WorksPhysics and Astronomy
Experimentally relevant witness for non-Markovianity in discrete time crystals
Abstract
Discrete time crystals (DTCs) are almost always analyzed under the Born-Markov approximation, in which the environment is memoryless. We show that a reservoir with a finite correlation time leaves an exact and strikingly simple imprint on the stroboscopic dynamics of DTCs, which can be used as an experimentally relevant witness for non-Markovianity. By analyzing a post-Markovian master equation we prove that the Floquet spectrum splits into the ordinary Markovian spectrum plus a replica of the purely unitary Floquet spectrum, uniformly contracted by , where is the drive period. Consequently, the DTC signal decay rate becomes independent of any system parameter once exceeds the Markovian lifetime of the DTC. The measured decay, and in particular its insensitivity to tunable system parameters, thus becomes a direct read-out of the reservoir correlation time. Below this threshold the dynamics is already non-Markovian, but the memory remains hidden in the decay. We confirm this in two structurally unrelated DTCs, a collective long-range model with incoherent pumping and a disorder-free Ising chain with dephasing, where a single estimator recovers . The time crystal thereby witnesses the non-Markovianity of its own environment.
Introduction
The spontaneous breaking of time-translation symmetry gives rise to time crystals, phases of matter in which a many-body system develops persistent temporal order [1, 2]. While no-go theorems preclude such order in thermal equilibrium [3, 4], periodically driven (Floquet) systems evade these constraints and host discrete time crystals (DTCs), in which the response locks to a subharmonic of the drive and the discrete time-translation symmetry is spontaneously broken down to a subgroup [5, 6, 7, 8]. Following these early proposals, DTC order has been studied across a wide range of platforms [9, 10, 11, 12] and observed experimentally with trapped ions, nitrogen-vacancy centers, and superconducting processors [13, 14, 15, 16].
Indeed, real experimental systems are never perfectly isolated, and understanding time-crystalline order in the presence of an environment has become a central theme. Counterintuitively, dissipation is not necessarily detrimental: engineered losses can themselves stabilize nonequilibrium order, as boundary time crystals [17] and dissipative discrete time crystals [18, 19], as well as continuous time crystals that break a continuous time-translation symmetry [20, 21, 22]. Such dissipative time crystals have recently been realized in driven atom-cavity and related setups [23, 24]. Dissipation nonetheless competes with the interactions that sustain temporal order, and sufficiently strong coupling to the environment eventually “melts” the time crystal, driving the system into a stationary or synchronized state with no subharmonic response.
Almost all studies of dissipative time crystals adopt the Markovian (Born–Markov) approximation, in which the environment is memoryless and the dynamics is generated by a time-local master equation. Physical reservoirs, however, possess a finite correlation time, and the resulting non-Markovian dynamics is characterized by a back-flow of information from the environment into the system [25, 26, 27, 28]. Previous studies such as the time-dependent dissipation in mean-field limit [29] and with strong symmetry [30] have suggested the beneficial effects of non-Markovian dynamics for preserving DTCs in the presence of strong dissipation . However, whether such memory effects help or hinder nonequilibrium order in more general systems remains largely unexplored, and in particular it is an open question whether environmental memory can counteract the dissipative melting of a time crystal.
In this Letter we show that environmental memory does not merely modify time-crystalline order. Rather, it “writes” its own correlation time into the observable response. We model the open dynamics with a post-Markovian master equation carrying an exponential memory kernel of tunable memory time and embed it exactly into a time-local evolution on an enlarged space. The embedded generator is block-triangular, and we prove that its Floquet spectrum is the ordinary Markovian spectrum together with a replica of the purely unitary Floquet spectrum contracted by , where is the period of the drive. Whenever exceeds the Markovian lifetime of the crystal, the replica sets the subharmonic decay and the measured signal envelope becomes a direct, calibration-free read-out of the reservoir correlation time, rigid against the drive, the interactions, the dissipation rate and the system size, while the subharmonic phase stays locked at . Below this threshold the memory is hidden: the dynamics is non-Markovian, yet the crystal decays as if the bath were memoryless. We verify this in two structurally unrelated DTCs, a collective Lipkin-Meshkov-Glick (LMG) model [31, 32, 33] with incoherent pumping, and a disorder-free short-range tilted Ising chain with dephasing, and show that a single estimator recovers in both.
