WorksPhysics and Astronomy
Thermodynamic cost of locally maintaining a steady state
Abstract
The Gibbs states of local Hamiltonians have short-range conditional correlations. To maintain a steady state with long-range conditional correlations, one might expect that a stochastic process must ``pay'' either in nonlocal moves or entropy production. Surprisingly, we find strictly local Gaussian processes that obey detailed balance and have unique steady states with arbitrarily long-range conditional correlations. We show that such processes must still produce entropy, if implemented using a local quadratic Hamiltonian coupled locally to equilibrium, rapidly relaxing baths. We lower-bound this entropy production in terms of the steady state's conditional mutual information (CMI); if the CMI decays exponentially with distance, the associated length-scale is called the Markov length . The entropy production per site per relaxation time obeys for large , and with an implementation-dependent constant. We also bound the tradeoff between coupling range and entropy production. We contrast these Gaussian processes with jump processes on discrete state spaces, for which, under certain irreducibility assumptions, steady states with long-range CMI are only possible if detailed balance is violated. We conjecture (and present evidence) that rapidly mixing local jump processes cannot stabilize steady states with long-range CMI.
Introduction
Gibbs states of local Hamiltonians are unusual states, and the canonical processes that stabilize them — Glauber dynamics and its variants — are unusual processes. In both cases what is unusual can be stated precisely. By the Hammersley–Clifford theorem, local Gibbs states are precisely the full-support distributions with the Markov property: if we divide the system into regions , , such that the Hamiltonian decomposes as , then conditioning on the state of removes all correlations between and — the conditional mutual information (CMI) vanishes. The analogous constraint for processes like Glauber dynamics is detailed balance with respect to the target steady state. Detailed-balance processes model the interaction of the system with a heat bath; in the steady state the system has reached equilibrium with the bath, and no entropy is produced. One might expect that any local process producing no entropy in its steady state is of this Glauber type — in detailed balance with respect to some local Hamiltonian — and hence has a Markov steady state. Processes that violate detailed balance generate entropy, quantifiable in terms of the violation [1]. In many settings, however, the process is a means of creating a steady state, so the question of most interest is: how much entropy must be produced to maintain a given steady state with a local process?
The present work addresses this question, both for Gaussian processes (where explicit calculations are available) and for interacting lattice systems (where our results are more heuristic). The answer has some surprising features. First, discrete and continuous state spaces seem fundamentally different. One might have expected, on intuitive grounds, that a local process satisfying detailed balance stabilizes a local Gibbs state. While true for discrete jump processes with irreducible frozen blocks and full-support steady states, this is not true for general Gaussian processes: we present an explicit counterexample — a hyperuniform steady state with exactly Coulomb conditional correlations. Even though the counterexample satisfies detailed balance — so its entropy production is invisible within stochastic thermodynamics — we show that any implementation of it as a complete, locally interacting Gaussian system with bounded quadratic stiffness and finite-range linear couplings to heat baths must nevertheless produce entropy, and we lower-bound that entropy production in terms of the steady state’s long-range CMI.1 Suppose the CMI decays at long distances with a characteristic *Markov length : with O(1) amplitude . For Gaussian field theories coupled to multiple baths at different temperatures we show that, in cutoff units (we set the ultraviolet cutoff throughout), , where is the coupling range. For our regulated counterexample at fixed range the entropy production is at least per site per characteristic relaxation time, and we give an explicit nearest-neighbor implementation attaining this scaling.2 In particular, the algebraic-CMI endpoint of this family requires infinite entropy production within this implementation class. That these processes are out of equilibrium is witnessed by a violation of the fluctuation-dissipation theorem: if they are local, their response to perturbations of the Hamiltonian necessarily reveals a mode-dependent temperature.
For lattice models with a finite local state space, the situation is strikingly different: assuming irreducibility of frozen blocks and full support, one can show that any process with a long-range CMI steady state must violate detailed balance; one can, moreover, quantify this violation (and thus the entropy production) under a uniform entropy-relaxation assumption for the frozen blocks. Whether this violation is achievable, we have not been able to prove. We conjecture that for strictly classical processes it is not: shot noise appears to fundamentally limit the achievable Markov length. We show, moreover, that for binary finite-state hidden Markov models with a strictly positive hidden transition matrix and an active odd-flip rule, a long-range CMI is not achievable; this suggests, but does not establish, that rapid-mixing processes cannot reach steady states with long-range CMI. Adapting a very recent result of Ref. [2], however, we observe that quantum spin systems with a discrete state space can violate this constraint, and, with a local open-boundary completion, maintain unique rapidly mixing steady states with long-range CMI. The thermal implementation we give has an infinite entropy production rate at this endpoint.
The rest of this work is structured as follows. § 2 briefly reviews entropy production and CMI. § 3 addresses linear stochastic field theories, for which all relevant quantities are Gaussian integrals. § 4 turns to local jump processes on discrete state spaces, focusing on processes that relax rapidly to their unique steady state. Finally, § 5 summarizes our results and lists open questions.
Context
What entropy production means
In the fundamental picture, the system is in contact with several thermal baths at distinct temperatures. In steady state the system itself is unchanged, but energy flows through it, from hot baths to cold ones. For two baths at temperatures coupled to a single overdamped degree of freedom in a unit-stiffness trap, with mobilities , the steady heat current is [we will define this carefully for many-body systems later], and the entropy production [3] is the Clausius rate3
This physical entropy production depends on the details of how the system is coupled to its baths. What we normally observe is the marginal process acting on the system alone, and this does not fix the implementation uniquely. Without further qualifications, therefore, “the entropy production needed to maintain a given state” is ill-defined. At one extreme, coupling the system to a mixture of two otherwise identical heat baths at different temperatures can give rise to a marginal process that is detailed balanced at some intermediate temperature: thus, a given physical implementation can produce “superfluous” entropy that can be eliminated without changing either the steady state or any of the observed rates. At the other extreme, any full-rank state is for some , and Glauber dynamics with respect to that maintains it with a single bath at zero entropy production. When we talk about the entropy production needed to maintain a steady state we really mean a constrained infimum over all local implementations that maintain as the system’s steady state.
