Baex J at Spin Networks in Nonperturbative Quantum Gravity

Spin Networks in Nonperturbative Quantum Gravity
John C. Baez
Department of Mathematics
University of California
arXiv:gr-qc/9504036 v1 21 Apr 1995
Riverside CA 92521
[email protected]
April 15, 1995
To appear in the proceedings of the AMS Short Course

on Knots and Physics, San Francisco, Jan. 2-3, 1995
Abstract
A spin network is a generalization of a knot or link: a graph
embedded in space, with edges labelled by representations of a Lie
group, and vertices labelled by intertwining operators. Such objects
play an important role in 3-dimensional topological quantum field the-
ory, functional integration on the space A/G of connections modulo
gauge transformations, and the loop representation of quantum grav-
ity. Here, after an introduction to the basic ideas of nonperturbative
canonical quantum gravity, we review a rigorous approach to func-
tional integration on A/G in which L2 (A/G) is spanned by states la-
belled by spin networks. Then we explain the ‘new variables’ for gen-
eral relativity in 4-dimensional spacetime and describe how canonical
quantization of gravity in this formalism leads to interesting applica-
tions of these spin network states.
1 Introduction
Spin networks are a generalization of knots and links, and in what follows we
would like to describe some of their recent applications to quantum gravity,
and also a little bit of their role in knot theory. Knot theory is a well-
established branch of mathematics, full of interesting theorems, and it is
easy as a mathematician to acquaint oneself with it by reading any of a num-
ber of texts. Quantum gravity, on the other hand, is more difficult for the
1
mathematician, consisting as it does mainly of unsolved and usually impre-
cisely posed questions. Thus it seems worthwhile to start with a thumbnail
sketch of quantum gravity, focusing on an approach known as ‘nonperturba-
tive canonical quantization’.
After this heuristic introduction, we give a mathematically rigorous ac-
count of spin networks in Section 2. For us, spin networks will be graphs
embedded in a manifold S, with edges labelled by representations of a Lie
group G and with vertices labelled by intertwining operators. Spin net-
works define gauge-invariant functions on the space A of connections on any
G-bundle over S, and we shall show that these functions span the gauge-
invariant subspace L2 (A/G) of a certain Hilbert space L2 (A). In Sections 3
and 4 we go back to the physics and give a more detailed description of the
role spin networks play in quantum gravity.
In the latter part of the twentieth century, much energy has been spent
in a struggle to reconcile two brilliant accomplishments of the early part of
the century: general relativity and quantum theory. So far there is remark-
ably little to show for all this work when it comes to verified predictions of
experimental results. Indeed, it is quite possible that the new predictions of
a theory of quantum gravity can only be tested at extremely small length
scales, far below those that can be probed by current experimental tech-
niques. The reason for this is simple dimensional analysis: from Planck’s
constant h̄, Newton’s gravitational constant κ, and the speed of light c one
can form a quantity with units of length, the Planck length
ℓP = (h̄κ/c3 )1/2 ,
in a unique way (up to dimensionless constant factors). This works out to

be about 10−35 meters. If we are hoping to get experimental evidence for
any theory of quantum gravity in the forseeable future, we have to hope that
somehow this simple argument is wrong. There are certainly many phenom-
ena already observed, but so far unexplained, that a theory of quantum grav-
ity might hope to ‘retrodict’: phenomena from particle physics, phenomena
from cosmology, even phenomena such as the 4-dimensionality of spacetime
that are taken for granted in all current theories of physics. Physicists have
suggested many ideas along these lines, but none command widespread ac-
ceptance at this time; they are all somewhere between controversial theories
and sheer lunacy.
2
In short, if there were several competing well-established theories of quan-
tum gravity, the lack of empirical evidence to decide between them could
easily be a serious problem. Luckily nature has been kind to us, in that
we have been unable to formulate even one theory combining general rela-
tivity and quantum theory in a manner that is mathematically consistent,
not in obvious contradiction with experiment, and elegant. The importance
of mathematical consistency and non-contradiction with experiment should
be obvious, but the role of elegance deserves some comment. First, it is al-
ways possible to find infinitely many consistent theories that fit a given finite
set of experimental results, just as there are infinitely many curves through
a finite set of points, and without the freedom to reject most of them as
appallingly ugly, we would never get anywhere. Second, there is the fact
that taken separately, general relativity and quantum theory are strikingly
esthetic as pure mathematics. This could be a mere coincidence, but in the
absence of evidence to the contrary it is probably best to take it as a hint,
and search for a theory combining the most beautiful aspects of both in an
integral manner. Indeed, though the ‘unreasonable effectiveness’ of elegant
mathematics is still a great mystery, it is also an empirical fact, so even as
hard-nosed empiricists we should heed it.
Many physicists, particularly in string theory, have argued that it is im-
possible to reconcile general relativity and quantum theory without taking
all the other forces and particles into account. Here however we concentrate
on a diametrically opposite approach, which tries to construct a quantum
version of the vacuum Einstein equations — the theory of gravity with no
matter around. We shall not attempt to discuss the relative merits of the two
approaches, but simply note the curious fact that some of the same mathe-
matics shows up in both, possibly for some deep reasons that would be good
to understand [11].
In what follows we need to assume a nodding acquaintance with some
differential geometry. Let us take as our spacetime a 4-manifold M diffeo-
morphic to R × S, where R represents time and the 3-manifold S (compact
and orientable to simplify the discussion) represents space. The vacuum Ein-
stein equation concerns a Lorentzian metric g on M. Given such a metric
there is a unique metric-preserving, torsion-free connection Γ, the Levi-Civita
connection. The curvature of this connection is described by the Riemann
tensor R, which in local coordinates is written Rα βγδ . A certain amount of
information about the curvature of spacetime is thus contained in the Ricci
3
tensor Rµν = Rα µαν , and the vacuum Einstein equation says simply that
Rµν = 0.
To understand why precisely these are the equations describing the cur-
vature of empty spacetime, though, it is necessary to step back and consider
the case where there is matter present. In the presence of matter, the density
of energy and momentum and their rate of flow in different spatial directions
are summarized by a symmetric tensor Tµν . The sense in which energy and
momentum are conserved in general relativity is quite subtle, but the simplest
way of stating this conservation principle is that Tµν is ‘divergence-free’:
∇µ Tµν = 0,
where ∇ is the operator of covariant differentiation corresponding to the

Levi-Civita connection. If we wish to say that spacetime is curved by energy
and momentum, it is natural to try some equation like
Cµν = Tµν
where C is built up from the Riemann tensor in some simple way. However,
are constrained by the fact that C must be symmetric and divergence-free.
The simplest thing to try is some multiple of
1
Gµν = Rµν − Rαα gµν ,
2
because this is symmetric and, by the Bianchi identity, also divergence-free.
In fact, Einstein’s equation in the presence of matter says that
Gµν = 8πκTµν .
But in a vacuum Tµν = 0, so we obtain
Gµν = 0,
which turns out to be equivalent to the vanishing of the Ricci tensor.

To see how the vacuum Einstein equation describes the dynamics of the
metric, it is useful to split spacetime into space and time. The manifold M
is diffeomorphic to R × S, but not in any canonical way. Nonetheless, let us
4
go ahead and pick a diffeomorphism between them. This lets us transfer the
usual coordinate function on R to a ‘time’ function x0 on M, so we can talk
about the metric on space at time t, that is, the metric g restricted to the
slice
St = {x0 = t} ⊆ M.
Assume we can choose the slices St to be spacelike, meaning that the re-
striction of g to each one is a Riemannian metric. We can only do this if g
is ‘globally hyperbolic’, but physically this is a reasonable condition. Then,
near any point p ∈ M, we can find local coordinates xµ such that x0 is the
above time function, the vector field ∂0 is normal to the slices, and the vector
fields ∂i corresponding to the remaining ‘space’ coordinates xi (i = 1, 2, 3)
are tangent to the slices. One does not really need to use coordinates with
these special properties, but it simplifies the discussion below.
Since Gµν is a symmetric 4 × 4 matrix, the vacuum Einstein equation
really has 10 components. Splitting spacetime into space and time lets us
interpret these components as follows. The equation G00 = 0 and the 3
equations G0i = 0 are constraints on the metric on space, gij , and its first
time derivative ġij . In other words, G00 and G0i can be computed on the
slice St in terms of the ‘initial data’ gij and ġij on that slice, and for there
to be a solution of Einstein’s equation having given initial data, the initial
data must satisfy the constraints G00 = G0i = 0. The remaining 6 equations
Gij = 0 are evolutionary equations involving the second time derivative g̈ij .
They allow us to compute how the metric evolves in time, given that the
constraints hold initially.
In fact, the constraints have a simple meaning, which turns out to be very
important. To understand this meaning, however, we need a brief detour into
classical mechanics. It is worth recalling the simplest classical mechanics
problem of all, the motion of a particle on the line in a potential V . The
particle’s position is a point q in the space R, and one calls this space of
possible positions the ‘configuration space’. The motion of the particle is
determined by its position q and velocity q̇ at time zero, but it is often
handier to work not with its velocity but with its momentum p = mq̇, where
m is its mass. The position and momentum determine a point (q, p) in the
‘phase space’ T ∗ R, since the momentum is most naturally regarded as a
cotangent vector to the configuration space. This phase space has a closed
5
nondegenerate 2-form, or ‘symplectic structure’, on it, given by
ω = dp ∧ dq.
The energy, or Hamiltonian, of the particle is a function on phase space,

given by the sum of kinetic and potential energy:
p2
H= + V (q).
2m
The Hamiltonian gives rise to a vector field vH on phase space by
ω(·, vH ) = dH.
This vector field generates a one-parameter group of diffeomorphisms of phase

