Empirical Entropy Production Rate

• Empirical entropy production rate is a measure of irreversible dynamics extracted from observable data, highlighting the breaking of detailed balance.
• It is formulated in various equivalent forms including entropy balance, path-space relative entropy, and quadratic current functionals in diffusive processes.
• Observable-based estimators adapt to diverse frameworks such as jump processes, Langevin dynamics, and coarse-grained systems to assess dissipation.

Empirical entropy production rate (EPR) is an operational measure of irreversibility extracted from observable data rather than assumed from complete microscopic knowledge of all transition mechanisms. In the cited literature it appears in several mathematically equivalent or complementary forms: as the nonnegative term in an entropy balance, as a relative-entropy rate between forward and time-reversed trajectories, as a quadratic current functional in diffusion settings, and as an estimator reconstructed from trajectories, waiting-time statistics, state sequences, or low-order moments. The concept has been developed for Markov jump processes, semi-Markov dynamics, overdamped and underdamped Langevin systems, quantum Gaussian systems, relativistic spin hydrodynamics, and systems with absorbing configurations or forbidden reverse moves (Zhang et al., 16 Jun 2025, Belenchia et al., 2019, Tome et al., 2024).

1. Thermodynamic meaning and formal definitions

A standard starting point is the entropy balance. For stochastic dynamics with probability distribution $P_i$ or density $\rho$, the Gibbs-Shannon entropy is written as

$S=-\sum_i P_i\ln P_i$

for discrete states, or as the corresponding integral form for continuous states. The entropy production rate is the nonnegative term that remains after separating the change of system entropy from entropy flux or heat dissipation. In the contact-process treatment this balance is written as

$\frac{dS}{dt}=\Pi-\Psi,$

while in jump diffusions it is written thermodynamically as

$e_p(t)=\frac{dS(t)}{dt}-h_d(t),$

with $h_d(t)$ the heat dissipation rate (Tome et al., 2024, Zhang et al., 16 Jun 2025).

For Markov jump processes, the classical internal entropy production is the Schnakenberg-type expression

$$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$

which vanishes under detailed balance. This formulation makes explicit that EPR quantifies the breaking of detailed balance and time-reversal symmetry, not merely the existence of temporal evolution (Cocconi et al., 2020).

A second foundational formulation is path-space. For stationary jump diffusions, the stationary EPR is the relative entropy rate between the forward path law $\mathcal P_{[0,T]}$ and the time-reversed path law $\mathcal P^R_{[0,T]}$: $$e_p^{ss}=\frac{1}{T}\mathcal H(\mathcal P_{[0,T]} \mid \mathcal P^R_{[0,T]}).$$ This identifies EPR directly with trajectory-level irreversibility. The same path-reversal viewpoint underlies the review of exactly solvable systems, where the difference of dynamical entropies of forward and reversed paths converges to the internal entropy production rate in the small-$\rho$0 limit (Zhang et al., 16 Jun 2025, Cocconi et al., 2020).

For diffusive systems, the rate frequently takes a current-quadratic form. In overdamped diffusion with probability current $\rho$1, one exact expression is

$\rho$2

For general jump diffusions on $\rho$3, the paper separates a diffusion contribution and a jump contribution: $\rho$4 The structure is the same in spirit: irreversible currents weighted by the local or nonlocal resistance metric produce a nonnegative rate (Zhang et al., 16 Jun 2025, Cocconi et al., 2020).

2. Observable-based empirical estimators

A major development in the literature is the replacement of formal current integrals by estimators written directly in terms of observables. In linear Langevin systems with mixed even and odd variables under time reversal, the entropy production rate can be written exactly in terms only of the mean vector $\rho$5 and covariance matrix $\rho$6. This makes $\rho$7, $\rho$8, and $\rho$9 empirically computable from observed first and second moments, without explicit estimation of the full probability current (Landi et al., 2015).

