Академический Документы
Профессиональный Документы
Культура Документы
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
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
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
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
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.
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
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
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,
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
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
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
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
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)
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:
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:
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)),
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
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
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
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
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)
|z|2
|z = exp ( + za ) |0
2
1 zn
= exp ( |z|2 ) |n, (2.14)
2 n=0 n!
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,
where E is a 3-vector valued function of x, giving the observed field strength at the
point x. Let be the density matrix,
(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)
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
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
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)
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
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
(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 ,
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
dq dp
d (q, p) = . (2.35)
2
The function
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,
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
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
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 /
2. Positivity,
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
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
(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)),
implying that the vectors in HK are all bounded functions. More importantly, this
also shows that the linear map
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)
Note that
aK ( ) = | (q,p) (q,p) | d (q, p)
= PK , (2.56)
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)
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
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
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
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!
f
(a f )(z) = (z), (a f )(z) = z f (z), f Hhol . (2.78)
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) ,
( , z) ei D(z) , (2.83)
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
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
un = S( )un , n = 0, 1, 2, . . . , (2.93)
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
|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
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!
|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.
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
(z , z) = z z, (2.108)
1 2 (z, z)
= dz dz. (2.109)
i z z
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
we get
= z , (2.113)
2.7 Outlook
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.
(0)
/ = 0. (3.1)
(iJ i ) = (i ). (3.2)
iJ
If in addition, satisfies
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
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
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
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
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
N
1
( ) = 2n n |a( )n , (3.13)
n=1
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
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)
= W H.
PH (3.18)
3.2 The Case of a Tight Frame 43
) | B(X), W H}
S = {P( (3.19)
= 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
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
) )(x) = (x),
(P(
H, (3.23)
44 3 Positive Operator-Valued Measures and Frames
= 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
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
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.
(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
P := W A1W = W W+ ,
= 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
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
W f |W f = f | f H , f , f RA .
3.3 Frames and Semi-frames 47
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
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
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
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)
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 , ...,
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.
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
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
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
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
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,
(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
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 | }.
Af = j | f j .
j
3.4 Discrete Frames and Semi-frames 55
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)
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
W = I.
converges strongly in H, i.e., W
Proof. That {
j } is the frame described in the statement results from the equalities
= | j |(W W )1 |2
j
= W (W W )1 2 = |(W W )1
| j j | = | j j | = I. (3.46)
j j
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 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
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
The family {k } is a weighted frame in H if there are positive weights v2k > 0
such that
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.
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
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.
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
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:
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
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
3.4.5 Discretization
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].)
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)
Gx := {y = gx : g G} X. (4.2)
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 ).
g ( ) = (g ), B(X). (4.6)
4.1 Homogeneous Spaces, Quasi-Invariant, and Invariant Measures 63
Now, let H1 be the subgroup of Gut , which consists of all elements of the type:
x0
h= . (4.12)
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
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
x = (x)x0 . (4.20)
Alternatively,
implying that
dg (x) f (g1 x)
(g, x) = = .
d (x) f (x)
f (x0 )
( (x), x) = ,
f (x)
so that
dg (x)
d =
( (x), x) d(x),
( (x), x) = .
d(x)
and, hence,
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 ) ,
d (g1 x)
= (g, x).
d (x)
Then,
But,
implying that
g = (b, a) G+ , (4.25)
g1 g2 = (b1 + a1 b2 , a1 a2 ). (4.26)
68 4 Some Group Theory
H := {g G+ : g = (0, a), a R+
}. (4.27)
gx = ax + b, g = (b, a) G+ . (4.30)
dg (x) ( ax ba )
(g, x) = = . (4.33)
d (x) a (x)
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 = .
(U(g) I) 0 as g 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
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).
It is straightforward to verify that the operators U(g), g G, defined by
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
writing V U(g) = WU(g)W 1 , a straightforward computation, using (4.37), (4.39),
P(
U(g) = P(g
)U(g) ), (4.47)
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
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
= |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.
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
[U(h), F] = 0, h H. (4.52)
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:
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
= a(g ).
Writing
is a unitary embedding of H onto a subspace H = W H of H.
by (4.54) and the unitarity of V (h), h H. Defining U(g) by
on all of H
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.
