betaFeedback

WorksPhysics and Astronomy

Active Jarzynski and Crooks Relations for Driven Jump-Diffusions

Abstract

Non-conservative active driving breaks detailed balance in non-Gaussian jump-diffusions, obscuring the connection between external protocol work, housekeeping dissipation, and free-energy differences. Here, we establish exact finite-time fluctuation relations for driven Markov jump-diffusions—encompassing infinite-activity Lévy processes—by comparing path probabilities under the forward and current-reversed dual dynamics relative to the instantaneous steady state (ISS). We derive a trajectory-level decomposition of total entropy production that cleanly separates transient state relaxation, housekeeping heat, and external protocol driving. When the ISS takes a Gibbsian form, gauge-tilting the dual path measure yields an Active Jarzynski equality, ⟨exp⁡[−(W+Δsact)]⟩π0=e−ΔF\langle\exp[-(W+Δs_{\mathrm{act}})]\rangle_{π_0}=e^{-ΔF}, which incorporates active housekeeping entropy ΔsactΔs_{\mathrm{act}} alongside mechanical work WW and coexists with the conventional Jarzynski equality. Under protocol reversal, this dual framework further reveals an Active Crooks relation, where the asymmetry between forward and reverse work distributions is governed by the conditional exponential average of active dissipation. Our results provide a framework for recovering free-energy differences in driven jump-diffusions with Gibbsian instantaneous steady states and for characterizing active irreversibility through trajectory statistics.

Introduction

Quantifying how external driving and internal fluctuations govern thermodynamic irreversibility is a central theme of modern statistical physics [2] [1]. In classical stochastic thermodynamics, a system coupled to a thermal bath (in thermal units kBTbath=1k_{\mathrm{B}}T_{\mathrm{bath}}=1) and driven from an initial equilibrium state π0\pi_0 is assumed to obey microscopic reversibility [15] [14]. When detailed balance holds at each frozen value of an external protocol, this time-reversal symmetry underpins two foundational results: the Jarzynski equality [3], ⟨e−W⟩π0=e−ΔF\langle e^{-W}\rangle_{\pi_0}=e^{-\Delta F}, which determines the equilibrium free-energy difference ΔF\Delta F from the exponential average of the protocol work WW, and the Crooks fluctuation theorem [4], pF(w)/pR(−w)=ew−ΔFp_{\mathrm{F}}(w)/p_{\mathrm{R}}(-w)=e^{w-\Delta F}, which compares the probability density of forward work ww with that of reverse work −w-w.

This classical paradigm does not extend directly to active systems undergoing discontinuous, non-Gaussian transitions [5, 6]. Across diverse scales—ranging from self-propelled colloids and molecular motors to transcriptional bursting circuits [7]—energy-consuming processes can generate non-conservative drift forces and jump noise. Driving such active jump-diffusions with a time-dependent protocol introduces three interlocking challenges. First, the system generally lacks a common time-independent steady state to serve as a baseline under a time-dependent protocol [12] [11]. Second, mechanical protocol work WW must be distinguished from continuous and jump-induced active dissipation (housekeeping entropy Δsact\Delta s_{\mathrm{act}}). Third, trajectory comparisons must account for jump-rate asymmetry and the associated compensators; forbidden reverse jumps can make forward and reversed path measures singular. Together, these effects disrupt standard work fluctuation symmetries, confounding external mechanical manipulation with autonomous active driving [8, 9].

To address these difficulties, we introduce the instantaneous steady state (ISS) πt\pi_t at each frozen protocol value and use it as a moving thermodynamic reference [10]. This preserves the familiar distinction between relaxation toward a stationary density and the dissipation sustained by its currents, while accounting for the changing protocol. To connect this thermodynamic separation to trajectory probabilities, we compare the original process with dual dynamics that preserve the same frozen density but reverse its stationary currents [11]. For a common initial distribution, the logarithmic ratio of the original and dual path probabilities gives the active entropy Δsact\Delta s_{\mathrm{act}}. Weighting the original trajectories by e−Δsacte^{-\Delta s_{\mathrm{act}}} therefore gives the dual path measure. Combining these dual dynamics with an exponential gauge transformation with the work weight e−We^{-W} (Figure 1) then enables us to reweight trajectory ensembles by mechanical work and establish exact finite-time work relations.

Reference comparisons and fluctuation relations

Figure 1. Reference comparisons and fluctuation relations. The left pair of arrows compares P\mathcal{P} with Pα\mathcal{P}^{\alpha} to identify path functionals. The right pair compares P\mathcal{P} with Pα,Γ\mathcal{P}^{\alpha,\Gamma} to obtain the weighted IFT. The upper arrows select conjugate dynamics and then apply the gauge weight.

This dual-gauge framework unifies non-equilibrium thermodynamics at the ensemble and trajectory levels.

When the frozen ISS admits a Gibbsian form πt(x)=Zt−1e−H(t,x)\pi_t(x)=Z_t^{-1}e^{-H(t,x)} for a prescribed, protocol-controlled Hamiltonian H(t,x)≡H(λt,x)H(t,x)\equiv H(\lambda_t,x), the macroscopic entropy production rate S˙tot(t)\dot{S}_{\mathrm{tot}}(t) obeys the decomposition

S˙tot(t)=W˙(t)+Q˙hk(t)−F˙t−ddtD(ρt∥πt)≥0,(1)\dot{S}_{\mathrm{tot}}(t)=\dot{W}(t)+\dot{Q}_{\mathrm{hk}}(t)-\dot{F}_t-\frac{\mathrm{d}}{\mathrm{d}t}D(\rho_t\Vert\pi_t)\geq0, \tag*{(1)}

where W˙\dot{W} and Q˙hk\dot{Q}_{\mathrm{hk}} denote the mean protocol power and housekeeping heat rate [12], W‾(t):=⟨W[0,t]⟩P\overline{W}(t):=\langle W[0,t]\rangle_{\mathcal{P}} is the mean work, Ft=−log⁡ZtF_t=-\log Z_t is the equilibrium free energy, and the Kullback–Leibler divergence D(ρt∥πt)D(\rho_t\Vert\pi_t) quantifies the lag of the evolving density ρt\rho_t behind πt\pi_t. At the trajectory level, total entropy production Δstot\Delta s_{\mathrm{tot}} accumulated over [0,T][0,T] decomposes into

Δstot=W+Δsact−ΔF+Δsrel,(2)\Delta s_{\mathrm{tot}}=W+\Delta s_{\mathrm{act}}-\Delta F+\Delta s_{\mathrm{rel}}, \tag*{(2)}

where ΔF=FT−F0=−log⁡(ZT/Z0)\Delta F=F_T-F_0=-\log(Z_T/Z_0) is the equilibrium free-energy difference, Δsact\Delta s_{\mathrm{act}} is the trajectory-level active housekeeping heat, WW is the mechanical work, and the endpoint relaxation term Δsrel\Delta s_{\mathrm{rel}} is related to the Kullback–Leibler divergence above through

⟨Δsrel⟩P=D(ρ0∥π0)−D(ρT∥πT).\langle\Delta s_{\mathrm{rel}}\rangle_{\mathcal{P}}=D(\rho_0\Vert\pi_0)-D(\rho_T\Vert\pi_T).

Crucially, comparing the forward path measure with the work-tilted dual path measure yields the finite-time Active Jarzynski equality,

⟨e−(W+Δsact)⟩π0=e−ΔF.(3)\left\langle e^{-(W+\Delta s_{\mathrm{act}})}\right\rangle_{\pi_0}=e^{-\Delta F}. \tag*{(3)}

Comparing forward paths with reversed records of the protocol-reversed experiment, initialized in πT\pi_T, establishes the Active Crooks relation,

pF(w)pR(−w)=ew−ΔF⟨e−Δsact∣W=w⟩π0,(4)\frac{p_{\mathrm{F}}(w)}{p_{\mathrm{R}}(-w)}=\frac{e^{w-\Delta F}}{\left\langle e^{-\Delta s_{\mathrm{act}}}\mid W=w\right\rangle_{\pi_0}}, \tag*{(4)}

where the conditional active factor describes the active contribution to the asymmetry between forward work ww and reverse work −w-w.