Theoretical framework
Memory-embedded dynamics
We take the reduced dynamics to obey a periodically driven extension of the Shabani–Lidar post-Markovian master equation (PMME) [34, 35, 36, 37, 38], which incorporates a finite reservoir-correlation time while retaining a time-local Markovian limit,
with , dissipator , Liouvillian and . The single-exponential kernel is setting the reservoir correlation time [39, 40]. As , becomes a delta function and Eq. (1) reduces to the Lindblad equation , so tunes continuously from memoryless to non-Markovian dynamics. The PMME can exhibit CP-indivisible dynamics, as demonstrated for specific kernels and parameter regimes in Ref. [36]. For exponential-kernel PMME dynamics, noncommutation between the Hamiltonian and dissipative superoperators can lead to information backflow [35, 37]. Both models considered below feature such noncommutation.
The single pole lets one trade the memory integral for a single auxiliary matrix (defined as the convolution itself, ) (for details see [41]), in the spirit of Lindblad rate equations [35]: the pair obeys the exact time-local equation
where is not a Lindbladian on a doubled space (). The auxiliary matrix is unnormalized, so it is not a density matrix. Because our system is periodic, , the dynamics is governed by the one-period propagator . Starting from the empty-memory state it advances the stroboscopic state as , where denotes the stroboscopic period number [42]. The lifetime of an observable is then set by the slowest expressed Floquet multiplier of , (in units of time; in driving periods) (see SM [41]). In the following we analyze the internal structure of the Floquet propagator .
Replica theorem
The embedding, Eq. (2), has an exact structural consequence. In the new variable , we have , . Thus, undergoes unitary evolution multiplied by a scalar decay at rate . The one-period propagator in the representation is block lower triangular, with -dimensional diagonal blocks and . Here is the unitary Floquet channel, and is the Markovian Floquet channel. Block triangularity implies that its spectrum is exactly the union of the spectra of these diagonal blocks:
Since gives , the measurable state carries a unitarily evolving replica of its own initial condition,
Physically, this exponential-kernel construction admits an exact representation in which dissipation switches on once, at a random time with probability density , consistent with Ref. [35]. The replica in Eq. (4) is the not-yet-switched contribution: it evolves unitarily with weight , which depends only on and . Complete positivity follows from the Lindblad dilation described in Ref. [41], where an auxiliary gate switches dissipation on irreversibly at rate .
Rigidity and information partition
Equation (3) has two physical consequences. First, because is unitary, every replica Floquet multiplier has modulus exactly . When an expressed replica mode controls the asymptotic subharmonic decay, the lifetime-setting Floquet multiplier locks to . The asymptotic decay rate of the subharmonic envelope then measures the kernel memory time : it is independent of the drive amplitude, the interactions, the dissipation rates and the system size at fixed . Second, each replica multiplier has the form , with an eigenvalue of : fixes its radius, while the coherent dynamics alone fixes its phase. Decay and oscillation are thus carried by distinct spectral data: the radius encodes the kernel memory time, while the phase encodes the oscillation frequency, with for a period-doubled response. The replica theorem alone does not establish robust phase locking. Only the rate is universal: the amplitude of the replica contribution depends on the initial state and on the measured observable as well as on the system dynamics.
Branch crossing and three regimes
Let denote the slowest expressed oscillating Floquet multiplier of the purely Markovian Floquet operator , and the corresponding Markovian lifetime. Because the two branches in Eq. (3) compete for the largest modulus, provided this Markovian mode and at least one relevant oscillating replica mode carry nonzero weight in the full embedded response, and no slower oscillatory mode becomes expressed, the modulus of the selected Floquet multiplier is
so the replica takes over above the sharp threshold , and the observable lifetime obeys .
Two questions classify the dynamics: is the dynamics memoryless, and which time sets the lifetime? The answers define three regimes,
In the Markovian regime the evolution reduces to the Markovian limit, and the crystal decays at its Markovian lifetime. In the hidden-memory regime the kernel has finite memory, but the replica decays at least as fast as the Markovian subharmonic mode and does not change the asymptotic exponential lifetime; the measured then provides only the upper bound . In the memory-dominated regime the replica controls the asymptotic oscillatory decay, and the kernel memory time becomes the lifetime itself, . The threshold thus marks where memory becomes visible in the asymptotic decay rate, not where it appears: this lifetime estimator cannot resolve memory times shorter than , although transient dynamics may reveal them.
The replica theorem constrains only the structure of Eq. (2), not the Hamiltonian, the jump operators, or the Hilbert space dimension. We therefore test it in two deliberately dissimilar DTCs: LMG Hamiltonian with incoherent pumping and a short-range tilted spin chain with dephasing (Sec. III B). Both are solved numerically at finite size without mean-field or cumulant truncations, and the one-period propagator is diagonalized numerically.
