{"id":"a098a56c-d7e8-43f3-885d-182eacfc5395","arxiv_id":"1908.01845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A specialized reweighting estimator for symmetric nonequilibrium protocols, Eq. 7, reduces bias and improves convergence over unidirectional and bidirectional estimators when only symmetric-protocol data are available.","lead":"This paper derives an estimator that combines each trajectory with its time-reversed twin to compute equilibrium averages from nonequilibrium simulations with symmetric pulling protocols. The estimator converges faster than existing approaches in tests on five model systems, though a standard bidirectional method is usually still better when available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 7's unbiasedness is unproven for the non-symplectic Brownian/Langevin integrators used in validation because Eq. 3's unit-Jacobian assumption is not established for stochastic path measures.","rationale":"The central estimator is a plausible consequence of Crooks' relation when the path-measure Jacobian is unity, and the paper is honest about the narrow regime in which the symmetric estimator is recommended. The strongest support is the clean derivation of Eq. 5 → Eq. 7 under the stated assumption, plus numerical demonstrations on five systems. However, the manuscript itself flags the Jacobian condition at Eq. 3 as requiring a symplectic integrator, and the validation uses non-symplectic stochastic integrators (Brownian dynamics and NAMD Langevin dynamics). This is the single most load-bearing weakness: the estimator is advertised as theoretically rigorous for the molecular systems where it is recommended, but the text only establishes that rigor for deterministic symplectic dynamics. The reader's weakest_assumption identified exactly this issue. I agree with that identification. The concern is not fatal because the numerical results are consistent with an unbiased estimator; if a Jacobian correction is needed, it may be small. But a conditional verdict is appropriate: the authors should either prove the Jacobian is unity for the stochastic integrators used (or, equivalently, show that Crooks' relation holds for the discretized path measure), or provide a large-N convergence test demonstrating unbiasedness. A secondary printed error is Eq. 9, which lacks the normalization denominator of the original MA estimator (Refs. 9, 10); as written it is not a valid estimator and should be corrected, but this does not bear on the derivation of Eq. 7 itself. The critique is technical and addressable, not a rejection of the paper's contribution.","tokens_in":13662,"tokens_out":13101,"duration_ms":146574,"concrete_test":"Re-run the symmetric 1D Brownian dynamics system (§III.A) with a very large number of trajectories (≥10^6) and apply Eq. 7 to estimate ΔF_t and the PMF using F=e^{-βW(t)}. Compare with the exact numerical references already used in the paper. If the estimator is unbiased, errors should decay as 1/√N and extrapolate to zero; a residual bias (for example, a systematic offset in the right half of the potential) would indicate J≠1 for the Euler–Maruyama integrator. To isolate the cause, also compute the ratio of forward to time-reversed path probabilities directly from the transition kernels of the same integrator and compare with e^{βW}; this analytically settles whether Eq. 5 holds for the discretized stochastic dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central derivation passes through Eq. 3, where a copy of the path integral is transformed under time reversal. The paper states that this substitution requires the Jacobian of Γ→Γ̃ to be unity, which it ties to a symplectic integrator and identical initial and final states. The validation systems, however, use Brownian dynamics (described as an in-house script in §III.B) and NAMD Langevin dynamics, neither of which is symplectic. No argument or numerical check is given that the discretized stochastic path measure is invariant under time reversal, or that Crooks' relation (Eq. 4) holds exactly for these integrators. If the Jacobian J is not unity, Eq. 6 should read [F(Γ)+J F(Γ̃)e^{-βW}]/[1+J e^{-βW}], and Eq. 7 is biased. Because the paper recommends the estimator for molecular simulations with Langevin dynamics, this gap directly affects the central claim of a theoretically rigorous estimator for the target applications. The reported RMSE improvements are suggestive but cannot certify unbiasedness; RMSE comparisons can mask persistent bias, and the bidirectional baseline printed in Eq. 9 is itself missing the normalization denominator of the original MA