Jump-diffusion dynamics and gauge-tilted path measures

We consider a driven jump-diffusion Xt∈RdX_t\in\mathbb{R}^d on [0,T][0,T] governed by the Itô equation

dXt=bt(Xt)dt+2 a dWt+∫Rd(y−Xt−)NX(dt,dy),(5)\mathrm{d}X_t=b_t(X_t)\mathrm{d}t+\sqrt{2}\,a\,\mathrm{d}W_t+\int_{\mathbb{R}^d}(y-X_{t-})N_X(\mathrm{d}t,\mathrm{d}y), \tag*{(5)}

where bt(x)b_t(x) is the time-dependent drift, WtW_t is a standard dd-dimensional Wiener process, A=aaT>0A=aa^{\mathsf{T}}>0 is a constant diffusion tensor characterizing thermal fluctuations from a bath at temperature TbathT_{\mathrm{bath}}, and NX(dt,dy)N_X(\mathrm{d}t,\mathrm{d}y) is the jump measure generated by Poisson noise, with state-dependent compensator qt(Xt−,y)dydtq_t(X_{t-},y)\mathrm{d}y\mathrm{d}t. The transition rate kernel qt(x,y)q_t(x,y) governs non-local jumps from xx to yy (e.g., transcriptional bursts [7]), encompassing infinite-activity Lévy processes with summable small jumps.

Let P\mathcal{P} denote the path probability measure of trajectories X={Xt}t∈[0,T]X=\{X_t\}_{t\in[0,T]} initialized in ρ0\rho_0, with instantaneous density ρt(x)\rho_t(x). An auxiliary reference process with drift btαb_t^\alpha, jump kernel qtαq_t^\alpha, and initial density ρ0α\rho_0^\alpha has path measure Pα\mathcal{P}^\alpha. The trajectory log-likelihood ratio between P\mathcal{P} and Pα\mathcal{P}^\alpha is

Iα[X]=log⁡dPdPα[X].(6)\mathcal{I}^\alpha[X]=\log\frac{\mathrm{d}\mathcal{P}}{\mathrm{d}\mathcal{P}^\alpha}[X]. \tag*{(6)}

Averaging the jump-diffusion Girsanov formula [13] [End Matter, Eq. (32)] gives the path-space relative entropy ⟨Iα⟩P=D(P∥Pα)≥0\langle\mathcal{I}^\alpha\rangle_{\mathcal{P}}=D(\mathcal{P}\Vert\mathcal{P}^\alpha)\geq0:

⟨Iα⟩P=D(ρ0∥ρ0α)+14∫0T∫ρt(bt−btα)TA−1(bt−btα) dxdt+∫0T∫∫ρt(x)[qtlog⁡qtqtα−qt+qtα]dydxdt,(7)\begin{aligned} \langle\mathcal{I}^\alpha\rangle_{\mathcal{P}} ={}&D(\rho_0\Vert\rho_0^\alpha)+\frac{1}{4}\int_0^T\int\rho_t(b_t-b_t^\alpha)^{\mathsf{T}}A^{-1}(b_t-b_t^\alpha)\,\mathrm{d}x\mathrm{d}t\\ &+\int_0^T\int\int\rho_t(x)\left[q_t\log\frac{q_t}{q_t^\alpha}-q_t+q_t^\alpha\right]\mathrm{d}y\mathrm{d}x\mathrm{d}t, \tag*{(7)} \end{aligned}

where the three terms correspond to initial distribution mismatch, drift mismatch, and jump-rate asymmetry, with −qt+qtα-q_t+q_t^\alpha accounting for the escape-rate (traffic) correction [24].

To incorporate thermodynamic observables such as mechanical work, we apply an exponential gauge tilt to Pα\mathcal{P}^\alpha with an additive trajectory functional Γ[X]\Gamma[X]. For a finite path-space partition function ZαΓ=⟨e−Γ⟩PαZ_\alpha^\Gamma=\langle e^{-\Gamma}\rangle_{\mathcal{P}^\alpha}, the tilted measure satisfies

dPα,ΓdPα=e−ΓZαΓ.(8)\frac{\mathrm{d}\mathcal{P}^{\alpha,\Gamma}}{\mathrm{d}\mathcal{P}^\alpha}=\frac{e^{-\Gamma}}{Z_\alpha^\Gamma}. \tag*{(8)}

By the chain rule for Radon–Nikodym derivatives,

dPα,ΓdP=dPα,ΓdPαdPαdP=e−(Iα+Γ)ZαΓ.(9)\frac{\mathrm{d}\mathcal{P}^{\alpha,\Gamma}}{\mathrm{d}\mathcal{P}}=\frac{\mathrm{d}\mathcal{P}^{\alpha,\Gamma}}{\mathrm{d}\mathcal{P}^\alpha}\frac{\mathrm{d}\mathcal{P}^\alpha}{\mathrm{d}\mathcal{P}}=\frac{e^{-(\mathcal{I}^\alpha+\Gamma)}}{Z_\alpha^\Gamma}. \tag*{(9)}

Normalization of Pα,Γ\mathcal{P}^{\alpha,\Gamma} immediately yields the master integral fluctuation theorem (IFT),

⟨e−Δsα,Γ⟩P=1,Δsα,Γ=Iα+Γ+log⁡ZαΓ,(10)\left\langle e^{-\Delta s^{\alpha,\Gamma}}\right\rangle_{\mathcal{P}}=1,\qquad\Delta s^{\alpha,\Gamma}=I^\alpha+\Gamma+\log Z^\Gamma_\alpha, \tag*{(10)}

which implies the generalized second-law inequality ⟨Δsα,Γ⟩P≥0\langle\Delta s^{\alpha,\Gamma}\rangle_{\mathcal{P}}\geq0 via Jensen’s inequality.

Because full trajectory functionals like IαI^\alpha are often less accessible experimentally than cumulative energy inputs, we also compare the marginal distributions of a trajectory observable O[X]\mathcal{O}[X] (e.g., protocol work WW). Conditioning the likelihood ratio dPα/dP=e−Iαd\mathcal{P}^\alpha/d\mathcal{P}=e^{-I^\alpha} on any measurable event O∈B\mathcal{O}\in B (with P{O∈B}>0\mathcal{P}\{\mathcal{O}\in B\}>0) gives the conditional relation

Pα{O∈B}P{O∈B}=⟨e−Iα∣O∈B⟩P.(11)\frac{\mathcal{P}^\alpha\{\mathcal{O}\in B\}}{\mathcal{P}\{\mathcal{O}\in B\}}=\left\langle e^{-I^\alpha}\mid\mathcal{O}\in B\right\rangle_{\mathcal{P}}. \tag*{(11)}

Selecting physically motivated reference dynamics Pα\mathcal{P}^\alpha and gauge weights Γ\Gamma in Eqs. (10) and (11) allows us to systematically derive both integral and detailed work relations.

Macroscopic free-energy dissipation. The density ρt(x)\rho_t(x) satisfies the Fokker–Planck equation ∂tρt=Lt∗ρt\partial_t\rho_t=\mathcal{L}_t^*\rho_t, where Lt∗\mathcal{L}_t^* is the adjoint of the jump-diffusion generator Lt\mathcal{L}_t. At each frozen time tt, the instantaneous steady state (ISS) πt\pi_t satisfies Lt∗πt=0\mathcal{L}_t^*\pi_t=0. Using πt\pi_t as a time-dependent reference, the macroscopic total entropy production rate S˙tot(t)\dot{S}_{\mathrm{tot}}(t) decomposes into nonadiabatic and active housekeeping contributions [12, 16] [see Eqs. (29)–(31) in the End Matter for explicit current–affinity expressions]:

S˙tot(t)=−ddtD(ρt∥πt)+ddt⟨YHS[0,t]⟩+S˙act(t)=S˙na(t)+S˙act(t)≥0.(12)\begin{aligned} \dot{S}_{\mathrm{tot}}(t)=-\frac{d}{dt}D(\rho_t\Vert\pi_t)+\frac{d}{dt}\langle Y_{\mathrm{HS}}[0,t]\rangle+\dot{S}_{\mathrm{act}}(t) \\ &=\dot{S}_{\mathrm{na}}(t)+\dot{S}_{\mathrm{act}}(t)\geq0. \tag*{(12)} \end{aligned}

Here, D(ρt∥πt)=∫ρtlog⁡(ρt/πt)dxD(\rho_t\Vert\pi_t)=\int\rho_t\log(\rho_t/\pi_t)dx is the relative entropy, YHS[0,t]=−∫0t(∂slog⁡πs)(Xs)dsY_{\mathrm{HS}}[0,t]=-\int_0^t(\partial_s\log\pi_s)(X_s)ds is the Hatano–Sasa protocol functional [10] with driving rate ddt⟨YHS[0,t]⟩=−∫ρt∂tlog⁡πt dx\frac{d}{dt}\langle Y_{\mathrm{HS}}[0,t]\rangle=-\int\rho_t\partial_t\log\pi_t\,dx, and

S˙na(t)=−ddtD(ρt∥πt)+ddt⟨YHS[0,t]⟩≥0(13)\dot{S}_{\mathrm{na}}(t)=-\frac{d}{dt}D(\rho_t\Vert\pi_t)+\frac{d}{dt}\langle Y_{\mathrm{HS}}[0,t]\rangle\geq0 \tag*{(13)}

is the nonadiabatic entropy production rate, while S˙act(t)=Q˙hk(t)≥0\dot{S}_{\mathrm{act}}(t)=\dot{Q}_{\mathrm{hk}}(t)\geq0 denotes the active housekeeping dissipation rate sustaining steady-state currents.