Trajectory-level estimators can be more local. In reconstituted actomyosin, individual filament shapes are decomposed into bending modes $S=-\sum_i P_i\ln P_i$0, and the cumulative entropy production is estimated from the observed mode-space trajectory as

$S=-\sum_i P_i\ln P_i$1

The entropy production rate is then obtained from the slope of $S=-\sum_i P_i\ln P_i$2 in time. The paper emphasizes that this is a model-free estimate of dissipation at the filament scale, based on the steady-state phase-space velocity field inferred from the data (Seara et al., 2018).

Another exact empirical route is the variance sum rule for overdamped nonequilibrium steady states. For

$S=-\sum_i P_i\ln P_i$3

the main practical formula is

$S=-\sum_i P_i\ln P_i$4

This expresses EPR through measurable drift, the short-time curvature of the displacement variance, and the force variance. The same work extends the method to active Brownian particles, optical-trap systems, and red-blood-cell flickering, and reports spatially heterogeneous $S=-\sum_i P_i\ln P_i$5-maps with finite correlation length in the microscopy data (Terlizzi et al., 2023).

Recent work has also produced explicitly information-theoretic estimators. For overdamped Langevin dynamics, an exact identity equates the total EPR rate to a mutual-information rate between the infinitesimal displacement and the time-symmetric midpoint, up to a bulk mean-flow term: $S=-\sum_i P_i\ln P_i$6 When the mean flow vanishes, irreversibility is entirely encoded in local time-symmetric fluctuation statistics. The same framework yields a nonnegative split into self and interaction EPR by the information chain rule (Cho et al., 31 Dec 2025).

For weakly damped underdamped systems, the total rate is decomposed into positive contributions from nonzero mean velocity, non-thermal velocity width, position-velocity correlations, and non-Gaussian velocity statistics. The paper emphasizes that the first three terms form a useful “Gaussian” estimate, since they depend only on first and second moments and dominate in the nanoparticle experiments considered (Ciampini et al., 27 Nov 2025).

3. Coarse-graining, waiting times, and partial observability

Empirical EPR becomes substantially harder when only a coarse-grained or partially observed process is available. Semi-Markov processes are a central framework for this regime because the relevant irreversibility can be encoded in both jump directions and waiting-time statistics. In the semi-Markov large-deviation approach, the empirical observable is the empirical semi-Markov kernel $S=-\sum_i P_i\ln P_i$7, and the EPR rate function is obtained by contraction: $S=-\sum_i P_i\ln P_i$8 For the waiting-time-based entropy estimator $S=-\sum_i P_i\ln P_i$9, the paper derives an explicit upper bound on $\frac{dS}{dt}=\Pi-\Psi,$0 via a family of rescaled waiting-time distributions, which in turn yields a lower bound on the variance of finite-time entropy production estimates (Maier et al., 18 Sep 2025).

When only state sequences are observed, several lower-bound strategies exist. First-passage methods use the mean first-passage time and the splitting probability between two thresholds of a single current. The first-passage ratio

$\frac{dS}{dt}=\Pi-\Psi,$1

is a universal lower bound in the large-threshold limit and is unbiased when the measured current is proportional to the stochastic entropy production. The same study shows that first-passage ratios can capture a finite fraction of the total EPR in regimes far from thermal equilibrium where thermodynamic uncertainty ratios capture a negligible fraction (Neri, 2022).

For partially observed continuous-time Markov chains, the hierarchy of bounds tightens as more information is retained. Full coarse graining yields the weakest bounds; semi coarse graining, which records intra-transitions inside hidden blocks, gives tighter bounds; and a transformed semi-CG representation using hidden-state run lengths and accumulated waiting times gives the tightest lower bound among the tested schemes. The most informative estimator in that setting is the KLD-based semi-Markov decomposition

$\frac{dS}{dt}=\Pi-\Psi,$2

where the affinity term measures directional irreversibility and the waiting-time term measures timing asymmetry (Kapustin et al., 2022).

