Eigenstate Thermalization Hypothesis in Conformal Field Theory

Nima Lashkari,1 Anatoly Dymarsky,2 and Hong Liu1

1
Center for Theoretical Physics, Massachusetts Institute of Technology, Cambridge, MA 02139
2
Department of Physics and Astronomy, University of Kentucky, Lexington, KY 40506, USA
Skolkovo Institute of Science and Technology, Skolkovo Innovation Center, Moscow 143026 Russia
We investigate the eigenstate thermalization hypothesis (ETH) in d+1 dimensional conformal field
theories by studying reduced density matrices in energy eigenstates. We show that if local probes of
high energy primary eigenstates satisfy ETH, then any finite energy observable with support on a
subsystem of finite size satisfies ETH. In two dimensions, we discover that if ETH holds locally, the
finite size reduced density matrix of states created by heavy primary operators is well-approximated
by a projection to the Virasoro identity block.

INTRODUCTION proving local ETH should amount to understanding how

arXiv:1610.00302v1 [hep-th] 2 Oct 2016

the corresponding statement regarding OPE coefficients

Whether and how an isolated interacting quantum sys-
tem initially out of equilibrium equilibrates is a complex
dynamical question about which we currently have little
analytic control. For a non-integrable system, there is
nevertheless a powerful conjecture–the Eigenstate Ther-
malization Hypothesis (ETH)–which states that expecta-
tion values of generic few-body observables in a finitely
excited energy eigenstate should coincide with those in
the micro-canonical ensemble up to corrections that are
exponentially small in entropy [1, 2]. Assuming ETH,
one can then readily deduce that in the thermodynamic
limit: (i) expectation values of few-body observables in
a generic state coincide with those in a thermal ensem-
ble; (ii) expectation values of few-body observables in a
non-equilibrium initial state will evolve towards those of
a thermal ensemble.
By now, there are many numerical evidences for ETH
in a variety of quantum systems [3]. While ETH was
often stated for few-body observables, recently numeri-
cal supports were also found for the full reduced density
matrix of any subsystem [4, 5].
ETH is powerful as instead of following time evolu-
tion of a general non-equilibrium state, we now face an
in-principle simpler problem of understanding properties
of finitely excited energy eigenstates. Unfortunately, de-
spite much numerical progress energy eigenstates of a
general interacting many-body system are still too com-
plex to be tractable by analytic methods. A proof of
ETH for general non-integrable systems appears out of
reach at this stage.
In this paper we show that progress can be made in
conformal field theories (CFT). We are able to prove that
for CFTs, a weakest form ETH which applies to local pri-
mary operators (to which we refer as local ETH) in fact
leads to a strongest form of ETH which applies to the
reduced density matrix of a general subsystem. Thus to
prove ETH for CFTs it is enough to prove that it applies
to local primary operators which in turn reduces to a
statement regarding coefficients in operator product ex-
pansions (OPE). Clearly not all CFTs satisfy ETH, thus
connects with non-integrability.
We believe that local ETH provides a powerful techni-
cal handle for obtaining many dynamical properties of a
CFT. As a simple application we use it to compute the
Renyi entropies for an interval in an energy eigenstate
in a (1+1)-dimensional CFT. Interestingly we find that
even though the reduced density matrix for the region
approaches that of the canonical ensemble in the ther-
modynamic limit, the Renyi entropies differ by a finite
amount.
Our discussion should be generalizable to general quan-
tum field theories which we will pursue elsewhere.
SETUP
Consider finitely excited energy eigenstate |Ei of a con-
formal field theory in a d-dimensional space of finite vol-
ume. For simplicity, we take it to be a sphere of radius L
(i.e. the total spacetime is Rt × S d ). A key simplification
of conformal symmetry is that in Euclidean signature any
state |ψi on S d can be mapped through conformal trans-
formation
ds2cylinder = dτ 2 + L2 dΩ2d = Λ2 (dr2 + r2 dΩ2d )
L
Λ= , τ = L log r (1)
r
to a local operator Ψ inserted at origin (r = 0) in Rd+1 .
Similarly, the conjugate state hψ| is mapped to a local
operator Ψ† inserted at r = ∞ in Rd+1 . Energy eigen-
states are in one-to-one correspondence to the operators
Pµ1 · · · Pµm Ψa , where Pµ is a momentum operator and
Ψa is a primary operator. Below, we will use state and
operator languages interchangeably.
Kinematics of conformal symmetry fixes the corre-
lators in descendant states Pµ1 · · · Pµm Ψa , in terms of
those of their corresponding primary Ψa . The dynam-
ical content of a CFT is the spectrum of its primaries
and OPE coefficients. As a dynamical statement about
energy eigenstates which are not related by symme-