When the ISS admits a Gibbsian form πt(x)=Zt−1e−H(t,x)\pi_t(x)=Z_t^{-1}e^{-H(t,x)} with partition function Zt=∫e−H(t,x)dxZ_t=\int e^{-H(t,x)}dx and equilibrium free energy Ft=−log⁡ZtF_t=-\log Z_t, the nonequilibrium free energy Fneq(t)=∫ρt[H(t,x)+log⁡ρt]dxF_{\mathrm{neq}}(t)=\int\rho_t[H(t,x)+\log\rho_t]dx satisfies Fneq(t)−Ft=D(ρt∥πt)F_{\mathrm{neq}}(t)-F_t=D(\rho_t\Vert\pi_t). In this Gibbs–ISS setting, the protocol driving rate becomes ddt⟨YHS[0,t]⟩=W˙(t)−F˙t\frac{d}{dt}\langle Y_{\mathrm{HS}}[0,t]\rangle=\dot{W}(t)-\dot{F}_t, where W˙(t)=∫ρt(x)∂tH(t,x)dx\dot{W}(t)=\int\rho_t(x)\partial_tH(t,x)dx is the mean mechanical power. Consequently, Eq. (13) implies F˙neq(t)=W˙(t)−S˙na(t)\dot{F}_{\mathrm{neq}}(t)=\dot{W}(t)-\dot{S}_{\mathrm{na}}(t), and Eq. (12) yields F˙neq(t)=W˙(t)+Q˙hk(t)−S˙tot(t)\dot{F}_{\mathrm{neq}}(t)=\dot{W}(t)+\dot{Q}_{\mathrm{hk}}(t)-\dot{S}_{\mathrm{tot}}(t), directly recovering the dynamic balance in Eq. (1).

In the Gibbs–ISS setting where YHS=W−ΔFY_{\mathrm{HS}}=W-\Delta F, this trajectory balance directly reduces to Eq. (2).

For a static protocol (W˙=F˙t=0\dot{W}=\dot{F}_t=0), YHSY_{\mathrm{HS}} vanishes and S˙na(t)=−ddtFneq(t)\dot{S}_{\mathrm{na}}(t)=-\frac{d}{dt}F_{\mathrm{neq}}(t) captures purely transient free-energy relaxation; once the system reaches a nonequilibrium steady state π\pi breaking detailed balance, this relaxation ceases (S˙na=0\dot{S}_{\mathrm{na}}=0) while housekeeping heat persists (S˙tot=S˙act>0\dot{S}_{\mathrm{tot}}=\dot{S}_{\mathrm{act}}>0).

Reference dynamics and entropy relations. Along an individual trajectory, the total entropy production Δstot[0,T]=Δssys+Q[0,T]\Delta s_{\mathrm{tot}}[0,T]=\Delta s_{\mathrm{sys}}+Q[0,T] combines the system-entropy change Δssys=−log⁡[ρT(XT)/ρ0(X0)]\Delta s_{\mathrm{sys}}=-\log[\rho_T(X_T)/\rho_0(X_0)] with the environment-entropy flow Q[0,T]Q[0,T] [15] [14]. Using the moving ISS πt\pi_t as a baseline, Δstot\Delta s_{\mathrm{tot}} admits the general trajectory-level decomposition Δstot=Δsrel+Δsact+YHS\Delta s_{\mathrm{tot}}=\Delta s_{\mathrm{rel}}+\Delta s_{\mathrm{act}}+Y_{\mathrm{HS}}, whose ensemble average recovers the time-integrated form of Eq. (12). Here, Δsrel=−log⁡[ρT(XT)/πT(XT)]+log⁡[ρ0(X0)/π0(X0)]\Delta s_{\mathrm{rel}}=-\log[\rho_T(X_T)/\pi_T(X_T)]+\log[\rho_0(X_0)/\pi_0(X_0)] captures endpoint state relaxation, while Δsact[0,t]\Delta s_{\mathrm{act}}[0,t] accumulates the local active affinity ftlocf_t^{\mathrm{loc}} along continuous displacements and the non-local affinity ftnlf_t^{\mathrm{nl}} across individual jumps [defined in Eqs. (28) and (34) in the End Matter], satisfying S˙act(t)=ddt⟨Δsact[0,t]⟩\dot{S}_{\mathrm{act}}(t)=\frac{d}{dt}\langle\Delta s_{\mathrm{act}}[0,t]\rangle and reducing to standard housekeeping dissipation for autonomous dynamics [16] [1].

To isolate the active contribution Δsact\Delta s_{\mathrm{act}} at the path-measure level, we construct conjugate dual dynamics with path measure Pπ\mathcal{P}^\pi and generator Ltπ\mathcal{L}_t^\pi that preserve the frozen ISS density πt\pi_t while reversing its stationary currents [16, 17] [10, 11]:

btπ=−bt+2A∇log⁡πt,qtπ(x,y)=πt(y)qt(y,x)πt(x).(14)b_t^\pi=-b_t+2A\nabla\log\pi_t,\qquad q_t^\pi(x,y)=\frac{\pi_t(y)q_t(y,x)}{\pi_t(x)}. \tag*{(14)}

At each fixed protocol value, the forward and dual generators share the same invariant measure,

Lt∗πt=(Ltπ)∗πt=0,(15)\mathcal{L}_t^*\pi_t=(\mathcal{L}_t^\pi)^*\pi_t=0, \tag*{(15)}

while the dual dynamics reverse both the continuous diffusive current [Eq. (33)] and the directional jump flux via πt(x)qtπ(x,y)=πt(y)qt(y,x)\pi_t(x)q_t^\pi(x,y)=\pi_t(y)q_t(y,x) (the dual and forward dynamics coincide under detailed balance). With both processes initialized in ρ0\rho_0, substituting the dual coefficients (14) into Girsanov’s formula (32) and using Lt∗πt=0\mathcal{L}_t^*\pi_t=0 gives dPπ/dP=e−Δsactd\mathcal{P}^\pi/d\mathcal{P}=e^{-\Delta s_{\mathrm{act}}} and Iπ=Δsact\mathcal{I}^\pi=\Delta s_{\mathrm{act}} [End Matter, Eq. (36)].

Capturing total entropy production and work relations further requires distinguishing direct path reversal from thermodynamic protocol reversal. The path-reversal operator Θ\Theta, defined by (ΘX)t=XT−t(\Theta X)_t=X_{T-t} (with standard left-limit conventions at jumps), maps the forward ensemble P\mathcal{P} to the time-reversed measure Θ#P\Theta_\#\mathcal{P} initialized in ρT\rho_T [21] [20]. Its effective generator retains the diffusion tensor AA with drift and jump coefficients governed by the evolving density ρT−t\rho_{T-t}:

btΘ(x)=−bT−t(x)+2A∇log⁡ρT−t(x),qtΘ(x,y)=ρT−t(y)qT−t(y,x)ρT−t(x).(16)b_t^{\Theta}(x)=-b_{T-t}(x)+2A\nabla\log\rho_{T-t}(x), \qquad q_t^{\Theta}(x,y)=\frac{\rho_{T-t}(y)q_{T-t}(y,x)}{\rho_{T-t}(x)}. \tag*{(16)}

The relative action IΘ\mathcal{I}^{\Theta} retains the escape-rate correction [Eq. (32)] and depends on the observation time TT through the coefficients in Eq. (16), which use the actual density ρT−t\rho_{T-t} rather than the frozen ISS πt\pi_t. In contrast, a physical protocol-reversal experiment RR drives the system under the time-reversed schedule btR=bT−tb_t^R=b_{T-t} and qtR=qT−tq_t^R=q_{T-t} from an initial distribution ρT\rho_T, generating a distinct path measure PR\mathcal{P}_R [1]. Comparing forward trajectories against the reversed records of this reverse experiment (Θ#PR\Theta_{\#}\mathcal{P}_R) eliminates the escape-rate mismatch and recovers the entropy production Δstot\Delta s_{\mathrm{tot}}. In the reset example in the End Matter, ⟨Δstot⟩P=0\langle\Delta s_{\mathrm{tot}}\rangle_{\mathcal{P}}=0 for every TT, whereas ⟨IΘ⟩P\langle\mathcal{I}^{\Theta}\rangle_{\mathcal{P}} varies with the observation time [22, 23]. Table I summarizes these three fundamental path comparisons (verified in the End Matter). The same path-measure construction also recovers the Hatano–Sasa equality for an initial steady-state distribution and detailed entropy fluctuation relations between paired forward and conjugate ensembles [10, 16].