A complementary optimization-based framework treats the true total EPR as the minimum compatible with observed coarse-grained statistics. The optimization variables are hidden steady-state probabilities and hidden mass rates, constrained to reproduce observed $\frac{dS}{dt}=\Pi-\Psi,$3, $\frac{dS}{dt}=\Pi-\Psi,$4, $\frac{dS}{dt}=\Pi-\Psi,$5, and selected waiting-time moments. The resulting hierarchical bounds

$\frac{dS}{dt}=\Pi-\Psi,$6

remain nontrivial even at stall, where current-based estimators may vanish (Nitzan et al., 2022).

An especially stripped-down estimator is available for continuous-time, finite-state, reversible Markov jump processes when only the state sequence and total observation time are known. In that construction, the logarithm of the ratio of forward and reverse cycle counts estimates cycle affinity, and the net chord traversals per unit time estimate the corresponding flux. The finite-time estimator is

$\frac{dS}{dt}=\Pi-\Psi,$7

No transition rates, dwell times, or transition timestamps are required beyond the total elapsed time (Biddle, 26 May 2026).

4. Forbidden reverse moves, absorbing states, and the contact process

A distinctive difficulty arises when the reverse transition rate vanishes while the forward transition is nonzero. In that case the standard Schnakenberg expression

$\frac{dS}{dt}=\Pi-\Psi,$8

becomes ill-defined. This situation occurs in the contact process for transitions involving absorbing configurations. To address it, the contact-process paper proposes a modified strictly nonnegative entropy-production functional based on the inequality

$\frac{dS}{dt}=\Pi-\Psi,$9

For a unidirectional transition $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$0 with $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$1 and $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$2, the entropy-production contribution is defined as

$e_p(t)=\frac{dS(t)}{dt}-h_d(t),$3

with the reverse contribution set to zero. The correction term ensures positivity and correct equilibrium behavior (Tome et al., 2024).

The same work separates the entropy balance so that the flux is linear in the probability distribution, which is crucial operationally because it can then be evaluated as an average over stationary probabilities. For single-site updates, one expression is used when both directions exist and another when the reverse rate vanishes; in the latter case,

$e_p(t)=\frac{dS(t)}{dt}-h_d(t),$4

This linearity is the feature that makes stationary entropy production directly computable from measured local configuration probabilities (Tome et al., 2024).

The one-dimensional contact process furnishes the main example. Lattice sites are binary, and the local transitions depend on a site and its nearest neighbors. For the only genuinely unidirectional transition, the entropy-flux contribution is

$e_p(t)=\frac{dS(t)}{dt}-h_d(t),$5

At stationarity $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$6, so $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$7. The required three-site probabilities $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$8 are estimated from long-time Monte Carlo simulations on periodic chains with $e_p(t)=\frac{dS(t)}{dt}-h_d(t),$9 to $h_d(t)$0, while suppressing annihilation of the last particle to avoid finite-size absorption (Tome et al., 2024).

The principal physical result is that the entropy production rate per site is finite for all annihilation rates $h_d(t)$1, but has a singularity at the critical point $h_d(t)$2. The reported scaling is

$h_d(t)$3

and

$h_d(t)$4

From finite-size scaling the paper estimates

$h_d(t)$5

The empirical observation that $h_d(t)$6 is numerically equal to the order-parameter exponent $h_d(t)$7 is attributed to the dominance of $h_d(t)$8, which is closely tied to the particle density. For $h_d(t)$9, the entropy flux per site vanishes as $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$0, consistent with the absorbing phase (Tome et al., 2024).

5. Continuous, quantum, and relativistic extensions

In continuous-state stochastic systems, empirical EPR retains its thermodynamic interpretation but acquires model-specific current structures. For general jump diffusions on $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$1, the rate is the sum of a diffusion contribution driven by the local probability current

$$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$2

and a jump contribution driven by the nonlocal current

$$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$3

The same paper proves the equivalence, for stationary Gibbs densities, among time-reversibility, zero stationary EPR, detailed balance, and a gradient structure for the drift and jump kernel (Zhang et al., 16 Jun 2025).