space, or ‘flow’, which describes the time evolution of the particle. For short,
one says that the Hamiltonian generates time evolution. Indeed, quite often
in classical mechanics the phase space is a symplectic manifold and time evo-
lution is generated by the Hamiltonian in this manner. And quite often the
phase space is the cotangent bundle of some configuration space; cotangent
bundles are equipped with a canonical symplectic structure.
In general relativity, the quantity analogous to the ‘position’ is the met-
ric gij on space, and we shall henceforth write this as qij to emphasize the
analogy. The configuration space of general relativity is thus the space M
of all Riemannian metrics on S. The quantity analogous to the ‘velocity’ is
the time derivative q̇ij . Following certain standard recipes, the analog of the
momentum works out to be
1
pij = (det qij )1/2 (q̇ ij − q̇kk q ij ).
2
The phase space of general relativity is the cotangent bundle T ∗ M, and the
pair (qij , pij ) determines a point in this phase space.
Since one can write the constraints G00 and G0i in terms of qij and pij ,
one can think of these constraints as functions on T ∗ M. From this point of
view, they play a dual role [27]. First, as already noted, for the pair (qij , pij )
to determine a solution of Einstein’s equation, it must lie on the subspace of
T ∗ M where the constraints vanish. Second, and more subtly, the constraints
generate physically important flows on phase space.
6
For example, if we integrate G00 over S, we obtain a function on T ∗ M
which generates time evolution with respect to the time coordinate x0 . For
this reason physicists call this constraint the ‘Hamiltonian constraint’, and
use the notation
H = G00
in this context. On the other hand, if we take a vector field N on S and
integrate N i G0i over S, we obtain a function on T ∗ M generating a flow which
is the same as that induced by the one-parameter group of diffeomorphisms
of S generated by N. Physicists call this constraint the ‘diffeomorphism
constraint’, and write it as
Hi = G0i .
It is important to note, however, that both H and Hi correspond to one-
parameter groups of diffeomorphisms of spacetime: H corresponds to diffeo-
morphisms that ‘push S forwards in time’, while Hi corresponds to diffeo-
morphisms that map S to itself.
Now let us turn the problem of quantizing general relativity, that is,
guessing a quantum theory that reduces to Einstein’s equation in the limit
h̄ → 0. Again, it is good to recall the example of a particle on the line. In
the quantum theory of a particle on a line, the particle’s state is given by a
‘wavefunction’, that is, an L2 function ψ on the configuration space R. If
we assume ψ is normalized so that kψk = 1, the probability of finding the
particle in any open subset U ⊆ R is then given by
Z
|ψ(x)|2 dx.
U
Now, classically the Hamiltonian is

p2
H(p, q) = + V (q),
2m
and to quantize the particle one simply replaces p and q in this formula with
certain operators p̂, q̂ on L2 (R):
h̄ d
(p̂ψ)(x) = ψ(x), (q̂ψ)(x) = xψ(x),
i dx
obtaining an operator Ĥ, the quantum Hamiltonian:
h̄2 d2
(Ĥψ)(x) = − ψ(x) + V (x)ψ(x).
2m dx2
7
The Hamiltonian Ĥ then describes the evolution of the wavefunction in time
via Schrödinger’s equation:
d
ih̄ ψ = Ĥψ.
dt
This recipe for replacing p and q by operators p̂ and q̂ is known as ‘canoni-
cal quantization’. If we try to copy this recipe in the case of general relativity,
we quickly notice some serious problems. In the case of the particle on the
line, the ‘configuration space’ in which q lies is just the real line, and we can
define L2 (R) using Lebesgue measure. In the case of general relativity, the
configuration space is the space M of all Riemannian metrics on S. This is
an infinite-dimensional manifold, and there is no good notion of ‘Lebesgue
measure’ on such spaces, so defining the Hilbert space L2 (M) is not so easy.
Of course, this problem will arise whenever we try to canonically quantize
a theory with infinite-dimensional configuration space. Thus it occurs, not
only in quantum gravity, but in other quantum field theories. When quantiz-
ing linear equations, where the configuration space is an infinite-dimensional
vector space V, one can use an infinite-dimensional analog of a Gaussian
measure to define L2 (V), and the resulting theories make fine mathematical
sense. This strategy is not sufficient by itself to deal with nonlinear equations
such as Einstein’s equation, however. Indeed, M is not even a vector space.
Another problem is that unlike the particle on the line, whose position and
velocity at a given time are arbitrary, general relativity involves constraints.
One should take these constraints into account when quantizing the theory,
but the correct way to do so is not at all clear! Indeed, much ink has been
spilled concerning the quantization of constrained systems, and we cannot
go into all the nuances of this issue here. Instead, let us simply give an
oversimplified account of Dirac’s approach to constraints [24], as applied to
gravity by DeWitt [23] in the late 1960s.
Suppose we could somehow get ahold of a Hilbert space L2 (M). The
idea is that not all vectors in this so-called ‘kinematical’ state space represent
physical states of quantum gravity, but only those satisfying certain quantum
versions of the constraints. To describe these quantized constraints, first we
try to define operators q̂ij and p̂ij on L2 (M) by the formulas
h̄ ∂
(p̂ij ψ)(q) = ψ(q), (q̂ij ψ)(q) = qij ψ(q),
i ∂qij
8
where q ∈ M. These formulas are merely heuristic, and part of the problem
is making good enough sense of them for the task at hand. For example, in
analogy with pij and qij in the first place, p̂ij and q̂ij are really something like
operator-valued functions on the subset of the spacelike slice contained in our
coordinate chart. When one tries to be rigorous one discovers that they are
operator-valued distributions. As noted above, one can work out formulas
for the constraints H and Hi in terms of pij and qij . The idea is then to take
these formulas, and replace all appearances of pij and qij in them by p̂ij and
q̂ij , obtaining operator-valued distributions Ĥ and Ĥi . For a wavefunction
ψ ∈ L2 (M) to represent a physical state of quantum gravity, the following
quantum versions of the constraint equations should hold:
Ĥψ = 0, Ĥi ψ = 0.
There are some deep conceptual problems associated with this approach
to quantum gravity. For example, instead of Schrödinger’s equation, in which
the Hamiltonian describes the time evolution of the wavefunction, all we
have is the so-called ‘Wheeler-DeWitt equation’ Ĥψ = 0. What does this
mean? Where has the dynamics of quantum gravity gone, if physical states
are annihilated by the operator whose classical counterpart generates time
evolution? Briefly, the answer appears to be this: the constraint equations
say that ψ only describes diffeomorphism-invariant information about the
world. In other words, we cannot use ψ to answer questions about what is
happening at a certain point whose location is specified using a coordinate
system, such as ‘what is the metric at the point xµ = aµ ?’. Unfortunately, we
are not used to doing physics without asking such questions! This is known
as the ‘problem of time’ in quantum gravity [33].
Just as bad as the conceptual problems are the sheer technical problems
involved in making mathematical sense out of DeWitt’s strategy for quantiz-
ing gravity. In fact, nobody has yet succeeded. It is worth noting that there
are other constrained nonlinear equations with infinite-dimensional config-
uration spaces where many of the same problems show up: for example,
the Yang-Mills equation, various versions of which appear to describe all the
forces except gravity. Quantizing these in a rigorous way is also extremely dif-
ficult, but physicists have had much more success with them at the practical
level. Let us briefly summarize the difference between the two cases.
Physicists often quantize nonlinear field theories by treating the non-
linearities as ‘small perturbations’ of some linear equation. There are also
9
perturbative methods for dealing with constraints. These methods can be
problematic, mathematically speaking, but unless afflicted by an unusual
desire for rigor, physicists are often happy to do perturbation theory using
formal power series in some coupling constant that measures the nonlinear-
ity. Due to the nonrigorous nature of these computations, the coefficients
of these power series usually come out to be ill-defined unless one performs
some clever maneuvers known as renormalization. One says the theory is
renormalizable if these maneuvers work, but renormalizability by itself does
not mean that the power series are convergent, or even asymptotic. Thus,
strictly speaking, renormalizability is not a sufficient condition for a nonlin-
ear quantum field theory to make rigorous sense. Nor, on the other hand,
is renormalizability a necessary condition to be able to quantize a nonlin-
ear wave equation, for there are also other ‘nonperturbative’ approaches. In
practice, however, particle physicists often restrict themselves to renormal-
izable theories, make physical predictions using the first few terms in the
power series, and compare these predictions with experiment to see if the
theory is on the right track.
For the Yang-Mills equation, and indeed for the whole ‘Standard Model’
of particles and forces other than gravity, this strategy has been quite suc-
cessful. But the strategy fails miserably with quantum gravity, for the theory
resists all attempts at renormalization. As it turns out, this is closely related
to the fact that the constraints are not polynomials in the basic ‘position’
and ‘momentum’ variables, qij and pij , and their derivatives. Instead, they
contain nasty factors of (det qij )−1/2 , essentially because the formula for pij
contains a factor of (det qij )1/2 . Standard methods for replacing qij with the
operator q̂ij lead to all sorts of problems when applied to such non-polynomial
expressions. (Similar problems arise in the path-integral approach which is
more commonly used in renormalization.) To overcome the nonrenormaliz-
ability of quantum gravity, particle physicists constructed ever more com-
plicated models containing gravity and other forces, in order to cancel out
unruly infinities. This led to the study of supergravity and eventually super-
string theory, which attempts to model all known forces and particles, and
unfortunately many unknown ones as well.
Meanwhile, many general relativists were suspicious of the whole idea of
perturbatively quantizing gravity. This amounts to treating all metrics as
small perturbations of a fixed ‘background metric’, usually taken to be the
flat Minkowski metric on R4 . Dimensional analysis, however, suggests that
10
one should expect more and more extreme quantum fluctuations of the metric
at smaller and smaller length scales, becoming very significant at about the
Planck length. So perhaps one should really adopt some nonperturbative
approach to canonical quantum gravity.
Saying this is easy: the problem is actually doing it. In principle, the most
direct way would be to make DeWitt’s approach into rigorous mathematical
physics without reference to a fixed background metric. But attempting
to do this led immediately into a quagmire which for two decades seemed
impassable. It was not until the late 1980s, when Ashtekar [1] found a clever
change of variables, that a way around began to seem possible.
We postpone a detailed discussion of these ‘new variables’ to Section 3.
For now, let us simply note that in terms of them, the configuration space
of general relativity space is not the space of metrics on S, but the space of
connections on some SL(2, C) bundle over S. Since SL(2, C) is a complex
Lie group, this configuration space is a complex manifold. It turns out that
in the quantum theory, kinematical states should be holomorphic functions
on this configuration space. However, using the fact that SL(2, C) has SU(2)
as a real form, one expects the kinematical state space to be isomorphic
to L2 (A), where A is the space of connections on a certain SU(2) bundle
over S. Thus, at least naively, one expects the physical states of quantum
gravity in the new variables formalism to be wavefunctions ψ ∈ L2 (A) an-
nihilated by certain constraints. These constraints include Hamiltonian and
diffeomorphism constraints as before, which express the invariance of ψ un-
der diffeomorphisms of spacetime, but also a ‘Gauss law’ constraint which
expresses the invariance of ψ under gauge transformations.
Now the configuration space of the Yang-Mills equation is also a space of
connections, and the only constraint in Yang-Mills theory is the Gauss law.
Thus in terms of the new variables, canonical quantum gravity is very similar
to SU(2) Yang-Mills theory with some extra constraints! This is one of the
main advantages of the new variables: they allow techniques from Yang-Mills
theory to be imported to quantum gravity. Another more technical advan-
tage is that in the new variables, the constraints are polynomial functions of
the analogs of position and momentum (and their derivatives). Because the
nonrenormalizability of quantum gravity in the traditional metric formula-
tion was closely related to the non-polynomial nature of the constraints, this
created a lot of excitement. Unfortunately, at the present time the Hamilto-
nian constraint still presents thorny problems.
11
Shortly after the discovery of the new variables, Rovelli and Smolin [45]
used them to develop a ‘loop representation’ of quantum gravity, and this
is when the relationship to knot theory became very apparent. The idea of
the loop representation of a gauge theory had been developed by Gambini
and Trias [31], and basically it consists of writing all the equations one can
in terms of ‘Wilson loops’. Wilson loops are certain functions on the space
of connections: at a point A ∈ A, the value of the Wilson loop is just the
trace of the holonomy of A around some loop α in S, taken in some (finite-
dimensional) representation ρ of G, written:
Z
tr(ρ(T exp A)).
α
One reason why physicists like them is that they are invariant under gauge
transformations, and one expects the physically observable aspects of a gauge
theory to be gauge-invariant.
Rovelli and Smolin argued that a state ψ of quantum gravity should give
rise to link invariant ψ̂, the ‘loop transform’ of ψ, whose value on the link
with components given by the loops α1 , . . . , αn is
Z n
Y Z
ψ̂(α1 , . . . , αn ) = tr(ρ((T exp A)) ψ(A) DA,
A i=1 αi
where ρ is the fundamental representation and DA is the purely formal

