Arxiv:1909.02145V3

Home , Lindbladian

arXiv:1909.02145v3 [quant-ph] 26 Aug 2020 h nepnnilgot fteOO—laigto leading OTOC— exponent and the Lyapunov chaoticity of Quantum the between term growth to the connect relation exponential direct [9] an a echo and Loschmidt [6–8], scrambling mation ensemble. thermal over taken ftecmuainrlto ewe prtr ttime square at zero, the operators time to between and referring relation commutation function the correlation of OTOC The 4-point Chaos. a Quantum of is a field of the focus in Out- revival the the recent in [4]— measure is 60s different late (OTOC)— the slightly of-Time-Order-Correlator in A introduced exponent already fi- Lyapunov was limit. the that classical how semi-classical the show the to to in connected [3]) Ref. is exper- in delity motivated echo work computation new (original quantum when theorists in later, interest years and 15 Chaos. iments about Quantum only characterized (fidelity) was 1984, to It protocol in quantity echo analog evolution, Loschmidt an measure quantum as a equivalent unitary suggested an [2] of and Peres lack sep- linear the trajectories the Given by in close time. initially exponen- in two the arate which by gives at measured that rate traditionally tial exponent is Lyapunov sensitivity largest This high the to conditions. [1]. refers initial Mechanics Chaos to Quantum classi- Classical in of definition, appear characteristics By systems the chaotic how cal understand to tempt ∗ [ A [email protected] eicasclapoiain 4 ] unu infor- quantum 5], [4, approximations Semi-classical h edo unu ho a onfo h at- the from born was Chaos Quantum of field The ( t ) B , us-nerbessesaeso otemlz u a eg be may but thermalize to slow are systems Quasi-integrable 0 ] 2 ∼ i aoaor ePyiu ttsiu,Deatmn ephys Département de Statistique, Physique de Laboratoire h aineo h admdie edn osotLauo t Lyapunov short to leading drive, is random (which the of variance the ueoeao qainwihdcae h TCdnmc.W dynamics. space tangent OTOC the se the to a dictates analogy which noise, In equation external case. an superoperator classical by the perturbed in example, weakly nating out-of-tim first are 4-point a the which of As systems evolution Syst the understood. Solar by well defined the exponent, less being example is clear counterparts most tum the – time ergodicity ii h unu ypnvepnn sgvnb h classica the by given is exponent Lyapunov quantum the limit utain,and fluctuations, rsn nartrta skicked is that essent rotor These a explain. in we present which effect quantum purely another htqaiitgal ytm r eaieygo scrambl and good exponent relatively Lyapunov are systems quasi-integrable that S eerhUiest;UiestePrsDdrt Sorbo Diderot, Université Paris University; Research PSL lsia us-nerbessesaekont aeLyapu have to known are systems quasi-integrable Classical h e [ .INTRODUCTION I. 2 A λ ( Q t t ) h eetwv fsuiso the on studies of wave recent The . B , ∼ ǫ 0 − PCUi.Prs0,CR;2 u hmn,705Prs Franc Paris, 75005 Lhomond, rue 24 CNRS; 06, Paris Univ. UPMC ] 2 1 i ). hr h vrg susually is average the where , ii) iii) ntehgl una eieteLauo ntblt ssup is instability Lyapunov the regime quantal highly the in o ucetysalprubtosthe perturbations small sufficiently for kT/ oe Goldfriend Tomer ~ a tyfiiea o temperature low a at finite stay may λ weakly Q whenever , Dtd etme ,2020) 2, September (Dated: but randomly t ∗ n og Kurchan Jorge and us-nerbessesqikyeov oalong-lived state: a initial to some evolve state from prethermalized quickly starting systems system quasi-integrable isolated thermalization the an follow re- to of has that is protocol attention quantum extensive A ceived one. classical the to akin of tori [27]. invariant chain on Toda live integrable passes the that equilibration states to quasi-static route through the where Fermi-Pasta-Ulam- chain— the Tsingou system— paradig- the quasi-integrable for matic illustrated to was This tangent little helps thermalization. mostly hence with and is tori, invariant instability high-dimensional the Lyapunov be- understood be separation the can The time-scales because ergodic another. and the chaotic to tween torus from ergodized dynamics slow one involves process relaxation resulting ytm 1,1](n eeecsteen.Teeset-ups These therein). references many-body concern (and 11] accompanied cold-atom on [10, is controlling bound for systems interest quantum abilities theoretical experimental a This new of with from [5]. the theo- growth in holographic its holes, of black context for the ries in initiated was OTOC in omaueteOO eesgetd[62]and [16–22] suggested realiza- were [23–25]. OTOC OTOC experimental preformed the many- the Several measure closed of to in tions systems. behavior processes quantum different the [15]. to body study chaotic relevance to or its and one 14], allow [13, They near-integrable [12], tegrable xml steSlrSse,wihhsaLauo time Lyapunov a well-known has which A System, of Solar latter. the the is phase-space with example the associated than time smaller much diffusion is charac- former time the Lyapunov terizing the be- slowly, chaotic thermalize strong but havior exhibit systems quasi-integrable) (or, ocrigqatmlmt ncas efind we chaos, on limits quantum Concerning . h eairo unu us-nerbessesis systems quasi-integrable quantum of behavior The nCasclMcais aybd near-integrable many-body Mechanics, Classical In ∼ r ntesneta h ai ewe the between ratio the that sense the in ers qed l de ique ysadsaiiytm of time stability and Myrs 5 -re orltr(TC,o integrable of (OTOC), correlator e-order n ai-i´;Sron Universités, Paris-Cité; Sorbonne nne ncasclsses edrv linear a derive we systems, classical in a etrso h rbe r already are problem the of features ial m–bttestainfrterquan- their for situation the but – em ǫ mscmae otedffso time diffusion the to compared imes 1 n:i clsas scales it one: l eeaieteqatmLyapunov quantum the examine we o ie uhsotrta their than shorter much times nov / tn hthspoe ob illumi- be to proven has that tting 3 n that find e ENS, eednei losprse – suppressed also is dependence ´ T . cl oml Supérieure, Normale Ecole ´ isolated eemndb h qatm quasi- (quantum) the by determined i) ntesemi-classical the in rse yquantum by pressed ǫ 1 ytm,wihmgtb in- be might which systems, / o scramblers ood 3 e with , > ǫ ys[6.The [26]. Gyrs 5 being 2 constants of motion, followed by a slow relaxation to equi- quantum model for quasi-integrable systems in Sec. IV. librium [13, 14]. In general, ‘thermalization time’ refers Then, Sec. V summarizes the solution to the classi- to the time it takes for a wavepacket to explore sequen- cal model [28], and is followed by Sec. VI, in which tially space with equilibrium probability, irrespective of we outline our main results concerning the quantum its size. When we refer to ‘scrambling’ / Ehrenfest time, model. In Sec. VII we derive the basic equations for we mean the time it takes for the packet to be spread obtaining the quantum Lyapunov exponent, and solve over all accessible space at each time. Clearly, the latter explicitly the semi-classical case described by the Bohr- may be infinite in a strictly classical situation, while the Sommerfeld quantization (Sec. VII D). Numerical simu- former is typically finite, even classically. In this paper, lations demonstrating the analytical predictions are pre- we apply the same difference to ‘prethermalization’ and sented in Sec. VIII. Finally, we discuss the results in ‘prescrambling’, where the space in question is the torus Secs. IX and X, where the former focuses on the implica- of prethermalizad Hilbert space. tions for the quantum bound on chaos. Technical details For the classical problem, the essence of the dynamics are given in four Appendices. of quasi-integrable systems may be understood with an analytically much simpler example: an integrable system which is weakly perturbed by an external noise [28] – ar- II. NOTATIONS guably, a stochastic drive can simulate the effect of the many-body integrability-breaking interactions. In par- In the next Sections we discuss classical along quantum ticular, Ref. [28] showed that the randomly driven sys- models. The mathematical language of the latter consists tem develops chaos that is almost tangent to the invari- of different objects and operations, such as matrices and ant high-dimensional torus, with a Lyapunov time which tensors. To facilitate the reading we now specify the is much smaller than the diffusion time. In addition, different notations that we use throughout the paper: chaos appears for any magnitude of the noise (no regular islands in phase-space for any value of perturbation), a) Superoperators– which operates on matrices and thus the stochastic model can mimic the behavior of clas- return matrices— are denoted by calligraphic let- sical quasi-integrable systems beyond the Kolmogorov- ters, e.g., . J Arnold-Moser (KAM) regime. It also does not contradict the KAM theorem, as an external random drive can be b) The superoperators (tensors of rank 4) act on op- thought as coupling to an infinite set of oscillators with erators (matrices) according to the following defi- all frequencies [29], thus ‘resonating with everything’. nition and notation In the current Paper the ideas described in the last ( O)nn′ nn′n n On n . paragraph are extended to quantum systems. We study F ⊙ ≡ F 1 2 2 1 n ,n in detail chaos in quantum integrable systems which X2 1 are weakly perturbed by (classical) noise. The initial conditions we have in mind will be linear combinations c) Matrix multiplication operates as usual of states α cα α having a set of quantum numbers 1 n | 1i n X X + Y Iα, ..., Iα Io , ..., Io that correspond to approxi- A⊙ B⊙ = A⊙ B⊙ { }∼P { } Y X + Y mately equal values of all the constants of motion – C⊙ D⊙ C⊙ D⊙ the quantum analogue of starting ‘near a torus’. Time evolution will dephase these contributions, even before d) We work in the Heisenberg picture where operators the breaking of integrability makes the amplitude′ norms are time-dependent. The subscript 0 refers to the iψα iψ change appreciably, i.e. cα e cα e α . The role of initial value. ‘chaos on the torus’ for classical| | systems→ | | is now played by ‘dephasing at constant cα ’ . Remarkably, the noise term has a strong effect on dephasing,| | even before the quan- III. QUANTUM LYAPUNOV EXPONENT AND tum numbers have changed substantially. An important THE QUANTUM TANGENT SPACE timescale in the present paper is the scrambling-on-the-′ iψ torus time— the prescrambling time— at which the e α In Classical Mechanics, the basic measure of chaos is are essentially random, but the c have not yet diffused: | α| the divergence of two initially close trajectories in the this is the time it takes for a wavepacket to cover the phase space, x (q, p). An exponential separation— whole torus. For these timescales, we shall focus on the defining the Lyapunov≡ exponent— signifies chaos. The evolution of the OTOC, and on the quantum bounds on standard procedure to calculate Lyapunov exponents is chaos. by considering the tangent space, which describes the The paper is organized as follows: First, we set the evolution of the distance between a pair of infinitesimally framework for the analysis by indicating notations in separated trajectories u x(t) x′(t). It is dictated Sec. II; defining the quantum Lyapunov exponent to- by the linear relation u˙ =≡M(x(t−))u, where the matrix gether with the quantum tangent space in Sec. III; M contains the second derivatives of the Hamiltonian and presenting our general stochastic one-dimensional evaluated along a reference trajectory x(t) in phase-space 3