estimator, which would disadvantage that baseline as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives a new statistical estimator for nonequilibrium path-ensemble averages in systems driven by symmetric protocols, i.e., protocols satisfying λ(t)=λ(τ−t). The estimator reweights each sampled trajectory by adding its time-reversed conjugate twin weighted by exp(−βW), and it is derived from the Crooks fluctuation theorem. The authors compare this estimator with unidirectional and bidirectional (Minh–Adib) estimators on five model systems: a symmetric 1D potential, an asymmetric 1D potential, deca-alanine, a CB7–hexafluorobenzene host–guest complex, and a gramicidin A ion channel. They report that the symmetric estimator is the best among estimators applied to symmetric-protocol data, but that the bidirectional estimator applied to asymmetric forward/reverse pairs usually outperforms it for the same total simulation time. The paper concludes that the symmetric estimator is recommended when bidirectional estimation is infeasible or expected to perform poorly.","tokens_in":13846,"tokens_out":9106,"duration_ms":85446,"significance":"If the estimator is unbiased under the applied dynamics, it has clear practical value for free-energy calculations and single-molecule pulling experiments where symmetric protocols can be implemented without additional equilibration. The paper's strengths include a clean derivation from the Crooks relation with no fitted parameters, a broad set of validation systems ranging from 1D potentials to a membrane ion channel, and independent reference free energies from numerical integration or umbrella sampling/WHAM. However, the central theoretical claim is conditional on a unit-Jacobian assumption that is not justified for the stochastic integrators used in the validation, and the printed bidirectional baseline formula is not normalized. These issues are addressable and do not invalidate the underlying idea if the derivation is properly extended and the presentation corrected.","major_comments":[{"comment":"The estimator is derived under the assumption that forward and reverse processes are identical, i.e., λ(t)=λ(τ−t). However, the text describes the symmetric process for the symmetric 1D potential as a single pull from λ=−1.5 to λ=1.5 in 750 steps, and the gramicidin A simulations as forward pulls from z=−1.3 nm to z=1.3 nm with separate reverse simulations. Neither of these described protocols satisfies λ(t)=λ(τ−t). If the actual protocols were round-trip (e.g., pull and return), the text should state this explicitly; if they were one-way protocols on a spatially symmetric system, then Eq. 5 (ρF[Γ]=ρR[Γ]) is not valid for the same trajectory Γ, and a separate derivation incorporating the spatial symmetry is required before Eq. 7 can be applied to these systems. This is a load-bearing issue for the empirical validation of the central claim.","section":"Section II.A and Section III.B/III.D"},{"comment":"The substitution in Eq. 3 assumes that the Jacobian of the transformation Γ→Γ̃ is unity, which the text ties to a symplectic integrator and identical initial and final states. The validation simulations in Section III, however, use Brownian dynamics (in-house script for 1D systems) and Langevin dynamics (NAMD for deca-alanine, host–guest, and gramicidin A), neither of which is symplectic in the sense of deterministic Hamiltonian flow. For these stochastic path measures, the time-reversal map need not preserve the discretized path measure, and the Crooks relation (Eq. 4) may not hold exactly for the discrete integrator. If the Jacobian deviates from unity, Eq. 6 would acquire a Jacobian factor and Eq. 7 would be biased. The authors should either prove the unit-Jacobian property for the specific stochastic integrators, numerically test the unbiasedness of Eq. 7 against the available reference free energies in a way that separates bias from variance, or restrict the claim of rigorous unbiasedness to symplectic dynamics.","section":"Section II.A, Eq. 3"},{"comment":"The printed bidirectional (Minh–Adib) estimator in Eq. 9 is not normalized: the sums in the numerator are not divided by the total weight in a denominator. As written, Eq. 9 does not estimate ⟨F⟩ even asymptotically, and using it as a baseline in Section IV would disadvantage the MA estimator. Please provide the correct normalized expression (which should include a denominator analogous to the one in Eq. 7) and confirm that the simulations in Section IV used the normalized version.","section":"Section II.A, Eq. 9"}],"minor_comments":[{"comment":"The sentence \"For one of the integrals in each of sums\" contains a grammatical error; it should be \"in each of the sums.