Models
Model I: LMG with pumping
Following Ref. [18], a robust dissipative time crystal can arise when the interplay of interactions and incoherent pumping stabilizes two -symmetry-broken states in the thermodynamic limit and a parity operation toggles between them each period. Our first model realizes
Figure 1. The memory witness at a glance. (a) By Eq. (3) the Floquet spectrum of the memory-embedded propagator is the Markovian cloud (grey) plus a replica of the unitary spectrum on a circle of radius (blue). As grows the replica circle inflates towards the unit circle (dashed); once it exceeds the largest expressed oscillating Markovian modulus, the lifetime-setting mode (red) locks onto it at angle . (b) The stroboscopic response is period-doubled; its decay is set by . In the hidden-memory regime () it inherits the fast Markovian decay (grey, “molten”), while in the memory-dominated regime the envelope follows (red, “revived”) and reads off the reservoir memory—independently of the drive and the dissipation. The enabling step, an exact embedding of the post-Markovian dynamics into a time-local generator coupling to an auxiliary memory sector at rate , is described in the text.
this with a collective spin subject to the pump and the periodic drive
with and . The longitudinal field takes two values within one period: a weak residual field during the first segment, and a strong pulse field during the second segment. In the strong-pulse limit , with , the second segment approaches an ideal -rotation about (a parity kick) that swaps the two symmetry-broken states. We monitor the stroboscopic magnetization from the spin-coherent state.
In Figure 2, we show the subharmonic dynamics and the underlying branch crossing. In panel (a) we plot the stroboscopic magnetization as a function of , for four memory times , at fixed . Under identical drive and dissipation the Markovian evolution suppresses the subharmonic response within a few periods, whereas sustains the subharmonic oscillations much longer. When the expressed replica modes dominate, the asymptotic exponential decay is governed by , with lifetime . The signal can additionally contain oscillatory modulations from the phases of these modes. We note that in the Markovian limit the subharmonic lifetime of this collective DTC grows with system size [Fig. 2(d)], consistent with increasing stability of the finite-size subharmonic response. Establishing persistent thermodynamic order and robustness against pulse errors requires further analysis; a detailed analysis of the Markovian dynamics, including the Floquet–Lindblad spectrum and finite-size scaling, will be presented separately [43]. The spectral replica theorem and the finite-size memory-time estimator do not require a divergent thermodynamic lifetime.
Figure 2. Memory-revived time crystal in the LMG model. (a) Stroboscopic magnetization for fixing : the Markovian response () melts within a few periods, whereas long memory sustains a period-doubled signal whose envelope tracks . (b) Floquet multiplier versus at fixed (colored diamonds): each curve is flat at its Markovian value up to the kink at and then locks onto (dotted). (c) versus the detuning for (colored circles) and the Markovian reference (solid); the memory branches sit on the parameter-free lines (dotted) while the Markovian branch drifts with the field. (d) versus system size fixing : the Markovian modulus climbs as the collective crystal stiffens toward its mean-field limit, whereas every memory plateau stays pinned on , independent of . Insets in (b,c): subharmonic phase , pinned to unity (shaded band: ). All the common parameters are , , and .
Panels (b) and (c) resolve the mechanism. In panel (b) we plot the modulus of the lifetime-setting Floquet multiplier, , as a function of , for ; the dotted line is the fit-free prediction of Eq. (3). Along the memory axis, remains pinned to its Markovian value up to the kink at (hidden-memory regime), beyond which it follows (memory-dominated regime). In panel (c) we plot as a function of the magnetic field , for memory times $\tau_m = 10, 20, 30, 50, 100, together with the Markovian reference (solid) and the corresponding levels (dotted). Along the field axis, the Markovian branch drifts with , a property of the driven system, while the replica-dominated segments form flat plateaus at , because at fixed and the drive enters the replica Floquet multiplier only through its phase and never through its modulus. The exact spectral union permits a sharp crossing of the branch moduli, producing a kink in their maximum: the kink occurs at in panel (b), and at fields satisfying in panel (c) [41]. Equality of the moduli need not imply equality of the complex multipliers, whose phases can differ. The insets of (b) and (c) show the phase of the same multiplier.
