90NH

Quantum Physics, Relativity, and Complex Spacetime:
Towards a New Synthesis

Gerald Kaiser
Department of Mathematics
University of Massachusetts at Lowell
Originally published in 1990 by North-Holland, Amsterdam

c Gerald Kaiser 2003. All rights reserved.

Email: [email protected]
Web: www.wavelets.com
To my parents, Bernard and Cesia

and to Janusz, Krystyn, Mirek and Renia
Unified field Theory

In the beginning there was Aristotle
And objects at rest tended to remain at rest
And objects in motion tended to come to rest
And soon everything was at rest
And God saw that it was boring.
Then God created Newton
And objects at rest tended to remain at rest
But objects in motion tended to remain in motion
And energy was conserved and momentum was conserved
and matter was conserved
And God saw that it was conservative.
Then God created Einstein
And everything was relative
And fast things became short
And straight things became curved
And the universe was filled with inertial frames
And God saw that it was relatively general
but some of it was especially relative.
Then God created Bohr
And there was the Principle
And the principle was Quantum
And all things were quantized
But some things were still Relative
And God saw that it was confusing.
Then God was going to create M
And M would have unified
And M would have fielded a theory
And all would have been one
But it was the seventh day
And God rested
And objects at rest tend to remain at rest.
Adapted from a poem by Tim Joseph
ii
CONTENTS
Preface
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Suggestions to the Reader
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Chapter 1. CoherentState Representations

1.1. Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2. Canonical coherent states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3. Generalized frames and resolutions of unity . . . . . . . . . . . . . . . . 14
1.4. Reproducingkernel Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5. Windowed Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6. Wavelet transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Chapter 2. Wavelet Algebras and Complex Structures
This chapter is not essential, hence omitted to save space; the same
material is available at http://www.wavelets.com/92Siam.pdf
Chapter 3. Frames and Lie Groups
3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2. Klauders groupframes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3. Perelomovs homogeneous Gframes . . . . . . . . . . . . . . . . . . . . . . . 50
3.4. Onofriss holomorphic Gframes . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5. The rotation group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6. The harmonic oscillator as a contraction limit . . . . . . . . . . . . . .82
Chapter 4. Complex Spacetime
4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2. Relativity, phase space and quantization . . . . . . . . . . . . . . . . . . . 90
4.3. Galilean frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4. Relativistic frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.5. Geometry and Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.6. The nonrelativistic limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
iii
Chapter 5. Quantized Fields
5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.2. The multivariate AnalyticSignal transform . . . . . . . . . . . . . . .154
5.3. Axiomatic field theory and particle phase spaces . . . . . . . . . . 161
5.4. Free KleinGordon fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.5. Free Dirac fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.6. Interpolating particle coherent states . . . . . . . . . . . . . . . . . . . . . 203
5.7. Field coherent states and functional integrals . . . . . . . . . . . . . 208
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Chapter 6. Further Developments
6.1. Holomorphic gauge theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
6.2. Windowed XRay transforms: Wavelets revisited . . . . . . . . . 235
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
iv
PREFACE
The idea of complex spacetime as a unification of spacetime and classical phase space, suitable as a possible geometric basis for the synthesis
of Relativity and quantum theory, first occured to me in 1966 while
I was a physics graduate student at the University of Wisconsin. In
1971, during a seminar I gave at Carleton University in Canada, it
was pointed out to me that the formalism I was developing was related to the coherentstate representation, which was then unknown
to me. This turned out to be a fortunate circumstance, since many
of the subsequent developments have been inspired by ideas related
to coherent states. My main interest at that time was to formulate
relativistic coherent states.
In 1974, I was struck by the appearance of tube domains in axiomatic quantum field theory. These domains result from the analytic
continuation of certain functions (vacuum expectaion values) associated with the theory to complex spacetime, and powerful methods
from the theory of several complex variables are then used to prove
important properties of these functions in real spacetime. However,
the complexified spacetime itself is usually not regarded as having
any physical significance. What intrigued me was the possibility that
these tube domains may, in fact, have a direct physical interpretation
as (extended) classical phase spaces. If so, this would give the idea of
complex spacetime a firm physical foundation, since in quantum field
theory the complexification is based on solid physical principles. It
could also show the way to the construction of relativistic coherent
states. These ideas were successfully worked out in 1975-76, culminating in a mathematics thesis in 1977 at the University of Toronto
entitled PhaseSpace Approach to Relativistic Quantum Mechanics.
Up to that point, the theory could only describe free particles.
The next goal was to see how interactions could be added. Some
progress in this direction was made in 1979-80, when a natural way
was found to extend gauge theory to complex spacetime. Further
progress came during my sabbatical in 1985-86, when a method was
developed for extending quantized fields themselves (rather than their
vacuum expectation values) to complex spacetime. These ideas have
so far produced no hard results, but I believe that they are on the
right path.
Although much work remains to be done, it seems to me that
enough structure is now in place to justify writing a book. I hope
that this volume will be of interest to researchers in theoretical and
v
mathematical physics, mathematicians interested in the structure of
fundamental physical theories and assorted graduate students searching for new directions. Although the topics are fairly advanced, much
effort has gone into making the book selfcontained and the subject
matter accessible to someone with an understanding of the rudiments
of quantum mechanics and functional analysis.
A novel feature of this book, from the point of view of mathematical physics, is the special attention given to signal analysis
concepts, especially timefrequency localization and the new idea of
wavelets. It turns out that relativistic coherent states are similar to
wavelets, since they undergo a Lorentz contraction in the direction
of motion. I have learned that engineers struggle with many of the
same problems as physicists, and that the interplay between ideas
from quantum mechanics and signal analysis can be very helpful to
both camps. For that reason, this book may also be of interest to
engineers and engineering students.
The contents of the book are as follows. In chapter 1 the simplest
examples of coherent states and timefrequency localization are introduced, including the original canonical coherent states, windowed
Fourier transforms and wavelet transforms. A generalized notion of
frames is defined which includes the usual (discrete) one as well as
continuous resolutions of unity, and the related concept of reproducing kernels is discussed.
In chapter 2 a new, algebraic approach to orthonormal bases of
wavelets is formulated. An operational calculus is developed which
simplifies the formalism considerably and provides insights into its
symmetries. This is used to find a complex structure which explains
the symmetry between the low and the highfrequency filters in
wavelet theory. In the usual formulation, this symmetry is clearly
evident but appears to be accidental. Using this structure, complex
wavelet decompositions are considered which are analogous to analytic coherentstate representations.
In chapter 3 the concept of generalized coherent states based on
Lie groups and their homogeneous spaces is reviewed. Considerable
attention is given to holomorphic (analytic) coherentstate representations, which result from the possibility of Lie group complexification. The rotation group provides a simple yet nontrivial proving
ground for these ideas, and the resulting construction is known as the
spin coherent states. It is then shown that the group associated
with the Harmonic oscillator is a weak contraction limit (as the spin
s ) of the rotation group and, correspondingly, the canonical co-
vi
herent states are limits of the spin coherent states. This explains why
the canonical coherent states transform naturally under the dynamics
generated by the harmonic oscillator.
In chapter 4, the interactions between phase space, quantum mechanics and Relativity are studied. The main ideas of the phase
space approach to relativistic quantum mechanics are developed for
free particles, based on the relativistic coherentstate representations
developed in my thesis. It is shown that such representations admit
a covariant probabilistic interpretation, a feature absent in the usual
spacetime theories. In the nonrelativistic limit, the representations
are seen to contract smoothly to representations of the Galilean
group which are closely related to the canonical coherentstate representation. The Gaussian weight functions in the latter are seen to
emerge from the geometry of the mass hyperboloid.
In chapter 5, the formalism is extended to quantized fields. The
basic tool for this is the AnalyticSignal transform, which can be applied to an arbitrary function on IRn to give a function on Cln which,
although not in general analytic, is analyticityfriendly in a certain sense. It is shown that even the most general fields satisfying
the Wightman axioms generate a complexification of spacetime which
may be interpreted as an extended classical phase space for certain
special states associated with the theory. Coherentstate representations are developed for free KleinGordon and Dirac fields, extending
the results of chapter 4. The analytic Wightman twopoint functions
play the role of reproducing kernels. Complexspacetime densities of
observables such as the energy, momentum, angular momentum and
charge current are seen to be regularizations of their counterparts
in real spacetime. In particular, Dirac particles do not undergo their
usual Zitterbewegung. The extension to complex spacetime separates,
or polarizes, the positive and negativefrequency parts of free fields,
so that Wick ordering becomes unnecessary. A functionalintegral
representation is developed for quantized fields which combines the
coherentstate representations for particles (based on a finite number
of degrees of freedom) with that for fields (based on an infinite number
of degrees of freedom).
In chapter 6 we give a brief account of some ongoing work, beginning with a review of the idea of holomorphic gauge theory. Whereas
in real spacetime it is not possible to derive gauge potentials and
gauge fields from a (fiber) metric, we show how this can be done in
complex spacetime. Consequently, the analogy between General Relativity and gauge theory becomes much closer in complex spacetime
than it is in real spacetime. In the holomorphic gauge class, the
vii
relation between the (nonabelian) YangMills field and its potential
becomes linear due to the cancellation of the nonlinear part which
follows from an integrability condition. Finally, we come full circle by
generalizing the AnalyticSignal transform and pointing out that this
generalization is a higherdimensional version of the wavelet transform which is, moreover, closely related to various classical transforms
such as the Hilbert, FourierLaplace and Radon transforms.
I am deeply grateful to G. Emch for his continued help and encouragement over the past ten years, and to John Klauder and Ray
Streater for having read the manuscript carefully and made many
invaluable comments, suggestions and corrections. (Any remaining
errors are, of course, entirely my responsibility.) I also thank D. Buchholtz, F. Doria, D. Finch, S. Helgason, I. Kupka, Y. Makovoz, J. E.
Marsden, M. OCarroll, L. Rosen, M. B. Ruskai and R. Schor for miscellaneous important assistance and moral support at various times.
Finally, I am indebted to L. Nachbin, who first invited me to write
this volume in 1981 (when I was not prepared to do so) and again in
1985 (when I was), and who arranged for a tremendously interesting
visit to Brazil in 1982. Quero tambem agradecer a todos os meus
colegas Brasileiros!
Suggestions to the Reader

The reader primarily interested in the phasespace approach to
relativistic quantum theory may on first reading skip chapters 13
and read only chapters 46, or even just chapter 4 and either chapters 5 or 6, depending on interest. These chapters form a reasonably
selfcontained part of the book. Terms defined in the previous chapters, such as frame, can be either ignored or looked up using the
extensive index. The index also serves partially as a glossary of frequently used symbols. The reader primarily interested in signal analysis, timefrequency localization and wavelets, on the other hand, may
read chapters 1 and 2 and skip directly to sections 5.2 and 6.2. The
mathematical reader unfamiliar with the ideas of quantum mechanics
is urged to begin by reading section 1.1, where some basic notions are
developed, including the Dirac notation used throughout the book.
1
Chapter 1
COHERENTSTATE REPRESENTATIONS
1.1. Preliminaries
In this section we establish some notation and conventions which will
be followed in the rest of the book. We also give a little background on
the main concepts and formalism of nonrelativistic and relativistic
quantum mechanics, which should make this book accessible to non
specialists.
1. Spacetime and its Dual
In this book we deal almost exclusively with flat spacetime, though
we usually let space be IRs instead of IR3 , so that spacetime becomes
X = IRs+1 . The reason for this extension is, first of all, that it involves
little cost since most of the ideas to be explored here readily generalize
to IRs+1 , and furthermore, that it may be useful later. Many models
in constructive quantum field theory are based on two or three
dimensional spacetime, and many currently popular attempts to unify
physics, such as string theories and KaluzaKlein theories, involve
spacetimes of higher dimensionality than four or (on the string world
sheet) twodimensional spacetimes. An event x X has coordinates
x = (x ) = (x0 , xj ),
(1)
where x0 t is the time coordinate and xj are the space coordinates.

Greek indices run from 0 to s, while latin indices run from 1 to s. If
we think of x as a translation vector, then X is the vector space of
all translations in spacetime. Its dual X is the set of all linear maps
k: X IR. By linearity, the action of k on x (which we denote by kx
instead of k(x)) can be written as
kx =
s
X
k x k x ,
(2)
=0
where we adopt the Einstein summation convention of automatically

summing over repeated indices. Usually there is no relation between
1. CoherentState Representations
x and k other than the pairing (x, k) 7 kx. But suppose we are given
a scalar product on X,
x x0 = g x x0
(3)
where (g ) is a nondegenerate matrix. Then each x in X defines

a linear map x : X IR by x (x0 ) = x x0 , thus giving a map
: X X , with
(x ) x = g x .
(4)
Since g is nondegenerate, it also defines a scalar product on X ,

whose metric tensor is denoted by g . The map x x establishes
an isomorphism between the two spaces, which we use to identify
them. If x denotes a set of inertial coordinates in free spacetime, then
the scalar product is given by
g = diag(c2 , 1, 1, , 1)
where c is the speed of light. X, together with this scalar product, is
called Minkowskian or Lorentzian spacetime.
It is often convenient to work in a single space rather than the
dual pair X and X . Boldface letters will denote the spatial parts of
vectors in X . Thus x = (t, x), k = (k0 , k) and
x x0 = c2 tt0 x x0 and kx = k x = k0 t k x,
(5)
where x x0 and k x denote the usual Euclidean inner products in

IRs .
2. Fourier Transforms
The Fourier transform of a function f : X Cl (which, to avoid analytical subtleties for the present, may be assumed to be a Schwartz
test function; see Yosida [1971]) is a function f: X Cl given by
Z
f (k) =
dx e2ikx f (x)
(6)
X
where dx dt ds x is Lebesgue measure on X. f can be reconstructed

from f by the inverse Fourier transform, denoted by and given by
Z
f (x) =
dk e2ikx f(k) (f)(x),
(7)
X
1.1. Preliminaries
where dk = dk0 ds k denotes Lebesgue measure on X X. Note that

the presence of the 2 factor in the exponent avoids the usual need
for factors of (2)(s+1)/2 or (2)s1 in front of the integrals. Physically, k represents a wave vector: k0 is a frequency in cycles per
unit time, and kj is a wave number in cycles per unit length. Then the
interpretation of the linear map k: X IR is that 2kx is the total
radian phase gained by the plane wave g(x0 ) = exp(2ikx0 ) through
the spacetime translation x, i.e. 2k measures the radian phase
shift. Now in prequantum relativity, it was realized that the energy
E combines with the momentum p to form a vector p (p ) = (E, p)
in X . Perhaps the single most fundamental difference between classical mechanics and quantum mechanics is that in the former, matter
is conceived to be made of dead sets moving in space while in the
latter, its microscopic structure is that of waves descibed by complex
valued wave functions which, roughly speaking, represent its distribution in space in probabilistic terms. One important consequence of
this difference is that while in classical mechanics one is free to specify position and momentum independently, in quantum mechanics a
complete knowledge of the distribution in space, i.e. the wave function, determines the distribution in momentum space via the Fourier
transform. The classical energy is reinterpreted as the frequency of
the associated wave by Plancks Ansatz,
p0 E = 2
h
(8)
where h
is Plancks constant, and the classical momentum is re
interpreted as the wavenumber vector of the associated wave by De
Broglies relation,
p = 2
hk.
(80 )
These two relations are unified in relativistic terms as p = 2

hk .
Since a general wave function is a superposition of plane waves, each
with its own frequency and wave number, the relation of energy and
momentum to the the spacetime structure is very different in quantum
mechanics from what is was in classical mechanics: They become
operators on the space of wave functions:
Z
(P f )(x) =
dk p e2ikx f(k) = i
h f (x),
(9)
x
X
or, in terms of x ,
P0 = i
h
and Pk = i
h
.
t
xk
(90 )
This is, of course, the source of the uncertainty principle. In terms

of energymomentum, we obtain the quantummechanical Fourier
transform and its inverse,
Z
f (p) =
dx eipx/h f (x)
X
Z
(10)
s1
ipx/
h
f (x) = (2
h)
dp e
f (p).
X
If f (x) satisfies a differential equation, such as the Schr

odinger
equation or the KleinGordon equation, then f (p) is supported on

an sdimensional submanifold P of X (a paraboloid or twosheeted
hyperboloid, respectively) which can be parametrized by p IRs . We
will write the solution as
Z
s
f (x) = (2
h)
d(p) eipx/h f(p),
(11)
P
where f(p) is, by a mild abuse of notation, the restriction of f to

P (actually, |f(p)|2 is a density on P ) and d(p) (p) ds p is an
appropriate invariant measure on P . For the Schr
odinger equation
(p) 1, whereas for the KleinGordon equation, (p) = | p0 | 1 .
Setting t = 0 then shows that f(p) is related to the initial wave
function by

f (x, 0) = (p)f (x),
(12)
where now denotes the the sdimensional inverse Fourier transform of the function f on P IRs .
We will usually work with natural units, i.e. physical units
so chosen that h
= c = 1. However, when considering the non
relativistic limit (c ) or the classical limit (h 0), c or h
will be
reinserted into the equations.
3. Hilbert Space
Inner products in Hilbert space will be linear in the second factor and
antilinear in the first factor. Furthermore, we will make some discrete
use of Diracs very elegant and concise braket notation, favored by
physicists and often detested or misunderstood by mathematicians.
1.1. Preliminaries
As this book is aimed at a mixed audience, I will now take a few paragraphs to review this notation and, hopefully, convince mathematicians of its correctness and value. When applied to coherentstate
representations, as opposed to representations in which the position
or momentum operators are diagonal, it is perfectly rigorous. (The
braket notation is problematic when dealing with distributions, such
as the generalized eigenvectors of position or momentum, since it tries
to take the inner products of such distributions.)
Let H be an arbitrary complex Hilbert space with inner product h, i. Each element f H defines a bounded linear functional
f : H Cl by
f (g) = hf, gi.
(13)
The Riesz representation theorem guarantees that the converse is also

true: Each bounded linear functional L : H Cl has the form L = f
for a unique f H. Define the bra hf | corresponding to f by
hf | = f : H Cl.
(14)
Similarly, there is a one-to-one correspondence between vectors g H

and linear maps
|gi: Cl H
(15)
defined by
|gi() = g,
Cl,
(16)
which will be called kets. Thus elements of H will be denoted alternatively by g or by |gi. We may now consider the composite map
braket
hf | gi : Cl Cl,
(17)
given by
hf | gi() = f (g) = f (g)
= hf, gi.
(18)
Therefore the braket map is simply the multiplication by the inner product hf, gi (whence it derives its name). Henceforth we will
identify these two and write hf | gi for both the map and the inner
product. The reverse composition
|gihf | : H H
may be viewed as acting on kets to produce kets:

|gihf | |hi = |gi hf |hi .
(19)
(20)
To illustrate the utility of this notation, as well as some of its

pitfalls, suppose that we have an orthonormal basis {gn } in H. Then
the usual expansion of an arbitrary vector f in H takes the form
X
X
f=
gn h gn | f i
| gn ih gn | f i,
(21)
n
from which we have the resolution of unity

X
| gn ih gn | = I
(22)
where I is the identity on H and the sum converges in the strong

operator topology. If {hn } is a second orthonormal basis, the relation
between the expansion coefficients in the two bases is
h hn | f i = h hn | If i =
h hn | gm ih gm | f i.
(23)
In physics, vectors such as gn are often written as | n i, which

can be a source of great confusion for mathematicians. Furthermore,
functions in L2 (IRs ), say, are often written as f (x) = h x | f i, with
h x | x0 i = (xx0 ), as though the | x is formed an orthonormal basis.
This notation is very tempting; for example, the Fourier transform is
written as a change of basis,
Z
h k | f i = dx h k | x ih x | f i
(24)
with the transformation matrix h k | x i = exp(2ikx). One of the
advantages of this notation is that it permits one to think of the
Hilbert space as abstract, with h gn | f i, h hn | f i, h x | f i and h k | f i
merely different representations (or realizations) of the same vector f . However, even with the help of distribution theory, this use
of Dirac notation is unsound, since it attempts to extend the Riesz
representation theorem to distributions by allowing inner products of
1.2. Canonical Coherent States
them. (The vector h x | is a distribution which evaluates test functions at the point x; as such, | x0 i does not exist within modernday
distribution theory.) We will generally abstain from this use of the
braket notation.
Finally, it should be noted that the term representation is used
in two distinct ways: (a) In the above sense, where abstract Hilbert
space vectors are represented by functions in various function spaces,
and (b) in connection with groups, where the action of a group on a
Hilbert space is represented by operators.
This notation will be especially useful when discussing frames, of
which coherentstate representations are examples.

We begin by recalling the original coherent-state representations
(Bargmann [1961], Klauder [1960, 1963a, b], Segal [1963a]). Consider a spinless non-relativistic particle in IRs (or s/3 such particles
in IR3 ), whose algebra of observables is generated by the position operators Xk and momentum operators Pk , k = 1, 2, . . . s. These satisfy
the canonical commutation relations
[Xk , Xl ] = 0,
[Pk , Pl ] = 0,
[Xk , Pl ] = ikl I,
(1)
where I is the identity operator. The operators iXk , iPk and iI

together form a real Lie algebra known as the Heisenberg algebra,
which is irreducibly represented on L2 (IRs ) by
Xk f (x) = xk f (x),
Pk f (x) = i
f (x),
xk
(2)
the Schr
odinger representation.
As a consequence of the above commutation relations between Xk
and Pk , the position and momentum of the particle obey the Heisenberg uncertainty relations, which can be derived simply as follows.
The expected value, upon measurement, of an observable represented
by an operator F in the state represented by a wave function f (x)
with kf k = 1 (where k k denotes the norm in L2 (IRs )) is given by
h F i = h f | F f i.
(3)
In particular, the expected position and momentum coordinates of

the particle are h Xk i and h Pk i. The uncertainties Xk and Pk in
position and momentum are given by the variances
2Xk = h (Xk h Xk i)2 i = h Xk2 i h Xk i2
2Pk = h (Pk h Pk i)2 i = h Pk2 i h Pk i2 .
(4)
Choose an arbitrary constant b with units of area (square length) and

consider the operators
Ak = Xk + ibPk = xk + b
.
xk
(5)
Notice that although Ak is nonHermitian, it is real in the Schr

odinger representation. Let
zk h f | Ak f i = h Xk i + ibh Pk i,
(6)
where zk denotes the complexconjugate of zk . Then for Ak Ak

zk I we have h Ak i = 0 and
0 kAk f k2 = 2Xk + b2 2Pk b.
(7)
The righthand side is a quadratic in b, hence the inequality for all b

demands that the discriminant be nonpositive, giving the uncertainty
relations
Xk Pk
1
.
2
(8)
Equality is attained if and only if Ak f = 0, which shows that the

only minimumuncertainty states are given by wave functions f (x)
satisfying the eigenvalue equations
Ak f = xk + b

f = zk f
xk
(9)
for some real number b (which may actually depend on k) and some
z Cls . But squareintegrable solutions exist only for b > 0, and then
there is a unique solution (up to normalization) z for each z Cls .
To simplify the notation, we now choose b = 1. Then Ak and Ak
satisfy the commutation relations
[Ak , Al ] = 0,
[Ak , Al ] = 2kl I,
(10)
and z is given by
2
z (x0 ) = N exp[
z 2 /4 + z x0 x0 /2],
(11)
where the normalization constant is chosen as N = s/4 , so that

kz k = 1 for z = 0. Here z2 is the (complex) inner product of z with
itself. Clearly z is in L2 (IRs ), and if z = x ip, then
hXk i = xk and hPk i = pk
(12)
in the state given by z . The vectors z are known as the canonical coherent states. They occur naturally in connection with the
harmonic oscillator problem, whose Hamiltonian can be cast in the
form
H=
1
1
s
P 2 + m2 2 X 2 = m 2 A A +
2m
2
2
(13)
with
Ak = Xk + iPk /m
(14)
(thus b = 1/m). They have the remarkable property that if the initial state is z , then the state at time t is z(t) where z(t) is the orbit
in phase space of the corresponding classical harmonic oscillator with
initial data given by z. These states were discovered by Schr
odinger
himself [1926], at the dawn of modern quantum mechanics. They were
further investigated by Fock [1928] in connection with quantum field
theory and by von Neumann [1931] in connection with the quantum
measurement problem. Although they span the Hilbert space, they
do not form a basis because they possess a high degree of linear dependence, and it is not easy to find complete, linearly independent
subsets. For this reason, perhaps, no one seemed to know quite what
to do with them until the early 1960s, when it was discovered that
what really mattered was not that they form a basis but what we
shall call a generalized frame. This allows them to be used in generating a representation of the Hilbert space by a space of analytic
functions, as explained below. The frame property of the coherent
states (which will be studied and generalized in the following sections
and in chapter 3) was discovered independently at about the same
time by Klauder, Bargmann and Segal. Glauber [1963a,b] used these
vectors with great effectiveness to extend the concept of optical coherence to the domain of quantum electrodynamics, which was made
10
necessary by the discovery of the laser. He dubbed them coherent

states, and the name stuck to the point of being generic. (See also
Klauder and Sudarshan [1968].) Systems of vectors now called coherent may have nothing to do with optical coherence, but there is
at least one unifying characteristic, namely their frame property (next
section).
The coherent-state representation is now defined as follows: Let
F be the space of all functions
Z
f (z) hz | f i = ds x0 z (x0 )f (x0 )

Z
(15)
s 0
2
0
02
0
= N d x exp[z /2 + z x x /2]f (x )
where f runs through L2 (IRs ). Because the exponential decays rapidly in x0 , f is entire in the variable z Cls . Define an inner product
on F by
Z
hf | giF
d(z) f(z)
g (z),
(16)
C
ls
where z x ip and
d(z) = (2)s exp(
z z/2) ds x ds p.
(17)
Then we have the following theorem relating the inner products in

L2 (IRs ) and F.
Theorem 1.1. Let f, g L2 (IRs ) and let f, g be the corresponding
entire functions in F. Then
hf | giF = hf | giL2 .
(18)
Proof. To begin with, assume that f is in the Schwartz space S(IRs )

of rapidly decreasing smooth test functions. For z = x ip, we have
z (x0 ) = N exp[
z 2 /2 + x2 /2 (x0 x)2 /2 + ip x0 ],
hence

f(x ip) = N exp[(x2 + p2 )/4 + ipx/2] exp[(x0 x)2 /2]f (p) (19)
11
where denotes the Fourier transform with respect to x0 . Thus by

Plancherels theorem (Yosida [1971]),
(2)
ds p exp(p2 /2)|f(x ip)|2

Z
(20)
2
2
s 0
0
2
0 2
= N exp(x /2) d x exp[(x x) ] |f (x )| .
Therefore
Z
d(z) |f(z) |2
Z
Z
2
s
=N
d x ds x0 exp[(x0 x)2 ] |f (x0 ) |2
Z
= ds x0 | f (x0 ) | 2
(21)
after exchanging the order of integration. This proves that

kfk2F = kf k2L2
(22)
for f S(IRs ), hence by continuity also for arbitrary f L2 (IRs ). By

polarization the result can now be extended from the norms to the
inner products.
The relation f f can be summarized neatly and economically
in terms of Diracs bra-ket notation. Since
f(z) = hz |f i = hf |z i,
theorem 1 can be restated as
Z
d(z) hf |z ihz |gi = hf |gi.
(23)
(24)
C
ls
Dropping the bra hf | and ket |gi, we have the operator identity
Z
d(z) |z ihz | = I,
(25)
C
ls
12
where I is the identity operator on L2 (IRs ) and the integral converges

at least in the sense of the weak operator topology,* i.e. as a quadratic
form. In Klauders terminology, this is a continuous resolution of
unity. A general operator B on L2 (IRs ) can now be expressed as an
on F as follows:
integral operator B
f)(z) (Bf )(z)
(B
= h z | Bf i
= h z | BIf i
Z
=
d(w) h z | Bw ih w | f i
C
ls
Z
w)
d(w) B(z,
f(w).
(26)
C
ls
Particularly simple representations are obtained for the basic position and momentum operators. We get
Xk z (x0 ) x0k z (x0 )

zk
=
+
z (x0 )
zk
2
(27)
Pk z = i(Ak Xk )z

zk
z ,
=i
zk
2
thus
k f =
X
zk
+
zk
2
f
(28)
Pk f = i
zk
zk
2
f.
* As will be shown in a more general context in the next section, under favorable conditions the integral actually converges in the strong
operator topology.
13
Hence Xk and Pk can be represented as differential rather than integral operators.

As promised, the continuous resolution of the identity makes it
possible to reconstruct f L2 (IRs ) from its transform f F:
Z
f = If =
d(z) | z ih z | f i,
(29)
C
ls
that is,
Z
f (x ) =
d(z) z (x0 )f(z).
(30)
C
ls
Thus in many respects the coherent states behave like a basis for
L2 (IRs ). But they differ from a basis in at least one important respect:
They cannot all be linearly independent, since there are uncountably
many of them and L2 (IRs ) (and hence also F) is separable. In particular, the above reconstruction formula can be used to express z in
terms of all the w s:
Z
z =
d(w) | w ih w | z i.
(31)
C
ls
In fact, since entire functions are determined by their values on some

discrete subsets of C
l s , we conclude that the corresponding subsets
of coherent states {z | z } are already complete since for any
function f orthogonal to them all, f(z) = 0 for all z and hence
f 0, which implies f = 0 a.e. For example, if is a regular lattice, a
necessary and sufficient condition for completesness is that contain
at least one point in each Planck cell (Bargmann et al., [1971]), in
the sense that the spacings xk and pk of the lattice coordinates
zk = xk + ipk satisfy xk pk 2
h 2. It is no accident that
this looks like the uncertainty principle but with the inequality going
the wrong way. The exact coefficient of h
is somewhat arbitrary
and depends on ones definition of uncertainty; it is possible to define measures of uncertainty other than the standard deviation. (In
fact, a preferablebut less tractabledefinition of uncertainty uses
the notion of entropy, which involves all moments rather than just
the second moment. See BialynickiBirula and Mycielski [1975] and
Zakai [1960].) The intuitive explanation is that if f gets sampled
at least once in every Planck cell, then it is uniquely determined since
the uncertainty principle limits the amount of variation which can
take place within such a cell. Hence the set of all coherent states
14
is overcomplete. We will see later that reconstruction formulas exist

for some discrete subsystems of coherent states, which makes them
as useful as the continuum of such states. This ability to synthesize
continuous and discrete methods in a single representation, as well as
to bridge quantum and classical concepts, is one more aspect of the
appeal and mystery of these systems.
1.3. Generalized Frames and Resolutions of Unity

Let M be a set and be a measure on M (with an appropriate
algebra of measurable subsets) such that {M, } is a finite measure
space. Let H be a Hilbert space and hm H be a family of vectors
indexed by m M .
Definition. The set
HM {hm | m M }
(1)
is a generalized frame in H if
1. the map h: m 7 hm is weakly measurable, i.e. for each f H
the function f(m) h hm | f i is measurable, and
2. there exist constants 0 < A B such that
Z
2
Akf k
d(m) | h hm | f i | 2 Bkf k2
f H.
(2)
M
HM is a frame (see Young [1980] and Daubechies [1988a]) in the special case when M is countable and is the counting measure on M
(i.e., it assigns to each subset of M the number of elements contained
in it). In that case, the above condition becomes
X
A kf k2
| h hm | f i | 2 B kf k2
f H.
(3)
mM
We will henceforth drop the adjective generalized and simply speak

of frames. The above case where M is countable will be refered to
as a discrete frame.
If A = B, the frame HM is called tight. The coherent states of the
last section form a tight frame, with M = Cls , d(m) = d(z), hm =
z and A = B = 1.
Given a frame, let T be the map taking vectors in H to functions
on M defined by

(T f )(m) h hm | f i f(m),
f H.
15
(4)
Then the frame condition states that T f is square-integrable with

respect to d, so that T defines a map
T : H L2 (d),
(5)
A kf k2 kT f k2L2 (d) B kf k2 .
(6)
with
The frame property can now be stated in operator form as

AI T T BI,
where I is the identity on H. In bra-ket notation,
Z
GT T =
d(m) | hm ih hm | ,
(7)
(8)
where the integral is to be interpreted, initially, as converging in the

weak operator topology, i.e. as a quadratic form. For a measurable
subset N of M , write
Z
G(N )
d(m) | hm ih hm |.
(9)
N
Proposition 1.2. If the integral G(N ) converges in the strong operator topology of H whenever N has finite measure, then so does the
complete integral representing G = T T .
Proof *. Since M is finite, we can choose an increasing sequence
{Mn } of sets of finite measure such that M = n Mn . Then the corresponding sequence of integrals Gn forms a bounded (by G) increasing
sequence of Hermitian operators, hence converges to G in the strong
operator topology (see Halmos [1967], problem 94).
If the frame is tight, then G = AI and the above gives a resolution
of unity after dividing by A. For non-tight frames, one generally has
to do some work to obtain a resolution of unity. The frame condition
means that G has a bounded inverse, with
B 1 I G1 A1 I.
* I thank M. B. Ruskai for suggesting this proof.
(10)
16
Given a function g(m) in L2 (d), we are interested in answering the

following two questions: (a) Is g = T f for some f H? (b) If so,
then what is f ? In other words, we want to:
(a) Find the range <T L2 (d) of the map T .
(b) Find a left inverse S of T , which enables us to reconstruct f from
T f by f = ST f .
Both questions will be answered if we can explicitly compute G1 .
For let
P = T G1 T : L2 (d) L2 (d).
(11)
Then it is easy to see that

(a) P = P
(b) P 2 = P
(12)
(c) P T = T.
It follows that P is the orthogonal projection onto the range of T ,
P : L2 (d) <T ,
(13)
for if g = T f for some f in H, then P g = P T f = T f = g, and

conversely if for some g we have P g = g, then g = T (G1 T g) T f .
Thus <T is a closed subspace of L2 (d) and a function g L2 (d) is
in <T if and only if

g(m) = (P g)(m) = T G1 T g (m)
= h hm | G1 T g iH
= h T G1 hm | g iL2 (d)
Z

=
d(m0 ) T G1 hm (m0 )g(m0 )
ZM
=
d(m0 ) h hm0 | G1 hm ig(m0 )
ZM
=
d(m0 ) h hm | G1 hm0 ig(m0 ).
(14)
The function
K(m, m0 ) h hm | G1 hm0 i
(15)
17
therefore has a property similar to the Dirac -function with respect

to the measure d, in that it reproduces functions in <T . But it differs
from the -function in some important respects. For one thing, it is
bounded by
| K(m, m0 ) | = | h hm | G1 h0m i |
kG1 k khm k khm0 k
(16)
A1 khm k khm0 k <

for all m and m0 . Furthermore, the test functions which K(m, m0 )
reproduces form a Hilbert space and K(m, m0 ) defines an integral
operator, not merely a distribution, on <T . In the applications to
relativistic quantum theory to be developed later, M will be a complexification of spacetime and K(m, m0 ) will be holomorphic in m and
antiholomorphic in m0 .
The Hilbert space <T and the associated function
K(m, m0 ) are an example of an important structure called a reproducingkernel Hilbert space (see Meschkowski [1962]), which is reviewed
briefly in the next section. K(m, m0 ) is called a reproducing kernel
for <T .
We can thus summarize our answer to the first question by saying
that a function g L2 (d) belongs to the range of T if and only if it
satisfies the consistency condition
Z
g(m) =
d(m0 ) K(m, m0 )g(m0 ).
(17)
Of course, this condition is only useful to the extent that we have

information about the kernel K(m, m0 ) or, equivalently, about the
operator G1 . The answer to our second question also depends on
the knowledge of G1 . For once we know that g = T f for some f H,
then

f = G1 Gf = G1 T T f = G1 T g.
(18)
Thus the operator

S = G1 T : L2 (d) H
is a left inverse of T and we can reconstruct f by
(19)
18
f = Sg = G1 T T f
Z
1
=G
d(m) | hm ih hm | f i
M
Z
=
d(m) G1 | hm ig(m).
(20)
This gives f as a linear combination of the vectors

hm G1 hm .
(21)
Note that
Z
d(m) | hm ih hm | = G1 GG1 = G1 ,
(22)
therefore the set

HM {hm | m M }
(23)
is also a frame, with frame constants 0 < B 1 A1 . We will

call HM the frame reciprocal to HM . (In Daubechies [1988a], the
corresponding discrete object is called the dual frame, but as we shall
see below, it is actually a generalization of the concept of reciprocal
basis; since the term dual basis has an entirely different meaning,
we prefer reciprocal frame to avoid confusion.)
The above reconstruction formula is equivalent to the resolutions
of unity in terms of the pair HM , HM of reciprocal frames:
Z
Z
m
d(m) | h ih hm | = I =
d(m) | hm ih hm |.
(24)
M
Corollary 1.3. Under the assumptions of proposition 1.2, the above

resolutions of unity converge in the strong operator topology of H.
The proof is similar to that of proposition 1.2 and will not be given.
The strong convergence of the resolutions of unity is important, since
it means that the reconstruction formula is valid within H rather
than just weakly. Application to f = hk for a fixed k M gives
Z
hk =
d(m) hm K(m, k),
(25)
M
which shows that the frame vectors hm are in general not linearly independent. The consistency condition can be understood as requiring
19
the proposed function g(m) to respect the linear dependence of the

frame vectors. In the special case when the frame vectors are linearly independent, the frames HM and HM both reduce to bases of
H. If H is separable (which we assume it is), it follows that M must
be countable, and without loss in generality we may assume that d
is the counting measure on M (renormalize the hm s if necessary).
Then the above relation becomes
X
hk =
hm K(m, k),
(26)
mM
and linear independence requires that K be the Kronecker : K(m, k)

k
= m
. Thus when the hm s are linearly independent, HM and HM
reduce to a pair of reciprocal bases for H. The resolutions of unity
become
X
X
| hm ih hm | = I =
| hm ih hm |,
(27)
mM
mM
and we have the relation

X
X
hm h hm | hk i
hm gmk
hk =
mM
(28)
mM
where
gmk h hm | hk i = h hm | Ghk i
(29)
is an infinite-dimensional version of the metric tensor, which mediates

between covariant and contravariant vectors. (The operator G plays
the role of a metric operator.) In this case, <T = L2 (d) `2 (M ) and
the consistency condition reduces to an identity. The reconstruction
formula becomes the usual expression for f as a linear combination of
the (reciprocal) basis vectors. If we further specialize to the case of a
tight frame, then G = AI implies that
k
h h m | h k i = A m
and
h hm | hk i = A1 km ,
(30)
so HM and HM become orthogonal bases. Requiring A = B = 1

means that HM = HM reduce to a single orthonormal basis.
Returning to the general case, we may summarize our findings
as follows: HM , HM , K(m, m0 ) and g(m, m0 ) h hm | Ghm0 i are
generalizations of the concepts of basis, reciprocal basis, Kronecker
20
delta and metric tensor to the infinitedimensional case where, in addition, the requirement of linear independence is dropped. The point
is that the all-important reconstruction formula, which allows us to
express any vector as a linear combination of the frame vectors, survives under the additional (and obviously necessary) restriction that
the consistency condition be obeyed. The useful concepts of orthogonal and orthonormal bases generalize to tight frames and frames with
A = B = 1, respectively. We will call frames with A = B = 1 normal.
Thus normal frames are nothing but resolutions of unity.
Returning to the general situation, we must still supply a way of
computing G1 , on which the entire construction above depends. In
some of the examples to follow, G is actually a multiplication operator,
so G1 is easy to compute. If no such easy way exist, the following
procedure may be used. From AI G BI it follows that
1
1
1
(B A)I G (B + A)I (B A)I.
2
2
2
(31)
Hence letting
=
BA
B+A
and
c=
2
,
B+A
(32)
we have
I I cG I.
(33)
Since 0 < 1 and c > 0, we can expand

G
X

1
(I cG)k
= c I (I cG)
=c
(34)
k=0
and the series converges uniformly since

kI cGk < 1.
(35)
The smaller , the faster the convergence. For a tight frame, = 0

and cG = I, so the series collapses to a single term G1 = c. Then
the consistency condition becomes
Z
g(m) = c
d(m0 ) h hm | hm0 ig(m0 ),
(36)
M
and the reconstruction formula simplifies to
21
Z
f =c
d(m) hm g(m).
(37)
If 0 < 1, the frame is called snug (Daubechies [1988a]) and

the above formulae merely represent the first terms in the expansions. However, it is found in practice that under certain conditions,
this first term gives a good approximation to the reconstruction for
snug frames. It appears to be an advantage to have much linear dependence among the frame vectors (precisely that which is impossible
when dealing with bases!), so the transformed function f(m) is oversampled. For such frames, the oversampling seems to compensate for
the truncation of the series. A good measure of the amount of linear
dependence among the hm s is the size of the orthogonal complement
<
T of the range of T , which is the null space N (T ) of T since
g <
T h g | T f i = 0
h T g | f i = 0
f H
f H
(38)
T g = 0.
Suppose we are given some function g(m) L2 (d) and apply
the reconstruction formula to it blindly, without worrying whether
the consistency condition is satisfied. That is, consider the vector
f G1 T g
(39)
in H. Since g can be uniquely written as an orthogonal sum g = g1 +g2
where g1 <T and g2 <

T = N (T ), we find that
f = G1 T g1
(40)
where g1 does satisfy the consistency condition. This means that

applying the reconstruction formula to an arbitrary g L2 (d) results
in a leastsquares approximation f to a reconstruction, in the sense
that the error g2 g g1 = g T f has a minimal norm in L2 (d).
For example, suppose we only have information about f(m) for
m in some subset of M . Let

g(m) = f (m) if m
(41)
0
if m
/ .
Then g does not belong to <T , in general, and f1 G1 T g is the
leastsquares approximation to the (unknown) vector f in view of our
22
ignorance. A similar argument applies if M is discrete and f must be

reconstructed from T f by numerical methods. Then we must confine
ourselves to a finite subset of M . The above procedure then gives
a leastsquares approximation to f by truncating the reconstruction
formula to a finite sum over .
A final note: The usual argument in favor of using bases rather
than overcomplete sets of vectors is that one desires a unique representation of functions as linear combinations of basis elements. When
a frame is not a basis, i.e. when the frame vectors are linearly dependent, this uniqueness is indeed lost since we may add an arbitrary
function e(m) <

T to f (m) in the reconstruction formula without
changing f (<T 6= {0} since the frame vectors are dependent). However, there is still a unique admissible coefficient function, i.e. one
satisfying the consistency condition. Moreover, as we shall see, it
usually happens in practice that the set M , in addition to being a
measure space, has some further structure, and the reproducing kernel K(m, m0 ) preserves this structure. For example, M could be a
topological space and K be continuous on M M , or M could be a
differentiable manifold and K be differentiable, or (as will happen in
our treatment of relativistic quantum mechanics) M could be a complex manifold and K be holomorphic. Furthermore, K could exhibit
a certain boundary or asymptotic behavior. In such cases, these
properties are inherited by all the functions f(m) in <T , and then of
all possible coefficient functions for a given f H, there is only one
which exhibits the appropriate behavior. In this sense, uniqueness is
restored. We will refer to frames with such additional properties as
continuous, differentiable, holomorphic, etc.
In addition to properties such as differentiability or holomorphy,
the kernels K we will encounter will usually have certain invariance
properties with respect to some group of transformations acting on
M . This, too, will induce a corresponding structure on the function
space <T .
1.4. ReproducingKernel Hilbert Spaces

The function K(m, m0 ) of section 1.3 is an example of a very general
structure called a reproducing kernel, which we now review briefly
since it, too, will play an important role in the chapters to come.
The reader interested in learning more about this fascinating subject
may consult Aronszajn [1950], Bergman [1950], Meschkowski [1962]
23
or Hille [1972].
Suppose we begin with an arbitrary set M and a set of functions
g(m) on M which forms a Hilbert space F under some inner product
h | i. In section 1.3, M happened to a measure space, F was <T
and the inner product was that of L2 (d). But it is important to
keep in mind that the exact form of the inner product in F does
not need to be specified, as far as the general theory of reproducing
kernels is concerned. Suppose we are given a complexvalued function
K(m, m0 ) on M M with the following two properties:
1. For every m M , the function Km (m0 ) K(m0 , m) belongs to
F.
2. For every m M and every g F, we have
g(m) = h Km | g i.
(1)
Then F is called a reproducingkernel Hilbert space and

K(m, m0 ) is called its reproducing kernel. Some properties of K follow immediately from the definition. For example, since Km F,
property (2) implies that
K(m0 , m) = Km (m0 ) = h Km0 | Km i.
(2)
Thus K must satisfy

(a) K(m, m0 ) = K(m0 , m)
(b) K(m, m) = kKm k2 0 m M
(3)
(c) |K(m, m0 )| kKm k kKm0 k.

