Contents

1 Introduction: The Harmonic Oscillator 5

2 The WKB Method 8
2.1 Some Hamilton-Jacobi preliminaries . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 The WKB approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Symplectic Manifolds 17
3.1 Symplectic structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Cotangent bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Mechanics on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 Quantization in Cotangent Bundles 36
4.1 Prequantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 The Maslov correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Phase functions and lagrangian submanifolds . . . . . . . . . . . . . . . . . . 46
4.4 WKB quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5 The Symplectic Category 64
5.1 Symplectic reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2 The symplectic category . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3 Symplectic manifolds and mechanics . . . . . . . . . . . . . . . . . . . . . . 79
6 Fourier Integral Operators 83
6.1 Compositions of semi-classical states . . . . . . . . . . . . . . . . . . . . . . 83
6.2 WKB quantization and compositions . . . . . . . . . . . . . . . . . . . . . . 87
7 Geometric Quantization 93
7.1 Prequantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Polarizations and the metaplectic correction . . . . . . . . . . . . . . . . . . 98
7.3 Quantization of semi-classical states . . . . . . . . . . . . . . . . . . . . . . . 109
8 Algebraic Quantization 111
8.1 Poisson algebras and Poisson manifolds . . . . . . . . . . . . . . . . . . . . . 111
8.2 Deformation quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.3 Symplectic groupoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
A Densities 119
B The method of stationary phase 121
C

Cech cohomology 124
D Principal T

bundles 125
2
Preface
These notes are based on a course entitled Symplectic geometry and geometric quantization
taught by Alan Weinstein at the University of California, Berkeley, in the fall semester of
1992 and again at the Centre Emile Borel (Institut Henri Poincare) in the spring semester
of 1994. The only prerequisite for the course (and for these notes) was a knowledge of
the basic notions from the theory of dierentiable manifolds (dierential forms, vector elds,
transversality, etc.). The aim of the course was to give students an introduction to the ideas of
microlocal analysis and the related symplectic geometry, with an emphasis on the role which
these ideas play in formalizing the transition between the mathematics of classical dynamics
(hamiltonian ows on symplectic manifolds) and that of quantum mechanics (unitary ows
on Hilbert spaces).
There already exist many books on the subjects treated here, but most of them provide
too much detail for the reader who just wants to nd out what the subject is about. These
notes are meant to function as a guide to the literature; we refer to other sources for many
details which are omitted here, and which can be bypassed on a rst reading.
The pamphlet [63] is in some sense a precursor to these notes. On the other hand, a much
more complete reference on the subject, written at about the same time, is [28]. An earlier
work, one of the rst to treat the connections between classical and quantum mechanics from
a geometric viewpoint, is [41]. The book [29] treats further topics in symplectic geometry
and mechanics, with special attention to the role of symmetry groups, a topic pretty much
ignored in the present notes. For more extensive treatment of the PDE aspects of the subject,
we refer to [43] for a physics-oriented presentation and to the notes [21] and the treatises
[32], [46], and [56]. For geometric quantization, one may consult [35], [53], [54], [60] or
[71]. For classical mechanics and symplectic geometry, we suggest [1], [2], [6], [8], [25], [38],
[59]. Finally, two basic references on quantum mechanics itself are [13] and [20].
Although symplectic geometry is like any eld of mathematics in having its denitions,
theorems, etc., it is also a special way of looking at a very broad part of mathematics and
its applications. For many symplecticians, it is almost a religion. A previous paper by
one of us [64] referred to the symplectic creed.
1
In these notes, we show how symplectic
geometry arises from the study of semi-classical solutions to the Schrodinger equation, and in
turn provides a geometric foundation for the further analysis of this and other formulations
of quantum mechanics.
These notes are still not in nal form, but they have already benetted from the comments
1
We like the following quotation from [4] very much:
In recent years, symplectic and contact geometries have encroached on all areas of mathemat-
ics. As each skylark must display its comb, so every branch of mathematics must nally display
symplectisation. In mathematics there exist operations on dierent levels: functions acting on
numbers, operators acting on functions, functors acting on operators, and so on. Symplec-
tisation belongs to the small set of highest level operations, acting not on details (functions,
operators, functors), but on all the mathematics at once. Although some such highest level
operations are presently known (for example, algebraisation, Bourbakisation, complexication,
superisation, symplectisation) there is as yet no axiomatic theory describing them.
3
and suggestions of many readers, especially Maurice Garay, Jim Morehead, and Dmitry
Roytenberg. We welcome further comments. We would like to thank the Centre Emile Borel
and the Isaac Newton Institute for their hospitality. During the preparation of these notes,
S.B. was supported by NSF graduate and postdoctoral fellowships in mathematics. A.W.
was partially supported by NSF Grants DMS-90-01089 and 93-01089.
4
1 Introduction: The Harmonic Oscillator
In these notes, we will take a spiral approach toward the quantization problem, beginning
with a very concrete example and its proposed solution, and then returning to the same
kind of problem at progressively higher levels of generality. Specically, we will start with
the harmonic oscillator as described classically in the phase plane R
2
and work toward the
problem of quantizing arbitrary symplectic manifolds. The latter problem has taken on a
new interest in view of recent work by Witten and others in the area of topological quantum
eld theory (see for example [7]).
The classical picture
The harmonic oscillator in 1 dimension is described by Newtons dierential equation:
m x = kx.
By a standard procedure, we can convert this second-order ordinary dierential equation
into a system of two rst-order equations. Introducing the phase plane R
2
with position
and momentum coordinates (q, p), we set
q = x p = m x,
so that Newtons equation becomes the pair of equations:
q =
p
m
p = kq.
If we now introduce the hamiltonian function H: R
2
R representing the sum of kinetic
and potential energies,
H(q, p) =
p
2
2m
+
kq
2
2
then we nd
q =
H
p
p =
H
q
These simple equations, which can describe a wide variety of classical mechanical systems
with appropriate choices of the function H, are called Hamiltons equations.
2
Hamiltons
equations dene a ow on the phase plane representing the time-evolution of the classical
system at hand; solution curves in the case of the harmonic oscillator are ellipses centered
at the origin, and points in the phase plane move clockwise around each ellipse.
We note two qualitative features of the hamiltonian description of a system:
1. The derivative of H along a solution curve is
dH
dt
=
H
q
q +
H
p
p = p q + q p = 0,
2
If we had chosen x rather than m x as the second coordinate of our phase plane, we would not have
arrived at this universal form of the equations.
5
i.e., the value of H is constant along integral curves of the hamiltonian vector eld.
Since H represents the total energy of the system, this property of the ow is interpreted
as the law of conservation of energy.
2. The divergence of the hamiltonian vector eld X
H
= ( q, p) = (
H
p
,
H
q
) is
X
H
=

2
H
q p


2
H
p q
= 0.
Thus the vector eld X
H
is divergence-free, and its ow preserves area in the phase
plane.
The description of classical hamiltonian mechanics just given is tied to a particular coordinate
system. We shall see in Chapter 3 that the use of dierential forms leads to a coordinate-
free description and generalization of the hamiltonian viewpoint in the context of symplectic
geometry.
The quantum mechanical picture
In quantum mechanics, the motion of the harmonic oscillator is described by a complex-
valued wave function (x, t) satisfying the 1-dimensional Schr odinger equation:
i

t
=

2
2m

x
2
+
k
2
x
2
.
Here, Plancks constant has the dimensions of action (energy time). Interpreting the
right hand side of this equation as the result of applying to the wave function the operator

H
def
=

2
2m

2
x
2
+
k
2
m
x
2,
where m
x
2 is the operator of multiplication by x
2
, we may rewrite the Schrodinger equation
as
i

t
=

H.
A solution of this equation does not represent a classical trajectory; instead, if is
normalized, i.e.

= 1,
then its square-norm
(x, t) = [(x, t)[
2
is interpreted as a probability density for observing the oscillator at the position x at time
t. The wave function (x, t) itself may be viewed alternatively as a t-dependent function of
x, or as a path in the function space C

(R, C). From the latter point of view, Schrodingers

equation denes a vector eld on C

(R, C) representing the time evolution of the quantum

system; a wave function satisfying Schrodingers equation then corresponds to an integral
curve of the associated ow.
Like Hamiltons equations in classical mechanics, the Schrodinger equation is a general
form for the quantum mechanical description of a large class of systems.
6
Quantization and the classical limit
The central aim of these notes is to give a geometric interpretation of relationships between
the fundamental equations of classical and quantum mechanics. Based on the present dis-
cussion of the harmonic oscillator, one tenuous connection can be drawn as follows. To the
classical position and momentum observables q, p we associate the dierential operators
q q = m
x
p p = i

x
.
The classical hamiltonian H(q, p) = p
2
/2m+kq
2
/2 then corresponds naturally to the operator

H. As soon as we wish to quantize a more complicated energy function, such as (1+q

2
)p
2
,
we run in to the problem that the operators q and p do not commute with one another, so
that we are forced to choose between (1 + q
2
) p
2
and p
2
(1 + q
2
), among a number of other
possibilities. The dierence between these choices turns out to become small when 0.
But how can a constant approach zero?
Besides the problem of quantization of equations, we will also treat that of quan-
tization of solutions. That is, we would like to establish that, for systems which are in
some sense macroscopic, the classical motions described by solutions of Hamiltons equa-
tions lead to approximate solutions of Schrodingers equation. Establishing this relation
between classical and quantum mechanics is important, not only in verifying that the theo-
ries are consistent with the fact that we see classical behavior in systems which are really
governed by quantum mechanics, but also as a tool for developing approximate solutions to
the quantum equations of motion.
What is the meaning of macroscopic in mathematical terms? It turns out that good
approximate solutions of Schrodingers equation can be generated from classical information
when is small enough. But how can a constant with physical dimensions be small?
Although there remain some unsettled issues connected with the question, How can
become small? the answer is essentially the following. For any particular mechanical
system, there are usually characteristic distances, masses, velocities, . . . from which a unit
of action appropriate to the system can be derived, and the classical limit is applicable
when divided by this unit is much less than 1. In these notes, we will often regard
mathematically as a formal parameter or a variable rather than as a xed number.
7
2 The WKB Method
A basic technique for obtaining approximate solutions to the Schrodinger equation from
classical motions is called the WKB method, after Wentzel, Kramers, and Brillouin. (Other
names, including Liouville, Green, and Jereys are sometimes attached to this method.
References [13] and [47] contain a discussion of some of its history. Also see [5], where the
method is traced all the way back to 1817. For convenience, nevertheless, we will still refer
to the method as WKB.) A good part of what is now called microlocal analysis can be
understood as the extension of the basic WKB idea to more precise approximations and
more general situations, so the following discussion is absolutely central to these notes.
2.1 Some Hamilton-Jacobi preliminaries
In this section we will carry out the rst step in the WKB method to obtain an approximate
solution to the stationary state eigenvalue problem arising from the Schrodinger equation.
The geometric interpretation of this technique will lead to a correspondence between classical
and quantum mechanics which goes beyond the one described in Chapter 1.
Consider a 1-dimensional system with hamiltonian
H(q, p) =
p
2
2m
+ V (q),
where V (x) is a potential (for example the potential kx
2
/2 for the harmonic oscillator).
Hamiltons equations now become
q =
p
m
p = V
t
(q).
For xed R
+
, Schrodingers equation assumes the form
i

t
=

H,
where

H =

2
2m

2
x
2
+ m
V
is the Schr odinger operator.
As a rst step toward solving the Schrodinger equation, we look for stationary states,
i.e. solutions of the form
(x, t) = (x) e
it
,
so-called because as time evolves, these solutions keep the same form (up to multiplication by
a complex scalar of norm 1). Substituting this expression for in the Schrodinger equation,
we obtain
(x) e
it
= (

H)(x) e
it
.
Eliminating the factor e
it
above, we arrive at the time-independent Schr odinger equa-
tion:
(

H E) = 0,
8
where E = . This equation means that is to be an eigenfunction of the linear dierential
operator

H; the eigenvalue E represents the energy of the system, which has a denite value
in this state.
Suppose for the moment that the potential V is constant, in which case the force V
t
(x)
is zero, and so we are dealing with a free particle. Trying a solution of the form (x) = e
ix
for some constant , we nd that
(

H E) = 0 ()
2
= 2m(E V ).
For V < E, the (real) value of is thus determined up to a choice of sign, and one has an
abundance of exact solutions of the Schrodinger equation which are oscillatory and bounded.
Such a wave function is not square-integrable and as such is said to be unnormalizable; it
represents a particle which is equally likely to be anywhere in space, but which has a denite
momentum (since it is an eigenfunction of the momentum operator p).
3
When E < V , the
constant is imaginary, and there are only real exponential solutions, which are unbounded
and admit no physical interpretation.
The basic idea at this stage of the WKB method is that, if V varies with x, then should
vary with x as well; a more general solution candidate is then
(x) = e
iS(x)/
,
for some real-valued function S known as a phase function. This proposed form of the
solution is the simplest version of the WKB ansatz, and in this case we have
(

H E) =

(S
t
(x))
2
2m
+ (V E)
i
2m
S
tt
(x)

e
iS(x)/
.
Since we will consider to be small, our rst-order approximation attempt will ignore the
last term in brackets; to kill the other two terms, we require that S satisfy the eikonal or
Hamilton-Jacobi equation:
H(x, S
t
(x)) =
(S
t
(x))
2
2m
+ V (x) = E,
i.e.
S
t
(x) =

2m(E V (x)).
To understand the phase function S geometrically, we consider the classical phase
4
plane
R
2
T

R with coordinates (q, p). The dierential dS = S

t
dx can be viewed as a mapping
dS: R T

R, where as usual we set p = S

t
. Then S satises the Hamilton-Jacobi equation
if and only if the image of dS lies in the level manifold H
1
(E). This observation establishes
a fundamental link between classical and quantum mechanics:
When the image of dS lies in a level manifold of the classical hamiltonian, the
function S may be taken as the phase function of a rst-order approximate solu-
tion of Schrodingers equation.
3
See [55] for a group-theoretic interpretation of such states.
4
These two uses of the term phase seem to be unrelated!
9
The preceding discussion generalizes easily to higher dimensions. In R
n
, the Schrodinger
operator corresponding to the classical hamiltonian
H(q, p) =

p
2
i
2m
+ V (q)
is

H =

2
2m
+ m
V
,
where denotes the ordinary Laplacian. As before, if we consider a WKB ansatz of the
form = e
iS/
, then
(

H E) =

|S|
2
2m
+ (V E)
i
2m
S

e
iS/
will be O() provided that S satises the Hamilton-Jacobi equation:
H

x
1
, . . . , x
n
,
S
x
1
, . . . ,
S
x
n

=
|S(x)|
2
2m
+ V (x) = E.
Since is of order zero in , while (

H E) = O(), the ansatz again constitutes a
rst-order approximate solution to the time-independent Schrodinger equation.
We will call a phase function S : R
n
R admissible if it satises the Hamilton-Jacobi
equation. The image L = im(dS) of the dierential of an admissible phase function S is
characterized by three geometric properties:
1. L is an n-dimensional submanifold of H
1
(E).
2. The pull-back to L of the form
n
=

j
p
j
dq
j
on R
2n
is exact.
3. The restriction of the canonical projection : T

R
n
R
n
to L induces a dieomor-
phism L R
n
. In other words, L is projectable.
While many of the basic constructions of microlocal analysis are motivated by operations
on these projectable submanifolds of T

R
n
R
2n
, applications of the theory require us to
extend the constructions to more general n-dimensional submanifolds of R
2n
satisfying only
a weakened version of condition (2) above, in which exact is replaced by closed. Such
submanifolds are called lagrangian.
For example, the level sets for the 1-dimensional harmonic oscillator are lagrangian sub-
manifolds in the phase plane. A regular level curve of the hamiltonian is an ellipse L. Since
L is 1-dimensional, the pull-back to L of the form p dq is closed, but the integral of p dq
around the curve equals the enclosed nonzero area, so its pull-back to L is not exact. It is
also clear that the curve fails to project dieomorphically onto R. From the classical stand-
point, the behavior of an oscillator is nevertheless completely described by its trajectory,
suggesting that in general the state of a system should be represented by the submanifold
L (projectable or not) rather than by the phase function S. This idea, which we will clarify
later, is the starting point of the geometrical approach to microlocal analysis.
10
For now, we want to note an important relationship between lagrangian submanifolds of
R
2n
and hamiltonian ows. Recall that to a function H : R
2n
R, Hamiltons equations
associate the vector eld
X
H
= q

q
+ p

p
=

j
H
p
j

q
j

H
q
j

p
j
.
A simple computation shows that X
H
and the form
n
are related by the equation
X
H
d
n
= dH,
i.e.
d
n
(X
H
, v) = dH(v)
for every tangent vector v. If L is a lagrangian submanifold of a level set of H, then TL
lies in the kernel of dH at all points of L, or, in other words, the 2-form d
n
vanishes on
the subspace of T
p
R
2n
generated by T
p
L and X
H
(p) for each p L. The restriction of d
n
to the tangent space T
p
R
2n
of R
2n
at any point p denes a nondegenerate, skew-symmetric
bilinear form, and thus, as we will see in the next chapter, subspaces of T
p
R
2n
on which d
n
vanishes can be at most n-dimensional. These remarks imply that X
H
is tangent to L, and
we have the following result.
Hamilton-Jacobi theorem . A function H: R
2n
R is locally constant on a lagrangian
submanifold L R
2n
if and only if the hamiltonian vector eld X
H
is tangent to L.
If the lagrangian submanifold L is locally closed, this theorem implies that L is invariant
under the ow of X
H
.
2.2 The WKB approximation
Returning to our WKB ansatz for a stationary-state solution of the Schrodinger equation,
we recall that if S: R
n
R is an admissible phase function, then (x) = e
iS(x)/
satises
(

H E) = O().
Up to terms of order , in other words, is an eigenfunction of

H with eigenvalue E.
There is no way to improve the order of approximation simply by making a better choice
of S. It is also clear on physical grounds that our ansatz for is too restrictive because
it satises [(x)[ = 1 for all x. In quantum mechanics, the quantity [(x)[
2
represents the
probability of the particle being at the position x, and there is no reason for this to be
constant; in fact, it is at least intuitively plausible that a particle is more likely to be found
where it moves more slowly, i.e., where its potential energy is higher. We may therefore hope
to nd a better approximate solution by multiplying by an amplitude function a
(x) = e
iS(x)/
a(x).
11
If S is again an admissible phase function, we now obtain:
(

H E) =
1
2m

aS + 2

j
a
x
j
S
x
j

+
2
a

e
iS/
.
If a is chosen to kill the coecient of on the right, then will be an eigenfunction of

H
modulo terms of order O(
2
). This condition on a is known as the homogeneous transport
equation:
aS + 2

j
a
x
j
S
x
j
= 0.
If S is an admissible phase function, and a is an amplitude which satises the homogeneous
transport equation, then the second-order solution = e
iS/
a is called the semi-classical
approximation.
Example 2.1 In 1 dimension, the homogeneous transport equation amounts to
aS
tt
+ 2a
t
S
t
= 0.
Solving this equation directly, we obtain
a
2
S
tt
+ 2aa
t
S
t
= (a
2
S
t
)
t
= 0
a =
c

S
t
for some constant c. Since S is assumed to satisfy the Hamilton-Jacobi equation, we have
S
t
=

2m(E V ), and thus

a =
c
[4(E V )]
1
4
.
If E > V (x) for all x R, then a is a smooth solution to the homogeneous transport equation.
Notice that a = [[ is largest where V is largest, as our physical reasoning predicted.
Since the expression above for a does not depend explicitly on the phase function S, we
might naively attempt to use the same formula when im(dS) is replaced by a non-projectable
lagrangian submanifold of H
1
(E). Consider, for example, the unbounded potential V (x) =
x
2
in the case of the harmonic oscillator. For [x[ <

E, the function a is still well-dened

up to a multiplicative constant. At [x[ =

E, however, a has (asymptotic) singularities;

observe that these occur precisely at the projected image of those points of L where the
projection itself becomes singular. Outside the interval [x[

E, the function a assumes

complex values.

To generate better approximate solutions to the eigenfunction problem, we can extend

the procedure above by adding to the original amplitude a = a
0
certain appropriately chosen
functions of higher order in . Consider the next approximation
= e
iS/
(a
0
+ a
1
).
12
Assuming that e
iS/
a
0
is a semi-classical approximate solution, we obtain:
(

H E) =
1
2m

i
2

a
1
S + 2

j
a
1
x
j
S
x
j
ia
0

+
3
a
1

e
iS(x)/
.
Evidently, will be a solution of the time-independent Schrodinger equation modulo terms
of order O(
3
) provided that a
1
satises the inhomogeneous transport equation
a
1
S + 2

j
a
1
x
j
S
x
j
= ia
0
.
In general, a solution to the eigenfunction problem modulo terms of order O(
n
) is given
by a WKB ansatz of the form
= e
iS/
(a
0
+ a
1
+ + a
n

n
),
where S satises the Hamilton-Jacobi equation, a
0
satises the homogeneous transport equa-
tion, and for each k > 0, the function a
k
satises the inhomogeneous transport equation:
a
k
S + 2

j
a
k
x
j
S
x
j
= ia
k1
.
This situation can be described in the terminology of asymptotic series as follows. By an
-dependent function f

on R
n
we will mean a function f : R
n
R
+
C, where is
viewed as a parameter ranging in R
+
. Such a function is said to be represented by a
formal asymptotic expansion of the form

k=0
a
k

k
, where each coecient a
k
is a smooth
complex-valued function on R
n
, if, for each K Z
+
, the dierence
f

k=0
a
k

k
is O(
K+1
) locally uniformly in x. When f

admits such an expansion, its coecients a

k
are
uniquely determined. It is obvious that any -dependent function which extends smoothly
to = 0 is represented by an asymptotic series, and a theorem of E.Borel (see [28, p.28])
tells us that, conversely, any asymptotic series can be summed to yield such a function.
The principal part of an asymptotic series

k=0
a
k

k
is dened as its rst term which is
not identically zero as a function of x. The order of a is the index of its principal part.
If we consider as equivalent any two -dependent functions whose dierence is O(

), i.e.
O(
k
) for all k, then each asymptotic series determines a unique equivalence class. A WKB
solution to the eigenfunction problem

H = E is then an equivalence class of functions
of the form
= e
iS/
a,
where S is an admissible phase function and a is an -dependent function represented by a
formal asymptotic series
a

k=p
a
k

k
13
with the property that its principal part a
p
satises the homogeneous transport equation
a
p
S + 2

j
a
p
x
j
S
x
j
= 0,
and for k > p, the a
k
satisfy the recursive transport equations:
a
k
S + 2

j
a
k
x
j
S
x
j
= ia
k1
.
This means that the -dependent function (or any function equivalent to it) satises the
Schrodinger equation up to terms of order O(

).
Geometry of the transport equation
In Section 2.1, we saw that a rst-order WKB approximate solution = e
iS/
to the time-
independent Schrodinger equation depended on the choice of an admissible phase function,
i.e., a function S satisfying the Hamilton-Jacobi equation H(x,
S
x
) = E. The generalized or
geometric version of such a solution was a lagrangian submanifold of the level set H
1
(E).
We now wish to interpret and generalize in a similar way the semi-classical approximation
with its amplitude included.
To begin, suppose that a is a function on R
n
which satises the homogeneous transport
equation:
aS + 2

j
a
x
j
S
x
j
= 0.
After multiplying both sides of this equation by a, we can rewrite it as:

x
j

a
2
S
x
j

= 0,
which means that the divergence of the vector eld a
2
S is zero. Rather than considering
the transport equation as a condition on the vector eld a
2
S (on R
n
) per se, we can lift all
of this activity to the lagrangian submanifold L = im(dS). Notice rst that the restriction
to L of the hamiltonian vector eld associated to H(q, p) =

p
2
i
/2 +V (q) is
X
H
[
L
=

S
x
j

q
j

V
q
j

p
j

.
The projection X
(x)
H
of X
H
[
L
onto R
n
(the (x) reminds us of the coordinate x on R
n
) there-
fore coincides with S, and so the homogeneous transport equation says that a
2
X
(x)
H
is
divergence-free for the canonical density [dx[ = [dx
1
dx
n
[ on R
n
. But it is better to
reformulate this condition as:
L
X
(x)
H
(a
2
[dx[) = 0;
14
that is, we transfer the factor of a
2
from the vector eld X
(x)
H
= S to the density [dx[. Since
X
H
is tangent to L by the Hamilton-Jacobi theorem, and since the Lie derivative is invariant
under dieomorphism, this equation is satised if and only if the pull-back of a
2
[dx[ to L
via the projection is invariant under the ow of X
H
.
This observation, together with the fact that it is the square of a which appears in the
density

(a
2
[dx[), suggests that a solution of the homogeneous transport equation should
be represented geometrically by a half-density on L, invariant by X
H
. (See Appendix A for
a discussion of densities of fractional order.)
In other words, a (geometric) semi-classical state should be dened as a lagrangian sub-
manifold L of R
2n
equipped with a half-density a. Such a state is stationary when L lies in
a level set of the classical hamiltonian and a is invariant under its ow.
Example 2.2 Recall that in the case of the 1-dimensional harmonic oscillator, stationary
classical states are simply those lagrangian submanifolds of R
2
which coincide with the
regular level sets of the classical hamiltonian H(q, p) = (p
2
+ kq
2
)/2. There is a unique (up
to a constant) invariant volume element for the hamiltonian ow of H on each level curve
H. Any such level curve L, together with a square root of this volume element, constitutes
a semi-classical stationary state for the harmonic oscillator.

Notice that while a solution to the homogeneous transport equation in the case of the 1-
dimensional harmonic oscillator was necessarily singular (see Example 2.1), the semi-classical
state described in the preceding example is a perfectly smooth object everywhere on the
lagrangian submanifold L. The singularities arise only when we try to transfer the half-
density from L down to conguration space. Another substantial advantage of the geometric
interpretation of the semi-classical approximation is that the concept of an invariant half-
density depends only on the hamiltonian vector eld X
H
and not on the function S, so it
makes sense on any lagrangian submanifold of R
2n
lying in a level set of H.
This discussion leads us to another change of viewpoint, namely that the quantum states
themselves should be represented, not by functions, but by half-densities on conguration
space R
n
, i.e. elements of the intrinsic Hilbert space H
R
n (see Appendix A). Stationary states
are then eigenvectors of the Schrodinger operator

H, which is dened on the space of smooth
half-densities in terms of the old Schrodinger operator on functions, which we will denote
momentarily as

H
fun
, by the equation

H(a[dx[
1/2
) = (

H
fun
a)[dx[
1/2
.
From this new point of view, we can express the result of our analysis as follows:
If S is an admissible phase function and a is a half-density on L = im(dS) which
is invariant under the ow of the classical hamiltonian, then e
iS/
(dS)

a is a
second-order approximate solution to the time-independent Schrodinger equation.
15
In summary, we have noted the following correspondences between classical and quantum
mechanics:
Object Classical version Quantum version
basic space R
2n
H
R
n
state lagrangian submanifold of R
2n
with half-density on R
n
half-density
time-evolution Hamiltons equations Schrodinger equation
generator of evolution function H on R
2n
operator

H on smooth
half-densities
stationary state lagrangian submanifold in level set eigenvector of

H
of H with invariant half-density
Proceeding further, we could attempt to interpret a solution of the recursive system of
inhomogeneous transport equations on R
n
as an asymptotic half-density on L in order to
arrive at a geometric picture of a complete WKB solution to the Schrodinger equation. This,
however, involves some additional diculties, notably the lack of a geometric interpretation
of the inhomogeneous transport equations, which lie beyond the scope of these notes. Instead,
we will focus on two aspects of the semi-classical approximation. First, we will extend the
geometric picture presented above to systems with more general phase spaces. This will
require the concept of symplectic manifold, which is introduced in the following chapter.
Second, we will quantize semi-classical states in these symplectic manifolds. Specically,
we will try to construct a space of quantum states corresponding to a general classical
phase space. Then we will try to construct asymptotic quantum states corresponding to
half-densities on lagrangian submanifolds. In particular, we will start with an invariant
half-density on a (possibly non-projectable) lagrangian submanifold of R
2n
and attempt to
use this data to construct an explicit semi-classical approximate solution to Schrodingers
equation on R
n
.
16
3 Symplectic Manifolds
In this chapter, we will introduce the notion of a symplectic structure on a manifold, moti-
vated for the most part by the situation in R
2n
. While some discussion will be devoted to
certain general properties of symplectic manifolds, our main goal at this point is to develop
the tools needed to extend the hamiltonian viewpoint to phase spaces associated to general
nite-dimensional conguration spaces, i.e. to cotangent bundles. More general symplectic
manifolds will reappear as the focus of more sophisticated quantization programs in later
chapters. We refer to [6, 29, 63] for thorough discussions of the topics in this chapter.
3.1 Symplectic structures
In Section 2.1, a lagrangian submanifold of R
2n
was dened as an n-dimensional submanifold
L R
2n
on which the exterior derivative of the form
n
=

p
i
dq
i
vanishes; to a function
H: R
2n
R, we saw that Hamiltons equations associate a vector eld X
H
on R
2n
satisfying
X
H
d
n
= dH.
Finally, our proof of the Hamilton-Jacobi theorem relied on the nondegeneracy of the 2-
form d
n
. These points already indicate the central role played by the form d
n
in the
study of hamiltonian systems in R
2n
; the correct generalization of the hamiltonian picture to
arbitrary conguration spaces relies similarly on the use of 2-forms with certain additional
properties. In this section, we rst study such forms pointwise, collecting pertinent facts
about nondegenerate, skew-symmetric bilinear forms. We then turn to the denition of
symplectic manifolds.
Linear symplectic structures
Suppose that V is a real, m-dimensional vector space. A bilinear form : V V R gives
rise to a linear map
: V V

dened by contraction:
(x)(y) = (x, y).
The -orthogonal to a subspace W V is dened as
W

= x V : W ker (x).
If is an isomorphism, or in other words if V

= 0, then the form is said to be nonde-

generate; if in addition is skew-symmetric, then is called a linear symplectic structure on
V . A linear endomorphism of a symplectic vector space (V, ) which preserves the form
is called a linear symplectic transformation, and the group of all such transformations
is denoted by Sp(V ).
17
Example 3.1 If E is any real n-dimensional vector space with dual E

, then a linear sym-

plectic structure on V = E E

is given by
((x, ), (x
t
,
t
)) =
t
(x) (x
t
).
With respect to a basis x
i
of E and a dual basis
i
of E

, the form is represented by

the matrix
=

0 I
I 0

.
It follows that if a linear operator on V is given by the real 2n 2n matrix
T =

A B
C D

,
then T is symplectic provided that A
t
C, B
t
D are symmetric, and A
t
D C
t
B = I. Note in
particular that these conditions are satised if A GL(E), D = (A
t
)
1
, and B = C = 0,
and so GL(E) is isomorphic to a subgroup Gl(E) of Sp(V ). More generally, if K : E F
is an isomorphism, then the association
(x, ) (Kx, (K
1
)

)
denes a linear symplectomorphism between E E

and F F

equipped with these linear

symplectic structures.

Since the determinant of a skew-symmetric mm matrix is zero if m is odd, the existence

of a linear symplectic structure on a vector space V implies that V is necessarily even-
dimensional and therefore admits a complex structure, i.e. a linear endomorphism J such
that J
2
= I. A complex structure is said to be compatible with a symplectic structure on
V if
(Jx, Jy) = (x, y)
and
(x, Jx) > 0
for all x, y V . In other words, J is compatible with (we also call it -compatible)
if J : V V is a linear symplectomorphism and g
J
(, ) = (, J) denes a symmetric,
positive-denite bilinear form on V .
Theorem 3.2 Every symplectic vector space (V, ) admits a compatible complex structure.
Proof. Let ' , ` be a symmetric, positive-denite inner product on V , so that is represented
by an invertible skew-adjoint operator K: V V ; i.e.
(x, y) = 'Kx, y`.
18
The operator K admits a polar decomposition K = RJ, where R =

KK
t
is positive-
denite symmetric, J = R
1
K is orthogonal, and RJ = JR. From the skew-symmetry of K
it follows that J
t
= J, and so J
2
= JJ
t
= id; i.e., J is a complex structure on V .
To see that J is -compatible, rst note that
(Jx, Jy) = 'KJx, Jy` = 'JKx, Jy` = 'Kx, y` = (x, y).
Also,
(x, Jx) = 'Kx, Jx` = 'JRx, Jx` = 'Rx, x` > 0,
since R and ' , ` are positive-denite.
2
Corollary 3.3 The collection . of -compatible complex structures on a symplectic vector
space (V, ) is contractible.
Proof. The association J g
J
described above denes a continuous map from . into the
space { of symmetric, positive-denite bilinear forms on V . By the uniqueness of the polar
decomposition, it follows that the map which assigns to a form ' , ` the complex structure J
constructed in the preceding proof is continuous, and the composition . { . of these
maps equals the identity on .. Since { is contractible, this implies the corollary.
2
If J is -compatible, a hermitian structure on V is dened by
', ` = g
J
(, ) + i(, ).
As is easily checked, a linear transformation T GL(V ) which preserves any two of the
structures , g
J
, J on V preserves the third and therefore preserves the hermitian structure.
In terms of the automorphism groups Sp(V ), GL(V, J), O(V ), and U(V ) of , J, g, and ', `,
this means that the intersection of any two of Sp(V ), GL(V, J), O(V ) equals U(V ).
To determine the Lie algebra sp(V) of Sp(V ), consider a 1-parameter family of maps e
tA
associated to some linear map A: V V . For any v, w V , we have
d
dt

t=0
(e
tA
v, e
tA
w) = (Av, w) + (v, Aw),
and so A sp(V) if and only if the linear map A: V V

is self-adjoint. Consequently,
dim(V ) = 2k implies dim(sp(V)) = dim(Sp(V)) = k(2k + 1).
Distinguished subspaces
The -orthogonal to a subspace W of a symplectic vector space (V, ) is called the symplectic
orthogonal to W. From the nondegeneracy of the symplectic form, it follows that
W

= W and dimW

= dimV dimW
19
for any subspace W V . Also,
(A + B)

= A

and (A B)

= A

+ B

for any pair of subspaces A, B of V . In particular, B

whenever A B.
Note that the symplectic orthogonal W

might not be an algebraic complement to W.

For instance, if dimW = 1, the skew-symmetry of implies that W W

. More generally,
any subspace contained in its orthogonal will be called isotropic. Dually, we note that if
codimW = 1, then W

is 1-dimensional, hence isotropic, and W

= W. In
general, spaces W satisfying the condition W

W are called coisotropic or involutive.

Finally, if W is self-orthogonal, i.e. W

= W, then the dimension relation above implies that

dimW =
1
2
dimV . Any self-orthogonal subspace is simultaneously isotropic and coisotropic,
and is called lagrangian.
According to these denitions, a subspace W V is isotropic if the restriction of the
symplectic form to W is identically zero. At the other extreme, the restriction of to certain
subspaces Z V may again be nondegenerate; this is equivalent to saying that ZZ

= 0
or Z + Z

= V . Such subspaces are called symplectic.

Example 3.4 In EE

with its usual symplectic structure, both E and E

are lagrangian
subspaces. It also follows from the denition of this structure that the graph of a linear map
B: E E

is a lagrangian subspace of E E

if and only if B is self-adjoint.

If (V, ) is a symplectic vector space, we denote by V V the vector space V V equipped
with the symplectic structure . If T : V V is a linear symplectic map, then the
graph of T is a lagrangian subspace of V V .
The kernel of a nonzero covector V

is a codimension-1 coisotropic subspace ker of

V whose symplectic orthogonal (ker )

is the distinguished 1-dimensional subspace of ker

spanned by
1
().