tries ETH should be restricted to primary energy eigen-
states. Moreover, we will restrict our attention to ho-
mogeneous energy eigenstates on S d which are invariant
under SO(d + 1) rotations. Such an energy eigenstate
with energy Ea corresponds to a spinless primary Ψa of
dimension ha = Ea /L. The energy density of the sys-
ha
tem is a = LEd ωa d = Ld+1 d
, where ωd is the volume of a
unit S d . For a CFT in a thermal state of temperature T ,
a ∼ CT d+1 where C is central charge, which motivates
us to define a “thermal” length scale associated with |Ea i
  − d+1
1
λT =
a
∼ T −1 .
C e−O(S(E)) . See supplementary material for a proof
We will be interested in the thermodynamic limit with
L → ∞, while keeping a finite energy density a and thus
a finite λT . In this limit the scaling dimension ha should
then scale with L as
 d+1
L ment hE|Op |Ei is controlled by that in the primary state
ha = Cωd . (3)
λT
The local ETH condition is defined as
p ha  l, from smoothness of function Op (E), Op (E − 2l
hEa |Op |Eb i = Op (E)δab + Oab ,
where Op is a local primary operator (with p labeling
different operators), the diagonal element Op (E) is a
smooth function of E = Ea +E b p
, and Oab ∼ e−O(S(E))
2
S(E)
with e the density of state at energy E. Due to ho-
mogeneity, it does not matter where Op is inserted, so we
left the location of Op implicit. In equation (4) OpR should
be understood as a smeared operator Op (f ) = S d f Op
with smearing function f such that the set of observables
{Op (f )} form an algebra. In particular, the operator Op2
should be well-defined and has finite expectation values
in physical states1 . Below for notational simplicity we
will continue to suppress f in Op (f ).
We now introduce subsystem ETH which states that
for a subregion B [5]
1. There exists a “universal” density matrix ρB (E)
(which depends only on B and energy E) such that
for any energy eigenstate |Ea i
||ρaB − ρB (E = Ea )|| ∼ e−O(S(E))
a
p ρB ≡ TrB1 P
where c |Ea ihEa |. In (5), kρ − σk =
1 2 =
2 Tr (ρ − σ) 2 i |λ i | is the trace distance of
two density matrices ρ and σ, and λi are the eigen-
values of the Hermitian operator ρ − σ.
1 Note that for an expectation value of O to be physically mean-
ingful, its fluctuations should be sufficiently small. This implies
that the corresponding expectation value of O2 must be finite.
Thus smeared operators must be used in formulating ETH in
quantum field theory.
2
2. Introducing σab ≡ TrB c |Ea ihEb |, then
keiα σab + e−iα σba k ∼ e−O(S(E)) . (6)
for all α.
Note that subsystem ETH (5)–(6) implies local ETH,
that is, for an operator O supported inside region B we
have
1 1
Tr((ρaB −ρB (E)O) ≤ kρaB −ρB (E)k 2 Tr((ρaB +ρB (E))O2 ) 2
(7)
(2) and thus is exponentially small. Similarly, Tr(Oσab ) =
of (7).2
To close our setup, let us briefly comment on the de-
scendant eigenstates. As an example consider |Ei corre-
sponding to a spinless operator (Pµ2 )l Ψa where l an inte-
ger. For such a state E = Ea,l = Ea + 2l L . The matrix ele-
Ψa , i.e. by Op (E − 2l
L ) rather than Op (E). Thus as stated
earlier, we should not include the descendant states in
either (4) or (5)–(6). Of course if one assumes (4), for
L)
(4)
is related to Op (E) only by corrections of order l/ha
and similarly the corresponding reduced density matrix
is close to ρB (E) in trace distance with corrections that
are polynomially suppressed in l/ha . But such state-
ments do not reflect any new dynamics beyond (4)–(6).
For completeness, in supplementary material we discuss
more explicitly the story for descendant states.
FROM LOCAL ETH TO SUBSYSTEM ETH
Let us now look at the implications of (4). Using con-
formal mapping (1) the matrix element hEb |Op |Ea i on
S d is mapped to the Euclidean three-point function
hEb |Op |Ea i (8)
L−hp hΨ†b (∞)Op (1)Ψa (0)i p −hp
=q = Cab L ,
† †
hΨa (∞)Ψa (0)ihΨb (∞)Ψb (0)i
p
(5) where Cab is the OPE coefficient for Op appearing in
the operator product Ψa and Ψ†b , when Ψa and Ψb are
primaries. We assume the thermodynamic limit exists
in (4) for any operator Op whose dimension hp does not
scale with L. From (8) and expressing L in term of ha
2 Expectation values depend on the normalization of the observ-
able and could be arbitrarily large for unbounded operators.
It is important to appreciate that density matrices with small
trace distance are physically indistinguishable, even though cor-
responding expectation values of certain operators might not be
close [6].
 3

†
(1) †
(1) More explicitly, denote the reduced density matrix for
✓0 B
region B coming from the energy eigenstate |Ea i in the
†