In panel (d) we plot as a function of the system size at fixed , for the Markovian case and for . This separates the Markovian relaxation time from the kernel memory time. The Markovian modulus climbs steadily with over the sizes studied, so the threshold increases with it. The replica plateaus themselves remain fixed at , independent of , because their modulus depends only on and ; their phases and observable weights can still depend on . The consequence is directly visible: the curve follows the replica plateau at small but switches to the Markovian branch once exceeds 20, i.e. the same kernel memory time passes from the memory-dominated to the hidden-memory regime as the crystal grows, while stay locked to their replica plateaus across the whole window. In the memory-dominated regime, the decay rate therefore reads out , whereas the threshold is set by the Markovian dynamics; this inference requires no thermodynamic limit. In the regime shown, a smaller crystal therefore has a lower memory-detection threshold: its shorter Markovian lifetime allows shorter kernel memory times to be resolved through the asymptotic decay rate, provided the replica contribution remains measurable.
Model II: short-range chain with dephasing
For the second model, we consider a finite-duration implementation of the disorder-free Floquet DTC studied in Refs. [44, 45] and couple it to a dephasing environment. In the unitary construction, a chain of spin- particles evolves under a Stark-like Ising interaction and a near- global spin rotation,
where is the global-pulse error and the Ising couplings vary linearly along an open chain, without quenched disorder. We take ; the unitary Floquet operator is then with , so that . This reproduces the unitary Floquet operator of the kicked construction at the same and , although the evolution within each period differs. At the resonant point and for even , for any , giving exact two-period recurrence independently of the pulse error [44, 45]. A nonzero period-doubled component additionally requires overlap of the initial state and observable with a eigenmode of . Our numerics operate close to this point.
We now open this chain within the PMME framework of Eq. (1), coupling it to local pure-dephasing channels with kernel memory time . The dephasing acts during both steps of the drive, including the finite-duration spin flip. We monitor the stroboscopic longitudinal magnetization starting from the fully polarized state .
Figure 3 repeats the analysis of Figure 2 for a structurally distinct DTC: nearest-neighbor instead of infinite-range coupling, local dephasing instead of collective pumping, a -dimensional Hilbert space, a different order parameter, and the pulse error as the tuning parameter of the Markovian lifetime. Since the replica theorem is exact for any such choice, the nontrivial test here is that the replica branch is expressed in the measured magnetization and sets its asymptotic exponential decay in the memory-dominated regime, exactly as in Model I.
Figure 3. Memory revived time crystal in a short-range chain with dephasing. (a) Stroboscopic magnetization for fixing ; memory revives the period-doubled response exactly as in the collective model. (b) Floquet multiplier versus at fixed (colored diamonds): the same plateau-then- structure, with the kink at . (c) versus the global-pulse error for (colored circles), the Markovian reference (solid), and the predictions (dotted); here the Markovian branch decreases with , the opposite trend to Figure 2(c), yet the memory plateaus land on identical values. (d) Regime boundary in the plane fixing : the threshold (green) separates a hidden-memory regime (dark, large readout error ) from a memory-dominated regime (light, error ); the memory-dominated regime widens as dissipation grows. Insets in (b,c): subharmonic phase (band: ). All panels: , , , initial state ; .
In panel (a) we plot the stroboscopic longitudinal magnetization as a function of , for memory times , at fixed pulse error . The traces decay substantially within periods ( lies in the hidden-memory regime and decays asymptotically at the Markovian rate), while persists across the plotted window. Here corresponds to approximately five drive periods, since . In panel (b) we plot as a function of , for pulse errors , with the prediction dotted. Each holds a plateau at its own Markovian value up to the kink at , beyond which all curves collapse onto : in the memory-dominated regime the lifetime-setting modulus becomes independent of and is fixed by at fixed .
In panel (c) we plot as a function of the pulse error , for memory times , together with the Markovian reference (solid) and the levels (dotted). The Markovian branch decreases with , while the replica-dominated segments form plateaus at , following the same fit-free prediction as in Model I. The threshold moves accordingly—for the memory branch overtakes the Markovian one near . A memory that is hidden in a more accurately driven crystal can become dominant as is reduced, provided the period-doubled response remains observable. The insets show that remains close to throughout the plotted DTC regime: the change in lifetime is accompanied by
an approximately period-doubled response. Thus two microscopically distinct DTCs with different Markovian responses recover the same memory-controlled modulus in the memory-dominated regime, allowing the kernel memory time to be inferred from the asymptotic decay rate.