Example 3.5 Suppose that (V, ) is a 2n-dimensional symplectic vector space and W V
is any isotropic subspace with dim(W) < n. Since 2n = dim(W) + dim(W

), there exists a
nonzero vector w W

`W. The subspace W

t
of V spanned by Ww is then isotropic and
dim(W
t
) = dim(W)+1. From this observation it follows that for every isotropic subspace W
of a (nite-dimensional) symplectic vector space V which is not lagrangian, there exists an
isotropic subspace W
t
of V which properly contains W. Beginning with any 1-dimensional
subspace of V , we can apply this remark inductively to conclude that every nite-dimensional
symplectic vector space contains a lagrangian subspace.

Various subspaces of a symplectic vector space are related as follows.

Lemma 3.6 If L is a lagrangian subspace of a symplectic vector space V , and A V is an
arbitrary subspace, then:
20
1. L A if and only if A

L.
2. L is transverse to A if and only if L A

= 0.
Proof. Statement (1) follows from the properties of the operation

and the equation L = L

.
Similarly, L + A = V if and only if (L + A)

= L A

= 0, proving statement (2).

2
Example 3.7 Note that statement (1) of Lemma 3.6 implies that if L A, then A

A,
and so A is a coisotropic subspace. Conversely, if A is coisotropic, then A

is isotropic,
and Example 3.5 implies that there is a lagrangian subspace L with A

L. Passing to
orthogonals, we have L A. Thus, a subspace C V is coisotropic if and only if it contains
a lagrangian subspace.
Suppose that V is a symplectic vector space with an isotropic subspace I and a lagrangian
subspace L such that I L = 0. If W L is any complementary subspace to I

L, then
I + L W + I

, and so W

I I

L. Thus, W

I I L = 0. Since I

W = 0
by our choice of W, it follows that I + W is a symplectic subspace of V .

A pair L, L
t
of transverse lagrangian subspaces of V is said to dene a lagrangian
splitting of V . In this case, the map denes an isomorphism L
t
L

, which in turn
gives rise to a linear symplectomorphism between V and LL

equipped with its canonical

symplectic structure (see Example 3.1). If J is a -compatible complex structure on V and
L V a lagrangian subspace, then L, JL is a lagrangian splitting. By Example 3.5, every
symplectic vector space contains a lagrangian subspace, and since every n-dimensional vector
space is isomorphic to R
n
, the preceding remarks prove the following linear normal form
result:
Theorem 3.8 Every 2n-dimensional symplectic vector space is linearly symplectomorphic
to (R
2n
,
n
).
Theorem 3.2 also implies the following useful result.
Lemma 3.9 Suppose that V is a symplectic vector space with a -compatible complex struc-
ture J and let T

: V V be given by T

(x) = x + Jx.
1. If L, L
t
are any lagrangian subspaces of V , then L

= T

(L) is a lagrangian subspace

transverse to L
t
for small > 0.
2. For any two lagrangian subspaces L, L
t
of V , there is a lagrangian subspace L
tt
trans-
verse to both L
t
and L.
Proof. It is easy to check that T

is a conformal linear symplectic map, i.e. an isomorphism

of V satisfying (T

x, T

y) = (1 +
2
) (x, y). Thus, L

is a lagrangian subspace for all

> 0. Using the inner-product g
J
on V induced by J, we can choose orthonormal bases
21
v
i
, w
i
of L
t
and L, respectively, so that for i = 1, , k, the vectors v
i
= w
i
span LL
t
.
Then w
i
+ Jw
i
form a basis of L

, and L
t
, L

are transverse precisely when the matrix

M = (v
i
, w
j
+ Jw
j
) = (v
i
, w
j
) + g
J
(v
i
, w
j
) is nonsingular. Our choice of bases
implies that
M =

id 0
0 A + B

where A = (v
i
, w
j
)
n
i,j=k+1
and B is some (nk)(nk) matrix. Setting I = spanv
i

n
i=k+1
and W = spanw
i

n
i=k+1
, we can apply Example 3.7 to conclude that A is nonsingular, and
assertion (1) follows.
To prove (2), observe that for small > 0, the lagrangian subspace L

is transverse to
L, L
t
by (1).
2
In fact, the statement of preceding lemma can be improved as follows. Let L
i
be a count-
able family of lagrangian subspaces, and let A
i
, B
i
be the matrices obtained with respect to L
as in the proof above. For each i, the function t det(A
i
+tB
i
) is a nonzero polynomial and
therefore has nitely many zeros. Consequently, the lagrangian subspace T
t
(L) is transverse
to all L
i
for almost every t R.
The lagrangian grassmannian
The collection of all unoriented lagrangian subspaces of a 2n-dimensional symplectic vector
space V is called the lagrangian grassmannian L(V ) of V . A natural action of the group
Sp(V ) on L(V ), denoted : Sp(V ) L(V ) L(V ) is dened by (T, L) =
L
(T) = T(L).
Lemma 3.10 The unitary group associated to an -compatible complex structure J on V
acts transitively on L(V ).
Proof. For arbitrary L
1
, L
2
L(V ), an orthogonal transformation L
1
L
2
induces a
symplectic transformation L
1
L

1
L
2
L

2
in the manner of Example 3.1, which in turn
gives rise to a unitary transformation L
1
JL
1
L
2
JL
2
mapping L
1
onto L
2
.
2
The stabilizer of L L(V ) under the U(V )-action is evidently the orthogonal subgroup of
Gl(L) dened with respect to the inner-product and splitting LJL of V induced by J (see
Example 3.1). Thus, a (non-canonical) identication of the lagrangian grassmannian with
the homogeneous space U(n)/O(n) is obtained from the map
U(V )

L
L(V ).
The choice of J also denes a complex determinant U(V )
det
2
J
S
1
, which induces a bration
L(V ) S
1
with 1-connected ber SU(n)/SO(n), giving an isomorphism of fundamental
groups

1
(L(V ))
1
(S
1
) Z.
22
This isomorphism does not depend on the choices of J and L made above. Independence of J
follows from the fact that . is connected (Corollary 3.3). On the other hand, connectedness
of the unitary group together with Lemma 3.10 gives independence of L.
Passing to homology and dualizing, we obtain a natural homomorphism
H
1
(S
1
; Z) H
1
(L(V ); Z).
The image of the canonical generator of H
1
(S
1
; Z) under this map is called the universal
Maslov class,
V
. The result of the following example will be useful when we extend our
discussion of the Maslov class from vector spaces to vector bundles.
Example 3.11 If (V, ) is any symplectic vector space with -compatible complex structure
J and lagrangian subspace L, then a check of the preceding denitions shows that
m
L
((T, L
t
)) = m
L
(L
t
) det
2
J
(T)
for any T U(V ) and L
t
L(V ). (Recall that : Sp(V )L(V ) L(V ) denotes the natural
action of Sp(V ) on L(V )).
Now consider any topological space M. If f
1
, f
2
: M L(V ) are continuous maps, then
the denition of the universal Maslov class shows that (f

1
f

2
)
V
equals the pull-back of
the canonical generator of H
1
(S
1
; R) by the map
(m
L
f
1
)(m
L
f
2
)
1
.
(Here we use the fact that when S
1
is identied with the unit complex numbers, the mul-
tiplication map S
1
S
1
S
1
induces the diagonal map H
1
(S
1
) H
1
(S
1
) H
1
(S
1
)
H
1
(S
1
S
1
) on cohomology). If T : M Sp(V ) is any map, we set (T f
i
) = (T, f
i
).
Since Sp(V ) deformation retracts onto U(V ), it follows that T is homotopic to a map
T
t
: M U(V ), and so ((T f
1
)

(T f
2
)

)
V
is obtained via pull-back by
(m
L
(T
t
f
1
))(m
L
(T
t
f
2
))
1
.
From the rst paragraph, it follows that this product equals (m
L
f
1
)(m
L
f
2
)
1
, from which
we conclude that
(f

1
f

2
)
V
= ((T f
1
)

(T f
2
)

)
V
.

Symplectic manifolds
To motivate the denition of a symplectic manifold, we rst recall some features of the
dierential form d
n
=
n
=

n
j=1
dq
j
dp
j
which appeared in our earlier discussion.
First, we note that

n
j=1
dq
j
dp
j
denes a linear symplectic structure on the tangent space
of R
2n
at each point. In fact:

q
j
,

p
k

=
jk

n

q
j
,

q
k

= 0
n

p
j
,

p
k

= 0
23
and so

n

q
j

= dp
j

n

p
j

= dq
j
,
from which it is clear that
n
is invertible.
Next, we recall that the hamiltonian vector eld associated via Hamiltons equations to
H : R
2n
R satises
X
H

n
= dH,
or in other words,
X
H
=
1
n
(dH),
so we see that the symplectic form
n
is all that we need to obtain X
H
from H. This
description of the hamiltonian vector eld leads immediately to the following two invariance
results. First note that by the skew-symmetry of
n
,
L
X
H
H = X
H
H =
n
(X
H
, X
H
) = 0,
implying that X
H
is tangent to the level sets of H. This again reects the fact that the ow
of X
H
preserves energy. Since
n
is closed, we also have by Cartans formula (see [1])
L
X
H

n
= d(X
H

n
) + X
H
d
n
= d
2
H = 0.
This equation implies that the ow of X
H
preserves the form
n
and therefore generalizes our
earlier remark that the hamiltonian vector eld associated to the 1-dimensional harmonic
oscillator is divergence-free.
We now see what is needed to do hamiltonian mechanics on manifolds. A 2-form on a
manifold P is a smooth family of bilinear forms on the tangent spaces of P. By assuming that
each of these bilinear forms is nondegenerate, we guarantee that the equation X
H
=
1
(dH)
denes a hamiltonian vector eld uniquely for any H. Computing the Lie derivative of H
with respect to X
H
L
X
H
H = X
H
H = (X
H
)(X
H
) = (X
H
, X
H
) = 0,
we see that the conservation of energy follows from the skew-symmetry of the form .
Finally, invariance of under the hamiltonian ow is satised if
L
X
H
= d(X
H
) + X
H
d = 0.
Here, the term d(X
H
) = d
2
H is automatically zero; to guarantee the vanishing of the
second term, we impose the condition that be closed.
Thus we make the following denition:
Denition 3.12 A symplectic structure on a manifold P is a closed, nondegenerate
2-form on P.
24
The condition that be nondegenerate means that denes an isomorphism of vector
bundles TP T

P, or equivalently, that the top exterior power of is a volume form on

P, or nally, that denes a linear symplectic structure on each tangent space of P.
An immediate example of a symplectic manifold is furnished by R
2n
with its standard
structure
n
=

n
j=1
dq
j
dp
j
(a dierential form with constant coecients and not just
a single bilinear form). Darbouxs theorem (Section 4.3) will tell us that this is the local
model for the general case. In the next section, we will see that the cotangent bundle of any
smooth manifold carries a natural symplectic structure.
Generalizing our earlier discussion of distinguished subspaces of a symplectic vector space,
we call a submanifold C P (co-)isotropic provided that each tangent space T
p
C of C is a
(co-)isotropic subspace of T
p
P. When C is coisotropic, the subspaces (
p
= (T
p
C)

comprise
a subbundle (TC)

of TC known as the characteristic distribution of C. It is integrable

because is closed. Of particular interest in our discussion will be lagrangian submanifolds
of P, which are (co-)isotropic submanifolds of dimension
1
2
dim(P). More generally, if L is a
smooth manifold of dimension
1
2
dim(P) and : L P is an immersion such that

= 0,
we will call the pair (L, ) a lagrangian immersion.
Example 3.13 If C P is a hypersurface, then C is a coisotropic submanifold. A simple
check of denitions shows that if H: P R is a smooth function having C as a regular level
set, then the hamiltonian vector eld X
H
is tangent to the characteristic foliation of C.
If (L, ) is a lagrangian immersion whose image is contained in C, then Lemma 3.6 implies
that for each p L, the characteristic subspace (
(p)
T
(p)
C is contained in

T
p
L, and thus
X
H
induces a smooth, nonsingular vector eld X
H,
on L. In view of the remarks above, this
assertion generalizes the Hamilton-Jacobi theorem (see the end of Section 2.1) to arbitrary
symplectic manifolds and lagrangian immersions.

New symplectic manifolds can be manufactured from known examples by dualizing and by
taking products. The symplectic dual of a manifold (P, ) consists of the same underlying
manifold endowed with the symplectic structure . Evidently P and its dual P share the
same (co-)isotropic submanifolds. Given two symplectic manifolds (P
1
,
1
) and (P
2
,
2
), their
product P
1
P
2
admits a symplectic structure given by the sum
1

2
. More explicitly, this
form is the sum of the pull-backs of
1
and
2
by the projections of P
1
P
2
to P
1
and P
2
. As
is easily veried, the product of (co-)isotropic submanifolds of P
1
and P
2
is a (co-)isotropic
submanifold of P
1
P
2
.
A symplectomorphism from (P
1
,
1
) to (P
2
,
2
) is a smooth dieomorphism f : P
1

P
2
compatible with the symplectic structures: f

2
=
1
. A useful connection among duals,
products, and symplectomorphisms is provided by the following lemma.
Lemma 3.14 A dieomorphism f : P
1
P
2
between symplectic manifolds is a symplecto-
morphism if and only if its graph is a lagrangian submanifold of the product P
2
P
1
.
The collection Aut(P, ) of symplectomorphisms of P becomes an innite-dimensional Lie
group when endowed with the C

topology (see [49]). In this case, the corresponding Lie

25
algebra is the space (P, ) of smooth vector elds X on P satisfying
L
X
= 0.
Since L
X
= d(X ), the association X X denes an isomorphism between (P, )
and the space of closed 1-forms on P; those X which map to exact 1-forms are simply the
hamiltonian vector elds on P. The elements of (P, ) are called locally hamiltonian
vector elds or symplectic vector elds.
Example 3.15 A linear symplectic form on a vector space V induces a symplectic struc-
ture (also denoted ) on V via the canonical identication of TV with V V . The symplectic
group Sp(V ) then embeds naturally in Aut(V, ), and the Lie algebra sp(V) identies with
the subalgebra of (V, ) consisting of vector elds of the form
X(v) = Av
for some A sp(V). Note that these are precisely the hamiltonian vector elds of the
homogeneous quadratic polynomials on V , i.e. functions satisfying Q(tv) = t
2
Q(v) for all
real t. Consequently, sp(V) is canonically identied with the space of such functions via the
correspondence
A Q
A
(v) =
1
2
(Av, v).

Symplectic vector bundles

Since a symplectic form on a 2n-manifold P denes a smooth family of linear symplectic
forms on the bers of TP, the frame bundle of P can be reduced to a principal Sp(n) bundle
over P. More generally, any vector bundle E B with this structure is called a symplectic
vector bundle. Two symplectic vector bundles E, F are said to be symplectomorphic if
there exists a vector bundle isomorphism E F which preserves their symplectic structures.
Example 3.16 If F B is any vector bundle, then the sum F F

carries a natural
symplectic vector bundle structure, dened in analogy with Example 3.1.

With the aid of an arbitrary riemannian metric, the proof of Theorem 3.2 can be gener-
alized by a berwise construction as follows.
Theorem 3.17 Every symplectic vector bundle admits a compatible complex vector bundle
structure.
Example 3.18 Despite Theorem 3.17, there exist examples of symplectic manifolds which
are not complex (the almost complex structure coming from the theorem cannot be made
integrable), and of complex manifolds which are not symplectic. (See [27] and the numerous
earlier references cited therein.) Note, however, that the Kahler form of any Kahler manifold
is a symplectic form.
26

A lagrangian subbundle of a symplectic vector bundle E is a subbundle L E such

that L
x
is a lagrangian subspace of E
x
for all x B. If E admits a lagrangian subbundle,
then E is symplectomorphic to LL

, and the frame bundle of E admits a further reduction

to a principal GL(n) bundle over B (compare Example 3.1).
Example 3.19 If L is a lagrangian submanifold of a symplectic manifold P, then the re-
stricted tangent bundle T
L
P is a symplectic vector bundle over L, and TL T
L
P is a
lagrangian subbundle. Also note that if C P is any submanifold such that TC contains a
lagrangian subbundle of T
C
P, then C is coisotropic (see Lemma 3.6).

In general, the automorphism group of a symplectic vector bundle E does not act tran-
sitively on the lagrangian subbundles of E. Nevertheless, a pair of transverse lagrangian
subbundles can be related as follows.
Theorem 3.20 Let E B be a symplectic vector bundle and suppose that L, L
t
are la-
grangian subbundles such that L
x
is transverse to L
t
x
for each x B. Then there exists a
compatible complex structure J on E satisfying JL = L
t
.
Proof. Let J
0
be any compatible complex structure on E. Since L
t
and J
0
L are both
transverse to L, we can nd a symplectomorphism T : L L
t
L J
0
L which preserves
the subbundle L and maps L
t
to J
0
L. A simple check of the denition then shows that
J = T
1
J
0
T is a compatible complex structure on E which satises JL = L
t
.
2
Example 3.21 If E is a symplectic vector bundle over M, then any pair L, L
t
of lagrangian
subbundles of E dene a cohomology class (L, L
t
) H
1
(M; Z) as follows.
Assuming rst that E admits a symplectic trivialization f : E M V for some
symplectic vector space V , we denote by f
L
, f
L
: M L(V ) the maps induced by the
lagrangian subbundles f(L), f(L
t
) of M V . Then
(L, L
t
) = (f

L
f

L
)
V
,
where
V
H
1
(L(V ); Z) is the universal Maslov class. From Example 3.11 it follows that
this class is independent of the choice of trivialization f.
For nontrivial E, we note that since the symplectic group Sp(V ) is connected, it follows
that for any loop : S
1
M, the pull-back bundle

E is trivial. Thus (L, L

t
) is well-
dened by the requirement that for every smooth loop in M,

(L, L
t
) = (

L,

L
t
).

27
Example 3.22 As a particular case of Example 3.21, we consider the symplectic manifold
R
2n
T

R
n
with its standard symplectic structure. Then the tangent bundle T(R
2n
) is
a symplectic vector bundle over R
2n
with a natural vertical lagrangian subbundle V R
n
dened as the kernel of

, where : T

R
n
R
n
is the natural projection.
If : L R
2n
is a lagrangian immersion, then the symplectic vector bundle

T(T

(R
n
))
has two lagrangian subbundles, the image L
1
of

: TL

T(T

(R
n
)) and L
2
=

V R
n
.
The class
L,
= (L
1
, L
2
) H
1
(L; Z) is called the Maslov class of (L, ).
A check of these denitions shows that the Maslov class of (L, ) equals the pull-back of
the universal Maslov class
n
H
1
(L(R
2n
); Z) by the Gauss map G: L L(R
2n
) dened
by G(p) =

T
p
L R
2n
. (See [3] for an interpretation of the Maslov class of a loop in
L as an intersection index of the loop G with a singular subvariety in the lagrangian
grassmannian).

3.2 Cotangent bundles

The cotangent bundle T

M of any smooth manifold M is equipped with a natural 1-form,

known as the Liouville form, dened by the formula

M
((x, b))(v) = b(

v),
where : T

M M is the canonical projection. In local coordinates (x

1
, , x
n
) on M and
corresponding coordinates (q
1
, , q
n
, p
1
, , p
n
) on T

M, the equations
q
j
(x, b) = x
j
(x) p
j
(x, b) = b

x
j

imply that

M
=
n

j=1
p
j
dq
j
.
Thus, d
M
=

n
j=1
dq
j
dp
j
in these coordinates, from which it follows that the form

M
= d
M
is a symplectic structure on T

M. Note that if M = R
n
, then
M
is just the
symplectic structure
n
on T

R
n
R
2n
discussed previously, and
M
=
n
.
Lagrangian immersions and the Liouville class
Given a lagrangian immersion : L T

M, we set
L
= , where : T

M M
is the natural projection. Critical points and critical values of
L
are called respectively
singular points and caustic points of L. Finally, (L, ) is said to be projectable if

L
is a dieomorphism. A nice property of the Liouville 1-form is that it can be used to
parametrize the set of projectable lagrangian submanifolds. To do this, we use the notation

to denote a 1-form on M when we want to think of it as a map from M to T

M.
28
Lemma 3.23 Let
1
(M). Then

M
= .
Proof. Because

is a section of T

M, it satises

= id
M
. By the denition of
M
, it
follows that for each v T
p
M,

M
(p)(v) =
M
(

(p))(

v) = '

(p),

v)` = '

(p), v`.
2
For this reason,
M
is often described as the tautological 1-form on T

M. Taking exterior
derivatives on both sides of the equation in Lemma 3.23, we get
d = d

M
=

d
M
=

M
.
From this equation we see that the image of is a lagrangian submanifold of T

M precisely
when the form is closed. This proves
Proposition 3.24 The relation (M,

) denes a natural bijective correspondence be-

tween the the vector space of closed 1-forms on M and the set of projectable lagrangian
submanifolds of T

M.
Generalizing our WKB terminology, we will call S : M R a phase function for a
projectable lagrangian embedding (L, ) T

M provided that (L) = dS(M). The preceding

remarks imply a simple link between phase functions and the Liouville form:
Lemma 3.25 If (L, ) T

M is a projectable lagrangian embedding, then S : M R is a

phase function for L if and only if d(S
L
) =

M
.
Thus, L is the image of an exact 1-form on M if and only if the restriction of the Liouville
form to L is itself exact. This motivates the following denition.
Denition 3.26 If L, M are n-manifolds and : L T

M is an immersion such that

M
is exact, then is called an exact lagrangian immersion.
If : L T

M is an exact lagrangian immersion, then Lemma 3.25 suggests that the

primitive of

M
is a sort of generalized phase function for (L, ) which lives on the manifold
L itself. We will return to this important viewpoint in the next chapter.
Example 3.27 A simple application of Stokes theorem shows that an embedded circle in
the phase plane cannot be exact, although it is the image of an exact lagrangian immersion
of R.
A general class of exact lagrangian submanifolds can be identied as follows. Associated
to a smooth submanifold N M is the submanifold
N

= (x, p) T

M : x N, T
x
N ker(p),
known as the conormal bundle to N. From this denition it follows easily that dimT

M =
2 dimN

, while the Liouville form of T

M vanishes on N

for any N.
If T is a smooth foliation of M, then the union of the conormal bundles to the leaves of T
is a smooth submanifold of T

M foliated by lagrangian submanifolds and is thus coisotropic

(see Example 3.19).
29

Although many lagrangian immersions : L T

M are not exact, the form

M
is
always closed, since d

M
=

M
= 0. The deRham cohomology class
L,
H
1
(L; R)
induced by this form will play an important role in the quantization procedures of the next
chapter and is known as the Liouville class of (L, ).
Example 3.28 To generalize the picture described in Example 3.27, we consider a smooth
manifold M, together with a submanifold N M and a closed 1-form on N. Then
N

= (x, p) T

M : x N p[
TxN
= (x)
is a lagrangian submanifold of T

M whose Liouville class equals [

N
] H
1
(N

; R), where

N
: N

N is here the restriction of the natural projection : T

M M.

Fiber-preserving symplectomorphisms
On each ber of the projection : T

M M, the pull-back of
M
vanishes, so the bers are
lagrangian submanifolds. Thus, the vertical bundle V M = ker

is a lagrangian subbundle
of T(T

M). Since
M
vanishes on the zero section Z
M
T

M, it follows that Z
M
is
lagrangian as well, and the subbundles TZ
M
and V M dene a canonical lagrangian splitting
of T(T

M) over Z
M
.
A 1-form on M denes a dieomorphism f

of T

M by ber-wise ane translation

f

(x, p) = (x, p + (x)).

It is easy to see that this map satises
f

M
=
M
+

,
so f

is a symplectomorphism of T

M if and only if is closed.

Theorem 3.29 If a symplectomorphism f : T

M T

M preserves each ber of the pro-

jection : T

M M, then f = f

for a closed 1-form on M.

Proof. Fix a point (x
0
, p
0
) T

M and let be a closed 1-form on M such that (x

0
) =
(x
0
, p
0
). Since f is symplectic, the form = f is also closed, and thus the map
h = f
1

f f

is a symplectomorphism of T

M which preserves bers and xes the zero section Z

M
T

M.
Moreover, since the derivative Dh preserves the lagrangian splitting of T(T

M) along Z
M
and equals the identity on TZ
M
, we can conclude from Example 3.1 that Dh is the identity
at all points of Z
M
. Consequently, the ber-derivative of f at the arbitrary point (x
0
, p
0
)
equals the identity, so f is a translation on each ber. Dening (x) = f(x, 0), we have
f = f

.
30
2
If is a closed 1-form on M, then the ow f
t
of the vector eld X

=
1
M
(

) is
symplectic; since V M ker

, the Hamilton-Jacobi theorem implies furthermore that the

ow f
t
satises the hypotheses of Theorem 3.29.
Corollary 3.30 For any closed 1-form on M, the time-1 map f = f
1
of the ow of X

equals f

.
Proof. By Theorem 3.29, the assertion will follow provided that we can show that f

M
=

M
+

. To this end, note that the denition of the Lie derivative shows that f satises
f

M
=
M
+

1
0
d
dt
(f

t

M
) dt =
M
+

1
0
f

t
(L
X

M
) dt.
By Cartans formula for the Lie derivative, we have
L
X

M
= d(X

M
) X

M
=

,
the latter equality following from the fact that X

V M ker
M
and d
M
=
M
.
Another application of Cartans formula, combined with the assumption that is closed
shows that
L
X

= 0,
and so f

for all t. Inserting these computations into the expression for f

M
above, we obtain
f

M
=
M
+

.
2
Using Theorem 3.29, we can furthermore classify all ber-preserving symplectomorphisms
from T

M to T

N.
Corollary 3.31 Any ber-preserving symplectomorphism F : T

M T

N can be realized
as the composition of a ber-translation in T

M with the cotangent lift of a dieomorphism

N M.
Proof. By composing F with a ber-translation in T

M we may assume that F maps the

zero section of T

M to that of T

N. The restriction of F
1
to the zero sections then induces
a dieomorphism f : N M such that the composition F (f
1
)

is a ber-preserving
symplectomorphism of T

N which xes the zero section. From the preceding theorem, we

conclude that F = f

.
2
31
The Schwartz transform
If M, N are smooth manifolds, then the map S
M,N
: T

M T

N T

(M N) dened in
local coordinates by
((x, ), (y, )) (x, y, , )
is a symplectomorphism which we will call the Schwartz transform.
5
An elementary, but
fundamental property of this mapping can be described as follows.
Proposition 3.32 If M, N are smooth manifolds, then the Schwartz transform S
M,N
satis-
es
(S
M,N
)

MN
=
M

N
.
In particular, S
M,N
induces a dieomorphism of zero sections
Z
M
Z
N
Z
MN
and an isomorphism of vertical bundles
V M V N V (M N).
Using the Schwartz transform, we associate to any symplectomorphism F : T

M T

N
the lagrangian embedding
F
: T

M T

(M N) dened as the composition of S

M,N
with
the graph
F
: T

M T

M T

N.
Example 3.33 By Corollary 3.31, a ber-preserving symplectomorphism F : T

M T

N
equals the composition of ber-wise translation by a closed 1-form on M with the cotangent
lift of a dieomorphism g : N M. A computation shows that if M N is the graph
of g and p : M is the natural projection, then the image of the composition of the
lagrangian embedding (T

M,
F
) with the Schwartz transform S
M,N
equals the submanifold

(M N) dened in Example 3.28. In particular, if F is the cotangent lift of g,

then the image of (T

M,
F
) equals the conormal bundle of .

Finally, we note that multiplying the cotangent vectors in T

M by 1 denes a sym-
plectomorphism T

M T

M which can be combined with the Schwartz transform S

M,N
to arrive at the usual symplectomorphism T

M T

N T

(M N). Thus in the special

case of cotangent bundles, dualizing and taking products leads to nothing new.
3.3 Mechanics on manifolds
With the techniques of symplectic geometry at our disposal, we are ready to extend our
discussion of mechanics to more general conguration spaces. Our description begins with a
comparison of the classical and quantum viewpoints, largely parallelling the earlier material
on the 1-dimensional harmonic oscillator given in the introduction. We then turn to the
semi-classical approximation and its geometric counterpart in this new context, setting the
stage for the quantization problem in the next chapter.
5
The name comes from the relation of this construction to the Schwartz kernels of operators (see Sec-
tion 6.2).
32
The classical picture
The hamiltonian description of classical motions in a conguration space M begins with the
classical phase space T

M. A riemannian metric g = (g
ij
) on M induces an inner product
on the bers of the cotangent bundle T

M, and a kinetic energy function which in local

coordinates (q, p) is given by
k
M
(q, p) =
1
2

i,j
g
ij
(q) p
i
p
j
,
where g
ij
is the inverse matrix to g
ij
. Regular level sets of k
M
are sphere bundles over M.
The hamiltonian ow associated to k
M
is called the co-geodesic ow due to its relation
with the riemannian structure of M described in the following theorem.
Theorem 3.34 If M is a riemannian manifold, then integral curves of the co-geodesic ow
project via to geodesics on M.
A proof of this theorem can be given in local coordinates by using Hamiltons equations
q
i
=
H
p
i
=

j
g
ij
p
j
p
i
=
H
q
i
=
1
2

u,v
g
uv
q
i
p
u
p
v
to derive the geodesic equation
q
k
+

i,j

k
ij
q
i
q
j
= 0.
For details, see [36].
In physical terms, Theorem 3.34 states that a free particle on a manifold must move along
a geodesic. A smooth, real-valued potential V : M R induces the hamiltonian function
H(q, p) = k
M
(q, p) + V (q)
on T

M. Integral curves of the hamiltonian ow of H then project to classical trajectories

of a particle on M subject to the potential V .
The quantum mechanical picture
For the time being, we will assume that the Schrodinger operator on a riemannian manifold
M with potential function V is dened in analogy with the at case of R
n
with its standard
metric. That is, we rst dene the operator on the function space C

(M) by

H =

2
2m
+ m
V
,
where denotes the Laplace-Beltrami operator. As before,

H induces a (densely dened)
operator

H on the intrinsic Hilbert space H
M
of M by the equation

H(a[dx[
1/2
) = (

Ha)[dx[
1/2
,
33
where [dx[ is the natural density associated to the metric on M, and the time-independent
Schrodinger equation on M assumes the familiar form
(

H E) = 0.
The advantage of this viewpoint is that both the classical state space T

M and the quantum

state space H
M
are objects intrinsically associated to the underlying dierential manifold
M. The dynamics on both objects are determined by the choice of metric on M.
The semi-classical approximation
The basic WKB technique for constructing semi-classical solutions to the Schrodinger equa-
tion on M proceeds as in Section 2.2. Specically, a half-density of the form e
iS/h
a is a
second-order approximate solution of the eigenvector problem (

H E) = 0 provided that
the phase function S: M R satises the Hamilton-Jacobi equation
H
dS
= E,
and the half-density a satises the homogeneous transport equation, which assumes the
coordinate-free form
aS + 2L
S
a = 0.
We can formulate this construction abstractly by considering rst a projectable, exact
lagrangian embedding : L T

M. By denition, this means that

L
= is a dieo-
morphism, where : T

M M is the natural projection, and Lemma 3.25 implies that for

any primitive : L R of

M
, the composition S =
1
L
is a phase function for (L, ).
Now if H : T

M R is any smooth function, then (L, ) satises the Hamilton-Jacobi

equation provided that E is a regular value of H and
H = E.
In this case, the embedding and hamiltonian vector eld X
H
of H induce a nonsingular
vector eld X
H,
on L (see Example 3.13). If a is a half-density on L, then the requirement
that (
1
L
)

a satisfy the homogeneous transport equation on M becomes

L
X
H,
a = 0.
If these conditions are satised, then the half-density e
i/
a on L can be quantized (i.e.
pulled-back) to yield a second-order approximate solution (
1
L
)

e
i/
a to the Schrodinger
equation on M, as above.
This interpretation of the WKB approximation leads us to consider a semi-classical state
as a quadruple (L, , , a) comprised of a projectable, exact lagrangian embedding : L
T

M, a generalized phase function : L R satisfying d =

M
, and a half-density a
on L. Our correspondence table now assumes the form
34
Object Classical version Quantum version
basic space T

M H
M
state (L, , , a) as above on M
time-evolution Hamiltons equations Schrodinger equation
generator of evolution function H on T

M operator

H on H
M
stationary state state (L, , , a) such that eigenvector of

H
H = E and L
X
H,
a = 0
Since the Hamilton-Jacobi and transport equations above make sense for any triple (L, , a)
consisting of an arbitrary lagrangian immersion : L T

M and a half-density a on L, it is
tempting to regard (L, , a) as a further generalization of the concept of semi-classical state in
T

M, in which we drop the conditions of projectability and exactness. Our goal in the next
chapter will be to determine when and how such geometric semi-classical states can be
used to construct analytical semi-classical approximate solutions to the time-independent
Schrodinger equation.
35
4 Quantization in Cotangent Bundles
This chapter deals with the problem of constructing semi-classical approximate solutions to
the time-independent Schrodinger equation from the data contained in a geometric semi-
classical state. The starting point will be a lagrangian immersion : L T

M whose image
in contained in a regular level set of the classical hamiltonian H associated to some metric
and potential on the conguration space M. Given a half-density a on L which is invariant
under the ow induced by H, our goal is to use this data in order to construct, in a more or
less systematic way, an approximate solution to Schrodingers equation on M.
This process will be referred to as quantization. (It is, of course, only one of many
operations which go by this name.) As we shall see, the question of whether or not a given
semi-classical state (L, , a) can be quantized is answered largely in terms of the geometry of
the lagrangian immersion (L, ). For this reason, we will often speak loosely of quantizing
lagrangian submanifolds or quantizable lagrangian submanifolds.
4.1 Prequantization
The simplest quantization procedure consists of pulling-back half-densities from projectable
lagrangian embeddings in T

M. As seen in Chapter 3, a triple (L, , a) for which : L T

M
is a projectable exact lagrangian embedding is quantized by choosing a primitive of

M
and forming the half-density
I

(L, , a)
def
= (
1
L
)

e
i/
a
on M. The choice of is unique only up to an additive constant, which leads to an ambiguity
in the overall phase of I

(L, , a). (This ambiguity was overcome in Chapter 3 by including

the choice of in the denition of a semi-classical state).
To generalize this procedure, suppose now that : L T

M is a projectable, but not

necessarily exact lagrangian embedding. Since the 1-form

M
on L is closed, it is locally
exact by the Poincare lemma, and so we can choose a good cover L
j
of L (see Appendix C)
and functions
j
: L
j
R such that d
j
=

M
[
L
j
. Given a half-density a on L, we set
a
j
= a[
L
j
and dene a half-density on
L
(L
j
) by quantizing (L
j
, [
L
j
,
j
, a
j
) in the sense
above:
I
j
= (
1
L
j
)

e
i
j
/
a
j
.
To quantize (L, , a), we must piece together the I
j
to form a well-dened global half-density
I

(L, , a) on M. This is possible for arbitrary a provided that the functions

j
can be chosen
so that the oscillatory coecients e
i
j
/
agree where their domains overlap; that is, we must
have

k
Z

def
= 2 Z
on each L
j
L
k
. According to the discussion in Appendix C, this is precisely the condition
that the Liouville class
L,
be -integral. At this point, however, we are forced to confront
once more the nature of . If it is a formal variable, the notion of -integrality is meaningless.
If denotes a number ranging over an interval of the real numbers, then the Liouville class
will be 2 times an integral class for all only if it is zero, i.e. only for an exact lagrangian
36
submanifold. Since we are specically trying to go beyond the exact case, this interpretation
is not acceptable either. Instead, we will make the following compromise.
Denition 4.1 A projectable lagrangian submanifold (L, ) T

M is quantizable if its
Liouville class
L,
is -integral for some R
+
. The values of for which this condition
holds will be called admissible for (L, ).
If L is quantizable but not exact, the set of all admissible forms a sequence converging to
zero, consisting of the numbers
0
/k, where
0
is the largest such number, and k runs over
the positive integers. If L is exact, all > 0 are admissible.
Example 4.2 Let M = S
1
and consider the closed 1-form = p d on S
1
, for p R. The
cohomology class represented by this form is -integral provided that
p

S
1
d Z

,
and so the admissible values of are the numbers p/k : k Z
+
. Geometrically this means
that all horizontal circles in the cylinder T

S
1
are quantizable in the sense dened above,
but with diering sets of admissible .
The situation changes as soon as we consider the closed 1-form = a d
1
+ b d
2
on the
torus S
1
S
1
. If is -integral for some R
+
, then a/ and b/ are both integers,
meaning that a/b is a rational number if b = 0. In general, the condition that closed 1-form
on a manifold M be quantizable is equivalent to the requirement that the ratio of any
two nonzero periods of be rational. (A period of a closed 1-form on M is any number
obtained by integrating around some closed loop in M). Thus, the class of projectable
quantizable lagrangian submanifolds of T

M is rather limited whenever dim(H

1
(M; R)) > 1.

