R2 ). Its left Haar
measure is dl = a3 da ds dt, its right Haar measure is dr = a1 da ds dt.
14.5 Generalizations of 2-D Wavelets 451

Note that Sh(R2 ) is an example of semidirect product of the type studied in Sect. 8.3.
Consider indeed the set
   
as a
H(2) = Ss Aa = : s R, a R+ .
0 a

Then Sh(R2 ) = R2  H(2), a closed subgroup of R2  GL(2, R), with multiplication

(8.59), namely,

(t1 , M1 )(t2 , M2 ) = (t1 + M1t2 , M1 M2 ), t R2 , M GL(2, R).

Besides Sh(R2 ), one considers also the full shearlet group, SH(R2 ), obtained by
taking a = 0 instead of a > 0 and the matrix
 
 a 0&
Aa =
0 sgn(a) |a|

l = |a|3 da ds dt, resp.

instead of Aa . In that case, the Haar measures become d
d 1
r = |a| da ds dt.
The group Sh(R2 ) has a natural action on signals L2 (R2 ), in position and
Fourier space, respectively:

a,s,t (x) = [U(a, s,t) ](x) = a3/4 (A1 1

a Ss (x t)) (14.91)

  s,t)
a,s,t (k) = [U(a,  ]k) = a3/4 eikt
 (ATa SsT k)

= a3/4 eikt
 (akx , a(ky + skx )). (14.92)

a .
The same holds for SH(R2 ), with the matrix A
Exactly as for wavelets (Theorem 14.2.1), we have a square integrable
representation of our group.
Theorem 14.5.9. The operator family U() defined in (14.91) is a unitary
irreducible representation of the full shearlet group SH(R2 ) in L2 (R2 ), and it
is unique up to unitary equivalence. This representation is square integrable: a
vector L2 (R2 ) is admissible and called a continuous shearlet iff it satisfies the
condition

|
 (k)|2
c = (2 )2 dk < , (14.93)
R2 kx2

The restriction of U to the reduced shearlet group Sh(R2 ) is unitary, but no longer
irreducible.
452 14 Multidimensional Wavelets and Generalizations

The situation with respect to irreducibility is exactly the same as for 1-D wavelets
(see Sect. 12.2), with the distinction between the full affine group Gaff and its
connected subgroup G+ .
In view of Theorem 14.5.9, we can now simply proceed as in the wavelet case
of Sect. 14.2. The continuous shearlet transform (CST) of a signal f L2 (R2 ) with
respect to the shearlet is defined as

(SH f )(a, s,t) := a,s,t | f . (14.94)

It turns out that the CST is a c -multiple of an isometry. Note that, in the CST
(SH f )(a, s,t), the scale parameter a measures the resolution, the shear parameter
s measures the directionality and the translation parameter t measures the position.
Exactly as in the case of wavelets, there is a weak reconstruction formula:
Theorem 14.5.10. Given a shearlet , any function f L2 (R2 ) may be repre-
sented as

da
f (x) = SH f (a, s,t) a,s,t (x) ds dt,
SH(R2 ) |a|3

the integral being understood in the weak sense.

As usual, this statement may be rephrased by saying that the shearlets a,s,t
constitute a resolution of the identity. Moreover, there is also a strong version of
the reconstruction formula, i.e., with L2 convergence, which is proved with help of
an approximate identity.
Now, what really distinguishes shearlets from general 2-D wavelets is the shape
of the basis function used, namely a function L2 (R2 ) such that
!k "
y

 (k) =
 (kx , ky ) =
1 (kx )
2 , (14.95)
kx

where 1 is a continuous wavelet, 1 C (R), and supp

1 [2, 12 ] [ 12 , 2],

and where 2 is such that 2 C (R) and supp 2 [1, 1]. It is easy to see that

this function is indeed an admissible shearlet. Then, given the function of (14.95),
it follows that the function
 a,s,t is supported on the set:

   
2 1 1 2 ky
supp
a,s,t k = (k ,
x yk ) : kx , , , s a .
a 2a 2a a kx

As illustrated in Fig. 14.8, each continuous wavelet a,s,t has a frequency support
on a pair of trapezoids, symmetric with respect to the origin, oriented along a line
of slope s. The support becomes increasingly thin as a 0.
The shearlets (14.95) are used in [431] to show that the CST precisely resolves
the wavefront set of a distribution and, in particular, curved singularities in images.
In addition, as shown in [221], the CST satisfies an uncertainty principle similar
14.5 Generalizations of 2-D Wavelets 453

Fig. 14.8 Support of the shearlets  a,s,t in the frequency domain for different values of a and s;
the shearlet is the function given in (14.95) (courtesy of G. Kutyniok)

to that of 2-D wavelets discussed in Sect. 14.3.1 and various shearlets of minimal
uncertainty may be derived explicitly.
The next step, is of course, to discretize the CST. Instead of the matrices Aa and
Ss of (14.89), we consider
   
2j 0 1 k2 j/2
Aj = , Sk j = , j, k Z. (14.96)
0 2 j/2 0 1

Also the translations t R2 are replaced by the discrete lattice m Z2 . Then one
considers the shearlet (14.95), assuming in addition that the wavelets 1 and 2
satisfy the conditions

|1 (2 j )|2 = 1, for a.e. R (the Caldern equation),

jZ

|2 ( + l)|2 = 1, for a.e. R.

lZ

There are many possible choices for the functions 1 , 2 . For any one of them, it
turns out that the function is well-localized, i.e.,

| (x)| KN (1 + |x|N ), for any N N.

Next consider the discrete shearlets jkm (x) = s3 j/4 (A1 1

j Sk j x m). Then it turns
out that the system { jkm , j, k Z, m Z } is a tight frame, that is,
2
454 14 Multidimensional Wavelets and Generalizations

| jkm | f |2 =  f 2 , for all f L2 (R2 ).

j,kZ,mZ2

The (rather technical) construction may be found in [358].

A slightly more general shearlet transform has been introduced recently in [406],
by following a more geometric approach. This version introduces another shearlet
6 2 ), which is the set R+ R R2 equipped with the multiplication
group, Sh(R

(a, s,t) (a , s ,t  ) = (a a, s + s a ,t  + Ss Aa t)

Note this amounts to permute the order of factors in (14.90) or, equivalently, to
replace the left action by a right action. However, this group is isomorphic to Sh(R2 )
and thus has a similar unitary representation U in L2 (R2 ), namely,

 s,t) ](x) = a3/4 (Ss Aa x t).

[U(a,

Comparing this formula with (14.91) clearly shows that the new group Sh(R 6 2 ) will
lead to simpler formulas. Next, to obtain a higher flexibility, one adds a weight
function w(a, s,t) > 0 in the definition of the (weighted) shearlet transform, defined
now as

6 ,w f )(a, s,t) = w(a, s,t)1/2 U(a,

(SH  s,t) | f ,

6 2 ), a1 da ds dt).
and taking its values in L2 (Sh(R
This new formulation provides a better geometric understanding of the dis-
cretization process of the CST. In particular, it allows to consider irregular shearlet
systems, defined as

6 , , w) = {w(a, s,t)1/2 a3/4 (Ss Aa t) : (a, s,t) },

SH(

where is any discrete subset of Sh(R6 2 ), whereas the regular systems described
above used only {(a , bk, cm) : a > 1, b, c > 0 and j, k Z, m Z2 }. For a full
j

analysis of the new formulation and future prospects, we refer to the original paper
[406].
In addition, wide-reaching generalizations of the discrete shearlet transform have
been introduced, under the general name of wavelets with composite dilations
[358, 433]. One of them consists of the so-called shearlets on the cone or cone-
adapted shearlets [357, 431]. The idea is to restrict the shear parameter s to a
compact interval instead of R. More precisely, the function to be analyzed is split
into a part with singularities of slope 1 (thus living in a cone) and a part with
singularities of slope 1. Then both parts can be analyzed with the parameter s
ranging in the interval [2, 2]. While this is achieved with very specific shearlets, it
has been shown in [345] that the technique can be considerably generalized, while
still yielding tight shearlet frames. Finally, a more recent evolution is the theory of
compactly supported shearlets, developed in [407, 432].
14.5 Generalizations of 2-D Wavelets 455

A natural conclusion for the group-theoretical analysis of the shearlet transform

is provided by a recent paper [204], in which the authors show that all the groups
relevant to signal analysis can be derived from the (1+1)-dimensional affine Galilei
group. This covers the affine group of the line (1-D wavelets), the WeylHeisenberg
group (Gabor), the reduced shearlet group and the Stockwell group. In view of the
interest of this analysis, we will give a detailed account of it in Sect. 16.2.2. Also,
the CST has been extended to arbitrary space dimensions, with the same properties.
It is derived from a square integrable representation of a specific shearlet group and
it may be applied successfully to the characterization of singularities in signals, in
particular to the resolution of the wavefront set of a distribution [223].
Finally, another point that distinguishes shearlets from the other post-wavelet
constructions is that they can be embedded in coorbit theory. This is a very general
framework developed by Feichtinger and Grchenig in several papers [278, 279].
The starting point is a (square)-integrable unitary representation U of a group G in
a Hilbert space H. From this, one constructs natural families of Banach smoothness
spaces, called coorbit spaces, around H, defined by the decay properties of the
matrix elements of the representation U. For instance, the usual representations
of the affine group and of the WeylHeisenberg group lead to classical Besov
or modulation spaces, respectively (see [Gr01, Chap. 011] or [Ant09, Chap. 8]
for precise definitions, too long to be given here). In addition, the theory yields
a systematic construction of frames in Hilbert or Banach coorbit spaces. This
approach applies well to shearlets and a detailed analysis of the shearlet coorbit
spaces is provided by [222]. For the compactly supported shearlets, in particular,
the coorbit theory yields embedding theorems into Besov spaces, analogous to the
classical Sobolev embeddings theorems [224]. It would be of course out of place
here to go in more detail about this. For the general coorbit theory, we refer to
the original papers [278, 279]. A summary may also be found in [Ant09, Sect. 8.5].

14.5.7 Geometrical Wavelets: Dictionaries, Molecules

All the techniques we have described so far aim at producing frames, preferably
tight, in order to obtain a good representation of a certain class of signals/images,
in particular, with a certain type of singularity. The basic idea is to focus on
the geometry of the image, by designing an analysis tool that is optimal for that
geometry. And the key word in that approach is sparsity: one tries to obtain the
representation of the signal at hand with as few terms as possible in the expansion.
The result, when the enterprise is successful, is rapidity of the algorithms. Note
that this idea of a dictionary was already introduced in the case of the affine
WeylHeisenberg group GaWH , as we will see in Sect. 16.1.
There is, however, another way of attacking the problem, namely, to produce
universal dictionaries for a given type of features in images, typically edges.
Several techniques have been developed, such as wedgelets [258], beamlets [259]
or platelets [608]. These techniques have been very successful in tasks like
456 14 Multidimensional Wavelets and Generalizations

compression, denoising, image segmentation or edge detection. Yet they all suffer
from the same defect. Namely, those atoms of the dictionary that will contribute for
analyzing a given image must be chosen adaptively from the image, that is, all atoms
in the dictionary must be tested and matched to each element of an image partition.
This is, of course, time consuming, so that the wedgelets-based methods are usually
too slow for using in real-time applications. Things may change according to a new
technique, based on the computation of moments in combination with the wedgelet
transform. The result is the so-called Fast Wedgelet Transform, that seems indeed
able to beat the bad time complexity of the usual transform [448].
On the other hand, traditional multiscale methods, such as Gabor or wavelet
frames, while efficient for representing most classes of pseudodifferential operators,
are not as effective when dealing with Fourier Integral Operators or the analysis
of wavefront sets. This has motivated the introduction of curvelets (Sect. 14.5.5)
and shearlets (Sect. 14.5.6). However, in the case of curvelets, the definition of the
basic elements is too restrictive. Indeed, operators of interest do not map curvelets
into curvelets, but rather into something more general, called a curvelet molecule,
introduced in [183]. In a nutshell, a curvelet molecule is a needle whose envelope
is supported near a ridge of length about 2 j/2 and width 2 j and which displays
an oscillatory behavior across the ridge. The curvelets described in Sect. 14.5.5 are
particular examples. These curvelet molecules are almost orthogonal and provide
optimally sparse representations of large classes of linear systems of hyperbolic
differential equations.
Now, shearlets are an improvement over curvelets, in the sense that they are more
flexible and grounded in the general group-theoretical CS formalism, including
the coorbit method. Thus it is not surprising that exactly the same situation as
above prevails. Indeed, here too, it is useful to consider a more general type of
waveform, called a shearlet molecule [356]. It is shown in that work that the matrix
representation of a Fourier Integral Operator with respect to a Parseval frame of
shearlets is optimally sparse and well-organized.
Actually, one can go further. Indeed, curvelets and shearlets are both based on
parabolic scaling width length2 and the developments into the corresponding
molecule systems are essentially parallel (the main difference is geometrical:
curvelets are described in a polar geometry, shearlets in a Cartesian one, in
fact a pseudo-polar one (also used for a frequency based implementation, see
Sect. 14.5.5). They both provide optimally sparse representations of functions
exhibiting singularities on lower dimensional submanifolds. Thus one can expect
the existence of a more general framework, encompassing curvelets and shearlets
and other related systems. This is precisely what has been achieved by Grohs and
Kutyniok [346], under the name of parabolic molecules. Once again one recovers
the almost orthogonality of these molecules, which then yields, at least for certain
values of the parameters, the required sparsity of representations. In addition, they
lead to new smoothness function spaces. This is probably the final touch in the
domain of sparse representations.
Chapter 15
Wavelets on Manifolds

Abstract In this chapter, we discuss the construction of wavelets related to other

groups than similitude groups. The first, and most important, case is that of wavelets
on the two-sphere S2 . We start with the continuous approach, based on the use of
stereographic dilations, i.e., dilations obtained by lifting to S2 ordinary dilations on
a tangent plane by an inverse stereographic projection. Next we describe briefly a
number of techniques for obtaining discrete wavelets on S2 . Then we extend the
analysis to wavelets on other manifolds, such as conic sections, a torus, general
surfaces of revolution or graphs.

15.1 Wavelets on the Two-Sphere

There are several applications where data to be analyzed are defined on a sphere. For
instance, in Earth and Space sciences (geography, geodesy, meteorology, astronomy,
cosmology, etc), in crystallography (texture analysis of crystals), in medicine (some
organs are regarded as sphere-like surfaces), or in computer graphics (modelling of
closed surfaces as the graph of a function defined on the sphere), but also in statistics
and other instances (then spheres of dimension higher than 2 might occur). If one
is interested only in very local features, one may ignore the curvature and work on
the tangent plane. But when global aspects become important (description of plate
tectonics on the Earth or analysis of the cosmic microwave background (CMB),
for instance), one needs a genuine generalization of wavelet analysis to the sphere.
Several authors have studied this problem, with various techniques, mostly discrete.
However, to preserve the rotational invariance of the sphere, a continuous approach
is clearly necessary.
Of course, Fourier analysis on the sphere S2 is standard, but cumbersome, since it
amounts to work with expansions in spherical harmonics. While the latter constitute
an orthonormal basis of L2 (S2 ), they are not localized at all on the sphere, so that
Fourier analysis is global. Actually, there are specific combinations of spherical
harmonics which are well localized, the so-called spherical harmonics kernels

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 457
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__15,
Springer Science+Business Media New York 2014
458 15 Wavelets on Manifolds

[510], but then one loses the simplicity of an orthonormal basis. Hence, alternative
solutions have been proposed by several authors. We may quote, for instance,
Gabor analysis on the tangent bundle [588]; frequential wavelets, based on spherical
harmonics [286]; or diffusion methods with a heat equation [171]. Also several
discrete methods have been proposed, but we will come back to these later on.
In any case, various problems plague those constructions, such as an inadequate
notion of dilation, the lack of wavelet localization, the excessive rigidity of the
wavelets obtained, the lack of directionality, etc. In this respect, the continuous
wavelet transform (CWT) has many advantages: locality is controlled by dilation,
the wavelets are easily transported around the sphere by rotations from SO(3),
efficient algorithms are available. A first solution has been proposed in [381], with
several ad hoc assumptions, but no geometrical feeling. In particular, it contains a
parameter that has to be interpreted as a dilation parameter, but whose geometrical
meaning is unclear. It turns out that the general formalism developed in this book
yields an elegant solution to the problem [6264, 80, 81]. Among other things, it
allows one to derive all the assumptions of [381]. Furthermore, this method may
be discretized and yields spherical tight frames [154]. In the next section, we will
sketch this spherical CWT method, following mostly the review paper [64]. Later
on, we will turn to the discrete methods.

15.1.1 Stereographic Wavelets on the Two-Sphere

15.1.1.1 The Stereographic CWT on the Two-Sphere

We start with the two-sphere S2 , and consider signals in L2 (S2 , d ), where

d = sin d d denotes the usual Lebesgue measure on the sphere. The natural
operations on such signals are motions on the sphere (translations) and local
dilations. The former are given by rotations from SO(3). As for the latter, one uses
a (radial) stereographic dilation on S2 , which is obtained in three steps (Fig. 15.1):
(i) given a point A S2 , different from the South Pole S, project it stereographically
from the South Pole (which corresponds to the point at infinity in the plane) to the
point B in the plane tangent to the sphere at the North Pole N; (ii) dilate B radially in
the usual way to B; and (iii) project back B to the sphere, which yields A. The map
A  A is the required spherical dilation around N. Thus a dilation by a becomes a
nonlinear map  a acting on the colatitude angle:

a
tan = a tan . (15.1)
2 2

As for dilations around any other point C S2 , it suffices to bring C to the North
Pole by a rotation SO(3), perform the dilation and go back by the inverse
rotation 1 .
15.1 Wavelets on the Two-Sphere 459

N d B ad B
A

A
a

Fig. 15.1 Visual meaning of the stereographic dilation on S2

Obviously translations and dilations do not commute. However, the only group
combining only SO(3) and the dilation group A R+ is their direct product.
Indeed, SO(3) has no outer automorphisms, which prevents the construction of a
nontrivial semidirect product with R+ R. A solution to this difficulty is to embed
the two groups into the Lorentz group SOo (3, 1), using the Iwasawa decomposition
SOo (3, 1) = SO(3) A N, where A R and N is a one-dimensional abelian group,
isomorphic to the complex plane C (see Sect. 4.5.2). The justification of this is that
the Lorentz group is the conformal group for the sphere S2 as well as for the tangent
plane, and both rotations and dilations are conformal transformations [63]. Thus
the parameter space of the CWT is X = SOo (3, 1)/N SO(3) R+ and a natural
(Iwasawa) section is I ( , a) = a 1. Furthermore, S2 = SOo (3, 1)/P, where
P = SO(2) A N, the minimal parabolic subgroup [Kna96, Lip74], is the isotropy
subgroup of the South Pole and SO(2) consists of rotations around the z-axis. Hence
SOo (3, 1) acts transitively on S2 .
In order to compute this action explicitly, two methods are available. The
first one, which extends in a straightforward way to the same problem in higher
dimensions [62], is to exploit the Iwasawa decomposition. The other one, specific to
the dimension 2, is to go to SL(2, C), the double covering of SOo (3, 1), and to use
the Gauss decomposition SL(2, C) = Z+ B , where Z+ C (see Sect. 4.5.2). Thus
S2 is now identified with the Riemann sphere and C = SL(2, C)/B . The restriction
to S2 of the natural projection is the stereographic projection p from S2 \ {S} onto
the tangent plane at the North Pole (hence it is bijective). In this way, the action
z  ( z + )( z + )1 of SL(2, C) on C is lifted to S2 by p1 , or, more properly,
that of SL(2, C)/Z2 SOo (3, 1), which is simply transitive. This shows that the
natural group to use is indeed the Lorentz group. For dilations, in particular, one
recovers (15.1), by both methods. Clearly, we are in the general situation described
460 15 Wavelets on Manifolds

in Sect. 7.3. Since N corresponds to translations in the plane, P cannot leave any
function invariant, unless it is constant. Thus the CS we are going to define are not
of the GilmorePerelomov type.
The next step is to find an appropriate UIR of SOo (3, 1) in the space of signals
L2 (S2 , d ). We choose the following principal series representation [576, Kna96,
Lip74], which is induced by the trivial character of A N:

[U(g) f ]( ) = (a, )1/2 f (g1 ), g = a n, = ( , ). (15.2)

In this relation, (a, ) is the correcting factor (RadonNikodym derivative) that

takes into account the fact that the measure d on S2 is not invariant under dilations.
For = ( , ), one gets simply

(a, ) = 4a2 [(a2 1) cos + (a2 + 1)]2 . (15.3)

Notice that this representation is infinite dimensional. Its restriction to SO(3) is the
quasi-regular representation

[Uqr ( ) f ]( ) = f ( 1 ), SO(3), (15.4)

which decomposes into the direct sum of all the familiar (2l + 1)-dimensional
representations, l = 0, 1, . . .. The key point is that U is square integrable mod (N, I ):
Proposition 15.1.1. The representation U given in (15.2) is square integrable
modulo the subgroup N and the Iwasawa section I . A nonzero vector
L2 (S2 , d ) is admissible mod(N, I ) if and only if there exists c > 0, independent of
l, such that:

8 2 l
da
s (l) :=
2l + 1 m=l 0
|Ylm | a |2
a3
< c, (15.5)

where Ylm denotes the usual spherical harmonic and a = U(I (e, a)) corresponds
to a pure dilation.
The proof, as usual, consists in an explicit calculation, using the properties of
a (l, m) is a Fourier coefficient).
Fourier analysis on the sphere (Ylm | a  =:
Thus, when is admissible, the family {I (x) , x X} is a continuous family of
CS. But in fact, we have more :
>
Proposition 15.1.2. For any admissible vector such that 02 ( , ) d = 0,
the family {I (x) , x X} is a continuous frame, that is, there exist constants m > 0
and M < such that

m f 
2
|I (x) | f |2 d (x) M  f 2 , f L2 (S2 , d ), (15.6)
X
15.1 Wavelets on the Two-Sphere 461

where d (x) = a3 da d , with d the Haar measure on SO(3). The frame operator
S is diagonal in Fourier space (i.e., it is a Fourier multiplier):

S?f (l, m) = s (l) f(l, m). (15.7)

This result applies, in particular, if is axisymmetric or zonal, that is, a function of

alone. The relation (15.6) means that the frame operator S and its inverse S1 are
both bounded. One notices, however, that the upper frame bound, which is implied
by the constant c, does depend on , whereas the lower frame bound does not, since
it follows from the asymptotic behavior of the function Ylm for large l.
These two propositions yield the basic ingredient for writing the CWT on S2 . The
wavelets on the sphere are the functions ,a = U(I ( , a)) , with admissible.
Then the CWT reads, with US ( , a) = U(I ( , a)):

F ( , a) = U(I ( , a)) | f 

= [US ( , a) ]( ) f ( ) d ( ) (15.8)
S2

= a ( 1 ) f ( ) d ( ).
S2

Notice that, if the wavelet is axisymmetric, the dependance on SO(2) drops

out, so that F ( , a) reduces to F ( , a), S2 .
The corresponding reconstruction formula is
 
da
f ( ) = d F ( , a) [S1 ,a ]( ) , f L2 (S2 , d ). (15.9)
0 a3 SO(3)

Correspondingly, instead of the familiar isometry property, one gets a Plancherel

relation:
 
da
 f 2 = d F ( , a) F ( , a), (15.10)
0 a3 SO(3)

where

F ( , a) := 
a, | f  = S1 a, | f . (15.11)

The new fact here is the occurrence of the inverse frame operator S1 in these
formulas. This results from the square integrability of the representation (15.2) over
the quotient space X, instead of the group itself.
Now, it was shown in [603] that the reconstruction formula (15.9) is valid
under the weaker condition 0 < s (l) < , l N. Since the behavior of s (l) is
arbitrary, this means exactly that the frame operator S is allowed to be unbounded.
The lower frame bound, being independent of , remains untouched, so that S1
462 15 Wavelets on Manifolds

stays bounded. In other words, if the vector is not admissible, in the sense of
Proposition 15.1.2, the family ,a is a lower semi-frame, as defined in (3.37).
This spherical CWT has all the desired properties, except covariance. Indeed, it
is easily shown by a direct calculation [63] that:
The spherical CWT (15.8) is covariant under motions on S2 : for any o SO(3),
the transform of the rotated signal f (o1 ) is the function F (o1 , a).
But it is not covariant under dilations. Indeed the wavelet transform of the dilated
signal (ao , )1/2 f (a1 1
o ) is U(g) | f , with g = ao a, and the latter, while
a well-defined element of SOo (3, 1), is not of the form I (  , a ).
This negative result of course reflects the fact that the parameter space X of the
spherical CWT is not a group, but only a homogeneous space.
Now the full admissibility condition (15.5) of a wavelet is somewhat com-
plicated to use in practice, since it requires the evaluation of nontrivial Fourier
coefficients. There is, however, a simpler, although only necessary, condition,
namely,

( , )
d ( , ) = 0 (15.12)
1 + cos

(hence must vanish sufficiently fast when one approaches the South Pole). This is
a zero mean condition, so that we have the filtering effect, as usual. Thus a genuine
CWT on the sphere has been obtained. One should notice that the poles do not play
any particular rle in this CWT, since the sphere S2 is a homogeneous space under
SO(3): all the points of S2 are really equivalent.
Typical wavelets satisfying the necessary condition (15.12) are difference
wavelets, as in the flat case (14.25):

( ) 1
= US (e, ) , > 1, (15.13)

where is a smoothing function and US (e, ) is again a pure dilation, namely,

 
US (e, ) ( , ) = ( , )1/2 (1/ , ).

Such a wavelet is fully admissible if is sufficiently regular at the poles. The

( )
simplest one, denoted by G , corresponds to G ( , ) = exp( tan2 2 ) and is
the spherical equivalent of the DOG. This is an axisymmetric wavelet, shown in
Fig. 15.2, which yields an efficient detection of discontinuities on the sphere .
An additional bonus is that this CWT on the sphere has the expected Euclidean
limit. By this we mean the following. Consider instead of the unit sphere S2 a
sphere S2R of radius R and let R . In technical terms, one performs a group
contraction, keeping fixed the subgroup SO(2) of rotations around the z-axis, using
the technique described in Sect. 4.5.4, with = R1 . Then S2R becomes the plane R2 ,
the group SO(3) becomes the Euclidean group of R2 and I (SO(3) R) becomes
15.1 Wavelets on the Two-Sphere 463

a b c

1
1
1

0.5 0.5
0.5
N
0 N 0
N 0

-0.5 -0.5
-0.5

-1 -1
-1
-1 -0.5 -1
-0.5 -1 -0.5 -0.5
-0.5 0
0 -0.5 0
0 0
0.5 0.5 0 0.5
0.5 0.5
1 0.5 1
Y X X X
Y Y

( )
Fig. 15.2 The spherical wavelet G wavelet, for = 1.25: (a) Original (a = 0.125); (b) Rotated;
(c) Rotated and scaled (a = 0.0625)

SIM(2), with corresponding transitive action. Furthermore, as R , the family

of representations UR USR contracts into the natural representation U of (14.4) of
SIM(2), as a strong limit on a dense set. The representation UR is realized in the
Hilbert space HR = L2 (S2R , R2 d ) and U in H = L2 (R2 , dx). For each R, we choose
DR D := C0 (R2 ), the space of continuous functions of compact support, which is
dense in H. The family of injections IR : HR H is defined as

4R2 ! r "
(IR f ) (r, ) = f 2 arctan , , f HR , (15.14)
4R2 + r2 2R

where we have used polar coordinates (r, ) in the plane. It is easily shown that
IR is in fact unitary from HR onto H. Then one shows that, for every F D and
g SIM(2), one has

lim IRUR (R (g)) IR1 F U(g)FH = 0, (15.15)

R

where R is the contraction map defined in (4.158). This means precisely that the
representation U of SIM(2) is a contraction of the family of representations UR of
SOo (3, 1) as R (see [63] for the technical details).
Finally, admissible vectors tend exactly to admissible vectors (also the necessary
admissibility condition (15.12) goes into the corresponding one (14.10) in the
plane). For instance the spherical DOG wavelet tends to the usual flat space DOG.
Therefore, the wavelet analysis on S2 goes into the usual wavelet analysis in the
plane.
In addition, one can show that the stereographic dilation is uniquely determined
if one requires that the map between S2 and its tangent plane be a conformal
diffeomorphism. As a result, one obtains uniquely the spherical CWT from the plane
(Euclidean) one, simply by lifting everything from the tangent plane to the sphere
by inverse stereographic projection, the wavelets, the admissibility conditions,
the directionality or steerability properties [603].
464 15 Wavelets on Manifolds

Thus one obtains an equivalence between the two wavelet formalisms. Let :
L2 (S2 , d ) L2 (R2 , dx) be the unitary map induced by the stereographic projec-
tion :
1
[ f ](x) = f (p1 (x)), f L2 (S2 , d ), (15.16)
1 + (r/2)2

with inverse
2
[ 1 F]( , ) = F(p( , )), F L2 (R2 , dx). (15.17)
1 + cos

Then every admissible Euclidean wavelet L2 (R2 , dx) yields an admissible

spherical wavelet 1 L2 (S2 , d ). In particular, if is a directional wavelet,
so is 1 . As an example, Fig. 15.3 presents the spherical Morlet wavelet,
obtained by lifting the usual 2-D Morlet wavelet onto the sphere. In order to
test the capabilities of the various wavelets, we show on Fig. 15.4 the analysis of
the characteristic function of a triangle with apex at the North Pole, 0
( )
50 , 0 90 . In Panel (a), the wavelet is the spherical DOG wavelet G
given in (15.13), for = 1.25 and scale 0.035. In (b) and (c), on the other hand,
one uses a spherical Morlet wavelet, oriented in two ways, = 0 and = 90 (
is the third Euler angle, which describes a rotation of the wavelet around its center).
As expected, this wavelet filters out the directions perpendicular to its orientation,
keeping the great circles = const. in the first case and the longitude circles =
const. in the second case.
In conclusion, the formulas (15.8) yield a genuine CWT on the sphere, entirely
derived from group theory, following the formalism of general coherent states
developed in this book. In addition, the Euclidean limit is valid, with a precise
group-theoretical formulation. Concrete tests with the spherical DOG wavelet and
the spherical Morlet wavelet show that it has the expected capability of detecting
discontinuities, whether or not they lie at one of the poles of the sphere [64].
Speaking about the sphere, it may be useful to emphasize that the wavelets
described here have nothing to do with the coherent states associated to SU(2), the
spin CS, whose parameter space is the Bloch sphere S2 , that we have described in
Sect. 7.2.1. Here the sphere is the manifold on which wavelets and signals live,
whereas the parameter space is the noncompact manifold X = SOo (3, 1)/N
SO(3) R.

15.1.1.2 Stereographic Wavelet Frames on the Two-Sphere

The only remaining problem is of a computational nature. Indeed the formula (15.8)
requires a pointwise convolution on the sphere, which is very time-consuming.
However, this is not specific to wavelet analysis, it simply reflects the lack of an
efficient convolution algorithm on the sphere, and in particular the difficulty of
finding an appropriate discretization of the latter.
15.1 Wavelets on the Two-Sphere 465

Fig. 15.3 The spherical Morlet wavelet at two scales, (a) a = 0.3 and (b) a = 0.03. Then displaced:
(c) a = 0.03, centered at ( /3, /3); and (d) The same, rotated by /2

Fig. 15.4 Analysis of a triangle with (a) the spherical DOG wavelet for = 1.25, at scale 0.035;
(b) and (c) the spherical Morlet wavelet, in two different orientations: = 0 , resp. = 90 ,
showing the directional selectivity of the wavelet
466 15 Wavelets on Manifolds

Actually, the spherical CWT can be discretized, but this requires to extend the
notion of frames to weighted frames and controlled frames, discussed in Sect. 3.4.2.
In a first step, we will build a half-continuous spherical frame, by discretizing the
scale variable only, while keeping continuous the position variable on the sphere
(this is exactly the approach adopted in Sect. 13.2 for designing the so-called
continuous wavelet packets).
Let us choose the half-continuous grid = {( , a j ) : S2 , j Z, a j > a j+1 },
where A = {a j : j Z} is an arbitrary decreasing sequence of scales, and j are
weights that mimic the natural (Haar) measure da/a3 .
If we start from the standard weighted frame condition given in (3.54), we do
obtain a weighted frame, but there is no way of getting a tight one. The reason
is obvious, the frame operator S has not been taken into account. We start instead
from the Plancherel formula (15.10) and write a modified frame condition (for an
axisymmetric wavelet)

m  f 2 j
S2
F ( , a j ) F ( , a j ) d (  ) M  f 2 . (15.18)
jZ

In that case, a tight frame might be obtained. Indeed the following proposition holds
true [154].
Proposition 15.1.3. Let A = {a j : j Z} be a decreasing sequence of scales. If
is an axisymmetric wavelet for which there exist two constants m > 0, M <
such that

m g (l) M, for all l N,

where
4
g (l) =
2l + 1 j |a j (l, 0)|2 ,
jZ

then any function f L2 (S2 , d ) may be reconstructed from the corresponding

family of spherical wavelets, as
 2 3 
f ( ) = j
S2
F (  , a j ) 1 
,a j ( ) d ( ), (15.19)
jZ

where  is the operator defined in Fourier space by

(
1 1
h(l, m) = g (l) h(l, m).

Note that the operator  is simply the discretized version of the continuous frame
operator S. Clearly (15.19) may be interpreted as a (weighted) tight frame controlled
by the operator 1
.
15.1 Wavelets on the Two-Sphere 467

One can also design a fully discrete spherical frame, by discretizing all the
variables. The scale variable is discretized as before: a A = {a j R+ : a j >
a j+1 , j Z}. As for the positions, we choose an equiangular grid G j indexed by the
scale level:
(2p+1) q
G j = { jpq = ( jp , jq ) S2 : jp = 4B j , jq = B j }, (15.20)

for p, q N j := {n N : n < 2B j } and some range of bandwidths B = {B j

2N : j Z}. Thus the complete discretization grid is (A , B) = {(a j , jpq ) : j
Z, p, q N j }.
Note that, in (15.20), the values { jp } constitute a pseudo-spectral grid, with
nodes related to the zeros of a Chebyshev polynomial of degree 2B j :

T2B j (cos jp ) = cos(2B j jp ) = 0.

For these values, there exists an exact quadrature rule for certain (explicit) weights
w jp > 0 and for every function f L2 (S2 , d ) of bandwidth B j (i.e., f(l, m) = 0 for
all l B j ), namely [263],


S2
f ( ) d ( ) = w jp f ( jpq ),
p,qN j

Then one gets a discrete weighted, nontight, frame controlled by the operator S1
[154] . Namely, for every function f L2 (S2 , d ), one has:

m  f 2 j w jp F ( jpq , a j ) F ( jpq , a j ) M  f 2 . (15.21)

jZ p,qN j

A sufficient condition for (15.21) may be given, but it is very complicated, involving
the determinant of an infinite dimensional matrix, unless f is band-limited.
There is an alternative that also leads to a half-continuous wavelet represen-
tation on S2 . It consists in using the so-called harmonic dilation instead of the
stereographic one. This dilation acts on the Fourier coefficients of a function f ,
that is, the numbers f,m := Ym | f S2 , where {Ym ,  N, m = , . . . , } is the
orthonormal basis of spherical harmonics in L2 (S2 ). The dilation da is defined by
the relation


(da f ),m := f a,m , a > 0. (15.22)

This technique, originally due to Holschneider [381] and FreedenWindheuser

[285], has recently been revived in the applications to astrophysics [605]. However,
although this definition leads to a well-defined, uniquely invertible wavelet repre-
sentation, with steerable wavelets and full rotation invariance, there is no proof so
far that it yields a frame. Hence one may question the stability of the reconstruction
process, since it is the lower frame bound that guarantees it.
468 15 Wavelets on Manifolds

15.1.1.3 Wavelet Approximations on the Sphere

The central theme of approximation theory is the representation of a function by a

truncated series expansion into a family of basis functions, for instance, the elements
of a frame. In the flat case, 1-D or 2-D, wavelets are widely used for approximation
in various function spaces [Mal99]. The crucial advantage is their multiresolution
character, which is optimally adapted to local perturbations. A natural framework
is given by the Lebesgue spaces L p (Rn ), 1 p < . One of the reasons is that
approximation is often formulated in terms of convolution with an approximate
identity, and many useful convolution identities are available in L p [Gaa73, Lie97].
Thus, if the function lives on the two-sphere, it is necessary to have a good notion
of convolution on S2 . For that purpose, it is useful to represent the sphere as the
quotient SO(3)/SO(2), since the convolution machinery extends almost verbatim
to locally compact groups, and then partly to homogeneous spaces [80, App. A].
The most useful identity, which has independent interest, applies to a zonal or
axisymmetric function.
Proposition 15.1.4. Let f and g be two measurable functions on S2 . If f is zonal,
the spherical convolution of f and g is a function on S2 , which can be written:

( f  g)(  ) = f ( ) g( ) d ( ) , (15.23)
S2

where   is the R3 scalar product of unit vectors of directions  and .

A proof of this identity may be found in [80], that we follow here.
In approximation theory, the technique used in the Euclidean case [Lie97, Ste71]
consists in performing a convolution with a smoothing kernel, that acts as an
approximate identity. The same approach was used for the sphere in [Fre97],
where various spherical kernels are constructed. As usual, the difficulty lies in the
definition of dilation on S2 . The problem gets a unique solution, however, if one
uses the spherical dilation (15.1).
Proceeding in that way, the main result of [80] is that the spherical CWT admits
a reconstruction formula, valid in the strong L2 topology, exactly as the usual CWT
in Rn (actually, the formula holds in any strong L p topology, for 1 p < ). As
in the flat case, one may choose between a bilinear and a linear formalism [268,
477, Tor95]. But there is a crucial difference. In the flat case, it is advantageous,
but not compulsory, to treat the large scales or low frequencies separately, in terms
of a scaling function. This was done in the context of the so-called infinitesimal
multiresolution analysis described in Sect. 13.2. Here, however, one is forced to
do it. The reason is that, geometrically, only small scales are relevant and lead to the
expected filtering behavior. One chooses arbitrarily a = ao as reference scale and
define the scales a > ao as large (for instance, put ao = 1).
To give the flavor of the result, let us quote the bilinear formulation from
[80, Theorem 4.7]. Note that one uses here the so-called L1 normalization, which
implies a suitable modification in the dilation operator, thus also in the admissibility
condition and in the CWT itself.
15.1 Wavelets on the Two-Sphere 469

Given a wavelet L1 (S2 ), we define the corresponding scaling function (ao )

by its Fourier coefficients :

 ?a (l, m)|2 da
| (ao ) (l, m)|2 = | , l 1, (15.24)
ao a
 1
| (ao ) (0, 0)|2 = . (15.25)
8 2

 a , where D
In (15.24), a := D  a is the modified dilation operator, and the integral
converges in virtue of the (modified) admissibility condition satisfied by . Of
course, (15.24) does not define the function uniquely. One can, for instance,
assume in addition that (ao ) (l, m) 0, l, m, as in [Fre97].

We define the large scale part of a signal s as


(ao ) (ao )
 ( , ao ) = ( ) s( ) d ( ), where ( ) := (ao ) ( 1 ). (15.26)
S2

Using these notations, the main result reads as follows [80]:

Theorem 15.1.5. Let L1 (S2 ) be a wavelet and let (ao ) , ao > 0, denote the
associated scaling function. Assume the following two conditions are satisfied:
for all l = 1, 2, . . .,

8 2 da

2l + 1 |m|l 0
?a (l, m)|2
|
a
= 1, (15.27)

for all (0, ao ), there is a constant M > 0, independent of , such that

 ao
da
 a 2 M. (15.28)
a

Then, for all s L2 (S2 ), we have the equality

 ao  
da (a )
s= S ( , a)a d +  ( , ao ) o d , (15.29)
0 a SO(3) SO(3)

where S is the spherical CWT of s with respect to the wavelet ,  is the large
scale part of s and the integral is understood in the strong sense in every L p (S2 ), 1
p < .
In the case of a zonal wavelet , the parameter space of the CWT reduces to S2 R+
instead of SO(3) R+ .
The necessity of treating separately the large scale part of a signal is made explicit
by considering a constant function, for instance, the unit function. Indeed constant
functions on the sphere are square integrable, and hence must be reconstructible,
470 15 Wavelets on Manifolds

but their CWT vanishes identically in the L1 normalization, so that only the second
term in (15.29) survives.
For a proof of Theorem 15.1.5 and for further details on the linear reconstruction
formula, we refer to the original paper [80].

15.1.2 Poisson Wavelets on the Sphere

Motivated by various applications in geophysics, Holschneider and his collaborators

have introduced a different kind of wavelets on the sphere, called Poisson wavelets
(see [382, 390] for a review and references to the original papers). We briefly sketch
their method.
There are two ways of proceeding. On one hand, one can define internal wavelets
and external wavelets, that is, wavelets supported inside, resp. outside the unit ball
B3 of R3 . In both cases, these are harmonic functions on Int B3 , resp. Ext B3 .
Restricting both of them to the sphere S2 = Int B3 Ext B3 , one gets spherical
wavelets and the two previous types are simply the harmonic continuation of the
latter into the respective domains. These internal wavelets solve the problem of
wavelet analysis on the solid ball B3 in a simpler way that those constructed in
[439].
Alternatively, one may define Poisson wavelets directly on S2 , as follows. Given
a > 0, S2 and d 0 , put

d ,a ( ) = (al)d eal Ql (cos ), (15.30)
l=1

where Ql = (2l + 1)Pl is the Legendre polynomial of degree l is and = ( , ).

