intellecton/papers/project_paper_6_holographic/references/MaldacenaStanford2016.txt

Comments on the Sachdev-Ye-Kitaev model

Juan Maldacena and Douglas Stanford

Institute for Advanced Study, Princeton, NJ 08540, USA

Abstract

We study a quantum mechanical model proposed by Sachdev, Ye and Kitaev. The
model consists of N Majorana fermions with random interactions of a few fermions
at a time. It it tractable in the large N limit, where the classical variable is a bilocal
fermion bilinear. The model becomes strongly interacting at low energies where it
develops an emergent conformal symmetry. We study two and four point functions
of the fundamental fermions. This provides the spectrum of physical excitations for
the bilocal field.
The emergent conformal symmetry is a reparametrization symmetry, which is
spontaneously broken to SL(2, R), leading to zero modes. These zero modes are
lifted by a small residual explicit breaking, which produces an enhanced contribution
to the four point function. This contribution displays a maximal Lyapunov expo-
nent in the chaos region (out of time ordered correlator). We expect these features
to be universal properties of large N quantum mechanics systems with emergent
reparametrization symmetry.
This article is largely based on talks given by Kitaev [1], which motivated us to
work out the details of the ideas described there.

arXiv:1604.07818v1  [hep-th]  26 Apr 2016


Contents

1
Introduction
2
1.1
Organization of the paper and summary of results . . . . . . . . . . . . . .
4

2
Two point functions
8
2.1
The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8

2.2
Summing the leading order diagrams . . . . . . . . . . . . . . . . . . . . .
8

2.3
The conformal limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10

2.4
Large q limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11

2.5
q = 2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12

2.6
Computing the entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13

2.7
Correction to the conformal propagator
. . . . . . . . . . . . . . . . . . .
15

3
Four point functions
15
3.1
The ladder diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16

3.2
Using conformal symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . .
17

3.2.1
The four point function as a function of the cross ratio . . . . . . .
19

3.2.2
Eigenfunctions of the casimir
. . . . . . . . . . . . . . . . . . . . .
20

3.2.3
The eigenvalues of the kernel kc(h)
. . . . . . . . . . . . . . . . . .
22

3.2.4
The inner products ⟨Ψh, Ψh⟩ and ⟨Ψh, F0⟩
. . . . . . . . . . . . . .
23

3.2.5
The sum of all ladders . . . . . . . . . . . . . . . . . . . . . . . . .
25

3.2.6
Operators of the model . . . . . . . . . . . . . . . . . . . . . . . . .
26

3.2.7
Analytic continuation to the chaos region . . . . . . . . . . . . . . .
28

3.3
Proper treatment of the h = 2 subspace . . . . . . . . . . . . . . . . . . . .
30

3.3.1
The h = 2 eigenfunctions and reparameterizations . . . . . . . . . .
31

3.3.2
The shift in the eigenvalues
. . . . . . . . . . . . . . . . . . . . . .
32

3.3.3
The enhanced h = 2 contribution . . . . . . . . . . . . . . . . . . .
35

3.3.4
Other terms from the h = 2 subspace . . . . . . . . . . . . . . . . .
38

3.4
More detail on the q = ∞ four point function
. . . . . . . . . . . . . . . .
39

3.5
Summary of the four point function . . . . . . . . . . . . . . . . . . . . . .
41

3.6
The chaos exponent at finite coupling . . . . . . . . . . . . . . . . . . . . .
42

3.6.1
The retarded kernel . . . . . . . . . . . . . . . . . . . . . . . . . . .
42

3.6.2
Large q
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43

3.6.3
General q
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43

4
The effective theory of reparameterizations
45

5
The density of states and the free energy
49

6
Towards a bulk interpretation
51
6.1
Comments on kinematic space.
. . . . . . . . . . . . . . . . . . . . . . . .
53

6.2
The fermions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54

1


6.3
Scrambling for near extremal black holes and its stringy corrections . . . .
55

7
Brief Conclusions
57

A The Schwinger-Dyson equations and the kernel
58

B The kernel as a function of cross ratios
59

C Representing F0 in terms of Ψh
60

D Writing Ψh(χ) in terms of Ψh,n(θ1, θ2)
60

E Direct approach to the shift in eigenvalue
61

F The first order change in h = 2 eigenvectors
63

G Numerical solution of the SD equations
65

H A model without the reparametrization symmetry
67

I
Further coments on Kinematic space
70
I.0.1
The kinematic space in the full model at q = ∞ . . . . . . . . . . .
71

1
Introduction

Studies of holography have been hampered by the lack of a simple solvable model that can
capture features of Einstein gravity. The simplest model, which is a single matrix quantum
mechanics, does not appear to lead to black holes [2] (see [3] for a review). N = 4 super
Yang Mills at strong ’t Hooft coupling certainly leads to black holes, and exact results are
known at large N for many anomalous dimensions and some vacuum correlation functions,
but at finite temperature the theory is difficult to study.
A system that reproduces some of the dynamics of black holes should be interacting,
but we might hope for a model with interactions that are simple enough that it is still
reasonable solvable.
Kitaev has proposed to study a quantum mechanical model of N Majorana fermions
interacting with random interactions [1]. It is a simple variant of a model introduced
by Sachdev and Ye [4], which was first discussed in relation to holography in [5]. The
Hamiltonian of [1] is simply

H =
�

iklm
jiklmψiψkψlψm
(1.1)

where the couplings jiklm are taken randomly from a Gaussian distribution with zero mean
and a width of order J /N 3/2.

2


One interesting feature of this model is that it develops an approximate conformal
symmetry in the infrared. Understanding how to deal with quantum mechanical theories
that develop such a conformal symmetry seems very important for both condensed matter
physics and gravity.
One naively expects a full Virasoro symmetry.
However, in the
model, the symmetry is both explicitly as well as spontaneously broken, so we end up with
“nearly conformal quantum mechanics,” or NCFT1. (We propose to use the term NCFT1
to denote systems that have one time dimension which are nearly invariant under a full
reparametrization (or Virasoro) symmetry 1.) The same situation arises in gravity, when
we consider very near extremal black holes. These are black holes that develop a nearly
AdS2 region, which we can call NAdS2, see [7] for a recent discussion. It is well known that
purely AdS2 gravity is not consistent, except for the ground states. So the right setting in
which to study holography for near extremal black holes is NAdS2/NCFT1.
Besides this structural similarity, it was noted in [8, 1] that the out of time order
correlators of the Sachdev-Ye-Kitaev model (1.1) (SYK) grow in a manner that reflects
an underlying chaotic dynamics. At relatively low energies this growth matches the one
expected in a theory of gravity [9, 10, 11], which saturates the chaos bound [12].
In this paper we study this model a bit further. We start by summarizing the com-
putation of the two point functions [4, 13] in the large N limit, following Sachdev, Ye,
Parcollet, and Georges. We will discuss this in a variant of the model where the interac-
tion involves q fermions at a time [1]. We will further show that the equations simplify
considerably in the large q limit. This allows us to connect analytically the free UV theory
to the interacting and nearly conformal IR theory. Further recent work in this or similar
models includes [14, 15, 16, 17, 7, 18, 19]. See also [20, 21] for a string motivated model
with disorder.
We then derive an explicit integral expression for the four point function in the infrared
limit.
This problem was also considered in [19].
The four point function is actually
infinite in the strict conformal limit, due to Nambu-Goldstone bosons associated to the
spontaneously broken reparameterization invariance. To remove the infinity we have to
take into account the explicit breaking of this symmetry, which lifts these modes by a small
amount. We expect that this should be a universal feature of large, but finite, entropy
NCFT1 systems. Namely, the systems cannot realize the conformal symmetry exactly2,
and that the small explicit breaking leads to a universal contribution that dominates the
four point function and saturates the chaos bound.3 In particular, AdS2 dilaton gravity

1This should be contrasted to what is usually called “conformal quantum mechanics”, such as [6], which
are only invariant under SL(2, R).

2The argument in [22] shows that an exact SL(2, R) symmetry is incompatible with a thermofield
interpretation with a finite number of states. Of course, in gravity the exact SL(2, R) symmetry is broken
by the presence of a dilaton field, see [7] for recent discussion.

3In 1 +1 dimensional CFT the conformal symmetry is also spontaneously broken (recall that L−2|0⟩ ̸=
0), but it is not explicitly broken. In that case we also have a universal (stress tensor) contribution to the
four point function. By itself this piece saturates the chaos bound [23, 24, 25], but only in special theories
does it dominate.

3


is an example with the same explicit breaking [26], leading to the same dominant term in
the four point function.
In addition to this term, the SYK four point function contains subleading pieces that
are finite in the low temperature limit. These contain information about the composite
operators that appear in the operator product expansion of �

i ψi(τ)ψi(0).
These get
anomalous dimensions at leading order in N and seem analogous to the single trace oper-
ators of the usual gauge theory examples of holography. One finds a tower of states with
an approximately integer spacing. This tower of states is reminiscent of the one appearing
in large N O(N) models, where we have one state for each spin. Here we get a similar
structure, but with dimensions which have O(1) corrections relative to the dimensions in
the free theory. This suggests that the bulk theory contains low-tension strings. These
extra states do not compete with the dilaton gravity piece, even though the strings are
light, simply because of the enhancement of gravity in NAdS2.
Much of the analysis in this paper, including the ladder diagrams, the spectrum of
the kernel kc(h), and the effective theory of reparameterizations, is simply what Kitaev
presented in his talks [1], and we are thankful to him for several further explanations.

1.1
Organization of the paper and summary of results

The article might seem a bit technical in some parts, so we will summarize below what
is done in various sections. The reader might want to jump directly the the sections that
look most interesting to him/her.
In section two we review the large N structure of the theory. The model has one
dimensionful parameter J , with dimensions of energy, which characterizes the size of the
interaction terms in the Hamiltonian. This implies that the interaction is relevant and
becomes strong at low energies. For large N the diagrams have a simple structure that is
reminiscent of the one for large N O(N) theories (see also the discussion in [18]). There is a
bilocal field ˜G(τ1, τ2) depending on two times which becomes classical in the large N limit.
On the classical solution, G, this field is equal to the two point function of the fermions
G(τ1, τ2) =
1
N
�N
i=1⟨ψi(τ1)ψi(τ2)⟩. The classical equation for G is non-local in time but
it can be solved numerically. G can be inserted in the action to compute the partition
function. We also show that in the variant of the model where q fermions interact at a
time, then the large q limit becomes analytically tractable and one can solve the classical
equations for any value of the coupling. Another simple solvable limit is the case q = 2.
In that case, the Hamiltonian has the form H = i �

kl jklψkψl which is a random mass-
like term. This can be diagonalized and we get a spectrum of masses, or energies, given
by the usual semi-circle law distribution for random matrices. This particular example
is integrable and some properties are different than the one for the generic q case. In
particular, we find that there is no exponentially growing contribution to the out of time
order four point function.
At low energies the model simplifies further due to the emergence of a conformal sym-
metry. In one dimension the conformal group is the same as the group of all reparame-

4


terizations. One can see this symmetry explicitly in both the low energy action, or the
low energy equations for the bilocal fields. One might expect a theory that has a full
reparameterization symmetry to be topological. This is not the case here because the
reparameterization symmetry is spontaneously broken down to an SL(2, R) subgroup. In
other words, the bilocal function G(τ1, τ2) = G(τ1 − τ2) becomes Gc ∝ τ −2∆
12
for large
values of J τ12 (with τ12 = τ1 − τ2). The partition function displays a zero temperature
entropy of order N. In addition, there is a finite temperature entropy which is linear in
the temperature, proportional to N/(βJ ). Generically the model is expected to have a
single ground state, but here we are considering temperatures that are fixed in the large
N limit. This means that we are accessing an exponentially large number of states.
In section three we discuss general features of the four point function of the fermions
⟨ψiψiψjψj⟩. This can also be viewed as a two point function of the bilocal fields. The final
form for the leading 1/N piece in for the four point function is displayed in (3.149).
This computation of the four point function is a bit technical and, for this reason, this
section is rather long. The diagrams that contribute have the form of ladder diagrams.
Therefore, they can be summed by defining a kernel K that corresponds to adding a
rung to a ladder. Then the full ladder has a form proportional to
1

1−KF0 where F0 is
a diagram with no rungs. This is conceptually easy. However, it is tricky to invert the
kernel since one has to understand in more detail the space of functions where it is acting.
Fortunately the problem partially simplifies at low energies due to the unbroken SL(2, R)
symmetry. This symmetry can be used to diagonalize the kernel and also to describe the
space of functions we should sum over. This leads to a relatively explicit expression for
the four point function in terms of a sum over intermediate states (3.88), (3.90), once we
exclude the Goldstone bosons which need to be treated separately. We can read off the
spectrum of operators that appear in the OPE expansion of two fermions.The spectrum
is given by the solutions hm to the equation kc(hm) = 1, with kc(h) in (3.73). We can
vaguely view this tower of operators as ψi∂1+2mψi. We say “vaguely” because the proper
dimensions we obtain from the above procedure display an order one correction from the
naively expected values (which would be 2∆ + 1 + 2m). This is an important clue for a
possible bulk interpretation. It is saying that the fermions cannot be associated to weakly
interacting particles in the bulk. Their interactions would have to be of order 1 rather
than 1/N.
We then give a proper treatment for the Goldstone modes that have Kc = 1 in the
conformal limit. These arise from reparametrizations of the conformal solution, Gc. These
fluctuations have zero action in the conformal limit, but get a non-zero action when we
take into account the leading corrections to the conformal answers. We first take a direct
approach and compute the leading correction to the classical solution G away from the
conformal limit G = Gc + δG. It turns out that the leading correction involves an extra
factor of 1/J . One can then proceed to compute the variation of the kernel K away from
the conformal limit K = Kc + δK. We then evaluate δK on reparametrizations of the
conformal solution δϵGc, where ϵ(τ) is an infinitesimal reparametrization. We get a non-
zero answer which can then be used to compute the four point function. Since δK ends

5


up in the denominator in the expression for the four point function, we get an enhanced
contribution with an additional factor of (βJ ) as compared to the conformal answer,
which is independent of βJ . This enhanced contribution is not conformally covariant.
However, it has a very simple form in the OPE limit which can be understood as follows.
The OPE gives rise to an energy operator of the model, which has quadratic fluctuations
⟨(δE)2⟩ in the thermal ensemble. These fluctuations are governed by the specific heat of
the system, which again is non-zero once we take into account the effects of the breaking
of the conformal symmetry. We also consider the contribution of these Goldstone modes
in the chaos limit. There they give the dominant term (βJ /N)e
2π
β t that saturates the
bound. The reparametrization symmetry of the model is essential to obtain this Pseudo
Goldstone boson.
In appendix H we discuss a model that has a low energy SL(2, R)
symmetry but without the conformal symmetry, by thinking of the couplings as dynamical
with an SL(2, R) invariant correlation function. In this case there is no Pseudo-Goldstone
mode, the low energy physics is SL(2, R) invariant and the chaos exponent is less than
maximal.
In section four, we give a discussion of the four point function from the perspective
of the large N effective action for the bilocal field ˜G, see also [18]. The intermediate states
that appear can be understood from the on-shell condition for fluctuations of this field.
The enhanced non-conformal part of the four point function arises from the functional
integral over ˜G configurations that are reparameterizations of the infrared saddle point
solution.
We give a simple effective field theory argument showing that the effective
action is given by the Schwarzian derivative, {f(τ), τ}, of the reparametrization (4.176),
with a coefficient of order (βJ )−1 [1]. This action constitutes an explicit breaking of the
conformal symmetry. It can be used to derive the enhanced contribution mentioned above,
and also to compute the specific heat.
In section five we discuss some features of the spectrum of the model. We start
by presenting a numerical computation of the spectrum for the case of N = 32. The
spectrum in this case looks reminiscent to that of a random matrix and, as expected, is
statistically symmetric under H → −H since the random couplings can be positive as well
as negative. At low temperature one is interested in the region near the bottom end of
the spectral distribution. We then look at the expression for the free energy at leading
order in N, which has the low-temperature expansion log Z = −βE0 + S0 +
c
2β where all
terms are of order N. The first term is the ground state energy, which is not interesting.
The second is the zero temperature entropy. The third, with the specific heat c ∝ N/J ,
arises from the breaking of the conformal symmetry and can be computed in terms of
the Schwarzian action for the reparametrizations, after noticing that we can change the
temperature by making a reparametrization of the Euclidean circle (or, equivalently, we
can go from the circle to the line by a reparametrization). We further consider the N 0

correction to log Z. This arises from the one loop correction to the effective action for
the bilocal fields. All the modes that have a non-vanishing action give a contribution
just to E0 and S0, since they are J independent (up to UV contributions to E0). The
reparametrization modes give a term that contributes a logarithm to the free energy [27],

6


specifically − 3

2 log(βJ ). This additional contribution has an interesting effect. It implies
that if we compute the spectral density ρ(E) by inverse Laplace transforming Z(β) we
obtain that ρ(E) ∝ J −1eS0+√

2c(E−E0), with no prefactor powers of (E − E0), in the
regime where we can trust the computation.
In section six we comment on the possible bulk interpretation. The enhanced non-
conformal contribution agrees with the four point function one expects in a theory of
dilaton gravity [28, 26].
This contribution is completely general in any situation with
a near extremal black hole with a NAdS2 region and it follows by the same pattern of
spontaneous plus explicit breaking of the conformal symmetry [26] (see [25] for a similar
discussion in the AdS3/CFT2 context). Therefore the information about other possible
bulk states comes from the contribution that is J independent and finite in the conformal
limit. These are the states that appeared in the OPE expansion of two fermions. This
looks like a single Regge trajectory with dimensions that are linearly increasing with “spin”,
though spin is hard to define in two dimensions. This implies that a dual description would
involve a string with low tension, with ls ∼ RAdS.
Part of our motivation to study this model arose from the observation that the four
point function was saturating the chaos bound, which is a necessary condition for a gravity
dual. It was shown in [11] that stringy corrections to the chaos exponent involve a factor
of 1 − l2
s/R2 where R is a suitable distance scale. This suggested that the condition might
also be sufficient to exclude models with light strings. However, examining the scale R
carefully, it is possible to show that for near extremal black holes this correction has the
form
�
1 −
l2
s

R2
AdS

(S−S0)

S0

�
, where S is the entropy and S0 is the zero temperature entropy.
Since the ratio of entropies is much less than one, we see that stringy corrections to the
Lyapunov exponent are very suppressed, suggesting that even in cases with ls ∼ RAdS2 we
could have a Lyapunov exponent close to the gravity value. In other words, for this case
a nearly saturated Lyapunov exponent is not a guarantee of a high string tension. And
indeed, in the SYK model we seem to have a low string tension. A related perspective on
this is the following: the four point function saturates the bound because it is dominated
by the universal “gravity” piece coming from the zero modes discussed above. This turns
out to be enhanced by a factor of S0/(S − S0).
Relative to this piece, the “stringy”
contributions to the four point function are small, so they have only a mild effect on the
chaos exponent.
We also comment on the bulk interpretation of the fermion fields. We speculate that
we should not have N fermions in the bulk, but rather one fermion with a string attached
to the boundary.
Finally, we note that the bilocal field can be viewed as a field in one more dimension.
At low energies the extra dimension defined in this way has a metric characterized by the
conformal group and can be viewed as a dS2 (or AdS2) space in accordance with recent
discussions of kinematic space [29, 30].
This follows simply from the structure of the
conformal group. Some terms in the action can be viewed as local terms in this space, but
others have a non-local expression.

7


In the appendices we give some more details on the computations.

2
Two point functions

2.1
The model

We consider a quantum mechanical model with N Majorana fermions with random interac-
tions involving q of these fermions at a time, where q is an even number. The Hamiltonian
is

H
=
(i)
q
2
�

1≤i1<i2<···<iq≤N
ji1i2···iqψi1ψi2 · · · ψiq
(2.2)

⟨j2
i1···iq⟩ = J2(q−1)!

N q−1
= 2q−1

q
J 2(q−1)!

N q−1
(no sum)
(2.3)

We take each coefficient to be a real variable drawn from a random gaussian distribution.
(2.3) indicates the variance of the distribution. It is characterized by a dimension one
parameter J, (or J , which is defined with an extra factor that makes the model more
uniform in q) which we take to be the same for all coefficients. The numerical factors, and
factors of N, are introduced to simplify the large N limit. A factor of i is necessary to
make the Hamiltonian Hermitian when q = 2 mod(4). This i means that the system is
not time reversal symmetric for odd q/2. Thus, if we restrict to time reversal symmetric
interactions the model with q = 4 represents the dominant interactions at low energy.
The others involve some degree of tuning. We assume that the system does not have a
spin glass transition [31] and we work to leading order in the 1/N expansion. Though
the model generically has a unique ground state, we work at temperatures which are fixed
in the large N expansion, implying that we access an exponentially large number of low
energy states, of order O(eαN), α > 0.

2.2
Summing the leading order diagrams

We will work first in Euclidean space. It is useful to define the Euclidean propagator as

G(τ) ≡ ⟨T(ψ(τ)ψ(0))⟩ = ⟨ψ(τ)ψ(0)⟩θ(τ) − ⟨ψ(0)ψ(τ)⟩θ(−τ)
(2.4)

For a free Majorana fermion this is very simple

Gfree(τ) = 1

2sgn(τ) ,
Gfree(ω) = − 1

iω =
�
dteiωτGfree(τ)
(2.5)

Gfree has the same expression at finite temperature, with τ ∼ τ + β. Notice that it is
correctly antiperiodic as τ → τ + β. . These equations also normalize the fermion fields
appearing in the interaction (2.2).
Recall that the free Majorana fermions are simply
described by operators that are essentially N dimensional Dirac γ matrices, see e.g. [17].

8


Using this free propagator we can then compute corrections due to the interaction. Let
us look at the first correction to the two point function, shown in figure 1. This arises
by bringing down two insertions of the interaction Hamiltonian and then averaging with
respect to the disorder. The disorder average is represented by a dotted line in figure 1. As
pointed out in [19], we can sometimes reproduce similar diagrams by considering ji1,··· ,iq to
be a dynamical field. Here we will stick to the disordered model. The disorder average links
the indices appearing in the two interaction Hamiltonians and we end up with a correction
that scales as J2 relative to the free two point function, with no additional factors of N,
since we get (q − 1) factors of N from the sum over the indices of the intermediate lines.

+
+
+
+

Figure 1: Diagrams representing corrections to the two point function, for the q = 4 case.
The free two point function is given by the straight line. The first correction involves also
an average over disorder, which is represented by a dashed line. We have also indicated a
couple more diagrams that also contribute at leading order in N.

=
+
+
+

=

Figure 2: Equations that define the summation of the leading large N contributions, for
the q = 4 case. The solid circle represents the one particle irreducible contributions. The
dotted circle represents the full two point function. This is a graphical representation of
the equations in (2.6).

