hep-th/yymmnnn

Three-Dimensional Gravity Reconsidered

arXiv:0706.3359v1 [hep-th] 22 Jun 2007

Edward Witten
School of Natural Sciences, Institute for Advanced Study
Princeton, New Jersey 08540

We consider the problem of identifying the CFT’s that may be dual to pure gravity in
three dimensions with negative cosmological constant. The c-theorem indicates that three-
dimensional pure gravity is consistent only at certain values of the coupling constant, and
the relation to Chern-Simons gauge theory hints that these may be the values at which
the dual CFT can be holomorphically factorized. If so, and one takes at face value the
minimum mass of a BTZ black hole, then the energy spectrum of three-dimensional gravity
with negative cosmological constant can be determined exactly. At the most negative
possible value of the cosmological constant, the dual CFT is very likely the monster theory
of Frenkel, Lepowsky, and Meurman. The monster theory may be the first in a discrete
series of CFT’s that are dual to three-dimensional gravity. The partition function of the
second theory in the sequence can be determined on a hyperelliptic Riemann surface of
any genus. We also make a similar analysis of supergravity.

June, 2007
 Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1. Relation To Gauge Theory . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2. What To Aim For . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3. A Non-Classical Restriction . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4. Plan Of This Paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2. Gauge Theory And The Value Of c . . . . . . . . . . . . . . . . . . . . . . . 9
2.1. Quantization Of Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2. Comparison To Three-Dimensional Gravity . . . . . . . . . . . . . . . . . . 15
2.3. Holomorphic Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4. Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5. Analog For Supergravity . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3. Partition Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1. The Bosonic Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2. The Supersymmetric Case . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3. Extremal SCFT’s With Small k ∗ . . . . . . . . . . . . . . . . . . . . . . . 47
4. k = 2 Partition Function On A Hyperelliptic Riemann Surface . . . . . . . . . . . . 58
4.1. Twist Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2. The E · E Operator Product . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3. Determining The Four-Point Function . . . . . . . . . . . . . . . . . . . . 66
4.4. The Genus One Partition Function Revisited . . . . . . . . . . . . . . . . . 70
4.5. Role Of The Conformal Anomaly . . . . . . . . . . . . . . . . . . . . . . . 72
Appendix A. More Extremal Partition Functions . . . . . . . . . . . . . . . . . . . 76

1. Introduction

Three-dimensional pure quantum gravity, with the Einstein-Hilbert action

Z  
1 3 √ 2
I= d x g R+ 2 , (1.1)
16πG ℓ
has been studied from many points of view (see [1-3] for some early developments and [4]
for a recent review and references), but its status is fundamentally unclear. The present
paper is devoted to a tentative attempt to reconsider it. Before giving an overview of the
paper, we begin with an introduction to the problem.
The first thought about this theory is that at the classical level it is “trivial,” in the
sense that there are no gravitational waves, and any two solutions are equivalent locally.
So perhaps it might be tractable quantum mechanically.
A second thought is that despite being “trivial,” the theory actually is unrenormaliz-
able by power counting, since the gravitational constant G has dimensions of length. So
perhaps the quantum theory does not exist.

1
 The claim about unrenormalizability, however, is fallacious, precisely because the
classical theory is trivial. In three dimensions, the Riemann tensor Rijkl can be expressed
in terms of the Ricci tensor Rij . In the case of pure gravity, the equations of motion express
the Ricci tensor as a constant times the metric. So finally, any possible counterterm can
R √
be reduced to a multiple of d3 x g and is equivalent on-shell to a renormalization of the
cosmological constant, which is parametrized in (1.1) via the parameter ℓ2 . A counterterm
that vanishes on shell can be removed by a local redefinition of the metric tensor g (of the
general form gij → gij + aRij + . . ., where a is a constant and the ellipses refer to local
terms of higher order). So a more precise statement is that any divergences in perturbation
theory can be removed by a field redefinition and a renormalization of ℓ2 .

1.1. Relation To Gauge Theory

The claim just made is valid regardless of how one formulates perturbation theory.
But actually, there is a natural formulation in which no field redefinition or renormalization
is needed. This comes from the fact that classically, 2 + 1-dimensional pure gravity can
be expressed in terms of gauge theory. The spin connection ω is an SO(2, 1) gauge field
(or an SO(3) gauge field in the case of Euclidean signature). It can be combined with the
“vierbein” e to make a gauge field of the group SO(2, 2) if the cosmological constant is
negative (and a similar group if the cosmological constant is zero or positive). We simply
combine ω and e to a 4 × 4 matrix A of one-forms:
 
ω e/ℓ
A= . (1.2)
−e/ℓ 0

Here ω fills out a 3 × 3 block, while e occupies the last row and column. As long as
e is invertible, the usual transformations of e and ω under infinitesimal local Lorentz
transformations and diffeomorphisms combine together into gauge transformations of A.
This statement actually has a close analog in any spacetime dimension d, with SO(d −1, 2)
replacing SO(2, 2). What is special in d = 3 is that [5,6] it is also possible to write the action
in a gauge-invariant form. Indeed the usual Einstein-Hilbert action (1.1) is equivalent to
a Chern-Simons Lagrangian1 for the gauge field A:
Z  
k ∗ 2
I= tr A ∧ dA + A ∧ A ∧ A . (1.3)
4π 3
1
Here tr∗ denotes an invariant quadratic form on the Lie algebra of SO(2, 2), defined by
tr∗ ab = tr a ⋆ b, where tr is the trace in the four-dimensional representation and ⋆ is the Hodge
star, (⋆b)ij = 21 ǫijkl bkl .

2
In dimensions other than three, it is not possible to similarly replace the Einstein-Hilbert
action with a gauge-invariant action for gauge fields.2
In the gauge theory description, perturbation theory is renormalizable by power count-
ing, and is actually finite, because there are no possible local counterterms. The Chern-
Simons functional itself is the only gauge-invariant action that can be written in terms of
A alone without a metric tensor; as it is not the integral of a gauge-invariant local density,
it will not appear as a counterterm in perturbation theory. The cosmological constant
cannot be renormalized, since in the gauge theory description, it is a structure constant of
the gauge group.
As we have already remarked, the Chern-Simons description of three-dimensional
gravity is valid when the vierbein is invertible. This is so for a classical solution, so it is
true if one is sufficiently close to a classical solution. Perturbation theory, starting with
a classical solution, will not take us out of the region in which the vierbein is invertible,
so the Chern-Simons description of three-dimensional gravity is valid perturbatively. The
fact that, in this formulation, the perturbation expansion of three-dimensional gravity is
actually finite can reasonably be taken as a hint that the corresponding quantum theory
really does exist.
However, nonperturbatively, the relation between three-dimensional gravity and
Chern-Simons gauge theory is unclear. For one thing, in Chern-Simons theory, nonpertur-
batively the vierbein may cease to be invertible. For example, there is a classical solution
with A = ω = e = 0. The viewpoint in [6] was that such non-geometrical configurations
must be included to make sense of three-dimensional quantum gravity nonperturbatively.
But it has has been pointed out (notably by N. Seiberg) that when we do know how to
make sense of quantum gravity, we take the invertibility of the vierbein seriously. For
example, in perturbative string theory, understood as a model of quantum gravity in two
spacetime dimensions, the integration over moduli space of Riemann surfaces that leads
to a sensible theory is derived assuming that the metric should be non-degenerate.
There are other possible problems in the nonperturbative relation between three-
dimensional gravity and Chern-Simons theory. The equivalence between diffeomorphisms
and gauge transformations is limited to diffeomorphisms that are continuously connected
to the identity. However, in gravity, we believe that more general diffeomorphisms (such as

2
In d = 4, it is possible to write the Hamiltonian constraints of General Relativity in terms
of gauge fields [7]. This has been taken as the starting point of loop quantum gravity.

3
modular transformations in perturbative string theory) play an important role. These are
not naturally incorporated in the Chern-Simons description. One can by hand supplement
the gauge theory description by imposing invariance under disconnected diffeomorphisms,
but it is not clear how natural this is.
Similarly, in quantum gravity, one expects that it is necessary to sum over the different
topologies of spacetime. Nothing in the Chern-Simons description requires us to make such
a sum. We can supplement the Chern-Simons action with an instruction to sum over three-
manifolds, but it is not clear why we should do this.
From the point of view of the Chern-Simons description, it seems natural to fix a
particular Riemann surface Σ, say of genus g, and construct a quantum Hilbert space
by quantizing the Chern-Simons gauge fields on Σ. (Indeed, there has been remarkable
progress in learning how to do this and to relate the results to Liouville theory [8-11].)
In quantum gravity, we expect topology-changing processes, such that it might not be
possible to associate a Hilbert space with a particular spatial manifold.
Regardless of one’s opinion of questions such as these, there is a more serious problem
with the idea that gravity and gauge theory are equivalent non-perturbatively in three
dimensions. Some years after the gauge/gravity relation was suggested, it was discovered
by Bañados, Teitelboim, and Zanelli [12] that in three-dimensional gravity with negative
cosmological constant, there are black hole solutions. The existence of these objects,
generally called BTZ black holes, is surprising given that the classical theory is “trivial.”
Subsequent work [13,14] has made it clear that three-dimensional black holes should be
taken seriously, particularly in the context of the AdS/CFT correspondence [15].
The BTZ black hole has a horizon of positive length and a corresponding Bekenstein-
Hawking entropy. If, therefore, three-dimensional gravity does correspond to a quantum
theory, this theory should have a huge degeneracy of black hole states. It seems unlikely
that this degeneracy can be understood in Chern-Simons gauge theory, because this es-
sentially topological theory has too few degrees of freedom. However, some interesting
attempts have been made; for a review, see [4].
The existence of the BTZ black hole makes three-dimensional gravity a much more
exciting problem. This might be our best chance for a solvable model with quantum black
holes. Surely in 3+1 dimensions, the existence of gravitational waves with their nonlinear
interactions means that one cannot hope for an exact solution of any system that includes

4
quantum gravity.3 There might be an exact solution of a 1+1-dimensional model with
black holes (interesting attempts have been made [16]), but such a model is likely to be
much less realistic than three-dimensional pure gravity. For example, in 1+1 dimensions,
the horizon of a black hole just consists of two points, so there is no good analog of the
area of the black hole horizon.

1.2. What To Aim For

So we would like to solve three-dimensional pure quantum gravity. But what would
it mean to solve it?
First of all, we will only consider the case that the cosmological constant Λ is negative.
This is the only case in which we know what it would mean to solve the theory.
Currently, there is some suspicion (for example, see [17]) that quantum gravity with
Λ > 0 does not exist nonperturbatively, in any dimension. One reason is that it does not
appear possible with Λ > 0 to define precise observables, at least none [18] that can be
measured by an observer in the spacetime.4 This is natural if a world with positive Λ (like
the one we may be living in) is always at best metastable – as is indeed the case for known
embeddings of de Sitter space in string theory [20]. If that is so, then pure gravity with
Λ > 0 does not really make sense as an exact theory in its own right but (like an unstable
particle) must be studied as part of a larger system. There may be many choices of the
larger system (for example, many embeddings in string theory) and it may be unrealistic
to expect any of them to be soluble.
Whether that is the right interpretation or not, we cannot in this paper attempt
to solve three-dimensional gravity with Λ > 0, since, not knowing how to define any
mathematically precise observables, we do not know what to try to calculate.
For Λ = 0, above three dimensions there is a precise observable in quantum gravity,
the S-matrix. However, in the three-dimensional case, there is no S-matrix in the usual
sense, since in any state with nonzero energy, the spacetime is only locally asymptotic to
Minkowski space at infinity [3]. More relevantly for our purposes, in three-dimensional
pure gravity with Λ = 0, there is no S-matrix since there are no particles that can be
3
An exact solution or at least an illuminating description of the appropriate Hamiltonian for
near-extremal black holes interacting with external massless particles is conceivable.
4
At least perturbatively, the de Sitter/CFT correspondence gives observables that can be mea-
sured by an observer who looks at the whole universe from the outside [19,18]. These observables
characterize the wavefunction of the ground state.

5
scattered. There are no gravitons in three dimensions, and there are also no black holes
unless Λ < 0. So again, we do not have a clear picture of what we would aim for to solve
three-dimensional gravity with zero cosmological constant.
With negative cosmological constant, there is an analog, and in fact a much richer
analog, of the S-matrix, namely the dual conformal field theory (CFT). This is of course
a two-dimensional CFT, defined on the asymptotic boundary of spacetime. Not only does
AdS/CFT duality make sense in three dimensions, but in fact one of the precursors of the
AdS/CFT correspondence was the discovery by Brown and Henneaux [21] of an asymptotic
Virasoro algebra in three-dimensional gravity. They considered three-dimensional gravity
with negative cosmological constant possibly coupled to additional fields. The action is
Z  
1 3 √ 2
I= d x g R+ 2 +... , (1.4)
16πG ℓ

where the ellipses reflect the contributions of other fields. Their main result is that the
physical Hilbert space obtained in quantizing this theory (in an asymptotically Anti de
Sitter or AdS spacetime) has an action of left- and right-moving Virasoro algebras with
cL = cR = 3ℓ/2G. In our modern understanding [15], this is part of a much richer structure
– the boundary conformal field theory.
What it means to solve pure quantum gravity with Λ < 0 is to find this dual conformal
field theory. We focus on the case Λ < 0 because this is the only case in which we would
know what it means to solve the theory. Luckily, and perhaps not coincidentally, this is
also the case that has black holes.

1.3. A Non-Classical Restriction

This formulation of what we aim to do makes it clear that we must anticipate a

restriction that is rather surprising from a classical point of view. In contemplating the
classical action (1.4), it appears that the dimensionless ratio ℓ/G is a free parameter.
But the formula for the central charge cL = cR = 3ℓ/2G shows that this cannot be the
case. According to the Zamolodchikov c-theorem [22], in any continuously varying family
of conformal field theories in 1 + 1 dimensions, the central charge c is constant. More
generally, the same is true for the left- and right-moving central charges cL and cR .
So the central charges of the dual CFT cannot depend on a continuously variable
parameter ℓ/G. It must be [18] that the theory only makes sense for specific values of ℓ/G.

6
 Of course, the c-theorem has an important technical assumption: the theory must
have a normalizable and SL(2, R) × SL(2, R)-invariant ground state. (The two factors of
SL(2, R) are for left- and right-moving boundary excitations.) This condition is obeyed by
three-dimensional gravity, with Anti de Sitter space being the classical approximation to
the vacuum. The desired SL(2, R) × SL(2, R) symmetry is simply the classical SO(2, 2)
symmetry of three-dimensional AdS space.
The statement that ℓ/G cannot be continuously varied is not limited to pure gravity.
It holds for the same reason in any theory of three-dimensional gravity plus matter that
has a sensible AdS vacuum. For example, in the string theory models whose CFT duals are
known, ℓ/G is expressed in terms of integer-valued fluxes; this gives a direct explanation
of why it cannot be varied continuously.

1.4. Plan Of This Paper

Now we can describe the plan of this paper.

We aim to solve three-dimensional gravity with negative Λ, at some distinguished
values of ℓ/G at which it makes sense.
We do not have any rigorous way to determine the right values. However, in section
2, taking at face value the Chern-Simons description of three-dimensional gravity, we will
use it to motivate certain values of ℓ/G. The values that emerge – with the help of a
small sleight of hand in the choice of the gauge group for the Chern-Simons theory –
are interesting. They are the values at which cL and cR are integer multiples of 24, and
complete holomorphic factorization of the dual CFT is conceivable.
Not wishing to look a gift horse in the mouth, we will assume that these are the right
values to consider. Relying on holomorphic factorization, to describe the solution of the
theory, we must describe a sequence of holomorphic CFT’s with c = 24k, k = 1, 2, 3, . . ..
For k = 1, it is believed [23] that there are precisely 71 holomorphic CFT’s with the
relevant central charge c = 24. Of these theories, 70 have some form of Kac-Moody or
current algebra symmetry extending the conformal symmetry. In the AdS/CFT corre-
spondence, these theories are dual to three-dimensional theories describing gravity plus
additional gauge fields (with Chern-Simons interactions). To describe pure gravity, we
need a holomorphic CFT with c = 24 and no Kac-Moody symmetry.
Such a model was constructed nearly twenty-five years ago by Frenkel, Lepowsky, and
Meurman [24], who also conjectured its uniqueness. The motivation for constructing the
model was that it admits as a group of symmetries the Fischer-Griess monster group M –

7
the largest of the sporadic finite groups. (The link between the monster and conformal field
theory was suggested by developments springing from an observation by McKay relating
the monster to the j-function, as we explain more fully in section 3.1.) Arguably, the
FLM model is the most natural known structure with M symmetry. A detailed and
elegant description is in the book [25]; for short summaries, see [26,27], and for subsequent
developments and surveys, see [28-30]. Assuming the (unproved) uniqueness conjecture of
FLM, we propose that their model must give the CFT that is dual to three-dimensional
gravity at c = 24.
For c = 24k, k > 1, we need an analog of requiring that there is no Kac-Moody
symmetry. A plausible analog, expressing the idea that we aim to describe pure gravity,
is that there should be no primary fields of low dimension other than the identity. A
small calculation shows that at c = 24k, the lowest dimension of a primary other than the
identity cannot be greater than k + 1, and if we assume that this dimension is precisely
k + 1, then the partition function is uniquely determined. Conformal field theories with
this property were first investigated by Höhn in [31,32] and have been called extremal
CFT’s; see also [33].
It is not known5 if extremal CFT’s exist for k > 1. If such a CFT does exist, it
is an attractive candidate for the dual of three-dimensional gravity at the appropriate
value of the cosmological constant. The primaries of dimension k + 1 and above would
be interpreted as operators that create black holes. The dimension k + 1 agrees well with
the minimum mass of a BTZ black hole. This statement may sound like magic, since the
value k + 1 is determined from holomorphy and modular invariance without mentioning
black holes; but the result is not so surprising if one is familiar with previous results on
the AdS/CFT correspondence in three dimensions [34].
Section 3 of this paper is devoted to describing the partition function of an extremal
CFT and discussing how such a theory could be related to three-dimensional gravity. In
both sections 2 and 3, we consider also the case of three-dimensional supergravity. More
precisely, we consider only minimal supergravity, corresponding to N = 1 superconformal
symmetry for the boundary CFT. In this case, holomorphic factorization is conceivable
at c = 12k ∗ , k ∗ = 1, 2, 3, . . .. Here there is a little ambiguity about exactly what we

5
Höhn’s definition of an extremal CFT allowed holomorphic factorization up to a phase, so
that c may be a multiple of 8, not 24. As a result, he discussed several examples of extremal
theories that we will not consider here. These examples have k non-integral and less than 2.

8
should mean by an extremal superconformal field theory (SCFT), but pragmatically, there
are good candidates at k ∗ = 1, 2. The k ∗ = 1 theory was constructed by Frenkel, Lep-
owsky, and Meurman, who also conjectured its uniqueness. Its discrete symmetries were
understood only recently in work by Duncan [35]. For k ∗ = 2, the extremal SCFT was con-
structed by Dixon, Ginsparg, and Harvey [26], by modifying the orbifold projection that
had been used [24] in constructing the k = 1 extremal CFT. Interestingly, the extremal
SCFT’s with k ∗ = 1, 2 both admit6 an action of very large discrete groups related to the
Conway group. This is a further indication that unusual discrete groups are relevant to
three-dimensional gravity and supergravity. In fact, we find some hints that supergravity
may have monster symmetry at k ∗ = 4 and baby monster symmetry at k ∗ = 6.
Regrettably, we do not know how to construct new examples of extremal conformal
or superconformal field theories. Section 4 is devoted to a calculation that aims to give
modest support to the idea that new extremal theories do exist. We consider an extremal
CFT with k = 2 and show that its partition function can be uniquely determined on a
hyperelliptic Riemann surface of any genus (including, for example, any Riemann surface
of genus 2). The fact that a partition function with the right properties exists and is
unique for any genus is hopefully a hint that an extremal k = 2 CFT does exist.
We make at each stage the most optimistic possible assumption. Decisive arguments
in favor of the proposals made here are still lacking. The literature on three-dimensional
gravity is filled with claims (including some by the present author [6]) that in hindsight
seem less than fully satisfactory. Hopefully, future work will clarify things.

Advice by J. Maldacena was essential at the outset of this work. I also wish to
thank J. Duncan, G. Höhn, G. Nebe, and J. Teschner for descriptions of their work and
helpful advice; T. Gannon, R. Griess, J. Lepowsky, and A. Ryba for correspondence about
the monster group and related matters; and many colleagues at the IAS and elsewhere,
especially A. Maloney, G. Moore, and S. Shenker, for helpful comments.

2. Gauge Theory And The Value Of c

The goal of the present section is to determine what values of the cosmological con-
stant, or equivalently of the central charge c of the boundary CFT, are suggested by the
relation between three-dimensional gravity and Chern-Simons gauge theory.
6
The fact that they have essentially the same symmetry group was pointed out by J. Duncan,
who also suggested the identification of the k ∗ = 2 theory.