The standard entropy production in stochastic thermodynamics is a different constrained infimum: over all implementations (local or not) that reproduce each rate of the marginal process. This infimum is [1, 4, 5, 6]
and is attained by assigning each inverse pair of transitions to a separate bath, whereupon the bath temperature and exchanged heat enter only through the combination (splitting a pair across several baths only adds to the total, by the log-sum inequality). The concept we focus on and the standard one are similar in both being infima, but under inequivalent constraints. A given marginal process that is promised to have a local implementation must satisfy both constraints, and must therefore produce at least as much entropy as the more stringent of these bounds.
Conditional mutual information
Almost everywhere in this work we will be concerned with the classical CMI, , which is simply the average mutual information of the conditional distribution: . In terms of Shannon entropies one can write , where subscripts denote subsystems. This latter definition carries over to quantum systems if one replaces Shannon entropies with von Neumann entropies. Our main partitioning geometry is the annular one shown in Figure 1(a), where separates and by some distance at its narrowest. For a full-support classical distribution on a finite graph, vanishing CMI for every separating tripartition is equivalent to a Gibbs representation by interactions on the graph cliques—this is the Hammersley-Clifford theorem. In particular, a range- Hamiltonian gives zero CMI for every complete shielding partition wider than . Hammersley-Clifford has not been fully generalized yet for quantum Hamiltonians, but see Refs. [7, 8, 9, 10]. In both classical and quantum settings, a small CMI implies efficient local reconstruction: specifically, if for fixed , one can approximately retrieve after is erased, by a recovery channel acting on , with error controlled by the CMI [7, 11, 12]. Thus the CMI plays a central part in the theory of quantum error correction, mixed-state phases of matter [13, 14], and learnability transitions [15, 16]—including retrieval in neural-network diffusion models [17]. The ubiquity of the CMI in these settings makes it natural to ask how one might control it, and what resources such control would require.
Figure 1. (a) The annular tripartition used throughout: a buffer of width separates a region from the rest of the system ; in the annulus degenerates to the two buffer segments of the ring geometry used in panel (b). (b) Exact Gaussian conditional mutual information of the near-Coulomb state (5), computed from covariance determinants, with an infinitesimal regulator (, ) keeping the CMI finite. : ring of sites, blocks and 16, two-sided buffer of width per side. Markers: exact evaluation; lines: the determinant formula (23) of Appendix A. : torus of sites, a disk of radius 5, an annulus of width .
An important distinction between CMI and other standard information-theoretic quantities is that CMI does not necessarily decrease under marginalizing: a Markov chain acting on can either increase or decrease [13, 18, 19, 20, 21]. In particular, marginalizing a Gibbs state will generically leave it non-Gibbs [22]: for example, the critical Gibbs states of a short-range Ising Hamiltonian has short-range CMI but long-range MI, so a Markov chain that simply deletes will replace the short-range CMI with the long-range MI. This is why it will be crucial in the rest of this work to focus on the case where one conditions completely on the state of .
What our results could not be
To orient the rest of the discussion, we note a few conjectures one might have been tempted to make but that are ruled out by elementary considerations or existing results.
Entropy production is a rate and the Markov length is not, so any relation between them involves entropy produced per timescale. For steady states the natural timescale is the one on which local deviations from the steady state (of the kind the process continually generates) relax, not the global mixing time: the stationary entropy budget concerns the continual relaxation of fluctuations, rather than preparation from an arbitrary initial state. Because the relaxation time diverges at an equilibrium critical point, our bounds constrain critical processes only weakly; indeed, arbitrarily weak detailed-balance-violating perturbations of a critical point — producing arbitrarily little entropy — generate long-range CMI, as shown very recently in Ref. [23].
Our results assume a unique steady state for the specified finite-volume dynamics; uniformity over volumes and frozen boundary conditions is stated separately. Dynamics with multiple steady states — conserving model-B dynamics, say — generate long-range steady-state CMI for free: under high-temperature charge-conserving equilibrium dynamics, any initial state of definite charge evolves to a distribution that is random apart from the global charge constraint, which suffices for long-range CMI, with no entropy produced.
Our results concern steady states, not transients. In the recent literature on finite-time phase transitions — e.g., spontaneous strong-to-weak symmetry breaking of product states under stochastic dynamics (see [24] for a recent review) — the CMI diverges at the finite-time transition, but not in the steady state of the associated process.
For discrete-time Markov chains it is straightforward to write down “prepare-and-replace” dynamics with long-range CMI in the steady state4: run stochastic dynamics from a fixed configuration up to the time at which the CMI diverges (as above), reset, and alternate indefinitely. The entropy accounting of such discrete-time protocols is deferred to future work.
Any bound must rely, not just on the Markov length, but also on the amount of CMI that is present at large distances. An illustrative example was very recently studied in Ref. [23], consisting of two non-reciprocally coupled Gaussian fields near criticality. Infinitesimal coupling suffices to mix the correlation and Markov lengths, but can clearly only produce entropy at an infinitesimal rate.
There can be no general converse, from detailed-balance violation to non-Gibbsian steady states: many processes violate detailed balance yet mix to the infinite-temperature state — for example, independent three-state spins whose only allowed transitions are , at equal rates.
Relation to previous work
The question addressed here—minimal entropy production needed to maintain a target state—was previously explored in Ref. [25]; our setup differs from this work, and other previous literature on entropy production, by explicitly imposing a spatial restriction on the allowed processes. The field theories we consider are closely related to those explored in Refs. [26, 27]; in particular, their “model AB” exhibits the phenomenon of a local process satisfying detailed balance with respect to a nonlocal parent Hamiltonian. The distinction between resolved and hidden entropy production is also central to coarse-grained stochastic thermodynamics [6, 28]. The steady state of the counterexample belongs to a family of states that have been explored in the context of hyperuniformity [29, 30, 31, 32]. The Markov property for quantum systems has seen considerable recent attention [9, 33, 34], though primarily for dynamics that is explicitly detailed balance with respect to a local parent Hamiltonian.
