Geometrical Formulation of Quantum Mechanics

Abhay Ashtekar1,2 and Troy A. Schilling1,3

1Center for Gravitational Physics and Geometry
Department of Physics, Penn State, University Park, PA 16802-6300, USA
2 Erwin Schrödinger International Institute for Mathematical Physics

Boltzmanngasse 9, A-1090 Vienna, Austria

3 Institute for Defense Analyses

1801 North Beauregard Street, Alexandria, VA 22311-1772

arXiv:gr-qc/9706069v1 23 Jun 1997

Abstract

States of a quantum mechanical system are represented by rays in a com-

plex Hilbert space. The space of rays has, naturally, the structure of a Kähler
manifold. This leads to a geometrical formulation of the postulates of quan-
tum mechanics which, although equivalent to the standard algebraic formula-
tion, has a very different appearance. In particular, states are now represented
by points of a symplectic manifold (which happens to have, in addition, a
compatible Riemannian metric), observables are represented by certain real-
valued functions on this space and the Schrödinger evolution is captured by
the symplectic flow generated by a Hamiltonian function. There is thus a re-
markable similarity with the standard symplectic formulation of classical me-
chanics. Features—such as uncertainties and state vector reductions—which
are specific to quantum mechanics can also be formulated geometrically but
now refer to the Riemannian metric—a structure which is absent in classical
mechanics. The geometrical formulation sheds considerable light on a num-
ber of issues such as the second quantization procedure, the role of coherent
states in semi-classical considerations and the WKB approximation. More
importantly, it suggests generalizations of quantum mechanics. The simplest
among these are equivalent to the dynamical generalizations that have ap-
peared in the literature. The geometrical reformulation provides a unified
framework to discuss these and to correct a misconception. Finally, it also
suggests directions in which more radical generalizations may be found.

I. INTRODUCTION

Quantum mechanics is probably the most successful scientific theory ever invented. It
has an astonishing range of applications—from quarks and leptons to neutron stars and
white dwarfs—and the accuracy with which its underlying ideas have been tested is equally
impressive. Yet, from its very inception, prominent physicists have expressed deep reser-
vations about its conceptual foundations and leading figures continue to argue that it is

1
incomplete in its core. Time and again, attempts have been made to extend it in a non-
trivial fashion. Some of these proposals have been phenomenological (see, e.g., [1–3]), aimed
at providing a ‘mechanism’ for the state reduction process. Some have been more radical,
e.g., invoking hidden variables (see, e.g., [4]). Yet others involve non-linear generalizations
of the Schrödinger equation [5–7]. Deep discomfort has been expressed at the tension be-
tween objective descriptions of happenings provided by the space-time geometry of special
relativity and the quantum measurement theory [8,9]. Further conceptual issues arise when
one brings general relativity in to picture, issues that go under the heading of ‘problem of
time’ in quantum gravity [10–12]. Thus, while there is universal agreement that quantum
mechanics is an astonishingly powerful working tool, in the ‘foundation of physics circles’
there has also been a strong sentiment that sooner or later one would be forced to generalize
it in a profound fashion [13–15].
It is often the case that while an existing theory admits a number of equivalent descrip-
tions, one of them suggests generalizations more readily than others. Furthermore, typically,
this description is not the most familiar one, i.e., not the one that seems simplest from the
limited perspective of the existing theory. An example is provided by Cartan’s formulation
of Newtonian gravity. While it played no role in the invention of the theory (it came some
two and a half centuries later!) at a conceptual level, Cartan’s framework provides a deeper
understanding of Newtonian gravity and its relation to general relativity. A much more
striking example is Minkowski’s geometric reformulation of special relativity. His emphasis
on hyperbolic geometry seemed abstract and abstruse at first; at the time, Einstein himself
is said to have remarked that it made the subject incomprehensible to physicists. Yet, it
proved to be an essential stepping stone to general relativity.
The purpose of this article is to present, in this spirit, a reformulation of the math-
ematical framework underlying standard quantum mechanics (and quantum field theory).
The strength of the framework is that it is extremely natural from a geometric perspective
and succinctly illuminates the essential difference between classical and quantum mechan-
ics. It has already clarified certain issues related to the second quantization procedure and
semi-classical approximations [16]. It also serves to unify in a coherent fashion a number of
proposed generalizations of quantum mechanics; in particular, we will see that generaliza-
tions that were believed to be distinct (and even incompatible) are in fact closely related.
More importantly, this reformulation may well lead to viable generalizations of quantum
mechanics which are more profound than the ones considered so far. Finally, our experience
from seminars and discussions has shown that ideas underlying this reformulation lie close to
the heart of geometrically oriented physicists. It is therefore surprising that the framework
is not widely known among relativists. We are particularly happy to be able to rectify this
situation in this volume honoring Engelbert and hope that the role played by the Kähler
geometry, in particular, will delight him.
Let us begin by comparing the standard frameworks underlying classical and quantum
mechanics. The classical description is geometrical: States are represented by points of a
symplectic manifold Γ, the phase space. The space of observables consists of the (smooth)
real-valued functions on this manifold. The (ideal) measurement of an observable f in a
state p ∈ Γ yields simply the value f (p) at the point p; the state is left undisturbed. These
outcomes occur with complete certainty. The space of observables is naturally endowed
with the structure of a commutative, associative algebra, the product being given simply by

2
pointwise multiplication. Thanks to the symplectic structure, it also inherits a Lie-bracket—
the Poisson bracket. Finally to each observable f is associated a vector field Xf called the
Hamiltonian vector field of f . Thus, each observable generates a flow on Γ. Dynamics
is determined by a preferred observable, the Hamiltonian H; the flow generated by XH
describes the time evolution of the system.
The arena for quantum mechanics, on the other hand, is a Hilbert space H. States of
the system now correspond to rays in H, and the observables are represented by self-adjoint
linear operators on H. As in the classical description, the space of observables is a real vector
space equipped with with two algebraic structures. First, we have the the Jordan product—
i.e., the anti-commutator—which is commutative but now fails to be associative. Second, we
have (1/2i times) the commutator bracket which endows the space of observables with the
structure of a Lie algebra. Measurement theory, on the other hand, is strikingly different. In
the textbook description based on the Copenhagen interpretation, the (ideal) measurement
of an observable  in a state Ψ ∈ H yields an eigenvalue of  and, immediately after the
measurement, the state is thrown into the corresponding eigenstate.specific outcome can
only be predicted probabilistically. As in the classical theory, each observable  gives rise to
a flow on the state space. But now, the flow is generated by the 1-parameter group exp iÂt
and respects the linearity of H. Dynamics is again dictated by a preferred observable, the
Hamiltonian operator Ĥ.
Clearly, the two descriptions have several points in common. However, there is also a
striking difference: While the classical framework is geometric and non-linear, the quantum
description is intrinsically algebraic and linear. Indeed, the emphasis on the underlying
linearity is so strong that none of the standard textbook postulates of quantum mechanics
can be stated without reference to the linear structure of H.
From a general perspective, this difference seems quite surprising. For linear structures
in physics generally arise as approximations to the more accurate non-linear ones. Thus,
for example, we often encounter non-linear equations which correctly capture a physical
situation. But, typically, they are technically difficult to work with and we probe properties
of their solutions through linearization. In the present context, on the other hand, it is the
deeper, more correct theory that is linear and the non-linear, geometric, classical framework
is to arise as a suitable limiting case.
However, deeper reflection shows that quantum mechanics is in fact not as linear as it is
advertised to be. For, the space of physical states is not the Hilbert space H but the space
of rays in it, i.e., the projective Hilbert space P. And P is a genuine, non-linear manifold.
Furthermore, it turns out that the Hermitian inner-product of H naturally endows P with
the structure of a Kähler manifold. Thus, in particular, like the classical state space Γ, the
correct space of quantum states, P, is a symplectic manifold! We will therefore refer to P
as the quantum phase space. Given any self-adjoint operator Ĥ, we can take its expectation
value to obtain a real function on H. It is easy to verify that this function admits an
unambiguous projection h to the projective Hilbert space P. Recall, now, that every phase
space function gives rise to a flow through its Hamiltonian vector field. What then is the
interpretation of the flow Xh ? It turns out [17] to be exactly the (projection to P of the)
flow defined by the Schrödinger equation (on H) of the quantum theory. Thus, Schrödinger
evolution is precisely the Hamiltonian flow on the quantum phase space!
As we will see, the interplay between the classical and quantum ideas stretches much

3
further. The overall picture can be summarized as follows. Classical phase spaces Γ are, in
general, equipped only with a symplectic structure. Quantum phase spaces, P, on the other
hand, come with an additional structure, the Riemannian metric provided by the Kähler
structure. Roughly speaking, features of quantum mechanics which have direct classical
analogues refer only to the symplectic structure. On the other hand, features—such as
quantum uncertainties and state vector reduction in a measurement process—refer also the
Riemannian metric. This neat division lies at the heart of the structural similarities and
differences between the (mathematical frameworks underlying the) two theories.
Section II summarizes this geometrical reformulation of standard quantum mechanics.
We begin in II A by showing that the quantum Hilbert space can be regarded as a (lin-
ear) Kähler space and discuss the roles played by the symplectic structure and the Kähler
metric. In II B, we show that one can naturally arrive at the quantum state space P by
using the Bergmann-Dirac theory of constrained systems. (This method of constructing
the quantum phase space will turn out to be especially convenient in section III while an-
alyzing the relation between various generalizations of quantum mechanics.) Section II C
provides a self-contained treatment of the various issues related to observables—associated
algebraic structures, quantum uncertainty relations and measurement theory—in an intrin-
sically geometric fashion. These results are collected in section II D to obtain a geometric
formulation of the postulates of quantum mechanics, a formulation that makes no reference
to the Hilbert space H or the associated linear structures. In practical applications, except
while dealing with simple cases such as spin systems, the underlying quantum phase space
P is infinite-dimensional (since it comes from an infinite-dimensional Hilbert space H). Our
mathematical discussion encompasses this case. Also, in the discussion of measurement
theory, we allow for the possibility that observables may have continuous spectra.
In section III, we consider possible generalizations of quantum mechanics. These general-
izations can occur in two distinct ways. First, we can retain the original kinematic structure
but allow more general dynamics, e.g., by replacing the Schrödinger equation by a suitable
non-linear one (see, e.g., [5]). In section III A we show that the geometrical reformulation of
section II naturally suggests a class of such extensions which encompasses those proposed by
Birula and Mycielski [5] and by Weinberg [7]. In section III B we consider the possibility of
more radical extensions in which the kinematical set up itself is changed. Although (to our
knowledge) there do not exist interesting proposals of this type, such generalizations would
be much more interesting. In particular, it is sometimes argued that the linear structure
underlying quantum mechanics would have to be sacrificed in a subtle but essential way
to obtain a satisfactory quantum theory of gravity and/or to cope satisfactorily with the
‘measurement problem’ [13]. To implement such ideas, the underlying kinematic structure
will have to be altered. A first step in this direction is to obtain a useful characterization
of the kinematical framework of standard quantum mechanics. Section III B provides a
‘reconstruction theorem’ which singles out quantum mechanics from its plausible general-
izations. The theorem provides powerful guidelines: it spells out directions along which one
can proceed to obtain a genuine extension.
Section IV is devoted to semi-classical issues. Consider a simple mechanical system, such
as a particle in R3 . In this case, the classical phase space Γ is six-dimensional while the
quantum phase space P is infinite-dimensional. Is there a relation between the two? In sec-
tion IV A we show that the answer is in the affirmative: P is a bundle over Γ. Furthermore,

4
the bundle is trivial. Thus, through each quantum state p ∈ P, there is a cross-section,
i.e. a copy of Γ. It turns out that the quantum states that lie on any one cross-section
are precisely the generalized coherent states [18,20,21]. In the remainder of section IV, we
use this interplay between P and Γ to discuss the relation between classical and quantum
dynamics. Section IV B is devoted to the correspondence in terms of Ehrenfest’s theorem
while IV C discusses the problem along the lines of the WKB approximation. Somewhat
interestingly, it turns out that WKB dynamics is an example of generalized dynamics of the
Weinberg type [7].
Our conventions are as follows. If the manifold under consideration is infinite-
dimensional, we will assume that it is a Hilbert manifold. (Projective Hilbert spaces are
naturally endowed with this structure; see, e.g., [16].) Riemannian metrics and symplectic
structures on these manifolds will be assumed to be everywhere defined, smooth, strongly
non-degenerate fields. (Thus, they define isomorphisms between the tangent and cotangent
spaces at each point). In detailed calculations we will often use the abstract index notation
of Penrose’s [22,23]. Note that, in spite of the appearance of indices, this notation is well-
defined also on infinite-dimensional manifolds. (For example, if V a denotes a contravariant
vector field; the subscript a does not refer to its components but is only a label telling us
that V is a specific type of tensor field, namely a contravariant vector field. Similarly, V a ωa
is the function obtained by the action of the 1-form ω on the contravariant vector field V .)
Finally, due to space limitation, we have not included detailed proofs of several technical
assertions; they can be found in [16]. Our aim here is only to provide a thorough overview
of the subject.
This work was intended to be an extension of a paper by Kibble [17] which pointed out
that the Schrödinger evolution can be regarded as an Hamiltonian flow on H. However, after
completing this work, we learned that many of the results contained in sections II and III B
were obtained independently by others (although the viewpoints and technical proofs are
often distinct.) In 1985, Heslot [24] observed that quantum mechanics admits a symplectic
formulation in which the phase space is the projective Hilbert space. That discussion was,
however, restricted to the finite-dimensional case and did not include a discussion of the
role of the metric, probabilistic interpretation and quantum uncertainties. Anandan and
Aharonov [25] rediscovered some of these results and also discussed some of the probabilistic
aspects. This work was also restricted to finite-dimensional systems and focussed on the
issue of evolution. Similar observations were made by Gibbons [26] who also discussed
density matrices (which are not considered here) and raised the issue of characterization of
quantum mechanics (which is resolved in section IV B). An essentially complete treatment
of the finite-dimensional case was given by Hughston [27]. (This work was done in parallel to
ours. However, it also contains some proposals for mechanisms for state reduction [28] which
are not discussed here.) The only references (to our knowledge) which treat the infinite-
dimensional case are [29,30], which also discuss the issue of characterization of standard
quantum mechanics. Finally, since the geometric structures that arise here are so natural,
it is quite possible that they were independently discovered by other authors that we are
not aware of.