(see e.g., Appendix A). The largest Lyapunov exponent H = p2 + q4, then we get is then defined by [[A, H],B ] = [[p,p2 + q4], q ]= 4i~[q3, q ]= 2 0 0 − 0 1 u 2 2 λcl lim lim ln || || = 4i~ [q, q0]q + q[q, q0]q + q [q, q0] , (6) ≡ t→∞ ||u(0)||→0 2t u(0) 2 − ||′ ′ || ′ ′ R t M(t )dt T R t M(t )dt λmax ( e ) e Note that the initial conditions for the ODE in Eq. (4) lim ln T T , (1) is ([A0,B0], 0), however as in the classical case we can t→∞ h i  2t  also consider commutators of the form [ , A ], for which · 0   we have the initial condition (0, [B0, A0]). Therefore, in where λ [A] is the maximal eigenvalue of A, and max T principle, it is sufficient to consider the above matrix denotes time ordering. of superoperators for general initial conditions, although Because we will be interested in the relation with quan- one should bare in mind that the magnitude of these must tum mechanics, it is natural to define the tangent space be bounded, as, e.g., [p0, q0] = ~. dynamics with Poisson brackets. They satisfy a chain As discussed in the| Introduction,| a quantum Lyapunov 2 2λQt rule: for any pair of conjugate variables, (p, q) and some exponent λQ can be defined via [A(t),B0] e . function F (q,p) we have In the current paper we mostly focush on a microcanon-i ∼ 2 ical version, taking ψ0 [A(t),B0] ψ0 with an initial F,p0 = p,p0 F, q + q,p0 F,p , (2) h | | i { } −{ }{ } { }{ } eigenfunction ψ0 . The canonical version is discussed | i where (q0,p0) corresponds to the value at time zero. In in Sec. IX. Eq. (2) we also make use of the fact that the the Poisson Equation (4) gives us a convenient framework to ex- brackets are canonical invariants; see explicit derivation plore the growth of the OTOC, and its analogy to the in Appendix A. Based on this chain rule, from the Hamil- classical Lyapunov separation. Ideally, we should com- ton equations one finds pute the average of the logarithm of the squared commutator. For simplicity, more often the average of the d p, q H,p , q H,p ,p p, q { 0} = {{ } } −{{ } } { 0} . squared commutator itself is computed, which thus con- dt q, q0 H, q , q H, q ,p q, q0 { } {{ } } −{{ } } { } stitutes an annealed average: this is what we shall do (3) in this paper. We focus on a general class of models: The matrix that appears in Eq. (3) is exactly M(t) that Integrable systems which are weakly perturbed by an ex- governs the dynamics of the displacement vector u in ternal noise. the tangent space. The initial condition for Eq. (3) is the vector (0, 1). One should also consider the other set of − Poisson brackets ,p0 . However, these are decoupled IV. ONE DIMENSIONAL MODEL from the set , q {· and} satisfy the same linear relation {· 0} with an initial condition (1, 0). Therefore it is sufficient Our goal is to understand the Lyapunov exponent of a to study the dynamics dictated by the matrix M(t) for quantum integrable Hamiltonian, Hint, which is weakly any initial condition. perturbed by additive noise. It turns out that the mech- Let us now turn to quantum mechanics. It is natural to anism whereby chaos is induced by noise is already well implement quantization by replacing Poisson brackets by ~ represented by a system of one degree of freedom. Trans- commutators , i [ , ], but now we have to take care forming to action-angle variables it reads: of factor orderings{· ·}. → One· thus· finds that the dynamics of an OTOC is dictated by a linear relation iΘ 1/2 iΘ H(N,e )= Hint(N)+ ǫ η(t)G(N,e ), (7) d [A, B ] [A, B ] i~ 0 = K1⊙ K2⊙ 0 , (4) where the action-like operator N counts the energy level dt [B,B0] 3 4 [B,B0] iΘ K ⊙ K ⊙ number of Hint, the operator e satisfies the commuta- iΘ iΘ where, for example, 1 and 2 come from [[A, H],B0]. tion relation [N,e ] = e , and η is a Gaussian white Equation (4) may beK seen asK a Lindbladian expression noise. Working with the operator eiΘ = cosΘ+ i sin Θ for the tangent-space evolution. In order to prove the allows us to easily relate the quantum problem to the relations playing the role of the chain-rule for the com- classical action-angle variables, while avoiding an explicit mutators, we use the fact that for any analytic function use of the problematic phase-opertor Θ. The operator r iΘ g(A) = dnA we have [g(A),B] = g [A, B], with e itself suffers from some perplexing properties [30, 31] n S ⊙ iΘ −iΘ −iΘ iΘ the superoperator g acting as such as e e = e e , but this will pose no prob- P S lem. Apart from6 the commutation relation stated above, ∞ n imΘ r−1 n−r we have e n = n + m . Thus, alternatively one can [X]= d A XA . (5) | i | i † Sg ⊙ n work with the more familiar ladder operator a , where n=0 r=1 X X the number operator reads N = a†a. The superoperation is just a combination of left and right The model in Eq. (7) may involve any functional form matrix multiplications, see Appendix B 1 for more de- for Hint and G. Nevertheless, in this paper we start with tails. As an example, if we take A = p, B = q and a concrete class of classical models— a particle in a power 4 potential weakly perturbed by a random field— and its H(N,eiΘ, η(t)), and then backward from T to 2T with quantum counterpart. The classical Hamiltonian reads: H(N,eiΘ, η(2T t)) + δH, where δH is some small −perturbation, and− the noise for the backward evolution p2 is the time-reversed of the one for the forward one. Next H = H + ǫ1/2qη(t) , H = + αqν , cl cl,int cl,int 2m we summarize the chaotic behavior of the classical model. (8) where m is the mass of the particle and 0 <ν< . Be- ∞ ing one-dimensional, Hcl,int is integrable, and the action V. SUMMARY OF THE CLASSICAL CASE coordinate can be calculated explicitly [32], giving The Lyapunov exponent of a general classical inte- γ 1/2 iΘ classical : Hcl = KI + ǫ q(I,e )η(t), (9) grable model perturbed by noise was derived in Ref. [28]. We now briefly present the analysis of this derivation as where γ = 2ν/(2 + ν), q(I,eiΘ) I2/(2+ν), and K is a applied to the Hamiltonian in Eq. (9). The different steps ∝ function of m, α and ν given in Eq. (C3). The form of the of the derivation shall be followed as closely as possible quantum version can be deduced directly from of Eq. (8) when we treat the quantum case in Sec. VII. by using scaling arguments: we look for a rescaling of A. Classical tangent space dynamics momentum p bp, such that coordinate rescales as q ~ → → ( /b)q. We find (see Appendix C) The motion in the tangent space is dictated by the Langevin dynamics iΘ quantum : H = ω0~H˜ (N,e )= 1/2 iΘ d uI 0 0 1/2 uI ω0~ H˜int(N)+˜ǫ q˜(N,e )η(t) , (10) = ′′ + ǫ η(t)K(t) dt uΘ H 0 uΘ ≈ cl,int h i 1/2 2 iΘ 1− γ − γ γ−1 0 ǫ ∂Θq(I0,e ) η(t) uI where ω0 = α 2 m 2 ~ has units of frequency, and ′′ , (11) ≈ H (I0) 0 uΘ ǫ˜ = ǫ/(m1−γα2−γ ~2γ−1) is adimensional. In Eq. (10) we cl,int 1/2 have also rescaled time by ω0, η(t) ω η(t), such that where ′ refers to derivative with respect to the action vari- → 0 all the quantities in the square brackets are separately a- able I, and K(t) is a matrix which depends on the second dimensional. derivatives of the perturbation. The structure of the final The Bohr-Sommerfeld approximation of the time- matrix results from power-counting in ǫ, after assuming independent part can be inferred by inserting the quanti- that the matrix is evaluated along unperturbed reference zation relation I = ~N in Eq. (9), which gives H˜int(N)= trajectory (along which the tangent space is measured) N γ andq ˜ N µ with µ = (2 γ)/2 [32][33]. This approx- I(t)= I , Θ(t)=Θ +H′ (I )t. An important remark: ∝ − 0 0 cl,int 0 imation is used to derive the semi-classical Lyapunov in at this point one can scale out ǫ by rescaling uI uI , −1/3 −1/3 → Sec. VII D. uΘ ǫ uΘ and t ǫ t, and thus immediately → → 1/3 The quantum Lyapunov exponent should be under- reach the conclusion that λcl ǫ . However, we shall stood for two copies that evolve and ‘feel’ the same noise continue without this rescaling∝ since it is not crucial for realization. In other words, if alternatively we consid- the purpose of this section. ered a Loschmidt echo, then we would do the follow- The Fokker-Planck equation, describing the evolution ing [9]: evolve the system forward up to time T with of the probability distribution of (uI ,uΘ), reads

∂P (u ,u ) ∂ ∂2 I Θ = H′′ u + ǫ ∂2 q(I ,eiΘ) u2 P (u ,u ). (12) ∂t − cl,int I ∂u Θ 0 Θ ∂u2 I Θ Θ I This equation is homogeneous , thanks to this we can derive a close set of ODEs describing averages over the noise of quadratic quantities:

2 2 iΘ 2 2 uI 0 ǫ ∂Θq(I0,e ) 0 uI d h 2 i ′′ h 2 i uΘ = 0 0 2Hcl,0 uΘ . (13) dt  h i   ′′   h i  uI uΘ H 0 0 uI uΘ h i cl,0 h i      

′ This set of equations describes the evolution of the an- time, λcl Hcl,int(I0), one can average over the angle nealed Lyapunov exponent. Under the assumption that Θ, deﬁning≪ the important parameter for what follows rotation around the torus is faster than the Lyapunov q¯ (∂2 q(I,eiΘ))2 . (14) ≡ h Θ iΘ q 5

Replacing the term by its average may be understood as where the inequality comes from the assumption of weak a first term in a Magnus expansion (see Appendix D 2). perturbation in Eq. (16). Thus, we confirm that small The resulting 3 3 eigenvalue problem gives the annealed perturbation that corresponds to very slow diffusion re- Lyapunov × sults in a Lyapunov separation with small projection along the action coordinates. All of these results are 2/3 1/3 ′′ 2/3 2/3 2λcl =2 ǫ Hcl,int q¯ (I0). (15) verified numerically in Sec. VIII, see Fig. 2. The above derivation relies on the assumption of weak perturbation— the diffusion time for action variables is VI. OUTLINE OF MAIN RESULTS FOR THE shorter than the Lyapunov time QUANTUM CASE

2 −1 I0 Let us now consider the quantum model in Eq. (10). λcl 2 . (16) ≪ ǫq¯ (I0) The model depends on two a-dimensional numbers: the typical energy level n0 and the adimensional perturbation 1/3 Since λcl ǫ , this inequality holds for small enough ǫ. strengthǫ ˜. In the current section we provide general ∼ 1/3 The λcl ǫ scaling was already found in the context arguments for the behavior of the quantum model, and ∝ of motion along a stochastic magnetic ﬁeld [34], and in outline our main results with a scheme in Fig. 1. the theory of products of random matrices [35, 36].

