{"id":"edfff91c-1e18-4f3c-867e-2a349da8670d","arxiv_id":"2411.08311","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A generalized work theorem shows exponentiated entropy production is a martingale conditional on any fixed initial state, for arbitrary backward processes, in both overdamped and underdamped Langevin systems.","lead":"This paper derives a generalized family of work theorems for Langevin dynamics by pairing forward trajectories with a freely chosen backward probability flow, using Ito calculus instead of path integrals. The central identity holds conditional on any starting point and at any stopping time, which removes the need to average over a prescribed initial distribution.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (24) is conditional on an unverified Novikov condition; for the paper's own Gaussian ψ-family, bounded stopping times near the terminal singularity violate it and should break E[e^{θτ}|x0]=1.","rationale":"Reader's weakest assumption is exactly the Novikov/positivity condition; I agree it is the load-bearing point. The overdamped algebra is correct and the conditional identity Eq. (24) is a solid martingale statement when the integrability hypothesis holds. But the paper presents Eq. (24) as holding for any ψ and any bounded stopping time, and the numerical/analytical example shows the hypothesis is non-vacuous: natural ψ's have terminal singularities, and for such ψ there exist bounded τ for which Novikov fails and the martingale is strict. This keeps the result useful only if the authors add explicit sufficient conditions (e.g., bounds on ∇lnψ up to the stopping horizon) and restrict τ to the interval on which ψ is nondegenerate. The underdamped Eq. (39) factor-2 error is real and should be fixed, but Appendix A.2 already contains the correct derivation, so it does not invalidate the final underdamped theorem. Thus the reader's CONDITIONAL verdict stands unchanged.","tokens_in":17828,"tokens_out":30103,"duration_ms":299614,"concrete_test":"Using the §3 constant-force example with γ=β=f=1, x0=0, choose ψ from Eq. (50) with t2=1, and take the deterministic bounded stopping time τ=1−ε for ε=10^{-2} and ε=10^{-4}. First check the Novikov integral: estimate E[exp(1/2∫_0^τ a_s² ds)] with a_s=sqrt(2)[1−(x_s−(1−s))/(2(1−s))], x_s=s+sqrt(2)B_s. As ε decreases the estimate should diverge or fail to converge. Then compute E[e^{θτ}|x0=0] by Monte Carlo (≥10^7 paths). If this expectation is significantly below 1 for small ε while the theorem claims all bounded stopping times, the missing Novikov hypothesis is load-bearing. An analytic version: show E[exp(c B_τ²/ε)]=∞ for c/ε>1/(2τ), which is implied by a lower bound on the integral over [τ−ε,τ].","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (24) is the central claim, and its proof reduces to showing e^{θ_t} is a true martingale. The paper invokes the Novikov condition (Eq. 11) but never verifies it for any ψ, nor states conditions on ψ that would guarantee it. This is not a purely technical gap. In the constant-force example of §3, ψ(x,t)=p(x,t2−t) has a singularity at t2. The martingale integrand is a_t=sqrt(2/(γβ))[β(∇H−f)+∇lnψ]; for γ=β=f=1 it is a_t=sqrt(2)[1−(x_t−(t2−t))/(2(t2−t))]. For a bounded stopping time τ=t2−ε, the quadratic variation contribution on a short interval before τ is at least of order (sup B²)/ε, so E[exp(1/2∫_0^τ a_s² ds)] is infinite for ε small enough (the Gaussian tail threshold is exceeded). Thus Novikov fails for some bounded stopping times, and the nonnegative local martingale can be strict, giving E[e^{θτ}|x0]<1. The paper acknowledges the Novikov requirement in the Discussion, but the theorem as framed ('any bounded stopping time') needs explicit extra hypotheses: ψ strictly positive and sufficiently smooth on [0,τ], and a verified exponential integrability condition. A secondary issue is the main-text underdamped derivation: Eq. (39) is missing the −γv·∇lnψ/m drift contribution from Eq. (36) (the divergence of the friction term), so the rearrangement to Eq. (40) is algebraically false; Appendix A.2 has the correct coefficient and repairs the theorem, but Eq. (39) should be corrected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a stochastic-calculus derivation of exponential martingales associated with entropy production for overdamped and underdamped Langevin dynamics. The central object is a backward density ψ(x,t) that solves a backward Fokker-Planck equation with an arbitrary initial condition, and the process θ_t = −βW_t + βΔH(x_t,t) + ln(ψ(x_t,t)/ψ(x_0,0)). The authors claim that e^{θ_t} is an exponential martingale and that, for any bounded stopping time τ, E[e^{θ_τ}|x_0] = 1 for every initial state x_0 (Eq. (24)). This is presented