B
S d ⇥ Rt
B
(sin ✓0 )
Rindler frame with ρ̃aB . The expectation values in this
(0) 1
( sin ✓0 )
state can be written as
( 1) µ µ
hΨ†a (X+ )Ψa (X− ) · · ·i
(a) (b) (c) Tr(ρ̃aB · · · ) =
h؆a (X+
µ
)Ψa (X− µ
)i
d X  h p
FIG. 1: (a) The cylinder S × Rt conformal frame (b) Radial p R
quantization Rd+1 conformal frame. (c) The conformal frame = Caa hOp (X0 ) · · ·i (11)
L
convenient for the study of the density matrix on subsystem hp
B.
where X0µ = (0, 1, X i>1 = 0). Note that the reduced den-
sity matrix can be considered a map from the algebra
p
using (3) we conclude that the OPE coefficient Cab must of observables on the subsystem to expectation values:
scale with ha → ∞ as ρ(O ∈ {O}) = Tr(ρO). For subregion B, we should re-
hp
strict to operators Op whose dimensions hp remain finite
p
Cab = had+1 δab fp (E) + Rpab (9) in the thermodynamic limit (i.e. does not scale L). The
summation in (11) should be understood as so. Also, it
hp
h
where fp (E) = λTp (N ωd )− d+1 Op (E) is a smooth func- should be understood that in the sum in (11) the descen-
tion of E independent of label a, and Rpab = Lhp Oabp
∼ dants of Op are included implicitly (similarly below).
d
d+1 hp
Using (8) and (4) we can write (11) in an operator form
e−O(ha )+ d+1 log ha as on general ground we expect that
S(E) ∝ Ld in the thermodynamic limit. Note that equa- ρ̃aB = ρ̃B (E) + RaB (12)
tion (9) implies the following: for operators Op whose
hp where
p
Caa grow slower than had+1 with ha cannot have a non- X
vanishing expectation value in the thermodynamic limit, ρ̃B (E) = Op (E)Rhp Op (X0 ) (13)
p
while it is impossible for an operator with Cab to grow hp
hp X
faster than had+1 as that would imply the thermodynamic RaB = p
Oaa Rhp Op (X0 ) . (14)
limit does not exist. hp
We are interested in reduced density matrices of a re-
gion B of a finite size in the thermodynamic limit. Con- Equation (13) (and similarly for RaB ) should be under-
sider B to be a ball around the north pole of S d with an- stood as follows. ρ̃B (E) is prepared in the Rindler frame
gular radius θ0 = tan−1 ( 2LR
). In the the thermodynamic via a Euclidean path-integrals over Rd+1 with boundary
limit of the angular size R/L → 0 keeping R finite, B conditions above and below the negative half-space and
becomes a ball of radius R in Rd . We will discuss other a local operator inserted at X0µ (see figure 2). In (13)
shapes or regions with disconnected components toward we denote the density matrix by the specific operator
the end. inserted at X0µ . Similarly we can write σab introduced
The trace distance is invariant under unitary rotation before (6) as (with a 6= b)
of states: X p
σ̃ab = Oab Rhp Op (X0 ) . (15)
kρ − σk = kρ̃ − σ̃k, Ã = U AU † . hp

Therefore, we are free to compare density matrices in The equations above are to be understood as operator
any conformal frame. A convenient conformal frame is equalities inside Euclidean path-integral in the Rindler
the Rindler frame frame. Mapping back to the radial quantization frame
the universal density matrix ρB (E) is
ds2cylinder = Λ2 (X)dX i dX i , (10) X
ρB (E) = Op (E)Rhp Ôp (0)
where the ball-shaped region B on S d is mapped to neg- hp

ative half-space, X 1 ≤ 0; see figure 1 and supplemen- †

tary material for details. Under this map Ψa and Ψ†a
that create the state vector and its dual are mapped to
µ
X± = (± sin θ0 , cos θ0 , X i>1 = 0), respectively. Thus,
in the thermodynamic limit θ0 → 0, the two operators
are colliding in this frame, and we can use their operator
product expansions to represent them.
Ôp = U Op U, (16)
and U is the unitarity corresponding to the conformal
transformation from the Rindler frame to radial quanti-
zation frame; see figure 2(c).
From (12) and the finite radius of convergence of OPE
we immediately find that for observables O1 , · · · Ok in
 4

where ρA (E) = TrB−A ρB (E); see figure 2(c).