Finally, two very recent works (posted while this paper was nearing completion) address similar themes: Ref. [23] showed that the steady-state Markov length can diverge for equilibrium critical points subject to weak nonequilibrium perturbations (which automatically produce arbitrarily little entropy), and Ref. [2] presented an example of a quantum steady state with long-range Markov length, which offers an interesting contrast to the classical discrete setting, where these states seem not to exist.
Single-component Gaussian processes
Setup
In this section we consider statistically translation-invariant Ornstein–Uhlenbeck theories of the form
where is an -component real field; and are matrix-valued; is stable, i.e., has only eigenvalues with positive real part; and is symmetric positive semidefinite. Equivalently, the process is specified by its MSR action, . Real-space locality enforces that polynomials in , and reality of together with parity symmetry kills odd powers of in the small- expansion. We define the drift . Wavenumbers run over the band , with the ultraviolet (lattice) cutoff; we set . The stationary covariance solves ; the stationary measure is with , and naturally specifies an effective Hamiltonian. We can write
with the circulation [35, 36]: iff detailed balance holds.
For Gaussian states, correlations and mutual information are read off from the covariance , while conditional correlations and CMI are read off from its inverse, the “precision matrix” (or “effective Hamiltonian”) . Poles of are zeroes of and vice versa. Rational with complex poles gives exponentially decaying CMI, with a length scale ; real-axis singularities give power-law or non-decaying CMI. We use for this precision-decay scale below; in the one-dimensional setting, the Markov length is .
The counterexample
Equipped with these definitions, we can introduce the model that will illustrate many of the themes of this paper: a one-component field on a chain, with ultralocal drift and conserved nearest-neighbor noise,
Since , the stationary covariance is , and is symmetric: , and detailed balance holds. (The mode is noiseless but damped, so it pins to zero, and the steady state is unique.) All modes relax at the same rate. Correlations vanish identically beyond one site but
has a real-axis pole at , so the Markov length is infinite. For consecutive finite intervals in the infinite chain, the CMI is (App. § 7)
so it scales as for fixed and small, large, and for . The case of the full system requires care: the state space is continuous, and the CMI naively diverges unless regulated by adding infinitesimal noise. Figure 1 plots the regulated ring and the analogous two-dimensional geometry. The effective Hamiltonian is the confining one-dimensional lattice Coulomb potential, [16]. This example illustrates the counterintuitive fact that a completely local process can satisfy detailed balance with respect to a nonlocal effective Hamiltonian.
There are two very different ways of reading this process. The first is the “obvious” one: is coupled to two baths at different temperatures; the cold bath is at , and obeys no constraints, while the hot bath is at nonzero , and is restricted to charge-conserving moves. (The conserving bath also gives a -dependent damping term, but this is subleading at small and can be neglected.) Also available, though, is an equilibrium reading, according to which the fundamental field is , its Hamiltonian is (nonlocal in the variables), and it undergoes single-site Glauber updates at temperature , , where is white noise. Then , precisely Eq. (5). Thus, the process (5) itself is equally compatible with strongly nonequilibrium local dynamics and equilibrium dynamics with respect to a nonlocal Hamiltonian.
These two interpretations of § 3.2 describe the same differential equation, so any passive correlation measurement will yield identical answers in the two cases. In particular, as . However, the response to perturbations of the Hamiltonian distinguishes between them. We sketch the main ideas here; a derivation in the Keldysh framework is presented in App. § 8. For a Gaussian theory, the response function of a mode probes its intrinsic dynamics, independent of the occupation factors [37]. The effective temperature of mode , . On the Coulomb interpretation, the low-momentum modes are increasingly stiff, so their response is suppressed as , similar to their fluctuations: thus, approaches a constant, as it should since the system is in equilibrium with respect to a nonlocal Hamiltonian. On the nonequilibrium interpretation, the stiffness of the low- modes is a -independent constant, so at small . This yields . Thus the fluctuation-dissipation theorem is violated on the nonequilibrium interpretation. Crucially, the Coulomb interpretation requires a singular Hamiltonian stiffness, . For a local quadratic Hamiltonian with uniformly bounded stiffness , the static response is . Matching the vanishing fluctuations then forces to vanish as . Thus any such local interpretation must be nonequilibrium.
Generalizations
Even for a single-component field, the mechanism outlined in the previous section can be extended to give increasingly nonlocal CMI, by forcing the noise to come with higher powers of , e.g., through dipole moment conservation [38]. This mechanism can lead to a ladder of hyperuniform steady states, with for any integer ; all these steady states have diverging CMI, and all the associated processes are local and obey detailed balance. More generally, for any rational function , one can straightforwardly construct a local process that satisfies detailed balance and has a steady state with precision , as follows. Take and to be even polynomials with on the band and on the band, and choose
Both are polynomials, so the process is local. Since , every mode is damped: the steady state exists and is unique, with covariance , as desired. And is symmetric, so : detailed balance holds by construction. The counterexample is the case , . The representation is not unique: multiplying and by a common positive even polynomial yields a different local reversible process with the same steady state. (The construction generalizes verbatim to -component fields: the common denominator, , and positive definite on the band.)
In continuum notation the counterexample reads
with the non-conserving relaxation rate and the conserved-noise mobility. We use this form whenever dimensional constants matter, as in the bounds of § 3.4.