One of the most important and useful aspects of reproducing
kernel Hilbert spaces is the fact that the kernel function itself virtually
generates the whole structure. For example, the function L(m)
kKm k dominates every g(m) in F in the sense that
|g(m)| = | h Km | g i | kKm k kgk = kgkL(m)
(4)
by the Schwarz inequality. In particular, all the functions in F inherit

the boundednes and growth properties of K. If K has singularities,
then some functions in F have similar singularities.
From the above it follows that for fixed m0 M ,
sup |g(m0 )| = L(m0 ),
kgk=1
(5)
24
the supremum being attained by gm0 (m) = Km0 (m)/L(m0 ), and this
is, up to an overall phase factor, the only function which attains the
supremum. This fact has an important interpretation when M is the
classical phase space of some system, F represents the Hilbert space
of the corresponding quantum mechanical system and |g(m)|2 is the
probability density in the quantum mechanical state g of finding the
system in the classical state m. The above inequality then shows that
the state which maximizes the probability of being at m0 is uniquely
determined (up to a phase factor) as gm0 . In other words, gm0 is a
wave packet which is in some sense optimally localized at m0 in phase
space.
Another example of how the kernel function K embodies the properties of the entire Hilbert space F is the following: If the basic set M
has some additional structure, such as being a measure space (as it
was in the setting of frame theory) or a topological space or a C k , C ,
realanalytic or complex manifold, and if K(m, m0 ), as a function of
m, preserves this structure (that is, K(m, m0 ) is measurable, continuous, C k , C , realanalytic or holomorphic in m, respectively), then
every function g(m) in F has the same property.
The question arises: If we are given a Hilbert space F whose
elements are all functions on M , how do we know whether this space
posesses a reproducing kernel? Clearly a necessary condition for K
to exist is that
| f (m) | = | h Km | f i | kKm k kf k
(6)
for all f F and all m M . But this means that for every fixed m,
the map Em : F Cl defined by
Em (f ) = f (m)
(7)
is a bounded linear functional on F. Em is called the evaluation map

at m. By the Riesz representation theorem, every bounded linear
functional E on F must have the form E(f ) = h e | f i for a unique
vector e F. Hence there exists a unique em F such that
f (m) = h em | f i
(8)
for all f F. Since also f (m) = h Km | f i, we conclude that em =

Km . Thus it follows that given only F such that all the evaluation maps Em are bounded (not necessarily uniformly in m), we can
construct a reproducing kernel by

K(m, m0 ) = h em | em0 i.
25
(9)
That is, F posesses a reproducing kernel if and only if all the evaluation maps are bounded.
In all of the applications we will encounter, the set M will have
a structure beyond those mentioned so far: it will be a Lie group G
or a homogeneous space of G. That is, each g G acts on M as an
invertible transformation preserving whatever other structure M may
have such as continuity, differentiability, etc, and these tranformations
form the group G under composition. Let us denote the action of g
on M by m 7 mg (i.e., G acts from the right). Then the operator
Tg : F F
(Tg f )(m) = f (mg)
(10)
gives a representation of G on F, since Tg Tg0 = Tgg0 . From f (m) =

h Km | f i it now follows that
Kmg = Tg Km ,
(11)
K(m0 g, mg) = h Km0 g | Kmg i = h Km0 | Tg Tg Km i.
(12)
hence
Therefore the reproducing kernel K(m0 , m) is invariant under the action of G if and only if all the operators Tg are unitary, i.e. the
representation g 7 Tg is unitary.
The group G usually appears in applications as a natural set of
operations such as translations in space and time (evolution), changes
of reference frame or coordinate system, dilations and frequency shifts
(especially useful in signal analysis), etc. Invariance under G then
means that the transformed objects (wave functions, signals) form a
description of the system equivalent to the original one, i.e. that the
transformation changes nothing of physical significance.
Finally, let us note that the concept of generalized frame, as
defined in the previous section, is a special case of a reproducing
kernel Hilbert space. For given such a frame {hm }, the function
K(m0 , m) = h hm0 | G1 hm i is a reproducing kernel for the space <T
since (a):

Km (m0 ) K(m0 , m) = T (G1 hm ) (m0 )
(13)
26
shows that Km = T (G1 hm ) belongs to <T , and (b):

Z
f (m) =
d(m0 ) K(m, m0 )f(m0 ) = h Km | f i<T .
(14)
1.5. Windowed Fourier Transforms

The coherent states of section 2 form a (holomorphic) frame. I now
want to give some other examples of frames, in order to develop a
better feeling for this concept. The frames to be constructed in this
section turn out to be closely related to the coherent states, but have a
distinct signal processing flavor which will lend some further depth
to our understanding of phase-space localization.
For simplicity, we begin with the study of functions f (t) of a single
real variable which, for motivational purposes, will be thought of as
time. Although in the model to be built below f is real-valued, we
allow it to be complex-valued since our eventual applications will be
quantummechanical. The same goes for the window function h. The
extension to functions of several variables, such as wave functions or
physical fields in spacetime, is straightforward and will be indicated
later. We think of f (t) as a timesignal, such as the voltage going
into a speaker or the pressure on an eardrum. We are interested in
the frequency content of this signal. The standard thing to do is to
find its Fourier coefficients (if f is periodic) or its Fourier transform
(if it is not periodic). But if f represents, say, a symphony, this
approach is completely inappropriate. We are not interested in the
total amplitude
Z
f() =
dt e2it f (t)
(1)
in f of each frequency . For one thing, both we and the musicians

would be long gone before we got to enjoy the music. Moreover,
the musical content of the signal, though coded into f, would be as
inaccessible to us as it is in f (t) itself. Rather, we want to analyze the
frequency content of f in real time. At each instant t = s we hear a
spectrum f(, s). To accomplish this, let us make a simple (though
not very realistic) linear model of our auditory system. We will speak
of the ear, though actually the frequency analysis appears to be
performed partly by the nervous system as well (Roederer [1975]).
27
The ears output at time s depends on the input f (t) for t in some
interval s t s, where is a lagtime characteristic of the ear.
In analyzing f (t) in this interval, the ear may give different weights
to different parts of f (t). Thus the signal to be analyzed for output
at time s may be modeled as
fs (t) = h(t s) f (t)
(2)
where h(t s) is the weight assigned to f (t). (As noted above, we are
allowing h to be complexvalued for future applications, though here
it should be realvalued; the bar means complex conjugation.) The
function h(t) is characteristic of the ear, with support in the interval
t 0. Such functions are known in communication theory as
windows.
Having localized the signal around time s, we now analyze its
frequency content by taking the Fourier transform:
Z
f (, s) = fs () =
dt e2it h(t s) f (t).
(3)
This function, representing our dynamical spectrum, is called a

windowed Fourier transform of f (Daubechies [1988a]). If the window
is flat (h(t) 1), f(, s) reduces to the ordinary Fourier transform.
On the other hand, if f (t) is a unit impulse at t = 0, i.e. f (t) =
(t), then f(, s) = h(s). Hence h(s) is the impulse response
(Papoulis [1962]) of the ear. Note that
Z
f(0, s) =
dt h(t s) f (t)
(4)
is nothing but the convolution of f with the impulse response. Thus

the windowed Fourier transform is a marriage between the Fourier
transform and convolution. Letting
h,s (t) = e2it h(t s),
(5)
f(, s) = h h,s | f i,
(6)
we have
the inner product being in L2 (IR). As our notation implies, we want

to make a frame indexed by the set
28
M = {(, s) | , s IR} IR2 ,
(7)
that is, the time-frequency plane. M corresponds to the phase space

in quantum mechanics in the sense that if t were the position coordinate, then would be the wavenumber and 2
h would be the momentum. Since we expect all times and all frequencies to be equally
important, let us guess that an appropriate measure on M is the
Lebesgue measure, d(, s) = ds d. Now

s f ()
f(, s) = h
(8)
s (t) = h(t s) is the translated window and denotes the
where h
Fourier transform. Thus for nice signals (say, in the space of
Schwartz test functions), Plancherels theorem gives
Z
ds d | f(, s) | 2 =
Z
ds
s f )() | 2
d | (h
IR2
dt | hs (t)f (t) | 2
Z
Z
2
= dt | f (t) |
ds | h(t s) | 2
=
ds
(9)
= khk2 kf k2 ,
where both norms are those of L2 (IR). This shows that the family of vectors h,s is indeed a tight frame, with frame constants
A = B = khk2 . Now h,s = exp(2it)hs (t) is just a translation of
hs in frequency, that is,
s ( 0 ).
,s ( 0 ) = h
h
(10)
Thus if the Fourier transform h()

of the basic window is concentrated
near = 0 (which, among other things, means that h(t) must be
fairly smooth, since any discontinuities or sharp edges introduce highfrequency components), then that of h,s is concentrated near 0 = .
In other words, h,s is actually a window in phase space or time
frequency. The fact that these windows form a tight frame means
that signals may be equally well represented in timefrequency as in
time. The resolution of unity is
29
khk
ds d | h,s ih h,s | = I,
(11)
IR2
and if the consistency condition is satisfied the reconstruction formula

gives
2
ds d e2it h(t s) f(, s).
f (t) = khk
(12)
IR2
As in section 1.2, we denote by HM the frame formed by the h,s s.

Incidentally, the windowed Fourier transform is quite symmetrical
with respect to the interchange of time and frequency. It can be re
written as

s f ()
f(, s) = h

= hs f ()
Z
0
2is
0 ) f( 0 ),
=e
d 0 e2i s h(
(13)
now plays the role of a window in frequency used to localize

where h
f . As will be seen in the next chapter, the reason for this symmetry is
that windowed Fourier transforms, like the canonical coherent states,
are closely related to the WeylHeisenberg group, which treats time
and frequency in a symmetrical fashion. (This is rooted in symplectic
geometry.)
The h,s s are highly redundant. We want to find discrete subsets
of them which still form a frame, that is we want discrete subframes.
The following construction is taken from Kaiser [1978c, 1984a]. Let
T > 0 be a fixed time interval and suppose we sample the output
signal f(, s) only at times s = nT where n is an integer. To discretize the frequency as well, note that the localized signal fnT (t) =
nT (t)f (t) has compact support in the interval nT t nT ,
h
hence we can expand it in a Fourier series
fnT (t) =
X
m
where F = 1/ and
e2imF t cmn
(14)
30
nT
dt e2imF t fnT (t)
cmn = F
nT
nT
dt e2imF t h(t nT ) f (t)
=F
(15)
nT
= F f(mF, nT ).
The only problem with this representation of fnT is that it only holds
in the interval nT t nT , since fnT vanishes outside this
interval while the Fourier series is periodic. To force equality at all
times, we multiply both sides by hnT (t):
hnT (t) fnT (t) = |hnT (t)|2 f (t)
X
=
e2imF t hnT (t) f(mF, nT )
m
(16)
hmF,nT (t) h h,mF,nT | f i.
Attempting to recover the entire signal rather than just pieces of it,
we now sum both sides with respect to n:
X
|h(t nT )|
f (t) =
hmF,nT (t)h hnT,mF | f i.
(17)
n.m
Recovery of f (t) is possible provided that the sum on the left converges
to a function
X
g(t)
|h(t nT )|2
(18)
n
which is bounded above and below by positive constants:

0 < A g(t) B.
(19)
In that case, we have

X
| h hmF,nT | f i | =
n,m
and hence the subset
dt g(t)|f (t)|2
(20)

T,F
HM
{hmF,nT | m, n ZZ}
31
(21)
forms a discrete subframe of HM with frame constants A and B.

This subframe is in general not tight, and the operator G is just
multiplication by g(t), so finding G1 presents no problem in this
case. The reconstruction formula is
X
f (t) = g(t)1
hmF,nT (t)f(mF, nT ).
(22)
n,mZZ
We are therefore able to recover the signal by sampling it in phase

space at time intervals t = T and frequency intervals = F .
What about the uncertainty principle? It is hiding in the condition
that the discrete subset hmF nT forms a frame! For we cannot satisfy
g(t) A > 0 unless the supports of h(t) and h(t T ) overlap, which
implies that T , or t = T / 1. In quantum mechanics, the
radian frequency is related to the energy by E = 2
h, so the above
condition is
t E 2
h.
(23)
This looks like the uncertainty principle going the wrong way. The
intuitive explanation for this has been discussed at the end of section
1.3.
Notice that the closer we choose T to , the more difficult it is
for the window function to be smooth. In the limiting case T = ,
h(t) must be discontinuous if the above frame condition is to be
obeyed. As noted earlier, this means that its Fourier transform h()
can no longer be concentrated near = 0, so the frequency resolution of the samples suffers. In concrete terms, this means that
whereas for nice windows h(t) we may hope to get a good approximation to the reconstruction formula by truncating the double sum
after an appropriate finite number of terms, this can no longer be
expected when T . In other words, it pays to oversample! Appropriate here means that we cover most of the area in the time
frequency plane where f lives. Clearly, if the sampling is done only
for |n| N , we cannot expect to recover f (t) outside the interval
N T t N T . If h()
is nicely peaked around = 0, say with a
spread of , and the signal f (t) is (approximately) bandlimited,
so that f() 0 for || W , then we can expect to get a good
approximation to f (t) by truncating the sum with m M , where
32
M is chosen to satisfy M F > W + . On the other hand, we may

actually not be interested in recovering the exact original signal f (t).
If we are only interested in a particular time interval and a particular
frequency interval (say, to eliminate some highfrequency noise), then
the appropriate truncation would include only an area in the time
frequency plane which is slightly larger than our area of interest. If
we sample a given signal f only in a finite subset of our lattice (as
we are bound to do in practice) and then apply the reconstruction
formula, truncating the sum by restricting it to , then the result is
a leastsquares approximation f1 to the original signal.
On a philosophical note, suppose we are given an arbitrary signal
without any idea of the type of information it carries. The choice of
a window h(t) then defines a scale in time, given by , and it is this
scale that distinguishes what will be perceived as frequency and what
as time. Thus, variations of f (t) in time intervals much smaller than
are reflected in the behavior of f, while variations at scales much
larger than survive in the sdependence of f. What an elephant
perceives as a tone may appear as a rhythm to a mouse.
Finally, I wish to compare the above reconstruction with the
Nyquist/Shannon sampling theorem (Papoulis [1962]), which gives
a reconstruction for band-limited signals ( f() 0 for || W )
in terms of samples of f (t) taken at times t = n/2W for integer n.
Note, first of all, that in the timefrequency reconstruction formula
above, it was not necessary to assume that f (t) is band-limited. In
fact, the formula applies to all squareintegrable signals. But suppose
that f (t) is bandlimited as above, and choose a window h(t) such
that h()
is concentrated around an interval of width about the
origin, with < W . Then we expect f(, s) 0 for || W +
. In order to reduce the double sum in our reconstruction formula to
a single sum over n as in the Nyquist theorem, choose F = W + .
Then f(mF, s) 0 whenever m 6= 0. To apply the reconstruction
formula, we must still choose a time interval T < . By the uncertainty principle, > 1/2. The above condition on therefore
implies that > 1/2W . It would thus seem that we could get away
with a slightly larger sampling interval T than the Nyquist interval
TN = 1/2W . Our reconstruction formula reduces to
X
f (t) g(t)1
h(t nT ) f(0, nT ).
(24)
n
The smaller the ratio /W , the better this approximation is likely
1.6. Wavelet Transforms
33
to be. But a small means a large , hence the samples f(0, nT )

are smeared over a large time interval.

The frame vectors for windowed Fourier transforms were the wave
packets
h,s (t) = e2it h(t s).
(1)
The basic window function h(t) was assumed to vanish outside of the
interval < t < 0 and to be reasonably smooth with no steep slopes,
so that its Fourier transform h()

was also centered in a small interval
about the origin. Of course, since h(t) has compact support, h()
is
the restriction to IR of an entire function and hence cannot vanish on
any interval, much less be of compact support. The above statement
simply means that h()

decays rapidly outside of a small interval containing the origin. At any rate, the factor exp(2it) amounted to a
translation of the window in frequency, so that h,s was a window in
the timefrequency plane centered about (, s). Hence the frequency
components of f (t) were picked out by means of rigidly translating the
basic window in both time and frequency. (It is for this reason that
the windowed Fourier transform is associated to the WeylHeisenberg
group, which is exactly the group of all translations in phase space
amended with the multiplication by phase factors necessary to close
the Lie algebra, as explained in chapter 3.) Consequently, h,s has
the same width for all frequencies, and the number of wavelengths
admitted for analysis is . For low frequencies with 1 this is inadequate since we cannot gain any meaningful frequency information
by looking at a small fraction of a wavelength. For high frequencies
with 1, too many wavelengths are admitted. For such waves,
a time-interval of duration seems infinite, thus negating the sense
of locality which the windowed Fourier transform was designed to
achieve in the first place. This deficiency is remedied by the wavelet
transform. The window h(t), called the basic wavelet, is now scaled
to accomodate waves of different frequencies. That is, for a 6= 0 let

ts
1/2
ha,s (t) = |a|
h
.
(2)
a
The factor |a|1/2 is included so that
34
kha,s k
dt | ha,s (t) | 2 = khk2 .
(3)
The necessity of using negative as well as positive values of a will

become clear as we go along. It will also turn out that h will need to
satisfy a technical condition. Again, we think of both h and f as real
but allow them to be complex. The wavelet transform is now defined
by

Z
ts
1/2
f (t).
(4)
f (a, s) = h H | f i =
dt |a|
h
a
Before proceeding any further, let us see how the wavelet transform localizes signals in the timefrequency plane. The localization
in time is clear: If we assume that h(t) is concentrated near t = 0
(though it will no longer be convenient to assume that h has compact support), then f(a, s) is a weighted average of f (t) around t = s
(though the weight function need not be positive, and in general may
even be complex). To analyze the frequency localization, we again
want to express f in terms of the Fourier transforms of h and f . This
is possible because, like the windowed Fourier transform, the wavelet
transform involves rigid timetranslations of the window, resulting in
a convolutionlike expression. The impulse response is now (setting
f (t) = (t))
ga (s) = |a|1/2 h(s/a),
(5)

f(a, s) = ga f (s) = ga f (s),
(6)
and we have
with
Z
ga () =
dt e2it |a|1/2 h(t/a) = |a|1/2 h(a).
(7)
Later we will see that discrete tight frames can be obtained with certain choices of h(t) whose Fourier transforms have compact support
in a frequency interval interval . Such functions (or, rather,
the operations of convolutiong with them) are called bandpass filters
in communication theory, since the only frequency components in f (t)
to survive are those in the band [, ]. Then the above expression
35
shows that f(a, s) depends only on the frequency component of f (t)

in the band /a /a (if a > 0) or /a /a (if a < 0).
Thus frequency localization is achieved by dilations rather than translations in frequency space, in contrast to the windowed Fourier transform. At least from the point of view of audio signals, this actually
seems preferable since it appears to be frequency ratios, rather than
frequency differences, which carry meaning. For example, going up
an octave is achieved by doubling the frequency. (However, frequency
differences do play a role in connection with beats and also in certain
nonlinear phenomena such as difference tones; see Roederer [1975].)
Let us now try to make a continuous frame out of the vectors ha,s .
This time the index set is
M = {(a, s) | a 6= 0, s IR} IR IR
(8)
where IR denotes the group of non-zero real numbers under multiplication. M is the affine group of translations and dilations of the
real line, t0 = at + s, and this fact will be recognized as being very
important in chapter 3. But for the present we use a more pedestrian
approach to obtain the central results. This will make the power and
elegance of the grouptheoretic approach to be introduced later stand
out and be appreciated all the more. At this point we only make the
safe assumption that the measure d on M is invariant under time
translations, i.e. that
d(a, s) = (a)da ds
(9)
where (a) is an as yet undetermined density on IR . Then, using

Plancherels theorem,
Z
M
d(a, s) | f(a, s) | 2 =
(a) da ds | (
ga f)(s) | 2
ZM
Z
=
(a)da
d | ga () | 2 | f() | 2
ZIR
ZIR
=
(a)da
d |a| | h(a)
| 2 | f() | 2
IR
ZIR
=
d H() | f() | 2
= h f | H f iL2 (IR) ,
(10)
36
where
Z
(a) | a | da | h(a)
|2
H() =
IR
i
2
2
ada (a) | h(a) | + (a) | h(a) | .
(11)
The frame condition is therefore A H() B for some positive

constants A and B. To see its implications, we analyze the cases
> 0 and < 0 separately. If > 0, let = a. Then
H() =
h
i
| 2 + (/) | h()
d (/) | h()
|2 .
(12)
If < 0, let = a. Then
H() =
h
i
|2 ,
d (/) | h()
| 2 + (/) | h()
(13)
giving the same expression. Therefore the frame condition requires

that

(/) = O (/)2
as / ,
(14)
unless h()
vanishes in a neighborhood of the origin. Note that the
above expression for H() shows that if both (a) and h()
vanish
for negative arguments then H() 0 and no frame exists. Hence to
support general complex-valued windows (such as bandpass filters for
a positivefrequency band), it is necessary to include negative as well
as positive scale factors a.
The general case, therefore, is that we get a (generalized) frame
whenever (a) and h(t) are chosen such that 0 < A H() B is
satisfied. The metric operator G and its inverse are given in terms
of Fourier transforms by

Gf = H f
and
G1 f = H 1 f .
(15)
Since G is no longer a multiplication operator in the time domain
(as it was in the case of the discrete frame we constructed from the
37
windowed Fourier transforms), the action of G1 is more complicated.

It is preferable, therefore, to specialize to tight frames. This requires
that H() be constant, so the asymptotic conditions on reduce to
the requirement that be piecewise continuous:
+ 2
c /a , if a > 0
(16)
(a) =
c /a2 , if a < 0
where c+ and c are non-negative constants (not both zero). Then
for > 0,
Z
i
d h +
H() =
c | h() | 2 + c | h()
|2
(17)
0
and for < 0,
Z
H() =
0
i
d h
2
+
2
c | h() | + c | h() | .
(18)
| 2 = | h()
Thus H() = A = B requires either that | h()

| 2 (which
+
holds if h(t) is real) or that c = c . Since we want to accomodate

complex wavelets, we assume the latter condition. Then we have
Z
d 2
+
A=B=c
h() .
(19)
IR ||
We have therefore arrived at the measure
d(a, s) = c+ da ds/a2
(20)
for tight frames, which coincides with the measure suggested by group
theory (see chapter 3). In addition, we have found that the basic
wavelet h must have the property that
Z
d 2
h() < .
(21)
ch
IR ||
In that case, h(t) is said to be admissible. This condition is also a special case of a grouptheoretic result, namely that we are dealing with
a squareintegrable representation of the appropriate group (in this
case, the affine group IR IR). To summarize, we have constructed
a continuous tight frame of wavelets ha,s provided the basic wavelet
is admissible. The corresponding resolution of unity is
38
c1
h
Z
IR IR
da ds
| ha,s ih ha,s | = I.
a2
(22)
The associated reproducingkernel Hilbert space <K is the space of

functions (Kf )(a, s) = h ha,s | f i f(a, s) depending on the scale
parameter a as well as the time coordinate s. As a 0, ha,s becomes
peaked around t = s and
f |a|1/2 cf (s)
(23)
R
where c = h(u) du. The transformed signal f is a smoothedout
version of f and a serves as a resolution parameter.
Ultimately, all computations involve a finite number of operations, hence as a first step it would be helpful to construct a discrete
subframe of our continuous frame. Toward this end, choose a fundamental scale parameter a > 1 and a fundamental time shift b > 0.
We will consider the discrete subset of dilations and translations
D = {(am , nam b) | m, n ZZ } IR IR.
(24)
Note that since am > 0 for all m, only positive dilations are included in
D, contrary to the lesson we have learned above. This will be remedied
later by considering h(t) along with h(t). Also, D is not a subgroup
of IR IR, as can be easily checked. The wavelets parametrized by
D are

t nam b
m/2
hmn = a
h
= am/2 h(am t nb).
(25)
am
To see that this is exactly what is desired, suppose h()

is con
centrated on an interval around = F (i.e., h is a bandpass filter).
mn is concentrated around = F/am . For given integer m,
Then h
the samples
fmn h hmn | f i,
n ZZ
(26)
therefore represent (in discrete time n) the behavior of that part of

the signal f (t) with frequencies near F/am . If m 1, fmn will vary
slowly with n, and if m 1, it will vary rapidly with n (if f (t) has
frequency components with F/am ). Now the timesamples are
separated by the interval m t = am b, so the sampling rate

Rm = 1/am b
39
(27)
is automatically adjusted to the frequency range F/am of the

output signal {fmn | n ZZ}: The highfrequency components get
sampled proportionately more often. This is an example of a topic in
mathematics which has recently attracted intense activity under the
banner of multiscale analysis (Mallat [1987], Meyer [1986]) and which
is in fact closely related to the subject of wavelet transforms.
Returning to the construction of a discrete frame, Parsevals formula gives
mn | f i
fmn = h hmn | f i = h h
Z
=
d kmn () f()
(28)
mn () is the Fourier transform of hmn (t),

where kmn () h
m ).
kmn (v) = am/2 exp(2inam b) h(a
(29)
We now assume that k() h()

vanishes outside the interval
I0 = {F/a F a},
(30)
where F > 0 is some fixed frequency to be determined below. The

width of the band I0 is W0 = (a a1 )F . Therefore the function
k(am )f() is supported on the compact interval

Im = F/am , F/am1
(31)
of width Wm = W0 /am , and we can expand it in a Fourier series in
that interval:
X
k(am ) f() =
exp(2in/Wm ) cmn ,
(32)
n
where
cmn =
1
Wm
Comparing this with
d exp(2in/Wm ) k(am ) f().
(33)
40
Z
d am/2 exp(2inbam ) k(am )f()
fmn =
(34)
suggests that we choose F so that Wm = 1/am b, which gives

F =
(a2
a
1) b
(35)
and
cmn = am/2 bfmn .
(36)
The Fourier series representation above only holds in the interval

Im , since the lefthand side vanishes outside this interval while the
righthand side is periodic. To get equality for all frequencies and
reconstruct f (t), multiply both sides by k(am ) and sum over m:
!
X
| k(am ) | 2
f() = b
=b
am/2 exp (2inam b) k(am ) fmn
m,n
kmn () fmn .
m,n
(37)
To have a frame we would need the series on the lefthand side to
converge to a function + () with
X
0 < A + ()
| k(am ) | 2 B
(38)
m
for some constants A and B. But this is a priori impossible, since
h()
is supported on an interval of positive frequencies and am > 0,
+
so () vanishes for 0. However, we can choose h(t), a and b
such that + () satisfies the frame condition for > 0. Negative frequencies will be taken care of by starting with the complexconjugate
of the original wavelet. We adopt the notation
h+ (t) h(t),
h (t) h(t).
(39)
Then the Fourier transforms k of h are related by

k () = k + (),
(40)
41
hence k () is supported on I0 = [F a, F/a]. A similar argument

to the above gives
!
X
X
m
2
k () fmn
(41)
| k (a ) |
f() = b
mn
m,n
and
()
| k (am ) | 2 = + ().
(42)
m
+
Hence if satisfies the frame condition for > 0, then

0 < A + () + () B
(43)
for all 6= 0. Since {0} has zero measure in frequency space, the
frame condition is satisfied by the joint set of vectors
a,b
+
HM
= {kmn
, kmn
| m, n ZZ}.
(44)
The metric operator

X

| kmn
ih kmn
|
(45)
(Gf )(t) = ( + )f (t)
(46)
= m,nZZ
is given by

and satisfies the frame condition 0 < AI G B. Since G is no

longer a multiplication operator in the time domain (as was the case
with the discrete frame connected to the windowed Fourier transform), the recovery of signals would be greatly simplified if the frame
was tight. The following construction is borrowed from Daubechies
[1988a]. Let F = a/(a2 1)b as above and let k be any nonnegative
integer or k = . Choose a realvalued function C k (IR) (i.e.,
is k times continuously differentiable) such that

0
for x 0
(47)
(x) =
/2 for x 1.
(Such functions are easily constructed; they are used in differential
geometry, for example, to make partitions of unity; see Warner [1971].)
Define h(t) through its Fourier transform k + () by
42
h
i
F/a
sin
for F
F F/a
+
k () =
h
i
cos F
, for F.
aF F
(48)
Note that k + () is C k since the derivatives of (x) up to order

k all vanish at x = 0 and x = /2. This means that the wavelets in
the frame we are about to construct are all C k . Also, k + vanishes
outside the interval I0 = [F/a, F a]. The width of its support is
W0 = (a a1 )F , and for each frequency > 0 there is a unique
integer M such that F/a < aM F , hence also F < aM +1 aF .
Therefore, for > 0,

a1 F
F
2
+ cos
= 1.
() = sin
F a1 F
aF F
+
(49)
Thus
+
() =
0
1
for 0
for > 0,
(50)
i.e., + () is the indicator function for the set of positive numbers.

It follows that () is the indicator function for the negative reals,
and
+ + = 1
a.e.
(51)
This choice of k + and k = k + gives us a tight frame,

X X

| kmn
ih kmn
| = I.
(52)
= m,nZZ
This frame is not a basis; if it were, it would have to be an orthonormal basis since it is a normal frame, hence the reproducing
kernel would have to be diagonal. But
0

K(, m, n; 0 , m0 , n0 ) h kmn
| km
0 n0 i
(53)
does not vanish for 0 = , n0 = n and m0 = m 1, due to the overlap

of wavelets with adjacent scales. However, it is possible to construct
orthonormal bases of wavelets which, in addition, have some other
43
surprising and remarkable properties. For example, such bases have

been found (Meyer [1985], Lemarie and Meyer [1986]) whose Fourier
transforms, like those above, are C with compact support and which
are, simultaneously, unconditional bases for all the spaces Lp (IR) with
1 < p < as well as all the Sobolev spaces and some other popular spaces to boot. Similar bases were constructed in connection with
quantum field theory (Battle [1987]) which are only C k for finite k but,
in return, are better localized in the time domain (they have exponential decay). The concept of multiscale analysis (Mallat [1987], Meyer
[1986]) provided a general method for the construction and study of
orthonormal bases of wavelets. This was then used by Daubechies
[1988b] to construct orthonormal bases of wavelets having compact
support and arbitrarily high regularity.
The mere existence of such bases has surprised analysts and made
wavelets a hot new topic in current mathematical research. They
are also finding important applications in a variety of areas such as
signal analysis, computer science and quantum field theory. They are
the subject of the next chapter, where a new, algebraic, method is
developed for their study.
44
3. Frames and Lie Groups

Chapter 3
FRAMES AND LIE GROUPS
3.1. Introduction
Although we have not sought to exploit it until now, it is clear that
all our frames so far have been obtained with the aid of group operations. The frames associated with the canonical coherent states
and the windowed Fourier transform were built using translations in
phase space (WeylHeisenberg group), while the wavelet frames used
translations and dilations (the affine group). In this chapter we look
for a unifying pattern in these constructions based on group theory.
We analyze the foregoing constructions in turn, and draw separate
lessons from each. It will be natural to work in reverse order. The
affine group, which is, in some sense, the simplest, will lead us to
the general method. Successive refinements will be suggested by the
windowed Fourier transform and the canonical coherent states.
3.2. Klauders GroupFrames

This was the first of the grouptheoretic constructions, pioneered by
J. R. Klauder [1960, 1963a, 1963b], who was also the first to apply it
to the affine group G = IR IR (Aslaksen and Klauder [1968, 1969]).
An element g = (a, s) of G acts on the real line (time) by dilation
followed by translation:
gt (a, s)t = at + s.
(1)
The groupcomposition law is given by

g 0 gt = (a0 , s0 )(a, s)t = a0 (at + s) + s0 = (a0 a, a0 s + s0 )t,
(2)
hence g 1 = (a1 , s/a). (The form of the composition law shows

that the subgroup of dilations acts on the subgroup of translations,
45
so that G is the semidirect product of the two.) The frame vectors

for the wavelet transform were obtained from a single basic wavelet
vector h by applying the transformation

ts
1/2
(U (a, s)h) (t).
(3)
h(t) 7 | a |
h
a
It it easy to see that with respect to the inner product in L2 (IR),

(a) (U (a, s) h) (t) = | a | 1/2 h(at + s) = U (a, s)1 h (t), and
(b) U (a0 , s0 )U (a, s)h = U (a0 a, a0 s + s0 )h.
In terms of the group operations, this means that U (g) = U (g)1
and U (g 0 g) = U (g 0 )U (g), respectively. That is, each U (g) is a unitary operator (on L2 (IR) ), and g 7 U (g) is a representation of G.
This means that U is a unitary representation of G on L2 (IR). The
theory of such representations for general groups is a deep and highly
developed subject, and is of fundamental importance in quantum mechanics, as was realized by Hermann Weyl and others long ago (Weyl
[1931]). In the broadest sense, group representations amount to a vast
generalization of the exponential function (think of the map z 7 eaz
from Cl to Cl ), and unitary group representations generalize the map
x 7 eiax from IR to the unit circle.
A representation U of a group G on a Hilbert space H is said to be
reducible if H has a nontrivial closed subspace S (i.e., S 6= {0} and
S 6= H) which is invariant under U (i.e., U (g)S S for every g G).
If U is unitary and S is invariant, then clearly so is its orthogonal
complement S . If no such S exists, then U is said to be irreducible.
It can be shown that the above representation of IR IR on L2 (IR)
is, in fact, irreducible.
The method to be described below assumes that we begin with
a given irreducible unitary representation U of a given group G on a
given Hilbert space H. In addition, we assume that the group is a
Lie group, meaning that it has a differetiable structure such that it
is essentially determined (in local terms) by its Lie algebra of left
(or right) invariant vector fields. (See Helgason [1978], Varadarajan
[1974] or Warner [1971] for background on Lie groups.) The affine
group and the WeylHeisenberg group are examples of Lie groups, as
is IRn (under vector addition as the group operation).
Given a general setup (U , G, H ) as above, choose an arbitrary
nonzero vector h in H. (Klauder dubbed h a fiducial vector; for
the affine group, this was the basic wavelet.) For every g G define
the vector
46
hg = U (g)h.
(4)
Since U is unitary, khg k = khk. The hg s are covariant under the

action of G on H, i.e.,
U (g 0 )hg = U (g 0 )U (g)h = U (g 0 g)h = hg0 g .
(5)
Hence the set of all finite linear combinations (span) of hg s is invariant under U , and therefore so is its closure S. Since U is irreducible
and S 6= {0}, it follows that S = H. This means that every vector in H can be approximated to arbitrary precision by finite linear
combinations of hg sa good beginning, if one is ultimately interested in reconstruction! To build a frame from the hg s, we need a
measure on G. Now every Lie group has an essentially unique (up
to a constant factor) leftinvariant measure, which we will denote by
d. This means that if E is an arbitrary (Borel) subset of G, and if
g1 E {g1 g | g E} is its left translate by g1 G, then
Z
Z
(g1 E)
d(g) = (E)
d(g).
(6)
g1 E
In local terms, d(g1 g) = d(g) for fixed g1 . [Similarly, there exists a

rightinvariant measure dR on G, which is in general different from
d; if dR is proportional by a constant to d, the group G is called
unimodular. Everything we do below can be repeated, with obvious
modifications, using dR , provided that hg is redefined as U (g 1 )h.]
To find d for the affine group, for example, we write it in the
form of a density,
d(a, s) = (a, s) da ds,
(7)
where da ds is a differential 2form denoting the area element in

IR2 G. (da ds becomes a positive measure upon choosing an
orientation in IR2 and orienting all subsets accordingly; see Warner
[1971].) Then, for fixed (a1 , s1 ) IR IR,
d((a1 , s1 )(a, s)) = (a1 a, a1 s + s1 ) d(a1 a) d(a1 s + s1 )
= a21 (a1 a, a1 s + s1 ) da ds,
hence leftinvariance implies
(8)
47
a21 (a1 a, a1 s + s1 ) = (a, s).
(9)
Setting a1 = 1/a and s1 = s/a then shows that

d(a, s) = da ds/a2 ,
(10)
where we have chosen the normalization (1, 0) = 1. This is precisely

the measure we obtained earlier using a more pedestrian approach.
Returning to the general case, consider the formal integral
Z
J=
d(g) | hg ih hg | .
(11)
G
We want to show that (a) J converges, in some sense, and (b) J is a

multiple of the identity, thus giving a tight frame in the generalized
sense defined in chapter 1. Before worrying about convergence, let us
formally apply U (g1 ) from the left:
Z
U (g1 )J =
d(g) U (g1 ) | hg ih hg |
G
Z
=
d(g) | hg1 g ih hg |
G
Z
(12)
=
d(g 0 ) | hg0 ih hg1 g0 |
1
ZG
=
d(g 0 ) | hg0 ih hg0 | U (g1 )
G
= JU (g1 ),
where we have used the leftinvariance of the measure and
h hg1 g0 | =
1

| hg1 g0 i = U (g11 ) | hg0 i = h hg0 | U (g1 ),
(13)
which follows from unitarity.

That is, J commutes with every representative U (g) of the group.
By Schurs Lemma (Varadarajan [1974]), it follows from the irreducibility of U that J is a multiple of the identity operator on H.
Since J, if it converges, is a positive operator, we arrive at the desired
frame condition
48

Z
d(g) | hg ih hg | = cI,
(14)
where c is a positive constant which can taken as unity by the appropriate normalization of d. Returning to the formal integral defining
J, the above argument shows that if the integral converges in some
sense, it must converge to cI, hence define a bounded operator. (Of
course c could be infinite!) Thus a necessary condition for convergence
in the weak sense (i.e., as a quadratic form) is that
Z
hh|J hi =
d(g) h h | hg ih hg | h i
G
Z
(15)
2
=
d(g) | h h | U (g) | h i | < .
G
We have already encountered this condition in the special case of the

affine group, where it was called the admissibility condition for h. The
same terminology is used in the present, general setting. The above
condition depends both on the representation U and the choice of h.
If it is satisfied for at least one nonzero vector h, the representation
U is called squareintegrable and h is called admissible. It turns out
that the existence of an admissible (nonzero) vector is also sufficient
for the weak convergence of the integral J. The following theorem
is due to Carey [1976], and Dufflo and Moore [1976]; G can be any
locally compact topological group, in particular any Lie group.
Theorem 3.1. Let U be a squareintegrable unitary irreducible
representation of G on a Hilbert space H. Then there exists a unique
selfadjoint (in general unbounded) operator C on H such that:
(a) The domain of C coincides with the set of all admissible vectors.
(b) If h1 and h2 are admissible, then for all f1 and f2 in H we have
Z
d(g) h f1 | U (g) h1 ih U (g) h2 | f2 i = h Ch2 | Ch1 ih f1 | f2 i. (16)
G
(c) If G is unimodular, then C is a multiple of the identity.

Choosing h1 = h2 h now leads to the earlier resolution of
unity, with c = kChk2 . As usual, an arbitrary vector f in H can
be presented as a function f(g) h hg | f i on G, from which f
may be reconstructed as a linear combination of hg s weighted by the
reproducing kernel K(g 0 , g) = h hg0 | hg i.
49
Note that the basic wavelet h corresponds to
h(g)
= h hg | h i = h h | U (g)h i,
(17)
so the admissibility condition is nothing but the requirement that h

2
belong to L (d).
A potentially interesting generalization of this scheme is actually
possible. By the foregoing theorem we can use two distinct admissible
vectors: the vector h2 is used to analyze f , i.e. present it as f(g) =
h (h2 )g | f i, while the vector h1 is then used to synthesize f from
f(g). There is no fundamental reason to use the same wavelets for
analysis and synthesis, provided only that they overlap in the sense
that
h Ch2 | Ch1 i =
6 0.
(18)
Absorbing the reciprocal of this constant into the group measure, the
map T : f 7 f is an isometry from H onto its range <T .
Due to the covariance of the hg s with respect to the action of
G, the representation of G on <T acquires the simple geometric form

(g1 )f (g) h hg | U (g1 )f i
U
= h hg1 g | f i
(19)
= f(g11 g),
showing that G acts on <T by merely translating the variable in the
as above is called
base space G. The realization of f by f and U by U
the coherentstate representation determined by the pair (U, h). We
will also refer to the frame { hg | g G} as the groupframe (G
frame) associated with (U, h). From a purely mathematical point of
view, one of the attractions of this scheme is that although we started
with an arbitrary representation of G on an arbitrary Hilbert space,
this construction brings it home to G itself and objects directly associated with it: the Hilbert space is a closed subspace of L2 (d), and
the representation is induced from the (left) action of G on itself, i.e.
g 7 g11 g. That is, Klauders construction exhibits U as a subrepresentation of the regular representation of G. (See Mackey [1968] for
the definition and discussion of the regular representation.)
50
3.3. Perelomovs Homogeneous GFrames

Let us now attempt to apply the grouptheoretic method to the windowed Fourier transform. As most of our applications will be to the
phasespace formulation of quantum mechanics, we shift gears and
replace the time by a space coordinate, t x (the sign is related to
the Minkowski metric; see section 1.1) and the frequency by a momentum coordinate, 2 p. Although it is a trivial matter to extend
everything we do in this section to an arbitrary (finite) number of degrees of freedom (where x and p belong to IRs ), we restrict ourselves
to a single degree of freedom to keep the notation simple. The self
adjoint generators of translations in space and momentum, P and X,
are defined by
f (x),
x
Expressed in the space domain,
P f (x) = i
(Xf )(p) = i

f (p).
p
(1)
Xf (x) = xf (x).
(2)
Starting with a basic window function h(x), the window centered at
(p, x) in phase space is given by
0
hp,x (x0 ) = eipx h(x0 x)

0
= eipx eixP h (x0 )

= eipX eixP h (x0 ).
(3)
That is, hp,x is obtained from h by a translation in space (by x)

followed by a translation in momentum (by p). Defining the corresponding unitary operators
U (p, x) = eipX eixP ,
(4)
let us see what happens when two such operations are applied in
succession:
(U (p1 , x1 )U (p, x)h) (x0 ) = (U (p1 , x1 )hp,x ) (x0 )
0
= eip1 x hp,x (x0 x1 )

0
= eip1 x eip(x x1 ) h(x0 x1 x)

= eipx1 hp1 +p,x1 +x (x0 )
= eipx1 U (p1 + p, x1 + x)h(x0 ).
(5)
51
Hence the operators U (p, x) do not form a group, since two successive
operations give rise to a multiplier exp(ipx1 ). The reason is that
translations in space do not commute with translations in momentum,
as can also be seen at the infinitesimal level by noting that their
respective generators obey the canonical commutation relations

[X, P ] = x, i
= iI.
(6)
x
The remedy (suggested by Hermann Weyl) is to include the identity
operator as a new generator (it generates phase factors ei which
can be used to absorb the multiplier). To see this in terms of unitary
representations of Lie groups, consider the abstract real Lie algebra
w with three generators {iS, iT, iE} and Lie brackets
[S, T ] = iE,
[S, E] = 0,
[T, E] = 0.
(7)
The corresponding threedimensional (simply connected) Lie group

W is known as the WeylHeisenberg group. Topologically, W is just
IR3 . (If configuration space is IRs , the corresponding WeylHeisenberg
group Ws IR2s+1 .) A general element in W, parametrized by
(p, x, ) IR3 , may be expressed as the product of three factors
g(p, x, ) = exp(iE) exp(ipS) exp(ixT ),
(8)
where exp denotes the exponential mapping from the Lie algebra
w to the Lie group W, whose group law is
g(p1 , x1 , 1 ) g(p, x, ) = g(p1 + p, x1 + x, 1 + + px1 ).
(9)
A unitary irreducible representation of W on L2 (IR) is obtained by

the correspondence S X, T P, E I. This is known as
the Schr
odinger representation. The unitary operator correponding
to g(x, p, ) is
U (p, x, ) = ei eipX eixP .
(10)
As expected, these operators are closed under multiplication, with the

composition law
U (p1 , x1 , 1 ) U (p, x, ) = U (p1 + p, x1 + x, 1 + + px1 ).
52
With this unitary irreducible representation of W on L2 (IR), we

have all the ingredients needed to attempt the construction of a group
frame for W. The prospective frame vectors are
hp,x, = U (p, x, )h = ei hp,x ,
(11)
where hp,x are the vectors defined earlier. The leftinvariant measure
on W (which, in this case, is also rightinvariant) is just Legesgue
measure on IR3 , given by the differential form dp dx d. This can
be seen by looking at the composition law: For fixed (p1 , x1 , 1 ),
d(p1 + p) d(x1 + x) d(1 + + px1 ) = dp dx d
(12)
since dp dp = 0. The method of section 2 then gives the following

candidate for a resolution of unity:
Z
J=
dp dx d | hp,x, ih hp,x, |
W
Z
(13)
=
dp dx d | hp,x ih hp,x | ,
W
where the phase factor cancels in the integrand. This integral clearly
diverges, since the integrand is independent of and the integration
is over all real . Equivalently, the representation U is not square
integrable, since for every nonzero h,
Z
ch dp dx d | h h | U (p, x, )h i | 2 =
(14)
due, again, to the constancy of the integrand in . We could get
around the problem by choosing a multiply connected version W1 of
W, say with 0 < 2. (W1 has the same Lie algebra as W).
This would indeed give a tight frame, but this frame is unnecessarily
redundant (as opposed to the beneficial sort of redundance associated with oversampling) since the vectors hp,x, are not essentially
different for distinct values of . More significantly, we would miss an
important lesson which this example promises to teach us. For other
important groups, as we will see, the problem cannot be circumvented
by compactifying the troublesome parameters. The following solution
was proposed by Perelomov [1972] (see also Klauder [1963b, p. 1068],
where this idea is anticipated). To get rid of the dependence, choose
a slice of W, = (p, x), and define

i(p,x)
h
hp,x .
p,x = hp,x,(p,x) = e
53
(15)
Integrating only over this slice with the measure dp dx, we get
Z
0
J dp dx | h
(16)
p,x ih hp,x | = ch I,
where ch = 2khk2 , since the integral reduces to the same one we had
for the windowed Fourier transform.
From a computational point of view this is, of course, trivial. But
to extend the technique to other groups we must understand the group
theory behind it. Suppose, then, that we return to the general setup
we had in the last section: Given a unitary irreducible representation
U of a Lie group G on a Hilbert space H, choose a nonzero vector h
in H and form its translates hg = U (g)h under the group action as
before. Consider now the set H of of all elements k of G for which
the action of U (k) on h reduces to a multiplication by a phase factor
(k) = exp[i(k)]:
U (k)h = (k)h,
k H.
(17)
In the case of W, H consists of all elements of the form k = (0, 0, )

and U (k) is, in fact, nothing but a phase factor. However, this is
deceptive, since in general U (k) may be a nontrivial operator, acting
trivially only on some vectors h. Hence, in general, H will in fact
depend on the choice of h and should properly be designated H(h).
Now for two elements k1 and k2 of H, we have
U (k1 k21 )h = U (k1 )U (k2 )1 h = (k1 )(k2 )1 h,
(18)
hence k1 k21 also belongs to H and it follows that H is a subgroup of

G. Furthermore, the above equation shows that the map k 7 (k) is
a grouphomomorphism of H into the unit circle, thus it is a character of H, i.e. a unitary representation on the onedimensional Hilbert
space Cl. H is called the stability subgroup for h, and its Lie algebra
is called the stability subalgebra. The reason for this terminology is
that H does not affect the quantummechanical state defined by h,
since all observable expectation values in that state are given in terms
of sequilinear forms in h. More precisely, if we choose khk = 1, the
corresponding state is by definition the rankone projection operator
P = | h ih h |
(19)
54
and the expected value of any observable A in this state is given by

h A i h h | A h i = trace (P A).
(20)
(This formulation, besides avoiding irrelevant phase facors, also permits states which are statistical mixtures of pure states as needed,
for example, in statistical quantum mechanics.) The translate of P
under a general group element g is then
Pg | hg ih hg | = U (g) P U (g) ,
(21)
hence for g in H we have Pg = P , i.e., P is stable under H as the

name implies. There is, therefore, some advantage to formulating
the theory as much as possible in terms of states rather than Hilbert
space vectors since this automatically eliminates the fictitious degree
of freedom represented by the overall phase. [Incidentally, a similar
situation appears to prevail in communication theory, since an overall
phase shift has no effect whatsoever on the informational content of
the signal.] However, the Hilbert space vectors will play an important
role in connection with holomorphy (recall that in the coherentstate
representation, f(z) is analytic, whereas | f(z) | 2 is not), hence we
work primarily with them.
If g G and k H, then
hgk = U (gk)h = U (g)U (k)h = (k)hg ,
(22)
hence Pgk = Pg . That is, Pg is the same for all members of the left
coset
gH {gk | k H}.
(23)
The set of all translates Pg is therefore parametrized by the left coset

space
M = G/H {gH | g G}.
(24)
Members of M will be alternatively denoted by m and by gH, and

will play a dual role: as points in M , and subsets of G. In the case
of W, for example, M is parametrized by (p, x) IR2 , i.e. it is phase
space, precisely the label space for the frame we obtained. The coset
(p, x)H is a straight line in W IR3 . (We will see in the next chapter
that W can be interpreted as a degenerate nonrelativistic limit of
phase spacetime and (p, x)H then corresponds to the trajectory of
a free classical particle in W.)
55
To build a frame, we take a slice of G by choosing a representative from each coset, i.e. choosing a map
: G/H G.
(25)
Note: This is actually a nontrivial process. The projection G

G/H defines a fiber bundle, and is a section of this bundle; in
general can be chosen smoothly only in a neighborhood of each
point of G/H, i.e., locally (see F. Warner [1971], p.120). However,
this is sufficient for our purposes, since we will ultimately deal only
with the state Pm , which is independent of the choice of . #
Thus has the form (gH) = g (g) for some function : G H.
In the case of W, we had (p, x) = (p, x, (p, x)). The frame vectors
corresponding to this choice are
hm h(m) = U ((m))h.
(26)
To build a frame, we need two more ingredients: an action of G on the

label space M , and a measure on M which is invariant with respect
to this action. The action is easy, since G acts naturally on G/H by
left translation:
g1 (gH) = (g1 g)H.
(27)
If m = gH M , we will denote (g1 g)H by g1 m. M is called a

homogeneous space of G. As for an invariant measure, it exists,
in general, only subject to a certain technical condition (Helgasson
[1962], p. 369). It does exist whenever G is a unimodular group (its
right and leftinvariant measures are proportional by a constant),
such as W. Let us assume that a Ginvariant measure does exist on
M (in which case it is unique, up to a constant factor) and denote it
by dM . Once more we consider the formal integral
Z
J=
M
dM (m) | hm
ih hm
Z
| =
dM (m) Pm .
(28)
If we can show that J commutes with every U (g1 ), then irreducibility

again forces J = cI. But
56

U (g1 ) hgH = U (g1 ) U ((gH))h
= U (g1 (gH)) h
= U (g1 g (g)) h
(29)
= U (g1 g (g1 g)(g1 g)1 (g)) h

= ((g1 g)1 (g)) hg1 gH .
That is, the vectors hm are almost covariant under the action of G,
with a residual phase factor (multiplier). This means that the states
Pm transform covariantly, i.e.,
U (g1 ) Pm U (g1 ) = Pg1 m .
(30)
Thus
1
U (g1 ) J U (g1 )
Z
=
dM (m) U (g1 ) Pm U (g1 )
ZM
=
dM (m) Pg1 m
M
Z
=
(31)
dM (m0 ) Pm0
= J,
using the unitarity of U (g1 ) and the invariance of dM . This proves
that J = cI, provided the integral converges in some sense.
Note: The above proof becomes shorter if one works directly with
the states Pm rather than the frame vectors; however, it is important
to see the action of G on the frame vectors in terms of the multipliers since this will play a role in the next section, where we look for
representations of G on spaces of holomorphic functions. #
A necessary condition for the weak convergence of the integral is
that
Z
hh|J hi =
dM (m) | h hm | h i | 2 < ,
(32)
M
i.e., that h(m)

h hm | h i be in L2 (dM ). As before, this condition
is also sufficient, and the same terminology is used: h is called admissible, and U is called squareintegrable. The action of U in the
3.4. Onofris Holomorphic GFrames
57
Hilbert space of functions f(m) = h hm | f i is somewhat more complicated than earlier because of the multiplier. (If we simply dropped
the multipliers we would no longer have a unitary representation of
G but a projective representation; see Varadarajan [1970].)
The above construction has the advantage that the trivial part of
the action of G is factored out, thereby improving the chances that the
integral J converges. As mentioned above, it depends on the existence
of the invariant measure dM on G/H, which is not guaranteed. Note
that for fixed g G, the stability subgroup for the vector U (g)h is
gHg 1 , i.e., a subgroup of G conjugate to H. In general, however,
the stability subgroups of two admissible vectors h1 and h2 may not
be conjugates. It may happen that G/H1 has an invariant measure
while G/H2 does not. It pays, therefore, to choose the vector h very
carefully. Intuition suggests that h should be chosen so as to maximize
the stability subgroup, since this will minimize the homogeneous space
M and improve the chances for convergence. Furthermore, maximal
use of symmetry would seem to make it more likely that an invariant
measure exists on the quotient. More will be said about this in the
next section, in connection with the weight of a representation.
We will refer to the frame { hm } as a homogeneous Gframe associated with (U, h). The dependence on will usually be suppressed,
since a change in (the local sections) gives an equivalent frame.

