{"id":"c8550f66-4180-47c9-b97c-774e47064d09","arxiv_id":"1908.10484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Gaussian-process regression with a physics-derived likelihood infers effective forces from Langevin time traces without binning, including credible intervals and joint inference of friction.","lead":"This paper fits a Gaussian process to time traces of a particle moving under an overdamped Langevin equation, inferring the effective force field without binning the data. The method supplies point estimates and credible intervals and can also learn the friction coefficient, with demonstrations on synthetic harmonic and multi-well potentials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Euler-discretized noiseless likelihood is the pivotal assumption; with observation noise or finite τ the inferred force and credible intervals are biased, and the paper does not test this breakdown.","rationale":"I read the paper as a proof-of-principle for a bin-free, GP-based effective-force estimator. The final predictive equations (12-13) are standard GP regression with a physics-derived noise variance and are likely correct, despite the flawed intermediate derivation at Eq. 10, which appears to confuse the posterior with a signal-plus-noise distribution. Because the derivation is wrong but the result is standard, the derivation issue is not the main threat to the central claim; a careful reader can re-derive the correct formulas. The more serious threat is the likelihood's dependence on an Euler-discretized, noiseless model. The paper's synthetic demonstrations appear to use exactly that model, so they cannot detect misspecification. Real experimental traces contain localization noise that violates the independent-increment assumption, and finite sampling intervals introduce Euler bias when the force varies on the diffusive step scale. The abstract and introduction claim broad applicability, so this untested assumption is load-bearing. The reader's weakest assumption already identified this, and the conditional verdict is appropriate: the paper should be accepted only if the authors add robustness tests against these violations. I therefore find no reason to change the vote. I also note the paper would benefit from code release and a comparison with the bin-free methods cited as refs. 23 and 37, but these affect reproducibility and novelty rather than the core correctness argument.","tokens_in":14109,"tokens_out":10990,"duration_ms":120291,"concrete_test":"Simulate a ground-truth overdamped Langevin process with a fine integration step dt << τ (e.g., dt = τ/100), subsample every τ-th position, and add independent Gaussian localization noise of standard deviation σ_loc comparable to or larger than the thermal step sqrt(2kTτ/ζ). Run the proposed GP inference using the Euler likelihood with the sampling interval τ, then compute the integrated error and the empirical coverage of the credible intervals against the true force. Repeat for increasing σ_loc and for τ values spanning the force-variation length scale. If the error grows substantially or coverage drops far below nominal, the method's applicability to noisy or coarsely sampled experimental data is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central predictive formulas, Eqs. 12-13, are only valid if the conditional likelihood Eq. 4 exactly describes the data: conditionally independent Gaussian increments with mean (τ/ζ) f(x_n) and variance 2ζkT/τ. This is the forward-Euler approximation to an overdamped Langevin equation with no observation error. The paper explicitly assumes noiseless positions and constant τ, but it never tests the method beyond synthetic data generated from this same discretized model. In experimental single-molecule traces, localization noise enters the measured positions; because each position appears in two successive increments, the increment noise becomes correlated and the diagonal likelihood is misspecified. For finite τ, the true transition density of the continuous-time process deviates from the Gaussian Euler form when the force varies appreciably over a diffusive step, biasing the posterior mean and making credible intervals overconfident. Since the abstract claims applicability to experimental and in silico molecular dynamics traces, this untested misspecification is the most load-bearing threat to the central claim. If the likelihood is wrong, the GP posterior is a Bayesian fit to the wrong model, and the inferred 'effective force' need not equal the true force. The reader's weakest assumption correctly identifies this gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian nonparametric method for inferring an effective force f(x) from a time trace of positions generated by an overdamped Langevin dynamics. The force is assigned a zero-mean Gaussian process prior with a squared-exponential kernel, and the Euler-discretized dynamics provide a Gaussian likelihood for the increments. Combining prior and likelihood yields a Gaussian posterior for the force at arbitrary test points, with predictive mean and covariance given in Eqs. (12)-(13). The friction coefficient ζ is inferred jointly by a Gibbs sampler that alternately updates the force and ζ. The method is demonstrated on synthetic harmonic, multi-well, and quartic potentials, and is compared to the binning method of Masson et al. and to residence-time analysis, with the claim that it converges with fewer data points, provides full credible intervals, and avoids binning artifacts.","tokens_in":14381,"tokens_out":9863,"duration_ms":98706,"significance":"If the claims hold, the method is a useful addition to the analysis toolbox for single-molecule and molecular-dynamics trajectories, removing binning artifacts and providing pointwise uncertainties. The final