Introduction
Our understanding of black hole physics and the holographic AdS/CFT correspondence have been significantly advanced by study of the Sachdev-Ye-Kitaev (SYK) model [5–12]. In a certain parameter regime corresponding to low energy and low temperature, the model exhibits an approximate conformal behavior holographically dual to Jackiw-Teitelboim (JT) two-dimensional gravity. Furthermore SYK correlators exhibit thermalization and scrambling behavior related to the physics of probing the near horizon region of a black hole [16 , 17 ] [13, 14, 15].
Equally insightful is the double-scaled version of the SYK model (DSSYK) [18 , 19 , 20 , 21 , 22 ] [2], which is fully solvable using the chord diagram method, and whose holographic dual has been identified as sine-dilaton gravity [23 ] .
It would be helpful to have a similarly tractable and insightful model to study the cosmological horizon of de Sitter space and make progress on the notoriously difficult problem of defining and understanding de Sitter holography [24 , 25 , 26 ] . Susskind and other authors have investigated the possibility that perhaps DSSYK itself, at infinite temperature, can be used to some extent as a testbed for de Sitter horizon physics [27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 ] . Harlow and Zhao advanced this prospect by proposing a concrete prescription for “folding” out-of-time-order (OTOC) 4-point correlators of a quantum system with bounded energy, so as to obtain OTOCs relevant to probing de Sitter cosmological horizons, and suggested testing their folding prescription in the SYK model.
Here we will implement the folding of [1] on OTOCs of the DSSYK model, whose bounded spectrum makes the backward Euclidean evolution required by the fold well-defined. We make one key modification of their protocol: instead of applying it to thermal Euclidean OTOCs at some inverse temperature β \beta β , we apply it to β = 0 \beta= 0 β = 0 OTOCs with the DSSYK “fake” inverse temperature β f \beta_f β f defining the Euclidean thermal circle. The fake temperature is a natural choice given the results of [27 , 28 , 29 , 30 , 31 , 32 ] where the fake temperature is identified with the de Sitter temperature.
Using chord diagrams we demonstrate explicitly that the modified DSSYK OTOCs exhibit “anti-scrambling”, a characteristic feature of probing near a de Sitter cosmological horizon, related to a time advance rather than the time delay characteristic of probing near a black hole horizon in AdS. The folded OTOCs match to the expected de Sitter eikonal behavior as one approaches the scrambling time, with the Lyapunov exponent of the β = 0 \beta= 0 β = 0 theory, λ L = 2 π / β f \lambda_L = 2\pi/\beta_f λ L = 2 π / β f , which is the value the chaos bound [35 ] would allow at a temperature 1 / β f 1/\beta_f 1/ β f . The sheet of the analytically continued eikonal amplitude is fixed by the direction of the fold as confirmed by the exact DSSYK evaluation.
The fold itself defines the fake period at finite λ \lambda λ and the result agrees at NLO with the renormalized period of the two-point function. We also study the folded OTOC’s behavior at later times, beyond twice the scrambling time, where the analysis of [1] observed a late time de Sitter crunch.
In Section II we review the basics of the de Sitter dynamics and the proposal of [1]. In Sections III and IV we review the DSSYK model, its holographic dual, and the idea of fake temperature. In Section V we describe how to obtain the folded DSSYK OTOCs, and in Sections VI and VII we analyze their properties. Section VIII has a comparison of the folded correlator with the two-shell classical analysis of [1]. In Section IX we explain how our results are consistent with the “no-go” constraints obtained in [3, 4], and suggest some topics for further investigation.
de Sitter cosmological horizons and anti-scrambling
The scattering of ingoing and outgoing particles with a large relative boost near a black hole horizon can be described in terms of the eikonal approximation as a particle interacting with a gravitational shock wave [36 ] . Suppose W W W is an operator creating a particle at some time t = 0 t = 0 t = 0 moving close to the horizon with lightcone momentum q − q_- q − , and let the operator V V V create a probe at a later time t t t with lightcone momentum p + p_+ p + . Looking at the 1 + 1 1+1 1 + 1 dimensional case, the probe crossing the boosted gravitational shock wave has the effect of abruptly shifting its Kruskal lightcone coordinate by some amount x e 2 π t / β q − x e^{2\pi t/\beta}q_- x e 2 π t / β q − , where x x x is positive and proportional to Newton’s constant G N G_N G N , and e 2 π t / β e^{2\pi t/\beta} e 2 π t / β is the near horizon exponential boost factor, whose rate is set by the Hawking temperature 1 / β 1/\beta 1/ β . This shift is called the time delay. Quantum mechanically the effect on the probe’s wavefunction is that it picks up the eikonal phase factor e i a ( t ) p + q − e^{ia(t)p_+q_-} e ia ( t ) p + q − where a ( t ) = x e 2 π t / β a(t) = x e^{2\pi t/\beta} a ( t ) = x e 2 π t / β ; this is known as the 1 + 1 1+1 1 + 1 dimensional eikonal S S S matrix. The effect is a time delay because the shock’s positive energy causes the black hole horizon size to grow; the near horizon probe finds itself closer to the horizon and its clock slows down. We note in passing that a negative energy shock wave, as is engineered for the Gao-Jafferis-Wall traversable wormhole [37 ] , shrinks the horizon and causes a time advance.
Continuing to Euclidean time converts the eikonal S S S matrix into a real exponential whose expectation value we could in principle compute:
Φ ( a ) = ⟨ e − a ( t ) p + q − ⟩ = ∫ 0 ∞ d p ∫ 0 ∞ d q ρ V ( p ) ρ W ( q ) e − a ( t ) p q (1) \Phi(a)=\left\langle e^{-a(t)p_{+}q_{-}}\right\rangle=\int_{0}^{\infty}dp\int_{0}^{\infty}dq\,\rho_{V}(p)\rho_{W}(q)e^{-a(t)pq} \tag*{(1)} Φ ( a ) = ⟨ e − a ( t ) p + q − ⟩ = ∫ 0 ∞ d p ∫ 0 ∞ d q ρ V ( p ) ρ W ( q ) e − a ( t ) pq (1)
where in the second equality we introduced explicit momentum wavefunctions ρ V , ρ W \rho_{V},\rho_{W} ρ V , ρ W .
Focusing now on JT gravity in the approximately conformal Schwarzian picture, we can write a ( t ) ∼ ( Δ V Δ W / C ) e 2 π t / β a(t)\sim(\Delta_{V}\Delta_{W}/C)e^{2\pi t/\beta} a ( t ) ∼ ( Δ V Δ W / C ) e 2 π t / β , where Δ V , Δ W \Delta_{V},\Delta_{W} Δ V , Δ W are the conformal dimensions of the probes, and C C C is the Schwarzian coupling. Taking Δ V = Δ W = 1 \Delta_{V}=\Delta_{W}=1 Δ V = Δ W = 1 , the Fourier/Legendre transform of the conformal 2-point correlators gives:
ρ V ( p ) = p e − p ; ρ W ( q ) = q e − q (2) \rho_{V}(p)=p e^{-p};\qquad\rho_{W}(q)=q e^{-q} \tag*{(2)} ρ V ( p ) = p e − p ; ρ W ( q ) = q e − q (2)
from which we get the eikonal function
Φ ( a ) = ∫ 0 ∞ d p p e − p ( 1 + a p ) 2 = 1 a 2 U ( 2 , 1 , 1 a ) = w 2 [ ( 1 + w ) e w E 1 ( w ) − 1 ] , w = 1 / a (3) \Phi(a)=\int_{0}^{\infty}dp\,\frac{p e^{-p}}{(1+ap)^{2}}=\frac{1}{a^{2}}U\left(2,1,\frac{1}{a}\right)=w^{2}\left[(1+w)e^{w}E_{1}(w)-1\right],\qquad w=1/a \tag*{(3)} Φ ( a ) = ∫ 0 ∞ d p ( 1 + a p ) 2 p e − p = a 2 1 U ( 2 , 1 , a 1 ) = w 2 [ ( 1 + w ) e w E 1 ( w ) − 1 ] , w = 1/ a (3)
where U U U is a confluent hypergeometric function, E 1 E_{1} E 1 is the exponential integral, and Φ \Phi Φ is normalized such that Φ ( a ) → 1 \Phi(a)\to1 Φ ( a ) → 1 as a → 0 a\to0 a → 0 ; its asymptotic expansion is Φ ( a ) = 1 − 4 a + 18 a 2 − 96 a 3 + … \Phi(a)=1-4a+18a^{2}-96a^{3}+\ldots Φ ( a ) = 1 − 4 a + 18 a 2 − 96 a 3 + … ; Φ ( a ) \Phi(a) Φ ( a ) is exactly the function F ( w ) ≡ − 4 w 2 I ( w ) F(w)\equiv-4w^{2}I(w) F ( w ) ≡ − 4 w 2 I ( w ) of [1] (their equation (4.32)), with w = 1 / a w=1/a w = 1/ a .
As reviewed in [1] when we probe de Sitter cosmological horizons the time delay discussed above becomes a time advance. This is because a de Sitter observer views the cosmological horizon “from the inside”, and a positive energy particle moving through a de Sitter horizon moves the horizon away from the observer rather than towards them.
In the eikonal analysis above, Harlow and Zhao as well as Cui and Kolchmeyer [3] observe that one can get the de Sitter version by simply taking Φ ( a ( t ) ) → Φ ( − a ( t ) ) \Phi(a(t))\to\Phi(-a(t)) Φ ( a ( t )) → Φ ( − a ( t )) . This implies a time advance in the shockwave picture rather than a time delay. The integral in equation (3) then has a double pole on the contour at p = 1 / a p=1/a p = 1/ a , and the double integral over p p p and q q q no longer converges: the “attractive” e + a p q e^{+apq} e + a pq beats the e − ( p + q ) e^{-(p+q)} e − ( p + q ) suppression at large momenta. Φ ( − a ) \Phi(-a) Φ ( − a ) is instead defined by continuing the closed form, w → − w w \to-w w → − w , with E 1 ( − w ) = − Ei ( w ) ∓ i π E_1(-w)=-\operatorname{Ei}(w)\mp i\pi E 1 ( − w ) = − Ei ( w ) ∓ iπ ; the two signs correspond to the two sides of the branch cut of E 1 E_1 E 1 along the negative real axis. The sign is not fixed by the replacement Φ ( a ) → Φ ( − a ) \Phi(a)\to\Phi(-a) Φ ( a ) → Φ ( − a ) alone; we will see in § 0.6 A that each direction of fold approaches the cut from a definite side, confirmed by the exact DSSYK numerics of Section VII.
Thus in the de Sitter case Φ \Phi Φ is complex; the real part is common to both sheets:
Re Φ ( − a ) = w 2 [ ( 1 − w ) e − w ( − Ei ( w ) ) − 1 ] = 1 + 4 a + 18 a 2 + 96 a 3 + ⋯ (4) \operatorname{Re}\Phi(-a)=w^2\left[(1-w)e^{-w}\left(-\operatorname{Ei}(w)\right)-1\right]=1+4a+18a^2+96a^3+\cdots \tag*{(4)} Re Φ ( − a ) = w 2 [ ( 1 − w ) e − w ( − Ei ( w ) ) − 1 ] = 1 + 4 a + 18 a 2 + 96 a 3 + ⋯ (4)
all terms positive. The imaginary part is ± π w 2 ( w − 1 ) e − w \pm\pi w^2(w-1)e^{-w} ± π w 2 ( w − 1 ) e − w , which is exponentially small when a a a is small. For sufficiently small a a a we can also approximate the de Sitter Φ \Phi Φ from the first terms in the asymptotic series 1 + 4 a + 18 a 2 + ⋯ 1+4a+18a^2+\cdots 1 + 4 a + 18 a 2 + ⋯ .
The consequence is that the leading correction at early times has the opposite sign from the black-hole case, i.e. it initially grows instead of decaying. Harlow and Zhao call this anti-scrambling. Cui and Kolchmeyer [3] (see also [38 ] [4]) showed that the time advance is incompatible with the KMS condition for the vacuum correlators of the observer’s matter operators, and that, once the observer’s clock is quantized, the Hartle–Hawking state is not a trace on the observer’s algebra.
A way around this was proposed in [1]. In a quantum system whose Hamiltonian is bounded above and below, backward Euclidean evolution e + τ H e^{+\tau H} e + τ H exists, and one can in principle define correlators whose Euclidean contour is “folded”. Their prescription: take the thermal four-point OTOC Tr [ e − β H V ( τ 4 ) W ( τ 3 ) V ( τ 2 ) W ( τ 1 ) ] \operatorname{Tr}[e^{-\beta H}V(\tau_4)W(\tau_3)V(\tau_2)W(\tau_1)] Tr [ e − β H V ( τ 4 ) W ( τ 3 ) V ( τ 2 ) W ( τ 1 )] , 0 < τ 1 < ⋯ < τ 4 < β 0<\tau_1<\cdots<\tau_4<\beta 0 < τ 1 < ⋯ < τ 4 < β , and replace
( τ 1 , τ 2 , τ 3 , τ 4 ) ⟶ ( − τ 1 , n β − τ 2 , β − τ 3 , ( n + 1 ) β − τ 4 ) , n ∈ Z . (5) (\tau_1,\tau_2,\tau_3,\tau_4)\longrightarrow(-\tau_1,n\beta-\tau_2,\beta-\tau_3,(n+1)\beta-\tau_4),\qquad n\in\mathbb{Z}. \tag*{(5)} ( τ 1 , τ 2 , τ 3 , τ 4 ) ⟶ ( − τ 1 , n β − τ 2 , β − τ 3 , ( n + 1 ) β − τ 4 ) , n ∈ Z . (5)
The integer n n n is undetermined in the semiclassical analysis; at finite λ \lambda λ different n n n correspond to different correlators; here we consider the two minimal folds n = 0 , 1 n=0,1 n = 0 , 1 , which turn out to be complex conjugates of each other (§ 0.5 ). They note that the prescription assumes the analytic continuation commutes with the semiclassical limit, that it could fail through a Stokes phenomenon (a point we address in § 0.6 B), and that it covers only the regime where the backreaction is “mild”. They suggested testing their prescription in the SYK model.
DSSYK and sine-dilaton gravity
The chord diagram approach to DSSYK is nicely reviewed by Berkooz and Mamroud [39 ] .
The p p p -local SYK Hamiltonian is
H = i p / 2 ∑ I J I Ψ I (6) H = i^{p/2} \sum_I J_I \Psi_I \tag*{(6)} H = i p /2 I ∑ J I Ψ I (6)
where the I I I denote ordered strings of p p p indices i 1 , i 2 , … i p i_1,i_2,\ldots i_p i 1 , i 2 , … i p ranging over N N N possible values, the J I J_I J I denote Gaussian random couplings times an overall energy scale J J J , and Ψ I \Psi_I Ψ I denotes a string of Majorana fermions with indices in I I I . DSSYK is defined as the double-scaled limit N , p → ∞ N,p \to\infty N , p → ∞ with λ = 2 p 2 / N \lambda= 2p^2/N λ = 2 p 2 / N fixed and q = e − λ q=e^{-\lambda} q = e − λ . The Gaussian couplings have variance chosen so that Tr [ H 2 ] = J 2 \operatorname{Tr}[H^2]=J^2 Tr [ H 2 ] = J 2 , where we use the normalized trace Tr [ 1 ] = 1 \operatorname{Tr}[1]=1 Tr [ 1 ] = 1 . The disorder-averaged moments are chord-diagram sums,
Tr [ H 2 k ] = J 2 k ∑ pairings of 2 k points q # crossings , (7) \operatorname{Tr}[H^{2k}] = J^{2k} \sum_{\text{pairings of }2k\text{ points}} q^{\#\text{crossings}}, \tag*{(7)} Tr [ H 2 k ] = J 2 k pairings of 2 k points ∑ q # crossings , (7)
The energy spectrum is given by
E ( s ) = − E 0 cos ( λ s ) , s ∈ [ 0 , π / λ ] , E 0 = 2 J 1 − q , (8) E(s)=-E_0\cos(\lambda s), \qquad s\in[0,\pi/\lambda], \qquad E_0=\frac{2J}{\sqrt{1-q}}, \tag*{(8)} E ( s ) = − E 0 cos ( λ s ) , s ∈ [ 0 , π / λ ] , E 0 = 1 − q 2 J , (8)
thus the spectrum is bounded from above and below by ± E 0 \pm E_0 ± E 0 ; this is what makes DSSYK a candidate for the construction of [1], which requires backward Euclidean evolution e + τ H e^{+\tau H} e + τ H to exist.
The holographic bulk dual of DSSYK has been identified as two-dimensional sine-dilaton gravity [23 ] . With our conventions, and setting 8 π G N = 1 8\pi G_N=1 8 π G N = 1 , the sine-dilaton action is given by:
1 λ [ 1 2 ∫ d 2 x g ( Φ s d R + 2 sin Φ s d ) + ( boundary terms ) ] (9) \frac{1}{\lambda}\left[\frac{1}{2}\int d^2x\sqrt{g}\left(\Phi_{\mathrm{sd}}R+2\sin\Phi_{\mathrm{sd}}\right)+(\text{boundary terms})\right] \tag*{(9)} λ 1 [ 2 1 ∫ d 2 x g ( Φ sd R + 2 sin Φ sd ) + ( boundary terms ) ] (9)
where Φ s d \Phi_{\mathrm{sd}} Φ sd is the dilaton field of [23 ] after their rescaling, and R R R is the Ricci scalar of the metric g g g . In the Schwarzschild gauge the classical solutions of sine-dilaton gravity are parameterized by an angle θ \theta θ and can be written:
d s 2 = − F d t 2 + d r 2 F , Φ s d = r , F ( r ) = 2 ( cos θ − cos r ) (10) ds^2=-F\,dt^2+\frac{dr^2}{F}, \qquad\Phi_{\mathrm{sd}}=r, \qquad F(r)=2(\cos\theta-\cos r) \tag*{(10)} d s 2 = − F d t 2 + F d r 2 , Φ sd = r , F ( r ) = 2 ( cos θ − cos r ) (10)
The ADM energy of these solutions is given by E = − E 0 cos θ E=-E_0\cos\theta E = − E 0 cos θ , with the boundary time normalized as for the DSSYK Hamiltonian above. These solutions have a black hole horizon at r = θ r=\theta r = θ and a cosmological horizon at r = 2 π − θ r=2\pi-\theta r = 2 π − θ .
Like other dilaton gravities, sine-dilaton gravity has a conserved charge, constant throughout a region without matter; it can be written as:
κ s d ( θ ) = ( ∇ Φ s d ) 2 + 2 cos Φ s d = 2 cos θ (11) \kappa_{\mathrm{sd}}(\theta) = (\nabla\Phi_{\mathrm{sd}})^2 + 2\cos\Phi_{\mathrm{sd}} = 2\cos\theta \tag*{(11)} κ sd ( θ ) = ( ∇ Φ sd ) 2 + 2 cos Φ sd = 2 cos θ (11)
DSSYK is a complete quantum theory of two-dimensional dilaton gravity whose spectrum is bounded above and below. In the low temperature limit with λ → 0 \lambda\to0 λ → 0 , s s s fixed, DSSYK reduces to the quantum Schwarzian system dual to JT gravity.
In a different limit λ → 0 \lambda\to0 λ → 0 , λ s \lambda s λ s fixed, β = x / λ \beta= x/\sqrt{\lambda} β = x / λ with x x x fixed, and J = 1 J=1 J = 1 , the DSSYK partition function can be estimated using a saddle point [2, 8]. The partition function is given by (see eqs. (3.8) and (3.9) of [2] [40]):
Z ( β ) = ∫ 0 π / λ d s ρ q ( s ) e − β E ( s ) (12) Z(\beta) = \int_{0}^{\pi/\lambda} ds\,\rho_q(s)e^{-\beta E(s)} \tag*{(12)} Z ( β ) = ∫ 0 π / λ d s ρ q ( s ) e − β E ( s ) (12)
where, to leading order in small λ \lambda λ , − β E ( s ) = ( 2 x / λ ) cos ( θ ) -\beta E(s) = (2x/\lambda)\cos(\theta) − β E ( s ) = ( 2 x / λ ) cos ( θ ) , where s = θ / λ s = \theta/\lambda s = θ / λ , with θ ∈ [ 0 , π ] \theta\in[0,\pi] θ ∈ [ 0 , π ] .
ρ q ( s ) ∝ e S ( θ ) / λ ; S ( θ ) = 2 ( π θ − θ 2 ) (13) \rho_q(s) \propto e^{S(\theta)/\lambda}; \qquad S(\theta) = 2(\pi\theta-\theta^2) \tag*{(13)} ρ q ( s ) ∝ e S ( θ ) / λ ; S ( θ ) = 2 ( π θ − θ 2 ) (13)
We see that the total exponent [ S ( θ ) + 2 x cos θ ] / λ [S(\theta) + 2x\cos\theta]/\lambda [ S ( θ ) + 2 x cos θ ] / λ is of order 1 / λ 1/\lambda 1/ λ , with a saddle point where its derivative with respect to θ \theta θ vanishes, giving:
v 2 cos π 2 v = x 2 π , v = 1 − 2 π θ (14) \frac{v}{2\cos\frac{\pi}{2}v} = \frac{x}{2\pi}, \qquad v = 1-\frac{2}{\pi}\theta \tag*{(14)} 2 cos 2 π v v = 2 π x , v = 1 − π 2 θ (14)
Expanding the exponent to second order about the saddle θ s \theta_s θ s , the thermal distribution of θ \theta θ is Gaussian with variance:
σ θ 2 = λ 4 + 2 x cos θ s (15) \sigma_\theta^2 = \frac{\lambda}{4+2x\cos\theta_s} \tag*{(15)} σ θ 2 = 4 + 2 x cos θ s λ (15)
We will be interested in the infinite temperature case β = 0 \beta= 0 β = 0 , where the saddle point is at the midpoint θ = π / 2 \theta= \pi/2 θ = π /2 of the energy spectrum, and the variance reduces to λ / 4 \lambda/4 λ /4 .