† Now consider the state |Ea,b,α i = √12 (|Ea i + eiα |Eb i)
B

1
B
1
B
X X
with |Ea i and |Eb i two energy eigenstates. The reduced
Op (E)Rhp Op Op (E)Ôp (0)
p p
density matrix in this state in the Rindler frame is
1 a
ρ̃a,b,α
(b) (c)
(ρ̃ + ρ̃bB + eiα σ̃ab + e−iα σ̃ba )
(a)
=
B
2 B
1
FIG. 2: (a) and (b) The Rindler frame path-integrals that = (ρ̃B (Ea ) + ρ̃B (Eb )) + Ra,b,α
B , (20)
prepare, respectively, ρ̃aB and ρ̃B (E) (c) The path-integral for 2
ρB (E) in the radial quantization frame.
where
1X p p p p
hp
region B, Ra,b,α = Oaa + Obb + eiα Oab + e−iα Oba R Op (X0 ),
2
hp
∆hO1 · · · Ok i Tr((ρ̃aB − ρ̃B )O1 · · · Ok )
≡ and we have used the local ETH assumption in (4). Re-
hO1 · · · Ok i Tr(ρ̃B O1 · · · Ok )
P p hp
peating the argument above one finds
h Oaa R hOp O1 · · · Ok i
=P p ∼ e−O(S(E)) . 1
hp
hp Op (E)R hOp O1 · · · Ok i keiα σab + e−iα σba k = kρ̃a,b,α
B − (ρ̃B (Ea ) + ρ̃B (Ea ))k
 2
In fact, we can prove that density matrices ρaB and ρB (E) a,b,α 1 2
≤ S ρ̃B k (ρ̃B (Ea ) + ρ̃B (Ea )) = O(ηab )
are close in trace distance. Trace distance is hard to 2
compute in continuum theories. Instead, we use another p p p

ηab = sup Oaa p
+ Obb + eiα Oab + e−iα Oba = e−O(S(E)) .
measure of distance called relative entropy which accord- p
ing to Pinsker’s inequality provides an upper bound on
trace distance of nearby states: By a similar Euclidean path-integral argument, the fol-
lowing two-norms are also small
kρaB − ρB (E)k2 = kρ̃aB − ρ̃B (E)k2 ≤ 2S(ρ̃aB kρ̃B (E)) . q
Tr ((ρaB − ρB (E))2 ) = O(ηa ),
Since the two density matrices are close, we only need r  
to compute relative entropy S(ρ̃aB kρ̃B ) perturbatively in † 2
Tr (σab σab ) = O(ηa ). (21)
small RaB . To the second order, we find that the relative
entropy is given by the quantum Fisher information
However, an explicit computation of the two-norm re-
Z∞ quires the propagator on a two-sheeted manifold that we
ds

S(ρ̃aB kρ̃B ) ' Tr (ρ̃B + s)−1 RaB (ρ̃B + s)−1 RaB do not have know.
2
0
= O(ηa2 ), FURTHER APPLICATIONS OF LOCAL ETH
p
ηa = sup Oaa = e−O(S(E)) , (17)
p
The expression in (12) is an operator equality. Ex-
where we have used the integral representation of the panding the Renyi entropies of the density matrix in the
logarithm of a positive operator. In supplementary ma- energy eigenstate it is easy to see that
terial, we expand relative entropy to all orders in ηa using
the replica trick in [7], and show (17) assuming that the Sn (ρaB ) − Sn (ρB (E)) = e−O(S(E)) .
n-point correlators on an n-sheeted manifold are finite.
That is to say that the Renyi entropies of reduced density
Invoking (17) it is evident that
matrices in energy eigenstates are universal.
kρaB − ρB (E)k2 ≤ 2S(ρ̃B kρ̃B (E)) = O(ηa2 ), (18) CFTs in 1 + 1-dimensions are special in that the ex-
pectation value of local operators vanishes in the thermal
which demonstrates the subsystem ETH. Note that sub- state. That is due to the fact that the thermal state on
system ETH holds for any finite ball-shaped subsystem a line is conformally flat. Then, the local ETH implies
p
of finite radius R. Monotonicity of trace distance un- that in the thermodynamic limit Caa → 0 for all quasi-
der partial trace then implies that subsystem ETH holds primaries, except for those made of stress tensor T and
for any subsystem of arbitrary shape and disconnected T̄ . In other words, the density matrices in eigenstates
regions that can be encircled in a ball of finite size: are well-approximated by their projection to the Vira-
soro identity block. Assuming local ETH one can com-
∀A such that ∃B : A ⊂ B, pute the universal density matrix ρ̃B (E) defined by (13).
kρaA − ρ(E)A k ≤ kρaB − ρ(E)B k = O(ηa ). (19) Since the quasi-primaries that appear in ρB (E) are all