A. Quantum tangent space dynamics B. Classical prescrambling time As we show in Sec. VII, it is sufficient to focus on Θ iΘ −iΘ In order to better understand the influence of the quan- the dynamics of two operators, C = [e , A0]e and N tum dispersion of the initial condition, we need to see C = i[N, A0], with A0 being some initial Hermitian first what happens classically when the initial separation operator. The time derivative of these operators follows of trajectories is finite. Let us thus consider two initial a linear super-operator equations, which are the analogue nearby trajectories at a non-infinitesimal initial separa- of those of Eq. (11), of the form tion (uI,0,uΘ,0). There is an initial time window, [0,tb], ˙ N N within which the uI stays small, while their angular sepa- C 0 0 1/2 C ′′ Θ = +˜ǫ η(t) M⊙ F⊙ Θ , ration grows ballistically uΘ(tb) uΘ,0 +H (I0)uI,0tb. C˙ i C ≈ cl,int L⊙ J⊙ N⊙ K⊙ This is followed by the Lyapunov regime, where the two (20) trajectories separate exponentially at a rate λcl. In gen- where , , , , , are super-operators easily ob- eral, it is expected that the exponential growth starts tained byL J usingM theF N ‘chain-rule’K for commutators. −1 after one Lyapunov time tb = λcl (we confirm this nu- The quantum Lyapunov exponent is adimensionalized λQ merically in Sec. VIII), and is expected to saturate af- as: λ˜Q = , with λ˜Q defined by the (adimensional) time ω0 ter some finite time, when the separation has grown to dependence of the OTOC associated with the evolution the size of the torus. Na¨ıvely, one would think that the generated by the adimensional Hamiltonian H/(ω0~): saturation time ts may be estimated in the usual way, λ t ˜ as the time u e cl s 2π. However, this is not quite 2 2 2λQt Θ,0 C (t) ψ0 CΘ ψ0 e . (21) true. What happens is∼ that it is not the initial time sep- ≡ h | | i∼ aration that is amplified by the exponential separation, The initial state, ψ = c n , is assumed to be con- | 0i n n| i but rather the separation after the ballistic regime, i.e. centrated around some n0, here we take directly ψ0 = −1 P | i uΘ(tb) uΘ(λcl ) we obtained above. We hence have: n0 . ∼ | Thei super-operator comes from factor reorderings, 2π 2π and thus vanishes in theJ classical case. When this is so, t λ−1 log λ−1 log (17) s cl cl −1 we may rescale CN CN , CΘ ˜ǫ−1/3CΘ and t ≡ uΘ(tb) ∼ uΘ(λcl ) ˜ǫ−1/3t, and conclude that→ , , →may be neglected for→ M N K Finally, let us check that during these times the diffu- smallǫ ˜, and that the Lyapunov exponent λ˜Q scales like sion of the action is small. At the saturation time ˜ǫ1/3. In the quantum case = 0, and time cannot be J 6 −1 rescaled withǫ ˜. 2πuI (λ ) −1 λclts cl As in the classical case, we assume weak perturbation, uI (ts)= uI (λcl )e = −1 . (18) uΘ(λcl ) such that there is a negligible diffusion of energy levels during one Lyapunov time, td tLyp. In adimensional −1 ≫ Plugging in the expression for uΘ(λcl ), and the solution units this reads: for λcl from Eq. (15) we find 2 3/2 ′′ ˜ǫq (n0) n0 H (n0), (22) 1/3 ≪ 2π ǫq¯2 u (t )= = π I , (19) where the energy diffusion rate is approximated hereafter I s ′′ −1 ′′ 0 2 2 Hcl,int(I0)λcl Hcl,int ! ≪ as n0 ǫ˜q (n0)td, and q (n0) is defined in Eq. (14). For ∼ 6 large values of ǫ, the ǫ1/3 scaling breaks down, e.g., [37, 38].

B. Quantum prescrambling time

Let us now see how what we have learned about saturation times in classical case (Sec. V B) affects the quantum picture. Following Refs. [39, 40], one shall imagine the semiclassical spreading of an initial wavepacket. The size of the packet prior to the exponential growth includes an initial ballistic regime, ℓ(t ) = ℓ + v t ℓ + v /λ . b 0 0 b ∼ 0 0 cl Now, the uncertainty principle implies that both ℓ0 and v0 are finite, for example considering a coherent state as an initial packet. Quantum effects thus saturate the exponential growth once the wavepacket spreads through- FIG. 1. A scheme summarizing our findings for the OTOC out the torus, λclt log(ℓ0 + v0/λcl). dynamics in quantum quasi-integrable systems described by Quantum mechanics∼− thus acts in two forms: for large Eq. (10). The analysis assumes weak perturbation, Eq. (22), Lyapunov exponents it is the initial wavepacket that thus excluding the blue regime. When J in Eq. (20) is negli- 1 3 spreads, just as in the usual Ehrenfest time estimate gible, a Lyapunov exponent which scales asǫ ˜ / is predicted – only that here it concerns spreading over the prether- based on scaling arguments (white area; this is verified in malization space (the quantum counterpart of the torus) Figs. 5 and 6). A semiclassical derivation based on the BS for rather than over the entire phase-space. For small Lya- n0 large enough is given in Sec. VII D. From the discussion on the prescrambling times in Sec. VI, we can predict that punov exponents, as we shall have when the perturba- the Lyapunov regime vanishes at low quantum numbers (red tion is weak, the ballistic time is long, and hence the area; supported by Fig. 5a and 3), and at very low perturba- quantum spread of initial velocities could even reach pre- tion (green area; supported by Fig. 5b and 6). scrambling. These two effects thus limit separately the conditions under which there is a Lyapunov time at all: if the prescrambling time is of the order of the (classical) regime as the time during which a packet has not spread. Lyapunov time, then the Lyapunov regime is finished be- It should be born in mind, however, that n has oscilla- fore it starts. | 0i tions in Θ of length 1/n0, and these dephase completely in times similar to that of a wave packet (see inset of Fig. 3). Finally, the implications of our results for ther- C. Overview of the results mal averages are discussed in the last section, Sec. IX. The current section is organized as follows: The A summary of our main conclusions is given schemat- equations for the tangent space dynamics and the an- ically in Fig. 1. The different regimes are based on the nealed Lyapunov exponent are derived in Sec. VII A and physical arguments presented in the current section, the Sec. VIIB respectively. Then, in Sec. VIIC we treat fur- complete analysis in Sec. VII, and its verification with ther a slightly simplified model and solve explicitly the numerical simulations in Sec. VIII. We indicate explic- semi-classical limit in Sec. VII D. itly what is the numerical evidence/analytical derivation from which we conclude each part of the diagram. We now move to the calculation of the quantum Lya- punov exponent. A. Evolution of the OTOCs

We focus on the dynamics of two operators VII. ANALYTICAL DESCRIPTION FOR THE Θ iΘ −iΘ N OTOCS DYNAMICS AND THE QUANTUM C [e , A0]e , C i[N, A0], (23) LYAPUNOV EXPONENT ≡ ≡

with A0 being some initial Hermitian operator. The The goal of this Section is to calculate the growth rate choice of normalization of the ﬁrst commutator with e−iΘ of the square of a commutator, ψ C2 ψ . The most nat- is analogous to the classical counterpart, there it compen- ural thing is to consider a linearh | combination| i of states sates for the fact that we are working with non-canonical iΘ around some n0 , corresponding to a wavepacket in the variables N and e (see Appendix A). The normaliza- angular variables.| i In fact, we shall use a single state tion also guarantees that CΘ is Hermitian. We ﬁnd that ψ = n0 , and check that the results correspond to those this choice of OTOCs facilitates the analytic derivation, |ofi a packet.| i The state n is spread over all angles, which however, the same dynamics is expected for other oper- | 0i seems at odds with the interpretation of the Lyapunov ators such as [cosΘ, A0]. 7

Let us look at the time derivatives of the above two 1. Integrable case operators (recall that time is rescaled by ω0): When there is no external noise, ˜ǫ = 0, the dynamics C˙ N = [[N, H˜ (N)+˜ǫ1/2q˜(N,eiΘ)η(t)], A ] O(˜ǫ1/2), int 0 ∼ of the OTOC follows ˙ Θ iΘ ˜ −iΘ iΘ −iΘ ˜ C = i[[e , H], A0]e i[e , A0][e , H]. (24) C˙ N 0 0 CN − − = . (32) C˙ Θ i CΘ The first term on the right hand side of Eq. (24) can be L⊙ J⊙ written as The above dynamics cannot yield an exponential growth for the OTOC. Only CΘ may grow exponentially, but iΘ −iΘ iΘ −iΘ Θ −iΩt Θ iΩt i[[e , H˜ ]e , A0]+ i[e , H˜ ][e , A0] since i = i[Ω(N), ] we get C (t) = e C e , an − J · 0 = i[[eiΘ, H˜ ]e−iΘ, A ] i[eiΘ, H˜ ]e−iΘCΘ, (25) oscillatory term. − 0 − where we have applied the relations [A, C]B = [AB, C] B. Annealed Lyapunov exponent A[B, C] and [e−iΘ, ] = e−iΘ[eiΘ, ]e−iΘ. In addition,− using this last relation,· the− second term· in Eq. (24) can be rewritten as The derivation which follows is done along the lines of the classical problem that was addressed in Ref. [28] and iΘ −iΘ Θ iΘ −iΘ briefly discussed in Sec. V. We can do this since the for- i[e , A0][e , H˜ ]= iC [e , H˜ ]e . (26) − mulation of the two problems is the same, the quantum In summary we have case is just in higher number (infinite) of degrees of freedom: the variables Cnn′ can be thought as vectors and Θ Θ 1/2 accordingly the superoperators can be thought as matri- C˙ = i[Ω(N), A0]+ i[Ω(N), C ]+ O(˜ǫ ), (27) ces. Nevertheless, there should be differences which come from quantum mechanics. where Ω(N) [eiΘ, H˜ (N)]e−iΘ = H˜ (N) ≡ − int int − The matrix of superoperators in Eq. (28) has elements H˜ (N 1). iΘ int − that depend on time through the operators N and e , The corresponding equations, which are equivalent to which evolve according to the full perturbed Hamilto- those of the tangent space in classical mechanics, are of nian in the Heisenberg picture. However, we can assume the form that the evolution is well approximated by the evolution unperturbed by noise, and the effect of noise is only im- C˙ N 0 0 CN = +˜ǫ1/2η(t) M⊙ F⊙ . portant at the level of the tangent space. Note that the C˙ Θ i CΘ L⊙ J⊙ N⊙ K⊙ same approximation is assumed in the classical case. For (28) small ˜ǫ (Eq. (22)) the perturbation then gives only small Let us now write explicitly the operators and , J L corrections and we have: which depend only on the integrable part H˜int. We work iΘ iH˜int(N0)t iΘ0 −iH˜int(N0)t in the eigenbasis n and denote the period of the torus (N,e )= (N0,e e e )+ O(˜ǫ). | i F F (33) ωn En En−1, (29) Now, we can employ power-counting in ˜ǫ to elimi- ≡ − nate several components in the matrix of superopera- with E the energy levels of the integrable model. tors. If we assume the scaling t ǫ˜−αt, CN CN n → → From Eq. (27) we have for : and CΘ ǫ˜−βCΘ, together with the fact that for white • J noise η(at→)= a−1/2η(t), we find ′ ( C)nn′ = i(ωn ωn′ )Cnn′ i j(n,n )Cnn′ , (30) J ⊙ − ≡ C˙ N =˜ǫ1/2−αη(t) CN +˜ǫ1/2−α−βη(t) CΘ, (34) M⊙ F ⊙ The superoperator can be represented as a sum • L of left and right matrix multiplication, using the rela- ˙ Θ −α+β 1/2−α+β N s s s−1 s−r C = ǫ˜ +˜ǫ η(t) C + tion [N , A0] = r=1 N [N, A0]N , alternatively, L N ⊙ h i we can use the relation in Eq. (B3) which leads to: iǫ˜−α +˜ǫ1/2−αη(t) CΘ. (35) P J K ⊙ ′ ωn ωn ′ h i ( C) ′ = − C ′ l(n,n )C ′ . (31) Then,ǫ ˜ 1 implies that can be neglected with re- L⊙ nn n n′ nn ≡ nn ≪ M − spect to , as well as compared to , and compared to . WeF have then: N L K In the semiclassical limit, when the density of levels is J ′ ′ 2 2 high, n n 1, one has that l(n,n ) ∂ H/∂n (see ˙ N 1/2 N ∼ ≫ → C 0ǫ ˜ η(t) (t) C = F ⊙ Θ . (36) Sec. VII D). C˙ Θ i C Before we proceed with the equations for the annealed L⊙ J⊙ Lyapunov exponent, we briefly discuss the integrable We keep both and , the latter is a factor-ordering case. term that disappearsL inJ the classical case. 8