Thus we have here again a harmonic dilation, as defined at the end of Sect. 15.1.1.2.
One should notice the analogy of this wavelet with the CauchyPaul wavelet on the
?d (a ) = (a )d ea .
line (12.20), which can be written, at scale a, in Fourier space,
Thus the parameter space of the Poisson wavelets is ( , a) S2 R+ . They have
properties very similar to those of the stereographic wavelets. For instance, they
have also the correct Euclidean limit.
Using these wavelets, one defines the Poisson WT as (Wd f )( , a) := d ,a | f ,
for f L2 (S2 , d ) - C (indeed, one has to remove the constant functions). Then the
map f  Wd f is an isometry
  
1 da
| f ( )|2 d ( ) = |(Wd f )( , a)|2 d ( ) ,
S2 (2d) S2 0 a

and the range of Wd is a Hilbert space with reproducing kernel

1
K( , a; , b) =  d | d .
(2d) ,a ,b
15.1 Wavelets on the Two-Sphere 471

Moreover, upon choosing a sufficiently dense grid in B3 , one obtains a tight frame.
Here, one identifies a point x in the interior of B3 with the point ( , a) S2 R+
by the map

x = ea , a = | log(|x|)|.

More precisely, exactly as for the stereographic wavelets, one can obtain both
semi-discrete frames (only the scale parameter is discretized) and fully discrete
ones.

15.1.3 Discrete Spherical Wavelets

Many authors have designed methods for constructing discrete spherical wavelets.
All of them have advantages and drawbacks. These may be characterized in terms
of several properties which are desirable for any efficient wavelet analysis, planar
or spherical (a thorough discussion of this topic may be found in [60]).
Basis: The redundancy of frames leads to nonunique expansions. Moreover, the
existing constructions of spherical frames are computationally heavy and often
applicable only to band-limited functions, as we have seen in Sect. 15.1.1.2. Thus
genuine bases are in general preferable, although difficult to obtain.
Orthogonality: This method leads to orthogonal reconstruction matrices, whose
inversion is trivial. Thus, orthogonal bases are good for compression, but this
is not always sufficient: sparsity of reconstruction matrices is still needed in the
case of large data sets.
Local support: This is crucial when working with large data sets, since it
yields sparse matrices in the implementation of the algorithms. Also, it prevents
spreading of tails during approximation. Note that we make here a distinction.
A wavelet has local support if it vanishes identically outside a small region.
It is localized if it is negligible outside a small region, so that it may have
(small, but nonzero) tails there. Since these tails may spread in the process of
approximation of data and spoil their good localization properties, local support
is definitely preferred (see the example in [535]).
Continuity, smoothness: These properties are always desirable in approximation,
but not easily achieved.

15.1.3.1 Spherical DWT from Multiresolution Analysis

As expected, many authors have tried to design discrete wavelets on the sphere, with
the aim of generating (bi)orthogonal bases. An early proposal, based on second-
generation wavelets [551] provided an efficient solution, altogether avoiding Fourier
472 15 Wavelets on Manifolds

transforms (i.e., spherical harmonics) and multiresolution analysis (MRA). However

this method resulted in a partition of S2 (by successive subdivision of triangles) into
subsets of inequal area.
Further, and numerous, attempts follow the standard procedure of a DWT based
on a MRA. However, the concept has to be adapted to the compactness of the sphere:
one can always refine a partition, introducing finer and finer details, but not coarsen
it, omitting more and more details. Therefore a MRA of L2 (S2 ) is defined as an
increasing sequence of closed subspaces

V0 V1 V2 . . . L2 (S2 ), (15.31)

  
 dense in L2 S2 , and for which there exists index sets K6j K6j+1 such
with j=0 V j / 0
that, for every j 0, one has a Riesz basis vj , v K6j of Vj . More precisely,
there exist constants 0 < A B < , independent of the level j, such that
4 5    4 5 
     
A 2 j  cvj vK6 2
j l (K j )
 cvj vj 
L2 (S2 )
B 2 j  cvj vK  2 !
j l K6j
".
vK6j

One does not require vj to be translations/dilations of the same function , since

this condition would not be easy to satisfy for spherical wavelet bases or frames.
Next one defines the wavelet spaces W6j as W6j = Vj+1 - Vj and then one constructs
6j .
a basis or a frame in each W
Let us quote a few of those methods, with focus on the properties just mentioned,
without being exhaustive. A more comprehensive review, with all references to
original papers, may be found in [60].
1. The spherical DWT using spherical harmonics
Various constructions of discrete spherical wavelets using spherical harmonics
may be found in the literature, leading to frames or bases. The advantages of
this method is that it produces no distortion (since no pole has a privileged
role) and that it preserves smoothness of the wavelets. However, the wavelets
so obtained are in general localized, but not locally supported, i.e., the support
covers the whole sphere. Since this implies full reconstruction matrices, the result
is not suitable for large amount of data. An example is [510], where spherical
harmonics kernels are used for localizing the support of the scaling functions and of
the wavelets. However, they obtain only frames. Bases have been constructed in
[284287], but their wavelets are defined as infinite convolutions of kernels of
spherical harmonics.
2. The spherical DWT via polar coordinates
The polar coordinate map : I = [0, ] [0, 2 ) S2 has the familiar form

: ( , )  (cos sin , sin sin , cos ) .

15.1 Wavelets on the Two-Sphere 473

A problem here is continuity. Indeed a continuous function f defined on I remains

continuous after mapping it onto S2 if and only if f ( , 0) = f ( , 2 ), for all
[0, ], and there exists two constants PN , PS such that f (0, ) = PN and f ( , ) =
PS , for all [0, 2 ). Unfortunately, these continuity conditions are not easily
satisfied by wavelets on intervals. The obvious advantage of this approach is that
many data sets are given in polar coordinates and thus one need not perform
additional interpolation when implementing. However, there are disadvantages.
First, no known construction gives both continuity and local support. Next, there
are distortions around the poles: maps the whole segment {(0, ), [0, 2 )}
onto the North Pole, and the whole segment {( , ), [0, 2 )} onto the South
Pole. Representative examples are [219], who constructs C0 wavelets, or [601], who
gets C1 wavelets.
3. The spherical DWT via radial projection from a convex polyhedron
Let S2 be the unit sphere centered in 0 and let be a convex polyhedron,
containing 0 in its interior and with triangular faces (if some faces are non-triangular,
one simply triangularizes them). The idea of the method is to obtain wavelets on S2
first by moving planar wavelets to wavelets defined on the faces of and then
projecting these radially onto S2 [526, 528, 529]. This proceeds as follows. Let
= denote the boundary of and let p : S2 denote the radial projection
from the origin:

p(x, y, z) = (x2 + y2 + z2 )1/2 (x, y, z)..

Let T denote the set of triangular faces of and consider the following weighted
scalar product on L2 (S2 ):

F | G = p(T )
F( ) G( ) wT ( ) d ( ),
T T

= (1 , 2 , 3 ) S2 , F, G L2 (S2 ). (15.32)

Here wT (1 , 2 , 3 ) = 2 dT2 |aT 1 + bT 2 + cT 3 |3 , with aT , bT , cT , dT the

coefficients of x, y, z, 1, respectively, in the determinant

x y z 1
x1 y1 z1 1
= aT x + bT y + cT z + dT 1,
x2 y2 z2 1
x3 y3 z3 1

where (xi , yi , zi ), i = 1, 2, 3, are the vertices of the planar triangle T T . Then one
1/2
proves that the norm  :=  |  is equivalent to the usual norm in L2 (S2 ), i.e.,
there exist constants m > 0, M < such that
474 15 Wavelets on Manifolds

m  f   f 2 M  f  , f L2 (S2 ).

Explicit expressions for optimal bounds m and M are given in [531].

Using this technique, one may obtain the following results:
1. Piecewise constant orthogonal wavelet bases on the faces of yield piecewise
constant wavelet bases on spherical triangulations [525]. Here the orthogonality
holds with respect to the scalar product ,  , not the usual scalar product in
L2 (S2 ), but the strength of the construction is that the scaling matrices M j are
orthogonal and sparse.
2. Locally supported piecewise linear wavelet bases on triangulations of R2 yield
piecewise rational semi-orthogonal wavelet bases on S2 , having local support
and satisfying the Riesz stability property, hence continuous [526]. Therefore the
matrix M j is sparse, but not orthogonal. Since there are fast algorithms adapted
to sparse matrices, these wavelets can be used successfully in spherical data
compression.
3. Taking a cube for , one may obtain weighted Haar wavelets on S2 by
constructing first wavelets on the faces from wavelets on an interval by the
tensor product method described in Sect. 14.4, then applying the construction
just described [528].
This method offers many advantages: no distortion around the poles, possible
construction of continuous and locally supported stable wavelet bases, local support
of the wavelets (leading to sparse matrices), easy implementation, possible exten-
sion to sphere-like surfaces [527] (see Sect. 15.2). As a disadvantage, we may note
the lack of smoothness of the wavelets.
Actually, for certain particular polyhedrons, the radial projection may be replaced
by an area preserving projection (see Sect. 15.2.3.1), so that one does not need the
inner product  |  , only the usual one  | 2 . This is the case for a cube [537] or a
regular octahedron [538].
(4) Needlets
A new class of discrete spherical wavelets, called needlets, has been introduced
recently by Narcowich et al. [116, 480]. These functions, which are actually special
spherical harmonics kernels, are derived by combining three ideas, namely, a
LittlewoodPaley decomposition, a suitable distribution of (finitely many) points
on the sphere, called centers, and an exact quadrature rule. The dilation takes place
in the space of spherical harmonics, effectively in Fourier space, i.e., it is a harmonic
dilation as described at the end of Sect. 15.1.1.2. The upshot is a new class of
tight frames on the sphere. The frame functions are both compactly supported in
the frequency domain (i.e., band-limited in l) and almost exponentially localized
around each center, i.e., their amplitude decreases exponentially away from each
center. When combined with a new statistical method, they offer a powerful tool
for analysing CMB (WMAP) data, e.g. for analysing the cross-correlation between
the latter and galaxy counts from sky surveys [462, 506]. They have also found nice
applications in statistics [115, 116].
15.1 Wavelets on the Two-Sphere 475

More recently, a variant of the construction, dubbed Mexican needlets, has been
introduced by Geller and Mayeli [324,325]. Although the frames generated by these
objects are only approximately tight, they offer certain advantages, in particular,
there are extremely well localized in the real domain, i.e., on the sphere, their
tails decay as Gaussians. As such they are very well adapted to data with noise
or holes, as encountered in CMB maps. Indeed, these Mexican needlets have shown
a great potential for CMB analysis [553]. Actually, they are only a special case of
more general wavelets on compact manifolds, such as a sphere or a torus in any
dimension. However, the general construction is very far from our point of view, so
we shall only refer to the original papers [324, 325].
4. Ridgelets and curvelets on the sphere
For the sake of completion, we ought to mention here a different method
introduced in [568] for the purpose of astrophysical applications. The idea is to
take an axisymmetric, band-limited spherical wavelet (i.e., vanishing for l > lc ) and
applying to it a sort of harmonic dilation lc  alc . Then, using the HEALPix grid
[340], the authors design a different discrete spherical wavelet on S2 . Furthermore,
they extend their approach to ridgelets and curvelets on the sphere, following the
formulation of [566].
As a matter of fact, no construction described so far has led to wavelet bases
on the sphere which are simultaneously continuous (or smoother), orthogonal and
locally supported, although any two of these three conditions may be met at the
same time. This suggests to try another approach.

15.1.3.2 Lifting the DWT from the Plane to the Sphere

A possible method consists in lifting wavelets from the tangent plane to the sphere
by inverse stereographic projection [60]. It yields simultaneously smoothness,
orthogonality, local support, vanishing moments. The disadvantage is that it gives
distortions around a pole. In addition, it is not suitable for the whole sphere S2 , but
only for data away from that pole. However, the latter can be taken anywhere on
the sphere, for instance, in a region where no data is given. To give an example,
European climatologists routinely put the North Pole of their spherical grid in the
middle of the Pacific Ocean. Therefore, this is in fact a minor inconvenient in
practice.
Let again p : S2 R2 be the stereographic projection from the South Pole S
onto the tangent plane at the North Pole, where S2 = S2 \ {S} is the pointed sphere.
Notice that L2 (S2 ) = L2 (S2 ), since the set {S} is of measure zero.
As mentioned in Sect. 15.1.1.1, the stereographic projection p induces a unitary
map : L2 (S2 ) L2 (R2 ), given in (15.16), with inverse 1 : L2 (R2 ) L2 (S2 )
given in (15.17), that is, 1 (F) = (F p), F L2 (R2 ). Here = (J p)1/2 is
a real-valued function on S2 , where J : R2 R is the Jacobian of the transformation
S2  (x, y) R2 , that is, d ( ) = J(x, y) dx dy, where d ( ) is the area element
on S2 . As a consequence, we have
476 15 Wavelets on Manifolds

F | GL2 (R2 ) =  (F p) | (G p)L2 (S2 ) , F, G L2 (R2 ). (15.33)

This equality allows one to construct orthogonal bases on L2 (S2 ) starting from
orthogonal bases in L2 (R2 ). More precisely, we will use the fact that, if the
functions F, G L2 (R2 ) are orthogonal, then the functions F s := (F p) and
Gs := (G p) will be orthogonal in L2 (S2 ). Thus, the construction of MRA and
wavelet bases in L2 (S2 ) is based on the equality (15.33).
Now we can proceed and lift to the sphere the MRA in the plane, described
in Sect. 14.4. To every function F L2 (R2 ), one may associate the function F s =
(F p) L2 (S2 ). In particular, with the notation of (14.54),
s
Fj,k = (Fj,k p) for j Z, k Z2 , (15.34)

and similarly for the spherical functions sj,k and j,ks , where , , =
j,k j,k
h, v, d, are the planar 2-D scaling functions and wavelets, respectively. For j Z,
we define V j as V j := { (F p), F V j }. Then we have:
1. V j V j+1 for j Z, and each V j is a closed subspace of L2 (S2 );
 
2. jZ V j = {0} and jZ V j is dense in L2 (S2 );
3. {0,k
s , k Z2 } is an orthonormal basis for V .
0

A sequence (V j ) jZ of subspaces of L2 (S2 ) satisfying these three conditions

constitutes a MRA of L2 (S2 ). Notice that here j Z , since S2 is not compact, unlike
the MRA (15.31) mentioned above, defined for the compact sphere S2 . This is why
we use here a different notation as compared to (15.31), V j , W j instead of Vj , W6j .

Define now the wavelet spaces W j by V j+1 = V j W j . Then { j,k , k Z , =
s 2

h, v, d} is an orthonormal basis for W j and { j,ks , j Z, k Z2 , , = h, v, d} is

A 
jZ W j = L (S ). This the orthonormal wavelet basis
an orthonormal basis for 2 2

on S .
2

Thus, an orthonormal (resp. Riesz) 2-D wavelet basis yields an orthonormal

(resp. Riesz) spherical wavelet basis. In addition, if has compact support in R2 ,
then sj,k has local support on S2 (and diam supp sj,k 0 as j ), and similarly
for the respective wavelets. Smooth 2-D wavelets yield smooth spherical wavelets.
In particular, Daubechies wavelets yield locally supported and orthonormal wavelets
on S2 . Thus the same tools as in the planar 2-D case can be used for the
decomposition and reconstruction matrices (so that existing toolboxes may be used).
Concrete examples show that a Daubechies wavelet db3 lifted on the sphere is
more efficient than a discretized spherical CWT for detecting a singularity along the
equator on the sphere. The computational load is smaller and the precision is much
better, in the sense that the width of the detected singular curve is narrower. For
other tests and comparisons, we refer to [60, 535] or [536].
15.2 Wavelets on Other Manifolds 477

15.2 Wavelets on Other Manifolds

The two-sphere is not the only manifold that deserves a wavelet analysis. Indeed,
some data live on manifolds more complicated than the sphere, for instance, a two-
sheeted hyperboloid or a paraboloid. In cosmology, for example, an Anti-de Sitter
model of the universe is a one-sheeted hyperboloid, but a particle moving on that
manifold has a phase space isomorphic to a two-sheeted hyperboloid. In optics also,
data on non-Euclidean manifolds are essential for the treatment of omnidirectional
images, which have numerous applications in navigation, surveillance, visualiza-
tion, or robotic vision, for instance. In the catadioptric image processing, a sensor
overlooks a mirror, whose shape may be spherical, hyperbolic or parabolic [155].
However, instead of projecting the data from that mirror onto a plane, one can
process them directly on the mirror, which then suggests to use wavelets on such
manifolds [156].
Another example is a closed sphere-like surface, that is, a surface obtained
from a sphere by a smooth deformation. Such manifolds are used in medicine,
for modelling some organs. Wavelets on such a surface are obtained by combining
the radial projection described in Sect. 15.1.3.1 (3) with a smooth deformation of
the sphere onto the surface [527]. More generally, data may be given on a two-
dimensional smooth manifold M , for instance a surface of revolution. Wavelet
transforms on such manifolds have been constructed in all these cases and we are
going to sketch some of them in the following sections.

15.2.1 Wavelets on Conic Sections

The sphere, the two-sheeted hyperboloid and the paraboloid constitute the so-called
conic sections, generated by intersecting a double cone by a plane. This was a major
discovery due to Apollonius of Perga (c. 262 BC c. 190 BC), a Greek geometer and
astronomer of the Alexandrian school (incidentally, it was Apollonius who gave the
ellipse, the parabola, and the hyperbola the names by which we know them [86]).
We will now sketch wavelet constructions on these manifolds. Additional details
may be found in [64, 81].

15.2.1.1 The Two-Sheeted Hyperboloid H2

The upper sheet H2+ = { = (1 , 2 , 3 ) R3 : 12 + 22 32 = 1, 3 > 0} of

the two-sheeted hyperboloid may be treated exactly as the sphere, replacing SO(3)
by the isometry group SOo (2, 1). Indeed, one has H2 SOo (2, 1)/SO(2), which
is the noncompact Riemannian symmetric space dual to the sphere S2 . For dilations,
however, a choice has to be made, since there are many possibilities, each type
being defined by some projection. Details may be found in [156, Bog05]. Given an
(admissible) hyperbolic wavelet , the hyperbolic CWT of f L2 (H2+ ) with respect
to is
478 15 Wavelets on Manifolds


(WH f )(g, a) := g,a | f  = a (g1 ) f ( ) d ( ), g SOo (2, 1), a > 0,
H2+
(15.35)
a formula manifestly analogous to its spherical counterpart (15.8). As in the
spherical case, a ( ) = (a, ) (d1/a ), with da an appropriate dilation, (a, )
is the corresponding RadonNikodym derivative, and is the SOo (2, 1)-invariant
measure on H2 .
As for the sphere, the key for developing the CWT is the possibility of performing
harmonic analysis on H2+ , including a convolution theorem, thanks to the so-called
FourierHelgason transform. As a consequence, the usual properties hold true, for
instance, an exact reconstruction formula. However, no result is known concerning
frames that would be obtained by discretization.
On the other hand, it is possible to construct wavelet orthonormal bases on H2+ by
lifting them from the equatorial plane 3 = 0 by inverse orthographic (i.e., vertical)
projection. In this case, no point has to be avoided, since only one pole is present, but
distortions will occur again if one goes sufficiently far away from the tip (pole). This
drawback can be avoided, however, if one uses instead an area preserving projection
that will be constructed in Sect. 15.2.3.1.

15.2.1.2 The Paraboloid

Among the three shapes for a catadioptric mirror, the parabolic one is the most
common (think of the headlights of a car). And this case brings us back to
the topic of Sects. 15.1.1 and 15.1.3.2. Indeed it has been shown in [329] that
the reconstruction of the orthographic projection from a parabolic mirror can be
computed as the inverse stereographic projection from the image plane onto the unit
sphere. Thus wavelet frames and wavelet orthogonal bases may be obtained from
the corresponding spherical constructions.
The (axisymmetric) paraboloid is the manifold of equation P2 = {(1 , 2 , 3 )
R3 : 3 = 12 + 22 }. This is a singular limit case between the sphere S2 and the
two-sheeted hyperboloid H2 . The paraboloid P2 does not have a constant curvature
and it also lacks a large isometry group, so that the general coherent state formalism
cannot be used. Therefore, designing a CWT on P2 , by a limiting procedure or
otherwise, is bound to be a difficult process.
A different approach has been suggested, however, which is based on a
homeomorphim from a cylinder onto the pointed paraboloid P2 := P2 \{0} [Hon07].
The construction follows the group-theoretical method, but the outcome is a time-
frequency transform, resembling rather a CWT in k-space (momentum space). In
particular, the formalism does not allow for a local dilation around an arbitrary
point, so it is not really a wavelet transform.
Alternatively, one may lift planar orthogonal wavelet bases onto the paraboloid
directly by inverse orthographic projection, as for the hyperboloid, with the same
danger of distortions far away. Here too, such distortions will be avoided is one uses
an area preserving projection from the paraboloid to the plane, as will be seen in
Sect. 15.2.3.1.
15.2 Wavelets on Other Manifolds 479

15.2.2 Wavelets on the n-Sphere and the Two-Torus

In fact the method developed in Sect. 15.1.1 may be generalized almost verbatim
to the n-sphere Sn , for any n 1 [62] and then used for defining a CWT on the
two-torus T2 . These manifolds occur indeed in various applications. For instance,
signals living on a n-sphere or on a torus are often relevant in statistics [115, 116].
As for the torus, it is, of course, the basic geometrical shape underlying a tokamak,
the device using a magnetic field to confine a plasma in the shape of a torus
and a candidate for achieving nuclear fusion [585]. Wavelets on a torus may
also be significant in Superstring/M-Theory. Indeed it is known that the lowest
order quantum corrections are calculated by using Riemann surfaces that have the
topology of tori [Bec06].
We will sketch here the main points of the construction of wavelets on these
manifolds.

15.2.2.1 Wavelets on the (n 1)-Sphere Sn1

Following [62], we consider finite energy signals f L2 (Sn1 , d ), where d ( )

is the usual (rotation invariant) measure on Sn1 . For constructing a CWT on Sn1 ,
we have to identify the appropriate affine transformations. For that purpose, we
note first the isomorphism Sn1 SO(n)/SO(n 1), from which we deduce
immediately that SO(n) acts transitively on Sn1 ; this action belongs to the motions
or displacements on the sphere, that we have analyzed in depth in Sects. 7.4 and 7.6.
As for dilations, we note that the conformal group of both Sn1 and Rn1 is the
pseudo-orthogonal group SOo (n, 1). The latter admits the Iwasawa decomposition
SOo (n, 1) SO(n) A N, where A SOo (1, 1) R+ R is the one-dimensional
subgroup of Lorentz boosts along the xn -axis; and N Rn1 is (n 1)-dimensional
and abelian, and corresponds under stereographic projection to translations in the
tangent hyperplane at the North Pole. From this, one computes the (transitive) action
of SOo (n, 1) on Sn1 . The result is that the transformation corresponding to a pure
dilation a A is exactly the usual Euclidean dilation in the tangent hyperplane lifted
on Sn1 by inverse stereographic projection.
Next we consider the following (class I) principal series representation of the
Lorentz group SOo (n, 1), acting in L2 (Sn1 , d ):
 
[U s (g) f ] ( ) = (g, )s/2 f g1 , Sn1 , (15.36)

where (g, ) is a cocycle or RadonNikodym derivative. The representation U s

is a strongly continuous representation of SOo (n, 1) in L2 (Sn1 , d ). It is reducible
if s = 0, 1, 2, . . . and cyclic otherwise. It is unitary and irreducible if and only
2 [576]. In the following, we set s = (n 1)/2 and write simply U
if Re s = n1
U (n1)/2 .
480 15 Wavelets on Manifolds

Since the parameter space is X SO(n) A SOo (n, 1)/N, we need a section :
X SOo (n, 1) in the principal fibre bundle defined by the Iwasawa decomposition.
As before, we take the natural (Iwasawa) section I ( , a) = a, SO(n), a A
SOo (1, 1). Using this section, we define the spherical wavelets (CS) as the elements
of the orbit of under SOo (n, 1):

I (x) ( ) = [U(I (x)) ] ( ), x = ( , a) X, (15.37)

provided L2 (Sn1 , d ) is admissible mod (N, I ). Indeed, it turns out that the
representation U of SOo (n, 1) is square integrable modulo (N, I ), that is, there
exists one (and in fact a dense set) nonzero admissible vector L2 (Sn1 , d )
such that
 
da
I= dn ( ) |U(I ( , a)) | |2 < , L2 (Sn1 , d ),
0 an SO(n)
(15.38)
where n is the invariant measure on SO(n). This means that the family
{I (x) , x X} is a continuous family of CS, but not necessarily a frame. For
n = 3, we have seen in Proposition 15.1.2 that the family indeed is a frame. The
proof presumably extends to higher dimensions, but this remains to be seen.
The results just described are the basic ingredient for writing the CWT on Sn1 .
Given an admissible vector (mother wavelet) L2 (Sn1 , d ), the wavelets on the
sphere are the functions ,a = U(I ( , a)) , and the CWT reads, with UI ( , a)
U(I ( , a)):

S( , a) = U(I ( , a)) | s

= [UI ( , a) ]( ) s( ) d ( ) (15.39)
Sn1

= a ( 1 ) s( ) d ( ).
Sn1

Note the exact analogy with (15.8).

As a final remark, we note that the Euclidean limit holds for this spherical CWT
exactly as for the case of the two-sphere S2 . For further details, including the explicit
admissibility condition, we refer to the original paper [62].
Let us now particularize to the case n = 2, that is, wavelets on the circle S1 . The
machinery applies, with S1 SO(2). The relevant Lorentz group is now SOo (2, 1) =
SL(2, R)/Z2 , with the Iwasawa decomposition SOo (2, 1) = SO(2) R+ R. This
case has been treated in a slightly different fashion in [177]. There the authors
start from SL(2, R) and consider the quotient X = SL(2, R)/N, where N is the
subgroup of upper triangular matrices with ones on the diagonal. They also use
a stereographic projection from the center of the circle onto the line tangent at the
pole, which entails that they must restrict themselves to the half-circle [ /2, /2].
Otherwise the treatment is entirely parallel, including the frame property and the
15.2 Wavelets on Other Manifolds 481

Euclidean limit. Actually, the rationale behind the choice of [177] is the desire to
treat wavelets on the circle and on the line in a unified manner. We may also mention
that another type of wavelet on the circle may be obtained by periodization [376].

15.2.2.2 Wavelets on the Two-Torus T2

An additional benefit of the S1 case is the possibility of constructing a CWT

on the two-torus T2 [179]. We parameterize T2 = S1 S1 by = (1 , 2 ), i
( , ), and consider signals living in L2 (T2 , d ), where d = d1 d2 . Next,
since T2 SO(2) SO(2), wavelets on T2 will be obtained by performing twice
the construction on S1 described above. This means embedding each factor SO(2)
into SOo (2, 1) via the Iwasawa decomposition of the latter. Moreover, one can
rely on SOo (2,
 2), since it is locally isomorphic to SOo (2, 1) SOo (2, 1), in fact
SOo (2, 2) = SOo (2, 1) SOo (2, 1) /Z2 . Thus the parameter space of the CWT
on T2 is X = T2 (R+ ) , which can be parameterized by (1 , 2 , a1 , a2 ), with
2

i ( , ), ai (0, ), i = 1, 2.
Translations on T2 are simply i  i i , i = 1, 2, which are represented in
L (T2 , d ) by
2

[U(1 , 2 ) f ](1 , 2 ) = f (1 1 , 2 2 ).

As for dilations, they are given by applying the relation (15.1) to each factor i . This
action is represented in L2 (T2 , d ) by the unitary operator
 
[Da1 ,a2 f ](1 , 2 ) = (a1 , 1 )1/2 (a2 , 2 )1/2 f (1 )1/a1 , (2 )1/a2 , (15.40)

where is the cocycle given by

2a
(a, ) = .
(a2 1) cos + a2 + 1

For the construction of the CWT itself, we simply follow the spherical case. First, a
nonzero function L2 (T2 , d ) is admissible if one has, for every f L2 (T2 , d ),

da1 da2
|a11,a,22 | f |2 d1 d2 < , (15.41)
X a21 a22

where a11,a,22 = U(1 , 2 )Da1 ,a2 . Going to momentum space (that is, Fourier
space, as for S2 ), one can see that L2 (T2 , d ) is admissible if there exists a
positive constant c < such that
 
da1 da2
an11,a,n22 |2
| < c, n1 , n2 Z2 ,
0 0 a21 a22
482 15 Wavelets on Manifolds

where an11,a,n22 = n1 ,n2 | a1 ,a2  is the Fourier transform of a1 ,a2 = Da1 ,a2 and
n1 ,n2 (1 , 2 ) := (2 )1 exp(in1 1 ) exp(in2 2 ), n1 , n2 Z, which constitute an
orthonormal basis in L2 (T2 , d ). In this case also, there exists a simpler, necessary,
mean zero condition for admissibility.
In conclusion, the wavelets on the torus are the functions {a11,a,22 , (1 , 2 , a1 , a2 )
X} and the CWT on T2 reads as

F(1 , 2 , a1 , a2 ) = a11,a,22 | f 

= a11,a,22 (1 , 2 ) f (1 , 2 ) d1 d2 , f L2 (T2 , d ). (15.42)
T2

We see an almost complete parallelism between the CWT on T2 and the CWT
on S2 . In particular, the family of functions {a11,a,22 , (1 , 2 , a1 , a2 ) X} constitutes
a continuous frame (presumably nontight) whenever the function is admissible.
In addition, the Euclidean limit holds true here also.
There remains the question of building admissible functions on the torus and
designing orthonormal bases among them. Because of the direct product property
of SOo (2, 2), it is enough to lift onto T2 , by inverse stereographic projection, tensor
product admissible functions on the tangent plane R2 . As for orthonormal bases,
one best proceeds, like for S2 and other surfaces of revolution, by the technique of
area preserving maps, that we shall describe in detail in the next section.
There is a problem, however. All the wavelet constructions rely on a single,
global, dilation parameter a > 0, whereas here we need two of them, a1 , a2 . Thus
one is tempted to add an additional restriction such as a2 = a1 or, more generally,
a2 = (a1 ), for some function . The problem is that, when such a relation is
imposed, the frame property is lost. The way out is to add another variable to the
parameter space X  , the restriction of X under the relation . To that effect, the
authors of [179] propose to add the action of the modular group SL(2, Z), that
is, the subgroup of SL(2, R) consisting of matrices with integer entries. While
this group indeed acts on the torus T2 , it is too big, so that one has to restrict
the additional parameters to the quotient X = SL(2, Z)/Np , where Np denotes
the isotropy subgroup of some point p T2 . In this way, one can define a new,
restricted admissibility condition, called modular admissibility, under which the
resulting wavelet family indeed constitutes a frame. For details, we refer to the
original paper [179].

15.2.3 Wavelets on Surfaces of Revolution

Let us go back to dimension 2, that is, surfaces. The sphere, the axisymmetric
hyperboloid or paraboloid, the torus are all surfaces of revolution, invariant under
rotations around the vertical axis. The sphere and the torus are compact, the other
two are not. One can also consider arbitrary surfaces of revolution, obtained by
15.2 Wavelets on Other Manifolds 483

rotating a smooth curve around the vertical axis. For such surfaces, the rotation
(longitude) degree of freedom is trivial, one has only to perform convolutions on
circles. So, in a sense, designing wavelets on a surface of revolution is basically a
one-dimensional problem.

15.2.3.1 Area Preserving Projections

In general, we assume data are given on a two-dimensional smooth manifold M ,

for which there is a bijective projection p from M (or a subset N M ) on a plane
R2 , for instance a tangent plane or an equatorial plane. If M is rotation invariant
and the projection p is radial, we have only to consider its action in a longitudinal
plane = const., where [0, 2 ) is the longitude angle. Such is the case for the
stereographic projection of Sects. 15.1.1 and 15.1.3.2 (see Fig. 15.1) and also for the
vertical or orthographic projection (1 , 2 , 3 ) M  (1 , 2 , 0) R2 . A systematic
analysis on wavelets on manifolds along these lines has been made in [82].
While these projections have nice properties, they suffer from a major drawback,
namely, they do not preserve areas. As a consequence, lifting the DWT via the
inverse projections results in severe distortions at large distances (e.g. close to the
South Pole in the case of the sphere). However, a simple method of constructing
a projection which preserves the area has been obtained in [534]. It applies to all
2-D surfaces of revolution M obtained by rotating a piecewise smooth plane curve
around a line in its plane, such that one end point of the curve is the only point
of intersection with the line and each plane perpendicular to the line intersects the
curve at most once. For the construction of a MRA of L2 (M ) and a CWT on M , we
also need to suppose that the curve that generates the surface has infinite length. This
will cover, for instance, the two-sheeted hyperboloid and the paraboloid. Moreover,
the method can be adapted to the case where only a finite portion of the curve is
considered, which then allows to treat a hemisphere or a truncated version of the
two other conic sections. We will now sketch that method.
Consider the surface of revolution M defined by the parametric equations

1 = cos ,
2 = sin , [0, 2 ), I = [0, b) or [0, ),
3 = ( ),

obtained by rotating the planar curve of equation 3 = (1 ) around O3 . We

suppose that 0, is piecewise smooth and increasing on I, thus bijective and
invertible.
In order to get an area preserving bijection p from M to a subset of the plane, we
impose that A (S ) = A (p(S )) for every portion S of M , where A (S ) denotes
the area of S . The intersection of M with the plane 3 = 0 , where 0 (I), 0 =
(0), will be a circle of radius 0 = 1 (0 ). In particular, for the portion M0 M
situated under the plane 3 = 0 , we require that A (M0 ) = A (p(M0 )). To that
484 15 Wavelets on Manifolds

Fig. 15.5 The projection

M  = p(M) (from [534])

effect, we calculate A (M0 ) and determine the radius R0 of the disk p(M0 ) with area
R20 equal to A (M0 ). Given a point M(1 , 2 , ( )) M , 1 , 2 = 0, its projection
M  = p(M) is obtained by defining its vertical projection N(1 , 2 , 0), then taking M 
on the half line ON, such that OM  = R0 (see Fig. 15.5). Then an explicit calculation
yields

A (M0 ) = h(0 ) = R20 , (15.43)

&
where h : I [0, ) satisfies the relation h ( ) = 2 1 + (  ( ))2 . It is obviously
increasing and continuous, hence bijective. Then, from the function h and its
inverse h1 , one can calculate the projection p and its inverse p1 . Clearly, the
relation (15.43) means that the projection p preserves the areas.
The possibility of applying this technique in concrete cases hinges on finding
explicitly h and h1 . Explicit formulas are given in [534] for the cases where M
is an axisymmetric paraboloid, the upper sheet of a hyperboloid or a cone, or a
truncated version of each. Contrary to the other cases, the expression of h1 could
not be obtained analytically for the hyperboloid, but this does not prevent the method
to work numerically in that case also.
As for the torus T2 , obviously it does not fit into the scheme just described,
being generated by the rotation of a circle around the 3 -axis, instead of an open
ended curve. Nevertheless, measure-preserving maps between S2 and T2 can be
obtained, thus allowing one to lift wavelet analysis from the plane to the torus, via
the sphere [179].

15.2.3.2 Multiresolution Analysis and DWT on M

Assume the curve has an infinite length. As before, the surface M is given by the
parametric equations
 
= (x, y) = 1 (x, y), 2 (x, y), 3 (x, y) , (x, y) R2 .
15.2 Wavelets on Other Manifolds 485

Then we construct the area preserving projection p : M R2 and its inverse

p1 (x, y) = (x, y). In particular, the projection p preserves the area elements of
M and R2 : d ( ) = dx dy, that is, 1 in (15.33). Hence we have, for any
f, g L2 (M ),

 f| gL2 (M ) = f( )
g( ) d ( )
M

= f(p1 (x, y))
g(p1 (x, y)) dx dy
p(M )

=  f p1 | g p1 L2 (R2 ) . (15.44)

In the same way, we have, for any f , g L2 (R2 ),

 f | gL2 (R2 ) =  f p | g pL2 (M ) . (15.45)

As in Sects. 15.1.1.1 and 15.1.3.2, we define the map : L2 (M ) L2 (R2 ) by

f = f p1 , and its inverse 1 : L2 (R2 ) L2 (M ) by 1 f = f p. Then the
relations (15.44) and (15.45) mean precisely that and 1 are unitary maps.
Thus we are essentially in the same situation as in Sect. 15.1.3.2, the difference
being that the correction factor , which leads to distortions, is now identically
equal to 1. Hence an MRA of L2 (R2 ) may be lifted by 1 to an MRA of
L2 (M ) and similarly for orthogonal wavelet bases, Riesz bases, etc. In addition,
the decomposition and reconstruction matrices needed for M are the same as for
the planar case R2 , since all calculations are done directly on R2 , using standard
toolboxes.

15.2.3.3 CWT on M

The same technique can be used for lifting the CWT from the plane to the
manifold M . According to the standard approach, the 2-D plane CWT is derived
from the unitary irreducible representation of the similitude group SIM(2), namely,
particularizing (14.4) to 2 dimensions,

[U(b, a, )s ](x) = a1 s(a1 r (x b)), s L2 (R2 ). (15.46)

Combining this operator with the projection p, we obtain a unitary irreducible

representation of the similitude group SIM(2) in L2 (M ):
 
[ U (b, a, ) s p1 ) (a1 r (p( ) b) , s L2 (M ).
s ]( ) = a1 ( (15.47)

Next we define wavelets on M . Given an admissible plane wavelet L2 (R2 ),

define its lifted counterpart as
 := p. The latter is admissible if
486 15 Wavelets on Manifolds


|
 p( )|2
c = c = d ( ) < .
M |p( )|2

b,a, := U (b, a, )
In that case, the transformed wavelets  are admissible as well.
Finally, the CWT of the signal s L (M ) with respect to the wavelet
 2  naturally
reads as

S (b, a, ) = U (b, a, )

 | sL2 (M

= U(b, a, )
 p1 | s p1 L2 (R2 )

= b,a, | sL2 (R2 = S (b, a, ),

where s = s p1 , that is, the usual 2-D plane CWT given in (14.63). In the same
way, one gets a reconstruction formula on M :

da
s( ) = c1


b,a, ( )S (b, a, ) db 3 d .

SIM(2) a

15.2.3.4 Uniform Grids on the Sphere and Surfaces of Revolution

The simple lifting operation described so far is not sufficient in practice. In many
applications one requires simple, uniform and refinable grids on the sphere. One
simple method to construct such grids is to transfer existing planar grids. As we have
seen, the standard manipulation of spherical data includes convolutions with local
and global kernels, Fourier analysis with spherical harmonics, wavelet analysis,
nearest neighbour search, etc. Some of these manipulations become very slow if
the sampling of functions on the sphere and the related discrete data set are not well
designed. Thus, the discrete data should have the following properties:
(a) Hierarchical tree structure (allowing construction of an MRA);
(b) Equal area for the discrete elements of the partition of the sphere;
(c) Isolatitudinal distribution for the discrete area elements (essential for fast
computations involving spherical harmonics).
Most grid constructions given in the literature do not provide an equal area partition
(see [537] for a detailed list). However, this property is crucial for a series of
applications, including statistical computations and wavelet constructions, since
non-equal area partitions can generate severe distortions at large distances.
Until the construction of the HEALPix grid [340], no grid satisfied simultane-
ous the requirements (a)(c). Recently, however, Rosca [532, 534] succeeded in
designing such grids, starting from a map that preserves areas and maps a square
onto a disc. Thus, any uniform grid of the square can be mapped bijectively onto
a uniform grid on the disc. By the same token, it is trivial to refine those grids,
15.2 Wavelets on Other Manifolds 487

Fig. 15.6 Two uniform grids on the unit disk (from [532])

Fig. 15.7 The grids of Fig. 15.6 projected on the sphere by Lamberts projection, separately for
the two hemispheres (from [532])

simply by refining the one on the square, by subdividing the cells. We show in
Fig. 15.6 a uniform grid on the unit disk and its refinement, both obtained from the
corresponding grids on the square.
Next we lift these two grids onto the sphere, via Lamberts azimuthal equal area
projection [435]. Actually, here we have two possibilities. Either we project the grids
on the whole sphere, or we project them on the northern and southern hemispheres
separately. We show in Fig. 15.7 the result of the latter operation. It is clear from the
figure that these grids satisfy the requirements (a)(c) above. In particular, all cells
have the same area.
Compared to the projections used for HEALPix grids and to the planar domains
mapped there, this method has the nice property that, besides the uniform grids and
sampling points on the sphere, one can transport any function from the square to the
sphere by the technique used in Sect. 15.1.3.2. In addition, if one has a set of real
functions defined on the square, the corresponding spherical functions preserve the
following properties: orthonormal basis, Riesz basis, frame, and local support. Thus
the construction is both simple and efficient in practice.
488 15 Wavelets on Manifolds

The same method may be applied in other situations. For instance, there is a
similar, area preserving, bijection from a rectangle to an ellipse. Here too, uniform
grids may be transported [533]. Another example is an area preserving map from
the cube to the sphere, obtained in two steps, first an area preserving map from a
square to a curved square, applied to each face successively, followed by an inverse
Lambert azimuthal projection to the sphere. In this way, one can construct other
uniform grids on the sphere, that may find practical applications in the geosciences
[537]. Exactly the same method may be used for mapping a regular octahedron to
the sphere, replacing squares by triangles. Thus one obtains still other regular grids
on the sphere [538]. We refer to the original papers for the details.
The same technique can be used for generating uniform grids on other surfaces
of revolution. One lifts a plane grid from a (large) square on the plane directly onto
the surface, using the appropriate area preserving projection. Explicit examples are
given in [534], namely, a paraboloid, the upper sheet of a two-sheeted hyperboloid
and a cone. For the hyperboloid, the analytical expression of h1 could not be
obtained, but the method still works well numerically. Indeed, in all cases, it
provides nice grids, adapted wavelets, and so on.
However, these grids have the disadvantage that, when we want to consider
only the portion of the surface located under a plane of equation 3 = 0 , then
some cells of the grid must be cut and the remaining cells do not have equal area.
Since such situations are the most frequent in practical applications, one needs also
the construction of uniform grids on (finite) surfaces of revolution M0 of equation
@ 1
3 = ( ), where = 12 + 22 ,
3 0 .