Besides this first diagram, there are many more “iterated watermelon” diagrams that
contribute at leading order in N. Two more are shown in figure 1. The set of diagrams
is sufficiently simple that they can be summed by writing self consistency equations for
the sum. First, it is convenient to define a self energy, Σ(τ, τ ′), which includes all the one
particle irreducible contributions to the propagator. By translation symmetry, Σ(τ, τ ′) =

9


Σ(τ − τ ′) and we can write the full two point function, and the definition of Σ as

1

G(ω) = −iω − Σ(ω) ,
Σ(τ) = J2 [G(τ)]q−1
(2.6)

Notice that the first equation is written in frequency space while the second in the original
(Euclidean) time coordinate. Here we have assumed translation symmetry. The possible
values of the frequency depend on whether we are at β = ∞, where it is continuous, or
at finite β where we have ω = 2π

β (n + 1

2). When we talk about zero temperature, we are
imagining taking the large N limit first and then the zero temperature limit.
As a side comment, note that we could consider a model with a Hamiltonian which is
a sum of terms with various q’s, and with random couplings with their own variance Jq.
The large N equations for such models would be very similar except that the right hand
side of (2.6) would be replaced by Σ = �

q J2
q [G(τ)]q−1. But we did not find any good use
for this.

2.3
The conformal limit

At strong coupling, the first equation in (2.6) can be approximated by ignoring the first
term on the right hand side. It is convenient to write these approximate equations as
�
dτ ′G(τ, τ ′)Σ(τ ′, τ ′′) = −δ(τ − τ ′′) ,
Σ(τ, τ ′) = J2 [G(τ, τ ′)]q−1
(2.7)

Written in this form, they are invariant under reparametrizations,

G(τ, τ ′) → [f ′(τ)f ′(τ ′)]∆ G(f(τ), f(τ ′)) ,
Σ(τ, τ ′) → [f ′(τ)f ′(τ ′)]∆(q−1) Σ(f(τ), f(τ ′))
(2.8)
provided that ∆ = 1/q.
We can then use an ansatz of the form

Gc(τ) =
b

|τ|2∆sgn(τ),
or
Gc(τ) = b

�
π

β sin πτ

β

�2∆
sgn(τ)
(2.9)

where we have given also the finite temperature version, which follows from (2.8) with
f(τ) = tan τπ

β .
We can determine b by inserting these expressions into the simplified
equations and obtain

J2bqπ =
�1

2 − ∆
�
tan π∆ ,
∆ = 1

q
(2.10)

We will use ∆ and 1/q interchangeably below. To derive the first equation here, it is
convenient to use the Fourier transform
� ∞

−∞
dτeiωτ sgn(τ)

|τ|2∆ = i 21−2∆√π Γ(1 − ∆)

Γ( 1

2 + ∆)|ω|2∆−1sgn(w)
(2.11)

10


From (2.9) it is possible also to compute the Lorentzian time versions by setting τ = it.
Since the correlator is not analytic at τ = 0 it is important to know whether we are doing
the analytic continuation of the τ > 0 or the τ < 0 Euclidean expressions.
The two
choices give different choices of ordering of the Lorentzian correlator. For example, the
continuation of the τ > 0 form of the Euclidean correlator gives

⟨ψ(t)ψ(0)⟩ = Gc,E(it + ϵ) = b
e−iπ∆

(t − iϵ)2∆
(2.12)

where we summarized the fact that we continue from τ > 0 by the t → t − iϵ prescription.
This equation is valid for any sign of t. Of course the other ordering can be obtained
by continuing from the τ < 0 version. We can also get the finite temperature version by
replacing (t − iϵ) → β

π sinh [π(t − iϵ)/β] in (2.12).
It is sometimes also convenient to introduce the retarded propagator defined as

Gc,R(t) ≡ ⟨ψ(t)ψ(0) + ψ(0)ψ(t)⟩ θ(t) = 2b cos(π∆)

�
π

β sinh πt

β

�2∆
θ(t)
(2.13)

where θ(t) is the step function. Of course, (2.13) also shows that the dimension ∆ sets the
quasinormal mode frequencies as ωn = −i 2π

β (∆ + n).
Here we have given the conformal limit of the expressions. For large βJ, it is possible to
solve the equations (2.6) numerically to obtain expressions that smoothly interpolate be-
tween the free UV limit and the infrared expressions given above, see figure 15 in appendix
G. In addition, in the next subsection we show how to do this interpolation analytically
in the large q limit.

2.4
Large q limit

One convenient feature of the model in (2.2) is the fact that it simplifies considerably for
large q.4 We can write

G(τ) = 1

2sgn(t)
�
1 + 1

qg(τ) + · · ·
�
,
Σ(τ) = J221−qsgn(τ)eg(τ)(1 + · · · )
(2.14)

where the dots involve higher order terms in the 1/q expansion. We will work in the regime
where g(τ) is of order one. In this regime we can approximate

1

G(ω) =
1

− 1

iω + [sgn×g](ω)

2q
= −iω + ω2[sgn × g](ω)

2q
= −iω − Σ(ω)
(2.15)

where in the first equality we fourier transformed the first equation in (2.14) and we
expanded in powers of 1/q in the second equality, keeping only the first nontrivial term.

4We are grateful to S.H. Shenker for discussions on this point.

11


Comparing this expression for Σ with the one in (2.14) we get the equation

∂2
t [sgn(τ)g(τ)] = 2J 2sgn(τ)eg(τ) ,
J ≡ √q J

2
q−1

2
(2.16)

This equation determines g(τ). It is well defined in the large q limit, when we scale J so
that J is kept fixed as q → ∞. Of course, since J is dimensionful, we can always go to
some value of τ where this equation will be valid. We are interested in a solution with
g(τ = 0) = 0. In other words, at short distances we should recover the free fermion result.
The derivation of this equation is valid both for zero temperature and finite temperature.
The general solution is

eg(τ) = c2

J 2
1

sin(c(|τ| + τ0))2
(2.17)

We can now impose the boundary conditions g(0) = g(β) = 0 to obtain

eg(τ)
=




cos πv

2

cos
�
πv( 1

2 − |t|

β )
�





2

(2.18)

βJ
=
πv

cos πv

2
.
(2.19)

The second equation determines the parameter v, which ranges from zero to one as βJ
ranges from zero to infinity. It is also possible to take the β = ∞ limit of the above
expressions to obtain

eg(τ) =
1

(|t|J + 1)2
(2.20)

Note that these results imply that Σ changes more rapidly than G. In fact, G is almost
constant, and almost equal to 1

2sgn(τ), when Σ is changing to its IR value.

2.5
q = 2

Another solvable example is the case of q = 2. In this case we can solve (2.6) as

G(ω) = −
2

iω + i sgn(ω)
√

4J2 + ω2
(2.21)

This is the same as the one studied in [32, 33, 20]. For positive euclidean time we get

G(τ)
=
sgn(τ)
� π

0

dθ
π cos2 θe−2J|τ| sin θ
(2.22)

=
1

πJτ −
1

4π(Jτ)3 + · · · ,
Jτ ≫ 1
(2.23)

For this particular case, we can simply diagonalize the Hamiltonian (2.2), since it is
quadratic. We get a set of fermionic oscillators with some masses. The masses have a

12


semicircle law distribution, since we are diagonalizing a random mass matrix. Near zero
frequencies the distribution is constant and we get the same as what we expect for a 1 + 1
dimensional fermion field (from θ ∼ 0, π above). The spacing between the frequencies goes
like 1/N, so this fermion is on a large circle. In this sense this example is a bit trivial since
it is the same as free fermions. However, it is useful to view it as an extreme example of
the more interesting models with q > 2. Therefore, in this model we indeed get a fermion
in an extra dimension. However, note that we get a single fermion, not N fermions in the
extra dimension. We can also simply obtain the finite temperature expression for the two
point function by summing over images in the zero temperature answer

Gβ(τ) =

∞
�

m=−∞
Gβ=∞(τ + βm)(−1)m =
� π

0

dθ
π cos2 θ
cosh[( τ

β − 1

2)2Jβ sin θ]

cosh(Jβ sin θ)
(2.24)

2.6
Computing the entropy

It is possible to write the original partition function of the theory as a functional integral
of the form [1, 14]

e−βF =
�
D ˜G˜Σ exp
�
N
�
log Pf(∂t − ˜Σ) − 1

2

�
dτ1dτ2

�
˜Σ(τ1, τ2) ˜G(τ1, τ2) − J2

q
˜G(τ1, τ2)q
���

(2.25)
It can be checked that the classical equations obtained from this reproduce the equations
in (2.6), when we vary with respect to ˜G and ˜Σ independently. Here the tildes remind us
that we are are thinking about the integration variables, while G, Σ without tildes are the
solutions of the classical equations from (2.26), obeying (2.6). Substituting those solutions
into (2.25) we get the leading large N approximation to the free energy:

−βF/N = log Pf(∂t − Σ) − 1

2

�
dτ1dτ2

�
Σ(τ1, τ2)G(τ1, τ2) − J2

q G(τ1, τ2)q
�
(2.26)

In the q = ∞ model we know the full solutions for G and Σ, so we can insert them
in (2.26) to obtain the free energy. In order to avoid evaluating the Pfaffian term, it is
convenient to take a derivative with respect to J∂J of the free energy (2.26).
Due to
the fact that G and Σ obey the equations of motion, the only contributing term is the
derivative of the explicit dependence on J, so that we obtain

J∂J(−βF/N) = J2β

q

� β

0
dτG(τ)q = −β

q ∂τG|τ→0+ = −βE
(2.27)

where we have used the equations (2.6) in position space. Since the partition function
only depends on the combination βJ, then J∂J is the same as β∂β. Therefore the above
expression gives us the energy.
As q → ∞, we can insert the solution (2.18) into (2.27). We can also use the equation
(2.16) to do the integral. Furthermore we can turn J∂J → J ∂J and use (2.19) to turn it

13


into a derivative with respect to v, always keeping q and β fixed. This gives

J ∂J (−βF/N)
=
v

1 + πv

2 tan πv

2
∂v(−βF/N)
(2.28)

=
β
4q2

� β

0
dτ2J 2eg(τ) = β

4q22(−g′(0)) = πv

q2 tan πv

2
(2.29)

−βF/N
=
1
2 log 2 + 1

q2πv
�
tan
�πv

2

�
− πv

4

�

(2.30)

where we fixed the integration constant using that for J → 0 we should recover the free
value, which is simply the log of the total dimension of the Hilbert space. The expansion
around weak coupling is simply an expansion in powers of v2, which translates into an
expansion in powers of (βJ )2, as expected. On the other hand, at strong coupling we can
use (2.19) to find

v = 1 −
2
βJ +
4

(βJ )2 − (24 + π2)

3(βJ )3 + · · ·
(2.31)

Then, the term of order 1/q2 in (2.30) behaves as

1
q2

�
2

1 − v − (2 + π2

4 ) + π2

3 (1 − v) + · · ·
�
= 1

q2

�
(βJ ) − π2

4 +
π2

2(βJ ) + · · ·
�
(2.32)

Here the first term can be interpreted as a correction to the ground state energy. The
second term is a correction to the zero temperature entropy, to which the 1

2 log 2 term in
(2.30) also contributes. Finally the third term is a temperature dependent correction to
the entropy, or near extremal entropy, which goes like T for low temperature.
The temperature independent piece can be compared with the result obtained in [1]
for general q (see the earlier [31] for the q = 4 case using the Sachdev-Ye model)

S0
N = 1

2 log 2 −
� ∆

0
dxπ(1

2 − x) tan πx ∼ 1

2 log 2 − π2

4q2 + · · ·
(2.33)

where the last expression is the approximate answer for large q, which agrees with the
temperature independent pice of (2.30) using (2.32).
It is also possible to compute the free energy at q = 2. Directly from the free fermion
picture, and subtracting the ground state energy, we find

log Z/N
=
� π

0

dθ
π cos2 θ log
�
1 + e−2Jβ sin θ�

∼
π

12βJ + · · · ,
for
βJ ≫ 1
(2.34)

We see that at small temperatures the entropy vanishes, in agreement with the first equality
in (2.33) with ∆ → 1

2. We can also see that for large temperatures this reproduces the
value S/N = 1

2 log 2.

14


We will later show that for general q the expression of the free energy has the form

log Z = −βE0 + S0 + c

2β
(2.35)

plus higher orders in 1/β. Here E0 is the ground state energy, S0 is the zero temperature
entropy and c/β is the specific heat. E0, S0 and c are all of order N. The exact large
N free energy can be computed numerically for general q. Appendix G contains some
discussion of this.

2.7
Correction to the conformal propagator

It is also interesting to consider the leading correction to the conformal two point function.
For large q the conformal answer is

Gc = b sgn(τ)

|τ|
2
q
= 1

2
1

|J τ|2∆ ,
J 2(2b)q = 1
(2.36)

Using (2.14) and (2.20), we find the leading correction

G(τ) = Gc(τ)
�
1 − 2

q
1

J |τ| + · · ·
�
.
(2.37)

At finite temperature, we use (2.18) to find

G(τ) = Gc(τ)

�

1 − 2

q
1
βJ

�

2 + π − 2π|τ|/β

tan π|τ|

β

�

+ · · ·

�

.
(2.38)

On the other hand, for q = 2 we see from (2.23) that the order 1/J correction vanishes.
We will later discuss general values of q.

3
Four point functions

In this section, we analyze the leading 1/N piece of the four point function, at strong
coupling βJ ≫ 1. In any correlation function, the average over disorder ji1,...,iq will give
zero unless the indices of the fermions are equal in pairs. This means that the most general
nonzero four point function is

⟨ψi(τ1)ψi(τ2)ψj(τ3)ψj(τ4)⟩.
(3.39)

We will consider the case in which we average over i, j. (The pure i = j and i ̸= j cases
are related in a simple way.) The averaged correlator

1
N 2

N
�

i,j=1
⟨T(ψi(τ1)ψi(τ2)ψj(τ3)ψj(τ4))⟩ = G(τ12)G(τ34) + 1

N F(τ1, ..., τ4) + · · ·
(3.40)

has a disconnected piece given by a contraction with the dressed propagators, plus a power
series in 1/N. We will analyze the first term in this series, F.

15


+
+
+
+
τ1

τ2

τ3

τ4

Figure 3: Diagrams representing the 1/N term in the index-averaged four point function,
for the q = 4 case. One should also include the diagrams with (τ3 ↔ τ4) and a relative
minus sign. The propagators here are the dressed two point functions discussed above.

=

Figure 4: The (n+1)-rung ladder Fn+1 can be generated from the n-rung ladder by “mul-
tiplication” with the kernel K, shown in blue. We call the vertical propagators a “rung”
and the horizontal ones a “rail”.

3.1
The ladder diagrams

The diagrams that one must sum to compute F are ladder diagrams with any number of
rungs, built from the dressed propagators discussed in the previous section. The first few
diagrams for F are shown in figure 3. We will use Fn to denote the ladder with n rungs,
so that F = �

n Fn. The first diagram, F0, is just a product of propagators

F0(τ1...τ4) = −G(τ13)G(τ24) + G(τ14)G(τ23).
(3.41)

This piece contributes at order 1/N because the propagators set i = j in the sum of (3.40).
The next diagram is a one-rung ladder, where we integrate over the locations of the ends
of the rung:

F1 = J2(q − 1)
�
dτdτ ′�
G(τ1 − τ)G(τ2−τ ′)G(τ−τ ′)q−2G(τ−τ3)G(τ ′−τ4) − (τ3 ↔ τ4)
�
.

(3.42)
In this expression, the factor of (q − 1) comes from the choice of which of the lines coming
out of the interaction vertex should be contracted into a rung, and which should continue
on as the side rail. This diagram also contributes at order 1/N, because the 1/N q−1 scaling
of the product of two couplings multiplies a factor of N q−2 from the sum over (q−2) indices
in the rung loops. One can check that all of the ladder diagrams (and only these!) are
proportional to 1/N.
The standard technique for summing a set of ladder diagrams is to use the fact that they
are generated by multiplication by a kernel K. This is illustrated in figure 4. Explicitly,

Fn+1(τ1, τ2, τ3, τ4) =
�
dτdτ ′ K(τ1, τ2; τ, τ ′)Fn(τ, τ ′, τ3, τ4),
(3.43)

where the kernel is

K(τ1, τ2; τ3, τ4) ≡ −J2(q − 1)G(τ13)G(τ24)G(τ34)q−2.
(3.44)

16


It is convenient to think about the integral transform in (3.43) as a matrix multiplication,
where the first two arguments of K form one index of the matrix, and the last two form
the other index. The sum of all ladder diagrams is then a geometric series that can be
summed by matrix inversion:

F =

∞
�

n=0
Fn =

∞
�

n=0
KnF0 =
1

1 − K F0.
(3.45)

To carry this out, we would like to understand how to diagonalize K. The way we have
defined it, K is not a symmetric operator under (τ1, τ2) ↔ (τ3, τ4).
However, we can
conjugate by a power of the propagator to get a symmetric version

�K(τ1, τ2; τ3, τ4) ≡ |G(τ12)|
q−2

2 K(τ1, τ2; τ3, τ4)|G(τ34)|
2−q

2
(3.46)

= −J2(q − 1)|G(τ12)|
q−2

2 G(τ13)G(τ24)|G(τ34)|
q−2

2 .
(3.47)

This is enough to show that K has a complete set of eigenvectors. We will consider this
kernel as acting on the space of anti-symmetric functions of two arguments, say τ3, τ4. We
will use both �K and K in what follows.

3.2
Using conformal symmetry

So far, what we have said is true for any value of the coupling βJ. In order to proceed
further, we will go to the conformal limit βJ ≫ 1. In this limit we can use the conformal
expressions for Gc(τ) (2.9). It is worth noting that the J dependence in K drops out in the
conformal limit. This is due to the factors of b in the infrared expressions for G (2.9) and
(2.10). In the conformal limit computations on the zero temperature line are equivalent
to computations on the finite temperature circle, after using the map

τline = f(τcircle) = tan πτcircle

β
.
(3.48)

This is a special case of the general reparametrization symmetry (2.8). The expressions
for the propagators are simpler when we consider the theory on the line, so we will work
there for most of this section. Substituting (2.9) into the kernel we get

Kc(τ1, τ2; τ3, τ4)
=
− 1

α0

sgn(τ13)sgn(τ24)

|τ13|2∆|τ24|2∆|τ34|2−4∆
(3.49)

α0
≡
2πq

(q − 1)(q − 2) tan π

q
=
1

(q − 1)J2bq .
(3.50)

It will turn out that we can safely compute some, but not all, of the large-βJ correlator
using this expression for K. The reason is that some of the eigenfunctions have eigenvalue
Kc = 1 in the conformal limit, leading to a divergence in the geometric series (3.45). When

17


the time comes, in section 3.3, we will treat those eigenfunctions in perturbation theory
outside the conformal limit. For now, we proceed with (3.49).
The key property that makes it possible to diagonalize (3.49) is conformal invariance.
This can be presented using the following generators of an SL(2) algebra

ˆD = −τ∂τ − ∆,
ˆP = ∂τ,
ˆK = τ 2∂τ + 2τ∆

[ ˆD, ˆP] = P,
[ ˆD, ˆK] = − ˆK,
[ ˆP, ˆK] = −2 ˆD.
(3.51)

Here ∆ = 1/q is the conformal dimension of the fermion. These generators commute with
the kernel Kc, in the sense that up to total derivatives with respect to τ3 and τ4, we have

( ˆD1 + ˆD2)Kc(τ1, τ2; τ3, τ4) = Kc(τ1, τ2; τ3, τ4)( ˆD3 + ˆD4)
(3.52)

and similarly for the ˆP and ˆK generators. (These are the generators appropriate for acting
on the non-symmetric kernel Kc. To get a set that commutes with the symmetric version
�Kc we should replace ∆ by 1/2.)
This symmetry is useful in two ways. First, it implies that the ladder diagrams Fn are
simple powers times a function of the SL(2) invariant cross ratio:

χ = τ12τ34

τ13τ24
.
(3.53)

This is because the function F0 in (3.41) transforms like a conformal four point function,
and this property is preserved by acting with an SL(2) invariant operator. This will allow
us to represent the kernel in the space of functions of a single cross ratio, rather than in
the space of functions of two times. In other words, we can consider Kc(χ; ˜χ) instead of
Kc(τ1, τ2; τ3, τ4). Second, it implies that the kernel commutes with the casimir operator
C1+2 built from the sum of the generators acting on the two times:

C1+2 = ( ˆD1 + ˆD2)2 − 1

2( ˆK1 + ˆK2)( ˆP1 + ˆP2) − 1

2( ˆP1 + ˆP2)( ˆK1 + ˆK2)

= 2(∆2 − ∆) − ˆK1 ˆP2 − ˆP1 ˆK2 + 2 ˆD1 ˆD2.
(3.54)

The casimir is a differential operator with a family of eigenfunctions given by simple powers
times functions Ψh(χ). Because the spectrum is nondegenerate, these must be exactly the
eigenfunctions of the kernel Kc(χ; ˜χ) acting in the space of cross ratios. This leads to a
recipe for the four point function:

1. Understand the properties of F and Fn as functions of the cross ratio.

2. Find the eigenfunctions of C1+2 with these properties. These are particular hyper-
geometric functions Ψh(χ), related to conformal blocks of weight h.

3. Determine the set of h to have a complete basis of functions. This turns out to be
h = 1

2 + is and h = 2, 4, 6, 8, ....

18


4. Compute kc(h), the eigenvalue of the kernel Kc as a function of h.

5. Determine the inner products ⟨Ψh, F0⟩ and ⟨Ψh, Ψh⟩.

6. Compute the four point function as

F(χ) =
1

1 − Kc
F0 =
�

h
Ψh(χ)
1

1 − kc(h)
⟨Ψh, F0⟩
⟨Ψh, Ψh⟩.
(3.55)

We now go through each of these steps in detail.

