Real observers solving imaginary problems

Juan Maldacena
arXiv:2412.14014v1 [hep-th] 18 Dec 2024

Institute for Advanced Study, Princeton, NJ 08540, USA

Abstract
The sphere partition function is one of the simplest euclidean gravity computations. It is
usually interpreted as count of states. However, the one loop gravity correction contains a
dimension depenent phase factor, iD+2 , which seems confusing for such an interpretation. We
show that, after including an observer, this phase gets cancelled for the quantity that should
correspond to a count of states.
Contents
1 Introduction 2

2 Review of Polchinski’s computation of the phase factor 3

3 An analogy: A massive particle on a sphere 4

3.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 The sign of the phase factor for gravity . . . . . . . . . . . . . . . . . . . . . . . . . 7

4 The partition function with the observer included 8

4.1 Partition function of the clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 Explaining the last factor of i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

5 Discussion 11

A Conformal Transformations of the sphere 12

B Result for the sphere partition function 12

C Comment on a factor of i in JT gravity 13

D Comment on the zero mode counting 14

1 Introduction
The sphere partition function is probably the simplest object we can compute in euclidean quantum
gravity with a positive cosmological constant.
The classical action is usually called de Sitter entropy, which suggests a state counting inter-
pretation [1]. A puzzle appears when we compute the one loop correction. For pure gravity, it was
observed by Polchinski, that we get a dimension dependent phase (in D dimensions)
 
Ac
ZS D = exp ZSgrav
D , ZSgrav
D = i
D+2
× (Real and positive) (1)
4GN

where Ac is the area of the cosmological horizon for an observer and ZSgrav D is the one loop gravity
contribution. As we will review in section 2, the factors of i are related to negative modes. The rest
of the determinants give a real and positive answer and their precise values can be found in [2], see
appendix B for a summary. This result was obtained for pure gravity, but the addition of stable
matter contributes only to the real and positive term, but no extra phase1 .
For generic D this seems to be an obstruction to a state counting interpretation of the sphere
partition function.
In this paper we will make two separate observations.
The first is to give a toy example where a similar phase appears from a computation that one
might have naively expected to give a real and positive answer. The toy example corresponds to a
massive particle propagating in a euclidean rigid sphere with no gravity.
The second is to include an observer, in the spirit of [3], and argue that with a suitable interpre-
tation the phase really drops out when we focus on the quantity that is relevant for the computation
1
There can be a divergence for some cases, such as a real massless field with a non-compact target space. We will
assume we do not have such cases here.

2
of the entropy. In other words, the factors of i in (1) are precisely those necessary to cancel against
factors of i that appear from two sources, one is related to factors of i in the partition function for
the trajectory of the observer and the other is related to a difference between the euclidean path
integral and the integral we need to do in order to impose the Hamiltonian constraint.
In summary, the i disappears from a more refined quantity, which is then a good candidate for
a state counting interpretation.

2 Review of Polchinski’s computation of the phase factor

In this section we review Polchinski’s computation [4]. We consider a pure gravity theory with the
action
√
Z
1
I=− g(R − 2Λ) , 2Λ = (D − 1)(D − 2) (2)
16πGN
where we picked Λ so that the radius of the sphere solution is one. We expand around the sphere
solution by setting ds2 = dŝ2sphere + hµν dxµ dxν . We further consider the gauge fixing function