The solution consists in combining the inverse projection p1 with an area

preserving map T from a square to a disk, thus the lifting is done with p1 T ,
instead of p1 , starting from a uniform planar square grid. The two resulting grids
are shown in Fig. 15.8 in the case of the paraboloid [533]. Another interesting
example is that of a hemisphere, where these improved grids have promising
applications for cartography.

15.2.4 Local Wavelets

Let us go back to a general smooth 2-D surface M (the same technique applies for
an arbitrary smooth manifold of dimension n 2). If M admits a global group of
isometries, which is large enough, as in the case of the two-sphere S2 and the two-
sheeted hyperboloid H2 , one may use the group-theoretical approach discussed at
length in Sects. 15.1.1.1 and 15.2.1.1, respectively. In that case, the natural measure
on M and the projection onto a tangent plane R2 are determined by geometry (group
15.2 Wavelets on Other Manifolds 489

25
25
20
20
15
15
10
10
5
5
0
0 5
5
5
5
0 0 0
0
-5 -5
-5 -5

Fig. 15.8 Uniform grids on the paraboloid obtained by applying the projection p1 (left) and
p1 T (right) to a planar grid (from [533])

theory). However, the measure is not dilation invariant and the projection does not
preserve areas, which forces one to introduce correction factors (RadonNikodym
derivatives).
Alternatively, one may in some cases calculate a projection M R2 that does
preserve areas and is bijective, as we did in Sect. 15.2.3.1. In that case, one in fact
makes all calculations on the plane and one may use the natural inner product on M ,
not a weighted one. There is a representation in L2 (M ) of the similitude group of
the plane, as we have seen in Sect. 15.2.3.3, and this representation yields precisely
the CWT on M described there. Then one can lift the whole wavelet machinery
from the plane to M , as we have seen in Sects. 15.2.3.2 and 15.2.3.3. A case in
point is the axisymmetric paraboloid P2 . It would be interesting to compare the two
methods when they are both applicable, for instance, in the case of the sphere or the
hyperboloid.
If the manifold M does not admit a global isometry group and one does not
have the required area preserving projection, the group-theoretical method cannot be
used. There is an alternative, however, namely, to define a local wavelet transform,
a technique introduced in [82], that we now sketch.
Take as usual L2 (M , d ), for an appropriate measure . We denote by T (M )
the tangent plane at M . For defining the wavelet analysis on M , we need
two ingredients, a wavelet function around a point M and a proper dilation
operator on M . A local wavelet is a square integrable function ( ) L2 (M ) that
is compactly supported in a neighborhood B of M , that is, ( ) L2 (B ).
The size of B will depend on the local geometry of M . For defining a suitable
local dilation operator on L2 (M ), one exploits dilations in R2 and equips M with
a local Euclidean structure by mapping B to T (M ). A simple way to implement
this mapping is by flattening B along the normal to M at . Let p be that local
flattening map, that we assume to be a diffeomorphism:

p : B T (M ). (15.48)
490 15 Wavelets on Manifolds

Let B . A local dilation of coordinates is defined in the usual way, by

transporting the corresponding dilation operator da , a > 0, acting on the plane
T (M ):

 a = p1
da p ( ). (15.49)

( )
Obviously, the definition of p determines a local maximum scale amax , in order
that (15.49) be well-defined.
We can now easily construct a dilation operator acting on L2 (B ), following the
same pattern as before. Given ( ) L2 (B ), the following operator maps L2 (B )
unitarily onto itself:

( )
Da : ( ) ( )  1/2 (a, ) ( ) (a1 ),

where, as usual,

d (a1 )
(a, ) =
d ( )

is the corresponding RadonNikodym derivative that takes care of the possible

( )
noninvariance of the measure under dilation. As usual, we will write a :=
( )
Da ( ) .
Given a signal f L2 (M ), we can now define its wavelet transform at M
and scale a as
( )
(Wloc f )( , a) = a | f L2 (M ) . (15.50)

The inverse transform will take the general form:

  a ( )
( ) da
f= d ( ) (Wloc f )( , a) a .
M 0 a

The bound a ( ) used in this formula is a cut-off ensuring that the support of
( ) ( )
the dilated wavelet a ( ) stays under control, i.e., supp(a ( ) ) B . As for the
exponent appearing in the measure, it depends on the manifold at hand and serves
the same purpose.
An easy 1-D example given in detail in [82] is that of a circle. It generalizes
immediately to the sphere S2 . Let us fix the point 0 = (0 , 0 ) S2 . Then the
required projection p0 along the normal at 0 means the projection along the axis
O0 onto T0 , the plane tangent at 0 . In polar coordinates in the tangent plane T0 ,
this projection is given by
 
p0 ( , ) = sin( 0 ), T0 , 0 = (0 , 0 ).
15.3 Wavelets on Graphs 491

For instance, if we take 0 = , we recover the usual vertical projection on the plane
TS tangent at the South Pole: p0 ( , ) = (sin , ) .
However, there remains an open problem, namely, how to extend the local WT
to a global one. In other words, how to go from one point to another one on M . In
the language of differential geometry, this means going from one local chart to the
next one, and the tools for that are available. For example, in the case of the sphere
S2 (or the circle S1 ), one needs two charts, one for each hemisphere, with an overlap
around the equator, a well-known problem familiar to physicists of the 1970s in
the context of a proper treatment of magnetic monopoles. In the general case, one
may think for instance of translating the point and the corresponding patch B
along geosesics, but that does not always work. In the case of the paraboloid P2 , for
instance, it does not work. There indeed, geodesics are curves coming from infinity,
winding down around P2 all the way to the tip and winding up again to infinity, so
that each geodesic intersects itself infinitely many times. Clearly, some geometrical
work remains to be done here before one gets a genuine wavelet transform on an
arbitrary manifold.

15.3 Wavelets on Graphs

As ultimate step of our analysis, we consider the case of a WT on a graph. This

problem was treated briefly in [82], that we follow closely, then in depth in [363].
A graph is a mathematical object that is well suited to model pairwise relations
between objects of a certain collection, such as the nodes of a sensor network or
points sampled out of a surface or manifold. A graph can thus be defined as a
collection V of vertices or nodes and a collection E of edges that connect pairs
of vertices. We shall only consider finite undirected graphs, hence both sets will be
finite and edges have no direction. Another way of defining an (undirected) graph is
provided by its adjacency matrix A, that is, a square matrix A = (ai j )i, j of dimension
d = |V |, whose entry ai j equals the number of edges between vertex i and vertex j.
In the sequel, we will use the notation i j to denote that vertices i and j are linked
by an edge. For undirected graphs, the adjacency matrix is of course symmetric.
There is a natural notion of distance between any two vertices, namely, the length of
the shortest path between them, where a path is a sequence of vertices v1 , v2 , . . . , vk
such that vl vl+1 is an edge, for l = 1, 2, . . . , k 1 and no vertex is repeated.
Let us assume that the signals of interest are functions f : V R, which can
be identified with d-dimensional real vectors f Rd . We want to design a wavelet
transform on the graph, in such a way that the scaling parameter of the wavelet is in
accordance with the notion of distance between vertices. Our construction depends
on two fundamental ingredients. First, a suitable notion of Fourier transform on
the graph, that will be used to construct explicitly a scaling. Then, we will need a
way to localize wavelets around any chosen vertex of the graph. The key notion for
492 15 Wavelets on Manifolds

both steps is the fundamental object of spectral graph theory, namely, the Laplacian
matrix. In our finite dimensional, undirected setting, the Laplacian is defined as the
matrix L with entries

deg(i), if i = j,
i, j = 1, if i = j and i j, (15.51)

0, otherwise,

where deg(i), the degree of i, is the number of edges incident to i. The Laplacian
is known to be a positive semi-definite matrix. Thus it has a well-defined eigen-
decomposition. We will denote by k and k = (k (i))i=1,...,d , k = 0, . . . , d 1, its
eigenvalues and corresponding eigenvectors, respectively. We order the eigenvalues,
assumed to be nondegenerate for simplicity, in increasing order, 0 = 0 1 . . .
d1 . The eigenvectors form an orthonormal system that can be used to decompose
any signal. They will play essentially the same role as traditional Fourier modes,
which are also defined as eigenfunctions of the Laplacian. The Fourier coefficients
of f Rd are defined as fm = m | f , with the traditional Euclidean scalar product.
The Laplacian eigenfunctions can be used to define operators via their spectral
decomposition. Let : R+ R be a square integrable function. Define the operator
valued function (L) by its action on functions on V :
d1
(L) f = (k ) fk k , for f Rd .
k=0

For instance, choosing (u) = u, we recover the spectral decomposition of the

Laplacian:
d1
Lf = k fk k .
k=0

With a slight abuse of language, we will call (k ) the Fourier transform of the
operator . Thus the class of operators we consider are Fourier multipliers. In
our finite dimensional setting, (L) is simply a d d matrix. Using this indirect
description, it is natural to dilate the operator-valued function (L) by properly
scaling its Fourier coefficients, that is, using a harmonic dilation as in (15.22):

  d1
(tL) f ( j) = (t k ) fk k ( j), t R+
, j = 1, . . . , d.
k=0

Of particular interest to us is the case where the signal of interest is a Kronecker

delta at vertex i:

  d1
(tL)i ( j) = (t k )k (i)k ( j).
k=0
15.3 Wavelets on Graphs 493

We interpret this equation as the definition of a function at scale t, localized at

vertex i. Both operations, scaling and localization, depend on . By using these
mechanisms at various scales and for all vertices of the graph, we construct a family
of real-valued functions {i,t } on V , indexed by scale and location, namely,

i,t := (tL)i , i V, t R+
.

We call wavelets such localized and scaled functions and, by extension, we define
the wavelet transform of a real-valued function on V as the set of coefficients
 
W f (i, a) = i,a | f  = i,a ( j) f ( j).
jV

The corresponding inversion formula should take the following form:


da

f = c1 W f (i, a) v , (15.52)
iV 0 a

where is an arbitrary parameter and c a constant to be determined. Indeed, one

shows easily that the relation (15.52) holds true for any f Rd provided L2 (R+ )
and the constant c is defined by the relation
 +
da
0 < c = | (a)|2 < +. (15.53)
0 a

The condition is satisfied whenever is a continuous function that vanishes at the

origin, i.e., (0) = 0, that is, a necessary admissibility condition, similar to the
familiar one for wavelets.
Further information on the WT on graphs may be found in [82] and in [363]
to which we refer. The latter paper, in particular, presents spectacular examples of
spectral graph wavelets on the so-called Swiss roll data cloud, the Minnesota road
network and the cerebral cortex.
Chapter 16
Wavelets Related to Affine Groups

Abstract In this chapter, we discuss in detail the wavelets corresponding to the

affine WeylHeisenberg group and to affine groups of spacetime. This includes the
affine Galilei group and the affine Poincar group, as well as the application of
wavelets to motion analysis. We end with some generalizations, viz. wavelets on
Riemannian symmetric spaces.

In this chapter, we discuss in detail the wavelets corresponding to groups obtained

by adding dilations to the standard groups studied in previous chapters, namely,
the WeylHeisenberg group, the Galilei group and the Poincar group, as well as
the application of wavelets to motion analysis. All the resulting groups are of the
semidirect product type, hence this chapter illustrates the abstract theory developed
in Chap. 9.

16.1 The Affine WeylHeisenberg Group

Combining dilations with the WeylHeisenberg group gives a very natural

transformation group for signal processing, the affine WeylHeisenberg group
GaWH . It contains as subgroups both the affine group, that leads to the wavelet
transform, and the WeylHeisenberg group, that gives the windowed Fourier
transform (see Sect. 12.1), so the hope is that GaWH might yield a tool that
combines the advantages of the two techniques. CS associated to GaWH have been
constructed by Torrsani [586, 587] and also by Hogan and Lakey [373]. We follow
mostly the former.
A generic group element g = (b, , a) GaWH is the product of a translation b,
a frequency shift and a dilation a = 0 (we have dropped an irrelevant phase). The
group law reads as

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 495
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__16,
Springer Science+Business Media New York 2014
496 16 Wavelets Related to Affine Groups

(b, , a)(b ,  , a ) = (b + ab , +  /a, aa ).

Thus GaWH is a semidirect product of the WeylHeisenberg group (translations in

position and in spatial frequency) by the group R of dilations.
The group is unimodular and its invariant Haar measure is |a|1 db d da. GaWH
has a unique unitary irreducible representation, acting in L2 (R, dx), namely:
!xb"
[U(b, , a) f ](x) = |a|1/2 ei x f (16.1)
a
(Actually this is a projective representation (i.e., a representation up to a phase),
since we have dropped the phase factor coming from the central extension in GWH
throughout). In any case, this representation is not square integrable. One way to
obtain (weighted) CS is to multiply the Haar measure by some density function
( ) (at the price of losing the covariance with respect to frequency translations)
and then indeed one gets a resolution of the identity [586].
Another possibility is to restrict oneself to suitable subgroups of GaWH . Along
this line, one may proceed in two essentially equivalent ways.
1. Find a suitable subgroup of GaWH such that the restriction of (16.1) to it
becomes square integrable. A physically natural choice is to consider frequency
translations as functions of the dilation parameter, that is, consider elements of
GaWH of the form (b, (a), a). Enforcing the subgroup property then gives (this
amounts to solve a cohomological problem, i.e., to find a cocycle)
!1 "
(a) = 1 , R, (16.2)
a

and the subgroup G = {(b, (a), a), a = 0, b R}. The restriction of U to G

is indeed square integrable and yields ordinary CS.
One can also drop the requirement that G be a subgroup and show by direct
calculation that the most general function that yields a tight frame is:

(a) = + , , R. (16.3)
a
However, in both cases, the result is only a trivial modification of the usual
wavelet analysis, which may nevertheless have some interest in practice, because
it leads to a substantial gain in computing time.
2. Apply the general theory of Sect. 7.3. Define the one-parameter subgroup H =
{(0, (a), a) GaWH } R and the corresponding homogeneous space X =
GaWH /H . With coordinates (b, ) R2 , X has the GaWH -invariant measure
d (b, ) = db d . Define a section : X GaWH as (b, ) = (b, , (b, )),
with : X R a piecewise differentiable function. Then a straightforward
calculation shows that the most general section yielding a resolution of the
identity (a tight frame) is given by a function of the form
16.1 The Affine WeylHeisenberg Group 497

(b, ) ( ) = (c + d)1 , c, d R, = d/c. (16.4)

For c = 0, one gets a constant section, which leads to a standard Gabor analysis,
whereas c = 0 gives wavelet analysis.
Thus, in order to get something new, one has to consider non-tight frames.
A natural solution has been found by Torrsani [587], and essentially the same result
has been recovered by Hogan and Lakey [373]. Given a vector L2 (R, dx), define
the positive function

( ) =  (( ) ( ))2 | ( )| d .
| (16.5)
R

Call the function : R R+ admissible for if the function  is bounded and

bounded away from 0, i.e., there exist two positive constants K1 , K2 such that

0 < K1 ( ) < K2 < , a.e.

Proposition 16.1.1. Let GaWH = {(b, , a)} be the affine WeylHeisenberg group
and U its natural unitary representation (16.1) in L2 (R, dx). Let : R R be a
piecewise differentiable function and : (b, )  (b, , ( )) the corresponding
section in the bundle GaWH X = GaWH /H R2 , where H is the subgroup of
GaWH of all elements of the form (0, (a1 1), a). Let L2 (R, dx) be such that
the function is admissible for . Then the representation U is square integrable
mod (H , ) and the vector generates a frame of CS:

(b, ) (x) = [U( (b, )) ](x),

 
= | ( )|1/2 ei x ( )1 (x b) , (b, ) X R2 . (16.6)

The resolution (or frame) operator A is the operator of convolution by a function

, which is the inverse Fourier transform of  given in (16.5). Finally, a tight frame
of weighted CS b is obtained explicitly, in Fourier space, as

( ( ) = ( )1/2  ( ).

(16.7)
b (b, )

The proof of this proposition reduces to a straightforward computation of the matrix

element
 
 f |A f  = |U( (b, )) | f |2 db d = ( ) | f( )|2 d .
R2 R

The rest is immediate.

In order to be admissible, a function must satisfy a rather mild technical
condition given in [587], which means essentially that it does not oscillate too
498 16 Wavelets Related to Affine Groups

wildly. Several explicit examples may be found in that paper, including of course
the affine function ( ) = (c + d)1 discussed previously. Another interesting
example [373, 587] is the function
@
1, for | | o ,
comp ( ) = (16.8)
| |o + , for | | > o ,

where and o are two positive constants. The associated CS, called composite
wavelet packets, behave as Gabor CS for low frequencies (| | < o ) and as wavelets
for high frequencies. In this way one truly obtains the interpolation between the two
standard methods of time-frequency analysis. It may be noted that such a mixed
behavior is quite common in physiological sensors, like the ear or the eye [DeV88,
Mar82].
The CS constructed in Proposition 16.1.1 offer interesting perspectives in signal
processing, for they constitute a continuous version of the widely used wavelet
packets [210, 211, Mey91] (see Sect. 13.1.3). (Notice that the objects described
here have nothing to do with what we have called continuous wavelet packets in
Sect. 13.2, following [268].)
An alternative approach, which has become popular in the signal processing
community, is to ignore the problem of square integrability and work directly with
functions
!xb"
hb, ,a (x) := (U(b, , a)h) (x) = |a|1/2 ei x h , (b, , a) GaWH , (16.9)
a
where the window h is a real C function in L2 (R, dx), with nonvanishing integral,
>
h(x) dx = 0. A typical example is a Gaussian window, and it is of course optimal
in the Gabor sense (i.e., with respect to uncertainty relations or phase space
localization). Such functions hb, ,a are called time-frequency atoms (one usually
takes a > 0, since it is meant as a scale variable). For a given function h, any family
of such atoms is called a dictionary, as we briefly mentioned in Sect. 14.5.7. The
goal is then to find a suitable dictionary {h , GaWH }, usually countable,
that allows to expand into it an arbitrary function f L2 (R, dx), in such a way that
the expansion is well adapted to the particular class of signals to be analyzed. An
efficient algorithm to that effect has been introduced by Mallat and Zhang, under
the name of matching pursuit [459], and it offers a good alternative to the best basis
algorithm [210, 211] mentioned in Sect. 13.1.3(2). However, since the dictionary is
discrete in general, this is really a discretization problem, so we will postpone the
discussion to Sect. 17.3.1.
The whole construction extends to n dimensions [400]. The n-dimensional
affine WeylHeisenberg group, also denoted GaWH , is a semidirect product of
the n-dimensional WeylHeisenberg group (translations in position and in spatial
frequency) by the group R+ SO(n) of dilations and rotations. The unitary irre-
ducible representations of GaWH may be constructed by an extension of Kirillovs
method of coadjoint orbits (GaWH is not solvable for n > 2). Among them is the
16.2 The Affine or Similitude Groups of Spacetime 499

(projective) Stone-von Neumann representation in L2 (Rn , dx):

! ! x q ""
[U(q, p, a, R) f ](x) = an/2 eip(xq) f R1 , (16.10)
a

with R SO(n) and the other variables as before. This representation is not square
integrable, hence one has to find appropriate homogeneous spaces.
As shown in [400], many tight frames may be constructed for the following three
spaces, by the same method as in the one-dimensional case:
X1 = GaWH /H1 , where H1 is the subgroup of dilations.
X2 = H2 \GaWH , where H2 is the subgroup of spatial frequency translations.
Notice that here GaWH acts on X2 on the right. It follows that the resulting CS
do not coincide with the multidimensional wavelets mentioned in Sect. 14.2.
X3 = GaWH /H3 , where H3 = R+ SO(n) is the subgroup of dilations and
rotations.
In addition, non-tight frames and the associated weighted CS may be obtained
explicitly in all three cases, by the same method as before.

16.2 The Affine or Similitude Groups of Spacetime

As mentioned at the end of Sect. 12.1, the similitude group associated to a given
geometry is obtained by adding dilations to the corresponding isometry group. Both
the ax + b group and SIM(n) are of this type, the latter being just the affine
Euclidean group, and both lead to interesting CS, namely wavelets. Thus it is natural
to look for CS associated to other similitude or affine groups.
However, the wavelets presented so far are designed for analyzing static signals
or images. When it comes to time-varying phenomena, such as moving objects
or successive images (such as on a TV screen or in a movie), an additional
time dependence must be included, that is, we have to look for affine groups in
spacetime. Here too, the general CS scheme is applicable, once one has identified
the appropriate group of spacetime transformations and its representation. We will
sketch first the simplest solution [267]. More sophisticated ones, based on the affine
Galilei group or the affine Poincar group will be discussed afterwards.

16.2.1 Kinematical Wavelets, Motion Analysis

In n space dimensions, a time-dependent signal of finite energy may be represented

by a function s L2 (Rn+1 , dx dt). On such signals, the required operations are
space translations (Rn ), time translations (R), space rotations (SO(n)) and two
independent dilations, for space and time variables, respectively. The appropriate
500 16 Wavelets Related to Affine Groups

form of the latter becomes evident if one notices that our visual system introduces
a correlation between the size and the speed of a moving object: in order to be
visible, fast moving objects have to be wide, and narrow objects have to be slow. The
following transformations (which are mathematically equivalent to two independent
dilations) reproduce that pattern:

x  a c1/n+1 x, t  a cn/n+1t. (16.11)

Here, a > 0 is a global scaling variable and c > 0 adjusts the speed. The
corresponding unitary operators in L2 (Rn+1 , dx dt) read:

Global dilation:
!x t "
(Da f )(x,t) = a(n+1)/2 f , , a > 0, x Rn , t R; (16.12)
a a
Speed tuning:
(Ac f )(x,t) = f (c1/n+1 x, cn/n+1t), c > 0, x Rn , t R. (16.13)

The two operations commute, so that scale analysis and speed analysis are
independent, as they should be.
Putting all this together, one arrives at the following group:
 
Gn,1 = Rn+1  R+ +
R SO(n) Z2 , (16.14)

with elements (b, ; a, c; R, ), where b Rn and R denote space and time

translations, respectively, a > 0 a global dilation, c > 0 the speed tuning, R SO(n)
a rotation and = 1 is time inversion (this additional parameter ensures the
irreducibility of the representation that will be used). Notice that Gn,1 is a semidirect
product of the type we have studied in Chap. 9.
The group law is

(b, ; a, c; R, )(b ,  ; a , c ; R ,  ) =

(b + ac1/n+1 Rb , + acn/n+1  ; aa , cc ; RR ,  ). (16.15)

The group Gn,1 is nonunimodular, with left and right Haar measures given
respectively, by

da dc db d dR da dc
d l = , d r = db d dR. (16.16)
a c an+1 a c
As in the previous cases, the group Gn,1 has a natural unitary representation U in the
space L2 (Rn+1 , dx dt) of finite energy signals:
16.2 The Affine or Similitude Groups of Spacetime 501

! c1/n+1 t "
[U(g) f ](x,t) fg (x,t) = a(n+1)/2 f R(x b), ,
a acn/n+1
g = (b, ; a, c; R, ) Gn,1 . (16.17)

This representation is irreducible and square integrable, and the admissibility

condition reads:

|
 (k, )|2
c := cn dk d < , (16.18)
Rn+1 |k|n | |

where  (k, ) is the usual Fourier transform of (x,t) and cn a dimensional

constant.
From there on, everything follows the traditional pattern. The wavelet transform
is defined as (T f )(g) = g | f  and it has all the properties we have encountered
previously. In particular, T is a multiple of an isometry (energy conservation):
 
c
|(T f )(g)|2 dl (g) =
2 Rn+1
| f (x,t)|2 dx dt, g = (b, ; a, c; R, ),
=1 Gn,1
(16.19)
and there is an inverse transform (reconstruction formula):

f (x,t) = 2c1
(T f )(g)g (x,t) dl (g). (16.20)
Gn,1

In practice, one should take an axisymmetric wavelet, as in the purely spatial case
(Sect. 14.2). Then the rotation R is replaced again by a point on the unit sphere, the
CWT lives on the space Gn,1 /SO(n 1) = Rn+1 R+ S Z2 , and the wavelets
n

so obtained are of the GilmorePerelomov type. The rest is as before.

Thanks to the filtering property in a and c, this CWT (called kinematical) is
efficient in detecting moving objects: the angular parameter detects the direction
of the target, the dilation parameter a catches its size, while the new parameter
c adjusts the speed of the wavelet to that of the target. In more than one space
dimension, the parameter c identifies the speed of the target, that is, the modulus
of its velocity, but the angular variables will detect the direction of the movement,
thus the full velocity vector. Thus the spatio-temporal CWT is a tool for motion
tracking or target detection. Clearly there are plenty of applications in which such a
technique might be used [436438, 475].
A case of high practical interest is n = 2, of course, that is, the analysis of video
or movie sequences, for which there are many applications, like traffic monitoring,
autonomous vehicle navigation, etc. Chapter 10 of our textbook [Ant04] is entirely
devoted to a systematic treatment of these problems, including the algorithms for
motion tracking. As for the choice of the wavelet, one uses a separable wavelet,
(x,t) = S (x)T (t). The choice made in [Ant04] is the product of two Morlet
wavelets, 2-D for the space component S , 1-D for the time component T .
502 16 Wavelets Related to Affine Groups

However, recent results indicate that a more efficient tool for speed capture is
obtained if one chooses instead a Gaussian conical wavelet (14.32) for the space
component [166].
Actually one may go slightly further. In the case of kinematical wavelets, the
motion (or relativity) group consists only of spacetime translations and rotations.
One can replace it by a more sophisticated one, either the Galilei or the Poincar
group, depending on which physical context one is working. If one then combines
either of these with spacetime dilations, one obtains the corresponding affine group.
Wavelets for these groups have been constructed also, and we shall discuss them
in the next section. Note that, in the nonrelativistic framework, it is not clear
which time-dependent wavelets are the most efficient, the Galilean ones discussed
in Sect. 16.2.2 or the purely kinematical ones described here (which are simpler).

16.2.2 The Affine Galilei Group

Let us begin by the most general case, that of the affine Galilei group in n space
dimensions, obtained by combining the Galilei group with independent space and
time dilations. But a choice must be made here.
Indeed, in quantum mechanics, the natural approach would be to consider the
semidirect product Gaff := G0 D2 of the pure Galilei group G0 (see Sect. 8.4.1) with
(n)
a two-dimensional dilation group D2 D2 (R+ ) R , and then to construct
2 2

projective representations of the resulting group, by the standard method of central

extensions [443], discussed in Sect. 4.5.3. However, a straightforward computation
[56] shows that the only central extensions of the group Gaff by R are of the form
G0  GWH , where GWH is the WeylHeisenberg group, itself a central extension
of D2 . In particular, the central extension procedure fails to generate mass, as it
does in the usual situation, without dilations. The alternative is to take first a central
extension of G0 , and then the semidirect product with D2 . The first step produces
the extended Galilei group G M , corresponding to the extension parameter M > 0
[443, 596], the next one the extended affine group Gaff M := G M  D . However, as
2
we shall see below, M should not be interpreted as mass, but rather as a mass unit
or mass scale (the physical mass varies under dilations, whereas M is fixed). Of
course, in the present wavelet context, the names mass or energy are purely
formal. Note that the central extension of G0 plays a rle in classical physics as
well [8].
However, before going to the most general case Gaff M , it is worth noting that, in

one dimension, the nonextended group Gaff (1, 1) is also very interesting. Indeed, it
has been shown in [204] that all the groups relevant to signal analysis can be derived
from Gaff (1, 1). This covers the affine group of the line (1-D wavelets), the Weyl
Heisenberg group (Gabor), the reduced shearlet group and the Stockwell group (that
will not be discussed in detail here). Thus we will start by outlining this hierarchy.
16.2 The Affine or Similitude Groups of Spacetime 503

An element g Gaff is written g = (b, a, v, , ) corresponding to the action

x  e x + e vt + a,
t  e t + b,

where b, a R index time, resp. space, translations and , R index time, resp.
space, dilations. In matrix form, we may write:

e ve a
g = 0 e b .
0 0 1

The corresponding group law reads as

gg = (b + e b , a + e a + e b v, v + e v , +  , +  ).

An interesting family of subgroups of Gaff (1, 1), called GHp , 1 < p 1, are
obtained by restricting the two dilations , to lie on a line = (p + 1)1 .
These groups are called extended Heisenberg groups. Among these, several cases
are distinguished:
1. The reduced shearlet group Sh(R2 ), corresponding to p = 1, i.e., = 12 ,
as defined in Sect. 14.5.6. Indeed, upon the identification (a, s, (t1 ,t2 ))
(e , v, (a, b)), one gets the matrix

a s a t1
0 a t2 ,
0 0 1

from which one recovers indeed the law (14.90).

2. The Schrdinger group, corresponding to p = 1/2, i.e., = 2 , discussed in
[56] (see also Sect. 16.2.3).
3. The wavelet group G+ or ax + b group may be identified with the subgroup of
Sh(R2 ) corresponding to b = v = 0.
4. The Stockwell group GSW consists of elements ( , , ), > 0, , R, with
multiplication rule
! 1 "
(1 , 1 , 1 )(2 , 2 , 2 ) = (1 + 2 + [1 1 (1 2 )], 1 2 , 1 + 2 .
1

This shows that the Stockwell group is a trivial central extension of the wavelet
group G+ .
Now we return to an arbitrary dimension n and consider the extended Galilei
group G M , M > 0, which is the (unique) central extension of G0 , with elements
g = ( , b, a, v, R) and group law
504 16 Wavelets Related to Affine Groups

 2 3 
gg = +  + M v Ra + 12 v2 b , b + b , a + Ra + b v, v + Rv , RR

This group is the one to use for describing a nonrelativistic system with mass M.
Finally, we will consider as affine Galilei group the semidirect product GaffM =

G  D2 , for M > 0 (actually this group is an extension of Gaff = G0  D2 by R, but

M

a noncentral one [56]). We shall denote a generic element of GaffM by

g = ( , b, a, v, R, , ), (16.21)

where R is the extension parameter in G M , b R and a Rn are the time

and space translations, respectively, v Rn is the boost parameter, R SO(n) is
a rotation and, as before, R, resp. R, index space and time dilations,
respectively. The action of g on spacetime is then

x  e Rx + e vt + a,
(16.22)
t  e t + b.

The group law of Gaff

M reads as:

!
gg = + e2  + M[e v Ra + 12 e v2 b ], b + e b , a + e Ra + e b v,
"
v + e Rv , RR , +  , +  . (16.23)

Moreover, the center of the group is trivial (that is, Gaff

M is a noncentral extension of

G ; it is also the automorphism group of G ).

M M

Quite naturally, the subgroup GSM of Gaff M corresponding to = 2 is called the

Galilei-Schrdinger group. It is studied in detail in Sect. 16.2.3 (see [56]).

In order to build spacetime wavelets associated to Gaff M , we consider a unitary

representation of spin zero [56]. It may be obtained by direct unitary imple-

mentation of the action (16.22) on spacetime, or by the Mackey method of
induced representations, discussed in Sect. 9.2.4. The Hilbert space is H = L2 (Rn
R2 , dk dE d ), and the representation reads:

[U(g) ](k, E, ) = e(n+2) /2 ei( +Ebka) (k , E  ,  ), (16.24)

with

k = e R1 (k M v), E  = e (E k v + 12 M v2 ),  = e2 . (16.25)

These relations, which are in fact a part of the coadjoint action of Gaff
M , show that

the real mass parameter is the combination M , and it varies under dilations, as it
should.
16.2 The Affine or Similitude Groups of Spacetime 505

 k2

Now, since the sign of E 2M and that of are both invariant under the
transformation (16.25), the representation U splits into the direct sum of four
irreducible subrepresentations U j ( j = ), corresponding to the decomposition of
Rn R2 into four disjoint subsets according to the signs of these two invariants.
In addition, the full representation U, and thus also each subrepresentation U j ,
is square integrable over Gaff
M . For definiteness we choose U , corresponding to
++
H++ = L (D++ , dk dE d ), where
2

/ k2 0
D++ = (k, E, ) Rn R2 : > 0, E >0 . (16.26)
2M

Then indeed an explicit calculation shows that:

 
| (k, E, )|2
|U++ (g) | |2 dg = c 2++   dk dE d , (16.27)
k2
Gaff
M D++ n+1 E 2M

where the left Haar measure on Gaff

M is

dg = e3 8 d db da dv dR d d , g = ( , b, a, v, R, , ),

where dR is the usual invariant measure on SO(3).

Thus, a vector H++ is admissible if and only if the last integral converges.
There is obviously a dense set of such admissible vectors. For any of them, , one
M , as = U (g) (g G M ),
gets a tight frame of Galilean wavelets, indexed by Gaff g ++ aff
with the expression (16.24) for the representation. These are of course CS of the
GilmorePerelomov type.
In addition, it is possible to construct sets of wavelets indexed by fewer
parameters by taking quotients X = Gaff M /H by some subgroups H and appropriate

sections : X Gaff . The simplest example is obtained if we take for H the

M

subgroup D0 R+
of time dilations e . The corresponding coset space X = Gaff /D0
M

is parametrized by points x = ( , b, a, v, R, ). We consider first the basic section

0 : X GaffM,

0 (x) = ( , b, a, v, R, , 0). (16.28)

A straightforward calculation shows that the admissibility condition reduces to

k2
setting E 2M = 1 (or a constant) in the integral (16.27), that is, a vector H++
is admissible mod(D0 , 0 ) if and only if the following integral converges:
 ! k2 " 2
(n+1) k, 1 + , dk d < . (16.29)
>0 2M

We consider now a general section:

(x) = 0 (x) (0, 0, 0, 0, I, 1, (x)) = ( , b, a, v, R, , (x)), (16.30)

506 16 Wavelets Related to Affine Groups

where : X D0 R+ is a Borel function and (x) D0 represents a time

dilation. Again this corresponds to a relation between the variables (k, E, ), which
may conveniently be written as f (k, E, ) = 0. Thus the admissibility condition
mod(D0 , ) reads:

| (k, E, )|2
2 ( f (k, E, )) dk dE d < . (16.31)
D++ n+1 (E 2M
k
)

For any such admissible vector , one obtains a dense set of CS, indexed by
M /D and given by
X = Gaff 0 (x) = U++ ( (x)) , where U++ ( (x)) is the
representation (16.24) with e replaced by (x) in (16.25). Of course, for (x) 1,
one recovers the basic section 0 .
An interesting example is given by (x) = e2 , i.e., the Schrdinger case. The
corresponding constraint relation is simply = const, that we normalize to = 1,
thus getting the admissibility condition:

| (k, E, 1)|2
k2 >0 2 dk dE < . (16.32)
E 2M E 2M
k

Notice that the -integration has disappeared in (16.32), because inserting the factor
( 1) in (16.31) is equivalent to quotienting out the subgroup of phase factors,
in addition to the time dilation subgroup D0 .

16.2.3 The (Restricted) GalileiSchrdinger Group

Another possibility is to impose from the beginning the relation = 2 , which

amounts to restrict oneself to the Galilei-Schrdinger group GSM = G M  DS ,
where DS is the corresponding one-dimensional subgroup of D2 . The rationale
of this restriction is that GSM is the maximal kinematical invariance group of
the free Schrdinger and heat equations, that is, the largest group of spacetime
transformations that map solutions of these equations into other solutions [484,504]
(in Niederers definition, the full invariance group, called the Schrdinger group,
contains in addition the so-called expansions, which are more or less the Galilean
equivalent of pure conformal transformations). It might be useful at this point to
recall the free Schrdinger equation:
! "
HS (x,t) := it + (x,t) = 0, (16.33)
2m
where HS is the free Schrdinger Hamiltonian describing a free quantum particle.
The reader may be surprised to see here quantum considerations entering a
chapter on wavelets, which are a signal processing tool, thus a classical concept.
16.2 The Affine or Similitude Groups of Spacetime 507

However, this is no coincidence, since wavelets are particular CS, a quantum notion.
Moreover, this fact illustrates the close connection between signal processing and
quantum physics.
The restriction to the Galilei-Schrdinger group leads to several drastic
modifications.
1. The group law changes; in particular, the addition of phase factors becomes
 
 = +  + M e v Ra + 12 e2 v2 b . (16.34)

As a consequence, the center of GSM is the one-dimensional subgroup of phase

factors, isomorphic to R.
2. Therefore, GSM is a central extension of G0  DS by R, and we have (
M
denotes a central extension with (mass) parameter M):

GSM = GM  DS = (R
M G0 )  DS = R
M (G0  DS ) (16.35)

Thus, in the Schrdinger case, the two operations
M and  commute. It is

interesting to note that DS is the only one-dimensional subgroup of D2 that admits
a central extension by R [56].
3. The parameter is now invariant, by (16.25), and the dependence on in the
representation reduces to a trivial phase. Accordingly it may be factored out,
exactly as for the WeylHeisenberg group. Thus we may fix = 1 as above, so
that M m becomes the mass, as it appears in the Schrdinger equation [596].
4. The restriction to GSM of the representation (16.24) cannot be square integrable,
since GSM has a noncompact center R. After quotienting out this center,
the representation space reduces to HS = L2 (Rn R, dk dE). In technical
terms, the original space is a direct integral over , and we are taking the
restriction to a single component, corresponding to = 1. The corresponding
reduced representation, US , splits into the direct sum of two irreducible ones,
US = U+S US , corresponding to the decomposition HS = HS+ HS , where
k2
HS = L2 (DS , dk dE), with DS = {(k, E) Rn R : E 2m 0} (the two
subspaces HS might be called Schrdinger-Hardy spaces, they are the analogues
of the usual Hardy spaces on R, i.e., the subspaces of progressive, resp. anti-
progressive, wavelets).
5. The restriction to the Galilei group G M of the representation US decomposes into
a direct integral of irreducible representations:

L2 (Rn R, dk dE) HE dE, HE L2 (Rn , dk)
R

US ( , b, a, v, R, 1, 1) = UE ( , b, a, v, R) dE,
R

where UE is equivalent to the usual unitary irreducible representation of G M .

508 16 Wavelets Related to Affine Groups

6. As expected from the discussion above, the two representations US are square
integrable modulo the center , with admissibility condition [compare (16.32)]

 (k, E)|2
|
2 dk dE < ,  HS ,
(16.36)
DS E 2m
k

where  (k, E) denotes the (n + 1)-dimensional Fourier transform of (x,t).

1/2
In other words, is admissible if and only if it belongs to the domain of HS ,
where HS is the free Schrdinger operator given in (16.33). There is a dense set
of admissible vectors , and each of them generates a tight frame of CS, of the
GilmorePerelomov type.
An interesting class of Schrdinger wavelets consists of functions of the form

(x,t) = HS (x,t), (16.37)

 S . Such are, for instance:

where is a suitable element of H
! " (x2 +t 2 )/2
The SchrdingerMarr wavelet: SM (x,t) = it + e .
2m
The SchrdingerCauchy wavelet:
! " n
SC (x,t) = it + (t + i)1 (x j + i)1 .
2m j=1

The choice (16.37) is based on the following property. Let be a wave function,
solution of the Schrdinger equation [HS V (x)] (x,t) = 0, with a potential V (x).
Then the wavelet transform of with respect to the wavelet (16.37) is given by

U+S (g) |  = U+S (g) |V , g = (b, a, v, R, ) GSM / . (16.38)

In particular, if is a solution of the free Schrdinger equation, its wavelet

transform with respect to is identically zero! The proof of (16.38) is immediate if
one notices the identity

U+S (g) HS U+S (g)1 = e2 HS ,

which expresses the fact that GSM is the invariance group of the free Schrdinger
equation. However, it remains to be seen whether Schrdinger wavelets of the
type (16.37) will be useful in solving actual problems in quantum mechanics.
A typical instance where the answer might be positive is the description of quasi-
classical (Rydberg) states of atoms, that is, states with a very high principal quantum
number, in which the atom is very loosely bound (see for instance [Coh77]).
16.2 The Affine or Similitude Groups of Spacetime 509

16.2.4 The Affine Poincar Group and the Conformal Group

In the relativistic case, the situation changes completely. First, by covariance, there
can be only one scale parameter, common to space and time variables. Hence
the (n + 1)-dimensional affine Poincar group, also called the WeylPoincar or
the similitude group, takes the form of a semidirect product:

SIM(n, 1) = Rn+1  (R+

SOo (n, 1)).