only made of stress tensor one can directly compute the
Renyi entropies of subsystem B order by order in R/λT .
In [5] we compute the vacuum subtracted Renyi entropies
in our universal density matrix and find
(1+n) 2 (1+n) 2
4 (n +11)
∆Sn (ρB (E)) = 12n (R/λT ) − 120cn (R/λT ) 12n2
2 2
+ (1+n)
630c2 (R/λT )
6 (4−n )(n +47)
144n4 + ··· (22)
which is to be compared to the Renyi entropies
in the thermal reduced densitry matrix ρB (T ) =
TrB c (e−βH /Z):
(1+n) 2 (1+n) 4
∆Sn (ρB (T )) = 12n (R/λT ) − 120cn (R/λT )
+ (1+n) 6
630c2 (R/λT ) + · · · (23)
We observe that while for n > 1 the Renyi entropies
do not match, entanglement entropies (n = 1) match
perfectly. Discrepancy between Renyi entropies is a con-
sequence of infinite dimensionality of the Hilbert space.
The Fannes-Audenaert inequality and its generalizations
for n > 1 restrict the difference between entropies to
be bound by the trace distance multiplied by a factor
proportional to the dimension of the Hilbert space. As
the latter diverges, entanglement and Renyi entropies
could be different for arbitrarily close ρB (E) and ρB (T ).
Equivalence of entanglement entropies though can be fur-
ther used to show that
kρB (E) − ρB (T )k2 ≤ S(ρB (E)kρB (T ))
= ∆hH(ρB (T ))i − ∆S = e−O(S(E)) . (24)
Here H(ρB (T )) = − log ρB (T ) is a local integral over
T00 in 1 + 1-dimensions, and we have tuned the two
states to have the same energy density. Therefore,
∆hH(ρB (T ))i = 0 by construction. For a discussion of
ETH in 1 + 1-dimensional CFT at large central charge
see [9].
DISCUSSIONS
In this paper, we have argued that chaotic CFTs are
special in that their reduced density matrix on any finite
subsystem of arbitrary shape in energy eigenstates are
well-approximated by a universal density matrix. In or-
der to prove this we assumed local ETH. All integrable
models we checked (free theories and minimal models in
1+1-dimensions) failed to satisfy our local ETH assump-
tion. Therefore, we can interpret the local ETH assump-
tion as our working definition of chaos in CFTs. It would
be interesting to connect the local ETH to more stan-
dard definitions of quantum chaos in the literature. In
particular, one might hope to use the exponential decay
of out-of-time order correlators and bootstrap equations
to prove a statement similar to local ETH.
5
ACKNOWLEDGEMENTS
We would like to thank John Cardy, Thomas Faulkner,
Liam Fitzpatrick, Daniel Harlow, Thomas Hartman,
Tarun Grover, Mark Srednicki and Sasha Zhiboedov for
valuable discussions. The research of NL is supported
in part by funds provided by MIT-Skoltech Initiative.
This paper has the preprint number Technical Report
MIT-CTP/4841. This work is supported by the Office
of High Energy Physics of U.S. Department of Energy
under grant Contract Number de-sc0012567.
[1] M. Srednicki, J. Phys. A 29, L75 (1996)
doi:10.1088/0305-4470/29/4/003 [chao-dyn/9511001].
[2] M. Srednicki, Phys. Rev. A 43.4, 2046 (1991)
[3] M. Rigol, V. Dunjko, M. Olshanii Nature,4̇52(7189), 854
(2008).
[4] J. R. Garrison and T. Grover, arXiv:1503.00729 [cond-
mat.str-el].
[5] A. Dymarsky and H. Liu, arXiv:1511.06680 [cond-
mat.stat-mech].
[6] N. Lashkari and J. Simn, JHEP 1406, 038 (2014)
doi:10.1007/JHEP06(2014)038 [arXiv:1402.4829 [hep-
th]].
[7] N. Lashkari, Phys. Rev. Lett. 117, no. 4, 041601 (2016)
doi:10.1103/PhysRevLett.117.041601 [arXiv:1508.03506
[hep-th]].
[8] A. Dymarsky, N. Lashkari, H. Liu, In preparation.
[9] A. L. Fitzpatrick, J. Kaplan, M. T. Walters and J. Wang,
JHEP 1509, 019 (2015) doi:10.1007/JHEP09(2015)019
[arXiv:1504.01737 [hep-th]].
[10] N. Lashkari, Phys. Rev. Lett. 113, 051602 (2014)
doi:10.1103/PhysRevLett.113.051602 [arXiv:1404.3216
[hep-th]].
SUPPLEMENTARY MATERIAL
RINDLER SPACE: CONVENIENT CONFORMAL
FRAME
Consider a (d + 1)-dimensional CFT in radial quanti-
zation with a ball-shaped subsystem of angular size θ0 on
S d at r = 1. According to the operator/state correspon-
dence the density matrix in the subsystem is given by a
path-integral over the (d + 1)-dimensional space with two
operators inserted, Ψ at r =  and Ψ† at r = 1/ with
 → 0, and a cut open at the location of the subsystem.
The initial metric in the radial quantization is
ds2 = dr2 + r2 dΩ2d (25)
 6