9
 Before proceeding to any calculation, we will dispose of a few preliminary points.
The first is that [36], as long as the three-dimensional spacetime is oriented, as we will
assume in this paper, three-dimensional gravity can be generalized to include an additional
interaction, the Chern-Simons functional of the spin connection ω:
Z  
k′ 2
∆0 I = tr ω ∧ dω + ω ∧ ω ∧ ω . (2.1)
4π W 3

Here we think of ω as an SO(2, 1) gauge field (or an SO(3) gauge field in the case of Eu-
clidean signature). Also, tr is the trace in the three-dimensional representation of SO(2, 1),
and k ′ is quantized for topological reasons (the precise normalization depends on some as-
sumptions and is discussed in sections 2.1 and 2.4). Equivalently, instead of ω, we could
use the SO(2, 2) gauge field A introduced in eqn. (1.2), and add to the action a term of
the form Z  
k′ 2
∆I = tr A ∧ dA + A ∧ A ∧ A , (2.2)
4π M 3
where now tr is the trace in the four-dimensional representation of SO(2, 2). Provided that
the conventional Einstein action (1.1) is also present, it does not matter which form of the
gravitational Chern-Simons interaction we use, since they lead to equivalent theories. If
one adds7 to ω a multiple of e, the Einstein action (1.1) transforms in a way that cancels
the e-dependent part of (2.2), reducing it to (2.1) (while modifying the parameters in the
Einstein action).
For our purposes, the SO(2, 2)-invariant form (2.2) is more useful. This way of writing
the Chern-Simons functional places it precisely in parallel with the Einstein-Hilbert action,
which as in (1.3) can similarly be expressed as a Chern-Simons interaction, defined with a
different quadratic form. We will use the fact that all interactions can be written as Chern-
Simons interactions to constrain the proper quantization of all dimensionless parameters,
including ℓ/G.
We start with the fact that the group SO(2, 2) is locally equivalent to SO(2, 1) ×
SO(2, 1). Moreover, we will in performing the computation assume to start with that
SO(2, 1)×SO(2, 1) is the right global form of the gauge group. (Then we consider covering

7
This operation is invariant under diffeomorphisms and local Lorentz transformations because
ω and e transform in the same way under local Lorentz transformations – a statement that holds
precisely in three spacetime dimensions.

10
groups8 that are only locally isomorphic to SO(2, 1) × SO(2, 1).) Thus, by taking suitable
linear combinations of ω and e, we will obtain a pair of SO(2, 1) gauge fields AL and AR .
These have Chern-Simons interactions
Z   Z  
kL 2 kR 2
I= tr AL ∧ dAL + AL ∧ AL ∧ AL − tr AR ∧ dAR + AR ∧ AR ∧ AR .
4π 3 4π 3
(2.3)
Both kL and kR are integers for topological reasons, and this will lead to a quantization of
the ratio G/ℓ that appears in the Einstein-Hilbert action, as well as the gravitational Chern-
Simons coupling (2.2). The minus sign multiplying the last term in (2.3) is convenient; it
will ensure that kL and kR are both positive.

2.1. Quantization Of Parameters

For completeness, we begin by reviewing the quantization of the Chern-Simons cou-

pling in gauge theory. The basic case to consider is that the gauge group is U (1). The
gauge field A is a connection on a complex line bundle L over a three-manifold W , which
for simplicity we will assume to have no boundary. Naively speaking, the Chern-Simons
action is Z
k
I= A ∧ dA (2.4)
2π W

with some coefficient k. If the line bundle L is trivial, then we can interpret A as a one-
form, and I is well-defined as a real-valued functional. If this were the general situation,
there would be no need to quantize k.
However, in general L is non-trivial, A has Dirac string singularities, and the formula
(2.4) is not really well-defined as written. To do better, we pick a four-manifold M of
boundary W and such that L extends over M . Such an M always exists. Then we pick
an extension of L and A over M , and replace the definition (2.4) with
Z
k
IM = F ∧ F, (2.5)
2π M

where F = dA is the curvature. Now there is no Dirac string singularity, and the definition
of IM makes sense. But IM does depend on M (and on the chosen extension of L, though
we do not indicate this in the notation). To quantify the dependence on M , we consider

8
For an early treatment of coverings in the context of Chern-Simons theory with compact
gauge group, see [37]

11
two different four-manifolds M and M ′ with boundary W and chosen extensions of L. We
can build a four-manifold X with no boundary by gluing together M and M ′ along W ,
with opposite orientation for M ′ so that they fit smoothly along their common boundary.
Then we get Z
k
IM − IM ′ = F ∧ F. (2.6)
2π X
R R
Now, on the closed four-manifold X, the quantity X
F ∧F/(2π)2 represents c (L)
X 1
2

(here c1 is the first Chern class) and so is an integer. In quantum mechanics, the action
function I should be defined modulo 2π (so that exp(iI), which appears in the path
integral, is single-valued). Requiring IM − IM ′ to be an integer multiple of 2π, we learn
that k must be an integer. This is the quantization of the Chern-Simons coupling for U (1)
gauge theory.9
Now let us move on to the case of gauge group SO(2, 1). The group SO(2, 1) is
contractible onto its maximal compact subgroup SO(2), which is isomorphic to U (1). So
quantization of the Chern-Simons coupling for an SO(2, 1) gauge field can be deduced
immediately from the result for U (1). Let A be an SO(2, 1) gauge field and define the
Chern-Simons coupling
Z  
k 2
I= tr A ∧ dA + A ∧ A ∧ A , (2.7)
4π W 3

where tr is the trace in the three-dimensional representation of SO(2, 1). Then, in order
for I to be part of the action of a quantum theory, k must be an integer. The reason that
the factor 1/2π in (2.5) has been replaced by 1/4π in (2.7) is simply that, when we identify
U (1) with SO(2) and then embed it in SO(2, 1), the trace gives a factor of 2.

Coverings
So we have obtained the appropriate quantization of the Chern-Simons coupling for
gauge group SO(2, 1). However, this is not quite the whole story, because SO(2, 1) is not
simply-connected. As it is contractible to SO(2) ∼= U (1), it has the same fundamental
group as U (1), namely Z. Hence it is possible, for every positive integer n, to take an

9
There is a refinement if the three-manifold W is endowed with a spin structure. In this
case, k can be a half-integer, as explained in [38]. This refinement is physically realized in the
quantum Hall effect with filling fraction 1. That effect can be described by an electromagnetic
Chern-Simons coupling with k = 1/2; the half-integral value is consistent because the microscopic
theory has fermions and so requires a spin structure.

12
n-fold cover of SO(2, 1). The most familiar of these is the two-fold cover, SL(2, R). In
addition, SO(2, 1) has a simply-connected universal cover.
We want to work out the quantization of the Chern-Simons interaction if SO(2, 1) is
replaced by one of these covering groups. Again, it is convenient to start with U (1). To
say that the gauge group of an abelian gauge theory is U (1) rather than R means precisely
that the possible electric charges form a lattice, generated by a fundamental charge that
R
we call “charge 1.” Dually, the magnetic fluxes are quantized, with C F/2π ∈ Z for any
two-cycle C. Replacing U (1) by an n-fold cover means that the electric charges take values
R
in n−1 Z, and dually, the magnetic fluxes are divisible by n, C F/2π ∈ nZ. As a result,
R
for a four-manifold X, we have C F ∧ F/(2π)2 ∈ n2 Z. So, in requiring that the Chern-
Simons function (2.4) should be well-defined modulo 2π, we require that k ∈ n−2 Z. This
is the appropriate result for an n-fold cover. In the case of the universal cover, with U (1)
replaced by R, the magnetic fluxes vanish and there is no topological restriction on k.
These statements carry over immediately to covers of SO(2, 1), whose covers are all
contractible to corresponding covers of U (1). So for an n-fold cover of SO(2, 1), we require

k ∈ n−2 Z, (2.8)

and for the universal cover of SO(2, 1), k is arbitrary and can vary continuously.

Diagonal Covers
There is more to say, because three-dimensional gravity is actually related to
SO(2, 1) × SO(2, 1) gauge theory, not just to gauge theory with a single SO(2, 1). So
we should consider covers of SO(2, 1) × SO(2, 1) that do not necessarily come from sepa-
rate covers of the two factors.
As SO(2, 1)×SO(2, 1) is contractible to SO(2)×SO(2) = U (1)×U (1), we can proceed
by first analyzing the U (1) × U (1) case. We consider a U (1) × U (1) gauge theory with
gauge fields A, B and a Chern-Simons action
Z Z
kL kR
I= A ∧ dA − B ∧ dB. (2.9)
2π W 2π W

To define I in the topologically non-trivial case, we pick a four-manifold M over which

everything extends and define
Z  
kL kR
IM = FA ∧ FA − FB ∧ FB , (2.10)
M 2π 2π

13
where FA and FB are the two curvatures. This is well-defined mod 2π if
Z  
kL kR
IX = FA ∧ FA − FB ∧ FB (2.11)
X 2π 2π

is a multiple of 2π for any U (1) × U (1) gauge field over a closed four-manifold X.
In U (1) × U (1) gauge theory, the charge lattice is generated by charges (1, 0) and
(0, 1), and the cohomology classes x = FA /2π and y = FB /2π are integral. For a cover of
U (1) × U (1), we want to extend the charge lattice. To keep things simple, we will consider
only the case that will actually be important in our application: a diagonal cover, in which
one adds the charge vector (1/n, 1/n) for some integer n. In this case, x and y are still
integral, and their difference is divisible by n: x = y + nz where n is an integral class. We
have Z Z
2
IX = 2π(kL − kR ) y + 2πkL (n2 z 2 + 2nyz). (2.12)
X X

The condition that this is a multiple of 2π for any X and any integral classes y, z is that

n−1 Z if n is odd
kL ∈ −1
(2n) Z if n is even (2.13)
kL − kR ∈ Z.

These are also the restrictions on kL and kR if the gauge group is a diagonal cover of
SO(2, 1) × SO(2, 1) with action (2.3). For example, the group SO(2, 2) is a double cover of
SO(2, 1) × SO(2, 1), and this cover corresponds to the case n = 2 of the above discussion.
So for SO(2, 2) gauge theory, the appropriate restriction on the Chern-Simons levels is

1
kL ∈ Z
4 (2.14)
kL − kR ∈ Z.

For a general n-fold diagonal cover, one should use (2.13).

It is also possible to form a “universal diagonal cover,” corresponding roughly to the
limit n → ∞ in the above formulas. With this gauge group, there is no restriction on
kL , but kL − kR is an integer. In terms of three-dimensional gravity, to which we return
next, this corresponds to letting ℓ/G be a freely variable parameter, while the gravitational
Chern-Simons coupling k ′ defined in eqn. (2.2) is an integer. As explained in section 1.3,
although this is the state of affairs classically in three-dimensional gravity, it cannot be
the correct answer quantum mechanically.

14
2.2. Comparison To Three-Dimensional Gravity

So far, we have understood the appropriate gauge theory normalizations for the Chern-
Simons action

I =kL IL + kR IR
Z   Z  
kL 2 kR 2
= tr AL ∧ dAL + AL ∧ AL ∧ AL − tr AR ∧ dAR + AR ∧ AR ∧ AR .
4π 3 4π 3
(2.15)
Our next step will be to express AL and AR , which are gauge fields of SO(2, 1)×SO(2, 1) (or
a covering group) in terms of gravitational variables, and thereby determine the constraints
on the gravitational couplings. We have

kL + kR (IL + IR )
I= (IL − IR ) + (kL − kR ) . (2.16)
2 2

The term in (2.16) proportional to IL − IR will gave the Einstein-Hilbert action (1.1),
while the term proportional to (IL + IR )/2 is equivalent to the gravitational Chern-Simons
coupling (2.2) with coefficient k ′ = kL − kR .
P
The spin connection ω ab = i ab
i dx ωi is a one-form with values in antisymmetric
3 × 3 matrices. The vierbein is conventionally a one-form valued in Lorentz vectors,
P
ea = i dxi eai . The metric is expressed in terms of e in the usual way, gij dxi ⊗ dxj =
P a b
ab ηab e ⊗e , where η = diag(−1, 1, 1) is the Lorentz metric; and the Riemannian volume
√
form is d3 x g = 16 ǫabc ea ∧eb ∧ec , where ǫabc is the antisymmetric tensor with, say, ǫ012 = 1.
All this has an obvious analog in any dimension. However, in three dimensions, a Lorentz
vector is equivalent to an antisymmetric tensor; this is the fact that makes it possible to
relate gravity and gauge theory. It is convenient to introduce ∗eab = ǫabc ec , which is a
one-form valued in antisymmetric matrices, just like ω. We raise and lower local Lorentz
indices with the Lorentz metric η, so 21 ǫabc ǫbcd = −δda , and ec = − 12 ǫabc ∗ebc .
We can combine ω and ∗e and set AL = ω − ∗e/ℓ, AR = ω + ∗e/ℓ. A small computation
gives Z Z
1 ∗ 1
IL − IR = − tr e(dω + ω ∧ ω) − tr (∗e ∧ ∗e ∧ ∗e). (2.17)
πℓ 3πℓ3
1
P
In terms of the matrix-valued curvature two-form Rab = (dω + ω ∧ ω)ab = 2 ij dxi ∧
ab ab
dxj Rij , where Rij is the Riemann tensor, and the metric tensor g, this is equivalent to
Z  
1 3 √ 2
IL − IR = d x g R+ 2 . (2.18)
πℓ ℓ

15
Remembering the factor of (kL + kR )/2 in (2.16), we see that this agrees with the Einstein-
Hilbert action (1.1) precisely if
ℓ
kL + kR = . (2.19)
8G
The central charge of the boundary conformal field theory was originally computed
by Brown and Henneaux [21] for the case that the gravitational Chern-Simons coupling
k ′ = kL − kR vanishes. In this case, we set k = kL = kR = ℓ/16G. The formula for the
central charge is c = 3ℓ/2G, and this leads to c = 24k. For the case k ′ = 0, the boundary
CFT is left-right symmetric, with cL = cR , so in fact cL = cR = 24k.
In general, the boundary CFT has left- and right-moving Virasoro algebras that can be
interpreted (for a suitable orientation of the boundary) as boundary excitations associated
with AL and AR respectively. So the central charges cL and cR are functions only of kL
and kR , respectively. Hence the generalization of the result obtained in the last paragraph
is
(cL , cR ) = (24kL , 24kR ). (2.20)

2.3. Holomorphic Factorization

In conformal field theory in two dimensions, the ground state energy is −c/24. More
generally, if there are separate left and right central charges cL and cR , the ground state
energies for left- and right-movers are (−cL /24, −cR /24). Modular invariance says that
the difference between the left- and right-moving ground state energies must be an integer.
In the above calculation, (cL − cR )/24 = kL − kR . According to (2.20), this is an integer
provided that the gravitational Chern-Simons coupling k ′ = kL − kR is integral.
Holomorphic factorization requires that the left- and right-moving ground state en-
ergies should be separately integral, so that there is modular invariance separately for
left-moving and right-moving modes of the CFT. Thus, for holomorphic factorization, cL
and cR must both be integer multiples of 24. This is the only constraint, since holomorphic
CFT’s with c = 24 do exist (and have been classified [23] modulo a conjecture mentioned
in section 1.4).
According to (2.20), the condition for cL and cR to be multiples of 24 is precisely that
kL and kR must be integers. As in our discussion of eqn. (2.7), this is the right condition
if the gauge group is precisely SO(2, 1) × SO(2, 1) rather than a covering group.
It is possible to give an intuitive explanation of why this is the right gauge group if the
boundary CFT is supposed to be holomorphically factorized. The Virasoro algebra is, of

16
course, infinite-dimensional, but it has a finite-dimensional subalgebra, generated by L±1
and L0 , that is a symmetry of the vacuum. It is customary to refer to the corresponding
symmetry group of the vacuum as SL(2, R), but in fact, in a holomorphic CFT, in which all
energies are integers, the group that acts faithfully is really SO(2, 1). So a holomorphically
factorized CFT has symmetry group SO(2, 1) × SO(2, 1), and it is natural that this is the
right gauge group in a gauge theory description of (aspects of) the dual gravitational
theory.
Now let us consider some other possible gauge groups. One possibility is to take a
double cover of each factor of SO(2, 1) ×SO(2, 1), taking the gauge group to be SL(2, R) ×
SL(2, R). The appropriate restriction on the gauge theory couplings was determined in
1
(2.8) (where we should set n = 2) and is that kL and kR take values in 4
Z. Hence the
central charges cL and cR are multiples of 6. In particular, SL(2, R)×SL(2, R) gauge theory
would allow us to consider values of the couplings that contradict modular invariance of
the boundary CFT.
There is perhaps a more intuitive argument suggesting that SL(2, R) ×SL(2, R) is not
the right group to consider. If the gauge group is SL(2, R) × SL(2, R), then upon taking
the two-dimensional representation of one of the SL(2, R)’s, we get a two-dimensional
real vector bundle over W which generalizes what in classical geometry is the spin bundle.
Thus, this would be a theory in which, in the classical limit, W is a spin manifold, endowed
with a distinguished spin structure (or even two of them). That is appropriate in a theory
with fermions, but not, presumably, in a theory of pure gravity.
We similarly lose modular invariance and have difficult to interpret geometric struc-
tures if we consider other non-diagonal covers of SO(2, 1) × SO(2, 1). So let us discuss the
diagonal covers of SO(2, 1) × SO(2, 1) that were considered at the end of section 2.1. We
know that the universal diagonal cover is not right, since then ℓ/G could vary continuously.
This leaves the possibility of an n-fold diagonal cover for some n. The best-motivated ex-
ample is perhaps the two-fold cover SO(2, 2), which is the symmetry of Anti de Sitter
spacetime (as opposed to a cover of that spacetime). In this case, according to (2.14), kL
and kR can take values in Z/4, as long as k ′ = kL − kR is integral. For the boundary CFT,
this means that cL and cR can be multiples of 6 (with their difference a multiple of 24).
For example, let us consider a hypothetical CFT of (cL , cR ) = (6, 6). The ground state
energies are (−1/4, −1/4). As these values are not integers, the symmetry group of the
ground state is not SO(2, 1) × SO(2, 1), but a four-fold diagonal cover, which is a double

17
cover of the gauge group SO(2, 2) that was assumed. Such a CFT cannot be holomorphi-
cally factorized, since the left- and right-moving ground state energies are not integers; it
cannot even be holomorphically factorized up to a phase.10 The structure is considerably
more complicated than in the holomorphically factorized case. Similar remarks apply to
other diagonal covers.
Pragmatically, the most important argument against trying to describe three-
dimensional gravity via a cover of SO(2, 1) × SO(2, 1) may be simply the fact that no
good candidates are known. For example, no especially interesting bosonic CFT (as op-
posed to a superconformal field theory) seems to be known at c = 6, 12, or 18, which are
values that we would expect if we use the gauge group SO(2, 2). By contrast, at c = 24,
which is natural for SO(2, 1) × SO(2, 1), the Frenkel-Lepowsky-Meurman monster theory
is a distinguished candidate, as we mentioned in section 1.4 and will explain in more detail
in section 3.
The simplest possible hypothesis is that the right gauge group to consider in study-
ing three-dimensional pure gravity is SO(2, 1) × SO(2, 1), with integral kL and kR and
holomorphic factorization of the boundary CFT. It would be highly unnatural to overlook
the fact that SO(2, 1) × SO(2, 1) gauge theory leads to precisely the values of the central
charge at which the drastic simplification of the boundary CFT known as holomorphic
factorization is conceivable. Moreover, the fact that the classical action (2.15) of the gauge
theory is a sum of decoupled actions for AL and AR – related respectively to left- and
right-moving modes of the boundary CFT – is a hint of holomorphic factorization of the
boundary theory.
At any rate, regardless of whether they give the whole story, it does seem well-
motivated to look for holomorphically factorized CFT’s with central charges multiples
of 24 that are dual to three-dimensional pure gravity at special values of the cosmological
constant. That will be our main focus in the rest of this paper.

10
The left-moving partition function would have to be Φ/∆1/4 , where ∆ = η(q)24 is the
discriminant, η being the Dedekind eta function, and Φ is a modular form of weight 3. The
power of ∆ was determined to get the right ground state energy; modular invariance implies that
the modular weight of Φ must equal that of ∆1/4 . As there is no modular form of weight 3,
such a theory does not exist. Even at (cL , cR ) = (12, 12), it is not possible to have holomorphic
factorization up to a phase. To get modular invariance and the right ground state energy, the
left-moving partition function would have to be E6 /∆1/2 , where E6 is the Eisenstein series of
weight 6. However, the coefficients in the q-expansion of this function are not positive.

18
2.4. Interpretation

A few further words of interpretation seem called for.

We do not claim that three-dimensional gravity is equivalent, nonperturbatively, to
Chern-Simons gauge theory. Some objections to this idea were described in section 1.1. We
know that Chern-Simons gauge theory is useful for perturbation theory, as was explained
in that section, and we hope that it is useful for understanding some nonperturbative
questions. We used the gauge theory approach to get some hints about the right values of
the cosmological constant (or equivalently of the central charge) simply because it was the
only tool available. We certainly do not claim to have a solid argument that the values of
(cL , cR ) suggested by SO(2, 1) × SO(2, 1) gauge theory are the only relevant ones.
Implicit in the gauge theory approach to quantization of the dimensionless parameter
ℓ/G is an assumption that at least some of the non-geometrical states that can be described
in gauge theory make sense. Indeed, when understood in classical geometry with the metric
assumed to be smooth and nondegenerate, the Einstein-Hilbert action (1.1) is well-defined
as a real-valued function. The topological problems that cause it to be multivalued when
interpreted in gauge theory as a Chern-Simons function, and lead to quantization of ℓ/G,
depend on allowing certain configurations that are natural in gauge theory, but singular
in geometry because the vierbein is not invertible.
Furthermore, holomorphic factorization most likely is possible only if the path integral
includes a sum over some non-geometrical configurations. In quantum gravity, we at
least expect to sum over all topologies of a three-manifold W , perhaps with some fixed
asymptotic behavior. The choice of topology should be expected to affect both the left-
and right-movers of the boundary CFT. To achieve holomorphic factorization, there must
presumably be some sort of separate topological sum for left-movers and right-movers. As
this does not occur in classical geometry, it must depend on some sort of non-geometric
contributions to the path integral – though these contributions may be exponentially small
in many circumstances.