Next, the natural representation, obtained once again by the Mackey method,
is of the Wigner type, but acts in H = L2 (Cn , dm(k)), where Cn = R+
(SOo (n, 1)/SO(n)) is the solid forward light cone in Rn+1 and dm(k) =
(k2 )(n+1)/2 dk0 dk, with k2 = k02 k2 . In the spin 0 case, this representation reads:
 
[U(b, a, ) f ](k) = eikb f (a )1 k , (b, a, ) SIM(n, 1),
(16.39)
f L2 (Cn , dm(k)).

The new fact is that this unitary irreducible representation is square integrable, so
that standard CS may be constructed. Explicit examples have been obtained in [160,
Unt87]. Moreover, as shown in the latter paper, one can construct discrete subsets
of that family of CS which still constitute a frame. We will discuss these results and
some of their generalizations in Chap. 17.
In addition, this work has an interesting geometrical content, that warrants a
detailed discussion. The starting point is to embed SIM(n, 1) into SOo (n + 1, 2),
the conformal group of Minkowski spacetime (note the analogy with the method
described in Sect. 15.1 for deriving wavelets on the sphere). Indeed the representa-
tion (16.39) is simply the restriction to SIM(n, 1) of a projective representation of
SOo (n + 1, 2) (exactly as the familiar massless representations of P+ (1, 3)). Then
consider the complex tubular domain

T = Rn+1 + iCn := {z = k + ip : k Rn+1 , p Cn }. (16.40)

The tubular domain T is a homogeneous space for SOo (n + 1, 2) since

T SOo (n + 1, 2)/SO(n + 1) SO(2),

and a Riemannian globally symmetric space, realizing Cartans classical

domain BD I [Hel78] (see also Sect. 7.6). Hence SIM(n, 1) acts on T , by
(b, a, )z = a1 z b. Fix = 0 + i(1, 0, . . . , 0) as base point in T and let
z = g , g = (b, a, ) SIM(n, 1). Let now be an element of H and assume
it is an admissible vector for the representation U of SIM(n, 1) given in (16.39).
Then, following the general pattern, generates a family of CS of SIM(n, 1),
indexed by the points z T :
510 16 Wavelets Related to Affine Groups

z g = U(g) , g = (b, a, ) SIM(n, 1). (16.41)

Moreover, these CS constitute a tight frame, for one has:


|z | f |2 d (z) = c f 2H , f H = L2 (Cn , dm(k)), (16.42)
T

where, for z = k + ip, d (z) = (p2 )n+1 dp dk is the SOo (n + 1, 2)-invariant measure
on T . In the case n = 1, the admissibility condition reads [160]:

| (k)|2
dm(k) < , (16.43)
C1 k2

and it is manifestly satisfied for a dense set of elements of H, which proves the
square integrability of the representation U. A similar admissibility condition holds
for n > 1.
A particularly interesting example of admissible vector, given in [Unt87], is the
(unnormalized) function

G (k) = (k2 )(n+1)/2 ek0 . (16.44)

This function and the CS generated from it have been used by Unterberger as a
basic tool for his comprehensive relativistic generalization of Weyls operational
calculus (under the name of Fuchs and Klein-Gordon calculus). One may notice
that the admissible vector G coincides, up to the scaling factor (k2 )(n+1)/2 , with
the Poincar Gaussian vector G defined in Sect. 10.1.2, and originally introduced in
[398,399] (see the discussion in [28] and [30, Sect. 7]). Also the connection between
CS of the affine Poincar group, the conformal group and the tubular domain T is
mentioned in [212, 452].
In dimension n = 3, the conformal group of Minkowski spacetime is isomorphic
to SOo (4, 2)/Z2 SU(2, 2)/Z4 . Wavelets for this group have been studied in great
detail by Calixto and Prez-Romero [178], in the two realizations of the tubular
domain T and the Cartan symmetric domain.
To conclude, let us mention that the 1 + 1 dimensional case has been studied
also in [149] in the context of the classification of three-parameter extensions of the
affine group, leading to suitable generalized Wigner functions.

16.3 Some Generalizations: Wavelets on Riemannian

Symmetric Spaces

Actually, Unterbergers approach suggests a class of spaces X for which our

general formalism applies, namely the Riemannian globally symmetric spaces of
noncompact type [Hel78, Per86]. Indeed, if X = G/H is such a space, in other
words, a noncompact classical domain, then there exists a smooth global section
16.3 Some Generalizations: Wavelets on Riemannian Symmetric Spaces 511

: X G [Hel78, Theorem VI.1.1], and hence the principal bundle G X = G/H

is trivializable. Typical examples, besides T itself, are SU(1, 1)/U(1), discussed at
length in Sects. 4.2.2, 7.6 and 10.3, or SU(2, 2)/S(U(2) U(2)). Actually, we have
already met wavelets on classical domains. Besides the SU(1,1) case, for both S2
and H2 , the unit disk, image of one sheet or one hemisphere, are classical domains,
over which wavelets have been described in Sects. 15.1 and 15.2.1.1, respectively.
In the general case, let G be a connected noncompact semisimple Lie group and
H the maximal compact subgroup of G. Assume G contains subgroups that can be
interpreted as motions and dilations of a certain manifold Y . In mathematical terms,
the existence of wavelets associated to G amounts to find an isometry from L2 (X)
to L2 (Y ), where X is either a subgroup Go of G or the quotient X = G/H. In the
first case, the central tool is a square integrable or discrete series representation
of Go . In the second one, it is a representation of the relative discrete series of
X, that is, a square integrable representation of G on the corresponding symmetric
space G/H. Actually, there is a considerable literature about these representations
(see for instance [280, 491, 539], and the references contained in those papers). In
addition, we have given in Sect. 7.3.1. some more details about the noncompact case,
more precisely, the irreducible Hermitian symmetric spaces, classified by Cartan.
On the other hand, in the case of a Riemannian globally symmetric space of
compact type, where X = G/H is compact, no smooth global section X G exists
(although a global Borel section always does), but then any unitary irreducible
representation of G is square integrable mod (H, ), for any measurable section .
At this stage, no general theory is available for the wavelet problem in the general
framework of representation theory of semisimple groups, but results have been
obtained in a number of cases, and we describe some of them below.
1. Let again G = KAN be the Iwasawa decomposition of G (Sect. 4.5.2), with
minimal parabolic subgroup P = MAN and let N = (N), where is the Cartan
involution of G [Hel78, Kna96]. In this case, under specific conditions that
involve only N, a wavelet transform exists for any measurable subset S of MA,
with X = N S and U a principal series representation of G (induced by P),
acting in L2 (N) [404].
A particular case is that of G = SL(n + 2, R), with X = N A1 Hn R,
where Hn GWH is the WeylHeisenberg group in n dimensions and A1 A.
Here also Y = N Hn [405].
2. Wavelets on Hn have been obtained using closed subgroups of SU(n + 1, 1),
either the affine WeylHeisenberg group [449] or the minimal parabolic subgroup
P [365]. More precisely, consider the Siegel upper half-plane

H n+1 = {w Cn , z C : Im z > |w|2 }.

The (Shilov) boundary of H n+1 is the WeylHeisenberg group Hn and the

minimal parabolic subgroup P of SU(n + 1, 1) may be considered as the affine
automorphism group of H n+1 , in the sense that it consists of operations that
may be interpreted as translations, dilations and rotations on H n+1 . Here again
512 16 Wavelets Related to Affine Groups

P has a natural unitary representation in L2 (Hn ), which decomposes into a direct

sum of square integrable representations, each of which thus generates a wavelet
transform on Hn (one recovers the case of the affine WeylHeisenberg group
simply by ignoring the rotations; we have treated this case in detail in Section
16.1 above).
3. As a particular case, take n = 0 in the previous example. Then SU(1, 1)/K, where
K is the maximal compact subgroup, may be identified with the open unit disk D
in the complex plane. Then, starting from the UIR U j in the Hilbert space H j =
L2 (D, (2 j 1) d j ), described in Sect. 4.2.2, and using the Gauss decomposition
of SU(1, 1), wavelets or, more properly, coherent states may be derived along the
standard lines [368].
4. More generally, one may consider the Siegel upper half-space

Hn = Sp(2n, R)/SU(n)
{Z Cnn : Z T = Z, Im Z Pn }
Dn = {W Cnn : W T = W, I WW Pn },

where Pn is the cone of positive nn real symmetric matrices. The isomorphism

between Hn and the generalized unit disk Dn is the result of the Cayley transform
W = (Z iI)(Z + iI)1 , Z Hn . Thus, Hn is the tubular domain mentioned in
Sect. 9.1. Its (Shilov) boundary is the set B of n n real symmetric matrices. The
symplectic group Sp(n, R) acts on Hn and it contains two automorphism groups
of B, called P0 and P. The elements of P0 may be identified with translations and
dilations, while P contains in addition rotations from SO(n). Both P0 and P have
UIRs on L2 (B), which are square integrable, and B essentially reduces to the
nilpotent component in the Iwasawa decomposition of Sp(n, R). Thus we obtain
wavelets on B [367].
5. The previous example has been generalized in [366] to the classical domain of
type one (see also Sect. 7.6)

DI = {Z C(n+m)m : I Z Z > 0},

where A > 0 means that A is a positive definite Hermitian matrix. Then a Cayley
transform maps DI onto the domain

Dn,m = {(w, z) : w Cnm , z Cmm , Im z w w > 0}.

The (Shilov) boundary of Dn,m is the set

Sn,m = {(w, z) : w Cnm , z Cmm , Im z = w w}

= {(w, x) : w Cnm , x Cmm and Hermitian}.
16.3 Some Generalizations: Wavelets on Riemannian Symmetric Spaces 513

With a natural multiplication law, this set Sn,m is a group of affine automor-
phisms of the domain Dn,m , henceforth denoted K. Next, one may identify the
domain Dn,m with the so-called Iwasawa subgroup P of the group of all affine
automorphisms of Dn,m . It turns out that the group P has a natural unitary
representation U on L2 (K). The representation U is reducible and decomposes
into a direct sum of irreducible, square integrable, representations in closed
subspaces H of L2 (K). In addition, the map W defined, as usual as W f =
 f |U() L2 (K) , for any H admissible, is unitary from L2 (K) into L2 (P).
In other words, one has obtained wavelets in L2 (K) according to the standard
pattern.
For m = 1, the domain Dn,1 reduces to the Siegel upper half-plane H n+1 and thus
K is just the WeylHeisenberg group Hn , so that we have again wavelets in L2 (Hn ).
For n = 0, on the other hand, the domain Dn,0 reduces to the Siegel upper half-
space Hn , with boundary B, i.e., we are back to Example (4), leading to wavelets
on B.
Finally, if we take both n = 0 and m = 1, one gets back the ax + b group and the
usual 1-D wavelet analysis.
For further details on this analysis, we refer to the paper of He and Liu [366]
and to the textbooks of Helgason [Hel78] or Terras [Ter88] for the underlying
mathematics.
In conclusion, all those examples yield a wavelet transform, either on the
WeylHeisenberg group Hn , or on the space B of real symmetric matrices, starting
from various transitive group actions.
Chapter 17
The Discretization Problem: Frames, Sampling,
and All That

Abstract This last chapter is devoted to the discretization problem : can one obtain
discrete wavelets by sampling continuous ones? In particular, what happens to
frames under that operation? We start with the WeylHeisenberg group underlying
canonical CS and discuss Gabor frames. Next we describe discrete frames asso-
ciated with affine semidirect product groups, such as the affine WeylHeisenberg
group or the affine Poincar group. Finally we turn to groups without dilations, in
particular, the Poincar groups in 1+1 and 1+3 dimensions.

In the preceding chapters, we have encountered both continuous and discrete frames
(the latter are more traditional). But in fact the two are necessarily linked. Suppose
one is dealing with a continuous frame, of rank one for simplicity,

| x x | d (x) = A. (17.1)
X

When it comes to numerical calculation, the integral has to be discretized, so that in

effect one always restricts oneself to a discrete subset of X. The question is whether
this restriction will imply a loss of information and hence, there arises the frame
discretization problem: Given a rank one frame F {x , A}, x X, can one find a
discrete subset X such that the vectors {x j , x j } constitute a discrete frame:

|x j x j | = A ? (17.2)
x j

Note that the frame operator A need not coincide with the original operator A. In
fact, often it does not, since a continuous tight frame may very well contain discrete
non-tight ones. Examples will appear soon.
When the frame comes from some group representation U, x = U( (x)) ,
x X, the existence of such a discrete subframe depends on two factors, the fiducial

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 515
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__17,
Springer Science+Business Media New York 2014
516 17 The Discretization Problem: Frames, Sampling, and All That

vector and the subset X. Some results in this direction have been obtained
for specific groups, and we will survey them in the next few sections.
In fact, the impetus for constructing good discrete frames comes from signal
processing, more precisely, from wavelets [234], as explained in Sect. 12.5. These
results revived interest in the question of discretization and soon it was extended to
the Gabor or Windowed Fourier Transform, in other words, the canonical CS. Thus,
once again, we are back into old problems of quantum mechanics!
On the other hand, the same problem may be addressed from the point of view
of signal processing. The question then is reformulated as follows: How does
one sample a continuous signal without losing information? This is again an old
problem, which has been revitalized in the context of Gabor or wavelet frames.

17.1 The WeylHeisenberg Group or Canonical CS

We will begin our discussion with the case of canonical CS, equivalently, Gabor
frames. This is a world in itself, with a vast literature. Fortunately the two volumes
[Fei98, Fei01] give a detailed and up-to-date survey of the field, to which we refer
for further information and original references.
In this case, the geometry of the parameter space, i.e., phase space, is Euclidean.
Typically the discrete parameter set = {(qn , pm ), m, n Z} is a square (or
rectangular) lattice in the (q, p)-plane and the frame theorem simply says (in the
notation of Sect. 2.2) that a frame {mn := (qn ,pm ) } is obtained if the density of
is larger than a critical value. This is the well-known result of von Neumann
concerning the density of canonical coherent states [Per86], which is closely
related to standard theorems known in different circles under the names of Fourier,
Heisenberg (uncertainty relations), Nyquist or Shannon. More precisely, let =
qo Z po Z and consider the corresponding family {mn }, where

mn (x) = eimpo x (x nqo ), m, n Z. (17.3)

(Note that, compared to (2.33), we have dropped the constant phases). Thus the area
of the unit cell of the lattice is S = qo po . Then [125, 501, Fei98, Per86]:
1. If qo po > 2 (undersampling), the family {mn } is not complete.
2. If qo po < 2 (oversampling), the family {mn } is overcomplete, and remains so
if one removes a finite number of points from .
3. If qo po = 2 (critical sampling), the family {mn } is complete, and remains so if
one removes any single point from , but becomes noncomplete if one removes
two or more points.
This is a purely mathematical result, but it has an immediate translation in physical
terms, if one views the (q, p)-plane as the phase space of a quantum system, as
discussed in Chap. 2. Restoring the conversion factors, involving the Planck constant
h, we see that a cell of area 2 in the (q, p)-plane corresponds to a cell of area 2 h
17.1 The WeylHeisenberg Group or Canonical CS 517

in the phase plane. Thus the result above means that the system {mn } is complete
if there is, on average, not less than one CS in a cell of area 2 h in the phase plane.
Note that a corresponding statement is less obvious in the case of wavelets, for lack
of an explicit physical interpretation of the latter.
Coming back to the language of frames, and forgetting physical units, this
result may be reformulated as follows [Dau92]: frames, even tight frames, with
good time-frequency localization exist for qo po < 2 , no frame exists for qo po >
2 ; in the critical case qo po = 2 , frames do exist, but with bad localization
properties, as a consequence of the celebrated BalianLow theorem (BLT) given
below [119, 451]. In particular, orthonormal bases exist only in the critical case, and
thus are necessarily poorly localized. This results from the following inequalities,
where m and M are the frame bounds of the Gabor frame generated by L2 (R):

2
m  2 M. (17.4)
qo po

Theorem 17.1.1 (BalianLow theorem). Let L2 (R) and let qo , po > 0 satisfy
qo po = 2 . If {mn } is a Gabor frame, then
 
either |x (x)|2 dx = , or |k
 (k)|2 dk = . (17.5)

The statement (17.5) means that and  cannot both have fast decrease simul-
taneously (remember fast decrease of  means smoothness of ), i.e., the Gabor
frame {mn } cannot be well localized in phase space. Actually this theorem
may be extended in various ways, including the generalization to irregular sets
{qn , pm , m, n Z}. A thorough discussion may be found in [Dau92, Chap. 4] or
in [Fei98, Chap. 2].
In his original paper [294], Gabor used a Gaussian as the basis function , on
the grounds that it minimizes the joint uncertainty in phase space (in other words, it
saturates the uncertainty relations, as we have seen in Sect. 2.1. However, although
this yields a frame in the oversampling case, it does not in the critical case [109].
In addition, the corresponding expansion

= mn | 
mn , (17.6)
m,nZ

does not converge in L2 , but only in the sense of distributions.

An interesting extension of the BLT has been given by Zak [614,615]. The set-up
is a (quasi)-planar gas of electrons in a strong magnetic field perpendicular to the
plane. As is well-known, the energy levels of an electron in such a system form
a discrete ladder, called the Landau levels, each of which is infinitely degenerate.
Thus arises the problem of finding a good orthonormal basis for each of those
levels, preferably made up of well localized states. Now the extended BLT of Zak
asserts that a good localization of these states, which are in fact Gabor states, is
518 17 The Discretization Problem: Frames, Sampling, and All That

impossible if they are taken orthogonal: An electron in a magnetic field (in the z-
direction) cannot be well localized both in the x- and in the y-directions. As a matter
of fact, exactly the same result was obtained in [51] with orthonormal wavelets,
which shows that we have here a genuine physical effect.
For many aspects of the theory of Gabor frames, including the proof of the BLT,
an essential tool is the Zak transform, also called the kq-representation [109,Dau92,
Fei98]. This transform is defined as follows:

(Z f )(s,t) = e2 itl f (s l), (17.7)

lZ

and it can be shown to be unitary from L2 (R) to L2 ([0, 1] [0, 1]). In the present
context, its main virtue lies in the relation

(Zgmn )(s,t) = e2 ims e2 int (Zg)(s,t), m, n Z , (17.8)

where, as above,

gmn (x) = e2 imx g(x n), m, n Z .

Thus, in the kq-representation, the commuting displacement operators are

simultaneously diagonalized [precisely, this representation was introduced
originally for the sake of defining symmetric coordinates k = 2a t (quasimomentum)
and q = as (quasicoordinate) in the dynamics of electrons in solids (a is the lattice
spacing)]. In other words, we are using here for constructing a basis the canonical
CS corresponding to an abelian subgroup of the WeylHeisenberg group.
A closely related approach has been developed in [165, Bou97] and [Gaz09,
Chap. 16]. The idea is to start from a discrete subgroup of the WeylHeisenberg
group, GNdWH , made up of commuting translations in position and momentum,
exactly as above. This subgroup is obtained by restricting the action of GWH
to the torus T2 = R2 / , with the quantization condition qo po = 2 hN, N N,
and consists of the elements {( , nqNo , mp
N ), R, m, n Z}. This quantized
o

torus may be taken as the phase space of a quantum particle, and the aim of
the work is to study the corresponding dynamical system, both in the classical
setup (where it can be chaotic) and in the quantum context. To that effect, one
constructs unitary representations of the discrete WeylHeisenberg group GNdWH and
the corresponding CS. These CS precisely provide the (quantization) link between
the two frameworks, and the wavelet or CS transform essentially reduces to the Zak
transform. Note that the restriction of the usual (Schrdinger) UIR of GWH to its
discrete subgroup GNdWH is highly reducible, and decomposes into a direct integral
of N-dimensional UIRs. Correspondingly, the restriction of the CS transform to any
of these subspaces of dimension N almost coincides with the familiar FFT. As a
last remark, we ought to mention again that the Zak transform is the key tool for
the construction of CS on the circle S1 , or more generally on a torus, as developed
in [338].
17.2 Wavelet Frames 519

17.2 Wavelet Frames

Clearly, the question of the existence of a frame obtained by discretizing a given

CS transform must take into account the geometry of the parameter space, that is,
the lattice must be invariant under discrete operations from the invariance group
(which in general do not form a discrete subgroup).1 In the WeylHeisenberg case,
this requirement leads to choosing a rectangular lattice in the phase plane, as we
have just seen. Similarly, for the ax + b group, corresponding to 1-D wavelets,
the geometry of the time-scale half-plane is non-Euclidean, and the lattice must
be invariant under discrete dilations and translations:
For the scales, one chooses naturally a j = aoj , j Z, for some ao > 1;
For the times, one takes bk = k bo aoj , j, k Z.
Thus
j/2
jk (x) = ao (a
o x kbo ),
j
j, k Z. (17.9)

The most common choice is ao = 2 (octaves!) and bo = 1, which results in

jk (x) = 2 j/2 (2 j x k), j, k Z. (17.10)

It is worth noticing that this so-called dyadic lattice {(k2 j , 2 j ), j, k Z} is exactly

the same that indexes the DWT, although the two approaches are totally different
(see Sect. 13.1). Actually other lattices may be used here too, for instance based on
the golden mean , as described (for the discrete WT) in Sect. 13.3 above.
For a given choice of , ao , one finds a range of values of bo such that { jk }, as
given in (17.9), is a frame. Detailed results may be found in [232, Dau92]. Here we
will restrict ourselves to the following simplified version.
Theorem 17.2.1. Let and ao be such that:

(i) inf
1 ao j=
|
 (aoj )|2 > 0;

 ( )| C | | (1 + | |) , > 0, > + 1.
(ii) |
Then there exists boo such that { jk } constitutes a frame for all choices bo < boo .
Both the Mexican hat and the Morlet wavelet satisfy the conditions of the
proposition for the dyadic case, ao = 2, bo = 1, thus they both generate discrete
frames on the dyadic lattice (and in fact, on a more general class of hyperbolic
lattices [231, 232]). Explicit values for the corresponding frame bounds m, M may

1 Here we use the term lattice somewhat loosely. In the strict mathematical sense, a lattice in a

locally compact topological group is a discrete subgroup with the property that the quotient space
has finite invariant measure. In the case of Rn , n > 1, the proper term is Delone set.
520 17 The Discretization Problem: Frames, Sampling, and All That

be found in [Dau92]. For the Mexican hat, one finds m = 3.223, M = 3.596, so that
M/m = 1.116. This is not terribly good, and the Morlet wavelet is even worse. This
defect is remedied in practice by densifying the lattice , for instance replacing
in (17.10) the exponent j by N j, = 0, 1, . . . , N 1 (N is called the number of
voices). Taking for instance N = 4, one gets M/m = 1.007 for the Mexican hat.
This of course improves the situation considerably. For further details, we refer the
reader to [Dau92]. However, if the speed is the determining criterion, one can do
better by using the pseudo-QMF algorithm described in Sect. 13.2. The latter is in
fact close in spirit to the procedure explained here, but more efficient.
It is interesting to notice that, in the wavelet case, the relation analogous to (17.4)
reads, for the wavelet :

2
m
bo | (aoj )|2 M, (17.11)
jZ

for all = 0. This implies, in particular, that the sum should be almost constant,
a strong requirement indeed, which is rarely satisfied. By contrast, the inequal-
ities (17.4) do not impose any restriction on . The difference simply reflects
the different properties of the two groups which underlie the construction. The
WeylHeisenberg group is unimodular, so that every vector is admissible, whereas
the ax + b group is not, hence the nontrivial constraints on in order for it to be
admissible.
As a final remark, we note that similar results hold true for n-dimensional
Mexican hat and Morlet wavelets associated to SIM(n) [476, Mur90].

17.2.1 Alternative Approaches to the ax + b Group

Of course, the simplest situation occurs when the lattice is the orbit of a discrete
subgroup G. Unfortunately this is seldom possible: already in the 1-D wavelet
case, the dyadic lattice is not of this type. However, in that particular case the
difficulty may be circumvented [417] by embedding the ax + b group G+ into
SL(2, R) and then restricting to the discrete subgroup SL(2, Z), in one of lowest
discrete series representations of SL(2, R), indexed by spin s = 12 , 1, 32 , 2 or 52 . These
results, and the technique used for proving them, strongly suggest that arithmetic
ideas are likely to play a rle in his context.
Another possibility that goes in the same direction is to consider the dyadic
lattice = {(k2 j , 2 j ), j, k Z} as a subgroup of the dyadic affine subgroup
{(k2 j , 2l ), j, k, l Z} of G+ , as mentioned in Sect. 13.4.1. However this remains
unexplored up to now. Still another possibility is to introduce dilation semigroups.
Indeed, as is clear from (13.3), the actual parameter space of a multiresolution
analysis is not the full dyadic lattice, but only half of it, namely (we put jo = 0 for
simplicity) + = {(k2 j , 2 j ), k Z, j = 0, 1, 2, . . .}. Then + is the semidirect product
17.3 Frames for Affine Semidirect Products 521

+ = Z  A of Z by the abelian multiplicative semigroup A = {2 j , j = 0, 1, 2, . . .}.

This point of view turns out to be extremely fruitful, as we have seen in detail in
Sect. 13.5.

17.3 Frames for Affine Semidirect Products

17.3.1 The Affine WeylHeisenberg Group

As mentioned in Sect. 16.1, the matching pursuit algorithm [459] is based on a

suitable discretization of the natural UIR (16.1) of the affine WeylHeisenberg
group GaWH , leading to a particular dictionary of time-frequency atoms. The
rationale of this approach is the following. An efficient representation of signals is
obtained by expanding them into a basis that reflects their essential characteristics.
For instance, in 1-D, a sharp impulse will be well decomposed into functions well
concentrated in time (or space), whereas spectral lines are better represented by
waveforms which have a narrow frequency support (the same argument is used in
selecting a wavelet for a given type of signals). When the signal contains both of
those elements, no single representation (frame, basis) will suffice completely, be it
Fourier, Gabor or wavelets. They are simply not flexible enough. Thus one can use
instead [459] time-frequency atoms {h , }, where (b j , k , am ) runs over
a countable subset of GaWH . In other words, one needs a suitable discrete family of
CS of GaWH .
For particular choices of , the family reduces to Gabor or wavelet frames. For
instance, if we fix the scale parameter at ao and set = mn = (nqo , mo , ao ), m, n
Z, we obtain a Gabor frame as in (17.3). On the other hand, choosing = jk =
(kbo aoj , a
o o , ao ), j, k Z, one gets the wavelet frame or basis (17.9). In that
j j

sense, time-frequency atoms truly interpolate between Gabor and wavelet frames,
as already noted in [587, Tor95].
This being said, the key idea of the matching pursuit algorithm is to choose
the successive vectors in the optimal family in a recursive and adaptive way, by
projecting the function to be analyzed on the linear span of the subfamily obtained at
the previous step and minimizing, in the sense of the norm, the remaining difference.
This algorithm converges and yields a very flexible and economical method of
representing signals and images that contain simultaneously a large spectrum of
conflicting characteristics.

17.3.2 The Affine Poincar Groups

We have seen in Sect. 16.2.4 of the preceding chapter that the (spin 0) represen-
tation U of the affine Poincar group in n + 1 dimensions, SIM(n, 1), acting in
522 17 The Discretization Problem: Frames, Sampling, and All That

H = L2 (Cn , dm(k)) and given in (16.39), is square integrable, with admissibility

condition (16.43). Starting from this condition, a family of discrete tight frames for
SIM(n) can be obtained [160], that we now describe. As in the case of the affine
WeylHeisenberg group, the corresponding lattice Cn is generated by dyadic
translations and Lorentz transformations, that is, operations of the underlying group.
Take first n = 1. Let u = u(k0 , k) be a real continuous function, supported in a
square of side b1
o contained in the cone C1 = {k R : k > 0, k0 > 0}. In order to
2 2

discretize the action of SIM(1, 1), fix a scale ao > 1 and choose a sequence of real
numbers {tl , l Z} such that

[u(aoj l k)]2
a2o j k2
= 1, k C1 , (17.12)
l, jZ

where
 
coshtl sinhtl
l = SOo (1, 1), l = l1 .
sinhtl coshtl

Using these ingredients, define the following countable set of wavelets:

j
ul jmn (k) = ei(ao l k) u(aoj l k), l, j Z, = (m, n)bo bo Z2 . (17.13)

Then an explicit calculation shows that the family {ul jmn , l, j, m, n Z} is a tight
frame whenever the function u satisfies the condition (17.12). Moreover one has:

|ul jmn | f |2 = b2

o f .
2
(17.14)
l, j,m,nZ

Notice that the factor b2

o is the area of the support of u.
In order to obtain explicit functions satisfying the condition (17.12), it suffices to
introduce hyperbolic coordinates

k0 = cosht, k = sinht, > 0, t R,

and to take a separable function u(k) u( ,t) = u1 ( ) u2 (t). Then

u(aoj l k) = u(aoj ,t tl ) = u1 (aoj ) u2 (t tl ),

and it is enough to choose C functions of compact support

v1 ( ) = 2 [u1 ( )]2 , v2 (t) = [u2 (t)]2 ,

such that
17.3 Frames for Affine Semidirect Products 523

v1 (aoj ) = 1, v2 (t tl ) = 1.
jZ lZ

It should be noted that these discrete wavelets are not exactly the discretized version
of the CS associated with the unitary representation (16.39) of the affine Poincar
group; they correspond rather to the discretization of the action
1 k]b  
 a, ) f ](k) = ei[(a )
[U(b, f (a )1 k ,
(b, a, ) SIM(1, 1), f L2 (C1 , dm(k)), (17.15)

that is,

 a, ) = U(0, a, ) U(b, 1, I).

U(b,

However this action differs from the true representation (16.39),

U(b, a, ) = U(b, 1, I) U(0, a, ),

only by a phase:

 a, ) f ](k) = eik(a
[U(b,
1 bb
) [U(b, a, ) f ](k).

Then an explicit calculation, similar to the one made in Sect. 10.1.1, shows that
this action is square integrable as well, only the admissibility condition is slightly
 a, ) generated by the action of U
modified. Actually the states U(b,  on an
admissible vector form a tight frame, and the result just discussed means this
remains true upon discretization. Strictly speaking, of course, these states are not
coherent states, but weighted CS, are described in Sect. 7.3, but they play the
same rle.
Exactly the same construction may be performed in dimension n > 1. One starts
again from a real valued continuous function u supported in an (n + 1)-dimensional
hypercube of side b1o , contained in the forward cone Cn . One fixes a scale ao > 1
and chooses a countable family of Lorentz matrices, {l SOo (n, 1), l L}, with
L a countable index set, such that

[u(aoj l1 k)]2
(n+1) j 2 (n+1)/2
= 1, k Cn , (17.16)
jZ, lL ao (k )

Next one introduces in the cone Cn hyperbolic coordinates ( , ), > 0, V1+ =

{k = (k0 , k) Rn+1 : k2 = k02 k2 = 1, k0 > 0}, and writes u(k) = u1 ( ) u2 ( ). If
the functions

v1 ( ) = (n+1) [u1 ( )]2 , v2 ( ) = [u2 ( )]2 ,

524 17 The Discretization Problem: Frames, Sampling, and All That

verify the identities

v1 (aoj ) = 1, v2 (l1 ) = 1, (17.17)

jZ lL

then the function u satisfies the condition (17.16).

In that case, the family of wavelets
j 1
ul j (k) = eiao (l k )
u(aoj l1 k), l L, j Z, bo Zn+1 , (17.18)

is again a tight frame and one has:

|ul j | f |2 = bo
(n+1)
 f 2 . (17.19)
l, j,

Here too, the factor on the r.h.s. is the volume of the support of u. The only difficulty
is to find a suitable function v2 , and the solution given in [160] is based on the
existence of a discrete subgroup of SOo (n, 1) such that the quotient SOo (n, 1)/
is compact. Then V1+ / is compact as well, so that it is possible to find a compact
subset (a cube or a ball) X of V1+ such that

V1+ = (X ).

Hence, if w is a continuous function with support in X , then the function

w( )
v2 ( ) =
w( 1 )

satisfies the required identity (17.17).

However this result is in fact general, and the technique developed in [160]
extends immediately to a whole class of semidirect product groups, as we now show.

17.3.3 Discrete Frames for General Semidirect Products

The affine Poincar group SIM(n, 1) = Rn  (R+ SOo (n, 1)) is a particular case of
a semidirect product of the type G = V  S, with V = Rn and S = R+ K, where R
+

corresponds to dilations and K is a semisimple connected Lie group. This class was
treated, in increasing generality, in [41, 42, 146, 237, 292], and we have discussed
it at length in Chaps. 8 and 9. As we have seen there, the DufloMoore operator C
is, in all cases, a multiplication operator by a positive function C(k), which may be
bounded or unbounded, and similarly for the inverse C(k)1 . Thus, in general we
get CS systems, but not necessarily frames. However, as shown in [41,42], these CS
17.3 Frames for Affine Semidirect Products 525

systems always contain a discrete frame, which may be tight or non-tight. The proof
is a direct generalization of that used in [160] for the affine Poincar group, and we
sketch it, following mainly [374].
The key step is the existence of a suitable discrete subgroup of S. As in ,
Sect. 8.3.2, we consider the square integrable unitary representation U of G induced
by Go = Rn  So , where So K S is the compact stabilizer of a fixed point
ko R  n . Let O be the orbit of ko , which necessarily has positive Lebesgue measure

in R . Then the basic topological facts underlying the construction of frames are
n

summarized in the following lemma.

Lemma 17.3.1.
(i) There exists a discrete subgroup S = {s , } of S and a compact subset
K of O such that

O = s (K );

(ii) For any nonempty subset X O,  there exists in S a finite set SJ = {s , J},
with J a finite index set, such that:

O = s s (X ).
, J

The connection between the statements (i) and (ii) of the lemma is that {s (X ),
J} is a finite covering of K , which always exists since the latter is compact.
In the most general case, the representation U = U(b, s), b Rn , s S, is induced
by a finite dimensional cyclic representation L of the compact group So K, of
dimension N. We consider first the case N = 1, that is, the equivalent of the spin
0 case for the Poincar group. In other words, L is the trivial representation of So .
Then one proceeds exactly as in the Poincar case.
Let X be an open subset of O,  a function in the
 containing ko , and L2 (O)
representation space of L, with support in X . Let SJ = {s , J} be the finite
subset of S determined by Lemma 17.3.1, and let bo be the side of the smallest
hypercube in O containing X . Then a straightforward calculation shows that the
vectors

= U(0, s s )U( , I) , , J, bo Zn , (17.20)

form a discrete frame (which may be taken as a tight frame by properly adjusting
the constants). Explicitly, the CS (or wavelets) (17.20) are given by the following
functions:
1 k]
(k) = ei[(s s ) u((s s )1 k), , J, bo Zn . (17.21)
526 17 The Discretization Problem: Frames, Sampling, and All That

Clearly these wavelets reduce to those of the affine Poincar group, (17.18), if we
take K = SOo (n, 1) and s = a in the definition of the group G .
In the general case, the inducing representation L of So is a finite dimensional
cyclic representation, of dimension N, acting in a Hilbert space K, and the induced
representation U of G is given generically by the form [see (9.80)]:

(U(b, s) )(k) = eikb L([h0 (s1 , k)]1 ) (h1 k),

where h0 (s, k) denotes the element of So (the cocycle) associated to s in the inducing
construction (see Sect. 9.2.4). In that case, the construction of a frame goes through
in exactly the same way. The only difference is that one has to take care of the
action of So on the support of the generating function . More precisely, given a
cyclic vector v K for V , there exists N elements {ti So , i = 1, . . . , N}, such that
{V (ti )v, i = 1, . . . , N} is an algebraic basis of K. Then one shows that, for any w K,
there are two positive constants m, M, such that, for all k in a neighborhood of ko ,

N
m wK |L([h0 (ti1 , k)]1 )v|w|2 M wK . (17.22)
i=1

From this relation, one then deduces, by the same explicit calculation as before, that
the family of CS

i = U(0, s s ti )U( , I) , , J, i = 1, . . . , N, bo Zn , (17.23)

is a frame. Whether this frame can be made tight has to be decided in a specific case,
in terms of the constants m, M, appearing in (17.22).
The remarkable fact about this result is that all groups of the type considered,
namely semidirect products of the form G = Rn  (R+ K), are amenable to the
same conclusion. In all cases, discrete frames may be constructed, by the same
discretization method, and this obviously is an important point for applications.

17.4 Groups Without Dilations: The Poincar Groups

For the affine groups discussed in the previous sections, the presence of dilations
was the crucial factor that made the relevant representation square integrable. If
we drop them, we are forced back to CS of the general type. The most important
of these are the CS of the relativity groups. Explicit results have been obtained
for the Poincar group P+ (1, 1) in one time and one space dimension, as well as
for the more physical P+ (1, 3) and we will survey them in this section. As will be
clear from the sequel, the technique for discretizing the CS is essentially the same.
However, the new feature here is the need of a section : P+ (1, 1)/T P+ (1, 1),
resp. : P+ (1, 3)/T P+ (1, 3). As we have seen in Sect. 10.1, the nature of the
17.4 Groups Without Dilations: The Poincar Groups 527

continuous frames depends on the choice of the section. Naturally we expect the
same behavior for the discrete frames, and this is indeed the case. More remarkably,
the results for the three standard sections are exactly the same for P+ (1, 1) and
P+ (1, 3).

17.4.1 The Poincar Group P+ (1, 1)

Let us consider first the Poincar group P+ (1, 1). Discrete frames for this group
have been obtained in [402, 403], which we describe now (see also [55] for a more
elaborate version). From the Wigner representation (10.65), the CS of P+ (1, 1) for
an arbitrary section , indexed by the function #, take the form:

(q,p) (k) = (U( (q, p) ))(k) = eik.q ( p1 k) = ei Xp (k)q ( p1 k), (17.24)

where

Xp (k) = k ( p1 k)0 #(p), (17.25)

and is an admissible vector in H = L2 (Vm+ , dk/k0 ), that is,


| (k) |2 dk < . (17.26)
R

We proceed now to discretize these CS. For n Z, let pn = (pn0 , pn ) be a

discretization of p. Then for an arbitrary section (q, p) and n, l Z, we write
the discretized version of the coherent states in (17.24) as

n,l (k) = eiXl (k) qn,l (n1 k), qn,l = l qn , (17.27)

where qn > 0 is to be fixed later, Xn (k) and n are respectively the discretized
forms of Xp (k) and p . Let

a = (a0 , a) Vm+ and b = (b0 , b) Vm+ , with a < b, (17.28)

and suppose that (k) = 0 if k

/ [a, b], i.e., the length of the support of (k) is
(b a). Then (n1 k) = 0 if

2 3 a0 pn + apn0 b0 pn + bpn0
/ n a, n b =
k ,
m m

and the length of the support of (n1 k), as a function of k, is

528 17 The Discretization Problem: Frames, Sampling, and All That

[(b0 a0 )pn + (b a)pn0 ]/m.

Let (n1 k) =
 (Xn ), where Xn = Xn (k) is a function of k, and

1
bn := n b = (b0 pn0 + bpn , b0 pn + bpn0 ) ,
m
1
an := n a = (a0 pn0 + apn , a0 pn + apn0 ) .
m

 (Xn ) is given by
Thus, the length of the support of

1
Ln = Xn (bn ) Xn (an ) = [(b0 a0 )pn + (b a)pn0 ] (b0 a0 )#(pn ), (17.29)
m

where #(pn ) is the discretized form of #(p).

Let us now formally define the discretized frame operator,

T= |n,l n,l |. (17.30)
n,l=

To examine the convergence of the operator sum in (17.30), we proceed as in

Sect. 10.1.1 and consider, for arbitrary , H, the sum

I , =  | T  =  |n,l n,l | 
n,l=

 dk dk
= Vm+ Vm+
e[i{Xn (k)Xn (k )}qn,l ] (k) (n1 k) (n1 k ) (k )
k0 k0
.
n,l=
(17.31)

 (Xn ) obtained
Writing qn = 2 /Ln , where Ln is the length of the support of
in (17.29), we can write

 1
I , = Vm+ Vm+
e[i{Xn (k)Xn (k )}2 nLn ] (k)
 (Xn (k))
n,l=

dk dk
 (Xn (k )) (k )
. (17.32)
k0 k0

Since
1
 
dXn (k ) = k ( 1 
k ) #(pn ) dk , (17.33)
k0 0 n
17.4 Groups Without Dilations: The Poincar Groups 529

changing variables k  Xn (k ) yields

 
 1 ] 1
I , = dXn
Vm+ l=
e[i{Xn (k)Xn (k )}2 lLn
k0 (n1 k)#(pn )
n=

dk
 (Xn (k )) (k) (k )
 (Xn (k)) . (17.34)
k0

Using the relation (Poisson formula)

1 (xx )
ei2 nL = L (x x ), (17.35)
n=

we get
 
Ln
I , = dXn
Vm+
(Xn (k) Xn (k )) 1
k0 (n k)#(pl )
n=

dk
 (Xl (k )) (k) (k )
 (Xn (k)) . (17.36)
k0

Assuming that the -integration can be performed, the above expression may be
rewritten as
 ! "
Ln 2 dk
I , =  |T  =
(k) (k) 1
| (n
1
k)| .
n= k0 (n k) #(pn )
+ k0
Vm
(17.37)
Thus, like its continuous counterpart, T is a multiplication operator:

(T ) (k) = T (k) (k), (17.38)

with

Ln
T (k) = 1
| (n1 k)|2 , (17.39)
n= k0 (n k)#(pn ))

Hence the boundedness and invertibility of T depend on the boundedness and strict
positivity of the function T (k).
For t, n , a , b R, let

k0 = m cosht, k = m sinht, pn0 = m cosh n , pn = m sinh n ,

(17.40)
a0 = m cosh a , a = m sinh a , b0 = m cosh b , b = m sinh b .