as a generalization of the Manzano et al. result, with a conditional version in Eq. (27), a stationary version in Eq. (28), and an underdamped counterpart in Eqs. (39)-(41). The paper also discusses violations of the fluctuation-dissipation relation and gives numerical examples based on a drift-diffusion model of kinetic proofreading and an Ornstein-Uhlenbeck process.","tokens_in":18122,"tokens_out":15797,"duration_ms":153125,"significance":"The conditional martingale identity is a genuine conceptual strengthening if it holds under stated conditions: it removes the initial-distribution average, allows decoupling of forward and backward processes, and gives a concrete route to work theorems for singular initial distributions. The Ito-calculus derivation is self-contained and avoids path integrals, and the appendix treatment of general diffusion coefficients is useful. The numerical examples illustrate the scope of the claimed identities. However, the theorem as stated is not fully proved: the Novikov condition is invoked but never verified, and the main-text underdamped derivation contains an algebraic inconsistency. These are repairable, but they are load-bearing for the central claims.","major_comments":[{"comment":"The main identity is stated for every bounded stopping time τ and every backward density ψ solving Eq. (16), but the proof only establishes that exp(θ_t) is a nonnegative local martingale. The Novikov condition (11) is asserted but never verified for any of the ψ families considered, and no sufficient conditions on H, f, and ψ are given. This is not a purely technical gap: for the Gaussian family ψ(x,t)=p(x,t_2−t) in Eq. (50), ψ is singular at t=t_2, and for a bounded stopping time such as τ=t_2−ε the integrand a_s = sqrt(2/(γβ))[β(∇H−f)+∇lnψ] has quadratic variation of order (sup_s B_s^2)/ε over a short interval, so E[exp(1/2 ∫_0^τ a_s^2 ds)] is infinite for ε sufficiently small. In that situation the optional stopping equality can fail; for a nonnegative local martingale one only has E[e^{θ_τ}|x_0] ≤ 1. The theorem should be restated with explicit hypotheses: ψ strictly positive and smooth on [0,τ], and a verified exponential integrability condition for the chosen ψ. The Discussion's acknowledgment that the Novikov condition is an \"inconvenient aspect\" does not repair the theorem statement, which currently claims \"∀ bounded stopping time\" without these qualifications.","section":"§2.2, Eqs. (11), (24); §3, Eq. (50)"},{"comment":"The underdamped rearrangement is algebraically incorrect. Equation (38) contains a drift term −(γ/m) v·∇_v lnψ dt after cancellation of the two nγ/m terms, but the claimed representation dθ_t = a_t·dB_t − (1/2)||a_t||^2 dt with a_t = sqrt(2γβ)v_t + sqrt(2γ/(βm^2))∇_v lnψ_t has a cross term −(2γ/m) v·∇_v lnψ dt. The correct Ito expansion of lnψ from Eq. (36) indeed gives the coefficient −2γ/m, as shown in Appendix A.2, Eq. (A21), at D=1/(βγ); Equation (38) omits one contribution coming from the ∂_tψ term. Consequently Eq. (39) is missing the factor 2, and Eq. (40) does not follow from Eq. (39) as written. The appendix calculation repairs the result, but the main-text derivation of the underdamped martingale must be corrected.","section":"§2.3, Eqs. (38)–(40)"}],"minor_comments":[{"comment":"The sentence mentioning the \"generalized Legendre-Frenchel transform\" contains a typo (Fenchel) and the reference [35] is listed twice as \"[35, 35, 36]\".","section":"§4, Discussion"},{"comment":"The explanation of the two nγ/m terms in Eq. (38) is terse; writing out the divergence ∇_v·[(−γv + ∇_xU − f)ψ] and the Ito correction from (dv)^2 explicitly would help the reader verify the cancellation.","section":"§2.3, Eq. (36)"},{"comment":"The numerical claim that E[e^{θ_τ}] = 1 for all choices of t_2 would be more convincing with error bars or a convergence criterion, since the expectation involves rare large deviations of the exponential.","section":"§3, Fig. 4"},{"comment":"The conditional statement Eq. (27) contains the ratio ρ(x_0,0) in the denominator; the paper should state explicitly that the identity is meaningful only when ρ(x_0,0)>0, and explain how the Dirac-delta application mentioned in the text avoids this issue.","section":"§2.2, Eq. (27)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising central idea and the overdamped derivation is clean, but the unconditional statement of the martingale identity is not supported without Novikov verification, and the underdamped main-text derivation has a concrete algebraic error that the appendix happens to repair. I recommend major revision rather than rejection because both issues are fixable within the manuscript's scope and the conditional-strengthening insight is worth publishing after the theorem is properly qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth a serious look. The overdamped martingale identity in Eq. (24) is correct as far as the stochastic calculus goes, and it genuinely extends earlier work by conditioning on the initial state x0 and allowing arbitrary initial conditions for the backward process. The stationary case is a nice practical improvement, and the underdamped extension is real. The derivation via Itô calculus is cleaner than the usual path-integral route and avoids some of the rigor gaps.