The Gravitational Chern-Simons Action Reconsidered

Our discussion of the quantization of the gravitational Chern-Simons coupling (2.1)
Z  
k′ 2
∆0 I = tr ω ∧ dω + ω ∧ ω ∧ ω (2.21)
4π W 3
has been based entirely on gauge theory. We now want to describe the similar but slightly
different answer that would come from classical differential geometry. As always, the

19
differences reflect the fact that the gauge theory analysis allows some non-geometrical
configurations.
We first briefly restate the gauge theory analysis. ω is a connection on an SO(2, 1) or
(in Euclidean signature) SO(3) bundle over a three-manifold W . As usual, to define ∆0 I
more precisely, we pick an oriented four-manifold M of boundary W with an extension of
ω over W . Then we define Z
k′
IM = tr F ∧ F. (2.22)
4π M

If M is replaced by some other four-manifold M ′ , and X = M − M ′ is a four-manifold

without boundary obtained by gluing together M and M ′ , then
Z Z
k′ ′
IM − IM ′ = tr F ∧ F = 2πk p1 (F ), (2.23)
4π X X
R
where p1 (F ) = (1/8π 2 )tr F ∧ F is the first Pontryagin form. In general, X
p1 (F ) can be
any integer, so the condition that the indeterminacy in IM is an integer multiple of 2π
means simply that k ′ is an integer.
This is the right answer if the right thing to do is to simply think of ω as an SO(3) or
SO(2, 1) gauge field, ignoring its classical relation to gravity. However, in classical gravity,
one can get a better answer. In classical gravity, ω is a connection on the tangent bundle
T W of W . Now replace T W by T W ⊕ ǫ, where ǫ is a trivial real line bundle. Then ω can
be regarded as a connection on T W ⊕ ǫ in an obvious way, and T W ⊕ ǫ extends over M
as the tangent bundle of M . With this choice, (2.22) becomes
Z
k′
IM = tr R ∧ R, (2.24)
4π M

where R is the curvature form of M , and (2.23) becomes

Z
′
IM − IM ′ = 2πk p1 (R). (2.25)
X

The effect of this is that instead of the first Pontryagin number of a general bundle over
X, as in (2.23), we have here the first Pontryagin number p1 (T X) of the tangent bundle of
X. This number is divisible by 3, because of the signature theorem, which says that for a
four-manifold X, p1 (T X)/3 is an integer, the signature of X. Hence, in the gravitational
interpretation, the condition on k ′ is
1
k′ ∈ Z. (2.26)
3
20
 There is also a variant of this. If W is a spin manifold and we are willing to define the
gravitational Chern-Simons coupling (2.21) in a way that depends on the spin structure of
W (this does not seem natural in ordinary gravity, but it may be natural in supergravity,
which we come to in section 3.2), the condition on k ′ can be further relaxed. In this
case, we can select M so that the chosen spin structure on W extends over M . If M ′ is
another choice with the same property, then X = M − M ′ is a spin manifold. But (as
follows from the Atiyah-Singer index theorem for the Dirac equation) the signature of a
four-dimensional spin manifold is divisible by 16. So in this situation p1 (T X) is a multiple
of 48, and the result for k ′ under these assumptions is
1
k′ ∈ Z. (2.27)
48

The Physical Hilbert Space

Now we shall discuss the meaning for gravity of the physical Hilbert space of Chern-
Simons gauge theory.
In three-dimensional Chern-Simons gauge theory, one can fix a Riemann surface C and
construct a Hilbert space HC of physical states obtained by quantizing the given theory on
C. This Hilbert space depends on the Chern-Simons couplings, so when we want to be more
precise, we will call it HC (kL , kR ). In [6], it was proposed that the physical Hilbert space
of three-dimensional gravity on a Riemann surface C should be obtained in essentially
this way, using SO(2, 1) × SO(2, 1) gauge theory (or something similar, depending on
whether the cosmological constant is positive, negative, or zero). Some serious objections
to the claim that pure gravity and Chern-Simons gauge theory are equivalent in three
dimensions were noted in section 1.1. Yet remarkable progress has made made [8,11] in
understanding the quantization of SO(2, 1) gauge theory, and even more remarkably, in
relating this quantization to Liouville theory. It therefore seems likely that the Hilbert
space obtained by quantizing Chern-Simons gauge theory of SO(2, 1) × SO(2, 1) means
something for three-dimensional gravity even if the proposal in [6] was premature. The
rest of this subsection will be devoted to an attempt (not used in the rest of the paper) to
reconcile the different points of view. See [39] for some useful background.
We may start by asking what we mean by the physical Hilbert space HC obtained
in quantum gravity by quantizing on a closed manifold C. What type of question is this
space supposed to answer? Quantization on, for example, an asymptotically flat spacetime
leads to a Hilbert space that can be interpreted in a relatively straightforward way, but the
physical meaning of a Hilbert space obtained by quantizing on a compact spatial manifold
(if such a Hilbert space can be defined at all) is not clear.

21
 C
W

Fig. 1:

The Hartle-Hawking wavefunction Ψ is computed by integrating over three-

manifolds W with a given boundary C. The dotted line labeled r denotes a path
in W connecting two points in C. It is possible to vary W so that the length of this
path goes to zero with no change in the geometry of C. In this limit, W becomes
singular even though C is smooth. This gives a “source term” in the Wheeler-de
Witt equation, because of which the Hartle-Hawking wavefunction does not obey
this equation.

One line of thought that is relatively close to working is to consider the Hartle-Hawking
wavefunction [40] and claim that it is a vector in HC . (The obvious idea that quantum
mechanical probabilities would be calculated in terms of inner products of vectors in a
Hilbert space HC of physical states is afflicted with similar but more serious problems.)
The Hartle-Hawking wavefunction is a functional of metrics on C. For every metric h
on C, we define Ψ(h) as the result of performing a path integral over three-manifolds W
whose boundary is C and whose metric g coincides with h on the boundary. Formally, one
can try to argue that Ψ(h) obeys the Wheeler-de Witt equation and thus is a vector in
a Hilbert space HC of solutions of this equation. Moreover, one can formally match the
Wheeler-de Witt equations of gravity with the conditions for a physical state in Chern-
Simons gauge theory. Though many steps in these arguments work nicely, one runs into

22
trouble because a Riemann surface can be immersed, rather than embedded, in a three-
manifold, and hence it is possible for W to degenerate without C degenerating (fig. 1). As
a result, the Hartle-Hawking wavefunction does not obey the Wheeler-de Witt equation
and is not a vector in HC .
In the case of negative cosmological constant, the boundary CFT gives a sort of cure
for the problem with the Hartle-Hawking wavefunction. Instead of thinking of C as an
ordinary boundary of W , we think of it as a conformal boundary at infinity. The partition
b
function Ψ(h) of the boundary CFT is defined by performing the path integral over all
choices of W with C as conformal boundary. This is well-behaved, because, with C at
b
conformal infinity, it is definitely embedded rather than immersed. Moreover, Ψ(h) is a
sort of limiting value of the Hartle-Hawking wavefunction. Indeed, let φ be a positive
b
function on C. Then Ψ(h) is essentially11 the limiting value of Ψ(eφ h) as φ → ∞.
b
This suggests that we should be able to think of Ψ(h) as a vector in the Hilbert space
HC associated with three-dimensional gravity and a two-manifold C. This viewpoint can
be explained more fully. The phase space of SO(2, 1) × SO(2, 1) Chern-Simons theory on
C is the space of SO(2, 1) × SO(2, 1) flat connections on C. The space of SO(2, 1) flat con-
nections on C has several topological components (labeled by the first Chern class, which
comes in because SO(2, 1) is contractible to SO(2) ∼= U (1)). One of these components,
the only one that can be simply interpreted in terms of classical gravity with negative
cosmological constant,12 is isomorphic to Teichmuller space T . Thus, this component of
the classical phase space M is a product of two copies of T , parametrized by a pair of
points τ, τ ′ ∈ T . One can quantize M (or at least this component of it) naively by using
the standard holomorphic structure of T . If we do this, the wavefunction of a physical
state is a “function” of τ and τ ′ that is holomorphic in τ and antiholomorphic in τ ′ . (An-
tiholomorphy in one variable reflects the relative minus sign in the Chern-Simons action
(2.15); we assume that kL and kR are positive.) Actually, a physical state wavefunction
Ψ(τ, τ ′ ) is not quite a function of τ and τ ′ in the usual sense, but is a form, of weights
determined by kL and kR . So it takes values in a Hilbert space HC (kL , kR ) that depends on
the Chern-Simons couplings. Such a wavefunction Ψ(τ, τ ′ ) is determined by its restriction

11
One needs some renormalization in this limit; the necessary renormalization reflects the
b (h) is not quite a function only of the conformal
conformal anomaly. Because of this anomaly, Ψ
structure of C but a function of the metric that transforms with a certain weight under conformal
rescalings. As a result, the precise choice of φ matters, but only in a rather simple way.
12
In the approach to quantization developed in [8,11], this is the only component considered.

23
to the diagonal subspace τ = τ ′ . Moreover, if we want to make a relation to gravity, it is
natural to require that Ψ should be invariant under the diagonal action of the mapping
class group on τ and τ ′ ; this condition is compatible with restricting to τ = τ ′ .
Similarly, the partition function of a CFT on the Riemann surface C is a not neces-
sarily holomorphic “function” (actually a form of appropriate weights) Ψ(τ, τ). Being real
analytic, Ψ can be analytically continued to a function Ψ(τ, τ ′ ) with τ ′ at least slightly
away from τ . It does not seem to be a standard fact13 that Ψ analytically continues to
a holomorphic function on T × T (with invariance only under one diagonal copy of the
mapping class group). However, this is true in genus 1, since the partition function can
be defined as Tr q L0 q ′L0 , where we can take q and q ′ to be independent complex variables
of modulus less than 1. It seems very plausible that the statement is actually true for all
values of the genus, since one can move on Teichmuller space by “cutting” on a circle and
inserting q L0 q ′L0 . If so, the partition function of the CFT can always be interpreted as a
vector14 in the Chern-Simons Hilbert space HC (kL , kR ).
If we are given a theory of three-dimensional gravity, possibly coupled to other fields,
the partition function of the dual CFT is a wavefunction Ψ(τ, τ ′ ) which, according to the
conjecture just stated, is a vector in HC (kL , kR ). Any gravitational theory of the same
central charges leads to another vector in the same space.
From this point of view, it seems that we should not claim, as was done in [6], that
HC (kL , kR ) is a space of physical states that are physically meaningful in pure three-
dimensional gravity. Rather, a particular bulk gravitational theory, such as pure gravity,
gives rise to a particular dual CFT whose partition function gives a definite vector in
HC (kL , kR ). Another gravitational theory, perhaps with matter fields, whose dual CFT
has the same values of the central charges, will lead to a dual partition function that is
another vector in the same space. Thus, HC (kL , kR ) is in a sense a universal target for
gravitational theories – with arbitrary matter fields – of given central charges.
We have formulated this for a particular Riemann surface C, but in either the gravita-
tional theory or the dual CFT, C can vary and there is a nice behavior when C degenerates.
So it is more natural to think of this as a structure that is defined for all Riemann surfaces.
In conformal theory, this perspective is described in [41].

13
However, G. Segal has obtained results in this direction.
14
It may be necessary here to extend HC (kL , kR ) to a space of forms on T ×T that are invariant
under the action of the mapping class group but not necessarily square-integrable.

24
2.5. Analog For Supergravity

Here we will discuss the extension of some of these ideas to three-dimensional super-
gravity.
We consider primarily minimal supergravity, corresponding to the case that the bound-
ary CFT has N = 1 supersymmetry for left-movers or right-movers or perhaps both. Thus,
for, say, the left-movers, the Virasoro algebra is replace by an N = 1 super-Virasoro alge-
bra. The symmetry algebra generated by L±1 and L0 is extended (in the Neveu-Schwarz
sector) to a superalgebra that also includes the fermionic generators G±1/2 . This is the Lie
superalgebra of the supergroup OSp(1|2), whose bosonic part is Sp(2, R), or equivalently
SL(2, R). In particular, since the operators G±1/2 transform in the two-dimensional rep-
resentation of SL(2, R), the relevant group is definitely SL(2, R) (or possibly a covering of
it), not its quotient SO(2, 1).
For definiteness, consider a two-dimensional CFT with (0, 1) supersymmetry, that is,
with N = 1 supersymmetry for right-movers and none for left-movers. Then left-movers
have an ordinary Virasoro symmetry and right-movers have an N = 1 super-Virasoro
symmetry. Such a theory can be dual to a three-dimensional supergravity theory, which
classically can be described by a Chern-Simons gauge theory in which the gauge supergroup
is SO(2, 1) × OSp(1|2), or possibly a cover thereof. For brevity, we will here assume that
the gauge group is precisely SO(2, 1) × OSp(1|2). The action is the obvious analog of
(2.15):

I =kL IL + kR IR
Z   Z  
kL 2 kR 2
= tr AL ∧ dAL + AL ∧ AL ∧ AL − str AR ∧ dAR + AR ∧ AR ∧ AR .
4π 3 4π 3
(2.28)
Here AL is an SO(2, 1) gauge field, AR is an OSp(1|2) gauge field, and str is the supertrace
in the adjoint representation of OSp(1|2).
We want to generalize the analysis of section 2.1 to determine the allowed values of kL
and kR . There is actually almost nothing to do. AL is simply an SO(2, 1) gauge field, so
kL must be an integer. As for AR , we can for topological purposes replace the supergroup
OSp(1|2) by its bosonic reduction SL(2, R), since the fermionic directions are infinitesimal
and carry no topology. So we can borrow the result of (2.8):

kL ∈ Z
1 (2.29)
kR ∈ Z.
4
25
 We still have (cL , cR ) = (24kL , 24kR ), since the Brown-Henneaux computation of the
central charge depends only on the bosonic part of the action. So cL must be a multiple
of 24, as before, but now it seems that cR should be a multiple of 6.
This is not the result that one might hope for, because holomorphic factorization in
N = 1 superconformal field theories requires that cR should be a multiple of 12, not 6.
So half of the seemingly allowed values of cR are difficult to interpret in the spirit of this
paper. Replacing SO(2, 1) × OSp(1|2) by a covering group would only make things worse,
as in the bosonic case.
There are many conceivable ways to interpret this result, including the possibility
that our assumptions have been too optimistic. However, one additional possibility seems
worthy of mention here. Part of the structure of a superconformal field theory is that there
are Ramond-sector vertex operators. They introduce a “twist” in the supercurrent, which
has a monodromy of −1 around a point at which a Ramond vertex operator is inserted.
Let us assume that the superconformal dual of three-dimensional supergravity should be
a theory in which Ramond vertex operators make sense. What is the gravitational dual of
an insertion of such a vertex operator?
As in fig. 2, points in the conformal boundary of spacetime at which a Ramond vertex
operator is inserted must be connected in the bulk by lines – such as the line labeled L in
the figure – around which the fermionic fields of OSp(1|2) (that is, the gravitinos) have a
monodromy −1. It is plausible to interpret these lines, which we will call Ramond lines,
as world-lines of Ramond-sector black holes. A spacetime history containing a black hole
trajectory is of course more complicated than a spacetime with a simple line drawn in
it. But for our present purposes, which are purely topological, the difference may not be
important. There can also be Neveu-Schwarz sector black holes (and black holes in purely
bosonic gravity), but we are about to discuss an effect which seems special to Ramond-
sector black holes.
So now we have a new problem. We want to define
Z  
1 2
IR = str AR ∧ dAR + AR ∧ AR ∧ AR (2.30)
4π W 3

in the presence of a Ramond world-line L on W . In the absence of the Ramond line, we

know that IR in OSp(1|2) gauge theory is defined modulo 4 · 2π, which is why as stated in
(2.29), kR can be a multiple of 1/4. What happens in the presence of the Ramond line?
Rather than solving a new topological problem, we can take the following shortcut.
Let W ′ be a new three-manifold obtained by taking a double cover of W branched over the

26
 Fig. 2:

C is a Riemann surface with a pair of insertions of Ramond-sector vertex

operators at points labeled p and p′ . They are connected by a line L that runs
in a three-manifold W whose boundary is C. Supergravity on W contributes to
conformal field theory on C. The fermion fields of OSp(1|2) receive a minus sign
in monodromy around L.

line L. We let IR (W ) be the action (2.30) and IR (W ′ ) be the corresponding action for the
gauge field AR pulled back to W ′ . When pulled back to W ′ , the singularity of AR along
the Ramond line disappears, so IR (W ′ ) is defined modulo 4 · 2π. There is no better way
to define IR (W ) in the presence of a Ramond line than to say that IR (W ) = IR (W ′ )/2.
So IR (W ) is defined modulo 2 · 2π.
This means that kR should be a multiple of 1/2, not 1/4. So in other words, if
including Ramond lines is the right thing to do, we get

kL ∈ Z
1 (2.31)
kR ∈ Z,
2
and hence cL and cR are multiples of 24 and 12, respectively.
Unfortunately, the above “derivation” is little more than a scenario to try to justify
the answer that we hoped for. However, a good pragmatic reason to focus on the case

27
that cR is a multiple of 12 is that there are interesting candidate superconformal field
theories (SCFT’s) in that case, as we discuss in section 3. There are no obvious interesting
candidates at cR = 6, 18, etc.
In the supersymmetric case it is convenient to express the Chern-Simons coupling k
as k = k ∗ /2, where we will focus on the case that k ∗ is an integer. In terms of k ∗ , the
central charge is c = 12k ∗ .

Some Generalizations
We conclude this section by briefly mentioning some simple generalizations.
First of all, (1, 1) supersymmetry in two dimensions, with N = 1 super-Virasoro
symmetry for both left-movers and right-movers, is dual to three-dimensional supergravity
theories related to OSp(1|2) × OSp(1|2) Chern-Simons gauge theory. If one wants the left-
or right-movers to have more than N = 1 supersymmetry, one simply replaces OSp(1|2)
by an appropriate supergroup with more fermionic generators. For example, OSp(2|2) is
related to N = 2 supersymmetry, P SU (2|2) is related to what is usually called N = 4
supersymmetry (with the “small” N = 4 superconformal algebra), and OSp(r|2) with
r > 2 is related to theories with “large” N = k superconformal algebras. (The most widely
studied case of these algebras is r = 4; for example, see [42].)
It is also possible to consider other extended chiral algebras, apart from superconfor-
mal algebras. For example, one can start in three dimensions with an SL(3, R) Chern-
Simons gauge theory, which plausibly may be related to some sort of three-dimensional W3
gravity theory in much the same (not fully understood) way that SL(2, R) Chern-Simons
theory is related to ordinary three-dimensional gravity. A dual CFT would very likely
then have a W3 chiral algebra. An analogous statement plausibly holds for many extended
chiral algebras.

3. Partition Functions

In this section, we will determine what we propose to be the exact spectrum of physical
states of three-dimensional gravity or supergravity with negative cosmological constant, in
a spacetime asymptotic at infinity to Anti de Sitter space. Equivalently, we will determine
the genus one partition function of the dual CFT.
In all cases, we work at the values of ℓ/G at which holomorphic factorization is pos-
sible. These values were related to gauge theory in section 2. We assume holomorphic

28
factorization, and mainly consider only the holomorphic sector of the theory. The full
partition function is the product of the function we determine and a similar antiholomor-
phic function. These partition functions have been studied before [31,32], with different
motivation.

3.1. The Bosonic Case

We begin with the bosonic case. What are the physical states of pure gravity in a
spacetime asymptotic at infinity to AdS3 ?
Since there are no gravitational waves in the theory, the only state that is obvious
at first sight is the vacuum, corresponding in the classical limit to Anti de Sitter space.
In a conformal field theory with central charge c = 24k, the ground state energy is L0 =
−c/24 = −k. The contribution of the ground state |Ωi to the partition function Z(q) =
Tr q L0 is therefore q −k .
Of course, there is more to the theory than just the ground state. According to Brown
and Henneaux [21], a proper treatment of the behavior at infinity leads to the construction
of a Virasoro algebra that acts on the physical Hilbert space. The Virasoro generators
Ln , n ≥ −1 annihilate |Ωi, but by acting with L−2 , L−3 , . . ., we can make new states of
Q∞ sn P
the general form n=2 L−n |Ωi, with energy −k + n nsn . (We assume that all but finitely
many of the sn vanish.) If these are the only states to consider, then the partition function
would be
∞
Y 1
Z0 (q) = q −k n
. (3.1)
n=2
1 − q
This cannot be the complete answer, because the function Z0 (q) is not modular-
invariant. There must be additional states such that Z0 (q) is completed to a modular-
invariant function.
Additional states are expected, because the theory also has BTZ black holes. The main
reason for writing the present paper, after all, is to understand the role of the BTZ black
holes in the quantum theory. We will assume that black holes account for the difference
between the naive partition function Z0 (q) and the exact one Z(q). To use this assumption
to determine Z(q), we need to know something about the black holes.
The classical BTZ black hole is characterized by its mass M and angular momentum
J. In terms of the Virasoro generators,
1
M = (L0 + L0 )
ℓ (3.2)
J =(L0 − L0 ),

29
so L0 = (ℓM + J)/2, L0 = (ℓM − J)/2. The classical BTZ black hole obeys
M ℓ ≥ |J|, or L0 , L0 ≥ 0. The BTZ black hole is usually studied in the absence of
the gravitational Chern-Simons coupling, that is for kL = kR = k. Its entropy is
√ √

S = π(ℓ/2G)1/2 M ℓ − J + M ℓ + J . (This entropy was first expressed in terms of
two-dimensional conformal field theory in [13].) With ℓ/G = 16k as in (2.19), this is
√ √ p 
equivalent to S = 4π k L0 + L0 . For the holomorphic sector, the entropy is there-
fore
p
SL = 4π kL L 0 , (3.3)

and similarly for the antiholomorphic sector.

There is no classical BTZ black hole with L0 < 0, and the entropy of such a black
hole is zero if L0 = 0. We will take this as a suggestion that quantum states corresponding
to black holes exist only if L0 > 0, that is L0 ≥ 1. This means that the exact partition
function Z(q) should differ from the function Z0 (q) in (3.1) by terms of order q:
∞
Y
−k 1
Z(q) = q + O(q). (3.4)
n=2
1 − qn

Admittedly, we are here trying to squeeze more information from the classical result
than is justified. But it turns out that a modular-invariant partition function of this form
exists and is unique. This result (which is due to Höhn [31]) follows from the fact that
the moduli space M1 of Riemann surfaces of genus 1 is itself a Riemann surface of genus
0, in fact parametrized by the j-function. If E4 and E6 are the usual Eisenstein series of
weights 4 and 6, then j = 1728E43 /(E43 − E62 ). Its expansion in powers of q is

j(q) = q −1 + 744 + 196884q + 21493760q 2 + 864299970q 3 + 20245856256q 4 + . . . . (3.5)

Actually, it is more convenient to use the function

J(q) = j(q) − 744 = q −1 + 196884q + 21493760q 2 + . . . , (3.6)

which likewise parametrizes the moduli space.