The frames associated with the windowed Fourier transform and the
canonical coherent states are similar in that both provide phase
space representations of functions in L2 (IRs ). The distinguishing feature is that the representation determined by the canonical coherent
states is in terms of holomorphic (analytic) functions. Neither of the
two general grouptheoretic methods covered so far explains the origin
of this analyticity, yet it has been found that all such grouprelated
phasespace representations have analytic counterparts. Moreover,
analyticity will play a key role in the coherentstate representations
of relativistic quantum mechanics (next chapter) and quantum field
theory (chapter 5). The method to be described in this section assumes we are dealing with a compact, semisimple group. Since the
coherentstate representations we will develop for relativistic quantum mechanics are based on the Poincare group, which is neither
compact nor semisimple, the present considerations do not apply di-
58
rectly to the main body of this book (chapters 4 and 5). Nevertheless, we describe them in considerable detail in the hope that they
may shed some light on our later constructions, which are still not
wellunderstood in general terms.
Let us return to the canonical coherent states in order to isolate
the property leading to analyticicy and find its generalization to other
groups. We follow an approach first advocated by Onofri [1975]. For
related developments, also see Perelomov [1986]. Again consider the
threedimensional WeylHeisenberg group W, whose (real) Lie algebra w has a basis {iS, iT, iE} which is represented on L2 (IR) by
S X, T P and E I. Recall that we arrived at the canonical coherent states z as eigenvectors of the nonhermitian operator
A = X + iP , which represents the complex combination S + iT of
generators in w. Since w is a real Lie algebra, we must complexify
it in order to consider such combinations of its generators. This is
done in the same way as complexifying a real vector space, namely by
taking the tensor product with the field of complex numbers. With
obvious notation, wc = w Cl = w + iw. As will be seen below,
complex combinations such as S iT play a very important role in
the theory of real Lie groups, exactly for the same reason that complex eigenvectors and eigenvalues are necessary in order to study real
matrices.
We begin by rederiving the canonical coherent states from an
algebraic point of view which shows their relation to the vectors hp,x
associated with the windowed Fourier transform and pinpoints the
property which makes them analytic. In the last section we found
that
hp,x = eipX eixP h.
(1)
Now if B and C are operators such that [B, C] commutes with both
B and C, then the BakerCampbellHausdorff formula (Varadarajan
[1974]) reduces to
1
eB eC = e 2 [B,C] eB+C .
(2)
Since [ipX, ixP ] = ipxI, we therefore have

hp,x = exp(ipx/2) exp (ipX ixP ) h.
(3)
Substituting
X = (A + A)/2
and
iP = (A A)/2
(4)
59
and defining z x ip, we get

hp,x = exp(ipx/2) exp [
z A /2 zA/2] h.
(5)
So far, all our manipulations have been justifiable since we have only
exponentiated skewadoint operators. The next one is more delicate
since the operators to be exponentiated are not skewadjoint; it will
be justified later. Using the BakerCampbellHausdorff formula again
with B = zA /2 and C = zA/2, write
hp,x = exp (ipx/2 zz/4) exp (
z A /2) exp (zA/2) h.
(6)
If we choose

h(x0 ) = N exp x0 2 /2 = 0 (x0 )
(7)
(where N = (2)1/4 and 0 is just the canonical coherent state z

with z = 0), then Ah = 0, hence exp (zA/2) h = h and
hp,x = exp (ipx/2 zz/4) exp (
z A /2) 0 .
(8)
We claim that
exp (
z A /2) 0 = z .
(9)
This can be seen by applying A to the lefthand side, then using

A0 = 0 and [A, A ] = 2I:
A exp (
z A /2) 0 = [A, exp (
z A /2)] 0
= z exp (
z A /2) 0 ,
(10)
which shows that exp (

z A /2) 0 equals z up to a constant factor,
which is easily shown to be unity. That is,
hp,x = exp (ipx zz/4) z ,
(11)
so the canonical coherent states are a special case (modulo the z

dependent factor in front) of the frame vectors hp,x for the choice
h = 0 . Note that the corresponding states are related by

| hp,x ih hp,x | = exp |z|2 /2 | z ih z | ,
(12)
60
giving the weight function on the righthand side in the bargain. The
reason for this is that the hp,x s were obtained by unitarily translating
h, whereas the operator exp (
z A /2) which translates 0 to z is not
unitary but results in kz k = exp |z|2 /4 k0 k; the weight function
then corrects for this.
We can now justify the fine point we glossed over earlier. The
expression
exp (
z A /2) exp (zA/2) 0
(13)
makes sense because

(a)
ezA/2 0 = 0
(14)
since A0 = 0, and
(b) 0 is an analytic vector (Nelson [1959]) for the operator A , since
X
X
|z|n
1
n
k (
z A /2) 0 k =
kA n 0 k
n
n!
2 n!
n=0
n=0
X
|z|n n/2
2
n! k0 k <
=
2n n!
n=0
(15)
z Cl ,
where kA n 0 k = 2n/2 n! follows from [A, A ] = 2I and A0 = 0.

As will be seen below, the key to constructing representations of
real Lie groups on spaces of analytic functions will be to
(a) complexify the Lie algebra;
(b) find a counterpart to A0 = 0;
(c) define counterparts to z = exp(
z A /2) 0 .
Before embarking on this task, we must make a brief excursion
into the structure theory of Lie algebras. For background and details,
see Helgason [1978] or Hermann [1966]. The Lie groups and algebras
we consider below are assumed to be real and semisimple. (A Lie
algebra is semisimple if it contains no proper abelian ideals; a Lie
group is semisimple if its Lie algebra is semisimple.) Since W is not
semisimple (the subspace spanned by iE is an ideal of w), these
results do not apply to it, strictly speaking. However, as will be
shown in section 3.6, W can be obtained as a (contraction) limit of a
simple Lie group (SU (2)), and this turns out to be sufficient for our
purpose. Thus, let G be an arbitrary real, semisimple Lie group and
g its Lie algebra. The following is known:
61
1. Every element X of g defines a linear map ad X on g by

ad X(Y ) = [X, Y ]. Similarly, every Z gc defines a complex
linear map (denoted by ad Z) on gc . It therefore makes sense to
look for eigenvalues and eigenvectors of ad Z.
2. g has a (Lie) subalgebra h called a Cartan subalgebra, defined
as a maximal abelian subalgebra of g satisfying an additional
technical condition. h is analogous to a maximal commuting
set of observables in quantum mechanics. Its complexification,
denoted by hc , is a maximal abelian subalgebra of gc . (The technical condition on h essentially ensures that each of the linear
maps ad H, with H hc , has a complete set of eigenvectors in
gc , hence that all the linear transformations ad H with H hc
can be diagonalized simultaneously.)
3. For every complexlinear form : hc Cl , let
g = {Z gc | [H, Z] = (H)Z
H hc }.
(16)
That is, g is the set of all vectors in gc which are common

eigenvectors of ad H for every H in hc , with corresponding eigenvalue (H). For generic , no such eigenvectors will exist, hence
g = {0}. When g 6= {0}, is called a root (of gc with respect
to hc ), each Z in gc is called a root vector and g is called a root
subspace of gc . Clearly, the zeroform (H) 0 is a root, since
the fact that hc is abelian means that g0 contains hc . The fact
that hc is maximalabelian means that actually g0 = hc , since
every Z in g0 commutes with all H in hc .
4. Let Z g and Z g , where and are arbitrary linear
forms on hc . Then the Jacobi identity implies that for every H
in hc ,
[H, [Z , Z ]] = [[H, Z ], Z ] + [Z , [H, Z ]]
(17)
= ((H) + (H)) [Z , Z ],
thus [Z , Z ] g+ . This statement is abbreviated as
, hc .

g , g g+
(18)
5. The set of all nonzero roots is denoted by . gc has a directsum

decomposition (root space decomposition) as
gc = hc +
g ,
(19)
62
and each g (with ) is a onedimensional subspace of gc .

6. The Killing form of gc is the bilinear symmetric form defined by
B(Z, Z 0 ) = trace (ad Z ad Z 0 ) .
(20)
It is nondegenerate on gc , and its restriction to hc is also non

degenerate. Since a nondegenerate bilinear form on a vector
space defines an isomorphism between that space and its dual,
the restriction of B to hc defines an isomorphism between hc and
hc . The vector in hc corresponding to hc is denoted by H .
It is defined by
B(H , H) = (H)
H hc .
(21)
Note that the vector corresponding to 0 is H0 = 0.

7. If , then and (H ) B(H , H ) 6= 0. Furthermore,

g ,g
= Cl H ,
(22)
i.e. the set of all brackets [Z , Z ] with Z g fills out the
onedimensional subspace spanned by H . (H 6= 0 since .)
8. Any two root subspaces g and g with + 6= 0 are orthogonal
with respect to B.
9. It is possible to choose (nonuniquely) a subset + of , called
a set of positive roots, such that
(a) If and belong to + and + is a root, then +
belongs to + .
(b) The set + is disjoint from + .
(c) = + .
It follows from (4) and (9) that
X
X
n+
g
and
n
g
(23)
+
are Lie subalgebras of gc , and by (5),

gc = hc + n+ + n .
(24)
Since there can only be a finite number of roots, (4) and (9) also imply
that after taking a finite number of brackets of elements in either n+
or n we obtain zero; that is, the subalgebras n are nilpotent. We
may choose a basis for gc as follows: Pick a nonzero vector Z from
63
each g with + . Then by (7) there is a unique vector Z in

g with
[Z , Z ] = 2H /(H ).
(25)
Since Z and Z are root vectors, they also satisfy

[H , Z ] = (H )Z
[H , Z ] = (H )Z ,
(26)
and (H ) 6= 0 (property (7)). If we define

C = H /(H ),
(27)
then each triplet (Z , C , Z ) spans a Lie subalgebra of gc which is

isomorphic to sl(2, Cl ):
[C , Z ] = Z ,
[Z , Z ] = 2C .
(28)
The root space decomposition then shows that gc is a direct sum of

copies of sl(2, Cl ), indexed by + .
The connection with the nonhermitian combinations X iP can
now be explained. The following argument is heuristic and has no
pretense to rigor. Its sole purpose is to motivate the construction
of general holomorphic frames. It will be made precise in section
3.6 at the level of unitary representations, where a clear geometric
interpretation will be given.
Begin with the group G = SU (2), which is a simple, real Lie group
whose Lie algebra g = su(2) has a basis {iJ1 , iJ2 , iJ3 } satisfying
[J1 , J2 ] = iJ3
[J2 , J3 ] = iJ1
(29)
[J3 , J1 ] = iJ2 .
We first show that w can be obtained as a contraction limit of g.*
Choose a positive number and define K1 = J1 , K2 = J2 and
K3 = 2 J3 . These form a new basis for g, with
* The idea of group contractions is due to In
on
u and Wigner [1953].
64
[K1 , K2 ] = iK3
[K2 , K3 ] = i2 K1
(30)
[K3 , K1 ] = i2 K2 .
In the limit 0, g becomes isomorphic to w. We will see in
section 3.6 that within a unitary irreducible representation of SU (2),
the operator K3 can be chosen so that K3 I as 0, hence we
may interpret the limits of K1 and K2 as X and P , respectively.
Now apply the above structure theory to gc , which is just
sl(2, Cl ). The direct sum of copies of sl(2, Cl ) therefore reduces to a
single term. For the Cartan subalgebra of g we can choose h = IR J3
(the onedimensional subspace spanned by J3 ), so that hc = Cl J3 .
Two linearly independent root vectors are given by J = J1 iJ2 ,
with
[J3 , J+ ] = J+ ,
[J3 , J ] = J .
(31)
We choose (J3 ) = 1 as the single positive root and J for the basis
vectors of the two onedimensional root subspaces. Then
[J+ , J ] = 2J3
(32)
shows that C corresponds to J3 . Setting K = J and K3 = 2 J3 ,

and taking the limit 0, we obtain the correspondence
K+ S + iT 7 X + iP = A
K S iT 7 X iP = A
(33)
K3 E 7 I,
where we have used 7 to denote the representation of w on L2 (IR).
[This correspondence is not unique; for example, K+ A , K A,
K3 E is equally good. As will be shown in section 3.6, both of
these correspondences actually occur as weak limits, due to the fact
that an irreducible representation of SU (2) contracts to a reducible
representation of w.] Note that the three roots {2 , 0, 2 } all merge
into a single root, zero, in the contraction limit. Thus the operators
A and A are interpreted as the contraction limits of root vectors.
Note: Another way of seeing the importance and naturality of A
and A is in the context of the KirillovKostantSouriau theory,
sometimes called Geometric Quantization, applied to the Oscillator
65
group (Streater [1967]; see section 3.6). For background on Geometric

Quantization, see Kirillov [1976], Kostant [1970] and Souriau [1970];
see also Guillemin and Sternberg [1984], Simms and Woodhouse [1976]
and Sniatycki
[1980].
We are, at last, ready to generalize the construction of group
representations on spaces of holomorhic functions to other groups.
We begin with a semisimple, real Lie group G and a unitary irreducible representation U of G on a Hilbert space H. To avoid technical difficulties, we will here assume that G is compact. (The reason for assuming compactness, as well as ways to get around it, will
be discussed below.) Then the irreducibility of U implies that H be
finitedimensional, hence the operators U (g) representing group operations are just unitary matrices and the operators U (X) representing
elements of g are skewadoint matrices. (Sometimes the operator
representing X is written as dU (X) to emphasize its infinitesimal
nature; we will write it as U (X) to keep the notation simple.) At the
Lie algebra level, U extends, by complexlinearity, to a representation
T of gc :
T (X + iY ) U (X) + iU (Y ).
(34)
T is the unique representation of gc that extends U ; that is, for all

c1 , c2 Cl and all Z1 , Z2 gc ,
T (c1 Z1 + c2 Z2 ) = c1 T (Z1 ) + c2 T (Z2 )
T ([Z1 , Z2 ]) = [T (Z1 ), T (Z2 )]
T (X) = U (X)
if
(35)
X g gc .
The matrices T (Z) are no longer skewadjoint, but because of the

finite dimensionality of H, there is no problem in exponentiating them
to give a representation of Gc on H, which we also denote by T . Thus,
the representative of the group element exp Z of Gc is defined by*
T (exp Z) = exp[T (Z)],
Z gc .
(36)
Note: If G is noncompact, exp Z is still welldefined but the right

hand side of the above equation is problematic since for any non
trivial representation U , H is infinitedimensional and T (Z) will, in
* Since G is compact, it is exponential; that is, every group element
can be written in the form exp Z for some Z.
66
general, be a nonskewadjoint, unbounded operator. (Even the definition T (Z) = T (X) + iT (Y ) becomes troublesome, since the skew
adjoint operators T (X) and T (Y ) may both be unbounded and their
domains may have little in common; additional assumptions must be
made.) In the case of W, this was resolved by restricting exp[T (Z)] to
act on analytic vectors. A similar approach is used in extending the
present construction to noncompact G. For the present, we continue
to assume that G is compact to avoid this problem. #
Lemma 3.2..
(a) g 7 T (g) is an irreducible (nonunitary) representation of Gc .
(b) The map Z 7 T (exp Z) is analytic as a map from the complex
vector space gc to the complex matrices on H.
Proof. If A is a matrix which commutes with all T (Z) for Z gc ,
then in particular it commutes with all U (X) for X g, hence must
be a multiple of the identity since U is irreducible. Therefore T is
irreducible. To prove (b), note that Z 7 T (exp Z) is, by definition,
the composite of the two analytic maps Z 7 T (Z) and T (Z) 7
exp[T (Z)].
It follows that the map T from Gc (considered as a complex manifold; see Wells [1980]) to the group GL(H) of nonsingular matrices
on H is also analytic; that is, T is a holomorphic representation of
Gc , obtained by analytically continuing the representation U of G.
Now the point of Onofris construction is this: We have seen that by
choosing a state which is stable under H, U can be reformulated as
a representation of G on a space of functions f(m) defined over the
homogeneous space G/H. In the case G = W, G/H was identified as
a phase space, but in general it is not clear that it can be interpreted
as such. Following Onofri, we will show that:
(a) The representation T induces a complex structure on the homogeneous space G/H, making it into a complex manifold on which G
acts by holomorphic transformations. (Such a manifold is called a
complex homogeneous space of G, or a holomorphic homogeneous
Gspace.)
(b) In addition, G/H has the (symplectic) structure of a classical
phase space, and the action of G on G/H is by canonical transformations. Thus it becomes possible to think of G/H as phase
space. To actually identify G/H as the phase space of a classical
physical system, i.e. as the set of dynamical trajectories followed
by that system, it is necessary for G to include the dynamics for
the system, i.e. its evolution group, of which nothing has been
said so far. This will be discussed in the next chapter.
67
The representatives U (H) of the elements H of h form a commuting set of skewadjoint matrices, hence can all be diagonalized
simultaneously. Let h be a common eigenvector:
U (H) h = (H) h,
(37)
where (H) is imaginary. Since U is linear at the Lie algebra level,

it follows that is a linear functional on h, called the weight of h.
(Roots are simply weights in the adjoint representation, where H is
replaced by gc and U (H) by adH.) For any nonzero element Z in
g with in + , we have (remembering that U (H) = T (H) since
H h)
T (H)T (Z ) h = [T (H), T (Z )] h + T (Z )T (H) h
= T ([H, Z ]) h + (H) T (Z ) h
(38)
= ((H) + (H)) T (Z ) h.
That is, T (Z ) raises the weight by . Similarly, for ,
T (Z ) lowers the weight by . Since nonzero vectors with different
weights are linearly independent and H is finitedimensional, it follows
that H must contain a nonzero vector with lowest weight, i.e. such
that
T (Z ) h = 0
(39)
T (Z) h = 0
Z n .
(40)
Equivalently,
For the group W, h was the ground state 0 and the above equations correspond to T (iE) 0 = i0 (so (iE) = i) and A0 =
0.
Consider the subalgebra
b = hc + n+
(41)
of gc , called a Borel subalgebra. If N n+ and we denote by N

its complexconjugate with respect to the real subalgebra g, then
n . Hence for arbitrary Z = H + N b, we have
N
= T (H)h
+ T (N
)h = T (H)h,
T (Z)h
(42)
belongs to hc since [H,
hc ] = 0. Extending by complex
and H
linearity to hc , we therefore have
68

= (H)
h,
T (Z)h
(43)

h = exp (H)
h.
exp T (Z)
(44)
hence
The subgroup B = exp(b) is called a Borel subgroup of gc . For Z as
above, let b = exp Z B, b = exp Z and (b) = exp[(H)].

Then
T (b) h = (b) h.
(45)
Since (H) is imaginary for H h, complexlinearity implies that

= (H) for H hc . Hence
(H)
(b) = (b) 1 ,
b B.
(46)
The map : B Cl satisfies (b1 b2 ) = (b1 ) (b2 ), i.e. it is

a character of B. Furthermore, since (b) is analytic in the group
parameters of b, Onofri calls it a holomorphic character of B.
Notice that the state corresponding to h (i.e., the onedimensional subspace spanned by it) is invariant under B. If we restrict
ourselves to the real group G, this means that the state is invariant
under the subgroup H. If H is the maximal subgroup of G leaving
this state invariant, then the weight is called nonsingular. In that
case, H plays the same role as it did in the last section: it is the
stability subgroup of the state. We will assume this to be the case; if
it is not (in which case is singular), the present considerations still
apply but in modified form. Note that in the nonsingular case, the
stability subgroup is abelian.
We now adapt the construction of the last section in a way which
respects the complexanalytic structure of Gc . Introduce the notation
(T (g) )1 T # (g),
g Gc .
(47)
Since the representation U of G is unitary, U (X) = U (X) for

It follows that
X g ; hence for Z gc , we have T (Z) = T (Z).
for group elements g = exp(Z) of Gc ,
T (g)# = T (
g ),
g Gc ,
(48)
and, in particular, T (g)# = T (g) = U (g) for g G.

where g = exp(Z)
Define the vectors

hg = T (g)# h,
g Gc ,
69
(49)
which, when restricted to g G, coincide with the earlier frame vectors but are antiholomorphic in the group parameters of Gc . An
arbitrary vector f H defines a holomorphic function on Gc by
f(g) h hg | f i.
(50)
hgb = T (g)# T (b)# h = (b) 1 hg ,
(51)
f(gb) = (b)1 f(g).
(52)
For arbitrary b B,
hence
Note:
The reader familiar with fiber bundles (Kobayashi and Nomizu [1963, 1969]) will recognize the above equation as the condition
defining a holomorphic section of the holomorphic line bundle associated to the principal bundle B Gc Gc /B by the character
: B Cl . We now proceed to construct this section in a naive way,
that is, without assuming any knowledge of bundle theory. #
The above shows that the state determined by hg depends only
on the left coset gB, which we denote by z. Let
Z = Gc /B
(53)
be the left coset space. Now

(a) Gc is a complex manifold;
(b) B, as a complex subgroup, is a complex submanifold of Gc ;
(c) the projection map Gc Z is holomorphic.
Hence it follows that Z is a complex manifold. This means that a
neighborhood of each point z = gB Z can be parametrized by a
local chart, i.e. a set of local complex coordinates (z1 , . . . , zn ) (say,
with (0, . . . , 0) corresponding to z), and the transformation from one
local chart to another on overlapping neighborhoods is a local holomorphic function. In the case of G = W, B corresponds (under the
contraction limit) to the complex subalgebra spanned by E and A,
and Z can be identified with Cl, hence only a single chart is needed
to cover all of Z. In general, more than one chart is necessary. For
G = SU (2), we will see that Z is the Riemann sphere S 2 , hence two
charts are needed; however, the north pole has measure zero, and a
single chart will do for S 2 \{} Cl.
70
Since Gc acts on itself by holomorphic transformations, its action

on Z is also by holomorphic transformations. This means the following: For g1 Gc and z = gB Z, let w(z) = g1 z (g1 g)B. If
(z1 , . . . , zn ) and (w1 , . . . , wn ) are local charts in neighborhoods of z
and w, respectively, then the mapping : (z1 , . . . , zn ) 7 (w1 , . . . , wn )
is holomorphic in a neighborhood of (0, . . . , 0). (It must, of course,
be locally invertible with holomorphic inverse.)
We could proceed as in section 3 and consider the states
Pz =
| hg ih hg |
,
h hg | hg i
z gB Z,
(54)
where we must now divide by khg k2 since the nonunitary operator

T (g) does not preserve norms. However, this would spoil the holomorphy which we are attempting to study. Instead, proceed as follows:
Choose an arbitrary reference point a in Gc and define the holomorphic coherent states
az =
hg
.
h ha | hg i
(55)
As indicated by the notation, the righthand side depends only on

the coset z = gB. There is no guarantee that the denominator on the
righthand side is nonzero, but certainly the open set
U = {g Gc | h ha | hg i =
6 0}
(56)
(which, as indicated, depends only on the coset = aB) contains a,

and its projection V to Z is an open set containing such that az
is defined for all z in V . Hence by choosing more than one reference
point a, if necessary, we can cover Z with patches V on which the
az s are defined.
An arbitrary vector f H can now be expressed as a local holomorphic function
fa (z) = h az | f i
(57)
of z in V , with transition functions

h hg | ha i a
fb (z) =
f (z)
h hg | hb i
ab (z)fa (z),
z V +1 V .
(58)
71
The reader may wonder where this is all leading, since we are
ultimately interested in the real group G and not in Gc . Here is
the point: It is known (Bott [1957]) that the complex homogeneous
space Z = Gc /B actually coincides with the real homogeneous space
M = G/H used in Perelomovs construction! For example, consider
the WeylHeisenberg group: G/H is parametrized by (x, p) while
Gc /B is parametrized by x ip, and they are the same set but with
the difference that the latter has gained a complex structure. The
identification of M with Z in general can be obtained by noting that
as a subgroup of Gc , G acts on Z by holomorphic transformations;
this action turns out to be transitive, and the isotropy subgroup at
the origin z0 = B is H, hence Z G/H. In other words, M
inherits a complex structure from Gc , and the natural action of G on
M preserves this structure. Because of this, we need not deal directly
with Gc to reap the benefits of the complex structure. Let us therefore
restrict ourselves to G. Then
hg = h ha | hg i az = U (g) h,
(59)
hence k hg k = k hk 1 and the state corresponding to hg is

2
| hg i h hg | = | h ha | hg i | | az i h az |
e(z,) | az i h az | ,
(60)
where (z, ) 2 ln |h ha | hg i| depends only on z = gH and = aH

and their complex conjugates (it is not analytic). Notice that in eq.
(60), the lefthand side, hence also the righthand side, is independent
of a. Only the three individual factors on the righthand side depend
on a. This becomes important if several patches are needed to cover
Z, since it means that we can change the reference point without
affecting the smoothness of the frame. With the above definitions,
the resolution of unity derived in section 3 becomes
Z
dM (z) e(z,) | az ih az | = I,
(61)
M
where we have assumed for simplicity that a single chart suffices.

If more than one chart is needed, partition M as a (disjoint) union
n Mn , where each Mn is covered by a single chart. Since, by the above
remark, the integrand is independent of a, the corresponding
integrals
P
In form a partition of unity in the sense that n In = I. Therefore,
the holomorphic coherent states form a tight frame which we call the
72
holomorphic Gframe associated with the representation U and the

lowestweight vector . In this connection, note that a choice of lowest
weight actually determines the representation U up to equivalence,
hence a more economical terminology would be to call the above the
holomorphic Gframe associated with . The corresponding inner
product is given by
Z
h f1 | f2 i =
dM (z) e(z,) f1a (z) f2a (z).
(62)
M
In the case G = W, since Z = Cl, everything can be done globally:

Just one chart is needed, and the reference point a can be fixed once
for all. Taking a = 1 (the identity element of G) and = 0, we find
that the z s reduce to the canonical coherent states and
2
e(z,0) = | h z | 0 i | = ezz/2 .
(63)
Hence in the general case, e takes the place of the Gaussian weight
function: it corrects for the fact that holomorphic translations do not
preserve the norm. The action of G on the space of local holomorphic
functions fa (z) has a multiplier:
h hg | U (g1 ) f i
h hg | ha i
h hg1 g | ha i h hg1 g | f i
1
1
=
h hg | ha i
h hg1 g | ha i
h az | U (g1 )f i =
(64)
(z, g1 ,
)fa (g1 1 z),
where z = gH, = aH and
(z, g1 ,
) =
=
h hg | U (g1 ) ha i
h hg | ha i
h hg1 g | ha i
1
h hg | ha i
h hg | hg 1 a i
=
h hg | h a i
(65)
is holomorphic in z and antiholomorphic in .

We have mentioned that M inherits a complex structure from
Gc . Actually, this is only part of the story. Let and denote the
3.5. The Rotation Group
73
external derivatives with respect to z and z, respectively, i.e., in local

coordinates,
f =
f
dzk ,
zk
= f d
f
zk
zk
(66)
(summation over k is implied). Consider the 2form

2
= i dz j d
= i
zk ,
zj zk
(67)
where is the function in the exponent of the weight function above.

Theorem.
= 0,
(a) is closed, i.e. =
(b) is independent of the reference point a,
(c) is invariant under the action of G, and
(d) is nondegenerate, if is nonsingular.
Proof. We prove (a), (b) and (c). For the proof of (d), see Onofri
[1975].
(a) follows from the fact that + = d is the total exterior derivative,
hence the identity d2 = 0 implies
2 = 0,
2 = 0,
+ = 0.
(68)
(c) follows from the fact that (z, g1 ,

) is holomorphic in z, hence
2 = 0. But eq. (65) shows that
||2 is harmonic and ||
|(z, g1 , )|2 =
exp[(g11 z, )]
,
exp[(z, )]
(69)
which implies that the pullback g1 of under g1 equals , i.e. that

is invariant.
(b) follows from (c) and (z, g1 ) = (g11 z, ).
The property (b) implies that is defined globally on Z, whereas
(a) and (d) mean that is a symplectic form (Kobayashi and Nomizu
[1969]) on Z, which makes Z a possible classical phase space. Finally,
(c) means that the symplectic structure defined by is Ginvariant,
hence G acts on Z by canonical transformations. (Actually, the 2
form together with the complex structure define a K
ahler structure
on Z, i.e. a Hermitian metric such that the complex structure is
invariant under parallel translations.)
74

A simple but important example of the foregoing methods is provided
by their application to the threedimensional rotation group SO(3).
The resulting frame vectors are known as spin coherent states. An
excellent and detailed account of this is given in Perelomov [1986]; the
treatment here will be fairly brief. SO(3) is locally isomorphic to the
group SU (2) of unitary unimodular 22 matrices, which we denote
by G in this section. This is the set of all matrices

g=
,
||2 + ||2 = 1,
(1)

hence G S 3 (the unit sphere in IR4 ) as a manifold. The Lie algebra

g has a basis {J1 , J2 , J3 } satisfying [J1 , J2 ] = iJ3 plus cyclic permutations (where, as usual, it is actually iJk which span the real algebra
g) which can be conveniently given as Jk = (1/2)k in terms of the
Pauli matrices

1 =
0
1
1
0

,
2 =
0 i
i 0

,
3 =
1 0
0 1

.
(2)
Root vectors in gc are given by J = J1 iJ2 , satisfying

[J3 , J ] = J ,
[J+ , J ] = 2J3 .
The vectors {J3 , J } form a complex basis for gc , with

0 1
0 0
J+ =
,
J =
.
0 0
1 0
(3)
(4)
Unitary irreducible representations of G are characterized by a single

number (highest weight) s = 0, 1/2, 1, 3/2, . . ., with the representation
space H having dimensionality 2s + 1. The generators Jk are represented by hermitian matrices Sk satisfying the irreducibility (Casimir)
condition
S2 S12 + S22 + S32 = (1/2)(S+ S + S S+ ) + S32 = s(s + 1).
(5)
A basis for H is obtained by starting with a highestweight vector vs ,

i.e.,
75
S+ vs = 0,
(6)
S3 vs = svs ,
and applying S repeatedly until the commutation relations imply

that the resulting vector vanishes. This results (after normalization)
in an orthonormal basis {vs , vs1 , . . . , vs } satisfying
S3 vm = mvm
p
S+ vm = (s m)(s + m + 1) vm+1
p
S vm = (s + m)(s m + 1) vm1 .
(7)
To build a homogeneous frame as in section 3.3, we use a decomposition of G in terms of Euler angles,
g(, , ) = exp(iJ3 ) exp(iJ2 ) exp(iJ3 ),
(8)
with 0 < 2, 0 and 0 < 4, which gives a

corresponding decomposition of U (g) as
U (, , ) = exp(iS3 ) exp(iS2 ) exp(iS3 ).
(9)
If 2s is odd, , and have the same ranges as before; if 2s is even,

then + 2 and give the same operators, hence 0 < 2. If
we choose one of the basis vectors vm as our initial vector h, then the
stability subgroup is H = {g(0, 0, )} S 1 . The homogeneous space
G/H S 3 /S 1 is parametrized by (, ), or by the unit vectors n =
(cos sin , sin sin , cos ), hence G/H can be identified with the
unit sphere S 2 . Choosing the section : S 2 G as (, ) = (, , 0),
we obtain the frame vectors
hn = eiS3 eiS2 h.
(10)
The Ginvariant measure on S 2 is just the area measure

dn = sin d d ,
(11)
and the tight frame is

Z
dn | hn ih hn | = cI
(12)
S2
for some number c. To find c, take the trace of both sides and use the
fact that | hn ih hn | is a rankone projection operator (and hence its
trace is unity). This gives
76
4 = c Tr I = c(2s + 1),
(13)
hence the resolution of unity is

Z
2s + 1
dn | hn ih hn | = I.
4
S2
(14)
The overlap between frame vectors can be shown to be

2s

1 + n n0
2
,
| h hn | hn0 i | =
2
(15)
hence they are orthogonal if and only if n0 = n.

To construct a holomorphic frame, we must consider the complex
Lie algebra gc and its Lie group Gc = SL(2, Cl ) of unimodular 2 2
complex matrices

g=
,
= 1.
(16)

The Lie algebra h of the subgroup H of G used above is a Cartan
subalgebra of g, and hc = ClJ3 . The corresponding rootspace decomposition is
gc = hc + n+ + n = ClJ3 + ClJ+ + ClJ ,
(17)
and this yields a Gaussian decomposition of (almost all) elements of

Gc as products of lowertriangular, diagonal and uppertriangular
matrices:
g(, d, ) = eJ e2dJ3 eJ+

d

1 0
e
0
1
=
1
0 ed
0

(18)
d + .
If we write N = exp(n ), then the above decomposition is Gc
N Hc N + . Comparison with the original form of g gives
= ed ,
= ed ,
= ed ,
= ed + ed .
(19)
We call (g), (g) and ed (g) the Gaussian parameters of g. Thus
(g) = ,
(g) = /,
(g) = /.
77
(20)
Remarks:
1. Matrices with = 0, i.e. of the form

g=
0
1

,
(21)
clearly do not have this decomposition. They form a 2dimensional complex submanifold of Gc , hence have (group) measure
zero.
which
2. Elements of G are distinguished by =
and = ,
implies
1 = (1 + )
1 .
= (1 + )
(22)
The Borel subgroup discussed in section 3.4 is B = Hc N + and
consists of all matrices of the form

b=
.
(23)
0 1
Its cosets B can therefore be parametrized by Cl. The unattainable matrices with = 0 form a single coset, corresponding to = .
Hence the homogeneous space Gc /B is the Riemann sphere:
Z Gc /B Cl {} S 2 ,
(24)
in agreement with our earlier G/H = S 2 but with an added complex

structure, as claimed in section 3.4. The action of Gc on Z is as
0
00
follows: If g
B =
B, i.e.

1 0
1 0
B=
B,
(25)

0 1
00 1
then
0
00 = (g
)=
+ 0
.
+ 0
(26)
That is, Gc acts on Z by M

obius transformations.
#
We defined hg = T (g) h, where T (g) was the analytic continua1
tion of U (g) to Gc and T (g)# = (T (g) ) . But
78

T ( d + )# = T (+ )# T (d)# T ( )#
+ ) exp(2dS
3 ) exp(S
).
= exp(S
(27)
Hence the unique state left invariant by B is that corresponding to

the vector of lowest weight, h = vs . For this choice, we get
+ ) h e2ds
hg = e2ds exp(S
h .
(28)
The holomorphic coherent states with respect to the reference point

g1 Gc are then
g1 =
e2sd1 h
hg
=
,
h hg 1 | hg i
h h1 | h i
(29)
with
+ ) h i.
h h1 | h i = h h | exp(1 S ) exp(S
(30)
+ ) in the opposite facTo evaluate this, express exp(1 S ) exp(S

+
torization N Hc N and use S h = 0. It suffices to do the computation at the level of 22 matrices since the result depends only on
the commutation relations, which are preserved by the representation.
Thus

1
1
0
1

1
0 1

d0
1 0
e
=
0 1
0

0
ed
1
0
0
1
(31)
0 0 0
+
d
which gives
0
ed = 1 + 1 ,
ed 0 = 1 ,
0 ed = .
(32)
Hence
0 0 0
h h1 | h i = h h | T (+
d ) h i
2s
0
= e2sd = 1 + 1 ,
and
(33)

2s
g1 = e2sd1 1 + 1
h .
79
(34)
Since a single chart covers all of Z S 2 except for the point at = ,

just one reference point g1 is needed in this case. The simplest choice
is g1 = 1, the identity of G, which gives
1 = h .
(35)
With this choice, the weight function is

e() = | h h | hg i | 2 ,
2s(d+d)
=e
g = d+
| h h | h i | 2
(36)
= e2s(d+d) .
But for g G we have the constraint
1 ,
e2s(d+d) = | | 2 = (1 + )
(37)
2s .
e() = (1 + )
(38)
hence
To find the invariant measure on Gc /B, recall that the 2form =

is invariant under Gc and nondegenerate. Hence it defines an
i
invariant measure, once we choose a positive orientation on Gc /B.
An easy computation gives
= 2is ln(1 + )
2isd d
=
2 .
(1 + )
(39)
d d
2s+2 | h ih h | = cI
(1 + )
(40)
Thus
Z
2si
C
l
for some c. Taking the trace and using

2s
h h | h i = (1 + )
gives, with = rei ,
(41)
80

Z
8s
0
rdr
= c(2s + 1),
(1 + r2 )2
(42)
from which c = 4s/(2s + 1). Thus we have the resolution of unity

Z
d2
2s + 1
(43)
2s+2 | h ih h | = I,
C
l (1 + )
where d2 is now Lebesgue measure on Cl.
What do the functions f() = h h | f i look like? Consider the
vectors
un =
1
(S+ )n h,
n!
n = 0, 1, 2, .
(44)
Then u0 h, u1 , . . . , u2s are linearly independent and u2s+1 = 0,

2s+1
since S+
= 0. Thus
h = eS+ h
1 + + ()
2s u2s ,
= u0 + u
(45)
f() = f0 + f1 + + 2s f2s ,
(46)
hence
where fn = h un | f i. Thus, f() is a polynomial of degree 2s in .

How are the two sets of frame vectors hn and h related? It
turns out that n is related to by stereographic projection of the
2sphere onto the complex plane. The exact relation depends on the
particular factorizations of G used to construct the frames. In the
case of hn , this was the Euler angle decomposition, whereas for h it
was the Gaussian decomposition. However, there is an intrinsic way
of relating the two sets of vectors, which goes as follows: Consider the
functions
h h | Sk h i
Sk () =
,
h h | h i
(47)
which are the expectations of the observables Sk in the state of h ,

and which correspond to the components of the classical angular momentum of a system whose only degrees of freedom are the rotational
81
motions given by G. Sk () is a quantummechanical version of a

+ )h, we have
statistical average. Since h = exp(S
S+ () = lnh h | h i
2s
=
.
1 +
(48)
n
n
To find S3 (), note that [S3 , S+ ] = S+ implies [S3 , S+
] = nS+
, hence
h
i
+
+ eS
S3 , eS+ = S
(49)
and
+
+ h + eS
S3 h = S
S3 h

+ + s h .
= S
(50)

1
S3 () = s S+ () = s
.
+ 1
(51)
Hence
The equator of S 2 corresponds to S3 = 0, hence to || = 1, and the

south pole (S3 = s) and north pole (S3 = s) to = 0 and = ,
respectively. The above equations imply that
2 () S1 ()2 + S2 ()2 + S3 ()2
S
= |S+ ()|2 + S3 ()2
(52)
= s2 ,
thus S()
belongs to the 2sphere of radius s centered at the origin. In
fact, the correspondence S()

is a bijection if we include = .
The relation
S+ ()
=
s S3 ()
(53)
shows that is just the stereographic projection of S()

from the
north pole to the complex plane tangent to the south pole. We could
choose s as a new independent variable ranging over the 2sphere of
82
radius s minus the north pole and define hs = h , where is the unique
point with S()

= s. Then the vectors hs are essentially equivalent to
the hn s. Note that they are eigenvectors of the operator s S with
eigenvalue s2 , i.e.,
s Shs = s2 hs ,
(54)
since the equation obviously holds for s = (0, 0, s), hence for all s
by symmetry.
3.6. The Harmonic Oscillator as a Contraction Limit

In section 3.4, we gave a heuristic argument suggesting that the Weyl
Heisenberg group W is a contraction limit of SU (2). This can
now be made precise and given a geometric interpretation at the representation level. Furthermore, imitating the analysis of the non
relativistic limit of KleinGordon theory (next chapter), we shall gain
an understanding of the relation between the Harmonic Oscillator and
the canonical coherent states in the bargain. This connection between
the rotation group and the harmonic oscillator has some potentially
important applications in quantum field theory, which I hope to explore in the future.
We will study the limit of the representation of G = SU (2) with
spin s as s . Since s is now a variable, we denote the representation space by Hs and the holomorphic coherent states by hs . The
matrices representing the generators Jk will be denoted by Sks . By
way of motivation, compare the resolution of unity on Hs ,
Z
2s + 1
2s2 | hs ih hs | = Is ,
d2 (1 + )
(1)
C
l
with that on the representation space H of W in terms of the canonical
coherent states,
Z
1
d2 z ezz/2 | z ih z | = I.
(2)
2 Cl
Note that if we set
z
=
2 s+1
(3)
83
and define sz hs , then the resolution of unity on Hs becomes

2s + 1
4(s + 1)
d z
C
l
zz
1+
2(2s + 2)
(2s+2)
| sz ih sz | = I.
(4)
If we now take the formal limit s , this coincides with the resolution of unity on H, provided we can show that sz z . Our task
is now (a) to find the sense in which this limit is to be taken, (b) to
show that the generators Sks of g go over to the generators A, A and
I of w and (c) to show that the coherent states sz go over to the
canonical coherent states z .
To properly study the limit s , we will first of all imbed all
the spaces Hs into H, so that the limit may be considered within H.
This is done most easily by using the orthonormal bases obtained by
s
applying S+
and A to the respective ground states. An orthonormal basis for H is given by
1/2
wn = (2n n!)
(A ) 0 ,
n = 0, 1, 2, . . . ,
(5)
where A0 = 0 and the normalization is determined by the commutation relation [A, A ] = 2I. The generators A and A act by
Awn = 2n wn1
(6)
A wn = 2n + 2 wn+1 .
An orthonormal basis for Hs is given by
s
(2s n)! s n s
wns =
S+ h ,
n = 0, 1, 2, . . . , 2s,
n!(2s)!
(7)
s s
where S
h = 0. We imbed Hs into H by identifying wns with wn and
defining Sks wn = 0 for n > 2s. Then for 0 n 2s,
p
s
S
wn = n(2s n + 1) wn1
p
s
(8)
S+
wn = (n + 1)(2s n) wn+1
S3s wn = (n s)wn .
To see how the generators Sks must be scaled, note that the relation
between and z implies that
84

s )hs = exp(
hs exp(S
z K+ /2)w0 sz ,
+
(9)
where
s
K+
=
s
S+
.
s+1
(10)
s
s
s
s
Define K
= S
/ s + 1, so that K
= K+
, and
K3s
1 s
S3s
s
[K+ , K
]=
.
2
s+1
(11)
Then for 0 n 2s,

r
n(2s n + 1)
wn1
s+1
r
(n + 1)(2s n)
s
wn+1
K+ wn =
s+1
ns
wn .
K3s wn =
s+1
s
K
wn =
(12)
If we now take the limit s while keeping n fixed, we obtain
s
K
wn 2n wn1
s
(13)
K+
wn 2n + 2 wn+1
K3s wn wn .
Comparing this with the action of A and A , we see that
s
K
A
s
K+
A
K3s
(14)
as s in the weak operator topology of H. (These limits are not

valid in the strong operator topology since it is necessary to keep n
fixed.) Thus we have shown that in the weak sense, the representation
of SU (2) goes over to a representation of W as s . Note that the
operator
s
N
= S3s + s,
(15)
85
which satisfies
s
N
wn = nwn ,
0 n 2s,
s
N
wn = swn ,
n > 2s,
(16)
has the weak limit

s
N =
wlim N
1
A A,
2
(17)
which is the Hamiltonian for the harmonic oscillator (minus the

groundstate energy). If we write A = X + iP with X and P self
adjoint, then
N=

1
X2 + P 2 I .
2
(18)
s
The fact that K+
A means that

s
sz exp zK+
/2 w0 exp(
z A /2)w0 z ,
(19)
where z are the canonical coherent states and the convergence is in

the weak topology of H.
We now examine the limit s from a global geometric point
of view, using the coherent states. Recall that the expectation vector
s ()
S
h hs | Ss hs i
h hs | hs i
(20)
ranges over the sphere of radius s centered at the origin, with the
north pole corresponding to = . The transformation from Sks to
Kks deforms this sphere to an ellipsoid,
s
+
|K
()|2
3s ()2 =
+K
s+1
s
s+1
2
.
(21)
s 1.
When s , this ellipsoid splits into the two planes K
3
s
Our weak limit K3 I only picked out the lower plane. We could
have picked out the upper plane by imbedding Hs into H differently,
namely by identifying w2sn with wn for n = 0, 1, 2, . . . , 2s. In that
case we would have obtained the weak limits
86

s
K
A
s
K+
A
(22)
K3s I.
In terms of the coherent states, this corresponds to using the highest
weight vector instead of the lowestweight vector or, equivalently,
using a chart centered about the north pole rather than the south
pole, for example, by using as reference point the element g1 = eiJ2 .
The corresponding harmonicoscillator Hamiltonian is the weak limit
of the operator
s
s
N+
s S3s = 2s N
.
(23)
s
s
The expectation values of N
and N+
,
2s
s
N
() =
1 +
2s
s
+
N
() =
,
1 +
(24)
s
s
+
N
() = N
(1/).
(25)
are related by inversion:
The splitting of the ellipsoid into the two planes can be understood from the point of view of representation theory by writing the
irreducibility condition in terms of the Ks:

s
1
2
s
s
s
s
K+
K
+ K
K+
+ (K3s ) =
Is .
2s + 2
s+1
2
(26)
When s , this implies formally that (K3s ) I. The subspaces

H on which K3s I are invariant in the limit, hence the limiting
representation of W is reducible. Evidently, by taking the limits in the
weak topology we are able to pick out just one irreducible component
at a time.
There is an interesting analogy between the above analysis and
the nonrelativistic limit of KleinGordon theory, which we now point
out for those readers already familiar with the latter. (The non
relativistic limit will be discussed in the next chapter.) The relation
S3s /(s+1) I corresponds to P0 /mc2 I in the limit mc2 ,
87
where P0 is the relativistic energy operator. As will be seen in the

next chapter, the Poincare group contracts, in this approximation, to
a semidirect product of the 7dimensional WeylHeisenberg group and
the rotation group. A firstorder correction is obtained by expanding
p

P0 = m2 c4 + P2 c2 mc2 + P2 /2m mc2 + H , (27)
where H is the nonrelativistic free Schr
odinger Hamiltonian. Similarly, it follows from

1 s s
2
s s
S+ S + S
S+ + (S3s ) = s(s + 1)
2

1 s s
s s
S+ S S
S+ S3s = 0
2
(28)
that
2
s s
(S3s 21 ) = (s + 12 ) S+
S ,
(29)
hence formally
S3s
1
2
q
2
s Ss
= (s + 12 ) S+
Ss Ss
(s + ) + ,
2s + 1
(30)
1
2
from which
s()
S3
s(+)
S3
()
1 s s
K ,
s + K+
2
1 s s
s + 1 K+
K ,
2
(31)
which corresponds to P0
(mc2 + H). The analogy between
the largespin and the nonrelativistic limits can be summarized as
follows: The Poincare group corresponds to SU (2), the energy P0 to
S3s , the rest energy mc2 to s, and the nonrelativistic Hamiltonian H
to the harmonic oscillator hamiltonian N . The analog of the central
extension of the Galilean group (which is the contraction limit of the
Poincare group, as discussed in the next chapter) is the Oscillator
group (Streater [1967]), whose generators are A, A , I and N , with
the commutation relations
88