The simple quantization technique described above does not generalize immediately to
non-projectable immersed lagrangian submanifolds (L, ) T

M, since
L
cannot be used to
push-forward half-densities from L to M. For the time being, however, we will focus on the
set of regular points of
L
in order to pass from half-densities on L to half-densities dened
near non-caustic points of L. Regardless of how this is to be carried out, it is desirable that
quantization be linear with respect to half-densities:
I

(L, , sa
1
+ a
2
) = s I

(L, , a
1
) + I

(L, , a
2
).
Certainly this condition holds for the procedure we have been using in the projectable case,
and it is reasonable to adopt as a general rule. One consequence is
I

(L, , 0) = 0.
If a semi-classical state is represented by the union of a disjoint pair (L
1
,
1
, a
1
), (L
2
,
2
, a
2
)
of lagrangian submanifolds carrying half-densities, then by linearity we should have
I

(L
1
L
2
,
1

2
, (a
1
, a
2
)) = I

(L
1
L
2
,
1

2
, (a
1
, 0)) + I

(L
1
L
2
,
1

2
, (0, a
2
))
= I

(L
1
,
1
, a
1
) + I

(L
2
,
2
, a
2
).
37
Now consider an arbitrary immersed lagrangian submanifold (L, ) T

M and half-
density a on L. If p
L
(L) is non-caustic and
L
is proper, then there is a contractible
neighborhood U M of p for which
1
L
(U) consists of nitely many disjoint open subsets
L
j
L such that each (L
j
, [
L
j
) is a projectable lagrangian submanifold of T

U. Choosing a
generalized phase function
j
: L
j
R for each L
j
, we note that by the preceding remarks,
the quantization of (L, , a) should look something like

j
(
1
L
j
)

e
i
j
/
a
on U. As before, the requirement that
j
be a generalized phase function for L
j
determines
each
j
only up to an additive constant, and so the meaning of the preceding sum is ambigu-
ous. To quantize half-densities on L consistently, we must therefore decide how to specify
the relative phases of the oscillatory coecients e
i
j
/
.
If L is exact, then we may use any function : L R satisfying d =

M
to x phases,
as in the following example.
Example 4.3 For L = R, the lagrangian embedding : L T

R given by (x) = (x
2
, x)
has a singular point at x = 0, and we denote by L
+
, L

the right and left projectable

components of L, respectively (these correspond to the upper and lower components of the
parabola (L)). A phase function : L R for (L, ) is given by
(x) = 2x
3
/3.
If a = B(x) [dx[
1/2
is any half-density on L, then since
(
1
L
)

q
= 2
1
q
1/2

x
,
the transformation rule for half-densities implies
(
1
L
+
)

a = 2
1/2
q
1/4
B(q
1/2
) [dq[
1/2
(
1
L

a = 2
1/2
q
1/4
B(q
1/2
) [dq[
1/2
.
Thus, the prequantization of (L, , a) is given for q > 0 by
I

(L, , a)(q) =

e
2iq
3/2
/3
B(q
1/2
) + e
2iq
3/2
/3
B(q
1/2
)

2
1/2
q
1/4
[dq[
1/2
.
The parabola (L) lies in the regular level set H
1
(0) of the hamiltonian for a constant
force eld
H(q, p) =
1
2
(p
2
q),
and it is easy to check that the induced vector eld X
H,
on L equals X
H,
= (1/2) /x.
Thus, a half-density a on L is invariant under the ow of X
H,
if and only if a = B[dx[
1/2
for some B R. From the expression above, we obtain
I

(L, , a) =

e
2iq
3/2
/3
+ e
2iq
3/2
/3

2
1/2
q
1/4
B[dq[
1/2
38
as a semi-classical approximate solution to the Schrodinger equation

2
2

x
2

x
2
= E.
Unfortunately, this solution blows up at q = 0 (and is not dened for q < 0), so we have more
work ahead of us. In particular, we have no way as yet to check that letting be continuous
at 0 as a function on L is the right way to assure that we have a good approximation in the
immediate vicinity of q = 0. In fact, we will see later that this is the wrong choice!

Since exactness is only used to insure that the function e

i/
is well-dened on L, we can treat
certain non-exact cases in a similar way. For this purpose, we make the following provisional
denition, generalizing Denition 4.1.
Denition 4.4 An immersed lagrangian submanfold (L, ) T

M is said to be prequan-
tizable if its Liouville class
L,
is -integral for some R
+
. The values of for which
this condition holds will again be called admissible for (L, ).
If is admissible for some prequantizable lagrangian immersion (L, ), then there exists a
good cover V
j
of the manifold L and functions
j
: V
j
R such that d
j
=

M
[
V
j
and

k
Z

on each V
j
V
k
. Consequently, the
j
describe a single function : L T

=
R/Z

which satises d =

M
and which denes a global oscillatory function e
i/
on L. If
a is a half-density on L, we can now quantize (L, , , a) by summing the pull-backs of e
i/
a
to M. In the previous notation, the value of I

(L, , , a) on U is dened as

j
(
1
V
j
)

e
i/
a.
Example 4.5 Consider the 1-dimensional harmonic oscillator with hamiltonian
H(q, p) =
1
2
(q
2
+ p
2
).
According to Denition 4.4, a number R
+
is admissible for the level set H
1
(E) provided
that
H
1
(E)

1
= 2E Z

.
Thus, energy levels of the 1-dimensional harmonic oscillator corresponding to level sets for
which a particular value of is admissible are given by E = n. As one can read in any
textbook on quantum mechanics, the actual quantum energy levels are E = (n + 1/2).
The additional 1/2 can be explained geometrically in terms of the non-projectability of the
classical energy level curves, as we shall soon see.

39
Prequantum bundles and contact manifolds
Prequantizability can be described geometrically in terms of principal T

bundles with con-

nection over T

M. It is customary to make the following denition.

Denition 4.6 For xed R
+
, the prequantum T

bundle associated to a cotangent

bundle (T

M,
M
) consists of the trivial principal bundle Q
M,
= T

M T

together with
the connection 1-form =

M
+ d.
Here, denotes the multiple-valued linear variable in T

and : Q
M,
T

M is the bundle
projection. If : L T

M is any lagrangian immersion, then the curvature of the induced

connection on

Q
M,
coincides with

M
and therefore vanishes. The holonomy of this
connection is represented by the mod-Z

reduction of the Liouville class

L,
.
From the prequantum standpoint, the basic geometric object representing a classical
state is therefore a quadruple (L, , a, ) consisting of a lagrangian immersion : L T

M,
a half-density a on L, and a parallel lift of L to the T

bundle

Q
M,
. If we associate
to Q
M,
the prequantum line bundle c
M,
by means of the representation x e
ix/
of T

in U(1), then induces a parallel section of

c
M,
corresponding to the (inverse of the)
oscillatory function e
i/
on L which appeared in the preceding section. This remark proves
the following geometric characterization of prequantizability.
Theorem 4.7 An immersed lagrangian submanifold (L, ) T

M is prequantizable if and
only if there exists a nonzero parallel section over L of the line bundle

c
M,
for some > 0.
Prequantum T

bundles and parallel lifts of lagrangian submanfolds constitute our rst

examples of the fundamental objects of contact geometry, the odd-dimensional counter-
part of symplectic geometry. Digressing briey from quantization, we assemble here a few
facts about contact manifolds.
Denition 4.8 A contact form on a (2n+1)-dimensional manifold Q is a 1-form such
that (d)
n
vanishes nowhere on Q. A manifold endowed with a contact form is called a
strict contact manifold.
To interpret the condition on we note rst that the kernel of any nowhere vanishing 1-form
denes a 2n-dimensional distribution in Q. If , are (local) vector elds lying in this
distribution, we have
d(, ) = () () ([, ]) = ([, ]).
This says that the distribution is integrable (in the sense of Frobenius) i d is zero on
ker(). The condition (d)
n
= 0, on the other hand, means that the kernel of d is
1-dimensional and everywhere transverse to ker(). Consequently, d is a linear symplectic
form on ker(), and the largest integral submanifolds of ker() are n-dimensional (i.e.
ker() is maximally non-integrable).
Denition 4.9 A legendrian submanifold of a 2n+1-dimensional strict contact manifold
(Q, ) is an n-dimensional integral submanifold for .
40
If is a contact form and f a nowhere-vanishing function on Q, then f is again a contact
form which is said to be equivalent to , since they have the same legendrian submanifolds.
This leads to the following denition.
Denition 4.10 A contact structure on a manifold M is a codimension one subbundle
E TM which is locally dened by contact forms, and a manifold endowed with such a
structure is called simply a contact manifold.
When the quotient TM/E is trivial, E is the kernel of a globally dened contact form. Such
a contact structure is called coorientable.
Our basic example of a strict contact manifold is furnished by a prequantum T

bundle
Q
M,
over a cotangent bundle T

M; the image of a parallel lift of a lagrangian immersion

: L T

M to Q
M,
is an immersed legendrian submanifold of Q
M,
. Although we will
not pursue the idea in these notes, a possible generalization of the quantization procedure
described above might therefore begin by re-interpreting geometric semi-classical states as
triples (R, , a) consisting of a legendrian immersion (R, ) in a contact manifold together
with a half-density a on R.
4.2 The Maslov correction
It turns out that the naive quantization procedure of the preceding section is incorrect,
since it ignores a certain structure which arises from the relation of L to the bers of the
projection T

M

M. This factor will be incorporated into our quantization procedure using
a procedure due to Maslov. To begin, we will illustrate this idea in the case of lagrangian
submanifolds of the phase plane.
Given a lagrangian immersion : L T

R R
2
, we denote by
p
the composition of
with the projection of R
2
onto the p-axis. If
p
is a dieomorphism, then (L, ) is said to be
p-projectable, in which case there exists an alternate generalized phase function : L R
satisfying d =

(q dp), obtained by thinking of R

2
as the cotangent bundle of p-space.
A simple example of an embedded lagrangian submanifold of T

R which does not project

dieomorphically onto the q-axis is a vertical line, or ber, of the form (x) = (q
0
, x) for x
R. Since the wave function corresponding to a constant half-density a on L should correspond
to a probability distribution describing the position of a particle at q
0
with completely
indeterminate momentum, it should be a delta function supported at q
0
. Following an idea
of Maslov, we analyze this situation by pretending that p is position and q momentum, and
then quantizing to obtain a function on p-space. Using the phase function (x) = q
0
x on
L, we obtain
(
1
p
)

e
i/
[dx[
1/2
= e
iq
0
p/
q
0
[dp[
1/2
.
The result is exactly the asymptotic Fourier transform (see Appendix B) of the delta function
which we guessed above!
In its simplest form, Maslovs technique is to suppose that (L, ) T

R is p-projectable,
so that d =

(q dp) for some phase function on L. If a is a half-density on L, we dene

a function B on p-space by the equation
B[dp[
1/2
= (
1
p
)

e
i/
a.
41
The Maslov quantization of (L, , , a) is then given by the half-density
J

(L, , , a)
def
= T
1

(B) [dq[
1/2
on q-space, where T

denotes the asymptotic Fourier transform. To relate this procedure to

our earlier quantization by pull-back, we must compare the results in the case of lagrangian
submanifolds L R
2
which are bi-projectable, i.e. projectable in both the q- and p-directions.
For simplicity, we begin with the example of linear lagrangian subspaces.
Example 4.11 For real k = 0, consider the lagrangian embedding : R T

R given by
(x) = (x, kx). Generalized phase functions on (L, ) for the forms p dq and q dp are given
by (x) = kx
2
/2 and (x) = kx
2
/2, respectively. If a is a constant half-density on L, then
the transformation rule for half-densities implies
(
1
L
)

a = A[dq[
1/2
(
1
p
)

a = [k[
1/2
A[dp[
1/2
for a real constant A determined by a. Quantization by pull-back therefore gives
I

(L, , , a) = e
ikq
2
/2
A[dq[
1/2
.
On the other hand, we have (
p
)

(p) = p
2
/2k, and a computation
6
shows that
T
1

((
1
p
)

e
i/
)(q) = [k[
1/2
e
isgn(k)/4
e
ikq
2
/2
,
and so Maslovs technique yields
J

(L, , , a) = e
isgn(k)/4
I

(L, , , a).
Thus, the half-density obtained from (L, , a) by Maslov quantization diers from the simple
pull-back by a constant phase shift.

To establish a similar correspondence in somewhat greater generality, consider an arbi-

trary bi-projectable lagrangian embedding (L, ) R
2
with phase functions and corre-
sponding to p dq and q dp, respectively. For simplicity, we will assume that the additive
constants in and are chosen so that
= +

(qp).
Next, let S(q) and T(p) be the functions dened on q- and p-space by pull-back:
S =
1
L
T =
1
p
,
so that S and T satisfy the Legendre transform relation (see [2])
S(q) = p(q) T
t
(p(q)) + T(p(q)),
6
See, for example, [32, Vol.1, Thm.7.6.1] .
42
where p(q) = S
t
(q). From this relation, it follows easily that
T
tt
(p(q)) = (S
tt
(q))
1
.
A half-density a on L determines functions A(q) and B(p) such that
(
1
L
)

a = A[dq[
1/2
(
1
p
)

a = B[dp[
1/2
and
A(q) = [S
tt
(q)[
1/2
B(p(q)).
For each q, we must now compare the Maslov half-density
J

(L, , , a) = (2)
1/2

R
e
i(pq+T(p))/
B(p) dp [dq[
1/2
with that obtained by pull-back:
I

(L, , , a) = e
iS/
A[dq[
1/2
.
To this end, we set k(q) = T
tt
(p(q)) and apply the principle of stationary phase (see Ap-
pendix B). The critical point of the exponent pq +T(p) occurs where q = T
t
(p), i.e. where
p = S
t
(q) = p(q). Hence,

R
e
i(pq+T(p))/
B(p) dp = (2)
1/2
e
isgn(k)/4
e
iS(q)/
[k(q)[
1/2
B(p(q)) + O(
3/2
).
Thus
J

(L, , , a) = e
isgn(k)/4
I

(L, , , a) +O().
For bi-projectable (L, ), we therefore conclude that Maslovs technique coincides with quan-
tization by pull-back up to a constant phase factor and terms of order .
The essential dierence between the naive prequantization of Section 4.1 and the Maslov
quantization of a p-projectable lagrangian embedding (L, ) which is not q-projectable lies
in the relative phase constants of the summands of I

(L, , a), as illustrated by the following

example.
Example 4.12 A phase function associated to q dp for the lagrangian embedding (x) =
(x
2
, x) of L = R into R
2
is given by (x) = x
3
/3. The Maslov quantization of a half-density
a = B(x) [dx[
1/2
on L is thus
J

(L, , , a) = T
1

e
ip
3
/3
B(p)

[dq[
1/2
,
since (
1
p
)

a = B(p) [dp[
1/2
. For each q > 0, critical points of the function R(p) = pq p
3
/3
occur precisely when q = p
2
, i.e. when (q, p) (L), and an application of the principle of
stationary phase therefore yields two terms corresponding to the upper and lower halves of
L. Specically, we have
J

(L, , , a) =

e
i/4
e
2iq
3/2
/3
B(q
1/2
) + e
i/4
e
2iq
3/2
/3
B(q
1/2
)

2
1/2
q
1/4
[dq[
1/2
+ O().
43
Compare this with the result of Example 4.3. The extra phase factors of e
i/4
make
J

(L, , , a) essentially dierent from the prequantization of (L, , a). However, while the
term of order 0 of J

(L, , , a) is, like I

(L, , , a), singular at the caustic point q = 0, the

full expression for J

(L, , , a) as an integral is perfectly smooth there, at least if a has

compact support. This smoothness at caustics is a clear advantage of Maslov quantization.

The relative phase factor in the preceding example can be attributed to the fact that the
function T(p) = p
3
/3 has an inection point at p = 0. More precisely, since T
t
is convex,
a factor of e
i/2
arises in passing from the upper to the lower half of the parabola; if T
t
were
concave, the situation would be reversed.
The preceding observation leads us to assign an index to closed, immersed curves in the
phase plane. A p-dependent phase function T for L R
2
will have inection points at
precisely those p for which (T
t
(p), p) is a singular point of L. Moreover, the sign of T
tt
at
nearby points depends only on L and not on the choice of T. With these remarks in mind,
suppose that (L, ) is a closed, immersed curve in R
2
which is non-degenerate in the sense
that if T is a p-dependent phase function for a subset of L, then T
t
has only non-degenerate
critical points. Under this assumption, sgn(T
tt
) changes by 2 in the vicinity of a critical
point of T
t
, and we can assign an index to (L, ) by summing these changes while traversing
L in a prescribed direction. The result is twice an integer known as the Maslov index m
L,
of (L, ). (Compare [3]).
Example 4.13 The computation of the Maslov index can be interpreted geometrically if
we rst observe that the non-degeneracy condition requires L to remain on the same side of
a ber
1
(q) near a singular point. Since the index only involves the sign of T
tt
, it follows
from Example 4.12 that the integer 1 should be assigned to a critical point of
L
which is
traversed in the p direction to the right of the ber. In other words, downward motion to
the right of the ber is positive, while the sign changes if either the direction of motion or
the side of the ber is reversed, but not if both are.
A circle in the phase plane has a right and a left singular point, both of which are positive
according our rule when the circle is traversed counterclockwise. The Maslov index of the
circle therefore equals 2. In fact, the same is true of any closed embedded curve traversed
counterclockwise. On the other hand, a gure-eight has Maslov index zero.
The reader is invited to check that if L is a circle with a xed orientation H
1
(L; Z),
then m
L,
equals '
L,
, `, where
L,
is the Maslov class of (L, ) dened in Example 3.22.

Our goal is now to modify the prequantization procedure for arbitrary lagrangian sub-
manifolds of the phase plane by incorporating the Maslov index. The basic idea is as follows.
Given an immersed lagrangian submanifold (L, ) R
2
, we rst choose a good cover L
j

of L such that the image of each L

j
under is either q- or p-projectable and no intersection
L
j
L
k
contains a critical point of
L
. Next, we x a partition of unity h
j
subordinate
44
to L
j
. To quantize a half-density a on L, we quantize each (L
j
, , a h
j
) to obtain a half-
density I
j
on R either by pull-back or by Maslovs technique. As before, we would then like
to dene the quantization of (L, , a) as the sum
I

(L, , a) =

j
I
j
.
In order to specify the relative phases of the I
j
, and to make this denition independent of
the choice of cover L
j
and partition of unity, we will require that I

(L, , a) coincide up
to order with the usual quantization by pull-back for any half-density a supported in a
projectable subset of L. This condition can be precisely formulated in terms of the Maslov
index and Liouville class of (L, ) as follows.
On an open interval U of non-caustic points, each half-density I
j
is the sum of the pull-
back of half-densities on each component of L
j

1
L
(U). If L
j
is quantized by pull-back, this
statement is obvious; if Maslovs technique is applied to L
j
, it follows from an application of
the principle of stationary phase as in Example 4.12. On L
j

1
L
(U), these half-densities
are of the form

I
j
= e
is
j
/4
e
i
j
/
a.
Here, is a real-valued function on L
j
satisfying d
j
=

1
, while s
j
are integers depending
only on the component of L
j

1
L
(U) in question. (More precisely, s
j
is zero if L
j
is
q-projectable and is quantized by pull-back; otherwise s
j
equals sgn(T
tt
) for a suitable p-
dependent phase function). For I

(L, , a) to be well-dened, we must choose the functions

j
so that

I
j
=

I
k
on each intersection L
j
L
k
regardless of the particular half-density a. In
other words, we require
e
i(
j

k
)/
e
i(s
j
s
k
)/4
= 1
at each point of L
j
L
k
. Since
j

k
is constant on L
j
L
k
, we can dene a
jk
as the
(constant) value of (
j

k
) (s
j
s
k
)/4 on L
j
L
k
, so that our requirement becomes
a
jk
Z

.
Evidently, this condition can be fullled on any arc of a curve in the phase plane. If L is
a circle, then this condition implies that the sum of any of the a
jk
lies in Z

, or, in other
words, that the Maslov index m
L,
of (L, ) satises

2
m
L,
+

1
Z

.
This is the simplest version of the Maslov quantization condition.
Example 4.14 Returning to the harmonic oscillator of Example 4.5, we see that the level
set H
1
(E) satises the Maslov condition provided that for some integer n,
E = (n + 1/2).
Allowable energy levels in this case therefore correspond to the Bohr-Sommerfeld condition,
which actually gives the precise energy levels for the quantum harmonic oscillator.

45
A general quantization scheme
Motivated by the simple results above, our aim in the next sections will be to develop a
systematic method for quantizing lagrangian submanifolds of cotangent bundles. The basis
of this method will again be Maslovs technique, which relies in this context on the concept
of generalized phase functions.
Local parametrizations of an immersed lagrangian submanifold (L, ) T

M dened by
such phase functions will rst enable us to quantize a given half-density locally on L by
means of a slightly more general version of the (inverse) asymptotic Fourier transform. The
result will be a collection of half-densities on M. In order for these half-densities to piece
together appropriately, it is necessary and sucient that L satisfy a general version of the
Maslov quantization condition, which we formulate in the next section using the Maslov
class.
4.3 Phase functions and lagrangian submanifolds
In this section, we generalize the concept of phase functions to non-projectable lagrangian
submanifolds of cotangent bundles. Roughly speaking, the idea is the following. As we
saw in Chapter 3, a projectable lagrangian submanifold of T

N can be locally parametrized

by the dierential of a function f on U N, viewed as a mapping df : U T

N. For
each p U, the meaning of df
p
as an element of T

p
N is that for any smooth curve in N
satisfying (0) = p, we have 'df
p
, (0)`
def
= (f)
t
(0). To parametrize more general lagrangian
submanifolds of T

N in a similar way, we can begin with a function : UR

m
R together
with a point p = (p, v) UR
m
and attempt to dene d
p
T

p
N by 'd
p
, (0)` = ( )
t
(0)
for any lift of to the product U R
m
such that (0) = p. In general, this fails, since
the value of the directional derivative ( )
t
(0) depends on the lift . If, however, the
ber-derivative / vanishes at p, the expression for d
p
produces a well-dened element
of T

p
N. As we shall see, the assumption that the map / : U R
m
R
m
is transverse
to 0 implies that the ber critical set

p U R
m
:

= 0

is a smooth submanifold of U R
m
, and the assignment p d
p
denes a lagrangian
immersion of

into T

N. In general, the restriction of the projection U R

m
U to

is non-injective, and thus the image of

is a non-projectable lagrangian submanifold. From

the point of view of the WKB method, this generalization amounts to replacing Maslovs
ansatz
(2)
n/2

R
n
e
i(/p,q)+T(p))/
a(p) dp [dq[
1/2
for the solution of Schrodingers equation by the more general form
(2)
m/2

R
m
e
i(q,)/
a(q, )[d[[dq[
1/2
,
46
where is an auxiliary variable in R
m
which may have nothing to do with the variable p dual
to q. An advantage of this generalization will be to allow a calculus which is more clearly
invariant under changes of coordinates, unlike the previous Fourier transform picture, which
requires linear structures on p- and q-space.
To begin, we x some notation and terminology.
7
Let M, B be smooth manifolds, and
let p
M
: B M be a smooth submersion. Dualizing the inclusion E = ker(p
M
)

TB gives
rise to an exact sequence of vector bundles over B
0 E

B E

0
where E

B denotes the annihilator of E. The ber-derivative of a function : B

R is the composition d

d, and its ber critical set is dened as

= (d

)
1
Z
E
.
(We will denote the zero section of a vector bundle F by Z
F
). The function is said to
be nondegenerate if its ber derivative is transverse to Z
E
, in which case

is a smooth
submanifold of B. At points of

, the section d

has a well-dened intrinsic derivative (see

[26]), denoted d

, which induces for nondegenerate an exact sequence of vector bundles

over

0 T

B
d

0.
The ber-hessian H of at p

is dened as the composition d

: E
p
E

p
.
The bundle E

may be identied with the pull-back p

M
T

M, giving rise to a natural

projection E

p
T

M. On the ber-critical set

, the dierential d denes a section of

E

whose composition with p we denote by

M.
Theorem 4.15 If is nondegenerate, then the map

M is an exact lagrangian
immersion.
Proof. Since E

is the union of the conormal bundles of the bers of p

M
, it follows from
Example 3.19 that E

is a coisotropic submanifold of T

B. Moreover, the characteristic

distribution (

of E

is tangent to the bers of the mapping E

M.
The nondegeneracy assumption on is equivalent to the requirement that the lagrangian
submanifold (d)(B) of T

B be transverse to E

. By Lemma 3.6(2), this implies that the

section (d)(B)[

= (d)(B) E

is nowhere tangent to the distribution (

and therefore
immerses into T

M. To complete the proof, we note that the equality

M
= d[

implies that L

is exact lagrangian.
2
7
Some readers may nd it instructive to follow the ensuing discussion by writing everything in local
coordinates.
47
From the preceding denitions, it follows that the nullity k of the ber-hessian at p

equals dim(T
p

E
p
), which in turn equals the dimension of the kernel of (

on T
p

.
This means in particular that the dimension of the bers of B must be at least k; if the ber
dimension of B equals k, then the phase function is said to be reduced at p. This occurs
when its ber-hessian vanishes at p.
Example 4.16 Suppose that M = R and consider the function (x, ) = x +
3
/3 on
B = RR. Here, d

=
2
+x, and (d

) = 2d +dx. Evidently the phase function is

nondegenerate, and its ber critical set consists of the parabola x =
2
, which fails to be
x-projectable precisely when x = 0. This is just the value of x for which (x, ) acquires a
degenerate critical point.

A phase function which generates a neighborhood of a point in a lagrangian submanifold

is not unique per se, but it is unique up to stable equivalence, a concept which we now
dene. To begin, we introduce the following terminology.
Denition 4.17 A triple (B, p
B
, ) is called a Morse family over a manifold M if p
B
:
B M is a smooth (possibly non-surjective) submersion, and is a nondegenerate phase
function on B such that

is an embedding. A Morse family is said to be reduced at p B

if is a reduced phase function at p.
We will say that the lagrangian submanifold im(

) = L

is generated by the Morse family

(B, p
B
, ). If : L T

M is a lagrangian immersion and p L, then we denote by

M(L, , p) the class of Morse families (B, p
B
, ) which generate (U) for some neighborhood
U L of p. For such Morse families, we denote by g

: U

the dieomorphism dened

by g

=
1

.
If (B, p
B
, ) M(L, , p), then the following operations produce further elements of
M(L, , p).
1. Addition: For any c R, (B, p
B
, + c) M(L, , p).
2. Composition: If p
B
: B
t
M is a second submersion and g : B
t
B is a ber-
preserving dieomorphism, then (B
t
, p
B
, g) M(L, , p).
3. Suspension: The suspension of (B, p
B
, ) by a nondegenerate quadratic form Q on
R
n
is dened as the Morse family comprised of the submersion p
B
: B R
n
M
given by composing p
B
with the projection along R
n
, together with the phase function

= + Q. Evidently the ber-critical set of

equals the product

0, and

(b, 0) =

(b) for all (b, 0)

. Thus (B R
n
, p
B
,

) M(L, , p).
4. Restriction: If B
t
is any open subset of B containing
1

(p), then the restrictions of

p
B
and to B
t
dene a Morse family on M which belongs to M(L, , p).
These operations generate an equivalence relation among Morse families called stable equiv-
alence. The central result of this section is the following.
48
Theorem 4.18 Let : L T

M be a lagrangian immersion, and let p L. Then:

1. The class M(L, , p) contains a reduced Morse family over M.
2. Any two members of M(L, , p) are stably equivalent.
The next two subsections are devoted to the proof of this theorem.
Symplectic normal forms
The purpose of this section is to develop several results needed to prove Theorem 4.18, all of
which more or less rely on the so-called deformation method. This method was introduced
by Moser in [45] (it probably has a longer history) and applied to a variety of problems of
symplectic geometry in [62]. We will prove the following theorem.
Theorem 4.19 Let (P, ) be a symplectic manifold, : L P a lagrangian immersion, and
E a lagrangian subbundle of

TP which is complementary to the image of

: TL

TP.
Then there is a symplectic immersion of a neighborhood U of the zero section Z T

L
into P such that = on Z and which maps the vertical subbundle V
Z
L T
Z
(T

L) onto
E.
As a result of the existence of compatible complex structures on

P, lagrangian subbundles
of

P complementary to the image of

: TL

TP always exist. Thus, Theorem 4.19

implies the following nonlinear normal form result, which states that, near any of its
lagrangian submanifolds L, a symplectic manifold looks like a neighborhood of the zero
section in T

L.
Corollary 4.20 If L is a lagrangian submanifold of P, then a neighborhood of L in P is
symplectomorphic to a neighborhood of the zero section in T

L, by a map which is the identity

on L.
Theorem 4.19 may be combined with Proposition 3.24 to yield:
Corollary 4.21 The map gives a 1-1 correspondence between a neighborhood of L in the
space of all lagrangian submanifolds of (P, ) and a neighborhood of the zero section in the
space Z
1
(L) of closed 1-forms on L.
The proof of Theorem 4.19 will rely on the following
Relative Darboux theorem [62] . Let N be a submanifold of a manifold P, and let

0
,
1
be two symplectic forms on P which coincide on T
N
P. Then there are neighborhoods
U and V of N and a dieomorphism f : U V such that
1. f

1
=
0
2. f[
N
= id and Tf[
T
N
P
= id.
49
Proof. Dene a 1-parameter family of closed forms on P by

t
=
0
+ t(
1

0
).
Since all
t
agree on the submanifold N, there is a neighborhood of N in P (which for our
purposes we may assume to be P itself) on which all
t
are nondegenerate.
To nd f satisfying f

1
=
0
, we will construct a time-dependent vector eld X
t
for
which the isotopy f
t
that it generates satises f

t

t
=
0
for all t [0, 1]. By the usual
properties of the Lie derivative, it is necessary and sucient that such a vector eld solve
the equation
0 =
d
dt
(f

t

t
) = f

d
t
dt

+ f

t
(L
Xt

t
) = f

t
(
1

0
+ d(X
t

t
)).
In order to x N, we also want X
t
[
N
= 0. These conditions will be satised if we set
X
t
=
1
t
() for a 1-form on a neighborhood of N in P, vanishing on N, such that
d =
1

0
. If the submanifold N consisted of a single point, the form would be easy
to nd, since the closed 2-form
1

0
is locally exact by the classical Poincare lemma. For
more general submanifolds, we use a more general version of this lemma:
Relative Poincare lemma . Let N be a submanifold of P, and let be a closed k-form on
P which vanishes on TN. Then there is a form on a neighborhood of N such that d =
and which vanishes on T
N
P. Furthermore, if vanishes on T
N
P, then can be chosen so
that its rst derivatives vanish along N.
Proof. Since the statement is local around N, we will identify P with a tubular neighborhood
of N in P and let h
t
: P P be a smooth isotopy such that
h
1
= id h
0
= berwise projection of P onto N.
Since vanishes on N, we have
= h

1
h

0
=

1
0
d
dt
(h

t
) dt =

1
0
h

t
(L
Yt
) dt,
where Y
t
is the (time-dependent) vector eld which generates the isotopy h
t
for t > 0. By
Cartans formula and the fact that is closed, the last expression is equal to

1
0
h

t
d(Y
t
) dt = d

1
0
h

t
(Y
t
) dt

.
Our assertion follows if we take =

1
0
h

t
(Y
t
) dt.
2
If (P, ) is a symplectic manifold and g
t
: P P a continuous family of dieomorphisms,
then homotopy invariance for deRham cohomology implies that the forms
t
= g

t
lie in
the same cohomology class for all t. By the same method of proof as in the relative Darboux
theorem, we can deduce the following converse result:
50
Corollary 4.22 (Moser [45]) If
t
is a family of symplectic structures on a compact
manifold P and
t
1

t
2
is exact for all t
1
, t
2
, then there is a dieomorphism f : P P
with f

1
=
0
.
By restricting to a neighborhood of a point in N, Givental [6] proves an equivalence theorem
like the relative Darboux theorem under the weaker hypothesis that
0
and
1
coincide on
TN, i.e. just for vectors tangent to N. In contrast to Corollary 4.22, McDu [40] describes
a family
t
of symplectic forms on a compact six-dimensional manifold P such that
1

0
is exact, but there is no dieomorphism f : P P at all satisfying f

1
=
0
. Of course,
for intermediate t, the form
t

0
is not exact.
Proof of Theorem 4.19. Beginning with the lagrangian immersion : L P, we let
f : Z P denote the composition , where Z is the zero section of T

L and : T

L L
is the natural projection. Using Theorem 3.20, we then choose a -compatible complex
structure J on f

TP which satises J(f

TZ) = E, where E is the lagrangian subbundle given

in the statement of the theorem. Similarly, we choose a
L
-compatible complex structure J
t
on T
Z
(T

L) which rotates TZ into V

Z
L. Now consider the symplectic bundle map
T
Z
(T

L)

f
f

TP
given by

f(v J
t
w) = f

v J(f

w) for v, w TZ. In a neighborhood U T

L of Z, we
can extend f to obtain an immersion F : U P which satises F

=

f on T
Z
(T

L). To
nish the proof of the normal form theorem, we apply the relative Darboux theorem to the
forms
L
and F

() on U.
2
Next, we turn to two generalizations of the classical Morse lemma (see Appendix B)
which require the following version of Taylors theorem.
Lemma 4.23 Let N B be a submanifold dened by the vanishing of functions g
1
, , g
k
whose dierentials are linearly independent along N. If f is any function such that f and
df vanish at all points of N, then there exist functions c
ij
such that
f =

i,j
c
ij
g
i
g
j
on a neighborhood of N.
Parametrized Morse lemma . Let f : M R
k
R satisfy the condition that for
each x M, (x, 0) is a nondegenerate critical point for f[
|xR
k. Then for each x
0
M,
there exists a neighborhood U of (x
0
, 0), a nondegenerate quadratic form Q on R
k
, and a
dieomorphism u of U xing M0 and preserving bers of the projection to M such that
f(u(x, )) = f(x, 0) + Q().
51
Proof. By a preliminary change of coordinates linear on bers, we may assume that
f(x, ) = f(x, 0) +Q() + a(x, )
where the error term a(x, ) is O([[
3
) for each x. To nd the dieomorphism u we apply
the deformation method. Dene f
t
= f (1 t)a. We seek a vector eld X
t
tangent to the
bers x = constant which generates an isotopy u
t
xing M 0 and satisfying f
t
u
t
= f
0
for all t [0, 1]. This means that X
t
should be of the form
X
t
=

i
h
i
t

i
,
for certain smooth functions h
i
t
which vanish on M 0, and it should satisfy
0 =
d
dt
(f
t
u
t
) = u

t
(X
t
f
t
+ a)
for all t [0, 1]. Evidently, the latter condition will be met if X
t
is chosen so that
X
t
f
t
+ a = 0.
To determine X
t
, we invoke Lemma 4.23 in order to nd smooth functions c
ij
t
which vanish
on M 0 and satisfy
a =

c
ij
t
f
t

i
f
t

j
.
The required condition

i
h
i
t
f
t

i
=

i,j
c
ij
t
f
t

i
f
t

j
will be satised if we set
h
i
t
=

j
c
ij
t
f
t

j
,
which vanishes (to second order) on M 0.
2
If is a phase function on M R
n+k
whose ber-hessian H has rank k at a point b

,
then by the lower-semicontinuity of rank H, there exists an integrable subbundle F

E =
ker(p
M
) such that the ber-hessian

: F F

is nondegenerate at each zero of

the section

in a neighborhood of b. A change of coordinates near b and an application

of the parametrized Morse lemma then proves:
Thom splitting theorem . Let be a function on B = M R
n+k
whose ber-hessian
H has rank k at a point b

. Then there exists a ber-preserving dieomorphism g of

B, a function : M R
n
R, and a nondegenerate quadratic form Q on R
k
such that
( g)(x, ,
t
) = (x, ) + Q(
t
)
in a neighborhood of b in B. The function is totally degenerate in the sense that its
ber-hessian is identically zero at b.
2
52
Existence and stable equivalence of Morse families
Returning to the proof of Theorem 4.18, we rst note the following reinterpretation of the
Thom splitting theorem in the terminology of Morse families.
Theorem 4.24 Any Morse family in M(L, , p) is stably equivalent to a reduced Morse
family in M(L, , p).
To show that M(L, , p) is nonempty, we rst prove
Theorem 4.25 Suppose that : L T

M is an exact lagrangian immersion such that

T(T

M) admits a lagrangian subbundle F transverse to both

V M and

TL. Then L is
generated by a Morse family over M.
Proof. By Theorem 4.19, there exists a symplectic immersion from a neighborhood U of
the zero section Z T

L into T

M which satises = on Z and which maps V

Z
L
onto the subbundle F. Since is symplectic, the 1-form

M

L
is closed on U. Its
restriction to Z is exact because
L
is zero on Z. By the relative Poincare lemma, there
exists a function on U satisfying d =

M

L
.
To complete the proof, we rst note that the restriction of d to a ber of the compo-
sition of with the projection
t
: T

M M equals
L
. Combined with the denition
of
L
, a computation (using the transversality hypothesis) now shows that there exists a
neighborhood B of Z within U such that (B,
t
((B)),
t
, ) is a Morse family such that

= Z and

= for the projection : T

L L.
2
Proof of Theorem 4.18. Using Lemma 3.9 we can construct a subbundle F which satises
the hypotheses of the preceding theorem over a neighborhood of any p in L. Consequently,
there exists a Morse family generating a neighborhood of p, and so the class M(L, , p) is
nonempty. Combined with Theorem 4.24, this proves part (1) of Theorem 4.18.
To prove part (2), it suces by Theorem 4.24 to show that any two reduced Morse
families in M(L, , p) are stably equivalent. Since this property is local near p, we may
assume that M = R
m
and and that ,

are both dened on R
m
R
k
. Consider the
mappings F,

F : R
m
R
k
R
m
R
m
dened by
F(x, y) =

x,

x
(x, y)

F(x, y) =

x,

x
(x, y)

.
By the assumption that ,

are nondegenerate phase functions, it follows that F and

F are
embeddings near q =
1

(p) and q =
1

(p) respectively. Moreover, F coincides with the

map

on the ber-critical set

near q, and similarly for

F. Since the Morse families
are reduced, it follows that the images of DF
q
and D

F
q
coincide in T
p
(R
m
R
m
). By an
application of the implicit function theorem, we can therefore nd a ber-preserving map g
53
of R
m
R
m
whose restriction to some neighborhood B of q is a dieomorphism which sends

to

and in particular g(q) = q.