The J-function has a pole at q = 0 and no other poles. The statement that J
parametrizes the moduli space means precisely that any modular-invariant function can
be written as a function of J. The partition function Z(q) has a pole at q = 0, that is
at J = ∞. This pole is of order k, and the partition function has no other poles. Any
holomorphic function of J that has no singularity except at J = ∞ is a polynomial in J.

30
In the present case, as the pole in Z(q) at q = 0 is of order k, Z must be a polynomial in
J of degree k. Thus
k
X
Z(q) = fr J r , (3.7)
r=0

with some coefficients fr . These k + 1 coefficients can be adjusted in a unique fashion to

ensure that Z(q) takes the form (3.4), or in other words to ensure that the terms in Z(q)
of order q −n , n = 0, . . . , k, coincide with the naive function Z0 (q).
When we do this, we get a function that we will call Zk (q), k = 1, 2, 3, . . .. This
function is our candidate for the generating function that counts the quantum states of
three-dimensional gravity in a spacetime asymptotic to AdS3 . For example, for k = 1 we
have simply Z1 (q) = J(q), and the next few examples are

Z2 (q) =J(q)2 − 393767

=q −2 + 1 + 42987520q + 40491909396q 2 + . . .
Z3 (q) =J(q)3 − 590651J(q) − 64481279
(3.8)
=q −3 + q −1 + 1 + 2593096794q + 12756091394048q 2 + . . .
Z4 (q) =J(q)4 − 787535J(q)2 − 8597555039J(q) − 644481279
=q −4 + q −2 + q −1 + 2 + 81026609428q + 1604671292452452276q 2 + . . . .

Following [31], we refer to a holomorphic CFT with c = 24k and partition function
Zk (q) as an extremal CFT. According to our proposals, the dual of three-dimensional
gravity should be an extremal CFT.
As was already noted in the introduction, Frenkel, Lepowsky, and Meurman con-
structed [24,25] an extremal CFT with k = 1, that is, a holomorphic CFT with c = 24 and
partition function J(q) = Z1 (q). They also conjectured its uniqueness. If that conjecture
as well as the ideas in the present paper are correct, then the FLM theory must be the
dual to quantum gravity for k = 1. Unfortunately, as also noted in the introduction, for
k > 1, extremal CFT’s are not known, though their possible existence has been discussed
in the literature [31-33] for reasons not related to three-dimensional gravity. Our reasoning
in this paper suggests that such theories should exist and be unique for each k.
The main point of the FLM construction was that their theory has as a group of
symmetries the Fischer-Griess monster group M, the largest of the sporadic finite groups.
Arguably, the FLM theory is the most natural known structure with M symmetry. The
coefficients in the q-expansion of the J-function are integers for number-theoretic reasons,

31
but the FLM construction gave a new perspective on why they are positive; this is a
property of the partition function of any CFT. It also gave a new perspective on why
these coefficients are so large; this follows from M symmetry, since M does not have small
representations.
Indeed, the original clue to the FLM construction was the observation by J. McKay
that the first non-trivial coefficient 196884 of the function J(q) is nearly equal to the
smallest dimension 196883 of a non-trivial representation of M. (This observation was
later greatly generalized [43,44].) The FLM interpretation is that 196884 is the number of
operators of dimension 2 in their theory. One of these operators is the stress tensor, while
the other 196883 are primary fields transforming in the smallest non-trivial representation
of M.
In our interpretation, the 196883 primaries are operators that (when combined with
suitable anti-holomorphic factors) create black holes. It is illuminating to compare the
number 196883 to the Bekenstein-Hawking formula. An exact quantum degeneracy of
196883 corresponds to an entropy of ln 196883 ∼
= 12.19. By contrast, the Bekenstein-
Hawking entropy at k = 1 and L0 = 1 is 4π ∼ = 12.57. We should not expect perfect
agreement, because the Bekenstein-Hawking formula is derived in a semiclassical approxi-
mation which is valid for large k.
Agreement improves rapidly if one increases k. For example, at k = 4, and again
taking L0 = 1, the exact quantum degeneracy of primary states is 81026609426, according
to eqn. (3.8). (Two of the states at this level are descendants.) This corresponds to an
entropy ln 81026609426 ∼
= 25.12, compared to the Bekenstein-Hawking entropy 8π ∼ = 25.13.
Shortly we will compute the entropy in the large k limit.
Our interpretation is that a primary state |Λi represents a black hole, while a de-
Q∞
scendant n=1 Ls−nn
|Λi describes a black hole embellished by boundary excitations. These
boundary excitations are the closest we can come in 2 + 1 dimensions to the gravitational
waves that a black hole can interact with in a larger number of dimensions. If this inter-
pretation is correct, then for an exact count of black hole states, we should count primaries
only. However, as the above examples illustrate, in practice this issue has only a very slight
effect on the black hole degeneracies. The boundary excitations contribute to the entropy
an amount that is independent of k and thus negligible in the regime where we compare
to the Bekenstein-Hawking entropy, and also, numerically, negligible even for small k. The
separation between the black hole and the boundary excitations is best-motivated for large
k.

32
Alternative Formula
Now we will present an alternative formula for the partition functions Zk (q). This
formula avoids the large coefficients involved in writing the Zk as a polynomial in J, and
will make some properties of the Zk more manifest.
Let q = exp(2πiτ ) with τ in the upper half plane. If p is a prime number, then the
Hecke operators acting on functions F (τ ) are defined essentially by
p−1
X
Tp′ F (τ ) = F (pτ ) + F ((τ + b)/p). (3.9)
b=0

(What we call Tp′ is p times the Hecke operator Tp as usually defined. See [45] for an
introduction to Hecke operators and [29,30] for their use in the present subject.) More
generally, for any positive integer t, we define
d−1
XX
Tt′ F (τ ) = F ((tτ + bd)/d2 ). (3.10)
d|t b=0

For t = 0, we define T0′ F (τ ) = 1. If F (τ ) is modular-invariant, then so is Tt′ F (τ ). The

definition of Hecke operators generalizes naturally to modular forms, but we will not need
this.
An immediate consequence of the definition is that, for any F (τ ) that is invariant
under τ → τ + 1, and so has a Laurent expansion in powers of q = exp(2πiτ ), the
coefficients in the q-expansion of Tn′ F (τ ) are linear combinations with positive integer
coefficients of the q-expansion coefficients of F (τ ). The FLM construction shows that the
q-expansion coefficients of the J-function are non-negative integers and the coefficients
of positive powers of q are dimensions of non-trivial (reducible) monster representations.
Hence the same is true of Tn′ J.
An important special case of how the Hecke operators act on q-expansion coefficients
is that if F (τ ) = q −1 + O(q), then

Tn′ F (τ ) = q −n + O(q). (3.11)

Now we can give an alternative description of the partition functions Zk (τ ). We recall

that these functions are uniquely determined by being modular-invariant and agreeing up
Q∞
to order q with the function Z0 = q −k n=2 (1 − q n )−1 . To find a function with these
properties, we simply expand
∞
X
Z0 = ar q r , (3.12)
r=−k

33
and then we let
k
X
Zk (τ ) = a−r Tr′ J(τ ). (3.13)
r=0

For example,

Z2 (τ ) = (T2′ + T0′ )J(τ ) = J(2τ ) + J(τ /2) + J((τ + 1)/2) + 1. (3.14)

This representation makes it clear that, just like the Tr′ J, the functions Zk share some
properties with the J-function: the q-expansion coefficients are non-negative, and the
coefficients of positive powers of q are dimensions of non-trivial monster representations.
This makes it tempting to speculate that the CFT’s dual to three-dimensional gravity
have M symmetry for all k. If so, the M symmetry is invisible in classical General Rela-
tivity and acts on the microstates of black holes. (However, the analogous conjecture for
supergravity, which would involve the Conway group, appears to be untrue, as we will see
in section 3.3.)
As an illustration of the usefulness of the expression for the partition function in terms
of Hecke operators, we will use it to compare with the Bekenstein-Hawking entropy. We
consider the semiclassical limit k, L0 → ∞, with r = L0 /k positive and fixed. We take
r = p/q to be a rational number and assume that k is divisible by q, so that L0 is an
integer n. If we write
∞
X
J(q) = cm q m , (3.15)
m=−1

then a formula of Petersson and Rademacher gives the asymptotic behavior

√ 3 1
ln cm ∼ 4π m − ln m − ln 2 + . . . . (3.16)
4 2
Let
∞
X
Zk (τ ) = bk,n q n . (3.17)
n=−k

We want to determine the behavior of bk,n for large k and n with r = n/k fixed. To evaluate
the partition function using the formula (3.13), let us first look at the contribution from
r = k, that is Tk′ J(τ ). In evaluating Tk′ J from the definition (3.10), the dominant term
(for large k and n) is the term with d = k. This contribution to bk,n is b0k,n ∼ kckn , and
hence, using the Petersson-Rademacher formula,
√ 1 3 1
ln b0k,n ∼ 4π kn + ln k − ln n − ln 2 + . . . . (3.18)
4 4 2
34
The first term is the Bekenstein-Hawking entropy (3.3). The additional terms in (3.18), as
well as the remaining contributions in (3.13) with r < k that we have omitted in deriving
(3.18), do not modify the Bekenstein-Hawking formula in the limit of large k, fixed n/k,
but give interesting subleading corrections that will be studied elsewhere [46].
Another method to get similar results (expressing black hole degeneracies in terms of
coefficients of the singular part of the partition function) is the Farey tail expansion [34].
See also [47].

The FLM Construction

Consider a holomorphic CFT with c = 24. If the number of primary fields of dimension
1 is s, then the expansion of the partition function near q = 0 begins Z(q) = q −1 +s+O(q),
and modular invariance implies that the genus 1 partition function is precisely Z(q) =
J(q) + s = q −1 + s + 196884q + . . .. Of the 196884 fields of dimension 2, one is a descendant
of the identity and s are descendants of dimension 1 primaries. Hence there are 196883 − s
primary fields of dimension 2, too few to furnish a non-trivial representation of the monster
group unless s = 0.
For this reason, Frenkel, Lepowsky, and Meurman [24] aimed to construct a holomor-
phic CFT of c = 24 with no primary fields of dimension 1, hoping that it would have
monster symmetry. Their construction was made as follows. The Leech lattice is an even,
unimodular lattice of rank 24 with no vector of length squared less than 4. One can con-
struct a holomorphic theory of c = 24 by compactifying 24 chiral bosons Xi , i = 1, . . . , 24
using any even unimodular lattice. If one uses the Leech lattice, then because it contains
no vector of length squared less then 4, a primary field of the form exp(ip · X) has di-
mension at least 2, as desired. However, this lattice theory has 24 primary fields ∂Xi of
dimension 1. Aiming to eliminate those fields, FLM considered an orbifold15 by the Z2
symmetry Xi → −Xi which eliminates the dimension 1 primaries. As it turns out, this
particular orbifold preserves modular invariance, does not add dimension 1 primaries in
twisted sectors, and gives the hoped-for theory with monster symmetry.
FLM conjectured that the model they constructed is the unique holomorphic CFT with
partition function J. If this conjecture is correct, it implies that many other constructions
give the same theory. There are [49] many orbifolds of the Leech lattice theory that give

15
Their work preceded the general study of orbifolds in string theory [48], though some related
stringy constructions, such as the Ramond and Neveu-Schwarz sectors of string theory, were
already known.

35
holomorphic CFT’s with no primary of dimension 1. According to the FLM uniqueness
conjecture, all of these theories are isomorphic to their monster theory. Moreover, the
Leech lattice theory itself can be obtained as an orbifold theory starting from another even
unimodular lattice of rank 24, so other lattices can be used as starting points as well.
Aiming to imitate the FLM construction for k = 2, one might start with an even
unimodular rank 48 lattice with no vector of length squared less than 6. Three such lattices
are known; two are described in [50] and a third has been constructed more recently [51].
(There is no reason to believe that these are the only three rank 48 lattices with this
property; there may be a vast number of them.) In the theory of 48 free chiral bosons Xi
compactified using such a lattice, a primary field exp(ip · X) has dimension at least 3, but
1
there are dimension 1 primaries ∂Xi and dimension 2 primaries ∂Xi ∂Xj − δ ∂X
48 ij
· ∂X.
A simple Z2 orbifold eliminates the dimension 1 primaries, as in the FLM case, but not
the dimension 2 primaries. One may attempt to find a more complicated orbifolding
construction to remove the unwanted primaries. Though all three lattices have interesting
discrete symmetry groups, it appears that there is no anomaly-free subgroup that removes
all dimension 2 primaries.16 Many other orbifolds can be considered, such as the symmetric
product of k copies of the k = 1 monster theory, but it appears difficult to remove all
primaries of low dimension.
Optimistically speaking, this situation might be compared to current algebra of a
simply-laced compact Lie group G. At level 1, one has the Frenkel-Kac-Segal construction
of current algebra of G via free bosons. At higher integer level, the theory still exists, but
generically has no equally straightforward realization in terms of free fields. Perhaps the
situation is somewhat similar for extremal CFT’s.

Further Remarks
We have emphasized here the partition function, because it is what we can determine
for general k. However, if one can actually describe the relevant CFT – as conjecturally we
can at k = 1 via the FLM construction – this gives much more than a partition function.
In this case, by computing matrix elements of primary fields, we get a detailed description
of the black hole quantum mechanics.
Above 2+1 dimensions, a black hole can form out of radiation and ordinary matter. In
pure gravity in 2 + 1 dimensions, the only thing that a black hole can form from is, roughly

16
I am grateful for assistance from G. Nebe in investigating this question.

36
speaking, smaller black holes. To be more precise, let Φi , i = 1, . . . , w be primary fields
of dimension not much greater than k that individually could, in acting on the vacuum,
create black holes of fairly small mass. A product of such fields acting on the vacuum at
prescribed points
w
Y
Φi (zi )|Ψi (3.19)
i=1

may create a black hole of large mass, that is, a primary state of large energy or a descen-
dant of such a state. Evaluating such matrix elements is the closest analog we can find,
in the present model, to describing the formation of a large mass black hole from matter.
Unfortunately, at the moment, we can perform such computations only for k = 1, since
for other values of k we do not know the CFT. Ideally, one would like to describe the CFT
for all k and investigate the semiclassical limit of large k.
If our hypotheses are correct, the following may be the most significant difference
between three-dimensional pure gravity and a more realistic theory of black holes in 3 + 1
dimensions. In the present model, as the holomorphic fields are all conserved currents,
and the energy levels are integers, the dynamics is integrable and periodic on a short time
scale. This is certainly not expected for black holes in general. An embedding of three-
dimensional gravity in a larger system, such as a string theory, would be expected to give
a non-integrable deformation of the model.

A Conundrum
We have here considered the holomorphic sector of a CFT, but the CFT dual of three-
dimensional gravity has both holomorphic and antiholomorphic degrees of freedom. Let
us therefore discuss a puzzle (stressed by J. Maldacena) that arises when one combines
e be the ground
the holomorphic and antiholomorphic degrees of freedom. We let |Ωi, |Ωi
e denote primary
states in the holomorphic and antiholomorphic sectors, and let |Φi and |Φi
e and its descendants correspond,
states other than the ground state. The state |Ωi ⊗ |Ωi
according to our picture, to Anti de Sitter space and its boundary excitations. A state
e or a descendant thereof, corresponds to a BTZ black hole, perhaps with boundary
|Φi⊗|Φi,
e or |Φi ⊗ |Ωi,
excitations. But what do we make of states |Ωi ⊗ |Φi e and their descendants?
Such states are trying to be Anti de Sitter space for holomorphic variables and black holes
for antiholomorphic variables, or vice-versa.
In classical three-dimensional gravity, there is no satisfactory solution that has this
interpretation. We can, however, see what is involved in trying to make one. First of all,

37
three-dimensional Anti de Sitter space AdS3 is simply the universal cover of the SL(2, R)
group manifold, with the symmetry SL(2, R) ×SL(2, R) (or rather a cover thereof) coming
from the left and right action of SL(2, R) on itself. The conformal boundary of the universal
cover of AdS3 is a cylinder R × S 1 . We can parametrize it by −∞ < τ < ∞, 0 ≤ σ ≤ 2π,
with conformal structure described by the metric ds2 = dτ 2 − dσ 2 . The BTZ black hole is
the quotient of (a cover of) SL(2, R) by a subgroup Z ⊂ SL(2, R) × SL(2, R) that is not
simply a subgroup of one factor or the other. This quotient has conformal boundary that
is again a cylinder, conformally equivalent to the boundary of Anti de Sitter space itself.
That is why the BTZ black hole can be regarded as an excitation of Anti de Sitter space.
To describe a state |Ωi ⊗ |Φi (or its parity conjugate) at the classical level, we want a
solution of Einstein’s equations that is invariant under precisely one of the two factors of
the SL(2, R) × SL(2, R) symmetry. This will reflect the fact that |Ωi is SL(2, R)-invariant
and |Φi is not. A space with just one of the SL(2, R) symmetries is AdS3 /Z, where Z is
a subgroup of one SL(2, R) factor. The quotient AdS3 /Z is a perfectly good solution of
Einstein’s equations in bulk, but its asymptotic behavior at infinity does not quite agree
with that of AdS3 . The boundary at infinity is a cylinder, but the compact direction
in the cylinder is lightlike rather than spacelike. For our approach to three-dimensional
gravity in the present paper to be correct, something like this quotient AdS3 /Z has to
make sense at the quantum level, perhaps because of a small quantum correction that
makes the compact direction in the cylinder effectively spacelike. Note that the metric
ds2ǫ = dτ 2 − ǫdσ 2 induces on the cylinder a conformal structure that is independent of ǫ
up to isomorphism as long as ǫ > 0.
What we have described is an analog for black holes of a statement made in section 2.4:
to ensure holomorphic factorization, the sum over topologies probably must be extended
to include configurations that are difficult to interpret classically. In the present case, the
analog of the sum over topologies is the sum over states with or without the presence of a
black hole.

3.2. The Supersymmetric Case

Now we consider the analog for supergravity. The analysis is a little more complicated
and in some ways the results are less satisfactory.
A rather similar problem has been treated by Höhn [31]. He considered not a super-
conformal field theory, but a more general holomorphic theory with bosonic operators of
integral dimension and fermionic operators of half-integral dimension. In that case, the

38
partition function can have a pole at the Ramond cusp, and is given by a Laurent series
in jθ = K − 24 rather than a polynomial.
We continue to assume holomorphic factorization, and we assume that the holomor-
phic part of the boundary CFT is an N = 1 superconformal field theory (SCFT). This
means (in whatever not well understood sense gravity is related to gauge theory) that the
gauge group of the bulk theory has a factor OSp(1|2). From a classical point of view,
the OSp(1|2) bundle over a three-manifold W endows W with a spin structure. In the
AdS/CFT correspondence, W has for conformal boundary a Riemann surface C, and the
spin structure on W determines one on C. One is instructed in the AdS/CFT correspon-
dence to specify C, including its spin structure, and sum over all choices of W . (Depending
on the theory, the antiholomorphic degrees of freedom may themselves be supersymmetric
and endowed with a choice of spin structure.) We will assume that it is physically appro-
priate to specify the spin structure on C rather than summing over it (just as one specifies
the complex structure of C), though at the end of section 3.3, we discuss what happens if
one sums over spin structures.

(a) (b) (c) (d)

τ τ τ τ
− + − +

0 − 1 0 − 1 0 + 1 0 + 1

Fig. 3:

We represent a genus 1 Riemann surface C as the quotient of the complex

plane by the lattice generated by complex numbers 1 and τ , where τ lies in the
upper half plane H. There are four possible spin structures, classified by whether
the fermions are periodic or antiperiodic around the “horizontal” and “vertical”
cycles. This is indicated by labeling the horizontal and vertical directions with a
+ sign for periodicity or a − sign for antiperiodicity.

We focus on the case that C has genus 1, in which case there are four possible spin
structures (fig. 3). The path integrals for these four spin structures can be interpreted in
terms of traces in two Hilbert spaces, known as the Neveu-Schwarz (NS) and Ramond (R)
Hilbert spaces HNS and HR , respectively.

39
 Traces in HNS are constructed using the spin structures, shown in (a) and (b) in
the figure, in which fermions are antiperiodic in the horizontal direction. If additionally
the fermions are antiperiodic in the vertical direction, as in (a), then the path integral
computes
F (τ ) = TrHNS q L0 , (3.20)

where q = exp(2πiτ ). We call this the NS partition function. If they are periodic in the
vertical direction, as in (b), the path integral computes instead

G(τ ) = TrHNS (−1)F q L0 , (3.21)

where (−1)F is the operator that is +1 on bosonic states, including the ground state of
the NS sector, and −1 on fermionic states.
Traces in HR are computed using spin structures, shown in (c) and (d) in the figure,
in which fermions are periodic in the horizontal direction. If the fermions are antiperiodic
in the vertical direction, the path integral computes an ordinary trace

H(τ ) = TrHR q L0 . (3.22)

If fermions are periodic in both directions, the path integral gives

χ = TrHR (−1)F q L0 = Tr (−1)F . (3.23)

Both H(τ ) and χ are severely constrained by the fact that in the R sector, one of the
superconformal generators, which we will G0 , commutes with L0 and obeys G02 = L0 . This
implies first of all that L0 ≥ 0 in the R sector, so H(τ ) has no pole at q = 0. Further,
the action of G0 implies that states of L0 > 0 are paired between bosons and fermions.
They cancel out of χ and make equal contributions to H(τ ). It follows that χ is simply
an integer, the trace of the operator (−1)F in the subspace with L0 = 0 (or alternatively
the index of the operator G0 mapping from bosonic states to fermionic ones). Moreover, if
we expand H(τ ) in powers of q
∞
X
H(τ ) = hn q n , (3.24)
n=0

then all coefficients hn are even except possibly h0 . In this expansion, only integral powers
of q appear, since all fields obey integral boundary conditions in the R sector.
In the NS sector, the ground state energy is −c/24. In the holomorphically factorized
case that we focus on, c = 12k ∗ for an integer k ∗ , so the ground state energy is −k ∗ /2.