This is a Langevin equation satisﬁed by the compo- ﬁed by the components of the commutator (whose com- nents of the commutator. Next, we may repeat the plete form is given in Eq. (D1)). Using the homogeneity steps we followed in the classical case, deducing from the in the same way, we obtain a closed equation for the Langevin equation (36) a Fokker-Planck equation satis- quadratic averages of components:

d N N Θ Θ C ′ C ′ =˜ǫ ′ (t) ′ (t) C C , dth nn mm i Fnn n3n4 Fmm m3m4 h n4n3 m4m3 i n3,n4 mX3,m4 d Θ Θ ′ N Θ ′ Θ N ′ ′ Θ Θ C ′ C ′ = l(n,n ) C ′ C ′ + l(m,m ) C ′ C ′ + i (j(n,n )+ j(m,m )) C ′ C ′ , dth nn mm i h nn mm i h nn mm i h nn mm i d N Θ ′ N N ′ N Θ C ′ C ′ = l(m,m ) C ′ C ′ + i j(m,m ) C ′ C ′ , (37) dth nn mm i h nn mm i h nn mm i

where the functions j(n,n′) and l(n,n′)= l(n′,n) appear Next, we write down a set of ODEs for the dynam- XY in Eqs. (30) and (31). This is a closed set of ODEs for ics of averages, according to Eq.(37). Defining Fnn′ X Y ≡ the averaged products of matrix elements. One can take Cnn′ Cn′n , with X and Y being N or Θ, we find: a subset of these equations and hope that they will form h i a closed set. Next we consider a simplified model for d NN Θ Θ F ′ = ǫ˜ nn′m m n n (t) C C , (40) which this can be done. In particular, we are interested in dt nn R 1 2 1 2 h n2 n1 m2m1 i n1,n2 quantities of the form C C which correspond mX1,m2 n1 nn1 n1n to evaluating expectationsh of square-commutatorsi at a P given eigenstate n . NΘ ΘN | i d ΘΘ ′ Fnn′ + Fnn′ F ′ =2l(n,n ) , (41) dt nn 2 C. Simplified model

NΘ ΘN d Fnn′ + Fnn′ ′ NN We now simplify the problem by employing = l(n,n )F ′ dt 2 nn q˜(N,eiΘ) = V (N)cosΘ + cosΘV (N), a particular NΘ ΘN case of the general model in Eq. (10). The main charac- F ′ F ′ ij(n,n′) nn − nn , teristics of the solution should hold for other functional − 2 forms, that is, taking higher harmonics cos(2Θ), sin(Θ), (42) etc. We thus treat the a-dimensional Hamiltonian:

˜ ˜ 1/2 NΘ ΘN NΘ NΘ H = Hint(N)+˜ǫ η(t) (V (N)cos(Θ) + cos(Θ)V (N)) . d F ′ F ′ F ′ + F ′ nn − nn = ij(n,n′) nn nn , (43) (38) dt 2 − 2 The OTOCs dynamics is dictated by Eq. (36) that contains the superoperators , , and . The former two where ′ ′ ′ (t) ′ ′ (t) ′ ′ (t), and we use L J F nn kk ll nn ll n nkk are related to the integrable part and already given in the factR that j(n,n≡′) F = j(Fn′,n) and l(n,n′) = Sec. VII A. l(n′,n). Finally, we preform− time-averaging for (see Let us calculate the superoperator , which corre- Appendix D 2), that is, we drop all of the oscillat-R Θ F sponds to the term proportional to C in the operation ing components. As in the classical case, this pro- [[N, V (N)cosΘ + cosΘV (N)], A0]. We shall thus con- cedure refers to the assumption that the Lyapunov sider time is much longer than the periods around the ′ Θ Θ torus. We ﬁnd nn m1m2n1n2 (t) Cn n Cm m = 1 Θ iΘ −iΘ Θ 2 1 2 1 1 ′ 2 ΘΘR ΘΘ h i [[N, cos Θ], A0]= (C e + e C ). (V (n)+ V (n )) F ′ + F ′ , and thus Eq. (40) 2 4 Pn,n +1 n+1,n can be replaced by Working in the eigenbasis of N, where m eiΘ n = iω t −iΘ −iω t h | | i e m δ and m e n = e n δ , we ﬁnd d NN ǫ˜ ′ 2 ΘΘ ΘΘ m,n+1 m,n−1 F ′ = (V (n)+ V (n )) F ′ + F ′ . that h | | i dt nn 4 n,n +1 n+1,n (44) XY iω ′ t −iω t This closes the equations for the F functions. ′ ′ = e n δ ′ δ ′ + e n δ ′ δ ′ Fnn kk n,k n +1,k n+1,k n ,k Following the classical case, the closed set of equations 1 (V (n)+ V (n′)) . (39) may be transformed to an equation for a single commu- × 2 tator: 9

3 d ΘΘ 2 ′ ′ 2 d ΘΘ ˜ǫ 2 ′ ′ 2 ΘΘ ΘΘ F ′ +2l (n,n )(n n ) F ′ 2l (n,n )(V (n)+ V (n )) F ′ + F ′ =0, (45) dt3 nn − dt nn − 4 n,n +1 n+1,n where we substitute j2(n,n′) = (n n′)2l2(n,n′). the dimension-full variables, this is equivalent to taking Equation (45) is the key outcome− of the calculation— the limit ~ 0 while fixing the energy (~N)γ ). it describes the growth of the norm of matrix elements First, the→ function l(n,n′) for a given∼n 1 and n′ = ΘΘ Θ 2 ≫ Fnn′ Cnn′ , in a simple model which tries to capture n + Z reads generic≡ properties| | of quasi-integrable systems. This third nγ (n 1)γ (n + Z)γ + (n + Z 1)γ order ODE should be accompanied with initial conditions l(n,n + Z)= − − − − . that correspond to some energy shell, as we describe be- Z low. Eq. (45) should contain (a) the exponential growth Since we are interested in operators which are localized of expectation values within the prescrambling time, and in energy space, we are focusing on the limit of Z n. (b) their spreading along energy levels. One might also Then we find l(n,n+Z)= γ(γ 1)nγ−2, which is simply≪ conjecture that this equation shall give (c) the satura- − ∂2H˜ (n)/∂n2. In addition, the expression nµ + (n + tion, i.e., the prescrambling time. In the current paper int Z)µ that appears in Eq. (45) is approximated as 2nµ. we focus on the exponential growth, providing an explicit Next, let us discuss the term which is proportional to result for the semi-classical limit. The other two dynam- ′ 2 dF (n n ) dt . Whenever this term is negligible, we can ical properties are left for future considerations. − ˜ λQt ˜3 2 2µ We already learn something from Eq. (45): if the term rescale time to find F e with λQ l ǫn˜ . Hence, ∼ ∼ 2 2µ with the first time-derivative were absent, we could ab- we can drop this term self-consistently if λ˜QZ ǫn˜ . 1 ≪ sorbǫ ˜ into time and conclude that we have an ˜ǫ 3 scal- This criterion is equivalent to ing of the Lyapunov exponent, just as in the classical d ˜ case (see remark in Sec. V). The term linear in , orig- l(n,n + Z)Z = ωn ωn+Z λQ. (46) dt | | | − | ≪ inates from the factor reordering induced by CΘ = [Ω(N), CΘ], and is of purely quantum origin. J ⊙ As we discuss for the classical problem in Sec. V and demonstrate explicitly in the next Section: when the inequality in Eq. (46), which depends on the initial condition, is not satisfied we do not expect to see a Lyapunov 1. Initial conditions regime. In summary, the semi-classical limit refers to large en- The definition of CΘ and CN in Eq. (23) concerns ergy n0 and not too small perturbation Z0 the commutation of a time-evolving operator with some 2µ+γ−2→ ∞ ≪ ˜ǫn0 . Inserting those limits into Eq. (45), summing initial Hermitian operator A0. We may choose A0 = ′ ΘΘ Θ 2 over n , and recalling that n′ Fnn′ = n (C ) n = ψ0 ψ0 as a projection on an initial wavepacket concen- Θ 2 h | | i (C )nn, we find trated| ih around| an eigenstate n , or in the extreme case, P | 0i just as n0 n0 . The corresponding commutators at time 3 d 2 t = 0 then| ih read| (CΘ)2 = 2˜ǫ γ(γ 1)nγ−2 n2µ dt3 nn − × Θ N Θ 2 Θ 2 C = n +1 n +1 n n , C =0. (C )nn + (C )n+1,n+1 . (47) 0 | 0 ih 0 | − | 0ih 0| Therefore, the solution to Eq. (45) shall be obtained for Finally, assuming that the initial condition is concen- Θ 2 Θ 2 ΘΘ trated around n0, such that (C ) (C ) , a given initial condition Fnn′ (t = 0) which is zero almost n0+1,n0+1 ∼ n0,n0 everywhere (this is a third order ODE, it has three initial we can find the exponential growth described by Eq. (47) ˜ conditions which are related to F NN and F ΘN through with the ansatz (CΘ)2 e2λQt to find the resulting n0,n0 ∼ Eqs. (41)-(44)). a-dimensional quantum annealed Lyapunov:

2λ˜ =22/3ǫ˜1/3(γ(γ 1)nγ−2)2/3n2µ/3. (48) Q − 0 0 D. Semi-classical Limit: Bohr-Sommerfeld approximation We verify the correspondence between the semiclassical Lyapunov and the classical one in Eq. (15) by putting ˜ ~ The aim of the current Section is to investigate Eq. (45) back the units λQ = ω0λQ. Then, taking n0 = I/ , and in the semi-classical limit, and show that it yields the inserting the deﬁnitions of ω0 andǫ ˜ we have classical Lyapunov exponent in Eq. (15). In this limit we − − − 3 1 2 γ 2 γ γ−2 ˜ γ λ = ǫα 2 m 2 I , have the Bohr-Sommerfeld approximation Hint(N)= N Q 2 and V (N)= N µ, where γ and µ are related to the power- law potential; see discussion after Eq. (10). This should which is exactly the Lyapunov we get for the classical be taken together with the limit n (going back to Hamiltonian in Eq. (8) withq ˜(eiΘ) = 2 cos Θ. → ∞ 10

VIII. NUMERICAL SIMULATIONS OF KICKED with the evolution operator U = Uτ (r3)Uτ (r2)Uτ (r1). SYSTEMS In all the examples below we··· choose a kicking rate which is fixed with respect to the Lyapunov exponent, We now move to verify with numerical simulations all τ 0.01tLyp(n0, ˜ǫ), to guarantee uncorrelated drive and ∼ the theoretical results derived in the previous sections. allow a reasonable number of timesteps for observing a In particular, working in the classical or quantum tan- Lyapunov regime. gent space allows us to derive the asymptotic exponential growth of trajectories separations or the OTOCs, but it says nothing explicitly on the saturation of this A. Classical and quantum weakly, but randomly, divergence— the prescrambling time. The saturation is kicked rotor expected when the angular separation reaches a value of order one, when a phase-space wavepacket would cover The case of γ = 2 (infinite potential well) can be con- the classical torus. Below we study numerical examples sidered as a rotor. The only difference is that for the of the classical and quantum problems. latter, the angular momentum operator N can assume One way to realize the external white noise is to treat negative values. In that case, one should modify the op- a kicked system eration cos Θ to account for negative values as well. We

∞ verified that the latter does not affect the results. η(t)= r δ(t kτ), (49) k − k=−∞ X 1. Classical −1 with some kicking rate τ and where rk are taken from a normal distribution of zero mean and variance τǫ˜. If the As a reference, and demonstration of the theoretical time between kicks is shorter than the unperturbed evo- description presented in Sec. V, we study the analog clas- −1 lution and the Lyapunov time τ ωn tLyp, then the sical problem external drive can be considered≪ as a white≪ noise. Stro- boscopic drive with a constant magnitude corresponds I2 H (I, Θ) = +2η(t)˜ǫ1/2 cosΘ. (52) to fundamental examples in the study of classical and cl 2 quantum chaos. A well-known system is the Standard (Chirikov) map and its quantum equivalent— the Quan- The Hamilton equation yields the random map tum Kicked Rotor [39, 41–44]. The case of random kicking is closely related to It+dt = It +2rt sin Θt, (53) Chirikov ‘Typical Map’ [45, 46], where the magnitudes of Θt+dt =Θt + Itdt (mod 2π), (54) the kicks are given by a set of T random variables which is repeated periodically. Frahm and Shepelyansky [46] where rt is taken from a normal distribution of zero mean studied in detail the classical and quantum version of and variance ˜ǫdt. We integrate this map for pairs of ini- this map. For the classical case they found a Lyapunov tial conditions, one initialized at I0 = 0 and some random exponent that scales asǫ ˜1/3. initial phase Θ0 = α0 and the other is at a distance u0 −8 In what follows we treat the adimensional model from it. We fix the initial norm of u0 = 10 . A pair of such initial conditions is integrated with the same real- H˜ (N,eiΘ)= N γ +2η(t)cosΘ, (50) ization of the noise. In Fig. 2(a) we present separately the quenched evo- where η(t) is given in Eq. (49). This randomly kicked sys- lution of the action separation uI (dashed curves) and (1) (2) tem can be integrated numerically by applying the uni- the angular separation ucos = cosΘ cosΘ (solid −iN γ τ −2ir cosΘ − tary operation Uτ (r) = e e between kicks, curves). The figure shows all the three regimes discussed where r is drawn from a normal distribution of zero in Sec. V: at short times, the angular separation roughly mean and variance ˜ǫτ. We work in the eigenbasis of N, grows in a linear fashion (ballistic regime), whereas the ′ where operations of cosΘ correspond to 2 n cosΘ n = action separation changes little. At later times, an expo- ′ h | | i δn,n′+1 + δn+1,n′ for n,n nonnegative. This can be done nential growth starts in both coordinates and saturates in Fourier space, as long as the system is far from the when uθ O(1). By that time, still uI 1, the more edges n = 0 and n = M, with M the size of the system. so the weaker∼ the perturbation. In Fig. 2(b)≪ we show, by We verified that this is indeed a good approximation, by collapsing the curves with rescaling time, that the rate of exact diagonalization of cos Θ. exponential growth is proportional toǫ ˜1/3, as expected. We consider the micro-canonical OTOC, where the system is initialized with ψ0 = n0 and focus on the 2 | i | i 2 evolution of C (t) = ψ0 [cos Θ(t),N0] ψ0 . For our lo- 2. Quantum calized initial wavefunctionh | we have that| i 2 C2(t)=2 (n n )2 n U † cosΘ U n , (51) Let us now move to the quantum problem. In Fig- − 0 h | 0 | 0i n ure 3 we show the evolution of the OTOC in Eq. (51) for X

0 0

-10 -10

0 -20 -20 -20 -30 -30 -40 (a) (b) 0 100 200 -40 -40 101 103 105 107 0 50 100 150 200

FIG. 2. The separation of two initially close by trajectories for the classical randomly kicked rotor (averaged over 1000 pairs of trajectories). The solid lines corresponds to angular separation whereas the dashed lines indicate the difference in the action variables. Different curves correspond to different noise magnitude,ǫ ˜1/2 outlined in the legend (smaller perturbations corresponds to longer saturation times). The initial separation is fixed. In panel (a) the time axis is in logarithmic scale, 1/3 whereas in panel (b) it is in linear scale and rescaled withǫ ˜ t ∼ λclt.

two different initial conditions with the same perturba- 6 tion strengthǫ ˜. Times are rescaled withǫ ˜1/3. The curves 10 show how the prescrambling time, measured in Lyapunov 10 5 times, decreases with decreasing n0. As explained in the beginning of Sec. VII A, we shall have in mind an ini- 4 tial condition of a wavepacket around an eigenstate n , 10 | 0i rather than strictly the eigenstate. We have thus verified 3 that— starting with ψ = n , at later times the off- 10 | 0i | 0i diagonal terms of the commutator square grow roughly 2 as the diagonal terms (see inset of Fig. 3). Fig. 4 shows 10 how energy diffuses little during the Lyapunov regime, 1 this is equivalent to the small diffusion of the tori in the 10 classical case. It would be interesting to check that a wavepacket in the coherent-state or Wigner representa- 0 2 4 6 8 tion indeed fills the torus in an Ehrenfest time, and diffuses subsequently [27, 28]. In Fig. 5(a) we show the growth of the OTOC for differ- FIG. 3. The OTOC growth for the quantum randomly kicked rotor, for two different initial conditions, n0 = 8191 (blue) ent initial energy levels n0 and fixed relative perturbation 1/2 −3 2 and n0 = 255 (red), and fixed magnitude of the perturbation ǫ =5 10 n0 (In this example we choose to keepǫ/n ˜ 0 1/2 1/3 ǫ˜ = 100. The time is rescaled with ˜ǫ . For smaller n0 the fixed rather× thanǫ ˜ since the former controls the small- Ehrenfest time (indicated by the dotted grey lines), measured ness of the perturbation, as the energy is almost constant in Lyapunov times, is shorter. The inset shows off-diagonal throughout the evolution). The OTOCs show an expo- components of the commutator Cn0,n0+10 in green solid line. nential growth with essentially the classical Lyapunov ex- The growth of the diagonal elements, shifted along the y-axis ponent for a time window— the Lyapunov regime— that for reference, are shown in red dashed line (n0 = 255). roughly starts at one Lyapunov time and ends at the prescrambling time tE. In the inset of Fig. 5(a) we show that the latter is proportional to log n0 Lyapunov times. with decreasing ˜ǫ, and for small enough perturbation it For log n0 1 there is no Lyapunov regime. This is the vanishes. This behavior resembles the one observed for ∼ usual situation for largeǫ ˜. For smallerǫ ˜ the Lyapunov the classical model in Fig. 2. time becomes large, and we must correct the initial size −1/3 ℓq(˜ǫ) ˜ǫ , according to the estimate above (see discussion∼ in Sec. VI). B. Other randomly kicked integrable models We check how the prescrambling time depends on the perturbation strength. In Fig. 6(b) we show the evolu- In Fig. 6 we present results for the case of γ =4/3 and tion of C2(t) for various magnitudes of external noise and γ = 3/2 that correspond to the Bohr-Sommerfeld ap- fixed initial condition. The Lyapunov regime gets shorter proximation of an integrable part with power-potential 12