In panel (d) we plot the relative spectral inference error as a color map over the plane at fixed , with the threshold curve superposed (green), marking a sharp branch-crossing boundary in the memory–dissipation plane. Here is obtained from the Floquet spectrum, rather than from a fit to a noisy time trace. Below it we have the hidden-memory regime [Eq. (6)], where the lifetime of the subharmonic response is determined by ; above it we have the memory-dominated regime, in which the spectral inference error vanishes under the max rule, up to numerical precision. The error also vanishes on the boundary, where . Because decreases with in the regime shown, the boundary bends downward and the memory-dominated region grows with . Thus, in this regime, stronger dissipation lowers the memory-detection threshold and enlarges the region in which the lifetime estimator operates. The memory-dominated response therefore occupies an extended region of the plane rather than an isolated parameter set.
Discussion
We have shown that a reservoir with a finite correlation time leaves an exact and universal fingerprint on the Floquet spectrum of a driven open time crystal. Embedding the master equation into a time-local generator on an enlarged operator space renders that generator exactly block-triangular, and the resulting Floquet spectrum is the Markovian spectrum plus a copy of the unitary Floquet spectrum contracted by . Once the memory time exceeds the Markovian lifetime of the crystal (the memory-dominated regime), the replica branch controls the asymptotic subharmonic decay, provided the relevant modes remain expressed in the observable. The resulting decay rate then measures the kernel memory time . Because at fixed drive period the drive enters only the phase of the replica Floquet multiplier and the memory time only its modulus, the two pieces of information are carried by distinct spectral data.
The practical consequence is a witness that is unusually well matched to what a time-crystal experiment already measures. The lifetime uses only the slope of the period-doubled decay and requires no tomography, no independent bath spectroscopy, and no detailed Hamiltonian reconstruction once the applicability of the exponential-kernel model and replica dominance have been established. Counterintuitively, stronger dissipation can lower the memory-detection threshold when it shortens the competing Markovian lifetime, as demonstrated in Model II. Existing trapped-ion, nitrogen-vacancy and superconducting-processor DTC platforms [13, 14, 15, 16] measure stroboscopic spin observables, providing a natural setting in which to test this estimator. Complementing these capabilities, Zhang et al. demonstrated that a phenomenological PMME accurately captures and predicts experimentally measured non-Markovian superconducting-qubit dynamics [37]. Vlachos et al. experimentally engineered non-Markovian dynamics using classical stochastic control of a superconducting transmon [46]. Together, these developments motivate testing the proposed memory-time estimator in driven systems with controlled environmental noise. The witness, however, operates only in the memory-dominated regime, : the kernel memory time must exceed the corresponding Markovian lifetime. In the hidden-memory regime the measured asymptotic decay yields only the upper bound .
Three caveats delimit the result. First, the replica branch must carry nonzero weight in the measured observable. This weight is set by the modal coefficients and is nonzero in the parameter regimes examined in both models studied here; a period-doubled replica contribution requires expressed coherent Floquet modes with phases at or near . Neither their observable weight nor robust phase locking follows from the replica theorem alone. Second, the exact spectral splitting of Eq. (3), and the state decomposition in Eq. (4), relies on the single-pole (exponential) memory kernel and the embedding used here. For kernels with several poles, or non-exponential kernels, this proof does not apply, and the decay need not equal the longest kernel time scale. Such extensions require a separate analysis. Third, the lifetime signature identifies a kernel memory time within the assumed model; it does not by itself certify information backflow or CP indivisibility. These require independent diagnostics [25, 26]. The exponential-kernel construction also admits a classical random-switching realization [35], so the inferred time scale does not identify the microscopic origin of the memory.
More broadly our results turn an environmental time scale into an observable feature of collective subharmonic response. Non-Markovianity is normally something to be diagnosed by process-tomographic measures, such as trace-distance backflow [25, 47] or divisibility breaking [48, 49], that are expensive to implement in a many-body setting [50]. Here the time crystal performs the measurement itself: its own rigidity converts the bath correlation time into a directly readable decay rate. This suggests that driven phases with a protected subharmonic response could serve as built-in spectrometers of their own environment, and, conversely, that reservoir engineering is a route to extending the lifetime of the nonequilibrium order.
Exact time-local embedding and Floquet replica structure
Nonlocal equation and exact local embedding
We begin from the modified post-Markovian equation [34, 35, 36, 37, 38]
where , , and . The memory kernel is chosen as , where is the reservoir correlation time. For this kernel it is convenient to define the operator , which we refer to as the pseudomode operator [35],
with initial condition . Differentiating Eq. (10) and using the Leibniz rule gives
Hence the nonlocal equation is exactly equivalent to
It is convenient to collect the physical state and pseudomode operator into the embedded object and write the dynamics in block form in doubled space
The embedded generator acts on the doubled operator space . Equations (A4)–(A5) provide an exact time-local linear embedding of the original nonlocal dynamics, generated by . The generator is not itself a GKLS generator on a doubled physical Hilbert space, since is a pseudomode operator and is generally not normalized.