40
Since fermions in the NS sector are antiperiodic in the spatial direction, NS excitations may
have either integer or half-integer energy above the ground state. The partition function
therefore takes the general form
∗ X
F (τ ) = q −k /2
(1 + aq 1/2 + bq + cq 3/2 + . . .) = fm q m . (3.25)
m∈Z/2, m≥−k∗ /2

In this expansion, the states in which m + k/2 is an integer are bosons, while the states
in which m + k/2 is half-integral are fermions. That is so because, in the NS sector,
bosonic fields are periodic and have integral excitation energies, while fermionic fields are
half-integral and have half-integral excitation energies. G(τ ) can therefore very explicitly
be expressed in terms of the same coefficients:
X ∗
G(τ ) = (−1)2m+k fm q m . (3.26)
m∈Z/2, m≥−k∗ /2

We have simply included a factor that is +1 on bosons and −1 on fermions. A more

succinct way to say the same thing is that
∗
G(τ ) = (−1)k F (τ + 1). (3.27)

Thus, G is simply determined in terms of F . The same is true for H. Consider the
modular transformation τ → −1/τ , which has the effect of exchanging the horizontal and
vertical directions in fig. 3 (with a reversal of orientation for one). This exchanges spin
structures (b) and (c), as a result of which

H(τ ) = G(−1/τ ). (3.28)

We can combine (3.27) and (3.28) to get

∗
H(τ ) = (−1)k F (−1/τ + 1). (3.29)

An important special case of this is

∗
lim H(τ ) = (−1)k F (1). (3.30)
τ →i∞

The limit on the left hand side exists, because H has no pole at q = 0. The limit is just
the coefficient h0 in (3.24):
∗
h0 = (−1)k F (1). (3.31)

41
h0 is the number of Ramond states of zero energy.
According to the above formulas, everything (except an integer χ, the supersymmetric
index) can be expressed in terms of F . So it is useful to understand the modular properties
of F . F is not invariant under τ → τ + 1, as we have already seen. But, since all energies
take values in Z/2, F is invariant under τ → τ + 2. In addition, since the modular
transformation τ → −1/τ maps spin structure (a) itself, F is invariant under τ → −1/τ .
The two transformations τ → τ + 2 and τ → −1/τ generate a subgroup of SL(2, Z) that
consists of 2 × 2 integral unimodular matrices
 
a b
(3.32)
c d

that are congruent mod 2 to one of

   
1 0 0 −1
, . (3.33)
0 1 1 0

These matrices of course act on τ in the usual fashion, τ → (aτ + b)/(cτ + d). This group
is sometimes called Γθ ; it is conjugate to the group Γ0 (2) characterized by requiring b to
be even. (A third conjugate group is characterized by the condition that c should be even.
Each of these three groups can be defined as the subgroup of SL(2, Z) that leaves fixed
one of the even spin structures, that is, one of the first three shown in fig. 3.)
The quotient of the upper half plane H by SL(2, Z) is a curve of genus zero with
one point missing. This can be seen by studying the fundamental domain (fig. 4(a)). The
missing point is called the “cusp” at τ = i∞. It is because the quotient has genus zero that
Riemann surfaces of genus 1 can be parametrized by a single holomorphic function, namely
J(τ ). As we have seen in section 3.1, this is useful because it means that the constraints of
modular invariance are equivalent to the statement that the partition function is a function
of J. Similarly, the quotient of the upper half plane H by Γθ is again a curve of genus
zero, this time with two missing points or cusps, as explained in fig. 4(b). This means
that there is a function K(τ ) such that any Γθ -invariant function F can be expressed as a
function of K.
The two cusps of H/Γθ are at τ → i∞ and τ = 1, and correspond respectively to the
NS sector and the R sector. For τ → i∞, the NS partition function F (τ ) has an expansion
in terms of L0 eigenvalues in the NS sector. Likewise, the Ramond partition function H(τ )
can be expanded for τ → i∞ in L0 eigenvalues in the Ramond sector; but according to
(3.29), the behavior of H(τ ) for τ near i∞ is the same as the behavior of F (τ ) for τ near 1.

42
 (a) (b)

1
− 21 2 −1 1

Fig. 4:

(a) A fundamental domain for the action of SL(2, Z) on the upper half plane.
By the action of τ → τ + 1, one can take |Re τ | ≤ 1/2, and by the action of
τ → −1/τ , one can take |τ | ≥ 1. The SL(2, Z) action identifies the left and right
hand halves of the boundary of the fundamental domain, making what topologically
is a Riemann surface of genus zero with one missing point at τ = i∞. This point
is called the cusp. (b) The analogous fundamental domain for the action of Γθ .
Here the symmetry τ → τ + 2 lets us reduce to |Re τ | ≤ 1, and τ → −1/τ still
lets us reduce to |τ | ≥ 1. The group action still identifies the left and right hand
halves of the boundary of the fundamental domain, and the result is a Riemann
surface of genus zero, now with two points omitted. The missing points are the
Neveu-Schwarz cusp at τ = i∞ and the Ramond cusp at τ = ±1 (those two points
are equivalent under τ → τ + 2). The Ramond cusp is missing in the quotient of
the upper half plane by Γθ because the points τ = ±1 are not in the upper half
plane, but on its boundary.

So in terms of the function F , the behavior at the two cusps is determined by the low-lying
spectrum in the NS and Ramond sectors, respectively.
We therefore call these the NS and R cusps. The natural uniformizing parameter at
the NS cusp is q 1/2 = exp(iπτ ), where τ is the argument of the function F in (3.25). The
right parameter is exp(iπτ ) since the symmetry is τ → τ + 2. At the R cusp, the natural
parameter is q = exp(2πiτ ), where now τ is the argument of H.
To get a clear picture of the nature of a holomorphic function F on H/Γθ , it is
necessary to describe the behavior near the cusps, or equivalently to describe what happens
when one tries to extend F to a function on the compactification of H/Γθ that is obtained
by adding the cusps. This compactification is a space Y ∼
= CP1 .

43
 We can explicitly describe a function K that parametrizes H/Γθ and has a pole only
at the NS cusp. This can be done in several ways. One formula is

∆(τ )2
K(τ ) = − 24, (3.34)
∆(2τ )∆(τ /2)
Q24
where ∆ = q n=1 (1 − q n )24 is the discriminant, a modular form of weight 24. In (3.34),
we have subtracted the constant 24 so that the expansion of K(τ ) in powers of q 1/2 has
no constant term:

K(τ ) = q −1/2 + 276q 1/2 + 2048q + 11202q 3/2 + 49152q 2 + 184024q 5/2 + 614400q 3
+ 1881471q 7/2 + 5373952q 4 + 14478180q 9/2 + . . . .
(3.35)
This is analogous to the definition of the J-function without a constant term. Another
formula for K is
∞ ∞
! ∞
q −1/2 Y Y Y
n−1/2 24 n−1/2 24
K= (1 + q ) + (1 − q ) + 2048q (1 + q n )24 . (3.36)
2 n=1 n=1 n=1

The product formula in (3.34), since it converges for all |q| < 1, shows that K is non-
singular as a function on H. As for the behavior at the cusps, either formula shows that
K has a simple pole at the NS cusp, that is, it behaves for q → 0 as q −1/2 . K is regular
at the Ramond cusp; in fact
K(τ = 1) = −24. (3.37)

This statement is equivalent to the statement that K + 24 = ∆(τ )2 /∆(2τ )∆(τ /2) vanishes
at τ = 1. Indeed, that function has a pole at the NS cusp, so it must have a zero somewhere.
Its representation as a convergent infinite product shows that it is nonzero for 0 < |q| < 1,
so the zero is at the Ramond cusp. We give another explanation of (3.37) in section 3.3 in
analyzing the k ∗ = 1 model. The fact that the holomorphic function K on Y has only one
pole, which is of first order, gives another way to prove that Y is of genus zero.
Now let us consider the Neveu-Schwarz partition function F of a holomorphic SCFT.
Any Γθ -invariant function F on H can be written as a function of K. The function F
arising in a holomorphic SCFT is actually polynomial in K. Indeed, since the definition
of F as Tr q L0 is convergent for 0 < |q| < 1, F is regular as a function on H. So the only
poles of F are at cusps, but the formula (3.31), which reflects the fact that the ground
state energy in the Ramond sector is zero, says that F has no pole at the Ramond cusp.

44
So the only pole of F is at the Neveu-Schwarz cusp, that is at K = ∞. Consequently, in
any holomorphic SCFT, the Neveu-Schwarz partition function F is a polynomial in K.
∗
The degree of this polynomial is precisely k ∗ , since F ∼ q −k /2
for q → 0. So
∗
k
X
F = fr K r . (3.38)
r=0

Thus, F depends on k ∗ + 1 coefficients.

Now, for the case of an SCFT that is dual to three-dimensional supergravity, what
would we expect F to be? To follow the same logic as in the bosonic case, we would argue
that since the minimum energy of a BTZ black hole is L0 = 0, and the entropy vanishes
at L0 = 0, primary states corresponding to black holes should be absent for L0 ≤ 0.
Equivalently, primary fields other than the identity should be absent for dimension less
than (k ∗ + 1)/2. Either statement would mean that up to terms of order q 1/2 , F would
coincide with the naive function

Y∞
∗ −k∗ /2 1 + q n−1/2
F0 (k ) = q (3.39)
n=2
1 − qn

that counts superconformal descendants of the identity.

For each positive integer k ∗ , there is a uniquely determined function Fk∗ that is a
polynomial in K and coincides with F0 (k ∗ ) up to order q 1/2 . This is a natural analog of
the partition function Zk that we defined for an extremal CFT without supersymmetry.
However, the same logic that makes us think that there should be no NS primaries of
L0 ≤ 0 also would make us believe that there should be no Ramond primaries of L0 = 0.
(There are automatically no Ramond states of L0 < 0, since L0 = G02 .) In either the NS
or Ramond sector, the classical black hole entropy vanishes for L0 ≤ 0.
The number of Ramond primaries of L0 = 0 in a theory with Neveu-Schwarz partition
∗
function Fk∗ is, according to (3.31), h0 = (−1)k Fk∗ (1). Let us call this number βk∗ . βk∗
is uniquely determined for each k ∗ . A practical way to determine it, using (3.37), is to
write ∗ ∗
k
X k
X
k∗ k∗ r k∗
βk∗ = (−1) Fk∗ (1) = (−1) fr K(1) = (−1) fr (−24)r . (3.40)
r=0 r=0

So one first evaluates the coefficients fr in (3.38) to ensure that there are no NS primary
states with L0 ≤ 0, and then one evaluates the sum in (3.40).

45
 The first ten values of βk∗ are given in Table 1. The first observation is that βk∗ never
vanishes in this range. Consequently, it is impossible to assume that there are no primary
operators other than the identity of dimension less than (k ∗ + 1)/2. If we assume that
there are no such primaries in the NS sector, then there are primaries of dimension k ∗ /2
in the Ramond sector.

k∗ βk∗
1 24
2 24
3 95
4 1
5 143
6 1
7 262
8 -213
9 453
10 -261

Table 1. Values of the index βk∗ for k ∗ = 1, . . . , 10.

The good news is that the numbers in the table are rather small compared to black
hole multiplicities that arise in the classically allowed region L0 > 0. So an optimistic view
is to interpret the numbers in the table as quantum corrections. For example, at k ∗ = 9,
where β takes the relatively large value 453, the multiplicity of the lowest mass classically
allowed black hole, namely L0 = 1/2 in the NS sector, turns out to be 135149371 if the
partition function F9 can be trusted. (As usual, we count only primaries, although this
involves only a small correction.) The lowest classically allowed black hole in the Ramond
sector, at L0 = 1, has multiplicity 381161020987.
However, it may not seem logical to assume that there are no NS primaries at dimen-
sion k ∗ /2 and allow Ramond primaries of that dimension. Moreover, it really does not
make sense to do this, since some values of β in the table are negative.
So we retreat from claiming that there are no NS primary fields of dimension less
than (k ∗ + 1)/2, and instead we consider the hypothesis that there are no such primaries

46
of dimension less than k ∗ /2. In this case, we are free to add an integer s to Fk∗ . If we do
so, the number of NS primaries of dimension k ∗ /2 becomes f0 = s, and the number h0 of
∗
Ramond primaries of that dimension changes by (−1)k s. The combination

∗
βk∗ = h0 − (−1)k f0 (3.41)

is unchanged, and so is still given by the numbers in the table.

Unfortunately, then, we have no way to determine the numbers f0 and h0 separately,
only the “index” β. We also have little insight about the more conventional index χ defined
in (3.23). Both of these quantities characterize the primary states of L0 = 0.
As in the bosonic case, one can construct useful formulas for the functions Fk∗ in
terms of Hecke operators acting on F1 . However, we will omit this.

Optimistic Conjecture
We define an extremal SCFT to be one with no primary fields other than the identity
of dimension less than k ∗ /2. Equivalently, the NS partition function is equal up to an
additive integer to the function Fk∗ that was defined above.
Generalizing what we have said about the bosonic case, the most optimistic conjecture
we can propose is that an extremal SCFT exists and is unique for every positive integer
k ∗ , and is dual to three-dimensional supergravity. The best evidence we can offer is that
extremal SCFT’s in this sense do exist for k ∗ = 1, 2 and there are results about uniqueness
[31,35] at least for k ∗ = 1.
In addition, as we will see, the functions Fk∗ with k ∗ = 3, 4 have interesting properties
suggestive of the existence of actual theories.
These matters are discussed in section 3.3. In addition, in the appendix, we describe
the functions Fk∗ for 5 ≤ k ∗ ≤ 10.

3.3. Extremal SCFT’s With Small k ∗

k ∗ =1
The first construction of an extremal SCFT was made by FLM at k ∗ = 1, that is
c = 12. They considered eight free bosons Xi compactified using the E8 root lattice,
combined with eight free fermions ψi to achieve superconformal symmetry. This theory
has NS primary fields of dimension 1/2, namely the ψi . To eliminate these fields, they
considered a Z2 orbifold, dividing by the operation that acts as −1 on all Xi and ψi .

47
They conjectured that this construction gave the unique SCFT that has c = 12 and no NS
primary of dimension 1/2.
While the construction is simple, it has the drawback of not making manifest the
global symmetry of the model. This was remedied much more recently by Duncan [35],
who described the same model in another way. In this construction, one begins with 24
free fermions λi , i = 1, . . . , 24, forming again a system with c = 12. In quantization,
one can require the λi to be either antiperiodic or periodic around the spatial direction,
giving what we will call NS0 and R0 sectors. (We reserve the name NS and R for another
construction that will appear shortly.)
With 24 fermions, the ground state energy in the NS0 sector is −1/2, while in the
R0 sector, the ground state energy is +1. The minimum dimension of a spin field is the
difference between these two energies, or 3/2.
The model has an O(24) symmetry that rotates the fields λi . The ground state in
the R0 sector is highly degenerate, because of zero modes of the fields λi . It consists of
212 = 4096 states, transforming in the spinor representation of O(24) (or rather its double
cover).
The spin fields of lowest dimension are therefore 4096 fields Wα of dimension 3/2
transforming as spinors. Dimension 3/2 is the correct dimension for a supercurrent, but
P α α
a generic linear combination W = α ǫ Wα , with c-number coefficients ǫ , does not
generate a superconformal algebra. Schematically, and without worrying about the precise
coefficients, the operator product W · W gives

ǫǫ ǫǫ T ǫΓij ǫ λi λj ǫΓijkl ǫ λi λj λk λl
W (x)W (0) ∼ + + + + regular. (3.42)
x3 x x2 x

(Here T is the stress tensor and Γi are the gamma matrices of O(24).) The condition for
W to generate a superconformal algebra is that the last two terms should be absent. This
is equivalent to
ǫΓij ǫ = ǫΓijkl ǫ = 0. (3.43)

It is shown in [35] that a spinor ǫ obeying these conditions exists and, if normalized to
have ǫǫ = 1, is unique up to an O(24) transformation. Another important fact is that any
such ǫ has one definite SO(24) chirality or the other.
Any choice of a solution of (3.43) turns the theory into an N = 1 superconformal
field theory, with W as the supercurrent. However, as ǫ is not O(24)-invariant, the theory
regarded as an N = 1 SCFT does not have O(24) symmetry. Rather, the choice of ǫ breaks

48
O(24) symmetry to a group that is known as the Conway group Co0 . It is the symmetry
group of the Leech lattice, and is a double cover of a sporadic finite group Co1 . Thus, as
a superconformal field theory, this model has symmetry Co0 , not O(24).
With this choice of superconformal algebra, we need to understand what are the
Neveu-Schwarz and Ramond vertex operators. A Ramond vertex operator O has a square
root singularity in the presence of the supercurrent W :

O′
W (x)O(x′ ) ∼ (3.44)
(x − x′ )n−1/2

for some integer n and some operator O′ . With W understood as a spin operator with
respect to the original fermions λ, those fermions have precisely this property. So they are
Ramond fields. This enables us to analyze all of the states in the original NS0 sector. The
ground state corresponds to the identity operator, whose OPE with W of course has no
branch cut, so it is an NS rather than Ramond operator. The operators in the NS0 sector
that have no branch cut with W are those that are products of an even number of λ’s and
their derivatives. The partition function that counts the corresponding states is
∞ ∞
!
q −1/2 Y Y
(1 + q n−1/2 )24 + (1 − q n−1/2 )24 . (3.45)
2 n=1 n=1

We recognize this as part of the formula (3.36) for the function K. In the R0 sector, of
the 4096 ground states, half have one chirality or fermion number and half have the other.
Any excitation at all can be combined with a ground state of properly chosen chirality to
get an operator that either does or does not have a branch cut with W , as desired. So for
each L0 eigenvalue in the R0 sector, precisely half the states contribute to the NS sector
and half to the R sector. The contribution of R0 states to the NS sector is therefore
∞
Y
2048q (1 + q n )24 . (3.46)
n=1

Adding up (3.45) and (3.46), we see that the total partition function F1 of the NS sector
in this model is precisely what we have called K:
∞ ∞
! ∞
q −1/2 Y Y Y
n−1/2 24 n−1/2 24
F1 = K = (1 + q ) + (1 − q ) + 2048q (1 + q n )24 . (3.47)
2 n=1 n=1 n=1

We can similarly compute the Ramond partition function H1 of this model. The
contribution of the NS0 sector is obtained from (3.45) by changing a sign so as to project

49
onto states of odd fermion number, rather than even fermion number. And the contribution
of the R0 sector is the same as (3.46). So
∞ ∞
! ∞
q −1/2 Y Y Y
H1 = (1 + q n−1/2 )24 − (1 − q n−1/2 )24 + 2048q (1 + q n )24
2 n=1 n=1 n=1 (3.48)
= 24 + 4096q + 98304q 2 + 1228800q 3 + 10747904q 4 + . . . .

Except for states of L0 = 0, the global supercharge G0 of the Ramond sector exchanges the
part of the Ramond sector coming from NS0 with the part coming from R0 . This implies
that we can alternatively write
∞
Y
H1 = 24 + 4096q (1 + q n )24 , (3.49)
n=1

where we have removed the NS0 contribution except for the ground states, and doubled
the R0 contribution. The equivalence of these formulas is not very obvious.
A consequence of the above formulas is that h0 , the number of Ramond states of
L0 = 0, is equal to 24. This gives another explanation for (3.37). Another consequence
of (3.49) and (3.47) is that for L0 > 0, the coefficients hn in the q-expansion of H1 are
precisely twice the corresponding coefficients fn in the expansion of F1 :

hn = 2fn . (3.50)

(Here hn is defined for integer n, while fn is defined for integer or half-integer n, so this
formula involves all coefficients hn but only half of the fn .) This is highly exceptional for
SCFT’s and reflects the particular way that this one was constructed. The relation (3.50)
is equivalent to

0 = F1 (τ ) + F1 (τ + 1) − H1 (τ ) = F1 (τ ) − G1 (τ + 1) − H1 (τ ). (3.51)

Although disguised in our way of presenting it, this is a standard relation in string theory.
In the FLM construction of the model via 8 free bosons and 8 free fermions, (3.51) amounts
to the statement of Gliozzi, Olive, and Scherk [52] that the Ramond-Neveu-Schwarz model
has equal partition function in the Ramond and Neveu-Schwarz sectors, because of space-
time supersymmetry.
The factor of 2 in (3.50) is consistent with asymptotic equality between the numbers
of NS and Ramond black holes because it only involves half of the coefficients of F1 . H1
has only half as many coefficients as F1 , but they are twice as big.

50
 As we have explained, the NS partition function Fk∗ of an extremal SCFT with any
k ∗ is a polynomial in K or equivalently in F1 ,

Fk∗ = f (F1 ), (3.52)

for some polynomial f . From (3.29), it follows that the Ramond partition function Hk∗ of
an extremal SCFT can be obtained from H1 using the same polynomial:
∗
Hk∗ = (−1)k f (−H1 ). (3.53)

A Model With k ∗ = 2
What in our language is an extremal SCFT of k ∗ = 2 was constructed [26] by Dixon,
Ginsparg, and Harvey (DGH), relatively soon after the work of FLM, by twisting the
FLM monster construction in a somewhat similar fashion. Note that an extremal SCFT
of k ∗ = 2 has c = 24, like the FLM monster theory.
We recall that the starting point of the FLM construction is 24 free bosons Xi that
are compactified via the Leech lattice. Then one performs a Z2 orbifold, dividing by the
symmetry Xi → −Xi . To construct an orbifold, one first constructs untwisted and twisted
sectors (by quantizing fields Xi (σ) that are assumed to be periodic or antiperiodic functions
of σ) and then one projects both sectors onto their Z2 -invariant subspaces.
DGH observed that the ground state energy in the untwisted sector is −1, while the
ground state energy in the twisted sector is +1/2. So a twist operator of lowest energy
has dimension 3/2, the right dimension for a supercurrent. They went on to show that it
is possible to pick a twist operator S that generates a superconformal algebra.
In the FLM construction, the field S is not present, since it is odd rather than even
under Z2 , and is projected out when one forms the orbifold theory. However, DGH showed
that it is possible to obtain an N = 1 SCFT by modifying the usual orbifold projection.
They defined the NS sector to consist of fields that do not have a cut in their OPE with
S, while the Ramond sector consists of fields that do have such a cut. Concretely, the NS
sector consists of the Z2 -even part of the untwisted sector of the orbifold plus the part of
the twisted sector that transforms under Z2 as does S (so that S, in particular, is an NS
field), while the Ramond sector consists of the Z2 -odd part of the untwisted sector plus
the Z2 -even part of the twisted sector.
From this information, one can of course compute the NS and Ramond partition
functions F2 and H2 . From our point of view, of course, F2 is a polynomial in F1 = K,

51
chosen so that F2 = q −1 + O(q 1/2 ). This gives F2 = K 2 − 552, and similarly, in view of
(3.53), H2 = H12 − 552. From this, we get

F2 = q −1 + 4096q 1/2 + 98580q + 1228800q 3/2 + 10745856q 2 + 74244096q 5/2 + 432155586q 3 + . . .