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)
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
z +
z z = gz = . (4.67)
z+
1 |z|2
1 |z |2 = , (4.70)
| z + |2
so that
2
1 1 |z |2
= . (4.71)
| z + |4 1 |z|2
1 2 (z, z) 1 dz dz
= dz dz = , (4.75)
i z z i (1 |z|2 )2
H j = L2 (D, (2 j 1) d j ), (4.77)
(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
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
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
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
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
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
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 ),
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
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
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
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)
Proof. Consider the adjoint map T : H2 H1 , and take the adjoint of the relation
(4.113), getting
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
U1 (g) 2T = U1 (g) 2H1 + TU1 (g) 2H2 = 2H1 + U2 (g)T 2H2
T 2H2 = 2H1 , 0.
1
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
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.
( f g) h = f (g h) , f g = g f.
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
: , x = 0 for all x }.
= { G (4.121)
(
: G/ and
: G/
G
f (x) d (x) =
G/
f (x + ) dG/ (x), f L1 (G). (4.122)
G
f ( ) dG ( ) =
G/
f ( + ) dG/ ( ),
f L1 (G). (4.123)
,t = ei t .
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,
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)
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)
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].
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
Let g be a complex Lie algebra, that is, a complex vector space with an
antisymmetric bracket [., .] that satisfies the Jacobi identity
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
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
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).
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),
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
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,
Since
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)
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
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
g = z+ hz = b+ z = z+ b , z Z , h H c, b B . (4.148)
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
and
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
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)
g g = ( , g)( , g ) = ( + + (g, g ), gg ),
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
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)
[ s , s ]2 s , [ v c , v c ]2 = 0 , [ s , v c ]2 v c . (4.157)
X = Xs + Xc , Xs s, Xc vc ,
(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
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]:
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.
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
N
ci fi (x) = 0, for all x X, ci = 0, for all i. (5.2)
i=0
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
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
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
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 ,
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)
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,
There follows the very useful property of the reproducing kernel Hilbert space, that
this map is continuous. Indeed, from (5.9) and (5.11),
Hence
The adjoint EK (x) : K HK of the evaluation map is easily computed. Since, for
all v K and HK ,
the last equality following from the reproducing property (5.11), we get
EK (x)v = xv , v K. (5.19)
which is often useful. Similarly, from (5.14), we get for the POV function
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
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
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).
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
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.
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
Zn i
fni (Z) = n (Z) i = & , i = 1, 2, . . . , N, n = 0, 1, 2, . . . , , (5.35)
ck k
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
N
|Zi Zi | d (Z) = I .
i=1 X
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.
In order to motivate the discussion, let us go back to the canonical coherent states,
s (q,p) = ei(pQqP) s ,
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)
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
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.
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.
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).
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).
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 ,
N
vi | K(xi , x j )v j 0. (5.48)
i, j=1
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,
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
is total in HK .
2. There exists a -selection : HK M such that the linear map
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
N M
[ N ] | [ M ]K = vi | K(xi , y j )w j xi , (5.54)
i=1 j=1
or, equivalently,
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
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
Then, by (5.56),
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
Hence,
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):
In view of Theorem 5.4.3, from now on, we shall write for elements [ ] HK , it
being understood that
We claim that
and, in particular,
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
The projection operator PK (x) is the support of FK (x), which implies that
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:
whence,
and
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
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
is unitary and maps FK (x) onto FK (x), for each x X; i.e.,
Furthermore, the respective evaluation maps and reproducing kernels are related by
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,
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
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
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
If
n
F(x) = | xi xi | , (5.94)
i=1
5.4 Reproducing Kernel Hilbert Spaces: General Construction 127
SK = {xi = WK xi : i = 1, 2, . . . , n, x X} (5.95)
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},
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)
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 ).
i (x j ) = xi j | , i = 1, 2, . . . , n, j = 1, 2, . . . , N , (5.103)
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 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 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
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
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
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 ,
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
min (x)
vx = ,
max
dim H
K(x, y) = | ui (x)ui (y)| . (5.115)
i=1
dim H
ui (x)x < . (5.116)
i=1
K(x, y) = FK (x)1/2 FK (y)1/2 [FK (x) FK (y)]1/2 , x, y X, (5.118)
[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
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
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
[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
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
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
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
[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
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)
dz dz 1
d (z, z) = = dy dx, z = x + iy, (6.18)
2 i
6.2 Holomorphic Kernels 137
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
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
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
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
2 r(K)| f (z)|2 ,
kernel KD : D D C,
such that
KD (z, w)KD (w, z ) d (w, w) = KD (z, z ). (6.24)
D
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
k
1
d (z, z) =
(2 i)k i=1
dzi dzi . (6.25)
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
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 ),
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.