Entropy production
We have seen that the counterexample violates the fluctuation-dissipation theorem. We now turn to lower-bounding this violation, and the associated entropy cost, more generally. For simplicity we will continue to focus on the one-component, translation-invariant setting. For the explicit bounds we take and finite-range linear couplings to Markovian baths. The first thing we need to establish is that long Markov lengths cannot arise for a local field in equilibrium with a local bath. This is slightly less trivial than it sounds: the composite system obeys the Markov property by assumption, but we are concerned with the marginal over the bath, and the Markov length can grow (indeed diverge) under marginalization. For a particular mode, after integrating out the bath, the precision matrix for the fields is , where is the bath’s correlation function at . Stability of the state forces for every . The Markov length of the system’s marginal is set by the pole of that lies closest to the real axis. By assumption, are local, and do not have poles near the real axis. The danger is that might: in other words, integrating out a gapless mode can generate effective long-range interactions. Provided the bath has a finite correlation length , however, the nearest pole of is also gapped away from the real axis, so has no singularities near the axis, and cannot have a large Markov length. To summarize, marginalizing over the bath cannot generate a Markov length that is larger than the bath’s correlation length. (Note that a similar result for local state spaces was proved in the high-temperature limit in Ref. [21] — if we take the system and its bath to be in equilibrium at sufficiently high temperature and apply the erasure channel to the bath, Theorem 2 of that work states that the system’s marginal retains a finite Markov length.)
Any remaining, strictly local, route to a large Markov length must come from nonequilibrium effects. To parameterize these we will couple the system to a set of arbitrarily many baths, indexed by . Each bath is in equilibrium at a temperature and a characteristic, momentum-dependent Golden Rule Rate at which it couples to the system: , where is the interaction between system mode and the bath, and is the bath density of states. Using each bath’s FDT, and suppressing mode indices, the marginal process has , . Matching the dissipation , we get that , where and denotes the average over baths weighted by their couplings to mode . Rearranging the Clausius entropy-production formula gives
which is non-negative for every by Jensen’s inequality.
In general, achieving a large Markov length requires to have a zero near the real axis. Once the range of the process is capped, the only possible origin for this zero is from . Assume for simplicity that the zero is on the imaginary axis, and the behavior around it is quadratic: , . The precision length is (with the CMI convention specified in § 3.1). It is clear that ; moreover, if the coupling range is , is a trigonometric polynomial of degree in . By Bernstein’s inequality, a degree- polynomial that is bounded on the band can have curvature at most : a range- coupling cannot bend faster than times its height. This gives the scaling of the bound. Keeping the full Taylor series and centering between the noise floor and ceiling gives the sharper result
The argument above bounds the achievable Markov length in terms of the bath temperatures. Next, we fix the range and Markov length, and estimate the minimum entropy production consistent with these. We proceed as follows. Define , so the mode-wise entropy production is , with . Any choice of bath defines some valid , and for range , must be an order- polynomial in , no matter how many baths we couple to the system. To get a bound, we relax the problem and optimize over all polynomials rather than optimizing over bath assignments: we compute the “majorant bound”
This lower-bounds the entropy production regardless of the choice of bath. The physical intuition is particularly simple when is small. The target
has a narrow, tall peak: height and width . A polynomial with only Fourier harmonics cannot follow an arbitrarily narrow peak. To stay above it, the polynomial must also be large elsewhere. That unavoidable excess is entropy production. A helpful way to rewrite this bound is to note that at . Minimizing subject to these constraints gives the closed-form bound
For fixed range and large ,
Thus fixed-range thermal implementations necessarily pay a cost growing as . In particular, at nearest-neighbor range,
and an explicit two-bath construction attains it. We choose the following two baths:
Their coupling amplitudes are proportional to and , so both are nearest-neighbor couplings. Their entropy production is
Here and . At nearest-neighbor range, this choice saturates the bound. Figure 2 compares the cost of the explicit implementation with the lower bound from optimization; the bounds coincide at . The positive lower bound at every finite assumes strictly finite-range Markovian baths. As discussed earlier in this section, exponentially decaying CMI can instead arise at equilibrium by marginalizing over a bath with finite correlation length.
Figure 2. The range–dissipation tradeoff for , at and . Dashed curves: the lower bound from optimizing over (14). Solid curves: entropy production of the explicit bath construction (17). The bounds coincide at nearest-neighbor range. Both curves use the lattice momentum measure; the lower bound is not asserted to be physically attainable at higher range.
Jump processes on discrete spaces
In this section we consider jump processes that are defined on discrete systems (e.g., Ising spins or lattice gases): the local state space has a finite number of states, and these are locally coupled by jump moves of bounded support: a configuration is updated to , where differs from only locally, and the probability of the jump happening only depends on the local neighborhood of (within some range ). We will see that results of the previous section are surprisingly fragile to discreteness.
Detailed balance + locality implies locality of . The counterexample cannot arise for full-support jump processes with irreducible frozen blocks [39] (i.e., a sufficiently large block has a unique steady state if one runs its dynamics freezing the boundaries): the locality of follows from that of the moves. Suppose the stationary measure is . Detailed balance gives . Since the lhs is local, so is the rhs: every -difference between pairs of connected configurations is local. If all locally different pairs of configurations were connected by local moves, then this would immediately imply the locality of . This last implication follows if we make one of two assumptions. If we assume that the model consists of single-site updates, conditioned on some finite window around the site, such that the rate is strictly positive for any state on the window, then any pair of configurations is connected by local moves and is manifestly local. More generally, assume that the chain of moves supported inside is irreducible for every fixed exterior. If shields from , the frozen generator on depends only on . Denote its stationary state as . Detailed balance means the generator of the dynamics is frustration-free, so the global steady state is also the steady state of the frozen chain, so the conditional distribution is also a steady state for each ; uniqueness then gives , and hence . Thus, there is a much tighter relation between detailed balance and Hammersley-Clifford in this setting: under this irreducibility assumption, a full-support local detailed-balance process must stabilize a local Hamiltonian.
If one tries to derive a continuum process as the limit of a jump process, the argument above fails because each continuum increment sums over a large number of jumps, so channel-by-channel detailed balance is lost, and the weaker notion that survives for the continuum increments is unable to guarantee a local parent Hamiltonian (as the counterexample shows).
Hyperuniform states cannot be reached rapidly. A natural attempt to discretize the counterexample might run as follows. Consider a system with three states per site, , indexed by their charge , subject to the following two baths: a non-conserving zero-temperature bath that favors the state , and drives both other states to it; and a finite-temperature bath that conserves , and allows nearest-neighbor moves such as and . This fails to generate a hyperuniform steady state: processes like generate shot-noise fluctuations of the total charge, so the structure factor at zero momentum (which measures ) is nonzero.