3.2.1
The four point function as a function of the cross ratio

In the conformal limit, the ladder diagrams Fn will transform under SL(2) like a four
point function of dimension ∆ fields,

Fn(τ1...τ4) = Gc(τ12)Gc(τ34)Fn(χ),
χ = τ12τ34

τ13τ24
,
Gc(τ) = b sgn(τ)

|τ|2∆ .
(3.56)

Using the antisymmetry under τ1 ↔ τ2 and under τ3 ↔ τ4, the symmetry under (τ1, τ2) ↔
(τ3, τ4) and an SL(2) transformation, we can arrange to have τ1 = 0, τ3 = 1, τ4 = ∞
and also τ2 > 0. This restricts the cross ratio χ = τ2 to be positive. Because of the time
ordering in (3.40), the ordering of the fermions and the overall sign depends on whether
χ is less than or greater than one:

Fn(χ) ∼

�
+⟨ψj(∞)ψj(1)ψi(χ)ψi(0)⟩
0 < χ < 1
−⟨ψj(∞)ψi(χ)ψj(1)ψi(0)⟩
1 < χ < ∞.
(3.57)

When χ < 1 we have an iijj configuration, and when χ > 1 we have ijij, see figure (10).
In the region χ > 1, the correlation function has an extra discrete symmetry. This is
easiest to see if we place the points on the circle using the somewhat nonstandard map

τ − 2

τ
= tan θ

2.
(3.58)

The three operators at 0, 1 and ∞ get sent to the points −π, − π

2 and π

2 as shown in figure
5. The final operator at τ2 = χ ends up at some coordinate θ. The obvious symmetry
under θ → −θ translates to χ →
χ

χ−1. This means that in the region χ > 1, we must
have F(χ) = F(
χ

χ−1). Notice that this transformation maps the interval 1 < χ < 2 to the
range 2 < χ < ∞, with a fixed point at χ = 2. The conclusion is that the full F(χ) is
determined once we know it in the region 0 < χ < 2, and also that F must have vanishing
derivative at the point χ = 2.
An obvious advantage of the cross ratio is that the ladder kernel becomes a function
of fewer variables. One can substitute the form (3.56) into the original expression for the
kernel (3.43) and then do one of the τ integrals. The result is an equation of the form

Fn+1(χ) =
� 2

0

d˜χ
˜χ2 Kc(χ; ˜χ)Fn(˜χ)
(3.59)

19


=
1

2

3

4

1

2

3

4

θ
-θ

Figure 5: The symmetry of the χ > 1 correlator under χ →
χ

χ−1 is manifest as θ → −θ
after mapping to the circle.

where Kc(χ; ˜χ) is a symmetric kernel that is given in terms of hypergeometric functions
in appendix B.

3.2.2
Eigenfunctions of the casimir

We now search for a complete set of eigenfunctions of the casimir C1+2 with the properties
just described. First we need to understand how C1+2 acts on functions of the cross ratio.
One can check directly from (3.54) that

C1+2
1

|τ12|2∆f(χ) =
1

|τ12|2∆Cf(χ)
(3.60)

C ≡ χ2(1 − χ)∂2
χ − χ2∂χ.

Writing the eigenvalue as h(h − 1), the equation we would like to solve is Cf = h(h − 1)f.
The general solution is a linear combination of

χh
2F1(h, h, 2h, χ),
χ1−h
2F1(1 − h, 1 − h, 2 − 2h, χ).
(3.61)

We need to select from this set a complete basis for the space of functions with f ′(2) = 0.
These functions should also be normalizable with respect to the inner product from (3.59)
that makes K symmetric,

⟨g, f⟩ =
� 2

0

dχ
χ2 g∗(χ)f(χ).
(3.62)

This is the same inner product that makes C hermitian, neglecting boundary terms. Since
the eigenfunctions of a hermitian operator are complete, we can determine the basis by
finding the conditions that make the boundary terms vanish, and then selecting the eigen-
functions from among (3.61) that satisfy these conditions.
The hermiticity condition is

0 = ⟨g, Cf⟩ − ⟨Cg, f⟩ =
� 2

0
dχ
�
g∗(1 − χ)f ′ − g∗′(1 − χ)f
�′.
(3.63)

At χ = 2 the boundary term vanishes due to the requirement f ′(2) = 0. At χ = 0 it
vanishes provided that we impose that f → 0 faster than χ1/2. Because the eigenfunctions

20


(3.61) have logarithmic singularities at χ = 1, there is another possible “boundary” con-
tribution from this point. In order for it to vanish, we need to impose that the logarithmic
and constant terms in f agree as we approach χ = 1 from the two sides. In other words,
if we have f ∼ A + B log(1 − χ) for χ → 1−, then we should have f ∼ A + B log(χ − 1)
for χ → 1+. This will cancel the boundary terms provided that we define the integral by
approaching one in the same way from 1− and 1+.
We now look for eigenfunctions with these properties. We can start in the region χ > 1
by imposing that f ′(2) = 0. This selects a linear combination of the functions (3.61) that
can be written using a special hypergeometric identity as

Ψh = Γ( 1

2 − h

2)Γ( h

2)
√π
2F1

�h

2, 1

2 − h

2, 1

2, (2 − χ)2

χ2

�
1 < χ,
(3.64)

where we have chosen a convenient normalization constant. Note that Ψh = Ψ1−h in a
manifest way. In the region χ < 1, we must match to a linear combination

Ψh = AΓ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ) + B Γ(1 − h)2

Γ(2 − 2h)χ1−h
2F1(1 − h, 1 − h, 2 − 2h, χ)
χ < 1,

(3.65)
by requiring that the logarithmic and constant terms at χ = 1 agree with (3.64). This
determines

A =
1

tan πh

2

tan πh

2
,
B = A(1 − h) = −tan πh

2
tan πh

2
.
(3.66)

The final condition to impose is that Ψh must vanish at least as fast as χ1/2 as χ → 0.
There are two types of solutions.

1. For h = 1

2 + is both terms in (3.65) are marginally allowable. These solutions are
monotonic for 1 < χ and oscillatory for χ < 1, with infinitely many oscillations.

2. For h = 2n, n = 1, 2, 3, · · · the B coefficient vanishes, so (3.65) is again allowable at
small χ. These solutions are monotonic for 0 < χ < 1 and oscillatory for 1 < χ (it
crosses zero n times).

Together, these two sets form a complete basis of normalizable functions with f ′(2) = 0.
We emphasize that in both cases, Ψh is given by (3.64) for 1 < χ and (3.65) for χ < 1. For
the continuum states h = 1

2 + is there is an integral representation that gives the correct
answer for all χ > 0,

Ψh(χ) = 1

2

� ∞

−∞
dy
|χ|h

|y|h|χ − y|h|1 − y|1−h.
(3.67)

This integral does not converge for the discrete states. Finally, we note for later use that
near χ = 1 the function Ψh has the expansion

Ψh ∼ −
�
log(χ − 1) + 2γ + 2ψ(h) − π tan πh

2

�
(χ > 1).
(3.68)

For χ < 1 we replace log(χ − 1) → log(1 − χ).

21


3.2.3
The eigenvalues of the kernel kc(h)

The eigenfunctions Ψh of the casimir C were nondegenerate. Because the casimir commutes
with the kernel Kc, these functions must also be eigenfunctions of Kc. In principle, we can
compute the eigenvalues kc(h) by integrating the functions Ψh(χ) with Kc(χ; ˜χ). However,
we can get the answer in a simpler way. We start by backing off of the cross ratio formalism
and thinking about the casimir acting on two times, C1+2. Eigenfunctions of this operator
with eigenvalue h(h − 1) have the form of conformal three point functions of two fermions
with a dimension h operator,

sgn(τ1 − τ2)

|τ1 − τ0|h|τ2 − τ0|h|τ1 − τ2|2∆−h.
(3.69)

For any value of τ0 and h, these are also eigenfunctions of the kernel Kc. The eigenvalue
kc(h) depends only on h, since we can use SL(2) to move τ0 around. In particular, we can
take it to infinity, so that the eigenvalue is, see (3.49),

kc(h) =
�
dτdτ ′Kc(1, 0; τ, τ ′) sgn(τ − τ ′)

|τ − τ ′|2∆−h

= − 1

α0

�
dτdτ ′sgn(1 − τ)sgn(−τ ′)sgn(τ − τ ′)

|1 − τ|2∆|τ ′|2∆|τ − τ ′|2−2∆−h .
(3.70)

This integral can be evaluated by dividing up the τ and τ ′ integrals into regions where the
sign functions are constant. A quicker way to get the answer is as follows. We use

sgn(τ)

|τ|a
=
� dω

2π e−iωτc(a)|ω|a−1sgn(ω) ,
c(a) = 2i2−a√π Γ(1 − a

2)

Γ( 1

2 + a

2)
(3.71)

to write the factor in (3.70) that depends on |τ −τ ′| as a fourier transform. Then the τ and
τ ′ integrals factorize. We can shift the integration variables and then use (3.71) again for
each factor. These two factors are equal up to an overall sign. Finally we get an integral
of the same form as (3.71). Thus we find that

kc(h) = − 1

α0

c(2 − 2∆ − h)

c(2∆ − h)
[c(2∆)]2(−1).
(3.72)

Using α0 from (3.49) and using Γ function identities, one finds [1]

kc(h) = −(q − 1)
Γ( 3

2 − 1

q)Γ(1 − 1

q)

Γ( 1

2 + 1

q)Γ( 1

q)

Γ( 1

q + h

2)

Γ( 3

2 − 1

q − h

2)

Γ( 1

2 + 1

q − h

2)

Γ(1 − 1

q + h

2).
(3.73)

We can apply this result to the eigenfunctions Ψh(χ) by using the representation

sgn(τ12)sgn(τ34)

|τ12|2∆|τ34|2∆ Ψh(χ) = 1

2

�
dτ0
sgn(τ12)

|τ10|h|τ20|h|τ12|2∆−h
sgn(τ34)

|τ30|1−h|τ40|1−h|τ34|2∆−1+h.
(3.74)

22


=
-(q-1)
=
(q-1)
-(q-1)

Figure 6: On the left we have the kernel acting on G(τ). This is equal to (q − 1)G ∗ Σ ∗ G.
Using the approximate Schwinger-Dyson equation (2.7), this becomes −(q − 1)G.

which holds for h = 1

2 + is. This follows from the SL(2) covariance of the right hand side
and from (3.67). The τ1, τ2 dependence here is a superposition of eigenfunctions of the form
(3.69), so the left hand side is an eigenfunction of Kc with eigenvalue kc(h). The eigen-
functions in the discrete case are analytic continuations of the continuum eigenfunctions,
so their eigenvalues are determined by the continuation of kc(h).
The eigenvalue kc(h) is real for all of the eigenvectors h = 1

2 + is and h = 2, 4, 6, ....
It is positive for the discrete states, and negative for the continuum. We will find the full
analytic function useful in what follows. This function satisfies kc(h) = kc(1 − h). For
generic q, it has poles at h = 1 + 2

q + 2n for n ≥ 0 and the corresponding h → (1 − h)
reflection. Some simple special cases are

kc(h) = −3

2
tan π(h−1/2)

2

(h − 1/2)
q = 4
(3.75)

kc(h) =
2

h(h − 1)
q = ∞
(3.76)

kc(h) = −1
q = 2.
(3.77)

We can understand kc(h) at some special values of h using the Schwinger-Dyson equa-
tion. When h = 0, we are acting the kernel on a multiple of the orginal Gc(τ). This should
give kc(0) = −(q − 1), as we argue in figure 6. We will see below that when h = 2, we
are acting with the kernel on a linearized reparameterization of Gc(τ). One can then use
the reparameterization invariance of (2.7) to make a similar argument that kc(2) = 1, see
(3.108) below.

3.2.4
The inner products ⟨Ψh, Ψh⟩ and ⟨Ψh, F0⟩

Next we consider the norms of the eigenfunctions Ψh, beginning with the continuum h =
1
2 + is, and taking s, s′ > 0. We continue to use the norm for functions of χ defined in
(3.62). We expect the inner product ⟨Ψh, Ψh′⟩ to be proportional to δ(s − s′). A singular
contribution of this type can only come from the small χ region of the inner product
integral, where we can replace the hypergeometric functions in (3.64) by one. Using ∼ to
denote agreement up to terms that are finite as s → s′, we have

⟨Ψh, Ψh′⟩ ∼ π tan πh

4h − 2

� ϵ

0

dχ
χ
�
χi(s−s′) + χ−i(s−s′)�
∼ π tan πh

4h − 2 2πδ(s − s′).
(3.78)

23


Based on this calculation, one might expect that the inner product has finite terms in
addition to the δ(s − s′).
In fact, this cannot be the case, since eigenfunctions with
different values of s must be orthogonal. We conclude that the RHS of (3.78) is the exact
answer.
For the discrete set, h = 2n, we have that Ψh(χ) = 2Re[Qh−1(y)], where y = (2 − χ)/χ
and Q is the Legendre Q function. After writing the inner product as an integral over y,
one can use standard integral formulas for Q to find

⟨Ψh, Ψh′⟩ = δhh′π2

4h − 2.
(3.79)

We also need to compute the inner product of these eigenfunctions with the zero-rung
ladder F0. As a function of the times τ1, ..., τ4, F0 is given in (3.41). Using the conformal
form of G(τ) and going to a function of χ using (3.56), we have

F0(χ) =






−χ2∆ +
�
χ

1−χ
�2∆
0 < χ < 1

−χ2∆ −
�
χ

χ−1
�2∆
1 < χ < ∞.
(3.80)

We consider the inner product with the continuum states; the case with the discrete states
will follow by analytic continuation in h. The inner product integral can be done inside
the integral representation (3.67). Notice that the integral representation extends to a
function on the entire line −∞ < χ < ∞ that satisfies Ψ(χ) = Ψ(
χ

χ−1). For χ > 1 we
have that the zero-rung ladder F0 is symmetric under the same transformation, while for
χ < 1 it is antisymmetric. Using these properties we can write the inner product as a
single integral over the whole line:

⟨Ψh, F0⟩ = −1

2

� ∞

−∞
dydχ
sgn(χ)

|χ|2−h−2∆|χ − y|h|1 − y|1−h|y|h.
(3.81)

The integration region can now be divided up and all integrals can be done using the Euler
beta function. It is convenient to write the answer in terms of the eigenvalue function kc(h)
as

⟨F0, Ψh⟩ = α0

2 kc(h).
(3.82)

We can understand the appearance of kc(h) here by realizing that F0 is proportional to
the action of Kc on a delta function, so it should have an expression involving an integral
of kc(h) over the basis elements. We discuss this further in appendix C.

24


3.2.5
The sum of all ladders

We can now write a slightly naive expression for the full sum of ladders as

F(χ) =
�

h
Ψh(χ)
1

1 − kc(h)
⟨Ψh, F0⟩
⟨Ψh, Ψh⟩
(3.83)

= α0

� ∞

0

ds
2π
(2h − 1)
π tan(πh)
kc(h)

1 − kc(h)Ψh(χ) + α0

∞
�

n=1

�(2h − 1)

π2
kc(h)

1 − kc(h)Ψh(χ)
�

h=2n
.

The problem with this formula is that the n = 1 term in the sum diverges, since the eigen-
value kc(2) = 1. Of course, the actual four point function is finite; what this means is that
we have to treat the contribution of the h = 2 eigenfunctions outside the conformal limit,
where the eigenvalues will be slightly less than one. This gives an enhanced contribution
that we will analyze in section 3.3 below5 For now we focus on the contribution of the
h ̸= 2 eigenfunctions, for which the conformal limit can be taken smoothly. We refer to
the contribution of these eigenfunctions as Fh̸=2:

Fh̸=2

α0
=
� ∞

0

ds
2π
(2h − 1)
π tan(πh)
kc(h)

1 − kc(h)Ψh(χ) +

∞
�

n=1

�(2h − 1)

π2
kc(h)

1 − kc(h)Ψh(χ)
�

h=2n
. (3.84)

This can be put into a more convenient form by substituting

2

tan πh =
1

tan πh

2
−
1

tan π(1−h)

2
,
(3.85)

and then combining terms by extending the region of integration to all values of s and
using the antisymmetry of the rest of the integrand under h → 1 − h. We get

Fh̸=2(χ)

α0
=
� ∞

−∞

ds
2π
(h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)Ψh(χ)+

∞
�

n=2
Res
� (h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)Ψh(χ)
�

h=2n
(3.86)
where now the integral runs over all s, and we’ve written the discrete sum as a sum over
residues of the poles of 1/ tan(πh/2).
A nice feature of this formula is that it can be understood as a single contour integral,
over a contour in the complex h plane defined as

1
2πi

�

C
dh =
� ∞

−∞

ds
2π +

∞
�

n=1
Resh=2n.
(3.87)

Note that Ψh has poles at h = 1 + 2n.
Howevever, these are cancelled by zeros of
1/ tan(πh/2) at the same values. Therefore the product has poles only at h = 2n. The

5In appendix H we discuss a model where we effectively replace 1 − kc(h) → 1 − gkc(h), with g < 1, in
(3.83), which removes the h = 2 divergence.

25


contribution of the explicit residues will imply that we do not end up picking up the poles
at these locations either when we shift the contour to the right.
Let us see how this work in more detail. First, we consider the case χ > 1. Then we
can push the contour from the s axis rightward to infinity. In the process, we cancel the
sum over residues, but we pick up poles at the locations where kc(h) = 1 (see figure 7).
We refer to these values as hm, and we will say more about them in the next section:

Fh̸=2(χ) = −α0

∞
�

m=0
Res
� (h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)Ψh(χ)
�

h=hm
χ > 1.
(3.88)

The case for χ < 1 is more delicate, since we cannot push the 2F1(1−h, 1−h, 2−2h, χ)
function in Ψh(χ) to large positive h. So we do the following: first, we use the h → (1−h)
antisymmetry of the rest of the integrand to replace the tan(πh/2) inside the integral by
tan(πh). This gives an integrand that is explicitly symmetric under h → (1 − h). Next,
we use this symmetry to replace the B term in (3.65) by another copy of the A term. This
gives

Fh̸=2(χ)

α0
=
� ds

2π
(h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)
Γ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ)

+

∞
�

n=2
Res
� (h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)
Γ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ)
�

h=2n
,
(3.89)

where, in the residue sum, we have also used that Ψh(χ) = Γ(h)2

Γ(2h)χh2F1(h, h, 2h, χ) for even
integer h. This integrand can now be pushed to the right as before, cancelling the explicit
residues and picking up the poles where kc(h) = 1:

Fh̸=2(χ) = −α0

∞
�

m=0
Res
� (h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)
Γ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ)
�

h=hm
χ < 1.

(3.90)

3.2.6
Operators of the model

An important region of the four point function is the OPE limit of small χ. The expansion
of the four point function in this region gives the coefficients and dimensions of the opera-
tors appearing in the product of two fermions, ψi(0)ψi(χ). We can read these off from the
expression (3.90).
The first solution to kc(h) = 1 is h0 = 2. Although we omitted the divergent h = 2
piece from the discrete sum in defining Fh̸=2, we still pick up a finite contribution from
the double pole at that location when we deform the contour. However, it turns out that
this piece cancels against other contributions that will be described in section 3.3.4 below.

26


h

2

x
x
x

4
6
8
10
12

x
x
x

h

h0=2

x
x
x

h1

x
x

h2
h3
h4
h

1/2+is

Figure 7: The continuum piece of the contour that defines Fh̸=2 can be pushed to the
right, canceling the residues of the poles of the 1/ tan(πh/2) (dots), and picking up poles
from the locations where kc(h) = 1 (crosses). We have a double pole at h = 2.

After h0 = 2, we have an infinite set of solutions h1, h2, ... that are associated to ordinary
poles. The sum over these has the expected form for an operator product expansion

⟨4pt⟩ =

∞
�

m=1
c2
m
�
χhm2F1(hm, hm, 2hm, χ)
�
,
(3.91)

where the hm are the dimensions of the operators appearing, the quantity in brackets is
the corresponding conformal block, and c2
m would be the square of the operator product
coefficient. In particular, c2
m should be positive. From (3.90) we get

c2
m = −α0

N ·
(hm − 1/2)

π tan(πhm/2)
Γ(hm)2

Γ(2hm) ·
1

−k′(hm)
(hm > 2).
(3.92)

In this expression, we have included the overall factor of 1/N that relates F(χ) to the four
point function (3.40). One can check that c2
m is positive, because k′(hm) is negative and
tan πhm/2 is also negative. The rest of the factors are positive.
We do not have an exact expression for the dimensions hm, but we can parameterize
the values as
hm = 2∆ + 1 + 2m + ϵm,
(3.93)

where we observe that ϵm becomes small at large m. Asymptotically,

ϵm =2Γ(3 − 2∆) sin(2π∆)

πΓ(1 + 2∆)
1

(2m)2−4∆ ,
m ≫ 1

ϵm =
3

2πm ,
for ∆ = 1/4

ϵm =2∆

m2 ,
for ∆ → 0
(3.94)

One would like to view these as arising from two particles in AdS with some interaction.
In general the correction to the energy is related to the scattering phase shift δ ∼ log S,

27


where S is the S matrix. This is related to the relativistically invariant amplitude by
δ ∼ A/s where s is the center of mass energy, or equal to s ∼ m2 in this case (for large m).
We see that, generically, we cannot get (3.94) from a local interaction, since those would
involve powers of m2. For example, an interaction mediated by a particle of spin J would
give δ ∼ m2J−2, while what we have here goes like δ ∼ 1/m (for q = 4). For the special
case of ∆ → 0, we have something consistent with an interaction mediated by a spin zero
field, but the interaction is going to zero as ∆ → 0.
Here we have emphasized that the hm values are the powers that appear in the OPE. By
conformal invariance, these are the same powers that determine the decay of perturbations
to the system after excitation by a fermion bilinear.