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]).
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=
These states are the unnormalized coherent states for the circle in the Fourier
representation, as shown in the next section.
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
= 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
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.
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
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
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
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
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
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.
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
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
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
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,
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
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
Indeed, expanding the determinant with respect to the last row, starting at the lower
right corner, we easily get
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 !
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
O f = PK fmult PK (6.86)
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
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)
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,
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
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)
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, :
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:
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.
So far the only physical quantity we have at our disposal is the Hamiltonian.
Suppose there exists an explicit functional form
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
we obtain
n
xn 1
= 0, 0 = 0, = nm , 0 = 1. (6.110)
n m=0 m
For instance,
x1 x1 x2 x1 x2 x3
1 = , 2 = , 3 =
1 x1 2 + 12 x1 2 + (x1 + x2 )1 2 + 13
n = xn ! . (6.112)
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 .
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
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)
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:
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
xn+1 /xn
< 1 Q < 0. (6.124)
(n + 1)/n
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,
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)
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
1 J n+1
|cn |2 = . (6.135)
n!(n + 2)! I2 (2 J)
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.
We now turn to the continuum case. Suppose the quantum system has a Hamiltonian
with positive non degenerate purely continuous spectrum,
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
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
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)
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)
where
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.
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)
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)
|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.
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
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
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):
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
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)
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.,
Letting X = G/H and taking any Borel section : X G, we may define another
family of covariant CS
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
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
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
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
The mapping,
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),
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
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),
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
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
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:
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
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,
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
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
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.
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
Choosing the lowest weight vector | j, j, with isotropy subalgebra b = {Jo , J },
one gets for the corresponding CS:
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.
H = H0 + HI , (7.38)
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
j = j i j ,
x y
(7.41)
01 0 i 1 0
xj = , yj = , zj = .
10 j
i 0 j
0 1 j
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,
where
= (4 )1 Tr
Z(J) F
e H(J, ) d ,
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.
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].
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
(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.
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
| K := |WK A1 1
WK Cn L2 (X, d ) , (7.51)
(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
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)
HK HK HK . (7.58)
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
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
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
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 :
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 :
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
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
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
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
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
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:
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
(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=
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=
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)
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 .
<
5
s2 + m2 v2 + w2 = l 2 ; sm = v w v w .
<
T 1 | T |2
z = 2il , s=l , m = 2il , (7.102)
( , ) ( , ) ( , )
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
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
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
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)
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) .
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
! " n
2 2 +1
(n)
Vol G = n .
2
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
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)
2 3 l+n/21
l (g, ) = jg g1 n
.
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)
;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.
(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)
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.
( = 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
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
P = W W , P L2 (G, d ) = H , (8.11)
as its integral kernel. Furthermore, W intertwines U and the left regular represen-
tation U ,
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,
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
W 2L2 (G, d ) I( )
= = = 1, (8.18)
2 c( ) 2
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
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 .
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)
= 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
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
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].
W W = ( , )I, ( , ) C (8.37)
( , ) = 1, A. (8.38)
8.2 Orthogonality Relations 213
Set
for all , A , and , H. Inserting (8.37) and (8.39) into this, we get
g | g | d (g) = q( , ) | H . (8.40)
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:
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
q( , ) = C |C H . (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( , ).
1
CU(g) |CU(g) H = C |C H . (8.47)
(g)
8.2 Orthogonality Relations 215
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
(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,
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
1
K (g, g ) = g |g , (8.54)
C |C
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:
H (M)
G (a, M) = , (a, M) Rn H. (8.61)
| det M|
218 8 Coherent States from Square Integrable Representations
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
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
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
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).