‘Lebesgue measure’ on A. The reason ψ̂ should be a link invariant is that
the diffeomorphism constraint says ψ is invariant under diffeomorphisms of
S! Of course, this argument is merely heuristic, owing to the mysterious
nature of DA, but it is no worse than much of the reasoning in quantum
gravity. Indeed, a similar sort of argument led Witten [48] to discover the re-
lation between knot theory and another quantum field theory, Chern-Simons
theory. In fact, the link invariant coming from SU(2) Chern-Simons theory,
namely the Kauffman bracket, appears to be the loop transform of a state
of quantum gravity ‘with cosmological constant’, meaning that the Einstein
equation has been modified to give the vacuum a nonzero stress-energy ten-
sor. We will not go into this here, since there are a number of expository
treatments already [11, 12, 14, 18, 41], but certainly it is one of the main
reasons for interest in the interface between knots and quantum gravity.
There is much here that needs to be made more precise in the following
sections. However, we are at least in a position now to describe what needs
12
to be done. First, we need to develop integration theory on A, in order
to escape the use of purely formal entities like the ‘Lebesgue measure’ DA.
In Section 2 we do this when A is the space of smooth connections on any
principal G-bundle P over a manifold S, where G is a compact connected Lie
group and S is real-analytic. There is an especially nice ‘generalized measure’
on A which is a substitute for the nonexistent Lebesgue measure. Using it
we can define L2 (A), and it turns out that the gauge-invariant subspace of
L2 (A) is spanned by ‘spin network states’. These are described by graphs
analytically embedded in S, with oriented edges labelled by representations
of G, and with vertices labelled by intertwining operators from the tensor
product of representations labelling ‘incoming’ edges to the tensor product of
representations labelling ‘outgoing’ edges. Equipped with this mathematical
technology, we then return to quantum gravity.
2 Spin Networks
In quantum field theory computations, physicists often try to do integrals
over a space A of connections using a strange thing they call ‘Lebesgue
measure’ on A, usually written DA. This exploits the fact that A is an affine
space, or, arbitrarily choosing one point as the origin, a vector space. The
idea is that one should be able to pick a basis for A, thus setting up an
isomorphism
A∼= R∞ ,
and then, working in the coordinates defined by this basis, let
∞
Y
DA = dxi .
i=1
There are many problems with this idea, however! First, while A can indeed
be identified with an infinite-dimensional vector space, a basis of A does not
really give an isomorphism between A and an infinite product of copies of
R. Second, there is no good theory of infinite products of arbitrary measure
spaces. Of course one can be more sophisticated about these issues, but the
fact remains that there is no such thing as ‘Lebesgue measure’ on an infinite-
dimensional vector space. Thus it is not surprising that when physicists
actually do integrals over A using ‘Lebesgue measure’ they often get infinite
or ill-defined answers until they perform various sneaky tricks.
13
On the other hand, there is a different way of doing these integrals which
is used in lattice gauge theory. Lattice gauge theory is an approximation
to gauge theory on Rn in which one replaces the continuum by an infinite
graph having the points of a lattice as vertices and the line segments between
neighboring vertices as edges. A ‘connection’ on the lattice is simply an
assigment of an element ge of the gauge group G to each edge e of the graph,
representing the effect of parallel transport along e. The space of connections
in this context is thus very different: it is
A∼
= G∞ ,
not an infinite product of copies of R, but an infinite product of copies of

G! Now, while an infinite product of arbitrary measure spaces is ill-defined,
an infinite product of probability measure spaces is well-defined. (Recall that
a probability measure space is a space X with positive measure µ such that
R
X µ = 1.) If G is compact, it is equipped with a very natural probability
measure, namely RHaar measure, the unique left- and right-invariant Borel
measure dg with G dg = 1. So in lattice gauge theory with compact gauge
group, one can work with the measure
Y
DA = dge
e
on A, instead of the nonexistent ‘Lebesgue measure’. This is precisely what

is done.
Now, while lattice gauge theory is easier to make rigorous, it has the
disadvantage of being only an approximation to the continuum theories we
are really interested in. Indeed, in practice much work in lattice gauge theory
is computational in nature. Here one uses not a lattice but a finite graph, so
that A becomes a finite product of copies of G, and one numerically calculates
integrals over A by Monte Carlo methods. To apply these results to gauge
theory on Rn one must then investigate not only the ‘continuum limit’ in
which the lattice spacing goes to zero, but also the ‘large-volume limit’.
Approximating the continuum by a lattice or a fixed finite graph is par-
ticularly distressing in the case of general relativity, which is so intimately
connected to the differential geometry of manifolds. It would be nice if one
could have ones cake and eat it too, working with connections in the contin-
uum context and preserving diffeomorphism-invariance, but still thinking of
14
the space of connections as something like a product of copies of G. This is
precisely what the theory of ‘generalized measures’ on the space of connec-
tions seeks to achieve.
The traditional approach to measure theory is to pick a σ-algebra of
‘measurable’ subsets of some space X, assign measures to them in a man-
ner satisfying some axioms, and then define a vector space of ‘integrable’
complex-valued functions on X. Then one figures out how to integrate func-
tions in this space, shows that integration is linear, and shows how to pass
limits through integrals under certain conditions. Then, typically, one forgets
the proofs of these results and simply uses them. A more modern approach
is to shortcut this process and simply choose a space Fun(X) of functions on
X, equip it with a topology, and define a ‘generalized measure’ µ on X to be
a continuous linear functional
µ: C → C,
R
writing µ(f ) as X f dµ solely out of deference to tradition. Actually, of
course, the modern approach complements the old approach rather than
replaces it; they are related by many useful theorems, some of which we have
summarized elsewhere [9].
Now suppose that P is a principal G-bundle over S, where G is compact
and connected, and let A be the space of smooth connections on P . To define
generalized measures on A we need to choose the space Fun(A) of functions
we want to integrate. Following the idea of the loop representation, for exam-
ple, we could choose Fun(A) to be the algebra of functions on A generated
by all Wilson loops, or a completion of this algebra in some topology. In
fact, this is how Ashtekar and Isham [5] proceeded in their original attempt
to use ideas from the loop representation to set up a rigorous integration
theory on A. Subsequent work by Ashtekar, Lewandowski, and the author
[6, 7, 9, 10, 37] improved things in a number of ways. For one, it turns out to
be simpler to let Fun(A) contain arbitrary continuous functions of the holon-
omy of A along paths in S. Also, working with smooth paths turns out to
be a nuisance, because they can intersect each other in horribly complicated
ways. For this reason we shall follow Ashtekar and Lewandowski and require
S to be a real-analytic manifold, and work with piecewise real-analytic paths.
In a sense this is not so drastic, because every compact smooth manifold can
be given a real-analytic structure. One pays a price, however, since one is
15
no longer doing differential geometry in the category of smooth manifolds.
Eventually it would be nice to understand the smooth case too.
In any event, let us define Fun(A) as follows. If γ: [0, 1] → S is a piecewise
analytic path, let Aγ denote the space of smooth maps F : Pγ(0) → Pγ(1) that
are compatible with the right action of G on P :
F (xg) = F (x)g.
Note that for any connection A ∈ A, the parallel transport map

Z
T exp A : Pγ(0) → Pγ(1)
γ
lies in Aγ . If we fix a trivialization of P at the endpoints of γ, we can identify

Aγ with the group G. This makes Aγ into a compact manifold in a manner
which one can check is independent of the trivialization. Let Fun0 (A) be the
algebra of functions on A generated by those of the form
Z Z
ψ(A) = f (T exp A, . . . , T exp A) (1)
γ1 γn
where γ1 , . . . , γn are piecewise analytic paths in S, and f is a continuous

function on Aγ1 × · · · × Aγn . Then let Fun(A) be the completion of Fun0 (A)
in the sup norm:
kψk∞ = sup |ψ(A)|.
A∈A
It is easy to check that the Wilson loops lie in this algebra; in fact, they lie
in Fun0 (A).
A generalized measure on A is then defined to be a continuous linear
functional on Fun(A). This may seem rather abstract, so let us see how to
get our hands on one. There exists a nice general recipe for constructing any
generalized measure on A, which we have described elsewhere [10, 12], but
for now let us concentrate on a simple and important example: the uniform
generalized measure µu . This is a kind of replacement for the nonexistent
‘Lebesgue measure’ on A, and as we shall see, it is closely modelled after the
measure DA in lattice gauge theory.
To define µu , first we define it as a linear functional on Fun0 (A). Because
we are working with real-analytic paths, it turns out that any ψ ∈ Fun0 (A)
can be written in the form given by equation (1) with the paths {γi} forming
16
an embedded graph in S. By this we mean that each path γi : [0, 1] → S is
one-to-one and restricts to an embedding of (0, 1), and the images γi [0, 1] and
γj [0, 1] intersect, if at all, only at their endpoints when i 6= j. For functions
ψ written in this special form we define
Z Z
ψ dµu = f (g1 , . . . , gn )dg1 · · · dgn .
A Gn
Here we are using trivializations of P at the endpoints of the paths γi to

identify Aγ1 × · · · × Aγn with Gn , but the right-hand side is independent of
the choice of trivializations, because Haar measure is right- and left-invariant.
All the real work goes into checking that the right-hand side does not depend
on how we wrote ψ in this special form involving an embedded graph. Given
that, it is easy to check that µu is a linear functional on Fun0 (A). The bound
Z
| ψ dµu | ≤ kψk∞
A
then holds because Haar measure is a probability measure. Since Fun0 (A) is
dense in Fun(A), this bound implies that µu extends uniquely to a continuous
linear functional on all of Fun(A), which we again call µu . This is the uniform
generalized measure on the space of connections!
If we examine what we have just done, the relationship to lattice gauge
theory should become clear. Instead of working with a fixed graph embedded
in S as one does in lattice gauge theory, we have considered all possible
embedded graphs γ = {γi} in S. Each one of these indeed has the topology
of a graph with the paths γi as edges and the endpoints γi (0), γi(1) as vertices.
Following the spirit of lattice gauge theory, we can define a finite-dimensional
space of ‘connections’ on γ,
Aγ = Aγ1 × · · · × Aγn ,
on which there is a natural measure, namely a product of copies of Haar