\n\nThe main soft spot is the Novikov condition. The authors assert it but never verify it for any ψ, and the theorem is stated for \"any bounded stopping time\" without restrictions on ψ. Their own Gaussian ψ-family in the drift-diffusion example has a singularity at t2; for a bounded stopping time that gets within ε of t2, the integrand a_t grows like 1/ε and the Novikov expectation is infinite. So the identity can fail for such stopping times. The paper should state explicit hypotheses—ψ positive and smooth on [0,τ], and an exponential integrability condition—and then the claim follows. This is a genuine gap in the theorem statement, but not a flaw in the core argument; the Discussion already acknowledges the Novikov requirement.\n\nThe underdamped derivation has a factor-2 typo in Eq. (39): the first line has −γ/m v·∇v lnψ dt, but completing the square with the noise term requires −2γ/m. The second line and Appendix A.2 have the correct coefficient, so the theorem is fine after correction. A referee should ask for this fix.\n\nThe \"violation of work theorems\" discussion is a bit loosely framed—when D≠1/(βγ), the model no longer satisfies the fluctuation-dissipation relation, so the failure of the equality is a failure of the theorem's assumptions, not a contradiction. That's a minor wording issue.\n\nOverall, the central argument holds up. The paper is for the stochastic thermodynamics community, specifically people who work on martingale fluctuation theorems. It deserves a serious referee and publication after revision. I'd engage with it.","headline":"A genuinely useful generalization of martingale work theorems, with a correct central proof but a needed fix for the Novikov hypotheses and a factor-2 typo in Eq. (39).","tokens_in":18714,"tokens_out":15271,"would_cite":true,"duration_ms":126783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G44","60H10","82C31"],"pacs":["05.40.-a","05.70.Ln"],"model":"deepseek-v4-flash","headline":"Pairing a Langevin trajectory with any positive backward density solving the reversed Fokker-Planck equation yields the conditional identity $E[e^{\\theta_\\tau}|x_0]=1$ for every bounded stopping time and every initial state.","keywords":["martingale","work theorem","fluctuation theorem","entropy production","Langevin dynamics","backward Fokker-Planck equation","stopping time","fluctuation-dissipation relation"],"falsifier":"Run the linear overdamped process in Eq. (51) with $k=\\gamma=1$, $\\beta=1$, and $D=1/(\\beta\\gamma)=1$, starting from a Gaussian distribution; take $\\psi$ as the backward evolution of any Gaussian initial condition under the backward Fokker-Planck equation, stopped at a time $t_1$ earlier than any singularity of $\\psi$, and estimate $E[e^{\\theta_{t_1}}|x_0]$ from $10^7$ simulated trajectories binned by initial value. The paper's Eq. (24) predicts every bin equals 1; a bin whose estimate deviates by several standard errors would falsify the conditional work theorem.","tokens_in":17551,"feed_emoji":"🎲","tokens_out":15115,"duration_ms":129615,"temperature":0.7,"pith_summary":"This paper establishes that the exponential work-fluctuation identity is a martingale property that does not require the standard time-reversed ensemble: any positive density $\\psi$ that solves the backward Fokker-Planck equation with an arbitrary initial condition yields $E[e^{\\theta_\\tau}|x_0]=1$ for every bounded stopping time $\\tau$ and every initial state $x_0$, where $\\theta_t = -\\beta W_t + \\beta \\Delta H(x_t,t) + \\beta \\ln(\\psi(x_t,t)/\\psi(x_0,0))$. The proof uses the stochastic-calculus chain rule instead of path integrals, so it carries over to underdamped dynamics and to initial states that have no smooth density, such as a delta peak. This generalizes and strengthens the earlier ensemble-averaged work theorem, and a stationary choice of $\\psi$ gives a simplified identity requiring only the stationary distribution. If correct, the result lowers the computational burden of free-energy estimates and makes work theorems applicable to a wider class of nonequilibrium and biological systems.","feed_headline":"Any backward path choice yields a work theorem for each