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
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
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.
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
d d|k| 1
d (k) = d = d = 2 dk.
|k| |k|
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
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)
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
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
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.
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,
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
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
(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
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)]
where,
I ( )
c ( ) = and (x) = U( (x)) , (8.96)
2
1
K (x, y) = | . (8.98)
c ( ) (x) (y)
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
I( )
cV ( ) = , (8.103)
2
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
P = W W , P LK
2
(X, d ) = H , (8.106)
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
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 (x,t) = (x ,t )
x = Rx + vt + a, t = t + b. (8.112)
232 8 Coherent States from Square Integrable Representations
R( )i j = Tr[ j 1 i ], i, j = 1, 2, 3, (8.114)
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)
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 thus,
= ,
m v2
E = E + + p v, (8.122)
2
p = R( 1 )(m v + p).
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
3 G:
Define the section : L/SU(2) R
! k "
(k) = 0, 0, 0, , I2 . (8.126)
m
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
( , a, v, ) G . (8.132)
(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
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)
( , ) = ei . (8.138)
( , a, v, ) = (0, a, v, I2 )( , 0, 0, ). (8.140)
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
which can be obtained using (8.136). The coset decomposition (8.140) implies the
following action of g = ( , a, v, ) G on :
j (g)W1 ,
U (g) = W U g G ,
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)
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|
j (0, 0, 0, )
U nJM = D J ( )
nJM , M = J, J + 1, . . . , J 1, J. (8.159)
V J ( , ) = ei D J ( ). (8.160)
KnJ
Then (8.159) implies that for any vector
j ( , 0, 0, )
U = V J ( , )
. (8.161)
J
PnJ = nJM
| nJM |, (8.162)
M=J
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
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
M = J, J + 1, . . . , J 1, J, (8.166)
J
1
(2 )3 (2J + 1) M=J
nJM
| nJM
(q,p) (q,p) | dq dp = I, (8.167)
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)
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
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)
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,
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
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.
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)
(1 , X 1 )(2 , X 2 ) = (1 + 2 + (X 1 , X 2 ), X 1 + X 2 ), (9.5)
(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 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
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
X = (Q1 , Q2 , . . . Qn , P1 , P2 , . . . , Pn ),
(Qi )(x) = xi (x), (Pi )(x) = i (x). (9.19)
xi
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
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
= 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
(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
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.
Then,
Nk Tk O = V, (9.32)
and N = kO Nk is a manifold, referred to as the normal bundle. Analogously we
define
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
= G/H0 , H0 = N0 S0 . (9.35)
Also, set
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
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),
while
1
s (J ) ; J >s ,s = < J ; s Js >s ,s ,
< Ad#S J s, (9.43)
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.
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)
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,
Next, writing
(Note that Pvk svk = svk .) Thus, under the action of g = (x, s) G, an element (k, vk )
T O transforms to
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
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,
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:
9.2.3 Measures on
m
v p = vip dpi , vip R.
i=1
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
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
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
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
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:
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,
(k)i := ki = k k0 ; ei , i = 1, 2, . . . , m, (9.70)
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
G/(V S0 ) O , (9.75)
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):
( L)(h(x, s)1 , p)) = exp[i < k ; x >] L(h0 (s1 , p)). (9.79)
( LU(x, s) )(k) = exp[i < k ; x >] L(h0 (s1 , k))1 (s1 k). (9.80)
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 ,
(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
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
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
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
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
i = 1, 2, . . . , N, (q, p) V0 O } (9.95)
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)),
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
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)
N
c() = (2 )m ( LU(0, (p)) i )(k)2 m( (p)1 k) (p) d (p),
K
i=1 O
(9.99)
(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).
Writing x = (x1 , x2 ) = (cos , sin ) for points on the unit circle, the cotangent
bundle is
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
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].
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:
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,
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
(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,
where the constant term i (p) (for fixed p), which could be zero, is given by
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
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
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.
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.