measure, one copy for each edge of γ. To do the integral of a function that
only depends on the holonomies along the edges of γ, we simply use this
natural measure. The key fact is that although a function ψ ∈ Fun0 (A) can
be expressed in terms of many different embedded graphs in this way — after
all, if it can be expressed in terms of a graph γ, it can also be expressed in
terms of any graph γ ′ containing γ — the answer we get for the integral of
17
ψ is independent of this choice. This is why we obtain a well-defined linear
functional µu on Fun0 (A), which then extends to all of Fun(A). In short, we
are getting a generalized measure on A as a kind of ‘limit’ of measures on
the spaces of connections on all graphs embedded in S! This can be made
perfectly precise using the language of projective limits [7].
We can now define the Hilbert space L2 (A) as the completion of Fun(A)
with respect to the norm
Z
kψk2 = ( |ψ|2 dµu )1/2 .
A
In the special case where G = SU(2) and S is a compact oriented 3-manifold,

this Hilbert space will be our space of ‘kinematical states’ for quantum gravity
in the new variables formalism. The Gauss law constraint then amounts to
requiring that the state ψ ∈ L2 (A) be invariant under gauge transformations,
so it is important to find gauge-invariant vectors in L2 (A). Finding these
vectors is actually of interest no matter what G and S happen to be. It
turns out that these vectors can also be thought of as wavefunctions on the
quotient of A by the group G of smooth gauge transformations [12], so we
will denote the space of gauge-invariant vectors in L2 (A) by L2 (A/G).
The most obvious examples of vectors in L2 (A/G) are the Wilson loops.
If we have an analytic loop α in S and a representation ρ of G, the function
Z
ψ(A) = tr(ρ(T exp A))
α
is gauge-invariant, and it lies in L2 (A). More generally, any product of

Wilson loops is a gauge-invariant element of L2 (A).
More generally still, we can get vectors in L2 (A/G) from ‘spin networks’
[13]. Take a graph γ embedded in S and label each of its edges e with a
representation ρe of G. Let
Z
He (A) = ρe (T exp A).
e
If we trivialize P at the vertices of γ, and pick a basis for ρe , we can think of

He (A) as a matrix He (A)ij . Now form the tensor product of all these matrices,
one for each edge of γ. We get a big tensor H(A) having one superscript and
one subscript for each edge; it is too ugly to bother writing down, but we
hope the reader gets the idea.
18
Next, for each vertex v of γ let S(v) be the set of edges having v as
‘source’ — where the source of an edge γi is defined to be γi (0) — and let
T (v) be the set of edges having v as ‘target’ — where the target of γi is γi (1).
For each vertex v, pick an intertwining operator
O O
Iv : ρe → ρe .
e∈T (v) e∈S(v)
We can think of Iv as a tensor with one superscript for each edge e ∈ T (v)
and one subscript for each edge e ∈ S(v). Then form the tensor product
of all these tensors Iv , one for each vertex. We get a big tensor I. Then
form the tensor product of H(A) ⊗ I. Note that each superscript of H(A)
corresponds to a particular subscript of I and vice versa, because each edge
of γ lies in S(v) for one vertex v and lies in T (w) for one vertex w. So we
can contract the tensor H(A) ⊗ I to get a number, which of course depends
on A. This is our ‘spin network state’ ψ(A). Note that a Wilson loop is just
a special case of a spin network with only one edge and one vertex, with the
intertwining operator taken to be the identity operator.
One can show directly from this explicit definition that the spin network
states are gauge-invariant and lie in L2 (A), in fact in Fun0 (A). But we want
to show more: we want to show they span the whole space of gauge-invariant
vectors in L2 (A). For this a more abstract approach is better.
The group G acts on A, and in fact there is a unitary representation of
G on L2 (A) given by
gψ(A) = ψ(g −1A).
To see this it is useful to introduce work ‘one embedded graph at a time’. One
can show that for any embedded graph γ, L2 (Aγ ) can be identified with the
smallest closed subspace of L2 (A) containing all functions of the form given
in equation (1). If we do a gauge transformation g ∈ G on the connection A,
the holonomy along any edge γi transforms to
Z Z
T exp gA = g(γi(1)) (T exp A) g(γi(0))−1 , (2)
γi γi
so if ψ lies in L2 (Aγ ), so does gψ. Thus we get a representation of G on

each subspace L2 (Aγ ), and these representations are unitary because Haar
measure is left- and right-invariant. Since the union of the subspaces L2 (Aγ )
is dense in L2 (A), we obtain a unitary representation of G on all of L2 (A).
19
Let L2 (Aγ /Gγ ) denote the gauge-invariant subspace of L2 (Aγ ). Since
the action of G on L2 (A) preserves each subspace L2 (Aγ ), the union of the
subspaces L2 (Aγ /Gγ ) must span L2 (A/G). So, what does L2 (Aγ /Gγ ) look
like? For this we need a very precise picture of the action of G on L2 (Aγ ).
Write E for the set of edges of γ, and V for the set of vertices. Then
picking a trivialization of P over the vertices of γ, we obtain an isomorphism
L2 (Aγ ) ∼
= L2 (GE )
O 2
∼
= L (G).
e∈E
To see how G acts on the right-hand side, we use the Peter-Weyl theorem.
Note that G × G acts on G by
(g1 , g2 )(h) = g1 hg2−1,
giving a unitary representation of G × G on L2 (G). The Peter-Weyl theorem

describes how L2 (G) decomposes into irreducible unitary representations of
G × G. These are all of the form ρ1 ⊗ ρ2 , where ρ1 and ρ2 are irreducible
unitary representations of G. Let R be a set containing one irreducible
unitary representation of G from each equivalence class. Then the Peter-
Weyl theorem says that
M
L2 (G) ∼
= ρ ⊗ ρ∗ .
ρ∈R
It follows that OM
L2 (Aγ ) ∼
= ρ ⊗ ρ∗ ,
e∈E ρ∈R
and in terms of this description, any g ∈ G acts on L2 (Aγ ) as the operator

OM
ρ(g(s(e)) ⊗ ρ∗ (g(t(e))).
e∈E ρ∈R
The reason is that the holonomy of A along each edge e transforms in precisely
the manner suited to applying the Peter-Weyl theorem, as we saw in equation
(2).
Next, using the associativity of tensor product over direct sum, we obtain
M O
L2 (Aγ ) ∼
= ρe ⊗ ρ∗e ,
ρ∈RE e∈E
20
where RE is the set of all labellings of edges e ∈ E by representations ρe ∈ R.
Grouping the edges by their source and target, this gives
 
M O O O
L2 (Aγ ) ∼
=  ρe ⊗ ρ∗e  .
ρ∈RE v∈V e∈S(v) e∈T (v)
In these terms any g ∈ G acts on L2 (Aγ ) as the operator

 
M O O O
 ρe (gv ) ⊗ ρ∗e (gv ) .
Thus we have
 
M O O O
L2 (Aγ /Gγ ) ∼
= Inv  ρe ⊗ ρ∗e  ,
where for any representation λ of G, Inv(λ) is the subspace of G-invariant

vectors. But Inv(λ2 ⊗ λ∗1 ) is isomorphic to the space Hom(λ1 , λ2 ) of inter-
twining operators from λ1 to λ2 , so finally we have
 
M O O O
L2 (Aγ /Gγ ) ∼
= Hom  ρe , ρe  .
ρ∈RE v∈V e∈T (v) e∈S(v)
In other words, a complete set of gauge-invariant vectors in L2 (Aγ ) is

obtained by labelling each edge of γ with an irreducible unitary representa-
tion of G and labelling each vertex of γ with an intertwining operator from
the tensor product of representations labelling incoming edges to the tensor
product of representations labelling outgoing edges. If one unravels the logic
of the proof, one sees that these are special cases of the spin network states
described more explicitly above!
Our work has shown that spin network states span L2 (A/G). But in fact
we have shown more. For each embedded graph γ we obtain an orthonormal
basis of L2 (Aγ /Gγ ) by letting ρ range over all labellings of edges by irreducible
unitary representations, and, for each ρ, picking an orthonormal basis of
 
O O
Hom  ρe , ρe 
e∈T (v) e∈S(v)
21
for each vertex v. While it is a nuisance to assemble all these orthonormal
bases into a single orthonormal basis for all of L2 (A/G), for practical com-
putations having an orthonormal basis for each graph is usually sufficient.
Of course, for really practical computations one might need a recipe for
picking an orthonormal basis of the space of intertwining operators
 
O O
Hom  ρe , ρe  .
e∈T (v) e∈S(v)
For example, Brügmann [19] has tried some computer simulations of quan-
tum gravity using spin networks, and computers are notorious for wanting
everything to be very explicit. To pick such an orthonormal basis, one needs
to understand the representation theory of G. Luckily, the representation
theory of SU(2) is quite simple, so let us consider that case. A more detailed
treatment can be found in the work of Rovelli and Smolin [46].
The irreducible unitary representations of SU(2) can be labelled by their
dimension d = 1, 2, 3, . . ., or equivalently by their ‘spin’ j = 0, 21 , 1, . . ., where
d = 2j + 1. The tensor product of two such representations decomposes as
follows:
j1 ⊗ j2 = |j1 − j2 | ⊕ |j1 − j2 | + 1 ⊕ · · · ⊕ j1 + j2 .
This means that for a trivalent graph γ we do not need to worry much about
labelling the vertices with intertwining operators. Suppose, for example,
that a vertex has two incoming edges labelled with spins j1 and j2 , and one
outgoing edge labelled j3 . Unless the Clebsch-Gordon condition
|j1 − j2 | ≤ j3 ≤ j1 + j2 , j1 + j2 + j3 ∈ Z
holds, the representation j3 will not appear as a summand of of the tensor