Fake temperature in DSSYK
The fake temperature of DSSYK appears when we consider two-point functions of “matter” operators. We introduce an unnormalized basis of chord states ∣ n ⟩ |n\rangle ∣ n ⟩ , n ≥ 0 n \ge0 n ≥ 0 , labeled by the number of open chords, and the transfer matrix
T ∣ n ⟩ = ∣ n + 1 ⟩ + [ n ] q ∣ n − 1 ⟩ , [ n ] q = 1 − q n 1 − q (16) T|n\rangle= |n+1\rangle+ [n]_q|n-1\rangle, \qquad[n]_q = \frac{1-q^n}{1-q} \tag*{(16)} T ∣ n ⟩ = ∣ n + 1 ⟩ + [ n ] q ∣ n − 1 ⟩ , [ n ] q = 1 − q 1 − q n (16)
where ∣ n + 1 ⟩ |n+1\rangle ∣ n + 1 ⟩ opens a chord (weight 1) and [ n ] q = ∑ j = 0 n − 1 q j [n]_q=\sum_{j=0}^{n-1}q^j [ n ] q = ∑ j = 0 n − 1 q j closes one of the n n n open chords, summed over which one, each crossing the j j j chords opened after it. The vacuum-to-vacuum matrix element ⟨ 0 ∣ T 2 k ∣ 0 ⟩ \langle0|T^{2k}|0\rangle ⟨ 0∣ T 2 k ∣0 ⟩ , with ⟨ 0 ∣ \langle0| ⟨ 0∣ the functional extracting the ∣ 0 ⟩ |0\rangle ∣0 ⟩ coefficient, is then exactly the chord-diagram sum, so the Hamiltonian restricted to the chord Hilbert space is H = J T H=\mathcal{J}T H = J T , with ⟨ 0 ∣ T 2 ∣ 0 ⟩ = 1 \langle0|T^2|0\rangle=1 ⟨ 0∣ T 2 ∣0 ⟩ = 1 .
Introduce now some matter operator V V V , which is a sum over strings of p ~ \tilde p p ~ Majoranas:
V = i p ~ / 2 ∑ I ′ J ‾ I ′ Ψ I ′ (17) V=i^{\tilde p/2}\sum_{I'}\overline{J}_{I'}\Psi_{I'} \tag*{(17)} V = i p ~ /2 I ′ ∑ J I ′ Ψ I ′ (17)
where the J ‾ I ′ \overline{J}_{I'} J I ′ are Gaussian random couplings. We now have to keep track of two kinds of chords: H-chords related to the Hamiltonian, and matter chords related to V V V . The intersection of two H-chords has weight q q q , but the intersection of an H-chord with a matter chord has weight q Δ q^\Delta q Δ , where Δ = p ~ / p \Delta=\tilde p/p Δ = p ~ / p and in the semiclassical conformal regime corresponds to the conformal dimension of V V V .
The Euclidean two-point function of V V V is
G ( τ ) = ⟨ 0 ∣ e − ( β − τ ) T q Δ n ^ e − τ T ∣ 0 ⟩ Z ( β ) (18) G(\tau)=\frac{\langle0|e^{-(\beta-\tau)T}q^{\Delta\hat n}e^{-\tau T}|0\rangle}{Z(\beta)} \tag*{(18)} G ( τ ) = Z ( β ) ⟨ 0∣ e − ( β − τ ) T q Δ n ^ e − τ T ∣0 ⟩ (18)
where n ^ \hat n n ^ is the chord number operator n ^ ∣ n ⟩ = n ∣ n ⟩ \hat n|n\rangle=n|n\rangle n ^ ∣ n ⟩ = n ∣ n ⟩ , Z ( β ) Z(\beta) Z ( β ) is the partition function and β \beta β is the inverse temperature. From now on we will be focused on DSSYK correlators in the infinite temperature limit β → 0 \beta\to0 β → 0 . Switching to Lorentzian time t t t and letting G 0 ( i t ) G_0(it) G 0 ( i t ) denote the β = 0 \beta=0 β = 0 correlator and using Z ( 0 ) = 1 Z(0)=1 Z ( 0 ) = 1 , we have:
G 0 ( i t ) = ⟨ 0 ∣ e i t T q Δ n ^ e − i t T ∣ 0 ⟩ (19) G_0(it)=\langle0|e^{itT}q^{\Delta\hat n}e^{-itT}|0\rangle \tag*{(19)} G 0 ( i t ) = ⟨ 0∣ e i tT q Δ n ^ e − i tT ∣0 ⟩ (19)
In the semiclassical limit λ → 0 \lambda\to0 λ → 0 at fixed J t \mathcal{J}t J t , G 0 ( i t ) G_0(it) G 0 ( i t ) reduces to the form [39 ] [8]
G 0 s c ( i t ) = cosh ( J t ) − 2 Δ (20) G_0^{\mathrm{sc}}(it)=\cosh(\mathcal{J}t)^{-2\Delta} \tag*{(20)} G 0 sc ( i t ) = cosh ( J t ) − 2Δ (20)
We can determine J \mathcal{J} J by expanding (19) to leading order in t t t and matching to (20) expanded to the same order:
J 2 = J 2 1 − q Δ Δ = J 2 λ ( 1 − Δ 2 λ + ⋯ ) (21) \mathcal{J}^2=\mathcal{J}^2\frac{1-q^\Delta}{\Delta}=\mathcal{J}^2\lambda\left(1-\frac{\Delta}{2}\lambda+\cdots\right) \tag*{(21)} J 2 = J 2 Δ 1 − q Δ = J 2 λ ( 1 − 2 Δ λ + ⋯ ) (21)
i.e. J = λ J \mathcal{J}=\sqrt{\lambda}\mathcal{J} J = λ J at leading order. From now on we will set J = 1 \mathcal{J}=1 J = 1 .
We can write G 0 ( i t ) G_0(it) G 0 ( i t ) as the Fourier transform of a spectral density ρ Δ \rho_\Delta ρ Δ :
G 0 ( i t ) = ∫ d ω 2 π ρ Δ ( ω ) e − i ω t , ρ Δ ( ω ) = 2 π D ∑ i j ∣ V i j ∣ 2 δ ( ω − E i + E j ) (22) G_0(it)=\int\frac{d\omega}{2\pi}\rho_\Delta(\omega)e^{-i\omega t},\qquad\rho_\Delta(\omega)=\frac{2\pi}{D}\sum_{ij}|V_{ij}|^2\delta(\omega-E_i+E_j) \tag*{(22)} G 0 ( i t ) = ∫ 2 π d ω ρ Δ ( ω ) e − iω t , ρ Δ ( ω ) = D 2 π ij ∑ ∣ V ij ∣ 2 δ ( ω − E i + E j ) (22)
where D D D is the dimension of the Hilbert space. In the semiclassical limit the Fourier transform is a Beta-function integral, giving:
ρ Δ s c ( ω ) = ∫ d t e i ω t cosh − 2 Δ ( J t ) = 2 2 Δ − 1 J Γ ( 2 Δ ) ∣ Γ ( Δ + i ω 2 J ) ∣ 2 . (23) \rho_{\Delta}^{\mathrm{sc}}(\omega)=\int dt\,e^{i\omega t}\cosh^{-2\Delta}(\mathcal{J}t)=\frac{2^{2\Delta-1}}{\mathcal{J}\,\Gamma(2\Delta)}\left|\Gamma\left(\Delta+\frac{i\omega}{2\mathcal{J}}\right)\right|^2. \tag*{(23)} ρ Δ sc ( ω ) = ∫ d t e iω t cosh − 2Δ ( J t ) = J Γ ( 2Δ ) 2 2Δ − 1 Γ ( Δ + 2 J iω ) 2 . (23)
Using Stirling’s formula ∣ Γ ( Δ + i x ) ∣ 2 ≃ 2 π ∣ x ∣ 2 Δ − 1 e − π ∣ x ∣ \left|\Gamma(\Delta+ix)\right|^2\simeq2\pi|x|^{2\Delta-1}e^{-\pi|x|} ∣ Γ ( Δ + i x ) ∣ 2 ≃ 2 π ∣ x ∣ 2Δ − 1 e − π ∣ x ∣ as ∣ x ∣ → ∞ |x|\to\infty ∣ x ∣ → ∞ with x = ω / 2 J x=\omega/2\mathcal{J} x = ω /2 J , we get an expression for the tails of the spectral density:
ρ Δ ( ω ) ∝ ∣ ω ∣ 2 Δ − 1 e − π ∣ ω ∣ / 2 J (24) \rho_{\Delta}(\omega)\propto|\omega|^{2\Delta-1}e^{-\pi|\omega|/2\mathcal{J}} \tag*{(24)} ρ Δ ( ω ) ∝ ∣ ω ∣ 2Δ − 1 e − π ∣ ω ∣/2 J (24)
In the variable p = π ω / ( 2 J ) p=\pi\omega/(2\mathcal{J}) p = π ω / ( 2 J ) this is ∝ p 2 Δ − 1 e − p \propto p^{2\Delta-1}e^{-p} ∝ p 2Δ − 1 e − p . Comparing with the eikonal momentum distributions of (2) we see that they are related to the tails of the spectral function.
Using (20) and (21) we can define a fake inverse temperature β f = π / J ≃ π / λ \beta_f=\pi/\mathcal{J}\simeq\pi/\sqrt{\lambda} β f = π / J ≃ π / λ (at finite λ \lambda λ the coupling that sets the fake period will differ from the short-time coupling J \mathcal{J} J of (21) at O ( λ ) O(\lambda) O ( λ ) ; we return to this below and write β f ≡ π / J f \beta_f\equiv\pi/\mathcal{J}_f β f ≡ π / J f for the former):
G 0 s c ( τ ) = cos ( π τ β f ) − 2 Δ = sin [ π β f ( τ + β f / 2 ) ] − 2 Δ (25) G_0^{\mathrm{sc}}(\tau)=\cos\left(\frac{\pi\tau}{\beta_f}\right)^{-2\Delta}=\sin\left[\frac{\pi}{\beta_f}\left(\tau+\beta_f/2\right)\right]^{-2\Delta} \tag*{(25)} G 0 sc ( τ ) = cos ( β f π τ ) − 2Δ = sin [ β f π ( τ + β f /2 ) ] − 2Δ (25)
Rewriting the tails of the spectral density as
ρ Δ ( ω ) ∝ ∣ ω ∣ 2 Δ − 1 e − β f ∣ ω ∣ / 2 (26) \rho_{\Delta}(\omega)\propto|\omega|^{2\Delta-1}e^{-\beta_f|\omega|/2} \tag*{(26)} ρ Δ ( ω ) ∝ ∣ ω ∣ 2Δ − 1 e − β f ∣ ω ∣/2 (26)
and examining the convergence for complex t = t r − i τ t=t_r-i\tau t = t r − i τ of the integrand in (22) , we see that convergence at ω → + ∞ \omega\to+\infty ω → + ∞ requires τ > − β f / 2 \tau>-\beta_f/2 τ > − β f /2 and at ω → − ∞ \omega\to-\infty ω → − ∞ requires τ < β f / 2 \tau<\beta_f/2 τ < β f /2 , so G 0 s c ( i t ) G_0^{\mathrm{sc}}(it) G 0 sc ( i t ) is analytic in the strip ∣ Im t ∣ < β f / 2 |\operatorname{Im}t|<\beta_f/2 ∣ Im t ∣ < β f /2 .
From (25) we see that, for integer Δ \Delta Δ , G 0 s c ( τ ) G_0^{\mathrm{sc}}(\tau) G 0 sc ( τ ) is periodic in τ \tau τ with period β f \beta_f β f , and for Δ \Delta Δ half-integer it is antiperiodic. Taking the operators at two real times G 0 s c ( i t 2 − i t 1 ) = Tr ( V ( i t 2 ) V ( i t 1 ) ) G_0^{\mathrm{sc}}(it_2-it_1)=\operatorname{Tr}(V(it_2)V(it_1)) G 0 sc ( i t 2 − i t 1 ) = Tr ( V ( i t 2 ) V ( i t 1 )) , we see that on the fake thermal circle the two operators are shifted by β f / 2 \beta_f/2 β f /2 , i.e. they are at antipodal locations. This is why the correlator is nonsingular when the two operator insertions are at coincident real times.
To summarize, in the semiclassical regime the DSSYK β = 0 \beta=0 β = 0 correlator Tr ( V ( τ 2 ) V ( τ 1 ) ) \operatorname{Tr}(V(\tau_2)V(\tau_1)) Tr ( V ( τ 2 ) V ( τ 1 )) behaves like a fake temperature thermal correlator
Tr ( e − β f H V ( τ 2 + β f / 2 ) V ( τ 1 ) ) = Tr ( e − β f H / 2 V ( τ 2 ) e − β f H / 2 V ( τ 1 ) ) (27) \operatorname{Tr}\left(e^{-\beta_f H}V(\tau_2+\beta_f/2)V(\tau_1)\right)=\operatorname{Tr}\left(e^{-\beta_f H/2}V(\tau_2)e^{-\beta_f H/2}V(\tau_1)\right) \tag*{(27)} Tr ( e − β f H V ( τ 2 + β f /2 ) V ( τ 1 ) ) = Tr ( e − β f H /2 V ( τ 2 ) e − β f H /2 V ( τ 1 ) ) (27)
In the DSSYK/de Sitter proposals of [27 , 28 , 29 , 30 , 31 , 32 ] it is this fake temperature that plays the role of the de Sitter temperature.
The fake temperature analysis above was, strictly speaking, only valid in the λ → 0 \lambda\to0 λ → 0 semiclassical limit. Indeed for finite λ \lambda λ the DSSYK β = 0 \beta= 0 β = 0 correlator Tr ( V ( τ 2 ) V ( τ 1 ) ) \operatorname{Tr}(V(\tau_2)V(\tau_1)) Tr ( V ( τ 2 ) V ( τ 1 )) can be evaluated exactly [20 ] , and has no Euclidean time periodicity in τ 2 − τ 1 \tau_2-\tau_1 τ 2 − τ 1 ; the quantum corrections also smooth out the singularity of (25) at τ 2 − τ 1 = ± β f / 2 \tau_2-\tau_1=\pm\beta_f/2 τ 2 − τ 1 = ± β f /2 , the coincident point on the fake circle, since the exact β = 0 \beta=0 β = 0 correlator ∑ i j ∣ V i j ∣ 2 e τ ( E i − E j ) / D \sum_{ij}|V_{ij}|^2e^{\tau(E_i-E_j)/D} ∑ ij ∣ V ij ∣ 2 e τ ( E i − E j ) / D is an entire function of τ \tau τ for a bounded spectrum. So it is worth investigating to what extent the fake thermal circle picture survives at all for small but finite λ \lambda λ .
We notice first from (21) that, beyond leading order in finite λ \lambda λ , the value of J \mathcal{J} J depends on Δ \Delta Δ and thus the operator size/dimension. With this caveat, we can write an expression for the “one-loop” finite λ \lambda λ corrections to the fake-thermal two-point function:
G 0 ( i t ) = cosh ( J t ) − 2 Δ [ 1 + λ g ( J t ) + O ( λ 2 ) ] , J 2 = 1 − q Δ Δ , (28) G_0(it)=\cosh(\mathcal{J}t)^{-2\Delta}\left[1+\lambda g(\mathcal{J}t)+O(\lambda^2)\right], \qquad\mathcal{J}^2=\frac{1-q^\Delta}{\Delta}, \tag*{(28)} G 0 ( i t ) = cosh ( J t ) − 2Δ [ 1 + λ g ( J t ) + O ( λ 2 ) ] , J 2 = Δ 1 − q Δ , (28)
where g ( J t ) g(\mathcal{J}t) g ( J t ) is the one-loop correction. Taking now Δ = 1 \Delta=1 Δ = 1 , and using the exact DSSYK correlator, we can produce an approximate closed form expression, valid for J t ≳ 1.5 \mathcal{J}t\gtrsim1.5 J t ≳ 1.5 in the limit λ → 0 \lambda\to0 λ → 0 :
g ( J t ) ≃ − 3 4 ( J t − 1 ) (29) g(\mathcal{J}t)\simeq-\frac{3}{4}(\mathcal{J}t-1) \tag*{(29)} g ( J t ) ≃ − 4 3 ( J t − 1 ) (29)
This shows that the leading order effect at these times is, up to an overall constant, to renormalize the coupling that sets the periodicity, rather than to spoil the periodicity. Denoting this renormalized coupling by J f \mathcal{J}_f J f , and combining (21) and (29) , we get
J f ≃ J ( 1 + 3 8 λ ) = λ ( 1 − 1 4 λ ) ( 1 + 3 8 λ ) = λ ( 1 + λ 8 ) (30) \mathcal{J}_f\simeq\mathcal{J}\left(1+\frac{3}{8}\lambda\right)=\sqrt{\lambda}\left(1-\frac{1}{4}\lambda\right)\left(1+\frac{3}{8}\lambda\right)=\sqrt{\lambda}\left(1+\frac{\lambda}{8}\right) \tag*{(30)} J f ≃ J ( 1 + 8 3 λ ) = λ ( 1 − 4 1 λ ) ( 1 + 8 3 λ ) = λ ( 1 + 8 λ ) (30)
cosh ( J t ) − 2 [ 1 + λ g ( J t ) ] ≃ ( 1 + 3 λ 4 ) cosh ( J f t ) − 2 \cosh(\mathcal{J}t)^{-2}\left[1+\lambda g(\mathcal{J}t)\right]\simeq\left(1+\frac{3\lambda}{4}\right)\cosh(\mathcal{J}_f t)^{-2} cosh ( J t ) − 2 [ 1 + λ g ( J t ) ] ≃ ( 1 + 4 3 λ ) cosh ( J f t ) − 2
and correspondingly β f = π / J f \beta_f=\pi/\mathcal{J}_f β f = π / J f .
We conclude that the fake-thermal analysis of the DSSYK two-point function is consistent for finite but small λ \lambda λ provided we restrict to sufficiently small times t t t such that
λ ∣ g ( J t ) ∣ ≃ 3 4 λ ( λ t − 1 ) ≪ 1 (31) \lambda|g(\mathcal{J}t)|\simeq\frac{3}{4}\lambda(\sqrt{\lambda}t-1)\ll1 \tag*{(31)} λ ∣ g ( J t ) ∣ ≃ 4 3 λ ( λ t − 1 ) ≪ 1 (31)
In the results below we satisfy this constraint out to twice the scrambling time: at t = 2 t ∗ t=2t_* t = 2 t ∗ the left side of (31) is 0.15, 0.064 for q = 0.9 q=0.9 q = 0.9 , 0.975 respectively.
How to fold in DSSYK
Consider the following β = 0 \beta= 0 β = 0 Euclidean OTOC in the DSSYK model:
Tr ( V ( τ 4 ) W ( τ 3 ) V ( τ 2 ) W ( τ 1 ) ) (32) \operatorname{Tr}\left(V(\tau_4)W(\tau_3)V(\tau_2)W(\tau_1)\right) \tag*{(32)} Tr ( V ( τ 4 ) W ( τ 3 ) V ( τ 2 ) W ( τ 1 ) ) (32)
where for simplicity we will assume Δ V = Δ W = 1 \Delta_V = \Delta_W = 1 Δ V = Δ W = 1 .
We can rewrite this OTOC in a form natural for the chord picture where the OTOC is computed from the action of the transfer matrix moving around the fake thermal circle:
Tr ( e − g a H V ( 0 ) e − g b H W ( 0 ) e − g c H V ( 0 ) e − g d H W ( 0 ) ) (33) \operatorname{Tr}\left(e^{-g_aH}V(0)e^{-g_bH}W(0)e^{-g_cH}V(0)e^{-g_dH}W(0)\right) \tag*{(33)} Tr ( e − g a H V ( 0 ) e − g b H W ( 0 ) e − g c H V ( 0 ) e − g d H W ( 0 ) ) (33)
where the “gaps” are given by:
( g a , g b , g c , g d ) = ( τ 14 + δ a , τ 43 + δ b , τ 32 + δ c , τ 21 + δ d ) (34) (g_a,g_b,g_c,g_d) = (\tau_{14}+\delta_a,\tau_{43}+\delta_b,\tau_{32}+\delta_c,\tau_{21}+\delta_d) \tag*{(34)} ( g a , g b , g c , g d ) = ( τ 14 + δ a , τ 43 + δ b , τ 32 + δ c , τ 21 + δ d ) (34)
where τ 14 = τ 1 − τ 4 \tau_{14} = \tau_1-\tau_4 τ 14 = τ 1 − τ 4 etc. and the δ i \delta_i δ i are additional contributions to the gaps in analogy to our fake thermal analysis of the two-point function. The disconnected part of this 4-point function is the product of Tr ( V ( τ 4 ) V ( τ 2 ) ) \operatorname{Tr}(V(\tau_4)V(\tau_2)) Tr ( V ( τ 4 ) V ( τ 2 )) and Tr ( W ( τ 3 ) W ( τ 1 ) ) \operatorname{Tr}(W(\tau_3)W(\tau_1)) Tr ( W ( τ 3 ) W ( τ 1 )) . Consistency with our fake thermal analysis of the V V VV V V two-point function requires the two V V V ’s to be antipodal on the fake circle, δ b + δ c = β f / 2 \delta_b+\delta_c=\beta_f/2 δ b + δ c = β f /2 (equivalently δ d + δ a = β f / 2 \delta_d+\delta_a=\beta_f/2 δ d + δ a = β f /2 ), and likewise for the W W WW W W two-point function, δ c + δ d = β f / 2 \delta_c+\delta_d=\beta_f/2 δ c + δ d = β f /2 (equivalently δ a + δ b = β f / 2 \delta_a+\delta_b=\beta_f/2 δ a + δ b = β f /2 ). These conditions fix δ a = δ c \delta_a=\delta_c δ a = δ c , δ b = δ d \delta_b=\delta_d δ b = δ d and δ a + δ b = β f / 2 \delta_a+\delta_b=\beta_f/2 δ a + δ b = β f /2 , leaving a one-parameter family. The symmetric choice δ a = δ b = δ c = δ d = β f / 4 \delta_a=\delta_b=\delta_c=\delta_d=\beta_f/4 δ a = δ b = δ c = δ d = β f /4 is singled out by the symmetry between V V V and W W W .