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
1
= Tr[A A ], , = 0, 1, 2, 3, (10.1)
2
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
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
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
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
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
a0 k
q = a , p = k. (10.13)
k0
274 10 CS of the Relativity Groups
1 ! "
q = p [a + (0, q)] p
[a0 + { (0, q)}0 ]
p0 0
(10.14)
&
p = p, p = ( m2 + p2 , p),
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) = (
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
p0 (p) m(p)
(p) = , so that #(p) = . (10.20)
m + p (p) p0 p (p)
p mVp (p)
(p) = , (10.21)
p0 p (p)
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
2s+1
A = R6
|i (q,p) i (q,p) | dq dp, i (q,p) = UWs ( (q, p)) i . (10.23)
i=1
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)
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
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
i = ei , i = 1, 2, . . . , 2s + 1, (10.34)
(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
On HW
s define the operators (P , P),
0
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
where denotes the L2 (Vm+ , dk/k0 ) expectation value with respect to the vector
. Hence for the operator A [see (10.39)]
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
(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
(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)
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:
(2 )3 k0 p0 k p
a (k, p) = a0 (k, p) = , (10.55)
m k0
(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.
This is the principal section introduced in the previous chapter. This time
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.
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.
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:
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:
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)
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
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)
2 2
P0 |P| A P0 + |P| , (10.75)
m m
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
I1 2 (2 , 1 ) = (1 , 2 ) 2 |1 , (10.80)
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
(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.
f , f = f1 2 f2 2 , f1 2 = f1 , f1 0, f2 2 = f2 , f2 0.
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)
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
and associating to the sesquilinear forms so defined the operators Ul(r) (a, ). The
final result is
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
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
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):
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.
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
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
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
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
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/ :
$ %
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)
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
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:
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
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 :
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
K1 K2 K0 , (10.139)
2
10.3 The (1+1)-Dimensional Anti-de Sitter Group SOo (1, 2) and Its Contraction(s) 299
1 z
(z) = c e . (10.141)
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
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
where
1
d (g) = (1 |z|2 )2 d2 z d
2 2
lim I U ( (a, ))b U(a, )bH = 0, for g = (a, ) P+ (1, 1). (10.158)
0
lim f f H = 0. (10.159)
0
302 10 CS of the Relativity Groups
contracts in our sense to Poincar coherent states of the form (10.69), defined
precisely with the symmetric section (see Sect. 10.1.2)
where
qp
q0 = p0 ,
q = p0 q,
p = p. (10.162)
p0 + 1
pq
q0 = , (10.164)
p0 + 1
(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
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
(z, z, ) := ( , ) = 0 (
q, p p) +
q 1 (k) + o( ), (10.169)
B = I B I1 . (10.173)
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
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
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].
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]).
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
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
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
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)
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)
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
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!
/ = 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
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.
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
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)
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 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
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
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:
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)
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
Formally, and in accordance with (11.31) and (11.33), the operator M is recovered
through the quantization of the Dirac distribution on C:
We will give this equality its full mathematical sense in Sect. 11.5.
The covariance property of the above quantization reads as
where P =
n=0 (1) |en en | is the parity operator defined in (2.85),
n
and
d2 z
tr(Af A f ) = | f (z)|2 , f | (z)| = 1, z . (11.39)
C
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
a + a a a
Aq = := Q , Ap = := P . (11.41)
2 i 2
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
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
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 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)
s+1 h h
ln = s = coth (11.50)
s1 kB T 2kB T
An elliptic regular quantization is isometric iff w(u) {1, +1}. A simple example
is given by the family :
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
Hence, one notices that the Wigner-Weyl (i.e. canonical) quantization (m(u) = 1) is
not the only one providing these relations.
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.
(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
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.
Adapting this definition to the present context, and given an operator A, the
corresponding Wigner function is defined as
WA f = f (11.63)
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).
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
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)
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
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
( 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
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.
:= Af nn
|en en | ,
n,n =0
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
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
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
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)!
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.
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
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
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 .
S = M IH .
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
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
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,
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
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
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 ) .
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.
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
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)
M
A f : H H, A f = i f (ai ) |an an | .
i=1
M
em |A f |en = f (ai ) m (ai ) n (ai ) .
i=1
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
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:
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:
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
d d d
M |ai |a j |2 = ||ai ||2 M |ai |ai |2 = 1 M ,
j=i
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
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)
1 ! "1/4
e 2 (Jn) ein |en .