Recall from the proof of Theorem 4.15 that the identity
d =

M
holds on the ber critical set

. Since

g, this identity implies that d = d(

g)
on

, and so the phase functions and

g dier by an additive constant (which we may
assume to be zero) at points of

. In order for g to dene an equivalence between and

, this property would have to be valid on all of B. While there is no reason to expect this
of g itself, we will construct a similar dieomorphism with this property by appealing once
again to the deformation method:
Let
0
= ,
1
=

g. Then

0
=

1
= and

0
=

1
. Moreover
1

0
vanishes
to second order along , and so there exist functions c
ij
dened near b such that

0
=

i,j
c
ij

j
in a neighborhood of

.
As before, we seek a vector eld X
t
generating an isotopy f
t
that satises f

t

t
=
0
for
all t [0, 1]; to insure that each f
t
xes and preserves bers, we must also require that X
t
be of the form
X
t
=

i,j
h
ij

j
for certain functions h
ij
which vanish on .
To arrive at an equation for these coecients, we note that the equation
0 =
d
dt
(f

t

t
) = f

t
(X
t

t
+
1

0
)
will be satised provided that
0 = X
t

t
+
1

0
,
i.e.
0 =

i,j
c
ij

j
+

i,j
h
ij

0
+ t

u,v
c
uv

.
This equation holds if
0 = C + H(I + S),
where C = (c
ij
), H = (h
ij
), and S is a matrix function which vanishes at b for all t, since
0
is reduced. Hence we can solve for H in a neighborhood of b. This completes the proof.
2
54
Maslov objects
If : L T

M is any lagrangian immersion, then the symplectic vector bundle

T(T

M)
over L has two lagrangian subbundles, L
1
=

TL and L
2
=

V M. The Maslov class of

(L, ) is dened as the degree-1 cohomology class

L,
= (L
1
, L
2
) H
1
(L; R)
described in Example 3.21. From this denition, it follows that, unlike the Liouville classes,
the Maslov classes of two immersions (L, ) and (L,
t
) are equal whenever and
t
are
homotopic through lagrangian immersions.
Associated to any Morse family (B, p
B
, ) over a manifold M is an index function ind

:
L

Z dened by
ind

(p) = index(H

(p)
),
where the index of a quadratic form is, as usual, the dimension of the largest subspace on
which it is negative-denite. Since the ber-hessian is nondegenerate where L

is projectable
(see the discussion following Theorem 4.15), the index function ind

is constant on any
connected projectable subset of L

. From Theorem 4.18, it follows furthermore that any two

index functions ind

, ind

dier by an integer on each connected component of L

.
Example 4.26 Consider (x, ) =
3
/3 + (x
2
1). We have / =
2
+ x
2
1; thus
the ber critical set

is the circle
2
+ x
2
= 1, and its image under

is the gure-eight.
The caustic set of the projection L R consists of the points (1, 0), (1, 0). To compute
the index function corresponding to , we note that

2
= 2;
consequently ind

(x, p) = 0 for (x, p) lying on the upper right part of the curve and
ind(x, p) = 1 for (x, p) on the upper left. Since this loop is generated by a single phase
function, there is a corresponding global index function, and the index of the loop is neces-
sarily zero.
A dierent situation occurs in the case of a circle. The circle has two caustics, and the
jump experienced by any index function as one passes through a caustic is given by the
example just computed: passing through the right caustic in the +p direction decreases the
index by 1, while passing through the left caustic in the same direction increases the index
by 1. Traversing the circle in a counterclockwise direction, we see that the total index must
change by 2. Consequently, the circle does not admit a global generating function.
Of course, the Liouville class of the circle is also nonzero. However, it is easy to deform
the circle to a closed curve with two transverse self-intersections which has zero Liouville
class, but still admits no global phase function, since its Maslov class is nonzero.

55
As a degree-1 cohomology class on L, the Maslov class determines via the exponential
map R U(1) an isomorphism class of at hermitian line bundles over L. A canonical
representative of this class can be constructed as follows. First, we consider the union
M(L, ) =

pL
M(L, , p)
with the discrete topology. On the subset of LM(L, )Z consisting of all (p, (B, p
B
, ), n)
such that (B, p
B
, ) M(L, , p), we now introduce an equivalence relation by setting
(p, (B, p
B
, ), n) ( p, (

B, p

B
,

), n) provided that p = p and
n + ind

(p) = n + ind

(p).
The quotient space with respect to this relation is a principal Z-bundle M
L,
over L which
we will call the Maslov principal bundle.
Associated to the Maslov principal bundle via the representation n e
in/2
of Z in U(1) is
a complex line bundle
L,
over L called the Maslov line bundle. Having discrete structure
group, this line bundle carries a natural at connection with holonomy in e
in/2
Z
4
.
Our main use of the Maslov line bundle will be to modify the half-densities on L in order to
incorporate the Maslov correction into our quantization scheme.
4.4 WKB quantization
In this section we will combine the tools assembled in the preceding sections into a technique
for quantizing half-densities on lagrangian submanifolds of arbitrary cotangent bundles.
The phase bundle associated to an immersed lagrangian submanifold : L T

M and
> 0 is dened as the tensor product

L,,
def
=
L,

c
M,
,
where we recall that c
M,
is the prequantum line bundle over T

M (see Section 4.1). Observe

that the product of the natural at connections on
L,
and

c
M,
denes a at connection
on the phase bundle whose holonomy is represented by the mod-Z

reduction of the real

cohomology class

L,
+
L,
/2

H
1
(L; R),
which we call the phase class of (L, ). The phase bundle
L,,
can be described explicitly
as the collection of all quintuples (p, t, (B, p
B
, ), n, z) where (p, t, n, z) L T

Z C
and (B, p
B
, ) M(L, , p), modulo the equivalence relation given by
(p, t, (B, p
B
, ), n, z) ( p,

t, (

B, p

B
,

), n, z)
whenever p = p and
z e
it/
e
i(n+ind

(p))/2
= z e
i

t/
e
i( n+ind

(p))/2
.
56
A Morse family (B, p
B
, ) which generates an open subset L

of L denes a nonvanishing
parallel section of
L,,
over L

by
s
,
(p) = [p, 0, (B, p
B
, ), 0, e
i(y)/
],
where

(y) = p. A check of these denitions shows that whenever (p) =

(y) =

( y),
s
,
(p) e
i(y)/
e
i ind

(p)/2
= s

,
(p) e
i

( y)/
e
i ind

(p)/2
.
For each R
+
, we denote by
par
(
L,,
) the space of parallel sections of
L,,
. If, for
a particular , the phase class of (L, ) is -integral, then
par
(
L,,
) is a complex vector
space isomorphic to C. Otherwise,
par
(
L,,
) consists of a single point (the zero section of

L,,
). The product

L,
=

>0

par
(
L,,
)
then has the structure of a C-module. An element s
L,
is then a (generally discontinuous)
function which assigns to each > 0 an element s

in
par
(
L,,
), so that the map p s

(p)
denes a parallel section of
L,,
. The symbol space of (L, ) is dened as the complex
vector space
o
L,
def
= [[
1/2
L
C

L,
.
The amplitude bundle /

associated to a Morse family (B, p

B
, ) over a smooth man-
ifold M is dened as the complex line bundle
/

= [[
1/2
B [[
1/2
E
over B, where E again denotes the subbundle ker(p
B
) of TB. An amplitude is a section a of
/

. We will say that a is properly supported provided that the restriction of p

B
: B M
to Supp(a) is a proper map. The purpose of the space of amplitudes is to dene a relation
between half-densities on M and symbols on the subset L

of L generated by . To describe
this relation, we begin by noting that from the exact sequence of vector bundles
0 E TB p

B
TM 0
over B, it follows that [[
1/2
B is naturally isomorphic to [[
1/2
p

B
TM [[
1/2
E. This in
turn gives rise to the natural isomorphism
/

[[
1/2
p

B
TM [[E.
The image of an amplitude a on B under this isomorphism can be written as p

B
[dx[
1/2
,
where is a family of 1-densities on the bers of p
B
, i.e.,
x
is a density on each nonempty
p
1
B
(x). By ber-integration we pass to a half-density on M:
I

(, a)(x) = (2)
n/2
e
in/4

p
1
B
(x)
e
i/

[dx[
1/2
,
57
where n = dim(p
1
B
(x)), setting I

(, a)(x) = 0 if p
1
B
(x) = . When a is properly supported,
we may dierentiate under the integral to conclude that I

(, a) is a smooth half-density on
M.
To pass geometrically from a to a symbol on L

, we rst recall from Section 4.3 that

the nondegeneracy of gives rise to the following exact sequence of vector bundles over the
ber-critical set

0 T

B
d

0.
Since [[
1/2
E is naturally isomorphic to [[
1/2
E

, this sequence induces an isomorphism

of the restriction of /

to

with [[
1/2

. If the restriction of a to

corresponds to a
half-density a on

under this isomorphism, the associated symbol on L

is dened as
s
a
def
= g

a s

.
(Here we recall from the discussion following Denition 4.17 that g

is a dieomorphism
from a neighborhood of p in L onto

dened by the composition

1

. Also, for each

, the section s
,
is the canonical element of
par
(
L

,,
) dened above). When (L, ) is
projectable, the symbol s
a
and the half-density I

(, a) are linked by the following theorem.

Theorem 4.27 Suppose that two Morse families (B, p
B
, ), (

B, p

B
,

) generate the same
projectable lagrangian embedding (L, ), and let a, a be amplitudes on B,

B, respectively.
Then s
a
= s
a
on L if and only if
[I

(, a) I

, a)[ = O()
locally uniformly on V .
Proof. By Theorem 4.18, the Morse families (B, p
B
, ), (

B, p

B
,

) are stably equivalent, and
so there exists a dieomorphism g :

dened as the composition g = g

g
1

. A
check of the denitions shows that the symbols s
a
and s
a
are equal precisely when
g

a e
i/
e
i ind

/2

= a e
i

/
e
i ind

/2
,
where a is the half-density on

induced by the amplitude a, and similarly for a. Since

= g, up to a constant, Lemma B.3 implies that this occurs precisely when

[ det
ax
H[
1/2
e
i/
e
i ind

/2
= [ det
ax
H

[
1/2
e
i

/
e
i ind

/2
. ()
Since (L, ) is projectable, we can apply the principle of stationary phase to each ber of
the projection p
B
: B M to obtain
I

(, a)(x) =
e
i/
e
i ind

/2
[ det
ax
H[
1/2
+ O()
and similarly for I

, a)(x). The theorem follows by comparing these expressions with ()

above.
58
2
Morse families provide a general means for locally quantizing symbols on an immersed
lagrangian submanifold (L, ) T

M. Suppose that (B, p

B
, ) is a Morse family such
that the phase function generates an open subset L

L, and consider a symbol s on

L supported in L

. Then there exists a unique half-density a supported in L

such that
s = s

a, and (g

a may be canonically identied with a section of the amplitude bundle

of B over the ber-critical set

, which we can extend to an amplitude a on B, compactly

supported in bers. We then set
I

(L, , s) = I

(, a)(x).
From Theorem 4.27 we draw two conclusions about this tentative denition when L

is
projectable. First, we note that if (

B, p

B
,

) is a second Morse family which generates L

,
and a is an amplitude on

B obtained as above, then s
a
= s
a
= s, and so
[I

(, a) I

, a)[ = O().
Thus, I

(L, , s) is well-dened, up to O() terms. Furthermore, we note that Theorem 4.27

asserts that this choice of I

(L, , s) coincides with pull-back.

To quantize an arbitrary symbol s on L, we rst x a locally nite cover L
j
of L, such
that each L
j
is generated by a Morse family (B
j
, p
B
j
,
j
), and choose a partition of unity
h
j
subordinate to L
j
. We then set
I

(L, , s) =

j
I

(L, , h
j
s).
Using the remarks above, it is easy to check that up to O() terms, this denition depends
only on the semi-classical state (L, , s).
Of course, although the technique above enables us to quantize all symbols on L in a
consistent way, the existence of nonzero symbols requires that the phase bundle
L,
admit
nontrivial parallel sections, i.e. that the phase class of (L, ) be -integral. For this reason,
we introduce the following terminology.
Denition 4.28 An immersed lagrangian submanifold : L T

M is called quantizable
if, for some R
+
, its phase bundle
L,
admits a global parallel section, or, equivalently,
if its phase class is -integral. The set of for which this condition holds will be called
admissible for L.
This denition is known as the Maslov quantization condition. Note that it is a straight-
forward generalization of the condition derived in the preceding section.
Example 4.29 Let N be a closed submanifold of a smooth manifold M, and let U be a
tubular neighborhood of N, i.e., U is the image of the normal bundle
N
T
N
M under an
embedding :
N
M satisfying = on the zero section of
N
, where :
N
N is
the natural projection. Consider the Morse family (r

, p
N
, ), where r =
1
is a
59
retraction of U onto N, p
N
denotes the natural submersion r

M, and : r

R
is dened by
(p) = 'p,
1
(p
N
(p))`.
Since is an embedding, a computation shows that the ber critical set of is given by

= p
1
N
(N) = N

, and the map

: N

M equals the inclusion. Thus, the conormal

bundle of N is a lagrangian submanifold of T

M which admits a global generating function,

and therefore both the Liouville and Maslov classes of N

are zero. In particular, this

implies that the conormal bundle of any submanifold of M satises the Maslov quantization
condition.

Quantum states as distributions

Unfortunately, the interpretation of I

(L, , s) at regular values of

L
is not valid at caustics.
Indeed, this remark is suggested by the fact that I

(L, , s) is smooth, whereas we saw in

Section 2.2 that classical solutions to the transport equation are singular at caustic points.
The basic technical diculty is that the principle of stationary phase (the basic underpinning
of Theorem 4.27) no longer applies to the integral

p
1
B
(x)
e
i/

x
when x is a caustic point, since the phase function has a degenerate critical point in the
ber p
1
B
(x).
Example 4.30 The ber critical set of the phase function (x, ) =
3
/3 + x consists of
the parabola x =
2
, at whose points the ber-hessian assumes the form
2
/
2
= 2x.
The origin is therefore a degenerate critical point for , and stationary phase cannot be used
to estimate the integral

R
e
i
3
/3
a(0, ) d.

A more appropriate way to interpret the expression I

(L, , s) in the presence of caustic points

is as a family of distributional half-densities on M dened as follows. For each Morse family
(B, p
B
, ) M(L, ) and compactly supported amplitude a on B, we dene a distributional
half-density on M by
'I

(, a), u` = (2)
b/2
e
ib/4

B
e
i/
a p

B
u,
where b = dim(B) and u [[
1/2
0
M. The family I

(L, , s) then consists of all distributional

half-densities I

on M obtained by choosing a locally nite open cover L

j
of L and a
60
partition of unity h
j
subordinate to L
j
. Then set
I

=
k

j=1
I

(
j
, a
j
).
The family I

(L, , s) consists of those distributional half-densities obtained in this way using

amplitudes a
j
such that s
a
j
= s over each L

j
.
Although the class I

(L, , s) may appear very large, a link among its members can be
described as follows. We rst prove
Theorem 4.31 Suppose that two Morse families (B, p
B
, ), (

B, p

B
,

) generate the same
lagrangian embedding (L, ), and let a, a be amplitudes on B,

B, respectively. If : M R
is a smooth function whose dierential intersects (L) at exactly one point (p) transversely,
then s
a
(p) = s
a
(p) if and only if

'I

(, a), e
i/
u` 'I

, a), e
i/
u`

= O().
Proof. Set R = p
B
. Since p
B
is constant on the bers of p
B
, the function R is a
phase function having the same ber critical set as .
As in the proof of Theorem 4.27, equality of the symbols s
a
and s
a
occurs precisely when
the dieomorphism g :

satises
g

a
y
e
i(y)/
e
i ind

(p)/2

= a
y
e
i

( y)/
e
i ind

(p)/2
,
where a is the half-density on

induced by the amplitude a, and similarly for a. Since

R = R g, up to a constant, Lemma B.3 shows that this condition is equivalent to

[ det
a
R
tt
(y)[
1/2
e
i(y)/
e
i ind

(p)/2
= [ det
a

R
tt
( y)[
1/2
e
i

( y)/
e
i ind

(p)/2
.
Since g is ber-preserving, we have R(y)

R( y) = (y)

( y). Moreover, it is easy
to check that stable equivalence of the Morse families (B, p
B
, ), (

B, p

B
,

) implies that
ind

(p) ind

( y) = ind R
tt
(y) ind

R
tt
( y), and so the preceding equation gives
e
iR(y)/
e
i ind R

(y)/2
[ det
a
R
tt
(y)[
1/2
=
e
i

R( y)/
e
i ind

R

( y)/2
[ det
a

R
tt
( y)[
1/2
. ()
The critical point y of R is nondegenerate precisely when the intersection of (L) with
im(d) is transverse. In this case, an application of the principle of stationary phase gives
'I

(, a), e
i/
u` =
e
iR(y)/
e
i ind R

(y)/2
[ det
a
R
tt
(y)[
1/2
+ O()
and similarly for 'I

, a), e
i/
u`. Comparing this expression with () above completes
the proof.
2
61
Now consider a semi-classical state (L, , s). As before, we let L
j
be a locally nite
cover of L such that each L
j
is generated by a Morse family (B
j
, p
B
j
,
j
) over M, and choose
a partition of unity h
j
subordinate to L
j
. We then dene
'I

(L, , s), e
i/
u`
def
=

j
'I

(
j
, a
j
), e
i/
u`,
where a
j
is the amplitude on B
j
obtained in the usual way from the symbol h
j
s on L
j
. When
the image of d is transverse to L, then, up to O(), the principal part of 'I

(L, , s), e
i/
`
depends only on the principal part of s. For a thorough exposition of this topic, we refer to
[21, 28, 31].
Example 4.32 Suppose that (B, p
B
, ) is a Morse family over a manifold M, and suppose
that S : V R is a smooth function. By setting

S = S p
B
, we obtain a new Morse
family (B, p
B
,

S) for which

S
and

= f
dS

S
, where f
dS
is the berwise
translation map dened by dS.
Note in particular that if dS(p
B
(x)) =

(x) for some x

, then x is a critical point

of

S. This critical point is nondegenerate provided that

S
is transverse to the zero
section of T

M near x, or, equivalently, if

is transverse to the image of dS near x.

By allowing I

(, a) to be a distribution, we can sometimes dene it even when a does

not have compact support.
Equivalent semi-classical states
A further important modication of our quantization picture is based on the conceptual
distinction between an immersion and an immersed submanifold. In terms of our dis-
cussion, this means the following. If : L T

M and
t
: L
t
T

M are lagrangian
immersions, we will say that (L, ) and (L
t
,
t
) are equivalent provided that there exists a
dieomorphism f : L L
t
such that =
t
f. In this way, any lagrangian immersion
: L T

M denes an equivalence class of lagrangian immersions in T

M which, for nota-

tional simplicity we will denote by L. We refer to the equivalence class L as an immersed
lagrangian submanifold of T

M.
A check of the denitions shows that the main objects in our quantization scheme behave
nicely with respect to this notion of equivalence. On the classical level, a dieomorphism
f : L L
t
as above induces an isomorphism of symbol spaces
o
L,
f
o
L

,
.
Thus, a symbol on the immersed lagrangian submanifold L is well-dened as an equivalence
class of symbols on the members of L. Moreover, if E is a regular value of some hamiltonian
function H on T

M, then the Hamilton-Jacobi equation

H = E
62
denes a condition on the class L which we denote by H(L) = E. In this case, the vector
elds X
H,
and X
H,
induced on (L, ), (L
t
,
t
) by the hamiltonian vector eld of H (see
Example 3.13) satisfy X
H,
= f

X
H,
, and therefore the homogeneous transport equation
L
X
H,
s = 0
is a well-dened condition on (L, s) which we denote by L
X
H
s = 0.
From these remarks, we are led to view the equivalence class (L, s) as a semi-classical
state in T

M. The state is stationary with respect to a classical hamiltonian H on T

M
provided that (L, s) satises the associated Hamilton-Jacobi and homogeneous transport
equations, as described above. On the quantum level, it is easy to check that for any
members (L, , s), (L
t
,
t
, s
t
) of the equivalence class (L, s), we have
I

(L, , s) = I

(L
t
,
t
, s
t
),
and thus we may dene the quantization of (L, s) as the (unique) distributional half-density
I

(L, s) on M obtained by quantizing any member of (L, s). If (L, s) is a stationary semi-
classical state, then I

(L, s) is a semi-classical approximate solution to the time-independent

Schrodinger equation on M.
Our list of classical and quantum correspondences now assumes the form:
Object Classical version Quantum version
basic space T

M H
M
state (L, s) as above distributional half-density
on M
time-evolution Hamiltons equations Schrodinger equation
generator of evolution function H on T

M operator

H on H
M
stationary state state (L, s) satisfying eigenvector of

H
H(L) = E and L
X
H
s = 0
With these correspondences between classical and quantum mechanics in mind, we are fur-
ther led to dene a semi-classical state in an arbitrary symplectic manifold (P, ) to be an
immersed lagrangian submanifold equipped with a half-density and possibly some phase
object corresponding to a parallel section of the phase bundle in the cotangent bundle case.
Note, however, that it is unclear what the quantum states are which are approximated by
these geometric objects, since there is no underlying conguration space on which the states
can live. Extending some notion of quantum state to arbitrary symplectic manifolds is one
of the central goals of geometric quantization.
63
5 The Symplectic Category
There are many symplectic manifolds which are not cotangent bundles. For instance, an
important process known as symplectic reduction generates many examples of such manifolds,
starting with cotangent bundles. We begin this section with a discussion of reduction and
then turn to the classical and quantum viewpoints in the context of general symplectic
manifolds, concluding with Diracs formulation of the quantization problem.
5.1 Symplectic reduction
The technique of symplectic reduction geometrizes the process in mechanics in which rst
integrals of the hamiltonian are used to eliminate variables in Hamiltons equations.
Linear symplectic reduction
Degenerate skew-symmetric bilinear forms yield symplectic vector spaces in the following
way.
Lemma 5.1 A skew-symmetric bilinear form on a vector space Y induces a symplectic
structure on Y/Y

.
Proof. First note that gives rise to a skew-symmetric bilinear form
t
on Y/Y

by the
equation

t
([x], [y]) = (x, y).
To prove the nondegeneracy of
t
, we use the following commutative diagram, where is
the projection.
Y

(Y/Y

Y

Y/Y

If
t
([x]) = 0, then (x) = 0, and so x Y

and hence [x] = 0.

2
The symplectic quotient space Y/Y

described in this lemma is known as the (linear) re-

duced space associated to Y . A special case of linear reduction arises when Y is a coisotropic
subspace of a symplectic vector space and equals the restriction of the symplectic form to
Y . Lagrangian subspaces behave remarkably well with respect to this reduction.
Lemma 5.2 Let V be a symplectic vector space and L, C V a lagrangian and a coisotropic
subspace respectively. Then
L
C
= L C + C

is a lagrangian subspace of V contained in C, and

L
C
= (L C)/(L C

)
is a lagrangian subspace of C/C

.
64
Proof. Note that
(L
C
)

= (L + C

) C.
Since by assumption C

C, the self-orthogonality of L
C
follows from the simple fact that
if E, F, G are any three subspaces of V , then (E+F) G = EG+F if and only if F G.
Next, we observe that since L
C
= L
C
/C

, the second assertion follows from the equality

L

C
= (L
C
)

/C

.
2
Example 5.3 If (V, ), (V
t
,
t
) are symplectic vector spaces and L is a lagrangian subspace
of V , then then C = V
t
L is a coisotropic subspace of the direct sum V
t
V (see Exam-
ple 3.4), and C

= 0 L. Consequently, the linear reduced space C/C

equals (V
t
,
t
).
Now suppose that V and V
t
are of equal dimension and T : V V
t
is a linear symplectic
map. If
T
V
t
V is the lagrangian subspace dened by the graph of T, then the reduced
lagrangian subspace (
T
)
C
of C/C

equals T(L) V
t
.

Observe that if C is a coisotropic subspace of (V, ), then C

V/C also carries a natural

symplectic structure induced by . This gives rise to the following decomposition of V .
Lemma 5.4 If (V, ) is a symplectic vector space with a coisotropic subspace C, then there
exists a linear symplectomorphism
V C

V/C C/C

.
Proof. Let J be a -compatible complex structure on V , and set A = JC

and B =
CJC. Then A, B are orthogonal to C, C

with respect to the inner-product g

J
on V . The
projections V V/C and C C/C

restricted to A, B give rise to an isomorphism

V = C

+ A + B C

V/C C/C

.
2
Recall that if L, L
t
are lagrangian subspaces of a symplectic vector space V and W V
is a lagrangian subspace transverse to both L and L
t
, then there exists a natural linear
symplectomorphism from V to LL

which sends W onto the subspace 0 L

and L
t
onto
the graph of some self-adjoint linear map T : L L

. Denoting by Q
T
the quadratic form
Q
T
on L induced by T, we dene
ind(L, L
t
; W) = index Q
T
sgn(L, L
t
; W) = signature Q
T
These quantities will be useful in our study of the Maslov bundle under reduction, and we
collect some useful facts about them in the following examples.
65
Example 5.5 To begin, we leave to the reader the job of checking the following elementary
identities for the case when the lagrangian subspaces L, L
t
are themselves transverse:
sgn(L, L
t
; W) = sgn(L
t
, L; W) = sgn(L, W; L
t
).

Example 5.6 Suppose that V is a nite-dimensional vector space with a subspace E V

and its algebraic orthogonal E

. Then C = V E

is a coisotropic subspace of V V

with its usual symplectic structure, and C

= E 0. The reduced symplectic vector space

is then
C/C

= (V/E) E

.
The composition of a self-adjoint linear map A : V

V with the projection V V/E

restricts to a self-adjoint linear map A
E
: E

V/E (note that V/E identies canonically

with (E

). Evidently, the lagrangian subspace W of V V

given by the graph of A passes

under reduction by C to the graph W
C
of A
E
. Now if we denote by L, L
t
the lagrangian
subspaces V 0 and 0 V

, respectively, then L
C
= (V/E) 0 and L
t
C
= 0 E

. If Q

denotes the quadratic form on V

dened by A, we therefore have

sgn(Q

) = sgn(L
t
, W; L) and sgn(Q

[
E
) = sgn(L
t
C
, W
C
; L
C
).
From [21, p.130] we recall that if Q

is nondegenerate and Q is the quadratic form it induces

on V , then
sgn(Q) = sgn(Q[
E
) + sgn(Q

[
E
).
Combined with the preceding equations, this formula gives
sgn(L
t
, W; L) = sgn(Q[
E
) + sgn(L
t
C
, W
C
; L
C
).

Presymplectic structures and reduction

By denition, a symplectic form is closed and nondegenerate. In some sense, the next best
structure a manifold M may possess along these lines is a closed two-form of constant
rank, i.e., with the dimension of the orthogonal (T
x
M)

the same for all x M. In this

case, is called a presymplectic structure on M with characteristic subbundle (TM)

.
Theorem 5.7 The characteristic subbundle of a presymplectic manifold (M, ) is integrable.
Proof. Leaving for the reader the verication that (TM)

is actually a subbundle of TM,

we recall that Lie brackets and inner products are related by the formula
[X, Y ] = L
X
(Y ) Y L
X
.
If the vector eld Y belongs to (TM)

, then the rst term on the right hand side of this

equation vanishes identically. If X belongs to (TM)

, then Cartans formula combined with

the assumption that is closed implies that is invariant under the ow of X. This means
that L
X
= 0, so the second term vanishes as well, and [X, Y ] lies in (TM)

.
66
2
The foliation

dened by the characteristic subbundle of M is known as the characteris-

tic foliation of M. If the quotient space M/

is a smooth manifold, then we say that M

is reducible. A pointwise application of Lemma 5.2, together with the fact that L
X
= 0
for characteristic X, shows that the presymplectic structure induces a smooth, nonde-
generate 2-form
/
on M/

. Since d = 0 by hypothesis, and since the quotient map

M M/

is a submersion, the form

/
is necessarily closed and therefore symplectic.
The symplectic manifold (M/

,
/
) is called the reduced manifold of M.
For the most part, we will be interested in presymplectic manifolds which arise as
coisotropic submanifolds of some symplectic manifold (P, ). Recall that a submanifold
C P is called coisotropic if, for each p C, the tangent space T
p
C contains its symplectic
orthogonal
(T
p
C)

= v T
p
P : (v, w) = 0 for all w T
p
C.
In this case, we can view C as an abstract manifold and note that if
t
is the pull-back
of the symplectic form by the natural inclusion C P, then ker(
t
) = (TC)

. Since
dim(TC)

= dim(P) dim(C) is constant, the form

t
denes a presymplectic structure on
P.
Example 5.8 As noted in Example 3.13, the hamiltonian ow associated to a function
H : R
2n
R generates the characteristic foliation of any regular energy surface H
1
(E).
The corresponding reduced symplectic manifold (when it exists) will be the symplectic model
for the space of quantum states of energy E.
If P, Q are symplectic manifolds, and L is a lagrangian submanifold of Q, then P L is a
coisotropic submanifold of P Q whose characteristic foliation consists of leaves of the form
p L for p P. The product P L is therefore reducible, and the reduction projection
coincides with the usual cartesian projection P L P.

Our goal for the remainder of this section is to describe two operations on immersed
lagrangian submanifolds dened by a reducible coisotropic submanifold; in eect, these op-
erations will be nonlinear analogs of the mappings L L
C
and L L
C
described in the
linear setting in Lemma 5.2. To begin, we recall that the ber product of two maps
: N V and : M V is dened as the subset
N
V
M = ( )
1

V
of N M, where : N M V V is the product map and
V
V V is the
diagonal. This gives rise to the following commutative diagram
N
V
M
r
M
M
r
N

N

V
where r
M
, r
N
denote the restrictions of the cartesian projections NM M, N to the ber
product N
V
M.
67
Denition 5.9 If N, M, V are smooth manifolds, then two smooth maps : N V and
: M V are said to intersect cleanly provided that their ber product N
V
M is a
submanifold of N M and

TN

TM = ( r
M
)

T(N
V
M)
as subbundles of ( r
M
)

TV .
For brevity, we will say that a map : N V intersects a submanifold W V cleanly if
and the inclusion map W V intersect cleanly.
Example 5.10 1. If the product of two smooth maps : N V and : M V
is transverse to the diagonal
V
, then and intersect cleanly. A particular case of this
situation occurs when is a submersion and is any smooth map.
2. Suppose that : N V and : M V are any smooth maps whose ber product N
V
M
is a smooth submanifold of NM. Then for any tangent vector (v, w) of N
V
M, the vector
( )

(v, w) is tangent to the diagonal in V V , and so

v =

w. Since r
M
(v, w) = w,
it follows that r
M
(v, w) = 0 only if w = 0, in which case

v =

w = 0. Thus, if is an
immersion, we can conclude that r
M
: N
V
M M is an immersion as well.
In particular, if and are cleanly intersecting immersions, then the maps r
N
and r
M
are themselves immersions.
3. A basic example of two maps which do not intersect cleanly is provided by two embedded
curves in the plane which intersect tangentially at a single point.