with (θ1 , · · · θd ) the coordinates on S d . We perform the class of spinless primaries: (P 2 )l Ψa . The argument gen-
following conformal transformation eralizes to arbitrary descendants. Our eigenstates of in-
terest are labelled by (a, l). In a conformal theory the
L(r2 − 1) X0 matrix element of a scaling operator Op in these states
=
1 + r2 + 2r cos θ1 1 − 2X 1 + X · X is given by
2Lr sin θ1 cos θ2 (1 − X · X)/2
= hE(b,l) |Op |E(a,m) i
1 + r2 + 2r cos θ1 1 − 2X 1 + X · X
2Lr sin θ1 sin θ2 · · · sin θi Xi 1 hΨb (∞)(Kµ K µ )l Op (1)(Pν P ν )m Ψa (0)i
= , i>1 = p ,(29)
2
1 + r + 2r cos θ1 1 − 2X 1 + X · X Lhp hXa,m ihXb,l i
1 where hXa,m i = hΨa (∞)(Kµ K µ )m (Pν P ν )m Ψa (0)i.
L = cot(θ0 /2). (26)
2 Note that in radial quantization (Pµ |Ψi)† = hΨ|Kµ .
that maps the subsystem at r = 0 and θ1 ≤ θ0 to the Here, we assume that l and m are much smaller than
negative half-space, i.e. (0, X 1 < 0, 0 · · · 0). Here L is ha and hb .
the radius of S d in units where R is set to one. The new First, consider the term hXa,m i in the denominator.
metric in the X-coordinates that we call Rindler frame Primary field Ψa is killed by Kµ , so we only need to
is given by compute [(Kν K ν )m , (Pµ P µ )m ]Ψa . This operator can be
simplified by the successive application of the following
ds2 = Λ(X)2 dX i dX i commutation relations of the generators of the conformal
 −1 group:
0 LV− V+
Λ(X) = X − −
2 8L [Kµ , Pν ] = 2(δµν D − Mµν )
1
V± = (1 ± 2X + X · X). (27) [Mµν , Pρ ] = P[µ δν]ρ
[Mµν , Kρ ] = K[µ δν]ρ
In these coordinates the path-integral without operator
insertions prepares the Rindler density matrix in vacuum. [Mµν , Mρσ ] = Mρ][µ δν][σ . (30)
The operators Ψ and Ψ† are now inserted at X− and X+ Every time the operator D appears it acts on its eigen-
respectively. state and we get a factor ha + i, where 0 ≤ i ≤ m. All
other terms carry no powers of ha . At large ha we find
X± = (± sin θ0 , cos θ0 , 0 · · · , 0),
Λ(X− ) = (2 sin θ0 )−1 , hXa,m i = (d + 1)m (2ha )2m (1 + O(h−1
a )). (31)
−2 −1
Λ(X+ ) =  (2 sin θ0 ) . (28) The numerator in (29) has two terms

Under this map a conformal primary transforms accord- hΨq Op (Kµ K µ )l (Pν P ν )m Ψs (0)i
ing to +hΨq [(Kµ K µ )l , Op ](Pν P ν )m Ψs (0)i (32)

hΨ(r = 0) · · ·iΛ(X)δij = Λ(X(r = 0))−h hΨ(X(r = 0) · · ·iδij With no loss of generality we assume m ≥ l. Here we
argue that the first term wins over the second term in
Therefore, the large ha limit. As we saw above, the first term at
large ha scales as
hΨ(1/)Ψ() · · ·iradial = (2 sin θ0 )2h hΨ(X+ )Ψ(X− ) · · ·iRind
(d + 1)l (2hs )2l hΨq (∞)Op (1)(Pν P ν )m−l Ψs (0)i
In the thermodynamic limit θ0  1 the distance between = (d + 1)l (2hs )2l lim (∂ 2 )m−l hΨq (∞)Op (1)Ψs (z)i
z
Ψ and Ψ† goes to zero: |X+ − X− | = 2 sin θ0  1, and z→0
l 2l p
we use the OPE to obtain = (d + 1) (2hs ) Cqs fm−l (hp ) (33)
X
hΨ(1/)Ψ() · · ·iradial = 2h (2 sin θ0 )hp hOp (X0 ) · · ·i. where fm−l (hp ) is a polynomial of degree 2(m − l) in hp .
p To compute the second term in (32) we work out the
commutator

SPINLESS DESCENDANT EIGENSTATES [(Kµ K µ )l , Op ] = l[Kµ , Op ]K µ (Kα K α )l−1

+l(l − 1)[Kµ , [Kν , Op ]]K µ K ν (Kα K α )l−2
Local probes + · · · + [Kµ1 , [Kµ2 , · · · [Kµl , Op ]] · · · ]K µ1 · · · K µl .
The special conformal transformation generated by vec-
An arbitrary descendant energy eigenstate in confor-
tor field ξ µ sends
mal field is created by the operator Pµ1 · · · Pµn Ψν1 ···νm .
In order to simplify the presentation and avoid unnec- xµ − ξ µ x2
x0µ = (34)
essary manipulation of indices we focus on a particular (1 − 2ξµ xµ + 2 ξ 2 x2 )
 7

which transforms the scaling operator Op according to Density matrix

µ ν
(1 − 2ξµ xµ + 2 ξ 2 x2 )hp O(x0µ ) = e−iξµ K O(x)eiξν K . Now consider the spinless energy eigenstate |Ea,l i
created with a path-integral over the unit ball with
n
Matching the coefficients of n! in a series expansion on (P µ Pµ )l Ψa (0) = l Ψa (0) in the center of radial quan-
both sides gives tization. In the Rindler frame the Laplacian is
1