5
 II. GEOMETRIC FORMULATION OF QUANTUM MECHANICS

The goal of this section is to show that quantum mechanics can be formulated in an
intrinsically geometric fashion, without any reference to a Hilbert space or the associated
linear structure. We will assume that the reader is familiar with basic symplectic geometry.

A. The Hilbert space as a Kähler space

Let us begin with the standard Hilbert space formulation of quantum mechanics. In
this sub-section we will view the Hilbert space H as a Kähler space and examine the role
played by the associated symplectic structure and the Riemannian metric. This discussion
will serve as a stepping stone to the analysis of the quantum phase space P in section II B.
The similarities between classical and quantum mechanics can be put in a much more
suggestive form with an alternative, but equivalent, description of the Hilbert space. We
view H as a real vector space equipped with a complex structure J. The complex structure
is a preferred linear operator which represents multiplication by i; hence J 2 = −I. Initially,
this change of notation seems rather trivial; the element which is typically written (a + ib)Ψ
is now denoted aΨ + bJΨ and (external) multiplication of vectors by complex numbers is not
permitted. However, this slight change of viewpoint will come with a reward—a symplectic
formulation of quantum mechanics.
Since H is now viewed as a real vector space, the Hermitian inner-product is slightly
unnatural. We therefore decompose it into real and imaginary parts,
1 i
hΦ, Ψi =: G(Φ, Ψ) + Ω(Φ, Ψ). (2.1)
2h̄ 2h̄
(The reason for the factors of 1/2h̄ will become clear shortly.) Properties of the Hermitian
inner-product imply that G is a positive definite, real inner-product and that Ω is a sym-
plectic form, both of which are strongly non-degenerate. Moreover, since hΦ, JΨi = ihΦ, Ψi,
one immediately observes that the metric, symplectic structure and complex structure are
related as

G(Φ, Ψ) = Ω(Φ, JΨ). (2.2)

That is, the triple (J, G, Ω) equips H with the structure of a Kähler space. Therefore, every
Hilbert space may be naturally viewed as a Kähler space.
Next, by use of the canonical identification of the tangent space (at any point of H) with
H itself, Ω is naturally extended to a strongly non-degenerate, closed, differential 2-form
H, which we will denote also by Ω. Any Hilbert space is therefore naturally viewed as the
simplest sort of symplectic manifold, i.e., a phase space. The inverse of Ω may be used to
define Poisson brackets and Hamiltonian vector fields. As we are about to see, these notions
are just as relevant in quantum mechanics as in classical mechanics.

6
 1. The symplectic form

In classical mechanics, observables are real-valued functions, and to each such function
is associated a corresponding Hamiltonian vector field. In quantum mechanics, on the other
hand, the observables themselves may be viewed as vector fields, since linear operators
associate a vector to each element of the Hilbert space. However, the Schrödinger equation,
which in our language is written as Ψ̇ = − h̄1 J ĤΨ, motivates us to associate to each quantum
observable F̂ the vector field
1
YF̂ (Ψ) := − J F̂ Ψ. (2.3)
h̄
This Schrödinger vector field is defined so that the time-evolution of the system corresponds
to the flow along the Schrödinger vector field associated to the Hamiltonian operator. (Note
that, if the Hamiltonian is unbounded, it is only densely defined and so is the vector field.
The (unitary) flow, however, is defined on all of H. See, e.g., [31].)
Natural questions immediately arise. Let F̂ be any bounded, self-adjoint operator on
H. Is the corresponding vector field, YF̂ , Hamiltonian on the symplectic space (H, Ω)? If
so, what is the real-valued function which generates this vector field? What is the physical
meaning of the Poisson bracket? In particular, how is it related to the commutator Lie
algebra?
The answers to these questions are remarkably simple. As we know from standard
quantum mechanics, F̂ generates a one-parameter family of unitary mappings on H. By
definition, YF̂ is the generator of this one-parameter family and therefore preserves both the
metric G and symplectic form Ω. It is therefore locally Hamiltonian, and, since H is a linear
space, also globally Hamiltonian! In fact, the function which generates this Hamiltonian
vector field is of physical interest; it is simply the expectation value of F̂ .
Let us see this explicitly. Denote by F : H −→ R the expectation value function,
1
F (Ψ) := hΨ, F̂ Ψi = G(Ψ, F̂ Ψ). (2.4)
2h̄
We will continue to use this notation; expectation value functions will be denoted by simply
“un-hatting” the corresponding operators. Now, if η is any tangent vector at Ψ, then1
d
(dF )(η) = dt
hΨ + tη F̂ (Ψ + tη)i|t=0 = hΨ, F̂ ηi + hη, F̂ Ψi
1
= h̄
G(F̂ Ψ, η) = Ω(YF̂ , η) = (iYF̂ Ω)(η), (2.5)

where we have used the self-adjointness of F̂ , Eq. (2.2) and the definition of YF̂ . Therefore,
the Hamiltonian vector field XF generated by the expectation value function F coincides
with the Schrödinger vector field YF̂ associated to F̂ . As a particular consequence, the

1 With our conventions for symplectic geometry, the Hamiltonian vector field Xf generated by
the function f satisfies the equation iXf Ω = df , and the Poisson bracket is defined by {f, g} =
Ω(Xf , Xg ).

7
time evolution of any quantum mechanical system may be written in terms of Hamilton’s
equation of classical mechanics; the Hamiltonian function is simply the expectation value of
the Hamiltonian operator. Schrödinger’s equation is Hamilton’s equation in disguise!
Next, let F̂ and K̂ be two quantum observables, and denote by F and K the respective
expectation value functions. It is natural to ask whether the Poisson bracket of F and K
is related in a simple manner to an algebraic operation involving the original operators.
Performing a calculation as simple as Eq. (2.5), one finds that
1
 
{F, K}Ω = Ω(XF , XK ) = [F̂ , K̂] . (2.6)
ih̄
Notice that the quantity inside the brackets on the right side of Eq. (2.6) is precisely the
quantum Lie bracket of F̂ and K̂. The algebraic operation on the expectation value func-
tions, which is induced by the commutator bracket is exactly a Poisson bracket! Note that
this is not Dirac’s correspondence principle; the Poisson bracket here is the quantum one,
determined by the imaginary part of the Hermitian inner-product.
The basic features of the classical formalism appear also in quantum mechanics. The
Hilbert space, as a real vector space, is equipped with a symplectic form. To each quantum
observable is associated a real-valued function on H, and the time-evolution is determined by
the Hamiltonian vector field associated to a preferred function. Moreover, the Lie bracket of
two quantum observables corresponds precisely to the Poisson bracket of the corresponding
functions.

2. Uncertainty and the real inner-product

Let us now examine the role played by the metric G. Clearly, G enables us to define a
real inner-product, G(XF , XK ) between any two Hamiltonian vector fields XF and XK . One
may expect that this inner-product is related to the Jordan product in much the same way
that the symplectic form corresponds to the commutator Lie bracket. It is easy to verify
that this expectation is correct. Operating just as in Eq. (2.6), we obtain
h̄ 1
 
{F, K}+ := G(XF , XK ) = [F̂ , K̂]+ . (2.7)
2 2
The operation defined by the first equation above will be called the Riemann bracket of
F and K. Up to the factor of h̄/2, the Riemann bracket of F and K is simply given by
the (real) inner-product of their Hamiltonian vector fields, and corresponds precisely to the
Jordan product of the respective operators.
Since the classical phase space is, in general, not equipped with a Riemannian metric,
the Riemann product does not have an analogue in the classical formalism; it does, however,
admit a physical interpretation. In order to see this, note that the uncertainty of the
observable F̂ at a state with unit norm is given by
(∆F̂ )2 = hF̂ 2 i − hF̂ i2 = {F, F }+ − F 2 . (2.8)
Thus, the uncertainty of an operator F̂ , when written in terms of the expectation value
function F , involves the Riemann bracket. Moreover, this expression for the uncertainty is
quite simple.

8
 In fact Heisenberg’s famous uncertainty relation also assumes a nice form when expressed
in terms of the expectation value functions. It is very well-known that the familiar uncer-
tainty relation between two quantum observables may be written in a slightly stronger form
(see, e.g., [32]):
2 2
1 1
 
2 2
(∆F̂ ) (∆K̂) ≥ [F̂ , K̂] + [F̂⊥ , K̂⊥ ]+ , (2.9)
2i 2

where F̂⊥ is the non-linear operator defined by

F̂⊥ (Ψ) := F̂ (Ψ) − F (Ψ)

so that F̂⊥ (Ψ) is orthogonal to Ψ, if kΨk = 1.

Using the above results, we may immediately rewrite Eq. (2.9) in the form
2
h̄

(∆F̂ )2 (∆K̂)2 ≥ {F, K}Ω + ({F, K}+ − F K)2 . (2.10)
2
Incidentally, the last expression in Eq. (2.10) may be interpreted as the “quantum covari-
ance” of F̂ and K̂; see Ref. [16] for an explanation.

B. The quantum phase space

We have seen that to each quantum observable is associated a smooth real-valued (ex-
pectation value) function on the Hilbert space. Further, the familiar operations involving
quantum operators correspond to simple “classical looking” operations on the corresponding
functions. These observations suggest a formulation of standard quantum mechanics in the
language of classical mechanics. However, there are two difficulties. First, although the
Hilbert space is a symplectic space, because two state vectors related by multiplication by
any complex number define the same state, it is not the space of physical states, i.e., the
quantum analog of the classical phase space. Second, the description of the measurement
process in a manner intrinsic to the Kähler structure on H turns out not to be natural.
The true state space of the quantum system is the space of rays in the Hilbert space, i.e.,
the projective Hilbert space, which we shall denote P. It should not be surprising that P is
a Kähler manifold, and hence, in particular, a symplectic manifold. After all, for the special
case in which H is Cn+1 , P is the complex projective space CP n —the archetypical Kähler
manifold. In this section, we present a particularly useful description of the projective Hilbert
space which is valid for the infinite-dimensional case and which illuminates the role of its
symplectic structure. These developments will enable us to handle the above complications
and allow an elegant geometric formulation of quantum mechanics.

1. Gauge reduction

The standard strategy to handle the ambiguity of the state vector is to consider only
those elements of the Hilbert space which are normalized to unity. We will adopt this

9
approach by insisting that the only physically relevant portion of the Hilbert space is that
on which the constraint function,
1
C(Ψ) := hΨ, Ψi − 1 = G(Ψ, Ψ) − 1, (2.11)
2h̄
vanishes. The attitude adopted here is one in which the the constraint surface—the unit
sphere, with respect to the Hermitian inner-product—is the only portion of the Hilbert space
which is accessible to the system. The rest of the Hilbert space is often quite convenient,
but is viewed as an artificial element of the formalism.
Let us consider the above restriction from the point of view of the Bergmann-Dirac theory
of constrained systems (see, e.g., [33]). In other words, we will pretend, for a moment, that we
are dealing with a classical theory with the constraint C = 0. First notice that since the time-
evolution preserves the constraint surface, no further (secondary) constraints arise. Since
we have only a single constraint, it is trivially of first-class in Dirac’s terminology; i.e., the
constraint generates a motion which preserves the constraint surface (£XC C = {C, C}Ω = 0).
Recall that to every first-class constraint on a Hamiltonian system is associated a gauge
degree of freedom; the associated gauge transformations are defined by the flow along the
Hamiltonian vector field generated by the constraint function. In our case, the gauge direc-
tions are simply given by
1 ab 1
XCa = Ωab Db C = Ω Ψb = − J a b Ψb , (2.12)
h̄ h̄
where Da denotes the Levi-Cività derivative operator. For later convenience, let us define

J a := h̄XC a |S = −J a b Ψb |S . (2.13)

Notice that J a is the generator of phase rotations on S. Therefore, the gauge transfor-
mations generated by the constraint are exactly what they ought to be; they represent the
arbitrariness in our choice of phase!
Thus, we now see the relevance of the description in terms of constrained Hamiltonian
systems. By taking the quotient of the constraint surface of any constrained system by the
action of the gauge transformations, one obtains the true phase space of the system—often
called the reduced phase space. The projective Hilbert space may therefore be interpreted as
the “reduced phase space” of our constrained Hamiltonian system. In order to emphasize
both its physical role and geometric structure, we will refer to the projective Hilbert space
P as the quantum phase space. One can explicitly show that if H is infinite-dimensional, P
is an infinite-dimensional Hilbert manifold [16].
As the terminology suggests, any reduced phase space is equipped with a natural sym-
plectic structure. This fact may be seen as follows. Denote by i : S → H and π : S → P the
inclusion mapping and projection to the quantum phase space, respectively. By restricting
the symplectic structure Ω to the constraint surface, one obtains a closed 2-form i∗ Ω on S.
This 2-form is degenerate, but only in the gauge direction. Fortunately, since gauge trans-
formations are defined by the Hamiltonian vector field generated by the constraint function,
i∗ Ω is constant along its directions of degeneracy. As a result, there exists a symplectic form
ω on P whose pull-back via π agrees precisely with i∗ Ω. This is a standard construction in

10
the theory of systems with first class constraints; the only novelty here lies in its application
to ordinary quantum mechanics. Finally, we note that this result applies to the typical case
of interest, in which the original Hilbert space is infinite-dimensional [16]. The symplectic
structure ω is then a smooth, strongly non-degenerate field (i.e., defines an isomorphism
from the tangent space to the cotangent space at each point.)
Before discussing the geometry the quantum phase space, we should point out that the
viewpoint adopted in this section in fact generalizes quantum mechanics, but in a very trivial
way. In section II A we observed that each quantum observable defines a real-valued function
on the entire Hilbert space, and that the quantum evolution is given by the Hamiltonian flow
defined by a preferred function. In viewing a quantum system as a constrained Hamiltonian
system, we must concede that it is only the constraint surface S that is of physical relevance.
In particular, the restriction F |S of the expectation value function F contains all gauge-
invariant information about the observable; one may extend F off the constraint surface in
any desirable manner without affecting the corresponding flow on the projective space. The
particular extensions defined by expectation values of (bounded) self-adjoint operators may
be viewed as mere convention. We will make use of this point in section III A.