The usefulness of clean intersection in symplectic geometry lies in its compatibility with
symplectic orthogonalization, as in the following lemma.
Lemma 5.11 Let P be a symplectic manifold. If a lagrangian immersion : L P inter-
sects a coisotropic submanifold C P cleanly, then the map (p
C
r
C
): L
P
C C/(

has
constant rank.
Proof. Let F, G denote the lagrangian and coisotropic subbundles r

L
TL and r

C
TC of the
symplectic vector bundle r

C
TP over the ber product L
P
C. The kernel of (p
C
r
C
)

then
equals
G

r
C
T(L
P
C).
The clean intersection hypothesis implies that r
C
T(L
P
C) = F G, and so the basic
properties of symplectic orthogonalization applied berwise to these bundles give
ker(p
C
r
C
)

= G

G F
= G

F
= (G+ F)

.
68
Since F G has constant rank, the internal sum F + G and its orthogonal (F + G)

also
have constant rank, indicating that the dimension of ker(p
C
r
C
)

is independent of the base

point in L
P
C. In fact, we have dim(ker(p
C
r
C
)

) = dim(L) + dim(L
P
C) dim(C).
2
If the quotient of L
P
C by the bers of the map p
C
r
C
is a smooth, Hausdor manifold,
then we obtain an induced immersion
C
from the reduced space L
C
of L
P
C into C/(

.
In this case, we dene L
C
as the ber product of
C
and the quotient map p
C
: C C/(

.
By
C
: L
C
C/(

we denote the restriction of the cartesian projection L

C
C C to
L
C
.
Recall from Chapter 4 that an immersed submanifold N of a manifold V is an equivalence
class N of immersions, where : N V is considered equivalent to
t
: N
t
V provided that
there exists a dieomorphism f : N N
t
which intertwines and
t
, i.e. =
t
f. Evidently,
if a member : N V of the equivalence class N intersects a smooth map : M V cleanly,
then the same is true of any other member of N. In this case, we will say that the immersed
submanifold N and the map : M V intersect cleanly.
With these remarks in mind, we will say that a coisotropic submanifold C and an im-
mersed lagrangian submanifold L in a symplectic manifold P form a reducible pair (C, L)
provided that C is reducible, and L, C intersect cleanly. We leave it to the reader to check
that any reducible pair (C, L) in P induces immersed submanifolds L
C
and L
C
of C/(

and
P, respectively.
Theorem 5.12 If (C, L) is a reducible pair in a symplectic manifold P, then L
C
and L
C
are immersed lagrangian submanifolds of C/(

and P, respectively.
Proof. Let : L P be a xed element of L. Since L
C
is the quotient of L
P
C by the
bers of p
C
r
C
, the induced map
C
: L
C
C/(

is a smooth immersion. A pointwise

application of Lemma 5.2 shows that this immersion is lagrangian.
Since
C
: L
C
C/(

is an immersion, Example 5.10(2) implies that the map

C
:
L
C
C is a smooth immersion, and a pointwise application of Lemma 5.2 implies that this
immersion is lagrangian.
Finally, since the immersed submanifolds L
C
and L
C
are well-dened by the reducible
pair (C, L), it follows that these are immersed lagrangian submanifolds.
2
Example 5.13 If (M, g) is a riemannian manifold, all of whose geodesics are closed and
have the same length, then for each E > 0, the constant energy hypersurface C = k
1
M
(E)
(see Section 3.2) is reducible, and the reduced manifold CM = C/(

is the space of oriented

geodesics on M. The tangent space to CM at a point p identies naturally with the space
of Jacobi elds normal to the geodesic represented by p. Moreover, it is easy to check that
for any pair J
1
, J
2
of normal Jacobi elds along p, the symplectic form on CM is given in
terms of the metric by
(J
1
, J
2
) = g(J
1
, J
t
2
) g(J
t
1
, J
2
).
69
For further details, see [12].
The conormal bundle N

to a smooth submanifold N M is a lagrangian submanifold

of T

M which intersects C cleanly. More precisely, N

C is a sphere bundle over N

transverse to the characteristic foliation of C. From Theorem 5.12 it therefore follows that
the space of geodesics in M normal to N at some point comprises an immersed lagrangian
submanifold of CM.
To give a concrete case of this example, we consider the sphere S
n
equipped with its stan-
dard metric. Each oriented geodesic in S
n
identies naturally with an oriented 2-dimensional
subspace of R
n+1
, and therefore CS
n
is represented by the grassmannian Gr
+
(2, n + 1).

Example 5.14 Note that the standard symplectic structure on R

2n+2
is equivalent to the
imaginary part of the standard hermitian metric on C
n+1
. The unit sphere S
2n+1
is a
coisotropic submanifold of C
n+1
, the reduction of which is complex projective space CP
n
.
Since the symplectic form on C
n+1
is invariant under unitary transformations, the resulting
symplectic form on CP
n
is invariant under transformations induced by unitary transforma-
tions.
The maximal real subspace R
n+1
C
n+1
is a lagrangian submanifold of C
n+1
whose inter-
section with S
2n+1
is clean and coincides with the real unit sphere S
n
. From Theorem 5.12,
it follows that the real projective space RP
n
is a lagrangian submanifold of CP
n
.

Conormal submanifolds
A particular example of symplectic reduction which we will use in the next chapter begins
with the following set-up. Suppose that M is a smooth manifold, and let N M be a smooth
submanifold equipped with a tangent distribution T, i.e., a subbundle of TN T
N
M. By
C
N
we denote the annihilator of T, that is
C
N
= (x, p) T

M : x N, T
x
ker(p).
By denition, C
N
is a subbundle of T

N
M, but we wish to consider various properties of C
N
when it is viewed as a smooth submanifold of the symplectic manifold (T

M,
M
). We begin
with the following lemma.
Lemma 5.15 In the notation above, the annihilator C
N
is a coisotropic submanifold of T

M
if and only if the distribution T is integrable.
Proof. If T is integrable, then we can consider its leaves as submanifolds of M via the
inclusion N M. The conormal bundles of the leaves dene a foliation of C
N
by lagrangian
submanifolds; by Lemma 3.6, this implies that C
N
is coisotropic.
Now let Z denote the intersection of C
N
with the zero section of T

M. Then the natural

projection T

M M maps Z dieomorphically onto N, and moreover, T

Z
C
N
= TN T

,
70
where T

is the subbundle of V
Z
M which annihilates T. From the natural lagrangian
splitting of T(T

M) along the zero section of T

M, it follows that
(T
Z
C
N
)

= T (TN)

,
and thus the intersections of Z with the isotropic leaves of C
N
are integral manifolds of the
distribution T. This completes the proof.
2
Lemma 5.16 If the leaf space N/T is a smooth, Hausdor manifold, then the reduced space
C
N
/(

is canonically symplectomorphic to T

(N/T).
Proof. Since, by denition, each (x, p) C
N
contains the subspace T
x
of T
x
N in its kernel,
it follows that (x, p) denes an element of T

(N/T). This gives a well-dened surjective

submersion f : C
N
T

(N/T), and it is easy to check that the level sets of this map are the
bers of the projection C
N
C
N
/(

. Thus we get a dieomorphism C

N
/(

(N/T).
To prove that the preceding dieomorphism is symplectic, it suces to note that by the
commutative diagram
C
N
f
T

(N/T)

N N/T
the Liouville forms of T

M and T

(N/T) are related by

M
[
C
N
= f

N/T
.
2
Example 5.17 Suppose that T is a foliation of a manifold N. By Lemma 5.15, it follows
that the annihilator C
N
T

N of T is a coisotropic submanifold which, by denition,

contains the conormal bundle F

to any leaf F of T. Since F

is a lagrangian submanifold
of T

N, Lemma 3.6 implies that F

is foliated by isotropic leaves of C

N
. Moreover, the
proof of Lemma 5.16 shows that each such leaf projects dieomorphically onto F, and so the
bundle F

is equipped in this way with a at Ehresmann connection (see Ehresmann [22])

known as the Bott connection associated to the foliation T.

Example 5.18 As a special case of Example 5.17, suppose that (B, p

B
, ) is a Morse family
over a manifold M, and let C
B
T

B be the conormal bundle to the bers of the submersion

p
B
: B V . If L T

B is the lagrangian submanifold given by the image of the dierential

d, then the nondegeneracy condition on implies that L intersects C
B
transversally along a
submanifold I which maps dieomorphically onto the ber critical set

under the natural

projection : T

B B. The lagrangian immersion

M equals the composition

1
L C
B
p
C
B
C
V
/(

.
In other words, Theorem 4.15 can be reformulated as an application of Theorem 5.12 when
the lagrangian submanifold im(d) is transverse to E, i.e., when is nondegenerate.
71

We remark that since C

N
is a bundle over N, the intersection TC
N
V M is a (constant
rank) subbundle of TC
N
. A check of the denitions shows that this subbundle is mapped
onto V (N/T) under the quotient map C
N
T

(N/T) described above.

Example 5.19 An important special case of Examples 5.6 and 5.18 arises when (B, V, p
V
, )
is a Morse family which generates a lagrangian submanifold in L T

M and x B is a
(nondegenerate) critical point of , so that p =

(x) lies in the zero section Z

M
of T

M.
Denoting by x
t
the point in the zero section of T

B lying over x, we recall that T

x
(T

B)
admits the natural lagrangian splitting T
x
B T

x
B. Let let E T
x
B be the kernel of the
linearized projection p
B
: TB TM.
Now suppose that Q is the nondegenerate quadratic form on T
x
B dened by the hessian
of . By denition, the restriction of Q to E equals the form induced by the ber-hessian
H
x
of at x, and so
ind(Q[
E
) = ind H
x
.
On the other hand, it is easy to check that the tangent space to the conormal submanifold
C
V
dened in Example 5.18 at x equals the coisotropic subspace C = T
x
B E

, so that
the last formula in Example 5.6 implies that the Morse index
M
(x, ) of the critical point
x satises

M
(x, ) = ind H
x
+ ind(V
p
M, T
p
L; T
p
Z
M
).

Reduction and the Maslov bundle

Using linear reduction we can give a useful alternative description of the Maslov bundle of
an immersed lagrangian submanifold L in T

M. Given two lagrangian subspaces L, L

t
of a
symplectic vector space V , we denote by L
L,L
the subset of the lagrangian grassmannian
L(V ) comprised of those lagrangian subspaces which are transverse to both L and L
t
. For
W, W
t
L
L,L
, the cross index of the quadruple (L, L
t
; W, W
t
) is dened as the integer
(L, L
t
; W, W
t
) = ind(L, L
t
; W) ind(L, L
t
; W
t
).
Immediately from this denition and Example 5.5, we obtain the following cocycle property
of the cross index.
Lemma 5.20 If L, L
t
are lagrangian subspaces of V and W, W
t
, W
tt
L
L,L
, then
(L, L
t
; W, W
t
) + (L, L
t
; W
t
, W
tt
) + (L, L
t
; W
tt
, W) = 0.
2
72
We denote by T
L,L
(V ) the space of functions f : L
L,L
Z such that
f(W) f(W
t
) = (L, L
t
; W, W
t
)
for all W, W
t
L
L,L
. Since any such function is determined up to an additive constant by
its value at a single point of L
L,L
, it follows that Z acts simply and transitively by addition
on T
L,L
(V ).
Now we may assign a principal Z bundle to a triple (E, ,
t
), where E is a symplectic
vector bundle and ,
t
are lagrangian subbundles. Associated to E is its lagrangian grass-
mannian bundle L(E) whose ber over x M is simply L(E
x
). The lagrangian subbundles
,
t
then correspond to smooth sections of L(E), and we denote by M
,
(E) the principal
Z bundle whose ber over x M equals T
x,

x
(E
x
).
A special case of this set-up occurs when L is an immersed lagrangian submanifold of
T

M so that E =

T(T

M) is a symplectic vector bundle over L. Natural lagrangian

subbundles of E are then given by =

TL and
t
=

V M.
Theorem 5.21 In the notation above, the Maslov bundle M
L,
is canonically isomorphic to
M
,
(E).
Proof. Let (B, p
B
, ) be a Morse family over a manifold M, x a point x in the ber
critical set

, and set p =

(x). From Example 4.32, we recall that if S : M R is a

smooth function such that the image of dS intersects

) transversely at p, then x is a
nondegenerate critical point of

S, where

S = S p
B
. From Example 5.19 it follows that
the Morse index
M
(x,

S) at x is given by

M
(x,

S) = ind H
x
+ ind(V
p
M, T
p
L
S
; T
p
Z
M
),
where p
t
= p dS(p
V
(x)) and L

S
is the lagrangian submanifold generated by

S.
Fiberwise translation by dS does not change the index on the right, and moreover it maps
T
p
L
S
to T
p
L

, the subspace V
p
M to the vertical subspace V
p
M at p, and T
p
Z
M
to the
tangent space Y of the image of dS at p. Thus, the preceding equation and Example 5.5
imply

M
(x, S) = ind H
x
+ ind(T
p
L

, V
p
M; Y ).
Now if Y is any lagrangian subspace of T
p
(T

M) transverse to V
p
M, then there exists a
function S: V R such that Y = imdS(p
V
(x)). A function f

: L
TpL

,VpM
Z is obtained
by setting
f

(Y ) =
M
(x, S
Y
).
From the discussion above, it follows that this function satises
f

(Y ) f

(Y
t
) = (T
p
L

, V
p
M; Y, Y
t
).
Using these remarks, we dene a map LM(L, ) Z M
,
(E) by sending the triple
(p, (B, p
B
, ), n) to the function (p, f

). From the preceding paragraph and the denition of

M
,
(E), it follows that this map passes to an isomorphism of principal Z bundles M
L,

M
,
(E).
73
2
We refer to [3] and [31] for a proof that the Maslov bundle described above is equivalent
to the Maslov bundle constructed using index functions in Chapter 4. For the remainder
of this section, our goal is to establish the following property of the bundle M
,
(E) under
symplectic reduction. The proofs are rather technical and can be passed over in a rst
reading.
Theorem 5.22 Suppose that E is a symplectic vector bundle over M with lagrangian sub-
bundles ,
t
and a coisotropic subbundle . If and
t
have constant rank, then
M
,
(E) is canonically isomorphic to M
,

(/

).
The starting point of the proof is the following special case of the theorem.
Lemma 5.23 Suppose that E is a symplectic vector bundle with lagrangian subbundles ,
t
and a coisotropic subbundle . If ,
t
, then M
,
(E) is canonically isomorphic to
M
,

(/

).
Proof. First consider a symplectic vector space V with lagrangian subspaces L, L
t
and a
coisotropic subspace C. If L, L
t
C, then there exists a natural linear symplectomorphism
of reduced spaces
(L
C
+ L
t
C
)/(L
C
+ L
t
C
)

(L + L
t
)/(L + L
t
)

which maps L
C
/(L
C
+ L
t
C
)

onto L/(L + L
t
)

and L
t
C
/(L
C
+ L
t
C
)

onto L
t
/(L + L
t
)

.
Moreover, if W L
L,L
, then L
C
is transverse to L
C
, L
t
C
and W
C
/(L
C
+ L
t
C
)

maps onto
W/(L + L
t
)

under the symplectomorphism above. Thus,

ind(L, L
t
; W) = ind(L
C
, L
t
C
; W
C
).
Using this remark, we can dene the isomorphism M
,
(E) M
,

(/

) berwise
by sending (x, f) M
,
(E) to (x, f

), where
f

(W
x
) = f(W)
for each W L
x,

x
.
2
Lemma 5.24 Suppose that E is a symplectic vector bundle over Y . If ,
t
are lagrangian
subbundles such that
t
has constant rank, then M
,
(E) is canonically isomorphic to
Y Z.
Proof. If ,
t
are berwise transverse, then there exists a vector bundle symplectomorphism
E

which maps onto 0 and

t
onto 0

. Any choice of a positive-denite

quadratic form on is induced by a section T of Hom(,

) comprised of self-adjoint maps.

74
The lagrangian subbundle
tt
of E which maps to the subbundle of

dened by the
graph of T is then transverse to ,
t
and clearly satises
ind(
x
,
t
x
;
tt
x
) = 0
for all x Y . A natural section s(x) = (x, f) of M
,
(E) is therefore given by
f(
tt
x
) = 0.
In general, our hypotheses imply that = +
t
is a coisotropic subbundle of E which
contains the lagrangian subbundles and
t
. By Lemma 5.23, the bundle M
,
(E) is
canonically isomorphic to M
,

(/

). Since

,
t

are transverse lagrangian subbundles

of /

, the latter bundle has a canonical trivialization, and the assertion follows.
2
Proof of Theorem 5.22. An application of the cocycle identity of Lemma 5.20 shows
that if ,
t
,
tt
are lagrangian subbundles of a symplectic vector bundle E, then M
,
(E) is
canonically isomorphic to the principal-bundle product (see Appendix D for the denition)
M
,
(E) M

,
(E). Applying this remark, we obtain the following canonical isomor-
phisms:
M
,

(E) M
,
(E) M

(E)
M
,

(E)

(E) M

(E)

.
From the assumption that has constant rank, it follows that

has constant rank, and

thus M
,
(E) is canonically trivial by Lemma 5.24. Similarly, M

(E) is canonically trivial.

Combined with the canonical isomorphisms above, this shows that M
,

(E) is canonically
isomorphic to M

(E). Since the lagrangian subbundles

of E are contained in the

coisotropic subbundle , Lemma 5.23 implies that M

(E) is in turn canonically isomorphic

to M
,

(/

), completing the proof.

2
Example 5.25 Suppose that E, E
t
are trivial symplectic vector bundles over M with la-
grangian subbundles E and
t
E
t
, and let T be a section of the bundle Hom

(E, E
t
)
of symplectic vector bundle maps. The graph of the section T denes a lagrangian subbundle

(T)
of E
t
E, while the restriction of T to the subbundle denes a lagrangian subbundle
T() of E
t
.
It is easy to check that the lagrangian subbundles
(T)
and
t
intersect the coisotropic
subbundle E
t
along constant rank subbundles, whereas their reductions by E
t
are
given by T() and
t
, respectively. Thus, Theorem 5.22 implies that M
L
(T)
,L

L
(E
t
E)
and M
T(L),L
(E
t
) are canonically isomorphic.

75
Example 5.26 Any symplectomorphism F : T

M T

N induces a lagrangian embedding

F
: T

M T

(M N), dened as the composition of the graph

F
: T

M T

N T

M
of F with the Schwartz transform S
M,N
: T

N T

M T

(M N). A second lagrangian

embedding z
F
: M T

N is dened by the composition of F with the zero section s

0
: M
T

M. Observe that these denitions imply that the Liouville class

M,z
F
of the embedding
(M, z
F
) identies with
T

M,
F
under the natural isomorphism H
1
(M; R) H
1
(T

M; R)
induced by the projection T

M M.
To see that a similar relation holds for the Maslov classes of (M, z
F
) and (T

M,
F
),
consider the symplectic vector bundles E = s

0
T(T

M) and E
t
= z

F
T(T

N) over M, along
with the section T of Hom

(E, E
t
) dened by the restriction of F

to T
Z
(T

M). Moreover, we
have the lagrangian subbundles = s

0
V M of E and
t
= z

F
V N of E
t
, and the lagrangian
subbundles
(T)
E
t
E and T() E
t
, as described in Example 5.25. Theorem 5.21
shows that s

0
M
T

M,
F
is canonically isomorphic to M

T
,

(E
t
E) and that M
M,z
F
is
canonically isomorphic to M
T(),
(E
t
). Thus, it follows from Example 5.25 that the Maslov
classes are also related by pull-back: s

M,
F
=
M,z
F
.

Example 5.27 We can apply the conclusion of the preceding example to ber-preserving
symplectomorphisms f : T

M T

N. From Section 3.2, we recall that any such f is equal

to a berwise translation of T

M by a closed 1-form on M composed with the cotangent

lift of some dieomorphism N M. Evidently, the phase class of the induced lagrangian
embedding (M, z
f
) equals []

H
1
(M; R), and thus R
+
is admissible for the lagrangian
embeddings (M, z
f
), (T

M,
f
) if and only if [] is -integral.

5.2 The symplectic category

To systematize the geometric aspects of quantization in arbitrary symplectic manifolds, we
now introduce the symplectic category S. As objects of S we take the class of all smooth,
nite-dimensional symplectic manifolds. Given two objects (P, ) (Q,
t
) of S, we dene
their product as the symplectic manifold (P Q,

1
+

t
), where
1
,
2
denote the
cartesian projections. The symplectic dual of an object (P, ) is the object (P, ).
From Lemma 3.14, we recall that a smooth dieomorphism from a symplectic manifold
P to a symplectic manfold Q is a symplectomorphism if and only if its graph is a lagrangian
submanifold of Q P. More generally, an immersed lagrangian submanifold of Q P is
called a canonical relation from P to Q. The morphism set Hom(P, Q) is then dened
to consist of all canonical relations in Q P. Since immersed lagrangian submanifolds of
the product QP coincide with those of its dual, we can therefore dene the adjoint of a
canonical relation L Hom(P, Q) as the element L

Hom(Q, P) represented by the same

equivalence class L of immersions into the product P Q.
Composition of morphisms is unfortunately not dened for all L
1
Hom(P, Q) and
L
2
Hom(Q, R), and so S is therefore not a true category. Nevertheless, we can describe
76
a sucient condition for the composability of two canonical relations as follows. By Exam-
ple 5.8, the product R
Q
P is a reducible coisotropic submanifold of R QQP,
where
Q
denotes the diagonal in QQ. The product L
2
L
1
of the canonical relations L
1
and L
2
is a well-dened immersed lagrangian submanifold of R Q Q P, and we will
call L
2
L
1
clean if (R
Q
P, L
2
L
1
) is a reducible pair. Applying Theorem 5.12, we
then obtain
Proposition 5.28 If L
2
L
1
is clean, then L
2
L
1
is an immersed lagrangian submanifold
of R P, i.e. L
2
L
1
Hom(P, R).
2
Associativity of compositions holds in the symplectic category in the sense that for canonical
relations L
1
, L
2
, L
3
, the equation
L
1
(L
2
L
3
) = (L
1
L
2
) L
3
is valid provided that both sides are dened.
Among the members of S, there is a minimal object Z, the zero-dimensional symplectic
manifold consisting of a single point equipped with the null symplectic structure. Mor-
phisms from Z to any other object P S identify naturally with immersed lagrangian
submanifolds of P. Thus, the elements of P are its immersed lagrangian submanifolds,
and in particular, we may identify the set Hom(P, Q) of morphisms with the elements of
QP for any P, Q S.
8
A canonical relation L Hom(P, Q) is said to be a monomorphism if the projection
of L onto P is surjective and the projection of L onto Q is injective, in the usual sense.
Equivalently, L is a monomorphism if L

L = id
P
. Dually, one denes an epimorphism
in the symplectic category as a canonical relation L for which L L

= id
P
. From these
denitions we see that a canonical relation is an isomorphism if and only if it is the graph
of a symplectomorphism.
Canonical lifts of relations
The full subcategory of S consisting of cotangent bundles possesses some special properties
due to the Schwartz transform S
M,N
: T

N T

M T

(M N), which enables us to

identify canonical relations in Hom(T

M, T

N) with immersed lagrangian submanifolds in

T

(M N).
Example 5.29 A smooth relation between two manifolds M, N is a smooth submanifold S
of the product M N. Under the Schwartz transform, the conormal bundle of S identies
with a canonical relation L
S
Hom(T

N, T

M) called the cotangent lift of S. In view of

Example 3.27, this denition generalizes tangent lifts of dieomorphisms. Note in particular
8
Another point of view is to dene Hom(P, Q), not as a set, but rather as the object Q P in the
symplectic category; then the composition Hom(P, Q) Hom(Q, R) Hom(P, R) is the canonical relation
which is the product of three diagonals.
77
that although a smooth map f : M N does not in general give rise to a well-dened
transformation T

N T

M, its graph in MN nevertheless generates a canonical relation

L
f
Hom(T

N, T

M).
Example 5.30 The diagonal embedding : M M M induces a canonical relation
L

Hom(T

(M M), T

M) which in local coordinates assumes the form

L

= ((x, ), (x, x, , )) : x M and , , T

x
M satisfy = + .
In other words, L

is the graph of addition in the cotangent bundle. If L

1
, L
2
Hom(, T

M)
and L
2
L
1
is identied with an element of Hom(, T

(M M)) by the usual symplecto-

morphism T

M T

M T

(M M), then the sum L

1
+ L
2
Hom(, T

M) is dened
as
L
1
+ L
2
= L

(L
2
L
1
).
Note that if L
1
, L
2
coincide with the images of closed 1-forms
1
,
2
on M, then L
1
+ L
2
equals the image of
1
+
2
.
The dual of the isomorphism T

M T

M T

(M M) identies L

with an element
L
t

Hom(T

M, Hom(T

M, T

M)). This canonical relation satises L

1
+L
2
= L
t

(L
1
)(L
2
),
and if L is the image of a closed 1-form on M, then
L
t

(L) = f

,
where f

is the berwise translation mapping introduced in Section 3.2.

Example 5.31 (The Legendre transform) As a particular example of this situation, let
V be a smooth manifold, and consider the ber product TV
V
T

V along with its natural

inclusion : TV
V
T

V TV T

V and evaluation function ev : TV

V
T

V R given
by
ev((x, v), (x, p)) = 'p, v`.
If L T

(TV
V
T

V ) is the lagrangian submanifold given by the image of the dierential

d(ev), then the push-forward L
leg
=

L denes an isomorphism in Hom(T

(TM), T

(T

M))
given in local coordinates by
((x, v), (, )) ((x, ), (, v)).
As noted in [57], this canonical relation can be viewed as a geometric representative of the
Legendre transform in the following sense: If L: TM R is a hyperregular Lagrangian
function, in the sense that its ber-derivative denes a dieomorphism TM T

M, then
the composition of L
leg
with the image of dL equals the image of dH, where H: T

M R
is the classical Legendre transform of L.

78
Morphisms associated with coisotropic submanifolds
If C P is a reducible coisotropic submanifold, then the reduction relation R
C

Hom(P, C/(

) is dened as the composition of the quotient map C C/(

and the inclu-

sion C/(

P. Somewhat more concretely, the relation R

C
is the subset
([x], x) : x C
of C/(

P, where [x] is the leaf of the characteristic foliation passing through x C.

Evidently R
C
is an epimorphism, and so R
C
R

C
= id
P
. On the other hand, R
C
is not a
monomorphism unless C = P, and we dene the projection relation K
C
Hom(P, P) as
the composition R

C
R
C
, i.e.
K
C
= (x, y) : x, y C, [x] = [y].
By associativity, we have K
C
= K
C
K
C
, and K

C
= K
C
. Thus, K
C
is like an orthogonal
projection operator.
These relations give a simple interpretation of Theorem 5.12 in the symplectic category.
If L is an immersed lagrangian submanifold of P, i.e. L Hom(Z, P), then L is composable
with both R
C
and K
C
, provided that it intersects C cleanly. In this case, we have
L
C
= R
C
L L
C
= K
C
L.
In particular, K
C
xes any lagrangian submanifold of C.
Example 5.32 Suppose that T is a foliation on a manifold B whose leaf space B
T
is smooth.
In the notation of Example 5.17, we then nd that if T

(B/T) is canonically identied with

the reduced space E/(

, then the reduction relation R

E
is the Schwartz transform of the
conormal bundle to the graph of the projection B B/T.

5.3 Symplectic manifolds and mechanics

In general, an arbitrary symplectic manifold has no associated conguration space, and
therefore the classical and quantum mechanical viewpoints must be adapted to this new
context based on the available structure.
The classical picture
The central objects in the classical picture of mechanics on an arbitrary symplectic manifold
(P, ) are the semi-classical states, represented as before by lagrangian submanifolds of P
equipped with half-densities, and the vector space of observables C

(P).
With respect to pointwise multiplication, C

(P) forms a commutative associative alge-

bra. Additionally, the symplectic form on P induces a Lie algebra structure on C

(P) given
by the Poisson bracket
f, g = X
g
f,
79
where X
g
=
1
(dg) denotes the hamiltonian vector eld associated to g. These structures
satisfy the compatibility condition
fh, g = f h, g +f, g h,
and are referred to collectively by calling C

(P) the Poisson algebra of P.

The classical system evolves along the trajectories of the vector eld X
H
associated to
the choice of a hamiltonian H : P R. If f is any observable, then Hamiltons equations
assume the form

f = H, f.
Note that in local Darboux coordinates, the Poisson bracket is given by
f, g =

f
q
j

g
p
j

f
p
j

g
q
j

.
Setting f = q
j
or f = p
j
in the Poisson bracket form of Hamiltons equations above yields
their familiar form, as in Section 3.3.
Two functions f, g C

(P) are said to be in involution if f, g = 0, in which case the

hamiltonian ows of f and g commute. An observable in involution with the hamiltonian
H is called a rst integral or constant of the motion of the system. A collection f
i
of functions in involution on P is said to be complete if the vanishing of f
i
, g for all i
implies that g is a function of the form g(x) = h(f
1
(x), , f
n
(x)).
The quantum mechanical picture
Quantum mechanical observables should be the vector space /(H
P
) of self-adjoint linear op-
erators on some complex Hilbert space H
P
associated to P. The structure of a commutative,
non-associative algebra is dened on / by Jordan multiplication:
A B =
1
2
(AB + BA),
representing the quantum analog of pointwise multiplication in the Poisson algebra of P.
Similarly, the -dependent commutator
[A, B]

=
i

(AB BA)
denes a Lie algebra structure on / analogous to the Poisson bracket on C

(P).
Quantum mechanical states are vectors in H
P
. The time-evolution of the quantum system
is determined by a choice of energy operator

H, which acts on states via the Schrodinger
equation:

=
i

H.
A collection A
j
of quantum observables is said to be complete if any operator B which
commutes with each A
j
is a multiple of the identity. This condition is equivalent to the
irreducibility of A
j
, in other words, no nontrivial subspace of H
P
is invariant with respect
to each j.
80
Quantization
Although neither an underlying conguration manifold nor its intrinsic Hilbert space is
generally available in the context of arbitrary symplectic manifolds, the basic goal of quan-
tization remains the same as in the case of cotangent bundles: starting with a symplectic
manifold P, we wish to identify a -algebra / of operators which give the quantum analog
of the system at hand. Early approaches to this problem were based on the principle
that, regardless of how the algebra / is identied, the correspondence between classical and
quantum observables should be described by a linear map from the Poisson algebra C

(P)
to / which satises the following criteria known as the Dirac axioms
Denition 5.33 A linear map : C

(P) S(H
P
) is called a quantization provided that
it satises
1. (1) = identity.
2. (f, g) = [(f), (g)]

3. For some complete set of functions f

1
, , f
n
in involution, the operators
(f
1
), , (f
n
) form a complete commuting set.
As was eventually proven by Groenwald and Van Hove (see [1] for a proof), a quantization
of all classical observables in this sense does not exist in general.
To make the basic quantization problem more tractable, we rst enlarge the class of
classical objects to be quantized, but then relax the criteria by which quantum and classical
objects are to correspond. Based on an idea of Weyl and von Neumann, we rst replace
the classical observables by the groups of which they are the innitesimal generators. The
basic classical objects are then symplectomorphisms, which should be represented by unitary
operators on quantum Hilbert spaces.
A general formulation of the quantization problem is then to dene a functor from
the symplectic category to the category of (hermitian) linear spaces. This means that to
each symplectic manifold P we should try to assign a Hilbert space H
P
in such a way that
H
P
is dual to H
P
, and H
PQ
is canonically isomorphic to a (completed) tensor product
H
P

H
Q
. However this is accomplished, each canonical relation L Hom(P, Q) must then
be assigned to a linear operator T
L
Hom(H
P
, H
Q
) H
Q
H

P
in a way which commutes
with compositions, i.e.,
T

LL
= T

L
T
L
for each

L Hom(Q, R). In these abstract terms, our classical and quantum correspondences
can be expressed as follows.
81
Object Classical version Quantum version
basic spaces symplectic manifold (P, ) hermitian vector space H
P
QP H
Q

H
P
QP Hom(H
P
, H
Q
)
point C
state lagrangian submanifold L element of H
P
space of observables Poisson algebra C

(P) symmetric operators on H

P
At this stage, two relevant observations are apparent from our earlier study of WKB
quantization. First, in addition to just a lagrangian submanifold L P, it may require
more data (such as a symbol) to determine an element of H
P
in a consistent way. Second,
the Hilbert space H
P
may carry some sort of ltration (e.g. by powers of or by degree
of smoothness), and the quantization may be correct only to within a certain degree of
accuracy as measured by the ltration. By correctness we mean that composition of
canonical relations should correspond to composition of operators. As we have already seen,
it is too much to require that this condition be satised exactly. The best we can hope for
is a functorial relation, rather than a mapping, from the classical to the quantum category.
82
6 Fourier Integral Operators
For our rst examples of a quantization theory which gives a functorial relation between
the classical (symplectic) category and the quantum category of hermitian vector spaces, we
return to WKB quantization in cotangent bundles. In this context, semi-classical states are
represented by pairs (L, s) consisting of an immersed lagrangian submanifold L in T

(MN)
equipped with a symbol s. Our discussion begins by dening a suitable notion of composition
for such states. The WKB quantization I

(L, s) of the state (L, s) is then regarded as the

Schwartz kernel of the corresponding operator in Hom(H
M
, H
N
). A particular concrete case
of this classical-quantum correspondence is given by the symbol calculus of Fourier integral
operators.
6.1 Compositions of semi-classical states
In Section 4.4, we dened a semi-classical state in a cotangent bundle T

M as a pair (L, s)
consisting of a quantizable lagrangian submanifold L T

M and a symbol s on L. In this

section, we study certain natural transformations of semi-classical states.
Reduction of semi-classical states
We will say that a reducible pair (C, L) in a symplectic manifold P is properly reducible
if the quotient of I = L
P
C by its characteristic foliation is a smooth, Hausdor manifold
and the map I L
C
is proper.
Lemma 6.1 If (C, L) is a properly reducible pair in a symplectic manifold P, then there
exists a natural linear map
[[
1/2
L [[
1/2
C [[
1/2
L
C
.
Proof. First note that if V is a symplectic vector space, together with a lagrangian subspace
L and a coisotropic subspace C, then the exact sequence
0 L C

L C L
C
0
gives rise to the isomorphism
[[
1/2
L
C
[[
1/2
(L C

) [[
1/2
(L C).
The linear maps v (v, v) and (x, y) x + y dene a second exact sequence
0 L C L C L + C 0,
from which we get
[[
1/2
L
C
[[
1/2
(L C

) [[
1/2
(L + C) [[
1/2
L [[
1/2
C.
83
Finally, the exact sequence
0 (L + C)

V

(L + C)

0
combined with the half-density on V induced by the symplectic form denes a natural
isomorphism
[[
1/2
(L + C) [[
1/2
(L + C)

.
Since L C

= (L + C)

, we arrive at
[[
1/2
L
C
[[(L C

) [[
1/2
L [[
1/2
C.
Now consider a properly reducible pair (C, L) in a symplectic manifold P. By the pre-
ceding computations, there is a linear map
[[
1/2
L [[
1/2
C [[
1/2
L
C
[[(T
I
).
Since the quotient map I L
C
is proper, integration over its bers is well-dened and gives
the desired linear map
[[
1/2
L [[
1/2
C [[
1/2
L
C
.
2
Let M be a smooth manifold and consider a submanifold N M equipped with a
foliation T such that the leaf space N
T
is a smooth, Hausdor manifold. From Section 5.1
we recall that the integrability of T implies that
C
N
= (x, p) T

M : x N, T
x
ker(p)
is a coisotropic submanifold of T

M whose reduced space is the cotangent bundle T

N
T
of
the leaf space N
T
. If L is an immersed lagrangian submanifold of T

M such that (C
N
, L)
form a reducible pair, then we denote by I the ber product L
T

M
C
N
of L and C
N
and
consider the following commutative diagram
L T

M
r
L

I
r
C
N
C
N

p
C
L
C

N
Our goal is describe how a symbol s on L naturally induces a symbol s
C
on L
C
.
Lemma 6.2 In the notation of the diagram above, there is a natural isomorphism
r

L,

L
C
,
.
84
Proof. From the proof of Lemma 5.16 it follows easily that the pull-back to I of the Liouville
forms on T

M and T

N coincide, and thus, the pull-back to I of the prequantum line bundles

over T

M and T

N are canonically isomorphic.