[Kµ1 , [Kµ2 · · · [Kµn , Op ] · · · ] = ∂i Λ(X)d−1 ∂i (40)

Λ(X)d+1
= ∂n (1 − 2ξµ xµ + 2 a2 ξ 2 )hp O(x0µ )
and the unit ball is mapped to the lower half-plane X 0 <
Terms that appear on the right-hand side of the equation 0. In these coordinates, the operators that create the
above have the form state and its conjugate are


l
f (xµ , ξ µ , hp )(∂ · · · ∂Op )Kµ1 · · · Kµj . (35) Ψ̃(X− ) = Λ(X )−d−1 ∂i Λ(X )d−1 ∂i Λ(X )−h Ψ(X )


l
Ψ̃† (X+ ) = lim −2(h+2l) Λ(X1/ )−d−1 ∂i Λ(X1/ )d−1 ∂i
Putting this back in the second term in (32) we obtain →0
terms that are × Λ(X1/ )−h Ψ(X1/ )

g(hp )(d + 1)j (2ha )2j hΨb (∂ · · · ∂Op )(1)Ψa i (36) where we have used the fact that under conjugation
X 0 → −X 0 . Note that
for j < l. Note that extra derivatives on Op do not lead to
Λ(X1/ ) ∼ −2 ,
any extra powers of ha . As a result, the first term in (32)
dominates. Putting all these terms back in equation (29) ∂in Λ(X1/ ) ∼ −(2+n) (41)
we obtain the matrix elements of Op in energy eigenbasis
and since ∂i Ψ will not carry any powers of  the conjugate
operator has the form
hE(b,l) |Op |E(a,m) i
p
Cab Ψ̃(X+ ) = f (L)Ψ(X+ ). (42)
= (d + 1)(l−m)/2 (2hs )l−m fm−l (hp )(1 + O(h−1
a )).
Lhp
Therefore, the density matrix in these coordinates is
In the case l = m the above expression becomes
hΨ̃(X− )Ψ(X+ ) · · ·i
tr(ρ · · · ) = (43)
p hhΨ̃(X− )Ψ(X+ )i
Cab
hE(b,l) |Op |E(a,m) i = (1 + O(h−1
a )). (37)
Lhp The operator Ψ̃Ψ simplifies further in the thermody-
The conformal algebra fixes their value in terms of p
Cab , namic limit L  1. To see this, apply ∂i to the OPE
ha , hb and hp . In appendices, we work out these matrix ∂ X ∂

elements and argue that at large ha and hb they are given i
Ψ(X)Ψ(X+ ) = i
(X − X+ )hp −2h O(X+ )
∂X p
∂X
by
= (−2h)(X − X+ )hp −2h−1 O(X+ ) + O(1/h), (44)
p
Cab
hE(b,l) |Op |E(a,m) i = g(m−l) (hp )hl−m
a Lhp (1 + O(hp /ha )) where we have used h  hp . Now, notice that
Lhp
Λ(X ) ∼ L,
where without loss of generality, we have assumed m ≥
l, and gk (hp ) is a polynomial of order 2k in hp with ∂in Λ(X ) ∼ Ln+1 .
g0 (hp ) = 1. Then, from equation (9) we find that
However when ∂i acts on Λ(X ) we get a factor of hL.
Therefore,
hE(a,m) |Op |E(a,m) i


Ψ̃(X− )Ψ(X+ ) = Λ(X− )−(h+2l) (∂i ∂i )2l Ψ (X− )Ψ(X+ ).
= hE(b,a−b+m) |Op |E(b,b−a+m) i 1 + O(h−1
a )


p
= Caa 1 + O(h−1
a ) , (38) Finally, the expression for the density matrix in these
coordinates becomes
which together with the assumption of local ETH for P hp p l −2h
primary energy eigenstates implies p (2/L) Cψψ (∂i ∂i ) (z − X+ ) z→X−
hOp · · ·i