Active Jarzynski and Crooks relations. For work protocols in the Gibbs–ISS setting πt(x)=Zt−1e−H(t,x)\pi_t(x)=Z_t^{-1}e^{-H(t,x)}, mechanical work is W=∫0T(∂tH)(t,Xt)dtW=\int_0^T(\partial_tH)(t,X_t)\mathrm{d}t and the equilibrium free-energy difference is ΔF=FT−F0=−log⁡(ZT/Z0)\Delta F=F_T-F_0=-\log(Z_T/Z_0). We initialize the forward (P\mathcal{P}) and dual (Pπ\mathcal{P}^{\pi}) processes in π0\pi_0, and the protocol-reversed experiment (PR\mathcal{P}_R) in πT\pi_T. By the trajectory decomposition in Eq. (2), W+Δsact−ΔF=Δstot−ΔsrelW+\Delta s_{\mathrm{act}}-\Delta F=\Delta s_{\mathrm{tot}}-\Delta s_{\mathrm{rel}}. Since the dual relative action Iπ=Δsact\mathcal{I}^{\pi}=\Delta s_{\mathrm{act}} captures the active term in this balance, applying the exponential work tilt Γ=W\Gamma=W to Pπ\mathcal{P}^{\pi} via Eq. (8) defines dPπ,W=(ZπW)−1e−WdPπ\mathrm{d}\mathcal{P}^{\pi,W}=(Z_{\pi}^{W})^{-1}e^{-W}\mathrm{d}\mathcal{P}^{\pi}, so that changing measure via Eq. (9) gives

⟨e−(Δsact+W)⟩P=⟨e−W⟩Pπ=ZπW.(17)\left\langle e^{-(\Delta s_{\mathrm{act}}+W)}\right\rangle_{\mathcal{P}}=\left\langle e^{-W}\right\rangle_{\mathcal{P}^{\pi}}=Z_{\pi}^{W}. \tag*{(17)}

To evaluate ZπWZ_{\pi}^{W}, we use the stationarity of the Gibbs density under the frozen dual dynamics and the time derivative of the Gibbs factor:

(Ltπ)∗e−H(t,⋅)=0,∂te−H(t,x)=−(∂tH)(t,x)e−H(t,x).(18)(\mathcal{L}_t^{\pi})^{*}e^{-H(t,\cdot)}=0,\qquad\partial_t e^{-H(t,x)}=-(\partial_tH)(t,x)e^{-H(t,x)}. \tag*{(18)}

Applying Eq. (18) to the Feynman–Kac equation (39) with Kt=LtπK_t=\mathcal{L}_t^{\pi} gives ZπW=ZT/Z0=e−ΔFZ_{\pi}^{W}=Z_T/Z_0=e^{-\Delta F} through the weighted endpoint identity (40) in the End Matter. Combining this value of the normalization factor with the reverse-path comparison in Eq. (38) gives the measure identity

dPπ,WdP=e−(W+Δsact−ΔF)=d(Θ#PR)dP.(19)\frac{\mathrm{d}\mathcal{P}^{\pi,W}}{\mathrm{d}\mathcal{P}}=e^{-(W+\Delta s_{\mathrm{act}}-\Delta F)}=\frac{\mathrm{d}(\Theta_{\#}\mathcal{P}_R)}{\mathrm{d}\mathcal{P}}. \tag*{(19)}

Thus Pπ,W=Θ#PR\mathcal{P}^{\pi,W}=\Theta_{\#}\mathcal{P}_R: work-tilting the dual process reproduces the path-reversed ensemble of the protocol-reversal experiment. Integrating Eq. (19) over all forward trajectories yields the Active Jarzynski equality:

⟨e−(W+Δsact)⟩π0=e−ΔF.(20)\left\langle e^{-(W+\Delta s_{\mathrm{act}})}\right\rangle_{\pi_0}=e^{-\Delta F}. \tag*{(20)}

Because the forward generator similarly satisfies Lt∗e−H(t,⋅)=0\mathcal{L}_t^{*}e^{-H(t,\cdot)}=0, applying Γ=W\Gamma=W directly to the forward reference Pα=P\mathcal{P}^{\alpha}=\mathcal{P} (Iα=0\mathcal{I}^{\alpha}=0) yields the conventional mechanical Jarzynski equality ⟨e−W⟩π0=e−ΔF\langle e^{-W}\rangle_{\pi_0}=e^{-\Delta F} [3, 18], which coincides with Eq. (20) when Δsact\Delta s_{\mathrm{act}} vanishes.

When the work distributions admit densities, setting O=W\mathcal{O}=W in Eq. (11) for a reference measure Pα\mathcal{P}^{\alpha} gives the density ratio

pF(w)pα(w)=[⟨e−Iα∣W=w⟩P]−1,(21)\frac{p_F(w)}{p^{\alpha}(w)}=\left[\left\langle e^{-\mathcal{I}^{\alpha}}\mid W=w\right\rangle_{\mathcal{P}}\right]^{-1}, \tag*{(21)}

where pFp_F and pαp^{\alpha} are the work densities under P\mathcal{P} and Pα\mathcal{P}^{\alpha}, respectively. Under the reversed protocol HR(t,x)=H(T−t,x)H^R(t,x)=H(T-t,x), work is odd under path reversal, WR(ΘX)=−W(X)W^R(\Theta X)=-W(X), so choosing Pα=Θ#PR\mathcal{P}^{\alpha}=\Theta_{\#}\mathcal{P}_R yields pα(w)=pR(−w)p^{\alpha}(w)=p_R(-w). Substituting the likelihood ratio from Eq. (19) into Eq. (21) then establishes the Active Crooks relation:

pF(w)pR(−w)=ew−ΔF⟨e−Δsact∣W=w⟩P.(22)\frac{p_F(w)}{p_R(-w)}=\frac{e^{w-\Delta F}}{\left\langle e^{-\Delta s_{\mathrm{act}}}\mid W=w\right\rangle_{\mathcal{P}}}. \tag*{(22)}

Physically, forward trajectories with identical mechanical work ww can accumulate different active entropies Δsact\Delta s_{\mathrm{act}} and thus carry distinct likelihood ratios relative to their time-reversed counterparts. The conditional factor in Eq. (22) averages these active weights over the constant-work subsensemble. It also equals the ratio of the dual to the forward work density at the same ww [Eq. (41) in the End Matter], quantifying how reversing stationary active currents alters work statistics. Solving Eq. (22) for pR(−w)p_R(-w) and integrating over ww yields the Active Jarzynski equality [Eq. (20)] by the law of total expectation, while vanishing active affinities (ftloc=ftnt=0f_t^{\mathrm{loc}}=f_t^{\mathrm{nt}}=0) reduce Eq. (22) to the classical Crooks relation pF(w)/pR(−w)=ew−ΔFp_F(w)/p_R(-w)=e^{w-\Delta F} [4]. Table II summarizes these three work relations.

Work relationReference before weightingTrajectory domain
Mechanical JarzynskiP\mathcal{P} (Iα=0I^\alpha=0)Full path space
Active JarzynskiPπ\mathcal{P}^{\pi} (Iα=ΔsactI^\alpha=\Delta s_{\mathrm{act}})Full path space
Active CrooksPπ\mathcal{P}^{\pi} (Iα=ΔsactI^\alpha=\Delta s_{\mathrm{act}})Fixed-work subensemble

Table II. Work relations for the Gibbsian ISS πt(x)=Zt−1e−H(t,x)\pi_t(x)=Z_t^{-1}e^{-H(t,x)} with Γ=W\Gamma=W. Forward (P\mathcal{P}) and dual (Pπ\mathcal{P}^{\pi}) processes start from π0\pi_0, and the protocol-reversed process (PR\mathcal{P}_{R}) starts from πT\pi_T. For both active relations, the work-tilted dual satisfies Pπ,W=Θ#PR\mathcal{P}^{\pi,W}=\Theta_{\#}\mathcal{P}_{R}.