Now substituting (17.29) and (17.40) in (17.39), and writing T (k) T (t), we obtain,
530 17 The Discretization Problem: Frames, Sampling, and All That

 
b a cosh(( b +2 a ) + (n ))
T (t) = 2 sinh
2 cosh(t n + (n )) | (t n )|2 , (17.41)
n=

where we have written

(pn ) (n ) = tanh (n ), (17.42)

with the dual speed (10.68) characterizing the section , and an arbitrary
(continuous) real function.
Let now n = n0 , where 0 > 0 (fixed). Then T takes the form:
 
b a cosh(( b +2 a ) + (n0 ))
T (t) = 2 sinh | (t n0 )|2 . (17.43)
2 n= cosh(t n0 + (n0 ))

For any compactly supported function , with support length L, the sum
in (17.43) contains at best L/0 + 1 terms, i.e., it is a finite sum. We observe that
each term in the sum is positive and bounded for any t R. Consequently, the
function T (t) is bounded and it is strictly positive, provided that 0 < L, i.e., that
the supports of any two consecutive terms in (17.43) overlap. Then the operator T
is bounded with bounded inverse. Thus, the discretized version n,l of the coherent
states in (17.24) form a discrete frame. Depending on the choice of and , this
frame could be tight or non-tight. In fact, the frame is tight iff T (t) is a constant
function.
We analyze the situation for the three specific choices of the section described
in Sect. 10.1.2.
1. The Galilean section 0 :
For this section, #(n ) = 0, which implies (n0 ) = n0 ; hence we can write
   
2 sinh b 2 a b + a
T (t) =
cosht cosh 2
+ n0 | (t n0 )|2 . (17.44)
n=

For different values of t, T (t) is different, so in this case the frame is never tight.
By contrast, in the corresponding continuous case, the frame is tight for certain
functions (see Sect. 10.1.2).
2. The Lorentz section  :
For this section #(n ) = sinh n , so that (n0 ) = 0 and
   
b a b + a | (t n0 )|2
T (t) = 2 sinh cosh . (17.45)
2 2 n= cosh(t n0 )

If | (t n0 )|2 = cosh(t n0 ) on the support of , that is, for

sinh a t n0 sinh b ,
17.4 Groups Without Dilations: The Poincar Groups 531

and zero outside, then

   
b a b + a
T (t) = 2 sinh cosh N(a, b) (17.46)
2 2

a b
where N(a, b) = 1 + . tsinh
0 / 0 tsinh
0 1 is the number of terms contributing
to the sum (17.43). Note that this is a constant function only if L = m0 , m N,
and then N(a, b) = m; otherwise N(a, b) is integer-valued, but nonconstant. Here
we have used

. m / = mo if mo m < mo + 1,
for some mo N. (17.47)
0 m 1 = mo + 1 if mo < n mo + 1,

Thus T (t) is a constant function, and consequently the frame is tight.

3. The symmetric section s :
sinh n n0
For this section #(n ) = , hence (n0 ) = and
1 + cosh n 2
 
b a cosh( b +2a +n0 )

T (t) = 2 sinh n 0 | (t n0 )|2 . (17.48)
2 n= cosh(t 2 )

In this case, T (t) is not a constant function and so the frame is never tight.
Note that if is a constant function, say, (n0 ) = , then
   
b a b + a | (t n0 )|2
T (t) = 2 sinh cosh + . (17.49)
2 2 n= cosh(t n0 + )

Here again, if we have | (t n0 )|2 = cosh(t n0 + ) on the support of

(t n0 ), T (t) becomes a constant operator and hence the frame is tight.
For a general section (characterized by the function ), the discretized
coherent states [see (17.24)] are
 
sinh(t n0 + (n0 ))
n,l (t) = exp i l (t n0 ) (17.50)
sinh( b 2 a ) cosh( b +2 a + (n0 ))

In that case, we can write the reconstruction formula, for any H,

(t) = n,l (t) |  T (t)1 | n,l (t). (17.51)
n,l=
532 17 The Discretization Problem: Frames, Sampling, and All That

17.4.2 The Poincar Group P+ (1, 3)

Next we turn to the Poincar group P+ (1, 3). We have seen in Sect. 10.1 that results
concerning continuous CS frames are essentially the same as for P+ (1, 1), for all
three standard sections. Thus one should expect to have the same situation in the
discrete case, and this is exactly what has been shown in [55]. Since the procedure
follows closely the one developed above in the 1+1 case, we will not give too many
details and refer the reader to the original paper instead.
As in Sect. 10.1.1, we consider the general case of mass m > 0 and spin
s 0, using the Wigner realization acting in the Hilbert space HW s = C2s+1
+ 1
L (Vm , k0 dk) given in (10.4). Then we define, for j = 1, . . . , 2s + 1 and for any
2

pair (n, l) Z Z3 , with n Z, l = (l ) Z3 ,

j
1. the position-momentum variables (qn,l , pn ) = P+ (1, 3)/(T SU(2)):

j
qn,l = ( qn,j l ) , qn,j > 0 , = 1, 2, 3
(17.52)
pn = (pn, ), = 1, 2, 3

[so the position variable lives on the regular lattice with unit cell ( qn,j )],
2. the momentum 4-vector
1
pn = (pn,0 , pn ) with pn,0 = m2 + pn 2 , (17.53)

3. the associated Lorentz matrix obtained from (10.8) by replacing k by pn .

Let us consider functions j HWs built on the model (10.34) and verifying the

rotation invariance condition (10.33). Then our coherent states are the discretized
version of the vectors (10.27), namely, the vectors

n,l
j
(k) = [UWs ( (qn,l
j
, pn )) j ](k) n Z, l Z3 ,
j
iXn (k) qn,l
=e D s (v(k, pn )) j ( (pn )1 k), (17.54)

Xn (k) and v(k, pn ) being the expressions obtained from (10.28) and (10.29) when
we substitute pn for p.
As in the 1+1 dimensional case, we look for conditions under which the discrete
set of coherent states (17.54) is a frame, tight or not. Here again, the key is to
consider the formal (frame) operator

2s+1
Ts = |n,lj n,lj |, (17.55)
j=1 nZ lZ3
17.4 Groups Without Dilations: The Poincar Groups 533

and study its convergence properties. Thus, for and in HW

s , we have to study the

convergence of the sum

2s+1
I s ( , ) =  | T s  =  | n,lj n,lj | . (17.56)
j=1 nZ lZ3

In order to proceed, one has to distinguish two cases, differing by the shape
of the supports of the j s: (1) Images of parallelepipedic boxes of R3 under the
homeomorphism R3 Vm+ ; and (2) images of spherical shells (including balls)
in R3 . For short, we will say that the supports are parallelepipedic boxes or spherical
shells.
Case 1: The supports of the j s are parallelepipedic boxes
We first assume that, for each j = 1, 2, . . . , 2s + 1, j is supported on an open
set S j Vm+ , which is the image of the open parallelepiped (a j , b j ) := {x R3 :
0 < aj < x < bj , = 1, 2, 3}, of R3 (i.e., a coordinate box), with closure S j .
For simplicity, the box (a j , b j ) is put in the first octant, but similar results may be
obtained for all the other positions, so that the analysis covers the case where the
support of each j is an arbitrary parallelepipedic box in R3 , and in fact any compact
and connected set. For this situation, the following result is obtained in [55].
Proposition 17.4.1. For j = 1, 2, . . . , 2s + 1, let
(1) (a j , b j ) R3 be an elementary nonvoid open parallelepiped, with closure
[a j , b j ] = S j such that

b0j b3j = a0j a3j and ma,> < mb,< where

2s+1 2s+1
ma,> := max a j , mb,< := min b j ,
j=1 j=1

1/2
(2) be a function from D(P0 ) HW0 satisfying the rotation invariance prop-
erty (10.33), and
(3) j = e j HWs be a function with support S j .
Then, given an arbitrary affine space-like (Borel) section : P+ (1, 3) and any
j j
discretization grid (qn,l , pn ), where qn,l is defined as in (17.52) and the 3-momenta
pn satisfy the inequality
 1 1
1 1
pn  m sinh ( 2 sinh
1
(mb,< m + ma,> ma,> m + mb,< ) ), (17.57)
2 2 2 2
m2
the set of vectors

{n,l
j
:= UWs ( (qn,l
j
, pn )) j : l Z3 , n Z, j = 1, 2, . . . , 2s + 1}

constitutes a discrete frame for the Poincar group P+ (1, 3).

534 17 The Discretization Problem: Frames, Sampling, and All That

Of course, the resulting frame may be tight or not, depending on the section .
The proof of this proposition follows the same line as in the (1+1)-dimensional
case, but it is of course much more complicated. For the full details, we refer to the
original article [55]. In particular, it is advisable to treat separately the simpler case
of spin s = 0 and the general case s > 0.
The next step is to particularize the result to the three standard sections
: P+ (1, 3), Galilean, Lorentzian and symmetric, discussed in Sect. 10.1
and characterized by the corresponding functions (p), (p) or #(p), given,
respectively, in (10.20)(10.22). It turns out that the results are exactly the same
as in 1+1 dimensions, as described in Sect. 17.4.1.
Case 2: The supports of the j s are spherical shells
We will now consider the rotation invariant situation, which amounts to replace
the parallelepipeds of the previous case by spherical shells in R3 . Thus, we
assume that
1. the j s are supported in spherical shells:

j (k) = 0 if k
/ [m sinh r1j , m sinh r2j ] k0
/ [m cosh r1j , m cosh r2j ]

for some r1j , r2j R+ , j = 1, . . . , 2s + 1 and r1,> < r2,< , where r1,> := max{r1j }
j
and r2,< := min{r2j } (if r1j = 0, the support of j is a ball, of course);
j
2. the j s are of the form (10.34) with verifying both (10.33) and the relation

| ( k)|2 = | (k)|2 , SO(3). (17.58)

Then the second main result may be summarized in the following proposition,
which closely follows Proposition 17.4.1.
Proposition 17.4.2. For j = 1, 2, . . . , 2s + 1, let
(1) (r1j , r2j ) R+ be an elementary nonvoid open interval, with closure [r1j , r2j ] = S j
such that

r1,> < r2,< where r1,> = max{r1j } and r2,< = min{r2j },

j j

1/2
(2) be a function from D(P0 ) HW0 satisfying the rotation invariance proper-
ties (10.33) and (17.58)
(3) j = e j HWs be a function with support S j .
Then, given an arbitrary affine space-like (Borel) section : P+ (1, 3) and any
j j
discretization grid (qn,l , pn ), where qn,l is defined as in (17.52) and the 3-momenta
pn satisfy the inequality
17.4 Groups Without Dilations: The Poincar Groups 535

Table 17.1 Type of Poincar frames obtained in the various cases, both in 1+1 or 1+3
dimensions (from [55])
Type of frames
Galilean Lorentzian Symmetric
Distribution section 0 section  section s
Continuous Tight frame Frame Frame
under particular always never
conditions tight tight
Discrete Frame Tight frame Frame
never under particular never
tight conditions tight

21 3
pn  m sinh 2 (r2,< r1,> ) ),

the set of vectors

{ n,l
j
:= UWs ( (qn,l
j
, pn )) j : l Z3 , n Z, j = 1, 2, . . . , 2s + 1}

constitutes a discrete frame for the Poincar group P+ (1, 3).

Here too, the detailed proof may be found in [55]. And, once again, the particular
results obtained for the three usual sections are the same as before.
Thus the outcome of the discussion is that the P+ (1, 3) frames obtained for each
section are of the same nature, whether the spin is zero or not. More striking yet,
the results are exactly the same for the two groups P+ (1, 1) and P+ (1, 3)! On the
other hand, in either case, the type of the frames changes from the continuous to the
discrete distribution, both for the Galilean and the Lorentzian sections, but it does
not for the symmetric section. The reason of these differences is unknown so far.
For the convenience of the reader, we have collected in Table 17.1 all the situations
observed for the various sections used, both in the discrete and in the continuous
cases.
In fact, this may vindicate one of the leitmotives of Eugene Wigner, accord-
ing to which the (1+1)-dimensional situation resembles much more the (3+1)-
dimensional one than the one in (2+1) dimensions. Indeed, several items point
in the same direction, namely, the different situations in odd vs. even number of
space dimensions. For instance, the form of the wave equation, the behavior of its
radial solutions (ordinary vs. spherical Bessel functions), the different structure of
the rotation groups SO(2n + 1), SO(2n), corresponding to the different types of Lie
algebras so(2n + 1) Bn , so(2n) Dn , and so on.
536 17 The Discretization Problem: Frames, Sampling, and All That

17.5 Generalities on Sampling

Discretization of the continuous wavelet transform, as for any CS transform, is

unavoidable for numerical reasons. On one hand, it leads to the theory of discrete
frames, and many examples have been given in this chapter. On the other hand,
purely discrete transforms are usually preferred in the digital world of signal
processing.
The signal processing approach to the discretization problem consists in the
sampling process, that is, evaluating a continuous signal on a discrete set of points,
actually, a Delone set. The goal is, of course, not to lose information in the process,
and criteria to that effect are given by various forms of Shannons sampling theorem.
However, there is basically only one answer, namely the density of sampling points
must be high enough.
Now, there are a priori two ways of sampling a signal. The most natural one is
to do it at regular intervals, tn = nT, n Z, where T1 is the sampling rate. As a
result, the sampled signal may retain some form of discrete translation invariance, at
least in a statistical sense (stationary signals). But, in practice, the sampling is often
irregular. Then, provided the sampling intervals in = tn+1 tn , are bounded below
(for instance, if the points tn constitute a Delone set), many results are known, and
this is a very active field of research [277, 342], in which number theory plays a
nonnegligible rle. However, since the group-theoretical approach is the backbone
of this volume, we will not discuss this topic of irregular sampling any further and
simply refer the reader to the literature [277, Fei98].
In the regular case, instead, we have seen in Chap. 13 that a group-theoretical
structure underlies the discrete WT. This is clear from the issues raised in the
treatment of wavelets on Z p = Z/pZ introduced in Sect. 13.4.1. Indeed these
wavelets on Z p give a fresh look to the problem, precisely because they import
some ideas from number theory. The same ideas have then be extended in Sect. 13.5
to wavelets on an arbitrary discrete abelian group. Here too, the (fast) algorithms
have a group-theoretical backbone, which opens the door to future generalizations.
Once more, group theory is a powerful unifying factor.
As a matter of fact, sampling is nowadays an active field of research, in particular
for the case of irregular sampling. For examples, one may refer to the compendiums
of BenedettoFerreira [Ben01] and Benedetto-Zayed [Ben04] or the very recent
collection [And13, vol. 1].
Chapter 18
Conclusion and Outlook

As must have become clear from the last few chapters of this book, research on
waveletsboth theoretical and appliedhas currently been gathering increasing
momentum, perhaps more so because of its numerous applications in todays
cutting-edge technology. The use of coherent states, as an applied and theoretical
tool in physical and mathematical research, is perhaps equally pervasive.

18.1 Present Status of CS and Wavelet Research

Let us just listof course, with no pretension of being exhaustivea number of

active directions in which work is presently moving, using as a guide the articles
contained in the special issue [38]. References to individual articles have been given
in the corresponding chapters.
(1) In physics:
Application of CS and squeezed states to various problems in atomic or
molecular physics, such as nonlinear oscillators, Morse quantum systems,
quantum particles in a magnetic-solenoid field, . . .
Analysis of various problems in dynamics, such as NambuGoldstone
dynamics, accelerated relativistic particles, action-angle variables, . . .
Developing systematic applications of time-frequency methods (e.g., Gabor
transforms, wavelets, WignerVille transforms) to laser-atom interactions
or in spectroscopy, in particular NMR spectroscopythis latter having
potential applications in medical physics.
Quantization: finite dimensional quantization, q-deformed CS in quantiza-
tion, integral quantization, . . .
Application of CS in quantum information theory: entanglement, intelligent
states, . . .
Application of CS in loop quantum gravity.

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 537
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3__18,
Springer Science+Business Media New York 2014
538 18 Conclusion and Outlook

Applications of wavelet analysis in geophysics, solar physics or astrophysics.

Applications of wavelet analysis in biophysics and psychophysicsfor
instance, in the analysis of EEG, ECG, as well as their use in medical
diagnostic techniques.
(2) At the interface between physics and mathematics:
The rle played by probabilistic aspects of coherent states and their various
generalizations in quantum physics and quantum probability
Application of CS in various particular problems, such as chaotic atomic
systems, quantum particle on a Mbius strip or a circle, supersymmetric CS,
semiclassical CS in periodic potentials, complexified CS with non-Hermitian
Hamiltonians, . . .
(3) In mathematics:
The use of coherent states associated to complex holomorphic structures,
and their applications to the study of the geometry of holomorphic vector
bundles. Equally, there is also the question of non-holomorphic extensions
of the theory.
The search for general square integrability results on homogeneous spaces
of groups and possible orthogonality relations in this setting.
Appearance of CS in various mathematical structures: pseudodifferential
operators and Weyl calculus, intertwining operators, extremal POV mea-
sures, Hilbert W*-modules Toeplitz operators, operator localization.
The interplay between nonlinear CS and families of orthogonal polynomials.
Wavelets on manifolds: extension of the work described in Sect. 15.1 to more
general (pseudo-)Riemannian manifolds, wavelets on graphs .
Development of shearlet theory and analogues, application to PDEs
(wavefront sets).
(4) In engineering and applied science:
Applications of time-frequency methods (Gabor, wavelets, WignerVille)
in mechanical systems, e.g., in the analysis of the behavior of materials
subjected to shocks, or for monitoring various types of installations.
Novel applications of continuous wavelets, in particular directional wavelets,
in imaging problems, e.g., watermarking of images (that is, inserting in an
image an invisible signature to guarantee its ownership), or detection and
modelling of (oriented) textures, a special case of shape recognition, with
applications in robotic vision.
Target detection: application of 3-D or space-time continuous wavelet
transforms for tracking and identification of targets.
Application of space-time wavelets to motion analysis, such as video
sequences, traffic control, . . .
18.2 What Is the Future from Our Point of View? 539

18.2 What Is the Future from Our Point of View?

In the spirit of the sort of mathematical analysis presented in this book, a number of
open areas of work are discernible:
Much work still remains to be done in the use of CS to study the quantization
process, both along the lines of geometric quantization including group repre-
sentation theory and also using Feynman path integralssubjects that we did not
have the occasion to touch upon in this book. In fact CS or integral quantization,
particularly in their probabilistic aspects, enables one to take a much wider
view of the quantization process than standard geometric quantization, or even
BerezinToeplitz quantization, opening up many more physical applications.
The extension of the theory of CS and wavelets to quaternionic Hilbert
spaces, exploiting recent work on the development of holomorphy theory
for quaternions. The expectation here is that representations of wavelet-like
transforms on quaternionic Hilbert spaces would enable one to deal efficiently
with multi-component wavelets. A second non-conventional approach is to build
coherent states on Hilbert C -modules, which are spaces analogous to Hilbert
spaces, but where the inner product takes values in a C -algebra. Such spaces
have been found useful in quantum stochastic processes and representations of
quantum groups.
A study of possible deformations of the various structures introduced here,
leading to q-deformed CS, CS for supersymmetric Lie algebras, p-adic CS or
wavelets, etc.
An adequate mathematical tool to analyze material structures using contact
microscopy still needs to be developed. In this kind of work a material probe
is brought and moved around in the immediate vicinity of the material surface
being studied and it would appear that a coherent state or wavelet type of
analysis would be far better adapted as a tool for such a study than the standard
Fourier transform, which is the natural mathematical tool for analyzing material
structures through diffraction experiments.
Discretization problems, in connection with the availability of very fast
computers, and parallelization of wavelet (pyramidal-like) algorithms.
Systematic study of decoherence and quantum nondemolition experiments using
coherent states.
Clearly, there is plenty of work waiting to be pursued, in all directions. The lesson
we might draw at this point is that group theory remains a valuable guide and a
source of mathematical inspiration, as well as a powerful unifying factor. Secondly,
that it always pays to develop mathematical tools with sufficient precision and
generality, not only for their intrinsic beauty, but also in view of the breadth of
applications which they afforda virtue that can be of immense practical value (and
indeed, economically promising as well). We conclude, therefore, by characterizing
the field that we have explored in this book as applied mathematical physics.
References

A. Books and Theses

[Abr97] P. Abry, Ondelettes et turbulences Multirsolutions, algorithmes de dcomposition,

invariance dchelle et signaux de pression (Diderot, Paris, 1997)
[Adl95] S.L. Adler, Quaternionic Quantum Mechanics and Quantum Fields (Oxford Univer-
sity Press, New York, 1995)
[Akh65] N.I. Akhiezer, The Classical Moment Problem and Some Related Questions in
Analysis (Oliver and Boyd, Edinburgh and London, 1965)
[Ala94] A. Alaux, Limage par rsonance magntique (Sauramps Mdical, Montpellier, 1994)
[Ald96] A. Aldroubi, M. Unser (eds.), Wavelets in Medicine and Biology (CRC Press, Boca
Raton, 1996)
[Ali07] S.T. Ali, M. Engli, Berezin-Toeplitz quantization over matrix domains, in Contribu-
tions in Mathematical Physics A tribute to Grard Emch, Ed. by S.T. Ali, K.B. Sinha
(Hindustan Book Agency, New Delhi, India, 2007)
[And13] T.D. Andrews, R. Balan, J.J. Benedetto, W. Czaja, K.A. Okoudjou (eds.), Excursions
in Harmonic Analysis, vol. 1, 2 (Birkhuser, Boston, 2013)
[Ant04] J-P. Antoine, R. Murenzi, P. Vandergheynst, S.T. Ali, Two-Dimensional Wavelets and
their Relatives (Cambridge University Press, Cambridge (UK), 2004)
[Ant09] J-P. Antoine, C. Trapani, Partial Inner Product Spaces Theory and Applications.
Lecture Notes in Mathematics, vol. 1986 (Springer, Berlin, Heidelberg, 2009)
[Arn95] A. Arnodo, F. Argoul, E. Bacry, J. Elezgaray, J.F. Muzy, Ondelettes, multifractales
et turbulences De lADN aux croissances cristallines (Diderot, Paris, 1995)
[Asc72] E. Ascher, Extensions et cohomologie de groupes. Lecture Notes, Enseignement du
troisime cycle de la physique en Suisse Romande (CICP) (1972)
[Bar77] A.O. Barut, R. Raczka, Theory of Group Representations and Applications (PWN,
Warszawa, 1977)
[Bec06] K. Becker, M. Becker, J.H. Schwarz, String Theory and M-Theory : A Modern
Introduction (Cambridge University Press, Cambridge, 2006)
[Ben01] J.J. Benedetto, P.J.S.G. Ferreira, Modern Sampling Theory, Mathematics and Appli-
cations (Birkhuser, Boston, Basel, Berlin, 2001)
[Ben04] J.J. Benedetto and A.I. Zayed, Sampling Theory, Wavelets, and Tomography
(Birkhuser, Boston, Basel, Berlin, 2004)
[Ben07] I. Bengtsson, K. Zyczkowski, Geometry of Quantum States: An Introduction to
Quantum Entanglement (Cambridge University Press, Cambridge, 2007)
[Ber66] S.K. Berberian, Notes on Spectral Theory (Van Nostrand, Princeton, NJ, 1966)

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 541
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3,
Springer Science+Business Media New York 2014
542 References

[Ber99] J.C. van den Berg (ed.), Wavelets in Physics (Cambridge University Press, Cambridge,
1999)
[Ber98] G. Bernuau, Proprits spectrales et gomtriques des quasicristaux. Ondelettes
adaptes aux quasicristaux. Thse de Doctorat, CEREMADE, Universit Paris IX
Dauphine, France, 1998
[Bie81] L. Biedenharn, J.D. Louck, The Racah-Wigner Algebra in Quantum Theory. Ency-
clopaedia of Mathematics, vol. 9 (Addison-Wesley, Reading, MA, 1981)
[Bis10] S. Biskri, Dtection et analyse des boucles magntiques solaires par traitement
dimages. Thse de Doctorat, UST Houari Boumedine, Alger, 2010
[Bog05] I. Bogdanova, Wavelets on non-Euclidean manifolds. Ph.D. thesis, EPFL, 2005
[Bog74] J. Bognar, Indefinite Inner Product Spaces (Springer, Berlin, 1974)
[Bor72] A. Borel, Reprsentations des groupes localement compacts. Lecture Notes in
Mathematics, vol. 276 (Springer, Berlin, 1972)
[Bou93] K. Bouyoucef, Sur des aspects multirsolution en reconstruction dimages: Applica-
tion au Tlescope Spatial de Hubble. Thse de Doctorat, Univ. P. Sabatier, Toulouse,
1993
[Bou97] A. Bouzouina, Comportement semi-classique de symplectomorphismes du tore quan-
tifis. Thse de Doctorat, Univ. Paris-Dauphine, 1997
[Bus91] P. Busch, P.J. Lahti, P. Mittelstaedt, The Quantum Theory of Measurement (Springer,
Berlin and Heidelberg, 1991)
[Can98] E.J. Cands, Ridgelets: Theory and applications. Ph.D. thesis, Department of Statis-
tics, Stanford University, 1998
[Chr03] O. Christensen, An Introduction to Frames and Riesz Bases (Birkhuser, Basel,
Boston, Berlin, 2003)
[Chu92] C.K. Chui, An Introduction to Wavelets (Academic, San Diego, 1992)
[Coh77] C. Cohen-Tannoudji, B. Diu, F. Lalo, Mcanique Quantique, Tome I (Hermann,
Paris, 1977)
[Com90] J.-M. Combes, A. Grossmann, P. Tchamitchian (eds.), Wavelets, Time-Frequency
Methods and Phase Space (Proc. Marseille 1987), 2nd edn. (Springer, Berlin, 1990)
[Com12] M. Combescure, D. Robert, Shape Analysis and Classification: Theory and Practice
(Springer, Dordrecht, Heidelberg, 2012)
[Cos01] L.F. Costa, R.M. Cesar Jr., Coherent States and Applications in Mathematical Physics
(CRC Press, Boca Raton, FL, 2001)
[Dau92] I. Daubechies, Ten Lectures on Wavelets (SIAM, Philadelphia, 1992)
[Dav90] E.B Davies, Heat Kernels and Spectral Theory (Cambridge University Press, Cam-
bridge, 1990)
[DeV88] R. De Valois, K. De Valois, Spatial Vision (Oxford University Press, New York, 1988)
[Dir01] P.A.M. Dirac, Lectures on Quantum Mechanics (Dover, New York, 2001)
[Dix64] J. Dixmier, Les C*-algbres et leurs reprsentations (Gauthier-Villars, Paris, 1964)
[Dod03] V.V. Dodonov, V.I. Manko (eds.), Theory of Nonclassical States of Light (Taylor and
Francis, London, New York, 2003)
[Duv91] M. Duval-Destin, Analyse spatiale et spatio-temporelle de la stimulation visuelle
laide de la transforme en ondelettes. Thse de Doctorat, Universit dAix-Marseille
II, 1991
[Fea90] J.-C. Feauveau, Analyse multirsolution par ondelettes non orthogonales et bancs de
filtres numriques. Thse de Doctorat, Universit Paris-Sud, 1990
[Fei98] H.G. Feichtinger, T. Strohmer (eds.), Gabor Analysis and Algorithms Theory and
Applications (Birkhuser, Boston-Basel-Berlin, 1998)
[Fei01] H.G. Feichtinger, T. Strohmer (eds.), Advances in Gabor Analysis (Birkhuser,
Boston, 2001)
[Fen94] D.H. Feng, J.R. Klauder, M. Strayer (eds.), Coherent States: Past, Present and Future
(Proc. Oak Ridge 1993) (World Scientific, Singapore, 1994)
[Fla98] P. Flandrin, Temps-Frquence (Herms, Paris, 1993); Engliah translation Time-
Frequency/Time-Scale Analysis (Academic, New York, 1998)
References 543

[Fol95] G.B. Folland, A Course in Abstract Harmonic Analysis (CRC Press, Boca Raton, FL,
1995)
[Fre97] W. Freeden, M. Schreiner, T. Gervens, Constructive Approximation on the Sphere,
with Applications to Geomathematics (Clarendon Press, Oxford, 1997)
[Fue05] H. Fhr, Abstract Harmonic Analysis of Continuous Wavelet Transforms. Lecture
Notes in Mathematics, vol. 1863 (Springer, Berlin, Heidelberg, 2005)
[Gaa73] S.A. Gaal, Linear Analysis and Representation Theory (Springer, Berlin, 1973)
[Gaz09] J-P. Gazeau, Coherent States in Quantum Physics (Wiley-VCH, Berlin, 2009)
[Gel64] I.M. Gelfand, N.Y. Vilenkin, Generalized Functions, vol. 4 (Academic, New York,
1964)
[Gen13] G. Gentili, C. Stoppato, D.C. Struppa, Regular Functions for a Quaternionic Variable.
Springer Monographs in Mathematics (Springer, Berlin, 2013)
[Gol81] H. Goldstein, C. Poole, J. Safko, Classical Mechanics, 3rd edn. (Addison-Wesley,
Reading, MA, 1981)
[Got66] K. Gottfried, Quantum Mechanics: Fundamentals, vol. I (Benjamin, New York and
Amsterdam, 1966)
[Gr01] K. Grchenig, Foundations of Time-Frequency Analysis (Birkhuser, Boston, 2001)
[Gun94] H. Gnther, NMR Spectroscopy, 2nd edn. (Wiley, Chichester, New York, 1994)
[Gui84] V. Guillemin, S. Sternberg, Symplectic Techniques in Physics (Cambridge University
Press, Cambridge, 1984)
[Hei06] C. Heil, D. Walnut (eds.), Fundamental Papers in Wavelet Theory (Princeton Univer-
sity Press, Princeton, NJ, 2006)
[Hel78] S. Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces (Academic,
New York, 1978)
[Hel76] C.W. Helstrom, Quantum Detection and Estimation Theory (Academic, New York,
1976)
[Her89] G. Herzberg, Molecular Spectra and Molecular Structure: Spectra of Diatomic
Molecules, 2nd edn. (Krieger Pub., Malabar, FL, 1989)
[Hil71] P. Hilton, U. Stammbach, A Course in Homological Algebra (Springer, Berlin, 1971)
[H0l01] A.S. Holevo, Statistical Structure of Quantum Theory (Springer, Berlin, 2001)
[Hol95] M. Holschneider, Wavelets, An Analysis Tool (Oxford University Press, Oxford, 1995)
[Hon07] G. Honnouvo, Gabor analysis and wavelet transforms on some non-Euclidean 2-
dimensional manifolds. Ph.D. thesis, Concordia University, Montreal, PQ, Canada,
2007
[Hua63] L.K. Hua, Harmonic Analysis of Functions of Several Complex Variables in the Clas-
sical Domains. Translations of Mathematical Monographs (American Mathematical
Society, Providence, RI, 1963)
[Hum72] J.E. Humphreys, Introduction to Lie Algebras and Representation Theory (Springer,
Berlin, 1972)
[In54] E. Inn, A study of the unitary representations of the Galilei group in relation to
quantum mechanics. Ph.D. thesis, University of Ankara, 1954
[Ino92] A. Inomata, H. Kuratsuji, C.C. Gerry, Path Integrals and Coherent States of SU(2)
and SU(1,1) (World Scientific, Singapore, 1992)
[Jac62] N. Jacobson, Lie Algebras (Interscience, New York and London, 1962)
[Jac04] L. Jacques. Ondelettes, repres et couronne solaire. Thse de Doctorat, Univ. Cath.
Louvain, Louvain-la-Neuve, 2004
[Jaf96] S. Jaffard, Y. Meyer, Wavelet Methods for Pointwise Regularity and Local Oscillations
of Functions. Memoirs of the American Mathematical Society, vol. 143 (American
Mathematical Society, Providence, RI, 1996)
[Kah98] J-P. Kahane, P.G. Lemari-Rieusset, Fourier Series and Wavelets (Gordon and Breach,
Luxembourg, 1995); French translation Sries de Fourier et ondelettes (Cassini, Paris,
1998)
[Kai94] G. Kaiser, A Friendly Guide to Wavelets (Birkhuser, Boston, 1994)
[Kat76] T. Kato, Perturbation Theory for Linear Operators (Springer, Berlin, 1976)
544 References

[Kem37] E.C. Kemble, Fundamental Principles of Quantum Mechanics (McGraw Hill, New
York, 1937)
[Kir76] A.A. Kirillov, Elements of the Theory of Representations (Springer, Berlin, 1976)
[Kla68] J.R. Klauder, E.C.G. Sudarshan, Fundamentals of Quantum Optics (Benjamin, New
York, 1968)
[Kla85] J.R. Klauder, B.S. Skagerstam, Coherent States Applications in Physics and
Mathematical Physics (World Scientific, Singapore, 1985)
[Kla00] J.R. Klauder, Beyond Conventional Quantization (Cambridge University Press, Cam-
bridge, 2000)
[Kla11] J.R. Klauder, A Modern Approach to Functional Integration (Birkhuser/Springer,
New York, 2011)
[Kna96] A.W. Knapp, Lie Groups Beyond an Introduction (Birkhuser, Basel, 1996; 2nd edn.
2002)
[Kut12] G. Kutyniok, D. Labate (eds.), Shearlets: Multiscale Analysis for Multivariate Data
(Birkhuser, Boston, 2012)
[Lan81] L. Landau, E. Lifchitz, Mechanics, 3rd edn. (Pergamon, Oxford,1981)
[Lan93] S. Lang, Algebra, 3rd edn. (Addison-Wesley, Reading, MA, 1993)
[Lie97] E.H. Lieb, M. Loss, Analysis (American Mathematical Society, Providence, RI, 1997)
[Lip74] R.L. Lipsman, Group Representations. Lecture Notes in Mathematics, vol. 388
(Springer, Berlin, 1974)
[Lyn82] P.A. Lynn, An Introduction to the Analysis and Processing of Signals, 2nd edn.
(MacMillan, London, 1982)
[Mac68] G.W. Mackey, Induced Representations of Groups and Quantum Mechanics (Ben-
jamin, New York, 1968)
[Mac76] G.W. Mackey, Theory of Unitary Group Representations (University of Chicago
Press, Chicago, 1976)
[Mad95] J. Madore, An Introduction to Noncommutative Differential Geometry and Its Physical
Applications (Cambridge University Press, Cambridge, 1995)
[Mae94] S. Maes, The wavelet transform in signal processing, with application to the extraction
of the speech modulation model features. Thse de Doctorat, Univ. Cath. Louvain,
Louvain-la-Neuve, 1994
[Mag66] W. Magnus, F. Oberhettinger, R.P. Soni, Formulas and Theorems for the Special
Functions of Mathematical Physics (Springer, Berlin, 1966)
[Mal99] S.G. Mallat, A Wavelet Tour of Signal Processing, 2nd edn. (Academic, San Diego,
1999)
[Mar82] D. Marr, Vision (Freeman, San Francisco, 1982)
[Mes62] H. Meschkowsky, Hilbertsche Rume mit Kernfunktionen (Springer, Berlin, 1962)
[Mey91] Y. Meyer (ed.), Wavelets and Applications (Proc. Marseille 1989) (Masson and
Springer, Paris and Berlin, 1991)
[Mey92] Y. Meyer, Les Ondelettes, Algorithmes et Applications (Armand Colin, Paris, 1992);
English translation Wavelets, Algorithms and Applications (SIAM, Philadelphia,
1993)
[Mey00] C.D. Meyer, Matrix Analysis and Applied Linear Algebra (SIAM, Philadelphia, 2000)
[Mey93] Y. Meyer, S. Roques (eds.), Progress in Wavelet Analysis and Applications (Proc.
Toulouse 1992) (Ed. Frontires, Gif-sur-Yvette 1993)
[Mur90] R. Murenzi, Ondelettes multidimensionnelles et applications lanalyse dimages.
Thse de Doctorat, Univ. Cath. Louvain, Louvain-la-Neuve, 1990
[vNe55] J. von Neumann, Mathematical Foundations of Quantum Mechanics (Princeton
University Press, Princeton, NJ, 1955) (English translated by R.T. Byer)
[Pap02] A. Papoulis, S.U. Pillai, Probability, Random Variables and Stochastic Processes, 4th
edn. (McGraw Hill, New York, 2002)
[Par05] K.R. Parthasarathy, Probability Measures on Metric Spaces (AMS Chelsea Publish-
ing, Providence, RI, 2005)
References 545

[Pau85] T. Paul, Ondelettes et Mcanique Quantique. Thse de doctorat, Univ. dAix-Marseille

II, 1985
[Per86] A.M. Perelomov, Generalized Coherent States and Their Applications (Springer,
Berlin, 1986)
[Per05] G. Peyr, Gomtrie multi-chelles pour les images et les textures. Thse de doctorat,
Ecole Polytechnique, Palaiseau, 2005
[Pru86] E. Prugovecki, Stochastic Quantum Mechanics and Quantum Spacetime (Reidel,
Dordrecht, 1986)
[Rau04] H. Rauhut, Time-frequency and wavelet analysis of functions with symmetry proper-
ties. Ph.D. thesis, TU Mnich, 2004
[Ree80] M. Reed, B. Simon, Methods of Modern Mathematical Physics I: Functional Analysis
(Academic, New York, 1980)
[Rud62] W. Rudin, Fourier Analysis on Groups (Interscience, New York, 1962)
[Sch96] F.E. Schroeck Jr., Quantum Mechanics on Phase Space (Kluwer, Dordrecht, 1996)
[Sch61] L. Schwartz, Mthodes mathmatiques pour les sciences physiques (Hermann, Paris,
1961)
[Scu97] M.O. Scully, M.S. Zubairy, Quantum Optics (Cambridge University Press, Cam-
bridge, 1997)
[Sho50] J.A. Shohat, J.D. Tamarkin, The Problem of Moments (American Mathematical
Society, Providence, RI, 1950)
[Ste71] E.M. Stein, G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces (Prince-
ton University Press, Princeton, NJ, 1971)
[Str64] R.F. Streater, A.S. Wightman, PCT, Spin and Statistics and All That (Benjamin, New
York, 1964)
[Sug90] M. Sugiura, Unitary Representations and Harmonic Analysis: An Introduction
(North-Holland/Kodansha Ltd., Tokyo, 1990)
[Suv11] A. Suvichakorn, C. Lemke, A. Schuck Jr., J-P. Antoine, The continuous wavelet
transform in MRS, Tutorial text Marie Curie Research Training Network FAST
(2011). http://www.fast-mariecurie-rtn-project.eu/#Wavelet
[Tak79] M. Takesaki, Theory of Operator Algebras I (Springer, New York, 1979)
[Ter88] A. Terras, Harmonic Analysis on Symmetric Spaces and Applications II (Springer,
Berlin, 1988)
[Tho98] G. Thonet, New aspects of time-frequency analysis for biomedical signal processing.
Thse de Doctorat, EPFL, Lausanne, 1998
[Tor95] B. Torrsani, Analyse continue par ondelettes (Interditions/CNRS ditions, Paris,
1995)
[Unt87] A. Unterberger, Analyse harmonique et analyse pseudo-diffrentielle du cne de
lumire. Astrisque 156, 1201 (1987)
[Unt91] A. Unterberger, Quantification relativiste. Mm. Soc. Math. France 4445, 1215
(1991)
[Van98] P. Vandergheynst, Ondelettes directionnelles et ondelettes sur la sphre. Thse de
Doctorat, Univ. Cath. Louvain, Louvain-la-Neuve, 1998
[Var85] V.S. Varadarajan, Geometry of Quantum Theory, 2nd edn. (Springer, New York, 1985)
[Vet95] M. Vetterli, J. Kovacevic, Wavelets and Subband Coding (Prentice Hall, Englewood
Cliffs, NJ, 1995)
[Vil69] N.J. Vilenkin, Fonctions spciales et thorie de la reprsentation des groupes (Dunod,
Paris, 1969)
[Wel03] G.V. Welland, Beyond Wavelets (Academic, New York, 2003)
[vWe86] C. von Westenholz, Differential Forms in Mathematical Physics (North-Holland,
Amsterdam, 1986)
[Wey28] H. Weyl, Gruppentheorie und Quantenmechanik (Hirzel, Leipzig, 1928)
[Wey31] H. Weyl, The Theory of Groups and Quantum Mechanics (Dover, New York, 1931)
[Wic94] M.V. Wickerhauser, Adapted Wavelet Analysis from Theory to Software (A. K. Peters,
Wellesley, MA, 1994)
546 References

[Wis93] W. Wisnoe, Utilisation de la mthode de transforme en ondelettes 2D pour lanalyse

de visualisation dcoulements. Thse de Doctorat ENSAE, Toulouse, 1993
[Woj97] P. Wojtaszczyk, A Mathematical Introduction to Wavelets (Cambridge University
Press, Cambridge, 1997)
[Woo92] N.J.M. Woodhouse, Geometric Quantization, 2nd edn. (Clarendon Press, Oxford
1992)
[Zac06] C. Zachos, D. Fairlie, T. Curtright, Quantum Mechanics in Phase Space: An Overview
With Selected Papers (World Scientific Publishing, Singapore, 2006)