Trace preservation and complete positivity
Trace preservation follows immediately from Eq. (A4a),
because and for any operator .
The pseudomode operator is not normalized. Defining and tracing Eq. (A4b) gives
and, with ,
Thus has trace smaller than unity at finite time and approaches unit trace asymptotically.
For the exponential-kernel construction used here, complete positivity can be shown directly through a Markovian dilation. Define
Subtracting Eq. (A4b) from Eq. (A4a) gives
Introduce a two-state ancilla and denote its matrix units by , with . The block-diagonal state is
With
the GKLS equation for reproduces Eqs. (A10a)–(A10b) for the two diagonal blocks and leaves the block-diagonal form (A11) invariant. Because , the initial dilated state is the fixed product
i.e. the system in and the ancilla in the fixed pure state . The generator is of GKLS form, so evolves completely positively; in particular and each diagonal block is positive, whence and separately. Thus with , so it may consistently be regarded as a subnormalized pseudomode operator. With the ancilla prepared in a state independent of and
the map is a fixed-dilation followed by a partial trace, and is therefore CPTP. The construction carries over verbatim to a driven Hamiltonian, with .
Embedded spectral problem and exact branch decomposition
In any fixed basis of , let and denote, respectively, the right eigenmode of and its left dual:
For a diagonalizable we choose the biorthogonal normalization
The right mode is written as
where and are the physical and pseudomode components. The right-eigenvalue equations are
Whenever the shifted resolvent exists, Eq. (A20b) gives
Substitution into Eq. (A20a) yields the reduced nonlinear equation
The same elimination can be expressed through the Schur-complement determinant identity,
whenever the Schur complement is well defined. This is a meromorphic representation of the characteristic determinant: zeros and poles of the two factors must not be interpreted separately at singular points of the resolvent.
The exact spectral structure is more transparent after the change of variables . Define
Then
Because is block lower triangular,
and therefore
This factorization is exact and requires no commutativity assumption between and . The Schur-complement form Eq. (35) and the triangular factorization Eq. (37) are two representations of the same characteristic determinant; the latter makes the two exact spectral families manifest.
Markovian limit and stationary states
The Markovian limit is obtained by taking , for which as a one-sided approximate identity. The triangular variables make this limit particularly transparent. Equation (A10a) has the exact solution
Using , Eq. (A4a) becomes . For every fixed , as , while its integrated contribution is in finite dimension. Consequently
for fixed finite times. Equation (A8) reflects the same initial boundary layer of width .
For finite , the stationary subspace is inherited exactly from the Markovian generator. From Eqs. (A10), a stationary embedded state satisfies
Since has purely imaginary spectrum, has no zero eigenvalue for finite . Hence
Thus every stationary state of the embedded dynamics is a stationary state of , and conversely every Markovian stationary state lifts to the embedded stationary pair
Driven extension, Floquet spectrum, and expressed modes
The derivation above was formulated for a time-independent . For an explicitly driven Hamiltonian, the ordinary propagator is replaced by the two-time propagator
The driven extension of Eq. (9) is
For the same exponential kernel, define the pseudomode operator
The exact local equations retain the same structure,
Accordingly,
The same time-independent similarity transformation introduced above triangularizes at every instant,
For a periodic drive, , define the embedded one-period propagator
and the purely unitary and Markovian Floquet channels
Because the same transformation applies at all times,
where is an off-diagonal feed-forward block and therefore does not affect the spectrum. Hence
Since is a unitary conjugation channel, every replica multiplier has modulus . Moreover, implies , and therefore
This identity is exact: the first term is the autonomous replica contribution, whose envelope is fixed solely by .
Floquet-mode expansion
Away from isolated defective degeneracies, we assume that is diagonalizable. Let denote its right column modes and the corresponding left dual modes,
with biorthogonal normalization . Write the physical and pseudomode components of a right Floquet mode as
For the initially empty pseudomode,
and its modal expansion is , . For an observable acting on the physical sector,
where . Thus is the weight with which Floquet mode is expressed in the measured signal; a mode contributes only when . For exactly degenerate, or numerically unresolved, multipliers, individual eigenvectors within the degenerate subspace are not unique. We therefore group such multipliers and sum their complex contributions before assigning an observable weight, making the result independent of the basis chosen within that subspace. The numerical threshold used to declare a mode expressed is specified below in Eq. (61). Writing
separates the decay rate per unit time from the oscillation phase . In particular, produces the period-doubled factor .