~ ˜ ~ −βω0 Hint(N) nT(βω0 ) = Tr N e /Z, (56) n o where we put back the energy and time scales, ~ω0 and 0.0091 −1 ω0 respectively. The long time limit in the annealed 0 averaging of Eq. (55) has to be taken with care, or alter- 0.0493 natively, one can make a ‘quenched’ calculation by taking 0 the expectation of the logarithm of the squared commutator. The canonical averaging in Eq. (56) imposes a 0.0755 relation β~ω k(n ), which is a decreasing function of 0 ≡ T 0 nT. Hence, if we evaluate the averaging in Eq. (55) with 0.1830 nT we obtain: 0 β~λ = k(n )λ˜ (n , ǫ˜) g(n , ǫ˜). (57) 8000 8100 8200 8300 8400 T T Q T ≡ T The fact that there should be at all a Lyapunov regime at a given finite value of n0 already implies that the adi- FIG. 4. The spreading of the initial wave-function U(t)|n0i ~ mensional quantity β λT scales as a finite number, and (blue) and hn| cos Θ(t)|n0i (red) that controls the OTOC growth (see Eq. (51)). Initially, the latter is narrower than the the system is a ‘rather good scrambler’ between Lya- former until they meet each other at later times. The initial punov and prescrambling times. We now derive a gen- 13 eigenstate refers to n0 = 2 − 1, where the size of the system eral semiclassical expression for the adimensional func- 14 is M = 2 . The curves are averaged over 76 realizations of tion g(nT, ǫ˜) and show that it grows, for a givenǫ ˜, as nT the noise. decreases: the quantization of nT will provide a bound. The system is then a relatively good scrambler, in the ~ sense that β λT reaches, at low T ,a finite (albeit small) q4 and q6 respectively; see Eq. (8) (the perturbation part ˜ǫ-dependent value. Note that althoughǫ ˜ depends on ~ is taken with µ = 0, as N is roughly fixed during the and ω0, we can change nT independently by varying the Lyapunov regime). Similar to the previous examples, temperature. the figure shows that the quantum Lyapunov exponent For a general Hamiltonian in Eq. (10), the scalings for ˜ 1/3 2(γ−2)/3 the adimensional Lyapunov and the assumption of weak follows the classical one, λQ ǫ˜ n0 . The exponential growth starts after ∼1 Lyapunov time until it perturbation are given respectively according Eq. (15) ∼ and Eq. (22), replacing I n . The adimensional rela- saturates at later time. The figure also illustrates how 0 → 0 the Lyapunov regime vanishes at sufficiently weak per- tion in Eq. (57) then reads turbation according to Eq. (46). The relevant quantity 1/3 γ−2 2 ′′ ∆ωn = γ(γ 1)n is fixed by the initial condition, λ˜(n , ǫ˜) ǫ˜q¯ (n ) n H˜ (n ) 0 − 0 β~λ = T = T T int T . whereas the Lyapunov time increases with decreasing ˜ǫ. T k−1(n ) ˜ ′′ 3 k−1(n ) T H (nT)nT ! T ! As the ratio ∆ω /λ˜ increases the exponential growth int n0 Q (58) saturates earlier. Once this ratio is O(1) we do not ob- Now, the first brackets cannot be too large according to serve an exponential growth. The ballistic∼ regime suffices Eq. (22), and, if we do not want to violate the bound [5], to scramble over the torus, as explained above. then the second brackets must be a decreasing function

of nT. Hence, the most chaotic system corresponds for

nT = O(1), at which the Lyapunov regime also vanishes. IX. THE QUANTUM BOUND ON CHAOS We can verify this explicitly for the BS of H˜int, for which we ﬁnd β~λ =ǫ ˜1/3k(n )n(γ−2)/3 with 0 < γ 2. We T T T ≤ In the current Paper we have focused on a micro- know, however, that when n0 becomes of O(1), the Lya- canonical version of the OTOC, namely, the expecta- punov regime shrinks to zero. Θ 2 In a many-body system, the mechanism for the quan- tion value of n0 (C ) n0 . Recently, it was shown that the quantumh Lyapunov| | exponenti (deﬁned by the growth tum bound may be hence understood as follows: consider a system consisting of M copies of our integrable model, rate of the OTOC) is bounded in thermal systems as (1) (M) (i) β~λ 2π [5]. We shall argue that, at least in our having values of ω0 = ω0 > ... > ω0 0, with ω0 T ∼ model,≤ the quantum limitation to chaos is imposed by spanning an interval that goes down to zero. The system blocking one by one the Lyapunov regimes of the degrees is at temperature T , so that the corresponding average (1) (M) of freedom that would yield the largest Lyapunov diver- quantum numbers are nT < ... < nT . The coupling in- gencies. troduces perturbations withǫ ˜(i). Importantly, the global Let us start by performing a canonical averaging: Lyapunov exponent is dominated by the largest of indi- vidual ones. Consider then choosing the adimensionalǫ ˜(i) =ǫ ˜(1), 1 2 −βω0~ H˜int(N) λT = ln Tr [A(t),B0] e /Z, (55) the same i. At each temperature some subsystems will t ∀ n o 13

(b) 106

100 4 6 10 4 10-1 2 102 0 0 4 (a) -2 10 10 10 100 10-1 100 101 5 10 15 20

1/2 FIG. 5. (a) The (quenched) growth rate of the OTOC for a fixed relative perturbation strength ˜ǫ /n0 as a function of time rescaled by the semiclassical Lyapunov exponent (averaged over 86 noise realizations). For larger initial energy level n0 the Lyapunov regime is longer, below a certain n0 the Lyapunov regime disappears. The inset shows how λ˜QtE , with tE being the prescrambling time (stars in the main figure), increases with the logarithm scale of n0, indicating a linear trend. (b) The growth of the OTOC (one noise realization) for different relative perturbation strength and fixed initial state n0 = 8191 (smaller perturbation corresponds to lower saturation levels).

5 10 10 5 (a) (b) 4 10 10 4

3 10 10 3

2 10 10 2

1 10 10 1

0 10 10 0 0 10 20 30 0 10 20 30

FIG. 6. The (quenched) growth rate of the OTOC under the evolution of the Hamiltonian in Eq. (50) with γ = 4/3 and γ = 5/2 in panels (a) and (b) respectively. The curves are averaged over 17 realizations of the random drive. The different 9 curves corresponds to fixed initial condition n0 = 2 − 1 and varying perturbationǫ ˜. For each curve the time is rescaled with ˜ 1/3 γ−2 2/3 the Lyapunov exponent λQ(˜ǫ) = 2˜ǫ γ(γ − 1)n0 . The gray dashed line is exp(t − 1) shows clearly the validity of the ˜ −1 −2 −3 −4 −5 γ theoretical prediction of λQ. The values ofǫ ˜ for the different curves are (10 , 10 , 10 , 10 , 10 ) × n0 , where largerǫ ˜ corresponds to longer (rescaled) saturation times.

(i) have nT < 1, and will thus not contribute with a Lya- weakly driven by an external white noise. For the clas- punov regime. Hence, the global Lyapunov exponent cor- sical counterpart [28], we know that the mechanism of ∗ responds to the one of the systems nT 1 that is just the exponential growth ‘over the torus’ and saturation is about to lose its Lyapunov regime by quantum∼ effects, in fact generic, and is also relevant in higher dimensions. which one depends on the value of T . This in turn means The same is expected for the quantum problem and the that, as T 0, the combination β~λT remains a number growth of the OTOC. The formalism of the quantum of O(1), albeit→ small (because ˜ǫ is small). tangent space lays the groundwork to study such gener- alization. Our main motivation is to understand the properties X. CONCLUSION AND OUTLOOK of an isolated quasi-integrable models by mimicking the many-body integrability breaking coupling by some ex- In the current manuscript we have focused on the case ternal noise. However, the model we study concerns only of a one degree of freedom integrable model, which is classical noise acting on a quantum system, which might 14 not be suitable to capture all the effects of quantum cou- operator, i.e. an analogous quantum picture for the solid plings. One, rather primitive, way to account for noise curves in Fig. 2. One interesting future direction is thus with quantum origins is to consider correlated instead to see how a phase-space wavepacket spreads throughout of white noise. This can be addressed theoretically, as the torus, when described in the coherent-state represen- was done for the classical problem [28], or numerically— tation. by reducing the kicking rate with respect to the unper- The formalism we derived— the quantum tangent turbed evolution and the Lyapunov time. A more serious space, e.g., Eq. (4)— might be useful to address oper- way to take into account the quantum origins of noise is ator growth in other set-ups. An interesting generaliza- to start with a 1-dimensional model which is coupled to tion might be to consider a quantum chain of bosons, ensemble of linear oscillators and employ the Feynman- where the tangent space is written in terms of the ladder Vernon, Caldeira Leggett method, e.g., as in Ref. [47]. operators a(α), replacing e−iΘ, and occupation numbers Within this formalizm, the assumption of Markovianity N (α) = (a(α))†a(α). For such systems, apart of the expo- shall lead to a Lindblad-like operator, which accounts for nential growth rate, one can also consider the butterfly the coupling to the bath [48]. We expect this term to (α) (β) 2 velocity, i.e., the rate at which [N ,N0 ] depends on enter within the tangent space formalizm, however, we time as a function of sites α and β. note that such a new term should not modify the results, Finally, it will be interesting to test the implications of as it can be derived from a classical noise on a quantum our model against isolated quasi-integrable systems. In system [49]. particular the scaling of the Lyapunov exponent with the effective perturbation strength, and its comparison to the Our results highlight the meaning of the Ehrenfest time thermalization time of the system. The latter has been as the time at which the wave character of a quantum recently measured in a cold-atom system with a tunbable system plays an important role. We have found that integrability breaking interactions [14]. for the case of very weak perturbations and that of very small quantum numbers, the mechanism for exponential growth is turned off by quantum effects that originate in ACKNOWLEDGMENTS the discreteness of the spectrum, ultimately the uncertainty principle. This phenomenon is absent in the clas- We thank Laura Foini, Eduardo Fradkin, Andrea Gam- sical case, where two initial conditions can be arbitrarily bassi, Stefano Ruffo, Dima Shepelyansky, Alessandro close to each other. Our results show how the energy Silva, and Denis Ullmo for helpful discussions and sug- (or any other quasi-constant of motion) does not diffuse gestions, and Doron Cohen for pointing out relevant ref- significantly during the Lyapunov regime (Fig. 4). We erences for this work. TG and JK are supported by the have not explicitly studied the scrambling of the cos Θ Simons Foundation Grant No. 454943.

Appendix A: Dynamics in the space of Poisson Brackets

Below we derive the equations which govern the dynamics in the space of Poisson brackets. This dynamics is equivalent to the one of the tangent space. We treat the case of canonical and non-canonical variables

1. Canonical variables

In relation to the problem studied in the paper, we consider action-angle variables (I, Θ). The derivations holds for any canonical variables, e.g., coordinates and momentum (q,p), and for many-degrees of freedom. Since the Poisson brackets act as derivatives ∂ ∂ = , Θ , = , I , (A1) ∂I −{· } ∂Θ {· } they also have a corresponding chain rule: for a function F (I, Θ) ∂F ∂I ∂F ∂Θ ∂F F, Θ0 = = = I, Θ0 F, Θ + Θ, Θ0 F, I (A2) { } −∂I0 −∂I0 ∂I − ∂I0 ∂Θ −{ }{ } { }{ }

∂F ∂I ∂F ∂Θ ∂F F, I0 = = + = I, I0 F, Θ + Θ, I0 F, I (A3) { } ∂Θ0 ∂Θ0 ∂I ∂Θ0 ∂Θ −{ }{ } { }{ } where the subscript 0 refers to values at initial time. In the last equality we use the fact that the Poisson brackets are canonically invariant— taking them with respect to the canonical variables (Θ0, I0) or (Θ, I) is the same. 15