starting state","feed_subtitle":"The exponential of negative work is a martingale for every starting state, not just the ensemble average.","key_machinery":"The machinery is an exponential martingale built from a backward density. Given a probability density $\\psi(x,t)$ that satisfies the backward Fokker-Planck equation $-\\partial_t\\psi = (1/\\gamma)\\nabla\\cdot[(\\nabla H-f)\\psi] + (1/(\\beta\\gamma))\\Delta\\psi$ with an arbitrary initial condition, stochastic calculus gives $d e^{\\theta_t} = e^{\\theta_t} a_t\\cdot dB_t$ with $a_t = \\sqrt{2/(\\beta\\gamma)}[\\beta(\\nabla H-f)+\\nabla\\ln\\psi]$, so the drift term cancels exactly when the diffusion coefficient equals $1/(\\beta\\gamma)$. The square-exponential condition on $a_t$ ensures that $e^{\\theta_t}$ is a true martingale rather than a local martingale, and optional stopping then delivers $E[e^{\\theta_\\tau}|x_0]=1$ for any bounded stopping time. The same construction extends to underdamped dynamics by working with the phase-space Fokker-Planck equation and adding space inversion to the backward process.","core_discovery":"The central claim is the conditional identity $E[e^{\\theta_\\tau}|x_0]=1$, valid for overdamped and underdamped Langevin dynamics, for any bounded stopping time $\\tau$, and for any choice of backward density $\\psi$ that solves the backward Fokker-Planck equation with an arbitrary initial condition. The exponent is $\\theta_t = -\\beta W_t + \\beta[H(x_t,t)-H(x_0,0)] + \\beta \\ln(\\psi(x_t,t)/\\psi(x_0,0))$, so it combines negative dimensionless work, dimensionless energy change, and the log-density ratio of the backward process along the forward trajectory. Because the equality holds for every starting $x_0$, it is strictly stronger than the usual averaged version in which the work theorem is integrated over the initial distribution, and it remains meaningful when the initial distribution is a delta function. Choosing $\\psi$ to be the stationary distribution of the forward dynamics yields a simplified stationary work theorem; choosing $\\psi$ to be the time-reversed forward density reproduces the earlier ensemble result. The derivation also shows that if the diffusion constant violates the fluctuation-dissipation relation $D = 1/(\\beta\\gamma)$, the exponential acquires a nonzero drift and the theorem fails, with the deviation expressed by an explicit formula.","pith_inferences":["Because $\\psi$ is arbitrary, the family of identities can be used as an importance-sampling tool: choosing $\\psi$ to reduce estimator variance would let one estimate free-energy differences from fewer trajectories than the standard choice requires.","The conditional form suggests a per-initial-state fluctuation theorem that may sharpen single-molecule measurements: binning repeated trajectories by their initial configuration should yield $E[e^{\\theta_\\tau}|x_0]=1$ in every bin, a testable prediction that the ensemble-averaged version cannot make.","The violation formula for $D \\neq 1/(\\beta\\gamma)$ offers a model-free probe of active or non-thermal systems: measuring $E[e^{\\theta_t}]$ as a function of time directly estimates how strongly the effective diffusion deviates from the fluctuation-dissipation relation, independent of any specific microscopic model.","The authors' observation that faster erasure costs more work suggests a concrete next step: apply the stationary-choice identity with a stopping time tied to erasure completion to obtain a finite-time bound on the erasure work, extending the classical $k_B T\\ln 2$ limit."],"forward_implications":["For any bounded stopping time $\\tau$ and any initial state $x_0$, $E[e^{\\theta_\\tau}|x_0]=1$, so work equalities need not be averaged over the initial ensemble; they apply to experiments that prepare a single initial state, including delta-peaked initial conditions where trajectory entropy is undefined.","Choosing $\\psi$ as the stationary distribution of the forward dynamics turns the identity into $E[e^{-\\beta(W_\\tau - \\Delta H_\\tau + T\\Delta S^*_\\tau)}]=1$, which requires solving only a time-independent problem and therefore has lower computational cost than solving the time-dependent Fokker-Planck equation.","The same martingale construction works for underdamped Langevin dynamics with phase-space variables, including the deterministic zero-friction limit, where the information-entropy production vanishes along trajectories.","When the fluctuation-dissipation relation fails ($D \\neq 1/(\\beta\\gamma)$), the exponential develops a nonzero drift and the generalized work theorem is violated; the explicit drift formula gives a quantitative measure of how far a process is from the equilibrium framework."],"supporting_citations":[{"why":"Provides