B. Articles

[1] P. Abry, R. Baraniuk, P. Flandrin, R. Riedi, D. Veitch, Multiscale nature of network traffic.
IEEE Signal Process. Mag. 19, 2846 (2002)
[2] M.D. Adams, The JPEG-2000 still image compression standard.
http://www.ece.uvic.ca/~frodo/publications/jpeg2000.pdf
[3] S.L. Adler, A.C. Millard, Coherent states in quaternionic quantum mechanics. J. Math.
Phys. 38, 21172126 (1997)
[4] G.S. Agarwal, K. Tara, Nonclassical properties of states generated by the excitation on a
coherent state. Phys. Rev. A 43, 492497 (1991)
[5] G.S. Agarwal, E. Wolf, Calculus for functions of noncommuting operators and general
phase-space methods in quantum mechanics I: Mapping theorems and ordering of func-
tions of noncommuting operators. Phys. Rev. D 2, 21612186 (1970)
[6] G.S. Agarwal, E. Wolf, Calculus for functions of noncommuting operators and general
phase-space methods in quantum mechanics II: Quantum mechanics in phase space. Phys.
Rev. D 2, 21872205 (1970)
[7] G.S. Agarwal, E. Wolf, Calculus for functions of noncommuting operators and general
phase-space methods in quantum mechanics III: A generalized Wick theorem and multi-
time mapping. Phys. Rev. D 2, 22062225 (1970)
[8] V. Aldaya, J. Guerrero, G. Marmo, Quantization on a Lie group: Higher-order polariza-
tions, in Symmetry in Sciences X, ed. by B. Gruber, M. Ramek (Plenum Press, Nerw York,
1998), pp. 136
[9] M. Alexandrescu, D. Gibert, G. Hulot, J.-L. Le Mouel, G. Saracco, Worldwide wavelet
analysis of geomagnetic jerks. J. Geophys. Res. B 101, 2197521994 (1996)
[10] G. Alexanian, A. Pinzul, A. Stern, Generalized coherent state approach to star products
and applications to the fuzzy sphere. Nucl. Phys. B 600, 531547 (2001)
[11] S.T. Ali, A geometrical property of POV-measures and systems of covariance, in Differ-
ential Geometric Methods in Mathematical Physics, ed. by H.D. Doebner, S.I. Andersson,
H.R. Petry. Lecture Notes in Mathematics, vol. 905 (Springer, Berlin, 1982), pp. 20722
[12] S.T. Ali, Commutative systems of covariance and a generalization of Mackeys imprimi-
tivity theorem. Canad. Math. Bull. 27, 390397 (1984)
[13] S.T. Ali, Stochastic localisation, quantum mechanics on phase space and quantum space-
time. Riv. Nuovo Cim. 8(11), 1128 (1985)
[14] S.T. Ali, A general theorem on square-integrability: Vector coherent states. J. Math. Phys.
39, 39543964 (1998)
[15] S.T. Ali, J-P. Antoine, Coherent states of 1+1 dimensional Poincar group: Square
integrability and a relativistic Weyl transform. Ann. Inst. H. Poincar 51, 2344 (1989)
[16] S.T. Ali, S. De Bivre, Coherent states and quantization on homogeneous spaces, in Group
Theoretical Methods in Physics, ed. by H.-D. Doebner et al. Lecture Notes in Mathematics,
vol. 313 (Springer, Berlin, 1988), pp. 201207
References 547

[17] S.T. Ali, H.-D. Doebner, Ordering problem in quantum mechanics: Prime quantization and
a physical interpretation. Phys. Rev. A 41, 11991210 (1990)
[18] S.T. Ali, G.G. Emch, Geometric quantization: Modular reduction theory and coherent
states. J. Math. Phys. 27, 29362943 (1986)
[19] S.T. Ali, M. Engli, J-P. Gazeau, Vector coherent states from Plancherels theorem and
Clifford algebras. J. Phys. A 37, 60676089 (2004)
[20] S.T. Ali, M.E.H. Ismail, Some orthogonal polynomials arising from coherent states. J.
Phys. A 45 125203 (2012) (16pp)
[21] S.T. Ali, U.A. Mueller, Quantization of a classical system on a coadjoint orbit of the
Poincar group in 1+1 dimensions. J. Math. Phys. 35, 44054422 (1994)
[22] S.T. Ali, E. Prugovecki, Systems of imprimitivity and representations of quantum mechan-
ics on fuzzy phase spaces. J. Math. Phys. 18, 219228 (1977)
[23] S.T. Ali, E. Prugovecki, Mathematical problems of stochastic quantum mechanics: Har-
monic analysis on phase space and quantum geometry. Acta Appl. Math. 6, 118 (1986)
[24] S.T. Ali, E. Prugovecki, Extended harmonic analysis of phase space representation for the
Galilei group. Acta Appl. Math. 6, 1945 (1986)
[25] S.T. Ali, E. Prugovecki, Harmonic analysis and systems of covariance for phase space
representation of the Poincar group. Acta Appl. Math. 6, 4762 (1986)
[26] S.T. Ali, J-P. Antoine, J-P. Gazeau, De Sitter to Poincar contraction and relativistic
coherent states. Ann. Inst. H. Poincar 52, 83111 (1990)
[27] S.T. Ali, J-P. Antoine, J-P. Gazeau, Square integrability of group representations on
homogeneous spaces. I. Reproducing triples and frames. Ann. Inst. H. Poincar 55, 829
855 (1991)
[28] S.T. Ali, J-P. Antoine, J-P. Gazeau, Square integrability of group representations on
homogeneous spaces II: Coherent and quasi-coherent states. The case of the Poincar
group. Ann. Inst. H. Poincar 55, 857890 (1991)
[29] S.T. Ali, J-P. Antoine, J-P. Gazeau, Continuous frames in Hilbert space. Ann. Phys.(NY)
222, 137 (1993)
[30] S.T. Ali, J-P. Antoine, J-P. Gazeau, Relativistic quantum frames. Ann. Phys.(NY) 222, 38
88 (1993)
[31] S.T. Ali, J-P. Antoine, J-P. Gazeau, U.A. Mueller, Coherent states and their generalizations:
A mathematical overview. Rev. Math. Phys. 7, 10131104 (1995)
[32] S.T. Ali, J-P. Gazeau, M.R. Karim, Frames, the -duality in Minkowski space and spin
coherent states. J. Phys. A: Math. Gen. 29, 55295549 (1996)
[33] S.T. Ali, M. Engli, Quantization methods: A guide for physicists and analysts. Rev. Math.
Phys. 17, 391490 (2005)
[34] S.T. Ali, J-P. Gazeau, B. Heller, Coherent states and Bayesian duality. J. Phys. A: Math.
Theor. 41, 365302 (2008)
[35] S.T. Ali, L. Balkov, E.M.F. Curado, J-P. Gazeau, M.A. Rego-Monteiro, L.M.C.S.
Rodrigues, K. Sekimoto, Non-commutative reading of the complex plane through Delone
sequences. J. Math. Phys. 50, 043517 (2009)
[36] S.T. Ali, C. Carmeli, T. Heinosaari, A. Toigo, Commutative POVMS and fuzzy observ-
ables. Found. Phys. 39, 593612 (2009)
[37] S.T. Ali, T. Bhattacharyya, S.S. Roy, Coherent states on Hilbert modules, J. Phys. A: Math.
Theor. 44, 275202 (2011)
[38] S.T. Ali, J-P. Antoine, F. Bagarello, J-P. Gazeau (Guest Editors), Coherent states: A
contemporary panorama, preface to a special issue on Coherent states: Mathematical and
physical aspects. J. Phys. A: Math. Gen. 45(24) (2012)
[39] S.T. Ali, F. Bagarello, J-P. Gazeau, Quantizations from reproducing kernel spaces. Ann.
Phys. (NY) 332, 127142 (2013)
[40] S.T.Ali, K. Grska, A. Horzela, F. Szafraniec, Squeezed states and Hermite polynomials in
a complex variable. Preprint (2013), arXiv:1308.4730v1 [quant-phy]
[41] P. Aniello, G. Cassinelli, E. De Vito, A. Levrero, Square-integrability of induced represen-
tations of semidirect products. Rev. Math. Phys. 10, 301313 (1998)
548 References

[42] P. Aniello, G. Cassinelli, E. De Vito, A. Levrero, Wavelet transforms and discrete frames
associated to semidirect products. J. Math. Phys. 39, 39653973 (1998)
[43] J-P. Antoine, Remarques sur le vecteur de Runge-Lenz. Ann. Soc. Scient. Bruxelles 80,
160168 (1966)
[44] J-P. Antoine, Etude de la dgnrescence orbitale du potentiel coulombien en thorie des
groupes I, II. Ann. Soc. Scient. Bruxelles 80 169184 (1966)
[45] J-P. Antoine, Etude de la dgnrescence orbitale du potentiel coulombien en thorie des
groupes I, II. Ann. Soc. Scient. Bruxelles 81, 4968 (1967)
[46] J-P. Antoine, Dirac formalism and symmetry problems in quantum mechanics I: General
dirac formalism. J. Math. Phys. 10, 5369 (1969)
[47] J-P. Antoine, Dirac formalism and symmetry problems in quantum mechanics II: Symme-
try problems. J. Math. Phys. 10, 22762290 (1969)
[48] J-P. Antoine, Quantum mechanics beyond Hilbert space, in Irreversibility and Causality
Semigroups and Rigged Hilbert Spaces, ed. by A. Bhm, H.-D. Doebner, P. Kielanowski.
Lecture Notes in Physics, vol. 504, (Springer, Berlin, 1998), pp. 333
[49] J-P. Antoine, Discrete wavelets on abelian locally compact groups. Rev. Cien. Math.
(Habana) 19, 321 (2003)
[50] J-P. Antoine, Introduction to precursors in physics: Affine coherent states, in Fundamental
Papers in Wavelet Theory, ed. by C. Heil, D. Walnut (Princeton University Press, Princeton,
NJ, 2006), pp. 113116
[51] J-P. Antoine, F. Bagarello, Wavelet-like orthonormal bases for the lowest Landau level.
J. Phys. A: Math. Gen. 27 24712481 (1994)
[52] J-P. Antoine, P. Balazs, Frames and semi-frames. J. Phys. A: Math. Theor. 44, 205201
(2011) (25 pages); Corrigendum: J-P. Antoine, P. Balazs, Frames and semi-frames. J. Phys.
A: Math. Theor. 44, 479501 (2011) (2 pages)
[53] J-P. Antoine, P. Balazs, Frames, semi-frames, and Hilbert scales. Numer. Funct. Anal.
Optim. 33, 736769 (2012)
[54] J-P. Antoine, A. Coron, Time-frequency and time-scale approach to magnetic resonance
spectroscopy. J. Comput. Methods Sci. Eng. (JCMSE) 1, 327352 (2001)
[55] J-P. Antoine, A.L. Hohouto, Discrete frames of Poincar coherent states in 1+3 dimen-
sions. J. Fourier Anal. Appl. 9, 141173 (2003)
[56] J-P. Antoine, I. Mahara, Galilean wavelets: Coherent states for the affine Galilei group.
J. Math. Phys. 40, 59565971 (1999)
[57] J-P. Antoine, U. Moschella, Poincar coherent states: The two-dimensional massless case.
J. Phys. A: Math. Gen. 26, 591607 (1993)
[58] J-P. Antoine, R. Murenzi, Two-dimensional directional wavelets and the scale-angle
representation. Signal Process. 52, 259281 (1996)
[59] J-P. Antoine, R. Murenzi, Two-dimensional continuous wavelet transform as linear phase
space representation of two-dimensional signals, in Wavelet Applications IV. SPIE Pro-
ceedings, vol. 3078 (SPIE, Bellingham, WA, 1997), pp. 206217
[60] J-P. Antoine, D. Rosca, The wavelet transform on the two-sphere and related manifolds
A review, in Optical and Digital Image Processing. SPIE Proceedings, vol. 7000 (2008),
pp. 70000B-115
[61] J-P. Antoine, D. Speiser, Characters of irreducible representations of simple Lie groups.
J. Math. Phys. 5, 12261234 (1964)
[62] J-P. Antoine, P. Vandergheynst, Wavelets on the n-sphere and related manifolds. J. Math.
Phys. 39, 39874008 (1998)
[63] J-P. Antoine, P. Vandergheynst, Wavelets on the 2-sphere: A group-theoretical approach.
Appl. Comput. Harmon. Anal. 7, 262291 (1999)
[64] J-P. Antoine, P. Vandergheynst, Wavelets on the two-sphere and other conic sections. J.
Fourier Anal. Appl. 13, 369386 (2007)
[65] J-P. Antoine, M. Duval-Destin, R. Murenzi, B. Piette, Image analysis with 2D wavelet
transform: Detection of position, orientation and visual contrast of simple objects, in
Wavelets and Applications (Proc. Marseille 1989) ed. by Y. Meyer (Masson and Springer,
Paris and Berlin, 1991), pp.144159
References 549

[66] J-P. Antoine, P. Carrette, R. Murenzi, B. Piette, Image analysis with 2D continuous wavelet
transform. Signal Process. 31 241272 (1993)
[67] J-P. Antoine, P. Vandergheynst, K. Bouyoucef, R. Murenzi, Alternative representations of
an image via the 2D wavelet transform: Application to character recognition, in Visual
Information Processing IV. SPIE Proceedings, vol. 2488 (SPIE, Bellingham, WA, 1995),
pp. 486497
[68] J-P. Antoine, R. Murenzi, P. Vandergheynst, Two-dimensional directional wavelets in
image processing. Int. J. Imaging Syst. Tech. 7, 152165 (1996)
[69] J-P. Antoine, D. Barache, R.M. Cesar Jr., L.F. Costa, Shape characterization with the
wavelet transform. Signal Process. 62 265290 (1997)
[70] J-P. Antoine, P. Antoine, B. Piraux, Wavelets in atomic physics, in Spline Functions and
the Theory of Wavelets, ed. by S. Dubuc, G. Deslauriers. CRM Proceedings and Lecture
Notes, vol. 18, (AMS, Providence, RI, 1999), pp.261276
[71] J-P. Antoine, P. Antoine, B. Piraux, Wavelets in atomic physics and in solid state physics,
in Wavelets in Physics, Chap. 8, ed. by J.C. van den Berg (Cambridge University Press,
Cambridge, 1999)
[72] J-P. Antoine, R. Murenzi, P. Vandergheynst, Directional wavelets revisited: Cauchy
wavelets and symmetry detection in patterns. Appl. Comput. Harmon. Anal. 6, 314345
(1999)
[73] J-P. Antoine, L. Jacques, R. Twarock, Wavelet analysis of a quasiperiodic tiling with
fivefold symmetry. Phys. Lett. A 261, 265274 (1999)
[74] J-P. Antoine, L. Jacques, P. Vandergheynst, Penrose tilings, quasicrystals, and wavelets, in
Wavelet Applications in Signal and Image Processing VII, SPIE Proceedings, vol. 3813
(SPIE, Bellingham, WA, 1999), pp. 2839
[75] J-P. Antoine, Y.B. Kouagou, D. Lambert, B. Torrsani, An algebraic approach to discrete
dilations. Application to discrete wavelet transforms. J. Fourier Anal. Appl. 6, 113141
(2000)
[76] J-P. Antoine, A. Coron, J.-M. Dereppe, Water peak suppression: Time-frequency vs. time-
scale approach. J. Magn. Reson. 144, 189194 (2000)
[77] J-P. Antoine, J-P. Gazeau, P. Monceau, J. R. Klauder, K. Penson, Temporally stable
coherent states for infinite well and Pschl-Teller potentials. J. Math. Phys. 42, 23492387
(2001)
[78] J-P. Antoine, A. Coron, C. Chauvin, Wavelets and related time-frequency techniques in
magnetic resonance spectroscopy. NMR Biomed. 14, 265270 (2001)
[79] J-P. Antoine, L. Demanet, J.-F. Hochedez, L. Jacques, R. Terrier, E. Verwichte, Application
of the 2-D wavelet transform to astrophysical images. Phys. Mag. 24, 93116 (2002)
[80] J-P. Antoine, L. Demanet, L. Jacques, P. Vandergheynst, Wavelets on the sphere: Imple-
mentation and approximations. Appl. Comput. Harmon. Anal. 13, 177200 (2002)
[81] J-P. Antoine, I. Bogdanova, P. Vandergheynst, The continuous wavelet transform on conic
sections. Int. J. Wavelets Multires. Inform. Proc. 6, 137156 (2007)
[82] J-P. Antoine, D. Rosca, P. Vandergheynst, Wavelet transform on manifolds: Old and new
approaches. Appl. Comput. Harmon. Anal. 28, 189202 (2010)
[83] P. Antoine, B. Piraux, A. Maquet, Time profile of harmonics generated by a single atom in
a strong electromagnetic field. Phys. Rev. A 51, R1750R1753 (1995)
[84] P. Antoine, B. Piraux, D.B. Miloevic, M. Gajda, Generation of ultrashort pulses of
harmonics. Phys. Rev. A 54, R1761R1764 (1996)
[85] P. Antoine, B. Piraux, D.B. Miloevic, M. Gajda, Temporal profile and time control of
harmonic generation. Laser Phys. 7, 594601 (1997)
[86] Apollonius: see Wikipedia. http://en.wikipedia.org/wiki/Apollonius_of_Perga
[87] F.T. Arecchi, E. Courtens, R. Gilmore, H. Thomas, Atomic coherent states in quantum
optics. Phys. Rev. A 6, 22112237 (1972)
[88] I. Aremua, J-P. Gazeau, M.N. Hounkonnou, Action-angle coherent states for quantum
systems with cylindric phase space. J. Phys. A: Math. Theor. 45, 335302 (2012)
550 References

[89] F. Argoul, A. Arnodo, G. Grasseau, Y. Gagne, E.J. Hopfinger, Wavelet analysis of

turbulence reveals the multifractal nature of the Richardson cascade. Nature 338, 5153
(1989)
[90] F. Argoul, A. Arnodo, J. Elezgaray, G. Grasseau, R. Murenzi, Wavelet analysis of the
self-similarity of diffusion-limited aggregates and electrodeposition clusters. Phys. Rev. A
41, 55375560 (1990)
[91] T.A. Arias, Multiresolution analysis of electronic structure: Semicardinal and wavelet
bases. Rev. Mod. Phys. 71, 267312 (1999)
[92] A. Arnodo, F. Argoul, E. Bacry, J. Elezgaray, E. Freysz, G. Grasseau, J.F. Muzy, B.
Pouligny, Wavelet transform of fractals, in Wavelets and Applications (Proc. Marseille
1989) ed. by Y. Meyer (Masson and Springer, Paris and Berlin, 1991), pp. 286352
[93] A. Arnodo, E. Bacry , J.F. Muzy, The thermodynamics of fractals revisited with wavelets.
Physica A 213, 232275 (1995)
[94] A. Arnodo, E. Bacry, J.F. Muzy, Oscillating singularities in locally self-similar functions.
Phys. Rev. Lett. 74, 48234826 (1995)
[95] A. Arnodo, E. Bacry, P.V. Graves, J.F. Muzy, Characterizing long-range correlations in
DNA sequences from wavelet analysis. Phys. Rev. Lett. 74, 32933296 (1996)
[96] A. Arnodo, Y. dAubenton, E. Bacry, P.V. Graves, J.F. Muzy, C. Thermes, Wavelet based
fractal analysis of DNA sequences. Physica D 96, 291320 (1996)
[97] A. Arnodo, E. Bacry, S. Jaffard, J.F. Muzy, Oscillating singularities on Cantor sets. A
grand canonical multifractal formalism. J. Stat. Phys. 87, 179209 (1997)
[98] A. Arnodo, E. Bacry, S. Jaffard, J.F. Muzy, Singularity spectrum of multifractal functions
involving oscillating singularities. J. Fourier Anal. Appl. 4, 159174 (1998)
[99] N. Aronszajn, Theory of reproducing kernels. Trans. Amer. Math. Soc. 66, 337404 (1950)
[100] A. Ashtekar, J. Lewandowski, D. Marolf, J. Mouro, T. Thiemann, Coherent state
transforms for spaces of connections. J. Funct. Analysis 135, 519551 (1996)
[101] A. Askari-Hemmat, M.A. Dehghan, M. Radjabalipour, Generalized frames and their
redundancy. Proc. Amer. Math. Soc. 129, 11431147 (2001)
[102] E.W. Aslaksen, J.R. Klauder, Unitary representations of the affine group. J. Math. Phys. 9,
206211 (1968)
[103] E.W. Aslaksen, J.R. Klauder, Continuous representation theory using the affine group.
J. Math. Phys. 10, 22672275 (1969)
[104] D. Astruc, L. Planti, R. Murenzi, Y. Lebret, D. Vandromme, On the use of the 3D
wavelet transform for the analysis of computational fluid dynamics results, in Progress
in Wavelet Analysis and Applications (Proc. Toulouse 1992), ed. by Y. Meyer, S. Roques
(Ed. Frontires, Gif-sur-Yvette 1993), pp. 463470
[105] P.W. Atkins, J.C. Dobson, Angular momentum coherent states. Proc. Roy. Soc. London, A
321, 321340 (1971)
[106] B.W. Atkinson, D.O. Bruff, J.S. Geronimo, D. P. Hardin, Wavelets centered on a knot
sequence: Piecewise polynomial wavelets on a quasi-crystal lattice, preprint (2011).
arXiv:1102.4246v1 [math.NA]
[107] I.S. Averbuch, N.F. Perelman, Fractional revivals: Universality in the long-term evolution
of quantum wave packets beyond the correspondence principle dynamics. Phys. Lett. A
139, 449453 (1989)
[108] H. Bacry, J-M. Lvy-Leblond, Possible kinematics, J. Math. Phys. 9, 16051614 (1968)
[109] H. Bacry, A. Grossmann, J. Zak, Proof of the completeness of lattice states in kq
representation. Phys. Rev. B 12, 11181120 (1975)
[110] L. Baggett, K.F. Taylor, Groups with completely reducible regular representation. Proc.
Amer. Math. Soc. 72, 593600 (1978)
[111] V.G. Bagrov, J-P. Gazeau, D. Gitman, A. Levine, Coherent states and related quantizations
for unbounded motions. J. Phys. A: Math. Theor. 45, 125306 (2012). arXiv:1201.0955v2
[quant-ph]
[112] P. Balazs, J-P. Antoine, A Grybos, Weighted and controlled frames: Mutual relationship
and first numerical properties. Int. J. Wavelets Multires. Inform. Proc. 8, 109132 (2010)
References 551

[113] P. Balazs, D.T. Stoeva, J-P. Antoine, Classification of general sequences by frame-related
operators. Sampling Theory Signal Image Proc. (STSIP) 10, 151170 (2011)
[114] P. Balazs, D. Bayer, A. Rahimi, Multipliers for continuous frames in Hilbert spaces. J.
Phys. A: Math. Gen. 45, 244023 (2012)
[115] P. Baldi, G. Kerkyacharian, D. Marinucci, D. Picard, High frequency asymptotics for
wavelet-based tests for Gaussianity and isotropy on the torus. J. Multivar. Anal. 99, 606
636 (2008)
[116] P. Baldi, G. Kerkyacharian, D. Marinucci, D. Picard, Asymptotics for spherical needlets.
Ann. of Stat. 37, 11501171 (2009)
[117] M.C. Baldiotti, J-P. Gazeau, D.M. Gitman, Coherent states of a particle in magnetic field
and Stieltjes moment problem. Phys. Lett. A 373, 19161920 (2009); Erratum: Phys. Lett.
A 373 2600 (2009)
[118] M.C. Baldiotti, J-P. Gazeau, D.M. Gitman, Semiclassical and quantum description of
motion on the noncommutative plane. Phys. Lett. A 373, 39373943 (2009)
[119] R. Balian, Un principe dincertitude fort en thorie du signal ou en mcanique quantique.
C.R. Acad. Sci.(Paris) 292, 13571362 (1981)
[120] M. Bander, C. Itzykson, Group theory and the hydrogen atom I, II. Rev. Mod. Phys. 38,
330345, 346358 (1966)
[121] D. Barache, S. De Bivre, J-P. Gazeau, Affine symmetry semigroups for quasicrystals.
Europhys. Lett. 25, 435440 (1994)
[122] D. Barache, J-P. Antoine, J-M. Dereppe, The continuous wavelet transform, a tool for NMR
spectroscopy, J. Magn. Reson. 128, 111 (1997)
[123] V. Bargmann, On a Hilbert space of analytic functions and an associated integral transform.
Part I. Commun. Pure Appl. Math. 14, 187214 (1961)
[124] V. Bargmann, On a Hilbert space of analytic functions and an associated integral transform.
Part II. A family of related function spaces. Application to distribution theory, Commun.
Pure Appl. Math. 20, 1101 (1967)
[125] V. Bargmann, P. Butera, L. Girardello, J.R. Klauder, On the completeness of coherent
states. Reports Math. Phys. 2, 221228 (1971)
[126] A.O. Barut, L. Girardello, New coherent states associated with non compact groups.
Commun. Math. Phys. 21, 4155 (1971)
[127] A.O. Barut, H. Kleinert, Transition probabilities of the hydrogen atom from noncompact
dynamical groups. Phys. Rev. 156, 15411545 (1967)
[128] A.O. Barut, B.W. Xu, Non-spreading coherent states riding on Kepler orbits. Helv. Phys.
Acta 66, 711720 (1993)
[129] G. Battle, Wavelets: A renormalization group point of view, in Wavelets and Their
Applications, ed. by M.B. Ruskai, G. Beylkin, R. Coifman, I. Daubechies, S. Mallat, Y.
Meyer, L. Raphael (Jones and Bartlett, Boston, 1992), pp. 323349
[130] P. Bellomo, C.R. Stroud Jr., Dispersion of Klauders temporally stable coherent states for
the hydrogen atom. J. Phys. A: Math. Gen. 31, L445L450 (1998)
[131] J. Ben Geloun, J. R. Klauder, Ladder operators and coherent states for continuous spectra.
J. Phys. A: Math. Theor. 42, 375209 (2009)
[132] J. Ben Geloun, J. Hnybida, J.R. Klauder, Coherent states for continuous spectrum operators
with non-normalizable fiducial states. J. Phys. A: Math. Theor. 45, 085301 (2012)
[133] J.J. Benedetto, T.D. Andrews, Intrinsic wavelet and frame applications, in Independent
Component Analyses, Wavelets, Neural Networks, Biosystems, and Nanoengineering IX
ed. by H. Szu, L. Dai. SPIE Proceedings, vol. 8058, (SPIE, Bellingham, WA, 2011), p.
805802
[134] J.J. Benedetto, A. Teolis, A wavelet auditory model and data compression. Appl. Comput.
Harmon. Anal. 1, 328 (1993)
[135] F.A. Berezin, Quantization. Math. USSR Izvestija 8, 11091165 (1974)
[136] F.A. Berezin, General concept of quantization. Commun. Math. Phys. 40, 153174 (1975)
[137] H. Bergeron, From classical to quantum mechanics: How to translate physical ideas into
mathematical language. J. Math. Phys. 42, 39834019 (2001)
552 References

[138] H. Bergeron, Rigorous bra-ket formalism and wave-function operator. J. Math. Phys. 47,
022105 (2006)
[139] H. Bergeron and J-P. Gazeau, Integral quantization with two basic examples. Preprint
(2013); arXiv:1308.2348v1
[140] H. Bergeron, A. Valance, Overcomplete basis for one dimensional Hamiltonians. J. Math.
Phys. 36, 15721592 (1995)
[141] H. Bergeron, J-P. Gazeau, P. Siegl, A. Youssef, Semi-classical behavior of Pschl-Teller
coherent states. Eur. Phys. Lett. 92, 60003 (2010)
[142] H. Bergeron, P. Siegl, A. Youssef, New SUSYQM coherent states for Pschl-Teller
potentials: A detailed mathematical analysis. J. Phys. A: Math. Theor. 45, 244028 (2012)
[143] H. Bergeron, J-P. Gazeau, A. Youssef, Are the Weyl and coherent state descriptions
physically equivalent? Phys. Lett. A 377, 598605 (2013)
[144] H. Bergeron, A. Dapor, J-P. Gazeau, P. Makiewicz, Wavelet quantum cosmology, preprint
(2013). arXiv:1305.0653 [gr-qc]
[145] S. Bergman, ber die Kernfunktion eines Bereiches und ihr Verhalten am Rande I. Reine
Angw. Math. 169, 142 (1933)
[146] D. Bernier, K.F. Taylor, Wavelets from square-integrable representations. SIAM J. Math.
Anal. 27, 594608 (1996)
[147] G. Bernuau, Wavelet bases associated to a self-similar quasicrystal. J. Math. Phys. 39,
42134225 (1998)
[148] A. Bertrand, Dveloppements en base de Pisot et rpartition modulo 1. C. R. Acad. Sci.
Paris 285, 419421 (1977)
[149] J. Bertrand, P. Bertrand, Classification of affine Wigner functions via an extended
covariance principle, in Group Theoretical Methods in Physics (Proc. Sainte-Adle 1988),
ed. by Y. Saint-Aubin, L. Vinet (World Scientific, Singapore, 1989), pp. 13801383
[150] Z. Biaynicka-Birula, I. Biaynicki-Birula, Space-time description of squeezing. J. Opt.
Soc. Am. B4, 16211626 (1987)
[151] E. Bianchi, E. Magliaro, C. Perini, Coherent spin-networks. Phys. Rev. D 82, 024012
(2010)
[152] S. Biskri, J-P. Antoine, B. Inhester, F. Mekideche, Extraction of Solar coronal magnetic
loops with the 2-D Morlet wavelet transform. Solar Phys. 262, 373385 (2010)
[153] R. Bluhm, V.A. Kostelecky, J.A. Porter, The evolution and revival structure of localized
quantum wave packets. Am. J. Phys. 64, 944953 (1996)
[154] I. Bogdanova, P. Vandergheynst, J-P. Antoine, L. Jacques, M. Morvidone, Stereographic
wavelet frames on the sphere,. Appl. Comput. Harmon. Anal. 19, 223252 (2005)
[155] I. Bogdanova, X. Bresson, J-P. Thiran, P. Vandergheynst, Scale space analysis and active
contours for omnidirectional images. IEEE Trans. Image Process. 16 18881901 (2007)
[156] I. Bogdanova, P. Vandergheynst, J-P. Gazeau, Continuous wavelet transform on the
hyperboloid. Appl. Comput. Harmon. Anal. 23 (2007) , 286306 (2007)
[157] P. Boggiatto, E. Cordero, Anti-Wick quantization of tempered distributions, in Progress in
Analysis, Berlin (2001), vol. I, II (World Sci. Publ., River Edge, NJ, 2003), pp. 655662
[158] P. Boggiatto, E. Cordero, K. Grchenig, Generalized Anti-Wick operators with symbols in
distributional Sobolev spaces. Int. Equ. Oper. Theory, 48, 427442 (2004)
[159] A. Bhm, The Rigged Hilbert Space in quantum mechanics, in Lectures in Theoretical
Physics, vol. IX A, ed. by W.A Brittin et al. (Gordon & Breach, New York, 1967), pp.
255315
[160] G. Bohnk, Treillis dondelettes associs aux groupes de Lorentz. Ann. Inst. H. Poincar
54, 245259 (1991)
[161] W.R. Bomstad, J.R. Klauder, Linearized quantum gravity using the projection operator
formalism. Class. Quantum Grav. 23, 59615981 (2006)
[162] V.V. Borzov, Orthogonal polynomials and generalized oscillator algebras. Int. Transform.
Spec. Funct. 12, 115138 (2001)
[163] V.V. Borzov, E.V. Damaskinsky, Generalized coherent states for classical orthogonal
polynomials. Day on Diffraction (2002). arXiv:math.QA/0209181v1 (SPb 2002)
References 553

[164] K. Bouyoucef, D. Fraix-Burnaix, S. Roques, Interactive deconvolution with error analysis

(IDEA) in astronomical imaging: Application to aberrated HST images on SN1987A, M87
and 3C66B. Astron. Astroph. Suppl. Ser. 121, 16 (1997)
[165] A. Bouzouina, S. De Bivre, Equipartition of the eigenfunctions of quantized ergodic maps
on the torus. Commun. Math. Phys. 178, 83105 (1996)
[166] P. Brault, J-P. Antoine, A spatio-temporal Gaussian-Conical wavelet with high aperture
selectivity for motion and speed analysis. Appl. Comput. Harmon. Anal. 34, 148161
(2012)
[167] A. Briguet, S. Cavassila, D. Graveron-Demilly, Suppression of huge signals using the
Cadzow enhancement procedure. The NMR Newslett. 440, 26 (1995)
[168] C.M. Brislawn, Fingerprints go digital. Notices Amer. Math. Soc. 42, 12781283 (1995)
[169] C.M. Brislawn, On the group-theoretic structure of lifted filter banks, in Excursions in
Harmonic Analysis, vol. 1, 2, ed. by T.D. Andrews, R. Balan, J.J. Benedetto, W. Czaja,
K.A. Okoudjou (Birkhuser, Boston, 2013), pp. 113135
[170] F. Bruhat, Sur les reprsentations induites des groupes de Lie. Bull. Soc. Math. France 84,
97205 (1956)
[171] T. Blow, Multiscale image processing on the sphere, in DAGM-Symposium (2002), pp.
609617
[172] C. Burdik, C. Frougny, J,-P. Gazeau, R. Krejcar, Beta-integers as natural counting systems
for quasicrystals. J. Phys. A: Math. Gen. 31, 64496472 (1998)
[173] K.E. Cahill, Coherent-state representations for the photon density. Phys. Rev. 138, B1566
1576 (1965)
[174] K.E. Cahill, R. Glauber, Density operators and quasiprobability distributions. Phys. Rev.
177, 18821902 (1969)
[175] K.E. Cahill, R. Glauber, Ordered expansions in boson amplitude operators. Phys. Rev. 177,
18571881 (1969)
[176] A.R. Calderbank, I. Daubechies, W. Sweldens, B.L. Yeo, Wavelets that map integers to
integers. Appl. Comput. Harmon. Anal. 5, 332369 (1998)
[177] M. Calixto, J. Guerrero, Wavelet transform on the circle and the real line: A unified group-
theoretical treatment. Appl. Comput. Harmon. Anal. 21, 204229 (2006)
[178] M. Calixto, E. Prez-Romero, Extended MacMahon-Schwingers master theorem and
conformal wavelets in complex Minkowski space. Appl. Comput. Harmon. Anal. 31, 143
168 (2011)
[179] M. Calixto, J. Guerrero, D. Rosca, Wavelet transform on the torus: A group-theoretical
approach, preprint (2013)
[180] E.J. Cands, Harmonic analysis of neural networks. Appl. Comput. Harmon. Anal. 6, 197
218 (1999)
[181] E.J. Cands, Ridgelets and the representation of mutilated Sobolev functions. SIAM J.
Math. Anal. 33, 347368 (2001)
[182] E.J. Cands, L. Demanet, Curvelets and Fourier integral operators. C.R. Acad. Sci. Paris,
Sr. I, Math. 336, 395398 (2003)
[183] E.J. Cands, L. Demanet, The curvelet representation of wave propagators is optimally
sparse. Commun. Pure Appl. Math. 58, 14721528 (2004)
[184] E.J. Cands, L. Demanet, The curvelet representation of wave propagators is optimally
sparse. Commun. Pure Appl. Math. 58, 14721528 (2005)
[185] E.J. Cands, D.L. Donoho, Curvelets A surprisingly effective nonadaptive representation
for objects with edges, in Curves and Surfaces, ed. by L.L. Schumaker et al. (Vanderbilt
University Press, Nashville, TN, 1999)
[186] E.J. Cands, D.L. Donoho, Ridgelets: A key to higher-dimensional intermittency? Phil.
Trans. R. Soc. Lond. A. 357, 24952509 (1999)
[187] E.J. Cands, D.L. Donoho, Curvelets, multiresolution representation, and scaling laws, in
Wavelet Applications in Signal and Image Processing VIII ed. by A. Aldroubi, A. Laine,
M. Unser. SPIE Proceedings, vol. 4119 (SPIE, Bellingham, WA, 2000), pp. 112
554 References

[188] E.J. Cands, D.L. Donoho, Recovering edges in ill-posed inverse problems: Optimality of
curvelet frames. Ann. Statist. 30, 784842 (2002)
[189] E.J. Cands, D.L. Donoho, New tight frames of curvelets and optimal representations of
objects with piecewise C2 singularities. Commun. Pure Appl. Math. 57, 219266 (2004)
[190] E.J. Cands, D.L. Donoho, Continuous curvelet transform. I. Resolution of the wavefront
set II: Discretization and frames. Appl. Comput. Harmon. Anal. 19, 162197, 198222
(2005)
[191] E.J. Cands, F. Guo, New multiscale transforms, minimum total variationsynthesis:
Applications to edge-preserving image reconstruction. Signal Proc. 82, 15191543 (2002)
[192] E.J. Cands, L. Demanet, D. Donoho, L. Ying, Fast discrete curvelet transforms. Multiscale
Model. Simul. 5, 861899 (2006)
[193] A.L. Carey, Square integrable representations of non-unimodular groups. Bull. Austr.
Math. Soc. 15, 112 (1976)
[194] A.L. Carey, Group representations in reproducing kernel Hilbert spaces. Rep. Math. Phys.
14, 247259 (1978)
[195] P. Carruthers, M.M. Nieto, Phase and angle variables in quantum mechanics. Rev. Mod.
Phys. 40, 411440 (1968)
[196] P.G. Casazza, G. Kutyniok, Frames of subspaces, in Wavelets, Frames and Operator The-
ory. Contemporary Mathematics, vol. 345 (American Mathematical Society, Providence,
RI, 2004), pp. 87113
[197] P.G. Casazza, D. Han, D.R. Larson, Frames for Banach spaces. Contemp. Math. 247, 149
182 (1999)
[198] P. Casazza, O. Christensen, S. Li, A. Lindner, Riesz-Fischer sequences and lower frame
bounds. Z. Anal. Anwend. 21, 305314 (2002)
[199] P.G. Casazza, G. Kutyniok, S. Li, Fusion frames and distributed processing. Appl.Comput.
Harmon. Anal. 25, 114132 (2008)
[200] D.P.L. Castrigiano, R.W. Henrichs, Systems of covariance and subrepresentations of
induced representations. Lett. Math. Phys. 4, 169175 (1980)
[201] U. Cattaneo, Densities of covariant observables, J. Math. Phys. 23, 659664 (1982)
[202] A. Cerioni, L. Genovese, I. Duchemin, Th. Deutsch, Accurate complex scaling of three
dimensional numerical potentials. Preprint (2013); arXiv:1303.6439v2
[203] S.-J. Chang, K-J. Shi, Evolution and exact eigenstates of a resonant quantum system. Phys.
Rev. A 34, 722 (1986)
[204] S.H.H. Chowdhury, S.T. Ali, All the groups of signal analysis from the (1+1) affine Galilei
group. J. Math. Phys. 52, 103504 (2011)
[205] S.L. Chown, Antarctic marine biodiversity and deep-sea hydrothermal vents. PLoS Biol.
10, 14 (2012)
[206] C. Cishahayo, S. De Bivre, On the contraction of the discrete series of SU(1,1). Ann. Inst.
Fourier (Grenoble) 43, 551567 (1993)
[207] M. Clerc and S. Mallat, Shape from texture and shading with wavelets, in Dynamical
Systems, Control, Coding, Computer Vision. Progress in Systems and Control Theory, 25
393417 (1999)
[208] L. Cohen, General phase-space distribution functions. J. Math. Phys. 7, 781786 (1966)
[209] A. Cohen, I. Daubechies, J.-C. Feauveau, Biorthogonal bases of compactly supported
wavelets. Commun. Pure Appl. Math. 45, 485560 (1992)
[210] R. Coifman, Y. Meyer, M.V. Wickerhauser, Wavelet analysis and signal processing,
in Wavelets and Their Applications, ed. by M.B. Ruskai, G. Beylkin, R. Coifman,
I. Daubechies, S. Mallat, Y. Meyer, L. Raphael (Jones and Bartlett, Boston, 1992),
pp. 153178
[211] R. Coifman, Y. Meyer, M.V. Wickerhauser, Entropy-based algorithms for best basis
selection. IEEE Trans. Inform. Theory 38, 713718 (1992)
[212] R. Coquereaux, A. Jadczyk, Conformal theories, curved spaces, relativistic wavelets and
the geometry of complex domains. Rev. Math. Phys. 2, 144 (1990)
References 555