Mode selection
We first define the set of dynamically expressed, nonstationary modes,
Here and are, respectively, the observable-weight and stationary-mode cutoffs used in the numerical analysis.
To identify the oscillatory sector relevant to the DTC response, we further define
where excludes modes whose phase is numerically indistinguishable from zero; the model-specific numerical tolerances are given with the corresponding calculations below. For Model I we take , the normalized transverse magnetization, with . For the finite-size parity-symmetric dynamics considered here, the stationary mode is even and therefore has zero overlap with this odd observable. The same construction is applied to the longitudinal DTC order parameter of Model II.
The lifetime-setting multiplier is the slowest expressed oscillating mode,
When several multipliers share this modulus within a relative tolerance (as for the replica branch, all of whose multipliers have modulus ), we select the one with the largest ; its phase defines .
Lifetime-setting branch and activation threshold
Within the DTC regime considered in the main text, let denote the slowest expressed period-doubled multiplier of the purely Markovian Floquet channel . Define the Markovian DTC lifetime
Provided the relevant Markovian and replica period-doubled modes are both expressed for the chosen initial-state–observable pair, and remain the lifetime-setting oscillatory candidates in the parameter regime under consideration, the two exact spectral families compete only through their moduli. The lifetime-setting DTC multiplier therefore obeys .
Consequently,
The activation threshold is . Below threshold (the hidden-memory regime of the main text, ) the measured DTC lifetime is pinned to and provides only the upper bound ; above threshold (the memory-dominated regime) the replica branch sets the decay and . At an isolated defective degeneracy, generalized Floquet modes replace the simple modal expansion, but the exact spectral radii and the threshold condition above are unchanged.
Acknowledgment
This research was supported by the Israel Science Foundation (grant No. 2120/24). RG warmly thank Subhajit Sarkar for his valuable insights. V.M. acknowledges support from Anusandhan National Research Foundation (ANRF) through ARG (Project No. ANRF/ARG/2025/002531/PS).
References
- [1]F. Wilczek, Phys. Rev. Lett. 109, 160401 (2012).
- [2]K. Sacha and J. Zakrzewski, Rep. Prog. Phys. 81, 016401 (2018).
- [3]P. Bruno, Phys. Rev. Lett. 111, 070402 (2013).
- [4]H. Watanabe and M. Oshikawa, Phys. Rev. Lett. 114, 251603 (2015).
- [5]D. V. Else, B. Bauer, and C. Nayak, Phys. Rev. Lett. 117, 090402 (2016).
- [6]V. Khemani, A. Lazarides, R. Moessner, and S. L. Sondhi, Phys. Rev. Lett. 116, 250401 (2016).
- [7]C. W. von Keyserlingk, V. Khemani, and S. L. Sondhi, Phys. Rev. B 94, 085112 (2016).
- [8]N. Y. Yao, A. C. Potter, I.-D. Potirniche, and A. Vishwanath, Phys. Rev. Lett. 118, 030401 (2017).
- [9]D. V. Else, C. Monroe, C. Nayak, and N. Y. Yao, Annual Review of Condensed Matter Physics 11, 467 (2020).
- [10]V. Khemani, R. Moessner, and S. L. Sondhi, arXiv preprint arXiv:1910.10745 (2019), arXiv:1910.10745 [cond-mat.str-el].arxiv.org/abs/1910.10745
- [11]M. P. Zaletel, M. Lukin, C. Monroe, C. Nayak, F. Wilczek, and N. Y. Yao, Rev. Mod. Phys. 95, 031001 (2023).
- [12]K. Sacha, Phys. Rev. A 91, 033617 (2015).
- [13]J. Zhang, P. W. Hess, A. Kyprianidis, P. Becker, A. Lee, J. Smith, G. Pagano, I.-D. Potirniche, A. C. Potter, A. Vishwanath, N. Y. Yao, and C. Monroe, Nature 543, 217 (2017).
- [14]S. Choi, J. Choi, R. Landig, G. Kucsko, H. Zhou, J. Isoya, F. Jelezko, S. Onoda, H. Sumiya, V. Khemani, et al., Nature 543, 221 (2017).
- [15]X. Mi, M. Ippoliti, C. Quintana, A. Greene, Z. Chen, J. Gross, F. Arute, K. Arya, J. Atalaya, R. Babbush, et al., Nature 601, 531 (2022).
- [16]P. Frey and S. Rachel, Science advances 8, eabm7652 (2022).