Let’s apply this now to the following β = 0 \beta= 0 β = 0 standard connected DSSYK OTOC:
F s t d ( t ) = Tr ( V ( i t ) W ( 0 ) V ( i t ) W ( 0 ) ) (35) F_{\mathrm{std}}(t)=\operatorname{Tr}\left(V(it)W(0)V(it)W(0)\right) \tag*{(35)} F std ( t ) = Tr ( V ( i t ) W ( 0 ) V ( i t ) W ( 0 ) ) (35)
With our gap mapping to the fake thermal circle this becomes:
F s t d ( t ) = Tr ( e − β f H / 4 V ( i t ) e − β f H / 4 W ( 0 ) e − β f H / 4 V ( i t ) e − β f H / 4 W ( 0 ) ) = Tr ( e − β f H V ( 3 β f 4 + i t ) W ( β f 2 ) V ( β f 4 + i t ) W ( 0 ) ) (36) \begin{aligned}
F_{\mathrm{std}}(t)&=\operatorname{Tr}\left(e^{-\beta_fH/4}V(it)e^{-\beta_fH/4}W(0)e^{-\beta_fH/4}V(it)e^{-\beta_fH/4}W(0)\right)\\
&=\operatorname{Tr}\left(e^{-\beta_fH}V\left(\frac{3\beta_f}{4}+it\right)W\left(\frac{\beta_f}{2}\right)V\left(\frac{\beta_f}{4}+it\right)W(0)\right) \tag*{(36)}
\end{aligned} F std ( t ) = Tr ( e − β f H /4 V ( i t ) e − β f H /4 W ( 0 ) e − β f H /4 V ( i t ) e − β f H /4 W ( 0 ) ) = Tr ( e − β f H V ( 4 3 β f + i t ) W ( 2 β f ) V ( 4 β f + i t ) W ( 0 ) ) (36)
where in the second expression we see that F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) looks like a standard thermal OTOC at inverse temperature β f \beta_f β f after applying the regulator prescription of Maldacena, Stanford, and Shenker (MSS) [35 ] .
Now we apply the folding prescription of [1], the n = 0 n=0 n = 0 version of (5) , using the fake temperature instead of the thermal temperature. We call the result the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) . The HZ prescription τ 1 → − τ 1 \tau_1 \to-\tau_1 τ 1 → − τ 1 , τ 2 → − τ 2 \tau_2 \to-\tau_2 τ 2 → − τ 2 , τ 3 → β f − τ 3 \tau_3 \to\beta_f-\tau_3 τ 3 → β f − τ 3 , and τ 4 → β f − τ 4 \tau_4 \to\beta_f-\tau_4 τ 4 → β f − τ 4 , applied to the second expression in (36) , gives:
F f o l d ( t ) = Tr ( e − β f H V ( β f 4 − i t ) W ( β f 2 ) V ( − β f 4 − i t ) W ( 0 ) ) (37) F_{\mathrm{fold}}(t)=\operatorname{Tr}\left(e^{-\beta_f H}V\left(\frac{\beta_f}{4}-it\right)W\left(\frac{\beta_f}{2}\right)V\left(-\frac{\beta_f}{4}-it\right)W(0)\right) \tag*{(37)} F fold ( t ) = Tr ( e − β f H V ( 4 β f − i t ) W ( 2 β f ) V ( − 4 β f − i t ) W ( 0 ) ) (37)
The four Euclidean gaps are now ( 3 β f / 4 + i t , − β f / 4 − i t , 3 β f / 4 + i t , − β f / 4 − i t ) (3\beta_f/4+it,-\beta_f/4-it,3\beta_f/4+it,-\beta_f/4-it) ( 3 β f /4 + i t , − β f /4 − i t , 3 β f /4 + i t , − β f /4 − i t ) : two of them are negative, which is the fold. We see that, in the fake thermal circle picture, the folded OTOC is obtained from the standard OTOC by letting t → − t t \to-t t → − t and moving each V V V backwards by half the fake circle, so that the operator string is no longer ordered in fake Euclidean time; this is illustrated in the left plot of Figure 1 . The equivalent folding with the n = 1 n=1 n = 1 prescription is shown in the right plot of Figure 1 . Since F f o l d ( 1 ) ( t ) = ( F f o l d ( 0 ) ( t ) ) ∗ F_{\mathrm{fold}}^{(1)}(t)=(F_{\mathrm{fold}}^{(0)}(t))^* F fold ( 1 ) ( t ) = ( F fold ( 0 ) ( t ) ) ∗ , we will confine our analysis to the n = 0 n=0 n = 0 case.
Figure 1. Illustration of the n = 0 n=0 n = 0 (left) and n = 1 n=1 n = 1 (right) folds of the fake-MSS configuration on the fake thermal circle. Solid arcs are forward Euclidean gaps, dashed arcs are backward ones; Euclidean time increases clockwise and the real-time displacements of the V V V ’s are suppressed. The n = 0 n=0 n = 0 folding prescription moves each V V V backwards by β f / 2 \beta_f/2 β f /2 along the circle (and by − 2 i t -2it − 2 i t in real time), to fake positions − β f / 4 -\beta_f/4 − β f /4 and β f / 4 \beta_f/4 β f /4 , while the operator ordering in the trace is unchanged. The Euclidean contour is therefore folded: starting from W ( τ 1 ) W(\tau_1) W ( τ 1 ) it runs backwards by β f / 4 \beta_f/4 β f /4 (dashed) to V ( τ 2 ) V(\tau_2) V ( τ 2 ) , forwards by 3 β f / 4 3\beta_f/4 3 β f /4 to W ( τ 3 ) W(\tau_3) W ( τ 3 ) , backwards by β f / 4 \beta_f/4 β f /4 to V ( τ 4 ) V(\tau_4) V ( τ 4 ) , and forwards by 3 β f / 4 3\beta_f/4 3 β f /4 back to W ( τ 1 ) W(\tau_1) W ( τ 1 ) . After the common β f / 4 \beta_f/4 β f /4 offsets are removed (eq. (39) ), the n = 1 n=1 n = 1 gaps are the n = 0 n=0 n = 0 gaps with every sign reversed and t → − t t \to-t t → − t . Reversing each gap is a symmetry of the β = 0 \beta=0 β = 0 chord expansion (non-vanishing chord diagrams contain an even number of H H H ’s), so that F f o l d ( 1 ) ( t ) = F f o l d ( 0 ) ( − t ) = ( F f o l d ( 0 ) ( t ) ) ∗ F_{\mathrm{fold}}^{(1)}(t)=F_{\mathrm{fold}}^{(0)}(-t)=(F_{\mathrm{fold}}^{(0)}(t))^* F fold ( 1 ) ( t ) = F fold ( 0 ) ( − t ) = ( F fold ( 0 ) ( t ) ) ∗ for any Δ V , Δ W \Delta_V,\Delta_W Δ V , Δ W ; the choice of n n n determines the side from which the continued eikonal amplitude approaches its branch cut.
We can rewrite F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) as:
F f o l d ( t ) = Tr ( e − β f H / 4 V ( − β f 2 − i t ) e − β f H / 4 W ( 0 ) e − β f H / 4 V ( − β f 2 − i t ) e − β f H / 4 W ( 0 ) ) (38) F_{\mathrm{fold}}(t)=\operatorname{Tr}\left(e^{-\beta_f H/4}V\left(-\frac{\beta_f}{2}-it\right)e^{-\beta_f H/4}W(0)e^{-\beta_f H/4}V\left(-\frac{\beta_f}{2}-it\right)e^{-\beta_f H/4}W(0)\right) \tag*{(38)} F fold ( t ) = Tr ( e − β f H /4 V ( − 2 β f − i t ) e − β f H /4 W ( 0 ) e − β f H /4 V ( − 2 β f − i t ) e − β f H /4 W ( 0 ) ) (38)
To regain a β = 0 \beta=0 β = 0 DSSYK OTOC we remove all the β f / 4 \beta_f/4 β f /4 gaps to obtain:
F f o l d ( t ) = Tr ( V ( − β f 2 − i t ) W ( 0 ) V ( − β f 2 − i t ) W ( 0 ) ) (39) F_{\mathrm{fold}}(t)=\operatorname{Tr}\left(V\left(-\frac{\beta_f}{2}-it\right)W(0)V\left(-\frac{\beta_f}{2}-it\right)W(0)\right) \tag*{(39)} F fold ( t ) = Tr ( V ( − 2 β f − i t ) W ( 0 ) V ( − 2 β f − i t ) W ( 0 ) ) (39)
As noted by [1], in the folded OTOC the roles of V V V and W W W have switched, since V V V acts before W W W rather than the other way around as was the case for F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) . In the present case this switch is equivalent to t → − t t\to-t t → − t . To see this, note that f ( τ ) ≡ Tr ( V ( τ ) W V ( τ ) W ) f(\tau)\equiv\operatorname{Tr}(V(\tau)WV(\tau)W) f ( τ ) ≡ Tr ( V ( τ ) W V ( τ ) W ) is even in τ \tau τ for any matter content: expanding V ( τ ) = e τ H V e − τ H V(\tau)=e^{\tau H}Ve^{-\tau H} V ( τ ) = e τ H V e − τ H , every term in the disorder average is a chord diagram in which the H H H 's are paired, so only even total powers of τ \tau τ survive. (Equivalently, conjugation by ( − 1 ) n ^ (-1)^{\hat n} ( − 1 ) n ^ sends T → − T T\to-T T → − T while commuting with the matter chord operations.) Hence F f o l d ( t ) = f ( − β f 2 − i t ) = f ( β f 2 + i t ) F_{\mathrm{fold}}(t)=f\left(-\frac{\beta_f}{2}-it\right)=f\left(\frac{\beta_f}{2}+it\right) F fold ( t ) = f ( − 2 β f − i t ) = f ( 2 β f + i t ) , and since f f f has real Taylor coefficients, f ( z ˉ ) = f ( z ) ∗ f(\bar z)=f(z)^* f ( z ˉ ) = f ( z ) ∗ , so that
F f o l d ( − t ) = f ( β f 2 − i t ) = F f o l d ( t ) ∗ (40) F_{\mathrm{fold}}(-t)=f\left(\frac{\beta_f}{2}-it\right)=F_{\mathrm{fold}}(t)^* \tag*{(40)} F fold ( − t ) = f ( 2 β f − i t ) = F fold ( t ) ∗ (40)
i.e. Re F f o l d \operatorname{Re}F_{\mathrm{fold}} Re F fold is even and Im F f o l d \operatorname{Im}F_{\mathrm{fold}} Im F fold is odd in t t t . Below we will always look at F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) for t > 0 t>0 t > 0 .
The n = 1 n=1 n = 1 fold is f ( β f 2 − i t ) f\left(\frac{\beta_f}{2}-it\right) f ( 2 β f − i t ) : its gaps are those of the n = 0 n=0 n = 0 fold with all signs reversed and t → − t t\to-t t → − t . By the same parity it equals F f o l d ( − t ) = F f o l d ( t ) ∗ F_{\mathrm{fold}}(-t)=F_{\mathrm{fold}}(t)^* F fold ( − t ) = F fold ( t ) ∗ .
Two remarks are in order. First, (39) is an ordinary β = 0 \beta=0 β = 0 correlator of the DSSYK model, but one in which the V V V operators have been evolved backwards in Euclidean time by half the fake period. This is well-defined because the DSSYK energy spectrum is bounded, the condition required in [1]. Second, the fold displaces each V V V from the W W W 's by β f / 2 \beta_f/2 β f /2 in Euclidean time, the half-width of the strip ∣ Im t ∣ < β f / 2 |\operatorname{Im}t|<\beta_f/2 ∣ Im t ∣ < β f /2 in which the semiclassical two-point function (25) is analytic. The folded operator V ~ = e − β f H / 2 V e β f H / 2 \widetilde V=e^{-\beta_f H/2}Ve^{\beta_f H/2} V = e − β f H /2 V e β f H /2 is therefore far from bounded in the semiclassical sense: Tr ( V ~ † V ~ ) = G 0 ( β f ) \operatorname{Tr}(\widetilde V^\dagger\widetilde V)=G_0(\beta_f) Tr ( V † V ) = G 0 ( β f ) is exponentially large in 1 / λ 1/\lambda 1/ λ . The OTOC itself is however well-behaved: we will show in the next section that its spectral weight falls off fast enough for the folded integral to converge with room to spare. It is still not obvious a priori that the semiclassical eikonal formula should match the exact correlator at the fold. That it does, at the percent level (Section VII), shows that we satisfy the criterion of [1] that the continuation should commute with the semiclassical limit. The large norm of V ~ \widetilde{V} V makes the numerical evaluation quite delicate: the intermediate states in the chord-space evaluation of (39) have norms growing like e const / λ e^{\operatorname{const}/\lambda} e const / λ even though the result is O ( 1 ) O(1) O ( 1 ) .
Long-time limit of the exact DSSYK OTOC
Berkooz–Isachenkov–Narovlansky–Torrents [20 ] derived an exact expression for the standard OTOC four-point function; using the notation of Mukhametzhanov [2] and introducing four Lorentzian times u 1 , u 2 , u 3 , u 4 u_1,u_2,u_3,u_4 u 1 , u 2 , u 3 , u 4 , we can write an exact β = 0 \beta=0 β = 0 expression for (32) :
Tr ( V ( i u 1 ) W ( i u 2 ) V ( i u 3 ) W ( i u 4 ) ) = ∫ 0 π / λ ∏ i = 1 4 ( d s i ρ q ( s i ) ) e − ( g a E 1 + g b E 2 + g c E 3 + g d E 4 ) ( Γ 12 Γ 23 Γ 34 Γ 41 ) 1 / 2 { Δ s 1 s 2 Δ s 3 s 4 } (41) \operatorname{Tr}\left(V(iu_1)W(iu_2)V(iu_3)W(iu_4)\right)
=
\int_{0}^{\pi/\lambda}\prod_{i=1}^{4}\left(\mathrm{d}s_i\,\rho_q(s_i)\right)e^{-(g_aE_1+g_bE_2+g_cE_3+g_dE_4)}\left(\Gamma_{12}\Gamma_{23}\Gamma_{34}\Gamma_{41}\right)^{1/2}\left\{\begin{matrix}\Delta&s_1&s_2\\\Delta&s_3&s_4\end{matrix}\right\} \tag*{(41)} Tr ( V ( i u 1 ) W ( i u 2 ) V ( i u 3 ) W ( i u 4 ) ) = ∫ 0 π / λ i = 1 ∏ 4 ( d s i ρ q ( s i ) ) e − ( g a E 1 + g b E 2 + g c E 3 + g d E 4 ) ( Γ 12 Γ 23 Γ 34 Γ 41 ) 1/2 { Δ Δ s 1 s 3 s 2 s 4 } (41)
where we have used the gap notation of (33) ,(34) with ( g a , g b , g c , g d ) = ( i u 41 , i u 12 , i u 23 , i u 34 ) (g_a,g_b,g_c,g_d)=(iu_{41},iu_{12},iu_{23},iu_{34}) ( g a , g b , g c , g d ) = ( i u 41 , i u 12 , i u 23 , i u 34 ) , u i j = u i − u j u_{ij}=u_i-u_j u ij = u i − u j . We also have:
ρ q ( s ) = 1 2 π Γ q ( 2 i s ) Γ q ( − 2 i s ) , Γ i j = ( Γ q ( Δ + i ( s i + s j ) ) Γ q ( Δ + i ( s i − s j ) ) Γ q ( Δ − i ( s i − s j ) ) Γ q ( Δ − i ( s i + s j ) ) Γ q ( 2 Δ ) ) (42) \begin{aligned}
\rho_q(s)&=\frac{1}{2\pi\Gamma_q(2is)\Gamma_q(-2is)},\\
\Gamma_{ij}&=\left(\frac{\Gamma_q(\Delta+i(s_i+s_j))\Gamma_q(\Delta+i(s_i-s_j))\Gamma_q(\Delta-i(s_i-s_j))\Gamma_q(\Delta-i(s_i+s_j))}{\Gamma_q(2\Delta)}\right) \tag*{(42)}
\end{aligned} ρ q ( s ) Γ ij = 2 π Γ q ( 2 i s ) Γ q ( − 2 i s ) 1 , = ( Γ q ( 2Δ ) Γ q ( Δ + i ( s i + s j )) Γ q ( Δ + i ( s i − s j )) Γ q ( Δ − i ( s i − s j )) Γ q ( Δ − i ( s i + s j )) ) (42)
where ρ q ( s ) \rho_q(s) ρ q ( s ) is the density of states, the Γ q \Gamma_q Γ q are q-Gamma functions, and the curly bracket object in (41) is a q-6j symbol. The s i s_i s i parameters come from using the DSSYK spectral formula (8) in our units with J = 1 J=1 J = 1 .
To obtain F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) we should use the gaps ( − i t , i t , − i t , i t ) (-it,it,-it,it) ( − i t , i t , − i t , i t ) , corresponding to u 1 = u 3 = t u_1=u_3=t u 1 = u 3 = t , u 2 = u 4 = 0 u_2=u_4=0 u 2 = u 4 = 0 . The folded OTOC (39) corresponds to the gaps ( β f / 2 + i t , − β f / 2 − i t , β f / 2 + i t , − β f / 2 − i t ) (\beta_f/2+it,-\beta_f/2-it,\beta_f/2+it,-\beta_f/2-it) ( β f /2 + i t , − β f /2 − i t , β f /2 + i t , − β f /2 − i t ) , i.e. u 1 = u 3 = − t + i β f / 2 u_1=u_3=-t+i\beta_f/2 u 1 = u 3 = − t + i β f /2 , u 2 = u 4 = 0 u_2=u_4=0 u 2 = u 4 = 0 . Thus F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) is computed from the same integrand as F s t d ( − t ) = F s t d ( t ) F_{\mathrm{std}}(-t)=F_{\mathrm{std}}(t) F std ( − t ) = F std ( t ) , but with an additional factor e − ( β f / 2 ) ( E 1 − E 2 + E 3 − E 4 ) e^{-(\beta_f/2)(E_1-E_2+E_3-E_4)} e − ( β f /2 ) ( E 1 − E 2 + E 3 − E 4 ) .
Following the analysis of Mukhametzhanov [2], we can simplify the pieces of the integrand in the spectral representation (41) by taking the limit λ → 0 \lambda\to0 λ → 0 with λ s \lambda s λ s and s i − s j s_i-s_j s i − s j fixed (his eq. (5.6)), and then the “long time” limit λ → 0 \lambda\to0 λ → 0 with z z z fixed (his eq. (5.23)), i.e. λ L t − log ( 1 / λ ) \lambda_Lt-\log(1/\lambda) λ L t − log ( 1/ λ ) fixed, with z z z defined below. The first limit defines the same regime where we saw that the partition function is dominated by the saddle (14) .
The 4-dimensional integral for the OTOC in (41) can be rewritten as an integral over the average energy s = ( s 1 + s 2 + s 3 + s 4 ) / 4 s=(s_1+s_2+s_3+s_4)/4 s = ( s 1 + s 2 + s 3 + s 4 ) /4 and integrals over energy differences. From Mukhametzhanov’s eq. (5.18) we see that in the unfolded case the integral over s s s is dominated by the same saddle as the partition function.
Now consider a saddle analysis in the β = 0 \beta=0 β = 0 folded case. For simplicity of the discussion let’s enforce s 1 = s 3 s_1=s_3 s 1 = s 3 , s 2 = s 4 s_2=s_4 s 2 = s 4 and look at the behavior of the saddle in the s 1 s_1 s 1 -s 2 s_2 s 2 plane. Denoting the saddle exponent as S ( θ 1 , θ 2 ) / λ S(\theta_1,\theta_2)/\lambda S ( θ 1 , θ 2 ) / λ , where s 1 = θ 1 / λ s_1=\theta_1/\lambda s 1 = θ 1 / λ , s 2 = θ 2 / λ s_2=\theta_2/\lambda s 2 = θ 2 / λ , we see from Mukhametzhanov’s eqs. (5.7), (5.11), and (5.17) that the total entropic contribution to S ( θ 1 , θ 2 ) S(\theta_1,\theta_2) S ( θ 1 , θ 2 ) is 1 2 ( S ( θ 1 ) + S ( θ 2 ) ) \frac{1}{2}(S(\theta_1)+S(\theta_2)) 2 1 ( S ( θ 1 ) + S ( θ 2 )) , with S ( θ ) S(\theta) S ( θ ) as defined in (13) .