ˆ µ hµ − 1 ∇
fν = ∇ ˆ νh , h ≡ hµµ (3)
ν
2
We add a term to the action of the form f 2 with a suitable coefficient, and also add the corresponding
1
ghost terms. We define the traceless part of the metric fluctuation via ϕµν = hµν − D ĝµν h. Then
find the quadratic action
Z p ( 2
)
1 ˆ 2 + 2)ϕµν − (1 − D ) ˆ 2 − 2(D − 1)]h + bµ [−∇ˆ 2 − (D − 1)]cµ
I= ĝ ϕµν (−∇ h[−∇
64πGN 4
(4)
where ∇ ˆ 2 is the laplacian on the unit radius sphere acting on the objects with the corresponding
indices. As it is well known [5], we get the wrong sign kinetic term in front of the fluctuations of
the overall scale factor of the metric. We deal with this by performing a contour rotation to the
imaginary direction, h → iĥ. This contour rotation produces a factor of i∞ in the partition function.
This is an ultralocal term that can be absorbed into a renormalization of the local terms we already
had in the action. Of course, we expect that the final renormalized values of such parameters are
all real. After this step the kinetic term for ĥ has the right sign.
The first point to make is that none of the bosonic operators has a zero mode. The ghost term
has zero modes related to the isometries of the sphere, and we just simply do not integrate over
them and divide by the volume of the sphere.
However, we find that, after the rotation h → iĥ, the quadratic operator for ĥ in (4) has some
negative modes. This happens for two modes. The angular momentum zero mode, ℓ = 0, which
is a single mode, and the angular momentum one mode, ℓ = 1, which is (D + 1)-fold degenerate.
The functional integrals for these modes should be rotated back to the real h axis and this gives
the factors of i in (1).
The traceless symmetric tensor part has only positive eigenvalues and it gives a real and positive
answer. The ghosts also give a real and positive answer. This is not obvious since the vector operator
has some negative modes. However, as argued in [4], one should consider the absolute value of the
ghost determinant. At this point we do not know whether we should have iD+2 or (−i)D+2 . We
will give a procedure for fixing this sign in later sections.
As a side comment, we should mention that the paper [6] claims to get a different answer than
[4], an answer with no phase. We think that this paper contains an mistake2 .
2
The mistake is in their treatment of what they call the field χ+ which is is not a full field, it is missing the ℓ = 0, 1

3
Figure 1: We consider a particle of mass m propagating on a Euclidean sphere. A possible classical
solution is a trajectory along a maximal circle. The sum over paths also involves shorter paths which are
the dominant ones.

3 An analogy: A massive particle on a sphere

In this section, we discuss a simpler problem where some surprising factors of i appear in a euclidean
computation. In addition, we will use the solution to this problem in section 4.
We consider a massive scalar particle propagating on a Euclidean S D . The logarithm of the
field theory partition function can be viewed as a sum over paths
X
log Zfield = e−mLpath , m≫1 (5)
paths

We stressed that we are interested in the semiclassical limit of large mass, m ≫ 1. (We are setting
the radius of the sphere to one, so m → mR where R is the radius.).
A particular classical solution consists of a path that wraps once around a great circle of the
sphere, see figure 1. In coordinates

ds2 = cos2 θdτ 2 + dθ2 + sin2 θdΩ2D−2 (6)

we are talking about the solution at θ = 0 and extended along the τ direction. Expanding around
that solution we find that the action has the form
Z
m ⃗ 2 − θ⃗2 ] + · · · ,
I = 2πm + dτ [(∂τ θ) θ≪1 (7)
2

where θ⃗ = θ⃗n, with ⃗n a point on S D−2 . We see that this action has some negative modes correspond-
ing to the case of constant θ.⃗ There are D − 1 of them. These negative modes are not surprising,
they correspond to moving the circle away from the maximum circle which would decrease their
length. The negative modes then give a contribution of the form

log Z ⊃ · · · + (±i)D−1 e−2πm × (power of m) + · · · (8)

where we are highlighting the phase factor and the notation suggests that we have not yet decided
on the overall sign.
modes. This field is rotated, χ+ → iχ+ near equation (27) of [6]. Once the missing modes are taken into account we
reproduce Polchinski’s answer. I thank Ted Jacobson for bringing [6] to my attention.

4
Figure 2: In red we see the defining contours for the integral (10). Note that the C− part of the contour
is oriented differently for the D even case relative to the D odd case. In (a) we discuss the D odd case.
Here we can add and subtract the small piece Cs which give us the simple terms in (11). Then we can
shift the contour to C ′ and move it to the upper half plane picking up the poles through Cp . In (b) we
discuss the D even case. In this case the orientations of the contours is such that we cannot move the
whole contour to the upper half plane. But we can move it to the direction where the exponential eiνt
decreases the fastest. This is the dotted line, when the phase of ν is deformed as indicated. In this
process we pick up the same poles as before, by the contours Cp . If we had chosen the other sign of
ϵ the dotted line would have been in the upper right quadrant and the orientation of the Cp contours
would have been the opposite.