2. Symplectic geometry

Our method of arriving at the quantum phase space P by the reduced phase space con-
struction of constrained systems immediately suggests further definitions and constructions.
Recall that to each bounded, self-adjoint operator F̂ on H, we have associated the
function F (Ψ) := hΨ, F̂ Ψi on the Hilbert space. In fact, we may go a short step further.
First, let us restrict the expectation value function F to the constraint surface, thereby
obtaining the function i∗ F : S → R. i∗ F is clearly gauge-invariant (i.e. independent of
phase), and therefore defines the function f : P → R for which π ∗ f = i∗ F . Therefore, to
each quantum observable is associated a smooth, real-valued function on the quantum phase
space. The functions obtained in this manner will represent the observables in the geometric
formulation of quantum mechanics. Let us therefore make

Definition II.1 Let f : P → R be a smooth function on P. If there exists a bounded,

self-adjoint operator F̂ on H for which π ∗ f = hF̂ i|S , then f is said to be an observable
function.

Note that we consider the set of quantum observables to consist of the bounded self-adjoint
operators on H. At first sight this appears to be a severe restriction. However, further
reflection shows that it is not. In any actual experiment, one deals only with a finite range
of relevant parameters and hence in practice one only measures observables of the type
considered here. Thus, there is by definition a one-to-one correspondence between quantum
observables and the observable functions on P. As we will see below, the set of observable
functions is a very small subset of the entire function space.
A natural question arises: What is the relationship between the Hamiltonian vector
fields XF (on H) and Xf (on P)? Given any point Ψ ∈ S, we may push-forward the vector
XF |Ψ to obtain a tangent vector at π(Ψ). Since F is gauge-invariant, it commutes with C;
therefore

11
 £XC XF = X{F,C}Ω ≡ 0.
As a consequence, XF is “constant along the integral curves of J ”. Thus, by pushing-
forward XF at each point of S, one obtains a well-defined (smooth) vector field on all of P.
As is known to those familiar with the analysis of constrained systems, this vector field is
also Hamiltonian; in fact, it agrees precisely with Xf . The flow on P, which is induced by
the Schrödinger vector field of F̂ corresponds exactly to the Hamiltonian flow determined
by the observable function f .
Next, consider the Poisson bracket {, }ω defined by the reduced symplectic structure ω.
Let F, K : H → R be expectation value functions of two quantum observables and denote
by f, k : P → R the corresponding observable functions on P. As a consequence of the
above result,
π ∗ {f, k}ω = π ∗ (ω(Xf , Xk )) = ω(π∗ XF , π∗ XK ) = Ω(XF , XK )|S = {F, K}Ω |S . (2.14)
Therefore, the Poisson bracket defined by ω exactly reflects the commutator bracket on the
space of quantum observables.
In summary, to each quantum observable F̂ is associated a real-valued function f : P →
R on the quantum phase space. The Schrödinger vector field determined by F̂ determines a
flow on P; this flow is generated by the Hamiltonian vector field associated to the observable
function f . Further, the mapping F̂ 7→ f is one-to-one and respects the Lie algebraic
structures provided by the commutator and Poisson bracket on P, respectively.

C. Riemannian geometry and measurement theory

Any quantum mechanical system may be described as an infinite-dimensional Hamil-

tonian system. However, the structure of the quantum phase space is much richer than
that of classical mechanics. P is also equipped with a natural Riemannian metric. As we
will see, the probabilistic features of quantum mechanics are conveniently described by the
Riemannian structure.
The quantum metric may be described in much the same way as the symplectic structure.
The restriction i∗ G of G to the unit sphere is a strongly non-degenerate Riemannian metric
on S. Recall that the gauge generator J is (up to the constant factor of h̄) the Schrödinger
vector field associated to the identity operator. Since any Schrödinger vector field preserves
the Hermitian inner-product, it preserves both the symplectic structure Ω and the metric
G. As a consequence, J is a Killing vector field on S;
£J (i∗ G) ≡ 0. (2.15)
Therefore, P may also be described as the Killing reduction [34] of S with respect to the
Killing field J .
A manifold which arises in this way is always equipped with a Riemannian metric of its
own.2 Although i∗ G is “constant” on the integral curves of J , it is not degenerate in that

2 One must require that the integral curves of the Killing vector field do not come arbitrarily close
to one another. This condition is satisfied in our case.

12
direction. However, by subtracting off the component in the direction of J ,
1
 

g̃ := G − (Ψ ⊗ Ψ + J ⊗ J )  , (2.16)
2h̄ S

we obtain a symmetric tensor field which agrees with i∗ G when acting on vectors orthogonal
to J , is constant along J and is degenerate only in the direction of J . Therefore g̃ defines
a strongly non-degenerate Riemannian metric g on P. It is a simple matter to verify that g,
when combined with the symplectic structure ω, equips the quantum phase space with the
structure of a Kähler manifold.

1. Quantum observables

According to Def. II.1, quantum mechanical observables may be represented by real-

valued functions on the quantum phase space. Unfortunately, these functions have still
been defined in terms of self-adjoint operators on the Hilbert space. Our goal, however,
is a formulation of quantum mechanics which is intrinsic to the projective space; we wish
to avoid any explicit reference to the underlying Hilbert space. We now explain how this
deficiency may be overcome.
Since the Schrödinger vector field YF̂ = XF generates a one-parameter family of unitary
transformations on H, XF preserves not only the symplectic structure Ω, but the metric G
as well; XF is also a Killing vector field. This fact also holds for the corresponding observable
function f ; the Hamiltonian vector field Xf associated to any observable function f is also
a Killing vector field on (P, g). We will see that it is this property which characterizes the
set of observable functions on the quantum phase space.
Let us begin by recalling a general property of Killing vector fields. Since the calculations
are somewhat involved, we will now use Penrose’s abstract index notation (which, as already
pointed out, is meaningful also on infinite-dimensional Hilbert manifolds.) Let X α be any
Killing vector field on P. Then, by definition, ∇α Xβ + ∇β Xα = 0, where ∇ denotes the
(Levi-Cività) derivative operator associated to the metric g. Therefore, Kαβ := ∇α Xβ is
necessarily skew-symmetric. As one can easily verify, [35] K satisfies the identity

∇α Kβγ = Rγβα δ Xδ . (2.17)

(Our conventions are such that for any 1-form kγ , Rαβγ δ kδ = (∇α ∇β − ∇β ∇α )kγ .) As
a consequence, the Killing vector field X is completely determined by its value and first
covariant derivative at a single point. (See [36] or Appendix C of Ref. [35] for a discussion
of this useful fact.)
Now suppose that the above Killing vector field is generated by the observable function
f . (Below, it will be understood that X = Xf .) Since

ωα γ Kγβ = −ωα γ ∇β Xγ = ∇α ∇β f, (2.18)

K satisfies the additional property that ωα γ Kγβ is symmetric. (It then defines a bounded,
skew-self-adjoint operator on each tangent space.) By considering the coupled differential
equations:

13
 ∇α f = ωγα X γ , (2.19)
∇α Xβ = Kαβ , (2.20)
∇α Kβγ = Rγβα δ Xδ , (2.21)

we see [16] that any observable function is completely determined by its value and first two
derivatives at a single point! This fact motivates
Definition II.2 For each point p ∈ P, let Sp consist of all triples, (λ, Xα , Kαβ ), where λ is
a real number, Xα is a covector at p, and Kαβ is a 2-form at p for which ωα γ Kγβ = ωβ γ Kγα .
We call Sp the algebra of symmetry data at p.
Thus, any observable function f determines an element (λ, X, K) ∈ Sp (λ provides the value
of f at p), and f is completely determined by this symmetry data. The algebra of quantum
observables is then isomorphic to a subset of Sp . The converse to this statement is provided
by [16]:
Theorem II.1 For any element (λ, X, K) of Sp , there exists an observable function f :
P → R such that f (p) = λ, (Xf )α = Xα and ∇α (Xf )β = Kαβ .
Therefore, the space of observable functions is isomorphic to the algebra of symmetry data
Sp for any p ∈ P. We will utilize this result in section III B. For the purposes of this section,
however, the main use of the above theorem is the following. Recall that the Hamiltonian
vector field of any observable function is a Killing vector field. Conversely, let f be any
smooth, real-valued function on P for which Xf is also a Killing vector field. Of course,
the value and first two covariant derivatives of f at p ∈ P determine an element of Sp .
Therefore, by Thm. II.1, f is an observable function. This important result is expressed in
Corollary 1 A smooth function f : P → R is observable if and only if its Hamiltonian
vector field is also Killing.
Therefore, as in classical mechanics, the space of quantum observables is isomorphic to the
space of smooth functions on the phase space whose Hamiltonian vector fields are infinites-
imal symmetries of the available structure. However, now the available structure includes
not just the symplectic structure but also the metric. Hence, unlike in classical mechanics,
this function space is an extremely small subset of the set of all smooth functions on P.
This should not be terribly surprising; as an example consider the finite-dimensional case,
for which the function space is infinite-dimensional while the algebra of Hermitian operators
is finite-dimensional.

2. Quantum uncertainty

In analogy with our considerations of the symplectic geometry of P, let us consider the
Riemann bracket defined by the quantum metric g. We denote the g-Riemann bracket of
f, k : P → R by
h̄ h̄
(f, k) := g(Xf , Xk ) = (∇α f )g αβ (∇β k). (2.22)
2 2

14
If f and k correspond to the expectation value functions F, K : H → R, then, by Eq. (2.16),

h̄ h̄ 1
{F, K}+ = G(XF , XK ) = g̃(XF , XK ) + G(J , XF )G(J , XK ) = π ∗ [(f, k) + f k], (2.23)
2 2 4

where we have used the fact that for any observable F̂ ,

G(J , XF )|Ψ = −G(JΨ, XF ) = Ω(XF , Ψ) = Ψ ◦ (dF ) = 2F. (2.24)

Therefore, the observable function which corresponds to the Jordan product of F̂ and K̂
is given not by the g-Riemann bracket of f and k, but by the quantity

{f, k}+ := (f, k) + f k, (2.25)

where we have utilized a minor notational abuse to emphasize the relationship with Eq. (2.7).
The above quantity will be called the symmetric bracket of f and k.
Note that the g-Riemann bracket of two observables, while not necessarily an observable
itself, is a physically meaningful quantity; it is simply the quantum covariance function (see
the comment following Eq. (2.10)). In particular, (f, f )(p) is exactly the squared uncertainty
of f at the quantum state labeled by p;

(∆f )2 (p) := (∆F̂ )2 (π −1 p) = (f, f )(p). (2.26)

In order to obtain a feeling for the physical meaning of the quantum covariance, notice
that the uncertainty relation of Eq. (2.10) assumes the form
2
h̄

(∆f )(∆k) ≥ {f, k}ω + (f, k)2 . (2.27)
2
As one can easily show, the standard uncertainty relation
2
h̄

(∆f )(∆k) ≥ {f, k}ω (2.28)
2
is saturated at the state p ∈ P if and only if (f, k)(p) = 0 and Xf ∝ j(Xk ), where j is
the complex structure on P (compatible with ω and g). The quantum covariance (f, k)(p)
therefore measures the “coherence” of the state p with respect to the observables f and k.
Let us conclude this sub-section by reproducing a result of Anandan and Aharanov [25].
Suppose that the Hamiltonian operator of a quantum system is bounded and let h be the
corresponding observable function. By definition of the Riemann bracket, the uncertainty
of h is given by (∆h)2 = h̄2 g(Xh , Xh ). Therefore, apart from the constant coefficient, the
uncertainty in the energy is exactly the length of the Hamiltonian vector field which generates
the time-evolution. Thus, the energy uncertainty can be thought of as ‘the speed with which
the system moves through the quantum phase space’; during its evolution, the system passes
quickly through regions where the energy uncertainty is large and spends more time in states
where it is small.

15
 3. The measurement process

Some of the most significant aspects of quantum mechanics involve the measurement
process and, without a complete geometric description of these issues, our program would
be incomplete. In this sub-section we will sketch the desired geometric description, including
the case when the spectrum of the operator is continuous. A more complete discussion may
be found in [16].
Let Ψ0 ∈ S be an arbitrary normalized element of the Hilbert space, and let p0 = π(Ψ0 )
be its projection to P. Of obvious interest, in the context of measurement, is the function
δ̃Ψ0 : Ψ ∈ S 7→ |hΨ0 , Ψi|2 . Since δ̃Ψ0 is independent of the phase of Ψ it defines a function
δp0 on P via

δp0 (p) := δ̃Ψ0 (π −1 p) = |hπ −1 (p0 ), π −1 (p)i|2 . (2.29)

If the quantum system is in the state labeled by p0 when a measurement is performed, the
relevant quantum mechanical probability distribution is determined by the function δp0 . We
therefore desire a description of this function which does not rely explicitly on the underlying
Hilbert space. This is provided by

Theorem II.2 Given arbitrary points p0, p ∈ P there

√ exists
 a (closed) geodesic which passes
2
through p0 and p. Further, dp0 (p) = cos σ(p0 , p)/ 2h̄ , where σ(p0 , p) denotes the geodesic
separation of p0 and p.

A few comments regarding Thm. II.2 are in order. First, if π −1 (p0 ) and π −1 (p) are
non-orthogonal, then the above-mentioned geodesic is unique, up to re-parameterization.
Next, since all geodesics on P are closed, the geodesic distance σ(p0 , p) is, strictly speaking,
ill-defined. Due to the periodicity of the cosine function, however, the “transition amplitude
function”, dp0 , is insensitive to this ambiguity. For the sake of precision, by σ(p0 , p), we will
mean the minimal geodesic distance separating p0 and p.
Suppose one is dealing with the measurement of an operator F̂ with discrete, non-
degenerate spectrum, and let f be the corresponding observable function. Each eigenspace
of F̂ is one complex-dimensional, and therefore determines a single point of P. Denote these
eigenstates by pi . Suppose the system is in the state labeled by the point p0 when an ideal
measurement of f is performed. We know that the system will ‘collapse’ to one of the states
pi . Theorem II.2 provides the corresponding probabilities. It is interesting to notice that
the probability of collapse to an eigenstate pi is a monotonically decreasing function of the
geodesic separation of p0 and pi ; the system is more likely to collapse to a nearby state than
a distant one.
We now have a description of the probabilities associated with the measurement process,
but there are two deficiencies to be remedied. The eigenstates pi above have been defined
in terms of the algebraic properties of the operator F̂ . For our program to be complete, we
require a definition of these eigenstates which does not refer to the Hilbert space explicitly.
Next, the above discussion was limited to the measurement of an observable with discrete,
non-degenerate spectrum. We will describe the generic situation in two steps. First, we
consider the measurement of observables with discrete, but possibly degenerate, spectra.