To understand the contribution from the alternating folds first note that since the four folds add to zero the symmetric saddle configuration s 1 = s 2 s_1=s_2 s 1 = s 2 is not affected by the folds. Looking then at asymmetric configurations with s 1 ≠ s 2 s_1\ne s_2 s 1 = s 2 , we see that the fold factor e − ( β f / 2 ) ( E 1 − E 2 + E 3 − E 4 ) e^{-(\beta_f/2)(E_1-E_2+E_3-E_4)} e − ( β f /2 ) ( E 1 − E 2 + E 3 − E 4 ) pushes in the direction of E 1 < E 2 E_1<E_2 E 1 < E 2 , contributing ( β f / 2 ) ( 2 λ ) ( cos θ 1 − cos θ 2 + cos θ 3 − cos θ 4 ) = 2 π ( cos θ 1 − cos θ 2 ) (\beta_f/2)(2\sqrt{\lambda})(\cos\theta_1-\cos\theta_2+\cos\theta_3-\cos\theta_4)=2\pi(\cos\theta_1-\cos\theta_2) ( β f /2 ) ( 2 λ ) ( cos θ 1 − cos θ 2 + cos θ 3 − cos θ 4 ) = 2 π ( cos θ 1 − cos θ 2 ) to S ( θ 1 , θ 2 ) S(\theta_1,\theta_2) S ( θ 1 , θ 2 ) , to leading order in small λ \lambda λ .
However asymmetric configurations in energy require that the matter operator insertions are transferring energy from one segment to the other. Denoting the energy transfer by ω \omega ω , we already saw in (26) that the spectral weight in ω \omega ω decays like e − β f ∣ ω ∣ / 2 e^{-\beta_f|\omega|/2} e − β f ∣ ω ∣/2 . Using ω = E ( θ 1 ) − E ( θ 2 ) ≃ ( 2 / λ ) ( θ 1 − θ 2 ) \omega=E(\theta_1)-E(\theta_2)\simeq(2/\sqrt{\lambda})(\theta_1-\theta_2) ω = E ( θ 1 ) − E ( θ 2 ) ≃ ( 2/ λ ) ( θ 1 − θ 2 ) at the energy midpoint, this becomes e − π ∣ θ 1 − θ 2 ∣ / λ e^{-\pi|\theta_1-\theta_2|/\lambda} e − π ∣ θ 1 − θ 2 ∣/ λ to leading order in small λ \lambda λ . Matching to the gamma function asymptotics of Mukhametzhanov’s eqs. (5.7) and (5.17), we find that the total contribution of this type to S ( θ 1 , θ 2 ) S(\theta_1,\theta_2) S ( θ 1 , θ 2 ) is − 3 π ∣ θ 1 − θ 2 ∣ -3\pi|\theta_1-\theta_2| − 3 π ∣ θ 1 − θ 2 ∣ : the four vertex factors give − 2 π ∣ θ 1 − θ 2 ∣ -2\pi|\theta_1-\theta_2| − 2 π ∣ θ 1 − θ 2 ∣ , and the factor Γ ( i ( ν 2 + ν 4 ) ) \Gamma(i(\nu_2+\nu_4)) Γ ( i ( ν 2 + ν 4 )) in the asymptotic form (5.7) of the 6j-symbol gives a further − π ∣ θ 1 − θ 2 ∣ -\pi|\theta_1-\theta_2| − π ∣ θ 1 − θ 2 ∣ on the slice s 1 = s 3 s_1=s_3 s 1 = s 3 , s 2 = s 4 s_2=s_4 s 2 = s 4 . These asymptotic forms are derived at fixed s i − s j s_i-s_j s i − s j , so they hold for λ ≪ ∣ θ 1 − θ 2 ∣ ≪ 1 \lambda\ll|\theta_1-\theta_2|\ll1 λ ≪ ∣ θ 1 − θ 2 ∣ ≪ 1 .
All together, the folded OTOC saddle behavior in the s 1 s_1 s 1 -s 2 s_2 s 2 plane is determined by maximizing
S ( θ 1 , θ 2 ) = ( π θ 1 − θ 1 2 ) + ( π θ 2 − θ 2 2 ) + 2 π ( cos θ 1 − cos θ 2 ) − 3 π ∣ θ 1 − θ 2 ∣ (43) S(\theta_1,\theta_2)=(\pi\theta_1-\theta_1^2)+(\pi\theta_2-\theta_2^2)+2\pi(\cos\theta_1-\cos\theta_2)-3\pi|\theta_1-\theta_2| \tag*{(43)} S ( θ 1 , θ 2 ) = ( π θ 1 − θ 1 2 ) + ( π θ 2 − θ 2 2 ) + 2 π ( cos θ 1 − cos θ 2 ) − 3 π ∣ θ 1 − θ 2 ∣ (43)
The maximum is at θ 1 = θ 2 = π / 2 \theta_1=\theta_2=\pi/2 θ 1 = θ 2 = π /2 , the β = 0 \beta=0 β = 0 saddle point of (14) . It is not a stationary point of the smooth part of (43) , whose gradient there is ( − 2 π , + 2 π ) (-2\pi,+2\pi) ( − 2 π , + 2 π ) , but a cusp. Moving off the diagonal by d = θ 2 − θ 1 > 0 d = \theta_2 - \theta_1 > 0 d = θ 2 − θ 1 > 0 , the fold gain is 2 π d 2\pi d 2 π d while the spectral weight cost is 3 π d 3\pi d 3 π d , so the folded integral converges with a margin. Note that the fold depth could be increased by a factor of 3/2 before the cusp is lost; we return to this observation in § 0.6 B.
In Mukhametzhanov’s long-time limit, eq. (41) , normalized by the product of two-point functions, reduces to (eq. 5.20 of [2]) (see also [41 ] ):
z − 2 Δ U ( 2 Δ , 1 ; 1 / z ) ; z = λ 3 / 2 2 λ L cosh ( λ L 2 ( u 1 − u 2 + u 3 − u 4 ) ) cosh ( λ L 2 ( u 1 − u 3 ) ) cosh ( λ L 2 ( u 2 − u 4 ) ) (44) z^{-2\Delta} U(2\Delta,1;1/z)\ ;\qquad z = \frac{\lambda^{3/2}}{2\lambda_L}\frac{\cosh\left(\frac{\lambda_L}{2}(u_1-u_2+u_3-u_4)\right)}{\cosh\left(\frac{\lambda_L}{2}(u_1-u_3)\right)\cosh\left(\frac{\lambda_L}{2}(u_2-u_4)\right)} \tag*{(44)} z − 2Δ U ( 2Δ , 1 ; 1/ z ) ; z = 2 λ L λ 3/2 cosh ( 2 λ L ( u 1 − u 3 ) ) cosh ( 2 λ L ( u 2 − u 4 ) ) cosh ( 2 λ L ( u 1 − u 2 + u 3 − u 4 ) ) (44)
where U ( 2 Δ , 1 ; 1 / z ) U(2\Delta,1;1/z) U ( 2Δ , 1 ; 1/ z ) is the confluent hypergeometric function, and λ L \lambda_L λ L is the Lyapunov exponent; λ L = 2 λ \lambda_L = 2\sqrt{\lambda} λ L = 2 λ at leading order for small λ \lambda λ . In the scaling limit only the growing exponential in the numerator survives, and Mukhametzhanov writes e λ L 2 ( u 2 + u 4 − u 1 − u 3 ) e^{\frac{\lambda_L}{2}(u_2+u_4-u_1-u_3)} e 2 λ L ( u 2 + u 4 − u 1 − u 3 ) in place of 2 cosh ( … ) 2\cosh(\ldots) 2 cosh ( … ) (in his time-ordering conventions); we have symmetrized it to the cosh, which is equivalent in the scaling limit and makes (44) even in t t t , as the exact β = 0 \beta= 0 β = 0 correlator is.
For Δ = 1 \Delta= 1 Δ = 1 , the integral representation U ( a , b ; x ) = 1 Γ ( a ) ∫ 0 ∞ d t e − x t t a − 1 ( 1 + t ) b − a − 1 U(a,b;x) = \frac{1}{\Gamma(a)}\int_0^\infty dt\,e^{-xt}t^{a-1}(1+t)^{b-a-1} U ( a , b ; x ) = Γ ( a ) 1 ∫ 0 ∞ d t e − x t t a − 1 ( 1 + t ) b − a − 1 gives
U ( 2 , 1 ; 1 z ) = ∫ 0 ∞ d t e − t / z t ( 1 + t ) 2 → t = z p z − 2 U ( 2 , 1 ; 1 z ) = ∫ 0 ∞ d p p e − p ( 1 + z p ) 2 (45) U\left(2,1;\frac{1}{z}\right)=\int_0^\infty dt\,e^{-t/z}\frac{t}{(1+t)^2}\xrightarrow{t=zp} z^{-2}U\left(2,1;\frac{1}{z}\right)=\int_0^\infty dp\,\frac{p e^{-p}}{(1+zp)^2} \tag*{(45)} U ( 2 , 1 ; z 1 ) = ∫ 0 ∞ d t e − t / z ( 1 + t ) 2 t t = z p z − 2 U ( 2 , 1 ; z 1 ) = ∫ 0 ∞ d p ( 1 + z p ) 2 p e − p (45)
and with 1 ( 1 + z p ) 2 = ∫ 0 ∞ d q q e − q e − z p q \frac{1}{(1+zp)^2}=\int_0^\infty dq\,q e^{-q}e^{-zpq} ( 1 + z p ) 2 1 = ∫ 0 ∞ d q q e − q e − z pq ,
F s t d ( t ) → ∫ 0 ∞ d p d q p q e − p − q − z p q = Φ ( z ) , i.e. z ≡ a (46) F_{\mathrm{std}}(t)\to\int_0^\infty dp\,dq\,pq\,e^{-p-q-zpq}=\Phi(z),\qquad\text{i.e. }z\equiv a \tag*{(46)} F std ( t ) → ∫ 0 ∞ d p d q pq e − p − q − z pq = Φ ( z ) , i.e. z ≡ a (46)
This is the eikonal function of (3) , with
a ( t ) = z = λ 3 / 2 2 λ L cosh ( λ L t ) → λ 4 cosh ( 2 λ t ) (47) a(t)=z=\frac{\lambda^{3/2}}{2\lambda_L}\cosh(\lambda_L t)\to\frac{\lambda}{4}\cosh(2\sqrt{\lambda}t) \tag*{(47)} a ( t ) = z = 2 λ L λ 3/2 cosh ( λ L t ) → 4 λ cosh ( 2 λ t ) (47)
The fold as a rotation of z z z
Since z z z depends on the times only through cosh ( λ L 2 ( u 1 − u 2 + u 3 − u 4 ) ) \cosh\left(\frac{\lambda_L}{2}(u_1-u_2+u_3-u_4)\right) cosh ( 2 λ L ( u 1 − u 2 + u 3 − u 4 ) ) (the two cosh’s in the denominator of (44) are unaffected, because u 1 − u 3 = u 2 − u 4 = 0 u_1-u_3=u_2-u_4=0 u 1 − u 3 = u 2 − u 4 = 0 ), the folding process, u 1 = u 3 = − t → − t + i β f / 2 u_1=u_3=-t\to-t+i\beta_f/2 u 1 = u 3 = − t → − t + i β f /2 , takes
z → z cosh ( λ L t e f f + i λ L β f / 2 ) cosh λ L t e f f , t e f f = 1 2 ( u 1 − u 2 + u 3 − u 4 ) . (48) z\to z\frac{\cosh(\lambda_L t_{\mathrm{eff}}+i\lambda_L\beta_f/2)}{\cosh\lambda_L t_{\mathrm{eff}}},\qquad t_{\mathrm{eff}}=\frac{1}{2}(u_1-u_2+u_3-u_4). \tag*{(48)} z → z cosh λ L t eff cosh ( λ L t eff + i λ L β f /2 ) , t eff = 2 1 ( u 1 − u 2 + u 3 − u 4 ) . (48)
To leading order in small λ \lambda λ , λ L β f / 2 = π \lambda_L\beta_f/2=\pi λ L β f /2 = π , so z f o l d = − z z_{\mathrm{fold}}=-z z fold = − z . The side from which the negative axis is approached follows from performing the Euclidean shift continuously: writing u 1 = u_1= u 1 =
u 3 = − t + i φ / 2 u_3 = -t + i\varphi/2 u 3 = − t + i φ /2 with φ \varphi φ increasing from 0 0 0 to β f \beta_f β f , (44) gives z ( φ ) ∝ cosh ( − λ L t + i λ L φ / 2 ) = cosh ( λ L t ) cos ( λ L φ / 2 ) − i sinh ( λ L t ) sin ( λ L φ / 2 ) z(\varphi) \propto\cosh(-\lambda_L t + i\lambda_L\varphi/2) = \cosh(\lambda_L t)\cos(\lambda_L\varphi/2) - i\sinh(\lambda_L t)\sin(\lambda_L\varphi/2) z ( φ ) ∝ cosh ( − λ L t + i λ L φ /2 ) = cosh ( λ L t ) cos ( λ L φ /2 ) − i sinh ( λ L t ) sin ( λ L φ /2 ) , whose imaginary part is negative for t > 0 t > 0 t > 0 and 0 < λ L φ / 2 < π 0 < \lambda_L\varphi/2 < \pi 0 < λ L φ /2 < π . The contour thus approaches the negative axis through the lower half of the z z z -plane:
z f o l d = − z − i 0 ⟹ F f o l d ( t ) → Φ ( − a − i 0 ) ( t > 0 ) , (49) z_{\mathrm{fold}} = -z - i0 \quad\Longrightarrow\quad F_{\mathrm{fold}}(t) \to\Phi(-a - i0) \qquad(t > 0), \tag*{(49)} z fold = − z − i 0 ⟹ F fold ( t ) → Φ ( − a − i 0 ) ( t > 0 ) , (49)
while for t < 0 t < 0 t < 0 the approach is from above, consistent with (40) . With w = 1 / a w = 1/a w = 1/ a and Φ ( a ) = w 2 [ ( 1 + w ) e w E 1 ( w ) − 1 ] \Phi(a) = w^{2}\left[(1+w)e^{w}E_{1}(w)-1\right] Φ ( a ) = w 2 [ ( 1 + w ) e w E 1 ( w ) − 1 ] , the fold sends
1 / a → 1 − a − i 0 = − w + i 0 (50) 1/a \to\frac{1}{-a-i0}=-w+i0 \tag*{(50)} 1/ a → − a − i 0 1 = − w + i 0 (50)
and on the principal branch E 1 ( − w + i 0 ) = − Ei ( w ) − i π E_{1}(-w+i0)=-\operatorname{Ei}(w)-i\pi E 1 ( − w + i 0 ) = − Ei ( w ) − iπ , hence
Φ ( − a − i 0 ) = w 2 [ ( 1 − w ) e − w ( − Ei ( w ) − i π ) − 1 ] ≃ 1 + 4 a + 18 a 2 + ⋯ , Im Φ ( − a − i 0 ) = π w 2 ( w − 1 ) e − w (51) \begin{aligned}
\Phi(-a-i0)&=w^{2}\left[(1-w)e^{-w}\left(-\operatorname{Ei}(w)-i\pi\right)-1\right]\simeq1+4a+18a^{2}+\cdots,\\
\operatorname{Im}\Phi(-a-i0)&=\pi w^{2}(w-1)e^{-w} \tag*{(51)}
\end{aligned} Φ ( − a − i 0 ) Im Φ ( − a − i 0 ) = w 2 [ ( 1 − w ) e − w ( − Ei ( w ) − iπ ) − 1 ] ≃ 1 + 4 a + 18 a 2 + ⋯ , = π w 2 ( w − 1 ) e − w (51)
using Ei ( w ) ≃ e w w ( 1 + 1 w + 2 ! w 2 + 3 ! w 3 + ⋯ ) \operatorname{Ei}(w)\simeq\frac{e^{w}}{w}\left(1+\frac{1}{w}+\frac{2!}{w^{2}}+\frac{3!}{w^{3}}+\cdots\right) Ei ( w ) ≃ w e w ( 1 + w 1 + w 2 2 ! + w 3 3 ! + ⋯ ) .
The thermal fold at high temperatures
We have examined the fold in the thermal state of DSSYK at high temperatures: β \beta β nonzero but ≪ O ( 1 / λ 3 / 2 ) \ll O(1/\lambda^{3/2}) ≪ O ( 1/ λ 3/2 ) , which is where the Schwarzian regime takes over. At finite β \beta β the Lyapunov exponent is λ L ( β ) = 2 π v / β \lambda_{L}(\beta)=2\pi v/\beta λ L ( β ) = 2 π v / β , with v v v the solution of (14) [8], and the rotation of z z z by π \pi π in § 0.6 A requires a Euclidean shift of π / λ L = β / ( 2 v ) \pi/\lambda_{L}=\beta/(2v) π / λ L = β / ( 2 v ) ; this is the fold depth that reaches the a → − a a\to-a a → − a sheet. Equivalently, at finite β \beta β the fake inverse temperature is 2 π / λ L = β / v 2\pi/\lambda_{L}=\beta/v 2 π / λ L = β / v , and the prescription is again to fold by half the fake period. For β → 0 \beta\to0 β → 0 we see from the saddle point equation (14) that v → β / β f v\to\beta/\beta_{f} v → β / β f to leading order in small λ \lambda λ , so the finite β \beta β prescription reduces to folding by β f / 2 \beta_{f}/2 β f /2 in the β → 0 \beta\to0 β → 0 limit.
Explicitly, starting from the MSS-symmetric thermal configuration and shifting each V V V by a fold depth s d s_{d} s d , the finite-β \beta β folded OTOC is
F f o l d ( t ) = 1 N β Tr ( e − g a H V e − g b H W e − g c H V e − g d H W ) , ( g a , g b , g c , g d ) = ( β 4 + s d + i t , β 4 − s d − i t , β 4 + s d + i t , β 4 − s d − i t ) (52) \begin{aligned}
F_{\mathrm{fold}}(t)&=\frac{1}{\mathcal{N}_{\beta}}\operatorname{Tr}\left(e^{-g_{a}H}V e^{-g_{b}H}W e^{-g_{c}H}V e^{-g_{d}H}W\right),\\
(g_{a},g_{b},g_{c},g_{d})&=\left(\frac{\beta}{4}+s_{d}+it,\frac{\beta}{4}-s_{d}-it,\frac{\beta}{4}+s_{d}+it,\frac{\beta}{4}-s_{d}-it\right) \tag*{(52)}
\end{aligned} F fold ( t ) ( g a , g b , g c , g d ) = N β 1 Tr ( e − g a H V e − g b H W e − g c H V e − g d H W ) , = ( 4 β + s d + i t , 4 β − s d − i t , 4 β + s d + i t , 4 β − s d − i t ) (52)
which reduces to (39) for β → 0 \beta\to0 β → 0 with s d = β f / 2 s_{d}=\beta_{f}/2 s d = β f /2 . Here N β = Z ( β ) G V ( β / 2 ) G W ( β / 2 ) \mathcal{N}_{\beta}=Z(\beta)G_{V}(\beta/2)G_{W}(\beta/2) N β = Z ( β ) G V ( β /2 ) G W ( β /2 ) normalizes by the partition function and by the two thermal two-point functions, which do not depend on s d s_{d} s d or t t t because the two V V V ’s, and the two W W W ’s, remain separated by β / 2 \beta/2 β /2 .
For finite β \beta β , we can repeat the saddle point analysis of the folded OTOC in the s 1 s_{1} s 1 -s 2 s_{2} s 2 plane. Instead of (43) we get the following expression:
S ( θ 1 , θ 2 ) = ( π θ 1 − θ 1 2 ) + ( π θ 2 − θ 2 2 ) + λ ( β + 4 β 2 v ) cos θ 1 + λ ( β − 4 β 2 v ) cos θ 2 − 3 π ∣ θ 1 − θ 2 ∣ (53) S(\theta_1,\theta_2)=(\pi\theta_1-\theta_1^2)+(\pi\theta_2-\theta_2^2)+\sqrt{\lambda}\left(\beta+4\frac{\beta}{2v}\right)\cos\theta_1+\sqrt{\lambda}\left(\beta-4\frac{\beta}{2v}\right)\cos\theta_2-3\pi|\theta_1-\theta_2| \tag*{(53)} S ( θ 1 , θ 2 ) = ( π θ 1 − θ 1 2 ) + ( π θ 2 − θ 2 2 ) + λ ( β + 4 2 v β ) cos θ 1 + λ ( β − 4 2 v β ) cos θ 2 − 3 π ∣ θ 1 − θ 2 ∣ (53)
where the Boltzmann terms couple only to the symmetric combination, while the fold couples only to the antisymmetric combination.