In addition, we have some zero modes. The zero modes arise from the geometric symmetries
that are spontaneously broken by the choice of path, so they give finite factors since the relevant
groups are all compact. They give the power of m in (8).
Now, the problem of a particle on a sphere can be solved exactly using quantum field theory
methods and the exact answer is, see eg. [2],

dt cosh 2t
∞
Z p
log Zfield = 2 cos νt , ν = m2 − (D − 1)2 /4 ∼ m = mR ≫ 1 (9)
ϵ t (2 sinh 2t )D
1 cosh 2t
Z Z
iνt
= dth(t) + dth(t) , with h(t) ≡ t De (10)
C+ C− t (2 sinh 2 )

where ϵ is a short distance cutoff that produces only terms that can be cancelled by local countert-
erms. These divergent terms are polynomial in ν. We have also expressed the integral as a contour
integral with the defining contours C± in figure 2.
Let us now consider first the case that D is odd. In this case, we can extend the integral to the
whole real line and write it as (see figure 2a)

dt cosh 2t
Z
log Z = (simple) + eiνt (11)
C′ t (2 sinh 2t )D

where the simple terms involve divergent terms proportional to 1/ϵ, as well as polynomial terms in
ν coming from the pole at t = 0. We can now shift the contour into the positive imaginary direction.

5
The first pole we pick up is at t = 2πi. Computing the residue we get that the final contribution is

ν D−1
log Z = (simple) + (±i)D−1 e−2πν + ··· , D = odd (12)
(D − 1)!

Since D is odd it does not matter whether we choose ±i. In this case, we simply have an overall
sign, alternating as we take D = 3, 5, 7, .... Thus, for odd D we have checked the prediction in
(8). In this case, the full expression (11) has some polynomial terms in ν and then the exponential
terms are cleanly separated. We have included the prefactor in the exponential (to leading order in
the 1/ν expansion). This arises from the zero mode integral of the particle path around the circle
solution.
We now consider the case of even D. In this case we expect a ±i from (8) while (9) is purely
real. Then, we seem to have a paradox. One comment is that, in this case, it is not possible to
combine the contour as in (11), see figure 2. In fact, the function has an infinite (asymptotic) series
expansion in powers of 1/ν. In these cases it might be hard to extract the exponential terms in ν.
Actually, what happens is that we are sitting precisely at a so called Stokes line when ν is real. A
Stokes line is precisely when an exponential correction in the large |ν| expansion , such as e−2πν ,
can appear or disappear, or change its overall sign [7].
The idea is that if we take

ν → |ν|(1 − iϵ) , or m = |m|(1 − iϵ) (13)

we get
X cn D−1
D−1 −2πν ν
log Zfield = + (−i) e + ··· (14)
n
νn (D − 1)!

while if we switch the sign of ϵ in (13) we would get iD−1 e−2πν . This result is obtained as follows.
We move the contour of integration as shown in figure 2(b). The exponential term in (14) is now
valid for both even and odd D, but for even D we should remember we are in the region (13).
Notice that the function contains a much larger term which is the one corresponding to small
paths, see figure 1. The new contours, C˜± in figure 2b, lie along a line which is close to the line
of maximum descent of the function, at least for large ν and fixed ϵ. We do do not expect any
additional saddle contribution from these integrals. So the exponential term is a small correction
to this much larger term.
The particular continuation (13) seems preferred if we think about the evolution in Lorentzian
time e−imt and we want to suppress terms for large t. For now we make this choice, but we later
comment about the other choice.
We can also get the particular sign in (14) from the particle path integral as follows. We have
seen that we encounter terms of the form
Z Z
πmθ2 2
dθe → dθeπ|m|(1−iϵ)θ (15)

We now need to decide whether we want to continue θ → ±iγ. The idea is that we want to rotate
the contour in the direction where we make the exponent decrease continuously, after we made the
iϵ deformation of the mass. In other words, we deform the contour avoiding the line of maximal
increase of the function, see figure 3a. This implies that we want to take it to the θ = −iγ direction,
for real γ, which produces
Z Z Z
πmθ2 π|m|(1−iϵ)θ2 2
dθe → dθe → −i dγe−π|m|γ ∝ −i (16)

6
Figure 3: (a) Rotation of the contour for the integral over the negative modes discussed in equation
(16). The dotted line denotes the direction of maximum increase. We rotate the contour in such a way
that we avoid this direction. The original contour is C and the new contour is C ′ . (b) Rotation of the
contour for the δβ = β − β0 integral discussed around equations (31) (32).

for each of the negative modes. Then we reproduce (14).