3.2.7
Analytic continuation to the chaos region

Another interesting region to consider is where we take the large real-time behavior of
an out-of-time-order product with the ordering ψi(t)ψj(0)ψi(t)ψj(0). The behavior of four
point functions in this limit is a probe of chaos. A convenient configuration is the correlator

Tr[y ψi(t)y ψj(0)y ψi(t)y ψj(0)]
y ≡ ρ(β)1/4
(3.95)

where we have split the thermal density matrix into four factors y as in [12]. In a conformal
theory, this can be obtained from the Euclidean correlator on the line, by mapping to
the finite temperature circle using (3.48) and then continuing to real time. To get the
configuration (3.95), the upshot is that we should study the four point function at a value
of the cross ratio equal to

χ =
2

1 − i sinh 2πt

β
.
(3.96)

Note that the χ → χ/(χ − 1) symmetry of the correlator takes t → −t in (3.96) and it
ensures the reality of (3.95). Notice that for t = 0 this is a value greater than one, so
we should start with the formula for χ > 1 and analytically continue it. For large values
of t, we will end up with a small and purely imaginary cross ratio. But because we are
continuing the χ > 1 expression to small χ, we do not end up with the OPE limit of small
χ.
The difference between these limits arises because the continuation to small χ of the
χ > 1 expression for Ψh is not the same as the function Ψh evaluated directly at small χ.
Indeed, for small χ, the continuation gives

Ψχ>1
h
(χ) ∼ Γ( 1

2 − h

2)Γ(h − 1

2)

21−hΓ( h

2)
(−iχ)1−h + (h → 1 − h).
(3.97)

If the real part of h is greater than one, this will be growing for small χ. By (3.96), this
translates to exponential growth as a function of t that is a diagnostic of many-body chaos.
Formally, the divergent term at h = 2 corresponds to a growth ∝ χ−1 ∝ e2πt/β that
saturates the chaos bound. We will see below that this rate of growth remains correct

28


h

2

x

4
6
8
10

h

2

x

4
6
8
10
=

Figure 8: To continue the sum over residues to the chaos region, we first replace the kc(h)
by kR(1 − h), and then pull the contour surrounding the poles back to the line 1/2 + is,
picking up the double pole at h = 2 but no other poles. In this form the function can
safely be continued. In addition we also have the original integral along h = 1/2 + is with
the function kc; we leave this piece alone because it can already be continued.

when we treat the enhanced h = 2 contribution outside the conformal limit. For now, we
consider the continuation of the rest of the correlator, Fh̸=2, but we emphasize that this
is a small correction to the h = 2 piece, in the chaos limit as well as elsewhere.
If Fh̸=2 were a finite sum of Ψh, we could analyze the chaos region by continuing each
of the terms separately. If we try this with (3.86), or with (3.88), we will find that the
residue sum does not converge after the continuation. So we have to first manipulate the
expression into a form that is safer to continue. We start by defining a function kR(h) by

kR(1 − h)

kc(h)
= cos π(∆ − h

2)

cos π(∆ + h

2).
(3.98)

This function has an interpretation in terms of the eigenvalues of the real-time ladder
kernel constructed from retarded propagators (3.152). However, for our purposes we only
need to know two properties. First, kR(1 − h) = kc(h) when h is an even integer, so we
can replace kc(h) → kR(1 − h) inside the residue sum of (3.86). Second, kR(1 − h) is equal
to one at only a single place in the complex plane, h = 2. This means that when we pull
the contour that circles the h = 4, 6, · · · poles back to the line h = 1

2 + is, as shown in
figure 8, we will only pick up a double pole at h = 2 plus the integral over the line. This
leads to

Fh̸=2(χ)

α0
=
� ds

2π
(h − 1/2)

π tan(πh/2)

�
kc(h)

1 − kc(h) −
kR(1 − h)

1 − kR(1 − h)

�
Ψh(χ)

− Res
� (h − 1/2)

π tan(πh/2)
kR(1 − h)

1 − kR(1 − h)Ψh(χ)
�

h=2
.
(3.99)

So far this is just another legal way to write the Euclidean correlator.
Now we consider the continuation of the χ > 1 expression to small χ. The integral over
s does not give anything growing as χ becomes small, because we can do the continuation

29


in such a way that the integral always remains convergent, and the integrand vanishes as
χ → 0. Therefore the only growing piece in Fh̸=2 comes from the second line of (3.99). This
is essentially a Regge pole. In our case it is a double pole, so we get a linear combination
of Ψ2(χ) and ∂hΨ2(χ)|h=2. Unlike the double pole in the OPE region, this does not cancel
against other contributions.
The term proportional to Ψ2 will saturate the chaos bound, but naively the second
term exceeds it, due to the extra logarithm in the small χ behavior:

∂hΨh(χ)|h=2 ∼ −
2π log
1

−iχ

−iχ
− 2π

−iχ.
(3.100)

This translates to something proportional to t e2πt/β at large t, which would violate the
bound. Also, the term comes with a sign that is forbidden by the argument of [12]. So,
by itself, Fh̸=2 would not be an allowable four point function. However, it is consistent as
a small correction to the enhanced h = 2 piece that we will study below. The t e
2πt

β term
then corresponds to a small finite-coupling shift (a decrease) in the growth exponent of
the large h = 2 contribution.

3.3
Proper treatment of the h = 2 subspace

We saw above that the conformal limit of the kernel has eigenfunctions with eigenvalue
kc(h) = 1, which lead to a divergence in the four point function. These eigenfunctions are
h = 2 eigenfunctions of the casimir operator C1+2. In order to get a finite answer for the
four point function, we have to treat these particular eigenfunctions outside the conformal
limit, by doing perturbation theory in the leading non-conformal correction to the kernel,
δK. This correction arises from the leading non-conformal correction δG to the correlators
that make up the kernel. The small parameter is the inverse coupling, 1/(βJ).
Since the perturbation δK breaks conformal symmetry, the line and the finite temper-
ature circle are inequivalent, and we have to study the problem directly on the circle. We
will use an angular coordinate θ = 2πτ/β, which runs from 0 ≤ θ < 2π on the circle.
(Equivalently, we can say that we work in units where β = 2π, and that θ is the periodic
Euclidean time variable.)
It will be slightly more convenient to use the symmetric version of the kernel �K in this
section. This was defined in (3.47)

�K(θ1, θ2; θ3, θ4) = −J2(q − 1)|G(θ12)|
q−2

2 G(θ13)G(θ24)|G(θ34)|
q−2

2 .
(3.101)

We will refer to the antisymmetric eigenfunctions of this kernel as Ψexact
h,n (θ1, θ2) = −Ψexact
h,n (θ2, θ1),
where h is an abstract label that will become clear below, and n describes the Fourier index
in the center of mass coordinate e−in(θ1+θ2)/2. The kernel �K is symmetric with respect to
the standard inner product

⟨Ψ, Φ⟩ ≡
� 2π

0
dθ1dθ2Ψ∗(θ1, θ2)Φ(θ1, θ2).
(3.102)

30


To get a formula for the four point function, we can use the fact that the zero-rung
ladder F0 is proportional to the kernel acting on the antisymmetric identity matrix, �K · I,
where

I(θ1...θ4) = −δ(θ13)δ(θ24) + δ(θ14)δ(θ23) = −2
�

h,n
Ψexact
h,n (θ1, θ2)Ψexact∗
h,n
(θ3, θ4).
(3.103)

Roughly, the sum of ladders is then F = (1 − �K)−1 �K · I. More precisely,

�
(q−1)J2G
q−2

2 (θ12)G
q−2

2 (θ34)
�
F(θ1...θ4) = 2
�

h,n

k(h, n)

1 − k(h, n)Ψexact
h,n (θ1, θ2)Ψexact∗
h,n
(θ3, θ4),

(3.104)
where k(h, n) is the exact eigenvalue associated to Ψexact
h,n (θ1, θ2). For the appropriate set
of eigenvectors, this formula is correct for any value of the coupling βJ.
In the conformal limit βJ ≫ 1 we can make contact with our previous analysis: the
exact eigenfunctions Ψexact
h,n
approach eigenfunctions Ψh,n of the casimir C1+2 with eigen-
value h(h − 1). The eigenvalue k(h, n) → kc(h) becomes a function of h only, and the sum
over n in (3.104) reproduces the previous expression in terms of the functions Ψh(χ), see
appendix D. The sum over h includes both the continuum and discrete pieces. We can
take the conformal limit smoothly for everything but h = 2. This gives the function Fh̸=2
that we studied previously, after mapping to the circle with τ = tan θ

2.
Before, we got an infinity in the conformal limit from the h = 2 contribution, which is
now given by a family of functions Ψ2,n for different Fourier index n. For these terms, we
have to retain the leading non-conformal correction to the eigenvalues k(2, n) = 1−O( 1

βJ ).
In the remainder of this section, we will do this in detail.

3.3.1
The h = 2 eigenfunctions and reparameterizations

We will start by working out the Ψ2,n functions on the circle. In the conformal limit, we
can substitute in for the propagators using (2.9) to get

�Kc(θ1, ..., θ4) = −α0
1

|2 sin θ12

2 |1−2∆
sgn(θ13)

|2 sin θ24

2 |2∆
sgn(θ24)

|2 sin θ34

2 |2∆
1

|2 sin θ13

2 |1−2∆.
(3.105)

with α0 defined in (3.50).
As on the line, this kernel commutes with a set of SL(2)
generators,
ˆP = e−iθ[∂θ − i/2],
ˆK = −eiθ[∂θ + i/2],
ˆD = i∂θ.
(3.106)

It follows that eigenfunctions of �Kc will be functions of two times that diagonalize the
casimir C1+2 = −1/2 − ˆK1 ˆP2 − ˆP1 ˆK2 + 2 ˆD1 ˆD2 and the translation operator ˆD1+2 =
ˆD1 + ˆD2. One can write the Casimir as a differential operator and directly find the h = 2
eigenfunctions.
We can get the answer another way by considering reparameterizations of the prop-
agator.
The Schwinger-Dyson equations in the conformal limit are reparameterization

31


invariant. This means that if we consider the change in G from a linearized reparameteri-
zation θ → θ + ϵ(θ), which is

δϵGc = [∆ϵ′(θ1) + ∆ϵ′(θ2) + ϵ(θ1)∂θ1 + ϵ(θ2)∂θ2] Gc,
(3.107)

then Gc + δϵGc will also solve the conformal Schwinger-Dyson equations (2.7). The first
equation in (2.7) then implies

0 = δϵGc ∗ Σc + Gc ∗ δϵΣc
−→
0 = δϵGc + Gc ∗ [(q − 1)J2Gq−2
c
δϵGc] ∗ Gc = (1 − Kc)δϵGc
(3.108)
where the star denotes the following product: (F ∗ G)(τ, τ ′′) =
�
dτ ′F(τ, τ ′)G(τ ′, τ ′′). We
conclude that δϵG is annihilated by (1 − K), so it is an eigenfunction of K with eigenvalue
one. For the symmetric kernel �K, the associated eigenfunction is |G|
q−2

2 δϵG.
To get a convenient basis, we can consider ϵ ∼ e−inθ. Plugging the conformal correlators
(2.9) into the reparameterization formula (3.107), evaluating |G|
q−2

2 δϵG and normalizing
with respect to (3.102), we get

Ψ2,n = γn
e−iny

2 sin x

2
fn(x),
fn = sin nx

2

tan x

2
− n cos nx

2 ,
(3.109)

x = θ1 − θ2,
y = θ1 + θ2

2
,
γ2
n =
3

π2|n|(n2 − 1).
(3.110)

These are eigenfunctions of �K with eigenvalue one, and eigenfunctions of the casimir C1+2
with h = 2. For the cases n = −1, 0, 1, the variation δϵG vanishes, because of the SL(2)
covariance of the conformal correlators. So we only have eigenfunctions for |n| ≥ 2. For
positive n, they organize into a single representation of SL(2), with highest weight vector
Ψ2,2. One can repeatedly apply P1 + P2 to this function to get all of the Ψ2,n, with n ≥ 2.
We similarly get a single lowest weight representation that describes n ≤ 2.
In section 4, we will use the reparameterization perspective to give a simple explanation
of why these eigenfunctions lead to a divergence in the four point function, and how it gets
regulated. For now we proceed in the most straightforward way, correcting the infinity by
finding the shift in k(2, n) that fixes the vanishing denominator in (3.104).

3.3.2
The shift in the eigenvalues

We will start by studying the correction to k(2, n) at large q, where the eigenvalue prob-
lem is quite simple for all values of the coupling βJ. The simplification is because the
propagators are equal to

G(θ) = sgn(θ)

2

�
1 + g(θ)

q
+ ...
�
.
(3.111)

At large q we can set the side rail propagators equal to the first term. To form the rung
function Gq−2, we exponentiate the 1

q correction as in (2.18). The eigenvalue equation

32


�KΨ = kΨ is

−J2q
�
dθ1dθ2
sgn(θ13)

2
sgn(θ24)

2
e
1
2 [g(θ12)+g(θ34)]

2q−2
Ψ(θ1, θ2) = k Ψ(θ3, θ4).
(3.112)

Because the side rail propagators are so simple, we can turn this integral equation into a
differential equation by applying the differential operator ∂θ3∂θ4e− 1

2 g(θ34) to both sides and
using ∂x sgn(x) = 2δ(x). Plugging in for eg using (2.18), parameterizing the eigenvalue as
k = 2/h(h − 1), and making a fourier ansatz, one finds

Ψ(θ1, θ2) =e−iny

sin ˜x

2
ψn(x)
˜x = vx + (1 − v)π
(3.113)
�
n2 + 4∂2
x − h(h − 1)v2

sin2 ˜x

2

�
ψn(x) = 0.
(3.114)

Here, v was defined in (2.19), and we are using the same notation for x, y as in (3.109).
At infinite coupling, v is close to one (2.31). When v is exactly equal to one, (3.114) is
the equation for an eigenfunction of the casimir C1+2. However, (3.114) gives the exact
eigenvectors of the large q model for any value of the coupling.
We would like to find eigenfunctions with the right symmetry properties. As functions
of the two angles θ1, θ2, the four point function has the symmetries

F(θ1, θ2) = −F(θ2, θ1) ,
F(θ1 + 2π, θ2) = −F(θ1, θ2) ,
F(θ1, θ2 + 2π) = −F(θ1, θ2)
(3.115)
We can combine the first two to obtain F(θ1, θ2) = F(θ2 + 2π, θ1). In terms of x and y
this means that
F(x, y) = F(2π − x, y + π).
(3.116)

The first symmetry in (3.115) can be used to restrict the range of x to be positive. Then
the periodicity condition implies (3.116). To compensate for the factor of e−iny in (3.113),
ψn(x) needs to be symmetric about x = π for even n and antisymmetric for odd n.
Solutions with these properties are

ψh,n(x) ∼ (sin ˜x

2)h
2F1(h − ˜n

2
, h + ˜n

2
, 1

2, cos2 ˜x

2)
˜n = n

v
(n even)
(3.117)

∼ cos ˜x

2(sin ˜x

2)h
2F1(1 + h − ˜n

2
, 1 + h + ˜n

2
, 3

2, cos2 ˜x

2)
(n odd)
(3.118)

The quantization condition on h comes from the boundary condition that ψ should vanish
at x = 0, which means ˜x = π(1 − v).
We are interested in the eigenfunctions that approach the h = 2 conformal eigenfunc-
tions at strong coupling. For a generic value of h near two, we get a divergence as ˜x goes to
zero. To the first two orders in (1−v), the correct condition is just that this diverging term

33


should not be present. This means the first or second argument of the hypergeometric func-
tion should be a negative integer. The solution near two is hn = 2 + |˜n| − |n| = 2 + |n| 1−v

v .
Converting to k = 2/h(h − 1), we get

k(2, n) = 1 − 3|n|

2 (1 − v) +
�7n2

4
− 3|n|

2

�
(1 − v)2 + ...
(3.119)

= 1 − 3|n|

βJ +
7n2

(βJ )2 + ....
(3.120)

Now we move to general q. We can’t solve the eigenvalue problem exactly, but we can
do first order perturbation theory in the kernel, computing the shift in the eigenvalues of
the h = 2 eigenfunctions by taking ⟨Ψ2,n, δ �K · Ψ2,n⟩ where δ �K is the leading correction to
the conformal form. More specifically, we will take the leading correction in the infrared;
this will be justified as long as the integrals we get for the matrix elements are convergent.
The correction to the kernel comes from substituting in the correction Gc + δG to
the conformal propagators, where δG is the leading correction in the infrared. For the
large q model, we found in (2.38) that the leading correction to the conformal answer is
proportional to the function

f0(θ) ≡ 2 + π − |θ|

tan |θ|

2
.
(3.121)

(Note that f0 is not the limit n → 0 of fn defined for n ≥ 2 in (3.109).) By solving the
Schwinger-Dyson equations numerically for different values of q, we found in all cases that

δG
Gc
= − αG

βJ f0
(3.122)

is a good approximation for large θβJ and for a suitable constant αG. We give a plot of
αG(q), from fitting against the numerical solution, in figure 9. One can also show directly
that δG is an eigenfunction of the conformal kernel with eigenvalue one (and therefore an
allowed perturbation in the infrared, by appendix A), up to a UV divergence that should
be interpreted as a local source in the Schwinger-Dyson equation. The required source is
proportional to the −iω term that we dropped in the conformal limit, but matching the
numerical coefficient would require us to know how the divergence is regularized, which
seems to require the exact solution, see appendix A. Of course, in the q = ∞ model we
have the exact solution, and one can check that the coefficient is αG = 2/q at large q.
In the large q model and also in the numerics at general q, the next correction in the
IR appears to be at order (βJ )−2. One expects the next correction after that to be at
order (βJ )1−h1 where h1 is the dimension of the next irrelevant operator, i.e. solution to
kc(h) = 1. When q = 4 we have h1 = 3.7735....
Getting the shift in the eigenvalue from the correction δG involves some work, which
we will defer to appendix E. One approach is to compute the shift by directly evaluating
the integrals in ⟨Ψ2,n, δ �K · Ψ2,n⟩. This can be simplified by using conformal symmetry to
show that the answer has to be proportional to n, and then doing the integrals at large n.
We give some details on this method in appendix E.

34


2
4
6
8
10
12
14
16
18
20
q

0

0.05

0.1

0.15

0.2

αG

2
4
6
8
10
12
14
16
18
20
q

2.8

2.85

2.9

2.95

3

3.05

3.1

αK

Figure 9: The functions αG(q) and αK(q), computed by solving the Schwinger Dyson
equations numerically for different values of q. The first three physical values are αG(2) =
0, αG(4) ≈ 0.1872, and αG(6) ≈ 0.1737. Analytically, we know that αG behaves as 2/q at
large q, and like π(q −2)/8 for q near two. αK never strays more than roughly five percent
from the large q value of three.

A quicker way to get the answer is to use the fact, shown in appendix F, that

1

qαG
· ⟨Ψh,n, δ �K · Ψ2,n⟩

1 − kc(h)
(3.123)

is independent of q, despite the fact that the components qαG, kc, and δ �K each depend on
q. Here, Ψh,n is the conformal eigenfunction with weight h. The expression (3.123) has a
pole at h = 2, with residue proportional to the eigenvalue shift. Equating this residue with
what we get in the q = ∞ model, and using some large q data (qαG = 2, k′
c(2) = −3/2,
and (3.120)), we find

k(2, n) = 1 − αK

βJ |n| + ...
(3.124)

αK ≡ −qk′
c(2)αG =

�
πq

sin 2π

q
+ q3(6 − q) − 6q2

2(q − 1)(q − 2)

�

αG.
(3.125)

This agrees with the more direct method in appendix E. We give a plot of the coefficient
αK in the right panel of figure 9. One finds that it stays reasonably close to three for all
values of q.

3.3.3
The enhanced h = 2 contribution

Because the eigenvalues of the h = 2 eigenvectors are close to one, they give an enhanced
contribution to the four point function, of order βJ. This piece comes from the h = 2 part
of (3.104), where we put in the conformal results for everything except the 1 − k(h, n) in

35


(a)
(b)

j
i
i

i
j
j
i

j

Figure 10: Two configurations for the fermions. In (a) we have the iijj configuration with
χ < 1 and on the right we have the ijij configuration with χ > 1.

the denominator, which we correct using the leading shift (3.124). The result is

Fbig(θ1...θ4)
G(θ12)G(θ34) = 6α0

π2αK
βJ
�

|n|≥2

ein(y′−y)

n2(n2 − 1)

�sin nx

2

tan x

2
− n cos nx

2

��sin nx′

2

tan x′

2
− n cos nx′

2

�

x = θ12
x′ = θ34
y = θ1 + θ2

2
y′ = θ3 + θ4

2
.
(3.126)

Because of the βJ enhancement, this term is parameterically large compared to the h ̸= 2
pieces we studied in the previous sections. It is not conformally invariant, in the sense
that the sum is not only a function of the cross ratio. This lack of conformal symmetry
arises because the eigenvalue shift (3.124) depends on the index n that labels the SL(2)
descendant in the h = 2 representation. Concretely, we have n2(n2−1) in the denominator,
instead of |n|(n2 − 1) which would have given a multiple of Ψ2(χ), see (D.214).
Using the fact that the h = 2 eigenfunctions are linearized reparameterizations of the
propagator, with reparameterization ϵn ∝ e−inθ, we can write (3.126) as

Fbig =
�

n
⟨ϵnϵ−n⟩δϵnG δϵ−nG,
⟨ϵnϵ−n⟩ =
�6α0q2

αKN

�
βJ

n2(n2 − 1).
(3.127)

This is the type of contribution on expects from a fluctuation integral over reparameteriza-
tions of the conformal saddle point, with an action given by the inverse of the ϵ propagator.
We will say more about this perspective in section 4 below. For now, we note that one
can fourier transform (3.127) to obtain

⟨ϵ(θ)ϵ(0)⟩ = 1

N
6(βJ )β2q2α0

(2π)4αK

�
−1

2(|θ| − π)2 + (|θ| − π) sin |θ| + 1 + π2

6 + 5

2 cos θ
�

(3.128)
The sum over n in (3.126) can be done by repeatedly integrating the geometric series,
or by using this propagator and (3.107). The result depends on whether the ordering of
the times corresponds to an iijj or ijij ordering of the fermions, see figure 10. In the

36


non-alternating configuration iijj, we have the very simple expression

iijj order :
Fbig(θ1, θ2, θ3, θ4)

G(θ12)G(π)
= 6α0

π2αK
βJ

�
θ12

2 tan θ12

2
− 1

� �
θ34

2 tan θ34

2
− 1

�

(3.129)

This correlator is produced by fluctuations in the total energy in the thermal ensemble. Let
us be a bit more explicit. We can compute the contribution of the energy fluctuations by
starting with the variation in the correlator produced by a small change in the temperature:

G(τ, β + δβ)

G(τ, β)
= 1 − 2∆

β

�

1 −
πτ

β tan πτ

β

�

δβ
(3.130)

Now we use the saddle point relation E = c/(2β2) to get δβ = −β3δE/c, see (2.35). From
the fluctuations in E we therefore expect a connected piece in the four point function that
is

1
N
Fbig(τ1, τ2, τ3τ4)

G(τ12)G(τ34)
= 4

q2

�

1 −
πτ12

β tan πτ12

β

� �

1 −
πτ34

β tan πτ34

β

�
β4

c2 ⟨(δE)2⟩.
(3.131)

The energy two point function can be computed from

⟨(δE)2⟩ = ∂2
β log Z = c

β3.
(3.132)

Inserting this, writing things in terms of θ = 2πτ/β, and converting c to αK using (5.181),
we find exact agreement with (3.129).
For small θ12, (3.129) goes as θ2
12, suggesting the presence of a dimension two operator,
or an operator product expansion of the form ψ(θ1)ψ(θ2) ∝ (θ12)−2∆+2T(θ2). If we took
also the θ34 → 0 limit, then we would expect to have a result proportional to ⟨T(θ2)T(θ4)⟩.
If this was in a conformal field theory, we would have expected this to go like (sin θ24/2)−4.
Instead we find that it is a constant. The the reason is that T is essentially the Hamiltonian
of the theory. As such it is conserved and its two point function in the thermal ensemble
simply measures the energy fluctuations. We can write an explicit expression for T in
terms of ϵ of the form

T = NαS

J

�
ϵ′′′ + (2π)2

β2 ϵ′
�
+ · · ·
(3.133)

with αS as in (4.173). The dots indicate possible higher order terms in ϵ. We can then
check using (3.128) that ⟨TT⟩ is indeed constant, and is given by (3.132)

⟨T(τ1)T(τ2)⟩ = c

β3 ∝
N

β2(βJ ).
(3.134)

Because of the factor of (βJ ), this becomes small in the conformal limit, seemingly in
keeping with the idea that the stress tensor should vanish in a one dimensional conformal

37


field theory. However, the contribution to the four point function also involves the three
point couplings ⟨Tψψ⟩, which are the other factors in (3.131). These give two factors of
(βJ ) in the numerator, which more than compensate for the suppression of ⟨TT⟩.
We now turn our attention to a configuration in the alternating configuration ijij. The
result is a bit more complicated but it simplifies nicely in the case that we take one of
the pairs of fermions to be diametrically opposed on the circle. For example, we can take
θ3 = 0, and θ4 = π:

ijij order :
Fbig(θ1, θ2, 0, π)

G(θ12)G(π)
= − 6α0

π2αK
βJ

�
θ12

2 tan θ12

2
− 1 − πsin θ1

2 sin θ2

2

| sin θ12

2 |

�

.
(3.135)

This is the configuration that is appropriate for continuing to the chaos limit. To get the
function defined in (3.95), we continue to θ2 = π

2 − 2πi

β t and θ1 = θ2 − π, finding

Fbig(t)

G(π)G(π) = 6α0

π2αK
βJ
�
1 − π

2 cosh 2πt

β

�
.
(3.136)

This saturates the chaos bound.