H2 = 24 + 196608q + 21495808q 2 + 864288768q 3 + . . . .
(3.54)
The number of states of L0 = 0 is 24 in the Ramond sector and 0 in the NS sector.
The global symmetry of the DGH model is closely related to the symmetry group Co0
of the Leech lattice (it is an extension of this by a finite abelian group that involves the
momenta of the Leech lattice). Since the k ∗ = 1 model also has Co0 symmetry, one might
optimistically conjecture that this group is relevant to supergravity at all k ∗ . However,
results below about k ∗ = 3 indicate that this is not the case.

Partition Function With k ∗ = 3

Unfortunately, this is the last case for which a suitable SCFT is known, but we can of
course determine the appropriate NS and Ramond partition functions for all k ∗ , at least
modulo an additive integer. We consider k ∗ = 3, 4 here because they turn out to have
unusual properties, and relegate the further cases k ∗ = 5, 6, . . . , 10 to an appendix.
For k ∗ = 3, we have
F3 = K 3 − 828K − 6143. (3.55)

This leads to

F3 = q −3/2 + 1 + 33606q 1/2 + 1843200q + 43434816q 3/2 + 648216576q 2

+ 7171304841q 5/2 + 63903727616q 3 + . . . (3.56)
H3 = 95 + 3686400q + 1296433152q 2 + 127807455232q 3 + . . . .

Here one finds an identity just like (3.50): the expansion coefficients hn of H3 are related
for n > 0 to analogous coefficients fn of F3 by

hn = 2fn . (3.57)

This expresses the fact that

F13 (τ ) + F13 (τ + 1) = H13 (τ ) − 1536, (3.58)

along with the similar linear relation (3.51).

52
 It seems that the identity hn = 2fn , n > 0, does not hold for any values of k ∗
except 1 and 3. In general, the differences hn − 2fn are relatively small, reflecting the fact
that asymptotically an NS or Ramond black hole has the same entropy, but they are not
zero. The fact that these degeneracies are actually equal for k ∗ = 3 may be a clue to the
construction of this model.
One important point about the k ∗ = 3 model is that it seems very unlikely to have
Co0 or Co1 symmetry. Indeed the leading coefficient 95 of H3 is not in any economical
way the dimension of a representation of Co0 . (The dimensions of the first few irreducible
representations of this group are 1, 24, 276, 299, and 1771; for Co1 , one must omit the
number 24 from this list.) Adding an integer n to F3 and therefore subtracting n from
H3 , does not help; we cannot pick n so that n and 95 − n are both dimensions of Co0
representations in an economical fashion. So it appears that, despite the tempting evidence
from the cases k ∗ = 1, 2, the group Co0 is probably not a general symmetry of three-
dimensional supergravity.

Monster Symmetry For k ∗ = 4?

However, in the next case, k ∗ = 4, there may well be a much larger discrete symmetry
group, namely the monster group M.
For k ∗ = 4, we have

F4 = K 4 − 1104K 2 − 8191K + 107545. (3.59)

This leads to the NS and Ramond partition functions

F4 = q −2 + q −1/2 + 1 + 196884q 1/2 + 21493760q + 864299970q 3/2 + 20246053140q 2

+ 333202640600q 5/2 + 4252023300096q 3 + 44656994071935q 7/2 + 401490908149760q 4 + . . .
H4 = 1 + 42987520q + 40491712512q 2 + 8504046600192q 3 + 802981773312000q 4 + . . . .
(3.60)
The first non-trivial coefficient in F4 is the famous number 196884 = 196883 + 1 whose
appearance in the J function gave the original hint of a connection between M and modular
functions and ultimately conformal field theory. Moreover, the leading coefficient in H4
is 1, which is compatible with any assumption about an automorphism group, since it is
the dimension of an irreducible representation, namely the trivial one. The contrasts with
what we found for k ∗ = 3, where the leading coefficient, namely 95, makes it difficult to
postulate a large discrete symmetry.

53
 To explore further the hypothesis of M symmetry, we attempt to express the coef-
ficients in F4 and H4 in terms of dimensions of representations of M. The dimensions
d1 , d2 , . . . , d12 of the first 12 monster representations, which we call Ri , i = 1, . . . , 12, are
given in the table.

d1 1
d2 196883
d3 21296876
d4 842609326
d5 18538750076
d6 19360062527
d7 293553734298
d8 3879214937598
d9 36173193327999
d10 125510727015275
d11 190292345709543
d12 222879856734249

Table 2. Presented here from [53] are the dimensions di of the ith
irreducible representation of the monster group M, for i = 1, . . . , 12.
We denote as Ri the representation of dimension di .

A little experimentation soon shows that the nontrivial coefficients in the NS partition
function F4 can be nicely written in terms of the di . The first 8 nontrivial coefficients
f1/2 , f1 , . . . , f4 are
196884 = d1 + d2
21493760 = d1 + d2 + d3
864299970 = 2d1 + 2d2 + d3 + d4
20246053140 = 3d1 + 4d2 + 2d3 + d4 + d6
333202640600 = 4d1 + 5d2 + 3d3 + 2d4 + d5 + d6 + d7
4252023300096 = 5d1 + 7d2 + 4d3 + 4d4 + 2d5 + 2d6 + d7 + d8
44656994071935 = 7d1 + 11d2 + 7d3 + 6d4 + 3d5 + 4d6 + 2d7 + 2d8 + d9
401490908149760 = 10d1 + 16d2 + 12d3 + 9d4 + 5d5 + 7d6 + 4d7 + 4d8 + d9 + d10 + d12 .
(3.61)

54
The coefficients on the right hand side are rather small given the numbers involved, and as
we will explain shortly, they do not grow much faster than is required by superconformal
symmetry. The first three numbers here equal the first three non-trivial coefficients of the
J function, but afterwards the two series diverge.

The coefficients hn in the q-expansion of H4 can similarly be expressed in terms of

dimensions of monster representations. In doing so, however, the following is a useful
shortcut and also possibly a clue to constructing the model. First of all, in contrast to the
case of k ∗ = 1, 3, it is not true that hn − 2fn = 0 for k ∗ = 4 and all n > 0. But some of
these differences vanish, and they are all relatively small. As we have already expressed
the fn for small n in terms of monster dimensions in (3.61), it suffices now to find similar
expressions for hn − 2fn . Actually, as the hn are all even for n > 0, we prefer to divide by
2, and we find
h1
− f1 = 0
2
h2
− f2 = −f1/2
2
h3
− f3 = 0 (3.62)
2
h4
− f4 = −f1
2
h5
− f5 = 0.
2
When this is combined with (3.61), we see that the hi /2, i ≤ 5 are linear combinations
of the di with small positive integer coefficients. So the initial coefficients of the Ramond
partition function are compatible with M symmetry, and in addition many of the differences
hn /2 − fn vanish.

T. Gannon has given a simple explanation of these results by showing that one can
express F4 and H4 in terms of the J-function via F4 (τ ) = J(2τ ) + J(τ /2) + 1, H4 (τ ) =
J(τ /2) + J((τ + 1)/2) + 1. Hence F4 and H4 inherit a relation to the monster from J. The
fact that the formulas have such a direct explanation may lessen the case for a new SCFT
with monster symmetry.

Given that the above formulas exist, they are not quite unique, because there are

55
some linear relations17 among the di with small coefficients:

d1 + d4 + d5 = d3 + d6
(3.63)
d3 + d8 + d12 = d2 + d7 + d9 + d11 .

We have made some choices to ensure that the above formulas are compatible with super-
conformal symmetry, in the following sense. The first formula in (3.61) indicates that at
L0 = 1/2, the NS sector has 196883 primary states |ρi i transforming in the representation
R2 . It follows that at L0 = 1, there are descendants G−1/2 |ρi i transforming in the same
representation. This continues at higher levels; the number of copies of R2 appearing as
descendants of the primary states |ρi i at L0 = s/2 is 1, 1, 1, 2, 3, 4, 5, 7 for s = 1, 2, . . . , 8.
If we count only primary states, we find, assuming that (3.61) is the right decomposition,
that primary states in the representation R2 occur at L0 = 1/2, 3/2, 2, 7/2, 4, each time
with multiplicity 1. Similar remarks apply to other representations. A special case is that,
for the range of L0 considered in (3.61), most of the states that transform in the trivial
representation R1 are actually descendants of the identity. If (3.61) is the right decom-
position, then the first primary field that is M-invariant is at L0 = 5/2. If we were to
rewrite (3.61) in terms of primary states only, the coefficients would be even smaller (but
still nonnegative), strengthening the case for monster symmetry.
In this particular example, we cannot add an integer to F4 without spoiling the hy-
pothesis of monster symmetry. If there were an NS primary with L0 = 2, it would have
a descendant at L0 = 5/2, and the number of primaries at L0 = 5/2 would be less than
196883.
In the appendix, we present somewhat similar though less extensive evidence for baby
monster symmetry at k ∗ = 6.

Sum Over Spin Structures

17
These relations can be explained as follows. The symmetric part of R2 ⊗ R2 decomposes as
2∼ R1 ⊕ R2 ⊕ R4 ⊕ R5 , and the antisymmetric part decomposes as ∧2 R2 ∼
Sym R2 = = R3 ⊕ R6 . On
the other hand, dim(Sym2 R2 ) − dim(∧2 R2 ) = dim R2 = d2 . Taken together, these facts imply
the first relation in (3.63), and the second follows similarly by considering the decomposition of
R3 ⊗ R3 .

56
 Given an SCFT with c an integer multiple18 of 24, by summing over spin structures,
one can make an ordinary bosonic CFT. The sum over spin structures projects both the
NS and Ramond sectors onto bosonic states of integer dimension, as is appropriate for a
bosonic CFT.
For c to be a multiple of 24, k ∗ must be even. It was already shown by DGH that the
sum over spin structures, applied to their model which in our language is k ∗ = 2, gives
the FLM model at k = 1. A natural question (raised by J. Duncan) is what this operation
does at higher even values of k ∗ .
The sum over spin structures removes fermionic operators such as the supercurrent
S, of dimension 3/2. However, it leaves operators such as S∂S that are even in S. This
operator has dimension 4, and is a Virasoro primary (though, of course, it is a descendant
in the N = 1 super-Virasoro algebra). A bosonic extremal CFT of k ≥ 4 should not have
a primary of dimension 4. Hence, the “sum over spin structures” operation applied to
an extremal SCFT of k ∗ ≥ 8 does not give an extremal CFT. However, it does give an
extremal CFT if applied to a theory of k ∗ = 4 or 6.
Hence, if extremal SCFT’s exist with k ∗ = 4, 6, they can be used to generate extremal
CFT’s of k = 2, 3. It is interesting that k ∗ = 4, 6 are precisely the values at which extremal
SCFT’s may have monster or baby monster symmetry. (The baby monster group is the
centralizer of an involution in the monster, and hence it is conceivable for an SCFT at
k ∗ = 6 to have baby monster symmetry while the CFT at k = 3 has monster symmetry.
The framework for this is described in [26].)
Part of what we have said can be understood in another way. Supergravity and
gravity are different in the semiclassical limit, and one cannot be obtained from the other
by merely summing over spin structures. So whatever is the SCFT dual of supergravity
(even if our assumptions are too optimistic and it is not extremal), the sum over spin
structures applied to this SCFT cannot give, for arbitrarily large k ∗ , the CFT that is dual
to gravity.

18
If c is an odd multiple of 12, the sum over spin structures is modular-invariant but projects
out the NS ground state and does not give a CFT in the usual sense. This is actually important
in superstring theory, where, with c = 12 in the light-cone treatment, the NS ground state is a
“tachyon” and is removed in the sum over spin structures.

57
4. k = 2 Partition Function On A Hyperelliptic Riemann Surface

The most important question raised by this paper is certainly whether appropriate
CFT’s and SCFT’s exist, beyond the few examples that are known.
In this section, we will perform a small computation that aims to give a hint that this
is the case, at least for the bosonic theory with k = 2 and hence c = 48. We will show that
the partition function of such a theory on a hyperelliptic Riemann surface of any genus can
be determined in a unique and consistent way. This includes, for example, any Riemann
surface of genus 2. The fact that we get a unique and consistent result in this situation
gives some encouragement for believing that the k = 2 model does exist and is unique.
In this paper, we merely demonstrate the consistency of an algorithm for determining
the partition function. Hopefully it will be possible in future work to get explicit formulas,
at least for genus 2. Our method also works for k = 1, though here the ability in principle to
determine the partition function comes as no surprise, since the model has been constructed
explicitly [24] and in fact the genus 2 partition function has been computed [54] by a quite
different method.
Perhaps it would help orient the reader to compare what we will do to another pos-
sible approach to determining the partition function. Genus 2 partition functions were
determined in [54] for a variety of holomorphic CFT’s with c = 24. The basic method was
to express the partition function in terms of a Siegel modular form, which depends on only
finitely many coefficients, and determine the coefficients by considering the behavior when
the Riemann surface C degenerates. There are two types of degeneration (C can break
up into two genus 1 curves joined at a point, or can reduce to a single genus 1 curve with
two points glued together). The partition function can be determined from the behavior
at just one degeneration, and there is a problem of consistency to show that one gets the
same result either way. One could attempt to demonstrate the consistency by studying
the appropriate Siegel modular forms, but we prefer to prove consistency by establishing
the associativity of a certain operator product algebra. One advantage is that this method
can be used to determine the partition function on a hyperelliptic Riemann surface of any
genus. On the other hand, if one could overcome the technicalities in genus 2 (see [55] for
one approach), one could possibly compute the genus 2 partition function of an extremal
CFT for all k.

58
4.1. Twist Fields

A hyperelliptic Riemann surface C is a double cover of the complex plane, for example
a double cover of the complex x-plane, which we will call C0 , described by an equation
2g+2
Y
2
y = (x − ei ). (4.1)
i=1

The 2 : 1 cover C → C0 is branched at the points e1 , . . . , e2g+2 . C is smooth if the ei are

distinct, and has genus g if the number of branch points is precisely 2g + 2.
Our approach to determining the partition function on such a Riemann surface is
based on an old idea [56-59]. The partition function of a conformal field theory W on the
hyperelliptic Riemann surface C can be determined by computing, in a doubled theory, the
correlation function of 2g + 2 copies of a “twist field” E, inserted at the points e1 , . . . , e2g−2
in C0 .
Consider any CFT W of central charge c. Away from branch points, the theory W
on the double cover C looks locally like the theory W × W on C0 . Here we have one copy
of W for each of the two branches of C → C0 .
In going around a branch point, the two copies of W are exchanged. So a more
complete description is to say that the theory on C0 is Sym2 W, the symmetric product
of two copies of W. The symmetric product theory is an orbifold of the product theory
W × W in which one divides by the Z2 symmetry that exchanges the two branches.
Let us describe what kind of operators the symmetric product theory has. First,
there are “untwisted” operators. These are simply operators of the theory W × W that are
invariant under the exchange of the two factors. For example, let T + and T − be the stress
tensors of the two factors. Any local operator constructed as a polynomial in T + and T −
and their derivatives and invariant under the exchange T + ↔ T − gives an operator in the
symmetric product theory. Operators of this particular type generate a chiral algebra that
we will call Sym2 V, where V denotes the chiral algebra generated by a single stress tensor.
In addition, there are “twisted sector” operators. These operators correspond (in
the operator/state correspondence of CFT) to states obtained by quantizing the theory
W ×W on a circle S 1 , in such a way that the two copies of W are exchanged in monodromy
around the circle. These states are the states of a single copy of W on a circle of twice
the circumference. The energies of the states in the twisted sector of the Sym2 W theory
are therefore precisely one-half the energies of the original W theory (in conformal field
theory, doubling the circumference of the circle divides the Hamiltonian by two).

59
 Operators related to states in the twisted sector are called twist fields. The dimension
of a twist field is the same as the difference in energy between the corresponding twisted
sector state and the untwisted ground state. The ground state energy in the untwisted
sector is −2 · c/24 (where the 2 comes from the two copies of W) and the ground state
energy in the twisted sector is −(1/2) · c/24 (where the factor of 1/2 was explained in the
last paragraph). The difference is dE = (c/24)(−1/2 − (−2)) = (3/2)c/24, and this is the
dimension of the twist field E of lowest energy. For our application, we are interested in
the case that c/24 is an integer k, and hence dE = 3k/2.
Now we can explain how the partition function of the theory W on the hyperelliptic
Riemann surface C defined in eqn. (4.1) can be interpreted as a genus zero correlation
function in the theory Sym2 W. From the standpoint of the Sym2 W theory on the x-plane,
the role of the branch points is just to exchange the two copies of W, via a twist field. On
the double cover, there is no operator insertion at the branch points except the identity;
the identity operator at a branch point corresponds in the downstairs description to the
ground state in the twisted sector and hence to the twist operator E of lowest dimension.
So the partition function on the double cover can be expressed in terms of the correlation
function E(e1 )E(e2 ) . . . E(e2g+2 ) on the x-plane. (The precise statement involves the
conformal anomaly; this is deferred to section 4.5.)
To compute this correlation function, we need to understand the chiral algebra that
is obtained by extending the symmetric product Sym2 V of two Virasoro algebras by the
field E. For this, we need to know what primary fields (of Sym2 V) appear in the operator
product E · E. To calculate the product E · E, we need to look at the behavior when a pair
of branch points approach each other. This situation is described locally by an equation

y 2 = (x − e)(x − e′ ), (4.2)

and we are interested in the behavior for e → e′ . In that limit, the equation becomes
y 2 = (x − e)2 , and the surface breaks up into two branches C± : y = ±(x − e). Each
branch is a copy of the complex plane. We can express the product E(e) · E(e′ ) for e → e′
as a sum of operators of the form U+ ⊗ U− , where U± is an operator in the original CFT
W on the branch C± . Moreover, the U± are descendants of some Virasoro primary fields
O± .

60
 (a) (b) O
C+

C− O

Fig. 5:

(a) The equation y2 = (x − e)(x − e′ ) describes two branches connected by

a tube. The tube collapses to a point as e → e′ . (b) Topologically, the tube
is equivalent to a cylinder. Propagation along the tube multiplies any energy
eigenstate by a c-number factor. If a primary field O flows in from the past, the
same field flows out in the future. Alternatively, if we think of all fields as flowing
out, then the two fields emerging at the two ends of the cylinder are conjugate.

The decomposition of E(e)·E(e′ ) as a sum of descendants of operators O+ ⊗O− can be

made in such a way that the operator O+ is conjugate to O− , meaning that the two-point
function hO+ O− i (in the original theory W) is nonzero in genus zero. To see this, go back
to the smooth double cover with e 6= e′ . For e near e′ , the two branches are smoothly
connected by a tube (fig. 5a). The surface thus has the topology of a cylinder (fig. 5b).
Propagation along this cylinder is the “identity operator” on primary fields. So if a given
primary state flows in at the bottom, it or flows out at the top. When we write E(e) · E(e′ )
as a sum of descendants of operators O+ ⊗ O− , the convention is that both O+ and O−
are outgoing, so they are conjugate, rather than being equal. At any rate, the important
conclusion for us is that O+ and O− have the same dimension.
According to our interpretation of three-dimensional gravity, it is described by a con-
formal field theory with c = 24k and no primary field other than the identity of dimension
less than k + 1. Hence, a primary operator O+ ⊗ O− of the symmetric product the-
ory, where O+ and O− have the same dimension, is either the identity operator or has
dimension at least 2(k + 1).

61
 To understand the chiral algebra or superalgebra19 generated by E together with
Sym2 V, we need to know all primary fields that appear with singular coefficients in the
operator product E(e) · E(e′ ). Since E has dimension 3k/2, an operator that will contribute
a singularity must have dimension less than 3k and hence at most 3k − 1. On the other
hand, we have just seen that all operators appearing in the OPE are either descendants of
the identity or have dimension at least 2(k + 1).
For k = 1 or 2, we have 3k − 1 < 2(k + 1). This leads to a drastic simplification, which
is the reason that we will restrict ourselves here to k = 1, 2: singularities in the product
E(e) · E(e′ ) come only from the identity operator and its Sym2 V descendants. Therefore,
to understand correlation functions of E, we need only understand the chiral algebra that
is obtained by adding to Sym2 V the primary field E of dimension 3k/2 with the OPE
schematically
E · E ∼ 1 + descendants. (4.3)

This is a chiral algebra that can be described in closed form and explicitly proved to
obey the Jacobi identity. The only Sym2 V primaries are 1 and E. The only genus zero three
point functions of primaries are h1 · 1 · 1i, which is trivial, and h1 · E · Ei, which is almost
trivial in the sense that it reduces to a two-point function. To understand correlation
functions of descendants, it suffices to understand the module for Sym2 V consisting of E
and its descendants. One way to construct this module explicitly is to observe that on the
double cover of the x-plane, E simply corresponds to the identity operator and Sym2 V to
the ordinary Virasoro algebra; so this module for Sym2 V can be deduced from the identity
module for V.
To get beyond the three-point function, the key step is to show that the chiral algebra
obtained by adding E to Sym2 V is consistent, that is, that the Jacobi identity is obeyed.
If so, this chiral algebra determines all genus zero correlation functions of E, and hence
determines the partition function of our CFT on a general hyperelliptic Riemann surface.
In such a problem of trying to extend a known chiral algebra Sym2 V by additional pri-
mary fields E (of integer or half-integer dimension, in which case one will get an extended
chiral algebra or a superalgebra), the Jacobi identity is equivalent [60] to the statement
that there exists a four-point function of the additional primary fields that has the appro-
priate singularities determined by the operator product expansion in all channels. (Such a

19
As E has dimension 3k/2, it is a fermionic operator if k is odd.

62
function automatically has the right symmetries.) In the case at hand, since the only pri-
mary field that we are trying to add is E, the only non-trivial primary four-point function
that we have to consider is E(e1 )E(e2 )E(e3 )E(e4 ) . In section 4.3, we explicitly construct
this four-point function and show that it behaves correctly in all channels.
In the following sense, this result is not at all surprising. The four-point function
E(e1 )E(e2 )E(e3 )E(e4 ) is essentially the partition function of the underlying CFT on the
hyperelliptic Riemann surface y 2 = (x − e1 )(x − e2 )(x − e3 )(x − e4 ). That surface has genus
1, and we already know from section 3 that for every k there is a unique, natural genus
1 partition function with the properties we want. In section 4.4, we show explicitly, for
k = 1, 2, that the four-point function E(e1 )E(e2 )E(e3 )E(e4 ) determined from the OPE’s
agrees with the genus 1 partition function as described in section 3.