the forward/reverse trajectory weight ratio that the new derivation generalizes and replaces with an arbitrary backward density.","marker":"[4]"},{"why":"Defines the trajectory-dependent entropy $S=-k_B\\ln\\rho$ that the backward-process entropy $\\Sigma$ generalizes.","marker":"[6]"},{"why":"Gives the path-integral formulation whose forward-backward coupling the alternative derivation avoids.","marker":"[8]"},{"why":"Supplies the ensemble-averaged martingale work theorem that the paper strengthens by conditioning on $x_0$ and generalizing the backward process.","marker":"[9]"},{"why":"Describes the martingale methods for stochastic thermodynamics that the proof mirrors and the stationary-choice identity extends.","marker":"[22]"},{"why":"Provides the stochastic-calculus results, including the square-exponential condition and optional stopping, that make $e^{\\theta_t}$ a true martingale.","marker":"[23]"},{"why":"Underlies the reverse-time diffusion equation that defines the backward Fokker-Planck equation used for $\\psi$.","marker":"[24]"}],"fun_headline_variants":["Work theorem now holds for each starting state, not just averaged","Exponential of negative work is a martingale for any initial state","Decoupled forward and backward paths yield stronger work theorem","Generalized work theorem: conditional equality for each starting state","Beyond ensemble: work theorem holds for every initial state"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the chosen backward density $\\psi$ stays positive, smooth, and satisfies the square-exponential condition up to the stopping time, so that $e^{\\theta_t}$ is a genuine martingale rather than a local martingale; if $\\psi$ acquires zeros or singularities (as in the Gaussian example of Eq. (50) at time $t_2$), or the drift grows too fast, the identity $E[e^{\\theta_\\tau}|x_0]=1$ can break down.","fun_headline_variants_meta":{"raw":{"variants":["Work theorem now holds for each starting state, not just averaged","Exponential of negative work is a martingale for any initial state","Decoupled forward and backward paths yield stronger work theorem","Generalized work theorem: conditional equality for each starting state","Beyond ensemble: work theorem holds for every initial state"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000456,"raw_usage":{"total_tokens":2265,"prompt_tokens":894,"completion_tokens":1371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1289}},"tokens_in":510,"tokens_out":1371,"duration_ms":31473,"temperature":1.0,"reasoning_tokens":1289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:46:38.116989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the linear overdamped process in Eq. (51) with $k=\\gamma=1$, $\\beta=1$, and $D=1/(\\beta\\gamma)=1$, starting from a Gaussian distribution; take $\\psi$ as the backward evolution of any Gaussian initial condition under the backward Fokker-Planck equation, stopped at a time $t_1$ earlier than any singularity of $\\psi$, and estimate $E[e^{\\theta_{t_1}}|x_0]$ from $10^7$ simulated trajectories binned by initial value. The paper's Eq. (24) predicts every bin equals 1; a bin whose estimate deviates by several standard errors would falsify the conditional work theorem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the forward/reverse trajectory weight ratio that the new derivation generalizes and replaces with an arbitrary backward density."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the trajectory-dependent entropy $S=-k_B\\ln\\rho$ that the backward-process entropy $\\Sigma$ generalizes."},{"cited_title":"Fluctuation theorem for stochastic dynamics","cited_arxiv_id":null,"evidence_quote":"Gives the path-integral formulation whose forward-backward coupling the alternative derivation avoids."},{"cited_title":"Pekola, and ´Edgar Rold´ an","cited_arxiv_id":null,"evidence_quote":"Supplies the ensemble-averaged martingale work theorem that the paper strengthens by conditioning on $x_0$ and generalizing the backward process."},{"cited_title":"Martingales for physicists: a treatise on stochastic the rmodynamics and beyond","cited_arxiv_id":null,"evidence_quote":"Describes the martingale methods for stochastic thermodynamics that the proof mirrors and the stationary-choice identity extends."},{"cited_title":"Brownian motion and stochastic calculus , volume 113","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic-calculus results, including the square-exponential condition and optional stopping, that make $e^{\\theta_t}$ a true martingale."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Underlies the reverse-time diffusion equation that defines the backward Fokker-Planck equation used for $\\psi$."}],"review_version":1}