[213] A. Coron, L. Vanhamme, J-P. Antoine, P. Van Hecke, S. Van Huffel, The filtering approach
to solvent peak suppression in MRS: A critical review. J. Magn. Reson. 152, 2640 (2001)
[214] N. Cotfas, J-P. Gazeau, Finite tight frames and some applications (topical review). J. Phys.
A: Math. Theor. 43, 193001 (2010)
[215] N. Cotfas, J-P. Gazeau, K. Grska, Complex and real Hermite polynomials and related
quantizations. J. Phys. A: Math. Theor. 43, 305304 (2010)
[216] N. Cotfas, J-P. Gazeau, A. Vourdas, Finite-dimensional Hilbert space and frame quantiza-
tion. J. Phys. A: Math. Gen. 44, 175303 (2011)
[217] E.M.F. Curado, M.A. Rego-Monteiro, L.M.C.S. Rodrigues, Y. Hassouni, Coherent states
for a degenerate system: The hydrogen atom. Physica A 371, 1619 (2006)
[218] S. Dahlke, P. Maass, The affine uncertainty principle in one and two dimensions. Comp.
Math. Appl. 30, 293305 (1995)
[219] S. Dahlke, W. Dahmen, E. Schmidt, I. Weinreich, Multiresolution analysis and wavelets on
S2 and S3 . Numer. Funct. Anal. Optim. 16, 1941 (1995)
[220] S. Dahlke, V. Lehmann, G. Teschke, Applications of wavelet methods to the analysis of
meteorological radar data - An overview. Arabian J. Sci. Eng. 28, 344 (2003)
[221] S. Dahlke, G. Kutyniok, P. Maass, C. Sagiv, H.-G. Stark, G. Teschke, The uncertainty
principle associated with the continuous shearlet transform. Int. J. Wavelets Multiresolut.
Inf. Process. 6, 157181 (2008)
[222] S. Dahlke, G. Kutyniok, G. Steidl, G. Teschke, Shearlet coorbit spaces and associated
Banach frames. Appl. Comput. Harmon. Anal. 27, 195214 (2009)
[223] S. Dahlke, G. Steidl, G. Teschke, The continuous shearlet transform in arbitrary space
dimensions. J. Fourier Anal. Appl. 16, 340364 (2010)
[224] S. Dahlke, G. Steidl, G. Teschke, Shearlet coorbit spaces: Compactly supported analyzing
shearlets, traces and embeddings. J. Fourier Anal. Appl. 17, 12321355 (2011)
[225] T. Dallard, G.R. Spedding, 2-D wavelet transforms: Generalisation of the Hardy space and
application to experimental studies. Eur. J. Mech. B/Fluids 12, 107134 (1993)
[226] C. Daskaloyannis, Generalized deformed oscillator and nonlinear algebras. J. Phys. A:
Math. Gen. 24, L789L794 (1991)
[227] C. Daskaloyannis, K. Ypsilantis, A deformed oscillator with Coulomb energy spectrum.
J. Phys. A: Math. Gen. 25, 41574166 (1992)
[228] I. Daubechies, On the distributions corresponding to bounded operators in the Weyl
quantization. Commun. Math. Phys. 75, 229238 (1980)
[229] I. Daubechies and A. Grossmann, An integral transform related to quantization. I. J. Math.
Phys. 21, 20802090 (1980)
[230] I. Daubechies, A. Grossmann, J. Reignier, An integral transform related to quantization. II.
J. Math. Phys. 24, 239254 (1983)
[231] I. Daubechies, Orthonormal bases of compactly supported wavelets. Commun. Pure Appl.
Math. 41, 909996 (1988)
[232] I. Daubechies, The wavelet transform, time-frequency localisation and signal analysis.
IEEE Trans. Inform. Theory 36, 9611005 (1990)
[233] I. Daubechies, S. Maes, A nonlinear squeezing of the continuous wavelet transform based
on auditory nerve models, in Wavelets in Medicine and Biology, ed. by A. Aldroubi, M.
Unser (CRC Press, Boca Raton, 1996), pp. 527546
[234] I. Daubechies, A. Grossmann, Y. Meyer, Painless nonorthogonal expansions. J. Math. Phys.
27, 12711283 (1986)
[235] E.R. Davies, Introduction to texture analysis, in Handbook of texture analysis, ed. by M.
Mirmehdi, X. Xie, J. Suri (World Scientific, Singapore, 2008), pp. 131
[236] R. De Beer, D. van Ormondt, F.T.A.W. Wajer, S. Cavassila, D. Graveron-Demilly, S. Van
Huffel, SVD-based modelling of medical NMR signals, in SVD and Signal Processing,
III: Algorithms, Architectures and Applications, ed. by M. Moonen, B. De Moor (Elsevier
(North-Holland), Amsterdam, 1995), pp. 467474
[237] S. De Bivre, Coherent states over symplectic homogeneous spaces. J. Math. Phys. 30,
14011407 (1989)
556 References

[238] S. De Bivre, J.A. Gonzalez, Semi-classical behaviour of the Weyl correspondence on the
circle, in Group-Theoretical Methods in Physics (Proc. Salamanca 1992), ed. by M. del
Olmo, M. Santander, J. Mateos Guilarte (CIEMAT, Madrid, 1993), pp. 343346
[239] S. De Bivre, J.A. Gonzlez, Semiclassical behaviour of coherent states on the circle, in
Quantization and Coherent States Methods in Physics, ed. by A. Odzijewicz et al. (World
Scientific, Singapore, 1993)
[240] S. De Bivre, A.E. Gradechi, Quantum mechanics and coherent states on the anti-de Sitter
space-time and their Poincar contraction. Ann. Inst. H. Poincar 57, 403428 (1992)
[241] J. Deenen, C. Quesne, Dynamical group of collective states I, II, III. J. Math. Phys. 23
878889, 20042015 (1982)
[242] J. Deenen, C. Quesne, Dynamical group of collective states I, II, III. J. Math. Phys. 25,
16381650 (1984)
[243] J. Deenen, C. Quesne, Partially coherent states of the real symplectic group. J. Math. Phys.
25 23542366 (1984)
[244] J. Deenen, C. Quesne, Boson representations of the real symplectic group and their
application to the nuclear collective model. J. Math. Phys. 26, 27052716 (1985)
[245] R. Delbourgo, Minimal uncertainty states for the rotation and allied groups. J. Phys. A:
Math. Gen. 10, 18371846 (1977)
[246] R. Delbourgo, J. R. Fox, Maximum weight vectors possess minimal uncertainty. J. Phys.
A: Math. Gen. 10, L233L235 (1977)
[247] V. Delouille, J. de Patoul, J.-F. Hochedez, L. Jacques, J-P. Antoine, Wavelet spectrum
analysis of EIT/SoHO images. Solar Phys. 228, 301321 (2005)
[248] N. Delprat, B. Escudi, P. Guillemain, R. Kronland-Martinet, P. Tchamitchian, B. Tor-
rsani, Asymptotic wavelet and Gabor analysis: Extraction of instantaneous frequencies.
IEEE Trans. Inform. Theory 38, 644664 (1992)
[249] Th. Deutsch, L. Genovese, Wavelets for electronic structure calculations. Collection Soc.
Fr. Neut., 12, 3376 (2011)
[250] B. Dewitt, Quantum theory of gravity I: The canonical theory. Phys. Rev. 160, 11131148
(1967)
[251] R.H. Dicke, Coherence in spontaneous radiation processes. Phys. Rev. 93, 99110 (1954)
[252] A.N. Dinkelaker, A.L. MacKinnon, Wavelets, intermittency and solar flare hard X-rays 1:
Local intermittency measure in cascade and avalanche scenarios. Solar Phys. 282 471481
(2013)
[253] A.N. Dinkelaker, A.L. MacKinnon, Wavelets, intermittency and solar flare hard X-rays 2:
LIM analysis of high time resolution BATSE data. Solar Phys. 282, 483501 (2013)
[254] M.N. Do, M. Vetterli, Contourlets, in Beyond Wavelets, ed. by G.V. Welland (Academic,
San Diego, 2003), pp. 83105
[255] M.N. Do, M. Vetterli, The contourlet transform: An efficient directional multiresolution
image representation. IEEE Trans. Image Process. 14, 20912106 (2005)
[256] V.V. Dodonov, Nonclassical states in quantum optics: A squeezed review of the first 75
years. J. Opt. B: Quant. Semiclass. Opot. 4, R1R33 (2002)
[257] D.L. Donoho, Nonlinear wavelet methods for recovery of signals, densities, and spectra
from indirect and noisy data, in Different Perspectives on Wavelets. Proceedings of
Symposia in Applied Mathematics, vol. 38, ed. by I. Daubechies (American Mathematical
Society, Providence, RI, 1993), pp. 173205
[258] D.L. Donoho, Wedgelets: Nearly minimax estimation of edges. Ann. Stat. 27, 859897
(1999)
[259] D.L. Donoho, X. Huo, Beamlet pyramids: A new form of multiresolution analysis, suited
for extracting lines, curves, and objects from very noisy image data, in SPIE Proceedings,
vol. 5914 (SPIE, Bellingham, WA, 2005), pp. 112
References 557

[260] A.H. Dooley, Contractions of Lie groups and applications to analysis, in Topics in Modern
Harmonic Analysis, vol. I (Istituto Nazionale di Alta Matematica Francesco Severi, Roma,
1983), pp. 483515
[261] A.H. Dooley, J.W. Rice, Contractions of rotation groups and their representations. Math.
Proc. Camb. Phil. Soc. 94, 509517 (1983)
[262] A.H. Dooley, J.W. Rice, On contractions of semisimple Lie groups. Trans. Amer. Math.
Soc. 289, 185202 (1985)
[263] J.R. Driscoll, D.M. Healy, Computing Fourier transforms and convolutions on the 2-sphere.
Adv. Appl. Math. 15, 202250 (1994)
[264] H. Drissi, F. Regragui, J-P. Antoine, M. Bennouna, Wavelet transform analysis of visual
evoked potentials: Some preliminary results. ITBM-RBM 21, 8491 (2000)
[265] R.J. Duffin, A.C. Schaeffer, A class of nonharmonic Fourier series. Trans. Amer. Math.
Soc. 72, 341366 (1952)
[266] M. Duflo, C.C. Moore, On the regular representation of a nonunimodular locally compact
group J. Funct. Anal. 21, 209243 (1976)
[267] M. Duval-Destin, R. Murenzi, Spatio-temporal wavelets: Application to the analysis of
moving patterns, in Progress in Wavelet Analysis and Applications (Proc. Toulouse 1992),
ed. by Y. Meyer, S. Roques (Ed. Frontires, Gif-sur-Yvette 1993), pp. 399408
[268] M. Duval-Destin, M.-A. Muschietti, B. Torrsani, Continuous wavelet decompositions,
multiresolution, and contrast analysis. SIAM J. Math. Anal. 24, 739755 (1993)
[269] S.J.L. van Eindhoven, J.L.H. Meyers, New orthogonality relations for the Hermite
polynomials and related Hilbert spaces. J. Math. Anal. Appl. 146, 8998 (1990)
[270] M. ElBaz, R. Fresneda, J-P. Gazeau, Y. Hassouni, Coherent state quantization of para-
grassmann algebras. J. Phys. A: Math. Theor. 43, 385202 (2010); Corrigendum: J. Phys.
A: Math. Theor. 45 (2012)
[271] A. Elkharrat, J-P. Gazeau, F. Dnoyer, Multiresolution of quasicrystal diffraction spectra.
Acta Cryst. A65, 466489 (2009)
[272] Q. Fan, Phase space analysis of the identity decompositions. J. Math. Phys. 34, 34713477
(1993)
[273] M. Fanuel, S. Zonetti, Affine quantization and the initial cosmological singularity. Euro-
phys. Lett. 101, 10001 (2013)
[274] M. Farge, Wavelet transforms and their applications to turbulence. Annu. Rev. Fluid Mech.
24, 395457 (1992)
[275] M. Farge, N. Kevlahan, V. Perrier, E. Goirand, Wavelets and turbulence. Proc. IEEE 84,
639669 (1996)
[276] M. Farge, N.K.-R. Kevlahan, V. Perrier, K. Schneider, Turbulence analysis, modelling
and computing using wavelets, in Wavelets in Physics, Chap. 4, ed. by J.C. van den Berg
(Cambridge University Press, Cambridge, 1999)
[277] H.G. Feichtinger, Coherent frames and irregular sampling, in Recent Advances in Fourier
Analysis and Its applications, ed. by J.S. Byrnes, J.L. Byrnes (Kluwer, Dordrecht, 1990),
pp. 427440
[278] H.G. Feichtinger, K.H. Grchenig, Banach spaces related to integrable group representa-
tions and their atomic decompositions I. J. Funct. Anal. 86, 307340 (1989)
[279] H.G. Feichtinger, K.H. Grchenig, Banach spaces related to integrable group representa-
tions and their atomic decompositions II. Mh. Math. 108, 129148 (1989)
[280] M. Flensted-Jensen, Discrete series for semisimple symmetric spaces. Ann. of Math. 111,
253311 (1980)
[281] K. Flornes, A. Grossmann, M. Holschneider, B. Torrsani, Wavelets on discrete fields.
Appl. Comput. Harmon. Anal. 1, 137146 (1994)
[282] V. Fock, Zur Theorie der Wasserstoffatoms. Zs. f. Physik 98, 14554 (1936)
[283] M. Fornasier, H. Rauhut, Continuous frames, function spaces, and the discretization
problem. J. Fourier Anal. Appl. 11, 245287 (2005)
558 References

[284] W. Freeden, M. Schreiner, Orthogonal and non-orthogonal multiresolution analysis, scale

discrete and exact fully discrete wavelet transform on the sphere. Constr. Approx. 14, 493
515 (1997)
[285] W. Freeden, U. Windheuser, Combined spherical harmonic and wavelet expansion A
future concept in Earths gravitational determination. Appl. Comput. Harmon. Anal. 4, 1
37 (1997)
[286] W. Freeden, T. Maier, S. Zimmermann, A survey on wavelet methods for (geo)applications,
Revista Mathematica Complutense 16 (2003) 277310
[287] W. Freeden, M. Schreiner, Biorthogonal locally supported wavelets on the sphere based on
zonal kernel functions. J. Fourier Anal. Appl. 13, 693709 (2007)
[288] W.T. Freeman, E.H. Adelson, The design and use of steerable filters. IEEE Trans. Pattern
Anal. Machine Intell. 13, 891906 (1991)
[289] L. Freidel, E.R. Livine, U(N) coherent states for loop quantum gravity. J. Math. Phys. 52,
052502 (2011)
[290] J. Froment, S. Mallat, Arbitrary low bit rate image compression using wavelets, in Progress
in Wavelet Analysis and Applications (Proc. Toulouse 1992), ed. by Y. Meyer, S. Roques
(Ed. Frontires, Gif-sur-Yvette 1993), pp. 413418 and references therein
[291] L. Freidel, S. Speziale, Twisted geometries: A geometric parameterisation of SU(2) phase
space. Phys. Rev. D 82, 084040 (2010)
[292] H. Fhr, Wavelet frames and admissibility in higher dimensions. J. Math. Phys. 37, 6353
6366 (1996)
[293] H. Fhr, M. Mayer, Continuous wavelet transforms from semidirect products: Cyclic
representations and Plancherel measure. J. Fourier Anal. Appl. 8, 375396 (2002)
[294] D. Gabor, Theory of communication. J. Inst. Electr. Engrg.(London) 93, 429457 (1946)
[295] J-P. Gabardo, D. Han, Frames associated with measurable spaces. Adv. Comput. Math. 18,
127147 (2003)
[296] E. Galapon, Paulis theorem and quantum canonical pairs: The consistency of a bounded,
self-adjoint time operator canonically conjugate to a Hamiltonian with non-empty point
spectrum. Proc. R. Soc. Lond. A 458, 451472 (2002)
[297] P.L. Garca de Lon, J-P. Gazeau, Coherent state quantization and phase operator. Phys.
Lett. A 361, 301304 (2007)
[298] L. Garca de Len, J-P. Gazeau, J. Quva, The infinite well revisited: Coherent states and
quantization. Phys. Lett. A 372, 35973607 (2008)
[299] T. Garidi, J-P. Gazeau, E. Huguet, M. Lachize Rey, J. Renaud, Fuzzy spheres from
inequivalent coherent states quantization. J. Phys. A: Math. Theor. 40, 1022510249 (2007)
[300] J-P. Gazeau, Four Euclidean conformal group in atomic calculations: Exact analytical
expressions for the bound-bound two-photon transition matrix elements in the H atom.
J. Math. Phys. 19, 10411048 (1978)
[301] J-P. Gazeau, On the four Euclidean conformal group structure of the sturmian operator.
Lett. Math. Phys. 3, 285292 (1979)
[302] J-P. Gazeau, Technique Sturmienne pour le spectre discret de lquation de Schrdinger.
J. Phys. A: Math. Gen. 13, 36053617 (1980)
[303] J-P. Gazeau, Four Euclidean conformal group approach to the multiphoton processes in the
H atom. J. Math. Phys. 23, 156164 (1982)
[304] J-P. Gazeau, A remarkable duality in one particle quantum mechanics between some
confining potentials and (R + L ) potentials. Phys. Lett. 75A, 159163 (1980)
[305] J-P. Gazeau, SL(2, R)-coherent states and integrable systems in classical and quantum
physics, in Quantization, Coherent States, and Complex Structures, ed. by J-P. Antoine,
S.T. Ali, W. Lisiecki, I.M. Mladenov, A. Odzijewicz (Plenum Press, New York and London,
1995), pp. 147158
[306] J-P. Gazeau, S. Graffi, Quantum harmonic oscillator: A relativistic and statistical point of
view. Boll. Unione Mat. Ital. 11-A, 815839 (1997)
[307] J-P. Gazeau, V. Hussin, Poincar contraction of SU(1,1) Fock-Bargmann structure. J. Phys.
A: Math. Gen. 25, 15491573 (1992)
References 559

[308] J-P. Gazeau, R. Kanamoto, Action-angle coherent states and related quantization, in
Proceedings of QTS7 Colloquium, Prague 2011. Journal of Physics: Conference Series,
vol. 343 (2012), p. 012038-1-9
[309] J-P. Gazeau, J.R. Klauder, Coherent states for systems with discrete and continuous
spectrum. J. Phys. A: Math. Gen. 32, 123132 (1999)
[310] J-P. Gazeau, P. Monceau, Generalized coherent states for arbitrary quantum systems, in
Colloquium M. Flato (Dijon, Sept. 99), vol. II (Klwer, Dordrecht, 2000), pp. 131144
[311] J-P. Gazeau, M. Novello, The question of mass in (Anti-) de Sitter space-times. J. Phys. A:
Math. Theor. 41, 304008 (2008)
[312] J-P. Gazeau, M. del Olmo, q-coherent states quantization of the harmonic oscillator. Ann.
Phys. (NY) 330, 220245 (2013). arXiv:1207.1200 [quant-ph]
[313] J-P. Gazeau, J. Patera, Tau-wavelets of Haar. J. Phys. A: Math. Gen. 29, 45494559 (1996)
[314] J-P. Gazeau, W. Piechocki, Coherent states quantization of a particle in de Sitter space.
J. Phys. A : Math. Gen. 37, 69776986 (2004)
[315] J-P. Gazeau, J. Renaud, Lie algorithm for an interacting SU(1,1) elementary system and its
contraction. Ann. Phys. (NY) 222, 89121 (1993)
[316] J-P. Gazeau, J. Renaud, Relativistic harmonic oscillator and space curvature. Phys. Lett. A
179, 6771 (1993)
[317] J-P. Gazeau, V. Spiridonov, Toward discrete wavelets with irrational scaling factor. J. Math.
Phys. 37, 30013013 (1996)
[318] J-P. Gazeau, F.H. Szafraniec, Holomorphic Hermite polynomials and non-commutative
plane. J. Phys. A: Math. Theor. 44, 495201 (2011)
[319] J-P. Gazeau, J. Patera, E. Pelantov, Tau-wavelets in the plane. J. Math. Phys. 39, 4201
4212 (1998)
[320] J-P. Gazeau, M. Andrle, C. Burdk, R. Krejcar, Wavelet multiresolutions for the Fibonacci
chain. J. Phys. A: Math. Gen. 33, L47L51 (2000)
[321] J-P. Gazeau, M. Andrle, C. Burdk, Bernuau spline wavelets and sturmian sequences.
J. Fourier Anal. Appl. 10, 269300 (2004)
[322] J-P. Gazeau, F.-X. Josse-Michaux, P. Monceau, Finite dimensional quantizations of the
(q, p) plane : new space and momentum inequalities. Int. J. Modern Phys. B 20, 1778
1791 (2006)
[323] J-P. Gazeau, J. Mourad, J. Quva, Fuzzy de Sitter space-times via coherent states
quantization, in Proceedings of the XXVIth Colloquium on Group Theoretical Methods
in Physics, New York, 2006, ed. by J. Birman, S. Catto, B. Nicolescu (Canopus Publishing
Limited, London, 2009)
[324] D. Geller, A. Mayeli, Continuous wavelets on compact manifolds. Math. Z. 262, 895927
(2009)
[325] D. Geller, A. Mayeli, Nearly tight frames and space-frequency analysis on compact
manifolds. Math. Z. 263, 235264 (2009)
[326] L. Genovese, B. Videau, M. Ospici, Th. Deutsch, S. Goedecker, J-F. Mhaut, Daubechies
wavelets for high performance electronic structure calculations: The BigDFT project. C.
R. Mecanique, 339, 149164 (2011)
[327] G. Gentili, C. Stoppato, Power series and analyticity over the quaternions. Math. Ann. 352,
113131 (2012)
[328] G. Gentili, D.C. Struppa, A new theory of regular functions of a quaternionic variable. Adv.
Math. 216, 279301 (2007)
[329] C. Geyer, K. Daniilidis, Catadioptric projective geometry. Int. J. Comput. Vision 45, 223
243 (2001)
[330] R. Gilmore, Geometry of symmetrized states. Ann. Phys. (NY) 74, 391463 (1972)
[331] R. Gilmore, On properties of coherent states. Rev. Mex. Fis. 23, 143187 (1974)
[332] J. Ginibre, Statistical ensembles of complex, quaternion, and real matrices. J. Math. Phys.
6, 440449 (1965)
[333] R.J. Glauber, The quantum theory of optical coherence. Phys. Rev. 130, 25292539 (1963)
560 References

[334] R.J. Glauber, Coherent and incoherent states of radiation field. Phys. Rev. 131, 27662788
(1963)
[335] J. Glimm, Locally compact transformation groups. Trans. Amer. Math. Soc. 101, 124138
(1961)
[336] R. Godement, Sur les relations dorthogonalit de V. Bargmann. C. R. Acad. Sci. Paris 255,
521523, 657659 (1947)
[337] C. Gonnet, B. Torrsani, Local frequency analysis with two-dimensional wavelet trans-
form. Signal Proc. 37, 389404 (1994)
[338] J.A. Gonzalez, M.A. del Olmo, Coherent states on the circle. J. Phys. A: Math. Gen. 31,
88418857 (1998)
[339] X.Gonze, B. Amadon et al., ABINIT: First-principles approach to material and nanosystem
properties. Computer Physics Comm. 180, 25822615 (2009)
[340] K.M. Grski, E. Hivon, A.J. Banday, B.D. Wandelt, F.K. Hansen, M. Reinecke, M.
Bartelmann, HEALPix: A framework for high-resolution discretization and fast analysis
of data distributed on the sphere. Astrophys. J. 622, 759771 (2005)
[341] P. Goupillaud, A. Grossmann, J. Morlet, Cycle-octave and related transforms in seismic
signal analysis. Geoexploration 23, 85102 (1984)
[342] K.H. Grchenig, A new approach to irregular sampling of band-limited functions, in Recent
Advances in Fourier Analysis and Its applications, ed. by J.S. Byrnes, J.L. Byrnes (Kluwer,
Dordrecht, 1990), pp. 251260
[343] K.H. Grchenig, Gabor analysis over LCA groups, in Gabor Analysis and Algorithms
Theory and Applications, ed. by H.G. Feichtinger, T. Strohmer (Birkhuser, Boston-Basel-
Berlin, 1998), pp. 211231
[344] H.J. Groenewold, On the principles of elementary quantum mechanics. Physica 12, 405
460 (1946)
[345] P. Grohs, Continuous shearlet tight frames. J. Fourier Anal. Appl. 17, 506518 (2011)
[346] P. Grohs, G. Kutyniok, Parabolic molecules, preprint TU Berlin (2012)
[347] M. Grosser, A note on distribution spaces on manifolds. Novi Sad J. Math. 38, 121128
(2008)
[348] A. Grossmann, Parity operator and quantization of -functions. Commun. Math. Phys. 48,
191194 (1976)
[349] A. Grossmann, J. Morlet, Decomposition of Hardy functions into square integrable
wavelets of constant shape. SIAM J. Math. Anal. 15, 723736 (1984)
[350] A. Grossmann, J. Morlet, Decomposition of functions into wavelets of constant shape, and
related transforms, in Mathematics + Physics, Lectures on recent results. I, ed. by L. Streit
(World Scientific, Singapore, 1985), pp. 135166
[351] A. Grossmann, J. Morlet, T. Paul, Integral transforms associated to square integrable
representations I: General results. J. Math. Phys. 26, 24732479 (1985)
[352] A. Grossmann, J. Morlet, T. Paul, Integral transforms associated to square integrable
representations II: Examples. Ann. Inst. H. Poincar 45, 293309 (1986)
[353] A. Grossmann, R. Kronland-Martinet, J. Morlet, Reading and understanding the contin-
uous wavelet transform, in Wavelets, Time-Frequency Methods and Phase Space (Proc.
Marseille 1987), ed. by J.-M. Combes, A. Grossmann, P. Tchamitchian, 2nd edn. (Springer,
Berlin, 1990), pp. 220
[354] P. Guillemain, R. Kronland-Martinet, B. Martens, Estimation of spectral lines with help
of the wavelet transform. Application in NMR spectroscopy, in Wavelets and Applications
(Proc. Marseille 1989) ed. by Y. Meyer (Masson and Springer, Paris and Berlin, 1991), pp.
3860
[355] H. de Guise, M. Bertola, Coherent state realizations of su(n + 1) on the n-torus. J. Math.
Phys. 43, 34253444 (2002)
[356] K. Guo, D. Labate, Representation of Fourier Integral Operators using shearlets. J. Fourier
Anal. Appl. 14, 327371 (2008)
References 561

[357] K. Guo, G. Kutyniok, D. Labate, Sparse multidimensional representations using

anisotropic dilation and shear operators, in Wavelets and Spines (Athens, GA, 2005)
(Nashboro Press, Nashville, TN, 2006), pp. 189201
[358] K. Guo, D. Labate, W.-Q. Lim, G. Weiss, E. Wilson, Wavelets with composite dilations
and their MRA properties. Appl. Comput. Harmon. Anal. 20, 202236 (2006)
[359] E.A. Gutkin, Overcomplete subspace systems and operator symbols. Funct. Anal. Appl. 9,
260261 (1975)
[360] G. Gyrgyi, Integration of the dynamical symmetry groups for the 1/r potential. Acta Phys.
Acad. Sci. Hung. 27, 435439 (1969)
[361] B.C Hall, The Segal-Bargmann Coherent State transform for compact Lie groups. J.
Funct. Analysis 122, 103151 (1994)
[362] B. Hall, J.J. Mitchell, Coherent states on spheres. J. Math. Phys. 43, 12111236 (2002)
[363] D.K. Hammond, P. Vandergheynst, R. Gribonval, Wavelets on graphs via spectral theory.
Appl. Comput. Harmon. Anal. 30, 129150 (2011)
[364] Y. Hassouni, E.M.F. Curado, M.A. Rego-Monteiro, Construction of coherent states for
physical algebraic system. Phys. Rev. A 71, 022104 (2005)
[365] J. He, H. Liu, Admissible wavelets associated with the affine automorphism group of the
Siegel upper half-plane. J. Math. Anal. Appl. 208, 5870 (1997)
[366] J. He, H. Liu, Admissible wavelets associated with the classical domain of type one.
Approx. Appl. 14 (1998) 89105
[367] J. He, L. Peng, Wavelet transform on the symmetric matrix space, preprint, Beijing (1997)
(unpublished)
[368] J. He, L. Peng, Admissible wavelets on the unit disk. Complex Variables 35, 109119
(1998)
[369] D.M. Healy Jr., F.E. Schroeck Jr., On informational completeness of covariant localization
observables and Wigner coefficients. J. Math. Phys. 36, 453507 (1995)
[370] C. Heil, D. Walnut, Continuous and discrete wavelet transforms, SIAM Review 31, 628
666 (1989)
[371] K. Hepp, E.H. Lieb, On the superradiant phase transition for molecules in a quantized
radiation field: The Dicke maser model. Ann. Phys. (NY) 76, 360404 (1973)
[372] K. Hepp, E.H. Lieb, Equilibrium statistical mechanics of matter interacting with the
quantized radiation field. Phys. Rev. A 8, 25172525 (1973)
[373] J.A. Hogan, J.D. Lakey, Extensions of the Heisenberg group by dilations and frames. Appl.
Comput. Harmon. Anal. 2, 174199 (1995)
[374] A.L. Hohouto, K. Thirulogasanthar, S.T. Ali, J-P. Antoine, Coherent states lattices and
square integrability of representations. J. Phys. A: Math. Gen.36, 1181711835 (2003)
[375] M. Holschneider, On the wavelet transformation of fractal objects. J. Stat. Phys. 50, 963
993 (1988)
[376] M. Holschneider, Wavelet analysis on the circle. J. Math. Phys. 31, 3944 (1990)
[377] M. Holschneider, Inverse Radon transforms through inverse wavelet transforms. Inv. Probl.
7, 853861 (1991)
[378] M. Holschneider, Localization properties of wavelet transforms. J. Math. Phys. 34, 3227
3244 (1993)
[379] M. Holschneider, General inversion formulas for wavelet transforms. J. Math. Phys. 34,
41904198 (1993)
[380] M. Holschneider, Wavelet analysis over abelian groups. Applied Comput. Harmon. Anal.
2, 5260 (1995)
[381] M. Holschneider, Continuous wavelet transforms on the sphere. J. Math. Phys. 37, 4156
4165 (1996)
[382] M. Holschneider, I. Iglewska-Nowak, Poisson wavelets on the sphere. J. Fourier Anal.
Appl. 13, 405419 (2007)
562 References

[383] M. Holschneider, R. Kronland-Martinet, J. Morlet, P. Tchamitchian, A real-time algorithm

for signal analysis with the help of wavelet transform, in Wavelets, Time-Frequency
Methods and Phase Space (Proc. Marseille 1987), ed. by J.-M. Combes, A. Grossmann, P.
Tchamitchian, 2nd edn. (Springer, Berlin, 1990), pp. 286297
[384] M. Holschneider, P. Tchamitchian, Pointwise analysis of Riemanns nondifferentiable
function, Invent. Math. 105, 157175 (1991)
[385] G.R. Honarasa, M.K. Tavassoly, M. Hatami, R. Roknizadeh, Nonclassical properties of
coherent states and excited coherent states for continuous spectra. J. Phys. A: Math. Theor.
44, 085303 (2011)
[386] M. Hongoh, Coherent states associated with the continuous spectrum of noncompact
groups. J. Math. Phys. 18, 20812085 (1977)
[387] A. Horzela, F.H. Szafraniec, A measure free approach to coherent states. J. Phys. A: Math.
Gen. 45, 244018 (2012)
[388] W.-L. Hwang, S. Mallat, Characterization of self-similar multifractals with wavelet
maxima. Appl. Comput. Harmon. Anal. 1, 316328 (1994)
[389] W.-L. Hwang, C.-S. Lu, P.-C. Chung, Shape from texture: Estimation of planar surface
orientation through the ridge surfaces of continuous wavelet transform. IEEE Trans. Image
Proc. 7, 773780 (1998)
[390] I. Iglewska-Nowak, M. Holschneider, Frames of Poisson wavelets on the sphere. Appl.
Comput. Harmon. Anal. 28, 227248 (2010)
[391] E. Inn, E.P. Wigner, On the contraction of groups and their representations. Proc. Nat.
Acad. Sci. U. S. 39, 510524 (1953)
[392] S. Iqbal, F. Saif, Generalized coherent states and their statistical characteristics in power-
law potentials. J. Math. Phys. 52, 082105 (2011)
[393] C.J. Isham, J.R. Klauder, Coherent states for n-dimensional Euclidean groups E(n) and
their application. J. Math.Phys. 32, 607620 (1991)
[394] L. Jacques, J-P. Antoine, Multiselective pyramidal decomposition of images: Wavelets with
adaptive angular selectivity. Int. J. Wavelets Multires. Inform. Proc. 5, 785814 (2007)
[395] L. Jacques, L. Duval, C. Chaux, G. Peyr, A panorama on multiscale geometric rep-
resentations, intertwining spatial, directional and frequency selectivity. Signal Proc. 91,
26992730 (2011)
[396] H.R. Jalali, M. K. Tavassoly, On the ladder operators and nonclassicality of generalized
coherent state associated with a particle in an infinite square well, preprint (2013).
arXiv:1303.4100v1 [quant-ph]
[397] C. Johnston, On the pseudo-dilation representations of Flornes, Grossmann, Holschneider,
and Torrsani. Appl. Comput. Harmon. Anal. 3, 377385 (1997)
[398] G. Kaiser, Phase-space approach to relativistic quantum mechanics I: Coherent state
representation for massive scalar particles. J. Math.Phys. 18, 952959 (1977)
[399] G. Kaiser, Phase-space approach to relativistic quantum mechanics II: Geometrical aspects.
J. Math.Phys. 19, 502507 (1978)
[400] C. Kalisa, B. Torrsani, N-dimensional affine Weyl-Heisenberg wavelets. Ann. Inst. H.
Poincar 59, 201236 (1993)
[401] W. Kaminski, J. Lewandowski, T. Pawowski, Quantum constraints, Dirac observables and
evolution: group averaging versus the Schrdinger picture in LQC. Class. Quant. Grav. 26,
245016 (2009)
[402] M.R. Karim, S.T. Ali, A relativistic windowed Fourier transform, preprint, Concordia
University, Montral (1997) (unpublished)
[403] M. R. Karim, S.T. Ali, M. Bodruzzaman, A relativistic windowed Fourier transform, in
Proceedings of IEEE SoutheastCon 2000, Nashville, Tennessee, pp. 253260 (2000)
[404] T. Kawazoe, Wavelet transforms associated to a principal series representation of semisim-
ple Lie groups I, II. Proc. Japan Acad. Ser. A Math. Sci. 71, 154157, 158160 (1995)
[405] T. Kawazoe, Wavelet transform associated to an induced representation of SL(n + 2, R).
Ann. Inst. H. Poincar 65, 113 (1996)
References 563

[406] P. Kittipoom, G. Kutyniok, W.-Q. Lim, Irregular shearlet frames: Geometry and approxi-
mation properties. J. Fourier Anal. Appl. 17, 604639 (2011)
[407] P. Kittipoom, G. Kutyniok, W.-Q. Lim, Construction of compactly supported shearlet
frames, Constr. Approx. 35, 2172 (2012)
[408] J. Kiukas, P. Lahti, K. Ylinenc, Phase space quantization and the operator moment problem.
J. Math. Phys. 47, 072104 (2006)
[409] J.R. Klauder, Continuous-representation theory I: Postulates of continuous-representation
theory. J. Math. Phys. 4, 10551058 (1963)
[410] J.R. Klauder, Continuous-representation theory II: Generalized relation between quantum
and classical dynamics. J. Math. Phys. 4, 10581073 (1963)
[411] J.R. Klauder, Path integrals for affine variables, in Functional Integration, Theory and
Applications, ed. by J-P. Antoine, E. Tirapegui (Plenum Press, New York and London,
1980), pp. 101119
[412] J.R. Klauder, Are coherent states the natural language of quantum mechanics?, in
Fundamental Aspects of Quantum Theory, ed. by V. Gorini, A. Frigerio. NATO ASI Series,
vol. B 144 (Plenum Press, New York, 1986), pp. 112
[413] J.R. Klauder, Quantization without quantization. Ann. Phys. (NY) 237, 147160 (1995)
[414] J.R. Klauder, Coherent states for the hydrogen atom. J. Phys. A: Math. Gen. 29, L293
L296 (1996)
[415] J.R. Klauder, An affinity for affine quantum gravity. Proc. Steklov Inst. Math. 272, 169176
(2011) and references therein
[416] J.R. Klauder, R.F. Streater, A wavelet transform for the Poincar group. J. Math. Phys. 32,
16091611 (1991)
[417] J.R. Klauder, R.F. Streater, Wavelets and the Poincar half-plane. J. Math. Phys. 35, 471
478 (1994)
[418] J.R. Klauder, K. Penson, J.-M. Sixdeniers, Constructing coherent states through solutions
of Stieltjes and Hausdorff moment problems. Phys. Rev. A 64, 013817 (2001)
[419] A. Kleppner, R.L. Lipsman, The Plancherel formula for group extensions. Ann. Ec. Norm.
Sup. 5, 459516 (1972)
[420] A.B. Klimov, C. Muoz, Coherent, isotropic and squeezed states in a N-qubit system. Phys.
Scr. 87, 038110 (2013)
[421] A. Klyashko, Dynamical symmetry approach to entanglement, in Physics and Theoretical
Computer Science: From Numbers and Languages to (Quantum) Cryptography - NATO
Security through Science Series: D - Information and Communication Security, vol. 7, ed.
by J-P. Gazeau, J. Nesetril, B. Rovan (IOS Press, Washington, DC, 2007), pp. 2554
[422] S. Kobayashi, Irreducibility of certain unitary representations. J. Math. Soc. Japan 20, 638
642 (1968)
[423] K. Kowalski, J. Rembielinski, L.C. Papaloucas, Coherent states for a quantum particle on
a circle. J. Phys. A: Math. Gen. 29, 41494167 (1996)
[424] K. Kowalski, J. Rembielinski, Quantum mechanics on a sphere and coherent states. J. Phys.
A: Math. Gen. 33, 60356048 (2000)
[425] K. Kowalski, J. Rembielinski, The Bargmann representation for the quantum mechanics
on a sphere. J. Math. Phys. 42, 41384147 (2001)
[426] C. Kristjansen, J. Plefka, G.W. Semenoff, M. Staudacher, A new double-scaling limit of
N = 4 super-Yang-Mills theory and pp-wave strings. Nuclear Phys. B 643, 330 (2002)
[427] M. Kulesh, M. Holschneider, M.S. Diallo, Geophysical wavelet library: Applications of
the continuous wavelet transform to the polarization and dispersion analysis of signals.
Comput. Geosci. 34, 17321752 (2008)
[428] R. Kunze, On the Frobenius reciprocity theorem for square integrable representations.
Pacific J. Math. 53, 465471 (1974)
[429] Y. Kuroda, A. Wada, T. Yamazaki, K. Nagayama, Postacquisition data processing method
for suppression of the solvent signal I. J. Magn. Reson. 84, 604610 (1989)
564 References

[430] Y. Kuroda, A. Wada, T. Yamazaki, K. Nagayama, Postacquisition data processing method

for suppression of the solvent signal II: The weighted first derivative. J. Magn. Reson. 88,
141145 (1990)
[431] G. Kutyniok, D. Labate, Resolution of the wavefront set using continuous shearlets. Trans.
Amer. Math. Soc. 361, 27192754 (2009)
[432] G. Kutyniok, W.-Q. Lim, Compactly supported shearlets are optimally sparse. J. Approx.
Theory 163, 15641589 (2011)
[433] D. Labate, W.-Q. Lim, G. Kutyniok, G. Weiss, Sparse multidimensional representation
using shearlets, in Wavelets XI (San Diego, CA, 2005), ed. by M. Papadakis, A. Laine, M.
Unser. SPIE Proceedings, vol. 5914 (SPIE, Bellingham, WA, 2005), pp. 254262
[434] P. Lahti, J-P. Pellonp, Continuous variable tomographic measurements. Phys. Lett. A
373, 34353438 (2009)
[435] Lamberts projection: see Wikipedia.
http://en.wikipedia.org/wiki/Lambert_azimuthal_equal-area_projection
[436] J-P. Leduc, F. Mujica, R. Murenzi, M.J.T. Smith, Missile-tracking algorithm using target-
adapted spatio-temporal wavelets, in Automatic Object Recognition VII. SPIE Proceedings,
vol. 5914 (SPIE, Bellingham, WA, 1997), pp. 400411
[437] J-P. Leduc, F. Mujica, R. Murenzi, M.J.T. Smith, Spatio-temporal wavelet transforms for
motion tracking, in IEEE ICASSP 1997, vol. 4 (1997), pp 30133016
[438] J-P. Leduc, F. Mujica, R. Murenzi, M.J.T. Smith, Spatio-temporal continuous wavelets
applied to missile warhead detection and tracking, in SPIE VCIP 97, vol. 3024, ed. by
J. Biemond, E.J. Delp (1997), pp. 787798
[439] B. Leistedt, J.D. McEwen, Exact wavelets on the ball. IEEE Trans. Signal Proc. 60, 1564
1589 (2012)
[440] P.G. Lemari, Y. Meyer, Ondelettes et bases hilbertiennes. Rev. Math. Iberoamer. 2, 118
(1986)
[441] C. Lemke, A. Schuck Jr., J-P. Antoine, D. Sima, Metabolite-sensitive analysis of magnetic
resonance spectroscopic signals using the continuous wavelet transform. Meas. Sci.
Technol. 22 (2011). Art. # 114013
[442] J-M. Lvy-Leblond, Galilei group and non-relativistic quantum mechanics. J. Math. Phys.
4, 776-788 (1963)
[443] J-M. Lvy-Leblond, Galilei group and Galilean invariance, in Group Theory and Its
Applications, vol. II, ed. by E.M. Loebl (Academic, New York, 1971), pp. 221299
[444] J-M. Lvy-Leblond, On the conceptual nature of the physical constants. Riv. Nuovo Cim.
7, 187214 (1977)
[445] E.H. Lieb, The classical limit of quantum spin systems. Commun. Math. Phys. 31, 327340
(1973)
[446] G. Lindblad, B. Nagel, Continuous bases for unitary irreducible representations of SU(1,1).
Ann. Inst. H. Poincar 13, 2756 (1970)
[447] W. Lisiecki, Khler coherent states orbits for representations of semisimple Lie groups.
Ann. Inst. H. Poincar 53, 857890 (1990)
[448] A. Lisowska, Moment-based fast wedgelet transform. J. Math. Imaging Vis. 39, 180192
(2011)
[449] H. Liu, L. Peng, Admissible wavelets associated with the Heisenberg group. Pacific J.
Math. 180, 101123 (1997)
[450] E. Livine, S. Speziale, Physical boundary state for the quantum tetrahedron. Class. Quant.
Grav. 25, 085003 (2008)
[451] F. Low, Complete sets of wave packets, in A Passion for Physics Essay in Honor of
Geoffrey Chew, ed. by C. DeTar (World Scientific, Singapore, 1985), pp. 1722
[452] G. Mack, All unitary ray representations of the conformal group SU(2,2) with positive
energy. Commun. Math. Phys. 55, 128 (1977)
[453] G.W. Mackey, Imprimitivity for representations of locally compact groups I. Proc. Nat.
Acad. Sci. 35, 537545 (1949)
References 565