For small enough β \beta β , there is a 2-dimensional symmetric saddle θ 1 = θ 2 = ( π / 2 ) ( 1 − v ) \theta_1=\theta_2=(\pi/2)(1-v) θ 1 = θ 2 = ( π /2 ) ( 1 − v ) , with v v v the solution of (14) , corresponding to a saddle point energy below the midpoint. With the fold depth β / ( 2 v ) \beta/(2v) β / ( 2 v ) , the slope of (53) off the diagonal at this saddle is 2 π − 3 π = − π 2\pi-3\pi=-\pi 2 π − 3 π = − π for every β \beta β , by (14) , exactly as at β = 0 \beta=0 β = 0 : the symmetric saddle remains a cusp maximum and never becomes locally unstable. Any failure has to come from elsewhere in the θ 1 \theta_1 θ 1 -θ 2 \theta_2 θ 2 plane. Close to the midpoint we can write θ 2 = π / 2 + ϵ \theta_2=\pi/2+\epsilon θ 2 = π /2 + ϵ , ϵ \epsilon ϵ small, and observe that, for θ 2 > θ 1 \theta_2>\theta_1 θ 2 > θ 1 :
∂ S ( θ 1 , θ 2 ) ∂ θ 2 ≃ − 2 ϵ + 4 λ b − 3 π (54) \frac{\partial S(\theta_1,\theta_2)}{\partial\theta_2}\simeq-2\epsilon+4\sqrt{\lambda}\,b-3\pi \tag*{(54)} ∂ θ 2 ∂ S ( θ 1 , θ 2 ) ≃ − 2 ϵ + 4 λ b − 3 π (54)
where we have defined b = β / ( 2 v ) − β / 4 b=\beta/(2v)-\beta/4 b = β / ( 2 v ) − β /4 as the effective backwards fold (more generally b = s d − β / 4 b=s_d-\beta/4 b = s d − β /4 for a fold depth s d s_d s d ). For b > 3 π / ( 4 λ ) b>3\pi/(4\sqrt{\lambda}) b > 3 π / ( 4 λ ) , corresponding to backward segments 1.5 times larger than the β = 0 \beta=0 β = 0 fold β f / 2 \beta_f/2 β f /2 , (53) therefore has a second, real stationary point in θ 2 \theta_2 θ 2 near the midpoint, away from the symmetric saddle; for smaller b b b the corresponding stationary points form a complex conjugate pair. Since this uses the form − 3 π ∣ θ 1 − θ 2 ∣ -3\pi|\theta_1-\theta_2| − 3 π ∣ θ 1 − θ 2 ∣ at ∣ θ 1 − θ 2 ∣ = O ( 1 ) |\theta_1-\theta_2|=O(1) ∣ θ 1 − θ 2 ∣ = O ( 1 ) , outside the regime in which it was derived, we take it only as an estimate of the threshold. We will confirm this threshold with our exact numerical results in the next section, and show that it corresponds to a pair of complex conjugate saddles emerging as dominant over the original symmetric saddle.
The scrambling time
To define the scrambling time t ∗ t_* t ∗ , we use the criterion of [1] that it is the time when the leading connected part of the OTOC equals the disconnected part. Using the expansion of Φ ( − a ( t ) ) \Phi(-a(t)) Φ ( − a ( t )) in (51) this implies a ( t ∗ ) = 1 / 4 a(t_*)=1/4 a ( t ∗ ) = 1/4 , and thus
t ∗ ≃ 1 2 λ log 2 λ (55) t_*\simeq\frac{1}{2\sqrt{\lambda}}\log\frac{2}{\lambda} \tag*{(55)} t ∗ ≃ 2 λ 1 log λ 2 (55)
To leading order in small λ \lambda λ this has the Harlow-Zhao form t ∗ = ( β f / 2 π ) log S d S t_* = (\beta_f/2\pi)\log S_{\mathrm{dS}} t ∗ = ( β f /2 π ) log S dS (their eq. (2.8)), with S d S = 2 / λ = N / p 2 S_{\mathrm{dS}}=2/\lambda= N/p^2 S dS = 2/ λ = N / p 2 playing the role of the de Sitter entropy, and β f \beta_f β f the inverse temperature. Although t ∗ t_* t ∗ is defined from the leading term of the asymptotic series, at a = 1 / 4 a = 1/4 a = 1/4 that series is useless (its first five terms sum to 7.0, versus Φ ( − 1 / 4 − i 0 ) = 1.26 + 2.76 i \Phi(-1/4-i0) = 1.26 + 2.76i Φ ( − 1/4 − i 0 ) = 1.26 + 2.76 i ); all comparisons below use the closed form (51) . It is also worth noting that ∣ F f o l d ( t ) ∣ |F_{\mathrm{fold}}(t)| ∣ F fold ( t ) ∣ peaks very close (within about a percent) to the scrambling time as defined by (55) , so this provides an alternate definition of the scrambling time self-contained within our analysis. With this estimate the reader can check that the numerical examples in the next section easily satisfy the consistency condition (31) out to twice the scrambling time.
Numerical analysis of the folded OTOC
We compute F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) and F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) numerically in the exact DSSYK model. We evaluate (35) and (39) directly in the chord Hilbert space with two matter chords, acting with the transfer matrix and the matter operators in sequence on the empty state; the states are labeled by the numbers of H H H -chords in the regions separated by the open matter chords, ∣ n 0 , n 1 , n 2 ⟩ \lvert n_0,n_1,n_2\rangle ∣ n 0 , n 1 , n 2 ⟩ , with n 0 + n 1 + n 2 ≤ n max n_0+n_1+n_2\leq n_{\max} n 0 + n 1 + n 2 ≤ n m a x , and an H H H -chord closing across a matter chord contributes q Δ q^{\Delta} q Δ while the crossing of the two matter chords contributes q Δ V Δ W q^{\Delta_V\Delta_W} q Δ V Δ W . Because of the backward Euclidean segments in (39) the intermediate vectors are exponentially large compared to the final answer (their norms grow like e O ( 1 / λ ) e^{O(1/\lambda)} e O ( 1/ λ ) ), so double precision arithmetic is insufficient; we used multiple-precision arithmetic with dps decimal digits. We increased n max n_{\max} n m a x and dps to achieve both convergence and the desired precision of the extracted results. The largest values used for the results shown here were n max = d p s = 440 n_{\max}=\mathrm{dps}=440 n m a x = dps = 440 ; the required values scale approximately like 1 / λ 1/\lambda 1/ λ . We have checked that, once we take n max n_{\max} n m a x large enough for the chord sums to converge, the numerics are extremely stable, and truncating the chord sum at that point has no effect within the precision of our analysis; this justifies calling our numerics “exact DSSYK results”.
Typical results are shown in Figure 2 for q = 0.975 q=0.975 q = 0.975 . The standard OTOC F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) has its maximum value at t = 0 t=0 t = 0 , and decreases monotonically thereafter. The real part of the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) has its minimum value at t = 0 t=0 t = 0 , and rises monotonically before peaking shortly before the scrambling time (the modulus ∣ F f o l d ( t ) ∣ \lvert F_{\mathrm{fold}}(t)\rvert ∣ F fold ( t )∣ peaks at t ∗ t_* t ∗ itself, cf. the remarks below (55) ); this is the anti-scrambling behavior.
Figure 2. Plots of the standard OTOC F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) and the real and imaginary parts of the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) in the exact DSSYK model, for q = 0.975 q = 0.975 q = 0.975 , implying λ = 0.0253 \lambda= 0.0253 λ = 0.0253 , β f ≃ π / λ = 19.7 \beta_f \simeq\pi/\sqrt{\lambda} = 19.7 β f ≃ π / λ = 19.7 , and scrambling time t ∗ = 13.74 t_* = 13.74 t ∗ = 13.74 . The lower figure shows the relative deviation of the complex amplitudes, δ = ∣ Φ ( − a ( t ) ) − F f o l d ( t ) ∣ / ∣ F f o l d ( t ) − 1 ∣ \delta= \lvert\Phi(-a(t)) - F_{\mathrm{fold}}(t)\rvert/\lvert F_{\mathrm{fold}}(t) - 1\rvert δ = ∣ Φ ( − a ( t )) − F fold ( t )∣ / ∣ F fold ( t ) − 1 ∣ (the δ \delta δ column of Table I ), where a ( t ) a(t) a ( t ) was computed by matching Φ ( a ( t ) ) \Phi(a(t)) Φ ( a ( t )) to F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) .
While the standard OTOC F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) is strictly real, the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) has a rapidly growing positive imaginary part near the scrambling time; this is the exponentially small term π w 2 ( w − 1 ) e − w \pi w^2(w-1)e^{-w} π w 2 ( w − 1 ) e − w of (51) , which becomes O ( 1 ) O(1) O ( 1 ) at a ∼ 1 / 4 a\sim1/4 a ∼ 1/4 ; it is consistent with the expected behavior in the mild backreaction regime where the de Sitter eikonal shockwave analysis is also valid.
The folded OTOC also has a very small negative imaginary part at early times; this has no counterpart in the eikonal formula, and is a combination of two effects that we will now describe.
The first effect is that the numerical results for F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) are sensitive to whether we perform the folding using the exact value of β f = π / J f \beta_f = \pi/J_f β f = π / J f , or an approximation to it. In the language of § 0.6 A, the results are sensitive to any over-rotation of z z z beyond π \pi π . With this criterion we can define an “exact” β f \beta_f β f by requiring F f o l d ( 0 ) F_{\mathrm{fold}}(0) F fold ( 0 ) to be stationary with respect to β f \beta_f β f : ∂ β f F f o l d ( t ) ∣ t = 0 = 0 \partial_{\beta_f}F_{\mathrm{fold}}(t)|_{t=0}=0 ∂ β f F fold ( t ) ∣ t = 0 = 0 . To see what this means, let f ( τ ) ≡ Tr V ( τ ) W V ( τ ) W f(\tau)\equiv\operatorname{Tr} V(\tau)WV(\tau)W f ( τ ) ≡ Tr V ( τ ) W V ( τ ) W , which is real for real τ \tau τ and even in τ \tau τ (§ 0.5 ), so that (39) reads
F f o l d ( t ) = f ( β f 2 + i t ) , (56) F_{\mathrm{fold}}(t)=f\left(\frac{\beta_f}{2}+it\right), \tag*{(56)} F fold ( t ) = f ( 2 β f + i t ) , (56)
since β f \beta_f β f and t t t enter only through the combination β f / 2 + i t \beta_f/2+it β f /2 + i t , and
∂ Im F f o l d ∂ β f = − 1 2 d Re F f o l d d t . (57) \frac{\partial\operatorname{Im}F_{\mathrm{fold}}}{\partial\beta_f}=-\frac{1}{2}\frac{d\operatorname{Re}F_{\mathrm{fold}}}{dt}. \tag*{(57)} ∂ β f ∂ Im F fold = − 2 1 d t d Re F fold . (57)
Expanding about t = 0 t=0 t = 0 , Im F f o l d ( t ) = f ′ ( β f / 2 ) t − f ′ ′ ′ ( β f / 2 ) t 3 / 6 + … \operatorname{Im}F_{\mathrm{fold}}(t)=f'(\beta_f/2)t-f'''(\beta_f/2)t^3/6+\ldots Im F fold ( t ) = f ′ ( β f /2 ) t − f ′′′ ( β f /2 ) t 3 /6 + … , so the condition we used to define the exact β f \beta_f β f , ∂ β f F f o l d ∣ t = 0 = 1 2 f ′ ( β f / 2 ) = 0 \partial_{\beta_f}F_{\mathrm{fold}}|_{t=0}=\frac{1}{2}f'(\beta_f/2)=0 ∂ β f F fold ∣ t = 0 = 2 1 f ′ ( β f /2 ) = 0 , is precisely the condition that Im F f o l d \operatorname{Im}F_{\mathrm{fold}} Im F fold has no term linear in t t t .
The same condition fixes the sign of the early growth. Since
Re F f o l d ( t ) = f ( β f / 2 ) − 1 2 f ′ ′ ( β f / 2 ) t 2 + O ( t 4 ) , (58) \operatorname{Re}F_{\mathrm{fold}}(t)=f(\beta_f/2)-\frac{1}{2}f''(\beta_f/2)t^2+O(t^4), \tag*{(58)} Re F fold ( t ) = f ( β f /2 ) − 2 1 f ′′ ( β f /2 ) t 2 + O ( t 4 ) , (58)
and observing that semiclassically f ( τ ) = Φ ( a ∞ cos λ L τ ) f(\tau)=\Phi(a_{\infty}\cos\lambda_{L}\tau) f ( τ ) = Φ ( a ∞ cos λ L τ ) and τ = β f / 2 \tau=\beta_f/2 τ = β f /2 is the maximum of f f f on the real axis, we see that f ′ ′ ( β f / 2 ) < 0 f''(\beta_f/2)<0 f ′′ ( β f /2 ) < 0 and so Re F f o l d \operatorname{Re}F_{\mathrm{fold}} Re F fold automatically rises from its value at t = 0 t=0 t = 0 . The initial rise of the folded correlator thus follows from the definition of the fold depth as a local maximum of f f f on the real axis; it is kinematic, and would hold in any model in which such a maximum exists. The dynamical content of anti-scrambling is not this initial rise but rather the details of the subsequent behavior: the exponent d log a / d t → 2 π / β f d\log a/dt\to2\pi/\beta_f d log a / d t → 2 π / β f , the shape Φ ( − a ( t ) − i 0 ) \Phi(-a(t)-i0) Φ ( − a ( t ) − i 0 ) with the same a ( t ) a(t) a ( t ) as F s t d F_{\mathrm{std}} F std , and the sign of the O ( 1 ) O(1) O ( 1 ) imaginary part, all quantities that are compared below.
An error δ \delta δ in β f \beta_f β f restores a linear term in Im F f o l d \operatorname{Im}F_{\mathrm{fold}} Im F fold ; from (57) and Re F f o l d ≃ 1 + 4 a + 18 a 2 \operatorname{Re}F_{\mathrm{fold}}\simeq1+4a+18a^2 Re F fold ≃ 1 + 4 a + 18 a 2 with a = λ 4 cosh ( λ L t ) a=\frac{\lambda}{4}\cosh(\lambda_Lt) a = 4 λ cosh ( λ L t ) ,
δ Im F f o l d ≃ − λ 3 / 2 δ sinh ( λ L t ) ( 1 + 9 a ) , (59) \delta\operatorname{Im}F_{\mathrm{fold}}\simeq-\lambda^{3/2}\delta\sinh(\lambda_Lt)(1+9a), \tag*{(59)} δ Im F fold ≃ − λ 3/2 δ sinh ( λ L t ) ( 1 + 9 a ) , (59)
which, from the analysis in § 0.6 A, is what we expect from an over-rotation of z z z by λ L δ / 2 \lambda_L\delta/2 λ L δ /2 beyond π \pi π , carried to second order in the asymptotic series for Φ \Phi Φ .
It is useful to compare this exact β f \beta_f β f to (30) , where we already have an expression for β f \beta_f β f accurate at next-to-leading order (NLO) in small λ \lambda λ :
β f = π λ ( 1 + λ / 8 + O ( λ 2 ) ) (60) \beta_f=\frac{\pi}{\sqrt{\lambda}(1+\lambda/8+O(\lambda^2))} \tag*{(60)} β f = λ ( 1 + λ /8 + O ( λ 2 )) π (60)
This agrees with the exact β f \beta_f β f computed numerically as an expansion in λ \lambda λ :
β f = π λ ( 1 + λ / 8 + c 2 λ 2 + c 3 λ 3 + O ( λ 4 ) ) (61) \beta_f=\frac{\pi}{\sqrt{\lambda}(1+\lambda/8+c_2\lambda^2+c_3\lambda^3+O(\lambda^4))} \tag*{(61)} β f = λ ( 1 + λ /8 + c 2 λ 2 + c 3 λ 3 + O ( λ 4 )) π (61)
where we find:
c 2 = − 0.09635 ± 0.00002 ; c 3 = 0.075 ± 0.002 (62) c_2=-0.09635\pm0.00002\ ;\qquad c_3=0.075\pm0.002 \tag*{(62)} c 2 = − 0.09635 ± 0.00002 ; c 3 = 0.075 ± 0.002 (62)
For the plots and tables shown here, we have defined the fold using the NNLO correction c 2 λ 2 c_2\lambda^2 c 2 λ 2 for β f \beta_f β f but not including c 3 λ 3 c_3\lambda^3 c 3 λ 3 ; this is very close to the exact fold, and its small residual effect is discussed below.
Equation (57) also quantifies how much the comparison with the eikonal amplitude depends on the choice of β f \beta_f β f : a shift δ \delta δ of β f \beta_f β f changes the full correlator by δ F f o l d = − i 2 δ ∂ t F f o l d \delta F_{\mathrm{fold}}=-\frac{i}{2}\delta\,\partial_tF_{\mathrm{fold}} δ F fold = − 2 i δ ∂ t F fold . For example at the scrambling time for q = 0.975 q=0.975 q = 0.975 , changing β f \beta_f β f by its NLO term (δ ≃ 0.06 \delta\simeq0.06 δ ≃ 0.06 , i.e. 0.3 % 0.3\% 0.3% ) shifts ∣ F f o l d ∣ \lvert F_{\mathrm{fold}}\rvert ∣ F fold ∣ by about 1.5 % 1.5\% 1.5% . The agreement with Φ ( − a − i 0 ) \Phi(-a-i0) Φ ( − a − i 0 ) at the scrambling time is therefore uncertain at the percent level from the definition of the fake period alone; the stationary β f \beta_f β f is the choice that is defined without reference to the eikonal formula.
The second effect arises because, even at the exact β f \beta_f β f , Im F f o l d \operatorname{Im} F_{\mathrm{fold}} Im F fold has a small negative imaginary part at early times. As we already observed in Section IV, at finite λ \lambda λ the exact β = 0 \beta= 0 β = 0 DSSYK correlators are only approximately periodic in Euclidean time with period β f \beta_f β f (cf. the two-point function in (25) ), so f ( τ ) f(\tau) f ( τ ) is not exactly symmetric about τ = β f / 2 \tau= \beta_f/2 τ = β f /2 and f ′ ′ ′ ( β f / 2 ) ≠ 0 f'''(\beta_f/2) \ne0 f ′′′ ( β f /2 ) = 0 . The residual Im F f o l d − Im Φ ( − a − i 0 ) \operatorname{Im} F_{\mathrm{fold}} - \operatorname{Im} \Phi(-a-i0) Im F fold − Im Φ ( − a − i 0 ) therefore starts at order t 3 t^3 t 3 .
Numerically we have modeled this early time non-eikonal behavior for q = 0.9 q = 0.9 q = 0.9 –0.975 0.975 0.975 , using the exact values of β f \beta_f β f . We find:
Im F f o l d ( t ) − Im Φ ( − a − i 0 ) ≃ − λ 3 8 sinh ( λ L t ) ( cosh ( λ L t ) − 1 ) ≃ − λ 9 / 2 2 t 3 ( λ L t ≪ 1 ) , (63) \operatorname{Im} F_{\mathrm{fold}}(t) - \operatorname{Im} \Phi(-a-i0) \simeq-\frac{\lambda^3}{8}\sinh(\lambda_L t)\left(\cosh(\lambda_L t)-1\right) \simeq-\frac{\lambda^{9/2}}{2}t^3 \qquad(\lambda_L t \ll1), \tag*{(63)} Im F fold ( t ) − Im Φ ( − a − i 0 ) ≃ − 8 λ 3 sinh ( λ L t ) ( cosh ( λ L t ) − 1 ) ≃ − 2 λ 9/2 t 3 ( λ L t ≪ 1 ) , (63)
i.e. the sinh cosh structure of the second-order term, with its sinh part fixed by the absence of a linear term. The t 3 t^3 t 3 coefficient divided by − λ 9 / 2 / 2 -\lambda^{9/2}/2 − λ 9/2 /2 is 1.33, 1.16, 1.08, 1.05 for q = 0.9 q = 0.9 q = 0.9 , 0.925, 0.95, 0.975, approaching one as λ → 0 \lambda\to0 λ → 0 ; see the right plot of Figure 3 . The functional form in (63) holds to within ∼ 5 % \sim5\% ∼ 5% for λ L t ≲ 1.5 \lambda_L t \lesssim1.5 λ L t ≲ 1.5 (t ≲ 0.41 t ∗ t \lesssim0.41\,t_* t ≲ 0.41 t ∗ ) at q = 0.95 q = 0.95 q = 0.95 . The dip in the left plot of Figure 3 is this term, cut off by the onset of the physical Im Φ ( − a − i 0 ) \operatorname{Im} \Phi(-a-i0) Im Φ ( − a − i 0 ) ; the NNLO β f \beta_f β f used in the plots adds a small linear piece (59) that deepens the dip by about 20% at q = 0.925 q = 0.925 q = 0.925 . Like the rest of the finite-λ \lambda λ corrections, it vanishes as λ → 0 \lambda\to0 λ → 0 .
Figure 3. Plot of the imaginary part of the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) in the exact DSSYK model, for q = 0.9 q = 0.9 q = 0.9 , 0.925 0.925 0.925 , 0.95 0.95 0.95 , and 0.975 0.975 0.975 , implying λ = 0.1054 \lambda= 0.1054 λ = 0.1054 , 0.078 0.078 0.078 , 0.0513 0.0513 0.0513 , and 0.0253 0.0253 0.0253 . The left plot shows the very small negative imaginary part that manifests at early times from folding using β f \beta_f β f computed to NNLO in λ \lambda λ ; this transitions to a much larger positive imaginary part closer to the scrambling time. The right plot shows a fit using the exact β f \beta_f β f to the non-eikonal early time behavior of Im F f o l d ( t ) \operatorname{Im} F_{\mathrm{fold}}(t) Im F fold ( t ) . The fit is consistent with Im F f o l d ( t ) − Im Φ ( − a − i 0 ) ≃ − κ 3 t 3 \operatorname{Im} F_{\mathrm{fold}}(t) - \operatorname{Im} \Phi(-a-i0) \simeq-\kappa_3 t^3 Im F fold ( t ) − Im Φ ( − a − i 0 ) ≃ − κ 3 t 3 , with κ 3 = λ 9 / 2 / 2 \kappa_3 = \lambda^{9/2}/2 κ 3 = λ 9/2 /2 .