[N, A ] = A .
[N, A] = A,
(32)
Finally, we note that the above analysis explains a wellknown

relationship between the harmonic oscillator and the canonical coherent states, namely that the latter evolve naturally under the harmonic
oscillator dynamics. We first derive the corresponding relation within
SU (2), i.e. for finite s. Consider the behavior of hs under the evos
lution operator exp(itN
). We wish to express
eitN sz = eits eitS3 eS+ hs
(33)
in terms of the coherent states sz , hence we need to display the operator on the right in the reverse Gaussian form
s
exp(0 S+
) exp(2d0 S3 ) exp( 0 S ).
(34)
As explained earlier, it suffices to do the computation on 22 matrices:

it/2

e
0
1
+
itJ3 J
e
e
=
0
eit/2
0 1
d0

(35)
0
0
1 0
1
e
0
.
0 ed
0 1
0
1
This gives d0 = it/2, 0 = eit and 0 = 0. Hence
0 S s
eitN sz = eits e
0 S s
= e
=
eitS3 h
h = h 0
(36)
sz(t) ,
where z(t) = eit z. (This is intuitively obvious, since exp(itS3s ) rotates the 12 plane clockwise by an angle t, hence it rotates the coordinate z counterclockwise by an angle t.) In the limit s , this
gives the wellknown result (Henley and Thirring [1962]) that the set
of canonical coherent states is invariant under the harmonic oscillator
time evolution, with individual coherent states moving along the classical trajectories z(t) determined by the initial conditions z = xip in
phase space. The above shows that the same is true within SU (2), i.e.
s
is essentially
for finite s, where it is a consequence of the fact that N
the generator of rotations about the 3axis.
Note: After this section was written, I learned from R. F. Streater
that a related construction was made by Dyson [1956].
89
Chapter 4
COMPLEX SPACETIME
4.1. Introduction
Relativistic quantum mechanics is a synthesis of special relativity and
ordinary (i.e., non-relativistic) quantum mechanics. The former is
based on the Lorentzian geometry of spacetime, while the latter is
usually obtained from classical mechanics by a somewhat mysterious
set of rules known as quantization in which classical observables,
which are functions on phase space, suddenly become operators on
a Hilbert space. Classical mechanics, in turn, can be formulated in
terms of Newtonian space-time (the Lagrangian approach) or it can
be based on the symplectic geometry of phase space (see Abraham
and Marsden [1978]). The latter, called the Hamiltonian approach,
is usually considered to be deeper and more powerful, and its study
has virtually turned modern classical mechanics into a branch of differential geometry. Yet, the standard formalism of relativistic quantum mechanics rests solely on the geometry of spacetime. Symplectic
geometry, so prominent in classical mechanics, seems to have disappeared without a trace.
In this chapter we develop a formulation of relativistic quantum
mechanics in which symplectic geometry plays an important role. This
will be done by studying the role of phase space in relativity and discovering its counterpart in relation to the Poincare group, which is
the invariance group of Minkowskian spacetime. It turns out that the
Perelomov construction fails for relativistic particles (the physically
relevant representations are not square-integrable), and an alternative route must be taken. The result is a formalism based on complex
spacetime which, we show, may be regarded as a relativistic extension
of classical phase space. As a by-product, two long-standing inconsistencies of relativistic quantum mechanics (in its standard spacetime
formulation) are resolved, namely the problems of localization and
covariant probabilistic interpretation. Rather than being sharply localizable in space (which leads to conflicts with causality; see Newton
90
4. Complex Spacetime
and Wigner [1949] and Hegerfeldt [1985]), particles in the new formulation are at best softly localizable in phase space. This is just a covariant version of the situation in the coherentstate representation. But
whereas for non-relativistic particles both the Schr
odinger representation and the coherentstate representation give equally consistent
theories, the spacetime formulation of relativistic quantum mechanics
is inconsistent because it lacks a genuine probabilistic interpretation,
a situation remedied by the phase-space formulation (section 4.5).
4.2. Relativity, Phase Space and Quantization

At first it appears that phase space and spacetime are mutually exclusive: The phase-space coordinates of a particle are in one-to-one
correspondence with the initial conditions which determine its classical motion, i.e. its worldline. Hence the phase space can be identified
with either the set of all initial conditions or the set of worldlines of
the classical particle. In either case, time is treated differently from
space: For initial conditions, an arbitrary initial time is chosen;
for world-lines, the dynamical flow is factored out. Furthermore,
locality in spacetime is lost in either case.
In this section we confine ourselves to the physical case s = 3, i.e.,
threedimensional space. Our approach will be to leave spacetime
intact and, instead, consider its group of symmetries, the Poincare
group, which is defined as follows: Let u be a four-vector. We denote
its time component (with respect to an arbitrary reference frame)
by u0 and its space components by u = (u1 , u2 , u3 ). The invariant
Lorentzian inner product of two such vectors is defined as
(u, v) uv c2 u0 v 0 u v g u v .
(1)
Later we will set c = 1 (which amounts to measuring time as the

distance traveled by light), but for the present it is important to
include c, since the non-relativistic limit c will be considered.
The Lorentz group L is the set of all linear transformations
: IR4 IR4 which leave the inner product invariant:
(u, v) = (u, v) u, v IR4 .
(2)
L includes transformations which reverse the orientation of time

and ones which reverse the orientation of space. Such space and time
reflections split L into four connected components. The component
91
connected to the identity (whose elements reverse neither the orientations of time nor of space) is called the restricted Lorentz group
and denoted by L0 . If we let u4 cu0 , so that uv = u v + u4 v 4 ,
we can identify L0 with SO(3, 1)+ (the plus sign indicates that the
orientation of time, hence also of space, is preserved separately.) The
Poincare group P is defined as the set of all Lorentz transformations
combined with spacetime translations:
P = {(a, )| L, a IR4 }
(3)
where (a, ) acts on IR4 by

(a, )u = u + a.
(4)
We will be dealing with the restricted Poincare group P0 where

L0 . The reason for our interest in P0 is that it parametrizes all
reference frames which can be obtained from a given reference frame
by a continuous motion. (By motion we mean any spacetime translation, rotation or boost, including physically impossible motions
such as spacelike translations and translations backwards in time.)
Now to specify a reference frame (a, ) relative to some fixed reference frame (located, say, at the origin in spacetime), we must give
its origin (namely, a), its velocity and its spatial orientation (all
relative to the fixed frame). Thus if we ignore the spatial orientation
by factoring out the rotation subgroup SO(3) of L+ , the resulting
seven-dimensional homogeneous space
C P0 /SO(3)
(5)
can be identified with the set of positions in spacetime (events) together with all possible velocities at these events. The set of all
(future-pointing) four-velocities is a hyperboloid
4
2
2
0
+
c {u IR | u (u, u) = c , u > 0}.
(6)
C IR4 +
c ,
(7)
Thus
which is an extension of classical phase space obtained by including

time along with the space coordinates. Such an object is usually called
a state space. Strictly speaking, C is an extended velocity phase space
rather than an extended momentum phase space; this appears to be
92
the price for retaining locality in spacetime. What matters for us

is that it will have the required symplectic (more precisely, contact)
structure. From its construction, it is clear that C is a relativistically
covariant object; it is not invariant since the choice of SO(3) in P0
is frame-dependent. We will see that C combines the geometries of
spacetime and phase space in a natural way.
Thus, rather than conflicting with relativity, the concept of phase
space actually follows from it: Appending time to the geometry of
space means appending velocity-changing transformations (which are
just rotations in a space-time plane) to the group of rigid motions,
hence the enlarged group contains velocity coordinates in addition to
space coordinates. In a sense, P0 itself is actually a super phase
space since it furthermore contains information on the spatial orientation, which is needed to include spin degrees of freedom along with
the translational degrees of freedom. P0 even has a natural generalization to the case of curved spacetime, namely the frame bundle of
all orthonormal frames (with respect to the given curved metric)
at all possible events. (For the definition and study of frame bundles,
see Kobayashi and Nomizu [1963, 1969].)
In our review of canonical coherent states (chapter 1), we saw that
the classical phase space resulted from the Weyl-Heisenberg group W,
which, unlike P0 was not a symmetry group of the theory but merely
a Lie group generated by the fundamental dynamical observables of
position and momentum at a fixed time. We will now see that W
is related to the non-relativistic limit of P0 in two distinct ways: as
a normal subgroup, and as a homogeneous space. This insight will
play a key role in generalizing the canonical coherent states to the
relativistic case. It turns out that the role of W as a group has no
relativistic counterpart, whereas its role as a homogeneous space does:
its relativistic generalization is C.
Consider the Lie algebra of P0 , which is spanned by the generators Pk of spatial translations, P0 of time translations, Jk of spatial
rotations and Kk of pure Lorentz transformations (the boosts). The
Lie brackets of are given by
[Jj , Jk ] = iJl
[P0 , Kr ] = iPr
[Kj , Kk ] = ic2 Jl
[Jj , Kk ] = iKl
[Jj , Pk ] = iPl
[Pr , Ks ] = ic2 rs P0
(8)
where (j, k, l) is a cyclic permutation of (1,2,3), r, s = 1, 2, 3 and all

unspecified brackets vanish. The physical dimensions of these generators are as follows: P0 is a reciprocal time, Pk is a reciprocal lenght,
93
Jk is dimensionless (reciprocal angle) and Kk is a reciprocal velocity.

Notice that so far nothing has been said about quantum mechanics. is simply the Lie algebra of the infinitesimal motions of
Minkowskian spacetime, a classical concept. (The unit imaginary i in
eq (9) can be removed by replacing P with iP , Jk with iJk and
Kk with iKk ; i is included because we anticipate that in quantum
mechanics these generators become self-adjoint operators.) Quantum
mechanics is now introduced through the following postulate:
(Q). The formalism of relativistic quantum theory is based on a unitary (though possibly reducible) representation of P0 .
That is, the representation provides the quantummechanical Hilbert
space, and the generators of , which by unitarity are represented by
self-adjoint operators, are interpreted as the basic physical observables: P0 as the energy, Pk as the momentum and Jk as the angular
momentum. One may then consider perturbations by introducing interactions or gauge fields. In fact, the assumption (Q) is very general
in scope; it serves as one of the axioms in axiomatic quantum field
theory (Streater and Wightman [1964]). Unlike the usual prescription
of quantization in non-relativistic quantum mechanics, (Q) is both
mathematically and physically unambiguous. Yet, (Q) implies and,
at the same time, supercedes the canonical commutation relations!
To see this, consider the non-relativistic limit c of . Letting
c2 P0 M,
(9)
we see that in the limit c , contracts to a Lie algebra g1

with generators M, Pk , Jk , Kk satisfying
[Jj , Jk ] = iJl
[M, Kr ] = 0
[Kj , Kk ] = 0
[Jj , Kk ] = iKl
[Jj , Pk ] = iPl
[Pr , Ks ] = irs M
(10)
and all other brackets vanishing. Note that (a) M is a central element
of g1 and (b) M, Pk and Kk span an invariant subalgebra w of g1
which is isomorphic to the WeylHeisenberg algebra, with M playing
the role of the central element E. Hence if G1 denotes the connected,
simply connected Lie group generated by g1 , then the invariant subgroup of G1 generated by w is isomorphic to the WeylHeisenberg
group W. The remaining generators Jk of g1 span the Lie algebra
so(3) of the spatial rotation group SO(3), so G1 is the semi-direct
product of W with SO(3) :
G1 = W
s SO(3).
(11)
94
Now suppose that the unitary representation of P+

in assumption
(Q) is irreducible. [Assumption (Q) means that we are dealing with
a general quantum system, possibly a system of interacting particles
or even quantum fields; it is the additional assumption of irreducibility which makes this system elementary, roughly a single particle.
Hence the concept of position, discussed below, is only now admissible.] Assuming that the formal limit c of Lie algebras rigorously
induces a corresponding limit at the representation level (and this is
indeed the case, as will be shown later), assumption (Q) implies
that we have a unitary irreducible representation of G1 in that limit.
Then M, Pk and Kk are represented by self-adjoint operators on some
Hilbert space, which we will denote by the same symbols. Irreducibility implies that the central element M has the form mI, where m is a
real number and I denotes the identity operator. Assume m > 0 (this
is physically necessary since m will be interpreted as a mass) and let
Xk = (1/m)Kk .
(12)
Then eq. (10) shows that Xk and Pk satisfy the canonical commutation relations:
[Xr , Ps ] = irs I,
(13)
thus Xk behaves like a position operator. This shows that the assumption (Q), which is conceptually simple, mathematically precise, relativistically invariant and very general, actually implies the much less
satisfactory quantization prescription in the non-relativistic limit,
under the additional assumption of irreducibility.
How does it happen that classical relativistic geometry, as represented by P0 , when combined with assumption (Q), yields the mysterious canonical commutation relations? To understand this, note
that eq. (10) came from the relativistic Lie bracket
[Pr , Ks ] = ic2 rs P0
(14)
which states that boosting (accelerating) in any given spatial direction does not commute with translating in the same direction. This, in
turn, is a consequence of the fact that Einsteinian space is not absolute
since in the accelerated frame there is a (Lorentz) contraction in the
direction of motion, so first translating and then boosting is not the
same as first boosting and then translating. In the non-relativistic
limit this gives the canonical commutation relations. No such easy
95
derivation of these relations would have been possible without invoking Relativity. For had we begun with Newtonian space-time, the
appropriate invariance group would have been not P0 but the Galilean
group G. Since Newtonian space is absolute and hence unaffected by
boosting to a moving frame, the Galilean boosts Kk 0 commute with
with the Galilean generators of translations Pk 0 , hence yield no canonical commutation relations and no associated uncertainty principle. In
the case of G1 , the canonical commutation relations are a remnant of
relativistic invariance. Thus the uncertainty principle originates, in
some sense, in classical Relativity theory!
Eq. (12) states that an acceptable set of position operators for a
non-relativistic particle is given in terms of the generators of Galilean
boosts (more precisely, the boosts of a central extension of the Galilean group, as explained below). It is interesting to see how this comes
about from a more intuitive, physical point of view, since position operators are problematic in relativistic quantum mechanics (as will be
discussed later) but the boosts have natural relativistic counterparts.
To gain insight, we will now give two additional rough but intuitive
arguments for the validity of eq. (12).
1. For a spinless particle, the generators of the Poincare group can be
realized as operators on a space of functions over spacetime (namely,
the space of solutions of the KleinGordon equation) by
x0
Pk = i
xk
Jk = xl Pm xm Pl
P0 = i
(15)
Kk = x0 Pk c2 xk P0
where (k, l, m) is a cyclic permutation of (1, 2, 3). In the non-relativistic limit, P0 mc2 , so
(1/m)Kk xk x0 Pk /m,
(16)
which displays (1/m)Kk as the operator of multiplication by xk (the

usual non-relativistic position operator) minus the distance the particle has traveled in time t = x0 while going at a velocity v = P/m.
This is just the initial position of the particle at time t = 0; that is, the
non-relativistic limit of (1/m)Kk is just the (non-relativistic) position operator in the Schr
odinger picture, where operators are constant
96
while states vary with time.

2. The second explanation of eq. (12) begins with the relativistic
state space C IR4 +
c and considers the non-relativistic limit.
As c , the hyperboloid +
c flattens out to a (three-dimensional)
hyperplane at infinity, the non-relativistic velocity space V IR3 .
In that limit, the boosts Kk commute and become the generators of
translations in V. If the cartesian coordinates of V are vk , then
Kk i
.
vk
(17)
Now the velocities vk are related to the momenta pk by vk = pk /m,

where m is the mass. Thus in the limit
Kk = mi
.
pk
(18)
But in the momentum representation of non-relativistic quantum mechanics , the position operators are represented by
Xk = i
,
pk
(19)
which again gives agreement with eq. (12).

Now that we have an acceptable relativistic generalization of the
classical phase space, let us return to our goal of extending the canonical coherentstate representation to relativistic particles. We have
seen that the WeylHeisenberg group, on which this representation is
based, is isomorphic to an invariant subgroup of the non-relativistic
limit of P+
. However, this subgroup is not the non-relativistic limit
of any subgroup of P0 , since the Lie brackets of Kk , Pk and P0 do
not close due to
[Kj , Kk ] = ic2 Jl , (j, k, l cyclic ).
(20)
That is, we cannot simply generalize the canonical coherent states by

choosing the right subgroup of P0 . However, eq. (11) shows that W
also plays another role in the group G1 , namely as a homogeneous
space:
W = G1 /SO(3).
(21)
97
In this form, it does have an obvious relativistic counterpart, namely

C. In retrospect, the role of W as a group is purely incidental to quantum mechanics , since there is no fundamental reason why the set of
all dynamical states should form a group. On the other hand, this set
should certainly be a homogeneous space under the group of motions,
since this group must transform dynamical states into one another
and this action can be assumed to be transitive (otherwise we may as
well restrict ourselves to an orbit). In either of its roles relative to G1 ,
W acquires a slightly different physical interpretation from the one
it had in relation to the canonical commutation relations: Since the
groupmanifold of W is generated by the vector fields Pk , Kk and M ,
the coordinates on W are the positions xk (generated by Pk ), the velocities vk (generated by Kk ) and the variable generated by M , which
is a degenerate form of time inherited form Relativity through the
limit c2 P0 M . By comparison, the coordinates on the original
WeylHeisenberg group were xk (generated by Pk ), the momentum
pk (generated by Xk ) and a dimensionless phase angle generated
by the identity operator. With this new interpretation, W is the
product of velocity phase space with time and truly represents the
non-relativistic limit of C.
We will construct a coherent-state representation of P0 by first
discovering its non-relativistic limit. This limit will be a representation of a quantum mechanical version G2 of the Galilean group G
and will be seen to be a close relative of the canonical coherentstate
representation of W. It is therefore necessary first to understand just
how the group G is related to the Poincare group P0 and its non
relativistic limit G1 . Again, we will do everything at the level of Lie
algebras. There is no problem with globalization.
The universal enveloping algebra of contains the mass-squared
operator
M 2 = c4 P0 2 c2 (P1 2 + P2 2 + P3 2 ) = c4 P0 2 c2 P2 ,
(22)
which is a Casimir operator, i.e. commutes with all generators in

. Assuming that both P0 and M are represented by positive operators and that M is invertible (and this will be the case for massive
particles), we have formally for large c:
c2 P0 = (M 2 + c2 P2 )1/2 = M + P2 /2M c2 + O(c4 ),
(23)
where we have used the fact that M commutes with Pk . The operator
98
H = P2 /2M
(24)
is just the Hamiltonian for the non-relativistic free particle, hence

generates its time translations and must be included in the Lie algebra
of the Galilean group. Let us therefore try to append it to the Lie
algebra g1 of G1 . Indeed, by eq. (10),
[H, Pk ] = 0
[H, Jk ] = 0
[H, M ] = 0
[H, Kk ] = iPk ,
(25)
g2 Span {Kk , Pk , Jk , M, H}
(26)
showing that
actually forms a Lie algebra with Lie brackets given by eqs. (10) and
(25). Clearly g2 contains g1 as a subalgebra. We will refer to the corresponding Lie group G2 as the quantum mechanical Galilean group.
It is the group of translations, rotations, boosts, dynamics (i.e., time
translations) and multiplications by constant phase factors (generated
by M ) acting on the wave functions of a non-relativistic quantum mechanical particle. The relation of G2 to the classical Galilean group G
is as follows: The subgroup generated by M , which can be identified
with the group of real numbers IR, is central; then G is obtained from
G2 by factoring out IR:
G = G2 /IR.
(27)
The action of IR on quantum mechanical wave functions, which amounts to a multiplication by a constant phase factor, is a necessary
part of G2 because of [Pr , Ks ] = irs M which, as we have seen, is
related to the uncertainty principle. Factoring out this action means
ignoring that phase factor, so it is reasonable that it should give the
classical Galilean group. On the Lie algebra level, it amounts to
setting M = 0. Had we included Plancks constant h in eq. (10), this
would have amounted to taking the classical limit h
0. The above
relation between IR, G2 and G in an example of a central extension
(Varadarajan [1969]). One says that G2 is a central extension of IR by
G.
The fact that W is a subgroup of G2 was noted by Bargmann
[1954]; that representations of G2 are contractions of representations
of the Poincare group was shown by Mackey [1955].*
* I thank R. F. Streater for these remarks.
4.3. Galilean Frames
99

Our object in this section is to construct coherent states which are
naturally associated with free nonrelativistic particles, just as the
canonical coherent states are associated with the WeylHeisenberg
group or the Harmonic oscillator and the spin coherent states are associated with with SU (2). An obvious starting point would be to
apply the KlauderPerelomov method (chapter 1) to the quantum
mechanical Galilean group G2 , since it is this group which describes
such particles. However, this method fails, due to the fact that all the
representations of physical interest are not squareintegrable. Therefore we will follow a more pedestrian route. Our main guides will be
analyticity (which, it turns out, follows from an important physical
condition) and physical intuition.
We return to the general case where the configuration space is IRs
instead of IR3 . For simplicity, we restrict ourselves to spinless particles. It is not difficult to include spin, as will be shown later. The
states of such particles are described by complexvalued wave functions f (x, t) of position x and time t which are are squareintegrable
with respect to x at any time t. Their evolution in time is given by
the Schr
odinger equation
i
f
= Hf,
t
(1)
where
1
(2)
2m
is the Hamiltonian operator, and is the Laplacian acting on
L2 (IRs ). Since H is selfadjoint, though unbounded, the solutions are
given through the unitary oneparameter group U (t) = exp(itH):
H=
f (x, t) = (U (t)f )(x)

Z
s
= (2)
ds p exp[itp2 /2m + ip x]f(p),
(3)
IRs
where it is assumed that f (x, 0), hence also its Fourier transform f(p),
is in L2 (IRs ).
The key to our approach will be to note that H is a nonnegative
operator, hence the evolution group U (t) can be analytically continued
to the lowerhalf complex time plane Cl as
100
U (t iu) = exp[i(t iu)H] = eitH euH ,
u > 0.
(4)
(Note also that since H is unbounded, no analytic continuation to

the upperhalf time plane is possible.) The operator euH is familiar
from two other contexts: it constitutes the evolution semigroup for
the heat equation (where u is time), and it is also the unnormalized
density matrix for the Gibbs canonical ensemble describing the statistical equilibrium of a quantum system at temperature T (where
u = 1/kT ). To get a feel for our use of this operator, let us be heuristic for a moment and consider what happens when a free classical
free particle of mass m is evolved in complex time = t iu. If its
initial position and momentum are x and p respectively, then its new
position will be
z( ) = x + (t iu)p/m
= (x + tp/m) i(u/m)p
(5)
= x(t) i(u/m)p.
Since x(t) is just the position evolved in real time t, we see that z(t)
is, in fact, a complex phase space coordinate of the same type we encountered in the construction of the canonical coherent states! Armed
with this intuition, let us now return to quantum mechanics and see if
this idea has a quantum mechanical counterpart. The operator euH ,
when applied to any function in L2 (IRs ), gives
fu (x) (euH f )(x)
Z
s
= (2)
ds p exp[up2 /2m + ip x]f(p).
(6)
IRs
If we replace x in the integrand by an arbitrary z Cls , the integral

still converges absolutely since the quadratic term in the exponent
dominates the linear term for large |p|. Clearly the resulting function
is entire in z (differentiating the integrand with respect to zk still
gives an absolutely convergent integral). This shows that the group
of Galilean spacetime translations,

U (x, t) = exp itH + ix P ,
(7)
101
extends analytically to a semigroup of complex spacetime translations

U (
z, ) = exp i
H + i
zP
(8)
defined over the complex spacetime domain
D = { (z, ) | z Cls , Cl }.
(9)
This translation semigroup can be combined with the rotations and

boosts to give an analytic semigroup G2c extending G2 .
Let Hu be the vector space of all the entire functions fu (z) as
f(p) runs through L2 (IRs ). Then
fu (z) = h euz | f i,
(10)
where
]
euz (p) = (2)s exp[up2 /2m ip z
h
i
u
= (2)s exp my2 /2u
(p my/u)2 ip x
2m
(11)
are seen to be Gaussian wave packets in momentum space with expected position and momentum given in terms of z x iy by
h Xk i = xk
and h Pk i = (m/u)yk .
(12)
The euz s are easily shown to have minimal uncertainty products. The
momentum uncertainty can be read off directly from the exponent
and is
p
Pk = m/2u,
(13)
hence
Xk =
p
u/2m.
(14)
We now have our prospective coherent states and their label space
M = Cls . To construct a coherentstate representation, we need a
measure on M which will make the euz s into a frame. Since the euz s
are Gaussian, the measure in not difficult to find:

du (z) = (m/u)s/2 exp my2 /u ds x ds y.
(15)
102
Defining
Z
h f | g iHu h fu | gu i
du (z)fu (z)gu (z),
(16)
we have
Theorem 4.1.
(a) h | iHu is an inner product on Hu under which Hu is a Hilbert
space.
(b) The map euH is unitary from L2 (IRs ) onto Hu .
(c) The euz s define a resolution of unity on L2 (IRs ) given by
Z
du (z) | euz ih euz | = I.
(17)
C
ls
Proof. We prove that k f k2Hu h f | f iHu = kfk2L2 . The inner

product can be recovered by polarization. To begin with, assume
that f is in the Schwartz space S(IRs ) of rapidly decreasing smooth
test functions. Then

fu (x iy) = exp up2 /2m + y p f (x),
(18)
hence by Plancherels theorem,
Z
ds x | fu (x iy) | 2
Z

s
= (2)
ds p exp up2 /m + 2y p | f(p) | 2
(19)
and
Z
du (z) | fu (z) | 2
s/2
= (m/u)
Z
s
= (2)
(2)

ds p exp up2 /m | f(p) | 2
(20)

ds y exp my2 /u + 2y p
ds p | f(p) | 2 = k f k2 ,
103
where exchanging the order of integration was justified since the integrals are absolutely convergent. This proves (b), hence also (a),
for f S(IRs ). Since the latter space is dense in L2 (IRs ), the proof
extends to f L2 (IRs ) by continuity. (c) follows by noting that
h f | g iL2 = h f | g iHu
Z
= du (z)fu (z)gu (z)
Z
= du (z)h f | euz ih euz | g i
(21)
and dropping h f | and | g i.

Using the map euH , we can transfer any structure from L2 (IRs )
to Hu . In particular, time evolution is given by

fu (z, t) = euH eitH f (z)
Z

s
= (2)
ds p exp i p2 /2m + iz p f(p)
(22)
h ez, | f i,
where = t iu and the wave packets
]
ez, (p) = (2)s exp[
p2 /2m ip z
(23)
are obtained from the euz s by evolving in real time t. They cannot be
of minimal uncertainty since the free-particle Schr
odinger equation is
neccessarily dissipative. Instead, they give the following expectations
and uncertainties:
h Pk i = (m/u)yk
h Xk i = xk (t/u)yk = xk (t/m)h Pk i
p
Pk = m/2u
s

t2
u
Xk =
1+ 2 .
2m
u
(24)
Since
ez, = eitH euz ,
(25)
104
it follows that
kf k2
du (z) | f (z, ) | 2
C
ls
itp2 /2m
= ke
(26)
fk = kfk ,
2
thus we have a frame { ez, | z Cls } at each complex instant =

t iu, with the corresponding resolution of unity
Z
du (z) | ez, ih ez, | = I.
(27)
C
ls
The space L2 (IRs ) carries a representation of the quantum mechanical Galilean group G2 . Since the ez, s were obtained from the
dynamics associated with this group, they transform naturally under
its action. A typical element of G2 has the form g = (R, v, x0 , t0 , ),
where R is a rotation, v is a boost, x0 is a spatial translation, t0 is
a timetranslation and is the phase parameter associated with
the central element M = m/
h m in our representation (see section
4.2). g acts on the complex spacetime domain D by sending the
point (z, ) to ( 0 , z0 ), where
x0 = Rx + tv + x0
y0 = Ry + uv
t 0 = t + t0
(28)
u0 = u.
The parameter has no effect on spacetime; it only acts on wave
functions by multiplying them by a phase factor. The representation
of G2 is defined by

Ug f (z, ) = eim f g 1 (z, ) .
(29)
Thus we have
Ug ez, = eim eg(z, ) ,
(30)
and the ez, s are projectively covariant under the action of G2 ; if we

define ez,, eim ez, , then this expanded set is invariant under
the action of G2 , with 0 = . Since ez, and ez,, represent
the same physical state, we wont be fussy and just work with the
105
ez, s. Anyway, this anomaly will disappear when we construct the

corresponding relativistic coherent states.
The above representation of G2 on L2 (IRs ) can be transfered to
Hu using the map euH . This map therefore intertwines (see Gelfand,
Graev and Vilenkin [1966]) the representations on G2 on L2 (IRs ) with
the one on Hu .
We conclude with some general remarks.
1. Since the euz s are spherical and therefore invariant under SO(n)
(which is, after all, why they describe spinless particles), they can be
parametrized by the homogeneous space W = G1 /SO(n) as long as we
keep u fixed (u is a parameter associated with the Hamiltonian, which
is a generator of G2 but not of G1 ). The action of W as a subgroup
of G1 on the euz s is preserved in passing to the homogeneous space,
hence W acts to translate these vectors in phase space. This explains
the similarity between the euz s and the canonical coherent states. On
the other hand, dynamics (in imaginary time) is responsible for the
parameter u. If we write k (m/u)y, then
h u
i
u
euz (p) = (2)s exp
k2
(p k)2 ip x
2m
2m
!
(31)
2
ipx
pk
exp uk /2m e
h p
.
2m/u
The measure du (z) is now

du (x, k) = (u/m)s/2 exp uk2 /m ds x ds k.
(32)
Hence the exponential factor exp[uk2 /2m] in euz , when squared in the
reconstruction formula, precisely cancels the Gaussian weight factor
in du (x, k), leaving the measure
u/m
s/2
ds x ds k
(33)
u
in
p phase space. It follows from the above form of ez that 2Pk =
2m/u plays the role of a scale factor in momentum space (as used
in
p the wavelet transforms of chapter 1), hence its reciprocal Xk =
u/2m acts as a scale factor in configuration space. Thus the Galilean coherent states combine the properties of rigid windows with
those of wavelets, due to the fact that their analytic semigroup G2c includes both phasespace translations and scaling, the latter due to the
106
heat operator euH . However, note that u is constant, though arbitrary, in the resolution of unity and the corresponding reconstruction
formula. Since there is an abundance of wavelets due to translations in phase space, only a single scale is needed for reconstruction.
(One could, of course, include a range of scales by integrating over
u with a weight function, but this seems unnecessary.) In the treatment of relativistic particles, u becomes the time component of a
fourvector y = (u, y), hence will no longer be constant. This is because relativistic windows shrink in the direction of motion, due to
Lorentz contractions, thus automatically adjusting to the analysis of
highfrequency components of the spectrum.
2. Notice that euz is essentially the heat operator euH applied to the
= x + iy. The fact
function at x, then analytically continued to z
that all the euz s have minimal uncertainties shows that the action
of the heat semigroup {U (iu)} is such that while the position undergoes the normal diffusion, the momentum undergoes the opposite
process of refinement, in just such a way that the product of the two
variances remains constant. This is reflected in the fact that the operator euH , whose inverse in L2 (IRs ) is unbounded, becomes unitary
when the functions in its range get analytically continued, and the
reconstruction formula is just a way of inverting euH . Hence no information is lost if one looks in phase space rather than configuration
space! It seems to me that this way of inverting semigroups must
be an example of a general process. If such a process exists, I am
unaware of it. In our case, at least, it appears to be possible because
of analyticity.
3. So far, it seems that coherentstate representations are intimately
connected with groups and their representations. However, there is a
reasonable chance that coherentstate representations similar to the
above can be constructed for systems which, unlike free particles, do
not possess a great deal of symmetry. Suppose we are given a system
of s/3 particles in IR3 which interact with one another and/or with
an external source through a potential V (x). We assume that V (x)
is timeindependent, so the system is conservative. (This means that
we do have one symmetry, namely under time translations. If, moreover, the potential depends only on the differences xi xj between
individual particles, we also have symmetry with respect to translations of the center of mass of the entire system; but we do not make
this assumption here.) This system is then described by a Schr
odinger
equation with the Hamiltonian operator H = H0 +V , where H0 is the
free Hamiltonian and V is the operator of multiplication by V (x). We
107
need to assume that this (unbounded) operator can be extended to a

selfadjoint operator on L2 (IRs ). How far can the above construction
be carried in this case? The key to our method was the positivity of
the free Hamiltonian H0 = P2 /2m. But a general Hamiltonian must
at least satisfy the stability condition:
(S) The spectrum of H is bounded below.
If H fails to meet this condition, then the system it describes is unstable, and a small perturbation can make it cascade down, giving
off an infinite amount of energy. For a stable system, the evolution
group U (t) = eitH can be analytically continued to an analytic semigroup U ( ) = ei H in the lowerhalf complex time plane as in the
free case. Depending on the strength of the potential, the functions
fu = U (iu)f may be continued to some subset of Cls . Formally, this
corresponds to defining

f (z, ) = eizP ei H f (0),

= eyP ei H f (x)
= t iu
(34)
for an initial function f (x) in L2 (IRs ). As mentioned, this expression

is formal since the operator eyP is unbounded and ei H f may not
be in its domain. But it can make sense operating on the range of
ei H , which coincides with the range of euH , provided y is not too
large. Let Yu be the set of all ys for which eyP
is defined

on the

range of euH and, furthermore, the function exp iz P exp i H f
is sufficiently regular to be evaluated at the origin in IRs , no matter
which initial f was chosen in L2 (IRs ). For many potentials, of course,
Yu will consist of the origin alone; in that case there are no coherent
states. We assume that Yu contains at least some open neighborhood
of the origin. Intuitively, we may think of Yu as the set of all imaginary positions which can be attained by the particle in an imaginary
timeinterval u, while moving in the potential V . In the free case,
Yu = IRs and there is no restriction on y provided only that u > 0.
This corresponds to the fact that there is no speed limit for free
nonrelativistic particles, hence a particle can get to any imaginary
position in a given positive imaginary time. For relativistic free particles, Yu is the open sphere of radius uc, where c is the speed of light.
Returning to our system of interacting particles, define the associated
complex spacetime domain
ZH = { (x iy, t iu) Cls+1 | u > 0, y Yu }.
(35)
108
This is the set of all complex spacetime points which can be reached
by the system in the presence of the potential V (x), and it is the label
space for our prospective coherent states. These are now defined as
evaluation maps on the space of analytically continued solutions:
f (z, ) = h eH
z, | f i,
(36)
the inner product being in L2 (IRs ). Then from the above expression,
again formally, we have the dynamical coherent states
i
H i
eH
e zP 0
z, = e
(37)
for (z, ) in ZH .
What is still missing, of course, is the measure dH
u . (Since the potential is tindependent, so will be the measure, if it exists.) Finding
the measure promises to be equally difficult to finding the propagator for the dynamics. The latter is closely related to the reproducing
kernel,
H
0 , 0 ) = h eH
KH (z, ; z
z, | ez0 , 0 i.
(38)
KH depends on and 0 only through the difference 0 , and is

the analytic continuation of the propagator to the domain ZH ZH .
It is related to the measure through the reproducing property,
Z
0 0
, ) KH (z, ; z
00 , 00 )
dH
u (z) KH (z , ; z
(39)
00
00
, ),
= KH (z , ; z
where the integration is carried out over a phase space in ZH
with a fixed value of = t iu. A reasonable candidate for dH
u (see
section 4.4) is
H
2 s
dH
d x ds y
u (z) = C kez, k
C eu (z) ds x ds y.
(40)
Rather than finding the measure explicitly, a more likely possibility is that its existence can be proved by functionalanalytic methods
for some classes of potentials and approximation techniques may be
used to estimate it or at least derive some of its properties. The theoretical possibility that such a measure exists raises the prospect of an
4.4. Relativistic Frames
109
interesting analogy between the quantum mechanics of a single system and a statistical ensemble of corresponding classical systems at
equilibrium with a heat reservoir. In the case of a free particle, if we
set k = (m/u)y as above (see remark 1) and define T by u = 1/2kT
where k is Boltzmanns constant, then it so happens that our measure du is identical to the Gibbs measure for a classical canonical
ensemble (see Thirring [1980]) of s/3 free particles of mass m in IR3 ,
at equilibrium with a heat reservoir at absolute temperature T . Thus,
integrating with du over phase space is very much like taking the classical thermodynamic average at equilibrium! It remains to be seen, of
course, whether this is a mere coincidence or if it has a generalization
to interacting systems.
There is also a connection between the expectation values of an
operator A in the coherent states eH
z, and its thermal average in the
Gibbs state,

h A i Z 1 Trace eH A = Z 1 Trace eH/2 A eH/2 ,
(41)

where Z Trace eH . Namely, if we have the resolution of unity
Z
H
H
dH
(42)
u (z) | ez, ih ez, | = I,
then
h A i = Z
= Z 1
Z
Z
H
H/2
dH
AeH/2 eH
u (z) h ez, |e
z, i
H
H
dH
u (z) h ez, i/2 A ez, i/2 i
(43)
dH
u (z) A(z, i/2),
where we have used the formula

Z
H
H
dH
Trace A =
u (z) h ez, | A ez, i,
(44)
which follows easily from eq. (42). Thus taking the thermal average
means shifting the imaginary part u by /2 in the integral.
110

We are at last ready to embark on the main theme of this book: A
new synthesis of Relativity and quantum mechanics through the geometry of complex spacetime. The main tool for this synthesis will
be the physically necessary condition that the energy operator of the
total system be nonnegative, also known in quantum field theory as
the spectral condition. The (unique) relativistically covariant statement of this condition gives rise to a canonical complexification of
spacetime which embodies in its geometry the structure of quantum
mechanics as well as that of Special Relativity. The complex spacetime also has the structure of a classical phase space underlying the
quantum system under consideration. Quantum physics is developed
through the construction of frames labeled by the complex spacetime
manifold, which thus forms a natural bridge between the classical and
quantum aspects of the system. It is hoped that this marriage, once
fully developed, will survive the transition from Special to General
Relativity.
As mentioned at the beginning of this chapter, the Perelomov
type constructions of chapter 3 do not apply directly to the Poincare
group since its time evolution (dynamics) is nontrivial. Pending a
generalization of these methods to dynamical groups, we merely use
the ideas of chapter 3 for inspiration rather than substance. In fact,
it may well be that a closer examination of the construction to be
developed here may suggest such a generalization.
We begin with the most basic object of relativistic quantum mechanics, the KleinGordon equation, which describes a simple relativistic particle in the same way that the Schr
odinger equation describes a nonrelativistic particle. The spectral condition will enable
us to analytically continue the solutions of this equation to complex
spacetime, and the evaluation maps on the space of these analytic
solutions will be bounded linear functionals, giving rise to a reproducing kernel as in section 1.4. Physically, the evaluation maps are
optimal wave packets, or coherent states, and it is this interpretation
which establishes the underlying complex manifold as an extension
of classical phase space. The next step is to build frames of such
coherent states. (Recall from section 1.4 that a frame determines a
reproducing kernel, but not vice versa.)
The coherent states we are about to construct are covariant under the restricted Poincare group, hence they represent relativistic
wave packets . As we have seen, such a covariant family is closely related to a unitary irreducible representation of the appropriate group,
111
in this case P0 . Such representations are called elementary systems,

and correspond roughly to the classical notion of particles, though
with a definite quantum flavor. (For example, physical considerations
prohibit them from being localized at a point in space, as will be
discussed later.) We will focus on representations corresponding to
massive particles. (A phasespace formalism for massless particles
would be of great interest, but to my knowledge, no satisfactory formulation exists as yet.) Such representations are characterized by two
parameters, the mass m > 0 and the spin j = 0, 1/2, 1, 3/2, . . . of the
corresponding particle. We will specialize to spinless particles (j = 0)
for simplicity. The extension of our construction to particles with spin
is not difficult and will be taken up later. Thus we are interested in
the (unique, up to equivalence) representation of P0 with m > 0 and
j = 0. A natural way to construct this representation is to consider
the space of solutions of the KleinGordon equation
(t
u + m2 c2 )f (x) = 0,
(1)
2
u
t =c
t2
=
(2)
where
2
is the DelAmbertian, or wave operator, is the usual spatial Laplacian and = /x . The function f is to be complexvalued (for
spin j, f is valued in Cl2j+1 ). We set c = 1 except as needed for future
reference. If we write f (x) as a Fourier transform,
Z
s1
f (x) = (2)
ds+1 p eixp f(p),
(3)
IRs+1
then the KleinGordon equation requires that f(p) be a distribution

supported on the mass shell
m = {p = (p0 , p) IRs+1 | p2 p20 p2 = m2 }.
(4)
m is a twosheeted hyperboloid,
m = +
m m ,
where
(5)
112
p
p0 = m2 + p2 (p) on
m.
(6)
f(p) = 2 (p2 m2 ) a(p)
(7)
Taking
for some function a(p) on m , and using

(p2 m2 ) = ((p0 )(p0 + ))
1
=
[(p0 ) + (p0 + )] ,
2
(8)
we get
Z
d
p eixp a(p)
f (x) =
m
(9)

d
p eixp a(p) + eixp a(p) ,
=
+
m
where
d
p = (2)s (2)1 ds p
(10)
is the unique (up to a constant factor) Lorentzinvariant measure

on m . (The factor 1 corrects for Lorentz contraction in frames at
momentum p.) For physical particles, we must require that the energy
be positive, i.e. that a(p) = 0 on
m . Hence the physical states are
given as positiveenergy solutions,
Z
f (x) =
d
p eixp a(p).
(11)
+
m
The function a(p) can now be related to the initial data by setting
x0 t = 0, which shows that
Z
f0 (x) f (x, 0) =
d
p eixp a(p)
+
(12)
m
1
= (2) a (x),
so
a(p) a(, p) = 2 f0 (p),
(13)
113
where denotes the spatial Fourier transform. In particular, f (x) is

determined by its values on the Cauchy surface t = 0. For general
solutions of the KleinGordon equation, we would also need to specify
f /t on that surface, but restricting ourselves to positiveenergy
solutions means that f (x) actually satisfies the firstorder pseudo
differential nonlocal equation
p
f
(14)
= m2 f (x)
t
(which implies the KleinGordon equation), hence only f (x, 0) is necessary to determine f . (We will see that when analytically continued
to complex spacetime, positiveenergy solutions have a local characterization.) The inner product on the space of positiveenergy solutions is defined using the Poincareinvariant norm in momentum
space,
Z
2
kf k
d
p | a(p) | 2 .
(15)
i
+
m
We will refer to the Hilbert space

L2+ (d
p) {a L2 (d
p) | a(p) = 0 on
m}
(16)
as the space of positiveenergy solutions in the momentum representation. It carries a unitary irreducible representation of P0 defined as
follows. The natural action of P0 on spacetime is
(b, )x = x + b,
(17)
where is a resticted Lorentz transformation ( L0 ) and b is a

spacetime translation. Since the KleinGordon equation is invariant
under P0 , the induced action on functions over spacetime transforms
solutions to solutions. Since the positivity of the energy is also invariant under P0 , the subspace of positiveenergy solutions is also left
invariant. P0 acts on solutions by

(U (b, )f ) (x) f 1 (x b) .
(18)
The invariance of the inner product on L2+ (d
p) then implies that the
induced action on that space (which we denote by the same operator)
is

(U (b, ) a) (p) = eibp a 1 p .
(19)
114
The invariance of the measure d

p then shows that U (b, ) is unitary,
thus (b, ) 7 U (b, ) is a unitary representation of P0 . It can be
shown that it is, furthermore, irreducible.
Neither of the function spaces {f (x)} and L2+ (d
p) are reproducingkernel Hilbert spaces, since the evaluation maps f 7 f (x) and
a 7 a(p) are unbounded. To obtain a space with bounded evaluation
maps, we proceed as in the last section. Due to the positivity of the
energy, solutions can be continued analytically to the lowerhalf time
plane:
Z
f (x, t iu) =
d
p exp (it u + ix p) a(p),
(20)
+
m
where u > 0. As in the nonrelativistic case, the factor exp(u)

decays rapidly as | p | , which permits an analytic continuation
of the solution
to complex spatial coordinates z = x iy. But since
p
2
(p) m + p2 is no longer quadratic in | p | , y cannot be arbitrarily large. Rather, we must require that the fourvector (u, y)
satisfy the condition
u y p > 0
(, p) +
m.
(21)
In covariant notation, setting y 0 u, we must have

yp > 0
p +
m,
(22)
so that the complex exponential exp [i(x iy)p] remains bounded

as p varies over +
m . This implies that yp > 0 for all p V+ , where
V+ {p = (p0 , p) IRs+1 | |p| < p0 /c}
(23)
is the open forward light cone in momentum space. In general, we

need to consider the closure of V+ , i.e. the cone
V + = {p IRs+1 | |p| p0 /c},
(24)
which contains the light cone {p2 = 0 | p0 > 0} (corresponding to

massless particles) and the point {p = 0} (corresponding in quantum
field theory to the vacuum state). The set of all ys with yp > 0 is
called the dual cone of V + , i.e.,
V+0 {y IRs+1 | yp > 0 p V + }.
(25)
115
It is easily seen that y belongs to V+0 if and only if | y | < cy 0 . Note

that as c , V + contracts to the nonnegative p0 axis while V+0
expands to the halfspace {(u, y) | u > 0, y IRs } which we have
encountered in the last section. V+0 coincides with V+ when c = 1,
but it is important to distinguish between them since they live in
different spaces (see section 1.1).
Thus for y V+0 , setting z = x iy, we define
Z
f (z) =
d
p eizp a(p).
(26)
+
m
The integral converges absolutely for any a L2+ (d

p) and defines a
function on the forward tube
T+ {x iy Cls+1 | y V+0 },
(27)
also known as the future tube and, in the mathematical literature, as

the tube over V+0 . Differentiation with respect to z under the integral
sign leaves the integral absolutely convergent, hence the function f (z)
is holomorphic, or analytic, in T+ . As y 0 in V+0 , f (z) f (x) in the
sense of L2+ (d
p). Thus f (x) is a boundary value of f (z). Clearly f (z)
is a solution of the KleinGordon equation in either of the variables
z or x. Let K be the space of all such holomorphic solutions:
K = {f (z) | a L2+ (d
p)}.
(28)
Then the map a(p) 7 f (z) is onetoone from L2+ (d

p) onto K. Hence
we can make K into a Hilbert space by defining
h f1 | f2 iK h a1 | a2 i,
(29)
where the inner product on the righthand side is understood to be

that of L2+ (d
p). We now show that K is a reproducingkernel Hilbert
space. Its evaluation maps are given by
Z
Ez (f ) f (z) =
d
p eizp a(p) h ez | a i,
(30)
+
m
where
ez (p) = eizp .
(31)
116
Lemma 4.2.
1. For each z T+ , ez belongs to L2+ (d
p), with
mc
2
1
kez k = (2)
K (2mc),
4
(32)
where = (s 1)/2,
p
p
y 2 = c2 (y 0 )2 y2 > 0
(33)
and K is a modified Bessel function (Abramowitz and Stegun

[1964]; the speed of light has been inserted for future reference.)
2. In particular, the evaluation maps on K are bounded, with
| f (z) | kez k kf k.
(34)
Proof. Set c = 1. (To recover c, replace m by mc in the end.) Then

Z
2
kez k =
d
p e2yp G(y).
(35)
+
m
Since G(y) is Lorentzinvariant and y V+0 , we can evaluate the

integral in a Lorentz frame in which y = (, 0):
Z
d
p e2
G(y) = G(, 0) =
+
m
Z
h
i
p
ds p
s
2
2
p
= (2)
exp 2 m + p .
2 m2 + p2
(36)
Set p = mq. Then

i
h
p
ds q
p
exp 2m 1 + q2
2 1 + q2
h
i
s/2 Z
p
rs1 dr
s s1
2
= (2) m
exp 2m 1 + r
(s/2) 0
1 + r2
(37)
Z
ms1
=
dt sinhs1 t exp [2m cosh t]
(4)s/2 (s/2) 0
m
= (2)1
K (2m).
4
s
G(y) = (2)
s1
117
The reproducing kernel can be obtained by analytic continuation

from kez k2 :
K(z 0 , z) h ez0 | ez i
Z
=
d
p exp [i(z 0 z)p]
+
m
= (2)1
mc
2
(38)

K (mc),
where
p
(z 0 z)2

1/2
= (y 0 + y)2 (x0 x)2 + 2i(y 0 + y)(x0 x)
(39)
is defined by analytic continuation from z 0 = z (when = 2) as

follows: The squareroot function is defined on the complex plane cut
along the negative real axis. Since y and y 0 both belong to V+0 , so
does y 0 + y. Now the argument of the square root is real if and only
if (y 0 + y)(x0 x) = 0, and this can happen only when (x0 x)2 0.
(Otherwise, either x0 x or xx0 belongs to V+0 , hence (y 0 +y)(x0 x)
is positive or negative, respectively.) But then,
(z 0 z)2 = (y 0 + y)2 (x0 x)2 (y 0 + y)2 > 0.
(40)
Thus for z 0 , z T+ , the quantity (z 0 z)2 cannot belong to the

negative real axis, so is welldefined.
The reproducing kernel is closely related to the analytically continued (Wightman) 2point function for the scalar quantum field of
mass m (Streater and Wightman [1964]):
K(z 0 , z) = i+ (z 0 z).
(41)
We will encounter this and other 2point functions again in the next
chapter, in connection with quantum field theory.
Note: We will be interested in the behavior of k ez k near the boundary of T+ , i.e. when 0. From the properties of K it follows
that
kez k2
()
2 when 0.
(4)+1
(42)
118
In particular, the evaluation maps are no longer bounded when = 0.