Quantum Gaussian systems admit a phase-space version of the same balance. For continuously monitored Gaussian quantum systems, the conditional entropy production rate differs from the unmonitored rate by an informational correction: $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$4 Because the entropy flux is linear in the conditional state and averages to the unconditional flux, the entire thermodynamic effect of continuous monitoring is captured by $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$5. This yields the sharpened second law

$$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$6

with $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$7 determined by measurement back-action and information gain (Belenchia et al., 2019).

In driven cavity magnomechanics, the steady-state EPR is evaluated from the covariance matrix of the Gaussian steady state. One expression given in the paper is

$$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$8

The study finds that $$\dot S_i(t)=\frac{1}{2}\sum_{m,n}\big(P_n w_{nm}-P_m w_{mn}\big) \ln\!\left(\frac{P_n w_{nm}}{P_m w_{mn}}\right)\ge 0,$$9 can increase or decrease depending on magnon-photon coupling and detuning, and that for small magnon-photon coupling the mutual information between magnon and phonon modes is linked to the irreversibility generated in the system (Edet et al., 2024).

Relativistic spin hydrodynamics introduces a different conceptual issue: the entropy current is not uniquely defined. The entropy current can be shifted by an entropy-gauge transformation, and local thermodynamic relations are not generally invariant under such transformations. Nonetheless, the divergence of the entropy current yields a universal entropy-production formula that is invariant under entropy-gauge transformations: $\mathcal P_{[0,T]}$0 This extends the van Weert–Zubarev formula by explicitly including the spin sector (Becattini et al., 2023).

6. Diagnostic uses, exemplary systems, and recurring misconceptions

Empirical EPR is often used not only as a scalar measure of dissipation but also as a diagnostic of the nature of a steady state. In chemical reaction networks, the relation between EPR and experimentally measurable reaction velocity is not generic. For the triangular and linear networks studied, near thermodynamic equilibrium one finds

$\mathcal P_{[0,T]}$1

whereas near a general nonequilibrium steady state

$\mathcal P_{[0,T]}$2

The simple quadratic law $\mathcal P_{[0,T]}$3 is therefore a signature of being close to thermodynamic equilibrium rather than a generic feature of arbitrary nonequilibrium steady states (Banerjee et al., 2013).

Climate diagnostics provide another example of nonuniqueness. The climate literature distinguishes three viable entropy production rates, each associated with a different system boundary: the material rate, the planetary rate, and the transfer rate. The reported ranges across three model climates are

$\mathcal P_{[0,T]}$4

This is a direct reminder that empirical EPR depends on the definition of the subsystem, the observable channels retained, and the processes counted as internal or external (Gibbins et al., 2020).

Exactly solvable stochastic models clarify several persistent misconceptions. A two-state Markov network cannot sustain a nonzero steady current and therefore cannot sustain nonequilibrium entropy production in the steady state. By contrast, a three-state cycle can sustain a circulating current and a nonzero steady EPR. Likewise, a uniform stationary distribution does not imply equilibrium: biased random walks on rings and cyclic Markov processes can be stationary and spatially uniform while still breaking detailed balance (Cocconi et al., 2020).

Experimental active matter underscores that larger mechanical activity does not necessarily imply larger entropy production in the most visibly driven regime. In reconstituted actomyosin, the entropy production rate is reported to be maximized in the stable non-contractile active state rather than in the contractile state, with statistically significant slope differences between regimes. The interpretation offered is that disordered myosin-actin geometry generates strong transverse “plucking” fluctuations and filament bending, which are more dissipative at the measured mesoscale than the more coherent sliding motions of the contractile phase (Seara et al., 2018).

A further recurring lesson from the partial-observation literature is that vanishing visible current does not imply zero total dissipation. At stalling, current-based estimators may vanish, while waiting-time asymmetry, higher-order transition statistics, or hidden-state optimization still yield nonzero lower bounds on the total EPR (Nitzan et al., 2022, Kapustin et al., 2022). This suggests that empirical EPR is best understood not as a single universal estimator, but as a hierarchy of exact formulas, operational reconstructions, and lower bounds whose tightness is controlled by how much of the underlying trajectory structure is experimentally accessible.