To see that the anti-scrambling of the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) is consistent with de Sitter shock wave behavior, we have performed the following exercise:
Assume F s t d ( t ) = Φ ( a ( t ) ) F_{\mathrm{std}}(t) = \Phi(a(t)) F std ( t ) = Φ ( a ( t )) and use the numerical results to determine the monotonically increasing function a ( t ) a(t) a ( t ) .
Use equation (3) to compute Φ ( − a ( t ) ) \Phi(-a(t)) Φ ( − a ( t ))
Compare Re [ Φ ( − a ( t ) ) ] − 1 \operatorname{Re}[\Phi(-a(t))]-1 Re [ Φ ( − a ( t ))] − 1 to Re [ F f o l d ( t ) ] − 1 \operatorname{Re}[F_{\mathrm{fold}}(t)]-1 Re [ F fold ( t )] − 1 , and compare Im [ Φ ( − a ( t ) ) ] \operatorname{Im}[\Phi(-a(t))] Im [ Φ ( − a ( t ))] to Im [ F f o l d ( t ) ] \operatorname{Im}[F_{\mathrm{fold}}(t)] Im [ F fold ( t )] .
This exercise tests the pure a → − a a \to-a a → − a content of the fold, independently of how accurately (47) describes a ( t ) a(t) a ( t ) at finite λ \lambda λ .
The results are shown in Table I , Table II and Table III for three values of λ \lambda λ . The real and imaginary part values agree well, and, as seen in the tables and in Figure 4 , the agreement improves as λ \lambda λ decreases, as expected. For the smallest λ \lambda λ value shown, the agreement at the scrambling time of the magnitudes and of the imaginary parts is at the percent level; the fractional agreement of the real parts is worse since the real parts are decreasing rapidly towards zero at the scrambling time. The good agreement of the imaginary parts confirms the sheet selection of § 0.6 A.
Figure 4. Left: λ \lambda λ dependence of the real part of the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) and the standard OTOC F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) , in the exact DSSYK model, for q = 0.9 q = 0.9 q = 0.9 , 0.925 0.925 0.925 , 0.95 0.95 0.95 , and 0.975 0.975 0.975 , implying λ = 0.1054 \lambda= 0.1054 λ = 0.1054 , 0.078 0.078 0.078 , 0.0513 0.0513 0.0513 , and 0.0253 0.0253 0.0253 . Right: dependence on λ \lambda λ of the magnitude of the fractional difference between the eikonal result Φ ( − a ( t ) ) − 1 \Phi(-a(t))-1 Φ ( − a ( t )) − 1 and F f o l d ( t ) − 1 F_{\mathrm{fold}}(t)-1 F fold ( t ) − 1 , where a ( t ) a(t) a ( t ) was computed by matching Φ ( a ( t ) ) \Phi(a(t)) Φ ( a ( t )) to F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) .
Table I. TABLE I. Fold-amplitude data for q = 0.975 q = 0.975 q = 0.975 (λ = 0.0253 \lambda= 0.0253 λ = 0.0253 , 2 λ = 0.318 2\sqrt{\lambda} = 0.318 2 λ = 0.318 , β f = 19.683 \beta_f = 19.683 β f = 19.683 , t ∗ = 13.74 t_* = 13.74 t ∗ = 13.74 ). δ = ∣ Φ ( − a ) − F f o l d ∣ / ∣ F f o l d − 1 ∣ \delta= |\Phi(-a) - F_{\mathrm{fold}}|/|F_{\mathrm{fold}} - 1| δ = ∣Φ ( − a ) − F fold ∣/∣ F fold − 1∣ is the relative deviation of the complex amplitudes.
Table II. Fold-amplitude data for q = 0.950 q = 0.950 q = 0.950 (λ = 0.0513 \lambda= 0.0513 λ = 0.0513 , 2 λ = 0.453 2\sqrt{\lambda} = 0.453 2 λ = 0.453 , β f = 13.786 \beta_f = 13.786 β f = 13.786 , t ∗ = 8.09 t_* = 8.09 t ∗ = 8.09 ). δ = ∣ Φ ( − a ) − F f o l d ∣ / ∣ F f o l d − 1 ∣ \delta= \lvert\Phi(-a) - F_{\mathrm{fold}}\rvert/\lvert F_{\mathrm{fold}} - 1\rvert δ = ∣ Φ ( − a ) − F fold ∣ / ∣ F fold − 1 ∣ is the relative deviation of the complex amplitudes.
Table III. Fold-amplitude data for q = 0.900 q = 0.900 q = 0.900 (λ = 0.1054 \lambda= 0.1054 λ = 0.1054 , 2 λ = 0.649 2\sqrt{\lambda} = 0.649 2 λ = 0.649 , β f = 9.563 \beta_f = 9.563 β f = 9.563 , t ∗ = 4.53 t_* = 4.53 t ∗ = 4.53 ). δ = ∣ Φ ( − a ) − F f o l d ∣ / ∣ F f o l d − 1 ∣ \delta= \lvert\Phi(-a)-F_{\mathrm{fold}}\rvert/\lvert F_{\mathrm{fold}}-1\rvert δ = ∣ Φ ( − a ) − F fold ∣ / ∣ F fold − 1 ∣ is the relative deviation of the complex amplitudes.
We also see in the tables that the extracted a ( t ) a(t) a ( t ) increases exponentially such that, as we approach the scrambling time, d log ( a ) / d t d\log(a)/dt d log ( a ) / d t approaches the fake temperature Lyapunov exponent chaos bound 2 π / β f ≃ 2 λ 2\pi/\beta_f \simeq2\sqrt{\lambda} 2 π / β f ≃ 2 λ ; the fold does not change the exponent, in agreement with the results found in [1] for de Sitter observer correlators.
Short time behavior
As t → 0 t \to0 t → 0 , F s t d ( t ) → q F_{\mathrm{std}}(t) \to q F std ( t ) → q ; this is exact, because the only remaining connected chord diagram is the two matter chords crossing with weight q Δ V Δ W = q q^{\Delta V\Delta W}=q q Δ V Δ W = q . Numerically we also find F f o l d ( 0 ) ≃ q − 1 F_{\mathrm{fold}}(0) \simeq q^{-1} F fold ( 0 ) ≃ q − 1 ; this is not exact, but holds through O ( λ 2 ) O(\lambda^2) O ( λ 2 ) , as the following matching shows.
Rewriting the long-time expression (47) for a ( t ) a(t) a ( t ) as a ( t ) = a ∞ cosh ( 2 λ t ) a(t)=a_{\infty}\cosh(2\sqrt{\lambda}t) a ( t ) = a ∞ cosh ( 2 λ t ) with a ∞ = λ / 4 a_{\infty}=\lambda/4 a ∞ = λ /4 , we can match F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) at short times to the expansion 1 − 4 a + 18 a 2 + … 1-4a+18a^2+\ldots 1 − 4 a + 18 a 2 + … , and match F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) at short times to the expansion 1 + 4 a + 18 a 2 + … 1+4a+18a^2+\ldots 1 + 4 a + 18 a 2 + … , to get
F s t d ( t ) ≃ Φ ( a s t d cosh ( 2 λ t ) ) ; a s t d = a ∞ ( 1 + 5 8 λ ) F f o l d ( t ) ≃ Φ ( − a f o l d cosh ( 2 λ t ) ) ; a f o l d = a ∞ ( 1 − 5 8 λ ) (64) \begin{aligned}
F_{\mathrm{std}}(t) &\simeq\Phi\left(a_{\mathrm{std}}\cosh(2\sqrt{\lambda}t)\right) \ ; \qquad a_{\mathrm{std}}=a_{\infty}\left(1+\frac{5}{8}\lambda\right) \\
F_{\mathrm{fold}}(t) &\simeq\Phi\left(-a_{\mathrm{fold}}\cosh(2\sqrt{\lambda}t)\right) \ ; \qquad a_{\mathrm{fold}}=a_{\infty}\left(1-\frac{5}{8}\lambda\right) \tag*{(64)}
\end{aligned} F std ( t ) F fold ( t ) ≃ Φ ( a std cosh ( 2 λ t ) ) ; a std = a ∞ ( 1 + 8 5 λ ) ≃ Φ ( − a fold cosh ( 2 λ t ) ) ; a fold = a ∞ ( 1 − 8 5 λ ) (64)
where indeed Φ ( a s t d ) = 1 − λ + λ 2 / 2 + O ( λ 3 ) = q + O ( λ 3 ) \Phi(a_{\mathrm{std}})=1-\lambda+\lambda^2/2+O(\lambda^3)=q+O(\lambda^3) Φ ( a std ) = 1 − λ + λ 2 /2 + O ( λ 3 ) = q + O ( λ 3 ) and Φ ( − a f o l d ) = 1 + λ + λ 2 / 2 + O ( λ 3 ) = q − 1 + O ( λ 3 ) \Phi(-a_{\mathrm{fold}})=1+\lambda+\lambda^2/2+O(\lambda^3)=q^{-1}+O(\lambda^3) Φ ( − a fold ) = 1 + λ + λ 2 /2 + O ( λ 3 ) = q − 1 + O ( λ 3 ) . Thus at short times it is not true that F s t d F_{\mathrm{std}} F std and F f o l d F_{\mathrm{fold}} F fold are simply related by a ( t ) → − a ( t ) a(t)\to-a(t) a ( t ) → − a ( t ) , but the discrepancy is at next-to-leading order in λ \lambda λ , and it shrinks towards the scrambling regime as seen in Table I –Table III .
Anti-Stokes behavior of the thermal fold at high temperatures
In § 0.6 B we saw from the integral defining the DSSYK OTOC that it should be possible to construct a folded OTOC with approximately de Sitter-like eikonal behavior for some range of finite β \beta β , by using a fold of approximately β / ( 2 v ) \beta/(2v) β / ( 2 v ) instead of β f / 2 \beta_f/2 β f /2 . However, we also saw from the saddle analysis that we expect a problem for folds exceeding roughly 1.5 β f / 2 1.5\,\beta_f/2 1.5 β f /2 . Here we will use numerical results to quantify this and identify the source of the breakdown for too-large folds.
The left plot in Figure 5 shows ∣ F f o l d ( t = 0 ) ∣ |F_{\mathrm{fold}}(t=0)| ∣ F fold ( t = 0 ) ∣ as a function of the effective backwards fold b = s d − β / 4 b=s_d-\beta/4 b = s d − β /4 , where the fold s d s_d s d is taken as a free parameter. In this example q = 0.925 q=0.925 q = 0.925 corresponding to λ = 0.078 \lambda=0.078 λ = 0.078 , and results are shown for β = 16 \beta=16 β = 16 and 40 40 40 . ∣ F f o l d ( 0 ) ∣ |F_{\mathrm{fold}}(0)| ∣ F fold ( 0 ) ∣ is O ( 1 ) O(1) O ( 1 ) for values of b < b c b<b_c b < b c , where b c ≃ 8.7 ≃ 1.6 β f / 2 b_c\simeq8.7\simeq1.6\beta_f/2 b c ≃ 8.7 ≃ 1.6 β f /2 , and then blows up exponentially, with an oscillation, for larger folds. The dashed line shows a fit to the form F f o l d ( t = 0 ) = F s m o o t h ( b ) + e c ( b ) cos θ ( b ) F_{\mathrm{fold}}(t=0)=F_{\mathrm{smooth}}(b)+e^{c(b)}\cos\theta(b) F fold ( t = 0 ) = F smooth ( b ) + e c ( b ) cos θ ( b ) . Assuming that the blowup corresponds to a complex conjugate pair of saddles, F f o l d ( t = 0 ) = F s m o o t h ( b ) + e S ( b ) + e S ∗ ( b ) F_{\mathrm{fold}}(t=0)=F_{\mathrm{smooth}}(b)+e^{S(b)}+e^{S^*(b)} F fold ( t = 0 ) = F smooth ( b ) + e S ( b ) + e S ∗ ( b ) , we extract Re S ′ ( b ) \operatorname{Re}S'(b) Re S ′ ( b ) and ∣ Im S ′ ( b ) ∣ |\operatorname{Im}S'(b)| ∣ Im S ′ ( b ) ∣ as shown in the right plot. We see that Re S ′ ( b ) \operatorname{Re}S'(b) Re S ′ ( b ) is slowly increasing with b b b and is only weakly dependent on the value of β \beta β . Since the integrand of (41) depends on the fold depth through e s d ( E 2 + E 4 − E 1 − E 3 ) e^{s_d(E_2+E_4-E_1-E_3)} e s d ( E 2 + E 4 − E 1 − E 3 ) , Re S ′ ( b ) \operatorname{Re}S'(b) Re S ′ ( b ) is the total energy splitting between the backward and forward faces at the saddle; it is heading towards its maximal value 4 E 0 4E_0 4 E 0 with the two pairs of faces at opposite edges of the spectrum. The decrease of ∣ Im S ′ ( b ) ∣ |\operatorname{Im}S'(b)| ∣ Im S ′ ( b ) ∣ with larger b b b indicates that the conjugate pair of saddles approach the real axis as b b b increases.
Figure 5. The blowup for q = 0.925 q=0.925 q = 0.925 , λ = 0.078 \lambda=0.078 λ = 0.078 . Left: Log of the DSSYK folded OTOC ∣ F f o l d ( t = 0 ) ∣ |F_{\mathrm{fold}}(t=0)| ∣ F fold ( t = 0 ) ∣ as a function of the effective backwards fold b = s d − β / 4 b=s_d-\beta/4 b = s d − β /4 , where the fold s d s_d s d is taken as a free parameter. The numerics are computed up to n max = 340 n_{\max}=340 n m a x = 340 with dps= 200 =200 = 200 and converge for all points shown. We see that ∣ F f o l d ( t = 0 ) ∣ |F_{\mathrm{fold}}(t=0)| ∣ F fold ( t = 0 ) ∣ is O ( 1 ) O(1) O ( 1 ) for values of b < b c b<b_c b < b c , where b c ≃ 8.7 ≃ 1.6 β f / 2 b_c\simeq8.7\simeq1.6\beta_f/2 b c ≃ 8.7 ≃ 1.6 β f /2 , and then blows up exponentially, with an oscillation, for larger folds. The dashed line shows a fit to the form F f o l d ( t = 0 ) = F s m o o t h ( b ) + e c ( b ) cos θ ( b ) F_{\mathrm{fold}}(t=0)=F_{\mathrm{smooth}}(b)+e^{c(b)}\cos\theta(b) F fold ( t = 0 ) = F smooth ( b ) + e c ( b ) cos θ ( b ) where, for each value of β \beta β , F s m o o t h ( b ) F_{\mathrm{smooth}}(b) F smooth ( b ) , c ( b ) c(b) c ( b ) , and θ ( b ) \theta(b) θ ( b ) are fit to quadratic functions of b b b . Right: Assuming that the blowup corresponds to a complex conjugate pair of saddles, F f o l d ( t = 0 ) = F s m o o t h ( b ) + e S ( b ) + e S ∗ ( b ) F_{\mathrm{fold}}(t=0)=F_{\mathrm{smooth}}(b)+e^{S(b)}+e^{S^*(b)} F fold ( t = 0 ) = F smooth ( b ) + e S ( b ) + e S ∗ ( b ) , we extract Re S ′ ( b ) \operatorname{Re}S'(b) Re S ′ ( b ) and ∣ Im S ′ ( b ) ∣ |\operatorname{Im}S'(b)| ∣ Im S ′ ( b ) ∣ .
For the β = 16 \beta=16 β = 16 example, if we compute the fold s d s_d s d from β / ( 2 v ) \beta/(2v) β / ( 2 v ) , or if we determine it by requiring ∂ s d F f o l d ( t ) ∣ t = 0 = 0 \partial_{s_d}F_{\mathrm{fold}}(t)|_{t=0}=0 ∂ s d F fold ( t ) ∣ t = 0 = 0 , we get b = 7.7 b=7.7 b = 7.7 or 6.9 6.9 6.9 , respectively; thus with our finite β \beta β folding prescription the β = 16 \beta=16 β = 16 folded OTOC, and by extension the folded OTOC for any smaller value of β \beta β , has b < b c b<b_c b < b c . For the β = 40 \beta=40 β = 40 folded OTOC, with the fold defined as β / ( 2 v ) \beta/(2v) β / ( 2 v ) , we get b = 13.6 b=13.6 b = 13.6 , well above b c b_c b c ; for this too-large β \beta β the eikonal behavior cannot be retrieved from the folded OTOC. With the fold β / ( 2 v ) \beta/(2v) β / ( 2 v ) , b b b reaches b c b_c b c at β ≃ 20 \beta\simeq20 β ≃ 20 , i.e. β λ ≃ 5.6 \beta\sqrt{\lambda}\simeq5.6 β λ ≃ 5.6 .
Figure 6 shows in more detail what is happening with the complex saddles. Here we have plotted for the β = 16 \beta=16 β = 16 example just the oscillating part F p a i r ( b ) F_{\mathrm{pair}}(b) F pair ( b ) of the folded OTOC F f o l d ( t = 0 ) F_{\mathrm{fold}}(t=0) F fold ( t = 0 ) , again taking the fold as a free parameter. The oscillating part is extracted using the finite difference method. Here we see that the oscillatory behavior is also present for b < b c b<b_c b < b c , over many orders of magnitude of ∣ F p a i r ( b ) ∣ |F_{\mathrm{pair}}(b)| ∣ F pair ( b ) ∣ , and with approximately the same coefficient and phase as we got by fitting the behavior for b > b c b>b_c b > b c . This indicates that the blowup is anti-Stokes behavior [42 ] : the coefficients of the saddles are approximately constant on both sides of the crossover line b = b c b=b_c b = b c , so the blowup is the pair of saddles becoming dominant over the original symmetric saddle discussed in § 0.6 .
Figure 6. Anti-Stokes behavior for q = 0.925 q=0.925 q = 0.925 , λ = 0.078 \lambda=0.078 λ = 0.078 , β = 16 \beta=16 β = 16 , taking the backward Euclidean segment b = s d − β / 4 b=s_d-\beta/4 b = s d − β /4 as a free parameter. The data are the exact F f o l d F_{\mathrm{fold}} F fold at t = 0 t=0 t = 0 on a uniform grid with spacing h = 0.05 h=0.05 h = 0.05 . Above the onset (shaded band, where the exponentially growing term rises from 1% of the O ( 1 ) O(1) O ( 1 ) correlator to equal it) the correlator is dominated by a complex-conjugate pair of saddles, F f o l d ( t = 0 ) ≃ Re F p a i r F_{\mathrm{fold}}(t=0)\simeq\operatorname{Re}F_{\mathrm{pair}} F fold ( t = 0 ) ≃ Re F pair , F p a i r ( b ) = 2 e S ( b ) F_{\mathrm{pair}}(b)=2e^{S(b)} F pair ( b ) = 2 e S ( b ) , with S ( b ) S(b) S ( b ) quadratic in b b b and fitted for b ≥ 8.3 b\geq8.3 b ≥ 8.3 . The complex saddle pair in the folded OTOC continues smoothly below its onset. Points: Δ 16 F f o l d / ∣ Δ 16 F p a i r ∣ \Delta^{16}F_{\mathrm{fold}}/\lvert\Delta^{16}F_{\mathrm{pair}}\rvert Δ 16 F fold / ∣ Δ 16 F pair ∣ , where Δ 16 \Delta^{16} Δ 16 is the 16th finite difference on the b b b grid. This reduces the slowly varying part of F f o l d F_{\mathrm{fold}} F fold to a residual floor of ∼ 7 × 10 − 10 \sim7\times10^{-10} ∼ 7 × 1 0 − 10 (for ∣ F f o l d ∣ ≃ 1.26 \lvert F_{\mathrm{fold}}\rvert\simeq1.26 ∣ F fold ∣ ≃ 1.26 ), while the pair, which grows by e h Re S ′ ≈ 2.4 e^{h\operatorname{Re}S'}\approx2.4 e h Re S ′ ≈ 2.4 per step, survives. Each point is placed where the pair’s contribution peaks. Dotted: cos arg Δ 16 F p a i r \cos\arg\Delta^{16}F_{\mathrm{pair}} cos arg Δ 16 F pair , i.e. the pair extrapolated from above the onset with unit coefficient. Below the onset the points follow the same oscillation down to b ≃ 7.7 b\simeq7.7 b ≃ 7.7 , where the pair is ∼ 4 × 10 − 7 \sim4\times10^{-7} ∼ 4 × 1 0 − 7 of the smooth part: a two-parameter fit Δ 16 F f o l d = A Re Δ 16 F p a i r + B Im Δ 16 F p a i r \Delta^{16}F_{\mathrm{fold}}=A\operatorname{Re}\Delta^{16}F_{\mathrm{pair}}+B\operatorname{Im}\Delta^{16}F_{\mathrm{pair}} Δ 16 F fold = A Re Δ 16 F pair + B Im Δ 16 F pair on 7.66 ≤ b ≤ 8.37 7.66\leq b\leq8.37 7.66 ≤ b ≤ 8.37 gives A = 1.02 A=1.02 A = 1.02 and B = − 0.04 B=-0.04 B = − 0.04 , compared with A = 1 A=1 A = 1 , B = 0 B=0 B = 0 above the onset by construction.