Examples. We illustrate the coexistence of active dissipation and the recovery of free-energy differences with a two-dimensional overdamped particle subject to chiral active drift and Poisson jumps in the harmonic potential H(t,x)=λt∣x∣2/2H(t,x)=\lambda_t|x|^2/2 with time-dependent stiffness λt\lambda_t. Its position Xt∈R2X_t\in\mathbb{R}^2 obeys

dXt=λt(Rrot−I2)Xtdt+2 dWt+∫R2(y−Xt−)NX(dt,dy),(23)\mathrm{d}X_t=\lambda_t(R_{\mathrm{rot}}-I_2)X_t\mathrm{d}t+\sqrt{2}\,\mathrm{d}W_t+\int_{\mathbb{R}^2}(y-X_{t-})N_X(\mathrm{d}t,\mathrm{d}y), \tag*{(23)}

where Rrotx=(−x2,x1)R_{\mathrm{rot}}x=(-x_2,x_1) is a 90∘90^\circ rotation. Jumps occur at rate one. Conditional on Xt−=xX_{t-}=x, the new position has distribution N(Rrotx/2,3I2/(4λt))\mathcal{N}(R_{\mathrm{rot}}x/2,3I_2/(4\lambda_t)). Both the diffusion and the jumps preserve the frozen density πt(x)=λte−λt∣x∣2/2/(2π)=e−H(t,x)/Zt\pi_t(x)=\lambda_t e^{-\lambda_t|x|^2/2}/(2\pi)=e^{-H(t,x)}/Z_t. We initialize with X0∼N(0,I2)X_0\sim\mathcal{N}(0,I_2). Here, tt denotes the evolution time. We evaluate all observables over [0,T][0,T], with T∈[0,1]T\in[0,1]. The free energy FTF_T is defined by the partition function of H(T,⋅)H(T,\cdot); the evolving density need not equal πT\pi_T. The End Matter gives the numerical method for both protocols.

For the fixed landscape λt=1\lambda_t=1, the actual density remains ρt=π\rho_t=\pi. Rotational drift and asymmetric jumps produce active entropy without protocol work or free-energy change:

W=ΔF=0,W=\Delta F=0,
⟨Δsact[0,T]⟩=10T3,⟨e−Δsact[0,T]⟩=1.(24)\langle\Delta s_{\mathrm{act}}[0,T]\rangle=\frac{10T}{3},\qquad\langle e^{-\Delta s_{\mathrm{act}}[0,T]}\rangle=1. \tag*{(24)}

Thus active entropy accumulates without a change in either the state distribution or free energy (Figure 2). The positive mean and the unit exponential average describe different statistics of the same active trajectories.

Active entropy statistics at fixed landscape

Figure 2. Active statistics at fixed landscape. The sample mean of Δsact\Delta s_{\mathrm{act}} is compared with 10T/310T/3. Here Jact(T)=−log⁡⟨e−Δsact[0,T]⟩π0=0J_{\mathrm{act}}(T)=-\log\langle e^{-\Delta s_{\mathrm{act}}[0,T]}\rangle_{\pi_0}=0; the open squares show its Monte Carlo estimates.

Keeping the same model, we apply the periodic stiffness protocol λt=32−12cos⁡(2πt)\lambda_t=\frac{3}{2}-\frac{1}{2}\cos(2\pi t). The trap compresses during the first half-cycle and expands during the second. The mechanical work and Gibbs free-energy difference are

W[0,T]=12∫0Tλ˙t∣Xt∣2dt,ΔFT=log⁡λT.(25)W[0,T]=\frac{1}{2}\int_0^T\dot{\lambda}_t|X_t|^2\mathrm{d}t,\qquad\Delta F_T=\log\lambda_T. \tag*{(25)}

The Gibbsian ISS πt\pi_t ensures that both the mechanical and Active Jarzynski relations hold for this protocol:

Jmech(T):=−log⁡⟨e−W⟩π0=ΔFT,(26)J_{\mathrm{mech}}(T):=-\log\langle e^{-W}\rangle_{\pi_0}=\Delta F_T, \tag*{(26)}
JAJ(T):=−log⁡⟨e−(W+Δsact)⟩π0=ΔFT.(27)J_{\mathrm{AJ}}(T):=-\log\langle e^{-(W+\Delta s_{\mathrm{act}})}\rangle_{\pi_0}=\Delta F_T. \tag*{(27)}

Figure 3 compares their sample estimates. At T=1T=1 the Hamiltonian returns to its initial value and ΔF1=0\Delta F_1=0. Since work fluctuates across trajectories, strict Jensen’s inequality applied to ⟨e−W⟩π0=1\langle e^{-W}\rangle_{\pi_0}=1 gives ⟨W⟩π0>0\langle W\rangle_{\pi_0}>0, while nonzero active currents yield positive mean active entropy. A spatial reflection M=diag⁡(1,−1)M=\operatorname{diag}(1,-1) maps the original dynamics to their dual while preserving ∣Xt∣2|X_t|^2, and hence WW. Consequently, ⟨e−Δsact∣W=w⟩=1\langle e^{-\Delta s_{\mathrm{act}}\mid W=w}\rangle=1, so this model has the classical Crooks work ratio despite positive mean active entropy.

Mechanical and Active Jarzynski estimates

Figure 3. Mechanical and Active Jarzynski estimates for λt=32−12cos⁡(2πt)\lambda_t=\frac{3}{2}-\frac{1}{2}\cos(2\pi t). Open squares and filled circles show Monte Carlo estimates of Jmech(T)J_{\mathrm{mech}}(T) and JAJ(T)J_{\mathrm{AJ}}(T), respectively. The dashed curve gives their common theoretical value ΔFT=log⁡λT\Delta F_T=\log\lambda_T.

Discussion. By combining current-reversed dual dynamics with exponential gauge tilts, we have established a general framework for finite-time fluctuation relations in driven jump-diffusions without detailed balance. A central physical insight of our work is the coexistence of the mechanical and Active Jarzynski equalities in systems admitting a Gibbsian ISS: while the conventional exponential work average ⟨e−W⟩\langle e^{-W}\rangle recovers the free-energy difference using the stationarity of the Gibbs density under the frozen forward dynamics, the active counterpart ⟨e−(W+Δsact)⟩\langle e^{-(W+\Delta s_{\mathrm{act}})}\rangle incorporates the housekeeping heat associated with the non-equilibrium currents using the corresponding stationarity under the frozen dual dynamics. In the corresponding Active Crooks relation [Eq. (22)], the conditional factor ⟨e−Δsact∣W=w⟩P\langle e^{-\Delta s_{\mathrm{act}}}\mid W=w\rangle_P measures how autonomous active fluctuations skew the work distribution away from the classical Crooks work relation. Beyond the Gibbs–ISS regime, our general weighted endpoint identity (detailed in the End Matter) remains valid for arbitrary reference Hamiltonians. Experimentally, our results are directly applicable to optically trapped active colloids, engineered microswimmers with run-and-tumble or Lévy-flight kinematics, and single-molecule manipulation in active viscoelastic media, offering a rigorous route to simultaneously reconstruct effective free-energy landscapes and quantify hidden housekeeping dissipation from finite-time trajectory measurements.

Acknowledgments. We thank Liu Hong, Qiao Huang, Yubin Lu, and Henri Orland for helpful discussions. S.F. and J.D. acknowledge support from the Guangdong Provincial Key Laboratory of Mathematical and Neural Dynamical Systems (Grant No. 2024B1212010004), the Dongguan Key Laboratory for Data Science and Intelligent Medicine, the Innovation Team Project for Higher Education Institutions in Guangdong Province, “Cross-disciplinary Research Team on Data Science and Intelligent Medicine” (Grant No. 2023KCXTD054), the International Collaboration Fund for Creative Research Teams of the National Natural Science Foundation of China (Grant No. W2541005), and the Natural Science Foundation of Guangdong Province (Grant No. 2025A1515011188). B.M. acknowledges support from the National Natural Science Foundation of China (NSFC) (Grant No. 12575045). We used GPT to assist with checking and improving the draft, and DeepSeek and GLM to assist with the numerical simulations.

End Matter

The compared path measures, including their initial distributions, satisfy Pα∼P\mathcal{P}^{\alpha}\sim\mathcal{P}, and their Girsanov density processes are assumed to be true martingales of unit expectation [13, 19]. We retain the main-text notation, with all state variables even under reversal and XcX^{c} denoting the continuous component, dXtc=bt(Xt)dt+2a dWt\mathrm{d}X^{c}_{t}=b_{t}(X_{t})\mathrm{d}t+\sqrt{2a}\,\mathrm{d}W_{t}. We write tjt_{j} for the jump times and use ∂t\partial_{t} for time differentiation at fixed position.

Currents and entropy production

The Fokker–Planck equation is

∂tρt=−∇⋅Jtloc−∫Jtnl(x,y) dy.\partial_{t}\rho_{t}=-\nabla\cdot J_{t}^{\mathrm{loc}}-\int J_{t}^{\mathrm{nl}}(x,y)\,\mathrm{d}y.

with currents

Jtloc=btρt−A∇ρt,J_{t}^{\mathrm{loc}}=b_{t}\rho_{t}-A\nabla\rho_{t},
Jtnl(x,y)=ρt(x)qt(x,y)−ρt(y)qt(y,x).J_{t}^{\mathrm{nl}}(x,y)=\rho_{t}(x)q_{t}(x,y)-\rho_{t}(y)q_{t}(y,x).