- [17]F. Iemini, A. Russomanno, J. Keeling, M. Schirò, M. Dalmonte, and R. Fazio, Phys. Rev. Lett. 121, 035301 (2018).
- [18]Z. Gong, R. Hamazaki, and M. Ueda, Phys. Rev. Lett. 120, 040404 (2018).
- [19]F. M. Gambetta, F. Carollo, M. Marcuzzi, J. P. Garrahan, and I. Lesanovsky, Phys. Rev. Lett. 122, 015701 (2019).
- [20]B. Buča, J. Tindall, and D. Jaksch, Nat. Commun. 10, 1730 (2019).
- [21]F. Carollo, K. Brandner, and I. Lesanovsky, Phys. Rev. Lett. 125, 240602 (2020).
- [22]A. Cabot, F. Carollo, and I. Lesanovsky, Phys. Rev. Lett. 132, 050801 (2024).
- [23]H. Keßler, P. Kongkhambut, C. Georges, L. Mathey, J. G. Cosme, and A. Hemmerich, Phys. Rev. Lett. 127, 043602 (2021).
- [24]P. Kongkhambut, J. Skulte, L. Mathey, J. G. Cosme, A. Hemmerich, and H. Keßler, Science 377, 670–673 (2022).
- [25]H.-P. Breuer, E.-M. Laine, and J. Piilo, Phys. Rev. Lett. 103, 210401 (2009).
- [26]Angel Rivas, S. F. Huelga, and M. B. Plenio, Reports on Progress in Physics 77, 094001 (2014).
- [27]H.-P. Breuer, E.-M. Laine, J. Piilo, and B. Vacchini, Rev. Mod. Phys. 88, 021002 (2016).
- [28]I. de Vega and D. Alonso, Rev. Mod. Phys. 89, 015001 (2017).
- [29]B. Das, N. Jaseem, and V. Mukherjee, Phys. Rev. A 110, 012208 (2024).
- [30]R. Ghosh, S. Sarkar, and Y. Dubi (2026), to be published.
- [31]H. J. Lipkin, N. Meshkov, and A. J. Glick, Nuclear Physics 62, 188 (1965).
- [32]N. Meshkov, A. J. Glick, and H. J. Lipkin, Nuclear Physics 62, 199 (1965).
- [33]A. J. Glick, H. J. Lipkin, and N. Meshkov, Nuclear Physics 62, 211 (1965).
- [34]A. Shahbani and D. A. Lidar, Phys. Rev. A 71, 020101 (2005).
- [35]A. A. Budini, Phys. Rev. E 89, 012147 (2014).
- [36]C. Sutherland, T. A. Brun, and D. A. Lidar, Phys. Rev. A 98, 042119 (2018).
- [37]H. Zhang, B. Pokharel, E. Levenson-Falk, and D. Lidar, Phys. Rev. Appl. 17, 054018 (2022).
- [38]J. G. Nila and T. A. Brun, Phys. Rev. A 113, 022442 (2026).
- [39]S. Maniscalco and F. Petruccione, Phys. Rev. A 73, 012111 (2006).
- [40]S. Maniscalco, Phys. Rev. A 75, 062103 (2007).
- [41]See supplemental information for details.
- [42]R. Ghosh, B. Das, and V. Mukherjee, Phys. Rev. B 113, 174302 (2026).
- [43]R. Ghosh and Y. Dubi (2026), manuscript in preparation.
- [44]R. Yousefjani, K. Sacha, and A. Bayat, Phys. Rev. B 111, 125159 (2025).
- [45]Y. Bar Lev and A. Lazarides, Phys. Rev. Lett. 133, 200401 (2024).
- [46]E. Vlachos, H. Zhang, V. Maurya, J. Marshall, T. Albash, and E. M. Levenson-Falk, Phys. Rev. A 106, 062620 (2022).
- [47]B.-H. Liu, L. Li, Y.-F. Huang, C.-F. Li, G.-C. Guo, E.-M. Laine, H.-P. Breuer, and J. Pillo, Nat. Phys. 7, 931 (2011).
- [48]Á. Rivas, S. F. Huelga, and M. B. Plenio, Phys. Rev. Lett. 105, 050403 (2010).
- [49]M. J. Hall, J. D. Cresser, L. Li, and E. Andersson, Physical Review A 89, 042120 (2014).
- [50]G. A. White, C. D. Hill, F. A. Pollock, L. C. Hollenberg, and K. Modi, Nature Communications 11, 6301 (2020).