We can further say that if we focus on the one particle state, by looking purely at the exponential
correction, we get the partition function for a single particle

ν D−1
Zparticle = (−i)D−1 e−2πν , with ν = mR ≫ 1 (17)
(D − 1)!

where R is the radius of the sphere.

3.1 Discussion
In this section we have discussed the computation of the path integral of a particle on S D . The
only semiclassical contribution corresponds to the particles wrapping the circle. The dominant
contribution corresponds to small paths which do not have any obvious associated semiclassical
solution, see figure 1. The particle wrapping the circle appears as a small correction, which for
even D sits precisely at a Stokes line and its precise phase can only be determined after analytically
continuing m to a slightly complex value.
Now, what are the lessons from this analogy for the gravity case?. In the gravity case we could
think that the sphere is perhaps a subleading contribution to something else, and that is the reason
it has this phase factor. It is not clear what the something else is. It could be perhaps some
contribution involving small universes, very large universes, or very non-smooth universes... We
will not propose something concrete here. Regardless of these comments, we will be able to use (17)
when we talk about the quantity with a state counting interpretation in the next section.

3.2 The sign of the phase factor for gravity

We now attempt to derive the precise phase for the gravitational case. If we start from the integral
of the variation of the scale factor h we can continue to h → ±iγ. Since this is a full field, we get
no phase, independently of whether we choose ±i. At this stage, we get an integral over γ which is
a well defined gaussian except for D + 2 modes. Now, these few modes need to be rotated to real

7
contours in h. In order to determine how we do this we will change the cosmological constant as

Λ → |Λ|(1 − iϵ) , (18)

which is a change similar to (13), given that the cosmological constant can be viewed as the vacuum
energy and the mass of the particle is also its rest energy.
We now note that each of the D + 2 modes will contain an integral of the schematic form
Z   Z   Z  
1 2 1 2 1 2
dγ exp γ → dγ exp (1 + iϵ)γ → i dh exp − h , γ = ih
GN Λ GN |Λ| GN |Λ|
(19)
where we ignored an overall positive constant in the exponent. This then gives the phase factor in
(1).

4 The partition function with the observer included

In this section we will apply the philosophy of [3] to the euclidean problem. This means that, in
addition to the sphere partition function, we also have an observer which will be simply modeled
as a particle with a clock. The full partition function can then be written as
 
Ac
Zobs = exp ZSgrav
D Zparticle Zclock (20)
4GN
where the partition function of the particle is given in (17).
At this stage we see that the factor of iD+2 in (1) combines with the factor of (−i)D−1 in (17)
to give an overall i3 = −i. Let us discuss in more detail the cancellation between the D − 1 factors
of i from the sphere partition function and the D − 1 factors of −i from the particle.
In order to make this cancellation more clear, it is useful to recall that the sphere partition
function has a factor of i from the ℓ = 0 mode and iD+1 from the ℓ = 1 mode. As Polchinski
discusses in [4], the ℓ = 1 mode comes exclusively from the conformal killing vectors of the sphere.
There are precisely D + 1 of them. Since they are a reparametrization, these are zero modes of the
original Einstein action, but they become non-zero (negative) modes once we add the gauge fixing
term. It is useful to understand how these modes act on the particle. Out of the D + 1 modes,
there are two of them which map the circle to the circle, acting as an SL(2) transformation on the
coordinate of the circle. More important for now are the D − 1 of them that move the circle away
from θ = 0. These change the position of the particle in the same way as the negative modes of the
particle action (7), see appendix A for more details.
Therefore we expect that an alternative gauge fixing scheme we should not have these reparametriza-
tion modes and we would not have the negative modes on the particle action either. In other words,
we expect that these two phase factors cancel precisely. This is an indirect way to check that the
sign of the i are as we stated in (1) (17), so that the i and the (−i) are cancelling for each of the
D − 1 modes. This also supports the correlation between the signs of ϵ in the deformations (13)
and (18).