product j1 ⊗ j2 , and there is no way to get a nonzero spin network state from
this labelling. If the Clebsch-Gordon condition does hold, j3 appears with
multiplicity 1 in j1 ⊗ j2 , so up to a constant factor which one must fix in
some standard way, there is no choice about which intertwining operator to
pick.
Graphs that are not trivalent can be reduced to the trivalent case as
follows. Take each n-valent vertex and replace it in ones mind by a n-leaved
tree each of whose vertices is trivalent. If the original vertex was n-valent,
this tree will have n−3 new ‘internal edges’. A basis of intertwining operators
22
for the original vertex is then given by all labellings of these internal edges
by spins satisfying the Clebsch-Gordon condition. Of course, there are many
different trees with n leaves, so there is a great deal of arbitrariness in this
choice of basis. However, to change from one basis to another simply requires
repeated use of some matrices known as the 6j symbols, which are familiar
in SU(2) representation theory and whose generalizations play an important
role in topological quantum field theory [26, 36].
Before returning to a more thorough treatment of the new variables and
the use of spin networks of quantum gravity, let us say a little about the
history of spin networks. Because of the immense difficulties they have had
in quantizing gravity, physicists have often considered the possibility that
our picture of spacetime as a manifold breaks down at the Planck scale.
This has led to various attempts to reformulate physics in terms of more
discrete, or combinatorial, ideas. In fact, spin networks were invented in the
early 1970s by Penrose [39] in one such attempt. However, while our spin
networks involve graphs embedded in a pre-existing manifold that represents
space, his spin networks were abstract graphs labelled by representations
and intertwining operators. In other words, rather than serving as a tool
for describing the geometry of a spacetime manifold, his spin networks were
intended as a purely combinatorial substitute for a spacetime manifold. The
problem with this sort of radical idea has always been bridging the gap
between it and existing theories of physics. Thus the more recent discovery
that spin networks also arise naturally in attempts to quantize Einstein’s
equation is quite intriguing.
If we succeeded in constructing quantum gravity as a field theory on a
manifold, we might simply decide to forget about more combinatorial ap-
proaches to quantum gravity. However, as we shall see, the Hamiltonian
constraint poses serious problems even in the new variables formalism, so-
called ‘ultraviolet problems’, which might be due to falsely extrapolating our
picture of spacetime as a manifold to arbitrarily small length scales. Thus
it is still worth keeping in mind the possibility that some combinatorial ap-
proach is the fundamental one. For this reason, any clues about the relation
between ‘abstract’ and ‘embedded’ spin networks are likely to be interesting.
Luckily, there are quite a few clues along these lines! While Penrose’s
ideas yield a purely combinatorial recipe for defining an invariant of abstract
SU(2) spin networks, this invariant can also be described in terms of spin
networks embedded in R3 . One can obtain a generalized measure µ on the
23
space A of SU(2) connections on R3 by taking a finite Borel measure µ0 on
the subspace A0 of flat connections, and defining
Z Z
ψ(A)dµ = ψ(A)dµ0 . (3)
A A0
Then if ψ is a spin network state associated to some graph γ, the integral

R
A ψ(A)dµ does not depend on the choice of embedding of γ, essentially
because the holonomy of a flat connection along a path does not change
when the path is deformed. Thus we obtain an invariant of abstract spin
networks. If µ0 is a probability measure, this is the same invariant that
one can define combinatorially following the ideas of Penrose (while carefully
adjusting his sign conventions [46]).
A striking generalization of this idea arises from the work of Witten [48]
on Chern-Simons theory and the Jones polynomial, and its subsequent rein-
terpretation using quantum groups by Reshtekhin and Turaev [42]. Heuris-
tically, Chern-Simons theory defines an isotopy invariant of spin networks
embedded in S 3 by means of a measure on the space of connections:
Z R
ik
tr(A∧dA+ 23 A∧A∧A)
ψ(A)e 4π S3 DA (4)
A
Indeed, Witten was able by formal manipulations to compute these invariants

in certain cases. Unfortunately, unlike equation (3), this formula is difficult
to make rigorous. It involves the purely formal ‘Lebesgue measure’ DA, and
replacing DA by the uniform generalized measure does not help much, since
the quantity integrated against it does not lie in Fun(A). In fact, one cannot
really compute these spin network invariants using a generalized measure in
our precise sense of the term [12]. One reason, though not the only one, is
that the invariants depend on a choice of extra structure known as a ‘framing’.
Nonetheless, Reshetikhin and Turaev were able to rigorously construct these
isotopy invariants of framed embedded spin networks by a combinatorial
procedure that generalizes Penrose’s recipe to quantum groups. In the case
of links labelled by the spin- 21 representation of SU(2), this invariant is just
the Kauffman bracket.
These ideas interact with the subject of quantum gravity in many ways.
In Section 1 we cited some expository accounts of the relation between Chern-
Simons theory and quantum gravity in 4-dimensional spacetime. However,
spin networks and Chern-Simons theory are also closely related to quantum
24
gravity with cosmological constant in 3-dimensional spacetime [28, 36, 40,
43, 44, 47, 49]. In this case, quantum gravity can be rigorously described
either by starting with Einstein’s equation on a smooth 3-manifold, phrased
in terms of an appropriate version of the new variables, or purely combinato-
rially in terms of a simplicial 3-manifold with edges labelled by spins. Better
yet, although there is still a bit of work left in proving this rigorously, it
appears that the two approaches are equivalent! In short, in this toy model
the question as to whether spacetime is continuous or discrete has no simple
yes-or-no answer; in some sense, the answer is ‘both’.
It is natural to wonder if the same might be true in 4 dimensions. There
are a few hints here. One the one hand, Einstein’s equation in the new
variables formulation has a simpler relative called BF theory [12], which is
much easier to canonically quantize. On the other hand, Crane and Yetter
have combinatorially constructed a relatively simple topological quantum
field theory in 4 dimensions using spin networks [22]. It is not known if these
are equivalent, though there are some clues suggesting it, such as the relation
of both to Chern-Simons theory. This would be worth understanding better,
even though we expect full-fledged quantum gravity in 4 dimensions to be
more complicated.
3 The New Variables

Now let us return to general relativity and describe more precisely how the
new variables make it look more like other gauge theories. In its original for-
mulation, general relativity is all about a metric on spacetime, while gauge
theories are all about a connection on some bundle over spacetime. Of course
there is a connection involved in general relativity too, the Levi-Civita con-
nection, but this is traditionally regarded as a subsidiary entity, since it can
be computed starting from the metric. To emphasize the gauge-theoretic
aspects of general relativity, one needs to rewrite it so a connection has the
starring role, and the metric appears as more of a minor character. There
have been many different attempts to do this, and for a more thorough ex-
ploration of them the reader will have to turn elsewhere [38]. Here we only
describe two: the Palatini formalism, and the Plebanski formalism. The lat-
ter is the one directly related to the ‘new variables’, but the former serves
as a good warmup exercise. This section makes somewhat greater demands
25
on the reader’s acquaintance with differential geometry, and also uses some
variational calculus. Luckily, there happens to be a textbook that explains
most of what we need [14]. We will skip over all sorts of important subtleties,
which are discussed in Ashtekar’s books [2, 3].
In the Palatini formalism there will be two basic fields, a connection and
a ‘soldering form’. The clever idea (in the modern version of this approach) is
to fix an oriented bundle T over M that is isomorphic to the tangent bundle
T M, but not canonically so. We can think of T as a kind of ‘imitation
tangent bundle’. Physicists usually call it, or any of its fibers, the ‘internal
space’. We assume T is equipped with a Lorentzian metric η — the ‘internal
metric’— and assume that the spacetime metric g is obtained from η via an
isomorphism e: T M → T . In other words,
g(v, w) = η(e(v), e(w)) (5)
for any vector fields v and w. We may regard e as a T -valued 1-form, and
then it is called the soldering form. We should add, in the interests of cultural
literacy, that e is also called the ‘coframe field’, while e−1 is known as the
‘frame field’, ‘tetrad field’, or ‘vierbein’.
In Palatini formalism the two basic fields are this soldering form e and
a connection A on T preserving the metric η, usually called a ‘Lorentz con-
nection’. One virtue of this formalism is that it makes sense even when
e: T M → T is not an isomorphism. Thus the Palatini formalism extends
general relativity to certain cases where the metric g is degenerate.
In computations it is handy to use differential forms on M taking values
in the exterior algebra bundle ΛT . These form an algebra, where the product
(written as ∧) is built from the exterior product in ΛT together with the usual
wedge product of differential forms. In complete analogy with the volume
form on an oriented Lorentzian manifold, the orientation and internal metric
on T give rise to an ‘internal volume form’, i.e. a section ν of Λ4 T . This
in turn yields a map from Λ4 T -valued forms to ordinary differential forms,
denoted by ‘tr’, and given by
tr(ν ⊗ ω) = ω
for any differential form ω on M. Now the curvature of the connection A can,
as usual, be regarded as an End(T )-valued 2-form, but the internal metric
provides an isomorphism T ∼ = T ∗ , so we may think of it as T ⊗ T -valued,
26
and then the fact that A is metric-preserving means the curvature actually
takes values in Λ2 T . We call this Λ2 T -valued 2-form F .
One can then obtain the vacuum Einstein equation from a variational
principle, starting with the ‘Palatini action’ given by
Z
SP al (A, e) = tr(e ∧ e ∧ F ).
M
Let us sketch how this goes. The idea is to compute the variation δSP al ,
demand that it vanish for all compactly supported variations δA and δe, and
see what this implies. We will need to use the wonderful formula
δF = dA δA, (6)
where dA denotes the exterior covariant derivative of ΛT -valued forms with

respect to the connection A, and δA is treated as a Λ2 T -valued 1-form. We
begin by computing the variation, or differential, of SP al as a function of A
and e:
Z
δSP al = δtr(e ∧ e ∧ F )
Z
= tr(2δe ∧ e ∧ F + e ∧ e ∧ δF ),
using the standard rules for differentiation. Using equation (6) we obtain
Z
δSP al = tr(2δe ∧ e ∧ F + e ∧ e ∧ dA δA).
Finally, integrating by parts as one always does in these variational compu-

tations, Z
δSP al = 2 tr(δe ∧ e ∧ F − e ∧ dA e ∧ δA).
This only vanishes for all compactly supported δA, δe if
e ∧ F = 0, e ∧ dA e = 0.
These are our ‘classical equations of motion’.