Similarly, let E be the symplectic vector bundle over I given by the pull-back of T(T

M).
Lagrangian subbundles ,
t
of E are then induced by the immersed lagrangian submanifold
L and the vertical subbundle V M of T(T

M). By denition, the pull-back of the Maslov

bundle of L to I is canonically isomorphic to the bundle M
,
(E).
Now the tangent bundle of C
N
induces a coisotropic subbundle C of E, and one must
check that the pull-back of the Maslov bundle of L
C
to I is canonically isomorphic to
M

C
,

C
(C/C

). From Theorem 5.22 we therefore obtain a canonical isomorphism of r

L
M
L
with

M
L
C
. By tensoring this isomorphism with the isomorphism of prequantum bundles
described in the preceding paragraph, we arrive at the desired isomorphism of phase bundles.
2
A parallel section s of
L,
pulls back to a parallel section of r

L
C
,
, which, under the
isomorphism of Lemma 6.2, identies with a parallel section of

L
C
,
. Since parallel
sections of

L
C
,
identify naturally with parallel sections of
L
C
,
, we obtain a map
(parallel sections of
L,
) (parallel sections of
L
C
,
).
By tensoring with the map of density spaces given above, we obtain, for each half-density
on C, a natural map of symbol spaces
S
L
S
L
C
.
We will denote the image of a symbol s on L under this map by s
C,
.
Composition of semi-classical states
If L
1
, L
2
are immersed lagrangian submanifolds of T

M, T

N, then their product L

2
L
1
gives
a well-dened immersed lagrangian submanifold of T

(M N) via the Schwartz transform

S
M,N
: T

NT

M T

(MN). By the fundamental properties of the Schwartz transform

described in Proposition 3.32, it follows easily that the phase bundle
L
2
L
1
,
is canonically
isomorphic to the external tensor product
L
2
,

1
L
1
,
. From the discussion in Appendix A
we also have a linear isomorphism of density bundles [[
1/2
(L
2
L
1
) [[
1/2
L
2
[[
1/2
L
1
,
and thus there is a natural linear map of symbol spaces
o
L
2
o
L
1
o
L
2
L
1
which we denote by (s
1
, s
2
) s
2
s

1
.
Denition 6.3 The product of semi-classical states (L
1
, s
1
), (L
2
, s
2
) in T

M, T

N is the
semi-classical state (L
2
L
1
, s
2
s

1
) in T

(M N).
2
85
Turning to compositions, we rst note that if M, V are smooth manifolds, then the
Schwartz transform S
M,V
: T

V T

M T

(M V ) identies each canonical relation in

Hom(T

M, T

V ) with an immersed lagrangian submanifold in T

(M V ). Our goal is now

to describe how semi-classical states (L, s) in T

(M N) and (L
t
, s
t
) in T

(N V ) dene
a semi-classical state in T

(M V ) when L, L
t
are composable as canonical relations in
Hom(T

M, T

N) and Hom(T

N, T

V ), respectively.
We begin with the following commutative diagram
T

V T

N T

N T

M T

(M N N V )

V T

M T

(M V )
where the upper horizontal arrow denotes the composition S
MN
(S
N,V
S
M,N
) of Schwartz
tranforms, and the lower horizontal arrow equals the Schwartz transform S
M,V
. The left
vertical arrow represents the reduction relation dened by the coisotropic submanifold C =
T

V
T

N
T

M, while the right vertical arrow is the reduction relation dened by

the image of C under the Schwartz transform. L
t
L Hom(T

M, T

V ) is dened as the
reduction of L
2
L
1
by the coisotropic submanifold T

V
T

N
T

M of T

V T

N
T

N T

M. Under the Schwartz transform,

T

V T

N T

N T

M T

(M N N V ),
the submanifold T

V
T

N
T

M maps to the conormal submanifold C

V,N,M
of T

(V
NNM) dened by V
N
M V NNM and its product foliation by subsets of
the form v
N
m for (v, m) V M. Thus, the Schwartz transform of L
2
L
1
reduces by
C
V,N,M
to yield the image of L
2
L
1
under the Schwartz transform T

V T

M T

(V M).
From Lemma 6.1, it follows that for any properly reducible pair L
1
, L
2
, we obtain a
natural linear map of symbol spaces
o
L
2
o
L
1
o
L
2
L
1
given by the composition of the product map above and the reduction map o
L
2
L
1
o
L
2
L
1
dened by the natural half-density on C
V,N,M
induced by the symplectic forms on T

V, T

N,
and T

M. We denote the image of s

2
s
1
under this map by s
2
s
1
.
We will say that semi-classical states (L
1
, s
1
) in T

(M N) and (L
2
, s
2
) in T

(N V )
are composable if the conormal submanifold C
V,N,M
and immersed lagrangian submanifold
L
2
L
1
form a properly reducible pair.
Denition 6.4 If a semi-classical state (L
1
, s
1
) in T

(M N) is composable with a semi-

classical state (L
2
, s
2
) in T

(N V ), then their composition is dened as the semi-classical

state (L
2
L
1
, s
2
s
1
) in T

(M V ).
2
86
Example 6.5 Consider a semi-classical state (L
F
, s) in T

(MN), where L
F
is the Schwartz
transform of the graph of a symplectomorphism F : T

M T

N. If L is an immersed la-
grangian submanifold of T

M, then Denition 6.4 provides that for each symbol s o

L
,
the semi-classical state (L, s) in T

M is transformed by F into the semi-classical state

(L
F
L, s s) in T

N.
To make this correspondence more explicit, let us choose a particular lagrangian immer-
sion : L T

M representing L (see the discussion in Section 4.4). We then note that the
natural half-density [
n
M
[
1/2
on L
F
T

M enables us to identify the symbol space o

L
F
with
the product

L
F

C
C

(L
F
, C).
Given a parallel section s and smooth complex-valued function h on L
F
, we nd by a com-
putation that the isomorphism o
L
o
L
F
L
determined by the symbol s = s h is given by
s a s
t
(a

h), where a is any half-density on L and s

t
is the unique parallel section of

L
F
L,
such that

s s
t
s
1
under the canonical isomorphism

L
F
,

L
F
L,

1
L,
.

Example 6.6 As a special case of Example 6.5, note that if the symplectomorphism F :
T

M T

N is the cotangent lift of a dieomorphism M N, then for each > 0, the

phase bundle of of L
F
admits a canonical parallel section, and so the symbol space o
L
F
identies naturally with R
+
C

(L
F
, C). If h C

(L
F
, C), then with this identication,
the composition of (L
F
, h) with a semi-classical state (L, s) equals the semi-classical state
(L
F
L, (h ) s).
Similarly, if is a closed 1-form on M and the symplectomorphism F : T

M T

M
equals berwise translation by , then R
+
is admissible for L
F
if and only if [] is
-integral. For such , each parallel section of the phase bundle
L
F
,
identies with an
oscillatory function of the form ce
iS/
for c R and some S : M T

satisfying S

d =
. A computation then shows that the composition of a semi-classical state of the form
(L
F
, e
iS/
h) with (L, , s) yields the semi-classical state (L
F
L, e
i(S
L
)/
(h ) s).

6.2 WKB quantization and compositions

To dene the composition of semi-classical states (L, s) Hom(T

M, T

N) and (

L, s)
Hom(T

N, T

V ) as a semi-classical state (

L L, s s) Hom(T

M, T

V ), we used the
Schwartz transform to identify the immersed lagrangian submanifold

L L T

V T

M
with an immersed lagrangian submanifold in T

(M V ), since it is the cotangent bundle

structure of the latter space which gives meaning to symbols (in the sense of Chapter 4)
on

L L. On the quantum level, an analogous correspondence is furnished by the Schwartz
kernel theorem.
Let M be a smooth manifold and let [[
1/2
0
M be equipped with the topology of C

convergence on compact sets. A distributional half-density on M is then a continuous,

C-linear functional on [[
1/2
0
M. We denote the space of distributional half-densities on M
87
by [[
1/2

M, and we equip this space with the weak

topology. If N is another smooth

manifold, a kernel is any element of [[
1/2

(M N). A kernel K denes a linear map

/: [[
1/2
0
M [[
1/2

N by duality via the equation

'/(u), v`
def
= 'K, u v` ()
Schwartz kernel theorem . Every K [[
1/2

(MN) denes a linear map /: [[

1/2
0
M
[[
1/2

N by () above, which is continuous in the sense that /(

j
) 0 in [[
1/2

N if
j
0
in [[
1/2
0
M. Conversely, to every such linear map, there is precisely one distribution K such
that () is valid.
2
Of course, each w [[
1/2
N denes an element w [[
1/2

N by the equation
' w, v` =

N
v w.
In all of the examples we will consider, the image of the map / will lie within the subspace
of [[
1/2

N represented by elements of [[
1/2
N in this way, and so we can consider / as
a continuous linear map [[
1/2
0
M [[
1/2
0
N, thereby giving a (densely dened) operator
H
M
H
N
on the intrinsic Hilbert spaces of M, N. In practical terms, the value of the half-
density Ku at each y N can be computed in these cases by, roughly speaking, integrating
the product K u over the submanifold y M of N M. Thus, in the same way that
the Schwartz transform provides a natural correspondence
Hom(T

M, T

N) T

(M N),
the Schwartz kernel theorem gives the identication
[[
1/2

(M N) Hom(H
M
, H
N
)
-dierential operators
With respect to linear coordinates x
j
on R
n
, we dene -dependent operators
D
j
def
= i

x
j
.
An -dierential operator of asymptotic order
9
k Z is then an asymptotic series of the
form
P

m=0
P
m

m+k
,
9
Note that this order generally diers from the order of a dierential operator.
88
where each P
m
is a polynomial in the operators D
j
. By formally substituting
j
for D
j
in
each term of P

, we obtain an -dependent function

P

on T

R
n
, known as the symbol of
P

, which is related to P

by the asymptotic Fourier inversion formula: for any compactly

supported oscillatory test function e
i/
u (we permit = 0), we have
(P

e
i/
u)(x) = (2)
n

e
i(/xy,))/

(x, ) u(y) dy d.
The principal symbol p

of the operator P

is dened as the symbol of the rst nonvanishing

term in its series expansion, i.e.,
p

(x, ) =
Pm
0
(x, )
m
0
+k
if P
m
= 0 for all m < m
0
and P
m
0
= 0. By replacing
P

by p

in the formula above, we

obtain
(P

e
i/
u)(x) = (2)
n

e
i(/xy,))/
p

(x, y, ) u(y) dy d + O(
m
0
+k+1
)
where p

is of course independent of y. (Later we will drop this assumption to obtain a more

symmetric calculus). For xed x, the function (y, ) 'x y, ` has a nondegenerate
critical point when y = x and = d(x). An application of the principle of stationary phase
therefore gives
(P

e
i/
u)(x) = e
i(x))/
u(x) p

(x, x, d(x)) h
m
0
+k
+ O(
m
0
+k+1
) ()
To interpret this expression in the context of WKB quantization, we rst note that the phase
function (x, y, ) = 'x y, ` on B = R
n
R
n
(R
n
)

, together with V = R
n
R
n
and
the cartesian projection p
V
: B V dene a Morse family (B, V, p
V
, ) which generates
the conormal bundle to the diagonal R
n
R
n
. The principal symbol p

, written as a
function of the variables (x, y, ) denes an amplitude a = p

[dx dy[
1/2
[d[ on B, whose
restriction to the ber-critical set

= (x, x, ) : 'x, ` = 0 induces a well-dened symbol

s
P
= p

(x, x, ) on L

.
Now, e
i/
u = I

(L, s), where L is the projectable lagrangian submanifold of T

R
n
dened by im(d), and s is obtained from the pull-back of u to L. By Example 6.6, we have
(L, s) = (L

L, s
P
s), that is, composition with (L

, s
P
) multiplies the symbol s by the
values of the principal symbol p

on L

. Combined with the preceding equation, this

gives
(P

e
i/
u)(x) = I

(L

L, s
P
s)
m
0
+k
+ O(
m
0
+k+1
).
The Schwartz kernel for the operator P

is given by the distribution family I

(L

, s
P
).
An -dierential operator on a manifold M is an operator P

on [[
1/2
0
M which coincides
in local coordinates with a series of the form P

as above. While the symbol of P

depends
on the choice of local coordinates, its principal symbol p

is a well-dened function on the

cotangent bundle T

M. Using a global generating function for the conormal bundle of

M M (see Example 4.29) and arguing as above, we obtain the following geometric
version of () above:
89
Theorem 6.7 If (L, s) is an exact, projectable semi-classical state in T

M and P

is an
-dierential operator of order k on M, then:
P

(I

(L, s)) = I

(L

L, s
P
s) + O(
m
0
+k+1
),
where s
P
is the symbol on L

induced by the principal symbol p

of P

.
Roughly speaking, this theorem asserts that polynomial functions on the identity relation L

quantize as the Schwartz kernel of dierential operators on M. As usual, this correspondence

is only asymptotic; many dierential operators share the same principal symbol, and their
actions on a given function coincide only up to terms of higher order in .
As a particular case of Theorem 6.7, the vanishing of the principal symbol p

on L implies
that
P

(I

(L, s)) = O(
m
0
+k+1
),
i.e., that I

(L, s) is a rst-order approximate solution to the equation P

u = 0. This remark
suggests the quantum analog of certain coisotropic submanifolds of T

M. The zero set of the

principal symbol p

is called the characteristic variety of the operator P

. If 0 is a regular
value of p

, then C
P

= p
1

(0) is a coisotropic submanifold of T

M, and semi-classical states

contained in C
P

represent solutions of the asymptotic dierential equation P

u = 0. In this
sense, C
P

, or, more properly, the reduced manifold of C

P

, corresponds to the kernel of P

in H
M
, and the projection relation K
C
P

quantizes as the orthogonal projection onto this

subspace.
A familiar illustration of these concepts is provided by the WKB approximation.
Example 6.8 Recall that the Schrodinger operator associated to a given potential V on a
riemannian manifold M is given by

H =

2
2m
+ m
V
.
For E > 0, the time-independent Schrodinger equation is then
P

= 0,
where P

=

H E is the zeroth-order asymptotic dierential operator on M with principal
symbol
p

(x, ) =
[[
2
2m
+ (V (x) E).
The characteristic variety of p

is simply the level set H

1
(E) of the classical hamiltonian
of the system, and as in previous sections we see that rst-order approximate solutions to
the time-independent Schrodinger equation arise from semi-classical states represented by
quantizable lagrangian submanifolds in C
P

90
The diculty of quantizing more general symbols on the identity relation L

lies in the
convergence of the integral
(P

u)(x) = (2)
n

e
i/xy,)/
a(x, y, ) u(y) d dy.
For dierential operators, integration over the phase variables was well-dened due to the
fact that the symbol of a dierential operator has a polynomial growth rate with respect to
these variables. Weakening this condition while still guaranteeing that the integral converges
leads to the denition of pseudodierential operators.
In R
n
, an -pseudodierential operator of order is given by
(Au)(x) = (2)
n

e
i/xy,)/
a

(x, ) u(y) d dy,

where the symbol a(x, ) is an asymptotic series in of the form
a

(x, )

a
j
(x, )
+j
.
Each coecient a
j
is a smooth function on R
n
R
n
` 0 and is positively homogeneous of
degree j, i.e.,
a
j
(x, c) = c
j
a
j
(x, )
for c > 0. As in the special case of dierential operators, the invariance of pseudodierential
operators under coordinate changes enables one to extend this theory to any smooth manifold
M. In this case, the principal symbol of a pseudodierential operator is a well-dened
function on T

M, and there is a corresponding version of Theorem 6.7 (see for example

[51]).
Fourier integral operators
In its simplest form, the theory of Fourier integral operators provides a means for quantizing
half-densities on more general lagrangian submanifolds L of T

M by replacing the function

(x, y, ) 'x y, ` in the denition of pseudodierential operators by a phase function
which generates L:
(Au)(x) = (2)
n

e
i(x,y,)/
a(x, y, ) u(y) d dy.
As before, the class of quantizable half-densities (and canonical relations) is constrained
by the necessary conditions for this integral to converge. In this section, we will only be
concerned with a few specialized cases; a detailed description of this theory can be found in
[21, 28, 31, 43, 56].
Perhaps the simplest generalization of the picture of (pseudo)dierential operators given
above is provided by quantizing semi-classical states whose underlying lagrangian subman-
ifolds are conormal bundles. Let M be a smooth manifold with a smooth submanifold N.
As constructed in Example 4.29, the conormal bundle L
N
T

N is generated by a single
91
Morse family (B, V, p
V
, ) with the properties that V is a tubular neighborhood of N in M
and B is a vector bundle over V with ber dimension equal to the codimension of N in M.
As in the case of (pseudo)dierential operators, an amplitude a on B having an asymptotic
expansion in terms which are positively homogeneous with respect to the natural R
+
action
on the bers of p
V
: B V gives rise to a well-dened distributional half-density family
I

(L, s) on M.
It is easy to see that the ber-critical set

is invariant under the R

+
action on B, and
that the identication L
N

dened by

is equivariant with respect to the natural

R
+
action on L
N
. Consequently, positively homogeneous symbols on L
N
induce positively
homogeneous amplitudes on B of the same order, and we can proceed to dene I

(L, s) as
before by requiring that s be homogeneous.
When M = X Y is a product manifold, the distributions I

(L, s) represent Schwartz

kernels for continuous linear operators [[
1/2
0
X [[
1/2

Y . Under certain additional restric-

tions on N which are always fullled, for example, when N is the graph of a dieomorphism
X Y , these operators map [[
1/2
0
X to [[
1/2
0
Y and extend continuously to [[
1/2

X. More-
over, they satisfy the composition law
I

L, s) I

(L, s) = I

L L, s s).
Example 6.9 If X, Y are smooth manifolds and f : X Y is a smooth dieomorphism,
then the procedure described above can be applied to the graph
f
of f, viewed as a
smooth submanifold of X Y . The family I

(L
f
, s) of distributions on X Y dened
above corresponds via the Schwartz kernel theorem to a family of continuous linear maps
[[
1/2
0
X [[
1/2

Y
A dieomorphism f : X Y induces a unitary operator on intrinsic Hilbert spaces given
by pull-back:
(f

)
1
: H
X
H
Y
.
At the same time, f gives rise to a symplectomorphism of cotangent bundles:
(T

f)
1
: T

X T

Y.
Of course, the more interesting Fourier integral operators from [[
1/2
X to [[
1/2
Y come
from quantizing canonical relations from T

X to T

Y which do not arise from dieomor-

phisms from X to Y . For instance, quantizing the hamiltonian ow
t
of a hamiltonian
H on T

X gives the solution operators exp(it

H/) of the Schrodinger equation.

92
7 Geometric Quantization
7.1 Prequantization
In the preceding chapter, we observed that the characteristic variety C = C
P

of a dierential
operator P

on a riemannian manifold M represents the classical analog of the kernel of P

in
H
M
, in the sense that semi-classical states contained in C quantize to rst-order approximate
solutions of the equation P

= 0. When C is reducible, the lagrangian submanifolds of M

contained in C correspond to lagrangian submanifolds of the reduced manifold C/(

. This
suggests that if C is reducible, then a quantum Hilbert space H
C
somehow associated to the
reduced manifold C/(

should map isometrically onto the kernel of P

via an appropriate
quantization of the adjoint reduction relation R

C
Hom(C/(

, T

M). Before turning to a

systematic means for dening H
C
, let us consider what conditions the reduced manifold C/(

must satisfy in order that quantizations of the reduction and projection relations be well-
dened for arbitrary quantizable semi-classical states. Specically, we ask what assumptions
on C guarantee that the classical projection operation K
C
preserves the class of quantizable
lagrangian submanifolds in T

M.
For the time being, we will ignore the Maslov correction and assume that C/(

is simply-
connected. A simple argument then shows that K
C
preserves the class of prequantizable
lagrangian submanifolds provided that the Liouville class of each (

leaf is -integral.
To interpret this condition in terms of C/(

, rst note that if C/(

is simply-connected,
any class [a] H
2
(C/(

; Z) is represented by a continuous map

(D, D)
f
(C/(

, [I])
for a xed leaf I of (

. By the homotopy lifting property, f lifts to a continuous map

(D, D)

f
(C, I).
Applying Stokes theorem, we then have

[a]

C
=

M
Z

by the assumption that the Liouville class of I is -integral. Thus, we conclude that K
C
preserves prequantizable lagrangian submanifolds provided that the reduced symplectic form

C
is itself -integral. In this case, the reduced manifold C/(

is said to be prequantizable.
Example 7.1 From Example 5.13 we recall that the standard metric on the unit n-sphere
S
n
induces a kinetic energy function k
n
whose constant energy surfaces C
E
are reducible
coisotropic submanifolds of T

S
n
for E > 0. Each leaf of the characteristic foliation of C
E
is a circle S which projects dieomorphically to a great circle in S
n
. If parametrizes a
such a geodesic, then its lift to a parametrization of S satises = (2E)
1/2
X
kn
. Since

M
(X
kn
) = 2E, the Liouville class of S is determined by the number

M
= 2 (2E)
1/2
.
Thus C
E
is prequantizable if and only if E = (n)
2
/2 (compare Example 4.2).
93

By Theorem D.2, this condition is equivalent to the existence of a principal T

bundle
Q over C/(

. In fact, the bundle Q can be constructed explicitly from the prequantum

T

bundle Q
M
over T

M by noting that for any leaf I of (

, the mod-Z

reduction of
I
represents the holonomy of Q
M
[
I
; if
I
is -integral, there exists a parallel section of Q
M
over I. If C/(

is prequantizable, parallel sections over the leaves of (

dene a foliation
O of Q
M
[
L
whose leaf space is a principal T

bundle Q over C/(

. The connection on Q
induced by the connection on Q
M
has curvature equal to the reduced symplectic form
C
.
A linear mapping : C

(P) S(H
P
) satisfying the rst two Dirac axioms will be
called a prequantization of the classical system represented by P. The basic example of
prequantization is that of cotangent bundles due to Segal [52], Koopman, and Van Hove.
In this case, the prequantum Hilbert space is taken to be the completion of the space of
smooth, complex-valued functions on the cotangent bundle P = T

M itself, with respect to

the inner-product
'f, g` =

P
f g
n
M
.
The mapping C

(P) S(H
P
) is then given explicitly by the formula
(f) = i X
f
+ m
L(f)
,
where as usual m
L(f)
denotes multiplication by the lagrangian L(f) = f
M
(X
f
). Clearly
the map is linear and satises the rst Dirac axiom; verication of the second is also
straightforward and will be carried out in somewhat more generality below.
The Segal prequantization is an important rst step towards attempts at (pre)-quantizing
more general symplectic manifolds. Roughly speaking, the idea is the following: Although
the symplectic form of an arbitrary symplectic manifold (P, ) is not exact, we can choose
a covering of P by open sets U
j
on which the restriction of the symplectic form satises
= d
j
for appropriately chosen 1-forms
j
on U
j
. In direct analogy with the Segal
prequantization, we can then associate to a function f C

(P) the operator

(f)
j
= iX
f
+ m
L
j
(f)
,
on C

(U
j
). In order to associate a global operator to f, we hope to piece these local
operators together. One way of doing this would be to impose the condition that each (f)
j
and (f)
k
coincide as operators on the function space C

(U
j
U
k
). This essentially requires
that the 1-forms
j
agree on overlaps, and we arrive at nothing new.
A true generalization of the Segal construction due to Kostant and Souriau is achieved
by rst reinterpreting (f) as an operator on sections of a line bundle over P = T

M. More
precisely, let E be the complex line bundle over P associated to the trivial principal T

bundle Q = P T

via the representation x e

ix/
of T

in U(1). The space of sections

of E identies canonically with C

(P) by means of the constant section s = 1, and we have

(f) = i
X
f
+ m
f
,
94
where denotes the connection on E induced by the connection 1-form =
M
+d on Q.
Returning to the general case, we nd that the operators (f)
j
enjoy a similar interpretation
provided that
j
+ d denes a local representative of a connection 1-form on a (possi-
bly nontrivial) principal T

bundle Q over P. The curvature of such a bundle necessarily

coincides with the symplectic form on P; according to the discussion in Appendix C, this
condition can be satised if and only if is -integral.
Denition 7.2 A prequantization of a symplectic manifold (P, ) is a principal T

bundle
Q over P equipped with a connection 1-form having curvature .
The upshot of the Kostant-Souriau construction is that prequantizable symplectic manifolds
have prequantizable Poisson algebras. While a direct proof of this fact follows the outline of
the preceding paragraph, we will study prequantizable manifolds in somewhat more detail
below. The prequantum Hilbert space in this case will be the completion of the vector
space of smooth sections of a hermitian line bundle associated to Q. Although several more
modications of this choice must be made in order to arrive at a reasonable substitute for
the intrinsic Hilbert space of the base of a cotangent bundle, this is an important rst step
in our general quantization program.
For the remainder of this section, we will focus on geometric properties of prequantum
circle bundles and prove that their existence coincides with the prequantizability of C

(P).
Automorphisms of (Q, )
Let P be a prequantizable symplectic manifold with prequantumT

bundle Q and connection

. To a function f C

(P), we associate an operator on C

(Q):

f
= X
f
fX,
where X
f
denotes the horizontal lift of the hamiltonian vector eld of f, and X is the
fundamental vector eld on Q dened by the equations
X = 1 X d = 0.
A direct computation shows that the connection form is invariant under the ow of
(k)
f
.
Conversely, if L

= 0 for some vector eld on Q, then we can decompose into its

horizontal and vertical parts:
= gX
for some real-valued function g on Q satisfying dg = d. From the denition of X it
follows that X g = 0 and [, X] = 0, and consequently, [, X] = 0. Thus, is the horizontal
lift of X
g
, and so =
g
. Moreover, the requirement that the curvature of equal the
symplectic form implies
[
f
,
g
] = [X
f
, X
g
] + (X
f
, X
g
)X 2f, gX =
|f,g
.
95
The association f
f
therefore denes a Lie algebra isomorphism between the Poisson
algebra C

(P) and the space (Q, ) of -preserving vector elds on Q with the standard
Lie bracket. This produces the exact sequence of Lie algebras
0 R (Q, ) (P, ) H
1
(P; R) 0,
where H
1
(P; R) is assigned the trivial bracket, and the image of (Q, ) in (P, ) consists
precisely of the hamiltonian vector elds on P. This sequence can be integrated to give
an exact sequence of automorphism groups as follows. Let (Q, ) be a prequantization of
(P, ), and let Aut(Q, ), Aut(P, ) denote those groups of dieomorphisms which preserve
, respectively. By the denition of X it follows that every F Aut(Q, ) preserves X and
is therefore T

-equivariant. In particular, this means that F is the lift of a dieomorphism f

of P; from the fact that

= d it follows furthermore that f

= , that is, f Aut(P, ).

The association F f denes a group homomorphism
Aut(Q, ) Aut(P, ).
To determine its kernel, we simply note that the identity map on P is covered by precisely
those automorphisms of Q given by the action of elements of T

on its connected components.

This implies that the kernel is isomorphic to H
0
(P, T

); if P (and hence Q) is connected,

this is just the circle T

and we have the exact sequence

0 T

Aut(Q, ) Aut(P, ),
i.e. Aut(Q, ) is a central extension of Aut(P, ) by T

. We interpret this observation to

mean that an automorphism of Q is determined up to phase by an automorphism of P.
If we equip the space C

(Q) of complex-valued functions on Q with the inner-product

'u, v` =

Q
uv ,
where denotes the volume form (d)
n
on Q, then each F Aut(Q, ) preserves , and
therefore denes a unitary operator U
F
on C

(Q) by composition:
U
F
(u) = u F.
Clearly the correspondence F U
F
denes a unitary representation of Aut(Q, ) on L
2
(Q);
from the exact sequence above, we therefore obtain a projective unitary representation of the
image of Aut(Q, ) in Aut(P, ).
Example 7.3 Translations of R
2n
are generated by the hamiltonian vector elds associated
to linear functionals on R
2n
. A basis for this space is given by the vector elds
X
q
i
=

p
i
X
p
i
=

q
i
,
which assume the form

q
i
=

p
i
q
i
X
p
i
=

q
i
96
when lifted to Q = R
2n
T

via the Segal prescription

f
= X
f
L(f)X, where again
L(f) = f
M
(X
f
). Our earlier results show that [
q
i
,
p
j
] =
|q
i
,p
j

=
ij
X, and so the
vector elds
q
i
,
p
i
generate a Lie subalgebra h
n
(Q, ) isomorphic to R
2n
R with
bracket given by
[(v, a), (w, b)] = (0, (v, w)).
By exponentiating, we nd that h
n
corresponds to a subgroup H
n
Aut(Q, ) which
comprises a central extension of the translation group:
0 T

H
n
R
2n
0.
The group H
n
is known as the Heisenberg group of R
2n
with its usual symplectic structure.
In explicit terms, H
n
is dieomorphic to R
2n
T

, with group multiplication given by

(Q, P, ) (Q
t
, P
t
,
t
) = (Q + Q
t
, P + P
t
,
t
) = (Q + Q
t
, P + P
t
, +
t
+

j
P
j
Q
t
j
).

Kostant-Souriau prequantization
To prequantize the Poisson algebra C

(P), we rst recall that complex line bundles E

k
associated to Q arise via representations of T

in U(1) of the form x e

ikx/
. Smooth
sections of E
k
identify with functions on Q satisfying the equivariance condition
f(p a) = e
ika/
f(p)
for a T

; in other words, the space of sections of E

k
is isomorphic to the ik/-eigenspace
c
k
of the fundamental vector eld X. Under this correspondence, covariant dierentiation
by a vector eld on P is given simply by the Lie derivative with respect to the horizontal
lift of to Q.
To each f C

(P) and integer k, we assign an operator on C

(Q) by

(k)
f
=
i
k

f
.
Since [
f
, X] = 0, the operator
(k)
f
restricts to an operator on each eigenspace c
k
of X having
the form

k
(f) =
i
k
X
f
+ m
f
.
Evidently the map
k
satises the rst Dirac axiom; moreover

k
(f, g) = k [
k
(f),
k
(g)]

.
To verify that
k
(f) is self-adjoint, is suces to prove that
k
(f) acts as a symmetric operator
on the real subspace of c
k
. For this purpose, we note that if the hamiltonian vector eld of
97
f is integrable (for example, if f is compactly supported), then the fact that
f
preserves
implies that
'
f
u, v` = 'u,
f
v`
for real-valued functions u, v on Q. Combined with the denition of
k
, conjugate-symmetry
of the inner-product implies that
'
k
(f) u, v` = 'u,
k
(f) v`.
We now dene the prequantum line bundle associated to Q as E = E
1
, equipped
with a hermitian metric ' , ` compatible with its induced connection, and let H
P
denote
the L
2
completion of the space of compactly supported sections of E with respect to the
inner-product
's
1
, s
2
` =

P
's
1
, s
2
`
n
.
For each f C

(P), the operator

1
(f) extends to an essentially self-adjoint operator on
H
P
; from the general remarks above, it follows that
1
denes a prequantization of the
Poisson algebra C

(P).
7.2 Polarizations and the metaplectic correction
The Kostant-Souriau prequantization of symplectic manifolds fails to satisfy the third Dirac
axiom and is therefore not a quantization. In fact, the position and momentum operators
(q) = i

p
+ m
q
(p) = i

q
on T

R both commute with /p and therefore do not form a complete set. According to
our earliest concepts of quantization (see the Introduction), the operator corresponding to q
should act by multiplication alone, whereas q involves the spurious /p term. If we restrict
to the p-independent subspace C

q
(R
2
) C

(R), however, this diculty is overcome: on

C

q
(R
2
),
(q) = m
q
(p) = i

q
.
This association agrees with our earlier heuristic quantization of the classical position and
momentum observables and suggests in general that the space of sections of a prequantum
line bundle E is too large for the third Dirac axiom to be satised. Indeed, in quantum
mechanics, wave functions depend on only half of the phase space variables, whereas the
space of sections of E has the size of the space of functions on P. For an interpretation of
this argument in terms of the Heisenberg uncertainty principle, see [53].
From the standpoint of WKB quantization, the appropriate quantum state space asso-
ciated with a classical conguration space M is (the L
2
completion of) [[
1/2
0
M, which we
temporarily identify with C

(M) using a metric on M. Since the Liouville form

M
vanishes
on each ber of the projection T

M

M, we can interpret C

(M) as the space of sections

of the prequantum line bundle E over T

M which are parallel along each ber of . In other

98
words, the correct adjustment to the size of the prequantum Hilbert space is determined by
the (canonical) foliation of T

M by lagrangian submanifolds.
This adjustment is generalized in the framework of geometric quantization by the intro-
duction of the following concept.
Denition 7.4 A polarization of a symplectic manifold (P, ) is an involutive lagrangian
subbundle T of T
C
P.
Here, T
C
P denotes the complexication of the tangent bundle of P, equipped with the
complex-valued symplectic form
C
given by the complex-linear extension of to T
C
P. In-
tegrability of T means that locally on P there exist complex-valued functions whose hamil-
tonian vector elds (with respect to
C
) span T. The quantum state space associated to P
is then given by the space of sections s of a prequantum line bundle E over P which are
covariantly constant along T. That is, for each complex vector eld X on P lying in T, we
have
X
s = 0.
Real polarizations
The standard polarization of a classical phase space T

M is given by the complexication

T = V M iV M of the vertical subbundle of T(T

M). In this case, the polarization T

satises T = T, meaning that T
p
is a totally real subspace of T
C,p
(T

M) for each p
T

M. In general, any polarization T of a symplectic manifold P which satises T = T

is the complexication of an integrable lagrangian subbundle of TP and is called a real
polarization of P.
Example 7.5 The most basic examples of real polarizations are given by either the planes
q=constant or p=constant in R
2n
. These correspond to the
q
and e
i/p,)/
bases of L
2
(R
n
)
respectively, in the sense that
f =

R
n
f(q) (q) dq = (2)
n/2

R
n
e
i/p,q)/

f dp.
Covariant dierentiation along vector elds X tangent to the q=constant polarization T
q
is given by the ordinary Lie derivative with respect to X, and so the space of T
q
-parallel
sections of the prequantum line bundle E identies with the space of smooth functions on
q-space. For the p=constant polarization T
p
, we consider covariant derivatives of the form

q
=

q
p

.
A T
p
-parallel section of E must therefore satisfy

q
2ip
and is therefore of the form (q, p) = v(p)e
i/p,q)/
. Note that the T
p
representation of the
free-particle H(q, p) = p
2
/2 is (H) = m
p
2
/2
, and so the p-polarization appears to be better
adapted to this case.
99

A real polarization T of a symplectic 2n-manifold P denes a subspace T(P) C

(P)
of functions constant along the leaves of T. According to the Hamilton-Jacobi theorem, the
hamiltonian vector eld of any member of T(P) is contained in T. Thus f, g = 0 for
any f, g T(P), i.e., T(P) is an abelian Poisson subalgebra of C

(P). By modifying the

normal form results of Chapter 4, one can furthermore prove
Theorem 7.6 Every real polarization is locally isomorphic to the q=constant polarization
on R
2n
.
By pulling back the position functions q
i
to P, we have
Corollary 7.7 For each p P, there exist f
i
T(P), i = 1, , n, such that the hamilto-
nian vector elds X
i
span T in a neighborhood of p.
An ane structure on a manifold M is a curvature- and torsion-free connection on TM.
Since any polarization-preserving symplectic transformation of T

R
n
R
2n
is ane on bers
(see Theorem 3.29), we have the following result.
Corollary 7.8 Each leaf of any real polarization carries a natural ane structure.
The main remark we wish to make for the moment is that our earlier prequantization
1
of C

(P) now represents the subalgebra T(P) C

(P) as multiplication operators on H

T
.
Since any f T(P) has the property that its hamiltonian vector eld X
f
lies completely
within the polarization T, it follows that
X
f
s = 0 for all s H
T
. By the denition of
1
,
this implies that
1
(f) = m
f
, the multiplication operator on H
T
. More generally, those
functions whose hamiltonian vector elds have ows which leave the polarization invariant
(not necessarily leaf by leaf) are those which are ane along the leaves.
Example 7.9 Following [28], we call a polarization T on P brating if each leaf of T is
simply-connected and the leaf space P
T
= P/T is a smooth manifold. In this case, the
quotient map p : P P
T
satises T = ker p

, and so p

P
T
identies with the normal
bundle T

P. Since T is a lagrangian distribution, the map : TP T

P sends T to
T

as well, thus dening a natural identication

T p

P
T
.
A function H: P R constant on the leaves of T induces a function H
T
on P
T
satisfying
dH = p

dH
T
.
From this remark, it follows that a section s of T over a leaf L is parallel with respect to
the connection described above if and only if ( s) denes a xed element of T

|L
P
T
.
Similarly, if E is a prequantum line bundle over P, we may use parallel sections of E over
the leaves of T to construct a foliation of the total space of E. The quotient of E by this
100
foliation denes a hermitian line bundle E
T
over the leaf space P
T
whose sections identify
with T-parallel sections of E.
The simplest example of a brating polarization is given by the vertical polarization T
M
of a cotangent bundle T

M. The basic geometry of this situation remains the same if the

symplectic structure on T

M is perturbed by a form on M. More precisely, given a closed

2-form on M, we can twist the standard symplectic structure on T

M as follows:
=
M
+

.
Since

vanishes on the bers of the projection, is a symplectic form on T

M and ker

is again a polarization. This general type of symplectic manifold is known as a twisted

cotangent bundle. A check of denitions shows that the standard and twisted symplectic
structures on T

M are equivalent precisely when their dierence is the pull-back of an exact

form on M. Moreover, the symplectic manifold (T

M,
M
+

) is prequantizable if and
only if the cohomology class of in H
2
(M; R) is -integral. Finally, an application of
Corollary 7.8 proves that if each leaf of a brating polarization T on a symplectic manifold
(P, ), is complete, then (P, ) is symplectomorphic to a twisted symplectic structure on
T

P
T
.