4.1 Partition function of the clock

In this section, we argue that the clock variable does not contribute any extra i. The clock variable,
T , and its conjugate energy E are described by a path integral of the form
Z  Z Z 
Zclock = DT DE exp i Ṫ E − E (21)

8
in euclidean signature. Notice that T is the lorentzian time clock variable and it is not analytically
continued as we go to Euclidean time. The lorentzian action would be similar but with an i in the
last term in the exponential. In order to do this path integral we can first expand in Fourier modes.
For all non-zero Fourier modes the integral over Tn gives a delta function for En , which makes the
integral over En trivial. Finally, we are left with an integral over the zero modes of each of the two
variables
Z Z ∞
dE0 dT0
Zclock = exp (−βE0 ) = dE0 ρclock (E0 ) exp (−βE0 ) , ρclock (E0 ) = eSclock (22)
2π 0

The integral over T0 is infinite in this model. However, we expect that for a physical clock this
integral is regularized and is given by the entropy of the clock degrees of freedom that give rise to
the clock time. This is how we regularized that infinity here. Note that the integral is only over
positive energies E0 > 0. Here β is the size of the circle, which we will eventually set to β = 2π, but
we will leave it general for now. (Of course, if we say that Sclock is finite, then the energies should
be discrete, and the continuum just an approximation.)
A more physical model for a clock could be a near extremal charged black hole, for example.
Here we assumed that the entropy of the clock is independent of its energy, but it is not difficult to
consider also a situation where it also has some energy dependence.

4.2 Explaining the last factor of i

So far, we have concluded that (20) is equal to

Zobs = −i × (positive) , with − i = i3 (23)

The three remaining factors of i arise as follows. One i from the overall size mode, and i2 from the
two of the conformal killing vectors that preserve the particle trajectory (reparametrizing its time,
see appendix A).
In order to relate (20) to an observable with a state counting interpretation, we should think
more carefully about the interpretation of this coputation. The idea is that we have a clock Hilbert
space Hclock and the Hilbert space of the de Sitter degrees of freedom HdS , which describes the rest
of the static patch except for the clock. In addition, we have the particle which has a trivial Hilbert
space, which only contributes with its energy, ν, when we gauge fix its center of mass degree of
freedom. On these we end up imposing a Hamiltonian constraint

0 = HdS + Hclock + Hparticle = HdS + Hclock + ν (24)

We view (24) as the Hamiltonian constraint at the position of the observer. We can then view the
partition function (23) as
Z
Zobs = dβZPatch (β)Zparticle (β)Zclock (β) (25)

The first factor is the partition function of the de Sitter static patch degrees of freedom at fixed
temperature. This is a partition function we can only compute gravitationally. It only differs from
the sphere partition function by the fact that we are fixing the length of the circle at the location of
the particle. The second is the particle partition function (17) where we modify just the exponential
e−2πν → e−βν . The third is the clock partition function (22). Of course, in the gravitational path
integral the integral over β is evaluated via saddle point, and the saddle is at β = 2π. Here we are
writing it explicitly because it will be useful for the physical interpretation.

9
 As a side comment, we can note that one alternative way to think about this would be to put
a Dirichlet boundary condition around the particle trajectory. The factor ZPatch would include
everything outside a small cylinder that surrounds the particle, as was discussed in [8], see also [9].
The second and third factors in (25) would describe everything inside the cylinder, including the
clock degrees of freedom.
We can think of each of the factors in (25) as given by some integral over energies of some
density of states, as in (22), and

ν D−1
Z
ZPatch (β) = dEρPatch (E)e−βE , ZParticle (β) = e−βν (26)
(D − 1)!