When e is an isomorphism, these equations are really just the vacuum
Einstein equation in disguise. First, one can show that in this case e∧dA e = 0
implies dA e = 0. Then, using the isomorphism e: T M → T to transfer A to
27
a metric-preserving connection on T M, say Γ, the equation dA e = 0 implies
that Γ is torsion-free, hence equal to the Levi-Civita connection of g. This
lets us translate the equation e ∧ F = 0 into an equation about the curvature
of Γ, i.e., the Riemann tensor. When we do so, we obtain the vacuum Einstein
equation!
The Plebanski formalism works in much the same way. Indeed, at first
it may seem like just an unnecessarily complicated version of the Palatini
formalism. As noted by Ashtekar, however, the constraints work out much
more nicely when one tries to canonically quantize the theory. The Plebanski
formalism is also called the ‘self-dual’ formalism, because it takes advantage
of self-duality, a very special feature of 4 dimensions. Recall that given a
4-dimensional real inner product space V , the Hodge star operator maps the
second exterior power of V to itself:
⋆: Λ2 V → Λ2 V.
Since ⋆2 = 1, we can split any element ω ∈ Λ2 V into a self-dual and an

anti-self-dual part:
ω = ω+ + ω− , ⋆ω± = ±ω± .
Another way to think of this is as follows. The inner product gives an iso-
morphism between End(V ) = V ⊗ V ∗ and V ⊗ V , which restricts to an
isomorphism between so(V ) and Λ2 V . The splitting of Λ2 V into self-dual
and anti-self-dual parts then turns out to correspond to a splitting of the 6-
dimensional Lie algebra so(V ) into two 3-dimensional ones. Taking V = R4
this gives
so(4) ∼
= so(3) ⊕ so(3). (7)
These simple facts have all sorts of ramifications for 4-dimensional physics
and topology, especially in gauge theory, where a connection on a Riemannian
manifold having self-dual curvature 2-form is automatically a solution of the
Yang-Mills equation, called an ‘instanton’ [29]. When M is compact, the
space of instantons modulo gauge transformations is finite-dimensional, and
one can extract powerful invariants of the smooth structure of M from this
‘moduli space’ using techniques developed by Donaldson and others [25].
For example, one can use these to show that R4 admits uncountably many
different smooth structures, unlike any other Rn ! Seiberg and Witten have
28
recently simplified some aspects of Donaldson theory by further recourse to
ideas from physics, but self-duality still plays a key role.
The sort of self-duality relevant to the Plebanski formalism is rather dif-
ferent, though with tantalizing relationships to the above [21]. First, given a
Lorentzian rather than Riemannian 4-manifold, one has ⋆2 = −1 on 2-forms.
This means that the eigenvalues of the Hodge star operator are ±i, so one
needs to work with complex-valued 2-forms to take advantage of self-duality.
When we tensor both sides of equation (7) with C, we obtain
so(4, C) ∼
= sl(2, C) ⊕ sl(2, C).
For these reasons, the Plebanski formalism applies most naturally to com-
plexified general relativity, and some extra work is needed to restrict to real-
valued metrics.
Second, in the Plebanski formalism self-duality first shows up not with
respect to ‘honest’ 2-forms, but with respect to sections of Λ2 CT , the second
exterior power of the complexified internal space. The internal metric η
extends by complexification to a bilinear pairing on CT , which we again call
η, and this together with the orientation give rise to an ‘internal’ Hodge star
operator ∗ acting on sections of Λ2 CT . We can thus split sections of this
bundle into self-dual and anti-self-dual parts:
ω = ω+ + ω− , ∗ω± = ±iω± .
Now, the orthonormal frame bundle of CT is a principal bundle P with

structure group SO(4, C). Assume we have a spin structure for CT , that
is, a double cover P̃ of P with structure group SO(4, g C) = SL(2, C) ⊕
SL(2, C). Then P̃ is the sum P+ ⊕P− of two principal bundles with structure
group SL(2, C). A metric-preserving connection on CT is usually called a
‘complex Lorentz connection’. By the above, a complex Lorentz connection
A is equivalent to a pair of connections A± on P± , which we call the self-
dual and anti-self-dual parts of A. The curvature of A± is a 2-form F±
taking values in the associated bundle AdP± , but sections of these bundles
can be identified with self-dual (resp. anti-self-dual) sections of Λ2 CT . The
curvature of A is a Λ2 CT -valued 2-form F with F = F+ + F− .
The two basic fields in the Plebanski formalism are a complex-valued
soldering form, that is, 1-form on M with values in CT , and a self-dual
29
Lorentz connection A+ , that is, a connection on P+ . The Plebanski action is
given by: Z
SP le (A+ , e) = tr(e ∧ e ∧ F+ )
M
This is very much like the Palatini action, and, much as before, the classical
equations of motion are
e ∧ F+ = 0, e ∧ dA+ e = 0.
These give Einstein’s equation for complex-valued metrics on M when e

defines an isomorphism between the complexified tangent bundle of M and
CT . To see this, first note that e gives rise to a complex-valued metric g on
M by equation (5). Moreover, we can use e to transfer A+ to a connection
Γ+ on the complexified tangent bundle of M. The equation e∧dA+ e = 0 then
implies Γ+ is torsion-free, hence equal to the self-dual part of the Levi-Civita
connection Γ of g. We can transfer Γ back to CT and get a connection A
having A+ as its self-dual part. Then it turns out that e ∧ F+ = e ∧ F , so as
in the Palatini formalism we obtain Einstein’s equation for g.
In terms of these ‘new variables’, as they are often called, the configu-
ration space of general relativity is no longer the space of all metrics on S.
Instead, it is the space A+ of all connections on P+ restricted to S, or more
precisely, to some fixed spacelike slice St ⊂ M. As in the metric formulation,
one can separate the equations of motion into evolutionary equations and
constraints. We shall not go through the calculations here, but simply state
some of the results [2, 3]. Recall that in the metric formulation of general rel-
ativity, the analogs of the position and momentum are the fields qij and pij on
S. In the new variables, the analogs of position and momentum are instead
the connection A+ and the field (e ∧ e)+ restricted to S. Since P+ is trivial
when restricted to S, A+ can be identified with an sl(2, C)-valued 1-form on
S, usually denoted simply by A. Similarly, (e ∧ e)+ can be identified with an
sl(2, C)-valued 2-form on S, but physicists often use the isomorphism
Λ2 T ∗ S ∼
= T S ⊗ Λ3 T ∗ S,
given by the interior product, to treat it as a ‘densitized vector field’ on S

with values in sl(2, C), usually denoted by Ẽ. Actually, using coordinates,
they often write A and Ẽ as Aai and Ẽ ia , where a = 1, 2, 3 indexes a basis of
sl(2, C) that is orthonormal with respect to the Killing form.
30
In these terms, the constraints in the new variables approach are
G a = ∂i Ẽ ia + [Ai , Ẽ i ]a , (8)
Hj = Fija Ẽai , (9)

and
H = ǫabc Fijc Ẽ ia Ẽ jb , (10)
where F is the curvature of A and ǫabc is the completely antisymmetric tensor
with ǫ123 = 1.
It is probably not obvious from our description where the advantage of
the self-dual formalism over the Palatini formalism lies. The key point is
the construction of the field Ẽ from (e ∧ e)+ . One can similarly construct
Ẽ from e ∧ e in the Palatini formalism, but then Ẽ winds up being subject
to extra constraints [3] which negate the advantages of this approach. In
the self-dual formalism, no conditions on Ẽ need hold for it to come from a
complex soldering form e.
4 Canonical Quantization
Now let us try to canonically quantize gravity in terms of the new variables,
with an eye to the importance of spin networks. The basic idea of DeWitt’s
approach goes over to this context with only a few small modifications. In
its naive form, the idea would be to define a Hilbert space L2 (A+ ) and define
operators on this space by
ˆ i ψ)(A) = h̄ δψ
(Âai ψ)(A) = Aai ψ(A), (Ẽ a (A).
δAai
(Again, we have suppressed the dependence of these operators on the point

in our spacelike slice.) We would then substitute these operators for A and Ẽ
in formulas (8-10), obtaining quantized constraints Ĥ, Ĥj , and Ĝ a . Physical
states should then be annihilated by these constraints.
There are, however, a number of subtleties that we need to address in
order to do things right. First there is the fact that the Plebanski formalism
most naturally describes complex general relativity, and needs some adjust-
ment to become a theory of honest real-valued Lorentzian metrics. This
31
issue has been a murky one for some time, and only now is it beginning to
become clear. At the classical level one can simply impose extra constraints
saying that the metric is Lorentzian. This would not be very nice in the
quantum theory, though, because the whole point of the new variables was
to simplify the constraints! Luckily it appears that at the quantum level we
can deal with the issue in quite a different way, namely by restricting our
attention to a space HL2 (A+ ) of holomorphic wavefunctions on A+ . Rather
than really explaining this here, let us merely note that it is closely related
to the Bargmann-Segal formulation of the quantum theory of a particle on
a line [3, 4]. The Bargmann-Segal formulation makes use of an isomorphism
between L2 (R) and the Hilbert space HL2 (C) consisting of holomorphic
functions that are square-integrable with respect to a Gaussian measure. A
similar theory for complex Lie groups and their compact real forms, due to
Hall [32], gives an isomorphism between a Hilbert space HL2 (SL(2, C)) and
the Hilbert space L2 (SU(2)) defined using Haar measure. A generalization
of this [8] gives an isomorphism, the ‘coherent state transform’, between a
Hilbert space HL2 (A+ ) and the space L2 (A), where A is the space of con-
nections on an SU(2) bundle over S. Thus the theory described in Section 2
is relevant to quantum gravity even though SL(2, C) is not compact.
Second, there is the issue of rigorously interpreting the constraints and
finding solutions to them. When we try to replace the fields A and Ẽ by Â and
ˆ in formulas (8-10), we run into difficulties. One problem is that operators
Ẽ
do not commute, so different orderings of the same polynomial in the classical
fields can have different meanings at the quantum level. The orderings given
here seem best when one wants operators on some space of functions on A+ ;
elsewhere one may see the opposite orderings [12], but that is the natural
consequence of working dually on some space of generalized measures. A
more profound problem is that Â and Ẽ ˆ are not really operator-valued func-
tions, but only operator-valued distributions — one must integrate them
against ‘test functions’ (really bundle sections) to obtain operators — and
like ordinary distributions, the product of operator-valued distributions is
only defined under special conditions, or using special tricks.
Luckily, the Gauss law constraint and diffeomorphism constraint have
simple geometrical interpretations which relieve us of the need for making
sense of the quantum Ranalogs of formulas (8) and (9). At the classical level,
functions of the form S Xa G a generate all the flows on the phase space T ∗ A
32
that correspond to the action of one-parameter groups of gauge transfor-
mations on T ∗ A. So at the quantum level we can interpret the Gauss law
constraint as saying that a state is invariant under ‘small’ gauge transfor-
mations, that is, those lying in the component of G containing the identity.
One can show that any state ψ ∈ L2 (A) that is invariant under small gauge
transformations is invariant under all gauge transformations. As we have
seen, there is an ample supply of such states, and the space of these, which
we write as L2 (A/G), is spanned by spin network states. R
Similarly, any function on phase space of the form S N i Hi generates
a flow corresponding to the action of a one-parameter group of diffeomor-
phisms. Thus at the quantum level we can interpret the diffeomorphism
constraint as saying that a state is invariant under the group Diff 0 (S) of
small analytic diffeomorphisms, that is, those lying in the connected compo-
nent containing the identity. Here we run into a problem. It is easy to see
that the only element of Fun0 (A) invariant under Diff 0 (S) is the function 1;
any other can be written as in equation (1) only for some nonempty graph γ,
and then there is a small diffeomorphism taking it to some other function, in
fact one orthogonal to it in L2 (A). A simple approximation argument then
shows that no ψ ∈ L2 (A) except ψ = 1 is invariant under Diff 0 (S)!
In fact this is not as bad as it may seem. Quite often in the quantization
of systems with constraints, the physical states are not really vectors in the
kinematical state space, but only vectors in some larger topological vector
space having the kinematical state space as a dense subspace. Consider a
simple example: the particle on the line, regarded as a particle on the plane
whose position (q1 , q2 ) is subject to the constraint q2 = 0. Naively, one would
start with the kinematical state space L2 (R2 ) and look for states ψ satisfying
the constraint q̂2 ψ = 0, i.e.,
q2 ψ(q1 , q2 ) = 0.
The only L2 function on the plane satisfying this equation is zero! However,
there are distributions on the plane satisfying this equation. The space of such
solutions is not itself a Hilbert space. However, it has a subspace isomorphic
to L2 (R), given by the distributions of the form ψ(q1 )δ(q2 ). This subspace
is correct Hilbert space for the particle on the line.
Similarly, while there are practically no solutions of the diffeomorphism
constraint living in L2 (A), there are plenty of ‘distributional’ ones. In fact,
33
there are plenty of generalized measures on A invariant under small diffeo-
morphisms of S. The most interesting of these are the gauge-invariant ones,
since we would like to solve the Gauss law constraint as well. Note that the
the inner product on L2 (A/G) sets up a chain of inclusions
Fun(A/G) ⊂ L2 (A/G) ⊂ Fun(A/G)∗ ,
where Fun(A/G) is the gauge-invariant subspace of Fun(A), and its Banach