For the ISS πt\pi_{t}, we define the active affinities

ftloc(x)=A−1bt(x)−∇log⁡πt(x),f_{t}^{\mathrm{loc}}(x)=A^{-1}b_{t}(x)-\nabla\log\pi_{t}(x),
ftnl(x,y)=log⁡πt(x)qt(x,y)πt(y)qt(y,x).(28)f_{t}^{\mathrm{nl}}(x,y)=\log\frac{\pi_{t}(x)q_{t}(x,y)}{\pi_{t}(y)q_{t}(y,x)}. \tag*{(28)}

The total thermodynamic affinities satisfy

A−1Jtlocρt=ftloc−∇log⁡ρtπt,A^{-1}\frac{J_{t}^{\mathrm{loc}}}{\rho_{t}}=f_{t}^{\mathrm{loc}}-\nabla\log\frac{\rho_{t}}{\pi_{t}},
log⁡ρt(x)qt(x,y)ρt(y)qt(y,x)=ftnl(x,y)+log⁡ρt(x)πt(x)−log⁡ρt(y)πt(y).\log\frac{\rho_{t}(x)q_{t}(x,y)}{\rho_{t}(y)q_{t}(y,x)}=f_{t}^{\mathrm{nl}}(x,y)+\log\frac{\rho_{t}(x)}{\pi_{t}(x)}-\log\frac{\rho_{t}(y)}{\pi_{t}(y)}.

At each frozen time tt, ftlocf_{t}^{\mathrm{loc}} is the continuous active force and ftnlf_{t}^{\mathrm{nl}} the jump affinity sustaining stationary currents. The terms involving ρt/πt\rho_{t}/\pi_{t} drive relaxation toward πt\pi_{t} under the frozen dynamics and give the nonadiabatic entropy-production rate when paired with the corresponding currents [16]. The total entropy-production rate has the current–affinity form

S˙tot(t)=∫Jtloc⋅A−1Jtlocρt dx+12∬Jtnl(x,y)log⁡ρt(x)qt(x,y)ρt(y)qt(y,x) dxdy.(29)\dot{S}_{\mathrm{tot}}(t)=\int J_{t}^{\mathrm{loc}}\cdot A^{-1}\frac{J_{t}^{\mathrm{loc}}}{\rho_{t}}\,\mathrm{d}x+\frac{1}{2}\iint J_{t}^{\mathrm{nl}}(x,y)\log\frac{\rho_{t}(x)q_{t}(x,y)}{\rho_{t}(y)q_{t}(y,x)}\,\mathrm{d}x\mathrm{d}y. \tag*{(29)}

Substituting the two affinity decompositions above into Eq. (29) gives the active and nonadiabatic rates S˙act(t)\dot{S}_{\mathrm{act}}(t) and S˙na(t)\dot{S}_{\mathrm{na}}(t) in Eq. (12),

S˙act(t)=∫Jtloc⋅ftloc dx+12∬Jtnlftnl dxdy,(30)\dot{S}_{\mathrm{act}}(t)=\int J_{t}^{\mathrm{loc}}\cdot f_{t}^{\mathrm{loc}}\,\mathrm{d}x+\frac{1}{2}\iint J_{t}^{\mathrm{nl}}f_{t}^{\mathrm{nl}}\,\mathrm{d}x\mathrm{d}y, \tag*{(30)}
S˙na(t)=S˙tot(t)−S˙act(t)=−∫(Lt∗ρt)log⁡ρtπt dx.(31)\dot{S}_{\mathrm{na}}(t)=\dot{S}_{\mathrm{tot}}(t)-\dot{S}_{\mathrm{act}}(t)=-\int(\mathcal{L}_{t}^{*}\rho_{t})\log\frac{\rho_{t}}{\pi_{t}}\,\mathrm{d}x. \tag*{(31)}

The last equality uses integration by parts and antisymmetry of JtnlJ_{t}^{\mathrm{nl}}. Differentiating D(ρt∥πt)D(\rho_{t}\Vert\pi_{t}) with respect to tt yields Eq. (13). The total rate is nonnegative because its continuous term is quadratic and (u−v)log⁡(u/v)≥0(u-v)\log(u/v)\geq0. For h=ρt/πth=\rho_{t}/\pi_{t}, concavity of the logarithm gives hLtlog⁡h≤Lthh\mathcal{L}_{t}\log h\leq\mathcal{L}_{t}h, while ∫πtLth dx=0\int\pi_{t}\mathcal{L}_{t}h\,\mathrm{d}x=0. Thus S˙na=−∫πthLtlog⁡h dx≥0\dot{S}_{\mathrm{na}}=-\int\pi_{t}h\mathcal{L}_{t}\log h\,\mathrm{d}x\geq0.

Path likelihoods

For reference coefficients btαb_{t}^{\alpha}, qtαq_{t}^{\alpha} and initial density ρ0α\rho_{0}^{\alpha}, write ut=(2a)−1(btα−bt)u_{t}=(\sqrt{2a})^{-1}(b_{t}^{\alpha}-b_{t}). Girsanov’s formula along the forward trajectory gives

Iα=log⁡ρ0(X0)ρ0α(X0)−∫0Tut(Xt)⋅dWt+12∫0T∣ut(Xt)∣2 dt+∑tj≤Tlog⁡qtj(Xtj−,Xtj)qtjα(Xtj−,Xtj)−∫0T∫[qt−qtα](Xt,y) dydt.(32)\mathcal{I}^{\alpha}=\log\frac{\rho_{0}(X_{0})}{\rho_{0}^{\alpha}(X_{0})}-\int_{0}^{T}u_{t}(X_{t})\cdot\mathrm{d}W_{t}+\frac{1}{2}\int_{0}^{T}|u_{t}(X_{t})|^{2}\,\mathrm{d}t+\sum_{t_{j}\leq T}\log\frac{q_{t_{j}}(X_{t_{j-}},X_{t_{j}})}{q_{t_{j}}^{\alpha}(X_{t_{j-}},X_{t_{j}})}-\int_{0}^{T}\int[q_{t}-q_{t}^{\alpha}](X_{t},y)\,\mathrm{d}y\mathrm{d}t. \tag*{(32)}

In the P\mathcal{P}-expectation of Eq. (32), the Brownian integral vanishes and the jump sum is evaluated using the compensator qt(Xt−,y)dydtq_{t}(X_{t-},y)\mathrm{d}y\mathrm{d}t, giving Eq. (7). The escape-rate correction integrates qt−qtαq_{t}-q_{t}^{\alpha} as a single difference. For Θ#P\Theta_{\#}\mathcal{P}, the initial contribution is log⁡[ρ0(X0)/ρT(X0)]\log[\rho_{0}(X_{0})/\rho_{T}(X_{0})], with mean D(ρ0∥ρT)D(\rho_{0}\Vert\rho_{T}).

ISS decomposition of entropy production

At the frozen density πt\pi_{t}, the dual coefficients give

bt†πt−A∇πt=−[btπt−A∇πt].(33)b_{t}^{\dagger}\pi_{t}-A\nabla\pi_{t}=-[b_{t}\pi_{t}-A\nabla\pi_{t}]. \tag*{(33)}

The environment-entropy flow in the entropy balance is

Q[0,T]=∫0TA−1bt(Xt)∘dXtc+∑tj≤Tlog⁡qtj(Xtj−,Xtj)qtj(Xtj,Xtj−).Q[0,T]=\int_{0}^{T}A^{-1}b_{t}(X_{t})\circ\mathrm{d}X_{t}^{c}+\sum_{t_{j}\leq T}\log\frac{q_{t_{j}}(X_{t_{j-}},X_{t_{j}})}{q_{t_{j}}(X_{t_{j}},X_{t_{j-}})}.

Here ∘\circ denotes Stratonovich integration. Using the affinities in Eq. (28), the active contribution is

Δsact[0,T]=∫0Tftloc(Xt)∘dXtc+∑tj≤Tftjnl(Xtj−,Xtj).(34)\Delta s_{\mathrm{act}}[0,T]=\int_{0}^{T}f_{t}^{\mathrm{loc}}(X_{t})\circ\mathrm{d}X_{t}^{c}+\sum_{t_{j}\leq T}f_{t_{j}}^{\mathrm{nl}}(X_{t_{j-}},X_{t_{j}}). \tag*{(34)}