This is just a definition of ρPatch (E). After a contour rotation

β → β0 − is (27)

we see that the integral over s just sets the total energy to be zero, thus imposing (24). (Below we
will fix the precise sign in front of the i in (27)). In other words, we get

ν D−1
Z
Zobs = −iZCount , ZCount ≡ 2π dEρPatch (−ν − E) × eSclock (28)
(D − 1)!
The quantity ZCount is defined by the equation on the right. It is a natural candidate for a count
over the number of degrees of freedom accessible to a static patch observer. We see that ZCount
differs from (20) (25) precisely by a factor of −i that comes from the change in measure from β to
s (27).
We can further write
ρdS (−E) = eSds e−2πE (29)
where eSdS is, by definition, the prefactor in this formula. The energy dependence is fixed by the
dependence of the full partition function Zobs on the mass of the particle.
Then we write
D−1 Z ∞ e−2πν ν D−1
Sclock ν
ZCount = e (2π)dEρdS (−ν − E) = eSdS eSclock (30)
(D − 1)! 0 (D − 1)!
This factor of i is also present when we go from the canonical ensemble to the microcanonical
one in usual thermodynamics. It is present in the sense that we should integrate over s in order
to go from the canonical to the microcanonical, rather than integrating over β. The two integrals
differ by an i. See appendix C for a comment on a similar factor of i in AdS JT gravity.
In the euclidean sphere partition function we have an integral over β, so there is a ±i when we
go the integral over s which extracts the microcanonical density of states.
In order to figure out the overall sign of the i in (27) and (30), we follow the following logic.
The integral over β is roughly like the integral over the overall scale factor, so we expect that it has
a form, recalling (18),
Z   Z  
1 2 1 2
dβ exp − (β − β0 ) → dβ exp − (1 + iϵ)(β − β0 ) , β0 = 2π (31)
GN Λ GN |Λ|

As we did previously, we rotate the contour avoiding the direction of maximum increase, see figure
3b. The bottom line is that Z Z ∞
dβ → −i ds (32)
−∞

10
 Therefore a −i in Zobs is precisely what is necessary for ZCount to be real and positive, so that
it could plausibly correspond to a count over states.
Using (30) we conclude that

e−2πν ν D−1 Sclock

ZCount = iZobs = eSdS e , ν = mR ≫ 1 (33)
(D − 1)!

with Ac Ac
eSdS ≡ e 4GN |ZS D | = e 4GN |ZSgrav QF T
D |ZS D (34)

where ZSgrav
D is the partition function for gravity computed in [2], see appendix (B). And ZSQF
D
T
is
the partition function of the quantum field theory that can be coupled to gravity, which is positive
by reflection positivity.

5 Discussion
We have shown that after including the observer we remove the dimension dependent factors of i.
They are removed by taking into account the degrees of freedom of the position of the observer. The
final factor of i is removed by correctly taking into account the connection between the integral over
the proper length along the observer and the similar integral along an imaginary direction which
imposes the Hamiltonian constraint at the observer location.
In order to obtain these results we had to analytically continue the parameters of the theory,
both the mass and the cosmological constant as in (13) (18). We would obtain the same final answer
(33) if we had changed the sign of ϵ. Everywhere that we had an i we would need to put a −i in the
intermediate steps. One important point is that the phase rotation in Λ and m should be correlated
as in (13) (18). This correlation can be justified a posteriori by noticing the cancellation between
the factors of i from the conformal killing vectors and the ones from the observer position.
An interesting possible computation for the future, which can check the cancellation of the D
dependent phase factors, is the following. We can consider the euclidean partition function of a
near extremal charged black hole in de-Sitter in the regime where the radius of the black hole is
significantly smaller than the radius of de Sitter. Such a black hole can be in thermal equilibrium
with the cosmological horizon leading to a smooth solution in Euclidean signature. We expect that
the straight partition function should only give a −i factor, which would then be interpreted as in
section 4.2.
In [3], a type II1 algebra of observables was defined for an observer with a clock. In that
description, the entropy is only defined up to an overall additive constant. This computation
suggests that there is some sense in which that discussion is an approximation to a computation
that involves a finite dimensional matrix type algebra, whose overall dimension is given by ZCount .
We should emphasize that [3] gives an actual construction of the algebra of observables they use
to compute their entropy. The object we defined here as ZCount is only defined gravitationally and
we have not given an explicit Hilbert space realization for it. We leave that as an exercise for the
reader!.
Previous work had identified the sphere partition function (1) as a count of states [1]. The
problem with that is the pesky factors of i in (1). We have shown here that these factors are
naturally removed if we include the observer and we think of (33) as the appropriate count of
states.
It is not clear to us whether this entropy should be viewed as a fine grained or coarse grained
entropy. Our only point is that by including the observer we get a nicer quantity from the Eucidean

11
computation. This was inspired by the observation in [3] that by including the observer we get a
well defined Hilbert space and algebra of observables.