(∂i ∂i )l (z − X+ )−2h z→X−

p X
hE(b,l) |Op |E(a,m) i = (Op (E)δab + Oab ) δlm + O(h−1
a ) . p
= (2/L)hp Cψψ hOp · · ·i (45)
(39) p
 8

which is the same as the density matrix in the primary RELATIVE ENTROPY REPLICA TRICK
state from which |Ea i descends. As a result, we find
(a,m)

kρB − ρaB k ∼ O (Ea L)−1 . (46) Similar to entanglement entropy there is a replica trick
that computes the relative entropy of arbitrary states
in quantum field theory [7, 10]. The relative entropy is
SUBSYSTEM ETH IMPLIES LOCAL ETH found from the analytic continuation in n of
P
Consider an observable A = a a|aiha| and the oper-  
1 Tr((ρ̃aB )n )Tr(ρ̃B )n−1
ator ρ − ρT in this basis: S(ρ̃aB kρ̃B ) = lim log
X n→1 n − 1 Tr(ρ̃aB ρ̃n−1 a n−1
B )Tr(ρ̃B )
ρ − ρT = cab |aihb|. (47)
ab
Now inserting (12) into the above expression we find that
The expectation value of A is
X X
Tr((ρ − ρT )A) = caa a ≤ |caa | a " #
(1) (2)
a a 1 + nAn + nAn + · · ·
!1/2 !1/2 S(ρ̃aB kρ̃B ) = ∂n log (1)
X X 1 + An n→1
2
≤ |caa | |caa |a Tr(RaB ρ̃n−1
B )
a a A(1)
n = n
!1/2 Tr(ρ̃B )
X n−2

≤ kΦ[ρ − ρT ]k1/2 |caa |a2 X Tr RaB ρ̃iB RaB ρ̃n−2−i
B
A(2)
n =
a Tr(ρ̃nB )
!1/2 i=1
X X Tr (Ra ρ̃m1 Ra ρ̃m2 · · · Ra ρ̃mk )
1/2 2
≤ kρ − ρT k |caa |a , (48) A(k)
n =
B B B B B B

a
Tr(ρ̃nB )
m1 +···mk =n−k
P
where Φ[ρ] = a ρaa |aiha| is the map that decoheres ρ in
the basis of A, and we have used the fact that this map
Consider the term A(k) . From the definition (13) we find
decreases the trace distance of operators. Note that caa
are real, but could have either sign. Denote by V and W X X
the projectors that project to caa that are, respectively, A(k)
n =
p1
(Oaa pk
· · · Oaa p1 ···pk
) fm Rhp1 +···hpk
1 ···mk
positive and negative. Then, m1 +···mk =n−k p1 ,···pk
X X

|caa |a2 = Tr V (ρ − ρT )V A2 p1 ···pk Tr (Op1 ρ̃m
B Op2 ρ̃B
1 m2
· · · Opk ρ̃m
B )
k

fm 1 ···mk
= n (51)
a a∈V Tr(ρ̃B )
X

− Tr W (ρ − ρT )W A2
p1 pk
a∈W

It is clear that (Oaa · · · Oaa ) = O(ηak ) where ηa =
2 p −O(S(E))
≤ Tr (ρ + ρT )A , (49) supp Oaa = e . Therefore, if we argue that
p1 ···pk
fm 1 ···mk is not entropically suppressed we have argued
where we used the fact that for a projector V :
A(k) = O(ηak ). Using (13) we have
Tr(V ρV A2 ) ≤ Tr(ρA2 ). Putting this back into (48)
we find

1/2 p1 ···pk
fm =
Tr((ρ − ρT )A) ≤ kρ − ρT k1/2 Tr (ρ + ρT )A2 (50) 1 ···mk
P m1 mk hp1 +···hpk
q1 ···qk (Oq1 (E) · · · Oqk (E)) hOp1 Oq1 · · · Opk Oqk iR
Repeating the argument above for ρ̃a,b,α
B − 21 (ρ̃B (Ea ) + P hq1 +···hqn
ρ̃B (Ea )) we find q1 ···qn (Oq1 (E) · · · Oqn (E)) hOq1 · · · Oqn iR

Tr((eiα σab + e−iα σba )A) ≤ keiα σab + e−iα σba k1/2 ×
 1/2 Neither the correlators, nor Op (E) have any entropic sup-
Tr (ρ̃a,b,α
B + 21 (ρ̃B (Ea ) + ρ̃B (Ea ))A2 . pressions. Therefore, as long as the correlators on the n-
for all α. By adding and subtracting the inequality above sheeted manifold are finite, the sums are convergent and
for values of α = 0 and α = π we find Tr(Aσab ) = A(k) = O(ηak ). Therefore, expanding (51) in ηa we have
e−O(S(E)) .
 
S(ρ̃aB kρ̃B ) = ∂n (n − 1)A(1) + O(ηa2 )
h Op 2
··· i n
n→1
−O(S(E))
Op 1 Op n
= O(ηa2 ) = e
log Op 1 Op n
. (52)

Op 2 ··· ⇢˜B (E)

⇢˜B (E) ⇢˜B (E)

FIG. 3: The relative entropy replica trick.