4.2. The E · E Operator Product

As a first step in that direction, we will calculate the details of the E · E operator
product. For this, we start with a double cover C of the x-plane branched at e and
e′ = −e, and so described by an equation y 2 = x2 − e2 . If u = x + y, v = x − y,
then the equation is uv = e2 . The two branches C+ and C− correspond respectively to
u → ∞, v → 0 and u → 0, v → ∞.
The path integral over C gives a quantum state Ψ in the theory W × W, that is, one
copy of W for each branch. This state is invariant under exchange of the two branches by
the symmetry y → −y, so it is really a state in the symmetric product theory Sym2 W.
We are really only interested in the part of Ψ proportional to the vacuum state and its
descendants. As we have seen, this part suffices to describe the desired chiral algebra if
k ≤ 2.
We will determine Ψ by using the fact that certain elements in the product V ×V of two
Virasoro algebras annihilate Ψ. This is so because there are globally-defined holomorphic
vector fields on C, of the form Vn = 2−n un+1 d/du = −2−n (e2 /v)n vd/dv. Let S be a
contour on the surface C that wraps once around the “hole.” If T is the stress tensor, the
R
contour integral S Vn T can be regarded, for any n, as an operator acting on the state Ψ.
This operator is invariant under deformation of the contour. It can be deformed (fig. 6)
to a contour S+ in the upper branch C+ or a contour S− in the lower branch C− . So we
have for all n Z Z !
Vn T − Vn T Ψ = 0. (4.4)
S+ S−

63
 S+ C+

C−
S−

Fig. 6:

A contour S that wraps once around the hole can be deformed to the contour
S+ in the upper branch or to the contour S− in the lower branch.

We want to express the two contour integrals that appear here in terms of Virasoro
generators on the two branches. To do this, we simply express Vn as a vector field on
the branch C+ or C− , either of which we identify with the x-plane. For example, on the
p
branch C+ , we write explicitly y = 1 − e2 /x2 = 1 − e2 /2x2 − e4 /8x4 + O(e6 ). We have
carried the expansion far enough to determine (for k ≤ 2) all singular terms in the product
E(e) · E(−e). So
   
n+1 e2 e4 n2 + n − 4 6 d
Vn = x 1 − (n + 2) 2 + 4 + O(e ) . (4.5)
4x x 32 dx
R
This means, if we ignore the conformal anomaly for the moment, that S+
Vn T corresponds,
on the branch C+ , to the operator
 
n+2 2 + n2 + n − 4
Q+
n = L+
n − e Ln−2 + e4 L+ 6
n−4 + O(e ). (4.6)
4 32

+
R
As a check, one can verify that [Q+ +
n , Qm ] = (n − m)Qn+m . Similarly, S−
Vn T corresponds
on the branch C− to the operator
 n      
e2 −n + 2 n2 − n − 4
Q−
n = L−
−n −e 2
L−
−n−2 +e 4
L−
−n−4
6
+ O(e ) . (4.7)
4 4 32

64
 The state Ψ is determined for each value of e (up to multiplication by a complex
b n Ψ = 0, where Q
scalar) by the condition that Q b n = Q+ − Q− . However, because of the
n n
Virasoro anomaly some c-number terms must be added to the above formulas, reflecting
the conformal anomaly in the mapping from u to x. There is no such c-number contribution
b 0 , since it cancels between the two branches. The c-numbers in Q
to Q b n for other n can be
bn , Q
conveniently determined by requiring that [Q b m ] = (n − m)Q
b n+m . For our purposes,
the only formulas we need are
   
b + e2 + e4 − + e2 − e4 −
Q0 = L0 − L−2 − L−4 − L0 − L−2 − L−4 + . . .
2 8 2 8
 2 4
 2
 2

b + 3e + e + e − e −
Q1 = L1 − L − L − L−1 − L−3 + . . . (4.8)
4 −1 16 −3 4 4
4 4
Qb 2 = L+ − e2 L+ − 3ke2 + e L+ − e L− + . . . .
2 0
16 −2 16 −2
Terms of order e6 have been omitted. The constant in Q b 2 was obtained from [Q
b0, Q
b2] =
−2Qb2 .
By requiring that Qb m Ψ = 0 for m = 0, 1, 2, and that Ψ converges to the Fock vacuum
|Ωi for e → 0, we now find Ψ to be

e2 e4 +
Ψ(e) = 1 + (L+ + L −
) + (L + L− −4 )
4 −2 −2
32 −4
 (4.9)
e4 + − 2 e4 + −
+ (L−2 + L−2 ) + L L + . . . |Ωi.
32 192k −2 −2
We can immediately use this formula to determine the singular part of the E · E OPE.
We normalize E so that the most singular term is E(x/2)E(−x/2) ∼ 1/x3k . Then, setting
e = x/2, we get

1 x2 x4
E(x/2)E(−x/2) ∼ 3k 1+ T + 10 ∂ 2 T
x 16 2
4
 (4.10)
x x4 + − −3k+6
+ 9T ⋆T + T T + O(x ).
2 3k · 210
All we have done is to write, on the right hand side, the operator that corresponds to
the state Ψ(e)/x3k , for e = x/2. (The factor of 1/x3k must be supplied by hand, since in
defining Ψ(e), we just normalized it so that the coefficient of the Fock vacuum is 1.) Also,
T + and T − are the stress tensors on the branches C+ and C− , respectively. T = T + + T −
is the diagonal or total stress tensor, and similarly Ln = L+ −
n + Ln . In addition, we use the
fact that in the correspondence between operators and states, the states L±
−2 |Ωi correspond
to the operators T ± , while L± 1 2 ±
−4 |Ωi correspond to 2 ∂ T . Finally, we have written T ⋆ T
for the operator corresponding to the state L2−2 |Ωi. T ⋆ T (0) can be obtained as the term
of order x0 in the operator product T (x) · T (0).

65
4.3. Determining The Four-Point Function

Our goal is to use the explicit formula (4.10) to show that the chiral algebra obtained
by adjoining E to Sym2 V obeys the Jacobi identity. The Jacobi identity is equivalent to
the existence of a four-point function for primary fields with the right symmetry properties
and the right OPE singularities in all channels. It is enough to consider primary fields with
respect to the Sym2 V algebra, which we already know to exist. The only non-trivial case
is the four-point function of E.
For k = 1, the singular part of the E · E operator product only involves the diagonal
stress tensor T = T + + T − . This is a substantial simplification; the extended chiral super-
algebra obtained by incorporating E can be viewed an extension of an ordinary Virasoro
algebra V, rather than a symmetric product Sym2 V. In fact, this algebra is a familiar one,
an N = 1 super Virasoro algebra:
 
1 x2
E(x/2)E(−x/2) ∼ 3 1 + T + O(x). (4.11)
x 16

The stress tensor T has c = 48 (as it is the diagonal stress tensor in the symmetric product
of two copies of a c = 24 theory). A superconformal field theory with this central charge
can be constructed from 32 chiral multiplets consisting of bosonic and fermionic fields φi
and ψi , where hφi (x)φj (0)i = −δij ln(x), and hψi (x)ψj (0)i = δij /x, for i, j = 1, . . . , 32. In
terms of these fields, we can take

32
i X
E= √ ψk ∂φk . (4.12)
4 2 k=1

With this free field realization, all correlation functions of E can be explicitly computed.
For example, the four-point function is

4
Y  
3/2 1
E(e1 )E(e2 )E(e3 )E(e4 ) = (dei ) + cyclic
i=1
(e1 − e2 )3 (e3 − e4 )3
 (4.13)
3 1
+ Q ,
32 i<j (ei − ej )

where additional terms obtained by cyclic permutations of e2 , e3 , e4 are to be added in the

first term on the right. We will eventually compare this formula to the known genus 1
partition function of the monster theory of k = 1.

66
 In (4.13), we included a factor of (de)3/2 for each operator E(e) to reflect the fact
that E has dimension 3/2. Thus the correlation “function” E(e1 )E(e2 )E(e3 )E(e4 ) is most
naturally understood as a 3/2-differential in each variable. This factor is often omitted
in writing such formulas, and we will do so as an abbreviation except when the factor is
important.
For k = 2, the singular part of E · E cannot be expressed in terms of the diagonal
stress tensor T only, and there is no way to avoid using Sym2 V. Correspondingly, the
chiral algebra for k = 2 seems to be unfamiliar, and it is considerably harder to get the
analog of (4.13) for k = 2. Since we do not have an explicit realization of the algebra
analogous to (4.12), we cannot compute the four-point function directly, and instead we
will determine it by requiring the appropriate OPE singularities.
The function E(e1 )E(e2 )E(e3 )E(e4 ) , with e1 , . . . , e4 ∈ C0 ∼
= CP1 , must be symmetric
in all arguments and must have the appropriate OPE singularities in all channels. In
addition, it is highly constrained by invariance under the action of SL(2, C) on C0 . The
most general “function,” actually a cubic differential in each variable, that is symmetric,
SL(2, C)-invariant, and consistent with at least the leading OPE singularity, is

4
Y  
3 1
E(e1 )E(e2 )E(e3 )E(e4 ) = (dei ) + cyclic
i=1
(e1 − e2 )6 (e3 − e4 )6
 Y 
1 1
+A + cyclic (4.14)
(e1 − e2 )4 (e3 − e4 )4 i=1,2, j=3,4 ei − ej
!
1
+B Q 2
. .
i<j (ei − ej )

Cyclic permutations of e2 , e3 , e4 are to be added as shown, and the coefficients A and B

must be determined to ensure that the OPE singularities are correct.
After a detailed calculation, one finds that the right values are

3 12k + 1
A= , B= (4.15)
16 216

where of course k = 2. To get these formulas, one may for instance set e1 = −e2 = e,
and use (4.10) to determine the singular behavior for e → 0. In this way, one ex-
presses the singular behavior of the four-point function in terms of three-point functions
X (0)E(e3 )E(e4 ) , where X is one of the descendants of the identity that appear in (4.10).

67
 The coefficient A appears in the coefficient of 1/e4 for e → 0, and in determining it,
the only important choices of X are 1 and T . The three-point functions that we need are

1
1 · E(z)E(w) =
(z − w)6
(4.16)
3
T (x)E(z)E(w) = .
(z − x) (w − x)2 (z − w)4
2

The first is immediate from the way we have normalized E. The second, since the three
fields involved are quasi-primary fields (they transform as primaries under SL(2, C)) is
determined by SL(2, C) invariance up to a constant multiple. The constant can be de-
termined using the T · E OPE (which reflects the fact that E is a primary of dimension
3 = 3k/2) or the E · E OPE.
To determine B, we need three more three-point functions, namely
 
2 1 1 18 24
∂ T (0)E(z)E(w) = + +
z2 w2 z 2 w2 (z − w)4 z 3 w3 (z − w)4
 
5 5 4 3
T ⋆ T (0)E(z)E(w) = 2
+ 2− (4.17)
z w zw z w (z − w)4
2 2
 2 
9k 3k 1
T + T − (0)E(z)E(w) = + .
16 64 z 4 w4 (z − w)2

The first of these results is obtained simply by differentiating the second formula in (4.16).
To get the second, we use the fact that

dx T (x)
T ⋆ T (0) = Resx=0 T (0). (4.18)
x

Now consider the expression

D dx T (x) E
T (0)E(z)E(w) . (4.19)
x

We view this as a differential form on the complex x-plane, with z and w kept fixed. As
such, it has poles precisely at x = 0, z, w. The residue at x = 0 is the desired three-point
function T ⋆ T (0)E(z)E(w) . The residues at at x = z, w can be computed in terms of
the correlator T (0)E(z)E(w) by using the fact that E is a primary of dimension 3, which
determines the singular behavior of the product T ·E. The sum of the residues must vanish,
and this gives our result.

68
 Finally, to get the last formula in (4.17), we note first that since the operators T + T −
and E are quasiprimary fields of dimension 4 and 3, we have
θ
T + T − (x)E(z)E(w) = , (4.20)
(x − z) (x − w)4 (z − w)2
4

for some constant θ. We have

T + T − (x)E(0)E(w) θ
lim = 4. (4.21)
w→∞ E(0)E(w) x
This is supposed to be computed in the theory Sym2 W on the x-plane, with insertions
of E at 0, w. This is equivalent to a computation in the theory W on a double cover of
the x-plane branched at 0, w. In the limit w → ∞, the double cover is described by the
equation y 2 = x, and it is convenient to transfer the computation to the y-plane.
We will now write the stress tensor explicitly as Txx or Tyy depending on which local
parameter is used in defining it. Allowing for the conformal anomaly, the relation between
the stress tensor Txx on the x-plane and the corresponding stress tensor Tyy on the y-plane
is  2
∂y c
Tyy = Txx − {y, x}, (4.22)
∂x 12
where {y, x} denotes the Schwarzian derivative of these two functions. With y 2 = x, we
get {y, x} = 3/8x2 . In the case of interest, c = 24k so
1 3k
Txx = Tyy + 2 . (4.23)
4x 4x
This formula enables us to express T + in terms of Tyy |y=√x and T − in terms of Tyy |y=−√x .
The limit in (4.21) becomes
D 1 3k

1

3k E
√
Tyy |y= x + 2 √
Tyy |y=− x + 2 . (4.24)
4x 4x 4x 4x
(The idea is that passing to the y-plane eliminates the twist operators E from the formalism,
√
and replaces T ± (x) by a stress tensor evaluated at y = ± x.) Using the fact that in
conformal field theory on the y-plane, hTyy i = 0 and Tyy (y1 )Tyy (y2 ) = c/2(y1 − y2 )4
with c = 24k, the correlation function in (4.24) can be evaluated to give θ/x4 with θ =
9k 2 /16 + 3k/64. This gives the last formula in (4.17).
With the aid of the three-point functions in (4.16) and (4.17) and the operator product
formula (4.10), one can verify that the candidate four-point function (4.14) has the right
singularities precisely if the coefficients A and B are as claimed in (4.15).
Our derivation has in fact been somewhat redundant, since the formulas for three-
point functions can largely be deduced from the formula for the E · E operator product,
and vice-versa.

69
4.4. The Genus One Partition Function Revisited

We will now explicitly use our formulas for the four-point function of the twist field E
to recover the partition functions of a k = 1 or k = 2 extremal CFT on a Riemann surface
of genus 1.
First we must describe the necessary formalism more precisely. The partition function
ZC (e1 , e2 , . . . , e2g+2 ) of any conformal field theory W on the hyperelliptic Riemann surface
C defined by
2g+2
Y
2
y = (x − ei ) (4.25)
i=1

can indeed be expressed in terms of the genus zero twist field correlation function

E(e1 )E(e2 ) . . . E(e2g+2 ) . (4.26)

However, to do so requires an important detail that we have not yet described.

There are at first sight several strange features in relating ZC to the twist field corre-
lation function:
(1) The dimension of the twist field E seems to be wrong. The Riemann surface C
degenerates as ei → ej for any i and j. As the complex structure of C is invariant under
the exchange ei ↔ ej , a natural local parameter near the degeneration locus on the moduli
space is actually w = (ei − ej )2 . (For example, this statement can be made precise using
eqn. (4.30) below.) One expects ZC to behave for ei → ej as w−k , since the ground state
energy is −k. However, as E has dimension 3k/2, the behavior of the twist field correlation
function as ei → ej is actually (ei − ej )−3k ∼ w−3k/2 .
(2) This brings in relief the fact that if k is odd, E actually has half-integral dimension
and so is fermionic. For odd k, the twist field correlation function (4.26) actually changes
sign under exchange of ei and ej , while ZC is invariant under this exchange.
(3) The twist field correlation function is most naturally understood as a 3k/2-
differential in each variable. We have written the formulas (4.13) and (4.14) in such a
way as to emphasize this. It may not be immediately clear what this means for the parti-
tion function ZC .
The answer to all of these questions has to do with the proper treatment of the
conformal anomaly in the holomorphic context. We postpone a full explanation to section

70
4.5 and for now we merely write down the precise relation between the genus 1 partition
function and the twist field correlation function:
 Y 
8k k
ZC (e1 , e2 , . . . , e4 ) = 2 (ei − ej ) E(e1 )E(e2 ) . . . E(e4 ) . (4.27)
1≤i<j≤4

Q
The factor of 28k just reflects the way we have normalized E. The factor of i<j (ei − ej )k
resolves problems (1) and (2) above. In section 4.5, we will explain the meaning of that
factor and its relation to problem (3) (as well as the generalization to higher genus). For
now, we merely mention that in eqn. (4.27), the twist field correlation function should be
Q
understood as a function, without the factor of i (dei )3k/2 .
Now we will use the general relation (4.27) along with our previous formulas for the
twist field correlation functions to analyze the genus 1 partition function for k = 1, 2. We
make an SL(2, C) transformation to set e4 = ∞ and to fix

e1 + e2 + e3 = 0. (4.28)

The genus 1 partition function at k = 1 is then, using (4.13),

28 3 3


ZC (k = 1) = Q 2
(e1 − e3 ) (e2 − e3 ) + cyclic + 24. (4.29)
i<j (ei − ej )

We want to express this in terms of q = exp(2πiτ ), the usual parameter on the genus 1
moduli space, and to show explicitly that ZC has the expected behavior near q = 0. The
classical formulas are
e1 − e2 = θ34 (0, τ )
e3 − e2 = θ14 (0, τ ) (4.30)
e1 − e3 = θ24 (0, τ ),
where X 2
θ1 (0, τ ) = q (n+1/2) /2

n∈Z
X 2
θ2 (0, τ ) = (−1)n q n /2
(4.31)
n∈Z
X 2
θ3 (0, τ ) = qn /2
.
n∈Z

Expanding (4.29) near q = 0, one can readily verify the expected behavior ZC (k = 1) =
q −1 + O(q). Since ZC is manifestly modular-invariant, this implies that ZC must equal

71
J. This can also be verified directly by comparing (4.29) to the classical formula for the
j-function

3
5 (e1 − e2 )2 + (e2 − e3 )2 + (e3 − e1 )2
j(q) = 2 Q 2
(4.32)
i<j (ei − ej )

and recalling that J = j − 744.

The case k = 2 can be treated similarly. Our previous formula (4.14) for the twist
field four-point function implies, via (4.27), a formula for ZC (k = 2). If we take e4 → ∞
and e1 + e2 + e3 = 0, this formula becomes
 
(e1 − e3 )2 (e2 − e3 )2
16
ZC (k = 2) =2 + cyclic
(e1 − e2 )4
  (4.33)
12 (e1 − e3 )(e2 − e3 )
+3·2 + cyclic + (12k + 1).
(e1 − e2 )2

One can expand this near q = 0 and verify the expected behavior ZC (k = 2) = q −2 + 1 +
O(q). Together with modular invariance, this implies that ZC (k = 2) = J 2 − 393767, as
one can also verify directly using (4.32).

4.5. Role Of The Conformal Anomaly

Naively speaking, the partition “function” of a conformal field theory on a Riemann

surface C depends only on the complex or conformal structure of C.
Actually, because of the conformal anomaly, things are more subtle. The usual way to
describe the situation is to endow C with a metric h, and to think of the partition function
as a function Z(h) that transforms in a certain prescribed way (depending on the central
charge) under conformal rescaling h → eφ h of the metric.
This description is natural in differential geometry. However, it has a few limitations.
One is that it really only gives an adequate framework if the left and right central charges
cL and cR are equal. A second limitation, more crucial for our purposes, is that it is
not well-adapted to holomorphic factorization. In a holomorphic CFT, everything will be
much simpler if we can incorporate the conformal anomaly in purely holomorphic terms.
An alternative approach is to think of the partition function not as a function but as a
section of a suitable line bundle R over M, the moduli space of complex Riemann surfaces.
This approach is developed in detail in [61]. Holomorphic factorization is incorporated
naturally in this approach by saying that in a holomorphic CFT, R is a holomorphic line
bundle over M.