3.3.4
Other terms from the h = 2 subspace

Because the factor
k(2,n)

1−k(2,n) in (3.104) is large, of order βJ , corrections of order (βJ )−1 from
the rest of the formula (3.104) will combine to give finite contributions in the conformal
limit. There are several sources of these terms. First, the δG correction to the propagators
on the LHS of (3.104) give a correction

F(θ1...θ4)

G(θ12)G(θ34) ⊃ −q

2

�δG(θ12)

Gc(θ12) + δG(θ34)

Gc(θ34)

�
Fbig(θ1...θ4)

Gc(θ12)Gc(θ34)
(3.137)

=
3α0

π2|k′
c(2)|
�
f0(θ12) + f0(θ34)
� �

|n|≥2

ein(y′−y)fn(x)fn(x′)

n2(n2 − 1)
.
(3.138)

Notice that this depends on q only through the factor α0/k′
c(2), which is a simple explicit
function of q. Next, we have a contribution from the first-order change in the h = 2
eigenvectors, δΨ2,n. In appendix F we show that δΨ2,n is independent of q except for a
coefficient of qαG. It is easy to check that this also leads to a term in F that depends on
q only through the prefactor α0/k′
c(2).
Third, we have contributions from the order (βJ )0 term in
k(2,n)

1−k(2,n).
This requires
knowledge of the second-order change in the eigenvalue, which we have not computed.
However, we have noticed that we get a very simple final answer if we assume

k(2, n) = 1 + k′
c(2)qαG|n|

βJ
+ k′′
c (2)
2

�qαG|n|

βJ

�2
+ · · · .
(3.139)

38


The first term is just a restating of (3.124). One can use (3.120) to check that the second
term is correct in the q = ∞ model. By diagonalizing the kernel constructed from the
numerical G(τ) (see appendix G), we have checked that the coefficient of the second term
in (3.139) is correct to within roughly percent-level multiplicative precision, for several low
values of q, n. We will assume that it is actually true.
The simplification that results from (3.139) is the following. One can use (D.214) to
show that the contribution to F from the order one term in
k(2,n)

1−k(2,n) combines with the
terms from the double pole at h = 2 in (3.88) and (3.90) to give, again, an expression
that depends on q only through the prefactor α0/k′
c(2). So we conclude that up to order
(βJ )0, the four point function will be given by the Fbig contribution, plus the residues of
the simple poles h1, h2, ... in (3.88) or (3.90), plus terms that are universal in q up to an
overall coefficient. We will compute these last terms by studying the q = ∞ four point
function in more detail.

3.4
More detail on the q = ∞ four point function

In the q = ∞ model, we can compute the four point function in way that simultaneously
treats all of the contributions we have been discussing so far. This is based on the fact that
F is a Green’s function for a simple differential operator. Because the side-rail propagators
in the large q kernel are proportional to sgn(θ), see (3.112), we have that

−
2

v2 ˜P 2∂θ1∂θ2K(θ1...θ4) = δ(θ13)δ(θ24),
˜P ≡
1

sin ˜x

2
,
˜x = vx + (1 − v)π.
(3.140)

In other words, the differential operator on the left hand side is the inverse of K. Roughly,
the four point function is given by F = (K−1 −1)−1. Multiplying both sides by (K−1 −1),
we get a differential equation for F with a delta function source.
To write the precise equation we get, it is convenient to use the coordinates x = θ12, y =
θ1+θ2

2
. These overcount physical configurations of points. We can reduce this overcounting
by restricting to x ≥ 0, x′ ≥ 0 and y ≥ y′. Then the correct equation is
�

−∂2
y
4 + v2∂2
˜x − v2 ˜P 2

2

�

F(x, y, x′, y′) = δ(y−y′)δ(x−x′)+δ(y−y′−π)δ(2π−x−x′). (3.141)

The second term on the RHS can be understood as the image of the first term under the
symmetry (x, y) → (2π − x, y − π), see the discussion above (3.117).
We can solve this equation by separation of variables. We expand in a complete set
of eigenfunctions of the operator −∂2
˜x + ˜P 2/2, with boundary conditions of zero at x = 0.
These eigenfunctions are just (3.117) and (3.118) with h = 2. The boundary condition
implies that ˜n should be an integer n ≥ 2 plus a correction of order (1 − v)3 that we will
neglect. Then (3.117) and (3.118) simplify to the functions fn defined in (3.109), with
eigenvalues n2/4. These functions satisfy a completeness relation
�

n>2

fn(˜x)fn(˜x′)
π(n2 − 1) = δ(˜x − ˜x′),
�

n>2
(−1)nfn(˜x)fn(˜x′)

π(n2 − 1) = δ(2π − ˜x − ˜x′).
(3.142)

39


So we can write

F(x, y, x′, y′) =
�

n>2
Hn(y − y′)fn(˜x)fn(˜x′)

π(n2 − 1)
(3.143)

�
−1

4∂2
y − v2n2

4

�
Hn(y) = v [δ(y) + (−1)nδ(y − π)] .
(3.144)

The factor of v on the right side came from δ(x − x′) = vδ(˜x − ˜x′) and δ(2π − x − x′) =
vδ(2π − ˜x − ˜x′).
The solution for Hn(y) should be continuous and 2π-periodic.
The
sources in (3.143) imply that we have a discontinuous derivative at y = 0 and y = π. The
discontinuity at zero is equivalent to a discontinuity between the derivative at 0+ and at
2π−. Solving these constraints, we find that the solution for 0 < y < 2π is

Hn(y) = −
2

n sin(nπv)
�
cos [nv(y − π)] + (−1)n cos [nv(|y − π| − π)]
�
(3.145)

= 4 cos(ny)

πn2(1 − v) + 4(y − π

2) sin(ny)
πn
+ O(1 − v)
(3.146)

=
�βJ

2
+ 1 − (y−π

2 )∂y

� 4 cos(ny)

πn2
+ O( 1

βJ ).
(3.147)

In the second line we expanded in 1 − v assuming 0 < y < π. (For π < y < 2π, we need
to replace the π/2 in the second term by 3π/2.) In the third line we used
1

1−v ≈ βJ

2 + 1.
Substituting (3.147) into (3.143) and also using fn(˜x) = fn(x) + (1 − v)(π − x)f ′
n(x) + ...,
we get the full q = ∞ four point function up to order (βJ )0:

F(x, y, x′, 0) =
�
βJ − 2
�
−1 + (y−π

2 )∂y + (x−π)∂x + (x′−π)∂x′
�� �

|n|≥2

e−inyfn(x)fn(x′)

π2n2(n2 − 1)
.

(3.148)
One can check that the term of order (βJ ) reproduces Fbig from (3.126) for the case
q = ∞ (α0 = 2, αK = 3, G = 1

2). Although we have not displayed it here, the next term,
at order (βJ )−1 can also be computed from (3.145). Beyond that order, one has to use
the hypergeometric functions (3.117) and (3.118) that generalize fn for non-integer n.
An interesting feature of the function Fbig was that it was independent of y in the
non-alternating configuration, which corresponds to y > |x + x′|/2. This persists at order
one, since the new y dependence of (3.148) is proportional to a y derivative of Fbig. This
is consistent with the idea that in the q = ∞ model the Hamiltonian is the only operator
that appears in the OPE. Notice that this is rather nontrivial from the way we set up
the calculation in the previous sections: in the OPE region, the y-dependent double pole
contribution in Fh̸=2 must be completely cancelled by some of the terms discussed in the
last section 3.3.4.

40


3.5
Summary of the four point function

In the previous section, we argued that the order-one terms in F coming from the double
pole and the various corrections to the h = 2 contributions add up to a function that
depends on q only through the prefactor α0/k′
c(2). We can then use the q = ∞ result
(3.148) to write the general four point function up to order one. When χ < 1, we have

F(x, y, x′, 0)
G(x)G(x′) =

= α0

�6βJ

αK
−
6

|k′
c(2)|

�
−1 + (y−π

2 )∂y + (x−π)∂x + (x′−π)∂x′
�� �

|n|≥2

e−inyfn(x)fn(x′)

π2n2(n2 − 1)

− α0

∞
�

m=1
Res
� (h − 1/2)

π tan(πh/2)
kc(h)

1 − kc(h)
Γ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ)
�

h=hm
(3.149)

For χ > 1 we have the same formula except that we need to replace Γ(h)2

Γ(2h)F(h, h, 2h, χ) →
Ψh(χ) on the second line, as in (3.88). We defined α0 in (3.50), kc in (3.73), x, x′, y in
(3.126), and fn in (3.109). αK is plotted in figure 9.
In the region χ < 1, the first line is actually independent of y. It encodes the contribu-
tion to the four point function from a conserved operator, T, essentially the Hamiltonian
of the theory. This term is not conformally invariant. The second line represents a tower
of other operators that do contribute in a conformally invariant way. The dimensions are
determined by kc(hm) = 1.6

The expression (3.149) is very convenient for analyzing the OPE limit.
If we are
interested in deriving the expression for the chaos limit, then we can start from the version
of (3.149) for χ > 1. We then replace the sum over residues by small circle contour integrals
around each point. Then we pull the contours out to h = 1

2 +is. In the process we pick up
residues at h = 2k, k ≥ 1. We have a double pole at h = 2 and single poles for k > 1. In
the case of the single poles, we replace kc(h) by the retarded kernel kR(1 − h), see (3.98).
Again, we now express those contributions in terms of small contour integrals around these
points and shift the contour to h = 1

2 +is. After we do this, we pick up a residue at h = 2.
Thus, the final contribution involves the difference between the residues of the kc kernel
and the retarded kernel

α0Res
�
(h − 1

2)

π tan(πh/2)

�
kc(h)

1 − kc(h) −
kR(1 − h)

1 − kR(1 − h)

�
Ψh(χ)
�

h=2
(3.150)

This expression contains terms going like χ−1 log χ and χ−1. These should be added to
similar terms that arise from the terms involving derivatives in (3.149). The total te
2πt

β

6We should also note that if (3.139) is not true, we will have further contributions, probably including
a “real” dimension two operator.

41


term, which contains the correction to the Lyapunov exponent, has the form (3.165) which
leads to (3.166). The logarithmic term that comes from the pole involving 1/(1−kc) cancels
the one coming from the derivatives in (3.149).

3.6
The chaos exponent at finite coupling

3.6.1
The retarded kernel

One way to compute the correlator in the chaos limit is to take the exact Euclidean answer
and continue it. This is the approach we took so far in the paper. If we are only interested
in getting the asympotic rate of growth, we can take a simpler approach, used by Kitaev
in [8]. We will consider an out-of-time-order correlation function in real time, where the
fermions are separated by a quarter of the thermal circle:

F(t1, t2) = Tr
�
y ψi(t1)y ψj(0)y ψi(t2)y ψj(0)
�
,
y ≡ ρ(β)1/4.
(3.151)

The 1/N piece of F is determined by a set of ladder diagrams on a time contour that
includes the thermal circle and also a pair of real-time folds for the operators ψ(t1) and
ψ(t2). As t1, t2 become large, these folds grow. The asymptotic growth rate of the 1/N
piece of F is determined only by the properties of the ladder diagrams on real-time part
of the contour. To analyze these ladders, we define a retarded kernel

KR(t1...t4) = J2(q − 1)GR(t13)GR(t24)Glr(t34)q−2.
(3.152)

Here, GR are the retarded propagators, which include the sum over insertions on the two
sides of the fold. The function Glr is a Wightman correlator with points separated by half
of the thermal circle in addition to the real time separation. In the conformal limit

GR(t) = 2b cos(π∆)θ(t)

�
π

β sinh πt

β

�2∆
,
Glr(t) = b

�
π

β cosh πt

β

�2∆
.
(3.153)

The growth rate of F is determined by the condition that adding one rung to the ladder will
not change the sum. This means that F must be an eigenfunction of KR with eigenvalue
one:
F(t1, t2) =
�
dt3dt4KR(t1...t4)F(t3, t4).
(3.154)

To solve this, we make a growth ansatz

F(t1, t2) = eλL(t1+t2)/2f(t12)
(3.155)

and then determine the values of λL such that we can find an f that gives an eigenfunction
of KR with eigenvalue one, solving (3.154). In the conformal limit, one can show by direct
integration that we have eigenfunctions and eigenvalues

e−h π

β (t1+t2)

�
cosh π

βt12
�2∆−h,
kR(h) =
Γ(3 − 2

q)Γ( 2

q − h)

Γ(1 + 2

q)Γ(2 − 2

q − h).
(3.156)

42


This agrees with the definition of kR(h), given previously in (3.98). As we noted there,
the only solution to kR(h) = 1 is h = −1, which gives λL = 2π

β . One might have expected
to also find subleading growth rates in the conformal limit, corresponding to different
families of eigenfunctions. Such eigenfunctions exist, but one can check that the next
largest allowed value of λL is zero. This explains why every growing term in the chaos
limit was growing at this rate, including various terms that were subleading to the enhanced
contribution Fbig.

3.6.2
Large q

In the large q model we can use this retarded kernel to find the growth exponent λL at all
values of the coupling. From (2.14) and the definition of GR in (2.13), together with the
analytic continuation to real times τ → β/2 + it of (2.18), we find that

GR(t) = θ(t),
qJ2Glr(t)q−2 =
2π2v2

β2 cosh2( πv

β t).
(3.157)

where v was defined in (2.19). v goes from zero at weak coupling to one at strong coupling.
Substituting (3.157) into the formula for the retarded kernel, and then taking derivatives
∂t1∂t2 of equation (3.154) with the ansatz (3.155), we get
�λ2
L
4 − ∂2
x

�
f(x) =
2π2v2

β2 cosh2( πv

β x)f(x).
(3.158)

After rescaling the x variable, the equation becomes

−
�λLβ

2πv

�2
˜f(˜x) =
�
−∂2
˜x −
2

cosh2 ˜x

�
˜f(˜x),
˜x = πv

β x,
˜f(˜x) = f(x).
(3.159)

This is a Schrodinger problem for a particle in the −2/ cosh2 ˜x potential. There is a single
bound state ˜f ∝ 1/ cosh ˜x. The energy of this state is minus one, which gives the exact
growth exponent

λL = 2π

β v.
(3.160)

At weak coupling we have λL ≈ 2J , and at strong coupling we have λL ≈ 2π

β [1 − 2/(βJ )].
We give a plot of λL in figure 11.

3.6.3
General q

For general q, we do not have an exact expression for λL at finite coupling. However,
we can relate the first (βJ)−1 correction to the parameter αG, and we can compute the
function at small and moderate βJ numerically. First we discuss the (βJ)−1 correction.
One way to compute this is to do first order perturbation theory in the retarded kernel. The
leading non-conformal correction to KR comes from plugging (3.122) into the definitions

43


10 -1
10 0
10 1
10 2

β J

0

0.2

0.4

0.6

0.8

1
q = 4

10 -1
10 0
10 1
10 2

β J

0

0.2

0.4

0.6

0.8

1

λ*β/2 π

large q

Figure 11: Left: the exact λL in the large q model. Right: λL for q = 4. The red
curve shows the formula (3.166) in a region of reasonable validity. The circles are exact
values, obtained by numerically solving real-time Schwinger-Dyson equations and then
diagonalizing the retarded kernel. Note that the x axis is βJ , not βJ.

GR(t) = [G(it+ϵ)−G(it−ϵ)]θ(t) and Glr(t) = G(it+β/2) to get the corrected propagators

δGR
GR
= − αG

βJ

�

2 −
π tan π

q + 2πt

β

tanh πt

β

�

,
δGlr
Glr
= − αG

βJ

�
2 − 2πt

β tanh πt

β

�
.
(3.161)

Rather than taking this direct approach, we will use the results derived earlier in this
section to get the answer by a different method. From (3.136), the leading term in the
chaos limit behaves like
F(t)

G(π)G(π) ≈ −3α0βJ

2παK
e
2π
β t.
(3.162)

If we correct the growth exponent to λL = 2π

β + δλL, and expand to linear order in δλL,
we expect a term linear in t times the growing exponential:

Fexpect(t)
G(π)G(π) = − (tδλL) · 3α0βJ

2παK
e
2π
β t.
(3.163)

We expect δλL to be of order (βJ )−1, so this term is of order one at large coupling. We
found a term exactly of this type when we analyzed the double pole in the chaos limit of
the Fh̸=2 function. The contribution that contains the log is

Fhave(t)
G(π)G(π) = −
3α0

k′
R(−1)π2∂hΨh(χ)|h=2
(3.164)

≈
3α0

2πk′
R(−1)
2π
β te
2π
β t
(3.165)

44


where we took the large t limit in the second line, using (3.96) and (3.100). Comparing
with (3.163), and rewriting αK in terms of αG using (3.125), we find

λL = 2π

β

�
1 − −k′(2)

k′
R(−1)
qαG
βJ + ...
�
.
(3.166)

We have checked that this agrees with the direct method described above.
The correction is always negative. It is consistent with the large q exact result. Evalu-
ating the derivatives and plugging in the numerical value for αG, we have that when q = 4

−k′(2)
k′
R(−1)
qαG
βJ ≈ 4.28

βJ ≈ 6.05

βJ ,
(q = 4).
(3.167)

When q approaches two, the correction diverges, like

−k′(2)
k′
R(−1)
qαG
βJ =
6π

(π2 − 6)(q − 2)
1
βJ ,
(q → 2).
(3.168)

This divergence seems to be consistent with the fact that the q = 2 the model is free, so
the chaos exponent must vanish for any value of βJ .
Another approach to computing λL is to numerically solve the real-time Schwinger-
Dyson equations to find GR and Glr, and then use binary search to find the largest value
of λL such that there exists an eigenfunction f(t12) that satisfies (3.154). This works well
for small and moderate βJ . In figure 11 we plot some data points for q = 4 computed
this way. They appear to match smoothly to the large (βJ ) result (3.166). We will give
a few more details about this approach in appendix G.

4
The effective theory of reparameterizations

In this section we will discuss the effective action of the model.
This gives a second
perspective on the computation of the four point function that makes some features clearer,
such as the physical interpretation of the enhanced h = 2 contribution. It also allows us
to connect the specific heat term in the free energy to the ladder kernel.
The effective action of the model is derived by starting with the original fermion path
integral and doing the Gaussian integral over the disorder. This gives a bilocal action for
the fermions. One can integrate out the fermions after introducing a field �G(τ1, τ2) and a
Lagrange multiplier field �Σ that sets �G equal to
1
N
�

j ψj(τ1)ψj(τ2). We are left with the
nonlocal action [1]

S
N = −1

2 log det(∂t − �Σ) + 1

2

�
dτ1dτ2

�
�Σ(τ1, τ2) �G(τ1, τ2) − J2

q
�G(τ1, τ2)q
�
(4.169)

for �G, �Σ. This is an exact rewriting of the theory. Because of the Lagrange multiplier
constraint, we can compute the four point function of fermions (3.40) in the �G, �Σ variables

45


as
1
N 2
�

ij
⟨ψi(τ1)ψi(τ2)ψj(τ3)ψj(τ4)⟩ =
�
d�Σd �G e−S �G(τ1, τ2) �G(τ3.τ4).
(4.170)

The action has a saddle point at the solutions G, Σ of the Schwinger Dyson equations
(2.6).
(Note that �G, �Σ denote the integration variables in (4.170), while G, Σ are the
classical solutions to the action (4.169)). Evaluating the integrand at this saddle point
gives the disconnected part of the four point function. We can also consider fluctuations.
It is convenient to define the fluctuations g, σ so that we have �G = G + |G|
2−q

2 g and
�Σ = Σ + |G|
q−2

2 σ. Notice that the measure is invariant d �Gd�Σ = dgdσ. Expanding the
action to second order in g, σ and using the saddle point equation G = (∂τ − Σ)−1 to
simplify, we find

S
N = −
1

4J2(q−1)

�
dτ1...dτ4 σ(τ1, τ2) ˜K(τ1...τ4)σ(τ3, τ4)

+ 1

2

�
dτ1dτ2

�
g(τ1, τ2)σ(τ1, τ2) − 1

2J2(q−1)g(τ1, τ2)2
�
.
(4.171)

Here, ˜K is the symmetric ladder kernel defined in (3.47). We can integrate out σ, getting
an action just for g:
S
N = J2(q−1)

4
g · ( ˜K−1 − 1)g.
(4.172)

We can use this to get the 1/N term in the four point function (4.170), by replacing
both factors of �G in the integrand by |G|
2−q

2 g and then doing the Gaussian integral with
an appropriately chosen contour. This immediately gives the expression (3.104) that we
previously derived from the Feynman diagrams.
The expressions written so far are valid at any energy. When we go to low energies and
we use the conformal expressions, Gc, Σc in order to evaluate the kernel ˜K, then we find
that the action is zero when evaluated on fluctuations that are reparameterizations of the
conformal correlator, as in (3.107). This is because these fluctuations are eigenfunctions of
the kernel with eigenvalue one (3.108). More conceptually, it is because the action (4.169)
is reparametrization invariant (under (2.8)) if we drop the ∂t term inside the determinant.
Notice that even though the action is reparametrization invariant, the solution Gc is only
invariant under the SL(2, R) subgroup. Thus we can view reparametrization invariance
as an emergent symmetry of the infrared theory which is spontaneously broken by the
conformal solution Gc. The zero modes in the action can be viewed as Nambu-Goldstone
modes for the spontaneous breaking of the full conformal symmetry down to SL(2, R).
Note that we could consider an alternative model which does not have a reparametrization
invariance, then we do not get any enhanced contribution in the IR, see appendix (H).
We can now include the leading non-conformal to the action (4.172), which is deter-
mined by the first order shift in the h = 2 eigenvalues of the kernel (3.124). This will
provide a non-zero action for these reparametrization modes. To compute this, we con-
sider a small reparameterization τ → τ + ϵ(τ), and evaluate the action on δϵGc. It is

46


2
4
6
8
10
12
14
16
18
20
q

0

0.005

0.01

0.015

αS

Figure 12: The coefficient of the Schwarzian action αS is plotted. The blue curve is the
value given by the previously computed αG. The circles are values inferred from (5.181)
and the numerical evaluation of the specific heat c. The agreement is a check that the
Schwarzian action is correct nonlinearly, not just for small reparameterizations.

convenient to work in frequency space for ϵ, and to use that reparameterizations δϵGc are
proportional to the h = 2 eigenfunctions of the kernel. We get an action proportional to
n2(n − 1), where n labels the Matsubara frequency. This factor arises from the product
of the |n| in the eigenvalue shift and the |n|(n2 − 1) in the normalization of the h = 2
eigenfunctions. The result, in position space, is