Complex polarizations
Associated to any polarization T are the distributions
D
C
= T T E
C
= T +T
which arise as the complexications of distributions D, E in TP. Although their dimensions
vary in general from point to point, D and E are pointwise -orthogonal, i.e.
D

x
= E
x
for all x P. From the denition of T, the distribution D is involutive. The polarization
T is called strongly admissible provided that E is involutive, the leaf spaces P
D
and P
E
are smooth manifolds, and the natural projection
P
D

P
E
is a submersion. In this case, each ber of carries a Kahler structure, and geometric quan-
tization attempts to construct a quantum state space for P from sections of an appropriate
line bundle over P which are parallel along D and holomorphic along the bers of . For a
discussion of quantization in this general setting, we refer to [53].
A polarization satisfying T T = 0 is called totally complex and identies with the
graph of a complex structure J on P, i.e.,
T = (v, iJv) : v TP
101
under the identication T
C
P TP iTP. We emphasize that J is a genuine complex
structure on P due to the integrability condition on the polarization T. Moreover, J is
compatible in the usual sense with the symplectic form on P, so the hermitian metric
' , ` = g
J
+ i is a Kahler structure on P. In this section, we will only cite two examples
involving totally complex T, in which case P
D
= P is Kahler, and P
E
= pt.
Example 7.10 Consider the complex plane C with its usual complex and symplectic struc-
tures. With respect to Darboux coordinates (q, p), the Cauchy-Riemann operator is dened
as

z
=
1
2

q
+ i

p

.
A function f : C C is holomorphic if f/z = 0.
If we identify sections of the (trivial) prequantum line bundle E with smooth, complex-
valued functions on C, then T-parallel sections correspond to those functions annihilated by
the covariant derivative

z
= 2

z
+ m
z
.
To determine the general form of these sections, we note that
z
= 0 if and only if for
some branch of the logarithm,

z
log =
z
2
,
or
log =
zz
2
+ h
for some holomorphic function h. Hence
(z) = e
[z[
2
/2
with holomorphic is the general form for T-parallel sections of E. The space H
C
thus
identies with the space of holomorphic : C C satisfying
|| =

C
[(z)[
2
e
[z[
2
dz < ,
or, in other words, the space of holomorphic functions which are square-integrable with
respect to the measure e
[z[
2
dz, known as the Fock or Bargman-Segal space. For a func-
tion to be quantizable in this picture, its hamiltonian ow must preserve both the metric
and symplectic structure of C and therefore consist solely of euclidean motions. Among
such functions are the usual position and momentum observables, as well as the harmonic
oscillator.

Example 7.11 If (P, ) is any prequantizable symplectic manifold with a totally complex
polarization T, then the prequantum line bundle E associated to can be given the structure
of a holomorphic line bundle by taking the (0, 1) component of the connection on E with
102
respect to the complex structure J arising from T. The space H
T
of T-parallel sections of E
then equals the space of holomorphic sections of E and is therefore completely determined
by the complex geometry of P.
Using some machinery from algebraic geometry (see [15]), it can be shown that for com-
pact P, the space H
P
is nite-dimensional and that its dimension is given asymptotically by
the symplectic volume of P. More precisely, we note that for k Z
+
, the line bundle E
k
denes a holomorphic prequantum line bundle associated with (P, k). For suciently large
k, the dimension of the quantum state space H
P
k
of holomorphic sections of E
k
is given by
the Hirzebruch-Riemann-Roch formula:
dimH
P
k
=

P
e
k
Td(P),
where Td denotes the Todd polynomial in the total Chern class of P. The rst conclusion
to be drawn from this fact is that for k large, dimH
P
k
is a symplectic invariant of P inde-
pendent of its complex structure. Roughly speaking, k plays the role of
1
, and so large
k means that we are approaching the classical limit. A second point to note is that dimH
P
k
is a polynomial in k (or
1
) whose leading-order term is

P
(k)
n
n!
= k
n
vol P.
Consequently the number of quantum states is determined asymptotically by the volume of
P.

Metalinear structures and half-forms

For the remainder of this chapter, we will focus on the quantization of symplectic manifolds
P equipped with a prequantum line bundle E and a suciently nice real polarization T.
Although the space H
T
of T-parallel sections of E appears to have the right size in the
simplest examples, there are still several problems to be resolved before we have a suitable
quantum state space. The rst arises as soon as we attempt to dene a pre-Hilbert space
structure on H
T
. On P, the square of an T-parallel section s is constant along the leaves of
T, and thus the integral
|s|
2
=

P
's, s`
n
diverges in general. On the other hand, there is no canonical measure on the leaf space P
T
with which to integrate the induced function 's, s`.
Regardless of how this rst diculty is resolved, we will again try to quantize the Poisson
algebra C

(P) by representing elements of Aut(P, ) as (projective) unitary operators on

H
T
. The most obvious quantization of a symplectomorphism f : P P (or more general
canonical relation), however, is an operator from H
T
to H
f(T)
. If f does not preserve the
polarization T, then these spaces are distinct. Thus, we will need some means for canonically
identifying the quantum state spaces H
T
associated to dierent polarizations of P.
103
Finally, the polarization T may have multiply-connected leaves, over most of which the
prequantum line bundle may admit only the trivial parallel section. A preliminary solution
to this problem is to admit distributional states, such as are given by T-parallel sections
of E whose restrictions to each leaf are smooth. Leaves on which E has trivial holonomy are
then said to comprise the Bohr-Sommerfeld subvariety of the pair (P, T).
Example 7.12 Consider the punctured phase plane

R
2
= R
2
` 0 polarized by level sets
of the harmonic oscillator H(q, p) = (q
2
+ p
2
)/2. The holonomy of the usual prequantum
line bundle on the level set H
1
(E) is given by the mod-Z

reduction of the area it encloses;

thus the Bohr-Sommerfeld variety consists of those circles of energy E = n. As in the
case of WKB prequantization, these energy levels do not correspond to actual physical
measurements, and so we will again need to incorporate a sort of Maslov correction.

The solution to manhy of these diculties relies on the use of a metaplectic structure on
P, a concept which we now introduce. Let P be a principal G bundle over a manifold M.
A meta G-bundle associated to P and a central extension
1 K

G

G 1
of G is a principal

G bundle

P over M together with a map :

P P satisfying the
equivariance condition
(p a) = (p) (a)
for all a

G. Two meta G-bundles (

P
1
,
1
), (

P
2
,
2
) over P are considered equivalent if
there exists a

G-equivariant dieomorphism :

P
1


P
2
such that
1
=
2
.
Example 7.13 A riemannian structure together with an orientation of an n-manifold M
denes a bundle of oriented orthonormal frames in TM, i.e. an SO(n)-structure on M. A
spin structure on M is then a meta SO(n)-bundle corresponding to the extension
1 K Spin(n) SO(n) 1
given by the double cover Spin(n) of SO(n). An orientable manifold admits a spin structure
if and only if its second Stiefel-Whitney class vanishes.

Example 7.14 Since

1
(Sp(n)) Z, there exists a unique connected double-covering group
of Sp(n), known as the metaplectic group Mp(n). A symplectic 2n-plane bundle F admits
a reduction of its structure group from GL(2n) to Sp(n); a metaplectic structure on F
then corresponds to a lifting of the symplectic frame bundle to a principal Mp(n) bundle. In
general, a symplectic vector bundle admits a metaplectic structure if and only if its second
Stiefel-Whitney class vanishes.
A symplectic manifold P whose tangent bundle is equipped with a metaplectic structure
is called a metaplectic manifold. If the Stiefel-Whitney class w
2
(P) is zero, metaplectic
104
structures are classied by the set H
1
(P; Z
2
) (see [28]). We emphasize that this classication
depends on the bundle of metaplectic frames and its covering map to the bundle of symplectic
frames. To see this explicitly in a special case, note that up to topological equivalence, the
punctured plane

R
2
admits only trivial Sp(1) and Mp(1) principal bundles, since both Sp(1)
and Mp(1) are connected. On the other hand, a metaplectic structure
:

R
2
Mp(1)

R
2
Sp(1)
is dened for any choice of continuous map s:

R
2
Sp(1) by
(p, A) = (p, s(p) (A))
for A Mp(1). Two such maps s
0
, s
1
dene equivalent metaplectic structures if and only if
(s
1
)
1
s
0
admits a continuous lift s:

R
2
Mp(1), in which case an equivalence is given by
the map :

R
2
Mp(1)

R
2
Mp(1) dened as
(p, A) = (p, s(p) A)
By the usual homotopy theory for continuous groups, this occurs precisely when [s
0
] = [s
1
]
as elements of
1
(Sp(1))/
#

1
(Mp(1)) Z
2
.

Example 7.15 Lying within the metaplectic group is a double-cover Ml(n) of the subgroup
Gl(n) preserving the usual lagrangian splitting of R
2n
. The double covering of GL(n) then
induced by the identication GL(n) Gl(n) (see Example 3.1) is called the metalinear
group ML(n). It is trivial as a topological covering, but as a group it is not the direct
product GL(n) Z
2
. More explicitly, the group ML(n) is isomorphic to the direct product
GL
+
(n) Z
4
with the covering map ML(n)

GL(n) given by
(A, a) = A e
ia
.
A metalinear lifting of the frames of an n-plane bundle E is called a metalinear structure
on E. If E is orientable, its structure group can be reduced to GL
+
(n), which can be
interpreted as the identity component of ML(n) in order to give a metalinear structure on
E. More generally, a vector bundle admits a metalinear structure if and only if the square
of its rst Stiefel-Whitney class is zero (see [28]).

The importance of the metalinear group for our purposes is that it admits 1-dimensional
representations which are not the lifts of representations of GL(n). Bundles associated to
metalinear structures via these representations will be the key to the metaplectic correction
of the prequantization procedure. First note that by means of the quotient homomorphism
ML(n) ML(n)/GL
+
(n) Z
4
,
105
we can associate a principal Z
4
bundle to any metalinear structure (MB(E), ) on a vector
bundle F. From the representation
a e
ia/2
of Z
4
in U(1), we then obtain a complex line bundle
1/2
F called the bundle of half-forms
associated to the triple (F, MB(F), ). The reason for this terminology is that the bundle

1/2
F can be constructed directly from the metalinear frame bundle via the representation
ML(n) C

given by the square-root of (det )

(A, a) (det A)
1/2
e
ia/2
.
A section of
1/2
F then identies with a complex-valued function on MB(F) satisfying
(e) = (det A)
1/2
e
ia/2
(e (A, a)),
and therefore represents, loosely speaking, the square-root of a volume form on F. By
inverting and conjugating the preceding representation of ML(n), one similarly denes the
bundles
1/2
F and
1/2
F of conjugate and negative half-forms on F. Evidently, the product
of two half-forms on F yields an n-form; similarly, a half-form and a conjugate half-form
can be multiplied to give a 1-density on the bundle F.
An important link between metalinear and metaplectic structures on symplectic mani-
folds can be described as follows.
Theorem 7.16 Let (P, ) be a symplectic 2n-manifold and L TP a lagrangian subbundle.
Then TP admits a metaplectic structure if and only if L admits a metalinear structure.
Proof. If J is any -compatible almost complex structure on P, then TP = LJL LL.
By the Whitney product theorem, we then have
w
2
(P) = w
1
(L)
2
,
and our assertion follows from the remarks in the preceding examples.
More explicitly, we may use L and J to reduce the structure group of TP to the subgroup
Gl(n) Sp(n) corresponding to GL(n). The resulting frame bundle is isomorphic to the
frame bundle of L, and thus, a metalinear structure on L can be enlarged to a metaplectic
structure on TP, while a metaplectic structure on TP reduces to a metalinear structure on
L.
2
Now let P be a metaplectic manifold equipped with a prequantum line bundle E and a
real polarization T. By the preceding theorem, T inherits a metalinear structure from the
metaplectic structure on TP, and thus
1/2
T is dened; it is equipped with a natural at
connection inherited from the one on T.
Denition 7.17 The quantum state space H
T
associated to P is the space of sections of
E
1/2
T which are covariantly constant and along each leaf of T.
106
In the following examples, we will sketch how in certain cases this denition enables us to
overcome the diculties mentioned at the beginning of this section.
Example 7.18 We rst return to Example 7.12 to illustrate how the introduction of half-
forms also enables us to incorporate the Bohr-Sommerfeld correction in the case of the
1-dimensional harmonic oscillator.
To quantize the punctured plane

R
2
with its polarization by circles centered at the origin,
we will use the trivial metaplectic structure on

R
2
(see Example 7.14). In this case, both the
trivial Mp(1) and Sp(1) bundles over

R
2
can be identied with the set of triples (r, , a e
i
),
where a, r > 0 and 0 , 2, and the covering map

R
2
Mp(1)

R
2
Sp(1) is given
by
(r, , a e
i
) (r, , a e
2i
).
The pull-back of

R
2
Sp(1) of T

R
2
to the circle S
r
of radius r can be reduced to the trivial
Gl(1) principal bundle over S
r
0
by means of the lagrangian splitting TS
r
0
JTS
r
0
. In terms
of the identications above, the subbundle

R
2
Gl(1) equal the subset
(r, , a e
i
) : r = r
0
, a > 0, 0 2
of

R
2
Sp(1) while its corresponding metalinear bundle equals the subset
(r, , a e
i(/2+c)
) : r = r
0
, a > 0, 0 2, c 0, 1
of

R
2
Mp(1). Note that this is a nontrivial Ml(1) bundle over S
r
0
.
From the expression for the covering map given above, it follows easily that the parallel
transport of a negative half-form associated to TS
r
0
around S
r
0
amounts to multiplication
by 1. Consequently, the bundle E
1/2
T admits a parallel section over S
r
0
precisely
when
r
2
0
=

Sr
0
p dq = (2n + 1)
for some integer n. That is, the Bohr-Sommerfeld variety consists of circles of energy E =
(n + 1/2), in accordance with the corrected Bohr-Sommerfeld quantization conditions
described in Chapter 4.

Example 7.19 Over each leaf L of the polarization T, the Bott connection on TL described
in Example 5.17 denes a foliation of the preimage of L in B(T). As proven in [28], the
quotient of B(T) by this leaf-wise foliation denes a metalinear structure on the leaf space
P
T
in such a way that half-forms on P
T
are in 1-1 correspondence with T-parallel negative
half-forms on T. In this way, the quantum state space H
T
identies with the space of
compactly supported sections of
E
T

1/2
P
T
.
The latter space is equipped with a natural inner-product
's
1

1
, s
2

2
` =

P
F
's
1
, s
2
`
1

2
.
Note that integration is well-dened, since
1

2
is a density on P
T
.
107

Blattner-Kostant-Sternberg kernels
If T
1
and T
2
are brating polarizations of P with the property that T
1,x
is transverse to
T
2,x
at each x P, then the quantum state spaces H
T
1
and H
T
2
can be related as follows.
By assumption, the symplectic form denes an isomorphism of T
1
with T

2
; similarly, the
metaplectic structure on P denes a natural isomorphism between the space of half-forms
on T
2
and the space of conjugate half-forms on T
1
(see [28]). Using this identication, we
can pair a half-form
1
on T
1
with a half-form
2
on T
2
to obtain a function (
1
,
2
) on P.
If we identify elements
1
,
2
of H
T
1
and H
T
2
with T
i
-parallel sections s
i

i
of E
1/2
T
i
on P, then a sesquilinear pairing H
T
1
H
T
2
C is dened by
''
1
,
2
`` =
1
(2)
n/2

P
's
1
, s
2
` (
1
,
2
)
n
,
where dim(P) = 2n. This pairing is nondegenerate and thus determines a linear operator
B: H
T
2
H
T
1
satisfying
''
1
,
2
`` = '
1
, B
2
`
for all
1
H
1
and
2
H
2
. This construction, which can be extended to more general pairs
of polarizations, is due to Blattner, Kostant, and Sternberg denes a sesquilinear pairing on
the associated quantum state spaces via a Blattner-Kostant-Sternberg kernel.
Example 7.20 The basic example of a Blattner-Kostant-Sternberg kernel arises from the
classical Fourier transform. From Example 7.5, we recall that for the q =constant polariza-
tion T
q
of R
2n
, the space H
Tq
consists of p-independent functions
q
(q, p) = u(q) on R
2n
,
which we may identify with the space of smooth functions on q-space. Similarly, if T
p
is the
p =constant polarization, then a smooth function v(p) on p-space identies with an element

p
(q, p) = v(p)e
i/q,p)/
of H
Tp
. The sesquilinear pairing described above is therefore given by
''
q
,
p
`` =
1
(2)
n/2

R
2n
e
i/q,p)/
u(q) v(p) [dq dp[.
Note that this corresponds to the usual association of a distribution v(q) on q-space to a
function v(p) on p-space via the inverse asymptotic Fourier transform.

Given a symplectic manifold P with the structures above, geometric quantization at-
tempts to represent the Poisson algebra C

(P) on H
T
as follows. First, if the hamiltonian
ow
t
of f C

(P) preserves T, then it lifts to a 1-parameter family

t
of operators on
smooth sections of E
1/2
T. The quantum operator associated to T is then
(f) = i
d
dt
(
t
)[
t=0
.
In particular, (f) acts as pointwise multiplication for any f T(P).
108
For those f C

(P) whose hamiltonian ow does not preserve T, we assume that

for t (0, ), the dierential D
t
maps T into a polarization T
t
for which there exists a
Blattner-Kostant-Sternberg kernel and a corresponding operator U
t
: H
Tt
H
T
. The ow

t
denes an operator
t
: H
T
H
Tt
, and the quantum operator (f) is dened as
(f) = i
d
dt
((U
t

t
) )[
t=0
.
We refer to [28] for further details.
7.3 Quantization of semi-classical states
In this brief section, we suppose that P is a metaplectic manifold with a prequantum line
bundle E and a brating polarization T. Our goal is to comment briey on how an ele-
ment of H
T
can be constructed from a semi-classical state in P consisting of a lagrangian
submanifold L P with some extra structure, such as a half-density or half-form with
coecients in a prequantum line bundle over P.
Example 7.21 If L P intersects each leaf of T transversely in at most one point, then
the section s of E
1/2
T over L corresponding to a section of E[[
1/2
L can be extended
to a section s over P which is covariantly constant on each leaf and which vanishes on leaves
which are disjoint from L. In this way, we can regard s as the quantization of the pair (L, s).
We can generalize this viewpoint by considering a lagrangian submanifold L which has
possibly multiple transverse intersections with the leaves of T (that is, each intersection of
L with a leaf of T is transverse). If s is again a section of E
1/2
T over L, then we
can construct and element of H
T
by superposition. More precisely, we think of L as the
union of lagrangian submanifolds L
j
, such that each L
j
intersects any leaf of T at most
once. Applying the procedure above to each (L
j
, s[
L
j
) and summing the results produces an
element of H
T
as long as this sum converges.
Finally, if L is a union of leaves of T and s is a section of E
1/2
T over L which is
covariantly constant along each leaf of T, then we can pair it with arbitrary elements of H
T
by integration over L, thus obtaining a linear functional on H
T
, which, via the inner product
on H
T
, can be considered as a generalized section.

Example 7.22 To indicate briey how distributional elements of H

T
arising as in the
preceding example are paired, we consider the following situation. Let (L, s) be a semi-
classical state in R
2n
such that L equals the zero section of R
2n
T

R
n
and s is any section
of E
1/2
T
q
over L. Similarly, let

L be the ber of T

R
n
over 0 R
n
along with a constant
section s of E
1/2
T
q
over

L. According to the preceding example, the quantization of
(L, s) is given by an ordinary smooth function on R
n
, whereas the quantization of (

L, s) is
concentrated at the origin of R
n
.
The meaning of this last statement is made more precise by the fact that if T : R
2n
R
2n
is a linear symplectomorphism, then the associated quantum operator H
Tq
H
T(Tq)
is
109
required to be unitary. By choosing T so that both L and

L are transverse to T(T
q
), we can
quantize (L, s) and (

L, s) with respect to this new polarization to obtain smooth functions

on the leaf space R
2n
T(Tq)
whose pairing is dened by integration. In terms of H
Tq
, this means
that, up to a normalization, the quantum state obtained from (

L, s) is a Dirac -function
concentrated at the origin of R
n
.

110
8 Algebraic Quantization
An approach to quantization going back to Dirac and recently revived in [9][10][11] is based
on the idea that the multiplicative structure of the -algebra of quantum observables is more
central to quantization than the representation of observables as operators on Hilbert space,
and that most of quantum mechanics can be done without regards for the precise nature of
observables. Quantum observables comprise a noncommutative algebra / which is assumed
to belong to a family /

of algebras, all with a common underlying vector space, but with

an -dependent multiplication

. As 0, the algebra /

approaches a commutative
algebra /
0
, which may be interpreted as the algebra of functions on a classical phase space.
Additionally, it is assumed that the

-commutator approaches the Poisson bracket in the

limit of large quantum numbers:
f, g = lim
0
(f

g g

f)/i.
The abstract goal of algebraic quantization is to construct the family /

of noncommutative
algebras from a Poisson algebra. In this chapter, we sketch two approaches towards this
goal, known as deformation quantization and the method of symplectic groupoids.
8.1 Poisson algebras and Poisson manifolds
The main objects in algebraic quantization theory are Poisson algebras and Poisson mani-
folds.
Denition 8.1 A Poisson algebra is a real vector space / equipped with a commutative,
associative algebra structure
(f, g) fg
and a Lie algebra structure
(f, g) f, g
which satisfy the compatibility condition
fg, h = f g, h +f, h g.
A Poisson manifold is a manifold P whose function space C

(P) is a Poisson algebra

with respect to the usual pointwise multiplication of functions and a prescribed Lie algebra
structure.
If P, Q are Poisson manifolds, a smooth map : P Q is called a Poisson map provided
that it preserves Poisson brackets, i.e. f, g = f , g for all f, g C

(Q). Similarly,
is called an anti-Poisson map if f, g = f , g for all f, g C

(Q).
On a Poisson manifold, the Leibniz identity implies that the Poisson bracket is given by
a skew-symmetric contravariant tensor eld via the formula
f, g = (df, dg).
The main examples of Poisson manifolds we will be concerned with are the following.
111
Example 8.2 1. Any smooth manifold M is a Poisson manifold when C

(M) is given the

trivial bracket f, g = 0.
2. Any symplectic manifold is a Poisson manifold with respect to its standard Poisson
structure
f, g = X
g
f.
3. Let g be a nite-dimensional Lie algebra with dual g

. The dierential of a smooth

function F : g

R is a map DF : g

g g whose composition with the Lie

bracket on g denes a smooth map [DF] : g

g. The Lie-Poisson operator on

C

(g

) is the dierential operator T dened by

TF =
1
2
'[DF]
(x,y)
, x + y`.
The dual space g

is then a Poisson manifold when equipped with the Lie-Poisson bracket

f, g =

T(f g),
where : g

is the diagonal and f, g C

(g

). More concretely,
f, g() = ([Df, Dg]),
where g

, and Df, Dg : g

g are the dierentials of f and g.

4. If V is a nite-dimensional vector space and is a skew-symmetric bilinear form on V

,
then
f, g
def
= (df, dg)
denes a Poisson algebra structure on C

(V ), making V a Poisson manifold.

8.2 Deformation quantization

The aim of deformation quantization is to describe the family

of quantum products on a
Poisson algebra / as an asymptotic series (in ) of products on /. In accordance with the
introductory remarks above, the zeroth and rst order terms of this series are determined
by the Poisson algebra structure of /; the role of the higher-order terms is roughly speaking
to give a more precise -dependent path from classical to quantum mechanics.
Denition 8.3 Let / be a complex vector space equipped with a commutative associative
algebra structure, and let

be a family of associative multiplications on / given by a formal

power series
f

g =

j=0
B
j
(f, g)
j
where each B
j
: // / is a bilinear map. Then

is called a -deformation of / if
112
1. B
0
(f, g) equals the product in /.
2. B
j
(g, f) = (1)
j
B(f, g)
3. B
j
(1, f) = 0 for j 1
4. B
j
is a dierential operator in each argument.
5. (f

g)

h = f

(g

h)
For condition (5) to make sense, we must extend the product

in the obvious way from /

to the space /[[]] of formal power series. We emphasize that the power series above is not
in general assumed to converge for = 0. Instead it should be viewed as an asymptotic
expansion for the product, and manipulations of this series will be purely formal.
In any case, conditions (1-5) imply that
f, g
def
=
1
2i
B
1
(f, g)
denes a Poisson algebra structure on /. This observation suggests the question of whether
every Poisson structure on a function algebra C

(P) can be realized as the rst order term

in a -deformation of /.
Example 8.4 If V is a nite-dimensional real vector space and : V

R is a
skew-symmetric bilinear form on V

, then the Hessian of at 0 V

is a linear map
A = H : V

V V . We dene the Poisson operator associated to as the

second-order dierential operator on C

(V V )
T

= 'A(/y

, /z

), (/y, /z)`,
where (y, z) are linear coordinates on V V arising from a single set of linear coordinates on
V , and (y

, z

) are dual to (y, z). The Moyal-Weyl operator is then the pseudodierential
operator given by exponentiation:

,
= e
iT/2
.
Pointwise multiplication of functions in f, g C

(R
2n
) can be dened as the pull-back
f g =

f g,
where : R
2n
R
2n
R
2n
is the diagonal embedding, and (f g)(y, z) = f(y) g(z) for
f, g C

(R
2n
). Similarly, a straightforward computation shows that the Poisson operator
is related to the Poisson bracket induced by via the equation
f, g =

T(f g).
The diagonal pull-back
f

g =

(f g)
113
of the Moyal operator denes a -deformation of the Poisson algebra C

(V ) called the
Moyal-Weyl product. In terms of the linear coordinates (y, z) on V V , the operator B
j
in the expansion of the Moyal-Weyl product is
B
j
(f, g) =
1
j!

i
2

r,s

r,s

y
r

z
s

j
f(y)g(z)

y=z=x
.
If is nondegenerate, then an application of the principle of stationary phase shows that
an integral expression for f

g is given by
(f

g)(x) =

e
iQ(yx,zx)
f(y) g(z) dy dz,
where Q is the skew-symmetric bilinear form on V induced by .

A Poisson manifold P is called regular if its Poisson tensor has constant rank, in
which case a theorem of Lie [39] asserts that P is locally isomorphic to a vector space with a
constant Poisson structure. In view of Example 8.4, any regular Poisson manifold is locally
deformation quantizable. To construct a -product on all of C

(P), one may therefore

attempt to patch together the local deformations to arrive at a global -product.
This technique has succeeded. A theorem of DeWilde and Lecomte [19] asserts that the
Poisson algebra of any nite-dimensional symplectic manifold admits a -deformation. Using
similar techniques, Melotte [44] extended their result to arbitrary regular Poisson manifolds.
A simplied proof of these results has recently been given by Fedosov [23] (see [68] for a
survey of deformation quantization, emphasizing Fedosovs construction).
8.3 Symplectic groupoids
The method of symplectic groupoids also attempts to directly construct a noncommutative
algebra /

of quantum observables without explicitly identifying a quantum state space.

Unlike deformation quantization, however, this approach involves a geometric procedure
which attempts to construct /

for a particular value of and in particular incorporates

geometric objects with certain quantum properties.
Groupoids
In this section we collect some basic denitions and examples of groupoids and their coun-
terparts in symplectic geometry.
Denition 8.5 A groupoid is a set endowed with a product map (x, y) xy dened
on a subset
2
called the set of composable pairs, and an inverse map :
satisfying the conditions
1.
2
= id
114
2. If (x, y), (y, z)
2
, then (xy, z), (x, yz)
2
, and (xy)z = x(yz).
3. ((x), x)
2
for all x , and if (x, y)
2
, then (x) (xy) = y.
4. (x, (x))
2
for all x , and if (z, x)
2
, then (zx)(x) = z.
Note that by (3), the map is bijective and thus inverses in are unique. An element x
of can be thought of as an arrow with source (x) = (x)x and target (x) = x(x); a
pair (x, y) then belongs to
2
if and only if the source of y equals the target of x. The set

0
of all sources (and targets) is called the base of , and is said to be a groupoid over

0
. Elements of
0
are units of in the sense that x(x) = x and (x)x = x for all x .
Finally, the multiplication relation of is the subset
m = (xy, x, y) : (x, y)
2

of . In abstract terms, a groupoid is a small category in which all morphisms have

inverses.
Example 8.6 1. Any group is a groupoid over its identity element, and conversely, any
groupoid whose base is a singleton comprises a group.
2. A disjoint union of groupoids is a groupoid over the union of their bases. If is a groupoid
with base
0
and
t
0

0
, then
t
= x : (x), (x)
t
0
is a groupoid over
t
0
.
3. Combining (1) and (2), we see that any vector bundle E denes a groupoid (E) over its
zero section.
4. The pair groupoid associated to a set X consists of = X X, endowed with the
multiplication (x, y) (y, z) = (x, z). Thus
0
is the diagonal, and , are the projections
(x, y) = (y, y) and (x, y) = (x, x). In this groupoid, there is exactly one arrow from any
object to another.

We will be interested in groupoids with some geometric structures. Maintaining the notation
above, we make the following denition.
Denition 8.7 A groupoid is called a Lie groupoid if
1.
0
is a submanifold of .
2. The mappings , :
0
are submersions.
3. Multiplication
2
and inversion

are smooth.
Condition (2) implies that the map is transverse to the diagonal in
0

0
, and so
both
2
= ( )
1
() and the multiplication relation m are smooth
submanifolds.
A submanifold L of is called unitary if the restriction of and to L are dieomor-
phisms L
0
. The unitary submanifolds form a group under the natural multiplication of
subsets.
115
Denition 8.8 A Lie groupoid is called a symplectic groupoid if is a symplectic
manifold and the multiplication relation m is a lagrangian submanifold of .
Two immediate consequences of this denition and the calculus of canonical relations are that

0
is lagrangian, and : is anti-symplectic, i.e., its graph is a lagrangian submanifold
of . Thus, a symplectic groupoid is characterized by the three canonical relations
(recall that Z is a point):

0
Hom(Z, ) L

Hom(, ) m Hom( , )
linked by the equation

0
= m L

.
As a consequence of the axioms and the assumption that m is lagrangian, there is a unique
Poisson structure on the base
0
of a symplectic groupoid such that :
0
is a Poisson
map and :
0
is anti-Poisson. This Poisson structure gives meaning to the following
concept.
Denition 8.9 A symplectic groupoid is said to integrate a Poisson manifold P if there
exists a Poisson isomorphism from the base
0
of onto P.
If there exists a groupoid which integrates P, we will say that P is integrable and refer
to as a symplectic groupoid over P.
Example 8.10 1. A simple example of a Lie groupoid is given by any Lie group. From
the requirement that the base of a symplectic groupoid be a lagrangian submanfold, only
discrete Lie groups can be symplectic groupoids.
2. If M is any manifold with the zero Poisson structure, then T

M, equipped with the group-

oid structure of a vector bundle and its standard symplectic structure, denes a symplectic
groupoid over M.
3. If P is any symplectic manifold, then the pair groupoid structure on P P denes a
symplectic groupoid over P. Its unitary lagrangian submanifolds are precisely the graphs of
symplectomorphisms of P.
4. If g is any Lie algebra, then the dual space g

with its Lie-Poisson structure is integrable.

For any Lie group G whose Lie algebra is g, we dene a groupoid structure on T

G by taking
and to be the right and left translations of covectors to the ber at the identity e G..
A simple computation shows that the base of T

G is its ber at the identity, while L

and m
identify with the conormal bundles to the inversion and multiplication relations of G under
the identications T

GT

G T

(GG) T

GT

GT

G T

(GGG). Equipped
with these operations and its usual symplectic structure, T

G is a symplectic groupoid over

g

116
Quantization via symplectic groupoids
From our discussion of geometric quantization, we know that certain symplectic manifolds
quantize to give vector spaces V , and lagrangian submanifolds correspond to elements in V .
If we wish V to be an associative -algebra with unit element, like the algebras of quan-
tum mechanics, then the underlying symplectic manifold must possess a groupoid structure
compatible with its symplectic structure.
The rst step in the quantization of a Poisson manifold P by the method of symplectic
groupoids is to construct a symplectic groupoid (,
0
, , m) over P. This is known as the
integration problem for Poisson manifolds. Using the techniques of geometric quantization
(prequantizations, polarizations), we then attempt to associate a vector space /

to in
such a way that the canonical relations
0
, L

, and m quantize as elements of /

, /

,
and /

which dene the structure of an associative -algebra on /

. If successful,
it is in this sense that the symplectic groupoid represents a classical model for the quantum
algebra /

.
To conclude this chapter, we will give several examples illustrating the spirit of the
symplectic groupoid method. Throughout, we will only deal with groupoids equipped
with reasonable (e.g. brating) polarizations, so that the construction of the vector space
/

follows unambiguously from the procedure dened in the preceding chapter.