P0 acts on T+ by a complex extension of its action on real spacetime, i.e.,
z 0 (b, ) z = z + b.
(43)
This means that x0 = x+b as before, and y 0 = y. (This is consistent

with the phasespace interpretation of T+ to be established below.)
The induced action on K is therefore

(U (b, )f ) (z) = f 1 (z b) .
(44)
This implies that the wave packets ez transform covariantly under P0 ,
i.e.
U (b, ) ez = ez+b .
(45)
We have now established that the space K of holomorphic positiveenergy solutions is a reproducingkernel Hilbert space. Recall
that picking out the positiveenergy part of f (x) was a nonlocal operation
in real spacetime, involving the pseudodifferential operator
2
m . However, when extended to T+ , such functions may be
characterized locally, as simultaneaous solutions of the KleinGordon
equation and the CauchyRiemann equations, since the negative
energy part of f (x) does not have an analytic continuation to T+ .
We now show that T+ may, in fact, be interpreted as an extended
phase space for the underlying classical relativistic particles. Clearly,
x < z are the spacetime coordinates. Their relation to the expectation values of the relativistic (NewtonWigner) position operators
will be discussed below. We now wish to investigate the relation
of the imaginary coordinates y = z to the energymomentum
vector. The bridge between the classical coordinates y and the
quantummechanical observables P will be, as usual, the (future)
coherent states ez . Before getting involved in computations, let us
take a closer look at these wave packets in order to get a qualitative picture. Since yp is Lorentzinvariant, it can be evaluated in a
reference frame where p = (mc2 , 0). Thus
p
yp = y 0 mc2 = 2 + y2 mc mc,
(46)
with equality if and only if y = 0, i.e. when y and p are parallel. This
is a kind of reverse Schwarz inequality which holds in V+0 V + under
119
the pairing provided by the Lorentzian scalar product. Thus for fixed
y V+0 and variable p +
m , we have
| ez (p) | emc ,
(47)
the maximum being attained when and only when p = (mc/)y py .

Hence we expect, roughly, that
h P i
h ez | P ez i
(mc/)y .
h ez | ez i
(48)
Therefore the vector y, while itself not an energymomentum, acts as

a control vector for the energymomentum by filtering out ps which
are far from py . The larger we take the parameter , the finer the
filter. The expected energymomentum can be computed exactly by
noting that
Z
h ez | P ez i =
d
p p e2yp
+
m
(49)
1 G(y)
=
2 y
where G(y) = kez k2 as before. Since G depends on y only through
the invariant quantity , we have
1
G
h P i = G1
2
y

K+1 (2mc) mc
=
y ,
K (2mc)
(50)
where we have used the recurrence relation (Abramowitz and Stegun

[1964])
K (2m) = 2m K+1 (2m).
(51)
This verifies and corrects the above qualitative estimate. In view of

the above relation, the hyperboloid
0
2
2
+
{y V+ | y = }
corresponds to the mass shell. Hence the submanifold
(52)
120
C = {x iy T+ | y 2 = 2 }
(53)
corresponds to the extended phase space C defined in section 4.2

which, we recall, was a homogeneous space of P0 that was interpreted
as spacetimevelocity space. Let us define the effective mass m of
the particle on +
by
(m c)2 h P ih P i h P i2 .
(54)
Then
m = m
K+1 (2mc)
,
K (2mc)
(55)
and
h P i =
m c
y .
(56)
We claim that m > m, which can be seen as follows. For all p, p0

0
2
+
m we have the reverse Schwarz inequality pp m , with equality
0
if and only if p = p . Hence
m2 = h P i2
ZZ
2
=G
d
p d
p0 pp0 exp(2yp 2yp0 )
(57)
> m2 .
This is a kind of renormalization effect due to the uncertainty, or
fluctuation, of the energymomentum in the state ez . It appears to go
in the wrong direction (i.e., h P i2 > h P 2 i) for the same reason as
does the inequality pp0 m2 , namely because of the Lorentz metric.
Thus h P i is proportional, by a ydependent but P0 invariant
factor, to y . We may therefore consider the y s as homogeneous
coodinates for the direction of motion of the classical particle in (real)
spacetime. Alternatively, the expectation of the velocity operator
P/P0 can be shown to be y/y0 . Thus of the s + 1 coordinates y ,
only s have a classical interpretation. It is important to understand
that the parameter has no relation to the mass; it can be chosen to
be an arbitrary positive number and has the physical dimensions of
length. It is the relativistic counterpart of the parameter u encountered in connection with the nonrelativistic coherent states, and its
121
significance will be studied later. At this point we simply note that

measures the invariant distance of z from the boundary of T+ . The
larger , the more smeared out are the spatial features and the more
refined are the features in momentum space. (Recall that the imaginary part u of the time played a similar role in the nonrelativistic
theory.)
Because the vector y is so fundamental to our approach, it deserves
a name of its own. We will call it the temper vector. The name is
motivated in part by the smoothing effect which y has on spacetime
quantities, and also by the fact that y plays a role similar to that
played by the inverse teperature = 1/kT in statistical mechanics;
the latter controls the energy.
From the asymptotic properties of the K s it follows that

h P i
(mc/)y , when mc ;
(/2 )y , when mc 0.
(58)
This can be understood as follows: When mc , e.g. c

for fixed m, we recover the nonrelativistic results. For example,
the expectations of the spatial momenta approach those in the non
relativistic coherent states:
h Pk i (mc/)yk (m/u)yk .
(59)
When mc 0, say 0 for fixed mc, then z approaches the boundary of T+ . In that case, fluctuations take over and the expectations
become independent of the mass m.
The relation of the spacetime parameters x <z to the
NewtonWigner position operators is as follows. Since a fixed f T+
describes the entire history of the particle, the associated state does
not change with time (i.e., the dynamics is already built in). This
means that we are in the socalled Heisenberg picture, and time
behavior must be described by evolving the observables A:
A(t) = eitP0 A(0) eitP0 .
(60)
The NewtonWigner position operators are uniquely determined by a

set of seemingly reasonable assumptions concerning the localizability
of the particle (Newton and Wigner [1949], Wightman [1962]), and
are given in the momentum representation at time x0 = 0 by
122
1/2
Xk (0) = 1/2 i
pk

pk
2 ,
=i
pk
(61)
k = 1, 2, . . . , s.
Now choose z = x iy T+ with x0 = 0. Then

Z
1/2 izp
e
pk
+
m
Z
1/2 izp
e
=<
d
p 1/2 eizp i
p
k
+
m
h ez | Xk (0) ez i =
d
p 1/2 eizp i
(62)
= xk kez k2 ,
hence
h Xk (0) i = xk .
(63)
It must be noted, however, that the concept of position for relativistic

particles is highly unsatisfactory. Not only are the position operators
noncovariant (this would seem to require a time operator on equal
footing with them, which would exclude the possibility of dynamics);
but the very concept of localizability for such particles is fraught with
difficulties. For example, eigenvectors of the NewtonWigner position
operators, known as localized states, spread out from a single point
at time x0 to fill the entire universe an arbitrarily small time later,
violating relativistic causality. (The same phenomenon in the non
relativistic theory presents no conceptual problem, since propagation
velocities are unrestricted there.) Even much weaker notions of localization give rise to problems with causality (Hegerfeldt [1985]). In my
opinion, it is best to admit that position is simply a nonrelativistic
concept, and in the relativistic theory events x should be regarded as
mere parameters of the spacetime manifold. As such, our formalism
extends them to z = x iy, with the new parameters y playing the
role of a control vector for the energymomentum observables. Thus,
in place of a set of pairs of canonically conjugate observables Xk , Pk ,
we have a set of observables P and a dual set of complex parameters
z . The symmetry between position and momentum operators in
the nonrelativistic theory was based on the WeylHeisenberg group,
123
and we have seen that this symmetry is accidental, being broken in

the transition to relativity.
Note: This further reinforces the idea discussed in section 4.2, that
grouptheoretic quantization, i.e. the quantization of classical systems obtained by requiring states and observables to transform under
unitary representations of the associated dynamical groups or central extensions thereof, is superior to canonical quantization (sending the classical phasespace observables to operators, with Poisson
brackets going to commutators). #
Let us now consider the uncertainties in the energy and momentum operators. More generally, we compute the correlation matrix
C h P P i h P ih P i,
(64)
of which the uncertainties 2P are the diagonal elements. From

h ez | P P ez i =
1 2G
4 y y
(65)
it follows that
2G
G G
G2
y y
y y
2 ln G
.
=
y y
4C = G1
(66)
(Incidentally, this shows that the function ln G is of some interest in

itself: its first partials are the expected momenta, while its second
partials are the correlations.) A computation similar to that for h P i
gives

m2 K+2 (2m) m2
m
C = y y 2
2 g
.
(67)
K (2m)
m
2
Although this expression does not appear to be enlightening in any
obvious way, the uncertainties P can be estimated from it in various
limits such as m and m 0.
The reproducing kernel by itself is of limited use. Although it
makes it possible to establish the interpretation of T+ as an extended
classical phase space, it does not provide us with a direct physical
interpretation of the function values f (z). The inner product in K is
borrowed from L2+ (d
p), hence a probability interpretation exists, so
124
far, only in momentum space. In the standard formulation of Klein

Gordon theory, it is possible to define the inner product in configuration space, but the corresponding density turns out not to be positive, thus precluding a probabilistic interpretation. This is one of the
wellknown difficulties with the firstquantized KleinGordon theory,
and is one of the reasons cited for the necessity to go to quantum
field theory (second quantization). We will see that the phasespace
approach does admit a covariant probabilistic picture of relativistic
quantum mechanics, thus making the theory more complete even before second quantization. These topics will be discussed further in
the next section and the next chapter. At this point we wish only
to define an autonomous inner product on K as an integral over a
phase space lying in T+ . This will provide us with a normal frame
of evaluation maps (chapter 1).
Recall that in the nonrelativistic theory of the last section, the
norm in the space Hu of holomorphic solutions was obtained by integrating | f (z, ) | 2 with respect to the Gaussian measure du (z) over
the phase space t iu = constant. But now, for given y 0 = u,
only those ys with | y | < cu belong to V+0 . That is, the particle can
only travel a finite imaginary distance in a finite imaginary time. In
view of the relation h P i y , the obvious candidate for a phase
space is the set defined for given t IR and > 0 by
t, = {x iy T+ | x0 = t, y 2 = 2 }.
(68)
Such sets are not covariant, but a covariant extension will be found in
the next section. As for the measure, a Gaussian weight function (such
as exp(my2 /u), which occured in du (y)) is no longer satisfactory
since it cannot be covariant. It turns out that we do not need a
weight function at all! This can be seen as follows: In the non
relativistic case, the shift to complex time was performed once and
for all by the operator euH . For fixed u > 0, the weight function
served to correct for the nonunitary translation from the real point
x in space to the complex point z = x iy. However, if we restrict
ourselves to the subset , then a translation to imaginary space is
necessarily accompanied by a translation in imaginary
time. The
p
2
analog of the above translation is (ti, x) 7 (ti + y2 , xiy).
The increase in y 0 , it turns out, precisely compensates for the shift
to complex space! This follows from the fact that the operator eyP ,
which affects the total shift to complex spacetime, is relativistically
invariant, hence the point y = 0 no longer plays a special role. We
will show later that in the nonrelativistic limit, we recover the weight
125
function naturally. Hence the Gaussian weight function associated

with the Galilean coherent states (which, as we have seen, is closely
related to that associated with the canonical coherent states) has its
origin in the geometry of the relativistic phase space, i.e. in the
curvature of the hyperboloid {y 2 = 2 }.
For t, as above and f K, define
Z
2
kf k =
d | f (z) | 2 ,
(69)
where we parametrize by (x, y) IR2s and the measure d is given

by
s
s
d(z) = A1
d xd y
(70)
with
1
A =

2
=
mc
mc
s
+1
K+1 (2mc)
m
G().
m
(71)
Then we have the folowing result.

Theorem 4.3. For all f K,
kf k = kf kK .
(72)
Proof. Assume, to begin with, that a(p) a(, p) belongs to the

Schwartz space S(IRs ) of rapidly decreasing test functions. Then
Z
s
f (z) = (2)
ds p (2)1 exp(ixp yp) a(p)
(73)

1
= (2) exp(it yp) a (x).
Hence, by Plancherels theorem,
Z
d x | f (x iy) | = (2)
x0 =t
ds p (4 2 )1 e2yp | a(p) | 2 .
IRs
(74)
Exchanging the order of integration in the integral representing
kf k2 , we obtain
126
kf k2
s
A1
(2)
2 1
d p (4 )
| a(p) |
ds ye2yp .
(75)
We now evaluate
Z
J(p)
ds y e2yp
(76)
as follows: Consider
all s + 1 components of p as independent and
p
2
define m(p) p . From the integral computed earlier, i.e.
Z
s
G(y) (2)
ds p (2)1 e2yp
(77)
m
1
= (2)
K (2m),
4
we obtain by exchanging p and y (as well as m and ):

Z
s
1 2yp
d y (2y0 ) e
=
K (2m).
m
(78)
Taking the partial derivative with respect to p0 on both sides gives

K (2m)
J(p) = 0
p
m

(79)
= (22 ) 0 K ()
p

= (22 ) 0
K () ,
p
where (p) 2m(p). Using again the recurrence relation for the
K s, we get
J(p) = 2p0 A .
(80)
Thus
kf k2
= (2)
ds p (2)1 | a(p) | 2 = kf k2K
(81)
127
for a(p) S(IRs ). By continuity, this extends to all a L2+ (d

p) since
S(IRs ) is dense in L2+ (d
p).
If we define
Z
h f1 | f2 i
d(z) f1 (z)f2 (z),
(82)
then by polarization we have the following immediate consequence of

the theorem.
Corollary 4.4. For all f1 , f2 K,
h f1 | f2 i = h f1 | f2 iK .
(83)
In particular, h | i defines a P0 invariant inner product on K, and

we have the resolution of unity
Z
d (z) | ez ih ez | = I,
(84)
making the vectors ez with z a normal frame.

Note: The above results extend to the case = 0 by continuity. The
normalization constant in the measure d is then given by
A0 lim A =
0
( + 1)
.
2(mc)s+1
(85)
The same formula applies for fixed > 0 and m 0, and shows that
A becomes unbounded as m 0. #
The vectors ez belong to L2+ (d
p), but correspond to vectors ez in
K defined by
ez (z 0 ) = h ez0 | ez i = K(z 0 z).
(86)
The norm kf k2 h f | f i on K provides us with an interpretation

of | f (z) | 2 as a probability density with respect to the measure d
on the phase space . Within this interpretation, the wave packets ez
have the following optimality property: For fixed z T+ let
ez (z 0 ) =
h ez0 | ez i
.
kez k
(87)
Proposition 4.5. Up to a constant phase factor, the function ez is

the unique solution to the following variational problem: Find f K
such that kf k = 1 and | f (z) | is a maximum.
128
Proof. This follows at once from the Schwarz inequality and theorem
4.3, since by eq. (26),
| f (z) | = | h ez | a i | = | h ez | f i |
k
ez k kf k = kez k kf k,
(88)
with equality if and only if f is a constant multiple of ez .

According to our probability interpretation of | f (z) | 2 , this means
that the normalized wave packet ez maximizes the probability of finding the particle at z.
Note: Unlike the nonrelativistic coherent states of the last section,
the ez s do not have minimum uncertainty products. In fact, since the
uncertainty product is not a Lorentzinvariant notion, it is a priori
impossible to have relativistic coherent states with minimum uncertainty products. The above optimality, which is invariant, may be
regarded as a reasonable substitute. Actually, there are better ways
to measure uncertainty than the standard one used in quantum mechanics, which is just the variance. From a statistical point of view,
the variance is just the second moment of the probability distribution. Perhaps the best definition of uncertainty, which includes all
moments, is in terms of entropy (BialynickiBirula and Mycielski
[1975], Zakai [1960]). Being necessarily nonlinear, however, makes
this definition less tractable.
4.5. Geometry and Probability

The formalism of the last section was based on the phase space t,
and the measure d, neither of which is invariant under the action of
P0 on T+ . Yet, the resulting inner product h | i is clearly invariant.
It is therefore reasonable to expect that and d merely represent
one choice out of many. Our purpose here is to construct a large natural class of such phase spaces and associated measures to which our
previous results can be extended. This class will include and will
be invariant under P0 . In this way our formalism is freed from its dependence on and becomes manifestly covariant. As a byproduct, we
find that positiveenergy solutions of the KleinGordon equation give
rise to a conserved probability current, so the probabilistic interpretation becomes entirely compatible with the spacetime geometry. As
is wellknown, no such compatibility is possible in the usual approach
to KleinGordon theory.
129
We begin by regarding T+ as an extended phase space (symplectic

manifold) on which P0 acts by canonical transformations. Candidates
for phase space are 2sdimensional symplectic submanifolds T+ ,
and P0 maps different s into one another by canonical transformations. A submanifold of the product form = S i+
, where S
(interpreted as a generalized configuration space) is an ssubmanifold
of (real) spacetime IRs+1 , turns out to be symplectic if and only if S is
given by x0 = t (x) with |t| 1, that is, if and only if S is nowhere
timelike. (This is slightly larger than the class of all spacelike configuration spaces admissible in the standard theory.) The original t,
corresponds to t (x) constant. The results of the last section are
extended to all such phase spaces of product form.
The action of P0 on T+ is not transitive but leaves each of the
(2s + 1)dimensional submanifolds
T+ {x iy T+ | y 2 = 2 }
(1)
invariant. Each T+ is a homogeneous space of P0 , with isotropy

subgroup SO(s), hence
T+ P0 /SO(s).
(2)
Thus T+ corresponds to the homogeneous space C of section 4.2

(where we had specialized to s = 3). In view of the considerations
in sections 4.2 and 4.4, each T+ can be interpreted as the product of
spacetime with momentum space. Phase spaces will be obtained
by taking slices to eliminate the time variable.
On the other hand, we also need a covariantly assigned measure
for each . The most natural way this can be accomplished is to begin
with a single P0 invariant symplectic form on T+ and require that its
restriction to each be symplectic. This will make each a symplectic
manifold (which, in any case, it must be to be interpreted as a classical
phase space) and thus provide it with a canonical (Liouville) measure.
Thus we look for the most general 2form on T+ such that
(a) is closed, i.e., d = 0;
(b) is nondegenerate, i.e., the (2s + 2)form s+1
vanishes nowhere;
(c) for every g = (a, ) P0 , g = , where g denotes the pull
back of under g (see Abraham and Marsden [1978]). Since every
P0 invariant function on T+ depends on z only through y 2 , the
most general invariant 2form is given by
130
= (y 2 ) dy dx + (y 2 ) y y dy dx .
(3)
Now the restriction (pullback) of the second term to T+ vanishes,

since it contains the factor y dy = d(y 2 )/2. Furthermore, the coefficient (y 2 ) of the first term is constant on T+ . Hence we may confine
our attention to the form
= dy dx
(4)
without any essential loss of generality. This form is symplectic as

well as invariant, hence it fulfills all of the above conditions. T+ ,
together with , is a symplectic manifold, and invariance means that
each g P0 maps T+ into itself by a canonical transformation.
A general 2sdimensional submanifold of T+ will be a potential
phase space only if the restriction, or pullback, of to is a symplectic
form. We denote this restriction by . Let be given by
= {z T+ | s(z) = h(z) = 0},
(5)
where s(z) and h(z) are two realvalued, C (or at least C 1 ) functions
on T+ such that ds dh 6= 0 on . For example, t, can be obtained
from s(z) = x0 t and h(z) = y 2 2 . The pullback depends only
on the submanifold , not on the particular choice of s and h.
Proposition 4.6.
The form is symplectic if and only if the
Poisson bracket
{s, h}
s h
h s

6= 0
x y
x y
(6)
everywhere on .
Proof. is closed since is closed and d( ) = (d) . Hence
is symplectic if and only if it is nondegenerate, i.e. if and only if its
s-th exterior power s vanishes nowhere on . Now ( )s equals the
pullback of s to , and a straightforward computation gives
c dx
c,
s = s! dy
(7)
where
c = (1)s dy0 dy1 dy1 dy+1 dys
dy
c = (1)s dxs dxs1 dx+1 dx1 dx0 .
dx
(8)
131
c and dx
c are essentially the Hodge duals (Warner [1971]) of dy
(dy
and dx , respectively, with respect to the Minkowski metric.)

Let {u1 , . . . , u2s , v1 , v2 } be a basis for the tangent space of T+ at z ,
with {u1 , . . . , u2s } a basis for the tangent space z of . Since ds and
dh vanish on the vectors uj ,
(s ds dh)(u1 , . . . , u2s , v1 , v2 )
= s (u1 , . . . , u2s ) (ds dh)(v1 , v2 )
(9)
= s (u1 , . . . , u2s ) (ds dh)(v1 , v2 ).

By assumption, (ds dh)(v1 , v2 ) 6= 0. Therefore is nondegenerate
at z if and only if s ds dh 6= 0 at z. But by eq. (7),
s ds dh = s! {s, h} dy dx,
(10)
where
dy = dy0 dys
dx = dxs dx0 .
(11)
Hence s 6= 0 at z if and only if {s, h} =

6 0 at z.
Let us denote the family of all such symplectic submanifolds by
0 .
Proposition 4.7. Let 0 and g P0 . Then g 0 and the
restriction g: g is a canonical transformation from (, ) to
(g, g ).
Proof. Let g denote the pullback map defined by g, taking forms
on g to forms on . Then the invariance of implies that
g g = .
(12)
Hence g is nondegenerate. It is automatically closed since is

closed. Thus g 0 . To say that g: g is a canonical transformation means precisely that and g are related as above.
We will be interested mainly in the special case where h(z) =
2
y 2 for some > 0 and s(z) depends only on x. Then the s
dimensional manifold
S {x IRs+1 | s(x) = 0}
(13)
is a potential generalized configuration space, and has the product

form
132
+
= S i+
{x iy T+ | x S, y }.
(14)
The following result is physically significant in that it relates the

pseudoEuclidean geometry of spacetime and the symplectic geometry of classical phase space. It says that is a phase space if and
only if S is a (generalized) configuration space.
Theorem 4.8.
Let = S i+
be as above. Then (, ) is
symplectic if and only if
s s
0,
x x
(15)
that is, if and only if S is nowhere timelike.

Proof. On , we have
{s, h} = 2
s
y 6= 0,
x
(16)
and we may assume {s, h} > 0 without loss. For fixed x S, the
0
above inequality must hold for all y +
, hence for all y V+ . This
0
implies that the vector s/x is in the dual V + of V+ .
We denote the family of all s as above (i.e., with S nowhere
timelike) by . It is a subfamily of 0 and is clearly invariant under
P0 . Note that admits lightlike as well as spacelike configuration
spaces, whereas the standard theory only allows spacelike ones.
We will now generalize the results of the last section to all .
The 2sform s defines a positive measure on , once we choose an
orientation (Warner [1971]) for . (This can be done, for example, by
choosing an ordered set of vector fields on which span the tangent
space at each point; the order of such a basis is a generalization of
the idea of a righthanded coordinate system in three dimensions.)
The appropriate measure generalizing d of the last section is now
defined as
d = (s! A )1 s .
(17)
d is the restriction to of a 2sform defined on all of T+ , which we

also denote by d. (This is a mild abuse of notation; in particular,
the d here must not be confused with exterior differentiation!) By
eq. (7), we have
c c
d = A1
dy dx .
(18)
133
We now derive a concrete expression for d. Since s obeys eq. (15)

and ds 6= 0 on , we can solve ds = 0 (satisfied by the restriction
c k . This (and a similar
of ds to ) for dx0 and substitute this into dx
procedure for y) gives

s
x0
1
s c
dx0
x
h
y0
1
h c
dy
y
c =
dx
c
dy
(19)
c
= (y /y 0 ) dy
on . Hence
d =
A1
s
y
x0
0
1
s
y
x

c
dy
c0.
dx
(20)
We identify with IR2s by solving s(x) = 0 for x0 = t (x) and

mapping
p
(t (x) i 2 + y2 , x iy) 7 (x, y).
(21)
c 0 dx
c 0 with the Lebesgue measure ds y ds x on
We further identify dy
2s
IR (this amounts to choosing a nonstandard orientation of IR2s ).
Thus we obtain an expression for d as a measure on IR2s . Now
s(x) = 0 on implies that
s (t (x), x)
xk
s t
s
=
+ k,
0
k
x x
x
0=
which can be substituted into the above expression to give

t y k
1
d = A
1 k 0 ds y ds x
x y

1
= A 1 t (y/y0 ) ds y ds x.
(22)
(23)
But eq. (15) implies that | t (x) | 1, hence for y V+0 ,

t (y/y0 ) < 1
(24)
134
and d is a positive measure as claimed. The above also shows that if

| t (x) | = 1 for some x, then d becomes asymptotically degenerate as | y | in the direction of t (x). That is, if is lightlike at
(t (x), x), then d becomes small as the velocity y/y0 approaches the
speed of light in the direction of t (x). This means that functions
in L2 (d) (and, in particular, as we shall see, in K) are allowed high
velocities in the direction t (x) at (t (x), x) S. This argument
is an example of the kind of microlocal analysis which is possible in
the phasespace formalism. (In the usual spacetime framework, one
cannot say anything about the velocity distribution of a function at a
given point in spacetime, since this would require taking the Fourier
transform and hence losing the spatial information.)
For , denote by L2 (d) the Hilbert space of all complex
valued, measurable functions on with
Z
2
kf k
d | f | 2 < .
(25)
If f is a C function on T+ , we restrict it to and define kf k as

above. Our goal is to show that kf k = kf kK for every f K. To
do this, we first prove that each f K defines a conserved current in
spacetime, which, by Stokes theorem, makes it possible to deform the
phase space t, of the last section to an arbitrary = S i+

without changing the norm. For f K, define
Z
1
c | f (x iy) | 2 ,
dy
(26)
J (x) = A
+
0
c0
where +
has the orientation defined by dy , so that J (x) is positive.
Then
Z
2
c J (x),
kf k =
dx
(27)
S
c 0 . (The restriction of dx
c0
where S has the orientation defined by dx
to S does not vanish since | t (x) | 1.)
Theorem 4.9. Let f(p) be C with compact support. Then J (x)
is C and satisfies the continuity equation
J
= 0.
x
Proof. By eq. (19),
(28)
J (x) =
A1
d
y y | f (x iy) | 2 ,
135
(29)
c 0 /y 0 . The function
where d
y dy
Fx (y, p, q) y exp [ix(p q) y(p + q)] f(p) f(q)
(30)
is in L1 (d
y d
p d
q ), hence by Fubinis theorem,
J (x) =
=
A1
A1
Z
d
y
+
+
m m
d
p d
q Fx (y, p, q)
d
p d
q exp [ix(p q)] f(p) f(q) H (p + q),
+
+
m m
(31)
where, setting k p + q,
and using the recurrence relation
for the K s given by eq. (51) in section 4.4, we compute
Z
H (k)
d
y y eyk
k2
=
k
d
y eyk

2
2
K ()
=
k

+1
2
= (k /)
K+1 ()
(32)
k H().
H() is a bounded, continuous function of for 2m, and
J (x) =
A1
d
p d
q exp [ix(p q)] f(p) f(q) (p + q ) H().
+
+
m m
(33)
Since f (p) has compact support, differentiation under the integral

sign to any order in x gives an absolutely convergent integral, proving
that J is C . Differentiation with respect to x brings down the
136
factor i(p q ) from the exponent, hence the continuity equation

follows from p2 = q 2 = m2 .
Remark. The continuity equation also follows from a more intuitive,
geometric argument. Let
B+ = {y V+0 | y 0 >
2 + y2 },
(34)
oriented such that

+
+
= B
(35)
(the outward normal on B+ points down, whereas +

is oriented
up). Then by Stokes theorem,
J (x) =
=
A1
Z
+
A1
c
dy
| f (x iy) | 2
+
B
c
d dy
| f (x iy) |
(36)
.
Here, d represents exterior differentiation with respect to y, and since

c contains all the dy s except for dy , we have
the sform dy
J (x) =
A1
Z
dy
+
B
| f (x iy) | 2 ,
y
(37)
where dy is Lebesgue measure on B+ . To justify the use of Stokes

theorem, it must be shown that the contribution from | y |
to the first integral vanishes. This depends on the behavior of f (z),
which is why we have given the previous analytic proof using the
Fourier transform. Then the continuity equation is obtained by differentiating under the integral sign (which must also be justified) and
using
2 | f | 2
= 0,
x y
(38)
which follows from the KleinGordon equation combined with analyticity, since

2
=
x y
z
z

i
z
z
2
2
=i
i
z z
z z
137

(39)
i (t
u z u
tz ) .
Incidentally, this shows that
j (z)
| f (z) | 2
y
h
i
= i f (z) f (z) f (z) f (z)
(40)
is a microlocal, spacetimeconserved probability current for each

fixed y V+0 , so the scalar function | f (z) | 2 is a potential for the
probability current. We shall see that this is a general trend in the
holomorphic formalism: many vector and tensor quantities can be
derived from scalar potentials.
Eqs. (37) and (40) also show that our probability current is a
regularized version of the usual current associated with solutions in
real spacetime. The latter (Itzykson and Zuber [1980]) is given by
h
i
Jusual
(x) = i f (x) f (x) f (x) f (x) ,
which leads to a conceptual problem since the time component, which
should serve as a probability density, can become negative even for
positiveenergy solutions (Gerlach, Gomes and Petzold [1967], Barut
and Malin [1968]). By contrast, eq. (36) shows that J 0 (x) is stricly
nonnegative. The tendency of quantities in complex spacetime to
give regularizations of their counterparts in real spacetime is further
discussed in chapter 5.
We can now prove the main result of this section.
Theorem 4.10. Let = S i+
and f K . Then kf k =
kf kK .
Proof. We will prove the theorem for f(p) in the space D(IRs ) of
C functions with compact support, which implies it for arbitrary
f L2+ (d
p) by continuity. Let S be given by x0 = t (x), and for
R > 0 let
138

DR = {x IRs+1 |x| < R, x0

ER = {x IRs+1 |x| = R, x0

S0R = {x IRs+1 |x| < R, x0

SR = {x IRs+1 |x| < R, x0
[0, t (x)] },
[0, t (x)] },
= 0},
(41)
= t (x)},
where [0, t (x)] means [t (x), 0] if t (x) < 0. We orient S0R and SR by
c 0 , ER by the outward normal
dx
r = R1
s
X
ck ,
xk dx
(42)
k=1
and DR so that DR = SR S0R + ER . Now let f(p) D(IRs ). Then

J (x) is C , hence by Stokes theorem,
Z
Z

c
c
J (x) dx =
d J dx
SR S0R +ER
DR
Z
(43)
J
s
= (1)
dx = 0.
x
DR
We will show that
Z
c 0 as R
J dx
(R)
(44)
ER
(i.e., there is no leakage to |x| ), which implies that

Z
2
c
kf k lim
J dx
R S
R
Z
c
= lim
J dx
R
kf k20
(45)
S0R
= kf k2K
by theorem 1 of section 4.4. To prove that (R) 0, note that on

c 0 = 0 and
ER , dx
c
c
c
c k = xk dx1 = xk dx2 = = xk dxs ,
dx
x1
x2
xs
(46)
139
each form being defined except on a set of measure zero; hence

c1
dx
.
x1
r = R
(47)
By eq. (29),
|J k (x)| J 0 (x),
(48)
hence
s

X
(R) =
k=1 ER
s Z
X
k=1

ck
J k dx
J k xk
ER
c1
dx

x1
(49)
c1
dx
s
J0 R
x1
E
Z R
=s
J 0 r a(R).
ER
Now by eqs. (31) and (32),

Z
0
J (x) =
ds p ds q eix(pq) (p, q),
(50)
IR2s
where
(p, q) = f(p) f(q) (p, q),
(51)
and C (IR2s ). Hence D(IR2s ). Let

p ,
D=x
(52)
= x/R, and observe that for x ER ,

where x

x0
ixp
v eixp
De = iR 1 x
R
ixp
iR (x, p) e
(53)
where v = p/p0 . Since has compact support, there exists a constant

< 1 such that |v| and |v0 | for all (p, p0 ) in the support
140
of . Furthermore, since |t (x)| 1, given any > 0 we have

|x0 | < R(1 + ) for ER for R sufficiently large; hence
|(x, p)| 1 (1 + ) for x ER and p supp .
(54)
Choose 0 < < 1 1, substitute

eixp =
i
D eixp ,
R (x, p)
x ER
(55)
into the expression for J 0 (x) and integrate by parts:

Z
(p, q)
0
1
s
s
ix(pq)
J (x) = (iR)
d pd q e
D
(x, p)
IR2s
Z
(iR)1
ds pds q eix(pq) 0x (p, q).
(56)
IR2s
This procedure can be continued, giving (for x ER )
J (x) = (iR)
IR2s
ds pds q eix(pq) (n)

x (p, q),
n = 1, 2, ,
(57)
where

1 n
(n)
(p, q)
x (p, q) = D
"

1 #n
x0
p 1
v
= x
x
(p, q).
R
(58)
n
Now D 1 is a partial differential operator in p whose coefficients are polynomials in Dk ( 1 ) with k = 0, 1, , n. We will show
that for x ER with R sufficiently large, there are constants bk such
that
k 1
D ( )< bk ,
k = 0, 1, ,
(59)
which implies that
k(n)
x kL1 (IR2s ) < cn ,
x ER , n = 1, 2,
(60)
141
for some constants cn , so that by eqs (49) and (57),

Z
a(R) = s
J 0 r
ER
Z
n
= sR
k(n)
r(x)
x kL1 (IR2s )
ER
sR
2 s/2 s1
R
cn
(s/2)
R(1+)
dx0
(61)
s/2
2s
cn Rsn (1 + )
(s/2)
0 as R
if we choose n > s. To prove eq. (59), note that it holds for k = 0 by

v. Then
eq. (54) and let u = x
Du (
x p ) (
x p/p0 ) =
1 u2
,
p0
(62)
and if for some k

Dk u =
Pk (u)
pk0
(63)
where Pk is a constantcoefficient polynomial, then

Pk0 (u)Du kPk (u)
k+1
pk0
p0
Pk+1 (u)
,
pk+1
0
Dk+1 u =
(64)
hence eq. (63) holds for k = 1, 2, by induction. Thus

Dk =
x0 k
x0 Pk (u)
D u=
,
R
R pk0
k = 1, 2, ,
(65)
which implies
k 1+

D
Pk (u).
max
mk |u|1
(66)
142
But Dk ( 1 ) is a polynomial in 1 and D, D2 , , Dk ; hence eq.

(59) follows from eqs. (54) and (66).
The following is an immediate consequence of the above theorem.
Corollary 4.11.
(a) For every , the form
Z
h f1 | f2 i =
d f1 (z)f2 (z)
(67)
defines a P0 invariant inner product on K, under which K is a

Hilbert space.
(b) The transformations (Ug f )(z) f (g 1 z), g P0 , form a unitary
irreducible representation of P0 under the above inner product,
and the map f 7 f from L2+ (d
p) to K intertwines this represen2
tation with the usual one on L+ (d
p).
(c) For each , we have the resolution of unity
Z
d | ez ih ez | = I
(68)
on L2+ (d
p) (or, equivalently, on K if ez is replaced by ez . )
Note: As in section 4.4, all the above results extend by continuity to
the case = 0. #
4.6. The NonRelativistic Limit

We now show that in the nonrelativistic limit c , the foregoing
coherentstate representation of P0 reduces to the representation of
G2 derived in section 4.3, in a certain sense to be made precise. As a
byproduct, we discover that the Gaussian weight function associated
with the latter representation (hence also the closely related weight
function associated with the canonical coherent states) has its origin
in the geometry of the relativistic (dual) momentum space +
.
That is,
p for large |y| the solutions in K are dampened by the factor
exp[ 2 + y2 ] in momentum space, which in the nonrelativistic
limit amounts to having a Gaussian weight function in phase space.
In considering the nonrelativistic limit, we make all dependence
on c explicit but set h = 1. Also, it is convenient to choose a coordinate system in which the spacetime metric is g = diag(1, 1, . . . , 1),
143
p
p
so that y 0 = y0 = 2 + y2 and p0 = p0 = m2 c2 + p2 . Fix u > 0
and let = uc. Then
p
p
y0 = u2 c2 + y2 m2 c2 + p2
(1)
up2
my2
= umc2 +
+
+ O(c2 ).
2m
2u
Working heuristically at first, we expect that for large c, holomorphic
solutions of the KleinGordon equation can be approximated by
Z
f (x iy)
+
m
d
p exp [it + ix p y0 + y p] f(p)

ds p
p2
2
exp it mc +
+ ix p
(2)2 2mc
2m
(2)

2
2
up
my
exp umc2
+ y p f(p)
2m
2u

(2mc)1 exp i mc2 my2 /2u fNR (x iy, ),
Z
where = t iu and fNR is the corresponding holomorphic solution of

the Schr
odinger equation defined in section 4.3. Note that the Gaussian factor exp[my2 /2u] is the square root of the weight function
for the Galilean coherent states, hence if we choose f(p) L2 (IRs )
L2+ (d
p), then
kemy
/2u
fNR k2L2 (Cl s ) = (u/m) kfNR k2Hu < .
(3)
We now rigorously justify the above heuristic argument. Let f (z)

be the function in K corresponding to f(p) and denote by fc its
restriction to x0 = t and y 2 = u2 c2 , for fixed u > 0.
Theorem 4.12. Let u > 0 and f(p) L2 (IRs ). Then
2
J(c) k2mcei mc fc emy
/2u
fNR k2L2 (Cl s )
0 as c .
(4)
Proof. Without loss of generality, we set u = m = 1 and t = 0 to

simplify the notation. Note first of all that
2
k2cec fc k2L2 (Cl s ) = 4c2 e2c Ac kf k2K

2
= 4c2 e2c Ac kfk2L2 (dp)
,
+
(5)
144
where Ac A ( uc = c). But

Ac = K+1 (2c2 )
2
= (2c)1 s/2 e2c
(6)

1 + O(c2 )
and
1
|fk2L2 (dp)
(2)s |fk2L2 (IRs ) .
(2c)
(7)

2
k2cec fc k2L2 (Cl s ) (4)s/2 kfk2L2 (IRs ) 1 + O(c2 ) ,
(8)
Thus
showing that 2cec fc approaches a limit in L2 (Cls ) as c . Now

ZZ
h c 2
2
i

c yp
(py)2 /2
J(c) =
d xd y
e
e
f (x)
Z
Z
c 2
2
s
2
s
c yp
(py)2 /2
e
e
.
= d p|f (p)|
d y
(9)
Choose , such that 1/2 < < < 1. Then
d p|f(p)|
s
J1
|p|>c1
Z
IRs
ds y
c
ec
yp
e(py)
4 s/2 kc fk2L2 (IRs )
/2
2
(10)
0 as c ,

where c is the indicator function of the set {p |p| > c1 }. Define
and by |y| = c sinh and |p| = c sinh . Then y 0 = cosh and
= c2 cosh , hence

yp c2 cosh( ) c2 1 + ( )2 /2 .
Thus for arbitrary a 0,
(11)
|y|>c cosh a
s s/2 Z
2c
(s/2)
a
1s s s/.2
2yp
ds y e2c
Ga (p)
c
(s/2)
21s cs s/.2
(s/2)
d sinhs1 cosh ec
145
()2

2
2
d e(s1) e + e ec ()
(12)
d esc
()2
21s cs s/.2 s+s2 /4c2

e
=
(s/2)
du eu .
c(a)s/2c
Let a = sinh1 (c ). Then for |p| < c1 ,

c(a ) s/2c c sinh1 (c ) sinh1 (c ) s/2c g(c). (13)
g(c) is independent of p and g(c) c1 as c . Also, < c
when |p| < c1 . Hence
Z
d p|f(p)|2
s
J2
|p|<c1
ds y e2c
2yp
|y|>c1
22s cs1 s/2 sc +s2 /4c2
e
(s/2)
2
eu du kfk2L2 (IRs )
(14)
g(c)
0 as c .
Now
2c2 2yp = y 2 + p2 2yp = (y p)2
= (y0 )2 /c2 (y p)2 (y p)2 .
(15)
Hence
Z
|p|<c1
d p |f(p)|2
s
Z
|y|>c1
ds y e(yp) J2 ,
(16)
146
and
Z
d p |f(p)|2
s
d y
|p|<c1
|y|>c1
c
c2 yp
(py)2 /2
2
(17)
4J2 0 as c .
Finally,
Z
d p |f(p)|2
s
J3
|p|<c1
d y
c
|y|<c1
d p |f(p)|2
s
|p|<c1
c2 yp
ds y e(yp)
(yp)2 /2
c
|y|<c1
2 2
ec
/2
2
2
1 ,
(18)
where
p
p

= 1 + y2 /c2 1 + p2 /c2

1
2 y2 p2
2c

1 2
c
+ c2
2
c2 .
We have used the estimate
p
Z v xdx
p

2
2

1 + u 1 + v =
2
1+x
Zuv

1

xdx = v 2 u2 .
2
u
(19)
(20)
Hence for sufficiently large c and |p| < c1 ,

c
2 2
ec
/2
2
2c 2 2
+ 1 ec /2

2 2
2 2
1 + 2c + 1 2 1 p2 /2c2 ec /2

2 2
2 2
2 1 ec /2 + 2c2 2 + c2 ec /2

2c2 2 + c2 1 + c2 2
2 2
ec
h(c) 0 as c .
(21)
Notes
147
Thus
Z
J3 h(c)
d p |f(p)|2
s
|p|<c1
ds y e(yp)
|y|<c1
h(c) s/2 kfk2L2 (IRs )
(22)
0 as c ,
which proves that J(c) 0 as c .
Notes
This chapter represents the main body of the authors mathematics
thesis at the University of Toronto (Kaiser [1977c]). All the theorems,
corollaries, lemmas and propositions (labeled 4.1-4.12) have appeared
in the literature (Kaiser [1977b, 1978a]). In 1966, when the idea of
complex spacetime as a unification of spacetime and phase space first
occurred to me, I had found a kind of frame in which both the bras
and the kets were holomorphic in z and the resolution of unity was
obtained by a contour integral, using Cauchys theorem. During a
seminar I gave in 1971 at Carleton University in Ottawa (where I was
then a postdoctoral fellow in physics), L. Resnick pointed out to me
that this wavepacket representation appeared to be related to the
coherentstate representation, which was at that time unknown to
me. The kets were identical to the canonical coherent states, but the
bras were not their Riesz duals; in the language of chapter 1, they
belonged to a (generalized) frame reciprocal to that of the kets, and
the resolution of unity was of the type given by eq. (24) in section 1.3,
which may be called a continuous version of biorthogonality. A version
of this result was reported at a conference in Marseille (Kaiser [1974]).
I was later informed by J. R. Klauder that a similar representation had
been developed by Dirac in connection with quantum electrodynamics
(Dirac [1943, 1946]).
The original idea of complex spacetime as phase space was to consider a complex combination of the (symmetric) Lorentzian metric
with the (antisymmetric) symplectic structure of phase space, obtaining a hermitian metric on the complex spacetime parametrized by local coordinates of the type x+ibp. (I have since learned that this structure, augmented by some technical conditions, is known as a K
ahler
148
metric; see Wells [1980].) The above wavepacket representation

indicated that this combination may in fact be interesting, but so far
it was ad hoc and lacked a physical basis. Also, the representation was
nonrelativistic, and it was not at all clear how to extend it to the relativistic domain, as pointed out to me by V. Bargmann in 1975. The
standard method of arriving at canonical coherent states is to use an
integral transform with a Gaussian kernel in the configurationspace
representation, and there is no obvious relativistic candidate for such a
kernel. The more general methods described in chapter 3 do not work,
since the representations of interest are not squareintegrable (section
4.3). An important clue came in 1974 from the study of axiomatic
quantum field theory, where I was fascinated by the appearance of
tube domains. These domains occur in connection with the analytic
continuation of vacuum expectation values of products of fields, and
are therefore extensions of such products to complex spcetime. However, the complexified spacetimes themselves are not taken seriously
as possible arenas for physics. They are merely used to justify the
application of powerful methods from the theory of several complex
variables, in order to obtain results concerning the restrictions of vacuum expectaion values to real spacetime. (However, the restrictions
to Euclidean spacetime do have important consequences for statistical
mechanics; see Glimm and Jaffe [1981].) I felt that if these tube domains could somehow be given a physical interpretation as extended
classical phase spaces, this would give the phasespace formulation
of relativistic quantum mechanics a firm physical foundation, since in
quantum field theory the extension to complex spacetime is based on
solid physical principles such as the spectral condition. This idea was
first worked out at the level of nonrelativistic quantum mechanics,
leading to the representation of the Galilean group given in section
4.3. That amounted to a reformulation of the canonical coherent
state representation in which the Gaussian kernel appears naturally
in the momentum representation, as a result of the analytic continuation of solutions of the Schr
odinger equation. This explained the
combination x + ibp (section 4.3, eq. (5)) and gave the coherentstate
representation a dynamical significance. It also cleared the way to
the construction of relativistic coherent states, since now the Gaussian kernel merely had to be replaced with the analytic Fourier kernel
eizp on the mass shell. An important tool was the use of groups to
compute certain invariant integrals, which I learned from a lecture by
E. Stein on Hardy spaces in 1975. The construction of the relativistic
coherent states given in sections 4.4 and 4.5 was carried out in 1975
76, culminating in the 1977 thesis. Related results were announced at
Notes
149
a conference in 1976 (Kaiser [1977a]) and at two conferences in 1977