S
N = αS

J

� β

0
dτ 1

2

�

(ϵ′′)2 −
�2π

β

�2
(ϵ′)2
�

,
αS ≡
αK

6q2α0
= q|k′
c(2)|αG
6q2α0
.
(4.173)

This action for ϵ is local, even though the original action is nonlocal. This is reasonable
because the breaking of reparameterization invariance is a UV effect. In fact, the action
that we get could have been guessed by standard effective field theory reasoning: it is
the simply the expression of lowest order in derivatives that vanishes for global SL(2)
transformations. It must vanish in that case because the correlator is SL(2) invariant,
δSL(2)Gc = 0 is zero. Notice that these SL(2) reparameterizations should not be thought
of as zero modes, they simply are not part of the functional integral over G.
Therefore the emergent conformal symmetry is both spontaneously broken by the in-
frared solution Gc as well as explicitly broken, which gives a small action (4.173). It is
small in the sense that it formally vanishes as J → ∞. On the other hand, notice that it
is large in the sense that it is of order N. To get a reasonable theory we need to include the
effects of this breaking. This pattern of symmetry breaking is reminiscent to the pions in
QCD, the chiral symmetry is both spontaneously and explicitly broken (by the quark mass
terms). Thus (4.173) turns the reparametrization modes into Pseudo-Nambu-Goldstone
bosons.
The enhanced contribution Fbig from (3.126) can now be understood in a simple way.
It is the result of the part of the functional integral (4.170) that consists of summing over

47


reparameterizations of the circle weighted by the action (4.173). This leads directly to
(3.127).
We would like to generalize the action (4.173) to finite reparameterizations τ → f(τ).
It is convenient to start with the zero temperature case where both f and τ are coordinates
on the line. f is a coordinate on the “straight” line where the IR correlator is a pure power,
and τ is a coordinate on the reparameterized line. Near any point (we take the origin),
one can write

f(τ) = f(0) + f ′(0)
�
τ + 1

2
f ′′(0)
f ′(0) τ 2 + ...
�
.
(4.174)

For small τ we have a small reparameterization with ϵ′ = 0 and ϵ′′ = f ′′/f ′, followed by
a scaling and translation. The scaling and translation have no effect on the correlator on
the zero temperature line, so we can generalize

1
2

�
dτ(ϵ′′)2 → 1

2

�
dτ
�f ′′

f ′

�2
(4.175)

which up to a total derivative implies that the action can be written as

S = −N αS

J

�
dτ{f, τ},
{f, τ} ≡ f ′′′

f ′ − 3

2

�f ′′

f ′

�2
.
(4.176)

In the second step we introduced the Schwarzian derivative {f, τ} and used integration by
parts. Note that the Schwarzian derivative is invariant under SL(2) symmetry f → (af+b)

(cf+d).
This is an exact symmetry since the zero temperature Gc is exactly invariant under this
transformation.
To get the action for reparameterizations of the circle, we consider the transformation

f(τ) = tan
�πτ

β

�
(4.177)

which maps the circle to the line. Already for this transformation we get an interesting
result. Inserting (4.177) into the Schwarzian action (4.176) we get a finite temperature
correction to the free energy

−βF ⊃ NαS

J

� β

0
dτ{tan πτ

β , τ} = 2π2αS
N
βJ .
(4.178)

At large q, we have αS =
1

4q2 (αK = 3 and α0 = 2), and this agrees with the term found
previously in (2.32). For q = 2 we get αS =
1

24π ( α0 = π2, αK = π), and again we
agree with (2.34). In figure 12 we give a plot of αS and indicate a numerical check of this
formula. Note that while this action nicely explains the near extermal entropy, it says
nothing about the zero temperature entropy.

48


If we are interested in further reparametrizations of the circle, τ → g(τ), we can use
the composition law for the Schwarzian derivative {f(g(τ)), τ} = (g′)2{f, g} + {g, τ} to
obtain

S
N = −αS

J

�
dτ{tan πg(τ)

β
, τ} = αS

2J

�
dτ

��g′′

g′

�2
−
�2π

β

�2
(g′)2
�

.
(4.179)

Writing g(τ) = τ + ϵ(τ) and expanding in ϵ we get both of the quadratic terms in (4.173).

5
The density of states and the free energy

The large N free energy is determined by evaluating the ˜G, ˜Σ action (4.169) on the saddle
point values G, Σ. In a low temperature expansion, we have

log Z = −βE0 + S0 + c

2β + · · · ,
(5.180)

where the ground state energy, entropy and specific heat are all proportional to N. The
ground state energy will not be important for our discussion. The zero temperature entropy
is given for general q by (2.33). The specific heat is determined by (4.178) as

c
2 = 2π2αS
N
J .
(5.181)

For the case q = 4 we have c ≈ 0.396 N/J.
All of the terms in (5.180) are proportional to N. There is an important order-one
correction to this free energy which we can compute from the determinant of the quadratic
action (4.171). The log determinant of this action gives a term7

−βF ⊃ −1

2

�

h,n
log[1 − k(h, n)].
(5.182)

We get an interesting log βJ term from the h = 2 modes, which have eigenvalues close to
one. Substituting in the corrected eigenvalues (3.124), we get

−βF ⊃ −

∞
�

n=2
log n

βJ + const. → #βJ − 3

2 log βJ + const.
(5.183)

This sum is divergent, but the divergence will presumably be cut off at n ∼ βJ, where
one expects higher order effects to make the eigenvalue k(2, n) small. This will lead to a
term proportional to βJ with an unknown coefficient; this is a correction to the ground

7This contribution to the free energy was first pointed out by J. Polchinski and A. Streicher, using
Feynman diagrams. They also noted that the sum over near-zero-modes would lead to a log term [27].

49


state energy. The special feature of the h = 2 sum is that we also get the finite log piece
indicated on the right. This can be extracted by zeta function regularization or the Euler
MacLaurin formula. The logarithm means that the partition function is proportional to
β−3/2 at large β.
We can also get this factor of β−3/2 from the action (4.173) as follows. We also do the
functional integral over ϵ(τ). However, we need to recall that we are not integrating over
SL(2, R) transformations. Thus, we need to divide the integral by the volume of SL(2, R),
since we should view SL(2, R) as a gauge symmetry. This will result in the insertion of
factors of the form δ(ϵ(0))δ(ϵ′(0))δ(ϵ′′(0)) in the functional integral. When we rescale ϵ to
get rid of the coefficient of the quadratic action in (4.173), we will find a factor of (βJ )−3/2

from the three delta functions.
The integral over all the non-zero modes with h ̸= 2 will produce a βJ divergence that
corrects the ground state energy, plus a β-independent factor which can be absorbed as a
1/N correction to S0 in (5.180).
With this information, we can now compute the density of states by doing an inverse
Laplace transform to the partition function. It is convenient to subtract the ground state
energy, so that from now on E indicates the energy above the ground state. The integral
is

ρ(E) =
1
2πi

�

γ+iR
dβZ(β)eβE ∝ eS0
�
dβ

(βJ)
3
2 eβE+c/2β ≈

�

2π
cJ3eS0+
√

2cE
.
(5.184)

In the final step we approximated the integral by saddle point, valid for cE ≫ 1. It is
interesting that the determinant from the saddle point integral cancels the factor β− 3

2 from
the one-loop free energy, so ρ(E) approaches a constant at low energy, in this approxima-
tion. We can also compute the integral for small cE, where it becomes ≈ 4
�

πE/J3eS0.
The
√

E vanishing is interesting, because it agrees with the behavior that one gets for the
spectrum of a random Hamiltonian. However, we cannot trust the computation for such
small values of E; at the crossover point cE ∼ 1, the saddle point in the β integral is at
β ∼ c ∼ N/J. At such low temperatures, the Schwarzian action discussed in the previous
section stops being semiclassical, and our analysis would need to be improved8.
A somewhat complementary approach to the spectrum is to exactly diagonalize the
Hamiltonian (2.2) for small values of N. The Majorana fermion operators ψi are just
matrices that satisfy {ψi, ψj} = δij. In other words, they are Dirac gamma matrices. To
represent the model with N fermions, we need a Hilbert space of dimension 2N/2. We were
able to study up to N = 32 without any special techniques. We give a plot of the binned
spectrum for N = 32 in figure 13.

8As a side remark, note that, as we discuss in[26], the effective action (4.173) also appears for near
extremal black holes. Thereore in such cases the computation of (5.184) will be valid. In that case, the
fact that the density is constant at low energies is consistent with the fact that BPS black holes can have
a large degeneracy at exactly zero energy. For non-BPS black holes we could have further corrections that
might remove the large degeneracy at exactly zero energy.

50


-2
-1
0
1
2
Energy, in units of J

0

50

100

150

200

250

300

350

400

Number of eigenvalues

Eigenvalues for N=32, plotted with 300 bins

Figure 13: The spectrum for a single realization of the q = 4 model with N = 32 fermions.

One obvious feature of the plot is that there is no scale-invariant divergence ρ(E) ∝ 1/E
or δ(E) at low energy. Instead, the density goes smoothly to zero. A naive reading of the
plot suggests that the spectrum vanishes as Ep with p near one. The zero temperature
entropy does not reflect any actual degeneracy, only a large density of states near the
ground state. From this perspective, a completely random Hamiltonian on a system of
N qubits also has a zero temperature entropy, S0 = N log 2, from the density ρ(E) ∝
�

E(E − 2)2N. This gives a low temperature free energy log Z = N log 2 − (3/2) log β.
In fact, from the plot (13), the density of states in SYK does not look too different
from the random matrix semicircle. It is important to note, though, that if we increase
N the density in the central region will be growing much faster than near the edges.
Near the center, we expect the density characteristic of the infinite temperature entropy,
ρ ∼ 2N/2 ≈ e0.35N, while near the edges we expect eS0N ≈ e0.23N. By diagonalizing the
Hamiltonian for different values of N between 24 and 32, and counting the number of levels
within bands of width 0.3J, we found the best fit e0.33N near the center and e0.24N near
the edge, in reasonable agreement with large N expectations. Note that the Hamiltonian,
while containing of order N 4 random elements is not as random as a general random matrix
in Hilbert space, which would contain 2N random elements.

6
Towards a bulk interpretation

A natural starting point for a bulk interpretation is the action (4.169). Due to the large
factor of N, this looks like a classical system for the fields �Σ and �G. One of them can be
easily eliminated, so we really have one field which is a function of two variables. Thus
we seem to have a field theory defined on a two dimensional space. It is natural to think
of the average of the two times as a time and the difference as a new dimension. The

51


solution to the Schwinger Dyson equations gives us a classical background for this system,
and then we have fluctuations governed by a quadratic action of the form (4.172). The
computation of the four point function of the fermions can be viewed as the computation
of the propagator for this bilocal field and it involved inverting the operator (1/ ˜K − 1)
(4.172).
In the large βJ limit, we have seen that there is a dominant mode associated to the
emergence of a conformal symmetry which is both spontaneously and explicitly broken. A
conformal symmetry is easily obtained if we consider AdS2 gravity. When we regularize
the space and we introduce some boundary conditions we get a boundary mode that is
the same as the one parametrized by the function f(τ) discussed above. The AdS2 metric
preserves explicitly an SL(2, R) group. This metric is spontaneously breaking the rest of
the reparametrizations. The boundary mode is the corresponding Goldstone boson and
it implies that pure AdS2 gravity is not well defined, if we want to have any non-trivial
excitation [34, 35, 7]. However, when AdS2 arises from a higher dimensional theory, there
is always a coupling to a dilaton which is not constant on AdS2. This explicitly breaks the
conformal symmetry and it gives rise to an action for the modes parametrized by f(τ).
The details of this will be discussed in a separate publication [26] and the discussion is
very closely related to the analysis in [7]. In summary, both the mode parametrized by
f(τ) and its action are reproduced by any near AdS2 (or NAdS2) geometry. This feature
is insensitive to the precise details about the type of matter we can have in AdS2. It results
purely from the emergence of the conformal symmetry and its slight breaking.
In order to elucidate the kind of matter we have in the dual of SYK we need to look at
the other propagating modes contained in the field G(τ1, τ2). Fortunately, for all the other
modes we can use the SL(2, R) symmetry to describe them. The propagating modes can
be read off from the OPE expansion of the fermion four point function. Their conformal
dimensions are the solutions to kc(hm) = 1 and were discussed in section 3.2.6. We get an
infinite tower of dimensions which asymptotes to (3.93) (3.94) at large values of m. This
asymptotic form of the dimensions has a structure that looks like a two particle state in
AdS2. However, it is important to note that the shift in dimensions is of order one, and
not order 1/N. Therefore, we cannot view these states as a two particle state of fermions
in the bulk with weak, gravitational strength, interactions.
This tower of particles is
reminiscent of a string theory with a string scale of order the AdS radius. In fact, it also
looks similar to what we would get in an O(N) model, where we get a state for each spin
and the number of single string states does not exhibit an exponential growth with energy
(Hagedorn behavior). Here all members of the tower are getting an order one shift in their
dimensions9. In two dimensions we do not have a clear notion of spin, but if we define
spin by the contribution to the correlator in the chaos region, then they have spins S > 2
(S = 2 is for the h = 2 states related to the reparametrizations).

9In the free O(N) models we get such a tower with dimensions given exactly by the sum of dimensions
of the two elementary free field components ψ∂2n+1ψ. In the case of the Gross Neveu model in 2 + 1
dimensions, interactions give the lowest member gets an anomalous dimension. Namely ψ2 has a shift
from ∆free = 2 → ∆ = 1.

52


6.1
Comments on kinematic space.

In this subsection we expand a bit on the comment on the relation between the two times
of G and the two variables of a bulk field. Recently, this was further explored in [18].
In general we can define

t = τ1 + τ2

2
,
σ = τ1 − τ2

2
(6.185)

At this point this is just a simple relabeling of the times. Written in this way, we see that
some of the terms in the action (4.169) become local in the t, σ space. But the Pfaffian
term is still non-local.
Furthermore, in the IR region, the Casimir operator acting on the two times τ1, τ2 has
the form
C12Φ = −(τ1 − τ2)2∂τ1∂τ2Φ = σ2(−∂2
t + ∂2
σ)Φ = ∇2Φ
(6.186)

where Φ(t1, t2) =
ˆδG
Gc where ˆδG is a small fluctuation around the classical solution of the
action (4.169), which in the IR is given by the conformal answer Gc (2.9). We see that
this looks like the laplacian in AdS2, in coordinates ds2 = −dt2+dσ2

σ2
, and in units where
the AdS radius is set to one. This space was defined in a purely kinematic way by using
general properties of the conformal group. It was called kinematic space in [29, 30]. Note
that even the quadratic action (4.172) for Φ is highly non-local. It involves 1/kc(h) − 1
which is a complicated function of the casimir, C = h(h − 1) = ∇2, see (3.73). We should
think of this Φ as describing many degrees of freedom since there are many solutions of
1/kc(h) − 1 = 0, describing the tower of states in section (3.2.6).
It is amusing to note that the expression for the energy given in (2.27) looks like an
ADM like expression for the energy in terms of a property of the solution at the boundary
of the geometry, namely τ1 = τ2 or σ = 0.
In the Euclidean theory we expect the bulk to be H2. However, the kinematic space
defined as above, through the Casimir operator, continues to be a Lorentzian signature
space. This is a general feature of the Casimir operator acting on bilocal fields as has
been used recently in [30]. We can view the space as dS2, or AdS2 with time periodically
identified, depending on the overall sign we choose for this metric. More explicitly, in
the Euclidean finite temperature theory, we have the times τ1 and τ2 which are periodic
variables. When we define the sum and the difference we get (for β = 2π)

τ = τ1 + τ2

2
,
σ = τ1 − τ2

2
,
C = sin2 σ(−∂2
τ + ∂2
σ)
(6.187)

which looks like the wave equation on global dS2, or AdS2 with time periodically identified.
In fact, the funny set of eigenfunctions that we needed to sum over in e.g. (3.83) has a
simple interpretation in AdS2 or dS2. They are the set of normalizable solutions of this
wave equation. We had a further anti-symmetry restriction on the wavefunctions, which
amounts to antisymmetry under an anti-podal transformation in this dS2 or AdS2 space.
It would be interesting to see if some variation of this model has a de-Sitter interpretation.

53


Finally, this relation to kinematic space suggests that the usual bulk of AdS/CFT requires
a further inverse X-ray or Radon transform. We make a few more comments on this in
appendix I.

6.2
The fermions

So far, we have avoided the “elephant in the room”, which are the N boundary fermions.
One can question whether these should correspond to N bulk fermions or not. Before
trying to answer this question, let us recall the special case of q = 2. In that case the N
boundary fermions give rise to just one “bulk” fermion as follows. After we diagonalize
the random mass matrix by an orthogonal transformation we find that φi = �

m rimψm,
where ψm is a fermion with a definite mass (or frequency)10. In the large N limit, the
distribution of masses is nearly continuous and we can view it as an extra dimension. So,
in this case we see that the different boundary fermions ψi give rise to different parts of the
bulk fermion field ψ. This should be the case, since a bulk fermion has many independent
creation and annihilation operators.
Before we continue, let us also make another general comment. One can imagine getting
rid of the fundamental fermions by viewing the couplings jijkl as dynamical with very slow
dynamics so that they are effectively constant [19]. At the order we are working, this gives
the same equations 11. Once we make the couplings jijkl dynamical, we can gauge the
O(N) symmetry. Naively this seems to remove the fermions from the spectrum so that we
do not need to discuss them further. However, we continue to have a related operator of
the form
O(τ, τ ′) = ψi(τ)
�
Pei
� τ
τ′ A� j

i ψj(τ ′)
(6.188)

The one point functions of this operator ⟨O(τ, τ ′)⟩ ∼ NG(τ, τ ′) continue to display an
SL(2, R) invariant form with dimensions ∆.
We can now wonder what the interpretation of such an operator in the bulk is. This
question was studied in detail for the related case of a matrix model in the non-singlet
sector in [37]. In that case, the corresponding state was a folded closed string coming from
the boundary into the bulk. In our case we can imagine a similar explanation in terms of
an open string that comes in from the boundary.
Then we can view the propagating states as strings oscillating in AdS2, see figure 14.
This is just a picture, since we are not displaying the precise string theory background.
We can now go back to the ungauged model. At the level that we treated it so far, it
has a global O(N) symmetry. And it is tempting to think that whether or not we gauge the
symmetry is some operation purely at the boundary. Therefore even for the O(N) model

10For positive m we have a complex creation operator and for negative m the corresponding annihilation
operator.

11Except that there is an additional contribution to the free energy from the j fields, of the form N q log β
from the j fields. We thank S.H. Shenker for this comment. This model is structurally reminiscent of the
2+1 dimensional O(N) theories with fundamental bosons and fermions studied in [36].

54


time

(a) 
(b)

Figure 14:
(a) Particle with a string going to the boundary. (b) Pair of particles in the
bulk with a string connecting them. They are oscillating in global AdS2. We can view the
string as fundamental or as a color electric flux of an O(N) gauge field.

we expect to see that the fermion contains a string going into the bulk. In that case, the
index i remains at the boundary of the bulk, we can view it as a Chan Paton index at
the boundary. But in the bulk we would have just a single string, and no uncontracted
indices. An alternative point of view, motivated by the global SO(N) symmetry, would be
to put O(N) gauge fields in the bulk and charged fermion fields in the interior. However,
in this case large N counting would give us a coupling gSO(N) ∼ 1, which, together with
the factor N, gives us a strong coupling. The fermions would be joined by a color electric
flux, which looks conceptually similar to the strings discussed above. In fact we would get
something like the ‘t Hooft model [38], but in AdS2.
Further work would be needed to check whether this is the right interpretation.
When the couplings are random but fixed, it would be interesting to understand the
corresponding bulk dual. Since the corrections to the leading answer would come at higher
orders in the 1/N expansion (at order 1/N q−1), it seems natural to suspect that they would
be associated to effects that are sensitive to quantum corrections.

6.3
Scrambling for near extremal black holes and its stringy
corrections

One of the original reasons for interest in the SYK model was the fact that it has the
maximal chaos exponent λL = 2π/β. This is a necessary condition to have a theory dual
to gravity, and it was thought that it might also be sufficient. A piece of evidence for this
idea was that stringy corrections decrease λL by an amount proportional to ℓ2
s/L2 [11]

λL = 2π

β

�
1 − ℓ2
s

L2 + · · ·
�
(6.189)

where L is a curvature scale at the horizon, so it seems that theories with maximal λL
should not have large strings. However, the operator dimensions hn that we found in the
OPE suggest that the bulk theory dual to SYK has a tower of light fields roughly similar

55


to a string spectrum with ℓs ∼ RAdS. This seems to be a counterexample to the idea that
maximal chaos implies a gravity dual. This motivates us to examine in more detail the
form of the scale L that was appearing in (6.189).
Let us briefly review the shock wave calculation that gives the chaos limit of the four
point function. We have an out of time ordered four point function with two pairs of
operators. One pair is at time zero, and the other is at time t. The growing part of the
correlator is given by the phase shift of the bulk field associated to the t = 0 pair as it
crosses a shock sourced by the other pair. The general form of the shock wave plus static
black hole metric is

ds2 = −a(uv)dudv + b(uv)dxidxi + h(x)δ(u)du2
(6.190)

where we have added D − 2 extra flat dimensions. Einstein’s equations give

1
2
�
− ∂2
i
b −
� ∂u∂vb

ab
�
(D − 2)
�
h(x)δ(u) = 8πGNTuu
(6.191)

1
2
�
−∇2 + φ′′(0)

φ(0)
�
h(x)δ(u) = 8πGNTuu
(6.192)

where φ ∝ b
D−2

2
is the “dilaton”, or the coefficient of the two dimensional curvature in
the action
�
φR(2) after dimensional reduction on the extra flat coordinates. And φ′′ is
the second derivative with respect to proper distance from the horizon, evaluated at the
horizon. Now we integrate this over the transverse space and also over u in a neighborhood
of the horizon. We get
1
2
φ′′

φ Ah = 8πGNPu.
(6.193)

where A is the area of the horizon, and h is the zero mode of the shock wave profile. Pu
is the momentum of the quantum associated to the pair of operators at time t, which is
Pu ∼ (∆/RAdS)e
2π
β t. Dividing both sides by 2GN gives

φ′′

φ Sh = 4πPu
(6.194)

where S is the entropy of the black hole. The phase shift for the other field crossing this
shock is

δ ∼
∆

RAdS
h ∼
� ∆2

R2
AdS

φ
φ′′

� 1

S e
2π
β t.
(6.195)

We should regard the quantity in brackets as being the βJ enhancement. In fact, for
near extremal black holes, we have that the profile of the dilaton at the horizon has the
form φ = φ0 + γ cosh
ρ

RAdS2 , where φ0 gives the extremal entropy and γ the near extremal
entropy, with γ ≪ φ0. So we can write

R2
AdS
φ′′

φ = S − S0

S0
−→
∝
1

q2βJ
(6.196)

56


where on the left side we have a gravity expression and on the right hand side we write the
quotient of entropies that we have in the SYK model. Here we have simply reproduced the
leading part of the answer from a gravity computation. We will connect it more clearly to
the reparametrizations in [26]. The prefactor enhancement of the butterfly effect for near
extremal black holes was noticed previously in [39, 40].
Now we consider stringy corrections to the chaos exponent, following [11]. We read
these off from (6.192). We reintroduce the extra dimensions, substitute

k2 → k2 + φ′′

φ
(6.197)

in the Regge behavior of the flat space string amplitude s2−ℓ2
sk2/2, and then take k to zero.
This gives an effective spin which is

j = 2 − ℓ2
s
2
φ′′

φ = 2 −
ℓ2
s

2R2
AdS

(S − S0)

S
.
(6.198)

So even if the string length is large, there is another parameter suppressing the correction
to

λL = 2π

β (j − 1) = 2π

β

�
1 −
ℓ2
s

2R2
AdS

S − S0

S
+ · · ·
�
.
(6.199)

Note that using the expression for the ratios of entropies in (6.196) we get an estimate for
the SYK model of a correction of order
1

q2βJ (provided ℓs ∼ RAdS). This is indeed what
we found in (3.166) up to q dependent factors.
It is a little surprising that the change in the Regge spin can be small, despite the
presence of light strings. The right interpretation seems to be that the gravity contribution
gets a βJ (or near-extremal) enhancement, but the higher stringy exchanges do not. So
gravity dominates and we have a spin near two.