Acknowledgments
I would like to thank D. Stanford for initial discussions on sphere partition functions which prompted
this work. I would also like to thank A. Herderschee, V. Ivo, T. Jacobson, D. Jafferis, E. Silverstein,
Z. Sun, and E. Witten for discussions.
This work was supported in part by U.S. Department of Energy grant DE-SC0009988.

A Conformal Transformations of the sphere

It is convenient to write the sphere as
2
−Y−1 + Y12 + · · · YD+1
2
=0, Y M ∼ λY M (35)

with the indicated identification under rescalings of Y M . The full conformal group is SO(1, D + 1)
acts linearly on Y M . The group of rotations is SO(D + 1) acting on the last D + 1 variables. The
sphere metric is given by
dYM dY M
ds2 = 2 (36)
Y−1
which can be seen most clearly by setting the “gauge condition” Y−1 = 1.
The special conformal generators are generated by the transformations are δY i = bi and δY−1 =
i
bi Y . We see that they change the metric of the sphere (36) by a scale factor.
The circle in question can be set as Yi = 0 for i = 3, · · · , D + 1. Then the transformations with
only b1 , b2 non-zero leave the circle invariant as a whole, but correspond to a special conformal
transformation of the τ coordinate in (6). On the other hand, the transformations with b1 = b2 = 0
move the circle in the θ directions. There are precisely D − 1 of these. We see that near θ = 0 they
act as a translation in the θ⃗ directions that move the circle away from the maximum circle line.

B Result for the sphere partition function

In this appendix, we summarize the result obtained in [2]3 for the one loop contribution to the
sphere partition function. They found that, up to (a power law) UV divergent term, the pure
gravity result is4
 21 dimSO(D+1)
32π 3 GN

1
|ZSgrav
D | = Z̃char (37)
Vol[SO(D + 1)]c AD−2

where the absolute value tells us that we are ignoring the phase factor. With

dimSO(D + 1) = (D + 1)D/2 (38)

D+1
Y 2π k/2
Vol[SO(D + 1)]c = (39)
Γ(k/2)
k=2
3
I thank Z. Sun for his help in producing this summary.
4
Note that D here is D = d + 1 where d is defined as in [2].

12
and
Z ∞      
dt (1 + q) 1 1 2 2
log Z̃char = χ̂bulk − χ̂edge + (D + 3) +q − + q + D − D − 4 (40)
ϵ 2t (1 − q) q q2

with q = e−t and

(D + 1)(D − 2) (q D−1 + 1) (q D + 1/q)

χ̂bulk = − (D − 1) (41)
2 (1 − q)D−1 (1 − q)D−1
(q D−2 + 1/q) (q D−1 + 1/q 2 )
χ̂edge = (D + 1) − (42)
(1 − q)D−3 (1 − q)D−3
The parameter ϵ in (40) regulates the UV divergence at t = 0 and should be viewed as
ϵphys
ϵ= , (43)
R
where R is the radius of the sphere and ϵphys is a physical proper distance cutoff. For D odd we
only have power law divergencies. For D even we also have a logarithmic divergence whose precise
subtraction procedure is more ambiguous and it is related to the precise definition of the coupling
of the Euler density counterterm that we can have in such dimensions. Note that the log ϵ gives
rise to some extra dependence on the dS radius R through (43).
In (37) the area is
D+1
2π 2
AD−2 = RD−2 ωD−2 , ωD−2 = (44)
Γ( D+1
2 )
where ωD−2 is the volume of a D − 2 dimensional sphere.
If we have gravity coupled to a quantum field theory, then the full answer is, of course, given by

|ZS D | = |ZSgrav QF T
D |ZS D (45)

where the ZSQF

D
T
is positive by reflection positivity of the quantum field theory.