(Kaiser [1977d, 1978b]). To my knowledge, this was the first successful formulation of relativistic coherent states, which have since then
gained some popularity (see De Bievre [1989], Ali and Antoine [1989]).
An earlier attempt to formulate such states was made by Prugovecki
[1976], but this was shown to be inadequate since the proposed states
were merely the Gaussian canonical coherent states in disguise, hence
not covariant under the Poincare group (Kaiser [1977c], remark 4 in
sec. II.5 and addendum, p. 133.) After the results of the thesis appeared in the literature, Prugovecki [1978; see also 1984] discovered
that they can be generalized by replacing the invariant functions eyp
with arbitrary (sufficiently regular) invariant functions. The price of
this generalization is that solutions of the KleinGordon equation are
no longer represented by holomorphic functions and the close connection with quantum field theory (chapter 5) appears to be lost. The
relation between the two formalisms and their history was discussed
at a conference in Boulder in 1983 (Kaiser [1984b]), where an inconsistency in Prugoveckis formalism was also pointed out.
The classical limit of solutions of the KleinGordon equation in
the coherentstate representation was studied in Kaiser [1979]. In
an effort to understand interactions, the notion of holomorphic gauge
theory was introduced (Kaiser [1980a, 1981]). This is reviewed in
section 6.1. An early attempt was also made to extend the theory
to the framework of interacting quantum fields (Kaiser [1980b]), but
that was soon abandoned as unsatisfactory. A more promising approach was developed later (Kaiser [1987a]) and is presented in the
next chapter.
Note that our phase spaces are not unique, since the configuration space S can be chosen arbitrarily as long as it is nowhere timelike
and > 0 can be chosen arbitrarily. The freedom in S is, in fact, related to the probabilitycurrent conservation, while the freedom in
, combined with holomorphy, allowed us to express the probability
current as a regularization of the usual current by the use of Stokes
theorem (section 4.5, eqs. (37) and (40)). By contrast, the phase
spaces obtained by De Bievre [1989] are unique. They are coadjoint
orbits of the Poincare group, related to geometric quantization
theory (Kirillov [1976], Kostant [1970], Souriau [1970]). Although
this uniqueness seems attractive, it involves a high cost: the dynamics must be factored out. This means that the ensuing theory is no
longer local in time. Since one of the attractions of coherentstate
representations is their pseudolocality in both space and momen-
150
tum, and since in a relativistic theory time ought to be treated like

space, it seems to me an advantage to retain time in the theory. Perhaps a more persuasive argument for this comes from holomorphic
gauge theory (section 6.1), where a theory describing a free particle
can, in principle, be perturbed by introducing a nontrivial fiber
metric to obtain a theory describing a particle in an electromagnetic
(or YangMills) field. This cannot be done in a natural way once time
has been factored out.
Some very interesting work done recently by Unterberger [1988]
uses coherent states which are essentially equivalent to ours to develop a pseudodifferential calculus based upon the Poincare group as
an alternative to the usual Weyl calculus, which is based on the Weyl
Heisenberg group. Since the Poincare group contracts to a group containing the WeylHeisenberg group in the nonrelativistic limit (section 4.2), Unterbergers KleinGordon calculus similarly contracts
to the Weyl calculus.
151
Chapter 5
QUANTIZED FIELDS
5.1. Introduction
We have regarded solutions of the KleinGordon equation as the quantum states of a relativistic particle. But such solutions also possess
another interpretation: they can be viewed as classical fields, something like the electromagnetic field (whose components, in fact, satisfy the wave equation, which is the KleinGordon equation with zero
mass). This interpretation is the basis for quantum field theory. The
general idea is that just as the finite number of degrees of freedom of a
system of classical particles was quantized to give ordinary (point)
quantum mechanics, a similar prescription can be used to quantize the
infinite number of degrees of freedom of a classical field. It turns out
that the resulting theory implies the existence of particles. In fact, the
asymptotic free in and outfields are represented by operators which
create and destroy particles and antiparticles, in agreement with the
fact that such creation and destruction processes occur in nature.
These particles and antiparticles are represented by positiveenergy
solutions of the asymptotic free wave equation, e.g. the KleinGordon
or Dirac equation. Thus the formalism of relativistic quantum mechanics appears to be, at least partially, absorbed into quantum field
theory.
In regarding solutions of the KleinGordon equation as the physical states of a relativistic particle, it was appropriate to restrict our
attention to functions having only positivefrequency Fourier components, since the energy of the particle must be positive. Even a
small negative energy can be made arbitrarily large and negative by
a Lorentz transformation, leading to instability. When the solutions
are regarded as classical fields, however, no such restriction on the
frequency is necessary or even justifiable. For example, in the case
of a neutral field (i.e., one not carrying any electric charge), the
solutions must be realvalued, hence their Fourier transforms must
contain negative as well as positivefrequency components. On the
152
5. Quantized Fields
other hand, the analytic extension of the solutions to complex spacetime appeared to depend crucially on the positivity of the energy. We
must therefore ask whether an extension is still possible for fields, or
if it is even desirable from a physical standpoint, since the connection between solutions and particles is not as immediate as it was
earlier. In this chapter we find an affirmative answer to both of these
questions. A natural method, which we call the AnalyticSignal transform, will be developed to extend arbitrary functions from IRs+1 to
Cls+1 , and when the functions represent physical fields, the double
tube T = T+ T in Cls+1 will be shown to have a direct physical significance as an extended classical phase space, not for the fields themselves but for certain particle and antiparticle coherent states e
z
associated with them. These states are related directly to the dynamical (interpolating) fields, not their asymptotic free in and outfields.
To be precise, they should be called charge coherent states rather
than particle coherent states, since they have a welldefined charge
whereas, in general, the concept of individual particles does not make
sense while interactions are present. If the given fields satisfy some
(possibly nonlinear) equations, the coherent states satisfy a Klein
Gordon equation with a source term. Hence they represent dynamical
rather than bare particles. For free fields, e+
z reduces to the state ez
defined in the last chapter and ez to its complex conjugate, which is a

negativeenergy solution of the KleinGordon equation holomorphic
in T .
Complex tube domains also appear in the contexts of axiomatic
and constructive quantum field theory, and our results suggest that
those domains, too, may have interpretations related to classical phase
space, a point of view which, to my knowledge, has not been explored
heretofore.*
While our extended fields are not analytic in general, they are
analyticityfriendly, i.e. have certain features which yield various
analytic objects under different circumstances. For example, their
twopoint functions are piecewise analytic, and the pieces agree with
the analytic Wightman functions. In the special case when the given
fields are free, the extended fields themselves are analytic in T . Furthermore, the fields in general possess a directional analyticity which
* R. F. Streater has recently told me that G. K
allen was informally advocating the interpretation of the holomorphic Wightman
twopoint function as a correlation function in phase space around
1957. Nothing appears to have been published on this, however.
5.2. The Multivariate AnalyticSignal Transform
153
looks like a covariant version of analyticity in time. Since the latter forms the basis for the continuation of the theory (in the form of
vacuum expectaion values) from Lorentzian to Euclidean spacetime
(see Nelson [1973a,b] and Glimm and Jaffe [1981]), it may be that our
extended fields, when restricted to the Euclidean region, bear some
relation to the corresponding Euclidean fields.
The formalism we are about to develop for fields is a natural
extension of the one constructed for particles in the last chapter. Like
its predecessor, it posesses a degree of regularity not found in the
usual spacetime formalism. Some examples of this regularity are:
(a) The extended fields (z) are, under reasonable assumptions, operatorvalued functions (rather than distributions, as usual) when
restricted to T .
(b) The theory contains a natural, covariant ultraviolet damping,
which is a permanent feature of the theory. This comes from
the possibility of working directly in phase space, away from real
spacetime. From the point of view of the usual (real spacetime)
theory, our formalism looks like a regularization. From our
point of view, however, no regularization is necessary since, it
is suggested, reality takes place in complex spacetime! In other
words, this regularization is permanent and is not to be regarded as a kind of trick, used to obtain finite quantities, which
must later be removed from the theory.
(c) In the case of free fields, the formalism automatically avoids zero
point energies without normal ordering, due to a polarization of
the positive and negative frequency components into the forward
and backward tubes, respectively. Observables such as charge,
energymomentum and angular momentum are obtained as conserved integrals of bilinear expressions in the fields over phase
spaces T . These expressions, which are densities for the
corresponding observables, look like regularizations of the corresponding expressions in the usual spacetime theory. The analytic
(Wightman) twopoint function acts as a reproducing kernel for
the fields, much as it did for the wave functions in chapter 4.
(d) The particles and antiparticles associated with the free Dirac field
do not undergo the random motion known as Zitterbewegung
(Messiah [1963]), again because of the aforementioned polarization.
154
5. Quantized Fields

As mentioned above, in dealing with physical fields such as the electromagnetic field, rather than quantum states, we can no longer justify
the restriction that frequencies must be positive. For one thing, as we
shall see, in the presence of interactions there is no longer a covariant way to eliminate negative frequencies. Hence the method used in
chapter 4 to analytically continue solutions of the KleinGordon equation to complex spacetime will no longer work directly. In this section
we devise a method for extending arbitrary functions from IRs+1 to
Cls+1 . When the given functions are positiveenergy solutions of the
KleinGordon or the Wave equation, this method reduces to the analytic extension used in chapter 4. But it is much more general, and
will enable us to extend quantized fields, whether Bose or Fermi, interacting or free, to complex spacetime. We begin by formulating the
method for functions of one variable, where it is closely related to the
concept of analytic signals. For motivational purposes, we think of
the variable as time (s = 0). In this chapter, Fourier transforms will
usually be with respect to spacetime (IRs+1 ) rather than just space
(IRs ). Hence we will denote them by f, reserving f for the spatial
Fourier transform, as done so far.
Suppose we are given a timesignal, i.e. a real or complex
valued function of a single real variable x. To begin with, assume that
f is a Schwartz test function, although most of our considerations will
extend to certain kinds of distributions. Consider the positive and
negative frequency parts of f , defined by
1
dp eixp f(p)
f+ (x) (2)
0
1
(1)
0
ixp
f (x) (2)
dp e
f(p).
Then f+ and f extend analytically to the lowerhalf and upperhalf

complex planes, respectively, i.e.
dp ei(xiy)p f(p),
f+ (x iy) = (2)
y>0
0
1
(2)
f (x iy) = (2)
i(xiy)p
dp e
f(p),
y < 0.
155
f+ and f are just the FourierLaplace transforms of the restrictions

of f to the positive and negative frequencies.
If f is complexvalued, then f+ and f are independent and the
original signal can be recovered from them as
f (x) = lim [f+ (x iy) + f (x + iy)] .
y0
(3)
If f is realvalued, then
f(p) = f(p),
(4)
hence f+ and f are related by reflection,

z ),
f+ (z) = f (
z Cl ,
(5)
and
f (x) = lim 2<f+ (x iy) = lim 2<f (x + iy).
y0
y0
(6)
When f is real, the function f+ (z) is known as the analytic signal

associated with f (x). A complexvalued signal would have two independent associated analytic signals f+ and f . What significance
do f have? For one thing, they are regularizations of f . The above
equation states that f is jointly a boundaryvalue of the pair f+
and f . As such, f may actually be quite singular while remaining
the boundaryvalue of analytic functions. Also, f provide a kind
of envelope description of f (see Klauder and Sudarshan [1968],
section 1.2). For example, if f (x) = cos ax (a > 0), then f (z) =
1
1
2 exp(iaz), so the boundary values are f (x) = 2 exp(iax).
In order to extend the concept of analytic signals to more than
one dimension, let us first of all unify the definitions of f+ and f by
defining
Z
1
f (x iy) (2)
dp (yp) ei(xiy)p f(p)
(7)
for arbitrary x iy Cl, where is the unit step function, defined by

( 0,
u<0
u=0
(u) = 1/2,
(8)
1,
u > 0.
156
5. Quantized Fields
Then we have
f+ (z),
f (z) = 21 f (x),
f (z),
y>0
y=0
y < 0.
(9)
[The apparent inconsistency f (x) = 21 f (x) for y = 0 is due to a mild

abuse of notation. It could be removed by redefining f (z) by a factor
of 2 or, more correctly but laboriously, rewriting it as (Sf )(z). We
prefer the above notation, since the boundaryvalues f (x) will not
actually be used in the phasespace formalism.]
Let us define the exponential step function by
(<) e ,
so that our extension is given by
Z
1
f (z) = (2)
Cl,
(10)
dp izp f(p).
(11)
The identity
(u) (u0 ) = (uu0 ) (u + u0 )
(12)
shows that has the pseudoexponential property

0
= (< < 0 ) + ,
(13)
which will be useful later.

Although this unification of f+ and f may at first appear to be
somewhat artificial, we shall now see that it is actually very natural.
Note first of all that for any real u, we have
Z
1
d i u
u
(u) e =
e ,
(14)
2i i
since the contour on the righthand side may be closed in the lower
halfplane when u < 0 and in the upper halfplane when u > 0. For
u = 0, the equation states that
Z
1
( + i) d
(0) =
2i 2 + 1
(15)
Z
1
d
1
= ,
=
2 2 + 1
2
157
in agreement with our definition, if we interpret the integral as the

limit as L of the integral from L to L. The exponential step
function therefore has the integral representation
Z
1
d i(x y)p
i(xiy)p
=
e
.
(16)
2i i
If this is substituted into our expression for f (z) and the order of
integrations on and p is exchanged, we obtain
Z
d
1
f (x y)
(17)
f (x iy) =
2i i
for arbitrary x iy Cl. We shall refer to the righthand side as the
AnalyticSignal transform of f (x). It bears a close relation to the
Hilbert transform, which is defined by
Z
1
du
(Hf )(x) = PV
f (x u),
(18)
u
where PV denotes the principal value of the integral. Consider the
complex combination

Z
1
1
f (x) i(Hf )(x) =
du i(u) + PV
f (x u)
i
u
Z
1
du
lim
f (x u)
=
i 0 u i
Z
1
d
=
lim
f (x )
i 0 i
(19)
= lim 2f (x i).
0
Similarly,
f (x) + i(Hf )(x) = lim 2f (x + i).
(20)
(Hf )(x) = i lim [f (x + i) f (x i)],
(21)
0
Hence
0
158
5. Quantized Fields
which, for realvalued f , reduces to

(Hf )(x) = lim 2=f (x + i) = lim 2=f (x i).
0
0
(22)
We are now ready to generalize the idea of analytic signals to an

arbitrary number of dimensions.
Definition. Let f S(IRs+1 ). The AnalyticSignal transform of f
is defined by
Z
d
1
f (x y).
(23)
f (x iy) =
2i i
The same argument as above shows that
Z
s1
f (z) = (2)
dp izp f(p)
s+1
ZIR
= (2)s1
dp (yp) eixpyp f(p).
(24)
IRs+1
We shall refer to the righthand side of this equation as the (inverse)

FourierLaplace transform of f in the halfspace
My {p IRs+1 | yp 0}.
(25)
The integral converges absolutely whenever f L1 (IRs+1 ), since

|izp | 1. Hence f (z) can actually be defined for some distributions, not only for test functions. The extension of the transform
to distributions is complicated by the fact that izp is not a test
function in the variable p, hence for a tempered distribution T ,
T (z) T(izp )
(26)
(defined through the Fourier transform T of T ) may not make sense

as a function on Cls+1 . It would be intersting to find a natural class
of distributions for which T (z) does make sense as a function. In
general, however, it may be necessary to consider distributions T such
that T (z) is some kind of distribution on Cls+1 . The solutions to
both of these problems are unknown to me, so a certain amount of
vagueness will be necessary on this point. In the next section, where
T is a quantized field, it will be assumed to satisfy some physically
reasonable conditions which imply that T (z) is welldefined in an
important subset T of Cls+1 .
159
Recall that for s = 0, f (z) was analytic in the upper and lower
halfplanes. In more than one dimension, f (z) need not be analytic,
even though, for brevity, we still write it as a function of z rather
than z and z. However, f (z) does in general possess a partial analyticity which reduces to the above when s = 0. Consider the partial
derivative of f (x iy) with respect to z , defined by
f
z
(27)
f
f
i .
x
y
Then f is analytic at z if and only if f = 0 for all . But using our
definition of f (z), we find that
Z
1
f
2 f (x iy) =
d
(x y).
(28)
2
x
2 f (z) 2
The righthand side is known as the XRay transform (Helgason

[1984]) of the function f /x , given in terms of the parameters x
and y defining the line x( ) = x y. It follows that the complex
derivative
in the direction of y vanishes, i.e.
Z
f
4y f (z) =
d y (x y)
x
(29)
=
f (x y)
d
= 0,
if f decays for large x (e.g., if f is a test function). Equivalently, using

2 (yp) eizp = 2 [(yp)] eizp
(yp) izp
= i
e
y
(30)
izp
= ip (yp) e
= ip (yp) eixp ,
we have
2i f (z) = (2)s1
Z
IRs+1
dp p (yp) eixp f(p),
(31)
160
5. Quantized Fields
so 4i f is the inverse Fourier transform of p f(p) in the hyperplane

Ny {p IRs+1 | yp = 0} = My .
(32)
Hence, if the intersection of the support of f with Ny has positive

Lebesgue measure in Ny , then f will not be analytic at x iy in
general. However, in any case,
2iy f (z)
s1
= (2)
dp yp (yp) eixp f(p) = 0,
(33)
IRs+1
in agreement with the above conclusion. In the onedimensional case

s = 0, this reduces to
f (z)
=0
z
y 6= 0,
(34)
which states that f (z) is analytic in the upper and lower half
planes. The point is that in one dimension, there are only two imaginary directions (up or down), whereas in s+1 dimensions, every y 6= 0
defines a direction. This motivates the following.
Definition. Let Y (z) be a vector field of type (0, 1) on Cls+1 , i.e.
Y = Y (z) . Then a function f on Cls+1 is holomorphic along Y if
Y f (z) Y (z) f (z) 0.
(35)
Thus, our AnalyticSignal transform f (z) is holomorphic along

Y (x iy) = y .
Note: The functorial way to look at f (z) is as an extension of f (x)
to the tangent bundle T (IRs+1 ). Then the above states that f (z)
is holomorphic along the fibers of this bundle. It makes sense that
f (z) ought to satisfy some constraints, since it is determined by a
function f (x) depending on half the number of variables. If IRs+1 is
replaced by a differentiable manifold, the line x( ) = x y would
have to be replaced by a geodesic. This gives a generalized transform
which can be used to extend functions on an arbitrary Riemannian
or PseudoRiemannian manifold, such as a curved spacetime, to its
tangent bundle. #
The multivariate AnalyticSignal transform is related to the Hilbert transform in the direction y (Stein [1970], p. 49), defined as
5.3. Axiomatic Field Theory and Particle Phase Spaces
1
(Hy f )(x) = PV
du
f (x uy),
u
x, y IRs+1 , y 6= 0.
161
(36)
Namely, an argument similar to the above shows that

f (x) + i(Hy f )(x) = lim 2f (x + iy),
(37)
(Hy f )(x) = i lim [f (x + iy) f (x iy)],
(38)
0
hence
0
which, for s = 0, reduces to the previous relation with the ordinary Hilbert transform. As in the onedimensional case, f (x) is the
boundaryvalue of f (z) in the sense that
f (x) = lim [f (x + iy) + f (x iy)].
0
(39)
For realvalued f , eqs. (38) and (39) reduce to

f (x) = lim 2< f (x + iy)
0
(Hy f )(x) = lim 2= f (x + iy).
(40)
0
Note: The AnalyticSignal transform is remarkable in that it combines in a single entity elements of the Hilbert, FourierLaplace and
XRay transforms. In fact, it is an example of a much more general transform which, furthermore, includes the Radon transform and
an ndimensional version of the wavelet transform as special cases!
(Kaiser [1990b,c]).

We begin with a study of general fields, i.e. fields not necessarily satisfying any differential equation or governed by any particular
model of interactions. The AnalyticSignal transform defines (at least
formally) a canonical extension of such fields to complex spacetime.
In this section we show that for general quantized fields satisfying
the Wightman axioms (Streater and Wightman [1964]), the proposed
162
5. Quantized Fields
extension is natural and interesting. In particular, the double tube

domain T T+ T (where y V0 in T ) can be interpreted as
an extended classical phase space for certain particle and antiparticle coherent states naturally associated with the quantized
fields. Although the extended fields are in general not analytic (they
need not even be functions, only ditributions), they are analyticity
friendly in the sense that various objects associated with them are
analytic functions. Consequently, they are more regular than the
original fields, which are their boundary values. In particular, we will
see that under some reasonable assumtions, the extended fields are
operatorvalued functions (rather than distributions) when restricted
to T . The extensions of free fields and generalized free fields are,
moreover, weakly holomorphic in T .
Recall (chapter 4) that a system with a finite number of degrees of
freedom, say a free classical particle, can be quantized in at least two
ways: by replacing the positionand momentum variables with operators satisfying the canonical commutation relations (this is called
canonical quantization), or by considering unitary irreducible representations of the underlying dynamical symmetry group (G2 for non
relativistic particles and P0 for relativistic particles). The second
scheme seems to be more natural, especially in the relativistic case,
where position variables are not a covariant concept. But the first
scheme has the advantage that interactions (potentials) can be introduced more easily, since it generates a kind of functional calculus for
operators.
Both schemes have counterparts in field quantization. In canonical quantization, the particle position is replaced with the initial configuration of the field , i.e. with the values of (x) at all space points
x at some initial time x0 = t; its momentum, according to Lagrangian
field theory, then corresponds to the timederivative of the field at
time t. Thus the phase space of the field is simply the set of all
possible initial data on the Cauchy surface x0 = t in real spacetime.
Quantization is implemented by requiring the initial field and its conjugate momentum to satisfy an infinitedimensional generalization of
the canonical commutation relations. Although this is the standard
approach to field quantization in the physics literature, its physical
and mathematical soundness is open to question. Whereas the two
schemes are equivalent when applied to nonrelativistic quantum mechanics, it is not clear that they remain equivalent when applied to
field quantization, in general. However, they are equivalent for the
quantization of free fields. We will apply canonical quantization to
163
the free KleinGordon and Dirac fields in sections 5.4 and 5.5.
The second approach to field quantization is subsumed in the
socalled axiomatic approach to quantum field theory. The unitary
representations of P0 under which the fields transform are no longer irreducible. This is a consequence of the fact that while P0 is transitive
on the phase space of a classical particle (i.e., any two locations, orientations and states of motion can be transformed into one another),
it is no longer transitive on the phase space of a classical field (two
sets of initial conditions need not be related by P0 ). The reducibility
of the representation corresponds to the presence of an infinite set of
degrees of freedom or, at the particle level, to an indefinite number
of particles. (Even two particles would result in reducibility, since
the Lorentz norm of their total energymomentum can be arbitrarily large.) Consequently, the representation is now characterized by
an infinite number of parameters and no longer uniquely determines
the theory, as it did (for a given mass and spin) when it was irreducible. On the other hand, the label space for the configuration
observables, which for N particles in IRs was {1, 2, 3, , sN }, is now
IRs , and Lorentz invariance means that the labels x mix with the
time variable. This gives an additional structure to quantized fields
not shared by ordinary quantum mechanics, and some of this structure is codified in terms of the Wightman axioms, partly making up
for the indeterminacy due to the reducibility of the representation.
For simplicity of notation we confine our attention to a single
scalar field. The results of this section extend to an arbitrary system
of scalar, spinor or tensor fields. Thus, let (x) be an arbitrary scalar
quantized field. Quantized means that rather than being a real
or complexvalued function on spacetime (like the components of the
classical electromagnetic field), (x) is an operator on some Hilbert
space H. Actually, as noted by Bohr and Rosenfeld [1950], quantized
fields are too singular to be measured at a single point in spacetime.
This led Wightman to postulate that is an operatorvalued distribution, i.e., when smeared over a test function f (x) it gives an operator
(f ) (unbounded, in general) on H. The axiomatic approach turns
out to provide a surprisingly rich mathematical framework common to
all quantum field theories; therefore, it is modelindependent. It was
followed by constructive quantum field theory, in which model field
theories are built and shown to satisfy the Wightman axioms. For
simplicity, we state the axioms in terms of the formal expression (x)
rather than its smeared form (f ). Also, we assume, to begin with,
that the field is neutral, which means that the operators (x) (or,
164
5. Quantized Fields
rather, their smeared forms (f ) with f realvalued) are essentially

selfadjoint. Charged fields, which are described by nonHermitian
operators, and spinor (Dirac) fields, will be considered later. The
axioms are:
1. Relativistic Invariance: Poincare transformations are implemented by a continuous unitary representation U of P0 acting on the
Hilbert space H of the theory. (For particles of halfintegral spin,
such as electrons, P0 must be replaced by its universal cover; see
section 5.5.) Thus,
(x + a) = U (a, ) (x) U (a, ) .
(1)
Let P and M be the selfadjoint generators of spacetime translations and Lorentz trnsformations, respectively. They are interpreted physically as the total energymomentum and angular
momentum operators of the field or, more generally, of the system of (possibly coupled) fields. They satisfy the commutation
relations of the Lie algebra of P0 . In particular, the P s must
commute with one another, thus have a joint spectrum IRs+1 .
(Actually, is a subset of the dual (IRs+1 ) of IRs+1 , which we

may identify with IRs+1 using the Minkowski metric; see section
1.1.)
2. Vacuum: The Hilbert space contains a unit vector 0 , called the
vacuum vector, which is invariant under the representation U , i.e.
U (a, )0 = 0 for all (a, ) P0 . 0 is unique up to a constant
phase factor, and it is cyclic, meaning that the set of all vectors
of the form
(f1 ) (f2 ) (fn )0
(2)
spans H. Invariance under P0 means that 0 is a common eigenvector of the generators P and M with eigenvalue zero:
P 0 = 0
M 0 = 0.
(3)
3. Spectral Condition: The joint spectrum of the energymomentum operators P is contained in the closed forward light cone:
V +.
(4)
This axiom follows from the physical requirement of stability,

which merely states that the energy is bounded below. To see
165
this, note first of all that must be invariant under L0 , by Axiom 1. Hence, if any physical state had a spectral component
with p
/ V + , that component could be made to have an arbitrarily large negative energy by a Lorentz transformation. Note
that the existence and uniqueness of the vacuum means that
contains the origin in its point spectrum with multiplicity one.
4. Locality: The field operators (x) and (x0 ) at points with spacelike separation commute, i.e.,
[(x), (x0 )] = 0
if (x x0 )2 < 0.
(5)
This corresponds to the physical requirement that measurements

of the field at points with spacelike separations must be independent, since no signal can travel faster than light. (For fields of
halfintegral spin, such as the Dirac field treated in section 5.5,
commutators must be replaced with anticommutators.)
5. Asymptotic Condition: As the time x0 , the field (x) has
weak asymptotic limits
(x) in (x)
(weakly) as x0
(x) out (x)
(weakly) as x0 .
(6)
in and out are free fields of mass m > 0, i.e. they satisfy the
KleinGordon equation:
(t
u + m2 ) in = (t
u + m2 ) out = 0.
(7)
Physically, this means that in the far past and future, all particles
are sufficiently far apart to be decoupled, and that interpolates
the in and out fields. Furthermore, the Hilbert spaces on which
in , out and operate all coincide:
Hin = Hout = H.
(8)
In addition to the above, there are axioms concerning the regularity

of the distributions (x) (they are assumed to be tempered, i.e. the
test functions belong to S(IRs+1 )) and the domains of the smeared
operators (f ), which we use implicitly, and clustering, which we will
not use here.
Let us now draw some conclusions from these axioms. Since the
energymomentum operators generate spacetime translations, it follows from eq. (1) that
166
5. Quantized Fields
i (x) = [(x), P ].
(9)
For the Fourier transform (p),

this means that
[(p),
P ] = p (p).
(10)
Consequently for any p IRs+1 , the vector
p (p)
0
(11)
(which is in general nonnormalizable) satisfies
P p = [P , (p)]
0 = p p .
(12)
That is, p is a generalized eigenvector of energymomentum with

eigenvalue p, unless it vanishes. The spectral condition therefore requires that
(p)
0 =0
p
/ 1 ,
(13)
where 1 is the intersection of the spectrum with the support of

More generally, if p is any generalized eigenthe distribution .
vector of energymomentum with eigenvalue p , then the above
commutation relations show that
0 ) p = (p p0 ) (p
0 )p ,
P (p
(14)
0 ) p vanishes or it is a generalized eigenvector of

thus either (p
energymomentum with eigenvalue p p0 , a necessary condition for
which is that p p0 . Thus we conclude that the operator (p),

when it does not vanish, removes an energymomentum p from the
field. This establishes the physical significance of the field operators,
as well as the connection between the (mathematical) Fourier variable p and the (physical) energymomentum operators P . From the
Jacobi identity, it follows that
0 )], P ] = (p + p0 ) [(p),
0 )],
[ [(p),
(p
(p
(15)
0 )] (if not zero) removes a total energymoshowing that [(p),

(p
mentum p + p0 from the field. The cyclic property of the vacuum,
furthermore, implies that the entire spectrum is generated by repeated applications of to the vacuum. Note that in general we
167
cannot draw any conclusions about the support of (p).

For example,
(p)
need not vanish for spacelike p, since may contain points p0
such that p0 p . In fact, very strong conclusions can be drawn
From Lorentz invariance and
from the nature of the support of .
0 ) = 0, where
(p) = (p) it follows that (p) = 0 if and only if (p
p0 = p for some L0 . We conclude that the support of must
be a union of sets of the form
m = {p | p2 = m2 },
m>0
0 = {p | p2 = 0, p 6= 0}
00 = {0}
(16)
2
im = {p | p = m },
m > 0,
which are, in fact, the various orbits of the full Lorentz group L.
Greenberg [1962] has shown that is a generalized free field (i.e., a
sum or integral of free fields of varying masses m 0) if vanishes
on any of the following types of sets:
A = im ,
B = 00
m>0
[
m ,
(17)
0m<M
C=
m ,
M >0
M > 0.
m>M
He has also shown, by giving counterexamples, that this conclusion

cannot be drawn if vanishes on sets of the type
[
D=
m ,
0 M1 < M2
M1 m<M2
(18)
E = 00 .
(See also DellAntonio [1961] and Robinson [1962].)
Note: Up to now, we have not assumed that the field satisfies the
canonical commutation relations (section 5.4), hence our conclusions
are quite general and should hold for an arbitrary (system of mutually) interacting field(s). The Lie algebra generated by the fields
(obtained by including, along with the fields, their commutators
168
5. Quantized Fields
0 )] as well as higherorder commutators) has a very inter[(p),

(p
esting formal structure, although it is not a Lie algebra in the usual
sense. (For one thing, it is uncountably infinitedimensional rather
than finitedimensional.) Namely, the above relations suggest that
the operators P be regarded as belonging to a Cartan subalgebra
is a root vector with asand that (p)

(with p in the support of )
sociated root p. The spectrum is therefore reminiscent of a set
+ of positive roots. In general, the Cartan subalgebra consists of
a maximal set of commuting observables. When considering charged
fields, the charge will also belong to the Cartan subalgebra, with root
values 0, , 2, . . ., where is a fundamental unit of charge. The
vacuum is a vector (not in the Lie algebra but in an associated representation space) of highest (or lowest) weight which, as in the
finitedimensional case, generates a representation of the algebra because of its cyclic property. To my knowledge, this important analogy
between the structures of general quantized fields (i.e., apart from the
canonical commutation relations or any particular models of interactions) and Lie algebras has not been explored, although the methods
of Liealgebra theory could add a powerful new tool to the study of
quantized fields. (In a somewhat different context, the structures of
quantized fields and infinitedimensional Lie algebras are united in
string theory; see Green, Schwarz and Witten [1987].) #
Let us now formally extend the quantized field (x) to Cls+1 , using the AnalyticSignal transform developed in section 5.2. Recall
that this transform was originally defined for Schwartz test functions.
In principle, we would like to define (z) by using its distributional
Fourier transform (p):

Z
s1
(z) = (2)
dp izp (p)
s+1
(19)
IR

izp

.
This presents us with a technical problem, as already noted in the last
section, since izp is not a Schwartz test function in p. One way out
is to smear (z) with a test function f (z) over Cls+1 . Although this is
the safest solution, it is not very interesting since not much appears
to have been gained by extending the field to complex spacetime: the
new field is still an operatorvalued distribution. However, we shall
see that there are reasons to expect (z) to be more regular than a
generic AnalyticSignal transform, due in part to the fact that (x)
satisfies the Wightman axioms. When is a (generalized) free field,
169
the restriction of (z) to the double tube T turns out to be a holomorphic operatorvalued function. We will see that even for general
Wightman fields, T is an important subset of Cls+1 . In the presence of
interactions, holomorphy is lost but some regularity in T is expected
to remain. We now proceed to find conditions which do not force to
be a generalized free field but still allow (z) to be an operatorvalued
function on T . The arguments given below have no pretense to rigor;
they are only meant to serve as a possible framework for a more precise analysis in the future. All statements and conditions concerning
convergence, integrability and decay of operatorvalued expressions
are meant to hold in the weak sense, i.e. for matrix elements between
fixed vectors. Since the operators involved are unbounded, we must
furthermore assume that the vectors used to form the matrix elements
are in their (form) domains.
For a fixed timelike temper vector y, izp fails to be a
Schwartz test function in two distinct ways: (a) It has a discontinuity on the spacelike hyperplane Ny = {p | yp = 0}, and (b) it has a
constant modulus on hyperplanes parallel to Ny , hence cannot decay
there. On the other hand, by relativistic covariance, the support of
must be smeared over the orbits of L, given by eq. (16). This gives a
stratification of as a sum of tempered distributions
= + + 0 + 00 +
with support properties
[
supp +
m +
(20)
m>0
supp 0 0
supp 00 00
[
supp
im .
(21)
m>0
Although + and contain 0 , the distributions + and

have no contributions from p2 = 0. (For example, the distribution
1 + (p) on IR has a decomposition T1 + T0 where T1 has support
IR and T0 has support {0} IR, but the p = 0 contribution to T1
vanishes.) Similarly, although 0 contains 00 , 0 has no contribution
we have,
from p = 0. Corresponding to the above decomposition of ,
formally,
170
5. Quantized Fields
(z) = + (z) + 0 (z) + 00 (z) + (z).
(22)
We will show that

1. + (z) and 0 (z) are holomorphic operatorvalued functions in T ;
2. 00 (z) must be a constant field, to be physically reasonable; and
3. under certain (hopefully not too restrictive) conditions, (z),
though not holomorphic, is an operatorvalued function on T .
First of all, we claim that each of the fields , ( = , 0, 00) is
still covariant under P0 .* To see this, take the Fourier transform of
eq. (1) and use the invariance of Lebesgue measure dp under L0 . This
leads to
0 p) U (a, ) ,
(p)
= eiap U (a, ) (
(23)
where 0 is the transpose of L0 . Since the different components

are essentially supported on disjoint subsets and these subsets are
invariant under L0 , we conclude that eq. (23) holds for each .
To show that + (z) and 0 (z) are holomorphic in T+ , let z T+ .
Since izp vanishes for p V \{0}, and since + and 0 have no
contribution from p = 0, we have
Z
(z) =
dp eizp (p) = +, 0.
(24)
V+
Now eizp may be regarded as the restriction to V + of a Schwartz

test function fz (p) which is of compact support in p for each fixed p0
and vanishes when p0 < E, for some E > 0. Thus + (z) and 0 (z)
make sense as operatorvalued functions on T+ , and they are cleary
holomorphic there. (This will be shown explicitly below.) A similar
analysis shows that the same can be done for z T .
Next, we consider 00 (z). Since 00 is supported on {p = 0},
it must have the form P () (p), where P () is a partial differential
operator. In the xdomain (i.e., in real spacetime), this corresponds to
a polynomial P (ix), for which the AnalyticSignal transform is not
welldefined, although it is possible that a regularization procedure
would cure this. But in any case, nonconstant polynomials in x
do not appear to be of physical interest since they correspond to
* However, need not satisfy other Wightman axioms such as
locality; for example, + need not be a generalized free field.
171
unbounded fields even in the classical sense (as functions of x). Hence
we assume that
(a) 00 (p) = 2A (p), where A is a constant operator.
This corresponds to a constant field (x) 2A and, correspondingly,
00 (z) A (see eq. (15) in section 5.2). In order that (z) be an
operatorvalued function on T , it therefore remains only for (z) to
be one. Note that so far, the only assumption we needed to make, in
addition to the Wightman axioms, was (a). To make (z) a function,
we now make our second assumption:
(b) (p) is integrable on all spacelike hyperplanes. Furthermore, the
integral of over the hyperplane Hy, {p | yp = } (y V 0 )
grows at most polynomially in .
It is not clear what specific minimal conditions on produce this
property. The integral occurring in (b) is known as the Radon trans
form (R)(y,
) of when y is a (Euclidean) unit vector, and will be
further discussed in section 6.2. (See also Helgason [1984].) Unlike the
Fourier transform, the Radon transform does not readily generalize
to tempered distributions (which were, after all, designed specifically
for the Fourier transform). However, it does extend to distributions
of compact support and can be further generalized to distributions
with only mild decay. Also, the relation of assumptions (a), (b) (or
their future replacements, if any) to the Wightman axioms needs to
be investigated.
In order to compute (z) for z T it suffices, by covariance, to
do so for x = 0 and y = (u, 0), for all u 6= 0. The analyses for u > 0
and u < 0 are similar, so we restrict ourselves to u > 0. Eq. (19) then
gives
s1
(iu, 0) = (2)
up0
ds p (p0 , p).
dp0 e
0
(25)
IRs
For fixed p0 0, condition (b) implies that the integral over p converges, giving an operatorvalued function F (p0 ) which is of at most
polynomial growth in p0 . (iu, 0) is then the Laplace transform
of F (p0 ), which is indeed welldefined.
Note: The behaviors of (z) and (p)

exhibit a certain duality which
s+1
reflects the dual nature of y IR
and p (IRs+1 ) (section 1.1).
We have just seen that when (p)

behaves reasonably for spacelike p,
then (z) behaves reasonably for timelike y. In the trivial case when
172
5. Quantized Fields
(p)
0 for p2 < 0 and (a) holds, (z) is holomorphic for y 2 > 0.
In fact, is then a generalized free field, hence may be said to be
trivial. This dual behavior also extends to p2 0 and y 2 < 0: For
any nonconstant field, (p)

is nontrivial for p2 0; for spacelike
y, the hyperplane Hy, (which contains timelike as well as spacelike
directions) therefore intersects the support of in a noncompact
set and we do not expect (z) to make sense as an operatorvalued
function outside of T . #
No claims of analyticity can be made for (z) in general. In fact,
Greenbergs results show that (z) may not be analytic anywhere in
Cls+1 unless is a generalized free field. For as in the classical case,
formal differentiation with respect to z gives
Z
s1
2i (z) = (2)
dp p (yp) eixp (p),
(26)
IRs+1
hence 4i is the inverse Fourier transform of p in the hyperplane

Ny . If is not a generalized free field, then, according to Greenberg,
the support of contains sets of timelike as well as sets of spacelike ps
with positive Lebesgue measure. Hence, for any nonzero y IRs+1 ,
the intersection of the support of with Ny has positive measure in
Ny , so will not be holomorphic at x iy in general. As in the
classical case, however, the above equation for implies that (z)
is holomorphic along the vector field y, i.e.,
y = 0.
(27)
This is a covariant condition which, when specialized to y = (y0 , 0),

states that (z) is holomorphic in the complex timedirection. As we
have seen, this result simply follows from the nature of the Analytic
Signal transform. A similar situation forms the basis of Euclidean
quantum field theory. However, there one is dealing not directly with
the field but with its vacuum expectation values, and the mathematical reason for the analyticity is the spectral condition, which would
appear to have little in common with the AnalyticSignal transform.
Incidentally, eq. (26) provides a simple formal proof that + (z) is
holomorphic in T . The support of + (p) is contained in V , hence for
any y in V 0 , its intersection with Ny is either empty or equal to 00 .
But the contribution from p = 0 vanishes, hence eq. (26) shows that
+ (z) = 0 in T . The same argument also shows that free fields and
173
generalized free fields are holomorphic in T . In the next two sections

we shall study free KleinGordon and Dirac fields in more detail.
Although (z) is not holomorphic in general, we will be able to
establish for it one essential ingredient of the foregoing phasespace
formalism, namely the interpretation of the double tube T as an extended classical phase space for certain particles and antiparticles
associated with the quantized field .
First, let us expand the above considerations to include charged
fields by allowing (x) to be nonHermitian (i.e., a nonHermitian
operatorvalued distribution). Then the extended field (z) need not
satisfy the reflection condition (z) = (
z ). The charge Q is defined
as a selfadjoint operator which generates overall phase translations
of the field, i.e.,
eiQ (x) eiQ = ei (x)
(28)
for real , where is a fundamental unit of charge. This implies

[(x), Q] = (x)
[(p),
Q] = (p),
(29)
showing that (x) and (p)

remove a unit of charge from the field,
while their adjoints add a unit of charge. We assume that phase
translations commute with Poincare transformations, and in particular with spacetime translations. Thus
[Q, P ] = 0,
(30)
so charge is conserved. Q can be included in the Cartan subalgebra

containing the P s, and the above commutation relations show that
are still root vectors, with Qroot values and ,

(p)
and (p)
respectively. We also assume that the vacuum is neutral, i.e. Q0 =
0. Repeated applications of and to 0 show that the spectrum
of Q is {0, , 2, . . .}.
Recall that the commutation relation between (p)

and P im
plied that (p) removes an energymometum p from the field. Similarly, its adjoint relation
] = p (p)

[P , (p)
(31)
174
5. Quantized Fields
adds an energymomentum p to the field. In place

shows that (p)
of the generalized eigenvectors p of energymomentum which we had
for the Hermitian field, we can now define two eigenvectors,

+
p (p) 0
(p)
0 ,
(32)
for each p 1 . For a nonHermitian field, these vectors are independent. They are states of charge and , respectively. We may
think of them as particles and antiparticles, although they do not
have a welldefined mass since p2 will be variable on 1 , unless is
a free field. Each p 6= 0 in 1 belongs to the continuous spectrum of
the P s, since it can be changed continuously by Lorentz transformations. Hence the vectors
p are nonnormalizable. Since the P s
+
are selfadjoint, and since p and
p belong to different eigenvalues
of the charge operator (which is also selfadjoint), we have (with the
usual abuse of Dirac notation, where inner products of distributions
are taken)
h +
p | q i = 0
2
s+1
h
(p q),
p | q i = (p ) (2)
(33)
where , a distribution with support in 1 , depends only on p2 by

Lorentz invariance. (Charge symmetry requires that be the same
for particles as for antiparticles.) If is the free field of mass m > 0,
(p2 ) = (p0 ) 2(p2 m2 ) = 2(p2 m2 ) in V + \{0}.
Now define the particle coherent states by
e+
z (z) 0
s1
0
dp izp (p)
= (2)
IRs+1
= (2)s1
(34)
dp izp +
p.
V+
Like the +
p s, these do not have a welldefined mass; in addition,
they are wave packets, i.e. have a smeared energymomentum, but
they still have a definite charge . Their spectral components are
given by
+
2
i
zp
h +
= (p2 )(yp) eizp .
p | ez i = (p )
(35)
175
If z belongs to the backward tube T , then yp < 0 on V + \{0}, hence

e+
z = 0. If z belongs to the forward tube T+ , then yp > 0 and the
vector e+
. For the free field, it reduces to
z is weakly holomorphic in z
the coherent state ez defined in chapter 4.
Similarly, define the coherent antiparticle states by
e
z (z)0
s1
0
dp izp (p)
= (2)
IRs+1
= (2)s1
(36)
dp izp
p.
V+
These are wave packets of charge for which
2
izp
h
= (p2 ) (yp) eizp .
p | ez i = (p )
(37)
Thus e
z vanishes in the forward tube and is weakly holomorphic in
the backward tube.
In the usual formulation of quantum field theory, particles are
associated not directly with the interacting, or interpolating, field
but with its asymptotic fields in and out , which are free. (We will
construct such freeparticle coherent states in the next two sections.)
However, the coherent states e
z are directly associated with the
interpolating field. We shall refer to them as interpolating particle
coherent states (section 5.6).
We are now ready to establish the phasespace interpretation of
T in the general case. We will show that T+ and T are extended
phase spaces associated with the particle and antiparticle coherent
states e+
z and ez , respectively, in the sense that they parametrize the
classical states of these particles.
We first discuss x as a position coordinate. In the case of interacting fields there is no hope of finding even a bad version of
position operators. Recall that position operators were in trouble
even in the case of a oneparticle theory without interactions! In the
general case of interacting fields, this problem becomes even more serious, since one is dealing with an indefinite number of particles which
may be dynamically created and destroyed. (As argued in section 4.2,
the generators M0k of Lorentz boosts qualify as a natural, albeit non
commutative, set of centerofmass operators; although I believe this
idea has merit, it will not be discussed here.) Since no position operators are expected to exist, we must not think of x as eigenvalues or
176
5. Quantized Fields
even expectation values of anything, but rather simply as spacetime

parameters or labels.
On the other hand, y will now be shown to be related to the
expectations of the energymomentum operators (which do survive
the transition to quantum field theory, as we have seen). For z, z 0
T+ , we have
+
0
h e+
z 0 | ez i = h 0 | (z ) (z) 0 i
Z
0
= (2)s1
dp ei(z z)p (p2 )
V+
Z
0
1
2
2
dm (m )
d
p ei(z z)p
=
2 0
+
m
Z
1
dm2 (m2 ) + (z 0 z; m),
=
2i 0
(38)
where we have set m2 p2 and used

(2)s1 dp = (2)s1 dp0 ds p = (2)1 dm2 d
p,
(39)
i1
p
2
2
d
p 2(2)
m +p
ds p
(40)
with
h
+
the Lorentzinvariant measure on +
m . (w; m) is the twopoint
function for the free KleinGordon field of mass m, analytically continued to w z 0 z T+ . In the limit y, y 0 0, this gives the
K
allenLehmann representation (Itzykson and Zuber [1980]) for the
usual twopoint function,
1
h 0 | (x ) (x) 0 i =
2i
0
dm2 (m2 ) + (x0 x; m),
(41)
which is a distribution. In Wightman field theory, such vacuum expectation values are analytically continued using the spectral condition,
and conclusions are drawn from these analytic functions about the
field in real spacetime. In our case, we have first extended the field
(albeit nonanalytically), then taken its vacuum expectation values
(which, due to the spectral condition, are seen to be analytic functions, not mere distributions). The fact that we arrived at the same
177
result (i.e., that the diagram commutes) indicates that our approach
is not unrelated to Wightmans. However, there is a fundamental difference: The thesis underlying our work is that the real physics
actually takes place in complex spacetime, and that there is no need
to work with the singular limits y 0.
The norm of e+
z is given by
Z
+ 2
s1
kez k = (2)
dp (p2 ) e2yp
V+
(42)
Z
1
2
2
=
dm (m ) G(y; m),
2 0
where G(y; m), computed in section 4.4, is given by
m
G(y; m) = (2)1
K (2m).
(43)
4
p
Recall that y 2 , (s 1)/2 and K is a modified Bessel
function. We assume that e+
z is normalizable, which means that the
spectral density function (m2 ) satisfies the regularity condition
Z
m
+ 2
2
kez k = (2)
K (2m)
dm2 (m2 )
4
(44)
0
F () < .
(This condition is automatically satisfied for Wightman fields, where
it follows from the assumption that is a tempered distribution; however, it is also satisfied by more singular fields since K decays exponentially.) It follows that
Z
+
s1
+
dp (p2 ) p e2yp
h ez | P ez i = (2)
V+
1 F ()
2 y
= y F 0 ()/2.
(45)
Using the recursion relation (Abramowitz and Stegun [1964])

K (2m) = 2m K+1 (2m),
we find that the state e+

z has an expected energymomentum
(46)
178
5. Quantized Fields
h P i =
m
y ,
y V+0 ,
(47)
where
F 0 ()
m
=
2F ()
R
dm2 (m2 ) m+1 K+1 (2m)
0R
.
2 (m2 ) m K (2m)
dm
(48)
We call m the effective mass of the particle coherent states; it generalizes the corresponding quantity for KleinGordon particles (section
4.4). The name derives from the relation
h P i2 h P ih P i = m2 .
(49)
It is important to keep in mind that in quantum field theory, the

natural picture is the Heisenberg picture, where operators evolve in
spacetime and states are fixed. Recall that for a free KleinGordon
particle (chapter 4), we interpreted ez as a wave packet focused about
the event x <z and moving with an expected energymomentum
(m /) y. This suggests that the above states e
z be given a similar
interpretation. Thus z becomes simply a set of labels parametrizing
the classical states of the particles.
This establishes the interpretation of T+ as an extended classical
phase space associated with the particle states e+
z . A similar computation shows that T acts as an extended classical phase space for
the antiparticle states e
z , whose expected energymomentum is
m
h P i =
(y ),
y V0 .
(50)
The expected angular momentum in the states e

z can be computed similarly. The angular momentum operator M is the generator of rotations in the plane, hence

i x
x
(x) = [(x), M ].
(51)
x
x
This implies for the Fourier transform

[(p), M ] = i p p (p).
p
p
(52)
5.4. Free KleinGordon Fields
179
Since the vacuum is Lorentzinvariant, we have M 0 = 0, hence

for z T+ ,
M e+
z
s1
] 0
dp eizp [M , (p)
= (2)
V+
s1
i
zp
= i(2)
dp e
V+
s1
= (2)
p p
p
p
0
(p)
(53)
dp (
z p z p ) eizp +
p,
V+
provided that (p)

vanishes on the boundary of V + (this excludes
massless fields). The expectation of M is therefore related to that
of P by
h M i = z h P i z h P i
= x h P i x h P i
(54)
m
=
(x y x y ).
Similarly, in e
z with z T ,
m
h M i =
(x y x y ).
(55)
This section can be summarized by saying that the vector y plays

a similar role for general quantized fields as it did for positiveenergy
solutions of the KleinGordon equation, namely it acts as a control vector for the energymomentum. In other words, the function
(yp) eyp acts as a window in momentum space, filtering out from
each mass shell m momenta which are not approximately parallel
to y. The step function (yp) makes certain that only parallel components of p pass through this filter by eliminating the antiparallel
ones (which would make the integrals diverge). Thus we may think
of (yp) eyp as a kind of ray filter in V , when y V 0 . We continue
to refer to y as a temper vector (section 4.4).
Note: The regularity condition given by eq. (44) for (m2 ), i.e. the
requirement that e+
z be normalizable, shows that acts as an effective
ultraviolet cutoff, since K (2m) decays exponentially as m ,
giving finite values to m and other quantities associated with the
field. #
180
5. Quantized Fields

In the context of general quantum field theory, we were able to show
that T plays the role of an extended phase space for certain particle states of the fields. The question arises whether the phasespace
formalism of chapter 4 can be generalized to quantized fields. There,
we saw that all freeparticle states in the Hilbert space could be reconstructed from the values of their wave function on any phase space
T+ . There are two possible ways in which this result might extend to quantized fields: (a) The vectors e
z belong the subspaces H1
with charge , hence we may try to get continuous resolutions, not
of the identity on H but of the orthogonal projection operators 1
to H1 , in terms of these vectors. (This can then be generalized to
the resolution of the projection operator n to the subspace Hn with
charge n, n ZZ.) (b) The global observables of the theory, such
as the energymomentum, the angular momentum and the charge
operators, are usually expressed as conserved integrals of the field operators and their derivatives over an arbitrary configuration space S in
spacetime (i.e., an sdimensional spacelike submanifold of IRs+1 ); our
approach would be to express them as integrals of the extended fields
over 2sdimensional phase spaces in T , much as the inner products
of positiveenergy solutions were expressed as such integrals. In this
section we do both of these things for the free KleinGordon field of
mass m > 0, which is a quantized solution of
(t
u + m2 ) (x) = 0.
(1)
We consider classical solutions at first. The Fourier transform (p)

has the form
(p)
= 2(p2 m2 ) a(p)
(2)
for some complexvalued function a(p) defined on the twosheeted
mass hyperboloid m = +
m m . Write
b(p) a(p),
p +
m.
(3)
If the field is neutral, then (x) is realvalued and b(p) a(p). For
charged fields, a(p) and b(p) are independent. At this point, we keep
both options open. Then
181
(x) = (2)
dp (p2 m2 ) eixp a(p)
s+1
IR
Z
=
d
p eixp a(p)
Zm
h
i
ixp
ixp
=
d
p e
a(p) + e b(p) .
(4)
+
m
The extension of (x) to complex spacetime given by the Analytic

Signal transform is
1
(x iy) =
2i
d
(x y)
i
Z
dp izp (p)
= (2)s1
IRs+1
Z
=
d
p izp a(p)
Zm
i
h
izp
izp
=
d
p (yp) e
a(p) + (yp) e b(p) .
(5)
+
m
If y V+0 , then yp > 0 for all p +

m , hence
Z
(z) =
d
p eizp a(p)
(6)
+
m
is analytic in T+ , containing only the positivefrequency part of the

field. Similarly, when y V0 , then yp < 0 for all p +
m and
Z
(7)
(z) =
d
p eizp b(p), z T .
+
m
Thus (z) is also analytic in T , where it contains only the negative

frequency part of the field. However, note that the two domains of
analyticity T+ and T do not intersect, hence the corresponding restrictions of (z) need not be analytic continuations of one another.
We are now ready to quantize (z). This will be done by first
quantizing (x) and then using the AnalyticSignal transform to extend it to complex spacetime. We assume, to begin with, that is a
182
5. Quantized Fields
neutral field, so b(p) a(p). According to the standard rules (Itzykson and Zuber [1980]) of field quantization, (x) becomes an operator
on a Hilbert space H such that at any fixed time x0 , the field configuration operators (x0 , x) and their conjugate momenta 0 (x0 , x)
obey the equaltime commutation relations
[(x), (x0 )]x00 =x0 = 0
[(x), 0 (x0 )]x00 =x0 = i(x x0 ).
(8)
This is an extension to infinite degrees of freedom of the canonical

commutation relations obeyed by the quantummechanical position
and momentum operators. Note that since time evolution is to be
implemented by a unitary operator, the same commutation relations
will then hold at any other time as well. For the nonHermitian
operators a(p), the corresponding commutation relations are
[a(p), a(p0 )] = 2p0 (p0 p00 ) (2)s (p + p0 )
(9)
for p, p0 m . Using the neutrality condition a(p) = a(p) , these

can be rewritten in their conventional form
[a(p), a(p0 )] = 0
[a(p), a(p0 ) ] = 2 (2)s (p p0 )
(10)
where now p, p0 +
m.
A charged field can be built up from a pair of neutral fields as
(x) =
1 (x) + i2 (x)
,
2
(11)
where the two fields 1 , 2 each obey the equaltime commutation

relations and commute with one another. Equivalently, the operators
a(p) and b(p) become independent and satisfy
[a(p), a(p0 )] = [b(p), b(p0 )] = [a(p), b(p0 )] = 0,
[a(p), a (p0 )] = [b(p), b (p0 )] = 2 (2)s (p p0 )
(12)
for p, p0 +
m . The canonical commutation relations for both neutral
and charged fields can be put in the manifestly covariant form
183
0 ) ] = 4 2 (p2 m2 ) (p0 2 m2 ) [a(p), a(p0 ) ]

[(p),
(p
2
= (2)s+2 2p0 (p0 p00 ) (p2 m2 ) (p0 p2 ) (p p0 )

2
= (2)s+2 2p0 (p0 p00 ) (p2 m2 ) (p00 p20 ) (p p0 )

= sign(p0 ) (2)s+2 (p2 m2 ) (p p0 )
0
for arbitrary p, p IR
mented by
s+1
(13)
. For charged fields, this must be supple-
0 )] = 0,
[(p),
(p
p, p0 IRs+1 .
(14)
Recall that in the general case we had

+
2
s+1
h +
(p p0 )
p | p0 i = (p ) (2)
(15)
for p, p0 V + , where (p2 ) is the spectral density for the twopoint

function (sec. 5.3, eq. (33)). For the free field now under consideration we have
+
0
h +
p | p0 i = h 0 | (p) (p ) 0 i
, (p
0 ) ] 0 i
= h 0 | [(p)
(16)
= (2)s+2 (p2 m2 ) (p p0 ),
which shows that the spectral density for the free field is
(p2 ) = 2 (p2 m2 ).
(17)
The spectral condition implies that

a(p)0 = 0,
b(p)0 = 0
p +
m,
(18)
since otherwise these would be states of energymomentum p.