16
This will require the aforementioned description of the eigenstates. We will then be prepared
to consider measurement of observables with continuous spectra.
Let us first examine the notions of eigenstates and eigenvalues of an observable operator
F̂ . Let F : H → R and f : P → R be the expectation value and corresponding observable
function, respectively. A vector Ψ ∈ H is an eigenstate of F̂ iff F̂ Ψ = λΨ, for some (real)
λ. Alternatively, by Eq. (2.3),
XF |Ψ = YF̂ |Ψ = (λ/h̄)J |Ψ . (2.30)

That is, Ψ is an eigenstate of F̂ if and only if the Hamiltonian vector field, XF is vertical
(i.e., purely gauge) at Ψ. This will be the case if and only if Xf vanishes at π(Ψ); π(Ψ) is
then a critical point of the function f . Evidently, the corresponding eigenvalue is exactly
the (critical) value of f at π(Ψ). In summary:
Definition II.3 Let f : P → R be an observable function. Critical point of f are called
eigenstates of f . The corresponding critical values are called eigenvalues.
We now consider the measurement of an observable f whose spectrum is discrete but
possibly degenerate (of course, the “spectrum of f ” coincides, by definition, with the spec-
trum of the corresponding operator F̂ ). Let λ be a degenerate eigenvalue of F̂ , and denote
by E˜λ the associated eigenspace of H. Associated to this eigenspace is a sub-manifold,
Eλ , of P, which we shall call the eigenmanifold associated to λ. Suppose that the system
is prepared in the state labeled by p0 and let Ψ0 ∈ π −1 p0 . The postulates of ordinary
quantum mechanics assert that measurement of f will yield the value λ with probability
hPλ Ψ0 , Pλ Ψ0 i = hΨ0 , PλΨ0 i = |hΨ0 , PλΨ0 /kPλ Ψ0 ki|2 , where Pλ is the projector onto the
relevant eigenspace of F̂ . From the above considerations, we know that this probability may
be expressed in terms of the geodesic separation of the points p0 and π (Pλ Ψ0 /kPλ Ψ0 k) ∈ Eλ .
We will denote the latter point by Pλ (p0 ).
What sets the point Pλ (p0 ) apart from all other elements of Eλ ? We need only notice
that for any point Φ ∈ Ẽλ ∩ S,
|hΨ0 , Φi|2 = |hΨ0 , Pλ Φi|2 = |hPλ Ψ0 , Φi|2 ≤ kPλ Ψ0 k2 . (2.31)
Therefore, of all elements Φ ∈ Ẽλ with unit normalization, that which maximizes the quantity
|hΨ0 , Φi|2 is simply Pλ Ψ0 /kPλ Ψ0 k, i.e., that to which the state Ψ0 will ‘collapse’ in the event
that measurement of F̂ yields the value λ. Therefore, by Thm. II.2, Pλ (p0 ) is simply that
point of Eλ which is nearest p0 !
We may now describe the measurement of an observable with discrete spectrum as fol-
lows. Suppose that immediately prior to measurement of f , the system is in the state p0 ∈ P.
Denote by λi the critical values of f and, by Eλi the corresponding eigenmanifolds. Interac-
tion with the measurement device causes the system to be projected to one of the eigenman-
ifolds, say Eλ0 . “Realizing that it collapsed” to the state Pλ0 (p0 ), the system returns what is
knows to be the value of the observable under consideration, i.e., f (Pλ0 (p0 )) = λ0 . Of course,
√ 
the probability that measurement causes reduction to Eλ0 is given by cos2 σ(p0 , Eλ0 )/ 2h̄ ,
where, σ(p0 , Eλ0 ) denotes the minimal geodesic separation of p0 and the sub-manifold Eλ0 .
Now let us study the generic case. Let F̂ be any observable operator on H, the spectrum
of which is allowed to be continuous. We first need a definition of the spectrum of F̂ in terms

17
of the corresponding observable function f : P → R. Recall the standard definition [37]:
λ is an element of the spectrum sp(F̂ ) if and only if the operator F̂ − λ1̂ is not invertible.
Equivalently, λ ∈ sp(F̂ ) iff given any positive ε ∈ R, ∃Ψ ∈ S such that kF̂ Ψ − λΨk < ε.
This condition guarantees that, to arbitrary precision, λ is an approximate eigenvalue of F̂ .
Using Eqs. (2.25) and (2.26), we may write the (square of the) above quantity as
h i
kF̂ Ψ − λΨk2 = {f − λ, f − λ}+ |π(Ψ) = (∆f )2 + (f − λ)2  . (2.32)

π(Ψ)

This equation allows us to define the spectrum of F̂ in terms of the function f : P → R;

Definition II.4 The spectrum sp(f ) of an observable f consists of all real numbers λ for
h i−1
which the function nλ : P → R ∪ {∞}, nλ : p 7→ (∆f )2 (p) + (f (p) − λ)2 is unbounded.

Of course, a point at which nλ = ∞ corresponds to an eigenstate of f .

The next step is a description of the spectral projection operators. Let Λ be a closed
subset of the spectrum sp(f ) of f , and denote by PF̂ ,Λ the projection operator associated
to F̂ and Λ. In analogy with the above, put E˜F̂ ,Λ = {PF̂ ,Λ Ψ | Ψ ∈ H − {0}}, and let Ef,Λ
denote the projection of E˜F̂ ,Λ to P. Note that the set E˜F̂ ,Λ —the analogue of the eigenspace
above— actually is the eigenspace of PF̂ ,Λ corresponding to the eigenvalue 1. Therefore, we
have Ef,Λ consists of the critical points of an expectation value function, associated to the
critical value 1. Unfortunately, this expectation value function is not directly expressible in
terms of f and Λ. We must look for an alternative description of the sub-manifold Ef,Λ .
In the representation defined by the operator F̂ , elements of ẼF̂ ,Λ have support on Λ.
Therefore, Ψ ∈ ẼF̂ ,Λ iff hF̂ iΨ ∈ Λn ∀n > 0, where Λn denotes the image of Λ under
the map λ 7→ λn . Recall that {f, f }+ is the (projection to P of the) expectation value
of F̂ 2 . In general, the expectation value of F̂ n projects to the n-fold symmetric product
{f, {f, {f, . . .}+ }+ }+ . Therefore, we have

Ef,Λ = {q ∈ P|{f, {f, {f, . . .}+ }+ }+ |q ∈ Λn ∀n > 0}, (2.33)

where there are n factors of f occurring above.

Having obtained a description of Ef,Λ , we may now define the spectral projections in
a manner intrinsic to the projective space. By precisely the same reasoning surrounding
Eq. (2.31), Pf,Λ maps a point p ∈ P to that element of Ef,Λ which is nearest p.
The measurement process may then be described as follows. Suppose the quantum
system is in the state labeled by the point p0 ∈ P at the instant an experimenter performs
a measurement of the observable f . Following the rules of quantum mechanics, she “asks
the system” whether the value of f lies in Λ—a closed subset of sp(f ), which she is free to
choose. The experimental apparatus drives the system to one of two states—either Pf,Λ (p0 )
or Pf,Λc (p0 ), where Λc is the (closure of the) complement, in sp(f ), of Λ. The system is
reduced to the former with probability
!
2 σ(p0 , Pf,Λ (p0 ))
cos √ ;
2h̄

18
in this event, the experiment yields the positive result (f ∈ Λ). Having precisely prepared
the system in the state p0 , the experimenter may then infer the value f (Pf,Λ (p0 )) of the
observable f . The probability of reduction to the latter state is obtained by replacing Λ by
Λc above.
Note that this description encompasses all measurement situations. In the event that the
spectrum of f is discrete and non-degenerate, the experimenter may choose to let Λi contain
the single eigenvalue λi . Moreover, she may measure all of the projections simultaneously.
In this way, one recovers the first familiar description of the measurement process. Note
also that, while the above discussion of the spectral projections may seem complicated and
somewhat unnatural at first, the definition of the spectral projection operators on the Hilbert
space has the same features. (Indeed, most text books simply skip this technical discussion.)
This is simply one of the technical complications that the geometric formalism inherits from
the Hilbert space framework.
To conclude, we wish to emphasize that the topic of our discussion has been ordinary
quantum mechanics. We have just restated the well-known quantum mechanical formalism
in a language intrinsic to the true space of states—the quantum phase space, P; no new
ingredients have been added to the physics. A particularly attractive feature of the formal-
ism, however, is the fact that slight modifications of the standard picture naturally present
themselves. For example, using many of these geometric ideas, Hughston [28] has explored
a novel approach to the measurement problem in terms of stochastic evolution.

D. The postulates of quantum mechanics

Let us collect the results obtained in the first three sub-sections.

We have formulated ordinary quantum mechanics in a language which is intrinsic to the
true space of quantum states—the projective Hilbert space P. As in classical mechanics,
observables are smooth, real-valued functions which preserve the kinematic structure. Being
a Kähler manifold, P is a symplectic manifold. The role of the quantum symplectic structure
is precisely that of classical mechanics; it defines both the Lie algebraic structure on the space
of observables and generates motions including the time-evolution.
There are, however, two important features of quantum mechanics which are not shared
by the classical description. First, the phase space is of a very particular nature; it is a Kähler
manifold and, as we will see in section III B, one of a rather special type—namely one of
constant holomorphic sectional curvature.3 The second difference lies in the probabilistic
aspects of the formalism, which is itself intimately related to the presence of the Riemannian
metric. More generally, this metric describes those quantum mechanical features which are
absent in the classical theory—namely, the notions of uncertainty and state reduction. For
example, the transition probabilities which arise in quantum mechanics are determined by
a simple function of the geodesic distance between points of the phase space.
These results are most easily summarized by stating the postulates in the geometric
language:

3 In this sense, the quantum framework is actually a special case of the classical one!

19
 (H) Physical states: Physical states of the quantum system are in one-to-one correspon-
dence with points of a Kähler manifold P, which is a projective Hilbert space.4

(U) Kähler evolution: The evolution of the system is determined by a flow on P, which
preserves the Kähler structure. The generator of this flow is a densely defined vector
field on P.

(O) Observables: Physical observables are represented by real-valued, smooth functions

f on P whose Hamiltonian vector fields Xf preserve the Kähler structure.

(P) Probabilistic interpretation: Let Λ ⊂ R be a closed subset of the spectrum of an

observable f , and suppose the system is in the state corresponding to the point p ∈ P.
The probability that measurement of f will yield an element of Λ is given by
!
2 σ(p, Pf,Λ (p))
δp (Λ) = cos √ , (2.34)
2h̄

where Pf,Λ (p) is the point, closest to p, in the space Ef,Λ , defined bye Eq. (2.33).

(RD ) Reduction, discrete spectrum: Suppose the spectrum of an observable f is discrete.

This spectrum provides the set of possible outcomes of the ideal measurement of f . If
measurement of f yields the eigenvalue λ, the state of the system immediately after
the measurement is given by the associated projection, Pf,λ (p), of the initial state p.

(RC ) Reduction, continuous spectrum: A closed subset Λ of the spectrum of f determines

an ideal measurement that may be performed on the system. This measurement
corresponds to inquiring whether the value of f lies in Λ. Immediately after this
measurement, the state of the system is given by Pf,Λ (p) or Pf,Λc (p), depending on
whether the result of the measurement is positive or negative, respectively.

In the last postulate, Λc is the closure of the complement, in the spectrum of f of the set
Λ. Although the first “reduction postulate” is a special case of the second, both have been
included for comparison with standard textbook presentations.
To conclude, although it is not obvious from textbook presentations, the postulates of
quantum mechanics can be formulated in an intrinsically geometric fashion, without any
reference to the Hilbert space. The Hilbert space and associated algebraic machinery pro-
vides convenient technical tools. But they are not essential. Mathematically, the situation
is similar to the discussion of manifolds of constant curvature. In practice, one often es-
tablishes their properties by first embedding them in Rn (equipped with a flat metric of
appropriate signature). However, the embedding is only for convenience; the object of in-
terest is the manifold itself. There is also a potential analogy from physics, alluded to in the

4 We will see in section III B that quantum phase spaces P can be alternatively singled out as
Kähler manifolds which admit maximal symmetries. This provides an intrinsic characterization
without any reference to Hilbert spaces.

20
Introduction. Perhaps the habitual linear structures of quantum mechanics are analogous to
the inertial rest frames in special relativity and the geometric description summarized here,
analogous to Minkowski’s reformulation of special relativity. Minkowski’s description paved
the way to general relativity. Could the geometric formulation of quantum mechanics lead
to a more complete theory one day?

III. A UNIFIED FRAMEWORK FOR GENERALIZATIONS OF QUANTUM

MECHANICS

There are three basic elements of the quantum mechanical formalism which may be
considered for generalization: the state space, the algebra of observables and the dynamics.
The framework developed in section II suggests avenues for each of these. First, while it
is not obvious how one might “wiggle” a Hilbert space, one may generalize the quantum
phase space by considering, say, the class of all Kähler manifolds {M, g, ω}. The geometric
language also suggests an obvious generalization of the space of quantum observables: one
might consider the space of all smooth, real-valued functions on the phase space. Finally,
whether or not one chooses either of these extensions, one may consider generalized dynamics
which, as in classical mechanics, preserves only the symplectic structure. Thus, one might
require the dynamical flow to preserve only the symplectic structure, and not necessarily
the metric.
While each of these structures may be extended separately, they are intimately related
and construction of a complete, consistent framework is a highly non-trivial task. Thus, for
example, if one allows all Kähler manifolds as possible quantum phase spaces, is seems very
difficult to obtain consistent probabilistic predictions for outcomes of measurements. More
generally, the problem of systematically analyzing viable, non-trivial generalizations of the
the kinematic structure would be a major undertaking, although the pay-off may well be
exceptional. Modification of dynamics, on the other hand, is easier at least in principle.
Therefore, we will first consider these in section III A and return to kinematics in section
III B.