The same obstruction applies whenever local charge-changing activity is nonzero and the local drift has summable correlations. Suppose that the jump process includes some charge-changing terms , each of which changes the charge by (which in the above example is ). Define the signed drift and the noise activity . For a steady state with summable correlations of the local drift observables, for some constant . In the steady state, and is invariant. There are two contributions to its evolution, coming from the deterministic drift and the accumulation of variance from random increments. Balancing these gives . By Cauchy–Schwarz, , and assembling these observations gives us the bound
so shot noise blocks hyperuniformity under these assumptions when . The shot-noise floor is set by the minimum charge increment , and one recovers the counterexample in the previous section for a suitable limit.
Is rapid mixing compatible with long-range CMI? Under the assumptions above, the argument ruled out the possibility of getting a long-range CMI through hyperuniformity. One might wonder whether there are other means to achieve a long-range CMI in the unique steady state of a jump process that mixes rapidly (as we were able to with a rapidly relaxing continuous process). We conjecture that this is impossible, based on the following heuristic argument. Since the process mixes rapidly and its steady state is unique, the steady state cannot have long-range order. All known examples where long-range conditional correlations persist in the absence of long-range order have some form of “string order,” i.e., there is a product of observables supported in that is ordered, though each individual factor is random. String order allows and to be correlated conditional on the state of , even at early times (the string is always extended so it is not subject to light cones). However, in the steady state of a system with asynchronous stochastic jumps, the string order parameter in region will be flipped by single events in unless it is protected by some global symmetry of the dynamics (as it is for spontaneous strong-to-weak symmetry breaking [24]), or by some other structure such as synchronous updates designed to ensure that these local events cancel. So any string decorrelates in its bulk, and long-range CMI cannot be achieved through this means. This step of the heuristic is in fact a theorem, proved in App. § 9: for a one-dimensional local jump process with a translation-invariant, spatially ergodic stationary law, persistent string order (with arbitrary bounded endpoint decorations) forces the string-flipping activity to vanish exactly; with full local support this is an exact conservation law of the dynamics, contradicting uniqueness of the stationary state on periodic chains. The remaining gap between this statement and the main conjecture is the first step of the heuristic — that long-range CMI at short times must arise through a string. App. § 10 closes this gap for binary steady states with an exact finite-state hidden-Markov representation, a strictly positive hidden transition matrix, and an active odd-flip rule, where exponential decay of the CMI itself can be proved; beyond that class, we are unable to rule out currently unknown mechanisms that avoid the “string order” route.
Discussion
In terms of entropy production, the discrete and continuous classical worlds are surprisingly different: in the former, assuming full support and irreducible frozen blocks, a local process can only be in detailed balance with respect to a local Hamiltonian; in the latter, this restriction is gone, and the entropy production mandated by locality is “hidden” from the standard definitions of stochastic thermodynamics. The two worlds also differ in the nonequilibrium steady states they admit: in the discrete world, long-range CMI seems unachievable in the unique steady state of any rapidly mixing process, while the continuous world admits many straightforward examples with rapid drift relaxation. Interestingly, in this respect the discrete quantum world looks more like the continuous classical one, admitting modes with no noise floor and rapid-mixing processes that generate long-range CMI [2]. (Interestingly, our thermal implementation of this quantum example also requires infinite entropy production to sustain its steady state; a thermally regularized version pays entropy logarithmically in the Markov length, — see App. § 11.)
Finally, we comment briefly on interacting mechanisms that evade the trade-offs of § 3.4. The linear noise floor from the cold bath can be circumvented by tightly coupling the system to a nonequilibrium ratchet, which in turn is coupled to heat baths [40], such that every “model A” damping event coincides with a heat quantum transported from a hot to a cold bath, as in autonomous Maxwell demons [41, 42]. In this setting, wrong-way transitions are suppressed by with — an effective temperature exponentially below the coldest reservoir, costing per corrective event: effectively, this mechanism replaces the scaling , allowing for an exponentially larger at fixed . However, all such strategies we have found require discrete events, so they cannot increase past the shot-noise floor of Eq. (19). Thus the general question of minimizing at fixed for interacting systems remains open.
Our results raise many natural questions for future work. The most immediate one is to settle the conjecture on achieving long-range CMI in discrete jump processes (and to understand entropy production in such processes, if they exist). Other natural questions include generalizing our results to find a relation between quantum CMI and KMS detailed balance (cf. the quantum Gibbs sampler literature [43]); identifying steady-state phase transitions at which the Markov length diverges, and understanding their thermodynamic signatures; extending our entropy-production floors beyond Gaussian theories to generic interacting field theories; and more generally incorporating our spatial locality restrictions (and Hammersley-Clifford) into the framework of stochastic thermodynamics.
Acknowledgements
I used Claude Fable 5 and ChatGPT Astra 6 to develop the calculations and proofs in this work. I thank Yu-Hsueh Chen, Matthew Fisher, Sam Garratt, Tarun Grover, Hyunsoo Ha, Jake Hauser, Tim Hsieh, David Huse, Jérôme Lloyd, Nick O’Dea, Shengqi Sang, Romain Vasseur, Cenke Xu, Carolyn Zhang, and Frank Zhang for helpful discussions.
CMI of the counterexample
This appendix computes the conditional mutual information plotted in Figure 1. The state is the regulated counterexample on a ring of sites, ; equivalently,
with and independent standard Gaussians — heights with unit prior variance, plus a white regulator. We take to consist of two segments (or an annulus) of width .
For jointly Gaussian variables, , where are the “canonical correlations,” i.e., the singular values of the matrix , and the ’s are conditional covariance matrices. For the equal-width buffers in App. A, both canonical correlations have simple exact formulas, including at nonzero regulator. Write , , , and define
For , the two nonzero canonical correlations are
All remaining canonical correlations vanish, and
In the small-regulator regime ,
We expect a similar decomposition to hold for the annular geometry in : the sum over two channels is replaced by one over multipoles of the enclosed charge. We have not evaluated this in closed form: the plotted CMI in the main text is computed directly from covariance determinants.