From the above relations we ﬁnd

I, Θ0 I,H , Θ0 d {Θ, Θ } {{Θ,H} , Θ } { 0} = {{ } 0} = dt  I, I0   I,H , I0  { } {{ } }  Θ, I0   Θ,H , I0   { }   {{ } }  I,H , Θ I,H , I 0 0 I, Θ0 −{{Θ,H} , Θ}{{Θ,H} , I} 0 0 {Θ, Θ } = 0 . (A4) −{{ 0} }{{ 0 } } I,H , Θ I,H , I  { I, I } −{{ } }{{ } } { 0}  0 0 Θ,H , Θ Θ,H , I   Θ, I0   −{{ } }{{ } }  { }  Since the upper and lower blocks are identical, and since the initial condition is ( I , Θ , Θ , Θ , I , I , Θ , I ) = ( 1, 0, 0, 1), { 0 0} { 0 0} { 0 0} { 0 0} − it is suﬃcient to consider only the ﬁrst two entries d I, Θ H, I , Θ H, I , I I, Θ { 0} = {{ } } −{{ } } { 0} . (A5) dt Θ, Θ0 H, Θ , Θ H, Θ , I Θ, Θ0 { } {{ } } −{{ } } { }

2. Non-canonical variables

We now consider the case when the pair of variables are not canonically conjugate. Instead of working in the action-angle space (I, Θ), we change coordinates to (I,g(Θ)). Then, the Poisson brackets are related to the derivatives according to ∂ 1 , Θ = = ,g , (A6) − {· } ∂I −g′ {· }

∂ 1 = , I , (A7) ∂g g′ {· } where g′ ∂g(Θ)/∂Θ. The chain rule relations are then ≡ 1 F,g = ( I,g F,g + g,g F, I ) (A8) { 0} g′ −{ 0}{ } { 0}{ }

1 F, I = ( I, I F,g + g, I F, I ) . (A9) { 0} g′ −{ 0}{ } { 0}{ } In analogy to Eq. (A5), we have now d I,g 1 H, I ,g H, I , I I,g { 0} = {{ } } −{{ } } { 0} . (A10) dt g,g0 g′ H,g ,g H,g , I g,g0 { } {{ } } −{{ } } { } We note that H, I ,g = g,H , I I,g ,H = H,g , I +g ˙ ′, (A11) {{ } } −{{ } }−{{ } } {{ } } that is, the matrix includes full-time-derivatives of g′. The matrix appearing in Eq. (A10) cannot imply symplectic dynamics, as the transformation (I, Θ) (I,g) is not a canonical one. We can get a symplectic dynamics by considering the vector →

I,g0 I,g0 ′ I, Θ0 { } ′ , which satisﬁes the relation { } ′ = g0 { } . g,g0 /g g,g0 /g Θ, Θ0 { } { } { } The time-derivative of this new vector is identical to the one in Eq. (A5), and can be written as

′ d I,g0 1 H, I ,g H, I , I g I,g0 { } ′ = {{ } }′ −{{ } } { } ′ , (A12) dt g,g0 /g g′ H,g ,g /g H, I ,g g,g0 /g { } {{ } } −{{ } } { } where we use the relation in Eq. (A11). 16

Appendix B: Chain-rule for commutators

In the current Appendix we consider a general statement for a chain rule for commutators and its implication on the operators evaluated in the eigenbasis of N

1. General relation

We prove the following general statement: for an analytic function (at some domain) g, and operators A and B we have g(B + s[A, B]) g(B) [A, g(B)] = lim − . (B1) s→0 s Proof: for an integer power g(x) = xk we have the known formula (readily proven by induction) [A, Bk] = k r−1 k−r r=1 B [A, B]B , which is equivalent to the expression in Eq. (B1): k P (B + s[A, B])k Bk Br−1[A, B]Bk−r = lim − . (B2) s→0 s r=1 X The general result now follows, since g is analytic then we have g(B)= a (B x )k, and k − 0 k k k (B + s[A,P B] x0) (B x0) [A, g(B)] = ak[A, (B x0) ]= ak lim − − − = − s→0 s X X a (B + s[A, B] x )k a (B x )k g(B + s[A, B]) g(B) = lim k − 0 − k − 0 = lim − . s→0 s s→0 s P P Eq. (B1) induces a linear relation between [A, g(B)] and [A, B].

2. Algebraic relations in the eigenbasis of N

The basic relations we have are eimΘ n = n + m and N,eimΘ = meimΘ, that is, in the eigenbasis of N we can | i | i iΘ ′ ′ ′ write e ′ = δ and N = nδ . We use the chain-rule for commutators above to calculate commutation n,n n,n +1 n,n n,n with N γ. Below we prove the relations: γ ′γ γ n n ([ ,N ]) ′ = − [ ,N] ′ , (B3) · nn n n′ · nn iΘ γ − γ γ e ,N ′ = ((n 1) n ) δ ′ . (B4) nn − − n,n +1 Proof: Since we know that in the action space (N)nn′ = nδnn′ we can employ ﬁrst order perturbation theory to write (N + s[ ,N]) = U −1DU, where · [ ,N]nn′ D = n + s[ ,N] ,U ′ =1+ s · . nn · nn nn n n′ − Note that the last term does not diverge as [ ,N] = 0 since N is diagonal. Therefore, to leading order in s we have · nn ′ ′ [ ,N]nn ′ [ ,N]nn (N + s[ ,N]) ′ = nδ ′ + s n · n · , (B5) · nn nn n n′ − n n′ − − and subsequently

′ ′ γ γ γ [ ,N]nn ′γ [ ,N]nn (N + s[ ,N]) ′ = n δ ′ + s n · n · , (B6) · nn nn n n′ − n n′ − − which gives Eq. (B3) γ ′γ γ n n ([ ,N ]) ′ = − [ ,N] ′ . (B7) · nn n n′ · nn − Finally, the representation of eiΘ in the eigenbasis of N gives the relation in Eq. (B4)

iΘ γ γ γ ′ e ,N ′ = ((n 1) n ) δ . (B8) nn − − n,n +1 17

Appendix C: Quantization of the power potential

1. Quantization of the integrable part

We look at the general classical Hamiltonian p2 H = + αqν . (C1) cl,int 2m The action variable of this Hamiltonian can be calculated explicitly [32]:

2+ν −1/ν 2ν I(Hcl,int)= s(ν)α √mHcl,int, (C2) √ Γ(1/ν+1) where s(ν)= 8π Γ(1/ν+3/2) with the Γ Euler function. This gives H (I)= sγ(ν)α1−γ/2m−γ/2Iγ K(m,α,ν)Iγ , (C3) cl,int ≡ with γ 2ν . Quantization of the classical Hamiltonian can be obtained by a rescaling procedure: we substitute ≡ 2+ν q (~/b)q, p bp, and we require that H = f(α, ν, m, ~)(p2 + qν ). One ﬁnds the rescaling parameter → → cl,int 2−γ ν 1 γ b = (mα~ ) 2+ν = (mα) 4 ~ 2 , (C4) and accordingly we can write f(α, ν, m, ~) ~ω H˜ (N) with ≡ 0 1− γ − γ γ−1 ω α 2 m 2 ~ , (C5) 0 ≡ having dimensions of time−1. Therefore, a quantization of the integrable Hamiltonian is simply

H = ~ω0H˜ (N). (C6) In the semi-classical limit, according to the Bohr-Sommerfeld quantization we shall substitute I = ~N in Eq. (C3), which gives H˜ (N)= N γ . Finally, let us note that since 0 <ν< we have that 0 <γ< 2. For Harmonic oscillator we have ν = 2, γ = 1, and for the inﬁnite potential well ν , γ∞= 2. → ∞

2. Quantization of the perturbation part

For the stochastic perturbation of the Hamiltonian we assume the classical form ǫ1/2qη(t), where η(t) has units of time−1. Therefore, the dimensions of ǫ1/2 are energy time1/2 length−1. Inserting the rescaling parameter b for the coordinate variable, we have · · 1/2 1/2 ~ ˜ ǫ Hint + ǫ qη(t)= ω0 Hint(N)+ − qη˜ (t) (C7) 2 3γ 3 (2−γ)~ 3γ −1 m 4 α 4 2 Finally, we rescale time t ω−1t, such that η(t) ω1/2η(t), to find → 0 → 0 ǫ 1/2 H + ǫ1/2qη(t)= ω ~ H˜ (N)+ qη˜ (t) , (C8) int 0 0 m1−γ α2−γ ~2γ−1 whereq ˜ =q ˜(N,eiΘ) is a a-dimensional operator. Let us check the dimensions of the factor that normalizes ǫ: by writing the dimension of m as energy time2 length−2 and recalling that the dimension of α is energy length−ν , we find that the factor scales as it should· be · · 2 2 −2 1−γ 2γ 2−γ 2γ−1 energy time (energy time length ) (energy length γ−2 ) (energy time) = · . · · · · length2 In the Bohr-Sommerfeld quantization we can find how the perturbation partq ˜(N,eiΘ) depends on N. From the derivation of the explicit action variable, given in Eq. (C3), we know that

− − 1/ν γ 2 2 γ µ q (H /α) (αm) 4 I 2 v(m,α,ν)I , (C9) ∝ cl,int ∝ ≡ 2 2−γ ~ with µ = 2+ν = 2 . Therefore, in the Bohr-Sommerfeld quantization, I = N, we have q˜(N,eiΘ) N µ. (C10) ∝ 18

Appendix D: Fokker Planck equation for and time-averaging

This appendix contains some detailed calculations which were used in the derivation of Eq. (45). In order to avoid confusion, we use the following summation law: all indices with enumerated subscript (n1,n2,n3 ... , except of n0 which is deﬁned in the text) are summed over, whereas all the others (n, n′, m, etc.) are free.