\"","section":"Section II.A"},{"comment":"The caption labels panel (b) as \"symmetric 1D system,\" but the text describes it as the asymmetric 1D potential; this should be corrected.","section":"Figure 1 caption"},{"comment":"The phrase \"at at 300 K\" contains a duplicated word; it should be \"at 300 K.\"","section":"Section III.B"},{"comment":"The RMSE comparisons are reported without statistical error bars on the RMSE itself. Since block averaging is used, the authors could report the standard error of the RMSE across blocks or another measure of uncertainty, so that the reader can judge whether differences between estimators are significant.","section":"Section IV"},{"comment":"The claim that \"the symmetric estimator has similar performance to a unidirectional protocol of half the length and twice the number of trajectories\" is stated in the abstract but not explicitly quantified in the results; please indicate which figures or tables support this conclusion.","section":"Abstract and Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising estimator with a sound formal derivation for symplectic dynamics, but the manuscript needs careful revision to address the stochastic-integrator justification, to clarify the exact protocols used in the symmetric 1D and gramicidin A applications, and to correct the printed bidirectional estimator formula. The failure to clearly describe the symmetric protocols in two of the five validation systems is particularly important, because if the protocols were not actually symmetric, the empirical demonstration would not test the theory. I recommend major revision rather than rejection, since these issues appear fixable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe new thing here is Eq. 7: an estimator for path-ensemble averages under symmetric nonequilibrium protocols, where each forward trajectory is combined with its time-reversed twin weighted by e^{-βW}. It is a specialization of the Minh-Adib bidirectional estimator to the case where forward and reverse protocols coincide and ΔF=0. The mathematical step is small—a direct application of Crooks relation plus the symmetry condition—but the practical value is real. The paper validates it on five systems, and the RMSE comparisons support the claim that it converges faster and with lower error than unidirectional or bidirectional estimators when data come from symmetric protocols.\n\nWhat I like is that the authors do not oversell. They explicitly say that with asymmetric forward/reverse pairs the old bidirectional estimator usually wins, and position the symmetric estimator for cases where bidirectional is infeasible. The empirical scope is broad: two 1D models, deca-alanine, a host-guest complex, and gramicidin A. The comparisons are honest, including regimes where the symmetric estimator does not help.\n\nSoft spots. The derivation of Eq. 3 invokes a symplectic-integrator assumption for the Jacobian of Γ→Γ̃, but the validation uses Brownian and Langevin dynamics. I think the Jacobian point is largely a red herring—time reversal is a coordinate relabeling with determinant ±1—but the real gap is that they never prove Crooks relation holds exactly for the discretized stochastic integrators. The empirical agreement with reference values is reassuring but does not certify unbiasedness. Eq. 9, the printed MA baseline, is also missing the normalization denominator; likely a typo, but as written it is not the correct estimator. Both are fixable.\n\nCitation pattern is fine: they build on Crooks and the MA estimator, and the gA data reuse is cited. No fitted parameters enter the derivation; reference free energies come from independent numerical integration or umbrella sampling. I'd send this to peer review and would cite it if I worked in steered molecular dynamics or single-molecule pulling. It's a good reading-group piece for the free-energy methods crowd.\n\nIn short: useful, honest, and mostly rigorous. The rigor gaps are at the edges, not in the center.","headline":"A clean, honestly-scoped specialization of the MA estimator for symmetric pulling protocols; small math, solid empirical validation, and a couple of rigor gaps that are fixable.","tokens_in":14395,"tokens_out":7804,"would_cite":true,"duration_ms":78910,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["87.15.kp","05.10.