Keldysh analysis of response in the counterexample
In this appendix we show, using the Keldysh formalism, how the two implementations of the counterexample naturally give rise to distinct response functions, even though the processes are identical absent an external perturbation. We use the fact that coupling a system to an equilibrium bath at temperature and mobility adds the following term to the Keldysh action:
The two readings of the counterexample have different Hamiltonians, and different system-bath couplings, so as to cancel out in the marginal process. The source is always added to as .
Implementation A: long-range Hamiltonian, one bath. Coulomb Hamiltonian , one current-coupled (conserving) bath at temperature . Assembling one block:
the conserving mobility cancels the Coulomb kernel in the drift (a local current driven by the gradient of a long-range potential has a local divergence), so , , , and the single bath is in detailed balance: .
Implementation B: trivial Hamiltonian, two baths. ; a cold density-coupled bath and a hot conserving bath :
In the limit , with fixed (the cold block loses its noise, the hot block its friction), and under the dictionary , , the parts of (26) and (27) become identical term by term.
Response. The source terms do not agree: implementation A carries , implementation B carries . Hence
with identical .
Persistent string order forces a conserved charge
This appendix proves the claim of § 4 that in the steady state of a local jump process, string order cannot survive unless it is protected by an exact conservation law of the dynamics. (If a conservation law exists, it prevents rapid mixing.) The strategy relies on translational invariance and spatial ergodicity, which is a consequence of the short-range correlations in the steady state. We also assume that the marginal distribution in the steady state has full support on the interaction range—i.e., for our two-site interaction window we assume that all four bitstrings occur with nonzero probability. Spatial ergodicity states that there are no long-range connected correlations in space. Translation-invariant, spatially ergodic distributions obey Birkhoff’s theorem, which states that for any bounded local observable , the space-average of any observable converges to its ensemble average. To keep the notation simple we will state the theorem for bits with two-site updates, and consider the parity string , but the argument manifestly works for any finite-size alphabet and any string that is a product over sites. The theorem states:
Theorem 1. (informal). Under the assumptions above, as in the steady state.
Proof. We sketch the proof. Again, to keep the combinatorics transparent we assume that all updates act on nearest-neighbor pairs of sites. Define as the probability that an update acting on that configuration will flip the value of (for inside the string), and define where the average is over the steady-state probability distribution (and is independent of by assumption). The Markov process generator , acting on , gives
where are boundary terms of order unity that can be bounded by a constant . Since by stationarity [we will drop the subscript from now on], we can bound
Define . Since the magnitude of is at most 1, we can also bound
By Birkhoff’s theorem the latter quantity goes to zero, so for large , by Eq. (30). So either or . Since is an average over non-negative rates, and the steady state has support on all configurations, if on average the rate at which parity flips must be identically zero. If this is true then parity is conserved and the dynamics is non-ergodic, contradicting our assumptions.
The value of this proof, and the reason it improves on the argument in the main text, is that we made no assumptions about the correlation between the value of the string and the local rates in the region it covers.
Short-range CMI for steady states with hidden Markov structure
This appendix states and proves the one-dimensional obstruction announced in § 4:
Theorem 2 (finite-state case). Let be a translation-invariant stationary law on of a finite-range jump process with bounded rates and at least one nonzero bulk rule flipping an odd number of spins, and suppose has an exact edge-emitting hidden-Markov representation with hidden states and strictly positive hidden transition matrix, . Then the two-sided CMI across a buffer of width obeys , with depending only on for deterministic readout — and only on for stochastic readout whose nonzero labeled transition probabilities are — uniformly in the jump rates, their range, and the entropy production.
Two-state hidden layer. First we consider a much simpler version of the general problem. We take a system with hidden bits on every edge, visible spins on every vertex (right of edge ), and a deterministic function coupling them (this is called “deterministic readout”). The underlying probability distribution is over , which obey a Markov chain in the spatial direction, but “reads” this Markov chain through the deterministic function. In this simple case there are only sixteen functions that could exist, and they fall into four classes: (i) two constant functions, which lead to product distributions; (ii) four cases where , where inherits the Markov property; (iii) eight functions for which at least one output reveals and , so (by the Markov property) conditioning on decouples the left and right states on and thus (by data processing) on , so the CMI vanishes; and (iv) two functions . These parity functions are the only case that might admit long-range CMI. The mechanism for long-range CMI is, however, precisely string order: . A zero-dimensional Markov chain with entirely positive rates certainly mixes, so the result in the previous section rules out this case, and (for a two-dimensional hidden state space) thus rules out long-range CMI.
General case: strategy. Either visible observations erase hidden-state ambiguity, or a hidden parity survives. In the first case, there are some configurations in which the hidden variables are revealed to the visible ones, and this decorrelates the hidden Markov chain and cuts off the CMI. This intuition is converted to an exponential bound by Travers’s result [46]. In the second case, either the hidden parity produces string order (in which case we rule it out as before) or it is completely hidden, in which case it does not affect any properties of the visible layer and thus does not contribute CMI.
Predictive memory. Set
The entropy chain rule gives
Thus exponential convergence of suffices. All these quantities are finite: past and future communicate through a hidden state, so their mutual information is at most .
The binary support dichotomy. Two hidden states are mergeable if they can emit the same word and reach the same hidden state. They are immediately incompatible if they cannot emit the same first symbol. Either every pair merges within two steps or is immediately incompatible, or
for a nonconstant hidden sign .
Indeed, take a pair violating the first alternative. Failure to merge in one step, together with and the binary alphabet, forces their labels to every destination to be deterministic and opposite, say and . Immediate compatibility makes nonconstant. Failure to merge in two steps then forces every row indexed by to be opposite every row indexed by . All positive rows therefore equal some row , and all negative rows equal . Comparing with row gives , hence
the claimed parity form.