The continuous term integrates the active force along displacements, and each jump contributes ftnl(x,y)f_{t}^{\mathrm{nl}}(x,y). Set Vtss=−log⁡πtV_{t}^{\mathrm{ss}}=-\log\pi_{t}. Substituting Eq. (28) into Eq. (34) and applying the chain rule to Vtss(Xt)V_{t}^{\mathrm{ss}}(X_{t}) gives Δsact=Q+VTss(XT)−V0ss(X0)−YHS\Delta s_{\mathrm{act}}=Q+V_{T}^{\mathrm{ss}}(X_{T})-V_{0}^{\mathrm{ss}}(X_{0})-Y_{\mathrm{HS}}. Using the main-text definitions of Δssys\Delta s_{\mathrm{sys}} and Δsrel\Delta s_{\mathrm{rel}} recovers the trajectory entropy balance. To evaluate the P\mathcal{P}-expectations of Δstot=Δssys+Q\Delta s_{\mathrm{tot}}=\Delta s_{\mathrm{sys}}+Q and Eq. (34), we use Stratonovich-to-Itô conversion and the jump compensator. Differentiating these expectations with respect to TT gives S˙tot(T)\dot{S}_{\mathrm{tot}}(T) and S˙act(T)\dot{S}_{\mathrm{act}}(T) in Eqs. (29) and (30). Averaging the trajectory entropy balance and differentiating with respect to TT gives Eq. (12). For a Gibbsian ISS, substituting Vtss(x)=H(t,x)−FtV_{t}^{\mathrm{ss}}(x)=H(t,x)-\mathcal{F}_{t} and YHS=W−ΔFY_{\mathrm{HS}}=W-\Delta\mathcal{F} into the trajectory identity for Δsact\Delta s_{\mathrm{act}} gives

Q+H(T,XT)−H(0,X0)=W+Δsact.(35)Q+H(T,X_{T})-H(0,X_{0})=W+\Delta s_{\mathrm{act}}. \tag*{(35)}

For transitions mediated by a single heat bath, the Einstein relation for diffusion and local detailed balance for jumps identify QQ as the heat released to the bath in thermal units [24] [14]. Equation (35) then expresses the first law, with supplied work W+ΔsactW+\Delta s_{\mathrm{act}}.

For forward and dual processes with the same initial density, substitution of Eq. (14) into Eq. (32) combines the continuous drift correction and the jump escape-rate correction into Lt∗πt/πt\mathcal{L}_t^* \pi_t/\pi_t. Hence

Iπ=Δsact+∫0TLt∗πtπt(Xt) dt=Δsact.(36)\mathcal{I}^{\pi}=\Delta s_{\mathrm{act}}+\int_0^T\frac{\mathcal{L}_t^*\pi_t}{\pi_t}(X_t)\,\mathrm{d}t=\Delta s_{\mathrm{act}}. \tag*{(36)}

The rate of path-space relative entropy in Eq. (7) is nonnegative for this comparison, giving S˙act(t)≥0\dot{S}_{\mathrm{act}}(t)\geq0.

For a protocol-reversed experiment with initial density ρ0R\rho_0^R, let ρsR\rho_s^R denote its density at time ss. The reversed-time density satisfies ∂t(ρT−tR)+LT−t∗ρT−tR=0\partial_t(\rho_{T-t}^R)+\mathcal{L}_{T-t}^{*}\rho_{T-t}^R=0. Applying Eq. (32) with the path-reversal coefficients built from ρT−tR\rho_{T-t}^R, followed by the chain rule for log⁡ρT−tR(Xt)\log\rho_{T-t}^R(X_t), gives

log⁡dPd(Θ#PR)=log⁡ρ0(X0)ρ0R(XT)+Q[0,T]+∫0T∂t(ρT−tR)+LT−t∗ρT−tRρT−tR(Xt) dt.(37)\log\frac{\mathrm{d}\mathcal{P}}{\mathrm{d}(\Theta_{\#}\mathcal{P}_R)} =\log\frac{\rho_0(X_0)}{\rho_0^R(X_T)}+Q[0,T] +\int_0^T\frac{\partial_t(\rho_{T-t}^R)+\mathcal{L}_{T-t}^{*}\rho_{T-t}^R}{\rho_{T-t}^R}(X_t)\,\mathrm{d}t. \tag*{(37)}

The integral vanishes by the equation for ρT−tR\rho_{T-t}^R. Taking ρ0R=ρT\rho_0^R=\rho_T gives Δstot\Delta s_{\mathrm{tot}}, as in Table I. With forward and protocol-reversed initial densities π0\pi_0 and πT\pi_T, respectively, the same formula gives

Reference dynamicsMeasure Pα\mathcal{P}^{\alpha}Relative action Iα\mathcal{I}^{\alpha}
Current-reversed dualPπ\mathcal{P}^{\pi}Δsact\Delta s_{\mathrm{act}}
Direct path reversalΘ#P\Theta_{\#}\mathcal{P}IΘ\mathcal{I}^{\Theta}
Protocol-and-path reversalΘ#PR\Theta_{\#}\mathcal{P}_RΔstot\Delta s_{\mathrm{tot}}

Table I. Path-measure comparisons on [0,T][0,T] and associated trajectory functionals Iα=log⁡(dP/dPα)\mathcal{I}^{\alpha}=\log(\mathrm{d}\mathcal{P}/\mathrm{d}\mathcal{P}^{\alpha}). Forward (P\mathcal{P}) and dual (Pπ\mathcal{P}^{\pi}) processes are initialized in ρ0\rho_0, whereas path-reversed (Θ#P\Theta_{\#}\mathcal{P}) and protocol-reversed (PR\mathcal{P}_R) ensembles are initialized in ρT\rho_T. Each comparison satisfies ⟨e−Iα⟩P=1\langle e^{-\mathcal{I}^{\alpha}}\rangle_{\mathcal{P}}=1.

log⁡dPd(Θ#PR)=log⁡π0(X0)πT(XT)+Q[0,T]=H(T,XT)−H(0,X0)+Q[0,T]−ΔF.(38)\begin{aligned} \log\frac{\mathrm{d}\mathcal{P}}{\mathrm{d}(\Theta_{\#}\mathcal{P}_R)} =\log\frac{\pi_0(X_0)}{\pi_T(X_T)}+Q[0,T] \\ &=H(T,X_T)-H(0,X_0)+Q[0,T]-\Delta F. \tag*{(38)} \end{aligned}

The second line uses the Gibbs densities. Substituting Q+H(T,XT)−H(0,X0)=W+ΔsactQ+H(T,X_T)-H(0,X_0)=W+\Delta s_{\mathrm{act}} from Eq. (35) into Eq. (38), then exponentiating the negative of this log ratio, gives the path-measure likelihood ratio in Eq. (19).

Weighted endpoint identity. For a prescribed Hamiltonian H(t,x)H(t,x), initialize the original process from e−H(0,⋅)/Z0e^{-H(0,\cdot)}/Z_0. For a smooth bounded test function φ\varphi, we evaluate ⟨φ(XT)e−Γ[0,T]⟩P\langle\varphi(X_T)e^{-\Gamma[0,T]}\rangle_{\mathcal{P}}, where

Γ[0,T]=Q[0,T]+H(T,XT)−H(0,X0).\Gamma[0,T]=Q[0,T]+H(T,X_T)-H(0,X_0).

When the Gibbs density πt=e−H(t,⋅)/Zt\pi_t=e^{-H(t,\cdot)}/Z_t satisfies Lt∗πt=0\mathcal{L}_t^*\pi_t=0, Eq. (35) gives Γ=W+Δsact\Gamma=W+\Delta s_{\mathrm{act}}. For the dual initialized from the same density π0\pi_0, Eq. (36) then gives e−Γ dP=e−W dPπe^{-\Gamma}\,\mathrm{d}\mathcal{P}=e^{-W}\,\mathrm{d}\mathcal{P}^{\pi}. For general H(t,x)H(t,x), set g(t,x)=⟨φ(XT)e−Γ[t,T]⟩t,xg(t,x)=\langle\varphi(X_T)e^{-\Gamma[t,T]}\rangle_{t,x}, where ⟨⋅⟩t,x\langle\cdot\rangle_{t,x} refers to the original process started at (t,x)(t,x). By the definition of gg, e−Γ[t,s]g(s,Xs)e^{-\Gamma[t,s]}g(s,X_s) is a martingale in ss, so its Itô drift vanishes, giving

∂tg+Ktg−(∂tH)(t,x)g=0,g(T,x)=φ(x).(39)\partial_t g+K_tg-(\partial_tH)(t,x)g=0,\qquad g(T,x)=\varphi(x). \tag*{(39)}

The tilted spatial operator KtK_t, generally with Kt1≠0K_t1\ne0, and its Lebesgue adjoint satisfy

Ktφ=eH(t,x)Lt∗(e−H(t,x)φ),K_t\varphi=e^{H(t,x)}\mathcal{L}_t^*\left(e^{-H(t,x)}\varphi\right),
Kt∗e−H(t,x)=e−H(t,x)Lt1=0.K_t^*e^{-H(t,x)}=e^{-H(t,x)}\mathcal{L}_t1=0.