[454] S. Majid, M. Rodriguez-Plaza, Random walk and the heat equation on superspace and
anyspace. J. Math. Phys. 35, 37533760 (1994)
[455] M. Mallalieu, C.R. Stroud Jr., Rydberg wave packets: fractional revivals and classical
orbits, in Coherent States: Past, Present and Future (Proc. Oak Ridge 1993), ed. by D.H.
Feng, J.R. Klauder, M. Strayer (World Scientific, Singapore, 1994), pp. 301314
[456] S.G. Mallat, Multifrequency channel decompositions of images and wavelet models. IEEE
Trans. Acoust. Speech, Signal Proc. 37, 20912110 (1989)
[457] S.G. Mallat, A theory for multiresolution signal decomposition: The wavelet representa-
tion. IEEE Trans. Pattern Anal. Machine Intell. 11, 674693 (1989)
[458] S. Mallat, W.-L. Hwang, Singularity detection and processing with wavelets. IEEE Trans.
Inform. Theory 38, 617643 (1992)
[459] S. Mallat, Z. Zhang, Matching pursuits with time frequency dictionaries. IEEE Trans.
Signal Proc. 41, 33973415 (1993)
[460] S. Mallat, S. Zhong, Wavelet maxima representation, in Wavelets and Applications (Proc.
Marseille 1989) ed. by Y. Meyer (Masson and Springer, Paris and Berlin, 1991), pp. 207
284
[461] V.I. Manko, G. Marmo, E.C.G. Sudarshan, F. Zaccaria, f -oscillators and non-linear
coherent states. Phys. Scr. 55, 528541 (1997)
[462] D. Marinucci, D. Pietrobon, A. Baldi, P. Baldi, P. Cabella, G. Kerkyacharian, P. Natoli, D.
Picard, N. Vittorio, Spherical needlets for CMB data analysis. Mon. Not. R. Astron. Soc.
383, 539545 (2008)
[463] D. Marion, M. Ikura, A. Bax, Improved solvent suppression in one- and two-dimensional
NMR spectra by convolution of time-domain data, J. Magn. Reson. 84, 425430 (1989)
[464] Y. Meyer, Principe dincertitude, bases hilbertiennes et algbres doprateurs. Sminaire
Bourbaki 662 (19851986)
[465] Y. Meyer, Principe dincertitude, bases hilbertiennes et algbres doprateurs. Astrisque
145146, 209223 (1987)
[466] Y. Meyer, H. Xu, Wavelet analysis and chirps. Appl. Comput. Harmon. Anal. 4, 366379
(1997)
[467] L. Michel, Invariance in quantum mechanics and group extensions, in Group Theoretical
Concepts and Methods in Elementary Particle Physics, ed. by F. Grsey (Gordon and
Breach, New York and London, 1964), pp. 135200
[468] J. Mickelsson, J. Niederle, Contractions of representations of the de Sitter groups.
Commun. Math. Phys. 27, 167180 (1972)
[469] M.M. Miller, Convergence of the Sudarshan expansion for the diagonal coherent-state
weight functional. J. Math. Phys. 9, 12701274 (1968)
[470] V. F. Molchanov, Harmonic analysis on homogeneous spaces, in Representation Theory
and Noncommutative Harmonic Analysis II, ed. by A.A. Kirillov (Springer, Berlin, 1995)
[471] M.I. Monastyrsky, A.M. Perelomov, Coherent states and symmetric spaces II. Ann. Inst.
H. Poincar 23, 2348 (1975)
[472] B. Moran, S. Howard, D. Cochran, Positive-operator-valued measures: A general setting
for frames, in Excursions in Harmonic Analysis, vol. 1, 2, ed. by T.D. Andrews, R. Balan,
J.J. Benedetto, W. Czaja, K.A. Okoudjou (Birkhuser, Boston, 2013), pp. 4964
[473] H. Moscovici, Coherent states representations of nilpotent Lie groups. Commun. Math.
Phys. 54, 6368 (1977)
[474] H. Moscovici, A. Verona, Coherent states and square integrable representations. Ann. Inst.
H. Poincar 29, 139156 (1978)
[475] F. Mujica, R. Murenzi, M.J.T. Smith, J-P. Leduc, Robust tracking in compressed image
sequences. J. Electr. Imaging 7, 746754 (1998)
[476] R. Murenzi, Wavelet transforms associated to the n-dimensional Euclidean group with
dilations: Signals in more than one dimension, in Wavelets, Time-Frequency Methods
and Phase Space (Proc. Marseille 1987), ed. by J.-M. Combes, A. Grossmann, P.
Tchamitchian, 2nd edn. (Springer, Berlin, 1990), pp. 239246
566 References

[477] M.A. Muschietti, B. Torrsani, Pyramidal algorithms for LittlewoodPaley decomposi-

tions. SIAM J. Math. Anal. 26, 925943 (1995)
[478] B. Nagel, Generalized eigenvectors in group representations, in Studies in Mathematical
Physics (Proc. Istanbul 1970), ed. by A.O. Barut (Reidel, Dordrecht and Boston, 1970),
pp. 135154
[479] M.A. Namark, Dokl. Akad. Nauk. SSSR 41, 359361 (1943); see also B. Sz-Nagy,
Extensions of linear transformations in Hilbert space which extend beyond this space.
Appendix to F. Riesz, B. Sz-Nagy, Functional Analysis (Frederick Ungar, New York, 1960)
[480] F.J. Narcowich, P. Petrushev, J.D. Ward, Localized tight frames on spheres. SIAM J. Math.
Anal. 38, 574594 (2006)
[481] M. Nauenberg, Quantum wave packets on Kepler elliptic orbits. Phys. Rev. A 40, 1133
1136 (1989)
[482] M. Nauenberg, C. Stroud, J. Yeazell, The classical limit of an atom. Scient. Amer. 270,
2429 (1994)
[483] H. Neumann, Transformation properties of observables. Helv. Phys. Acta 45, 811819
(1972)
[484] U. Niederer, The maximal kinematical invariance group of the free Schrdinger equation.
Helv. Phys. Acta 45, 802881 (1972)
[485] M.M. Nieto, L.M. Simmons Jr., Coherent states for general potentials I: Formalism, II:
Confining one-dimensional examples, III: Nonconfining one-dimensional examples. Phys.
Rev. D 20, 13211331, 13321341, 13421350 (1979)
[486] M.M. Nieto, L.M. Simmons Jr., Coherent states for general potentials. Phys. Rev. Lett. 41,
207210 (1987)
[487] A. Odzijewicz, On reproducing kernels and quantization of states. Commun. Math. Phys.
114, 577597 (1988)
[488] A. Odzijewicz, Coherent states and geometric quantization. Commun. Math. Phys. 150,
385413 (1992)
[489] A. Odzijewicz, Quantum algebras and q-special functions related to coherent states maps
of the disc. Commun. Math. Phys. 192, 183215 (1998)
[490] A. Odzijewicz, M. Horowski, A. Tereszkiewicz, Integrable multi-boson systems and
orthogonal polynomials. J. Phys. A: Math. Gen. 34, 43534376 (2001)
[491] G. lafsson, B. rsted, The holomorphic discrete series for affine symmetric spaces. J.
Funct. Anal. 81, 126159 (1988)
[492] G. lafsson, H. Schlichtkrull, Representation theory, Radon transform and the heat
equation on a Riemannian symmetric space. Contemp. Math. 449, 315344 (2008)
[493] E. Onofri, A note on coherent state representations of Lie groups. J. Math. Phys. 16, 1087
1089 (1975)
[494] E. Onofri, Dynamical quantization of the Kepler manifold. J. Math. Phys. 17, 401408
(1976)
[495] D. Oriti, R. Pereira, L. Sindoni, Coherent states in quantum gravity: A construction based
on the flux representation of loop quantum gravity. J. Phys. A: Math. Theor. 45, 244004
(2012)
[496] L.C. Papaloucas, J. Rembielinski, W. Tybor, Vectorlike coherent states with noncompact
stability group. J. Math. Phys 30, 24062410 (1989)
[497] Z. Pasternak-Winiarski, On the dependence of the reproducing kernel on the weight of
integration. J. Funct. Anal. 94, 110134 (1990)
[498] Z. Pasternak-Winiarski, On reproducing kernels for holomorphic vector bundles, in
Quantization and Infinite Dimensional Systems (Proc. Biaowieza, Poland, 1993), ed. by J-
P. Antoine, S.T. Ali, W. Lisiecki, I.M. Mladenov, A. Odzijewicz (Plenum Press, New York
and London, 1994), pp. 109112
[499] T. Paul, Affine coherent states and the radial Schrdinger equation I, preprint CPT-
84/P.1710 (1984) (unpublished)
References 567

[500] T. Paul, K. Seip, Wavelets in quantum mechanics, in Wavelets and Their Applications, ed.
by M.B. Ruskai, G. Beylkin, R. Coifman, I. Daubechies, S. Mallat, Y. Meyer, L. Raphael
(Jones and Bartlett, Boston, 1992), pp. 303322
[501] A.M. Perelomov, On the completeness of a system of coherent states. Theor. Math. Phys.
6, 156164 (1971)
[502] A.M. Perelomov, Coherent states for arbitrary Lie group. Commun. Math. Phys. 26, 222
236 (1972)
[503] A.M. Perelomov, Coherent states and symmetric spaces. Commun. Math. Phys. 44, 197
210 (1975)
[504] M. Perroud, Projective representations of the Schrdinger group. Helv. Phys. Acta 50, 233
252 (1977)
[505] J. Phillips, A note on square-integrable representations. J. Funct. Anal. 20, 8392 (1975)
[506] D. Pietrobon, P. Baldi, D. Marinucci, Integrated Sachs-Wolfe effect from the cross
correlation of WMAP3 year and the NRAO VLA sky survey data: New results and
constraints on dark energy. Phys. Rev. D 74, 043524 (2006)
[507] W.W.F. Pijnappel, A. van den Boogaart, R. de Beer, D. van Ormondt, SVD-based
quantification of magnetic resonance signals, J. Magn. Reson. 97, 122134 (1992)
[508] V. Pop, D. Rosca, Generalized piecewise constant orthogonal wavelet bases on 2D-
domains. Appl. Anal. 90, 715723 (2011)
[509] G. Pschl, E. Teller, Bemerkungen zur Quantenmechanik des anharmonischen Oszillators.
Z. Physik 83, 143151 (1933)
[510] D. Potts, G. Steidl, M. Tasche, Kernels of spherical harmonics and spherical frames, in
Advanced Topics in Multivariate Approximation, ed. by F. Fontanella, K. Jetter, P.J. Laurent
(World Scientific, Singapore, 1996), pp. 287301
[511] E. Prugovecki, Consistent formulation of relativistic dynamics for massive spin-zero
particles in external fields. Phys. Rev. D 18, 36553673 (1978) (Appendix C)
[512] E. Prugovecki, Relativistic quantum kinematics on stochastic phase space for massive
particles. J. Math.Phys. 19, 22612270 (1978)
[513] C. Quesne, Coherent states of the real symplectic group in a complex analytic parametriza-
tion I, II. J. Math. Phys. 27, 428441, 869878 (1986)
[514] C. Quesne, Generalized vector coherent states of sp(2N, R) vector operators and of
sp(2N, R) u(N) reduced Wigner coefficients. J. Phys. A: Math. Gen. 24, 26972714
(1991)
[515] J.M. Radcliffe, Some properties of spin coherent states. J. Phys. A: Math. Gen. 4, 313323
(1971)
[516] A. Rahimi, A. Najati, Y.N. Dehghan, Continuous frames in Hilbert spaces. Methods Funct.
Anal. Topol. 12, 170182 (2006)
[517] H. Rauhut, M. Rsler, Radial multiresolution in dimension three. Constr. Approx. 22, 193
218 (2005)
[518] J.H. Rawnsley, Coherent states and Khler manifolds. Quart. J. Math. Oxford 28(2), 403
415 (1977)
[519] J. Renaud, The contraction of the SU(1,1) discrete series of representations by means of
coherent states. J. Math. Phys. 37, 31683179 (1996)
[520] A. Rnyi, Representations for real numbers and their ergodic properties. Acta Math. Acad.
Sci. Hungary 8(34), 477493 (1957)
[521] D. Robert, La cohrence dans tous ses tats. SMF Gazette 132 (2012)
[522] J.E. Roberts, The Dirac bra and ket formalism. J. Math. Phys. 7, 10971104 (1966)
[523] J.E. Roberts, Rigged Hilbert spaces in quantum mechanics. Commun. Math. Phys. 3, 98
119 (1966)
[524] S. Roques, F. Bourzeix, K. Bouyoucef, Soft-thresholding technique and restoration of
3C273 jet. Astrophys. Space Sci. Nr. 239, 297304 (1996)
[525] D. Rosca, Haar wavelets on spherical triangulations, in Advances in Multiresolution for
Geometric Modelling, ed. by N.A. Dogson, M.S. Floater, M.A. Sabin (Springer, Berlin,
2005), pp. 407419
568 References

[526] D. Rosca, Locally supported rational spline wavelets on the sphere. Math. Comput. 74,
18031829 (2005)
[527] D. Rosca, Wavelets defined on closed surfaces, J. Comput. Anal. Appl. 8, 121132 (2006)
[528] D. Rosca, Weighted Haar wavelets on the sphere. Int. J. Wavelets Multiresol. Inf. Proc. 5,
501511 (2007)
[529] D. Rosca, Wavelet bases on the sphere obtained by radial projection. J. Fourier Anal. Appl.
13, 421434 (2007)
[530] D. Rosca, Piecewise constant wavelets on triangulations, obtained by 13 splitting. Int. J.
Wavelets Multiresolut. Inf. Process. 6, 209222 (2008)
[531] D. Rosca, On a norm equivalence on L2 (S2 ). Results Math. 53, 399405 (2009)
[532] D. Rosca, New uniform grids on the sphere. Astron. Astrophys. 520 (2010). Art. A63
[533] D. Rosca, Uniform and refinable grids on elliptic domains and on some surfaces of
revolution. Appl. Math. Comput. 217, 78127817 (2011)
[534] D. Rosca, Wavelet analysis on some surfaces of revolution via area preserving projection.
Appl. Comput. Harmon. Anal. 30, 262272 (2011)
[535] D. Rosca, J-P. Antoine, Locally supported orthogonal wavelet bases on the sphere via
stereographic projection. Math. Probl. Eng. 2009, 124904 (2009)
[536] D. Rosca, J-P. Antoine, Constructing wavelet frames and orthogonal wavelet bases on the
sphere, in Recent Advances in Signal Processing, ed. by S. Miron (IN-TECH, Vienna,
Austria, and Rijeka, Croatia, 2010), pp. 5976
[537] D. Rosca, G. Plonka, Uniform spherical grids via area preserving projection from the cube
to the sphere. J. Comput. Appl. Math. 236, 10331041 (2011)
[538] D. Rosca, G. Plonka, An area preserving projection from the regular octahedron to the
sphere. Results Math. 63, 429444 (2012)
[539] H. Rossi, M. Vergne, Analytic continuation of the holomorphic discrete series for a semi-
simple Lie group. Acta Math. 136, 159 (1976)
[540] D.J. Rotenberg, Application of Sturmian functions to the Schroedinger three-body prob-
lem: Elastic e+ H scattering. Ann. Phys. (NY) 19, 262278 (1962)
[541] C. Rovelli, Zakopane lectures on loop gravity, in Proceedings of 3rd Quantum Gravity
and Quantum Geometry School, 28 Feb13 March 2011 (Zakopane, Poland, 2011).
arXiv:1102.3660v5
[542] C. Rovelli, S. Speziale, A semiclassical tetrahedron, Class. Quant. Grav. 23, 58615870
(2006)
[543] D.J. Rowe, Coherent state theory of the noncompact symplectic group. J. Math. Phys. 25,
26622271 (1984)
[544] D.J. Rowe, Microscopic theory of the nuclear collective model. Rep. Prog. Phys. 48, 1419
1480 (1985)
[545] D.J. Rowe, Vector coherent state representations and their inner products. J. Phys. A: Math.
Gen. 45, 244003 (2012). (This paper belongs to the special issue [38])
[546] D.J. Rowe, J. Repka, Vector-coherent-state theory as a theory of induced representations.
J. Math. Phys. 32, 26142634 (1991)
[547] D.J. Rowe, G. Rosensteel, R. Gilmore, Vector coherent state representation theory. J. Math.
Phys. 26, 27872791 (1985)
[548] A. Royer, Phase states and phase operators for the quantum harmonic oscillator. Phys. Rev.
A 53, 70108 (1996)
[549] J. Saletan, Contraction of Lie groups. J. Math. Phys. 2, 121 (1961)
[550] G. Saracco, A. Grossmann, P. Tchamitchian, Use of wavelet transforms in the study of
propagation of transient acoustic signals across a plane interface between two homoge-
neous media, in Wavelets, Time-Frequency Methods and Phase Space (Proc. Marseille
1987), ed. by J.-M. Combes, A. Grossmann, P. Tchamitchian, 2nd edn. (Springer, Berlin,
1990), pp. 139146
[551] P. Schrder, W. Sweldens, Spherical wavelets: Efficiently representing functions on the
sphere, in Computer Graphics Proceedings (SIGGRAPH95) (ACM Siggraph, Los Angeles,
1995), pp. 161175
References 569

[552] E. Schrdinger, Der stetige bergang von der Mikro- zur Makromechanik. Naturwiss. 14,
664666 (1926)
[553] S. Scodeller, . Rudjord, F.K. Hansen, D. Marinucci, D. Geller, A. Mayeli, Introducing
Mexican needlets for CMB analysis: Issues for practical applications and comparison with
standard needlets. Astrophys. J. 733 (2011). Art. 121
[554] H. Scutaru, Coherent states and induced representations. Lett. Math. Phys. 2, 101107
(1977)
[555] B. Simon, Distributions and their Hermite expansions. J. Math. Phys. 12, 140148 (1971)
[556] B. Simon, The classical moment problem as a self-adjoint finite difference operator. Adv.
Math. 137, 82203 (1998)
[557] R. Simon, E.C.G. Sudarshan, N. Mukunda, Gaussian pure states in quantum mechanics
and the symplectic group. Phys. Rev. A37, 30283038 (1988)
[558] S. Sivakumar, Studies on nonlinear coherent states. J. Opt. B Quant, Semiclass. Opt. 2,
R61R75 (2000)
[559] E. Slezak, A. Bijaoui, G. Mars, Identification of structures from galaxy counts. Use of the
wavelet transform. Astron. Astroph. 227, 301316 (1990)
[560] A. Solomon, A characteristic functional for deformed photon phenomenology. Phys. Lett.
A 196, 2934 (1994)
[561] S.B. Sontz, Paragrassmann algebras as quantum spaces, Part I: Reproducing kernels, in
Geometric Methods in Physics. XXXI Workshop 2012 (Trends in Mathematics, ed. by
P. Kielanowski et al., Birkhuser Verlag, Basel, 2013), pp. 4763
[562] S.B. Sontz, Paragrassmann algebras as quantum spaces, Part II: Toeplitz operators. J. Oper.
Th. (2013, to appear); arXiv:1205.5493
[563] S.B. Sontz, A reproducing kernel and Toeplitz operators in the quantum plane. Preprint
(2013); arXiv:1305.6986 [math-ph]
[564] S.B. Sontz, Toeplitz quantization of an algebra with conjugation. Preprint (2013);
arXiv:1308.5454 [math-ph]
[565] M. Spera, On a generalized Uncertainty Principle, coherent states, and the moment map.
J. Geom. Phys. 12, 165182 (1993)
[566] J.-L. Starck, E.J. Cands, D.L. Donoho, The curvelet transform for image denoising. IEEE
Trans. Image Proc. 11, 670684 (2002)
[567] J.-L. Starck, D.L. Donoho, E. J. Cands, Astronomical image representation by the curvelet
transform. Astron. Astroph. 398, 785800 (2003)
[568] J.-L. Starck, Y. Moudden, P. Abrial, M. Nguyen, Wavelets, ridgelets and curvelets on the
sphere. Astron. Astroph. 446, 11911204 (2006)
[569] M.B. Stenzel, The Segal-Bargmann transform on a symmetric space of compact type. J.
Funct. Analysis 165, 4458 (1994)
[570] E.C.G. Sudarshan, Equivalence of semiclassical and quantum mechanical descriptions of
statistical light beams. Phys. Rev. Lett. 10, 277279 (1963)
[571] A. Suvichakorn, H. Ratiney, A. Bucur, S. Cavassila, J-P. Antoine, Toward a quantitative
analysis of in vivo magnetic resonance proton spectroscopic signals using the continuous
Morlet wavelet transform. Meas. Sci. Technol. 20 (2009). Art. #104029
[572] W. Sweldens, The lifting scheme: A custom-design construction of biorthogonal wavelets.
Applied Comput. Harmon. Anal. 3, 11861200 (1996)
[573] W. Sweldens, The lifting scheme: A construction of second generation wavelets. SIAM J.
Math. Anal. 29, 511546 (1998)
[574] F.H. Szafraniec, The reproducing kernel Hilbert space and its multiplication operators,
in Complex Analysis and Related Topics, ed. by E. Ramirez de Arellano et al. Operator
Theory: Advances and Applications, vol. 114 (Birkhuser, Basel, 2000), pp. 254263
[575] F.H. Szafraniec, Multipliers in the reproducing kernel Hilbert space, subnormality and
noncommutative complex analysis, in Reproducing Kernel Spaces and Applications, ed.
by D. Alpay. Operator Theory: Advances and Applications, vol. 143 (Birkhuser, Basel,
2003), pp. 313331
570 References

[576] R. Takahashi, Sur les reprsentations unitaires des groupes de Lorentz gnraliss. Bull.
Soc. Math. France 91, 289433 (1963)
[577] M.C. Teich, B.E.A. Saleh, Squeezed states of light. Quantum Opt. 1, 152191 (1989)
[578] R. Terrier, L. Demanet, I.A. Grenier, J-P. Antoine, Wavelet analysis of EGRET data, in
Proceedings of the 27th International Cosmic Ray Conference (ICRC 2001) (Copernicus
Gesellschaft, DE 2001), pp. 29232926
[579] T. Thiemann, Gauge field theory coherent states (GCS) 1: General properties. Class. Quant.
Grav. 18, 20252064 (2001)
[580] T. Thiemann, O. Winkler, Gauge field theory coherent states (GCS) 2: Peakedness
properties. Class. Quant. Grav. 18, 25612636 (2001)
[581] T. Thiemann, O. Winkler, Gauge field theory coherent states (GCS) 3: Ehrenfest theorems.
Class. Quant. Grav. 18, 46294682 (2001)
[582] T. Thiemann, O. Winkler, Gauge field theory coherent states (GCS) 4: Infinite tensor
product and thermodynamical limit. Class. Quant. Grav. 18, 49975054 (2001)
[583] K. Thirulagasanthar, S.T. Ali, Regular subspaces of a quaternionic Hilbert space from
quaternionic Hermite polynomials and associated coherent states. J. Math. Phys. 54,
013506 (2013)
[584] K. Thirulogasanthar, G. Honnouvo, A. Krzyzak, Coherent states and Hermite polynomials
on quaternionic Hilbert spaces. J. Phys. A: Math. Theor. 43, 385205 (2010)
[585] Tokamak : see Wikipedia. http://en.wikipedia.org/wiki/Tokamak
[586] B. Torrsani, Wavelets associated with representations of the affine Weyl-Heisenberg
group. J. Math. Phys. 32, 12731279 (1991)
[587] B. Torrsani, Time-frequency representation: Wavelet packets and optimal decomposition.
Ann. Inst. H. Poincar 56, 215234 (1992)
[588] B. Torrsani, Position-frequency analysis for signals defined on spheres. Signal Proc. 43,
341346 (2005)
[589] D.A. Trifonov, Generalized intelligent states and squeezing. J. Math. Phys. 35, 22972308
(1994)
[590] A.S. Trushechkin, I.V. Volovich, Localization properties of squeezed quantum states in
nanoscale space domains, preprint (2013). arXiv:1304.6277v1 [quant-ph]
[591] M. Unser, N. Chenouard, A unifying parametric framework for 2D steerable wavelet
transforms. SIAM J. Imaging Sci. 6, 102135 (2013)
[592] P. Vandergheynst, J.-F. Gobbers, Directional dyadic wavelet transforms: Design and
algorithms. IEEE Trans. Image Proc. 11, 363372 (2002)
[593] P. Vandergheynst, J-P. Antoine, E. Van Vyve, A. Goldberg, I. Doghri, Modelling and
simulation of an impact test using wavelets, analytical and finite element models. Int. J.
Solids Struct. 38, 54815508 (2001)
[594] L. Vanhamme, R.D. Fierro, S. Van Huffel, R. de Beer, Fast removal of residual water in
proton spectra. J. Magn. Reson. 132, 197203 (1998)
[595] J. Ville, Thorie et applications de la notion de signal analytique. Cbles et Trans. 2, 6174
(1948)
[596] J. Voisin, On some unitary representations of the Galilei group I: Irreducible representa-
tions. J. Math. Phys. 6, 15191529 (1965)
[597] A. Vourdas, Analytic representations in quantum mechanics. J. Phys. A: Math. Gen. 39,
R65R141 (2006)
[598] D.F. Walls, Squeezed states of light. Nature 306, 141146 (1983)
[599] Y.K. Wang, F.T. Hioe, Phase transition in the Dicke maser model. Phys. Rev. A 7, 831836
(1973)
[600] P.S.P. Wang, J. Yang, A review of wavelet-based edge detection methods. Int. J. Patt.
Recogn. Artif. Intell. 26, 1255011 (2012)
[601] I. Weinreich, A construction of C1 -wavelets on the two-dimensional sphere. Appl. Comput.
Harmon. Anal. 10, 126 (2001)
[602] H. Weyl, Quantenmechanik und Gruppentheorie. Z. Phys. 46, 146 (1927)
References 571

[603] Y. Wiaux, L. Jacques, P. Vandergheynst, Correspondence principle between spherical and

Euclidean wavelets. Astrophys. J. 632, 1528 (2005)
[604] Y. Wiaux, L. Jacques, P. Vielva, P. Vandergheynst, Fast directional correlation on the sphere
with steerable filters. Astrophys. J. 652, 820832 (2006)
[605] Y. Wiaux, J.D. McEwen, P. Vandergheynst, O. Blanc, Exact reconstruction with directional
wavelets on the sphere. Mon. Not. R. Astron. Soc. 388, 770788 (2008)
[606] W.M. Wieland, Complex Ashtekar variables and reality conditions for Holsts action. Ann.
Henri Poincar 13, 425448 (2012)
[607] E.P. Wigner, On the quantum correction for thermodynamic equilibrium. Phys. Rev. 40,
749759 (1932)
[608] R.M. Willette, R.D. Nowak, Platelets: A multiscale approach for recovering edges and
surfaces in photon-limited medical imaging. IEEE Trans. Med. Imaging 22, 332350
(2003)
[609] W. Wisnoe, P. Gajan, A. Strzelecki, C. Lempereur, J.-M. Math, The use of the two-
dimensional wavelet transform in flow visualization processing, in Progress in Wavelet
Analysis and Applications (Proc. Toulouse 1992), ed. by Y. Meyer, S. Roques (Ed.
Frontires, Gif-sur-Yvette 1993), pp. 455458
[610] K. Wdkiewicz, On the quantum mechanics of squeezed states. J. Modern Optics 34, 941
948(1987)
[611] J.A. Yeazell, M. Mallalieu, C.R. Stroud Jr., Observation of the collapse and revival of a
Rydberg electronic wave packet. Phys. Rev. Lett. 64, 20072010 (1990)
[612] J.A. Yeazell, C.R. Stroud Jr., Observation of fractional revivals in the evolution of a
Rydberg atomic wave packet. Phys. Rev. A 43, 51535156 (1991)
[613] H.P. Yuen, Two-photon coherent states of the radiation field. Phys. Rev. A 13, 22262243
(1976)
[614] J. Zak, Balian-Low theorem for Landau levels. Phys. Rev. Lett. 79, 533536 (1997)
[615] J. Zak, Orthonormal sets of localized functions for a Landau level. J. Math. Phys. 39, 4195
4200 (1998) and references quoted there
[616] A.A. Zakharova, On the properties of generalized frames. Math. Notes 83, 190200 (2008)
[617] W.-M. Zhang, D.H. Feng, R. Gilmore, Coherent states: Theory and some applications. Rev.
Mod. Phys. 26, 867927 (1990)
[618] I. Zlatev, W.-M. Zhang, D.H. Feng, Possibility that Schrdingers conjecture for the
hydrogen atom coherent states is not attainable. Phys. Rev. A 50, R1973R1975 (1994)
[619] W.H. Zurek, Decoherence, einselection, and the quantum origins of the classical. Rev. Mod.
Phys. 75, 715775 (2003)
[620] W.H. Zurek, S. Habib, J.P. Paz, Coherent states via decoherence. Phys. Rev. Lett. 70, 1187
1190 (1993)
[621] K. Zyczkowski, Squeezed states in a quantum chaotic system. J. Phys. A: Math. Theor. 22,
L1147L1151 (1989)
Index

Symbols coherent states (CS), 124

-selection, 119 action-angle, 157, 325
atomic, 243
Barut-Girardello, 3
A
of SU(1,1), 299
action of semigroup on a lattice, 395
canonical, 2, 15, 19, 169
algorithm
classical theory, 174
best basis, 384
covariant, 166, 180
fast wavelet, 385, 439
deformed, 146
matching pursuit, 498, 521
for infinite square well and Pschl-Teller
pyramidal, 382, 393
potentials, 158
angular selectivity, 427, 443
for semidirect products, 260
Apollonius, 477
Gazeau-Klauder, 151
applications of CS
generalized, 11
in atomic physics, 175, 243
Gilmore-Perelomov, 10, 168
in nuclear physics, 178
of SU(1,1), 81, 296, 297
applications of CWT
HilbertC -modules, 164
in 1-D, 368
holomorphic, 140
in 2-D, 429
nonholomorphic, 151
applications of DWT
nonlinear, 146
in 1-D, 385
of affine Galilei group, 505
in 2-D, 436
of affine Poincar group, 509
area preserving projections, 483
of affine Weyl-Heisenberg group GaWH ,
497
B of compact semisimple Lie groups, 174
bundle of Euclidean group E(n), 266
(co)tangent, 251 of Galilei group G(1, 1), 290
normal, 251 of isochronous Galilei group, 237
parallel, 268 of non-semisimple Lie groups, 179
of noncompact semisimple Lie groups, 177
C of Poincar group P+ (1, 1), 283
canonical commutation relations (CCR), 16 massless case, 288
coboundaries, 403 of Poincar group P+ (1, 3), 279
cochains, 402 of Schrdinger group, 508
cocycle, 63, 69, 72, 260, 393, 403 quasi-coherent states, 186
cocycle equation, 398, 404 quaternionic, 164

S.T. Ali et al., Coherent States, Wavelets, and Their Generalizations, Theoretical 573
and Mathematical Physics, DOI 10.1007/978-1-4614-8535-3,
Springer Science+Business Media New York 2014
574 Index

coherent states (CS) (cont.) discrete WT

spin or SU(2), 175 in 1-D, 379
square integrable covariant, 166, 180, 264 applications, 385
vector (VCS), 73, 108, 112, 170 generalizations, 384
weighted, 184, 523 in 2-D, 434
of affine Weyl-Heisenberg group GaWH , applications, 436
497 on the sphere S2 , 471
of Poincar group P+ (1, 1), 284 Duflo-Moore operator, 215, 219, 222, 224, 226,
cohomology group, 393, 403 239, 286
coboundaries, 403 dyadic lattice, 381, 519
cochains, 402
cocycle, 403
continuous WT E
in n-D, 414 evaluation map, 25, 121, 125
in 1-D, 353
applications, 368
continuous wavelet packets, 385 F
discretization, 365, 404, 407 Fibonacci numbers, 388
localization properties, 363 filters, 382
in 2-D, 414 compatible, 399
applications, 429 on Z p , 393
as a symmetry scanner, 432 pseudo-QMF, 387
continuous wavelet packets, 439 QMF, 382
interpretation, 426 Fourier transform, 351, 412
representations, 426 relativistic, 223
resolving power, 427, 443 fractals, 364
on conic sections, 477 frame, 40
on manifolds, 510 affine Weyl-Heisenberg CS, 497
on surfaces of revolution, 482 discrete, 54, 365
on the sphere Sn , 479 for affine Poincar group, 521
on the sphere S2 , 458 for affine Weyl-Heisenberg group, 521
Euclidean limit, 462 for Poincar group P+ (1, 1), 527
on the torus T2 , 481 for Poincar group P+ (1, 3), 534
contourlet transform, 450 for semidirect products, 524
coorbit theory, 455 Gabor or canonical CS, 516
curvelet transform wavelet, 519
continuous, 447 discretization, 515, 519
discrete, 449 finite, 334
for semidirect product, 269
fusion, 57
D Galilei CS, 291
decomposition Littlewood-Paley, 442
Cartan, 76, 96 Poincar CS, 279
Gauss, 97, 459 tight, 43, 356
Iwasawa, 97, 459, 479 frame operator, 40
dictionary, 455, 498 free orbit, 217
dilation
natural, 396
pseudodilation G
associated, 398 Gabor analysis, 5
on Z p , 392 Gabor transform, 349
principal, 397 gaborettes, 349
stereographic, 458 gaussons, 17, 246
Index 575

golden mean , 387, 432 holomorphic map, 129

-Haar basis, 388 hypergroup, 423
-integers, 388
group (abstract)
direct product, 98 I
extension, 99 intermittency measure, 365
hypergroup, 423
locally compact abelian (LCA), 86
lattice in, 88 L
sampling in, 89 Lie algebra, 91
semidirect product, 98, 217 contraction, 101, 293
unimodular, 62 root, 91
group (explicit) weight, 92
GL(n, R), 217 Lie group, 93
SL(2, R) SOo (1, 2) SU(1, 1), 3 (co)adjoint action, 94, 251, 253
affine, 12, 351 of Sp(2n, R), 248
affine Galilei, 504 coadjoint orbit, 95, 252, 297
affine Poincar = SIM(n, 1), 509 of SIM(n), 415
affine Poincar = SIM(1,1), 222 contraction, 102
affine Weyl-Heisenberg GaWH , 495 SOo (3, 1) to SIM(2), 462
Anti-de Sitter SOo (1, 2), 292 SU(1, 1) to P+ (1, 1), 294
Bessel-Kingman hypergroup, 423, 437 exponential map, 94
connected affine or ax + b, 12, 67, 209, 355 lifting scheme, 384
discrete Weyl-Heisenberg GNdWH , 518 localization, 358, 363
Euclidean E(2), E(n), 266 localization operators, 25
extended Heisenberg, 503
Gut (upper-triangular matrices), 63
Galilei, 231, 291 M
Galilei-Schrdinger, 506 manifold
isochronous Galilei, 233 Khler, 34
metaplectic Mp(2n, R), 246 symplectic, 34, 251
Poincar P+ (1, 1), 282 measure
Poincar P+ (1, 3), 271 (quasi-)invariant, 62, 78, 180
Schrdinger, 503 on cotangent bundle, 256
shearlet groups, 450, 454, 503 Borel, 38
similitude group SIM(n), 225, 412 Haar, 204
similitude group SIM(2), 221 positive operator-valued (POV), 25, 39
Stockwell, 503 commutative, 40
symplectic Sp(2n, R), 246 examples of, 41
Weyl-Heisenberg GWH , 21, 22, 246, 511 Namark extension theorem, 42
projection-valued (PV), 40
scale-angle, 432
H minimal uncertainty states, 16
Hankel transform, 437 model
Hilbert space Hepp-Lieb, 175
direct integral of, 116 nuclear collective, 178
measurable field of, 115, 121 modular function, 62
reproducing kernel, 10 molecule
reproducing kernel , 180, 206, 227, 230 curvelet, 456
reproducing kernel , 23, 108, 115, 117, parabolic, 456
121, 126, 130 shearlet, 456
rigged , 183 moment problem, 143
rigged , 48 motion analysis , 499
576 Index

multiresolution analysis, 380, 393, 395, 400, square integrable, 204, 206, 219, 260, 352,
484 413, 451
Haar, 402 square integrable mod (H), 167
on the sphere S2 , 471 square integrable mod (H, ), 167, 180,
spline, 402 227, 264, 269, 290, 460
reproducing kernel, 10, 80, 124, 125, 180, 206,
227, 230, 356, 415
N holomorphic, 138
NMR spectroscopy, 370 square integrable, 134
resolution of the identity, 3, 20, 81, 204, 206,
227, 230, 241, 264, 356, 415
O resolution operator, 40, 180
observables, 307 for Poincar P+ (1, 3) CS, 275
orthogonality relations, 212, 219, 228, 238, ridgelet transform, 445
285, 286 ridges in WT, 367
overcomplete set, 24

S
P sampling, 391, 407, 536
phase space, 25, 273, 282, 284, 289291, 297, in LCA group, 89
362 scaling function, 380, 405
polynomials Schrdinger, 1, 506
Hermite, 149 section, 4, 63, 166, 180
Pollaczek, 149 affine, 274
POV function, 108, 122, 126 affine admissible, 267
POV measure (POVM), 39, 320 Galilean, 274, 280, 285, 290
pseudodilations, 392 Lorentz, 280, 285
principal, 256, 280
quasi-section, 186, 289
Q symmetric, 281, 285
quantization, 35, 306, 518 semi-frame, 46, 183
bounded motions, 342 continuous, 183
coherent state (CS), 309 discrete, 54
covariant integral, 314 fusion, 60
Weyl-Heisenberg, 316, 323 lower, 46, 49, 462
frame, 333 upper, 46, 49
integral, 313 shearlet transform
motion on the circle, 339 continuous, 452, 454
quantizable distributions, 328 discrete, 453
quantizable functions, 325 Short-Time Fourier Transform, 349
skeleton of WT, 367
space
R coset or homogeneous, 4
reconstruction coset or homogeneous , 166, 180
formula, 208, 215, 356, 415 coset or homogeneous , 22, 62
operator, 356, 360, 361 Fock-Bargmann, 28, 296
representation Hardy, 355
discrete series, 204 Krein, 287
of SU(1,1), 76 square integrability, 9
Fock-Bargmann, 28 squeezed states, 17, 246
induced, 70, 227, 230, 260, 264 subband coding scheme, 383, 400
of SU(1,1), 78 subspace
of semidirect products, 259 V -admissible, 229
regular, 83, 206, 207 cyclic, 171
Index 577

symbols, 6 1-D discrete (DWT), 379

system of covariance, 72 2-D continuous (CWT), 423
system of imprimitivity, 72 2-D discrete (DWT), 434
integer, 384
local, 489
T of distributions, 361
tempered distributions, 357 of Schwartz functions, 360
theorem on Z p , 391, 392
Balian-Low , 517 on a graph, 493
Mackey imprimitivity , 72 wavelet(s), 5, 211, 352
Namark extension , 42 -wavelets, 387
Pontrjagin duality , 88 n-D Mexican hat, 416
Schur lemmas, 84 n-D Morlet, 417
time-frequency representation, 348 1-D Mexican hat, Marr, 353
transform 1-D Morlet, 354
Berezin, 324 2-D Cauchy, 419
contourlet, 450 3-D Cauchy, 420
curvelet, 447 on Z p , 390
Gabor, 349 pseudodilations, 392
ridgelet, 445 on LCA group, 394, 406
shearlet, 452 cohomological interpretation, 402
wavelet, 349 compatible filters, 395, 400
wedgelet, 455 algebraic, 387
Zak, 518 Cauchy, 418
Cauchy-Paul, 354, 419, 425
conical, 418
U difference, 417, 462
uncertainty relations, 16, 423, 516 directional, 416, 440
uniform grids on surfaces of revolution, 486 kinematical, 499
local, 488
metabolite-based, 355, 374
V minimal uncertainty, 425
vector multiselective, 440
-admissible, 227 on a graph, 493
admissible, 204, 352, 413, 451 on the sphere S2 , 458
admissible mod (H, ), 167, 291, 460 on the sphere Sn , 479
on the torus T2 , 481
steerable, 439
W von Mises, 440
wavefront set, 452 wedgelet transform, 455
wavelet bases Wigner function, 322
biorthogonal, 384 Wigner-Ville transform, 349
orthonormal, 381 Windowed Fourier or Gabor transform, 349
wavelet packets, 384
continuous, 439
wavelet transform (WT), 349
n-D continuous (CWT), 414 Z
1-D continuous (CWT), 353 Zak transform, 518