In the nonexceptional case, Travers’s theorem gives exponential convergence of [46 Theorem 2 and Remark 3]. For uniformity, his finite-block coupling uses a block length bounded by some : there are only finitely many labeled support patterns. Positive block entries are at least , and stationary hidden weights are at least . His explicit estimates therefore give uniform constants.
An asymmetric parity sector produces string order. Absorb into the visible spins and write
If the law of is not invariant under , some bounded odd function of a finite block has mean . The functions and are sign-even, so they depend only on the corresponding visible endpoint blocks. Consequently there are bounded decorations such that
Mixing of the positive hidden chain gives .
Let the update window have width , including every spin read or flipped, and let be the total rate of odd-spin flips in its translate beginning at . Write . Interior odd flips negate , while only finitely many windows touch its endpoints. Stationarity therefore gives
The visible process is spatially ergodic, so in . Since is uniformly bounded, division by gives . But every finite visible word has positive probability in the parity representation: choose an initial sign, propagate it using the visible word, and choose hidden states in the required sign classes. Thus the nonzero odd rule implies , a contradiction. The law of must be sign-symmetric.
An invisible symmetry does not affect anything. The only case that remains is the trivial one: the hidden layer has an exact symmetry that nothing in the visible layer ever detects (a “strong symmetry” of the spatial evolution). In this case it produces no structure in the visible layer and can be ignored.
Entropy production of a quantum steady state with long-range CMI
Model. Write and , with Majoranas. The open-chain fermionic model of Ref. [2], with boundary damping completed, has jumps
where and . For a local spin realization, introduce auxiliary fermions , replace each jump anchored at by , and add unit-rate jumps (), (), and . Jordan–Wigner ordering gives jumps on at most four consecutive spins; is the leftmost Pauli and removes the parity degeneracy. The presence of boundary terms ensures a unique steady state in the spin representation; rapid mixing and long-range CMI were already established. The origin of the long-range CMI is that the Lindbladian contains no terms repopulating the mode of the fermions. Indeed, in steady state, , so the steady state is hyperuniform, analogous to our classical example.
Implementation. We implement the system-bath coupling in the “collision model” framework: system fermionic operators are alternately coupled to ancillas coupling respectively to the and fermions. Each ancilla has Hamiltonian , and is initialized at temperature , i.e., excited-state probability . The ancilla couples to the system via , where raises the ancilla. The time-dependent drive is needed to allow jump operators to try to “cool” fermions in distinct bases. In the weak-coupling limit the baths have the following jump operators:
At , both channels empty the mode: pumping vanishes there, but loss remains finite. For the Markov length scales as . Given this data and the steady-state occupation factors, the energy current into each bath is just the change in its occupation, and the entropy production per site and relaxation time is
This argument was for an explicit bath but it manifestly applies to any bath that cools the mode to vacuum at unit rate: the combination of rapid dissipation and occupation near zero forces the cold bath to rapidly absorb heat, leading to a diverging entropy production rate in general.
References
- [1]U. Seifert, Stochastic thermodynamics, fluctuation theorems and molecular machines, Rep. Prog. Phys. 75, 126001 (2012).DOI
- [2]A.-R. Negari, Y. Zou and T. H. Hsieh, Instability of Markov chains under infinitesimal coherent dissipation, ArXiv:2610.18983 (2026).arxiv.org/abs/2610.18983
- [3]C. Godrèche and J.-M. Luck, Characterising the nonequilibrium stationary states of Ornstein–Uhlenbeck processes, J. Phys. A 52, 035002 (2019).DOI
- [4]J. Schnakenberg, Network theory of microscopic and macroscopic behavior of master equation systems, Rev. Mod. Phys. 48, 571 (1976).DOI
- [5]M. Esposito and C. Van den Broeck, Three faces of the second law. I. Master equation formulation, Phys. Rev. E 82, 011143 (2010).DOI
- [6]A. Gomez-Marin, J. M. R. Parrondo and C. Van den Broeck, Lower bounds on dissipation upon coarse graining, Phys. Rev. E 78, 011107 (2008).
- [7]P. Hayden, R. Jozsa, D. Petz and A. Winter, Structure of states which satisfy strong subadditivity of quantum entropy with equality, Commun. Math. Phys. 246, 359 (2004).
- [8]T. Kuwahara, K. Kato and F. G. S. L. Brandão, Clustering of conditional mutual information for quantum Gibbs states above a threshold temperature, Phys. Rev. Lett. 124, 220601 (2020).DOI
- [9]C.-F. Chen and C. Rouzé, Quantum gibbs states are locally markovian, arXiv preprint arXiv:2504.02208 (2025).arxiv.org/abs/2504.02208
- [10]A. Bakshi, A. Liu, A. Moitra and E. Tang, A dobrushin condition for quantum markov chains: Rapid mixing and conditional mutual information at high temperature, In Proceedings of the 58th Annual ACM Symposium on Theory of Computing, pp. 1508–1516 (2026).arxiv.org/abs/2510.08542
- [11]O. Fawzi and R. Renner, Quantum conditional mutual information and approximate markov chains, Communications in Mathematical Physics 340(2), 575 (2015).
- [12]M. Junge, R. Renner, D. Sutter, M. M. Wilde and A. Winter, Universal recovery maps and approximate sufficiency of quantum relative entropy: M. junge et al., In Annales Henri Poincaré, vol. 19, pp. 2955–2978. Springer (2018).
- [13]S. Sang and T. H. Hsieh, Stability of mixed-state quantum phases via finite Markov length, Phys. Rev. Lett. 134, 070403 (2025).