72
 More specifically, R ∼
= Lc/2 , where c is the central charge and L is a fundamental line
bundle over M that is known as the determinant line bundle. L is defined as follows. Let
p be a point in M corresponding to a Riemann surface C. Then Lp , the fiber of L at p, is
the top exterior power of the vector space H 0 (C, KC ) of holomorphic differentials on C.
So if C has genus g and ω1 , . . . , ωg are holomorphic differentials on C, then the expression
ω1 ∧ω2 ∧. . .∧ωg defines a vector in Lp . In our problem, c = 24k, and the partition function
is a section of L12k .
On a hyperelliptic Riemann surface

2g+2
Y
2
y = (x − ei ), (4.34)
i=1

we can make this completely explicit. A basis of the space of holomorphic differentials is
given by
dx x dx x2 dx xg−1 dx
, , ,..., . (4.35)
y y y y
A section of L12k over the space of hyperelliptic equations is hence an expression of the
form  12k
dx x dx x2 dx xg−1 dx
Θ= ∧ ∧ ∧...∧ F (e1 , . . . , e2g+2 ). (4.36)
y y y y
To get a section of L12k over the moduli space of hyperelliptic Riemann surfaces, Θ must
be invariant under the action of SL(2, C) on the space of hyperelliptic equations. This
action takes the form
aei + b
ei →
cei + d
ax + b
x→ (4.37)
cx + d
y
y→ Q2g+2 ,
(cx + d)g+1 i=1 (cei + d)1/2
and the condition for Θ to be invariant is that F transforms by

2g+2
Y
F (e1 , . . . , e2g+2 ) → F (e1 , . . . , e2g+2 ) (cei + d)−6kg . (4.38)
i=1

The twist field correlation function, on the other hand, is of the form

2g+2
Y
E(e1 )E(e2 ) . . . E(e2g+2 ) = (dei )3k/2 G(e1 , . . . , e2g+2 ) (4.39)
i=1

73
for some “function” G. SL(2, C) invariance of the correlation function means that G
transforms by
2g+2
Y
G(e1 , . . . , e2g+2 ) → G(e1 , . . . , e2g+2 ) (cei + d)3k . (4.40)
i=1

Comparing (4.38) and (4.40), we see that the two transformation laws are consistent if the
relation between F and G is

Y
F (e1 , . . . , e2g+2 ) = Ag G(e1 , . . . , e2g+2 ) (ei − ej )3k (4.41)
1≤i<j≤2g+2

with some constant Ag . Moreover, this is the most general transformation between F and
G that is holomorphic, invariant under permutations of the ei , and an isomorphism as
long as the ei are all distinct. So it must be correct with some choice of the constant.
(The relation between F and G is actually part of a general theory that is briefly indicated
below.)
This is not the whole story. The partition function has another important property in
genus 1: it can be regarded as an ordinary function, rather than a section of a line bundle.
Indeed, the genus 1 partition function can be defined as a trace, Z(q) = Tr q L0 , and in this
form it manifestly is equal to an ordinary function of q. Yet in the above presentation, the
genus 1 partition function is a section of L12k . What reconciles the two points of view is
that the line bundle L12 is trivial in genus 1. If we describe a genus 1 Riemann surface C
Q4
by the hyperelliptic equation y 2 = i=1 (x −ei ), then as explained in [61], L12 is trivialized
by the section
 12 Y
dx
s= (ei − ej )2 . (4.42)
y
1≤i<j≤4

The point of this formula is that s is holomorphic, SL(2, C)-invariant, invariant under
permutations of the ei , and nonzero as long as the ei are all distinct (so that C is smooth).
A power Ln of the line bundle L can be trivialized in this fashion precisely if n is an
integer multiple of 12 (and this gives one way to understand the fact that holomorphic
factorization is possible only if c is an integer multiple of 24). The genus 1 partition
function Z understood as an ordinary function is obtained from the section Θ of L12k by
dividing by sk . Thus
F
Z=Q 2k
. (4.43)
1≤i<j≤4 (ei − ej )

74
Combining this with (4.41), we get the relation (4.27) between the genus 1 partition func-
tion, understood as a trace, and the correlation function of twist fields:
 Y 
Z = A1 (ei − ej ) E(e1 )E(e2 )E(e3 )E(e4 ) . (4.44)
1≤i<j≤4

The numerical value of the constant A1 depends on how the twist field E has been nor-
malized.
What about g > 1? What is the best formula depends on how one prefers to treat
the conformal anomaly, and there seems to be in the physics literature no standard recipe
for doing so in the holomorphic context. To make contact with the genus 2 formulas of
Tuite [54], we may proceed as follows. A genus 2 Riemann surface is always a hyperelliptic
Q6
curve, y 2 = i=1 (x − ei ). In genus 2, we cannot trivialize the line bundle20 L12 , but we
can trivialize L10 by an obvious analog of (4.42):
 10 Y
dx x dx
se = ∧ (ei − ej )2 . (4.45)
y y
1≤i<j≤6

Hence, instead of regarding the genus 2 partition function as a section Θ of L12k , we can
sk of L2k . If we do this, then the relation between Z and the
regard it as a section Z = Θ/e
twist field correlation function is the obvious analog of (4.44):
 Y 
Z = A2 (ei − ej ) E(e1 )E(e2 ) . . . E(e6 ) . (4.46)
1≤i<j≤6

It can be shown that, in genus 2, a section of Lr is the same as a Siegel modular function
of weight r. So the formula (4.46) is the one we should use if we wish to regard the genus
2 partition function at c = 24k as a Siegel modular function of weight 2k, as was done in
[54]. This probably gives the most intuitively appealing formulas in genus 2. For genus
greater than 2, we leave the choice of the most useful formalism to the reader.
We conclude by briefly sketching the theoretical context for the formula (4.41). In
doing so, we will not limit ourselves to the case of a hyperelliptic Riemann surface, but
will consider an arbitrary base curve C0 and a two-fold cover C → C0 branched at points
e1 , . . . , es . Let Ki be the cotangent bundle to C0 at the point ei . Let LC0 and LC be the
determinant lines of C0 and C respectively. Then as has been shown by P. Deligne, using
results in [62], there is a natural isomorphism L8 ∼
= L16 ⊗i KC . If C0 is of genus zero,
C C0 i

then LC0 is naturally trivial, and the 3k/2 power of this isomorphism is the map in (4.41).
(The isomorphism has a natural square root, relevant if k is odd.)
20
This fact is consistent with holomorphic factorization at c = 24k, since the genus 2 partition
function cannot be written as a trace and so need not be definable as an ordinary function.

75
Appendix A. More Extremal Partition Functions

F In section 3.3, we have already described the extremal superconformal partition

functions Fk∗ for k ∗ ≤ 4. The purpose of this appendix is to briefly describe these functions,
and the corresponding Ramond partition functions Hk∗ , for 5 ≤ k ∗ ≤ 10.
For k ∗ = 5, we have

F5 = K 5 − 1380K 3 − 10239K 2 + 324871K + 2579929

= q −5/2 + q −1 + q −1/2 + 1 + 924492q 1/2 + 185710868q + 11953414737q 3/2 (A.1)
+ 416513865728q 2 + 9727092442216q 5/2 + . . . .

The corresponding Ramond partition function is

H5 = 143+371027968q+832984743936q 2+340658459983872q 3+56418463131631616q 4+. . . .

(A.2)
The coefficients in these series cannot be expressed as linear combinations of dimensions
of monster representations with reasonably small coefficients. But strangely, at least one
difference can be so represented; the difference between the coefficient of q in F5 and half
the corresponding coefficient in H5 is 196884 = 1 + 196883. (The coefficients in H5 , except
the first one, are even because of supersymmetry, so it is natural to divide by 2.) Also, the
first coefficient in H5 , namely 143, is the dimension of the smallest nontrivial representation
of a sporadic finite group known as the Suzuki group. Unfortunately, the other coefficients
are too large, and the representations of the Suzuki group too small, for it to be easy to get
convincing evidence of whether the Suzuki group is a symmetry of supergravity at k ∗ = 5.
For k ∗ = 6, the partition functions are

F6 = K 6 − 1656K 4 − 12287K 3 + 618373K 2 + 6487237K − 12026567

= q −3 + q −3/2 + q −1 + q −1/2 + 1 + 3724378q 1/2 + 1298410586q
+ 127852130050q 3/2 + 6378693040128q 2 + 204402839559265q 5/2 + . . . (A.3)
H6 = 1 + 2589372416q + 12754796707840q 2
+ 9529193701720064q 3 + 2622934801904828416q 4 + . . . .

The fact that the leading coefficient in the Ramond sector is 1 suggests that the model
may be invariant under a very large discrete symmetry group. The monster M does not
work, since the first non-trivial NS coefficient 3724378 cannot be expressed nicely in terms
of dimensions of monster representations. It turns out that another large sporadic group,

76
the baby monster B, is a better candidate. The dimensions r1 , . . . , r12 of the first 12
irreducible representations of B are listed in Table 3. In terms of those dimensions, we find
that the first two non-trivial NS coefficients in the partition functions can be expressed as
follows
3724378 = 7r1 + 4r2 + 3r3 + 3r4
(A.4)
1298410586 = 14r1 + 16r2 + 7r3 + 8r4 + 4r6 + 3r7 + r8 + 2r9 .

These formulas are not unique, as the ri obey linear relations with small coefficients, for
example r1 + r3 + r5 + r8 = r2 + r4 + r9 . If we let h1 denote the first non-trivial coefficient
in H6 , then it turns out that h1 /2 = f1 − f1/2 = 1298410586 − 3724378. So by subtracting
the formulas in (A.4), h1 /2 can also be expressed as a positive linear combination of the ri
with fairly small coefficients. If is also true that h2 /2 = f2 − f1 + f1/2 , and so is consistent
with B symmetry if f2 can be suitably expressed in terms of the ri .
These results are suggestive of baby monster symmetry at k ∗ = 6, though they are
perhaps a little less striking than the evidence presented in section 3.3 for monster sym-
metry at k ∗ = 4, since the required coefficients are larger and the dimensions involved are
smaller. Also, it is harder to continue this analysis beyond the first few coefficients, as
many representations of B come in.

r1 1
r2 4371
r3 96255
r4 1139374
r5 9458750
r6 9550635
r7 63532485
r8 347643114
r9 356054375
r10 1407126890
r11 3214743741
r12 4221380670

Table 3. Presented here from [53] are the dimensions ri of the ith irre-
ducible representation of the baby monster group B, for i = 1, . . . , 12.

77
 We present the remaining cases with little comment. For k ∗ = 7, we have
F7 = K 7 − 1932K 5 − 14335K 4 + 988051K 3 + 11525041K2 − 75824563K − 840705550
= q −7/2 + q −2 + q −3/2 + q −1 + q −1/2 + 2 + 13404883q 1/2
+ 7740446996q + 1126452195714q 3/2 + 78170641348884q 2 + . . . .
H7 = 262 + 15394525184q + 156260255891456q 2 + 203203950584774656q 3 + . . . .
(A.5)
For k ∗ = 8, the analogous formulas are
F8 = K 8 − 2208K 6 − 16383K 5 + 1433905K 4 + 17693341K 3 − 213343055K 2
− 3164679732K − 2780845557
= q −4 + q −5/2 + q −2 + q −3/2 + q −1 + 2q −1/2 + 3 + 44146598q 1/2
+ 40700662036q + 8516908978515q 3/2 + . . .
H8 = −213 + 80651894784q + 1605169778655232q 2 + 3496922597386551296q 3 + . . . .
(A.6)
This is the first case in which the leading coefficient in H is negative, so that it is necessary
to add an integer to F and H.
For k ∗ = 9, we find
F9 = K 9 − 2484K 7 − 18431K 6 + 1955935K 5 + 24992137K 4 − 445606619K 3
− 7443672266K 2 + 1774761946K + 223898812203
= q −9/2 + q −3 + q −5/2 + q −2 + q −3/2 + 2q −1 + 3q −1/2 + 3 + 135149374q 1/2
+ 193216791918q + 56847816152503q 3/2 + . . .
H9 = 453 + 381161021440q + 14282018665201664q 2 + 50519288391656243200q 3 + . . .
(A.7)
Finally, for k ∗ = 10, we find
F10 = K 10 − 2760K 8 − 20479K 7 + 2554141K 6 + 33421429K 5 − 793639831K 4
− 14145708496K 3 + 30831695165K 2 + 1166011724825K + 2482063616019
= q −5 + q −7/2 + q −3 + q −5/2 + q −2 + 2q −3/2 + 3q −1 + 3q −1/2 + 3
+ 389274233q 1/2 + 842231630010q + 341925580784341q 3/2 + . . .
H10 = −261 + 1652836102144q + 112692628289650688q 2 + 630520566901614002176q 3 + . . . .
(A.8)
This gives a second example in which the leading coefficient of H is negative, so that it is
necessary to add an integer.

78
 References

[1] H. Leutwyler, “A 2+1 Dimensional Model For The Quantum Theory Of Gravity,”
Nuovo Cim. 42A (1966) 159.
[2] E. J. Martinec, “Soluble Systems In Quantum Gravity,” Phys. Rev. D30 (1984) 1198.
[3] S. Deser, R. Jackiw, and G. ’t Hooft, “Three-Dimensional Einstein Gravity: Dynamics
of Flat Space,” Annals Phys. 152 (1984) 220.
[4] S. Carlip, “Conformal Field Theory, (2 + 1)-Dimensional Gravity, and the BTZ Black
Hole,” gr-qc/0503022.
[5] A. Achúcarro and P. Townsend, “A Chern-Simons Action for Three-Dimensional anti-
De Sitter Supergravity Theories,” Phys. Lett. B180 (1986) 89.
[6] E. Witten, “(2+1)-Dimensional Gravity as an Exactly Soluble System,” Nucl. Phys.
B311 (1988) 46.
[7] A. Ashtekar, “New Variables For Classical And Quantum Gravity,” Phys. Rev. Lett.
57 (1986) 2244-2247.
[8] J. Teschner, “On The Relation Between Quantum Liouville Theory And The Quantum
Teichmuller Spaces,” Int. J. Mod. Phys. A19S2 (2004) 459-477, hep-th/0303149.
[9] J. Teschner, “From Liouville Theory To The Quantum Gravity Of Riemann Surfaces”
(International Congress on Mathematical Physics 2003), hep-th/0308031.
[10] J. Teschner, “A Lecture On The Liouville Vertex Operators,” Int. J. Mod. Phys.
A19S2 (2004) 436-458, hep-th/0303150.
[11] J. Teschner, “An Analog Of A Modular Functor From Quantized Teichmuller Theory,”
math/0510174.
[12] M. Bañados, C. Teitelboim, and J. Zanelli, “The Black Hole In Three-Dimensional
Spacetime,” Phys. Rev. Lett. 69 (1992) 1849-1851, hep-th/9204099.
[13] A. Strominger, “Black Hole Entropy From Near Horizon Microstates,” JHEP 9802:009
(1998), hep-th/9712251.
[14] D. Birmingham, I. Sachs, and S. Sen, “Entropy Of Three-Dimensional Black Holes In
String Theory,” Phys. Lett. B424 (1998) 275-280, hep-th/9801019.
[15] J. Maldacena, “The Large N Limit Of Superconformal Field Theories And Super-
gravity,” Adv. Theor. Math. Phys. 2 (1998) 231-252, hep-th/9711200.
[16] C. G. Callan, Jr., S. B. Giddings, J. A. Harvey, and A. Strominger, “Evanescent Black
Holes,” Phys. Rev. D45 (1992) 1005-1009, hep-th/9111056.
[17] N. Goheer, M. Kleban, and L. Susskind, “The Trouble With de Sitter Space,” JHEP
0307:066 (2003), hep-th/0212209.
[18] E. Witten, “Quantum Gravity In de Sitter Space,” hep-th/0106109.
[19] A. Strominger, “The dS/CFT Correspondence,” JHEP 0110:034 (2001), hep-th/0106113.
[20] S. Kachru, R. Kallosh, A. Linde, and S. Trivedi, “de Sitter Vacua In String Theory,”
Phys. Rev. D68:046005,2003, hep-th/0301240.

79
[21] J. D. Brown and M. Henneaux, “Central Charges In The Canonical Realization Of
Asymptotical Symmetries: An Example From Three-Dimensional Gravity,” Commun.
Math. Phys. 104 (1986) 207-226.
[22] A. B. Zamolodchikov, “Irreversibility of the Flux of the Renormalization Group in a
2D Field Theory,” JETP Lett. 43 (1986) 430.
[23] A. N. Schellekens, “Meromorphic c = 24 Conformal Field Theories,” Commun. Math.
Phys. 153 (1993)159-186, hep-th/9205072.
[24] I. B. Frenkel, J. Lepowsky, and A. Meurman, “A Natural Representation of the
Fischer-Griess Monster With the Modular Function J As Character,” Proc. Natl.
Acad. Sci. USA 81 (1984) 3256-3260.
[25] I. B. Frenkel, J. Lepowsky, and A. Meurman, Vertex Operator Algebras And The
Monster (Academic Press, Boston, 1988).
[26] L. Dixon, P. Ginsparg, and J. Harvey, “Beauty And The Beast: Superconformal Sym-
metry In A Monster Module,” Commun. Math. Phys. 119 (1988) 221-241.
[27] L. Dolan, P. Goddard, and P. Montague, “Conformal Field Theory, Triality, and the
Monster Group,” Phys. Lett. B236 (1990) 165-172.
[28] R. Borcherds, “Automorphic Forms And Lie Algebras,” in Current Developments In
Mathematics (International Press, Boston, 1997).
[29] J. McKay, “The Essentials of Monstrous Moonshine,” Groups and Combinatorics - in
Memory of M. Suzuki, Adv. Studies in Pure Math. 32 (2001) 347-343.
[30] T. Gannon, “Moonshine Beyond The Monster: The Bridge Connecting Algebra, Mod-
ular Forms, and Physics” (Cambridge University Press, 2006).
[31] G. Höhn, “Selbstduale Vertexoperatorsuperalgebren und das Babymonster,” Ph.D.
thesis (Bonn 1995), Bonner Mathematische Schriften 286 (1996), 1-85, arXiv:0706.0236.
[32] G. Höhn, “Conformal Designs based on Vertex Operator Algebras,” arXiv:math/0701626.
[33] M. Jankiewicz and T. Kephart, “Modular Invariants And Fischer-Griess Monster,”
arXiv:math-ph/0608001.
[34] R. Dijkgraaf, J. Maldacena, G. Moore, and E. Verlinde, “A Black Hole Farey Tail,”
hep-th/0005003.
[35] J. F. Duncan, “Super-Moonshine For Conway’s Largest Sporadic Group,” arXiv:math/0502267.
[36] S. Deser, R. Jackiw, and S. Templeton, “Topologically Massive Gauge Theories,”
Annals Phys. 140 (1982) 372-411.
[37] G. Moore and N. Seiberg, “Taming The Conformal Zoo,” Phys. Lett. B220 (1989)
422.
[38] E. Witten,“SL(2, Z) Action On Three-Dimensional Conformal Field Theories With
Abelian Symmetry,” hep-th/0307041.
[39] S. Carlip “Quantum Gravity In 2+1 Dimensions: The Case Of A Closed Universe,”
Living Rev. Rel. 8 (2005) 1, gr-qc/0409039.

80
[40] J. Hartle and S. B. Hawking, “Wavefunction Of The Universe,” Phys. Rev. D28 (1983)
2960.
[41] D. Friedan and S. Shenker, “The Analytic Geometry Of Two-Dimensional Conformal
Field Theory,” Nucl. Phys. B281 (1987) 509.
[42] S. Gukov, E. Martinec, G. W. Moore, and A. Strominger, “An Index for 2-D Field
Theories with Large N = 4 Superconformal Symmetry,” hep-th/0404023.
[43] J. Thompson, “Some Numerology Between the Fischer-Griess Monster and the Elliptic
Modular Function,” Bull. London Math. Soc. 11 (1979), 352-353.
[44] J. H. Conway and S. P. Norton, “Monstrous Moonshine,” Bull. London Math. Soc.
11 (1979) 308-339.
[45] N. Koblitz, Introduction To Elliptic Curves And Modular Forms (Springer-Verlag,
1984).
[46] A. Maloney and E. Witten, to appear.
[47] D. Birmingham and S. Sen, “Exact Black Hole Entropy Bound In Conformal Field
Theory,” hep-th/0008051.
[48] L. Dixon, J. A. Harvey, C. Vafa, and E. Witten, “Strings On Orbifolds,” Nucl. Phys.
B261 (1985) 678-686.
[49] M. P. Tuite, “On The Relation Between Generalized Moonshine And The Unique-
ness Of The Moonshine Module,” Commun. Math. Phys. 166 (1995) 495-532, hep-
th/9305057.
[50] J. Conway and N. A. Sloane, Sphere Packings, Lattices, and Groups (Grundlehren der
Mathematischen Wissenschaften, 1998).
[51] G. Nebe, “Some Cyclo-Quaternionic Lattices,” J. Algebra 199 (1998) 472-298.
[52] F. Gliozzi, D. I. Olive, and J. Scherk, “Supersymmetry, Supergravity Theories, And
The Dual Spinor Model,” Nucl. Phys. B122 (1977) 253-290.
[53] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, Atlas Of
Finite Groups (Oxford University Press, 1985).
[54] M. P. Tuite, “Genus Two Meromorphic Conformal Field Theory,” math.qa/9910136.
[55] G. Mason and M. P. Tuite, “On Genus Two Riemann Surfaces Formed From Sewn
Tori,” Commun. Math. Phys. 270 (2007) 587-634, math.qa/0603088.
[56] Al. B. Zamolodchikov, “Correlation Functions Of Spin Operators In The Ashkin-
Teller Model And The Scalar Field On A Hyperelliptic Surface,” Soviet Phys. JETP
63 (1986) 1061-1066.
[57] V. G. Knizhnik, “Analytic Fields On Riemann Surfaces,” Commun. Math. Phys. 112
(1987) 567.
[58] L. Dixon, D. Friedan, E. Martinec, and S. Shenker, “The Conformal Field Theory Of
Orbifolds,” Nucl. Phys. B282 (1987) 13-73.
[59] S. Hamidi and C. Vafa, “Interactions On Orbifolds,” Nucl. Phys. B279 (1987) 465.

81
[60] A. B. Zamolodchikov, “Infinite Additional Symmetries In Two-Dimensional Conformal
Quantum Field Theory,” Theor. Math. Phys. 65 (1985) 1205-1213.
[61] G. Segal, “The Definition Of Conformal Field Theory,” in U. Tillmann, ed. Topology
Geometry, and Quantum Field Theory (Cambridge University Press, 2004).
[62] P. Deligne, “Le Déterminant de La Cohomologie,” in Current Trends In Arithmetical
Algebraic Geometry, Contemp. Math. 67 (Amer. Math. Soc., Providence, 1987).

82