A. Generalized dynamics

Let us then suppose that we continue to use a projective Hilbert space for the quantum
phase space, and let dynamics be generated by a preferred Hamiltonian function. However,
let us only require that time-evolution should preserve the symplectic structure ω (as in
classical mechanics), and not necessarily the metric g. From the viewpoint of the geometric
formulation, this is the simplest and most obvious generalization of the standard quantum
dynamics.5 The idea is reminiscent of the “non-linear Schrödinger equations” that have been

5 Recallthat the Hamiltonian need not be an observable function in the sense of section II B. We
could extend the kinematical set up as well and regard any smooth function on P as an observable
function. (This would be analogous to Weinberg’s [7] proposal which, however, was made at the

21
considered in the past. Therefore, it is natural to ask if these there is a relation between
the two. We will see that the answer is in the affirmative. Furthermore, the geometric
framework provides a unified treatment of these proposals and makes the relation between
them transparent, thereby enabling one to correct a misconception.
Let us begin by defining the the class CH of Hamiltonians we now wish to consider. CH
will consist of densely defined functions f on P satisfying the following properties: i) f is
smooth on its domain of definition; and ii) the Hamiltonian vector field it defines generates
a flow on all of P. In particular, The Hamiltonian functions we considered in section II—
expectation values of a possibly unbounded self-adjoint operator—belong to CH but they
constitute only a ‘small subset’ of CH .
The existing proposals of non-linear dynamics refer to flows in the full Hilbert space H
rather than in the quantum phase space P. To compare the two, we need to lift our flows to
H. Let us begin by recalling that it is natural to regard P as a reduced phase space, resulting
from the first class constraint C(Ψ) := hΨ, Ψi − 1 = 0 on H (see Eq. 2.11). Therefore, a
function f on P admits a natural lift F to S, the unit sphere in H. This function F on S
is constant along the integral curves of the vector field J which generates phase rotations;
£J f = 0. Denote the space of these lifts by C˜H .
To discuss dynamics on H, we need to extend these functions6 off S. From the reduced
phase space viewpoint, the extension is completely arbitrary. For, we can construct the
Hamiltonian vector field on H generated by any extension Fext . The restriction to S of this
vector field does depend on the extension but the the horizontal part of the restriction—i.e.,
the part orthogonal to J —does not. Hence, the projection of the vector field to P agrees
with Xf , irrespective of the choice of the initial extension. Thus, the generalized dynamics
generated by a given Hamiltonian function f on P can be lifted to a whole family of flows
on H, all of which, however, evolve the physical quantum states—elements of P—in the
same way. Because of this, apparently distinct proposals for non-linear evolutions on H can
in fact be physically equivalent. This point is rather trivial from the viewpoint of geometric
quantum mechanics. The reason for our elaboration is that—as we will see below—it has
not been appreciated in the Hilbert space formulations.
While the extension off S of elements of C˜H is completely arbitrary, one can use the
standard quantum mechanical framework to select a specific rule. Consider, to begin with,
a bounded, self-adjoint operator F̂ on H, and let F be the restriction to S of the corre-
sponding expectation value function. There is then an obvious extension of F to all of H:
set Fext (Ψ) := hΨ, F̂ Ψi (which we denoted by F in section II A). One can restate this rule

level of the Hilbert space H rather than the quantum phase space P.) We have refrained from
doing this because the required extension of measurement theory is far from obvious.

6 Strictlyspeaking, we only need to extend the dynamical flow off S. However, to compare our
results with Weinberg’s [7] we need to consider extensions of Hamiltonians. In discussions on
generalized dynamics [7,5], the issue of domains of definition of operators is generally ignored.
Our treatment will be at the same level of rigor. In particular, we will ignore the fact that our
Hamiltonian functions and vector fields are only densely defined.

22
as:

Fext (Ψ) := kΨk2 F (Ψ/kΨk). (3.1)

The advantage is that this equation may now be used to extend any element F of C˜H to all
of H× = H − {0}. Note that, with this preferred extension, the Hamiltonian vector field
XFext is homogeneous of degree one on H× :

XFext (cΨ) = cXFext (Ψ) ∀c ∈ C. (3.2)

Hence, the flow on H× which is generated by Fext is homogeneous, but fails to be linear
unless F is the restriction to S of the expectation value function defined by a self-adjoint
operator. Next, it is easy to verify that these Fext strongly commute with the constraint
function C(Ψ). Hence the flow along XFext preserves the constraint. Therefore, the specific
extension considered above has the property that the corresponding flow preserves not only
the symplectic structure Ω but also the norm on H× . However, unless f is an observable
function, it does not preserve the metric G.
Note that, even if we consider just these preferred extensions, the set of possible Hamil-
tonian functions on H× has been extended quite dramatically. To see this, consider the
case when H is finite-dimensional. Then, the class of Hamiltonian functions on H allowed
in standard quantum mechanics forms a finite-dimensional real vector space; it is just the
space of expectation value functions constructed from self-adjoint operators. The space C˜H ,
on the other hand, is infinite-dimensional since its elements are in one-to-one correspondence
with smooth functions on P. And each element of C˜H admits an unique extension to H× via
Eq. (3.1). It is natural to ask if one can do something ‘in-between’. Can we impose more
stringent requirements to select a class of potential Hamiltonians which is larger than that
of observable functions of section II but smaller than CH ? For example, one might imagine
looking for the class of functions on H× whose Hamiltonian flows preserve not just the norms
but also the inner-product. It turns out, however, that this class consists precisely of the
expectation value functions defined by self-adjoint operators; there is no such thing as a
non-linear unitary flow on H [16]. Despite the magnitude of our generalization, it seems to
be the only available choice.
We are now ready to discuss the relation between this generalization and those that
have appeared in the literature. Note first that Eq. (3.1) implies that there is a one-to-one
correspondence between elements of CH and smooth functions on H× which are gauge-
invariant (i.e. insensitive to phase) and homogeneous of degree two. (If H is viewed as a
vector space over complex numbers, this corresponds to homogeneity of degree one in both Ψ
and Ψ̄.) It turns out that this is precisely the class of permissible Hamiltonians that Weinberg
[7] was led to consider while looking for a general framework for non-linear generalizations
of quantum mechanics. Let us therefore call functions on H× satisfying Eq. (3.1) Weinberg
functions, and denote the space of these functions by OW . Our discussion shows that there
is a one-to-one correspondence between smooth functions on the projective Hilbert space
P and Weinberg functions on the punctured Hilbert space H× ;the homogeneity restriction
simply serves to eliminate the freedom in the extension of the function on H× . Thus the
extension of quantum dynamics that is immediately suggested by the geometrical framework
reproduces key features of Weinberg’s proposal.

23
 There are, however, considerable differences in the motivations and general viewpoints of
the two treatments. In particular, Weinberg works with the Hilbert space H (and, without
explicitly saying so, sometimes with H× ). However, he does assume at the outset that
elements Ψ and cΨ of H define the same physical state of the quantum system for all
complex numbers c. Thus, although it is not explicitly stated, his space of physical states is
also P. Therefore, it is possible to translate his constructions to the the geometric language.
As we will see below, the geometric viewpoint is often clarifying.
Next, let us consider two specific examples of non-linear dynamics that have been con-
sidered in the literature. Each of these involves the non-relativistic mechanics of a point
particle moving in Rn and the Hilbert space consists of square-integrable functions on Rn .
Therefore, it will be useful to reinstate a complex notation for the remainder of this sub-
section. In this notation, an element of OW is a real-valued function F (Ψ, Ψ̄) of both Ψ
and its conjugate, which is homogeneous of degree one in each argument. The Hamiltonian
vector field generated by such a function corresponds to
1 δF
XF [Ψ](x) = (x). (3.3)
ih̄ δΨ∗
The simplest example of this type is provided by is the so-called “non-linear Schrödinger
equation”. This equation is given by
∂Ψ
ih̄ (x, t) = (Ĥ0 Ψ)(x, t) + ǫ|Ψ(x, t)|2 Ψ(x, t), (3.4)
∂t
where Ĥ0 = P̂ 2 /2m + V̂ is the standard Hamiltonian operator describing a non-relativistic
particle under the influence of a conservative force with potential V̂ . Note that the quantity
|Ψ(x, t)| is the modulus of Ψ(x, t), not the norm kΨk of the state-vector Ψ. It is easy to
verify that (3.4) induces a flow on P. This flow is Hamiltonian and the generating function
is the projection to P of the function hΨ, H0Ψi + Hǫ (Ψ) on H, where,
ǫ
Z
Hǫ (Ψ) := dnx [Ψ∗ (x, t)Ψ(x, t)]2 . (3.5)
2
Thus, Eq. (3.4) is indeed a specific example of our generalized dynamics.
Note, however, that the dynamical vector field X on H defined directly by the non-linear
Schrödinger equation is given by

X[Ψ](x) = XH0 [Ψ](x) + Xǫ [Ψ](x), (3.6)

where Xǫ [Ψ](x) = (ǫ/ih̄)|Ψ(x, t)|2 Ψ(x, t). Clearly, it is not homogeneous in the sense of
Eq. (3.2), and hence is not generated by a Weinberg function. Therefore, Weinberg was
led to state that the “results obtained by the mathematical studies of this equation are
unfortunately of no use” to the generalization he considered. However, we just saw that
the non-linear Schrödinger equation does correspond to generalized dynamics on P and is
therefore of ‘Weinberg type’. Hence, the statement in quotes is somewhat misleading.
Let us elaborate on this point. If we first focus on physical states, what matters is just
the projected flow on P. This in turn is completely determined by the restriction Hǫ |S of
Hǫ to S. Therefore, we may feel free to ignore the behavior of Hǫ off of S. Of course,

24
to construct a vector field on S which projects to the relevant one on P, we can extend
Hǫ |S arbitrarily and compute the associated Hamiltonian vector field. In particular, we may
extend it in the way which Weinberg would suggest:

Hǫ′ (Ψ) := kΨk2 Hǫ |S (Ψ/kΨk). (3.7)

Thus, we have seen explicitly that the flow on P which is defined by Hǫ may also be described
by a Weinberg function! The emphasis on the true space of states P of the geometric
treatment clarifies this point which seems rather confusing at first from the Hilbert space
perspective.
The non-linear Schrödinger equation is a fairly simple example since the generating
function Hǫ is itself homogeneous (but of the “wrong” degree to be a Weinberg function).
Let us now consider an example which is more sophisticated.
In an effort to address the problem of combining systems which are subject to a non-
linear equation of motion, Bialynicki-Birula and Mycielski were led to a logarithmic equation
of motion [5]. They began with a general equation of motion of the form
∂Ψ
ih̄ (x, t) = (Ĥ0 Ψ)(x, t) + α(|Ψ(x, t)|2)Ψ(x, t), (3.8)
∂t
and showed that physical considerations, particularly the requirement that α-term should
not introduce interactions between otherwise non-interacting subsystems, imposes severe
restrictions the functional form of α. The only possibility is to have: α(ρ) = −b ln(an ρ) for
some constants a and b. (For details see [5]). Choosing units7 with respect to which a = 1,
the vector field along which the system evolves may be written as

X[Ψ](x) = XH0 [Ψ](x) + X1 [Ψ](x), (3.9)

where X1 [Ψ](x) = −(b/ih̄) ln(|Ψ|2 )Ψ(x). Again, the extra term may be seen to be Hamilto-
nian; X1 = XH1 , where
Z
H1 (Ψ) = b dnxΨ∗ (x)Ψ(x)[1 − ln(Ψ∗ (x)Ψ(x)]. (3.10)

Since H1 is phase-invariant, this is also an example of our generalized dynamics.

Again, we can carry out the procedure used for the non-linear Schrödinger equation to
see that the corresponding motion on the projective space may be described by use of a
Weinberg function. It is not difficult to show that the corresponding homogeneous function
is given by

H1′ (Ψ) = H1 (Ψ) + bkΨk2 ln(kΨk2 ). (3.11)

Therefore, the logarithmic equation induces a flow on P which may be described by a

function of the Weinberg type.

7 The constant a is a length scale of no physical significance, since it may be altered by addition
of a constant to the Hamiltonian operator.

25
 In retrospect, essentially any generalized dynamics of the form specified by Eq. (3.8)
may be written as a Hamiltonian flow on H× by “integrating” the function F . The resulting
Hamiltonian function is not likely to satisfy Weinberg’s homogeneity condition. However,
one may restrict it to the unit sphere, then extend this restriction as in Eq. (3.1). The
resulting function will generate a Hamiltonian vector field which differs from the original
one in general, but agrees with it on the unit sphere. Hence, both will generate the same
flows on the projective space. This may be done with any Hamiltonian function on H×
which preserves the unit sphere.

B. Characterization of the standard quantum kinematics

Let us now turn to the issue of generalizations of the kinematic framework itself. As
noted above, these extensions may well turn out to be profound. However, they also appear
to be much more difficult to carry out. Therefore, as a first step, it is natural to ask what
sets ordinary quantum mechanics apart from its natural generalizations. The purpose of
this sub-section is provide such a characterization of the standard kinematic framework.
Let us consider an arbitrary Kähler manifold {M, g, ω}, which is to represent the phase
space of a (radically!) generalized quantum theory. (As before, M will be assumed to be a
Hilbert manifold, and g and ω to be strongly non-degenerate.) The question is then: Are
there natural conditions which will single out ordinary quantum mechanics from this class
of generalizations? As we saw in section II, the observables of ordinary quantum mechanics
are smooth functions on the phase space whose Hamiltonian vector fields satisfy Killing’s
equation; we therefore let the observables of the generalized framework consist of all smooth
functions on M whose Hamiltonian flows are isometries. Denote this set as

O := {f : M 7→ R | £Xf g = 0}. (3.12)

(Note that while we call elements of O observables, we do not claim to possess a complete
and consistent formalism incorporating measurement. Indeed, this seems an impossible feat
at the present level of generality.)
It should be emphasized that there are Kähler manifolds which do not admit a single
observable function (other than constants); the torus is an example. Let us amplify this point
briefly. Since the discussion leading to Def. II.2 is valid for an arbitrary Kähler manifold,
any observable is completely determined by its value and first two derivatives at a single
point. As before, we will denote by Sp the set of symmetry data at the point p ∈ M (if λ,
X and K are 0-, 1- and 2-forms at p for which ωα γ Kγβ is symmetric, then (λ, X, K) ∈ Sp ).
Given an arbitrary p ∈ M, each observable function then determines an element of Sp .
According to Thm. II.1, the converse holds for ordinary quantum mechanics: there, the
algebra of symmetry data at each point p is integrable. In this sense, the phase space of
standard quantum mechanics admits “as many observables as possible.” This useful idea is
captured in

Definition III.1 If, for each p ∈ M, every element (λ, X, K) ∈ Sp is integrable, we say
that O is maximal.