-a"],"model":"deepseek-v4-flash","headline":"Time-reversed twin trajectories give a faster, lower-error estimator for symmetric nonequilibrium protocols.","keywords":["nonequilibrium processes","path-ensemble averages","symmetric protocols","free energy estimation","potential of mean force","bidirectional estimator","conjugate twin trajectories","nonequilibrium work relations"],"falsifier":"Run symmetric-protocol simulations and compare the Eq. (7) estimate against a very long unidirectional estimate or an exact reference; a bias that grows with work variance in the stochastic systems would indicate the Jacobian-unity assumption is violated. More directly, estimate the ratio of forward-path to twin-path probabilities from sampled twins and check whether it equals $e^{\\beta W}$ across the work range.","tokens_in":13434,"feed_emoji":"⚛️","tokens_out":7735,"duration_ms":66901,"temperature":0.7,"pith_summary":"The paper develops a statistically rigorous estimator for nonequilibrium path-ensemble averages that is specialized to symmetric protocols—processes whose forward and reverse driving are the same, so $\\lambda(t)=\\lambda(\\tau-t)$. The idea is that each sampled trajectory brings with it its conjugate twin, the time-reversed path, and the two are combined with a weight $e^{-\\beta W}$ derived from the forward–reverse path-density relation. Because the initial and final thermodynamic states coincide, the free energy difference is zero and no separate free energy estimate is needed to reweight the twin. In tests on five model systems, the estimator converges faster and with smaller root-mean-square error than unidirectional and bidirectional estimators when only symmetric-protocol data are used. The practical payoff is that the return leg of a pulling experiment or simulation is not wasted: symmetric data analyzed this way match the accuracy of twice as many forward-only trajectories of half length.","feed_headline":"Pairing trajectories with time-reversed twins lowers free-energy error","feed_subtitle":"Symmetric-protocol data converge faster and with less error when each trajectory is paired with its reverse twin.","key_machinery":"The load-bearing object is the conjugate-twin reweighting identity for symmetric protocols, $\\rho_F[\\Gamma]/\\rho_F[\\tilde\\Gamma]=e^{\\beta W}$, derived from the general forward/reverse path-density relation with $\\Delta F=0$. It is inserted into a doubled path average, Eq. (3), whose Jacobian is taken to be unity because the integrator is symplectic and the initial and final states coincide. This produces a ratio estimator in which the twin contribution is weighted by $e^{-\\beta W[\\Gamma]}$; the denominator $\\sum (1+e^{-\\beta W})$ renormalizes automatically. The same identity underlies all of the paper's applications: free energies come from setting $F[\\Gamma]=e^{-\\beta W[\\Gamma(t)]}$ and potentials of mean force come from setting $F[\\Gamma]=\\delta[z-z[\\Gamma(t)]] e^{-\\beta W[\\Gamma(t)]}$.","core_discovery":"On the paper's own terms, the central result is the estimator of Eq. (7): for $N$ sampled trajectories $\\Gamma_n$ with work $W[\\Gamma_n]$ and conjugate twins $\\tilde\\Gamma_n$, an arbitrary path-ensemble average is estimated by $$\\hat F = \\frac{\\sum_{n=1}^N (F[\\Gamma_n] + F[\\tilde\\Gamma_n] $e^{{-\\beta W[\\Gamma_n]}}$)}{\\sum_{n=1}^N (1 + $e^{{-\\beta W[\\Gamma_n]}}$)}.$$ The step that makes this valid is Eq. (5), the symmetric-protocol specialization of the forward–reverse path-density relation: $\\rho_F[\\Gamma] / \\rho_F[\\tilde\\Gamma] = e^{\\beta W}$, which follows because forward and reverse processes coincide and $\\Delta F=0$. Substituting this into a doubled form of the path average replaces the unidirectional average with a ratio in which every trajectory and its twin contribute, and the twin's contribution is automatically suppressed when the work is large. The paper claims this estimator is asymptotically unbiased for symmetric protocols and demonstrates on 1D potentials, deca-alanine, a host–guest complex, and gramicidin A that it gives lower RMSE than unidirectional and bidirectional estimates computed from the same symmetric-protocol data. It also reports the scope condition: when data from a forward/reverse pair of asymmetric protocols are available, the bidirectional estimator usually performs better, so the symmetric estimator is recommended where bidirectional estimation is infeasible or expected to fail.","pith_inferences":["Inference: the same ratio construction should transfer to any dynamics for which the twin map is measure-preserving; a natural test is to check whether the Jacobian is unity for a stochastic integrator with the paper's time-reversal convention, which the paper leaves open.","Inference: one could turn the denominator of Eq. (7) into a convergence