7
Brief Conclusions

The SYK model is an interesting quantum mechanical model displaying a spontaneously
and explicitly broken reparametrization symmetry. These features dominate the low energy
properties of the model and are expected to be universal for any large N system with
emergent reparametrization symmetry. One motivation to study this model is that near
extremal black holes also display this pattern of symmetry breaking [26]. We also expect
that this will be relevant to other condensed matter physics models.
This symmetry
breaking pattern gives rise to several features of the low energy dynamics. First, it gives
rise to a specific heat that is linear in the temperature.
It also gives rise to a large
contribution to the four point function, which saturates the chaos bound in the out of
time ordered configuration. All these features are expected to be universal features of
systems with emergent reparametrization symmetry or NCFT1s.

57


We also studied several features that are special to this particular model, such as
the spectrum of dimensions of fermion bilinear operators. These suggest that the dual
description should contain a single Regge trajectory with low tension strings in nearly
AdS2 space. We also gave a detailed description of the non-enhanced parts of the four
point function. Several questions remain about the proper holographic interpretation of
this particular model.

Acknowledgements

We thank D. Anninos, A. Kitaev, J. Polchinski, S. Shenker and Z. Yang for discussions.
J.M. is supported in part by U.S. Department of Energy grant de-sc0009988.
D.S. is
supported by the Simons Foundation grant 385600.

A
The Schwinger-Dyson equations and the kernel

We consider the Schwinger Dyson equations (2.6). We treat the iω term as a pertubation.
In fact for an arbitrary perturbation we can write the equations as

G ∗ Σ + G ∗ s = −1 = −δ(t − t′′) ,
Σ = J2Gq−1
(A.200)

where the ∗ stands for G ∗ Σ =
�
dt′G(t, t′)Σ(t′, t′′) and G is the full solution with the
source. The source induced by the −iω term in (2.6) is s = −δ′(τ − τ ′) We can now write
G = Gc + δG, where δG is the perturbation to the conformal solution induced by the
presence of the source.
Expanding the equations to first order, using the second equation to express δΣ in
terms of δG, and convolving with G on the right, we obtain

δG − (q − 1)J2Gc ∗ Gq−2
c
δG ∗ Gc = Gc ∗ s ∗ Gc
(A.201)

(1 − Kc)δG = −
�
dτ ′∂τGc(τ − τ ′)Gc(τ ′ − τ ′′)
(A.202)

where we have used the homogeneous equation Gc ∗ Σc = −δ(τ − τ ′′).
The right hand side has the form sgn(τ−τ ′′)

|τ−τ ′′|4∆ . This has the form of a function like (3.69)
with t0 → ∞ and h → −2∆. Therefore seems that all we need would be to invert 1 − Kc.
However, kc(−2∆) = ∞ (see (3.73)). This would lead to δG = 0. However, we could
have terms that obey (1 − Kc)δG = 0. A formal solution would be an h = −1 mode
which gives the correction δG ∝ Gc
1

|τ−τ ′| (see again (3.69) with τ0 → ∞). This is formally
annihilated by 1−Kc, but it is not actually annihilated because of UV divergences. These
UV divergences arise from the region where the two times in the integral are very close to
each other. This region would be regulated in the full theory and has the form of the right
hand side of (A.202). Furthermore, it has the right J dependence to match the right hand
side. This also shows that in order to compute the full coefficient, we need to know the

58


whole flow, in order to know how the coincident point divergence of the conformal case is
regulated.
Another point of this appendix is to show that the same kernel that appears in the
four point function also appears when we want to compute the corrections around the IR
solution.

B
The kernel as a function of cross ratios

The kernel gives the (n+1)-ladder diagram in terms of the n-ladder diagram as

sgn(τ12)sgn(τ34)

|τ12|2∆|τ34|2∆ Fn+1 (χ) = − 1

α0

�
dτadτb
sgn(τ1a)sgn(τ2b)

|τ2b|2∆|τ1a|2∆|τab|2−4∆ · sgn(τab)sgn(τ34)

|τab|2∆|τ34|2∆ Fn (˜χ)

χ = τ12τ34

τ13τ24
˜χ = τabτ34

τa3τb4
.
(B.203)

We would like to use conformal symmetry to turn this into a one-dimensional integral
equation. We can take τ1 = 0, τ3 = 1, τ4 = ∞, so that ˜χ =
τab
τa−1 and χ = τ2, and then
replace the τb integration variable by ˜χ. The measure is dτadτb = dτad˜χ(1−τa). One finds

Fn+1(χ) = 1

α0

� ∞

−∞

d˜χ
|˜χ|2

� |χ||˜χ|

|χ − ˜χ|

�2∆
sgn(χ˜χ)m(χ, ˜χ)Fn(˜χ)
(B.204)

m(χ, ˜χ) = sgn(χ − ˜χ)
� ∞

−∞
dτ
sgn(τ)sgn(1 − τ)sgn(1 − 1−˜χ

χ−˜χτ)

|τ|2∆|1 − τ|1−2∆|1 − 1−˜χ

χ−˜χτ|2∆ .
(B.205)

The integral over τ can be done by dividing up the region of integration and using
� 1

0

dτ

(1 − xτ)aτ b(1 − τ)c = Γ(1 − b)Γ(1 − c)

Γ(2 − b − c)
2F1(a, 1 − b, 2 − b − c, x).
(B.206)

The answer is

m(χ, ˜χ) ≡








2π

sin 2π∆F(1 − 2∆, 2∆, 1, z) − B2∆
�
1

1−z
�
− B1−2∆
�
1

1−z
�
z ≤ 0
−
2π

z2∆ sin 2π∆F(2∆, 2∆, 1, z−1

z ) +
2π

sin 2π∆F(2∆, 1 − 2∆, 1, z)
0 ≤ z ≤ 1
−
2π

sin 2π∆F(2∆, 1 − 2∆, 1, 1 − z) + B2∆(z−1) + B1−2∆(z−1)
1 ≤ z.

z ≡ 1 − min(χ, ˜χ)

|χ − ˜χ|
,
Bh(x) = Γ(h)2

Γ(2h)xh
2F1(h, h, 2h, x).
(B.207)

We are interested in applying this integral kernel to functions F(χ) with the symmetry
of the four point function. This means that we should have F(χ) = F(χ/(χ − 1)). This
transformation maps the interval between zero and two into the complement on the real
line, so we can restrict our attention to f(χ) with 0 ≤ χ ≤ 2. Using the invariance and
changing integration variables in (B.204), we get a closed equation in this interval:

Fn+1(χ) = 1

α0

� 2

0

d˜χ
˜χ2 Fn(˜χ)
� χ2∆ ˜χ2∆

|χ − ˜χ|2∆m(χ, ˜χ) + sgn(˜χ − 1)
χ2∆ ˜χ2∆

|χ + ˜χ − χ˜χ|2∆m(χ,
˜χ

˜χ − 1)
�
.

(B.208)

59


The expression in brackets times 1/α0 is the kernel Kc(χ, ˜χ) described in (3.59).

C
Representing F0 in terms of Ψh

Using contour manipulations very similar to the one we used to derive (3.88) and (3.90),
it is possible to write a formula for F0 as a sum over residues. There are two differences:
first, we do not have a divergent term at h = 2, so we do not need to subtract it. Second,
since we have kc(h) in the integrand instead of kc(h)/[1 − kc(h)], we are interested in the
poles of kc(h), which occur at values h = 2

q + 1 + 2n. We get the following formulas for
F0. When χ > 1,

F0(χ) = −α0

∞
�

n=0
Res
� (h − 1/2)

π tan(πh/2)kc(h)Ψh(χ)
�

h= 2

q +1+2n
χ > 1.
(C.209)

and when χ < 1 we have

F0(χ) = −α0

∞
�

n=0
Res
� (h − 1/2)

π tan(πh/2)kc(h)Γ(h)2

Γ(2h)χh
2F1(h, h, 2h, χ)
�

h= 2

q +1+2n
χ < 1.

(C.210)
These expressions (C.209) and (C.210) can be checked by numerically evaluating the
residue sums and comparing to (3.80).

D
Writing Ψh(χ) in terms of Ψh,n(θ1, θ2)

By solving the casimir differential equation, one finds that the antisymmetric eigenfunc-
tions of C1+2 with weight ∆ = 1/2 and symmetry under (x, y) → (2π − x, y + π) are

Ψh,n(θ1, θ2) = γh,n
e−iny

2 sin x

2
ψh,n(|x|),
x = θ12,
y = θ1 + θ2

2
.
(D.211)

where the functions ψh,n are the ones appearing in (3.117) and (3.118), but with v = 1 so
that ˜n = n. The norms of the continuum eigenfunctions h = 1/2 + is can be determined
by assuming that ⟨Ψh,n, Ψh′,n⟩ = 2πδ(s−s′), where the inner product is defined in (3.102),
and analyzing the integral near x = 0 and x = 2π, as in (3.78). One finds an expression
involving a product of gamma functions. With these normalizations, we have that

2h − 1

π tan(πh)
Ψh(χ)

(2 sin θ12

2 )(2 sin θ34

2 ) = 2
�

n
Ψ∗
h,n(θ1, θ2)Ψh,n(θ3, θ4),
χ = sin θ12

2 sin θ34

2

sin θ13

2 sin θ24

2
.

(D.212)
which shows that the continuum part of the formulas (3.83) and (3.104) agree. When
we go to the discrete case where h is an even integer, the continuum normalization γ2
h,n

60


diverges for |n| ≥ h, and the factor of 1/ tan(πh) in (D.212) also diverges. The coefficient
of this divergence gives the relation

2h − 1

π2
Ψh(χ)

(2 sin θ12

2 )(2 sin θ34

2 ) = 2
�

|n|≥h
Ψ∗
h,n(θ1, θ2)Ψh,n(θ3, θ4)
(D.213)

where the Ψh,n are now defined with discrete norms so that ⟨Ψh,n, Ψh′,n′⟩ = δhh′δnn′. This
establishes the equivalence of the discrete parts of (3.83) and (3.104). A special case that
we use in the main text of the paper is

Ψ2(χ) = 2
�

|n|≥2

ein(y−y′)fn(x)fn(x′)

|n|(n2 − 1)
.
(D.214)

where x = θ12, y = θ1+θ2

2
, x′ = θ34, y′ = θ3+θ4

2
and χ is defined as in (D.212). The functions
fn were defined in (3.109).

E
Direct approach to the shift in eigenvalue

In this appendix we sketch a second derivation of (3.125), which consists of substituting
G+δG in for the propagators in the kernel, with δG given in (3.122), and then analyzing the
integrals to compute ⟨Ψ2,n, δ �K ·Ψ2,n⟩. When we correct the propagator, we get corrections
to �K of two types. One type is a correction to the rung propagators, see figure 4. In that
case, we can use the fact that Ψ2,n is an eigenfunction of the unperturbed kernel to do
two of the integrals. This gives an expression that is independent of q, up to an overall
multiple (q − 2)αG. Comparing to the q = ∞ case, one finds that in general

δrungk(2, n) = −(q − 2)αG

βJ
3|n|

2 .
(E.215)

The corrections to the rail propagators are not as simple. The change in the kernel is

δrail �K = −J2(q − 1)|G(θ12)|
q−2

2 δG(θ13)G(θ24)|G(θ34)|
q−2

2 + (13 ↔ 24).
(E.216)

Our first goal is to show that ⟨Ψ2,n, δrail �K · Ψ2,n⟩ is proportional to |n|. We can do this
using conformal symmetry. It will be useful to represent the function f0 appearing in
(3.122) as an integral

f0 =
� π

−π
dθ0
| sin θ10

2 sin θ20

2 |

| sin θ12

2 |
.
(E.217)

This implies that δG has the form of an integrated conformal three point function of two
fermions with an operator of dimension minus one. Another useful identity is based on

1
8
sin2 θ12

2

sin2 θ10

2 sin2 θ20

2
=

∞
�

n=2
einθ0e−inyfn(x),
for |eiθ0| < 1,
(E.218)

61


which implies that Ψ2,n is proportional to an integrated conformal three point function of
two fermions with a dimension two operator:

Ψ2,n(θ1, θ2) = γn

4π

� 2π

0
dθ0 e−inθ0
2 sin θ12

2

(2 sin θ10

2 )2(2 sin θ20

2 )2,
γ2
n =
3

π2|n|(n2 − 1).
(E.219)

Here the integral is defined by giving θ0 a small imaginary part iϵ sgn(n).
The shift ⟨Ψ2,n, δrail �K · Ψ2,n⟩ is an integral over four times θ1, ..., θ4 of a product of
propagators and eigenfunctions. The idea is to represent the eigenfunctions Ψ2,n and the
change in the propagator δG using the integral formulas (E.219) and (E.217). This adds
three new integration variables, θa, θb, θc. The complete expression is proportional to the
integral over all seven θ variables of

γ2
nein(θa−θb)
�����
sin θ12

2 sin θ34

2

sin θ13

2 sin θ24

2

�����

2∆
sgn(θ12θ34θ13θ24)

sin2 θ1a

2 sin2 θ2a

2 sin2 θ3b

2 sin2 θ4b

2

�����
sin θ1c

2 sin θ3c

2

sin θ13

2

�����
(E.220)

plus a similar term with (13 ↔ 24). First we consider holding θa, θb, θc fixed and doing
the integral over θ1, ...θ4. The θa and θb variables are the integration parameters in the
representation (E.219) of Ψ2,n. They should be understood as having small imaginary
parts of opposite sign. With this prescription, the integral over θ1...θ4 is convergent, and
has analytic dependence on θa and θb. (The naive divergence of the integral θ13 = 0 is not
present becuase of the sgn(θ13) factor.) Now, the important point is that the integral is
SL(2) covariant, with external weights h = 2 for the θa, θb variables and weight h = −1
for the θc variable. So the answer must be proportional to

γ2
nein(θa−θb)sin θac

2 sin θbc

2

sin5 θab

2
.
(E.221)

We cannot have absolute value signs or sgn functions in this expression, because it must
have analytic dependence on θa, θb. Finally, we integrate over the last three variables. The
integral over θc turns the numerator into cos θab/2. In the integral over θa, the opposite iϵ
prescriptions for θa, θb imply that we pick up the residue of the fifth order pole at θa = θb.
This is proportional to n2(n2 − 1). Combining with the factor γ2
n defined in (3.109) we
conclude that ⟨Ψ2,n, δrail �K · Ψ2,n⟩ is indeed proportional to |n|.
To determine the coefficient of proportionality, one can compute the ratio of the rung
and rail corrections by analyzing the integrals at large n. More precisely, we take n large
and β large, with Ω = 2πn/β held fixed. In this limit it is better to use a proper time
coordinate on the circle, τ, rather than the angle θ = 2πτ/β. The h = 2 eigenfunctions
(3.109) are proportional to

Ψ(τ1, τ2) ∝ eiΩ(τ1+τ2)/2

τ12
f(Ωτ12/2),
f(ρ) = cos ρ − sin ρ

ρ .
(E.222)

62


For large n, all integrals will be dominated by the UV, where the propagator and correction
are
Gc = b sgnτ

|τ|2∆,
δG
Gc
∝ 1

|τ|.
(E.223)

The frequency Ω scales out, so we can choose the value Ω = 2. Then the rung and rail
contributions to the eigenvalue are proportional to the integrals

Irail =
�
dτ2dτ3dτ4eiτ2−iτ3−iτ4f(τ2) sgn(τ2)

|τ2|2−2∆
sgn(τ34)
|τ34|2−2∆
sgn(τ3)
|τ3|1+2∆
sgn(τ24)
|τ24|2∆ f(τ34)
(E.224)

Irung = q − 2

2

�
dτ2dτ3dτ4eiτ2−iτ3−iτ4f(τ2) sgn(τ2)

|τ2|3−2∆
sgn(τ34)
|τ34|2−2∆
sgn(τ3)
|τ3|2∆
sgn(τ24)
|τ24|2∆ f(τ34)

(E.225)

where the proportionality constant is the same in both cases. Since we know the normalized
rung contribution, we can get the full answer by computing the ratio of the above integrals
and using (E.215):

δk(2, n) =
�
1 + Irail

Irung

�
δrungk(2, n),
(E.226)

The rung integral is easy to evaluate using the fact that we started with eigenvectors of
the original kernel. The rail integral takes more work (it is convenient to represent some
of the factors in the integrand as fourier transforms) but the integrals can be done, and
one eventually finds agreement with (3.125).

F
The first order change in h = 2 eigenvectors

In this appendix we show that the the first order shift in the h = 2 eigenvectors Ψexact
2,n
=
Ψ2,n + δΨ2,n + ... is independent of q up to an overall multiple:

δΨ2,n = qαG

2 δΨq=∞
2,n .
(F.227)

Morally, the reason is the following. The h = 2 eigenvectors are given by reparameteriza-
tions of Gc, and the first order corrections are related to reparameterizations of δG, which
itself is univesal in q up to a coefficient. However, we will not need this interpretation. To
give the actual argument, we start by considering the reparameterization δϵI, where

I(τ1, τ2) =
�
dτadτbG(τ1, τa)Σ(τa, τb)G(τ2, τb).
(F.228)

Here, reparameterizations are defined to act as in (3.107), and we consider the function
I to have weight ∆ = 1/q. With this definition, I is reparameterization covariant, in
the sense that the reparameterization of the answer for the integral is the same as the

63


reparameterization of the various parts that go inside the integral. Writing this statement
out for linearized reparameterizations and using the exact Schwinger-Dyson equations
�
dtaG(t1, ta)Σ(ta, t2) = −δ(t12) + ∂t2G(t1, t2),
(F.229)

we find

(1 − K) · δϵG = 1

qHϵ,
Hϵ(τ1, τ2) ≡
�
dτ ϵ′(τ)G(τ1, τ)∂τG(τ2, τ) − (1 ↔ 2).
(F.230)

This is true for any value of the coupling, provided that K and G are the exact kernel
and propagator. Using �K = |G|
q−2

2 K|G|− q−2

2 , and taking a matrix element with one of the
conformal eigenvectors Ψh,n, we get (the inner product is as in (3.102))

⟨Ψh,n, (1 − �K) · |G|
q−2

2 δϵG⟩ = 1

q⟨Ψh,n, |G|
q−2

2 Hϵ⟩.
(F.231)

Naively, the leading piece of the LHS of (F.231) is at order (βJ)−1, where we use the
conformal answers for everything.
However, this gives zero because |Gc|
q−2

2 δϵGc is an
eigenvector of �Kc with eigenvalue one. In fact, the leading IR terms are at order (βJ)−2.
We get these by substituting in either δG or δK into the left side. The RHS has no terms
at this order, so these contributions must cancel:

⟨Ψh,n, (1 − �Kc) · δϵ(|Gc|
q−2

2 δG)⟩ − ⟨Ψh,n, δ �K · |Gc|
q−2

2 δϵGc⟩ = 0.
(F.232)

The integral defining the LHS of (F.232) has a UV divergence; we define the integral by
taking only the cutoff-independent (βJ)−2 piece and discarding the power divergence. In
the exact theory, UV divergences in this expression and on both sides of (F.231) will be
regulated to terms at order (βJ)−h and (βJ)−h−1. Depending on h these might dominate
over the IR term we are interested in, but as long as h ̸= 2 they can be separated.
Let us examine (F.232) in more detail. We can act with the kernel to the left, giving
(1−kc(h). Now, |Gc|
q−2

2 δϵGc is proportional to (1/q) times an h = 2 conformal eigenvector,
and the quantity being reparameterized in the LHS is independent of q, up to a multiple
αG. So we conclude that
1

qαG

⟨Ψh,n, δ �K · Ψ2,n⟩

1 − kc(h)
(F.233)

is independent of q. Apart from the prefactor, this expression is the first order perturbation
theory formula for the matrix element of ⟨Ψh,n, δΨ2,n⟩, so we conclude (F.227). Although
we did not need explicit formulas for the corrected eigenvectors in this paper, one can get
them by expanding and normalizing (3.117) and (3.118).

64


1
2
3
4
5
6
θ

0

0.1

0.2

0.3

0.4

0.5

0.6

G

β J = 10

1
2
3
4
5
6
θ

0

0.1

0.2

0.3

0.4

0.5

0.6

G

β J = 50

Figure 15: The exact G(θ) in the q = 4 model is shown in solid lines, for βJ = 10 (left)
and βJ = 50 (right). We also plot the conformal answer Gc in dash-dotted lines, and the
conformal answer plus the first correction Gcf0 in dashed lines.

G
Numerical solution of the SD equations

In this appendix, we discuss the numerical solution of the Schwinger-Dyson equations at
finite βJ. The euclidean solutions give us the coefficient αG (and thus also αK, αS). One
can also use these solutions to directly compute the large N free energy. The real-time
solutions were used to compute the blue circles in (11).
We will begin by discussing the euclidean equations, at finite temperature:

G(ωn)−1 = −iωn − Σ(ωn),
Σ(τ) = J2G(τ)q−1.
(G.234)

Here ωn = 2π(n+1/2)/β is a Matsubara frequency. One can solve these equations just by
iterating them, starting with the free correlator and using a numerical fourier transform to
switch betwen frequency ωn and time θ = 2πτ/β. In order to get the iteration to converge,
one should take a weighted update12

Gj(ωn) = (1 − x)Gj−1(ωn) + x
1

−iωn − Σj−1(ωn).
(G.235)

where the weighting x is a parameter. One can set it by beginning with x = 0.5 and
then monitoring the difference
�
|Gj − Gj−1|2 between successive steps. If this begins to
increase, one divides x by a half and continues the iteration. Some exact solutions are
shown for different values of βJ in figure 15.
For large values of βJ, the difference between the exact and conformal correlators is
fit very well by
G ≈ Gc − αG

βJ Gcf0
(G.236)

12We are grateful to A. Kitaev for suggesting this.