26
The set of observables of ordinary quantum mechanics is maximal. As we will see, it is
essentially this property of the quantum observables which allows us to recover the standard
formalism.
In order to illustrate the importance of the maximality of the space of observables, it is
necessary to introduce one mathematical concept. It is fairly easy to verify that the Riemann
curvature tensor of a projective Hilbert space is of the special form
Ch i
Rαβγδ = gγ[α gβ]δ + ωαβ ωδγ − ωγ[α ωβ]δ , (3.13)
2
where C = h̄/2 and [. . .] denotes skew-symmetrization. If the curvature tensor of our
generalized phase space M satisfies this equation at a point p, it is said to be of constant
holomorphic sectional curvature (CHSC) at p. In that case, the real number C is the value
of the holomorphic sectional curvature at p. If for some real C, the Riemann tensor assumes
the form written in Eq. (3.13) (for all p ∈ M), M is called a manifold of CHSC= C. It may
be useful to note that if a Kähler manifold M is of CHSC at each point, then it must be of
overall CHSC; that is, if R satisfies Eq. (3.13) at each p, then C must be a constant [38].
Holomorphic sectional curvature, in the context of complex manifolds, is analogous to
the scalar curvature of real manifolds. Since there is a very strong relationship between
the number of independent Killing vector fields on a real manifold and the form of the
Riemann curvature tensor [36], one may expect that the maximality of the space of quantum
observables to be strongly related to Eq. (3.13). We will therefore use an approach analogous
to that presented in Ref. [36] to consider more closely the interplay between the geometry
of the phase space and the algebra of observables. The results in this section are stated
without proof; for details, see Ref. [16].
Since the commutator of two Killing vector fields is another Killing vector field, the set
of observables on an arbitrary Kähler manifold is closed under the Poisson bracket {, }. The
Poisson bracket also satisfies the Jacobi identity; this equips O with the structure of a Lie
algebra. On the entire function space, we may also define the commutative operation
h̄
{f, k}+ := (f, k) + f k, (f, k) := (∇α f )g αβ (∇β g), (3.14)
2
which we call the symmetric bracket. The symmetric bracket is not necessarily closed on the
space of observables on an arbitrary Kähler manifold.
Suppose that f1 and f2 are any two observable functions on M, and let f3 = {f1 , f2 }.
Since each fi generates a Killing vector field, it determines an element (fi , Xi , Ki)|p ∈ Sp
for any p ∈ M, where Xi = Xfi and Kiαβ = ∇α (Xfi )β . Of course, f3 = ω(X1 , X2 ) and
X3α = −gαβ [Xf1 , Xf2 ]β |p = X2β K1βα − X1β K2βα , where [, ] denotes the commutator of vector
fields. It is straight-forward to derive the corresponding expression for the 2-form K3 ; the
result is

K3αβ = K2α γ K1γβ − K1α γ K2γβ + X1µ X2ν Rαβ µν . (3.15)

This fact suggests that we define the following operation on the set Sp of symmetry data:

27
 
[(f1 , X1 , K1 ), (f2 , X2 , K2 )]p := ω(X1 , X2 ),
X2β K1βα − X1β K2βα , 
(3.16)
K2α γ K1γβ − K1α γ K2γβ + X1µ X2ν Rαβ µν .

This bracket on Sp is defined only to mirror the Poisson bracket on O. That is, to obtain
the symmetry data corresponding to the Poisson bracket of two observables, one may simply
apply the above “bracket” operation to the symmetry data of the initial observables.
For an arbitrary point p on an arbitrary Kähler manifold, Sp is closed under [ , ]p . How-
ever, the Jacobi identity will, in general, fail. It is natural to ask for circumstances under
which Sp forms a Lie algebra. The answer to this question is provided by
Lemma III.1 [ , ]p is a Lie bracket on Sp if and only if the Riemann tensor is of CHSC at
p.
If O is maximal, then [, ]p is a Lie bracket on Sp for any p ∈ M. As a consequence of
Lemma III.1, the Riemann tensor is of CHSC at each point of M. Therefore M is a
manifold of CHSC; this is summarized by
Corollary 2 Suppose O is maximal. Then M is a manifold of CHSC.
Let us make the analogous construction for the symmetric bracket. Again, let f1 , f2 ∈ O
and let f4 denote the symmetric bracket, f4 = {f1 , f2 }+ = f1 f2 + (h̄/2)g(X1, X2 ). One can
easily show the following:

ωα β ∇β f4 = f1 X2α + f2 X1α + (h̄/2)ωα β (K1βγ X2γ + K2βγ X1γ ) , (3.17)

h i
∇α X4β = f1 K2αβ + f2 K1αβ + h̄K1γ[α ω K2β]δ + γδ
X1µ X2ν γ
h̄Rα(µν)γ − 2gα(µ gν)γ ωβ . (3.18)

Therefore it is natural to define the following commutative operation on Sp :


((f1 , X1 , K1 ), (f2 , X2 , K2 ))p := f1 f2 + h̄/2g(X1, X2 ),
f1 X2α + f2 X1α + (h̄/2)ωα β (K1βγ X2γ + K2βγ X1γ ) ,
(3.19)
f1 K2αβ + f2 K1αβ + h̄K1γ[α ω γδ K2β]δ
+
h i
X1µ X2ν h̄Rα(µν)γ − 2gα(µ gν)γ ωβ γ .

We know that if M is a projective Hilbert space this operation produces the symmetry data
determined by the symmetric (i.e. Jordan) bracket of the corresponding observables. For
the generic case however, O will not be closed under the symmetric bracket; hence we do
not expect Sp to be closed under its symmetric bracket. Therefore, the symmetric bracket
defined by Eq. (3.19) should be viewed as an operation on the space of all triples (f, X, K),
without the symmetry condition imposed on K.
The condition that Sp be closed under (, )p is even stronger than that found in
Lemma III.1:
Lemma III.2 The set Sp of symmetry data at p is closed under the symmetric bracket (, )p
if and only if the Riemann tensor is of CHSC= 2/h̄ at p.

28
The difference between this condition and that specified in Lemma III.1 is that here the
actual value of the holomorphic sectional curvature is determined by the coefficient appearing
in the definition of the symmetric bracket. Lemma III.1 states that the holomorphic sectional
curvature be constant, but does not specify its value.
Note that each of these lemmas involves only a single point of M. If the set of observables
is maximal, then by definition its elements are in one-to-one correspondence with elements
of Sp for any p ∈ M. Therefore, the closure of a maximal set of observables under the
symmetric bracket is equivalent to the closure of Sp under (, )p for any p ∈ M. If we apply
Lemma III.2 to each point of M, we immediately obtain

Corollary 3 Suppose O is maximal. Then O is closed under the symmetric bracket if and
only if the Riemann tensor is of CHSC= 2/h̄.

Corollary 2 does not specify the value of the holomorphic sectional curvature, but merely
states that the Riemann tensor assumes the special form written in Eq. (3.13) for some
constant C. However, combining it with Lemma III.2, one obtains

Theorem III.1 Suppose O is maximal and that Sp is closed under (, )p for a single point
p ∈ M. Then M is a manifold of CHSC= 2/h̄.

Of course, we would like to go one step further to say that M must be a projective
Hilbert space. Now, it is known that any two finite-dimensional Kähler manifolds which
are complete, simply-connected and of CHSC= 2/h̄ are isomorphic [38]. Therefore, in the
finite-dimensional case, we have obtained a characterization of the structure that picks out
the standard quantum kinematics from possible generalized frameworks: If the generalized
quantum phase space is a complete, simply-connected Kähler manifold and the set of observ-
ables is maximal and closed under {, }+ , then one is dealing with the structure of ordinary
quantum mechanics. This characterization should be useful in the search for genuine gener-
alizations. In a generalized theory, one or more of the conditions must be violated. Thus, the
characterization systematizes the search for generalizations and suggests concrete directions
to proceed.
The case when M is infinite-dimensional is much more interesting physically. However,
in this case, we do not know if we can conclude that a (complete) Kähler manifold on which
O is maximal and closed under {, }+ is isomorphic to a projective Hilbert space. Indeed
there may exist many different infinite-dimensional Kähler manifolds (satisfying the above
completeness requirements) of the same constant holomorphic sectional curvature. This is
an important open problem. If the situation turns out to be the same as in the finite-
dimensional case, we will again have a theorem characterizing ordinary quantum mechanics.
If the situation is different and we have many such Kähler manifolds, it would be even more
interesting. For, we would then be presented with viable generalizations of the standard
quantum formalism. Such examples would be very interesting because they are likely to
admit a consistent measurement theory and thus lead to physically complete generalizations
of a rather subtle type.

29
 IV. SEMI-CLASSICAL CONSIDERATIONS

One of the most striking features of the geometric approach to quantum mechanics is
its resemblance to the classical formalism. One might therefore expect that the geometric
framework may shed some light on the relation between quantum and classical physics and,
in particular, semi-classical approximations. We will see that this expectation is correct.
In section IV A we consider the relation between the two theories at a kinematical level
and elucidate the special role played by coherent states. These results are then used in the
remaining section to analyze dynamical issues. In IV B we consider systems such as harmonic
oscillators and free fields; in IV C, we discuss dynamics in the WKB approximation.

A. Kinematics

Let us now suppose that we are given a classical phase space (Γ, α), where Γ is as-
sumed to be a finite-dimensional vector space and α is the symplectic form thereon. Let
(P, g, ω) be the quantum phase space which results from the application of the text-book
quantization procedure to (Γ, α). Thus, on the quantum Hilbert space H, there are position
and momentum operators (Q̂i and P̂i , i = 1, ..., n) corresponding to the classical position
and momentum observables. We are all accustomed to this direction of construction. Our
goal now is to do things in “reverse”: Given the quantum phase space, we will provide a
construction of the classical one.
We will continue with our previous convention and write the projections to P of the
expectation values of these operators as qi and pi . It should be emphasized, however, that
these functions are the elementary quantum observables, not the classical ones; they are
functions on the infinite-dimensional space P and not on the finite-dimensional space Γ. Note
that by standard construction the Poisson brackets of these elementary quantum observables
satisfy the same “commutation relations” as do the classical variables:

{qi , pj } = δij and {qi , qj } = 0 = {pi, pj }. (4.1)

One of the most common approaches to the classical limit comes from a simple theo-
rem of Ehrenfest’s which suggests that expectation values of quantum operators are to be
approximated, in some sense, by the corresponding classical observables. The geometric
formulation is particularly well-suited to implement these ideas. To each quantum state
x ∈ P, let us associate the classical state (qi (x), pj (x)) ∈ Γ. This association defines the
obvious mapping ρ : P → Γ. In fact, one might view this as the definition of the classical
phase space. (The astute reader will object, correctly pointing out that one must specify
the elementary quantum observables qi and pi before such a “definition” of the classical
phase space can be made. Recall, however, that this is just the counterpart of the fact
that elementary classical observables must be specified before construction of the quantum
theory.)
For convenience, let us denote the generic elementary observables by fr , r = 1 . . . 2n (i.e.,
{fr } = {qi , pi , i = 1 . . . n}). Now let us define an equivalence relation on P: x1 ∼ x2 ⇔
fr (x1 ) = fr (x2 ) ∀r. This equivalence relation fibrates the quantum phase space, which may
now view as a bundle over the classical phase space, Γ = P/ ∼.

30
 Associated to any bundle is a “vertical” distribution; i.e., a special class of (vertical)
tangent vectors at each state x ∈ P. That a tangent vector v ∈ Tx P is vertical simply
means that v(fr ) = 0 ∀r = 1 . . . 2n. Equivalently the vertical subspace may be defined as

Vx := {v ∈ Tx P | ω(Xfr |x , v) = 0 ∀r = 1 . . . 2n}. (4.2)

The vertical directions are simply those in which the elementary quantum observables assume
constant values.
Let Vx⊥ denote the ω-orthogonal complement of the vertical subspace at x. One can show
[16] that each tangent space Tx P may be written as the sum Tx P = Vx ⊕Vx⊥ . Let us therefore
call elements of Vx⊥ horizontal. If Y is a vector field which is everywhere horizontal (resp.
vertical), we will simply write Y ∈ V ⊥ (resp. Y ∈ V). Note in particular that any algebraic
combination of the Xfr is necessarily horizontal everywhere. Let us make three preliminary
observations that will be used later. First, the distribution V is integrable since for any
v1 , v2 ∈ V, [v1 , v2 ](fr ) = 0 ⇒ [v1 , v2 ] ∈ V. Next, since Xfr (fs ) = {fs , fr }, we must have
[Xfr , v](fs ) = 0 ∀v ∈ V. Therefore, the Xfr preserve the vertical distribution. Similarly, the
Hamiltonian vector field of any algebraic function of the fr also preserves the distribution.
Lastly, since [Xfr , Xfs ] = 0 ∀r, s the horizontal spaces are also integrable. Therefore, there
exist global horizontal sections of our quantum bundle over Γ.
We are now prepared to reconstruct the classical symplectic structure from the geometry
of the quantum phase space. Let ξ and ζ be two vector fields on Γ and denote by ξ˜ and ζ̃
their horizontal lifts to P. Then the classical symplectic structure is defined by
˜ ζ̃).
α(ξ, ζ) := ω(ξ, (4.3)