diagnostic, since it estimates $\\langle 1+e^{-\\beta W}\\rangle$ and thereby monitors whether the sampled trajectories cover the rare, low-work events that dominate the average.","Inference: for membrane-permeation studies, where equilibrating a penetrant in the bilayer interior is difficult, a symmetric pull across the membrane plus Eq. (7) is a directly testable application that the paper suggests but does not implement."],"forward_implications":["For any symmetric pulling protocol, the new estimator is the best rigorous choice among the estimators compared: it lowers RMSE for both free energies and potentials of mean force relative to unidirectional and bidirectional analysis of the same trajectories.","Symmetric data analyzed with Eq. (7) are as accurate as a unidirectional protocol of half the length with twice as many trajectories, so the return leg of a single-molecule pulling or steered simulation contributes real statistical value.","When a forward/reverse pair of asymmetric protocols with equilibration at both ends is available, the bidirectional estimator usually still wins; the symmetric estimator is for cases where that setup is not feasible or one direction is far from equilibrium.","Because $\\Delta F=0$ for symmetric protocols, the estimator does not need a free energy estimate to compute dissipated work, simplifying the reweighting compared with the bidirectional estimator.","In slow pulling of gramicidin A, the symmetric estimator reconstructs the PMF barrier shape more accurately than bidirectional-plus-histogram analysis, although faster pulling favors the bidirectional approach."],"supporting_citations":[{"why":"Supplies the nonequilibrium work equality used to turn work into free energies and path-average functionals.","marker":"[1,2]"},{"why":"Supplies the forward/reverse path-density relation from which the symmetric-protocol twin weight $e^{\\beta W}$ is derived.","marker":"[3,29]"},{"why":"Supplies the time-dependent density and PMF reconstruction formulas used in the applications.","marker":"[4,5]"},{"why":"Defines the bidirectional estimator that serves as the main comparison baseline.","marker":"[9,10]"},{"why":"Provides the insight that conjugate twin trajectories dominate work-weighted path ensembles, motivating the symmetric-pairing construction.","marker":"[20]"},{"why":"Supplies the gramicidin A pulling data and umbrella-sampling reference PMFs used for validation.","marker":"[24]"}],"fun_headline_variants":["Symmetric-protocol estimator lowers free-energy error","Time-reversed twins improve symmetric-protocol estimates","New estimator pairs twins to cut free-energy error","Faster convergence, lower error for symmetric protocols"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimator is unbiased only if the transformation from a trajectory to its time-reversed twin preserves the path measure—that is, the Jacobian is unity—which the derivation ties to a symplectic integrator with identical initial and final states, while the validation simulations use stochastic dynamics where this measure preservation is asserted rather than verified.","fun_headline_variants_meta":{"raw":{"variants":["Symmetric-protocol estimator lowers free-energy error","Time-reversed twins improve symmetric-protocol estimates","New estimator pairs twins to cut free-energy error","Faster convergence, lower error for symmetric protocols"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001085,"raw_usage":{"total_tokens":4617,"prompt_tokens":1110,"completion_tokens":3507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":3448}},"tokens_in":726,"tokens_out":3507,"duration_ms":25320,"temperature":1.0,"reasoning_tokens":3448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:02:18.522132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run symmetric-protocol simulations and compare the Eq. (7) estimate against a very long unidirectional estimate or an exact reference; a bias that grows with work variance in the stochastic systems would indicate the Jacobian-unity assumption is violated. More directly, estimate the ratio of forward-path to twin-path probabilities from sampled twins and check whether it equals $e^{\\beta W}$ across the work range.","supporting_citations":[{"cited_title":"Jarzynski ,\\ @noop journal journal Phys","cited_arxiv_id":null,"evidence_quote":"Provides the insight that conjugate twin trajectories dominate work-weighted path ensembles, motivating the symmetric-pairing construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gramicidin A pulling data and umbrella-sampling reference PMFs used for validation."}],"review_version":1}