- [14]S. Sang, L. A. Lessa, R. S. K. Mong, T. Grover, C. Wang and T. H. Hsieh, Mixed-state phases from local reversibility, Phys. Rev. X 16, 041014 (2026), doi:10.1103/1kr5-jlcb.DOI
- [15]T. A. Kumar, Y. Zou, A.-R. Negari, R. G. Melko and T. H. Hsieh, Unlearnable phases of matter, ArXiv:2602.11262 (2026).arxiv.org/abs/2602.11262
- [16]J. Hauser, K. Su, H. Ha, J. Lloyd, T. G. Kiely, R. Vasseur, S. Gopalakrishnan, C. Xu and M. Fisher, Strong-to-weak symmetry breaking in open quantum systems: From discrete particles to continuum hydrodynamics, arXiv preprint arXiv:2602.16045 (2026).arxiv.org/abs/2602.16045
- [17]F. Hu, G. Liu, Y. F. Zhang and X. Gao, Local diffusion models and phases of data distributions, arXiv preprint arXiv:2508.06614 (2025).arxiv.org/abs/2508.06614
- [18]S. Sang, Z. Li, T. H. Hsieh and B. Yoshida, Ultrafast entanglement dynamics in monitored quantum circuits, PRX Quantum 4, 040332 (2023), doi:10.1103/PRXQuantum.4.040332.DOI
- [19]Y. Zhang and S. Gopalakrishnan, Nonlocal growth of quantum conditional mutual information under decoherence, Phys. Rev. A 110, 032426 (2024), doi:10.1103/PhysRevA.110.032426.DOI
- [20]S.-u. Lee, C. Oh, Y. Wong, S. Chen and L. Jiang, Universal spreading of conditional mutual information in noisy random circuits, Phys. Rev. Lett. 133, 200402 (2024), doi:10.1103/PhysRevLett.133.200402.DOI
- [21]Y. Zhang and S. Gopalakrishnan, Conditional mutual information and information-theoretic phases of decohered Gibbs states, ArXiv:2502.13210 (2025).arxiv.org/abs/2502.13210
- [22]A. G. D. van Enter, R. Fernández and A. D. Sokal, Regularity properties and pathologies of position-space renormalization-group transformations: scope and limitations of Gibbsian theory, J. Stat. Phys. 72, 879 (1993).DOI
- [23]Y.-H. Chen and T. Grover, Markov length can diverge in systems whose universal physics is spatially markovian, arXiv:2610.11235 (2026).arxiv.org/abs/2610.11235
- [24]C. Wang, Strong-to-weak spontaneous symmetry breaking, arXiv:2606.02555 (2026).arxiv.org/abs/2606.02555
- [25]J. M. Horowitz and J. L. England, Information-theoretic bound on the entropy production to maintain a classical nonequilibrium distribution using ancillary control, Entropy 19, 333 (2017).arxiv.org/abs/1707.00367
- [26]C. Nardini, É. Fodor, E. Tjhung, F. van Wijland, J. Tailleur and M. E. Cates, Entropy production in field theories without time-reversal symmetry, Phys. Rev. X 7, 021007 (2017).DOI
- [27]Y. I. Li and M. E. Cates, Steady state entropy production rate for scalar Langevin field theories, J. Stat. Mech. p. 013211 (2021).DOI
- [28]M. Esposito, Stochastic thermodynamics under coarse graining, Phys. Rev. E 85, 041125 (2012).pubmed.ncbi.nlm.nih.gov/22680437
- [29]D. Hexner and D. Levine, Noise, diffusion, and hyperuniformity, Phys. Rev. Lett. 118, 020601 (2017).DOI
- [30]S. Torquato, Hyperuniform states of matter, Phys. Rep. 745, 1 (2018).DOI
- [31]R. Maire and L. Chaix, Hyperuniformity and conservation laws in non-equilibrium systems, J. Chem. Phys. 163, 214507 (2025), ArXiv:2509.04242.DOI
- [32]M. Casiulis, S. Anand and S. Martiniani, Hyperuniform systems are maximally irreversible, ArXiv:2607.07411 (2026).arxiv.org/abs/2607.07411
- [33]A. Bakshi, A. Liu, A. Moitra and E. Tang, A dobrushin condition for quantum markov chains: Rapid mixing and conditional mutual information at high temperature, ArXiv:2510.08542 (2025).arxiv.org/abs/2510.08542
- [34]P. Rosa-Ruiz, M. Scandi, Á. Capel and Á. M. Alhambra, Static features from mixing in short- and long-range lindbladians: Markov property and correlations, ArXiv:2606.28054 (2026).
- [35]K. Tomita and H. Tomita, Irreversible circulation of fluctuation, Prog. Theor. Phys. 51, 1731 (1974).
- [36]R. Graham, Macroscopic potentials, bifurcations and noise in dissipative systems, In Noise in Nonlinear Dynamical Systems. Cambridge University Press (1989).
- [37]A. Kamenev, Field Theory of Non-Equilibrium Systems, Cambridge University Press, Cambridge (2011).
- [38]A. Gromov, A. Lucas and R. M. Nandkishore, Fracton hydrodynamics, Phys. Rev. Research 2, 033124 (2020).DOI
- [39]F. Martinelli, Lectures on glauber dynamics for discrete spin models, In Lectures on probability theory and statistics: Ecole d’été de probabilités de saint-flour xxvii-1997, pp. 93–191. Springer (2004).
- [40]C. Stahl, E. Lake and V. Khemani, Brownian ratchets and pumps universally simulate many-body active dynamics, ArXiv:2607.01231 (2026).arxiv.org/abs/2607.01231
- [41]T. L. Hill, Free Energy Transduction in Biology, Academic Press, New York (1977).DOI
- [42]P. Strasberg, G. Schaller, T. Brandes and M. Esposito, Thermodynamics of a physical model implementing a Maxwell demon, Phys. Rev. Lett. 110, 040601 (2013).pubmed.ncbi.nlm.nih.gov/25166147
- [43]C.-F. Chen, M. J. Kastoryano and A. Gilyén, An efficient and exact noncommutative quantum Gibbs sampler, ArXiv:2311.09207 (2023).arxiv.org/abs/2311.09207
- [46]N. F. Travers, Exponential bounds for convergence of entropy rate approximations in hidden Markov models satisfying a path-mergeability condition, Stoch. Proc. Appl. 124, 4149 (2014), Theorem and remark numbering as in arXiv:1211.6181v2.arxiv.org/abs/1211.6181