We have simply defined α as the “horizontal part” of the quantum symplectic structure.
Since the classical phase space is linear, it is obvious that Eq. (4.3) correctly defines the
classical symplectic structure. However, in order to allow more general constructions, let us
see why this definition of α provides a symplectic structure on Γ. First of all, it is easy to
show that ω(ξ, ˜ ζ̃) is constant along the fibres of P. Therefore, the definition is self-consistent.
To see that α is non-degenerate, one need only notice that α(ξ, ζ) = 0 ∀ζ ⇒ ω(ξ, ˜ ζ̃) = 0 ∀ζ̃ ∈
⊥
V ⇒ ζ̃ ∈ V; by construction, ζ must therefore vanish. Finally, that α is closed is obvious
since it is the pull-back, via any horizontal section, of a closed form.
Let us summarize these results. Quantization of a classical theory with a linear phase
space is canonical, given the elementary position and momentum observables. For this case,
the quantum phase space may be naturally viewed as a bundle over the classical phase space.
The classical phase space is simply the base space of this bundle. The symplectic structure
on P naturally defines a notion of horizontality, and the classical symplectic structure is
simply the horizontal part of ω.
Incidentally, note that our arguments go a short step toward the more general theory for
which the classical phase space is not necessarily linear. Suppose that V is an integrable,
symplectic distribution on P of finite co-dimension 2n which may be specified locally by the
constancy of functions fr , r = 1 . . . 2n. Suppose also that these functions may be chosen in
such a way that their Poisson algebra closes (up to constants). Then the quotient of P by
V inherits a natural symplectic structure. It is desirable to eliminate the requirement that
the co-dimension of V be finite; otherwise treatment does not apply to, e.g., quantum field

31
theory. We believe that this assumption can be removed although a detailed analysis is yet
to be carried out.
Note that our construction provides not only a projection from P to Γ but also a class
of preferred embeddings of classical phase space into the quantum one. One may guess that
these embeddings are related, in some way, to coherent states. After all, it is well-known
that spaces of coherent states are endowed with natural symplectic structures [39,18].
Let us therefore examine the nature of the coherent state spaces from the geometric
point of view. There are a number of different constructions of coherent states. We consider
the attractive and fairly general approach introduced by Perelomov [19], Gilmore [20] and,
to some extent much earlier, by Klauder [21]. Consider the the Heisenberg-Weyl group,8
obtained by exponentiating the Lie algebra generated by Q̂i , P̂j and the identity operator
on H. One defines Perelomov’s space of generalized coherent states by the action of this
group on an arbitrary element Ψ0 ∈ H. In this manner, one obtains generalized coherent
states Ψ(qi′ ,p′i ) := exp[− h̄i i (qi′ P̂i − p′i Q̂i )] Ψ0, labeled by pairs of parameters (qi′ , p′i) with
P

dimensions of position and momenta, respectively. If one chooses for Ψ0 the ground state
of the oscillator, one recovers the space of standard coherent states. However, one may
apply the above construction using arbitrary Ψ0 as the fiducial state. Hence, the notion of
generalized coherent states is a viable one for the generic system.
Let us now return to the relation between the quantum and classical phase spaces. Given
an arbitrary element x0 ∈ P of the quantum phase space, one can obtain a sub-manifold of
generalized coherent states as follows: Choose a state Ψ0 which projects to x0 and construct
the generalized coherent states via the standard method described above. To begin with,
these states constitute a genuinely non-linear subspace of the Hilbert space. Nonetheless,
we may project the entire space to P. An understanding of the nature of the resulting
sub-manifold of P is provided by the following facts. First, the expectation values of the
basic operators at the coherent states are given by

qi (x(q′ ,p′ ) ) = qi (x0 ) + qi′ and pi (x(q′ ,p′) ) = pi (x0 ) + p′i , (4.4)

(so that if we choose for Ψ0 the ground state of the harmonic oscillator, the expectation
values are precisely q ′ and p′ .) Second, any two coherent states generated by the same
fiducial element x0 possesses the same uncertainties. Therefore, the uncertainties in qi and
pj are constant on the generalized coherent state space, and hence on the sub-manifold of
P they define. Finally, note that the definition of the Heisenberg-Weyl group implies:
∂ 1
Ψ (q ′ ,p′ ) = P̂ Ψ(q′ ,p′) , (4.5)
∂q ′ ih̄
∂ 1
′
Ψ(q′ ,p′ ) = − Q̂Ψ(q′ ,p′ ) . (4.6)
∂p ih̄
Projecting this result to P, one immediately verifies that each space of generalized coherent
states is everywhere horizontal. Thus, our horizontal sections on P are precisely the gen-

8 Although this group is kinematical in origin, in the literature on coherent states, it is often
referred to as the dynamical group associated with the simple harmonic oscillator.

32
eralized coherent state spaces! Hence, it follows in particular that the uncertainties in the
elementary quantum observables are constant on the horizontal sections.
Finally, it is interesting to note that our construction of the horizontal sections did not
refer to the Heisenberg group. Given a general system, Perelomov’s generalized coherent
state spaces are constructed as above, but by use of a different (“dynamical”, see footnote
8) group; an entirely different set of generalized coherent state spaces will then result, which
will not correspond to horizontal sections. This seems an a point worth further investigation.

B. Dynamics: Oscillators

Let us now consider dynamical issues. Since the horizontal spaces are integrable, the
classical phase space may be embedded into P in infinitely many ways. Is there a preferred
horizontal section? The answer to this question will involve the dynamics of the system. In
general, we expect this issue to be far from trivial. The case of the harmonic oscillator is,
as one may guess, fairly simple. In this sub-section we will restrict ourselves to this case.
Recall the elementary treatment of the classical limit in terms of Ehrenfest’s theorem.
The rate of change of the expectation value hF̂ i of any observable operator F̂ is given by
d 1 Dh iE
hF̂ i = F̂ , Ĥ , (4.7)
dt ih̄
where Ĥ is the Hamiltonian operator. As usual, let f and h denote the corresponding
observable functions on P. Then as a result of Eq. (2.14) and its subsequent discussion,
Eq. (4.7) directly translates to
df
= {f, h}. (4.8)
dt
The fact that this equation exactly mirrors the classical expression is, however, deceiving
because the functional form of the quantum Hamiltonian h is typically entirely different
from that of the classical Hamiltonian. For example, in general, one may not even be able
to express h in terms of qi and pi alone. In fact, this is the case already for the harmonic
oscillator. At first, this seems puzzling because, as is well-known, the qi and pi “follow the
classical trajectories” if the Hamiltonian operator is a quadratic function of Q̂i and P̂i .
Let us therefore pursue this a bit further. For simplicity, consider the one-dimensional
case. First, we must obtain the form of the quantum Hamiltonian. This is easy. Since Ĥ =
(1/2m)P̂ 2 + (mω 2/2)Q̂2 = (1/2m){P̂ , P̂ }+ + (mω 2 /2){Q̂, Q̂}+ , the Hamiltonian function on
P must be of the form
1 2 mω 2 2 1 mω 2
h= p + q + (∆p)2 + (∆q)2 , (4.9)
2m 2 2m 2
where we have used only the definition and properties of the symmetric bracket. Cross-terms
involving the uncertainties—and hence the total Hamiltonian h—can not be expressed in
terms of qi and pi alone. More generally, such cross terms are by-products of the quantization
process and provide one of the significant differences between the forms of the classical and
quantum dynamics.

33
 Notice that the Hamiltonian may be decomposed into two parts; hence, so may the
corresponding Hamiltonian vector field. Xh = Xh0 + Xh∆ where h0 takes the form of
the classical Hamiltonian and h∆ = (1/2m)(∆p)2 + (mω 2 /2)(∆q)2 . Now, recall that the
uncertainties ∆q and ∆p are constant in the horizontal directions. As a consequence, Xh∆
must be purely vertical. This should not be at all surprising since, as we already know, time-
dependence of q and p is the same as for the classical observables. Since Xh∆ is everywhere
vertical, the pull-back s∗ h of the quantum Hamiltonian, via any horizontal section s : Γ → P,
to the classical phase space is precisely the classical Hamiltonian, up to a physically irrelevant
overall constant. This is essentially the reason why the evolution of the basic quantum
observables agrees with the classical evolution. However, the quantum evolution is quite
different from the classical, for it does not generally preserve the horizontal sections. A
natural question arises: is there a horizontal section which is preserved by the Hamiltonian
evolution?
Note that we are asking for much more than a dynamical trajectory confined to one
horizontal section. The question is whether there exists an entire horizontal section of P
which is preserved by the quantum evolution. A section s0 : Γ → P is of this special nature
if and only if Xh∆ vanishes on the entire image of s0 . Equivalently, we may search for a
section s0 on which the uncertainty term h∆ attains a local extremum. It seems a rather
natural guess that this will be the case at states which saturate the uncertainty relation
between q and p. This expectation is correct. One can see this by writing
 q q 2
(1/ωh̄)h∆ = ∆q mω/2h̄ − ∆p 1/2mωh̄ + (1/h̄)∆p∆q ≥ 1/2, (4.10)

where the inequality is due to Heisenberg. Therefore, any state x at which h∆ (x) = ωh̄/2
extremizes the uncertainty term. It is now easy to see that the only values of ∆p and ∆q
(subject, of course, to Heisenberg’s uncertainty relation) which extremize h∆ are given by
(∆q)2 = h̄/2mω and (∆p)2 = mωh̄/2. There is therefore a single horizontal section which is
preserved by the Hamiltonian evolution; it is the section on which the quantum evolution is
“most classical”. As one might suspect, this preferred section corresponds to the standard
coherent state space, generated by the oscillator’s ground state. These results clearly hold for
any finite-dimensional oscillator, and likely for the infinite-dimensional case (e.g., quantum
Maxwell theory).

C. Dynamics: WKB approximation

The previous discussion of dynamics has been, to some extent, in the context of the
Ehrenfest approach to semi-classical dynamics. Let us also consider the problem from the
point of view of Hamilton-Jacobi theory—the context in which one often introduces the
WKB approximation. While the discussion of section IV B has been limited to the special
case of the oscillator, we now consider more general systems and obtain two main results.
First, we will present interesting condition for the validity of the WKB approximation which,
to our knowledge, has not been discussed in the literature. Second, we will show that the
WKB equation actually corresponds to a Hamiltonian evolution on the projective space,
and therefore defines a generalized quantum dynamics of the Weinberg type.

34
 We will consider the dynamics of a non-relativistic particle moving in Rn under the
influence of a general conservative force (see footnote 6). The wave function is therefore
assumed to satisfy the Schrödinger equation
h̄2
!
∂
ih̄ Ψ(x, t) = − ∆ + V(x) Ψ(x, t), (4.11)
∂t 2m
where ∆ is the Laplace operator on Rn . In the spirit of our approach, we decompose
the state vector√into real and imaginary parts. This defines two real fields φ and π via
Ψ = (φ + iπ)/ 2h̄. In terms of these fields, the metric and symplectic structure on H
assume the forms
Z
G((φ1 , π1 ), (φ2 , π2 )) = dnx [φ1 (x)φ2 (x) + π1 (x)π2 (x)] , (4.12)
Z
Ω((φ1 , π1 ), (φ2 , π2 )) = dnx [φ1 (x)π2 (x) − φ2 (x)π1 (x)] . (4.13)

The second equation may look familiar from classical field theory. It implies that the fields
φ and π are canonically conjugate; one may view them, respectively, as the ‘field’ and
‘momentum density’ of the classical field theory under consideration. This is not surprising
because, after all, any quantum theory can be regarded as a field theory.
Indeed, one may calculate the expectation value of the above Hamiltonian operator,
h̄2  ~ 2
" #
1
Z
~ 2 + V (x) φ2 (x) + π 2 (x) ,
  
n
H(φ, π) = dx (∂φ) + (∂π) (4.14)
2h̄ 2m
and verify that the fields evolve according to the canonical equations of motion:
∂φ δH ∂π δH
= and =− . (4.15)
∂t δπ ∂t δφ
We are interested in the relationship between these equations of motion provided by quantum
dynamics and the Hamilton-Jacobi equation.
√
For this it is convenient to express Ψ as Ψ = ρ exp(iS/h̄) and rewrite Schrödinger
dynamics as equations of motion for ρ and S:
√
∂S 1 ~ 2 h̄2 ∆ ρ
+ (∂S) + V (x) = √ , (4.16)
∂t 2m 2m ρ
∂ρ ~ ~ = 0.
m + ∂ · (ρ∂S) (4.17)
∂t
The second equation is the conservation equation, ∂ρ/∂t + ∂~ · J~ = 0, where J~ = ρ∂S/m ~ is
the probability current density. The familiar observation is the fact that Eq. (4.16) becomes
the Hamilton-Jacobi equation upon dropping the term involving h̄; in a loose sense, one
obtains the classical limit upon taking h̄ → 0. (For details, see standard texts, e.g., [40,41].)
Let us explore the standard approach in which one simply drops the “quantum correc-
tion” to the Hamilton-Jacobi equation. To be concrete, let us refer to the corresponding
evolution as WKB evolution. First, one may express the Hamiltonian function in terms of
ρ and S:

35
 h̄2
" #
~ 2 + 1 ρ(x)(∂S)
Z
H(ρ, S) = dxn
(∂ρ) ~ 2 + ρ(x)V (x) . (4.18)
8mρ(x) 2m

It may be seen that ρ and S are also canonically related and that their equations of motion
may be obtained from Eqs. (4.15) by making the replacements φ → ρ and π → S. One may
also verify that the WKB evolution corresponds to dropping the first term in Eq. (4.18).
That is, if Hh̄ denotes this first term, and HW KB := H − Hh̄ , then the WKB evolution
corresponds to integration of the Hamiltonian vector field XW KB generated by HW KB .
Let us first examine the condition of validity of the the WKB approximation. We require
conditions under which the Hamiltonian vector field generated by Hh̄ is small compared to
that generated by H. To this end, we compute the functional derivatives of Hh̄ with respect
to the fields φ and π. We obtain
δHh̄ 1
 
~ 2 φ;
~ − π ∂φ
 
=− 2h̄ρ (φ∆φ + π∆π) + φ ∂π (4.19)
δφ 8mh̄ρ2
the corresponding expression for δH/δπ is obtained by making the replacement φ ↔ π
above. This expression may be written in terms of quantities which are somewhat more
physical. One need only notice that
~ ~ − π ∂φ),
J~ = 21 (φ∂π ~
D E
P̂ = dnx mJ(x),
R
D E (4.20)
P̂ 2 = dnx K(x), K = h̄2 (φ∆φ + π∆π).
R