1 2
|x |J, = & (11.118)
N (J) nZ
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 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)
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
Let us now examine in which way the action-angle material presented above can be
extended to more generic mechanical systems.
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 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
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)
(iv) Since it is trivially verified that in both cases the quantum Hamiltonian is
exactly what we expect:
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
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.
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
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
1 !t b"
ba (t) = . (12.3)
a a
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):
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)
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
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
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)
+
d
c := 2 |
( )|2 < . (12.12)
| |
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
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)
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)
?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
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):
and
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
da db
c1
|ba ba | = I. (12.26)
G+ a2
K(b , a ; b, a) = c1
b a |ba . (12.27)
(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
da db
s(x) = c1
ba (x) (T s)(b, a) , (12.31)
G+ a2
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
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
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 :
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
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
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:
(+)
| < 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),
and the duality form < , > may be seen as an extension/restriction of the inner
product of L2 (R+ ):
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
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
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
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.
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
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)
|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).
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}.
(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
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 :
where the amplitude Al (x) varies slowly (in a logarithmic sense) with respect to the
phase l (x), namely,
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
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].
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
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])
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):
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 )
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
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 |
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.
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
. . . 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
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
1
(2 ) = m0 ( ) ( ), with m0 ( ) = hn ein . (13.5)
2 nZ
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)
m1 ( ) m0 ( ) + m1 ( + ) m0 ( + ) = 0, a.e.. (13.9)
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).
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
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:
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
13.1.4 Applications
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
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].
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
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)
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)
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.
and that
1
1 1
h (x) dx = 1/2 1/2 2 = 0. (13.34)
0
!nb"
(U(b, a) f )(n) = fba (n) := f . (13.37)
a
p1
E = { f 2 (Z p ) : f (n) = 0}, (13.38)
n=0
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)
1
f (n) =
c (T f )(b, a)ba (n), f E. (13.41)
(b,a)G p
Da = Ka a , (13.42)
p1
f(k) = e2 ink/p f (n).
n=0
Da = Fa a , (13.43)
Faa = Fa a Fa (13.44)
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):
(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
1 1, 3 3
2 , 4 2
( ( )
r1
Fa (k) = ( j v)(k), k Z p ,
j=0
( )
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.
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].
a( ) = (a )(a ), a A , , . (13.49)
396 13 Discrete Wavelet Transforms
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.
(
a u( ) = u
(a ), . (13.51)
a a = aa , a, a A , (13.52)
a T = Ta a , where T is the translation by . (13.53)
(T a )(T a ) = T +a aa , (13.54)
( , a)( , a ) = ( + a , aa ),
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)
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)
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)
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 ; 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
D 1,a; ,
a; := D a A , a = 1, = 1, . . . d(a) 1
(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
(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 ;
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.
(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.
1 1 ein
Hn ( ) = . (13.75)
n 1 ei
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.
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
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
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 ),
Im (r r1 ) = {1}.
We define
Z r (A , M ) = Ker r , B r (A , M ) = Im r1 .
H r (A , M ) = Z r (A , M )/B r (A , M )
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)
( )
Haa ( )
[1 H] (a, a )( ) = , . (13.83)
Ha ( )Ha (a )
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,
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
(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
(
U
a f (a ) = Ha ( ) f ( ), V
a; f (a ) = Ga; ( ) f ( ),
.
G (13.88)
d(a)1
Ua Ua + Va; Va; = 1.
=1
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.
Then it follows from (13.90) that for all [compare with (13.8)],
(n; )
n; (t) = (Vn; ) (t) = n gl (at l) . (13.92)
l
! " ! "
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
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)
(
s( ) = f ( + ) = ( + 2 l) f( + 2 l) . (13.98)
Z lZ
408 13 Discrete Wavelet Transforms
1 n1
(
D s( ) =
n Hn (n1 ( + 2 l)) s(n1 ( + 2 l)) .
n l=0
(13.101)
In other words
1 1
snk = f (t) (n1t k) dt = (S f )(nk, n) . (13.102)
n R n
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
n; (n ) = Gn; ( )( ) ,
R. (13.106)
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):
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
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
for k, x Rn and k x = k1 x1 + k2 x2 + + kn xn .