Example 8.11 Consider a trivial Poisson manifold M and its associated symplectic group-
oid T

M. As described in the preceding example, the identity relation

0
coincides with
the zero section of T

M; under the identications T

M T

M T

(M M) and T

M
T

MT

M T

(MMM), the relations L

and m identify with the conormal bundles

of the diagonals
2
M M and
3
M M M respectively.
As in Chapter 7, we may identify the quantum Hilbert spaces associated to T

M, T

(M
M), and T

(MMM) with (completions of) the function spaces C

(M), C

(MM) and
C

(MMM). The relation

0
then quantizes as the function 1 on M. Since L

and m are
the conormal bundles of the diagonals in MM and MMM, respectively, our heuristic
discussion in Section 7.3 shows that, after an appropriate normalization, these relations are
quantized by the -functions (x, y) and (x, z)(y, z) supported on the diagonals
2
and

3
respectively. Thus, the quantization of the groupoid yields the usual identity element,
complex conjugation, and pointwise multiplication in the associative -algebra C

(M, C).

Example 8.12 Arguing as in the preceding example, we nd that the quantization of the
canonical relations
0
, L

, and m associated to the groupoid T

G of Example yields the

distributions (e) on G, (g
1
, g
1
1
) on GG, and (g
1
g
2
, g
1
, g
2
) on GGG. If the Haar
measure on G is used to identify the quantum Hilbert space H
G
with C

(G, C), then the

relations
0
, L

and m quantize as evaluation (at e G), anti-involution f(g) f(g

1
), and
convolution.
There is actually a aw in the preceding two examples, since geometric quantizaton
produces half-densities rather than functions, and furthermore, the natural domain of the
117
convolution operation on a group consists not of functions but of densities. It appears then
that the construction of V itself, and not just the multiplication, should depend on the
groupoid structure on .
Although the two multiplications associated with the two groupoid structures on T

G
described above live on dierent spaces, it is possible to relate them more closely by dualizing
one of them, say convolution. In this way, we obtain the coproduct on C

(G) dened as
a map from C

(G) to C

(G) C

(G) = C

(GG) by the formula (f)(x, y) = f(xy).

The coproduct satises a coassociative law and is related to pointwise multiplication by the
simple identity (fg) = (f)(g), making C

(G) into a Hopf algebra. The compatibility

between the two structures on C

(G) reects a compatibility between the two groupoid

structures on T

G.

Example 8.13 If we take P = R

2n
with its standard symplectic structure, there is a po-
larization of P P which does not depend on any polarization of P but only on the ane
structure. In fact, the polarization comes from the isomorphism of = R
2n
R
2n
with
T

R
2n
given by
(x, y) ((x + y)/2,
n
(y x)).
Using this polarization to quantize , we get as V the space of smooth functions on the
diagonal, which we identify with R
2n
itself, and the multiplication on V turns out to be the
Moyal product (see Example 8.4).

Example 8.14 If V is any nite-dimensional vector space, then a skew-adjoint linear map
: V

V denes a skew-symmetric bilinear form on V

and thus a translation-invariant

Poisson structure on V . If T is a torus equal to the quotient of V by some lattice, then
T inherits a translation-invariant Poisson structure from V . A symplectic groupoid which
integrates T is given by the cotangent bundle T

T, with
0
equals the zero section of T

T,
and L

equals the conormal bundle of the diagonal in T T. To describe the multiplication

relation, we identify T

T with T V

and let m T

(T T T) consist of all triples

(q, q
t
, q
tt
, p, p
t
, p
tt
) such that q
tt
= q
t
+
1
2
T(p
t
+ p
tt
) and
q = q
t
+
1
2
T(p
tt
) p = p
t
+ p
tt
.
When the map T is zero, the Poisson structure on T is trivial, and the groupoid product
quantized to the usual pointwise multiplication of functions on T. Otherwise one gets a
noncommutative multiplication on C

(T) which is precisely that of (the functions on) a

noncommutative torus, one of the basic examples of noncommutative geometry. (The
relation between Poisson tori and noncommutative tori was studied from the point of view
of deformation quantization in [50]).

118
A Densities
An n-form on an n-dimensional manifold M can be viewed as a scalar function on the
space of bases in the tangent bundle which satises
(eA) = (e) det(A),
where e = (e
1
, , e
n
) is any frame in a tangent space of M and A = (a
ij
) is any invertible
n n matrix. Because the change of variables formula for integration involves absolute
values of Jacobians, integration of n-forms on M requires a choice of orientation. The use of
densities instead of forms circumvents this need.
A density on a real vector space V of dimension n is a complex-valued function , dened
on the set B(V ) of bases in V , which satises (eA) = (e) [ det(A)[. The collection of such
functions is denoted [[V . This concept can be generalized as follows.
Denition A.1 For C, an -density on V is a map : B(V ) C such that
(eA) = (e) [ det(A)[

.
We denote the vector space of -densities on V by [[

V. Since GL(V ) acts transitively on

B(V ), an -density is determined by its value on a single basis. As a result, [[

V is a
1-dimensional complex vector space.
Operations on densities
1. If [[

V and C, then

is a well-dened -density on V .
2. A linear map T : V V

induces a real-valued 1-density |T| on V given by

|T| e
def
= [ det('Te
i
, e
j
`)[
1/2
.
Equivalently, any real bilinear form on V induces a 1-density | | on V .
3. Multiplication of densities is dened by multiplication of their values and gives rise to a
bilinear map:
[[

V [[

V [[
+
V.
4. If W is a subspace of V , then a basis of V , unique up to transformation by a matrix of
determinant 1, is determined by a choice of bases for W and V/W. Consequently, there is
a natural product
[[

W [[

(V/W) [[

V
In particular, if V = V
1
V
2
, then there is a natural product
[[

V
1
[[

V
2
[[

V.
119
5. Operations 3 and 4 induce natural isomorphisms
[[

V [[

V [[
+
V
[[

W [[

(V/W) [[

V.
[[

V
1
[[

V
2
[[

(V
1
V
2
)
6. If T : V V
t
is an isomorphism, there is a well-dened isomorphism
T

: [[

V [[

V
t
.
7. A natural map B(V )

B(V

) is dened by associating to each basis e of V its dual

basis e

. Since (eA)

= e

(A

)
1
for any A GL(V ), the map gives rise to a natural
isomorphism
[[

V [[

.
Example A.2 Suppose that A, C are vector spaces and
0 A A C
F
C

0
is an exact sequence such that F[
C
= T. By the operations described above, we obtain an
isomorphism
[[

A [[

A [[
2
C
given explicitly by
|T|
2
= [ det

T[

2
,
where is any real-valued 1-density on C, and the positive real number [ det

T[ is dened
by the equation
[ det

T[
1/2
= |T|.

It is easy to check that the association V [[

V denes a dierentiable functor, so we

can associate the -density bundle [[

E to any smooth vector bundle E over a manifold

M. The remarks above imply that if
0 E F G 0
is an exact sequence of vector bundles over M, then there is a natural density-bundle iso-
morphism
[[

E [[

G [[

F.
We denote by [[

E the vector space of smooth sections of [[

E and by [[

c
E the space of
smooth, compactly-supported sections. The density spaces associated to the tangent bundle
of M are denoted [[

M. An n-form on an n-manifold M induces an -density [[

in the
manner of (4) above.
120
A natural mapping [[
1
c
M C is dened by integration

M
.
Similarly, a pre-Hilbert space structure on the space [[
1/2
c
M of smooth, compactly-supported
half-densities on M is dened by
', ` =

M
.
The completion H
M
of this space is called the intrinsic Hilbert space of M.
B The method of stationary phase
We begin with a few useful formulas involving the asymptotic Fourier transform. Let o
denote the usual Schwartz space of rapidly decreasing, complex valued functions on R
n
. For
u o, the asymptotic Fourier transform and its inverse are dened respectively as
(T

u)() = (2)

n
2

R
n
e
i/x,)/
u(x) dx
(T
1

v)(x) = (2)

n
2

(R
n
)

e
i/x,)/
v() d
To see that these transforms are actually inverse to one another, note that the change of
variables = / gives
(2)

n
2

(R
n
)

e
i/x,)/
(T

u)() d = (2)
n

e
i/xy,)
u(y) dy d,
which equals u(x) by the usual Fourier inversion formula. This observation also veries the
asymptotic inversion formula:
u(x) = (2)
n

e
i/xy,)/
u(y) dy d.
A simple application of this formula shows that the asymptotic dierential operator D
j
=
i
j
satises the familiar equations
T

(D
j
u) =
j
T

u T

(x
j
u) = D
j
T

u.
A similar check of denitions proves the asymptotic Parseval formula:

R
n
uv dx =

(R
n
)

uT

v d.
We study next the asymptotic behavior of integrals of the form
I

R
n
e
iR(x)/
a(x) [dx[ a C

0
(R
n
), R C

(R
n
)
as 0. As a rst step, we will prove that if the critical point set of R is not contained in
the support of a, then I

is rapidly decreasing in :
121
Lemma B.1 If dR = 0 on Supp(a), then I

= O(

) as 0.
Proof. Suppose for the moment that R
x
1
= R/x
1
= 0 on Supp(a). Then
e
iR/
a = i
a
R
x
1

x
1
e
iR/
,
and so integration by parts with respect to the x
1
-variable gives
[I

[ =

R
n
e
iR/

x
1

a
R
x
1

[dx[

,
implying that I

is O(). Our assertion follows by noting that (a/Q

x
1
)
x
1
C

0
(R
n
) and
repeating the same argument.
For the general case, we can use a partition of unity to break up Supp(a) into nitely
many domains as above and then applying the same argument (with x
1
possibly replaced by
another coordinate) to each piece.
2
The upshot of this lemma is that the main (asymptotic) contribution to the integral I

must come from the critical points of R.

Lemma B.2 If the quadratic form Q is nondegenerate, then for each nonnegative integer
K,

R
n
e
iQ/
a [dx[ = (2)
n/2
e
i sgn(Q)/4
[ det
[dx[
T[
1/2
K

k=0
1
k!
(D
k
a)(0)
k
+ O

K+1+n/2

,
where T : R
n
(R
n
)

is the self-adjoint map associated to Q and D is the second-order

dierential operator given by
D =
i
2

j,k
T
1
jk

x
j

x
k
.
Proof. From [32, Vol.1, Thm.7.6.1], we recall that for (R
n
)

R
n
e
i/x,)
e
iQ(x)/
dx = (2)
n/2
e
isgn(Q)/4
e
iQ

()
[ det
[dx[
T[
1/2
,
where Q

() = 'T
1
, `/2 and the determinant det
[dx[
T is dened as in Appendix A. Con-
sequently, the asymptotic Fourier transform of the function x e
iQ(x)/
equals

e
isgn(Q)/4
e
Q

()/
[ det
[dx[
T[
1/2
.
Combining this expression with the asymptotic Parseval formula, we obtain
I

R
n
e
iQ/
a [dx[ =
e
isgn(Q)/4
[ det
[dx[
T[
1/2

(R
n
)

e
iQ

/
T

a () d.
122
A simple computation shows that T

a () = T

a (); the change of variables

doesnt aect the integral, and consequently we have
I

=
e
isgn(Q)/4
[ det
[dx[
T[
1/2

(R
n
)

e
iQ

()/
T

a () d.
To deal with the integral on the right, we use the the Taylor series expansion (with remainder)
of e
iQ

/
to write

(R
n
)

e
iQ

/
T

a() d =

k=0
1
k!

(R
n
)

(Q

())
k
T

a() d
= (2)
n/2
K

k=0
1
k!

k
T
1

((Q

)
k
T

a)(0) + O

K+1+
n
2

= (2)
n/2
K

k=0
1
k!
(D
k
a)(0)
k
+ O

K+1+
n
2

,
the last expression on the right following from the asymptotic Fourier inversion formula.
2
We now wish to apply this lemma to evaluate integrals of the form
I

M
e
iR/
,
where M is a smooth n-manifold equipped with a compactly supported density , and
R: M R is a smooth function. To this end, we require the following two lemmas. Recall
that the hessian of R : M R at a critical point p M is a well-dened self-adjoint
linear map R
tt
(p) : T
p
M T

p
M. The critical point p is called nondegenerate if R
tt
(p) is an
isomorphism. In this case, the function R has the following normal form near p.
Morse Lemma . If p is a nondegenerate critical point of a function R : M R, then
there exists a nondegenerate quadratic form Q on R
n
and an embedding g : U M, where
U is a neighborhood of 0 in R
n
, such that g(0) = p and
(R g)(x) = R(p) + Q(x).
2
(This theorem is a special case of the Parametrized Morse Lemma, proven in Section 4.3).
If g : M
t
M is any dieomorphism, then the change of variables formula states that

e
iR/

M
e
iR/
.
The same role in the stationary phase formula will be played by the following lemma.
123
Lemma B.3 In the notation above, suppose that p M is a critical point of R and set
p
t
= g
1
(p). If
t
is any density on M
t
such that
t
p
= (g

)
p
, then
[ det

(R g)
tt
(p
t
)[ = [ det

R
tt
(p)[.
2
Finally, we mention that denition of [ det

(R)
tt
(p)[ implies furthermore that
f(p) [ det
f
R
tt
(p)[
1/2
= [ det

R
tt
(p)[
1/2
for any function f on M. Combining these observations, we obtain
Principle of Stationary Phase . Let M be a smooth n-manifold and [[
c
M. If
R: M R has only nondegenerate critical points p
j
, j = 1, , k in Supp(), then

M
e
iR/
= (2)
n/2
k

j=1
e
iR(p
j
)/
e
isgn(R

(p
j
))/4
[ det

(R
tt
(p
j
))[
1/2
+ O(
1+n/2
).
2
C

Cech cohomology
This appendix will give some of the basic denitions of

Cech cohomology for manifolds and
describe its relation to deRham cohomology. A more general treatment is available in [14].
An open cover U = U

I
of a manifold M is said to be good if every intersection of
nitely many members of U is either contractible or empty. For ease of notation, we will
denote by U

0
..
k
the intersection

k
i=0
U

i
.
If is an abelian group, a -valued

Cech cochain with respect to the cover U is then a
rule which assigns an element c

0
,..,
k
of to every list (
0
, ..,
k
) for which the intersection
U

0
..
k
is nonempty. The group of all such cochains is denoted C
k
U
(M, ), and a coboundary
operator

k
: C
k
U
(M, ) C
k+1
U
(M, )
is dened by

k
(c)(
0
, ..,
k+1
) =
k+1

j=0
(1)
j
c(
0
, ..,
j
, ..,
k+1
).
where the symbol indicates which member of the list is to be deleted. The groups of
degree-k

Cech cocycles and coboundaries are dened respectively by

Z
k
U
(M; ) = ker(
k
)

B
k
U
(M; ) = im(
k1
),
and the k-th

Cech cohomology group of M with coecients in and relative to the covering
U is the quotient

H
k
U
(M; ) =

Z
k
U
(M; )/

B
k
U
(M; ).
124
Now consider a closed k-form Z
k
DR
(M). Since each U

is contractible, there exists

on each U

a (k 1)-form

satisfying d

= . For any indices , , we have

d(

) = 0
on U

. Since U

is itself contractible, there exist (k 2)-forms

dened on U

such
that
d

and
d(

) = 0
on the set U

for any indices , , . Continuing in this way, we see that determines a

Cech k-cocycle with coecients in R. This association denes for each k Z

+
a homomor-
phism
Z
k
DR
(M)
w

H
k
U
(M; R).
Theorem C.1 ([61]) The map w induces an isomorphism between the deRham cohomology
of M and the

Cech cohomology of M with real coecients.
One consequence of this theorem is that

H
k
U
(M; ) does not depend on the choice of U.
A group homomorphism
t
induces a homomorphism

H
k
U
(M; )

H
k
U
(M;
t
)
in the obvious way. Of particular interest in these notes is the subgroup Z

= 2 Z of R.
D Principal T

bundles
In this appendix, we record some standard facts about principal bundles over paracompact
manifolds, referring to [17] for more details. Throughout this section, denote by Z

the group
2 Z and set T

= R/Z
h
.
A principal T

bundle over a manifold P is a locally trivial T

bundle Q

P together
with a nonsingular, ber-preserving action T

Q Q. Two principal T

bundles Q

P
and Q
t

P are said to be isomorphic if there exists a smooth map f : Q Q

t
which is
equivariant with respect to the T

actions, i.e. f(a p) = a f(p) for all a T

and p Q,
and satises =
t
f.
Local triviality of a T

ber bundle Q

P implies that for any good cover U of P,
there exist homeomorphisms h
j
: U
j
T

1
(U
j
) such that h
j
(x, t +s) = s h
j
(x, t) and
(h
j
(x, t)) = x for all (x, t) U
j
T

and s T

. These maps give rise to the transition

functions g
jk
: U
jk
T

of Q, dened by the requirement that

h
j
(x, t) = h
k
(x, t + g
jk
(x))
for all x U
jk
. This equation implies that for each x U
ijk
,
h
i
(x, t) = h
i
(x, t + g
ij
(x) + g
jk
(x) + g
ki
(x)),
125
and so the transition functions satisfy the cocycle condition g
ij
+g
jk
+g
ki
= 0 (mod Z

). If
g
jk
: U
jk
R is any lift of g
jk
, then the numbers
c
ijk
= g
ij
+ g
jk
+ g
ki
are therefore elements of Z

which dene a

Cech cocycle [c
ijk
]. The corresponding class
[Q]

H
2
(P; Z

) is known as the Chern class of Q. The fundamental theorem describing

this space is the following (see [17, Cor.2.1.4] for a proof).
Theorem D.1 Two principal T

bundles over P are isomorphic if and only if their Chern

classes are equal. Moreover, the assignment Q [Q] induces a bijective map from the space
of isomorphism classes of principal T

bundles over P to

H
2
(P; Z

).
Corresponding to the abelian group structure on

H
2
(P; Z

) are the following operations on

principal T

bundles. If Q P is a principal T

bundle having transition functions g

jk

with respect to some good cover of P, then the inverse of Q is dened as the principal
T

bundle Q over P obtained from the transition functions g

jk
. Similarly, if Q, Q
t
are principal T

bundles over P with transition functions g

jk
and g
t
jk
respectively, then
the product of Q and Q
t
is dened as the principal T

bundle Q
P
Q
t
over P having
transition functions g
jk
+g
t
jk
. From the denitions above, it follows easily that the Chern
classes of inverses and products of principal T

bundles are given by [Q] = [Q] and

[Q
P
Q
t
] = [Q] +[Q
t
]. On the level of the bundles themselves, we can describe the product
Q
P
Q
t
as the quotient of the usual ber-product Q
P
Q
t
(which in this case is a T

-
bundle over the base), modulo the anti-diagonal action of T

, i.e. t (p, p
t
) = (t p, t p
t
).
T

bundles with connection

The innitesimal generator of the T

action on a principal bundle Q is a vector eld X on

Q dened by the equation
X(p) =
d
dt

t=0
t p.
A connection on Q is a T

-invariant 1-form on Q such that (X) = 1. In terms of a

good cover U of P, the form satises h

j
= d +

j
, where d denotes the usual form
on T

and the
j
are 1-forms on the U
j
satisfying

k
= d g
jk
,
where g
jk
are again R-valued lifts of the transition functions of Q. The curvature of the
connection is the unique closed 2-form on P such that
d =

.
From the compatibility condition for the
j
, it follows that the Chern class of Q is the

Cech
representative of the deRham cohomology class [].
Theorem D.2 A closed 2-form on a manifold P is the curvature form of a connection
on a principal T

bundle Q over P if and only if ', a` Z

for any a H
2
(P; Z).
126
Proof. Most published proofs of this result (e.g. [37]) use

Cech cohomology and the deRham
isomorphism. We prefer to give the following direct proof by P.Iglesias; see [33] for further
details. Let A(P, p
0
) denote the space of smooth paths : [0, 1] P such that (0) = p
0
, and
let e: A(P, p
0
) P be the endpoint map e() = (1). Since the interval [0, 1] is contractible,
there exists a natural contraction of A(P, p
0
) onto the constant map [0, 1] p
0
. If Y
t
is the
vector eld which generates this contraction, we dene the 1-form K on A(P, p
0
) by
K =

1
0
(Y
t
e

) dt.
Then d + K is a connection form on the product A(P, p
0
) T

having curvature e

.
To complete the proof, we will dene the T

bundle (Q, ) over P with curvature as an

appropriate quotient of A(P, p
0
)T

. For this purpose, we call two elements ,

t
A(P, p
0
)
homologous if their dierence is the boundary of a singular 2-chain in P. The quotient
of A(P, p
0
) by this equivalence relation is the covering space

P of P corresponding to the
commutator subgroup of
1
(P), and thus, H
1
(

P, Z

) = 0. An equivalence relation on the

product A(P, p
0
) T

is then dened by the condition that (, t) (

t
, t
t
) if ,
t
are
homologous and
t t
t
=

,
where =
t
. The quotient of A(P, p
0
) T

by this equivalence relation is a principal

T

bundle

Q over

P with connection having curvature

, where :

P P is the natural
projection.
If H
1
(P; Z

) = 0, then

P = P, and the proof is complete. Otherwise, let s be any map
from H
1
(P; Z

) into the space of loops based at p

0
which assigns a representative to each
homology class. A T

-valued group cocycle on H

1
(P; Z

) is then dened by
(h, h
t
) =

,
where = s(h + h
t
) (s(h) + s(h
t
)) . Since this cocycle is symmetric, it denes a central
extension of H
1
(P; Z

) which acts naturally on

Q by
(h, )[, z] = [s(h) , z + ].
Since T

is divisible, the extension is isomorphic to the product T

H
1
(P; Z

), and any
choice of isomorphism denes an action of H
1
(P; Z

) on

Q which preserves the connection
. The quotient of

Q is the desired principal T

bundle Q over P.
2
A connection on a principal T

bundle Q

P induces a connection on the inverse Q
of Q whose local representatives are of the form d
j
, where is locally represented by
d+
j
. Similarly, connections and
t
on the T

bundles Q, Q
t
over P induce a connection
+
t
on the product Q
P
Q
t
dened locally by d +
j
+
t
j
.
127
Two principal T

bundles with connection (Q, ), (Q

t
,
t
) over P are said to be isomorphic
provided that there exists an isomorphism f : Q Q
t
of the underlying principal bundles
which satises f

t
= . A check of the denition shows that (Q, ), (Q
t
,
t
) are isomorphic
if and only if (Q
P
Q
t
,
t
) is isomorphic to the trivial bundle (P T

, dS). Since
curvature is obviously invariant under isomorphisms of principal T

bundles with connection,

the classication of such bundles reduces to the classication of at connections on trivial
T

bundles over P.
A connection on a principal T

bundle Q

P is said to be at if d = 0. In this
case, the local representatives h

j
= d +
j
have the property that
j
= d s
j
for functions
s
j
: U
j
R. Denoting by s
j
the composition of s
j
with the projection R T

, we nd
that the functions s
jk
= s
j
s
k
dene transition functions for a trivial T

principal bundle
over P. On the other hand, the compatibility condition for the
j
implies that s
jk
g
jk
is
constant for each j, k, and so there exists an isomorphism f : Q P T

. It is easy to
check that if
0
is the trivial connection on P T

, then
f

0
=

for some closed 1-form on P. The cohomology class [] induced by this form in

H
1
(P; T

)
is called the holonomy of the (at) connection .
Theorem D.3 Two at connections ,
t
on the trivial principal T

bundle Q = P T

over
P are isomorphic if and only if they have equal holonomy. Moreover, the map (Q, ) [

]
induces a bijection from the space of isomorphism classes of at connections on the trivial
T

bundle with

H
1
(P; T

).
A section s of a principal T

bundle Q over P with connection is called parallel provided

that s

= 0. The map s
Q
: Q Q associated to a parallel section denes an isomorphism
of (Q, ) with the trivial T

bundle equipped with the trivial connection = d. Thus, we

have:
Corollary D.4 A principal T

bundle Q over P with a at connection admits a parallel

section if and only if (Q, ) has zero holonomy.
2
Associated line bundles
A representation : T

U(1) enables us to associate to any principal T

bundle Q P a
complex line bundle E P dened explicitly as the quotient of QC by the T

action
t (p, z) = (t p,
1
(t)z).
The space of functions g : Q C satisfying the condition
g(t p) =
1
(t)g(p)
128
is identied with the space of sections of E by the assignment g s
G
, where the section
s
g
: P E is dened by the requirement that
s
g
(x) = [(p, g(p))],
for any element p of
1
(x).
129
References
[1] Abraham, R., and Marsden, J.E., Foundations of Mechanics, 2nd edition, Ben-
jamin/Cummings, Reading, 1978.
[2] Arnold, V.I., Mathematical Methods of Classical Mechanics, Graduate Texts in
Math. 60, Springer-Verlag, New York, 1978.
[3] Arnold, V.I., On a characteristic class entering into conditions of quantization,
Funct. Anal. Appl. 1 (1967) 1-13.
[4] Arnold, V.I., Catastrophe theory (3rd edition), Springer-Verlag, New York, 1992.
[5] Arnold, V.I., Symplectic geometry and topology, preprint 1993.
[6] Arnold, V.I., and Givental, A.B., Symplectic geometry, Dynamical systems IV,
V.I. Arnold and S.P. Novikov, eds., Encyclopedia of Mathematical Sciences, Vol. 4,
Springer-Verlag, New York, 1990, pp. 1-136.
[7] Atiyah, M.F., The Geometry and Physics of Knots, Cambridge Univ. Press, Cam-
bridge, 1990.
[8] Audin,M., The Topology of Torus Actions on Symplectic Manifolds, Progress in
Math. vol. 93, Birkhauser, Basel, 1991.
[9] Bayen, F., Flato, M., Fronsdal, C., Lichnerowicz, A., and Sternheimer, D., Defor-
mation theory and quantization, I and II, Ann. Phys. 111, (1977), 61-151.
[10] Berezin, F.A., Some remarks about the associated envelope of a Lie algebra, Funct.
Anal. Appl. 1, (1967), 91-102.
[11] Berezin, F.A., Quantization, Math USSR Izv. 8 (1974), 1109-1165.
[12] Besse, A. Manifolds All of Whose Geodesics are Closed, Ergebnisse der Math. 93
Springer-Verlag, Berlin, 1978.
[13] Bohm, D., Quantum Theory, Prentice-Hall, New York, 1951.
[14] Bott, R. and Tu, L, Dierential Forms in Algebraic Topology, Springer-Verlag, New
York, 1982.
[15] Boutet de Monvel, L., and Guillemin, V., The Spectral Theory of Toeplitz Operators,
Annals. of Math. Studies 99, Princeton University Press, Princeton, 1981.
[16] Br uning, J. and Guillemin, V, Mathematics Past and Present: Fourier Integral Op-
erators, Springer-Verlag, Berlin, 1994.
[17] Brylinski, J.-L., Loop Spaces, Characteristic Classes, and Geometric Quantization,
Progress in Mathematics vol. 107, Birkhauser, Boston, 1992.
130
[18] Connes, A., Noncommutative Geometry, Academic Press, San Diego, 1995.
[19] De Wilde, M., and Lecomte, P., Formal deformations of the Poisson Lie algebra
of a symplectic manifold and star-products. Existence, equivalence, derivations, in
M. Hazewinkel and M. Gerstenhaber, eds., Deformation Theory of Algebras and
Structures and Applications, Kluwer Acad. Pub., Dordrecht (1988), 897-960.
[20] Dirac, P. A. M., The Principles of Quantum Mechanics, Clarenden Press, Oxford,
1930.
[21] Duistermaat, J. J., Fourier Integral Operators, Courant Institute, NYU, New York,
1973. (New edition to be published by Birkhauser Boston in 1995.)
[22] Ehresmann, C., Les connexions innitesimales dans un espace bre dierentiable,
Colloque de Topologie (Espaces Fibres) tenu a Bruxelles du 5 au 8 juin, 1950,
Georges Thone, Li`ege and Masson et Cie., Paris, 1951, pp. 29-55.
[23] Fedosov, B., A simple geometrical construction of deformation quantization, J. Di.
Geom. 40 (1994) 213-238.
[24] Folland, G.B., Harmonic Analysis in Phase Space, Annals of mathematics studies
122, Princeton University Press, 1989.
[25] Fomenko, A.T., Symplectic Geometry, Gordon and Breach, New York, 1988.
[26] Golubitsky, M., Guillemin, V., Stable Mappings and Their Singularities, Graduate
Texts in Math., vol. 14, Springer-Verlag, (1973).
[27] Gompf, R., New constructions of symplectic manifolds, Ann. Math., to appear.
[28] Guillemin, V., and Sternberg, S., Geometric Asymptotics, Math. Surveys 14, Amer.
Math. Soc., Providence, 1977.
[29] Guillemin, V., and Sternberg, S., Symplectic Techniques in Physics, Cambridge Uni-
versity Press, Cambridge, 1984.
[30] Heler, B., Semi-classical analysis for the Schrodinger operator and applications,
Lecture Notes in Math. 1336 (1988).
[31] Hormander, L., Fourier integral operators I., Acta Math. 127 (1971) 79-183.
[32] Hormander, L., The Analysis of Linear Partial Dierential Operators, Vols. 1-4,
Springer-Verlag, New York, 1983-90.
[33] Iglesias, P., La trilogie du moment, Ann. Inst. Fourier (to appear).
[34] Karasev, M.V. and Maslov, V.P., Nonlinear Poisson Brackets : Geometry and Quan-
tization Translations of mathematical monographs vol. 119, American Mathematical
Society, Providence, R.I., 1993.
131
[35] Kirillov, Geometric quantization, Dynamical Systems IV, V.I. Arnold and S.P.
Novikov, eds., Encyclopedia of Mathematical Sciences, Vol. 4, Springer-Verlag, New
York, 1990, pp. 137-172.
[36] Klingenberg, W., Riemannian Geometry, deGruyter Studies in Math., Berlin, 1982.
[37] Kostant, B., Quantization and unitary representations, Lecture Notes in Math. 170
(1970), 87-208.
[38] Libermann, P., and Marle, C.-M., Symplectic Geometry and Analytical Mechanics,
Reidel, Dordrecht, 1987.
[39] Lie, S., Theorie der Transformationsgruppen, (Zweiter Abschnitt, unter Mitwirkung
von Prof. Dr. Friedrich Engel), Leipzig, Teubner, 1890.
[40] McDu, D., Examples of symplectic structures, Invent. Math. 89 (1987), 13-36.
[41] Mackey, G., Mathematical Foundations of Quantum Mechanics, W.A.Benjamin,
Reading, 1963.
[42] Marsden, J., Lectures on Mechanics, London Math. Soc. Lecture Note Series 174,
Cambridge Univ. Press, Cambridge, 1992.
[43] Maslov, V.P., Theory of Perturbations and Asymptotic Methods, Moskov. Gos. Univ.,
Moscow, 1965.
[44] Melotte, D., Invariant deformations of the Poisson Lie algebra of a symplectic man-
ifold and star products, in M. Hazewinkel and M. Gerstenhaber, eds., Deformation
Theory of Algebras and Structures and Applications, Kluwer Acad. Pub., Dordrecht
(1988), 961-972.
[45] Moser, J., On the volume elements on a manifold, Trans. Amer. Math. Soc. 120
(1965), 280-296.
[46] Nazaikinskii, V., Sternin, B., and Shatalov, V., Methods of Noncommutative Analy-
sis, Theory and Applications, Berlin, New York, W. de Gruyter (to appear).
[47] Olver, F.W.J., Asymptotics and Special Functions, Academic Press, New York, 1974.
[48] Omohundro, S., Geometric Perturbation Theory in Physics, World Scientic, Singa-
pore, 1986.
[49] Omori, I., Innite-dimensional Lie Transformation Groups, Lecture Notes in Math-
ematics, vol.427, Springer-Verlag, New York, 1975.
[50] Rieel, M., Deformation quantization of Heisenberg manifolds, Comm. Math. Phys.
122 (1989), 531-562.
132
[51] Robert, D., Autour de lApproximation Semi-classique, Progress in Mathematics,
vol. 68, Birkhauser, Boston, 1987.
[52] Segal, I., Quantization of nonlinear systems, J. Math. Phys. 1 (1960), 468-488.
[53] Sniatycki, J., Geometric Quantization and Quantum Mechanics, Springer-Verlag,
New York, 1980.
[54] Souriau, J.M., Structure des Syst`emes Dynamiques, Dunod, Paris, 1970.
[55] Souriau, J.M., Groupes dierentiels, Lecture Notes in Math. 836 (1980), 91-128.
[56] Treves, F., Introduction to Pseudodierential and Fourier Integral Operators.. Vols
I,II. Plenum Press, New York, 1980.
[57] Tulczyjew, W.M., The Legendre transform, Ann. Inst. H. Poincare, 27 (1977) 101-
114.
[58] Unterberger, A., Quantication et analyse pseudo-dierentielle, Ann. Sci.

Ec. Norm.
Sup. 21 (1988), 133-158.
[59] Vaisman, I., Symplectic Geometry and Secondary Characteristic Classes, Birkh auser,
Boston, 1987.
[60] Wallach, N. R., Symplectic Geometry and Fourier Analysis, Math. Sci. Press, Brook-
line, 1977.
[61] Weil, A., Sur les theor`emes de de Rham, Comm. Math. Helv. 26 (1952) 119-145.
[62] Weinstein, A., Symplectic manifolds and their lagrangian submanifolds, Advances
in Math. 6 (1971), 329-346.
[63] Weinstein, A., Lectures on Symplectic Manifolds, Regional conference series in math-
ematics 29, Amer. Math. Soc., Providence, 1977.
[64] Weinstein, A., Symplectic geometry, Bulletin A.M.S. (new series) 5 (1981), 113.
[65] Weinstein, A., Coisotropic calculus and Poisson groupoids, J. Math. Soc. Japan 40
(1988), 705-727.
[66] Weinstein, A., Noncommutative geometry and geometric quantization, in Symplectic
Geometry and Mathematical Physics: actes du colloque in lhonneur de Jean-Marie
Souriau, P.Donato et al. eds., Birkhauser, Basel, 1991.
[67] Weinstein, A., Symplectic groupoids, geometric quantization, and irrational rotation
algebras, in Symplectic Geometry, Groupoids, and Integrable Systems, P.Dazord and
A.Weinstein, eds., MSRI Publications, Springer, New York, 1991.
[68] Weinstein, A., Deformation quantization, Seminaire Bourbaki, no. 789, Juin 1994.
133
[69] Weinstein, A., and Xu, P., Extensions of symplectic groupoids and quantization, J.
Reine Angew. Math. 417 (1991), 159-189
[70] Weyl, H., The Theory of Groups and Quantum Mechanics, Dover Publications, New
York, 1950.
[71] Woodhouse, N.M.J., Geometric Quantization, 2nd Ed., Oxford U. Press, Oxford,
1992.
[72] Yano, K., and Ishihara, S., Tangent and Cotangent Bundles; Dierential Geometry
Dekker, New York, 1973.
134