Hence the particle coherent states defined in the last section are now
given by
Z
+
ez (z) 0 =
d
p eizp a(p) 0
+
m
Z
(19)
i
zp +
d
p e p ,
z T+ ,
+
m
184
5. Quantized Fields
+
where the vectors
p are generalized eigenvectors of energymomen+
tum p m with the normalization
0
+
+
h
p | p0 i = h 0 | a(p) a(p ) 0 i
= h 0 | [a(p), a(p0 ) ]0 i
(20)
= 2(2) (p p ).
The wave packets e+
z span the oneparticle subspace H1 of H and
have the momentum representation
+
i
zp
+
h
.
p | ez i = e
(21)
A dense subspace of H1 is obtained by applying the smeared operators

(f ) (f)
s1
dx (x) f (x) = (2)
f(p)
dp (p)
IRs+1
(22)
to the vacuum, where f is a test function. This gives
+
f
(f ) 0 = (2)
dp f(p) (p) 0
s+1
IR
Z
Z
s1
+
+,
= (2)
dp f (p) p =
d
p f(p)
p
s1
(23)
+
m
IRs+1
where f is the restriction of f to +

m . Hence the functions
Z
+
h e+
d
p eizp f(p)
z | f i =
(24)
+
m
are exactly the holomorphic positiveenergy solutions of the Klein

Gordon equation discussed in section 4.4, with e+
z corresponding to
the evaluation maps ez . The space K of these solutions can thus be
identified with H1 , and the orthogonal projection from H to H1 is
given by
(1 )(z) = h e+
z | i.
(25)
Consequently, the resolution of unity developed in chapter 4 can now

be restated as a resolution of 1 :

Z
+
d | e+
z ih ez | ,
1 =
185
(26)
where + , earlier denoted by , is a particle phase space, i.e. has the

form
+ = {x iy | x S, y +
}
(27)
for some > 0 and some spacelike or, more generally, nowhere timelike (see section 4.5) submanifold S of real spacetime. As in section
4.5, the measure d is given in terms of the Poincareinvariant symplectic form = dy dx by restricting s to + and
choosing an orientation:
c c
d = (s!A )1 s = A1
dy dx .
(28)
Similarly, the antiparticle coherent states for the free field are
given by
Z
ez (z) 0 =
d
p eizp b(p) 0
+
m
Z
(29)
izp
d
p e p ,
z T .
+
m
Since for p +
m and z T we have
izp
+
+
h
= h
p | ez i = e
p | ez i,
(30)
+
it follows that e
z has exactly the same spacetime behavior as ez ,
confirming the interpretation of an antiparticle as a particle moving
backward in time. An antiparticle phase space is defined as a submanifold of T given by
= {x iy | x S, y
},
where S is as above. The resolution of 1 is then given by
Z
1 =
d | e
z ih ez | .
(31)
(32)
Manyparticle or antiparticle coherent states and their corresponding phase spaces can be defined similarly, and the commutation
186
5. Quantized Fields
relations imply that such states are symmetric with respect to permutations of the particles complex coordinates. For example,
+
e+
z1 z2 (z1 ) (z2 ) 0 = ez2 z1 ,
since (z1 ) and (z2 ) commute. In this way, a phasespace formalism can be buit for an indefinite number of particles (or charges),
analogous to the grandcanonical ensemble in classical statistical mechanics. This idea will not be further pursued here. Instead, we
now embark on option (b) above, i.e. the construction of global,
conserved field observables as integrals over particle and antiparticle
phase spaces.
The particle number and antiparticle number operators are given
by
Z
N+ =
d
p a(p) a(p)
+
Z m
(33)
d
p b(p) b(p),
N =
+
m
N+ and N generalize the harmonicoscillator Hamiltonian A A to

the infinite number of degrees of freedom possessed by the field, where
normal modes of vibration are labeled by p +
m for particles and p
for
antiparticles.
The
total
charge
operator
is Q = (N+ N ),
m
as can be seen from its commutation relations with a(p) and b(p). But
the resolution of unity derived in chapter 4 can now be restated as
Z
+
+
+
d exp(i
z p izp0 ) = (2)s 2(p) (p p0 ) = h
p | p0 i
(34)
d exp(izp i
z p0 ) = (2)s 2(p) (p p0 ) = h
p | p0 i
for p, p0 +
m , where the second identity follows from the first by replacing z with z and + with . It follows that N can be expressed
as phasespace integrals of the extended field (z):
Z
N+ =
d (z) (z)
Z +
(35)
N =
d (z) (z) .
187
Hence the charge is given by

Z
d (z) (z)
Q=
+
d (z) (z) .
(36)
The two integrals can be combined into one as follows: Define the
total phase space as = + , where the minus sign means that
enters with the opposite (negative) orientation to that of + ,
in the sense of chains (Warner [1971]). The reason for this choice of
orientation is that B+ and B are both open sets of IRs+1 , hence
must have the same orientation, and we orient +

and so that
= B .
(37)
Since the outward normal of B+ points down and that of B points

up, this means that
must have the opposite orientation to that
+
of . Thus, setting B B+ + B and +

, we have
= B .
(38)
This gives the orientation opposite to that of + , and we have

S = + .
(39)
Next, define the Wickordered product (or normal product) by

(z) (z),
z T+
: (z) (z) :
(40)
(z) (z) ,
z T .
This coincides with the usual definition, since in T+ , is a creation
operator and is an annihilation operator, while in T these roles are
reversed. The charge can now be written in the compact form
Z
Q = d : (z) (z) :
(41)
We may therefore interpret the operator

(z) : (z) (z) :
(42)
as a scalar phasespace charge density with respect to the measure

d.
188
5. Quantized Fields
The Wick ordering can be viewed as a special case of imaginary

time ordering, if we define (z) (
z ) :
: (z 0 ) (z) :: (
z 0 ) (z) := Ti [ (
z 0 ) (z)] ,
(43)
where
Ti [ (z 0 ) (z)]
(=(z00 z0 )) (z 0 ) (z) + (=(z0 z00 )) (z) (z 0 )
(44)
for z, z 0 T . This definition is Lorentzinvariant, since

Ti [(z 0 ) (z)] = (z 0 ) (z) = (z) (z 0 ) z, z 0 T
(45)
Ti [ (z 0 ) (z)] = (z 0 ) (z) = (z) (z 0 )
(46)
and
when z and z 0 are in the same half of T , whereas if they are in opposite
halves of T , the sign of =(z00 z0 ) is invariant.
Note: For the extended fields, the Wick ordering is not a necessity
but a mere convenience, allowing us to combine the integrals over
+ and into a single integral. Each of these integrals is already
in normal order, since the extension to complex spacetime polarizes
the free field into its positiveand negativefrequency parts. Also,
the extended fields are operatorvalued functions rather than distributions, hence products such as (z) (z) are welldefined, which is
not the case in the usual formalism. A similar situation will occur in
the expressions for the other observables (energymomentum, angular
momentum, etc.) as phasespace integrals. Hence the phasespace
formalism resolves the problem of zeropoint energies without the
need to subtract infinite terms by hand! In this connection, see the
remarks on p. 21 of Henley and Thirring [1962]. #
The above expression for the charge can be related to the usual
one in the spacetime formalism, which is
Z
c : (x) :
Qusual = i dx
x
x
Z S
(47)
c J (x),
dx
S
by using = B and applying Stokes theorem:

Z
c
c : :
Q=
dx
dy
S
Z
Z
1
c
= A
dx
dy
: :
y
S
B
Z
Z
c
A1
dx
dy j (x iy),
189
A1
(48)
where
(z)
y
is the phasespace current density. Using the notation

1
+i
z
2 x
y
we have
j (z)
= i( ).
y
Hence, by the holomorphy of ,
: :
y

= i : :
= i : :
(49)
(50)
(51)
j (z)
:.
x
x
Our expression for the charge is therefore
Z
c J (x),
Q=
dx
()
(52)
= i :
(53)
where
J()
(x)
A1
dy j (x iy)
i A1
dy :
:
x
x
B
(54)
190
5. Quantized Fields
is seen to be a regularized version of the usual current density J (x)

obtained by first extending it to T and then integrating it over B .
The vector field j (z) is conserved in real spacetime for each fixed
y V 0 , since
2
j
=
: :
x
x y
= i ( + ) ( ) : :
= i (t
uz u
tz) : :
(55)
= i : ( u
tz u
tz ) :
= 0,
by virtue of the KleinGordon equation combined with the holomor
phy of in T . This implies that J()
(x) is also conserved, hence the
charge does not depend on the choice of S or .
Note: In using Stokes theorem above, we have assumed that the
contribution from |y0 | vanishes. (This was implicit in writing
the noncompact manifold as B .) This is indeed the case, as
has been shown rigorously in the context of the oneparticle theory
in chapter 4 (theorem 4.10). Also, we see another example of the
pattern, mentioned before, that in the phasespace formalism vector
and tensor fields can often be derived from scalar potentials. Here,
(z) acts as a potential for j (z). Note also that the KleinGordon
equation can be written in the form
(t
uz + m2 )(z) = : (t
uz + m2 ) := 0,
(56)
which is manifestly gaugeinvariant. #

Recall now that (z) is a root vector of the charge with root
value :
[(z 0 ), Q] = (z 0 )
z 0 Cls+1 .
Substituting our expression for Q, we obtain the identity
(57)

Z
(z ) =
Z
=
191
d(z) [(z 0 ), : (z) (z) : ]

d(z) [(z 0 ), (z) ] (z)
(58)
d(z) K(z 0 , z) (z),
where, by the canonical commutation relations,

K(z 0 , z) [(z 0 ), (z) ] = [(z 0 ), (
z )]
Z
Z
0 0
0 ), (p)
]
= (2)2s2 dp0 dp iz p izp [(p
Z
Z
0 0
s
0
= (2)
dp
dp iz p izp sign(p0 ) (p2 m2 ) (p0 p)
Z
=
d
p sign(p0 ) (yp) (y 0 p) exp[i(z 0 z)p].
m
(59)
K is a distribution on Cls+1 Cls+1 which is piecewise analytic in
T T , with
i+ (z 0 z; m),
z 0 , z T+
0
i (z z; m),
z 0 , z T
K(z 0 , z) =
(60)
z 0 T+ , z T
0,
0,
z 0 T , z T+ .
The twopoint functions i+ and i are analytic in T+ and T ,
respectively, and act as reproducing kernels for the subspaces with
charge and . Because of the above property, it is reasonable
to call K(z 0 , z) a reproducing kernel for the field (z), though this
differs somewhat from the standard usage of the term as applied to
Hilbert spaces (see chapter 1). Note that K propagates positive
frequency components of the field into the forward (future) tube
and negativefrequency components into the backward (past) tube.
This is somewhat reminiscent of the Feynman propagator, but K is
a solution of the homogeneous KleinGordon equation in the real
spacetime variables rather than a Green function.
The energymomentum and angular momentum operators may be
likewise expressed as conserved phasespace integrals of the extended
field:
192
5. Quantized Fields
Z
d : :
P = i
Z
M = i
(61)
d : (x x ) : .
Like Q, these may be displayed as regularizations of the usual, more

complicated expressions in real spacetime. Note first that P can be
rewritten as
Z
Z
1
c
c : :
dy
P = i A
dx
S
Z
Z
i
c
c : :
= A1
dx
dy
(62)
2 S
Z
Z
1 1
c
c : : .
= A
dx
dy
2
y
S
The angular momentum can be recast similarly as
M =
i A1
c
dx
i
= A1
2
c : (x x ) :
dy

c x ( ) x ( ) : :
dy
S

Z
Z
1 1
c
c
dx
= A
dy x x : : .
2
y
y
S
(63)
Using = B and applying Stokes theorem, we therefore have
c
dx
Z
Z
1 1
2
c
P = A
dx
dy : :
2
y y
S
B
Z
c T () (x),
=
dx
(64)
where
()
T
(x)
1
A1
2
Z
dy
B
2
: :
y y
(65)
5.5. Free Dirac Fields
193
is a regularized energymomentum density tensor which, incidentally,

is automatically symmetric. Similarly,

Z
Z
1 1
2
2
c
= A
dx
dy x x : :
2
y y
y y
S
B
Z
c () (x),
=
dx
(66)
where
()
(x)
1
A1
2
=

dy x
B
()
x T (x)
2
2
y y
y y
: :
(67)
()
x T
(x)
is a regularized angular momentum density tensor.

For simplicity, we specialize in this section (only) to the physical case
of three spatial dimensions, s = 3. The proper Lorentz group L0 is
then SO(3, 1)+ , where the plus sign indicates that 00 > 0, so that
preserves the orientations of space and time separately. Its universal
covering group can be identified with SL(2, Cl) as follows (Streater and
Wightman [1964]): An event x IR4 is identified with the Hermitian
2 2 matrix

0
x + x3 x1 ix2
X = x =
(1)
x1 + ix2 x0 x3
where 0 = I (2 2 identity) and k (k = 1, 2, 3) are the Pauli spin
matrices. Note that det X = x2 x x. The action of SL(2, Cl) on
Hermitian 2 2 matrices given by
X 0 = AXA ,
A SL(2, Cl),
(2)
induces a linear transformation on IR4 which we denote by (A):

x0 = (A) x
(3)
194
5. Quantized Fields
From
2
x0 = det X 0 = | det A|2 det X = det X = x2
(4)
it follows that (A) is a Lorentz transformation, and it can easily be

seen to be proper. Hence defines a map
: SL(2, Cl) L0 ,
(5)
which is readily seen to be a group homomorphism. Clearly, (A) =

(A), and it can be shown that if (A) = (B), then A = B. Since
SL(2, Cl) is simply connected, it follows that SL(2, Cl) is the universal
covering group of L0 , the correspondence being twotoone.
The relativistic transformation law as stated in section 5.3,
(x + a) = U (a, ) (x) U (a, ) ,
(6)
applies to scalar fields, i.e. fields without any intrinsic orientation or

spin. To generalize it to fields with spin, note first of all that since
the representing operator U (a, ) occurs quadratically, the law is invariant under U U . This means that U could, in fact, be a representation, not of P0 , but of the inhomogeneous version of SL(2, Cl),
P2 IR4
s SL(2, Cl),
(7)
(a, A) x = (A) x + a.
(8)
which acts on IR4 by
P2 is the twofold universal covering group of P0 . A field (x) of

arbitrary spin is a distribution taking its values in the tensor product
L(H) V of the operator algebra of the quantum Hilbert space H
with some finitedimensional representation space V of SL(2, Cl). The
transformation law is
U (a, A) (x) U (a, A) = S(A1 ) ((A)x + a),
(9)
where S is a given representation of SL(2, Cl) in V. S determines the

spin of the field, which can take the values j = 0, 1/2, 1, 3/2, 2, . . ..
The locality condition for the scalar field (axiom 4) can be extended to nonscalar fields as
[ (x), (x0 )] = 0
if (x x0 )2 < 0
(10)
195
where are the components of . Now it follows from the axioms

that if the field has halfintegral spin (j = 1/2, 3/2, . . .), then the
above locality condition implies that it is trivial, i.e. that (x) 0.
A nontrivial field of halfintegral spin can be obtained, however,
if we modify the locality condition by replacing commutators with
anticommutators:
{ (x), (x0 )} (x) (x0 ) + (x0 ) (x)
=0
if (x x0 )2 < 0.
(11)
Replacing the commutators with anticommutators means that changing the order in which (x) and (x0 ) are applied to a state vector
in Hilbert space merely changes the sign of the vector, which has no
observable effect. Hence the physical interpretation that events at
spacelike separations cannot influence one another is still valid.
Similarly, for fields of integral spin (j = 0, 1, . . .), the locality
condition with anticommutators gives a trivial theory, whereas a non
trivial theory can exist using commutators.
The choice of commutators or anticommutators in the locality
condition does, however, have an important physical consequence. For
we have seen that the free asymptotic fields can be written as sums
of creation and destruction operators for particles and antiparticles.
If x1 , x2 , xn are n distinct points in the hyperplane x0 = 0 and
+ (x) denotes the positivefrequency part of the field (which can be
obtained from (x iy) by taking y 0 in V+0 ), then
+ (x1 ) + (x2 ) + (xn ) 0
(12)
is a state with n particles located at these points. Since any two of

these points are separated by a spacelike distance, the locality condition implies that this state is symmetric with respect to the exchange
of any two particles if commutators are used and antisymmetric with
respect to the exchange if anticommutators are used. Particles whose
states are symmetric under exchange are called Bosons, and ones antisymmetric under exchange are called Fermions. The choice of symmetry or antisymmetry crucially affects the largescale statistical behavior of the particles. For example, no two Fermions can occupy the
same state due to the antisymmetry under exchange; this is the Pauli
exclusion principle. Hence the choice of commutators or anticommutators is known as the choice of statistics, and the above theorem
correlating this choice with the spin is known as the SpinStatistics
theorem of quantum field theory (Streater and Wightman [1964]).
196
5. Quantized Fields
This theorem is fully supported by experiment, and represents one of

the successes of the theory.
The Dirac field is a quantized field with spin 1/2 whose associated particles and antiparticles are typically taken to be electrons
and positrons, though it is also used (albeit less accurately) to model
neutrons and protons. Our treatment follows the notation used in
Itzykson and Zuber [1980], with minor modifications. The free Dirac
field is a solution of the Dirac equation
(i/ m)(x) = 0,
(13)
where
/
(14)
is the Dirac operator and the s are a set of 4 4 Dirac matrices, meaning they satisfy the Clifford condition with respect to the
Minkowski metric:
{ , } + = 2g .
(15)
The components of satisfy the KleinGordon equation, since

/ 2 = u
t , and the solutions can be written as
Z
(x) =
+
m

d
p eixp u (p) b (p) + eixp v (p) d (p) ,
(16)
where u and v are positive and negativefrequency fourcomponent spinors and summation over the polarization index = 1, 2 is
implied. b and d are operators satisfying the canonical anticommutation relations
{b (p), b (q)} = {d (p), d (q)} = {b (p), d (q)} = 0,

{b (p), b (q)} = {d (p), d (q)} = 2(2)3 (p q)
(17)
for all p, q +
m . The spinors satisfy the orthogonality and completeness relations
197
u
(p) u (p) =
v (p) v (p) =
p/ + m
u (p) u
(p) =
2m
p
/
m
v (p) v (p) =
,
2m
(18)
where a summation on is implied in the last two equations and the

adjoint spinors are defined by
u
(p) = u (p) 0 ,
v (p) = v (p) 0 .
(19)
In addition,
u
(p) u (p) = v (p) v (p) =
p
.
m
(20)
b (p) and d (p) are interpreted as annihilation operators for particles and antiparticles, respectively, while their adjoints are creation
operators. The adjoint field is defined by
(x) = (x) 0
(21)
= m.
x
(22)
and satisfies
i
The particle and antiparticle number operators are now

Z
N+ =
d
p b (p) b (p)
+
Z m
N =
d
p d (p) d (p),
(23)
+
m
and the charge operator is

Q = (N+ N ).
(24)
As for the KleinGordon field, we wish to give a phasespace representation of Q. The first step is to extend (x) to Cl4 using the
AnalyticSignal transform, which gives
198
5. Quantized Fields
Z
(z) =
+
m

d
p izp u (p) b (p) + izp v (p) d (p) .
(25)
Again, the extended field is analytic in T , with the parts in T+ and T

containing only positive and negative frequncies, respectively. Using
the above orthogonality relations, as well as
Z
d e
i
z pizq
d eizpizq = 2(p) (2)3 (p q)
(26)
for p, q +
m , we obtain the following expressions for the particle
and antiparticle number operators as phasespace integrals:
Z
Z
N+ =
d (z) (z)
d : (z) (z) :
+
+
Z
(27)
N =
d : (z) (z) :,
where the fields in the first integral are already in normal order and
the second integral involves two changes of sign: one due to the normal
ordering, and another due to the orthogonality relation for the v s.
The charge operator can therefore be given the following compact
expression as a phasespace integral over the oriented phase space
= + :
Z
Z
Q = d : := d (z),
(28)
: is the scalar phasespace charge density. The usual

where :
expression for the charge as an integral over a configuration space S
is
Z
Z
c
c J (x).
Qusual = dx : (x) (x) :
dx
(29)
S
To compare these two expressions, we again use = B and

invoke Stokes theorem:

Z
c
c : :
Q=
dx
dy
S
Z
Z
1
c
: : .
= A
dx
dy
y
S
B
A1
199
(30)
Define the phasespace current density

j (z) 2m : (z) (z) :,
(31)
where the factor 2m is included to give j the correct physical dimensions, given our normalization. Note that j (z) is conserved in
spacetime, i.e.
j
= ( + )j
x

:
= 2m : +
z
z
=0
(32)
by the Dirac equation combined with the analyticity of in T . The

same combination also implies
1
i :
j (z) = :
2
:
= i :
i ) :
= i : (g
: + :
:,
= i :
(33)
where
=
i
[ , ]
2
(34)
are the spin matrices. The real part of this equation gives a phase
space version of the Gordon identity
: +( + ) :
:
j (z) = i( ) :
: .
=
+ :
y
x
(35)
200
5. Quantized Fields
The two terms are conserved separately, since

2
:= 0
= i(t
uz u
tz ) :
x y
2
: = 0,
:
x x
(36)
and the second term, which is due to spin, does not contribute to the
total charge since it is a pure divergence with respect to x. Thus
Z
Z
Z
1
c
c J (x),
Q = A
dx
dy j (z) =
dx
(37)
()
S
where
J()
(x)
A1
dy j (x iy)
(38)
is a regularized spacetime current.

Note: The Dirac equation can also be written in the manifestly
gaugeinvariant form
: = 2m2 (z). #
i j (z) = 2m :
(39)
Again, is a root vector of the charge operator, since it removes

a charge from any state to which it is applied:
[(z 0 ), Q] = (z 0 )
z 0 Cl 4 .
(40)
Substituting for Q the above phasespace integral and using the commutator identity
[A, BC] = {A, B}C B{A, C}
(41)
and the canonical anticommutation relations, we obtain

0
d {(z ), (z)} (z)
(z ) =
d KD (z 0 , z)(z),
(42)
where the reproducing kernel for the Dirac field is a matrixvalued

distribution on Cl4 Cl4 given by
KD (z 0 , z) = {(z 0 ), (z)}
Z
h
0
=
d
p (y 0 p) (yp) ei(z z)p u u
+
m
i
0
+ (y 0 p) (yp) ei(z z)p v v
0

i/ + m
=
K(z 0 , z).
2m
201
(43)
Here, K is the reproducing kernel for the KleinGordon field and / 0

is the Dirac operator with respect to the real part x0 of z 0 . Like K,
KD is piecewise holomorphic in z 0 z for z 0 , z T . Another form
of the reproducing relation can be obtained by substituting the more
complicated expression for Q given by eq. (37) into eq. (40):
Z
Z
1
0
c
(z ) = 2mA
dx
dy KD (z 0 , z) (z).
(44)
S
This form is closer to the usual relation.

The energymomentum and angular momentum operators for the
Dirac field can likewise be represented by phasespace integrals as
Z
P =
d : i :
Z
(45)
1
M =
d : (ix ix + 2 ) : .
More generally, let (z) represent either a KleinGordon field (in

which case will mean ) or a Dirac field, and let Ta be the local
generators of an arbitrary internal or external symmetry group, so
that the infinitesimal change in (z) is given by
(z) = ia Ta (z).
(46)
For example, Ta is multiplication by for U (1) gauge symmetry, T =

i for spacetime translations (where the derivative is with respect
to x ), etc. (In case the theory has an internal symmetry higher
than U (1), of course, must have extra indices since it must be
valued in a representation space of the corresponding Lie algebra.)
The generators satisfy the Lie relations
202
5. Quantized Fields
c
[Ta , Tb ] = Cab
Tc ,
(47)
c
are the structure constants. Then we claim that the conwhere Cab
served global field observable corresponding to Ta is
Z
Qa =
d : Ta : .
(48)
For this implies

Z
[(z ), Qa ] =
d KD (z 0 , z) Ta (z),
(49)
where KD is replaced by K if is a KleinGordon field. Since Ta

generates a symmetry, it follows that Ta (z) is also a solution of the
appropriate wave equation, hence it is reproduced by KD :
Z
d KD (z 0 , z)Ta (z) = Ta (z 0 ).
(50)
Therefore Qa has the required property

[(z 0 ), Qa ] = Ta (z 0 ).
It can furthermore be checked that
Z
c
[Qa , Qb ] =
d : [Ta , Tb ] := Cab
Qc ,
(51)
(52)
hence the mapping Ta 7 Qa is a Lie algebra homomorphism.

Finally, we show that due to the separation of positive and negative frequencies in T , the interference effect known as Zitterbewegung
does not occur for Fermions in the phasespace formalism. Let St be
the configuration space defined by x0 = t. Then the components of
the regularized threecurrent at time t are
Z
Z
1
k
3
k :,
J() (t) = 2mA
d x
dy :
(53)
St
and a straightforward computation gives

k
Z
p
k
J() (t) =
d
p
[b (p) b (p) d (p) d (p)] .
+
m
m
(54)
5.6. Interpolating Particle Coherent States
203
The righthand side is independent of t, hence no Zitterbewegung

occurs. In real spacetime, Zitterbewegung is the result of the inevitable interference between the positive and negativefrequency
components of . Its absence in complex spacetime is due to the polarization of the positive and negative frequencies of into T+ and
T , respectively.
In the usual theory, Zitterbewegung is shown to occur in the
singleparticle theory; the above computation can be repeated for the
classical (i.e., firstquantized) Dirac field, with an identical result
except for a change in sign in the second term due to the commutation of d and d . Alternatively, the above argument also implies the
absence of Zitterbewegung for the oneparticle and oneantiparticle
states of the Dirac field.

We now return to the interpolating charged scalar field . The asymptotic fields satisfy the KleinGordon equation,
(t
u + m2 ) in (x) = 0,
(t
u + m2 ) out (x) = 0
(1)
and have the same vacuum expectation values as the free Klein
Gordon field discussed in section 5.4. Hence, by Wightmans reconstruction theorem (Streater and Wightman [1964]), these three fields
are unitarily related. We identify the free field of section 5.4 with in .
Then there is a unitary operator S such that
out (x) = S in (x) S .
(2)
S is known as the scattering operator.

Define the source field j(x) by
j(x) (t
u + m2 ) (x).
(3)
It is a measure of the extent of the interaction at x, and by axiom 5,

j(x) 0
(weakly) as x0 .
(4)
Note that we are not making any additional assumptions about j. If j

is a known function (i.e., if it is a multiple of the identity on H for each
x), then it acts as an external source for . If, on the other hand, j
is a local function of such as : 3 :, it represents a selfinteraction of
204
5. Quantized Fields
. In any case, the above equations can be solved using the Green
functions of the KleinGordon operator, which satisfy
(t
ux + m2 ) G(x) = (x).
(5)
In general, we have formally

Z
(x) = 0 (x) +
dx0 G(x x0 ) j(x0 ),
(6)
where 0 is a free field determined by the initial or boundary conditions at infinity used to determine G. The retarded Green function
(we are back to s spatial dimensions) is defined as
Z
eixp
s1
Gret (x) = (2)
dp 2
,
(7)
(p+ m2 )
IRs+1
where
p+ (p0 + i, p)
(8)
with > 0 and the limit 0 is taken after the integral is evaluated.
Gret propagates both positive and negative frequencies forward in
time, which means that it is causal, i.e. vanishes when x0 < 0. Since
it is also Lorentzinvariant, it follows that
Gret (x x0 ) = 0
unless x x0 V+0 .
(9)
Gret (x x0 ) is interpreted as the causal effect at x due to a unit

disturbance at x0 . The corresponding choice of free field 0 is in ,
hence
Z
(x) = in (x) + dx0 Gret (x x0 ) j(x0 ).
(10)
If j is a known external source, this gives a complete solution for (x).
If j is a known function of , it merely gives an integral equation which
must satisfy.
Similarly, the advanced Green function is defined by
0
s1
Gadv (x x ) = (2)
ei(xx )p
dp 2
,
(p m2 )
IRs+1
(11)
205
with p (p0 i, p) and 0, and propagates both positive and

negative frequencies backward in time, which means it is anticausal.
The corresponding free field is out , hence
Z
(x) = out (x) + dx0 Gadv (x x0 ) j(x0 ).
(12)
Let us now apply the AnalyticSignal transform to both of these
equations:
Z
(z) = in (z) + dx0 Gret (z x0 ) j(x0 )
Z
(13)
0
0
0
(z) = out (z) + dx Gadv (z x ) j(x ),
where (with z = x iy)
1
Gret (z x )
2i
0
s1
= (2)
d
Gret (x y x0 )
i
Z
0
(yp) ei(zx )p
dp
(p2+ m2 )
IRs+1
(14)
d
Gadv (x y x0 )
i
Z
0
(yp) ei(zx )p
dp
.
(p2 m2 )
IRs+1
(15)
and
1
Gadv (z x )
2i
0
s1
= (2)
Since the AnalyticSignal transform involves an integration over the

entire line x( ) = x y, the effect of Gret (z x0 ) is no longer
causal when regarded as a function of z and x0 . Rather, it might
be interpreted as the causal effect of a unit disturbance at x0 on the
line parametrized by z. (Note that only those values of for which
x y z 0 V+0 contribute to the integral.) A similar statement goes
for Gadv (z x0 ).
Whereas in (z) and out (z) are holomorphic in T , (z) is not (unless j(x) 0), since Gret (z x0 ) and Gadv (z x0 ) are not holomorphic.
This breakdown of holomorphy in the presence of interactions is by
206
5. Quantized Fields
now expected. Of course , Gret and Gadv are all holomorphic along
the vector field y, as are all AnalyticSignal transforms.
out
In Wightman field theory, the vacua in
and 0 of the
0 , 0
in, out and interpolating fields all coincide (the theory is already
renormalized). Let us define the asymptotic particle coherent states
by
e+
in,z = in (z) 0
e
in,z = in (z)0
(16)
e+
out,z = out (z) 0
e
out,z = out (z)0 .
We will refer to
e+
z = (z) 0 ,
e
z = (z)0
(17)
as the interpolating particle coherent states. By eq. (13),

Z
+
+
ez = ein,z + dx0 Gret (z x0 ) j(x0 ) 0
Z
+
= eout,z + dx0 Gadv (z x0 ) j(x0 ) 0
(18)
and
e
z
e
in,z
e
out,z
+
Z
+
dx0 Gret (z x0 ) j(x0 ) 0

(19)
0
dx Gadv (z x ) j(x ) 0 .
From the definitions it follows that

Gret (z x0 ) = Gadv (x0 z)
Gadv (z x0 ) = Gret (x0 z),
hence eq. (18) can be rewritten as
Z
+
+
ez = ein,z + dx0 Gadv (x0 z) j(x0 ) 0
Z
+
= eout,z + dx0 Gret (x0 z) j(x0 ) 0 .
(20)
(21)
207
Eqs. (19) and (21) display the interpolating character of e

z . Note
that when j(x) is an external source, then the interpolating particle
coherent states differ from the asymptotic ones by a multiple of the
vacuum.
As in the case of the free theory, a general state with a single
positive charge can be written in the form
+
f = (f ) 0 .
(22)
For interacting fields, this may, in general, no longer be interpreted as

a oneparticle state, since no particlenumber operator exists.* But
the charge operator does exist since charge (unlike particlenumber) is
conserved in general, due to gauge invariance; hence +
f makes sense
+
as an eigenvector of charge with eigenvalue . f can be expressed
in terms of particle coherent states as
+
f(z) h e+
z | f i
=h
e+
in,z
=h
e+
out,z
+
f
|
Z
i+
+
f
Z
i+
dx0 Gret (z x0 ) h 0 | j(x0 ) +

f i
(23)
0
dx Gadv (z x ) h 0 | j(x
) +
f
i.
f(z) satisfies the inhomogeneous equations
(t
ux + m ) f(z) =
2
dx0 (t
ux + m2 ) Gret (z x0 ) h 0 | j(x0 ) +
f i
dx0 (t
ux + m2 ) Gadv (z x0 ) h 0 | j(x0 ) +
f i.
(24)
But from the definitions it follows that
* If the spectrum contains an isolated mass shell +

m and f (p)
+
+
is concentrated around m , then f is, in fact, a oneparticle state.
This is the starting point of the HaagRuelle scattering theory (Jost
[1965]). I thank R. F. Streater for this remark.
208
5. Quantized Fields
(t
ux + m2 ) Gret (z x0 ) = (t
ux + m2 ) Gadv (z x0 )
Z
0
s1
= (2)
dp (yp) ei(zx )p
(25)
IRs+1
(z x0 ),
where the last equation is a definition of (z x0 ) as the Analytic
Signal transform with respect to x of (x x0 ). The above is easily
seen to reduce to
(t
ux + m2 ) f(z) = h 0 | j(z) +
f i,
(26)
where j(z) is the AnalyticSignal transform of j(x). Equivalently,

eq. (3) can be extended to Cls+1 by applying the AnalyticSignal
transform, giving
(t
ux + m2 ) (z) = j(z),
(27)
hence
+
(t
ux + m2 ) f(z) = h 0 | (t
ux + m2 ) (z) | +
f i = h 0 | j(z) f i. (28)
For a known external source, this is a perturbed KleinGordon

equation for f(z); if j depends on , it appears to be of little value.
5.7. Field Coherent States and Functional Integrals

So far, all our coherent states have been states with a single particle or
antiparticle. In this section, we construct coherent states in which the
entire field participates, involving an indefinite number of particles.
We do so first for a neutral free KleinGordon field (or a generalized
free field; see section 5.3), then for a free charged scalar field. A similar construction works for Dirac fields, but the functions labeling
the coherent states must then anticommute instead of being classical functions and a generalized type of functional integral must be
used (Berezin [1966], Segal [1956b, 1965]). We also indulge in some
speculation on generalizing the construction to interpolating fields.
209
An extended neutral free KleinGordon field satisfies the canonical commutation relations
[(z), (z 0 )] = 0
[(z), (z 0 ) ] = K(z, z0 ) = i+ (z z0 )
(1)
for all z, z 0 T+ , as well as the reality condition (z) = (

z ). The
basic idea is that since all the operators (z) (z T+ ) commute, it
may be possible to find a total set of simultaneous eigenvectors for
them. This is not guaranteed, since (z) is not selfadjoint (it is not
even normal, by eq. (1)) and, in any case, it is unbounded and thus
may present us with domain problems. However, this hope is realized
by explicitly constructing such eigenvectors. This construction mimics
that of the canonical coherent states in section 3.4, which used the
lowering and raising operators A and A . As in the case of finitely
many degrees of freedom, the canonical commutation relations mean
that acts as a generator of translations in the space in which
is diagonal. The construction proceeds as follows: Let f(p) be a
function on IRs , which will also be regarded as a function on +
m . To
simplify the analysis, we assume to begin with that f is a (complex

valued) Schwartz test function, although this will be relaxed later.
f determines a holomorphic positiveenergy solution of the Klein
Gordon equation,
Z
f (z) =
d
p eizp f(p).
(2)
+
m
Define
(f )
d
p a (p) f(p) =
+
m
d (z) f (z),
(3)
where + is any particle phase space and the second equality follows
from theorem 4.10 and its corollary. (Note: this is not the same as
the smeared field in real spacetime, since the latter would involve
an integration over time, which diverges when f is itself a solution
rather than a test function in spacetime.) The canonical commutation
relations imply that for z T+ ,
Z
[(z), (f )] =
d(z 0 )K(z, z0 ) f (z 0 ) = f (z),
(4)
+
210
5. Quantized Fields
and for n 1,
[(z), (f )n ] = f (z) n (f )n1 .
(5)
We now define the field coherent states of by the formal expression
E f = e
(f )
(6)
0 .
Then if z T+ , so that (z) 0 = 0, eq. (5) implies that
(z) E f = [(z), e
(f )
] 0
= f (z) E f .
(7)
Hence E f is a common eigenvector of all the operators (z), z T+ .

This eigenvalue equation implies that the state corresponding to E f
is left unchanged by the removal of a single particle, which requires
that E f be a superposition of states with 0, 1, 2, particles. Indeed,
X
1 n
E =
(f ) 0 .
n!
n=0
f
(8)
The projection of E f to the oneparticle subspace can be obtained by

using the particle coherent states ez :
h ez | E f i = h 0 | (z) E f i
= f (z)h 0 | E f i = f (z),
(9)
where the last equality follows from (f) 0 ( (f )) 0 = 0. More

generally, the nparticle component of E f is given by projecting to
the nparticle coherent state
ez1 z2 zn (z1 ) (z2 ) (zn ) 0 ,
(10)
h ez1 z2 zn | E f i = f (z1 )f (z2 ) f (zn ),
(11)
which gives
so all particles are in the same state f and the entire system of particles is coherent! Similar states have been found to be very useful
in the analysis of the phenomenon of coherence in quantum optics
(Glauber [1963], Klauder and Sudarshan [1968]), where the name coherent states in fact originated. In the usual treatment, the positive
frequency components have to be separated out by hand using their
211
Fourier representation, since one is dealing with the fields in real

spacetime. For us, this separation occured automatically though the
use of the AnalyticSignal transform, i.e. (f ) can be defined directly as an integral of f (z) over + . (This would remain true even if
f had a negativefrequency component, since the integration over +
would still restrict f to positive frequencies.)
The inner product of two field coherent states can be computed
as follows. Note first that if g(z) is another positiveenergy solution,
then
Z
g
(f ) E
d f (z) (z) E g
Z +
(12)
=
d f (z) g(z) E g
+
= h f | g i Eg ,
where, by theorem 4.10,
Z
Z
hf |gi
d f (z) g(z) =
d
p f(p) g(p).
(13)
+
m
Hence
h E f | E g i = h 0 | e(f ) E g i
= eh f | g i .
(14)
Thus E f belongs to H (i.e., is normalizable) if and only if f(p) belongs

to L2+ (d
p) or, equivalently, f (z) belongs to the oneparticle space K
of holomorphic positiveenergy solutions. If we suppose this to be
the case for the time being, then the field coherent states E f are
parametrized by the vectors f L2+ (d
p) or f K. Next, we look for a
resolution of unity in H in terms of the E f s. The standard procedure
(section 1.3) would be to look for an appropriate measure d(f ) on
K. Actually, it turns out that due to the infinite dimensionality of K,
a larger space K00 K will be needed to support d. Thus, for the
time being, we leave the domain of integration unspecified and write
formally
Z
d(f ) | E f ih E f | = IH ,
(15)
K
212
5. Quantized Fields
where d is to be found. Taking the matrix element of this equation

between the states E h and E g , we obtain
Z
d(f ) eh h | f i+h f | g i = eh h | g i .
(16)
K
With h = g this gives

Z
d(f ) eh f | g ih g | f i = eh g | g i S[g].
(17)
The lefthand side is an infinitedimensional version of the Fourier

transform of d, as becomes apparent if we decompose f and g into
their real and imaginary parts. The Fourier transform of a measure is
called its characteristic function. Hence we conclude that a necessary
condition for the existence of d is that its characteristic function be
S[g]. In turn, a function must satisfy certain conditions in order to be
the characteistic function of a measure. In the finitedimensional case,
Bochners theorem (Yosida [1971]) guarantees the existence of the
measure if these conditions are satisfied. If the infinitedimensional
space of f s is replaced by Cln , the above relation would uniquely
determine d as a Gaussian measure. For the identity
Z
det A1 d2n exp[( A) A1 ( A)] = 1,
(18)
C
ln
where A is a positivedefinite matrix, implies

Z
d() e( + ) = e A ,
(19)
C
ln
with
d() = det A1 exp[ A1 ] d2n .
(20)
The integral in eq. (19) is entire in the variables and separately,

hence it can be analytically continued to , giving
Z
d() e( ) = e A .
(21)
C
ln
If = + i and = u + iv with , , u, v IRn , then
d(, ) e2i(vu) = e(uAu+vAv) S(v, u),
213
(22)
IR2n
showing that S(v, u) is the Fourier transform of d(, ).

If A is merely positivesemidefinite, i.e., if it is singular, then d
still exists but is concentrated on the range of A, and A1 makes
sense as a map from this range to the orthogonal complement of the
kernel of A.
Eq. (19) is a finitedimensional version of eq. (16) (with g = h).
In going to the infinitedimensional case, two separate complications
arise. First of all, recall that in finite dimesions, the Fourier transform
takes funcions on IRn to functions on the dual space, (IRn ) (section
1.1). One usually identifies these two spaces by choosing an inner
product on IRn , e.g., the Euclidean inner product. In the infinite
dimensional case, it is tempting to extend Bochners theorem by letting IRn go to a Hilbert space and looking for a measure on this space.
However, Segal [1956a, 1958] has shown that the ensuing d cannot
be a Borel measure, since it is only finitely additive. To obtain a Borel
measure, the domain of integration must be expanded to a space of
distributions, its dual then being a space of test functions. Minlos
theorem (Gelfand and Vilenkin [1964]; Glimm and Jaffe [1981]) states
that if a functional S[g] defined on the Schwartz space of test functions
S(IRs ) satisfies appropriate conditions (positivedefiniteness, normalization and continuity), a Borel probability measure d exists on the
dual space of tempered distributions S 0 (IRs ) such that eq. (17) is
satisfied for all g S(IRs ), the integration being over S 0 (IRs ). This
resolves the problem of infinite dimansionality. In our case, however,
the situation is further complicated by the fact that the space of f s
over which we wish to integrate, even if it is enlarged, consists not
of free functions but of solutions of the KleinGordon equation. This
difficulty can be overcome by first applying Minlos
theorem in
the
2
momentum space representation, where S[
g ] exp k
g kL2 (dp)
sat
s
isfies the necessary conditions for g S(IR ). This gives a probability
measure d
(f) on S 0 (IRs ). We then define the spaces

Z
izp
K0 g(z) =
d
pe
g(p) | g S
+
m

(23)
Z
0
izp
0
K0 f (z) =
d
pe
f (p) | f S .
+
m
214
5. Quantized Fields
These may be regarded as mutually dual, under the sesquilinear pairing
h f, g i h f, g i
d
p f(p) g(p),
f S 0 , g S.
(24)
+
m
Together with K, they form a triplet

K0 K K00 .
(25)
We now use the map f 7 f to transfer the measure from S 0 to K00 ,

obtaining a probability measure d on K00 . This results, finally, in the
resolution of unity
Z
d(f ) | E f ih E f | = I,
(26)
K00
where the integral converges, as usual, in the weak operator topology.

d is Gaussian in the sense that its restrictions to finitedimensional
cylinder sets in K00 are all Gaussian measures. It is, therefore, an
infinitedimensional version of the Gaussian measure on Cls which
gave the resolution of unity for the canonical coherent states in section
1.2.
One might well ask what is the point of insisting that the integration take place over K00 rather than S 0 (IRs ), the momentum space
representation. One reason is esthetic: The vectors E f combine the
finitedimensional (particle) coherentstate representation with the
infinitedimensional (field) coherentstate representation. Another
reason is that whereas the sample points f in S 0 (IRs ) are merely
distributions, the elements f in K00 are holomorphic functions, since
the decaying exponential eyp dominates the singular behavior of f,
just as it did when f L2 (d
p). Note, however, that nonnormalizable
f
field coherent states E now enter the resolution of unity. In fact, the
Hilbert space K has measure zero with respect to d, since L2 (d
p) has
measure zero with respect to d
. This is remedied by the fact that
only vectors h, g in the test function space K0 are now allowed in eq.
(16).
With the resolution of unity provided by the field coherent states,
the inner product in H can be represented by the functional integral

Z
h | iH =
d(f ) h | E f ih E f | i
K00
(27)
215
d(f ) [f ] [f ].
K00
The above construction was for a neutral scalar field. If were

charged, its coherent states would take the form
E f,g = e
(f )+(
g)
0 ,
(28)
where (f ) is as before and (

g ) is an antiparticle creation operator,
Z
(
g)
d (z) g(z),
(29)
which commutes with (f ). g(z) is a negativeenergy solution of the

KleinGordon equation, holomorphic in T , or, equivalently, g(z) =
g(
z ) with g K00 . Thus we write g K00 . The resolution of unity for
charged fields is therefore
Z
d(f, g) | E f,g ih E f,g | = IH ,
(30)
K00 K00
where d(f, g) = d(f ) d(

g ) is the tensor product of two Gaussian
measures defined as above.
Finally, it is reasonable to ask whether coherent states exist for an
interpolating field, by analogy with the interpolating particle coherent
states studied in sections 5.3 and 5.6. A necessary condition would
seem to be that the first half of the canonical commutation relations
still be valid, i.e.,
[(z), (z 0 )] = 0
(31)
for z, z 0 T+ , since one would like to find simultaneous eigenvectors

of (z) for all z T+ . Recall that the extended free neutral scalar
field had the form
Z

(z) =
d
p izp a(p) + izp a (p)
(32)
+
m
216
5. Quantized Fields
for arbitrary z Cls+1 , and the commutation relation given by eq.