The quantities mJ~ and K which appear in Eq. (4.19) may be interpreted physically as the
“density of momentum” and the “density of squared-momentum”.
In terms of these quantities, the above functional derivative may be written as
δH 1 h ~2 i
=− (mJ ) − ρK φ. (4.21)
δφ 2mh̄ρ2
Therefore,
!
δHh̄ δHh̄ i  ~ 2 
ih̄ −i (mJ)=− ρK (φ + iπ). (4.22)
δπ δφ 2mρ2
√
The WKB evolution is obtained by subtracting ((1/ 2h̄ times) this term from the
Schrödinger equation. The WKB equation of motion may therefore be written
∂ 1 h ~ 2 i
ih̄ Ψ = ĤΨ + (mJ) − ρK Ψ, (4.23)
∂t 2mρ2

where Ĥ is the unaltered Hamiltonian operator. Hence, the condition of validity of the
WKB equation is that the second term on the right side of Eq. (4.23) be small; i.e. that
the “density of squared-momentum” (weighted by the probability density) be comparable
in magnitude to the square of the “density of momentum”.
We have just arrived at an interesting form of the WKB equation of motion. One
should compare it with the general form given by Eq. (3.8) of the non-linear Schrödinger

36
equation. Note that the additional term α appearing in that equation is a function of only
the amplitude ρ. Hence, the WKB equation appears to be considerably more complicated
than the explicit forms of generalized dynamics that are generally considered.
However, we have already seen (c.f. the discussion following Eq. (4.18)) that the WKB
evolution is generated by a Hamiltonian function on H. Surprisingly, this evolution actually
induces a Hamiltonian flow on the projective Hilbert space. To see this, in the light of section
III A, we need only show that XW KB preserves the unit sphere. This is the case if and only
if the Poisson bracket of HW KB and the constraint function, C = dnxρ(x) − 1 vanishes
R

at each normalized state. This fact may be proven as follows: Since Hh̄ is independent of
S, {Hh̄ , C} = 0. Moreover, since H generates the Schrödinger evolution, which preserves
the unit sphere, {H, C} = 0. Taking the difference between these two Poisson brackets,
we obtain the desired result: {HW KB , C} = 0. This Poisson bracket actually vanishes
strongly; therefore, the WKB evolution actually preserves the norm of all state-vectors (not
just the unit vectors). Therefore, the dynamics of the WKB approximation is an example
of generalized dynamics of the Weinberg type; in particular, it defines a Hamiltonian flow
on the projective Hilbert space.

V. DISCUSSION

Let us begin with a summary of the main results.

We first showed that ordinary quantum mechanics can be reformulated in a geometric
language. In particular, as in classical mechanics, the space of physical states—the “quantum
phase space”—is a symplectic manifold and dynamics is generated by a Hamiltonian vector
field. However, unlike in classical mechanics, the quantum phase space is equipped also
with a Kähler structure. As one might suspect, the Riemannian metric, which is absent in
the classical description, governs uncertainty relations and state vector reduction which are
hallmarks of quantum mechanics. The geometric formulation shows that the linear structure
which is at the forefront in text-book treatments of quantum mechanics is, primarily, only a
technical convenience and the essential ingredients—the manifold of states, the symplectic
structure and the Riemannian metric—do not share this linearity. Therefore, the framework
can serve as a stepping stone for non-linear generalizations of quantum mechanics.
One can consider such generalizations in various directions. A “conservative” approach
would retain the kinematical structure and generalize only the dynamics. This is the view-
point that underlies the various non-linear generalizations of the Schrödinger equation that
have been considered in the literature. The strategy is also natural and easy to implement in
the geometric formulation: While in the standard formulation, dynamics can be generated
by a very restricted class of functions on the quantum phase space, we can generalize the
framework by allowing the Hamiltonian flow to be generated by any smooth function on the
quantum phase space. We saw that the known proposals of generalized dynamics fall in this
class. Furthermore, by concentrating on the quantum phase space—rather than the fiducial
Hilbert space—one can separate essential features of dynamics from inessential ones. In
particular, this naturally led to a clarification of the relation between Weinberg’s framework
and non-linear Schrödinger equations and corrected some misconceptions.
The kinematical generalizations are more difficult. However, we were able to streamline

37
the search for such generalizations by providing a characterization of ordinary quantum me-
chanics using structures that have direct physical interpretation: In essence, what singles
out quantum mechanics is that the underlying Kähler manifold has maximal symmetries.
From the geometric viewpoint then, restricting oneself to ordinary quantum mechanics is
rather analogous, in (pseudo-)Riemannian geometry, to working only with manifolds of con-
stant curvature. Now, in (pseudo-)Riemannian geometry, a rich theory remains even if one
drops the restriction of maximal symmetries. Indeed, in the context of general relativity,
the manifolds of constant curvature are only the ‘vacua’ and in most physically interesting
situations, there are significant departures from these geometries. Is the situation perhaps
similar in the case of quantum mechanics? Have we been restricting our attention only to
the most elementary of viable theories?
The geometrical formulation is also useful to probe semi-classical issues. We saw in
particular, that the quantum phase space has a natural bundle structure and the horizontal
cross-sections correspond precisely to families of generalized coherent states. It also provides
a succinct and clear condition for validity of the WKB approximation. Furthermore, it
turned out that dynamics in the WKB approximation yields a well-defined flow on the
quantum phase space which corresponds to a “generalized quantum dynamics” in the sense
of Weinberg.
The quantum (Hilbert space and hence) phase space is finite-dimensional only in ex-
ceptional cases, such as spin systems. Most work in the literature in the area of geometric
formulations of quantum mechanics deals only with this case. By working with Hilbert
manifolds, we were able to treat the generic case—such as particles moving in Rn —where
the quantum phase space is infinite-dimensional. Finally, most of our results go through
also in quantum field theory (although the measurement postulates are, as is usual, geared
to non-relativistic quantum mechanics.) Indeed, the geometric treatment sheds new light
on the second quantization procedure. Because the space of quantum states is itself a sym-
plectic manifold equipped with a Kähler structure, it turns out that one can use (a natural
infinite dimensional generalization of) the machinery of geometric quantization to carry out
quantization again. The resulting theory is precisely the second quantized one. Thus, second
quantization is indeed (quantization)2 ! (For details, see [16].)
We will conclude by listing a few of the important open problems. First, we have given
an intrinsically geometric formulation of the five postulates of quantum mechanics that deal
with kinematics, unitary evolution and measurements of observables (possibly with contin-
uous spectra). However, we did not include the spin-statistics postulate. The reason is that
we do not have a succinct formulation of this postulate which refers only to the essential
geometric structures. Obtaining such a formulation is a key open problem. The remaining
issues deal with generalizations of the standard framework. We saw that it is rather straight-
forward to extend dynamics by allowing the Hamiltonian to be any densely-defined function
on the quantum phase space. However, unless it is an observable function in the sense of
definition II.1, we may not have a consistent measurement theory for it. Whether this is a
problem is not so clear. Indeed, this feature arises already for the non-linear Schrödinger
equations considered in the literature and there it is generally not perceived as a problem.
However, a systematic analysis of this issue should be carried out. Next, as indicated in
section III B, there is possibility that there exist infinite-dimensional Kähler manifolds with
constant holomorphic sectional curvature (= 2/h̄) which are not isomorphic to a projective

38
Hilbert space. If this does turn out to be the case, we will obtain viable, non-trivial general-
izations of quantum kinematics for which even the measurement theory could be developed
in detail. Therefore, it is important to settle this issue. Also, even in the finite-dimensional
case, we do not know if there exist any Kähler manifolds other than projective Hilbert spaces
for which a satisfactory measurement theory can be developed. Even isolated examples of
such manifolds would be very illuminating. The final issue stems from the fact that the
space of quantum states shares several features with the classical phase space. Is there then
a quantization procedure to arrive directly at (P, ω, g) without having to pass through the
Hilbert space? Not only may the answer be in the affirmative but the new procedure may
even be useful in cases when the standard quantization procedure encounters difficulties.
That is, such a procedure may itself suggest generalizations of quantum mechanics.

ACKNOWLEDGMENTS

We would like to thank Lane Hughston for correspondence, Ted Newman for suggesting
that we examine the WKB approximation and Domenico Guilini for pointing out some
early references. This work was supported in part by the NSF grants PHY93-96246 and
PHY95-14240 and by the Eberly fund of The Pennsylvania State University.

39
 REFERENCES
[1] G.C. Ghirardi, A. Rimini and T. Weber, “Unified dynamics for microscopic and macro-
scopic systems”, Phys. Rev. D34, 470–491 (1986).
[2] P. Pearle, “Combining stochastic dynamical state reduction with spontaneous localiza-
tion”, Phys. Rev. A39, 2277-2289 (1989).
[3] G.C. Ghirardi, R. Grassi and A. Rimini, “continuous spontaneous reduction involving
gravity”, Phys. Rev. A42, 1057-1064 (1990).
[4] J.S. Bell, “Introduction to the hidden-variable question” in Proceedings of the interna-
tional school of physics Enrico Fermi, course IL: Foundations of quantum mechanics,
(Academic Press, New York, 1971).
[5] I. Bialynicki-Birula and J. Mycielski, “Nonlinear wave mechanics”, Ann. Phys. (N.Y.)
100, 62–93 (1976).
[6] P. Pearle, “Reduction of a state vector by a non-linear Schrödinger equation”, Phys.
Rev.D13, 857-868 (1976).
[7] S. Weinberg, “Testing quantum mechanics”, Ann. Phys. (N.Y.) 194, 336–386 (1989).
[8] P. Pearle, “Towards a relativistic theory of state vector reduction”, in Sixty-two years
of uncertainty, Ed A. I. Miller (Plenum Press, New York, 1990).
[9] J. S. Bell Speakable and unspeakable in quantum mechanics, (Cambridge University
Press, Cambridge, 1987).
[10] A. Ashtekar and J. Stachel, Conceptual problems of quantum gravity, (Birkhäuser,
Boston, 1991).
[11] K. V. Kuchař, “Time and interpretations of quantum gravity”, in Proceedings of the 4th
Canadian conference on general relativity and relativistic astrophysics, Ed G. Kunstat-
ter, D. Vincent and J. Williams (World Scientific, Singapore, 1992).
[12] C. J. Isham, “Canonical quantum gravity and the problem of time”, in Integrable sys-
tems, quantum groups and quantum field theory, Ed. L.A. Ibart and M.A. Rodrigues
(Kluwer, Dordrecht, 1992).
[13] R. Penrose, Emperor’s new mind: Concerning computers, minds and the laws of physics
(Oxford University Press, Oxford 1989).
[14] M. Gell-Mann and J.B. Hartle, “Classical equations for quantum systems”, Phys. Rev.
D47, 3345-3382 (1993).
[15] C. J. Isham, “Quantum logic and the histories approaches to quantum theory”, J. Math.
Phys. 35, 2157-2185 (1994).
[16] Troy A. Schilling, Geometry of quantum mechanics, doctoral thesis, (The Pennsylvania
State University 1996).
[17] T.W.B. Kibble, “Geometrization of Quantum Mechanics”, Commun. Math. Phys. 65,
189–201 (1979).
[18] A.M. Perelomov, Generalized coherent states and their applications, (Springer-Verlag,
Berlin, New York, 1986).
[19] A.M. Perelomov, Commun. Math. Phys. 26, 222 (1972).
[20] R. Gilmore, Ann. Phys. (N.Y.) 74, 391 (1972).
[21] J.R. Klauder, J. Math. Phys. 4, 1055 (1963), J. Math. Phys. 4, 1058 (1963).
[22] R. Penrose and W. Rindler, Spinors and space-time, vol 1 (Cambridge University Press,
Cambridge, 1985).

40
[23] A. Ashtekar, G.T. Horowitz and A. Magnon-Ashtekar, “A generalization of tensor cal-
culus and its applications to physics”, Gen. Rel. Grav. 14, 411–428 (1982).
[24] A. Heslot, “Quantum mechanics as a classical theory”, Phys, Rev. D31, 1341–1348
(1985).
[25] J. Anandan and Y. Aharonov, “Geometry of quantum evolution”, Phys. Rev. Lett. 65,
1697–1700 (1990).
[26] G.W. Gibbons, “Typical states and density matrices”, Jour. Geom. Phys. 8, 147–162
(1992).
[27] L.P. Hughston, “Geometric aspects of quantum mechanics”, in Twistor Theory, Ed.
S.A. Huggett, (Marcel Dekker, 1995).
[28] L.P. Hughston, “Geometry of stochastic state vector reduction”, Proc. R. Soc. Lond.
A452, 953–979 (1996).
[29] R. Cirelli, A. Manià and L. Pizzocchero, “Quantum mechanics as an infinite-dimensional
Hamiltonian system with uncertainty structure: Part I”, J. Math. Phys. 31, 2891–2897
(1990).
[30] R. Cirelli, A. Manià and L. Pizzocchero, “Quantum mechanics as an infinite-dimensional
Hamiltonian system with uncertainty structure: Part II”, J. Math. Phys. 31, 2898–2903
(1990).
[31] P.R. Chernoff, J.E. Marsden, Properties of infinite-dimensional Hamiltonian systems,
(Springer-Verlag, Berlin, Heidelberg, 1974).
[32] R. Shankar, Principles of quantum mechanics, Chapter 9, (Plenum Press, New York,
1980).
[33] P.A.M. Dirac, Lectures in quantum mechanics (Yeshiva University Press, New York,
1964).
[34] R. Geroch, “A method for generating solutions of Einstein’s equations”, J. Math. Phys.
12, 918–924 (1971).
[35] R.M. Wald, General relativity, (The University of Chicago Press, Chicago, 1984).
[36] A. Ashtekar and A. Magnon-Ashtekar, “A technique for analyzing the structure of
isometries”, J. Math. Phys. 19, 1567–1572 (1978).
[37] M. Reed and B. Simon, Methods of modern mathematical physics, v. I, Functional anal-
ysis, (Academic Press, 1980).
[38] K. Yano, Structures on manifolds, (World Scientific, Singapore, Philadelphia, 1984).
[39] W.-M. Zhang, “Coherent states: Theory and some applications”, Rev. Mod. Phys. 62,
867–927 (1990).
[40] H. Goldstein, Classical Mechanics, (Addison-Wesley, 1981).
[41] C. Lanczos, The variational principles of mechanics, (Dover, 1970).

41