C Comment on a factor of i in JT gravity

The reader might wonder why we do not need to worry about the factors of i in usual black hole
thermodynamics. In that case, β is fixed and we do not integrate over β when we compute the
standard canonical ensemble. We do need to worry about it when we integrate over β to go to the
microcannonical ensemble. That integral needs to be taken along the imaginary direction. In other
words, we set β = β0 − is and integrate over s.
An interesting example is the case of JT gravity. In that case, we have a real microcanonical
partition function or density of states. However, the path integral is related (but not equal) to
the integral of a particle in a magnetic field with a mass related to the energy [10, 11]. This
path integral gives a −i because this is a Schwinger like process describing pair creation in an
electric field5 . However, this −i is precisely the −i that arises from the difference between the path
integral over β (or real lengths of the trajectory of the particle which is what appears in the pair
creation computation) versus the integral over s, with β = β0 − is that we need to do to go to
the microcanonical ensemble. In other words, the density of states differs from the path integral
describing the Schwinger process precisely by this −i.
5
See [12] for a discussion for the −i that appears for unstable systems, and [13] for a more recent discussion of
vacuum decay.

13
D Comment on the zero mode counting
N/2
In gravitational path integrals the zero modes that arise from isometries contribute a factor of GN
where N is the number of zero modes. We can see this factor clearly in the sphere partition function
in (37) where the number of zero modes is given in (38). On the other hand, in the case that the
particle can be viewed as giving rise to a new classical solution we would expect a smaller symmetry
group which is SO(2) × SO(D − 1) with dimension
1
N ′ = (D − 1)(D − 2) + 1 (46)
2
If we naively extrapolate the particle computation (17) to this case, by taking m ∼ 1/GN , then
we see that the total partition function (20) will contain the factor of GN from gravity, as in (37),
N ′ /2
plus a factor coming from the mD−1 factor in (17), so that the total power of GN is indeed GN
as expected.

14
References
[1] G. W. Gibbons and S. W. Hawking, “Cosmological Event Horizons, Thermodynamics, and
Particle Creation,” Phys. Rev. D 15, 2738–2751 (1977)

[2] Dionysios Anninos, Frederik Denef, Y. T. Albert Law, and Zimo Sun, “Quantum de Sitter hori-
zon entropy from quasicanonical bulk, edge, sphere and topological string partition functions,”
JHEP 01, 088 (2022), arXiv:2009.12464 [hep-th]

[3] Venkatesa Chandrasekaran, Roberto Longo, Geoff Penington, and Edward Witten, “An algebra
of observables for de Sitter space,” JHEP 02, 082 (2023), arXiv:2206.10780 [hep-th]

[4] Joseph Polchinski, “The phase of the sum over spheres,” Phys. Lett. B 219, 251–257 (1989)

[5] G. W. Gibbons, S. W. Hawking, and M. J. Perry, “Path Integrals and the Indefiniteness of the
Gravitational Action,” Nucl. Phys. B 138, 141–150 (1978)

[6] Pawel O. Mazur and Emil Mottola, “Absence of phase in the sum over spheres,” (6 1989)

[7] M. V. Berry, “Stokes’ phenomenon, smoothing a victorian discontinuity,” Publications

Mathématiques de l’IHÉS 68, 211–221 (1988), http://www.numdam.org/item/PMIHES_1988_
_68__211_0/

[8] Batoul Banihashemi and Ted Jacobson, “Thermodynamic ensembles with cosmological hori-
zons,” JHEP 07, 042 (2022), arXiv:2204.05324 [hep-th]

[9] Eva Silverstein and Gonzalo Torroba, “Timelike-bounded dS4 holography from a solvable sector
of the T 2 deformation,” (9 2024), arXiv:2409.08709 [hep-th]

[10] Alexei Kitaev and S. Josephine Suh, “The soft mode in the Sachdev-Ye-Kitaev model and its
gravity dual,” JHEP 05, 183 (2018), arXiv:1711.08467 [hep-th]

[11] Zhenbin Yang, “The Quantum Gravity Dynamics of Near Extremal Black Holes,” JHEP 05,
205 (2019), arXiv:1809.08647 [hep-th]

[12] Sidney Coleman, Aspects of Symmetry: Selected Erice Lectures (Cambridge University Press,
Cambridge, U.K., 1985) ISBN 978-0-521-31827-3

[13] Anders Andreassen, David Farhi, William Frost, and Matthew D. Schwartz, “Direct Approach
to Quantum Tunneling,” Phys. Rev. Lett. 117, 231601 (2016), arXiv:1602.01102 [hep-th]

15