(31) was due to the polarization of the positive and negative frequencies into T+ and T , respectively. If interactions are introduced, the
positive and negativefrequency components get inextricably mixed
together, hence it is highly unlikely that the above commutation relation survives. However, a charged free field has the form
Z

(z) =
d
p izp a(p) + izp b (p) ,
(33)
+
m
where b commutes with a, hence

[(z), (z 0 )] = 0
z, z 0 Cls+1 .
(34)
I believe that this relation does have a chance of holding for interpolating charged scalar fields. It would be a consequence, for example, of
the physical requirement that the Lie algebra generated by the field
has no operators which remove (or add) a double charge 2. This
commutation relation is the weaker half of the freefield canonical
commutation relations, the stronger half (which we do not assume)
being that [(z), (z 0 ) ] is a cnumber, i.e. a multiple of the identity. If [(z), (z 0 )] = 0 for all z and z 0 , then it makes sense to look
for common eigenvectors of all the (z)s, which would be coherent
states of the interpolating field.
Notes
Most of the results in sections 5.25.5 were announced in Kaiser
[1987b] and have been published in Kaiser [1987a]. An earlier attempt to describe quantized fields in complex spacetime was made
in Kaiser [1980b] but was found to be unsatisfactory. The Analytic
Signal transform is further studied in Kaiser [1990c].
Segal [1963b] proposed a formulation of quantum field theory in
terms of the symplectic geometry of the phase space of classical fields.
This phase space corresponds, roughly, to the space K00 defined in section 5.7. An attempt to study quantized fields as (operatorvalued)
functions on the Poincare groupmanifold has been made by Lurcat
[1964]; see also Hai [1969]. As mentioned in section 4.2, this manifold
may be regarded as an extended phase space which includes spin degrees of freedom in addition to position and velocity coordinates. In
Notes
217
the context of classical field theory, this point of view has been generalized to curved spacetime by replacing the Poincare groupmanifold
with the orthogonal frame bundle over a Lorentzian spacetime (Toller
[1978]). These efforts have not, however, utilized holomorphy. It may
be interesting to expand the point of view advocated here to a complex
manifold containing the Poincare groupmanifold in order to account
naturally for spin. This might result in a total coherentstate representation where the classical phase space coordinates range over T
and the spin phase space coordinates range over the Riemann sphere,
as in section 3.5.
I owe special thanks to R. F. Streater for many important comments and corrections in this chapter.
218
6. Further Developments
Chapter 6
FURTHER DEVELOPMENTS
6.1. Holomorphic Gauge theory

In this section we give a brief, notverytechnical but hopefully intuitive, discussion of gauge theory and indicate how it may be modified
in order to make sense in complex spacetime. This represents work
still in progress, and our account is accordingly incomplete. One obstacle is the absence, so far, of a satisfactory Lagrangian formulation.
This is due in part to the fact that fields in complex spacetime are
constrained since they can be derived from local fields in IRs+1 (chapter 5). Our treatment is as elementary as possible, with a geometrical
emphasis. We ignore global questions and work within a single chart.
For more details, see Kaiser [1980a, 1981].
Gauge theory is a natural, geometric way of introducing interactions. It applies some of the ideas of General Relativity to quantum
mechanics and arrives at a class of theories which are generalizations
of classical electrodynamics, the latter being the simplest case. The
power of Einsteins theory of gravitation owes much to the fact that it
actually assumes less, initially, than its predecessor, Newtons theory.
By dropping the assumption that spacetime is flat, we lose the ability
to transport tangent vectors from one point in spacetime to another,
as is done when differentiating a vector field or even finding the acceleration of a particle moving in spacetime. To regain it, we need a
connection. We need to know how a vector transforms in going from
any given point x0 to a neighboring point along a curve x(t). The
tangent vector to the curve at the point x0 is
X = x .
(1)
(The partialderivative operators = /x form a basis for the

tangent space at x0 ; see Abraham and Marsden [1978].) An infinitesimal transport must have a linear effect, thus a vector Y at x0 should
change by
Y = (X)Y t,
219
(2)
where (X) is a linear transformation on the tangent space at x0 . Furthermore, (X) must be linear in X since the latter is an infinitesimal
(i.e., linearized) description of the curve at x0 . If Y = Y , this gives
Y = x Y ( ) t.
(3)
Since ( ) is a linear transformation, we have

Y = x Y t,
(4)
where are a set of (locally defined) funcions on spacetime, known

as the connection coefficients. This gives the rate of change of Y due
to transport along X.
Now suppose we are given a metric g on spacetime. (In Relativity, the metric is indefinite, i.e. Lorentzian rather than Riemannian; with this understood, we continue to call it a metric.) If two
vectors are transported along a curve, then their inner product must
not change since it is a scalar. This gives a relation betweeen and
g which determines the symmetric part (with respect to exchange of
X and Y ) of . The antisymmetric part is the torsion, which in the
standard theory is assumed to vanish. Hence the metric uniquely
determines a torsionless connection known as the Riemannian connection.
General Relativity (see Misner, Thorne and Wheeler [1970]) relates the Riemannian connection to Newtons gravitational potential,
thus giving gravity a geometric interpretation. This is reasonable,
since the connection determines the acceleration of a particle moving
freely (i.e., along a geodesic) in spacetime, which is related to gravity.
By assuming less to begin with, one discovers what must necessarily be added to even compute the acceleration, and this additional
structure turns out to include gravity! In Newtons theory, gravity
must be added in an adhoc fashion. The two theories coincide in the
nonrelativistic limit.
Now consider the wave function of a (scalar) quantum particle, for
example a solution of the free KleinGordon equation, in real spacetime IRs+1 . Suppose we drop the usually implicit assumption that
f can be differentiated by simply taking the difference between its
values at neighboring points. Instead of regarding f (x) as a complex number, we now regard it as a onedimensional complex vector
attached to x, analogous to the tangent vectors in Relativity. This
220
means assuming less structure to begin with, since f (x) now belongs
to Cl as a vector space rather than as an algebra. The complex plane
attached to x will be denoted by Clx and is called the fiber at x, and
the set of all fibers is called a complex line bundle over IRs+1 (Wells
[1980]). To differentiate f , we must know how it is affected by transport. The situation is similar to the one above, only now (X) must
be a 1 1 complex matrix, i.e. a complex number. An infinitesimal
transport gives
f = (X)f t = x f t,
(5)
where (x) is a complexvalued function. The total rate of change

along X is therefore
x f + x f DX f,
(6)
where DX is called the covariant derivative along X. Equivalently,

the differential change is
Df = df + f,
where
dx .
(7)
The 1form is called the connection form. Df is the sum of a horizontal part df (which measures change due to the dependence of f
on x) and a vertical part (which measures change due to transport).
A gauge transformation is represented locally by a multiplication by
a variable phase factor, i.e.
f (x) 7 ei(x) f (x).
(8)
This is a linear map on each fiber Clx , which corresponds in Relativity to the linear map on tangent spaces induced by a coordinate
transformation. In fact, since there is no longer any natural way to
identify distinct fibers, a gauge transformation is a coordinate transformation of sorts. We therefore require that Df be invariant under
gauge transformations, which implies that transforms as
7 id.
(9)
Suppose now that we try to complete the analogy with Relativity

by deriving from a Hermitian metric on the fibers such that the
scalar product of two vectors remains invariant under transport. If
we assume the metric to be positivedefinite, it must have the form
(f (x), g(x)) = f(x) h(x) g(x),
(10)
221
where h(x) is a positive function. It will suffice to consider the inner

product of f (x) with itself, i.e. the quantity
(x) (f (x), f (x)) = f (x) h(x) f (x).
(11)
We require that be invariant under transport. This means that

(f, f ) changes only due to its dependence on x, i.e.
(Df, f ) + (f, Df ) = d(f, f ).
(12)
h + h = dh,
(13)
It follows that
which constrains the real part of but leaves the imaginary part
arbitrary. Writing = R + iA, where R and A are real 1forms, we
have
2R = d log h.
(14)
The real part R of the connection can be transformed away by defining

1, which gives = and
f = h1/2 f and h
f.
Df = h1/2 (df + iAf) h1/2 (d + )
(15)
Since = f f, we may as well assume from the outset that = |f |2

and = iA is purely imaginary.
Note: The mapping f 7 f is not a gauge transformation in the
usual sense; in the standard gauge theory the metric is assumed to be
constant (h(x) 1), hence only phase translations are allowed. This
corresponds to having already transformed away R. It turns out that
in phase space, it will be natural to admit nonconstant metrics.
To make the KleinGordon equation invariant under gauge transformations, we now replace by D = + iA . The result is
(D D + m2 )f = ( + iA )( + iA )f + m2 f = 0.
(16)
This equation was known (even before gauge theory) to be a relativistically covariant description of a KleinGordon particle in the presence of the electromagnetic field determined by the vector potential
A (x). Hence the connection, which describes a geometric property
of the the complex line bundle, acquires a physical significance with
222
respect to electrodynamics, just as did the connection with respect

to gravitation. When f is differentiated in the usual way, it is unconsciously assumed that the connection vanishes. Coupling to an
electromagnetic field then has to be put in by hand, through the
substitution + iA , which is known as the minimal coupling
prescription. Gauge theory gives this adhoc prescription a geometric
interpretation. But note that in this case, the fiber metric h(x) did
not determine the connection. This is due to the complex structure:
iAf cancels in the inner product because it is imaginary. The electromagnetic field generated by the potential A is given by the 2form
F = dA,
(17)
which in fact measures the nontriviality of the connection form A:

if F = 0, then A is closed and therefore (locally) exact, i.e. it is
due purely to a choice of gauge. This is analogous in Relativity to
choosing an accelerating coordinate system, which gives the illusion
of gravity.
Note the complementary nature of the two theories: in Relativity,
the skew part of the connection, which is the torsion, is assumed to
vanish. In gauge theory, the inner product becomes Hermitian, and
the symmetric and antisymmetric parts correspond to its real and
imaginary parts, respectively. It is the real part of the connection
which is assumed to vanish in gauge theory. Were required to
be real, it could in fact be transformed away as above, giving rise
to no gauge field. That is, only the trivial part of the connection
is determined by the metric. The nontrivial (imaginary) part is
arbitrary. We will see that when the theory is extended to complex
spacetime, the metric does determine a nontrivial connection.
Gauge theory usually begins with a Lagrangian invariant under
phase translations f 7 ei f , which form the group U (1). The equations satisfied by f and A are derived using variational principles
(see Bleeker [1981]). There is a natural generalization where f (x) is
an ndimensional complex vector. In that case, the group of phase
tanslations is replaced by a nonabelian group G, usually a subgroup
of U (n). G is called the gauge group, and the correponding gauge
theory is said to be nonabelian or of the YangMills type. Such
theories have in recent years been applied with great success to the
two remaining known interactions (aside from gravity and electromagnetism), namely the weak and the strong forces, which involve
nuclear matter (Appelquist et al. [1987]).
Let us now see how nonabelian gauge theory may be extended
223
to complex spacetime. (This will include the abelian case of electrodynamics when n = 1.) Consider a field f on complex spacetime,
say on the double tube T , whose values are ndimensional complex
vectors. The set of all possible values at z T is a complex vector
space Fz Cln called the fiber at z. The collection of all fibers is
called a vector bundle. We assume that this bundle is holomorphic
(Wells [1980]), so that holomorphic sections z 7 f (z) Fz , represented locally by holomorphic vectorvalued functions, make sense.
Upon transport along a curve z(t) having the complex tangent vector
Z, f changes by
f = (Z)f t,
(18)
where (Z) is a linear map on each fiber. The total differential change
is
Df = (d + ) f.
(19)
d = + ,
(20)

1
+i
= dz = dz
2
x
y

1
= d
z = d
z
i .
2
x
y
(21)
If z = x iy, then
where
Since a general tangent vector has the form

Z = Z + Z ,
(22)
(Z) = dz + d
z,
(23)
we have
where = ( ) and = ( ).
Let us now try again to derive the connection from a fiber metric,
as we have failed to in the case of real spacetime. A positivedefinite
metric on the fibers Fz must have the form
(f (z), g(z)) = f (z) h(z) g(z)
(24)
224
where h(z) is a positivedefinite matrix. Again, it will suffice to consider the (squared) fiber norm
(z) = (f (z), f (z)).
(25)
We define a holomorphic gauge transformation to be a map of the

form
f (z) 7 f 0 (z) = (z)1 f (z),
(26)
h(z) 7 h0 = (z) h(z) (z),
where (z) is an invertible nn matrixvalued holomorphic function.
Clearly is invariant under holomorphic gauge transformations. The
corresponding gauge group acting on a single fiber Fz is the general
linear group G = GL(n, Cl), which includes the usual gauge group
U (n). However, analyticity correlates the values of (z) at different
fibers. Invariance of the inner product under transport gives
(Df, f ) + (f, Df ) = d(f, f ),
(27)
from which we obtain the matrix equation

+ h.
h + h = dh = h
(28)
As in the case of real spacetime, this only determines the Hermitian

part (relative to the metric) of . But if we make the Ansatz
h = h,
(29)
then the resulting connection = h1 h satisfies the above constraint.

We now show that in general, the connection is nontrivial, i.e.
cannot be transformed away by a holomorphic gauge transformation.
Under such a transformation, becomes
0 = ( h)1 ( h)
= 1 h1 h + 1
(30)
= 1 + 1 ,
since = 0 by analyticity. It follows that
hence
D0 f 0 (d + 0 )f 0 = 1 D(f 0 ),
(31)
(D0 )2 f 0 = 1 D2 (f 0 ).
(32)
225
If were trivial, then for some gauge we would have 0 = 0, hence

(D0 )2 = d2 = 0, so D2 = 0. But
D2 f = (d + )(d + ) f = d( f ) + df + f
= (d) f + f f,
(33)
where the 2form

+ +
= d + =
(34)
is the curvature form of the connection , analogous to the Riemann

curvature tensor in Relativity. Using the matrix equation
dh1 = h1 dh h1
(35)
= 0, we find
and 2 = 2 = +

+ = h1 h h1 h + h1 h h1 h = 0. (36)
This is an integrability condition for , being a consequence of the fact
that can be derived from h. One could say that h is a potential for
. Therefore the quadratic term cancels in eq. (34) and the curvature
form reduces to
= .
(37)

Hence if h is such that h1 h 6= 0, then the connection is non
trivial. The form is the complex spacetime version of a YangMills
field, and corresponds to the YangMills potential.
For n = 1, and are the complex spacetime versions of the
electromagnetic potential and the electromagnetic field, respectively.
Since h(z) is a positive function, it may be written as
h(z) = e(z)
(38)
where (z) is real. Then

= ,
= .
(39)
To relate and to the electromagnetic potential A and the electromagnetic field F , one performs a nonholomorphic gauge transformation similar to that in eq.(15): let
226
f(z) = e(z)/2 f (z),
h(z)
1.
(40)
Then the transformed potential becomes purely imaginary,

i
dx ,
=
2 y
(41)
1
2 y
(42)
giving
A (z) =
for the complex spacetime version of the electromagnetic potential.

Note that although A (z) is a pure gradient in the ydirection, the
corresponding electromagnetic field is not trivial, since
A
A
x
x

1
2
2
=

2 x y
x y
F (z)
(43)
need not vanish, in general.

Incidentally, there is an intriguing similarity between the inner
product using the fiber metric,
Z
hf |gi =
d(z) f (z) e(z) g(z),
(44)
and that in Onofris holomorphic coherent states representation (sec.

3.4, eq. (62)). Note that our coincides with Onofris symplectic
form . This possible connection remains to be explored.
Remarks.
1. The relation between the Yangmills potential A and the Yang
Mills field F in real spacetime is F = dA + A A, which is
quadratic in the nonabelian case (n > 1), since the wedge product AA involves matrix multiplication. In our case, however, the
connection satisfies the integrability condition given by eq. (36),
hence the quadratic term cancels and the relation becomes linear,
just as it is normally in the abelian case. The nonholomorphic
gauge transformation f 7 f in eq. (40) can be generalized
to the nonabelian case as follows: Since h(z) is a postive matrix, it can be written as h(z) = k(z) k(z), where k(z) is an
6.2 Windowed XRay Transforms: Wavelets Revisited
227
n n matrix. (k(z) need not be Hermitian; the holomorphic case

= 0 corresponds to a pure gauge field, i.e. = 0.) Setting
k
f(z) = k(z)f (z) and h(z)

1 brings us to the unitary gauge,
where the new gauge transformations are given by unitary matices f(z) 7 U (z)f(z). This amounts to a reduction of the gauge
group from GL(n, Cl) to U (n). In the unitary gauge, the relation
between the connection and the curvature becomes nonlinear, as
it is in the usual YangMills theory. See Kaiser [1981] for details.
2. has a symmetric real part and an antisymmetric imaginary
part. The antisymmetric part corresponds to the usual Yang
Mills field, whereas the symmetric part does not seem to have an
obvious counterpart in real spacetime.
3. In complex differential geometry, = h1 h is known as the
canonical connection of type (1, 0) determined by h. (Wells
[1980]). The functions f are assumed to be (local representations
of) holomorphic sections of the vector bundle. We do not make
this assumption, since it appears that analyticity may be lost in
the presence of interactions. However, it is possible that f (z) is
holomorphic, and it is the nonholomorphic gauge transformation
f 7 f which spoils the analyticity. That is, the nonanalytic part
of the theory may be all contained in the fiber metric h.

In this final section we generalize the idea behind the AnalyticSignal
transform (section 5.2) and arrive at an ndimensional version of the
Wavelet transform (chapters 1 and 2). In a certain sense, relativistic
wave functions and fields in complex spacetime may be regarded as
generalized wavelet transforms of their counterparts in real spacetime.
This is related to the fact, mentioned earlier, that relativistic windows
shrink in the direction of motion, due to the Lorentz contraction associated with the hyperbolic geometry of spacetime. This contraction
is like the compression associated with wavelets.
Other generalizations of the wavelet transform to more than one
dimension have been studied (see Malat [1987]), but they are usually
obtained from the onedimensional one by taking tensor products,
hence are not natural with respect to symmetries of IRn (such as rotations or Lorentz transformations), and consequently do not lends
themselves to analysis by grouptheoretic methods. The generaliza-
228
tion proposed here, which we call a windowed XRay transform, does

not assume any prefered directions, hence it respects all symmetries
of IRn . For example, the transforms of functions over real spacetime
IRs+1 will transform naturally under the Poincare group. If these
functions form a representation space for P0 (whether irreducible, as
in the case of a free particle, or reducible, as in the case of systems of
interacting fields), then so do their transforms.
Let us start directly with the windowed XRay transform (Kaiser
[1990b]). Fix a window function h: IR Cl, which will play a role
similar to a basic wavelet. For a given (sufficiently wellbehaved)
function f on IRn , define fh on IRn IRn by
Z
fh (x, y) =
dt h(t) f (x + ty).
(1)
Note that for n = 1 and y 6= 0,

1
fh (x, y) = |y|
dt h
1/2
= |y|
t0 x
y
f (t0 )
(2)
(W f )(x, y),
where W f is the usual Wavelet transform based on the affine group

A = IR
s IR , defined in section 1.6. On the other hand, for arbitrary
n but
h(t) =
1
,
2(1 it)
(3)
fh coincides with the AnalyticSignal transform defined in section

5.2. The name windowed XRay transform derives from the fact
that for the choice h 1, fh (x, y) is simply the integral of f along
the line x(t) = x + ty, and if |y| = 1 this is known as the XRay
transform (Helgason [1984]), due to its applications in tomography
(Herman [1979]).
Returning to n dimensions and an arbitrary window function h(t),
we wish to know, first of all, whether and how f can be reconstructed
from fh . Note that the transform is trivial for y = 0, since
Z
fh (x, 0) = f (x) dt h(t),
(4)
so we rule out y = 0 and let y range over IRn IRn \{0}. Note also
that fh has the following dilation property for a 6= 0:

Z
dt |a|1 h(t/a) f (x + ty) = fha (x, y),
fh (x, ay) =
229
(5)
where ha (t) |a|1 h(t/a).

To reconstruct f , begin by formally substituting the Fourier representation of f into fh :
Z
Z
fh (x, y) = dt
dn p e2ip(x+ty) h(t) f(p)
Z
(6)
= dn p e2ipx h(py)
f(p)
x,y | f iL2 = h hx,y | f iL2 ,
hh
x,y is defined by
where h
x,y (p) = e2ipx h(py),
(7)
so that
0
hx,y (x ) =
dn p e2ip(x x) h(py).
(8)
x,y , and hence also hx,y , is not squareintegrable for n > 1,

Note that h
since its modulus is constant along directions orthogonal to y.* This
means that eq. (6) will not make sense for arbitrary f L2 (IRn ). We
therefore assume, initially, that f is a test function, say it belongs to
the Schwartz space S(IRn ). Then eq. (6) makes sense with h | i as
the (sesquilinear) pairing between distributions and test functions. If
h is also sufficiently wellbehaved, then we can substitute
Z
h(py)
= d h()
(py )
(9)
and change the order of integration in eq. (8), obtaining
Z
Z
0
0
hx,y (x ) = d h() dn p e2ip(x x) (py ).
(10)
* So far, we have not assumed any metric structure on IRs+1 . Recall

(sec. 1.1) that the natural domain of f(p) is the dual space (IRn ) .
Orthogonal here means that p(y) = 0 as a linear functional. This
will be important when IRn is spacetime with its Minkowskian metric.
230
To reconstruct f , we look for a resolution of unity in terms of

the vectors hx,y (chapter 1). That is, we need a measure d(x, y) on
IRn IRn such that
Z
d(x, y) |fh (x, y)|2 = kf k2L2 .
(11)
For then the map T : f 7 fh is an isometry onto its range in L2 (d),
and polarization gives
h g | f i = h T g | T f i = h g | T T f i,
(12)
showing that f = T (T f ) in L2 (IRn ), which is the desired reconstruction formula. There are various ways to obtain a resolution of unity,
since f is actually overdetermined by fh , i.e. giving the values of fh
on all of IRn IRn amounts to oversampling, so fh will have to
satisfy a consistency condition. We have seen several examples of this
in the study of the windowed Fourier transform (section 1.5) and the
onedimensional wavelet transform (section 1.6 and chapter 2), where
discrete subframes were obtained starting with a continuous resolution of unity. However, for n > 1, there are other options than discrete
subframes, as we will see. In the spirit of the onedimensional wavelet
transform, our first resolution of unity will involve an integration over
all of IRn IRn . Note that

fh (x, y) = h(py)
f (x),
(13)
so Plancherels theorem gives
Z
Z
n
2
2
d x |fh (x, y)| = dn p |h(py)|

|f (p)|2 .
(14)
We therefore need a measure d(y) on IRn such that

Z
2
H(p) d(y) |h(py)|

1 for almost all p.
(15)
The solution is simple: Every p 6= 0 can be transformed to q

(1, 0, . . . 0) by a dilation and rotation of IRn . That is, the orbit of q
(in Fourier space) under dilations and rotations is all of IRn . Thus we
choose d to be invariant under rotations and dilations, which gives
d(y) = N |y|n dn y,
(16)
231
where N is a normalization constant and |y| is the Euclidean norm of

y. Then for p 6= 0,
Z
1 )|2
H(p) = H(q) = N |y|n dn y |h(y
Z
Z
(17)
dy2 dyn
2
1 )|
= N dy1 |h(y
.
(y12 + + yn2 )n/2
But a straightforward computation gives
Z
dy2 dyn
n/2
.
=
|y1 | (n/2)
(y12 + + yn2 )n/2
(18)
This shows that the measure d(x, y) dn x d(y) gives a resolution

of unity if and only if
Z
d
|h()|2 < ,
(19)
ch
||
which is precisely the admissibility condition for the usual (one
dimensional) Wavelet transform (section 1.6). If h is admissible, the
normalization constant is given by
N=
(n/2)
n/2 ch
(20)
and the reconstruction formula is then
f (x ) = (T T f )(x ) = N
dn x d n y
hx,y (x0 ) fh (x, y).
|y|n
(21)
The sense in which this formula holds depends, of course on the behavior of f . The class of possible f s, in turn, depends on the choice
of h. A rigorous analysis of these questions is not easy, and will not
be attempted here. Note that in spite of the factor |y|n in the denominator, there is no problem at y = 0 since
fh (x, 0) = h(0)
f (x) = 0
(22)
by the admissibility condition. The behavior of fh for small y can

be analized using the dilation property, since eq. (5) implies that for
> 0,
232
Z
fh (x, y/) =
dt h(t) f (x + ty).
(23)
Thus if h(t) decays rapidly, say if

h(t) 0
as
(24)
then we expect fh (x, y/) 0 as .

Since eq. (11) holds for admissible h, we can now allow f
2
L (IRn ). We would like to characterize the range <T of the map
T : f 7 fh from L2 (IRn ) to L2 (d). The relation
fh (x, y) = h hx,y | f iL2
(25)
shows that hx,y acts like an evaluation map taking fh L2 (d) to

its value at (x, y). These linear maps on <T are, however, not
bounded if n > 1, since then hx,y is not squareintegrable. (In general,
the value of fh at a point may be undefined.) Hence <T is not a
reproducingkernel Hilbert space (chapter 1). But in any case, the
distributional kernel
K(x, y; x0 , y 0 ) h hx,y | hx0 ,y0 i
Z
0
0)
= dn p e2ip(xx ) h(py)
h(py
(26)
represents the orthogonal projection from L2 (d) onto <T . Thus a

given function in L2 (d) belongs to <T if and only if it satisfies the
consistency condition
Z
g(x, y) = d(x0 , y 0 ) K(x, y; x0 , y 0 ) g(x0 , y 0 ),
(27)
where the integral is the symbolic representation of the action of K
as a distribution.
Remarks.
1. For n = 1, the reconstruction formula is identical with the one for
the continuous onedimensional wavelet transform W f , since by
Eq. (2),
Z
dx dy
|fh (x, y)|2 =
|y|
dx dy
2
|(W f )(x, y)| .
y2
(28)
233
2. In deriving the resolution of unity and the related reconstruction

formula, we have tacitly identified IRn as a Euclidean space, i.e.
we have equipped it with the Euclidean metric and identified the
pairing px in the Fourier transform as the inner product. The exact place where this assumption entered was in using the rotation
group plus dilations to obtain IRn from the single vector q, since
rotations presume a metric.
Having established fh as a generalization of the onedimensional wavelet transform, let us now investigate it in its own right. First, note
that for n = 1 there were only two simple types of candidates for
generalized frames of wavelets: (a) all continuous translations and
dilations of the basic wavelet, or (b) a discrete subset thereof. For
n > 1, any choice of a discrete subset of vectors hx,y spoils the invariance under continuous symmetries such as rotations, and it is
therefore not obvious how to use the above grouptheoretic method
to find discrete subframes. In fact, the discrete subsets {(am , nam b)}
which gave frames of wavelets in section 1.6 and chapter 2 do not form
subgroups of the affine group. One of the advantages of using tensor
products of onedimensional wavelets is that they do generate discrete frames for n > 1, though sacrificing symmetry. However, other
options exist for choosing generalized (continuous) subframes when
n > 1, and one may adapt ones choice to the problem at hand. Such
choices fall between the two extremes of using all the vectors hx,y and
merely summing over a discrete subset, as seen in the examples below.
1. The XRay Transform
The usual XRay transform is obtained by choosing h(t) 1, which
is not admissible in the above sense; hence the above wavelet reconstruction fails. The reason is easy to see: Note that now fh has
the following symmetries:
fh (x, ay) = |a|1 fh (x, y) a IR
fh (x + sy, y) = fh (x, y) s IR.
(29)
Together, these equations state that fh depends only on the line of

integration and not on the way it is parametrized. The first equation
shows that integration over all y 6= 0 is unnecessary as well as undesirable, and it suffices to integrate over the unit sphere |y| = 1. The
second equation shows that for a given y, it is (again) unnecessary and
undesirable to integrate over all x, and it suffices to integrate over the
234
hyperplane orthogonal to y. The set of all such (x, y) does, in fact,

correspond to the set of all lines in IRn , and the corresponding set of
hx,y s forms a continuous frame which gives the usual reconstruction
formula for the XRay transform (Helgason [1984]). The moral of the
story is that sometimes, inadmissibility in the wavelet sense carries
a message: Reduce the size of the frame.
2. The Radon Transform
Next, choose IR and
h (t) = e2it .
(30)
Like the previous function, this one is inadmissible, hence the

wavelet reconstruction fails. Again, this can be corrected by understanding the reason for inadmissibility. Eq. (6) now gives
Z
fh (x, y) = dn p e2ipx (py ) f(p).
(31)
For any a 6= 0, we have
fh (x, ay) = |a|1 fh (x, y),
(32)
where = /a. Hence it suffices to restrict the yintegration to the

unit sphere, provided we also integrate over IR. Also, for any
IR,
Z
fh (x + y, y) =
dn p e2ipx e2i py (py ) f(p)
2i
=e
(33)
fh (x, y).
Fixing x = 0, the function

(Rf)(y, ) = fh (0, y)
(34)
is called the Radon transform of f (Helgason [1984]). It may be

regarded as being defined on the set of all hyperplanes in the Fourier
space (IRn ) , and f can be reconstructed by integrating over the set
of these hyperplanes.
3. The FourierLaplace Transform
Now consider
h(t) =
1
2i(t i)
235
(35)
which gives rise to the AnalyticSignal transform. (We have adopted

a slightly different sign convention than is sec. 5.2. Also, note that
we have reinserted a factor of 2 in the exponent in the Fourier
is the exponential
transform, which simplifies the notation.) Then h
step function (sec. 5.2)
h()
= () e2 = 2 ,
(36)
and eq. (6) reads

Z
fh (x, y) =
dn p (py) e2ip(xiy) f(p)
Z
=
(37)
n
2ip(xiy)
d pe
f(p),
My
where My is the halfspace {p | py > 0}. This is the FourierLaplace

transform of f in My . For n = 1 and y > 0, it reduces to the usual
FourierLaplace transform.
This h, too, is not admissible. f (x) can be recovered simply by
letting y 0, and fh (x, y) may be regarded as a regularization of
f (x). If the support of f is contained in some closed convex cone
(IRn ) , then fh (x, y) f (x iy) is holomorphic in the tube T
over the cone dual to , i.e.
= {y IRn | py > 0 p }
T = {x iy Cln | y }.
(38)
(Note that no metric has been assumed.) In that case, f (x) is a

boundary value of f (xiy). This forms the background for the theory
of Hardy spaces (Stein and Weiss [1971]). We have encountered a
similar situation when IRn was spacetime (n = s + 1), = V + , and
f (x) was a positiveenergy solution of the KleinGordon equation;
then = V 0 and T = T+ . But in that case, f (x) was not in L2 (IRs+1 )
due to the conservation of probability. There it was unnecessary and
undesirable to integrate |f (z)|2 over all of T+ since it was determined
by its values on any phase space + T+ , and reconstruction was
then achieved by integrating over + (chapter 4).
236
As seen from these examples, the windowed XRay transform has

the remarkable feature of being related to most of the classical integral transforms: The XRay, Radon and FourierLaplace transforms.
Since the AnalyticSignal transform is a close relative of the multivariate Hilbert transform Hy (sec. 5.2), we may also add Hy to this
collection.
References
237
REFERENCES
Abraham, R. and Marsden, J. E.
[1978] Foundations of Mechanics, Benjamin/Cummings, Reading,
MA.
Abramowitz, M. and Stegun, L. A., eds.
[1964] Handbook of Mathematical Functions, National Bureau of
Standards, Applied Math. Series 55.
Ali, T. and Antoine, J.P.
[1989] Ann. Inst. H. Poincare 51, 23.
Applequist, T. Chodos, A. and Freund, P. G. O., eds.
[1987] Modern KaluzaKlein Theories, AddisonWesley, Reading,
MA.
Aronszajn, N.
[1950] Trans. Amer. Math. Soc. 68, 337.
Aslaksen, E. W. and Klauder, J. R.
[1968] J. Math. Phys. 9, 206.
[1969] J. Math. Phys. 10, 2267.
Bargmann, V.
[1954] Ann. Math. 59, 1.
[1961] Commun. Pure Appl. Math. 14, 187.
Bargmann, V., Butera, P., Girardello, L. and Klauder, J. R.
[1971] Reps. Math. Phys. 2, 221.
Barut, A.O. and Malin, S.
[1968] Revs. Mod. Phys. 40, 632.
Battle, G.
[1987] Commun. Math. Phys. 110, 601.
Berezin, F. A.
[1966] The Method of Second Quantization, Academic Press, New
York.
Bergman, S.
[1950] The Kernel Function and Conformal Mapping, Amer.
Math. Soc., Providence.
BialinickiBirula, I. and Mycielski, J.
[1975] Commun. Math. Phys. 44, 129.
Bleeker, D.
[1981] Gauge Theory and Variational Principles, Addison Wesley,
Reading, MA.
Bohr, N. and Rosenfeld, L.
[1950] Phys. Rev. 78, 794.
Bott, R.
[1957] Ann. Math. 66, 203.
238
References
Carey, A. L.
[1976] Bull. Austr. Math. Soc. 15, 1.
Daubechies, I.
[1988a] The wavelet transform, timefrequency localization and signal analysis, to be published in IEEE Trans. Inf. Thy.
[1988b] Commun. Pure Appl. Math. XLI, 909.
De Bièvre, S.
[1989] J. Math. Phys. 30, 1401.
DellAntonio, G. F.
[1961] J. Math. Phys. 2, 759.
Dirac, P. A. M.
[1943] Quantum Electrodynamics, Commun. of the Dublin
Inst. for Advanced Studies series A, #1.
[1946] Developments in Quantum Electrodynamics, Commun. of
the Dublin Inst. for Advanced Studies series A, #3.
Dufflo, M. and Moore, C. C.
[1976] J. Funct. Anal. 21, 209.
Dyson, F. J.
[1956] Phys. Rev. 102, 1217.
Fock, V. A.
[1928] Z. Phys. 49, 339.
Gelfand, I. M., Graev, M. I. and Vilenkin, N. Ya.
[1966] Generalized Functions, vol 5, Academic Press, New York.
Gelfand, I. M. and Vilenkin, N. Ya.
[1964] Generalized Functions, vol 4, Academic Press, New York.
Gerlach, B., Gomes, D. and Petzold, J.
[1967] Z. Physik 202, 401.
Glauber, R. J.
[1963a] Phys. Rev. 130, 2529.
[1963b] Phys. Rev. 131, 2766.
Glimm, J. and Jaffe, A.
[1981] Quantum Physics, Springer, New York.
Green, M. W., Schwarz, J. H. and Witten, E.
[1987] Superstring Theory, vols. I, II, Cambridge Univ. Press,
Cambridge.
Greenberg, O. W.
[1962] J. Math. Phys. 3, 859.
Guillemin, V. and Sternberg, S.
[1984] Symplectic Techniques in Physics, Cambridge Univ.
Press, Cambridge.
Hai, N. X.
References
239
Halmos, P.
[1967] A Hilbert Space Problem Book, Van Nostrand, London.
Hegerfeldt, G. C.
[1985] Phys. Rev. Lett. 54, 2395.
Helgason, S.
[1962] Differential Geometry and Symmetric Spaces,
Academic Press, NY.
[1978] Differential Geometry, Lie Groups, and Symmetric Spaces,
Academic Press, New York.
[1984] Groups and Geometric Analysis, Academic Press, New York.
Henley, E. M. and Thirring, W.
[1962] Elementary Quantum Field Theory,
McGrawHill, New York.
Herman, G. T., ed.
[1979] Image Reconstruction from Projections: Implementation
and Applications, Topics in Applied Physics 32, Springer
New York.
Hermann, R.
[1966] Lie Groups for Physicists, Benjamin, New York.
Hille, E.
[1972] Rocky Mountain J. Math. 2, 319.
In
on
u, E. and Wigner, E. P.
[1953] Proc. Nat. Acad. Sci., USA, 39, 510.
Itzykson, C. and Zuber, J.B.
[1980] Quantum Field Theory, McGrawHill, New York.
Jost, R.
[1965] The General Theory of Quantized Fields, Amer. Math. Soc.,
Providence.
Kaiser, G.
[1974] Coherent States and Asymptotic Representations, in Proc.
of Third Internat. Colloq. on GroupTheor. Meth. in
Physics, Marseille.
[1977a] Relativistic CoherentState Representations, in
GroupTheoretical Methods in Physics, R. Sharp and B.
Coleman, eds., Academic Press, New York.
[1977b] J. Math. Phys. 18, 952.
[1977c] PhaseSpace Approach to Relativistic Quantum Mechanics,
Ph. D. thesis, Mathematics Dept, Univ. of Toronto.
[1977d] PhaseSpace Approach to Relativistic Quantum
Mechanics, in proc. of Eighth Internat. Conf. on General
Relativity, Waterloo, Ont.
[1978a] J. Math. Phys. 19, 502.
240
References
[1978b] Hardy spaces associated with the Poincare group, in Lie

Theories and Their Applications, W. Rossmann, ed.,
Queens Univ. Press, Ont.
[1978c] Local Fourier Analysis and Synthesis, Univ. of Lowell
preprint (unpublished).
[1979] Lett. Math. Phys. 3, 61.
[1980a] Holomorphic gauge theory, in Geometric Methods in
Mathematical Physics, G. Kaiser and J. E. Marsden, eds.,
Lecture Notes in Math. 775, Springer, Heidelberg.
[1980b] Relativistic quantum theory in complex spacetime, in Differential Geometrical Methods in Mathematical Physics,
P.L. Garcia, A. PerezRendon and J.M. Souriau, eds.,
Lecture Notes in Math. 836, Springer, Berlin.
[1981] J. Math. Phys. 22, 705.
[1984a] A sampling theorem for signals in the joint time frequency
domain, Univ. of Lowell preprint (unpublished).
[1984b] On the relation between the formalisms of Prugovecki and
Kaiser, Univ. of Lowell preprint, presented as a poster at
the VII-th International Congress of Mathematical Physics,
in Mathematical Physics VII, W. E. Brittin, K. E. Gustafson
and W. Wyss, eds., NorthHolland, Amsterdam
[1987a] Ann. Phys. 173, 338.
[1987b] A phasespace formalism for quantized fields, in
GroupTheoretical Methods in Physics, R. Gilmore, ed.,
World Scientific, Singapore.
[1990a] Algebraic Theory of Wavelets. I: Operational Calculus and
Complex Structure, Technical Repors Series #17, Univ. of
Lowell.
[1990b] Generalized Wavelet Transforms. I: The Windowed XRay
Transform, Technical Repors Series #18, Univ. of Lowell.
[1990c] Generalized Wavelet Transforms. II: The Multivariate
AnalyticSignal Transform, Technical Repors Series #19,
Univ. of Lowell.
Kirillov, A. A.
[1976] Elements of the Theory of Representatins, Springer
Verlag, Heidelberg.
Klauder, J. R.
[1960] Ann. Phys. 11, 123.
[1963a] J Math. Phys. 4, 1055.
[1963b] J Math. Phys. 4, 1058.
Klauder, J. R. and Skagerstam, B.S., eds.
[1985] Coherent States, World Scientific, Singapore.
References
241
Klauder, J. R. and Sudarshan, E. C. G.

[1968] Fundamentals of Quantum Optics, Benjamin, New York.
Kobayashi, S. and Nomizu, K.
[1963] Foundations of Differential Geometry, vol. I, Interscience,
New York.
[1969] Foundations of Differential Geometry, vol. II,
Interscience, New York.
Kostant, B.
[1970] Quantization and Unitary Representatins, in Lectures in
Modern Analysis and Applications, R. M. Dudley et al., eds.
Lecture Notes in Math. 170, Springer, Heidelberg.
Lemarie, P. G. and Meyer, Y.
[1986] Rev. Math. Iberoamericana 2, 1.
Lurcat, F.
[1964] Physics 1, 95.
Mackey, G.
[1955] The Theory of GroupRepresentations, Lecture Notes, Univ.
of Chicago.
[1968] Induced Representations of Groups and Quantum Mechanics, Benjamin, New York.
Mallat, S.
[1987] Multiresolution approximation and wavelets,
preprint, GRASP Lab., Dept. of Computer and Information
Science, Univ. of Pennsylvania, to be published in Trans.
Amer. Math. Soc.
Marsden, J. E.
[1981] Lectures on Geometric Methods in Mathematical Physics,
Notes from a Conference held at Univ. of Lowell, CBMS
NSF Regional Conference Series in Applied Mathematics vol.
37, SIAM, Philadelphia, PA.
Meschkowski, H.
[1962] Hilbertsche R
aume mit Kernfunktion, Springer, Heidelberg.
Messiah, A.
[1963] Quantum Mechanics, vol. II, NorthHolland, Amsterdam.
Meyer, Y.
[1985] Principe dincertitude, bases hilbertiennes et algèbres
doperateurs, Seminaire Bourbaki 662.
[1986] Ondelettes et functions splines, Seminaire EDP,
Ecole Polytechnique, Paris, France (December issue).
Misner, C., Thorne, K. S. and Wheeler, J. A.
[1970] Gravitation, W. H. Freeman, San Francisco.
242
References
Nelson, E.
[1959] Ann. Math. 70, 572.
[1973a] J. Funct. Anal. 12, 97.
[1973b] J. Funct. Anal. 12, 211.
Neumann, J. von
[1931] Math. Ann. 104, 570.
Newton, T. D. and Wigner, E.
[1949] Rev. Mod. Phys. 21, 400.
Onofri, E.
[1975] J. Math. Phys. 16, 1087.
Papoulis, A.
[1962] The Fourier Integral and its Applications, McGrawHill.
Perelomov, A. M.
[1986] Generalized Coherent States and Their Applications,
Springer, Heidelberg.
Prugovecki, E.
[1976] J. Phys. A9, 1851.
[1978] J. Math. Phys. 19, 2260.
[1984] Stochastic Quantum Mechanics and Quantum Spacetime,
D. Reidel, Dordrecht.
Robinson, D. W.
[1962] Helv. Physica Acta 35, 403.
Roederer, J. G.
[1975] Introduction to the Physics and Psychophysics of Music,
Springer, New York.
Schr
odinger, E.
[1926] Naturwissenschaften 14, 664.
Segal, I. E.
[1956a] Trans. Amer. Math. Soc. 81, 106.
[1956b] Ann. Math. 63, 160.
[1958] Trans. Amer. Math. Soc. 88, 12.
[1963a] Illinois J. Math. 6, 500.
[1963b] Mathematical Problems of Relativistic Physics, Amer.
Math. Soc., Providence.
[1965] Bull. Amer. Math. Soc. 71, 419.
Simms, D. J. and Woodhouse, N. M. J.
[1976] Lectures on Geometric Quantization, Lecture Notes in Physics 53, Springer Verlag, Heidelberg.
References
243
Sniatycki,
J.
[1980] Geometric Quantization and Quantum mechanics,
Springer, Heidelberg.
Souriau, J.M.
[1970] Structure des Systèmes Dynamiques, Dunod, Paris.
Stein, E.
[1970] Singular Integrals and Differentiability Properties of
Functions, Princeton Univ. Press, Princeton.
Strang, G.
[1989] SIAM Rev. 31, 614.
Streater, R. F.
Streater, R. F. and Wightman, A. S.
[1964] PCT, Spin & Statistics, and All That, Benjamin, New York.
Thirring, W.
[1980] Quantum Mechanics of Large Systems, Springer, New York.
Toller, M.
[1978] Nuovo Cim. 44B, 67.
Unterberger, A.
[1988] Commun. in Partial Diff. Eqs. 13, 847.
Varadarajan, V. S.
[1968] Geometry of Quantum Theory, vol. I, D. Van Nostrand,
Princeton, NJ.
[1970] Geometry of Quantum Theory, vol. II, D. Van Nostrand,
Princeton, NJ.
[1974] Lie Groups, Lie Algebras, and their Representations,
PrenticeHall, Englewood Cliffs, NJ.
Warner, F.
[1971] Foundations of Differentiable Manifolds and Lie Groups,
Scott, Foresman and Co., London.
Wells, R. O.
[1980] Differential Analysis on Complex Manifolds, Springer, New
York.
Weyl, H.
[1919] Ann. d. Physik 59, 101.
[1931] The Theory of Groups and Quantum Mechanics, translated
by H. P. Robertson, Dover, New York
Wightman, A. S.
[1962] Rev. Mod. Phys. 34, 845.
Yosida, G.
[1971] Functional Analysis, Springer, Heidelberg.
244
References
Young, R. M.
[1980] An Introduction to NonHarmonic Fourier Series, Academic
Press, New York.
Zakai, M.
[1960] Information and Control 3, 101.

90NH

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

90NH

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

90NH

Uploaded by

Copyright:

Available Formats

Quantum Physics, Relativity, and Complex Spacetime:

Towards a New Synthesis

Originally published in 1990 by North-Holland, Amsterdam

To my parents, Bernard and Cesia

Unified field Theory

Suggestions to the Reader

Chapter 1. CoherentState Representations

Suggestions to the Reader

where x0 t is the time coordinate and xj are the space coordinates.

where we adopt the Einstein summation convention of automatically

where (g ) is a nondegenerate matrix. Then each x in X defines

Since g is nondegenerate, it also defines a scalar product on X ,

where x x0 and k x denote the usual Euclidean inner products in

where dx dt ds x is Lebesgue measure on X. f can be reconstructed

where dk = dk0 ds k denotes Lebesgue measure on X X. Note that

These two relations are unified in relativistic terms as p = 2

This is, of course, the source of the uncertainty principle. In terms

If f (x) satisfies a differential equation, such as the Schr

equation or the KleinGordon equation, then f (p) is supported on

where f(p) is, by a mild abuse of notation, the restriction of f to

The Riesz representation theorem guarantees that the converse is also

Similarly, there is a one-to-one correspondence between vectors g H

To illustrate the utility of this notation, as well as some of its

from which we have the resolution of unity

where I is the identity on H and the sum converges in the strong

In physics, vectors such as gn are often written as | n i, which

1.2. Canonical Coherent States

1.2. Canonical Coherent States

where I is the identity operator. The operators iXk , iPk and iI

In particular, the expected position and momentum coordinates of

Choose an arbitrary constant b with units of area (square length) and

Notice that although Ak is nonHermitian, it is real in the Schr

where zk denotes the complexconjugate of zk . Then for Ak Ak

The righthand side is a quadratic in b, hence the inequality for all b

Equality is attained if and only if Ak f = 0, which shows that the

1.2. Canonical Coherent States

where the normalization constant is chosen as N = s/4 , so that

necessary by the discovery of the laser. He dubbed them coherent

f (z) hz | f i = ds x0 z (x0 )f (x0 )

Then we have the following theorem relating the inner products in

Proof. To begin with, assume that f is in the Schwartz space S(IRs )

1.2. Canonical Coherent States

where denotes the Fourier transform with respect to x0 . Thus by

ds p exp(p2 /2)|f(x ip)|2

after exchanging the order of integration. This proves that

for f S(IRs ), hence by continuity also for arbitrary f L2 (IRs ). By

where I is the identity operator on L2 (IRs ) and the integral converges

1.2. Canonical Coherent States

Hence Xk and Pk can be represented as differential rather than integral operators.

d(z) z (x0 )f(z).

In fact, since entire functions are determined by their values on some

is overcomplete. We will see later that reconstruction formulas exist

1.3. Generalized Frames and Resolutions of Unity

We will henceforth drop the adjective generalized and simply speak

1.3. Generalized Frames and Resolutions of Unity

Then the frame condition states that T f is square-integrable with

The frame property can now be stated in operator form as

where the integral is to be interpreted, initially, as converging in the

If 0 < 1, the frame is called snug (Daubechies [1988a]) and