65


where f0 was defined in Eq. (3.121) and αG is a fitting parameter. More precisely, this
holds as long as τJ is large. We determine αG from the numerical solutions by fitting
for the coefficient in the region π/2 ≤ θ ≤ π. In the numerics we have a finite frequency
cutoff and finite J, but we take both large and look for convergence. For small q < 3 to
get accurate results we have to extrapolate in both variables, first in the cutoff and then
in J.
The function αG(q) was plotted in figure 9. Some explicit values are αG(2) = 0, αG(4) ≈
0.1872, αG(6) ≈ 0.1737, αG(8) ≈ 0.1522, and αG(10) ≈ 0.1336. A Pade approximant that
stays within approximately one percent of the numerical answer is

αG(q) ≈
2(q − 2)

16/π + 6.18(q − 2) + (q − 2)2.
(G.237)

With the solution to the Schwinger-Dyson equations, we can also compute the free en-
ergy using (2.26). In terms of the correlators and the self energy at Matsubara frequencies,
we have

log Z

N
= 1

2 log 2 + 1

2

∞
�

n=−∞
log
�
1 + Σ(ωn)

iωn

�
− β

2

� β

0

�
Σ(τ)G(τ) − J2

q G(τ)q
�
.
(G.238)

To get this expression from (2.26), we have used the free answer log Z = N

2 log 2 in the
case J = 0 to set the constant. The effect was to replace
�

n
log(−iωn) → log 2.
(G.239)

The answer we expect for the free energy is an expansion in powers of 1/(βJ):

log Z

N
= a1βJ + a2 + a3

βJ + ...
(G.240)

where −a1J is the ground state energy density, a2 is the zero temperature entropy density,
and 2a3 is the specific heat density. We can remove the ground state energy by considering

log Z − J∂J log Z

N
= a2 + 2 a3

βJ + ...
(G.241)

The derivative term can be evaluated using (2.27). Evaluating the sum of these terms on
the numerical solution to the Schwinger Dyson equations for moderately large βJ, we find
very good agreement with the S0(q) given in (2.33). The agreement is good enough that
we can subtract S0 and study the remainder for different values of βJ in order to compute
a3. This was used to compute the circles in figure 12.
We can also continue the equations (G.234) to get the retarded and Wightman correla-
tors in real time, following [13]. For this it is important to use the spectral function ρ(ω).
Here, ω with no subscript is a real-time frequency, which takes continuous values. We are

66


using conventions where the spectral function can be defined as the real part of the fourier
transform of the retarded propagator:

ρ(ω) ≡ 2Re GR(ω) = G>(ω)(1 + e−βω),
G>(t) ≡ ⟨ψ(t)ψ(0)⟩ = G(it + ϵ).
(G.242)

The Matsubara propagator G(ωn) can be written in terms of ρ as

G(ωn) =
� dω′

2π
ρ(ω′)

−iωn + ω′.
(G.243)

In this form, one can easily continue to complex frequency. The continuation to real-time
frequency is essentially the retarded propagator: GR(ω) = −iG(−iω + ϵ). To get the
real-time Schwinger-Dyson equation we also have to understand how to continue Σ(ωn).
Writing the second equation (G.234) in frequency space and using (G.242) we have

Σ(ωn) = J2
� β

0
eiωnτG(τ)q−1,
G(τ) =
� dω

2π e−ωτ
ρ(ω)

1 + e−βω .
(G.244)

After doing the τ integral we get an equation that can be continued to complex frequency,

Σ(ωn) = J2
� �q−1
�

j=1

dωj
2π
ρ(ωj)

1 + e−βωj

�
1 + e−β �
j ωj

−iωn + �

j ωj
.
(G.245)

Now we have a closed set of equations for ρ that can be iterated. First, we compute the
retarded propagator from the continuation of the first equation in (G.234):

GR(ω)−1 = [−iG(−iω + ϵ)]−1 = −iω + ϵ − iΣ(−iω + ϵ)
(G.246)

Next, we compute the spectral function by taking twice the real part. Finally, we substitute
ρ into (G.245) to get the new self energy. An appropriately weighted iteration of this
procedure converges. In implementing these equations numerically, we have to put both
an IR cutoff and a UV cutoff on the frequencies. This makes the problem more challenging
than the Euclidean problem, but we still get good agreement with the conformal answer
and the leading correction. See figure 16 for a plot.
The only place we used these real-time solutions in the main text was to compute the
circles in figure 11. To evaluate these we solve the above equations to get GR and Glr,
which can also be written in terms of ρ. We then assume an ansatz (3.155). This turns
(3.154) into a one-dimensional integral equation for f(t12). This can be discretized and
represented as a matrix equation. λL is determined by the condition that this matrix
should have an eigenvalue equal to one. We find this by doing binary search.

H
A model without the reparametrization symmetry

It is natural to ask whether there is a model where instead of 1/(1 − K) in the expression
for the four point function (3.45) we get 1/(1 − gK), with a g < 1. This would move

67


0
2
4
6
8
10
12
2π t/ β

0

0.05

0.1

0.15

0.2

0.25

0.3

Glr

2
4
6
8
10
2π t/ β

0

0.2

0.4

0.6

0.8

1

GR

Figure 16: The retarded propagator GR (left) and the half-circle Wightman correlator Glr
(right) are plotted in the q = 4 model with βJ = 10. The solid curve is the numerical
answer, the dash-dotted curve is the conformal answer, and the dashed curve is the con-
formal answer plus the leading correction (3.161). The behavior of the numerical GR near
t = 0 is somewhat contaminated by finite-cutoff wiggles.

the pole away from h = 2 and would lead to finite expression in the conformal limit. It
is clear from our discussion in section 4 that this can only be true in a model without
reparametrization symmetry.
A simple model with these properties arises if we assume that the couplings ji1···iq are
time dependent fields with a two point function

⟨ji1···in(t)ji1···in(0)⟩ = J2(q − 1)!

N q−1
×
1

|t|2α
(H.247)

The new factor is the last one. In the limit α → 0 we recover the original model (to leading
orders in the 1/N expansion).
With this modification we can still write the Schwinger Dyson equations as
1

G(ω) = −iω − Σ(ω) ,
Σ(τ) = J2[G(τ)]q−1
1

|τ|2α
(H.248)

In the low energy limit, we can now make a scale invariant ansatz as before

Gc = b sgn(τ)

|τ|2 ˆ∆
(H.249)

With this ansatz we can solve the low energy limit of (H.248) (dropping the iω term) and
we find that
ˆ∆Σ = ˆ∆(q − 1) + α = 1 − ˆ∆ ,
ˆ∆ = 1 − α

q
(H.250)

where we have denoted the dimension of Gc by ˆ∆, since it is not equal to 1/q. The overall
coefficient has exactly the same expression as before (2.10) in terms of ˆ∆

J2bqπ = (1

2 − ˆ∆) tan π ˆ∆
(H.251)

68


We can now consider the kernel that appears in the four point function computation.
It has an expresssion similar to the one before (3.44),

ˆKc(τ1, τ2; τ3, τ4)
=
−(q − 1)Gc(τ13)Gc(τ24)Σc(τ34)

Gc(τ34)

=
−(q − 1)bqJ2sgn(τ13)

|τ13|2 ˆ∆
sgn(τ24)
|τ24|2 ˆ∆
1

|τ34|2−4 ˆ∆

=
− (q − 1)

( 1

ˆ∆ − 1)Kc, ˆ∆
(H.252)

where Kc, ˆ∆ is the usual kernel but with ∆ → ˆ∆. Namely, in (3.49) (3.50) we replace
∆ → ˆ∆ and q → 1/ ˆ∆.
The eigenvalues of the new kernel are then equal to

ˆkc(h) = gkc, ˆ∆(h) ,
g ≡ (q − 1)

( 1

ˆ∆ − 1)
(H.253)

where k ˆ∆(h) is the usual expression in terms of ∆ (i.e. we replace 1/q → ˆ∆ everywhere in
(3.73).) Now if 0 < α < 1, then we see that ˆ∆ < 1/q which means that that g < 1. This
implies that now the sum that appears in the computation of the four point function is
regular and of the form
1

1 − ˆK
=
1

1 − gK ˆ∆
(H.254)

Therefore now we do not have to worry about the h = 2 contribution. For h = 2
we find that ˆK = g < 1 and the sum is finite. In this case, the expression analogous to
(3.83) is finite. Since k′
c(h = 2) < 0, the first pole is at a value hp < 2. Something similar
happens with the retarded kernel, ˆKR where the pole moves to a value −1 < hchaos < 0.
More explicitly, using the formula (3.98) we find

ˆkR(1 − h) = cos π( ˆ∆ − h

2)

cos π( ˆ∆ + h

2)
kc, ˆ∆(h)
(H.255)

We can easily check from here that ˆkR(h = 0) = q − 1 > 1 and that ˆkR(h = −1) = g < 1.
Therefore there is always a solution for ˆkR(hchaos) = 1 for −1 < hchaos < 0, leading to the at
behavior e(−hchaos) 2π

β t. This means that we have a growing contribution but growing more
slowly than the bound. Here we are assuming that when we go to the finite temperature
theory we also change the two point function (H.247) to its finite temperature version.
As α → 0, it seems clear that we will get a divergence that will go like 1/α. The
coefficient of this divergence would be a function of cross ratios. This is different than the
function that multiplies
1

βJ that we discussed in section (3.3.3).
As we take the α → 0 the sum over the normalizable h = 2 modes, in Fourier space,
involves a factor of the form 1/(1 − ˆK) ∝ 1/(α +
n

(βJ)), where we also included the terms

69


that would break the conformal symmetry when α = 0. Then depending on whether α or
1/(βJ) is larger, we go from one regime to the other.
We can then derive an effective action for reparametrizations which would reproduce
the above kernel. We find that it should have the schematic form
�

n

� 1

Jβ n2(n2 − 1) + α(n2 − 1)|n|
�
|ϵn|2
(H.256)

the last term in the action is non-local. It should come from the variation of the modified
term in the effective action
�
dθ1dθ2
J2

|2 sin θ12

2 |2αGc(θ12)q
(H.257)

when we make a reparameterization of Gc and then expand to quadratic order in ϵ. When
α is zero, the term is reparameterization invariant, but one can check that if we expand
to linear order in α it does give the second term in (H.256).
We could view the two point function of the j’s as arising from a higher dimensional
conformal field theory.
If that field theory has a holographic dual, then we would be
describing something that lives on an AdS2 subspace of a higher dimensional bulk. Such
a theory would not have a purely dynamical two dimensional gravity. This setup arises
naturally in the Kondo model and its holographic duals.
See [41] for a Kondo model
example that inspired the SYK model studied in this paper, and [42] and references therein
for holographic examples.

I
Further coments on Kinematic space

In this appendix we expand a bit more on the comments in section 6.1, where we explored
properties of the two dimensional space characterized by two times t1, t2 of a bilocal field.
We can consider the finite temperature Lorentzian theory. After defining the following
coordinates the Casimir becomes (setting β = 2π)

t = t1 + t2

2
,
σ → t1 − t2

2
,
→
C → sinh2 σ(−∂2
t + ∂2
σ)
(I.258)

where we now have the wave equation on the outside of the Lorentzian black hole. We see
that the two point function Gc is determining the metric of the space we should consider.
We can easily get to the interior by taking t1 → t1 + iβ/4, t2 → t2 − iβ/4 so that now
we get
C ∼ cosh2 σ(∂2
t − ∂2
σ)
(I.259)

which is the wave operator in the interior region. The fact that we have a complex shift
in the two times by t1 − t2 → t1 − t2 + iβ/2 is related to the fact that we can easily create
particles in the interior if we have access to both sides of the thermofield double, or if we
perform small perturbations of the thermofield double state [43].

70


Finally, there is an elegant relation between bulk and boundary using embedding co-
ordinates.
Points on the boundary can be written in terms of projective coordinates
XM = (X−1, X0, X1), with (X.X) ≡ −X2
−1 − X2
0 + X2
1 = 0 and X ∼ λX. If we have a pair
of such points on the boundary, Xa
M and Xb
M, then we can define

Y L = ϵMNLXa
MXb
N

(Xa.Xb)
(I.260)

where (Xa.Xb) is simply the inner product using the SL(2) metric. This obeys (Y.Y ) = −1.
Of course, conceptually, this is the same as what we have discussed above, except that
now we see that the formulas are SL(2) covariant.

I.0.1
The kinematic space in the full model at q = ∞

As a final comment, we will consider the case of of q → ∞. This case looks simpler because
the only state in the singlet spectrum is the h = 2 state. As we approach the q → ∞
limit the other states are decoupling, but their enegies are not becoming large. In this
sense it is different than the very large ‘t Hooft coupling limit of a gauge theory, where
the decoupling of the string states happens because they become heavy. An observation is
that in this case we get an interesting picture in terms of the “bulk” coordinates defined
in (6.185). In this limit the kernel is given by

K(t1, t2; t3, t4) = sgn(t13)sgn(t24)
1

(|t34| + ϵ)2 − (3 ↔ 4) ,
ϵ = 1

J
(I.261)

We then find that it obeys the equation

− (|t12| + ϵ)2∂t1∂t2K(t1, t2; t3, t4)
=
[δ(t1 − t3)δ(t2 − t4) − (3 ↔ 4)]
=
(σ + ϵ)2(−∂2
t + ∂2
z)K(t, z; t′, z′)
(I.262)

where we defined t, σ as in (6.185). After defining z = σ + ϵ we find that we get the
wave operator in AdS2 (parametrized by t, z) with a cutoff at z = ϵ. In particular, here
z ≥ ϵ always and, for the spectral problem we are putting boundary conditions that set
the normalizable functions to be zero at ˜z = ϵ. So in this case the kernel is really K =
2
∇ϵ,
see also (3.76), where ∇ϵ is the laplacian in AdS2 with dirichlet boundary conditions at
z = ϵ. Now the quadratic term in (4.172) becomes 1

K −1 = 1

2∇2
ϵ −1 and has a simple local
form. This is interesting because popular way to regularize AdS computations consists
in setting a cutoff a ˜z = ϵ. However, it was unclear which kind of regularization of the
boundary theory would give such a cutoff. Here we see an example, where we get the
same kind of regularized AdS2 problem. In this theory we flow rather quickly from the
topological theory in the UV to the IR AdS2-like theory. This needs to be taken with a
grain of salt given that we do not know whether there is a way to think about the model
as a local theory in AdS. Also the above construction seems more related to regulating
the kinematic space than the actual bulk.
In the finite temperature case we also get a dS2, or AdS2 with periodic time. In this
case we also need to rescale the size of the circles relative to their naive values, see (3.114).

71


References

[1] A. Kitaev, “A simple model of quantum holography.”
http://online.kitp.ucsb.edu/online/entangled15/kitaev/,http:
//online.kitp.ucsb.edu/online/entangled15/kitaev2/. Talks at KITP, April
7, 2015 and May 27, 2015.

[2] J. L. Karczmarek, J. M. Maldacena, and A. Strominger, “Black hole non-formation
in the matrix model,” JHEP 01 (2006) 039, arXiv:hep-th/0411174 [hep-th].

[3] I. R. Klebanov, “String theory in two-dimensions,” in Spring School on String
Theory and Quantum Gravity (to be followed by Workshop) Trieste, Italy, April
15-23, 1991. 1991. arXiv:hep-th/9108019 [hep-th].

[4] S. Sachdev and J.-w. Ye, “Gapless spin fluid ground state in a random, quantum
Heisenberg magnet,” Phys. Rev. Lett. 70 (1993) 3339, arXiv:cond-mat/9212030
[cond-mat].

[5] S. Sachdev, “Holographic metals and the fractionalized Fermi liquid,” Phys. Rev.
Lett. 105 (2010) 151602, arXiv:1006.3794 [hep-th].

[6] V. de Alfaro, S. Fubini, and G. Furlan, “Conformal Invariance in Quantum
Mechanics,” Nuovo Cim. A34 (1976) 569.

[7] A. Almheiri and J. Polchinski, “Models of AdS2 backreaction and holography,”
JHEP 11 (2015) 014, arXiv:1402.6334 [hep-th].

[8] A. Kitaev. http://online.kitp.ucsb.edu/online/joint98/kitaev/. KITP
seminar, Feb. 12, 2015.

[9] S. H. Shenker and D. Stanford, “Black holes and the butterfly effect,” JHEP 03
(2014) 067, arXiv:1306.0622 [hep-th].

[10] A. Kitaev. Talk given at the Fundamental Physics Prize Symposium, Nov. 10, 2014.

[11] S. H. Shenker and D. Stanford, “Stringy effects in scrambling,” JHEP 05 (2015)
132, arXiv:1412.6087 [hep-th].

[12] J. Maldacena, S. H. Shenker, and D. Stanford, “A bound on chaos,”
arXiv:1503.01409 [hep-th].

[13] O. Parcollet and A. Georges, “Non-fermi-liquid regime of a doped mott insulator,”
Phys. Rev. B 59 (Feb, 1999) 5341–5360.
http://link.aps.org/doi/10.1103/PhysRevB.59.5341.

[14] S. Sachdev, “Bekenstein-Hawking Entropy and Strange Metals,” Phys. Rev. X5
no. 4, (2015) 041025, arXiv:1506.05111 [hep-th].

72


[15] P. Hosur, X.-L. Qi, D. A. Roberts, and B. Yoshida, “Chaos in quantum channels,”
JHEP 02 (2016) 004, arXiv:1511.04021 [hep-th].

[16] W. Fu and S. Sachdev, “Numerical study of fermion and boson models with
infinite-range random interactions,” arXiv:1603.05246 [cond-mat.str-el].

[17] Y.-Z. You, A. W. W. Ludwig, and C. Xu, “Sachdev-Ye-Kitaev Model and
Thermalization on the Boundary of Many-Body Localized Fermionic Symmetry
Protected Topological States,” arXiv:1602.06964 [cond-mat.str-el].

[18] A. Jevicki, K. Suzuki, and J. Yoon, “Bi-Local Holography in the SYK Model,”
arXiv:1603.06246 [hep-th].

[19] J. Polchinski and V. Rosenhaus, “The Spectrum in the Sachdev-Ye-Kitaev Model,”
arXiv:1601.06768 [hep-th].

[20] D. Anninos, T. Anous, and F. Denef, “Disordered Quivers and Cold Horizons,”
arXiv:1603.00453 [hep-th].

[21] D. Anninos, T. Anous, P. de Lange, and G. Konstantinidis, “Conformal quivers and
melting molecules,” JHEP 03 (2015) 066, arXiv:1310.7929 [hep-th].

[22] N. Goheer, M. Kleban, and L. Susskind, “(1+1)-dimensional compactifications of
string theory,” Phys. Rev. Lett. 92 (2004) 191601, arXiv:hep-th/0310120
[hep-th].

[23] D. A. Roberts and D. Stanford, “Two-dimensional conformal field theory and the
butterfly effect,” Phys. Rev. Lett. 115 no. 13, (2015) 131603, arXiv:1412.5123
[hep-th].

[24] S. Jackson, L. McGough, and H. Verlinde, “Conformal Bootstrap, Universality and
Gravitational Scattering,” Nucl. Phys. B901 (2015) 382–429, arXiv:1412.5205
[hep-th].

[25] G. Turiaci and H. Verlinde, “On CFT and Quantum Chaos,” arXiv:1603.03020
[hep-th].

[26] J. Maldacena, D. Stanford, and Z. Yang, “To appear,”.

[27] J. Polchinski and A. Streicher. Private communication.

[28] A. Almheiri, D. Marolf, J. Polchinski, and J. Sully, “Black Holes: Complementarity
or Firewalls?,” JHEP 02 (2013) 062, arXiv:1207.3123 [hep-th].

[29] B. Czech, L. Lamprou, S. McCandlish, and J. Sully, “Integral Geometry and
Holography,” JHEP 10 (2015) 175, arXiv:1505.05515 [hep-th].

73


[30] B. Czech, L. Lamprou, S. McCandlish, B. Mosk, and J. Sully, “A Stereoscopic Look
into the Bulk,” arXiv:1604.03110 [hep-th].

[31] A. Georges, O. Parcollet, and S. Sachdev, “Quantum fluctuations of a nearly critical
heisenberg spin glass,” Phys. Rev. B 63 (Mar, 2001) 134406.
http://link.aps.org/doi/10.1103/PhysRevB.63.134406.

[32] N. Iizuka and J. Polchinski, “A Matrix Model for Black Hole Thermalization,”
JHEP 10 (2008) 028, arXiv:0801.3657 [hep-th].

[33] B. Michel, J. Polchinski, V. Rosenhaus, and S. J. Suh, “Four-point function in the
IOP matrix model,” arXiv:1602.06422 [hep-th].

[34] T. M. Fiola, J. Preskill, A. Strominger, and S. P. Trivedi, “Black hole
thermodynamics and information loss in two-dimensions,” Phys. Rev. D50 (1994)
3987–4014, arXiv:hep-th/9403137 [hep-th].

[35] J. M. Maldacena, J. Michelson, and A. Strominger, “Anti-de Sitter fragmentation,”
JHEP 02 (1999) 011, arXiv:hep-th/9812073 [hep-th].

[36] S. Giombi, S. Minwalla, S. Prakash, S. P. Trivedi, S. R. Wadia, and X. Yin,
“Chern-Simons Theory with Vector Fermion Matter,” Eur. Phys. J. C72 (2012)
2112, arXiv:1110.4386 [hep-th].

[37] J. M. Maldacena, “Long strings in two dimensional string theory and non-singlets in
the matrix model,” JHEP 09 (2005) 078, arXiv:hep-th/0503112 [hep-th]. [Int.
J. Geom. Meth. Mod. Phys.3,1(2006)].

[38] G. ’t Hooft, “A Two-Dimensional Model for Mesons,” Nucl. Phys. B75 (1974)
461–470.

[39] S. Leichenauer, “Disrupting Entanglement of Black Holes,” Phys. Rev. D90 no. 4,
(2014) 046009, arXiv:1405.7365 [hep-th].

[40] A. P. Reynolds and S. F. Ross, “Butterflies with rotation and charge,”
arXiv:1604.04099 [hep-th].

[41] O. Parcollet, A. Georges, G. Kotliar, and A. Sengupta, “Overscreened multichannel
SU(N) Kondo model: Large-N solution and conformal field theory,” Phys. Rev. B
58 (1998) 3794, arXiv:cond-mat/9711192 [cond-mat].

[42] J. Erdmenger, M. Flory, C. Hoyos, M.-N. Newrzella, A. O’Bannon, and J. Wu,
“Holographic impurities and Kondo effect,” in The String Theory Universe, 21st
European String Workshop and 3rd COST MP1210 Meeting Leuven, Belgium,
September 7-11, 2015. 2015. arXiv:1511.09362 [hep-th].
http://inspirehep.net/record/1407205/files/arXiv:1511.09362.pdf.

74


[43] J. M. Maldacena, “Eternal black holes in anti-de Sitter,” JHEP 04 (2003) 021,
arXiv:hep-th/0106112 [hep-th].

75