For I(t)=∫e−H(t,x)g(t,x) dxI(t)=\int e^{-H(t,x)}g(t,x)\,\mathrm{d}x, differentiation with respect to tt using Eq. (39) cancels the terms containing ∂tH\partial_tH. Integration by parts then gives I′(t)=−∫gKt∗e−H(t,⋅) dx=0I'(t)=-\int gK_t^*e^{-H(t,\cdot)}\,\mathrm{d}x=0. Evaluating I(0)=I(T)I(0)=I(T) and dividing by Z0Z_0 yields

⟨φ(XT)e−Γ[0,T]⟩e−H(0,⋅)/Z0=1Z0∫φ(x)e−H(T,x) dx.(40)\left\langle\varphi(X_T)e^{-\Gamma[0,T]}\right\rangle_{e^{-H(0,\cdot)}/Z_0} =\frac{1}{Z_0}\int\varphi(x)e^{-H(T,x)}\,\mathrm{d}x. \tag*{(40)}

Equation (37) with reverse initial density e−H(T,⋅)/ZTe^{-H(T,\cdot)}/Z_T gives e−ΓdP=(ZT/Z0)d(Θ#PR)e^{-\Gamma}\mathrm{d}\mathcal{P}=(Z_T/Z_0)\mathrm{d}(\Theta_{\#}\mathcal{P}_R), so Eq. (40) also holds for bounded measurable φ\varphi. Taking φ=1\varphi=1 gives ⟨e−Γ⟩P=ZT/Z0=e−ΔF\langle e^{-\Gamma}\rangle_{\mathcal{P}}=Z_T/Z_0=e^{-\Delta F}. When πt=e−H(t,⋅)/Zt\pi_t=e^{-H(t,\cdot)}/Z_t is the ISS, Kt=LtπK_t=\mathcal{L}_t^\pi and e−ΓdP=e−WdPπe^{-\Gamma}\mathrm{d}\mathcal{P}=e^{-W}\mathrm{d}\mathcal{P}^{\pi} give ZT/Z0Z_T/Z_0 and Eq. (20).

For forward and dual processes initialized from π0\pi_0 under the same protocol, conditioning dPπ/dP=e−Δsact\mathrm{d}\mathcal{P}^{\pi}/\mathrm{d}\mathcal{P}=e^{-\Delta s_{\mathrm{act}}} from Eq. (36) on W=wW=w gives

⟨e−Δsact∣W=w⟩P=pπ(w)pF(w).(41)\left\langle e^{-\Delta s_{\mathrm{act}}}\mid W=w\right\rangle_{\mathcal{P}}=\frac{p^{\pi}(w)}{p_F(w)}. \tag*{(41)}

Here pFp_F and pπp^{\pi} are the densities of the same work functional WW under P\mathcal{P} and Pπ\mathcal{P}^{\pi}, respectively. The equality holds almost everywhere on the support of pFp_F.

Examples and numerical methods. For each protocol of Eq. (23), we use jump-adapted Euler–Maruyama [25 Chap. 8] with a maximum step 10−310^{-3} and a total of 5×1065\times10^6 trajectories from five independent seeds. Jump waiting times are sampled from an exponential distribution with unit rate. Observables are recorded at intervals of 0.010.01 over [0,1][0,1]. The continuous contribution to the active entropy is evaluated using left-point Itô sums, since the rotational field has zero Stratonovich correction, and jump affinities are summed separately. Work is computed using trapezoidal quadrature on each jump-free segment, taking the terminal position before any jump. Exponential weights are pooled across all trajectories before taking the negative logarithm.

To compare IΘ\mathcal{I}^{\Theta} with Δstot\Delta s_{\mathrm{tot}}, take A=IA=I, b=−∇Vb=-\nabla V, qt(x,y)=γtpeq(y)q_t(x,y)=\gamma_t p_{\mathrm{eq}}(y), and X0∼peqX_0\sim p_{\mathrm{eq}} in Eq. (5), where peq=Z−1e−Vp_{\mathrm{eq}}=Z^{-1}e^{-V} [22, 23]. These jumps redraw from peqp_{\mathrm{eq}} and satisfy detailed balance at each time, unlike unidirectional resetting to a fixed point. Then ρt=peq\rho_t=p_{\mathrm{eq}} and qtΘ(x,y)=γT−tpeq(y)q_t^\Theta(x,y)=\gamma_{T-t}p_{\mathrm{eq}}(y). Equations (7) and (29) give ⟨IΘ⟩P=∫0T[γtlog⁡(γt/γT−t)−γt+γT−t] dt\langle\mathcal{I}^{\Theta}\rangle_{\mathcal{P}}=\int_0^T[\gamma_t\log(\gamma_t/\gamma_{T-t})-\gamma_t+\gamma_{T-t}]\,\mathrm{d}t and ⟨Δstot⟩P=0\langle\Delta s_{\mathrm{tot}}\rangle_{\mathcal{P}}=0. For γt=2+sin⁡t\gamma_t=2+\sin t, ⟨IΘ⟩P=0\langle\mathcal{I}^{\Theta}\rangle_{\mathcal{P}}=0 when T=πT=\pi, whereas it is positive when T=2πT=2\pi.

References

  1. [2]G. Gallavotti and E. G. D. Cohen, Dynamical Ensembles in Nonequilibrium Statistical Mechanics, Phys. Rev. Lett. 74, 2694 (1995).pubmed.ncbi.nlm.nih.gov/10057994
  2. [3]C. Jarzynski, Nonequilibrium equality for free energy differences, Phys. Rev. Lett. 78, 2690 (1997).DOI
  3. [4]G. E. Crooks, Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences, Phys. Rev. E 60, 2721 (1999).DOI
  4. [5]G. Szamel, Stochastic thermodynamics for self-propelled particles, Phys. Rev. E 100, 050603(R) (2019).DOI
  5. [6]P. Pietzonka and U. Seifert, Entropy production of active particles and for particles in active baths, J. Phys. A: Math. Theor. 51, 01LT01 (2018).DOI
  6. [7]I. Golding, J. Paulsson, S. M. Zawilski, and E. C. Cox, Real-time kinetics of gene activity in individual bacteria, Cell 123, 1025 (2005).pubmed.ncbi.nlm.nih.gov/16360033
  7. [8]G. Szamel, Single active particle in a harmonic potential: Question about the existence of the Jarzynski relation, Phys. Rev. E 107, 054602 (2023).DOI
  8. [9]L. Dabelow, S. Bo, and R. Eichhorn, Irreversibility in active matter systems: Fluctuation theorem and mutual information, Phys. Rev. X 9, 021009 (2019).DOI
  9. [12]S. Fan and Q. Zhang, Free energy dissipation and a decomposition of general jump diffusions on $\mathbb{R}^n$ without detailed balance, J. Stat. Phys. 193, 91 (2026), doi:10.1007/s10955-026-03667-2.DOI
  10. [13]P. Cheridito, D. Filipović, and M. Yor, Equivalent and absolutely continuous measure changes for jump-diffusion processes, Ann. Appl. Probab. 15, 1713–1732 (2005).DOI
  11. [15]U. Seifert, Entropy production along a stochastic trajectory and an integral fluctuation theorem, Phys. Rev. Lett. 95, 040602 (2005).pubmed.ncbi.nlm.nih.gov/16090792
  12. [16]M. Esposito and C. Van den Broeck, Three detailed fluctuation theorems, Phys. Rev. Lett. 104, 090601 (2010).pubmed.ncbi.nlm.nih.gov/20366974
  13. [17]V. Y. Chernyak, M. Chertkov, and C. Jarzynski, Path-integral analysis of fluctuation theorems for general Langevin processes, J. Stat. Mech. (2006), P08001.DOI
  14. [18]C. Hartmann, C. Schütte, and W. Zhang, Jarzynski’s equality, fluctuation theorems, and variance reduction: Mathematical analysis and numerical algorithms, J. Stat. Phys. 175, 1214–1261 (2019).DOI
  15. [19]Jacob, J., & Shiryaev, A. N. (2003). Limit Theorems for Stochastic Processes.
  16. [21]G. Conforti and C. Léonard, Time reversal of Markov processes with jumps under a finite entropy condition, Stochastic Process. Appl. 144, 85–124 (2022), doi:10.1016/j.spa.2021.10.002.DOI
  17. [22]J. Fuchs, S. Goldt, and U. Seifert, Stochastic thermodynamics of resetting, Europhys. Lett. 113, 60009 (2016).DOI
  18. [23]A. Pal and S. Rahav, Integral fluctuation theorems for stochastic resetting systems, Phys. Rev. E 96, 062135 (2017).DOI
  19. [24]C. Maes, Local detailed balance, SciPost Phys. Lect. Notes 32 (2021), doi:10.21468/SciPostPhysLectNotes.32.DOI
  20. [25]E. Platen and N. Bruti-Liberati, Numerical Solution of Stochastic Differential Equations with Jumps in Finance (Springer, Berlin, Heidelberg, 2010).DOI

Paper details

vlnc:W_fdy8xrkw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

GPT (multiple versions), DeepSeek (multiple versions), GLM (multiple versions)

View statement
Looking for related papers…
Contents