space dual Fun(A/G)∗ may be identified with the space of gauge-invariant
generalized measures on A. The group Diff 0 (S) acts in a consistent way on
all these spaces. A natural candidate for a space of simultaneous solutions
of the Gauss law and diffeomorphism constraints is the space
Fun(A/G)∗inv = {µ ∈ Fun(A/G)∗ : ∀g ∈ Diff 0 (S) gµ = µ}.
This space is not itself a Hilbert space, but it may have some subspace deserv-
ing to be called the ‘Hilbert space of diffeomorphism-invariant states’. On the
other hand, perhaps there are physically relevant diffeomorphism-invariant
states that are not contained in Fun(A/G)∗ , but only in some still larger
space containing L2 (A/G). An example would be the ‘Chern-Simons state’.
To understand these would require further study of generalized functions on
A/G.
Anyway, at least a general recipe for finding elements of Fun(A/G)∗inv is
known, as are many interesting examples [7, 9]. Among the most interesting
are the ‘knot states’, which appeared already in a nonrigorous way in the
pioneering work of Rovelli and Smolin [45]. These are most easily described
using spin networks, although they were not originally constructed that way
[4]. Fix an isotopy class C of analytic knots and an irreducible unitary
representation ρ of SU(2). This determines a set S of spin network states,
R
namely the Wilson loops tr(ρ(T exp α A)) with α ∈ C. Since linear combina-
tions of spin network states are dense in Fun(A/G), any generalized measure
µ ∈ Fun(A/G)∗ is determined by its values on spin network states. We define
the knot state µ by setting µ(ψ) = 1 if ψ ∈ S and µ(ψ) = 0 if ψ is a spin
network state not in S. One must check, of course, that µ is a well-defined
generalized measure, which involves proving a certain bound. By construc-
tion µ is invariant under small diffeomorphisms, so we have µ ∈ Fun(A/G)∗inv .
Since a Wilson loop is just a special sort of spin network, it is natural
to ask if this procedure generalizes to yield ‘diffeomorphism-invariant spin
34
network states’. The idea would be to use an arbitrary isotopy class of
spin networks to obtain a set S of spin network states, and to define µ ∈
Fun(A/G)∗inv by setting µ(ψ) = 1 for ψ ∈ S and µ(ψ) = 0 if ψ is a spin
network state not in S. The problem is simply to prove that µ is a well-
defined generalized measure. This seems to be true when we start with an
ambient isotopy class of links in S labelled with representations, but the
general spin network case is more subtle and still under investigation.
Now let us say a bit about the Hamiltonian constraint. This is perhaps
the most controversial aspect of the whole subject, and certainly one of the
most important ones: if we find a way to rigorously treat the Hamiltonian
constraint, we will be quite close to a rigorous quantization of Einstein’s
equation, but if not, the whole approach described here may be fundamen-
tally flawed, or at least in need of very new ideas. A lot of work has been
done on the Hamiltonian constraint, which we cannot really do justice to
here [15, 16, 17, 20, 30, 34]. Naively, the problem is to make sense of
∂ ∂
Ĥ(x) = h̄2 ǫabc Fijc (x)
∂Aia (x) ∂Ajb (x)
as a distribution on S taking values in operators on some space of holomor-
phic functions on A+ . Eventually, however, we want to find states that are
annihilated by this constraint together with the Gauss law and diffeomor-
phism constraints. Thus we would be perfectly happy if we could first use
the coherent state transform to transfer the constraint to L2 (A), and then
find some subspace of Fun(A/G)∗inv — or some related space of solutions of
the Gauss law and diffeomorphism constraints — that we could argue was
annihilated by the Hamiltonian constraint. The problem is that, in constrast
to the Gauss law and diffeomorphism constraints, there is no easy geometri-
cal interpretation of the Hamiltonian constraint in terms of connections on
S to fall back upon.
At first it might seem foolish to even hope for a simple geometrical 3-
dimensional interpretation of the Hamiltonian constraint. After all, the
Hamiltonian constraint expresses the 4-dimensional diffeomorphism invari-
ance of general relativity; or in other words, it encodes the dynamics of the
theory. It is all the more tantalizing, therefore, that there are some hints of
such an interpretation. Rovelli and Smolin’s discovery of these was one of
the early successes of the loop representation [45], but making them precise
has proved to be very difficult.
35
Stripped of all nuance, the observation of Rovelli and Smolin reduces to
the following. Given a knot α in S, the Wilson loop
Z
ψ(A) = tr(ρ(T exp A))
α
is a holomorphic function on A+ . Heuristically speaking, when one applies

the functional derivative ∂/∂Aai (x) to ψ one brings down a factor of the
tangent vector α̇i (t) if α(t) = x. So the double functional derivative in Ĥ(x)
brings down a factor of α̇i(t)α̇j (t), which is symmetric in i and j. Since Fijc
is antisymmetric in i and j, one obtains
Ĥ(x)ψ = 0.
In fact, when one goes through the argument more carefully one discov-
ers that the double functional derivative of a Wilson loop is very singular,
so the result Ĥ(x)ψ = 0 is a purely formal one unless one can devise some
regularization procedure to make the argument rigorous. This has not yet
been achieved. Still, upon examination, the argument seems to suggest a
3-dimensional geometrical interpretation of the Hamiltonian constraint as
some kind of ‘shift operator’ generating the motion of a Wilson loop, or
more generally the edges of a spin network, along its tangent vectors [45].
For this reason, much effort has been expended to understand the argument
and find a context in which it can be made rigorous. One would be very
happy, for example, to find by some heuristic argument a general formula
for the action of the Hamiltonian constraint on spin network states, or per-
haps ‘diffeomorphism-invariant spin network states’, which one could then
subsequently justify by means of its good properties.
In short, quantum gravity continues to provide mathematical physics with
challenging — indeed quite frustrating — problems. However, let us con-
clude on a more upbeat note! Even at the level of the kinematical state
space L2 (A/G), there are some very intriguing applications of spin networks
to physics. Classically, in the metric representation a state of gravity is de-
scribed by the metric on S and its first time derivative. We can rewrite
interesting functions of the metric in terms of the new variables, and then
attempt to quantize them by replacing A and Ẽ by their quantum versions
in these expressions, obtaining operators on L2 (A). These operators should
commute with the action of G, hence give rise to operators on L2 (A/G). As
36
the example of the Hamiltonian constraint shows, carrying this out is by no
means straightforward in all cases. However, Rovelli and Smolin [46] have
considered the area of a surface and the volume of a region in S, which are
technically simpler, and obtained explicit formulas for their quantum ver-
sions as operators on L2 (A/G), in terms of the spin network basis. These
operators turn out to have discrete spectrum: certain multiples of Planck
length squared for the area operator, and certain multiples of the Planck
length cubed for the volume operator. In quantum theory, the spectrum of a
self-adjoint operator corresponds to the values the corresponding observable
can assume, so this is an indication that area and volume are ‘quantized’ in
the very literal sense of assuming a discrete set of values! Moreover, there is
a simple geometrical reason for this fact. To speak in rather oversimplified
terms, the area operator applied to a given spin network state counts the
number of points at which an edge of q the spin network intersects the surface
in question, weighted by a factor of j(j + 1)ℓ2p , where j is the spin labelling
that edge. Similarly, the volume operator applied to a spin network state
counts the number of vertices of the spin network contained in the region in
question, weighted by some function of their labellings by intertwiners and
the geometry of the incident edges.
Until we have a fully working theory of quantum gravity, and under-
stand how to take the other forces in account, it is dangerous to take too
seriously any physical predictions a partial theory might make. Moreover,
this prediction of discreteness of area and volume at the Planck scale is ab-
surdly hard to test with present technology! Nonetheless, the idea that the
marriage of Einstein’s equation and quantum theory could make such a fas-
cinating prediction should serve as a kind of inspiration for mathematicians
and physicists working on quantum gravity.
References
[1] A. Ashtekar, New Hamiltonian formulation of general relativity, Phys.
Rev. D36 1587-1602.
[2] A. Ashtekar and invited contributors, New Perspectives in Canonical

Gravity, Bibliopolis, Napoli, Italy, 1988. (Available through the Amer-
37
ican Institute of Physics; errata available from the Center for Gravita-
tional Physics and Geometry at Pennsylvania State University.)
[3] A. Ashtekar, Lectures on Non-perturbative Canonical Quantum Gravity,
Singapore, World Scientific, 1991.
[4] A. Ashtekar, Mathematical problems of non-perturbative quantum gen-
eral relativity, in Proceedings of the 1992 Les Houches Summer School
on Gravitation and Quantization, ed. B. Julia, North-Holland, Amster-
dam, 1993.
[5] A. Ashtekar and C. Isham, Representations of the holonomy algebra of
gravity and non-abelian gauge theories, Class. Quan. Grav. 9 (1992),
1069-1100.
[6] A. Ashtekar and J. Lewandowski, Representation theory of analytic
holonomy C*-algebras, in Knots and Quantum Gravity, ed. J. Baez,
Oxford, Oxford U. Press, 1994, pp. 21-61.
[7] A. Ashtekar and J. Lewandowski, Projective techniques and functional
integration for gauge theories, to appear in Jour. Math. Phys..
[8] A. Ashtekar, J. Lewandowski, D. Marolf, J. Mourão and T. Thiemann,
Coherent state transforms for spaces of connections, to appear in Jour.
Funct. Analysis.
[9] J. Baez, Diffeomorphism-invariant generalized measures on the space of
connections modulo gauge transformations, in Proceedings of the Confer-
ence on Quantum Topology, ed. D. Yetter, World Scientific, Singapore,
1994.
[10] J. Baez, Generalized measures in gauge theory, Lett. Math. Phys. 31
(1994), 213-223.
[11] J. Baez, editor, Knots and Quantum Gravity, Oxford U. Press, Oxford,
1994.
[12] J. Baez, Knots and quantum gravity: progress and prospects, to appear
in the proceedings of the Seventh Marcel Grossman Meeting on General
Relativity, University of California at Riverside preprint available as
gr-qc/9410018.
38
[13] J. Baez, Spin networks in gauge theory, to appear in Adv. Math..
[14] J. Baez and J. P. Muniain, Gauge Fields, Knots and Gravity, World
Scientific, Singapore, 1994. (Errata available from the authors.)
[15] C. Di Bartolo, R. Gambini, J. Griego and J. Pullin, The space of states

in quantum gravity in terms of loops and extended loops: some remarks,
Pennsylvania State University preprint CGPG-95/3-5, available as gr-
qc/9503059.
[16] M. Blencowe, The Hamiltonian constraint in quantum gravity, Nucl.