14.2 Mathematical Analysis 413
n1
2 k/2
where |k|2 = k k, and An1 = is the volume of SO(n 1).
k=2 (k/2)
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
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
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)
da
d (x) = db d ( ), (14.15)
an+1
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
K(b , a , ; b, a, ) = c1
b ,a , | b,a, , (14.21)
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
(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:
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)
, ), 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)
(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:
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
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
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
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
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
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:
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.
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
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
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:
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
(D + i1 P1 )
(k) = (D + i1 P1 )
(k)
(1 , 2 > 0). (14.45)
(J + i2 P2 )
(k) = (J + i2 P1 )
(k)
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
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.
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
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:
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]).
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)
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
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.
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 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
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.
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.
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
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.
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.
iso (k) l ( ),
i,l (k, ) = (14.59)
i
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
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)
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:
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
>
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
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
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)
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
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
selectivity 2 . In particular, for each level l [0, L], Proposition 14.5.1 shows that
l
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,
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
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
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):
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])
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
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
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)
da
f (x) = ( f )(a,b, ) a,b, (x) 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
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
2
| f = (2 )2 f 2L2 . (14.88)
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
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
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|
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
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
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
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.,
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,
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 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
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
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
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.
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
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:
Notice that this representation is infinite dimensional. Its restriction to SO(3) is the
quasi-regular representation
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):
F ( , a) = U(I ( , a)) | f
= [US ( , a) ]( ) f ( ) d ( ) (15.8)
S2
= a ( 1 ) f ( ) d ( ).
S2
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)
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)
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
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
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
where
4
g (l) =
2l + 1 j |a j (l, 0)|2 ,
jZ
(
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 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:
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)
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].
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].
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.
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.
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
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)
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 ).
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.
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
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
on S .
2
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.
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].
(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.
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
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.
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):
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
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].
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)
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].
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.
1 = cos ,
2 = sin , [0, 2 ), I = [0, b) or [0, ),
3 = ( ),
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
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
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,
|
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
= U(b, a, )
p1 | s p1 L2 (R2 )
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
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 .
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
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 ( )
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.
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
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
d1
(tL)i ( j) = (t k )k (i)k ( j).
k=0
15.3 Wavelets on Graphs 493
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
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
(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
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
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:
( ( ) = ( )1/2 ( ).
(16.7)
b (b, )
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
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.
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.
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:
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)
(b, ; a, c; R, )(b , ; a , c ; R , ) =
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)
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
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).
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
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
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
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
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 = ( , b, a, v, R, , ), (16.21)
x e Rx + e vt + a,
(16.22)
t e t + b.
!
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)
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
dg = e3 8 d db da dv dR d d , g = ( , b, a, v, R, , ),
subgroup D0 R+
of time dilations e . The corresponding coset space X = Gaff /D0
M
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 .
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)
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
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
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
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:
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
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
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.
Hn = Sp(2n, R)/SU(n)
{Z Cnn : Z T = Z, Im Z Pn }
Dn = {W Cnn : W T = W, I WW Pn },
where A > 0 means that A is a positive definite Hermitian matrix. Then a Cayley
transform maps DI onto the domain
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
|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.
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
(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
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:
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
where, as above,
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
( )| 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].
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
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.
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
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
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:
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,
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 )
|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 ).
w( )
v2 ( ) =
w( 1 )
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
(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
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)]:
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
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.
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).
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:
where
where qn > 0 is to be fixed later, Xn (k) and n are respectively the discretized
forms of Xp (k) and p . Let
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
(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
dk
(Xn (k )) (k) (k )
(Xn (k)) . (17.34)
k0
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:
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
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=
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 )
sinh a t n0 sinh b ,
17.4 Groups Without Dilations: The Poincar Groups 531
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,
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 + )
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
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)
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
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
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}
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
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
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,> ) ),
{ n,l
j
:= UWs ( (qn,l
j
, pn )) j : l Z3 , n Z, j = 1, 2, . . . , 2s + 1}
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.
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
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
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
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
[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
[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
[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
[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
[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
[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
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
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