1. Fokker Planck equation

The derivation of a Fokker Planck equation from a Langevin equation is a standard procedure. We ﬁnd in the Stratonovitch convention

∂P ∂ ∂ = CN i CΘ ∂t −Ln1n2n3n4 n4n3 ∂CΘ − Jn1n2n3n4 ∂CΘ n4n3 n1n2 n1n2 2 ǫ˜ Θ Θ ∂ + m m m m (t) n n n n (t)C C P. (D1) 2F 1 2 3 4 F 1 2 3 4 n4n3 m4m3 ∂CN ∂CN n1n2 m1m2

Since the equation is homogeneous in Cnn′ , we can multiply it by Cnn′ Cmm′ and take average over the noise to ﬁnd a closed set of equations:

d N N Θ Θ C ′ C ′ =˜ǫ nn′n n (t) mm′m m (t) C C , (D2) dt h nn mm i F 3 4 F 3 4 h n4 n3 m4m3 i

d Θ Θ N Θ Θ N C ′ C ′ = ′ C C ′ + ′ C ′ C dth nn mm i Lnn n3n4 h n4n3 mm i Lmm m3m4 h nn m4m3 i Θ Θ Θ Θ + i ′ C C ′ + i ′ C ′ C , (D3) Jnn n3n4 h n4n3 mm i Jmm m3m4 h nn m4m3 i

d N Θ N N N Θ C ′ C ′ = ′ C ′ C + i ′ C ′ C . (D4) dth nn mm i Lmm m3m4 h nn m4m3 i Jmm m3m4 h nn m4m3 i

2. Magnus expansion

The growth of vectors and operators in the classical and quantum tangent spaces are governed by linear relations— Eq. (13) and Eqs. (40)-(43) respectively. These equations are of the form x˙2 = M(t)x2. We might relax the time- dependency of M(t) by employing time-averaging if the resulting growth rate, i.e., the Lyapunov exponent, is much smaller than typical rate of M(t), i.e., the frequency of motion around the torus in the classical case. Technically, the elimination of high-frequency terms is made in a systematic way with the Magnus expansion, of which we need here only the ﬁrst order correction. Quantum mechanically this is also possible, as we show now. The main conclusion ˜ is the following: time-averaging approximation is valid when λQ ωn0 , and results in neglecting all the oscillating terms of (t) (t) in Eq. (40). ≪ F F Formally, Eqs. (40)-(43) can be written as C˙ = S(t) C, where C and S are respectively vector and matrix of superoperators. A formal solution to this equation is ⊙

′ ′ R t S(t )⊙dt C(t)= e 0 C(0). T n o The superoperators oscillate over time through the quantum unitary evolution of eiΘ, given by the unperturbed ω ~H˜ N − † iΘ i 0 int( ) ω 1t imΘ Hamiltonian U e U, with U e ~ 0 . In principle, whenever the term e appears it gives rise to ≡ m−1 − − imΘ i(En−En m)t ′ i Pr=0 ωn rt ′ enn′ = e δn,n +m = e δn,n +m when evaluated in the unperturbed eigenbasis of N. For simplicity, let us assume that there is only one frequency ω, S(t) = S(ωt). Using the Magnus expansion, the solution up to some ﬁnite time mT , with the period T =2π/ω, is given by

m Sav⊙ Πi=1e , 19 where the averaged propagator is

T T t1 S 1 e av⊙ =1+ dt S(ωt ) + dt dt [S(ωt ) , S(ωt ) ]+ . (D5) 1 1 ⊙ 2 1 2 1 ⊙ 2 ⊙ ··· Z0 Z0 Z0 If we rescale the time in the integral with ω, then the outer integral runs from 0 to 2π and the n th term gives −n Sav⊙ − a factor of ω . Therefore, for ω λQ we can approximate e 1+ T S(t) , and the corresponding general solution ≫ ≈ ⊙

C(t)= eS(t)⊙C(0), where the overline indicates taking only the non-oscillating terms of the operation.

[1] G. Casati and B. Chirikov, Quantum chaos: between order and disorder (Cambridge University Press, 2006). [2] A. Peres, Stability of quantum motion in chaotic and regular systems, Phys. Rev. A 30, 1610 (1984). [3] R. A. Jalabert and H. M. Pastawski, Environment-independent decoherence rate in classically chaotic systems, Phys. Rev. Lett. 86, 2490 (2001). [4] A. Larkin and Y. Ovchinnikov, Quasiclassical method in the theory of superconductivity, Zh. Eksp. Teor. Fiz. 55, 2262 (1969), [Sov. Phys. JETP 28, 1200 (1969)]. [5] J. Maldacena, S. H. Shenker, and D. Stanford, A bound on chaos, J. High Energy Phys. 2016, 106. [6] A. Kitaev, Talk given at kitp program: Entanglement in strongly-correlated quantum matter, http://online.kitp.ucsb.edu/online/entangled15/kitaev/ (2015). [7] J. Polchinski and V. Rosenhaus, The spectrum in the Sachdev-Ye-Kitaev model, J. High Energy Phys. 2016. [8] J. Maldacena and D. Stanford, Remarks on the Sachdev-Ye-Kitaev model, Phys. Rev. D 94, 106002 (2016). [9] J. Kurchan, Quantum bound to chaos and the semiclassical limit, J. Stat. Phys. 171, 965 (2018). [10] I. Bloch, J. Dalibard, and W. Zwerger, Many-body physics with ultracold gases, Rev. Mod. Phys. 80, 885 (2008). [11] J. Eisert, M. Friesdorf, and C. Gogolin, Quantum many-body systems out of equilibrium, Nature Phys. 11, 124 (2015). [12] T. Kinoshita, T. Wenger, and D. S. Weiss, A quantum newton’s cradle, Nature 900, 440 (2006). [13] T. Langen, T. Gasenzer, and J. Schmiedmayer, Prethermalization and universal dynamics in near-integrable quantum systems, J. Stat. Mech-Theory E. 2016, 064009 (2016). [14] Y. Tang, W. Kao, K.-Y. Li, S. Seo, K. Mallayya, M. Rigol, S. Gopalakrishnan, and B. L. Lev, Thermalization near integrability in a dipolar quantum newton’s cradle, Phys. Rev. X 8, 021030 (2018). [15] L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, From quantum chaos and eigenstate thermalization to statistical mechanics and thermodynamics, Adv. Phys. 65, 239 (2016). [16] A. Bohrdt, C. B. Mendl, M. Endres, and M. Knap, Scrambling and thermalization in a diffusive quantum many-body system, New J. Phys. 19, 063001 (2017). [17] B. Swingle, G. Bentsen, M. Schleier-Smith, and P. Hayden, Measuring the scrambling of quantum information, Phys. Rev. A 94, 040302 (2016). [18] N. Y. Yao, F. Grusdt, B. Swingle, M. D. Lukin, D. M. Stamper-Kurn, J. E. Moore, and E. A. Demler, Interferometric approach to probing fast scrambling, arXiv preprint arXiv:1607.01801 [quant-ph] (2016). [19] G. Zhu, M. Hafezi, and T. Grover, Measurement of many-body chaos using a quantum clock, Phys. Rev. A 94, 062329 (2016). [20] N. Yunger Halpern, Jarzynski-like equality for the out-of-time-ordered correlator, Phys. Rev. A 95, 012120 (2017). [21] K. X. Wei, C. Ramanathan, and P. Cappellaro, Exploring localization in nuclear spin chains, Phys. Rev. Lett. 120, 070501 (2018). [22] C. B. Da˘gand L.-M. Duan, Detection of out-of-time-order correlators and information scrambling in cold atoms: Ladder- XX model, Phys. Rev. A 99, 052322 (2019). [23] J. Li, R. Fan, H. Wang, B. Ye, B. Zeng, H. Zhai, X. Peng, and J. Du, Measuring out-of-time-order correlators on a nuclear magnetic resonance quantum simulator, Phys. Rev. X 7, 031011 (2017). [24] M. Gärttner, J. G. Bohnet, A. Safavi-Naini, M. L. Wall, J. J. Bollinger, and A. M. Rey, Measuring out-of-time-order correlations and multiple quantum spectra in a trapped-ion quantum magnet, Nature Phys. 13, 781 (2017). [25] K. A. Landsman, C. Figgatt, T. Schuster, N. M. Linke, B. Yoshida, N. Y. Yao, and C. Monroe, Verified quantum information scrambling, Nature 567, 61 (2019). [26] J. Laskar, Chaotic diffusion in the solar system, Icarus 196, 1 (2008). [27] T. Goldfriend and J. Kurchan, Equilibration of quasi-integrable systems, Phys. Rev. E 99, 022146 (2019). [28] K.-D. N. T. Lam and J. Kurchan, Stochastic perturbation of integrable systems: A window to weakly chaotic systems, J. Stat. Phys. 156, 619 (2014). [29] R. Zwanzig, Nonlinear generalized langevin equations, J. Stat. Phys. 9, 215 (1973). [30] P. Carruthers and M. M. Nieto, Phase and angle variables in quantum mechanics, Rev. Mod. Phys. 40, 411 (1968). [31] M. M. Nieto, Quantum phase and quantum phase operators: some physics and some history, Phys. Scripta , 5 (1993). 20

[32] J. F. Cariñena, C. Farina, and C. Sigaud, Scale invariance and the Bohr-Wilson-Sommerfeld (BWS) quantization for power law onedimensional potential wells, Am J. Phys. 61, 712 (1993). [33] A proof that the Bohr-Sommerfeld approximation is valid for the case of one-dimensional power potential can be found in Ref. [50]. [34] A. B. Rechester, M. N. Rosenbluth, and R. B. White, Calculation of the kolmogorov entropy for motion along a stochastic magnetic field, Phys. Rev. Lett. 42, 1247 (1979). [35] C. Anteneodo and R. O. Vallejos, Scaling laws for the largest lyapunov exponent in long-range systems: A random matrix approach, Phys. Rev. E 65, 016210 (2001). [36] R. O. Vallejos and C. Anteneodo, Generalized lyapunov exponents of the random harmonic oscillator: Cumulant expansion approach, Phys. Rev. E 85, 021124 (2012). [37] B. V. Chirikov, A universal instability of many-dimensional oscillator systems, Phys. Rep. 52, 263 (1979). [38] E. B. Rozenbaum, S. Ganeshan, and V. Galitski, Lyapunov exponent and out-of-time-ordered correlator’s growth rate in a chaotic system, Phys. Rev. Lett. 118, 086801 (2017). [39] B. Chirikov, F. Izrailev, and D. Shepelyansky, Quantum chaos: Localization vs. ergodicity, Physica D 33, 77 (1988). [40] G. Berman and G. Zaslavsky, Condition of stochasticity in quantum nonlinear systems, Physica A 91, 450 (1978). [41] G. Casati, B. V. Chirikov, F. M. Izraelev, and J. Ford, Stochastic behavior of a quantum pendulum under a periodic perturbation, in Stochastic Behavior in Classical and Quantum Hamiltonian Systems, edited by G. Casati and J. Ford (Springer Berlin Heidelberg, Berlin, Heidelberg, 1979) p. 334. [42] F. M. Izrailev, Simple models of quantum chaos: Spectrum and eigenfunctions, Phys. Rep. 196, 299 (1990). [43] S. Fishman, D. R. Grempel, and R. E. Prange, Chaos, quantum recurrences, and anderson localization, Phys. Rev. Lett. 49, 509 (1982). [44] D. R. Grempel, S. Fishman, and R. E. Prange, Localization in an incommensurate potential: An exactly solvable model, Phys. Rev. Lett. 49, 833 (1982). [45] B. V. Chirikov, Report no. 267 (1969), [English Translation CERN Trans. 7140 (1971)]. [46] K. M. Frahm and D. L. Shepelyansky, Diffusion and localization for the chirikov typical map, Phys. Rev. E 80, 016210 (2009). [47] J. Tuziemski, Out-of-time-ordered correlation functions in open systems: A Feynman-Vernon influence functional approach, arXiv preprint arXiv:1903.05025 [quant-ph] (2019). [48] C. A. Brasil, F. F. Fanchini, and R. d. J. Napolitano, A simple derivation of the Lindblad equation, Revista Brasileira de Ensino de FAsica˜ 35, 01 (2013). [49] D. A. Rowlands and A. Lamacraft, Noisy coupled qubits: Operator spreading and the fredrickson-andersen model, Phys. Rev. B 98, 195125 (2018). [50] A. Voros, Exact quantization condition for anharmonic oscillators (in one dimension), J. Phys. A-Math. Gen. 27, 4653 (1994).