Phys. B341 (1990), 213-251.
[17] R. Borissov, Regularization of the Hamiltonian constraint and closure of

the constraint algebra, Temple University preprint TU-94-11, available
as gr-qc/9411038.
[18] B. Brügmann, Loop representations, Max-Planck-Institute preprint

MPI-Ph/93-94, available as gr-qc/9312001.
[19] B. Brügmann, personal communication.
[20] B. Brügmann and J. Pullin, On the constraints of quantum gravity in

the loop representation, Nucl. Phys. B390 (1993), 399-438.
[21] L. Chang and C. Soo, BRST cohomology and invariants of four-

dimensional gravity in Ashtekar’s variables, Phys. Rev. D46 (1992),
4257-4262.
[22] L. Crane, L. Kauffman and D. Yetter, State-sum invariants of 4-

manifolds, I, Kansas State University preprint, available as hep-
th/9409167.
[23] B. DeWitt, Quantum theory of gravity, I-III, Phys. Rev. 160 (1967)
1113-1148, 162 (1967) 1195-1239, 1239-1256.
[24] P. A. M. Dirac, Lectures on Quantum Mechanics, Yeshiva University,

New York, 1964.
[25] S. Donaldson and P. Kronheimer, The Geometry of Four-Manifolds,

Oxford U. Press, Oxford, 1990.
39
[26] B. Durhuus, H. Jakobsen and R. Nest, Topological quantum field theo-
ries from generalized 6j-symbols, Rev. Math. Phys. 5 (1993), 1-67.
[27] A. Fischer and J. Marsden, The initial value problem and the dynami-
cal formulation of general relativity, in General Relativity, an Einstein
Centenary Survey, eds. S. Hawking and W. Israel, Cambridge U. Press,
Cambridge, 1979, pp. 138-211.
[28] T. Foxon, Spin networks, Turaev-Viro theory and the loop representa-
tion, Class. Quan. Grav. 12 (1995), 951-964.
[29] D. Freed and K. Uhlenbeck, Instantons and Four-Manifolds, Springer-

Verlag, Berlin, 1984.
[30] R. Gambini, Loop space representation of quantum general relativity

and the group of loops, Phys. Lett. B255 (1991), 180-188.
[31] R. Gambini and A. Trias, Gauge dynamics in the C-representation, Nucl.

Phys. B278 (1986), 436-448.
[32] B. Hall, The Segal-Bargmann coherent state transform for compact Lie
groups, Jour. Funct. Analysis 122 (1994), 103-151.
[33] C. Isham, Canonical quantum gravity and the problem of time, in Inte-
gral Systems, Quantum Groups, and Quantum Field Theories, eds. L.
A. Ibort and M. A. Rodriguez, Kluwer, Dordrecht, 1993, pp. 157-207.
[34] T. Jacobson and L. Smolin, Nonperturbative quantum geometries, Nucl.

Phys. B299 (1988) 295-345.
[35] L. Kauffman, Knots and Physics, World Scientific, Singapore, 1991.
[36] L. Kauffman and S. Lins, Temperley-Lieb Recoupling Theory and In-

variants of 3-Manifolds, Princeton U. Press, Princeton, 1994.
[37] J. Lewandowski, Topological measure and graph-differential geometry

on the quotient space of connections, International Journal of Theoret-
ical Physics, 3 (1994) 207-211.
[38] P. Peldan, Actions for gravity, with generalizations: a review, Class.

Quan. Grav. 11 (1994), 1087-1132.
40
[39] R. Penrose, Angular momentum; an approach to combinatorial space
time, in Quantum Theory and Beyond, ed. T. Bastin, Cambridge Uni-
versity Press, Cambridge, 1971.
[40] G. Ponzano and T. Regge, Semiclassical limits of Racah coefficients, in

Spectroscopic and Group Theoretical Methods in Physics, ed. F. Bloch,
Wiley, New York, 1968.
[41] J. Pullin, Knot theory and quantum gravity: a primer, in Proceedings

of the Vth Mexican School of Particles and Fields, ed. J. Lucio, World
Scientific, Singapore, 1993.
[42] N. Reshetikhin and V. Turaev, Ribbon graphs and their invariants de-
rived from quantum groups, Comm. Math. Phys. 127 (1990), 1-26.
Invariants of 3-manifolds via link polynomials and quantum groups, In-
vent. Math. 103 (1991), 547-597.
[43] J. Roberts, Skein theory and Turaev-Viro invariants, to appear in Topol-

ogy.
[44] C. Rovelli, The basis of the Ponzano-Regge-Turaev-Viro-Ooguri model

is the loop representation basis, Phys. Rev. D48 (1993), 2702-2707.
[45] C. Rovelli and L. Smolin, Loop representation for quantum general rel-
ativity, Nucl. Phys. B331 (1990), 80-152.
[46] C. Rovelli and L. Smolin, Discreteness of area and volume in quantum

gravity, University of Pittsburgh preprint available as gr-qc/9411005.
Spin networks in quantum gravity, University of Pittsburgh preprint.
[47] V. Turaev and O. Viro, State sum invariants of 3-manifolds and quantum
6j symbols, Topology 31 (1992), 865-902.
[48] E. Witten, Quantum field theory and the Jones polynomial, Comm.
Math. Phys. 121 (1989), 351-399.
[49] E. Witten, 2+1 dimensional gravity as an exactly soluble system, Nucl.

Phys. B311 (1988) 46-78.
41

Baex J at Spin Networks in Nonperturbative Quantum Gravity

Uploaded by

Copyright:

Available Formats

Baex J at Spin Networks in Nonperturbative Quantum Gravity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Baex J at Spin Networks in Nonperturbative Quantum Gravity

Uploaded by

Copyright:

Available Formats

Spin Networks in Nonperturbative Quantum Gravity

To appear in the proceedings of the AMS Short Course

in a unique way (up to dimensionless constant factors). This works out to

where ∇ is the operator of covariant differentiation corresponding to the

But in a vacuum Tµν = 0, so we obtain

which turns out to be equivalent to the vanishing of the Ricci tensor.

The energy, or Hamiltonian, of the particle is a function on phase space,

This vector field generates a one-parameter group of diffeomorphisms of phase

Now, classically the Hamiltonian is

where ρ is the fundamental representation and DA is the purely formal

not an infinite product of copies of R, but an infinite product of copies of

on A, instead of the nonexistent ‘Lebesgue measure’. This is precisely what

Note that for any connection A ∈ A, the parallel transport map

lies in Aγ . If we fix a trivialization of P at the endpoints of γ, we can identify

where γ1 , . . . , γn are piecewise analytic paths in S, and f is a continuous

Here we are using trivializations of P at the endpoints of the paths γi to

on which there is a natural measure, namely a product of copies of Haar

In the special case where G = SU(2) and S is a compact oriented 3-manifold,

is gauge-invariant, and it lies in L2 (A). More generally, any product of

If we trivialize P at the vertices of γ, and pick a basis for ρe , we can think of

so if ψ lies in L2 (Aγ ), so does gψ. Thus we get a representation of G on

(g1 , g2 )(h) = g1 hg2−1,

giving a unitary representation of G × G on L2 (G). The Peter-Weyl theorem

and in terms of this description, any g ∈ G acts on L2 (Aγ ) as the operator

In these terms any g ∈ G acts on L2 (Aγ ) as the operator

where for any representation λ of G, Inv(λ) is the subspace of G-invariant

In other words, a complete set of gauge-invariant vectors in L2 (Aγ ) is

holds, the representation j3 will not appear as a summand of of the tensor

Then if ψ is a spin network state associated to some graph γ, the integral

Indeed, Witten was able by formal manipulations to compute these invariants

3 The New Variables

g(v, w) = η(e(v), e(w)) (5)

where dA denotes the exterior covariant derivative of ΛT -valued forms with

Finally, integrating by parts as one always does in these variational compu-

These are our ‘classical equations of motion’.

Since ⋆2 = 1, we can split any element ω ∈ Λ2 V into a self-dual and an

Now, the orthonormal frame bundle of CT is a principal bundle P with

These give Einstein’s equation for complex-valued metrics on M when e

given by the interior product, to treat it as a ‘densitized vector field’ on S

Hj = Fija Ẽai , (9)

(Again, we have suppressed the dependence of these operators on the point

Fun(A/G) ⊂ L2 (A/G) ⊂ Fun(A/G)∗ ,

where Fun(A/G) is the gauge-invariant subspace of Fun(A), and its Banach

Fun(A/G)∗inv = {µ ∈ Fun(A/G)∗ : ∀g ∈ Diff 0 (S) gµ = µ}.

is a holomorphic function on A+ . Heuristically speaking, when one applies

[2] A. Ashtekar and invited contributors, New Perspectives in Canonical

[15] C. Di Bartolo, R. Gambini, J. Griego and J. Pullin, The space of states

[16] M. Blencowe, The Hamiltonian constraint in quantum gravity, Nucl.

[17] R. Borissov, Regularization of the Hamiltonian constraint and closure of

[18] B. Brügmann, Loop representations, Max-Planck-Institute preprint

[19] B. Brügmann, personal communication.

[20] B. Brügmann and J. Pullin, On the constraints of quantum gravity in

[21] L. Chang and C. Soo, BRST cohomology and invariants of four-

[22] L. Crane, L. Kauffman and D. Yetter, State-sum invariants of 4-

[24] P. A. M. Dirac, Lectures on Quantum Mechanics, Yeshiva University,

[25] S. Donaldson and P. Kronheimer, The Geometry of Four-Manifolds,

[29] D. Freed and K. Uhlenbeck, Instantons and Four-Manifolds, Springer-

[30] R. Gambini, Loop space representation of quantum general relativity

[31] R. Gambini and A. Trias, Gauge dynamics in the C-representation, Nucl.