In Figure 7 we construct the folded OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) with the fold determined by requiring ∂ s d F f o l d ( t ) ∣ t = 0 = 0 \partial_{s_d}F_{\mathrm{fold}}(t)|_{t=0}=0 ∂ s d F fold ( t ) ∣ t = 0 = 0 , and plot it versus time for three q = 0.925 q=0.925 q = 0.925 examples: β = 1 \beta=1 β = 1 , 4 4 4 , and 10 10 10 . Here we have defined the scrambling time as t ∗ = arccosh ( λ L / 2 λ 3 / 2 ) / λ L t_*=\operatorname{arccosh}(\lambda_L/2\lambda^{3/2})/\lambda_L t ∗ = arccosh ( λ L /2 λ 3/2 ) / λ L , with λ L = 2 π v / β \lambda_L=2\pi v/\beta λ L = 2 π v / β , consistent to leading order with [1] and with (55) for β → 0 \beta\to0 β → 0 . For these examples b b b is well below b c = 8.7 b_c=8.7 b c = 8.7 , e.g. for β = 10 \beta=10 β = 10 we have b = 6.0 b=6.0 b = 6.0 . For all three examples a numerical estimate of the contribution from the complex saddles predicts them to be exponentially suppressed over the entire time range 0 < t < 2 t ∗ 0<t<2t_\ast 0 < t < 2 t ∗ , and indeed no sign of them is visible in the plots of the exact points. The first example β = 1 \beta=1 β = 1 tracks the expected de Sitter-like eikonal behavior out to t = 2 t ∗ t=2t_\ast t = 2 t ∗ at about the same level of precision as our previous β = 0 \beta=0 β = 0 results. However the same plot for β = 10 \beta=10 β = 10 (and to a lesser extent for β = 4 \beta=4 β = 4 ) shows substantial deviations.
Figure 7. High temperature examples of the folded DSSYK OTOC F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) for q = 0.925 q=0.925 q = 0.925 , λ = 0.078 \lambda=0.078 λ = 0.078 . For each plot we use the finite β \beta β folding prescription, with the fold s d s_d s d determined by requiring ∂ s d F f o l d ( t ) ∣ t = 0 = 0 \partial_{s_d}F_{\mathrm{fold}}(t)|_{t=0}=0 ∂ s d F fold ( t ) ∣ t = 0 = 0 . For these examples with β = 1 , 4 , 10 \beta=1,4,10 β = 1 , 4 , 10 , we set s d = 5.62 , 6.24 , 8.49 s_d=5.62,6.24,8.49 s d = 5.62 , 6.24 , 8.49 ; by comparison the leading order in small λ \lambda λ formula s d = β / ( 2 v ) s_d=\beta/(2v) s d = β / ( 2 v ) gives s d = 5.68 , 6.39 , 8.88 s_d=5.68,6.39,8.88 s d = 5.68 , 6.39 , 8.88 . The real and imaginary parts of F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) are plotted separately (solid blue and orange lines); also shown is the unfolded OTOC F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) (gray line), and the real and imaginary parts of the eikonal function Φ ( − a ( t ) − i 0 ) \Phi(-a(t)-i0) Φ ( − a ( t ) − i 0 ) (dashed blue and orange lines) where a ( t ) a(t) a ( t ) was computed by matching Φ ( a ( t ) ) \Phi(a(t)) Φ ( a ( t )) to F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) .
Figure 8 shows the origins of the deviations in the β = 10 \beta=10 β = 10 example. Here we fix β λ = 2.79 \beta\sqrt{\lambda}=2.79 β λ = 2.79 , corresponding to the β = 10 \beta=10 β = 10 , λ = 0.078 \lambda=0.078 λ = 0.078 case of Figure 7 , and compute the exact folded OTOC for four different pairs of β , λ \beta,\lambda β , λ values, all of them corresponding to b < b c b<b_c b < b c . We can then compare these with the folded version of the long-time limit of DSSYK, which gave the eikonal-like result (44) . Here we observe that eq. (5.20) of [2] was derived making the approximation of fixing the average energy s = ( s 1 + s 2 + s 3 + s 4 ) / 4 s=(s_1+s_2+s_3+s_4)/4 s = ( s 1 + s 2 + s 3 + s 4 ) /4 to its saddle point value.
Figure 8. The folded OTOC at finite temperature, for fixed β λ = 2.79 \beta\sqrt{\lambda}=2.79 β λ = 2.79 , compared with the long-time limit of [2]. Left: q = 0.925 q=0.925 q = 0.925 , β = 10 \beta=10 β = 10 (λ = 0.078 \lambda=0.078 λ = 0.078 ), s d = 8.49 s_d=8.49 s d = 8.49 ; Middle: q = 0.95 q=0.95 q = 0.95 , β = 12.33 \beta=12.33 β = 12.33 (λ = 0.051 \lambda=0.051 λ = 0.051 ), s d = 10.63 s_d=10.63 s d = 10.63 . Points: the exact Re F f o l d \operatorname{Re} F_{\mathrm{fold}} Re F fold (blue) and Im F f o l d \operatorname{Im} F_{\mathrm{fold}} Im F fold (orange). Dashed: the eikonal-like result of (44) with the average energy fixed to its saddle point. Solid: the same expression averaged over the thermal distribution of the energy parameter s = θ / λ s=\theta/\lambda s = θ / λ across which the Lyapunov exponent λ L \lambda_L λ L varies, using λ L ( θ ) = 2 λ sin θ \lambda_L(\theta)=2\sqrt{\lambda}\sin\theta λ L ( θ ) = 2 λ sin θ . Right: the remaining deviation rms ( e x a c t − t h e r m a l ) \operatorname{rms}(\mathrm{exact}-\mathrm{thermal}) rms ( exact − thermal ) , with the rms taken over 0 ≤ t ≤ 1.5 t ∗ 0\le t\le1.5t_\ast 0 ≤ t ≤ 1.5 t ∗ , as a function of λ \lambda λ at fixed β λ = 2.79 \beta\sqrt{\lambda}=2.79 β λ = 2.79 , for q = 0.875 q=0.875 q = 0.875 , 0.9 0.9 0.9 , 0.925 0.925 0.925 and 0.95 0.95 0.95 . The dotted line is a power law fit to the four points.
At finite temperature this energy s = θ / λ s=\theta/\lambda s = θ / λ has a thermal distribution that affects the Lyapunov exponent, i.e. λ L ( θ ) = λ E 0 sin θ → 2 λ sin θ \lambda_L(\theta)=\lambda E_0\sin\theta\to2\sqrt{\lambda}\sin\theta λ L ( θ ) = λ E 0 sin θ → 2 λ sin θ at leading order, which in turn affects the fold rotation of z z z in (48) which is λ L ( θ ) s d \lambda_L(\theta)s_d λ L ( θ ) s d with λ L ( θ s ) s d = π \lambda_L(\theta_s)s_d=\pi λ L ( θ s ) s d = π at the saddle value θ s \theta_s θ s . Thus for a fixed fold s d s_d s d , we expect a thermal spread in the fold rotation angle of order the saddle standard deviation σ θ \sigma_\theta σ θ from (15) times d λ L / d θ ∣ θ = θ s s d d\lambda_L/d\theta|_{\theta=\theta_s}s_d d λ L / d θ ∣ θ = θ s s d , which is π cot θ s σ θ \pi\cot\theta_s\,\sigma_\theta π cot θ s σ θ , of order λ \sqrt{\lambda} λ .
The solid lines in the left and middle panels of Figure 8 were obtained by performing the thermal average over ρ q ( s ) e − β E ( s ) \rho_q(s)e^{-\beta E(s)} ρ q ( s ) e − β E ( s ) on a grid ± 8 \pm8 ± 8 sigma around the saddle, weighted by sin 4 ( λ s ) \sin^4(\lambda s) sin 4 ( λ s ) and normalized by ⟨ sin 2 ( λ s ) ⟩ 2 \langle\sin^2(\lambda s)\rangle^2 ⟨ sin 2 ( λ s ) ⟩ 2 from the two-point functions, to account for the prefactors in eqs. (5.17) and (4.12) of [2]. For the part of the thermal distribution with λ L ( θ ) s d > π \lambda_L(\theta)s_d>\pi λ L ( θ ) s d > π , we continued analytically through the cut on the negative real axis as required by analyticity in the fold depth at fixed energy. Clearly the solid lines agree much better with the exact points. The small remaining deviation of the exact points from the thermally averaged form falls roughly as λ 1.1 \lambda^{1.1} λ 1.1 at fixed x x x (right panel of Figure 8 ), consistent with additional finite λ \lambda λ effects.
We conclude that, taking into account the thermal spread in the fold rotation angle, the β = 0 \beta=0 β = 0 folding prescription successfully extends to finite β \beta β for sufficiently high temperatures such that we satisfy the condition b < b c ≃ 1.6 β f / 2 b<b_c\simeq1.6\,\beta_f/2 b < b c ≃ 1.6 β f /2 on the fold depth. Above b c b_c b c the conjugate saddle pair dominates already at t = 0 t=0 t = 0 , which is the anti-Stokes crossover of Figure 5 and Figure 6 .
Late time behavior and the two-shell dilaton invariant
Harlow and Zhao derive the eikonal formula from the time advance of a probe crossing a shell, linearized in the shell energy (their eq. (3.34)); the neglected terms become O ( 1 ) O(1) O ( 1 ) at twice the scrambling time, thus their four-point function is stated to hold for “milder” (in terms of backreaction) anti-scrambling in the regime 1 ≪ ∣ t 1 − t 2 ∣ < 2 t ∗ 1\ll|t_1-t_2|<2t_* 1 ≪ ∣ t 1 − t 2 ∣ < 2 t ∗ (their eq. (4.29)).
The late time behavior of F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) is shown in Figure 9 and Figure 10 . As before we extract a ( t ) a(t) a ( t ) from F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) to compute Φ ( − a ( t ) ) \Phi(-a(t)) Φ ( − a ( t )) . We find that the extracted a ( t ) a(t) a ( t ) remains exponential: d log a / d t d\log a/dt d log a / d t settles to within 0.4 % 0.4\% 0.4% of 2 π / β f 2\pi/\beta_f 2 π / β f , with no change at 2 t ∗ 2t_* 2 t ∗ . We see that F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) continues to track Φ ( − a ( t ) ) \Phi(-a(t)) Φ ( − a ( t )) at late times; its magnitude decreases monotonically after peaking at t ∗ t_* t ∗ and it has the same asymptotic behavior as F s t d ( t ) F_{\mathrm{std}}(t) F std ( t ) .
Figure 9. Late-time behavior of the folded OTOC. Plotted are ∣ F f o l d ∣ |F_{\mathrm{fold}}| ∣ F fold ∣ (solid line), F s t d F_{\mathrm{std}} F std (dashed line), and eikonal ∣ Φ ( − a ) ∣ |\Phi(-a)| ∣Φ ( − a ) ∣ (circles), with a ( t ) a(t) a ( t ) extracted from F s t d F_{\mathrm{std}} F std , versus time for q = 0.95 q=0.95 q = 0.95 , corresponding to λ = 0.0513 \lambda=0.0513 λ = 0.0513 , β f = 13.786 \beta_f=13.786 β f = 13.786 , and scrambling time t ∗ = 8.09 t_\ast=8.09 t ∗ = 8.09 .
Figure 10. Late-time behavior of the folded OTOC in the complex plane. Plotted are the trajectory of F f o l d ( t ) F_{\mathrm{fold}}(t) F fold ( t ) (solid line, colored by t t t ) and the eikonal Φ ( − a ( t ) − i 0 ) \Phi(-a(t)-i0) Φ ( − a ( t ) − i 0 ) (circles), with a ( t ) a(t) a ( t ) extracted from F s t d F_{\mathrm{std}} F std , for q = 0.95 q = 0.95 q = 0.95 , corresponding to λ = 0.0513 \lambda= 0.0513 λ = 0.0513 , β f = 13.786 \beta_f = 13.786 β f = 13.786 , and scrambling time t ∗ = 8.09 t_* = 8.09 t ∗ = 8.09 . The inset zooms on t ≥ 11.5 t \ge11.5 t ≥ 11.5 , where Im F f o l d \operatorname{Im} F_{\mathrm{fold}} Im F fold has turned negative and approaches zero as − π / a 2 -\pi/a^2 − π / a 2 .
Harlow and Zhao in Section 3 of [1] mapped 2 + 1 2+1 2 + 1 -dimensional gravity with Λ > 0 \Lambda>0 Λ > 0 and spherical dust shells onto JT gravity with positive cosmological constant, and solved the problem of two colliding shells at the classical level. They expressed their solution in terms of κ J T \kappa_{\mathrm{JT}} κ JT , the JT gravity version of the dilaton invariant defined for sine-dilaton gravity in (11) . κ J T \kappa_{\mathrm{JT}} κ JT takes different values in each of the four regions cut out by the shells, and in one of the post-collision regions their solution contained a term growing exponentially with the time separation t 2 t_2 t 2 of the two shells,
κ J T ≃ κ J T 0 + 2 κ J T 0 ω 2 − 2 ω 1 ω 2 cosh t 2 , (65) \kappa_{\mathrm{JT}} \simeq\kappa_{\mathrm{JT}}^0+2\sqrt{\kappa_{\mathrm{JT}}^0}\omega_2-2\omega_1\omega_2\cosh t_2, \tag*{(65)} κ JT ≃ κ JT 0 + 2 κ JT 0 ω 2 − 2 ω 1 ω 2 cosh t 2 , (65)
their eq. (3.45) expanded to second order in ω i / κ J T 0 \omega_i/\sqrt{\kappa_{\mathrm{JT}}^0} ω i / κ JT 0 , where κ J T 0 \kappa_{\mathrm{JT}}^0 κ JT 0 is the unperturbed value of the dilaton invariant. The first-order term is the shift produced by a single shell, of relative size O ( ω / κ J T 0 ) O(\omega/\sqrt{\kappa_{\mathrm{JT}}^0}) O ( ω / κ JT 0 ) ; the other regions have only shifts of this kind. The bilinear term is the one that grows: at − t 2 = 2 t s c r + O ( 1 ) -t_2 = 2t_{\mathrm{scr}} + O(1) − t 2 = 2 t scr + O ( 1 ) , with t s c r = log S d S t_{\mathrm{scr}} = \log S_{\mathrm{dS}} t scr = log S dS (their eqs. (3.33) and (3.46)), it overwhelms κ J T 0 \kappa_{\mathrm{JT}}^0 κ JT 0 , κ J T \kappa_{\mathrm{JT}} κ JT changes sign, the dilaton becomes timelike, and this post-collision region ends in a big crunch. The growing term is the same gravitational time advance that produces anti-scrambling in the DSSYK folded OTOC; the crunch is its backreaction followed to the end.
Since DSSYK has a bounded spectrum, we expect that the two-shell computation should have a DSSYK counterpart where the would-be space-time singularity is replaced with some other kind of behavior. In this section we identify this counterpart and study it for the case of both light and heavy probes.
We can extract the unperturbed dilaton invariant κ J T 0 \kappa_{\mathrm{JT}}^0 κ JT 0 from the sine-dilaton discussion of Section III. Assume a geodesic observer at r = π r = \pi r = π and expand the classical solution of the sine-dilaton field Φ s d \Phi_{\mathrm{sd}} Φ sd near the observer as Φ s d = π + ρ \Phi_{\mathrm{sd}} = \pi+ \rho Φ sd = π + ρ . Then, ignoring boundary terms, the sine-dilaton action (9) reduces to:
1 2 λ [ ∫ d t d r ( π R + ρ R + 2 sin ( π + ρ ) ) ] (66) \frac{1}{2\lambda}\left[\int dt\,dr\,\left(\pi R+\rho R+2\sin(\pi+\rho)\right)\right] \tag*{(66)} 2 λ 1 [ ∫ d t d r ( π R + ρR + 2 sin ( π + ρ ) ) ] (66)
Ignoring the topological term π R \pi R π R , this is the action of two-dimensional JT gravity with a positive cosmological constant:
1 2 [ ∫ d t d r Φ J T ( R − 2 ) ] (67) \frac{1}{2}\left[\int dt\,dr\,\Phi_{\mathrm{JT}}(R-2)\right] \tag*{(67)} 2 1 [ ∫ d t d r Φ JT ( R − 2 ) ] (67)
where we have identified the JT dilaton as ρ / λ \rho/\lambda ρ / λ . We can now compute the JT dilaton invariant:
κ J T ( θ ) = ( ∇ Φ J T ) 2 + Φ J T 2 = F λ 2 + ρ 2 λ 2 ≃ 2 λ 2 ( 1 + cos θ ) = 2 λ 2 ( 1 − E E 0 ) (68) \kappa_{\mathrm{JT}}(\theta)=(\nabla\Phi_{\mathrm{JT}})^2+\Phi_{\mathrm{JT}}^2=\frac{F}{\lambda^2}+\frac{\rho^2}{\lambda^2}\simeq\frac{2}{\lambda^2}(1+\cos\theta)=\frac{2}{\lambda^2}\left(1-\frac{E}{E_0}\right) \tag*{(68)} κ JT ( θ ) = ( ∇ Φ JT ) 2 + Φ JT 2 = λ 2 F + λ 2 ρ 2 ≃ λ 2 2 ( 1 + cos θ ) = λ 2 2 ( 1 − E 0 E ) (68)
Although we computed κ J T ( θ ) \kappa_{\mathrm{JT}}(\theta) κ JT ( θ ) in the region ρ ≪ 1 \rho\ll1 ρ ≪ 1 near the observer, it differs from the sine-dilaton invariant κ s d ( θ ) \kappa_{\mathrm{sd}}(\theta) κ sd ( θ ) only by constants, so in fact (68) applies in the entire bulk region; it is the dilaton invariant of [1]. Notice that the big crunch behavior corresponds to κ J T \kappa_{\mathrm{JT}} κ JT decreasing to zero and becoming negative, which requires the energy becoming larger than E 0 E_0 E 0 [23 , 43 ] .
The two-shell analysis in DSSYK corresponds to a correlator with four matter insertions and thus four gaps:
F ( g 1 , g 2 , g 3 , g 4 ) = Tr [ e − g 1 H V e − g 2 H W e − g 3 H V e − g 4 H W ] (69) F(g_1,g_2,g_3,g_4)=\operatorname{Tr}\left[e^{-g_1H}V e^{-g_2H}W e^{-g_3H}V e^{-g_4H}W\right] \tag*{(69)} F ( g 1 , g 2 , g 3 , g 4 ) = Tr [ e − g 1 H V e − g 2 H W e − g 3 H V e − g 4 H W ] (69)
From the exact DSSYK expression (41) we see that the chord expansion assigns an energy E n E_n E n with Boltzmann weight e − g n E n e^{-g_nE_n} e − g n E n to each of the four faces of the crossed chord diagram, with the expectation values given by a spectral representation of the form:
⟨ E n ⟩ = − ∂ ∂ g n log F = ∫ 0 π d θ E ( θ ) I n ( θ ) ∫ 0 π d θ I n ( θ ) (70) \langle E_n\rangle=-\frac{\partial}{\partial g_n}\log F=\frac{\int_0^\pi d\theta\,E(\theta)I_n(\theta)}{\int_0^\pi d\theta\,I_n(\theta)} \tag*{(70)} ⟨ E n ⟩ = − ∂ g n ∂ log F = ∫ 0 π d θ I n ( θ ) ∫ 0 π d θ E ( θ ) I n ( θ ) (70)
where I n ( θ ) I_n(\theta) I n ( θ ) is the exact energy distribution, which in general is complex.
For our DSSYK OTOC F f o l d F_{\mathrm{fold}} F fold we take the gaps ( g 1 , g 2 , g 3 , g 4 ) = ( − β f / 2 − i t , β f / 2 + i t , − β f / 2 − i t , β f / 2 + i t ) (g_1,g_2,g_3,g_4)=(-\beta_f/2-it,\beta_f/2+it,-\beta_f/2-it,\beta_f/2+it) ( g 1 , g 2 , g 3 , g 4 ) = ( − β f /2 − i t , β f /2 + i t , − β f /2 − i t , β f /2 + i t ) . These are the gaps of § 0.6 with all signs reversed, which gives the same correlator by the evenness of f f f (§ 0.5 ); with this labeling faces 1 and 3 are the backward-evolved ones, with Boltzmann weight e + ( β f / 2 + i t ) E e^{+(\beta_f/2+it)E} e + ( β f /2 + i t ) E . Thus:
⟨ E 1 ⟩ = ⟨ E 3 ⟩ = − ⟨ E 2 ⟩ = − ⟨ E 4 ⟩ ≡ ⟨ E f o l d ⟩ ( t ) , ⟨ E f o l d ⟩ ( t ) = − i 4 d d t log F f o l d ( t ) (71) \langle E_1\rangle=\langle E_3\rangle=-\langle E_2\rangle=-\langle E_4\rangle\equiv\langle E_{\mathrm{fold}}\rangle(t),\qquad\langle E_{\mathrm{fold}}\rangle(t)=-\frac{i}{4}\frac{d}{dt}\log F_{\mathrm{fold}}(t) \tag*{(71)} ⟨ E 1 ⟩ = ⟨ E 3 ⟩ = − ⟨ E 2 ⟩ = − ⟨ E 4 ⟩ ≡ ⟨ E fold ⟩ ( t ) , ⟨ E fold ⟩ ( t ) = − 4 i d t d log F fold ( t ) (71)
In the regime a ≪ 1 a\ll1 a ≪ 1 , where F f o l d F_{\mathrm{fold}} F fold tracks Φ ( − a ( t ) ) \Phi(-a(t)) Φ ( − a ( t )) , restoring the dependence on the conformal dimensions Δ V = Δ W = Δ \Delta_V=\Delta_W=\Delta Δ V = Δ W = Δ and working to leading order in small λ \lambda λ :
log Φ ( − a ( t ) ) ≃ λ Δ 2 cosh ( λ L t ) ≃ 1 2 λ Δ 2 e 2 λ t (72) \log\Phi(-a(t))\simeq\lambda\Delta^2\cosh(\lambda_Lt)\simeq\frac{1}{2}\lambda\Delta^2e^{2\sqrt{\lambda}t} \tag*{(72)} log Φ ( − a ( t )) ≃ λ Δ 2 cosh ( λ L t ) ≃ 2 1 λ Δ 2 e 2 λ t (72)
Using (71) and writing κ ( t ) = κ 0 + Δ κ ( t ) \kappa(t)=\kappa_0+\Delta\kappa(t) κ ( t ) = κ 0 + Δ κ ( t ) , with κ 0 = 2 / λ 2 \kappa_0=2/\lambda^2 κ 0 = 2/ λ 2 , we can compare the DSSYK result to the two-shell result of [1]:
Δ κ κ 0 = − ⟨ E ⟩ f o l d E 0 ≃ i λ 2 Δ 2 8 e λ L t ⟷ Δ κ J T κ J T 0 ≃ − ω 1 ω 2 κ J T 0 e ∣ t 2 ∣ (73) \frac{\Delta\kappa}{\kappa_0}=-\frac{\langle E\rangle_{\mathrm{fold}}}{E_0}\simeq\frac{\mathrm{i}\lambda^2\Delta^2}{8}e^{\lambda_L t}\quad\longleftrightarrow\quad\frac{\Delta\kappa_{\mathrm{JT}}}{\kappa_{\mathrm{JT}}^0}\simeq-\frac{\omega_1\omega_2}{\kappa_{\mathrm{JT}}^0}e^{\lvert t_2\rvert} \tag*{(73)} κ 0 Δ κ = − E 0 ⟨ E ⟩ fold ≃ 8 i λ 2 Δ 2 e λ L t ⟷ κ JT 0 Δ κ JT ≃ − κ JT 0 ω 1 ω 2 e ∣ t 2 ∣ (73)
Matching terms we get ∣ t 2 ∣ = λ L t = ( 2 π / β f ) t \lvert t_2\rvert=\lambda_Lt=(2\pi/\beta_f)t ∣ t 2 ∣ = λ L t = ( 2 π / β f ) t to leading order in small λ \lambda λ , consistent with S d S = 2 / λ S_{\mathrm{dS}}=2/\lambda S dS = 2/ λ and our definition (55) of the scrambling time; the shell energies match as ω 1 ω 2 ↔ λ 2 Δ 2 κ 0 / 8 = Δ 2 / 4 \omega_1\omega_2\leftrightarrow\lambda^2\Delta^2\kappa_0/8=\Delta^2/4 ω 1 ω 2 ↔ λ 2 Δ 2 κ 0 /8 = Δ 2 /4 , i.e. ω ≃ Δ / 2 \omega\simeq\Delta/2 ω ≃ Δ/2 , thermal-scale shells with ω / κ 0 = O ( λ Δ ) \omega/\sqrt{\kappa_0}=O(\lambda\Delta) ω / κ 0 = O ( λ Δ ) , the regime of [1]. The putative crunch time in DSSYK time units is given by:
t c r u n c h = 2 t ∗ + 1 λ L log 2 Δ 2 (74) t_{\mathrm{crunch}}=2t_*+\frac{1}{\lambda_L}\log\frac{2}{\Delta^2} \tag*{(74)} t crunch = 2 t ∗ + λ L 1 log Δ 2 2 (74)
which is the classical extrapolation of the scrambling-regime growth and, e.g. for q = 0.95 q=0.95 q = 0.95 , lies about 10% past 2 t ∗ 2t_* 2 t ∗ .
From this exercise we see that the DSSYK analysis of F f o l d F_{\mathrm{fold}} F fold matches up structurally with the classical two-shell but with crucial differences. The DSSYK quantity in (73) is imaginary, while the two-shell Δ κ J T \Delta\kappa_{\mathrm{JT}} Δ κ JT is real and negative. Writing F f o l d = ∣ F f o l d ∣ e i φ F_{\mathrm{fold}}=\lvert F_{\mathrm{fold}}\rvert e^{\mathrm{i}\varphi} F fold = ∣ F fold ∣ e i φ , (71) gives Re ⟨ E ⟩ f o l d = 1 4 d φ / d t \operatorname{Re}\langle E\rangle_{\mathrm{fold}}=\frac{1}{4}\,d\varphi/dt Re ⟨ E ⟩ fold = 4 1 d φ / d t and Im ⟨ E ⟩ f o l d = − 1 4 d log ∣ F f o l d ∣ / d t \operatorname{Im}\langle E\rangle_{\mathrm{fold}}=-\frac{1}{4}\,d\log\lvert F_{\mathrm{fold}}\rvert/dt Im ⟨ E ⟩ fold = − 4 1 d log ∣ F fold ∣ / d t : the growing term comes from the growth of the modulus of the folded correlator, and appears as an imaginary part, i.e. a phase, of κ \kappa κ . The match is therefore of the growth exponent and of the magnitude of the two-shell term. The real part of κ \kappa κ , which in the classical analysis decreases through zero at 2 t s c r 2t_{\mathrm{scr}} 2 t scr , here only dips to a positive minimum, Re κ / κ 0 = 1 − O ( λ Δ ) \operatorname{Re}\kappa/\kappa_0=1-O(\lambda\Delta) Re κ / κ 0 = 1 − O ( λ Δ ) . In the eikonal regime, to leading order in small λ \lambda λ ,
min t Re κ / κ 0 ≃ 1 − 0.40 λ (75) \min_t\operatorname{Re}\kappa/\kappa_0\simeq1-0.40\lambda \tag*{(75)} t min Re κ / κ 0 ≃ 1 − 0.40 λ (75)
for Δ = 1 \Delta=1 Δ = 1 . With the identification of ⟨ E ⟩ f o l d \langle E\rangle_{\mathrm{fold}} ⟨ E ⟩ fold as the dilaton invariant, the folded correlator for thermal-scale probes does not reproduce the two-shell backreaction of [1]; the growing term is present as a phase of κ \kappa κ , and its real part never exceeds the single-shell size.
Heavy probes: a negative-temperature ensemble and the Nariai point instead of a singularity
Here we consider the case λ Δ = O ( 1 ) \lambda\Delta=O(1) λ Δ = O ( 1 ) , i.e. shells whose energy ω ≃ Δ / 2 \omega\simeq\Delta/2 ω ≃ Δ/2 is a finite fraction of κ 0 \sqrt{\kappa_0} κ 0 , the scale at which the two-shell expansion of [1] stops applying.
DSSYK simplifies further when λ Δ \lambda\Delta λ Δ is large. An H H H -chord crossing a matter chord of dimension Δ \Delta Δ costs q Δ q^{\Delta} q Δ ; for q Δ = e − λ Δ → 0 q^{\Delta}=e^{-\lambda\Delta}\to0 q Δ = e − λ Δ → 0 the H H H -chords of each face must open and close within the face, the vertex factors between adjacent faces go to one, and the four-point function factorises into a product over faces, each a canonical ensemble at inverse temperature equal to its gap.
With Z ( g ) = ∫ d μ e g E Z(g)=\int d\mu\,e^{gE} Z ( g ) = ∫ d μ e g E and given the E → − E E\to-E E → − E symmetry of the measure, each of the four faces contributes Z ( β f / 2 + i t ) Z(\beta_f/2+it) Z ( β f /2 + i t ) , so F f o l d → q Δ V Δ W Z ( β f / 2 + i t ) 4 F_{\mathrm{fold}}\to q^{\Delta_V\Delta_W}Z(\beta_f/2+it)^4 F fold → q Δ V Δ W Z ( β f /2 + i t ) 4 , where the t t t -independent prefactor is the crossing of the two matter chords and drops out of the face energies. This is a complex value whose phase winds with t t t . In the same limit the heavy two-point function is G 0 ( τ ) → Z ( τ ) 2 G_0(\tau)\to Z(\tau)^2 G 0 ( τ ) → Z ( τ ) 2 , so Z ( β f / 2 + i t ) 4 Z(\beta_f/2+it)^4 Z ( β f /2 + i t ) 4 is the product of the V V VV V V and W W WW W W two-point functions at separation β f / 2 + i t \beta_f/2+it β f /2 + i t . There is no small connected correction to expand in: the standard OTOC goes to q Δ 2 → 0 q^{\Delta^2}\to0 q Δ 2 → 0 at t = 0 t=0 t = 0 , and the eikonal regime a ≪ 1 a\ll1 a ≪ 1 is not available.
The folded face energy is
⟨ E ⟩ f o l d ⟶ λ Δ ≫ 1 ∫ d μ E e ( β f / 2 + i t ) E ∫ d μ e ( β f / 2 + i t ) E (76) \langle E\rangle_{\mathrm{fold}}
\underset{\lambda\Delta\gg1}{\longrightarrow}
\frac{\int d\mu\,E e^{(\beta_f/2+it)E}}{\int d\mu\,e^{(\beta_f/2+it)E}} \tag*{(76)} ⟨ E ⟩ fold λ Δ ≫ 1 ⟶ ∫ d μ e ( β f /2 + i t ) E ∫ d μ E e ( β f /2 + i t ) E (76)
the thermal energy of DSSYK at inverse temperature − β f / 2 -\beta_f/2 − β f /2 , continued in real time. For t ≠ 0 t\ne0 t = 0 the weights are complex, so this is the logarithmic derivative of the partition function at complex β \beta β rather than an ensemble average, which is why its real part is not confined to ∣ E ∣ ≤ E 0 |E|\le E_0 ∣ E ∣ ≤ E 0 . A negative-temperature form of the fold was already noted in [1] (their eqs. (6.11), (6.12)); in DSSYK the H → − H H\to-H H → − H symmetry of the spectrum turns their (6.11) into the evenness of f f f used in § 0.5 . In the bulk the heavy probes are walls that cut the geometry into static patches with no interaction between them.
At t = 0 t=0 t = 0 the Boltzmann factor e + β f E / 2 e^{+\beta_f E/2} e + β f E /2 is already pushing towards the upper spectral edge, where, however, the density of states is vanishing like sin 2 θ \sin^2\theta sin 2 θ . The balance is determined by the saddle of S ( θ ) − π cos θ S(\theta)-\pi\cos\theta S ( θ ) − π cos θ , at θ = π / 2 + u \theta=\pi/2+u θ = π /2 + u with 4 u = π cos u 4u=\pi\cos u 4 u = π cos u , giving E / E 0 = sin u = 0.592 E/E_0=\sin u=0.592 E / E 0 = sin u = 0.592 as λ → 0 \lambda\to0 λ → 0 ; the exact numerical measure gives 0.585 at q = 0.95 q=0.95 q = 0.95 . The folded face thus sits more than half-way to κ = 0 \kappa=0 κ = 0 before the heavy shells have moved.
Figure 11 illustrates the behavior of the canonical face for q = 0.95 q=0.95 q = 0.95 , along with exact DSSYK results for Δ = 1 , 8 , 32 \Delta=1,8,32 Δ = 1 , 8 , 32 , and 48 48 48 , corresponding to λ Δ = 0.05 \lambda\Delta=0.05 λ Δ = 0.05 , 0.4 0.4 0.4 , 1.6 1.6 1.6 , and 2.5 2.5 2.5 . The canonical curve crosses Re κ = 0 \operatorname{Re}\kappa=0 Re κ = 0 at t = 1.09 β f t=1.09\,\beta_f t = 1.09 β f , goes slightly negative, then returns to zero from below as Re κ / κ 0 ≃ − π 2 / ( 2 λ t 2 ) \operatorname{Re}\kappa/\kappa_0\simeq-\pi^2/(2\lambda t^2) Re κ / κ 0 ≃ − π 2 / ( 2 λ t 2 ) , for β f ≪ t ≪ λ − 3 / 2 \beta_f\ll t\ll\lambda^{-3/2} β f ≪ t ≪ λ − 3/2 . The reason it can go slightly negative in spite of only having support for E ≤ E 0 E \le E_0 E ≤ E 0 is because of the complex weights. The two light examples shown in the figure Δ = 1 \Delta= 1 Δ = 1 and 8 8 8 that have λ Δ < 1 \lambda\Delta< 1 λ Δ < 1 both exhibit the light probe behavior and do not cross zero. The two heavy examples, Δ = 32 \Delta= 32 Δ = 32 and 48 48 48 , track the canonical curve up to corrections of order O ( q Δ ) O(q^\Delta) O ( q Δ ) .
Figure 11. The DSSYK folded face energy at q = 0.95 q = 0.95 q = 0.95 (λ = 0.051 \lambda= 0.051 λ = 0.051 ) for light and heavy probes. Plotted are the exact DSSYK values for Re ⟨ E ⟩ f o l d / E 0 = 1 − Re κ / κ 0 \operatorname{Re}\langle E\rangle_{\mathrm{fold}}/E_0 = 1-\operatorname{Re}\kappa/\kappa_0 Re ⟨ E ⟩ fold / E 0 = 1 − Re κ / κ 0 for Δ = 1 , 8 , 32 , 48 \Delta= 1, 8, 32, 48 Δ = 1 , 8 , 32 , 48 , (λ Δ = 0.05 , 0.4 , 1.6 , 2.5 \lambda\Delta= 0.05, 0.4, 1.6, 2.5 λ Δ = 0.05 , 0.4 , 1.6 , 2.5 ), against the canonical face (76) , which reaches the edge, Re κ = 0 \operatorname{Re}\kappa= 0 Re κ = 0 , at t c ≃ 1.1 β f t_c \simeq1.1\beta_f t c ≃ 1.1 β f (1.86 t ∗ 1.86t_* 1.86 t ∗ at this coupling), overshoots it by at most 5% and returns to it as π 2 / ( 2 λ t 2 ) \pi^2/(2\lambda t^2) π 2 / ( 2 λ t 2 ) for t ≳ 5 β f t \gtrsim5\beta_f t ≳ 5 β f . Shells with λ Δ < 1 \lambda\Delta< 1 λ Δ < 1 (Δ = 1 , 8 \Delta= 1, 8 Δ = 1 , 8 ) keep the light-shell form, a bounded peak followed by relaxation; for λ Δ = 1.6 \lambda\Delta= 1.6 λ Δ = 1.6 and 2.5 2.5 2.5 the folded face starts at ( 1 − q Δ ) (1-q^\Delta) ( 1 − q Δ ) times the canonical value and follows the canonical curve to the edge. The numerical results shown here required multiple precision with n max n_{\max} n m a x as high as 480 and dps as high as 360.
The folded face and the two heavy examples thus reach the upper edge of the spectrum, θ = π \theta= \pi θ = π , E = E 0 E = E_0 E = E 0 , which is the Nariai point of the sine-dilaton solutions (10) , where the black-hole and cosmological horizons coincide. They reach it as the band edge of a negative-temperature ensemble, at a time set by β f \beta_f β f and unrelated to the scrambling time, and with no collision of probes. This is the sense in which the bounded spectrum replaces the classical crunch: the quantity that would diverge is the energy of a thermal ensemble, but it saturates.
Discussion and Outlook
We have shown that the Euclidean-fold prescription of Harlow and Zhao, implemented in the double-scaled SYK model at infinite temperature with the fake temperature β f = π / J f \beta_f = \pi/\mathcal{J}_f β f = π / J f defining the fold, produces a correlator that anti-scrambles: the connected part grows rather than decays. We have shown that the exact DSSYK evaluation approximately tracks the eikonal amplitude Φ ( − a ) \Phi(-a) Φ ( − a ) , with the shockwave time delay replaced by a time advance, through the scrambling time t ∗ t_* t ∗ and beyond, out to 2.5 t ∗ 2.5t_* 2.5 t ∗ at q = 0.95 q = 0.95 q = 0.95 .
Because DSSYK is exactly solvable, the sheet selection can be checked with no free parameter: the continued eikonal amplitude is complex, the n = 0 n = 0 n = 0 fold approaches the cut from below, and the exact correlator carries the exponentially small term π w 2 ( w − 1 ) e − w \pi w^2(w-1)e^{-w} π w 2 ( w − 1 ) e − w with the positive sign that this predicts, growing into an O ( 1 ) O(1) O ( 1 ) imaginary part at the scrambling time; the n = 1 n = 1 n = 1 fold gives the conjugate.
Using the face energies of the chord expansion we have compared the folded correlator with the two-shell dilaton invariant of [1]. The growing term appears with the same exponent, but as an imaginary part of the invariant, driven by the growth of the modulus of the folded correlator. Its real part, which carries the backreaction in the classical analysis, only moves by O ( λ ) O(\lambda) O ( λ ) for thermal-scale probes, the size of the single-shell terms, so the fold does not reproduce the approach to the crunch. For heavy probes the correlator factorizes and the folded face is a negative-temperature ensemble whose energy saturates at the band edge, the
Nariai point of the sine-dilaton dual.
At finite β \beta β the natural generalization is to fold by half the fake period, β / ( 2 v ) \beta/(2v) β / ( 2 v ) . Modulo the effect of thermal spreading of the fold rotation, a folding prescription based on this idea gives reasonable agreement with the de Sitter eikonal form for sufficiently high temperatures such that we satisfy the condition that the backward Euclidean segment is less than $b_c \simeq 1.5– 1.6 β f / 2 1.6\,\beta_f/2 1.6 β f /2 ; at lower temperatures where this bound is exceeded, the correlator is dominated by a pair of complex conjugate saddles, an anti-Stokes crossover that prevents retrieval of the eikonal behavior from the folded OTOC.
Our results are consistent with Chen, Stanford, Tang and Yang [4], who show that the static-patch eikonal OTOC with observer recoil is not a trace of Hermitian operators on a Hilbert space: for such a trace the regularized OTOC cannot exceed its factorized value, and the static-patch correlator does. DSSYK at β = 0 \beta=0 β = 0 is such a trace. Its standard OTOC obeys ∣ F s t d ∣ ≤ 1 \lvert F_{\mathrm{std}}\rvert\le1 ∣ F std ∣ ≤ 1 and decreases from q Δ V Δ W q^{\Delta_V\Delta_W} q Δ V Δ W with exponent 2 π / β f 2\pi/\beta_f 2 π / β f . The growth we find is in F f o l d ( t ) = Tr [ V ~ ( − t ) W V ~ ( − t ) W ] F_{\mathrm{fold}}(t)=\operatorname{Tr}[\widetilde{V}(-t)W\widetilde{V}(-t)W] F fold ( t ) = Tr [ V ( − t ) W V ( − t ) W ] , V ~ = e − β f H / 2 V e β f H / 2 \widetilde{V}=e^{-\beta_f H/2}Ve^{\beta_f H/2} V = e − β f H /2 V e β f H /2 , V ~ ( t ) = e i H t V ~ e − i H t \widetilde{V}(t)=e^{iHt}\widetilde{V}e^{-iHt} V ( t ) = e i H t V e − i H t , a cyclic and positive trace of a non-Hermitian probe whose Cauchy–Schwarz bound is Tr [ V ~ † V ~ ] = G 0 ( β f ) \operatorname{Tr}[\widetilde{V}^{\dagger}\widetilde{V}]=G_0(\beta_f) Tr [ V † V ] = G 0 ( β f ) , exponentially large in 1 / λ 1/\lambda 1/ λ (G 0 ( β f ) ≃ e 3.2 / λ G_0(\beta_f)\simeq e^{3.2/\lambda} G 0 ( β f ) ≃ e 3.2/ λ for q = 0.9 q=0.9 q = 0.9 –0.95 0.95 0.95 ); the O ( 1 ) O(1) O ( 1 ) value of F f o l d F_{\mathrm{fold}} F fold is a cancellation fixed by the fold being half the fake period. The map from the de Sitter field to DSSYK is V ↦ V V\mapsto V V ↦ V at two points and V ↦ V ~ V\mapsto\widetilde{V} V ↦ V at four points. It preserves the two-point KMS structure and breaks it at four points, as in [1, 3], and it is not a ∗ \ast ∗ -map; thus it does not contradict the tracial no-go.
Several questions are left open:
The folded correlator is not the expectation value of a product of operators in a state, so it is natural to ask if it can be given an operational meaning, for example in a two-sided version of the construction. A two-sided approach could also connect to the de Sitter OTOC analyses of [38 , 44 ] .
Acknowledgements
We thank Alex Zlokapa for useful discussions. Madeline Mooney is supported by the Research Corporation for Science Advancement and by the Brinson Foundation. This work has been authored by Fermi Forward Discovery Group, LLC under Contract No. 89243024CSC000002 with the U.S. Department of Energy, Office of Science, Office of High Energy Physics. This research was partly supported by the Quantum Communication Channels for Fundamental Physics (QCCFP) QuantISED program grant, U.S. Department of Energy, Office of Science, Office of High Energy Physics, DE-SC0019219.
References