predictive formulas are the standard Gaussian-process regression results, and the synthetic demonstrations are plausible. However, the derivation as written is incorrect, and the method is validated only on data generated from the same discretized model used in the likelihood; the advertised experimental applicability is therefore not yet supported. With corrected derivations and additional robustness experiments, the paper could be a solid methods contribution.","major_comments":[{"comment":"The derivation of the posterior predictive distribution is not valid. Eq. (9) is not an equivalent rewriting of Eq. (5): the likelihood P(x2:N|f(·),ζ,x1) does not involve the test-point values f*, and inserting f* with a degenerate variance εI does not create a likelihood for f*. Taking ε→0 forces f* to zero in the joint Gaussian, rather than producing a predictive distribution. Moreover, the marginal of the Gaussian in Eq. (10) over f* is N(0,K**), not Eq. (11). The correct route is to treat y_n = ζv_n as noisy observations of f(x_n) with noise variance 2ζkT/τ; the standard conditional Gaussian formulas then give exactly Eqs. (12)-(13) as the posterior mean and covariance of f* given the data. Please replace the erroneous epsilon-limit and marginalization argument with this (or an equivalent correct) derivation.","section":"GP prior and posterior for force (Eqs. 8-13)"},{"comment":"The method is tested only on synthetic traces generated by the same Euler-discretized forward model that defines the likelihood, Eqs. (2)-(4). Since the Abstract and Introduction claim applicability to experimental traces, the manuscript should analyze or at least quantify the two main misspecifications: (i) localization/observation noise, under which consecutive increments become correlated and the diagonal likelihood is misspecified; and (ii) finite step size τ, where the Euler transition density deviates from the true transition density when the force varies appreciably over a diffusive step. Please add numerical experiments that measure bias and credible-interval coverage under these misspecifications, or explicitly restrict the experimental-applicability claims to the noiseless, small-τ regime and state this limitation in the abstract and conclusions.","section":"Results and Discussion (noiseless Euler model)"},{"comment":"The Gibbs sampler is described only schematically, which prevents reproduction and assessment of the simultaneous inference claim. The Metropolis-Hastings update for ζ is not specified (proposal distribution, step size, tuning), and the paper does not state the number of iterations, burn-in, thinning, or the criterion used to select the MAP pair from the sampled sequence. To support Fig. 6 and the reported value ζ = 101.03 pg/µs, please provide these details and include convergence diagnostics or acceptance rates.","section":"Gibbs sampling for friction coefficient"}],"minor_comments":[{"comment":"The notation for the hyperparameter sweep is confusing: 'σ = 10 in the figure corresponds to σ = 10ατ(vmax−vmin)' mixes the dimensionless figure label with the physical value. Please clarify the exact mapping used.","section":"SI, Hyperparameters; Fig. 9 caption"},{"comment":"In Eq. (26), the kernel is defined as K(x,x′) but the right-hand side uses subscripts i and j (xi and x*j). Please correct the index notation to use x and x′.","section":"SI, Eq. (26)"},{"comment":"The stiffness h = ζ/τ is labeled with units 'pg/µs' in Fig. 5, but from Eq. (15) h has units of pg/(µs·µs) if ζ is in pg/µs and τ in µs. Please correct the units in the figure and text.","section":"Fig. 5 and text on stiffness"},{"comment":"The statement 'After many iterations the sampled pairs converge to the true value' is imprecise: Gibbs sampling converges to the posterior distribution, not to a single true value; the MAP estimate is then extracted from the samples. Please rephrase.","section":"Gibbs sampling section"},{"comment":"Reference [37] (Frishman and Ronceray, 2018) is incomplete: it lacks a title and a journal, book, or arXiv identifier. Please complete the citation.","section":"Reference [37]"},{"comment":"The claim 'So far no method satisfies all three criteria' is strong and is supported only by comparison to two binning-based baselines. Please consider acknowledging other binning-free or GP-based approaches, or broaden the comparison to justify the claim.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward application of standard Gaussian-process regression to force inference from overdamped Langevin trajectories. The main technical weakness is the incorrect derivation of the posterior predictive; although the final formulas are standard and fixable, the current text would mislead readers. The untested misspecification of the noiseless Euler model is a concern given the paper's stated experimental scope. I would recommend major revision, with the expectation that the authors can correct the derivation and add robustness experiments or carefully limit the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the central idea is simple and correct: under forward-Euler overdamped Langevin dynamics, each measured increment (x_{n+1}-x_n) is a noisy observation of the force at x_n, with noise variance 2ζkT/τ. So the GP posterior predictive is standard GP regression with a physics-derived noise model. Second, the paper has a real derivation typo in the epsilon limit that is confusing but does not invalidate the final formulas.\n\nWhat's actually new: applying Gaussian processes to force inference from single-particle or MD time traces, with a Gibbs sampler that draws both the force profile and the friction coefficient. The novelty is modest—this is a direct application of GP regression—but the packaging is useful. The synthetic tests are sensible: harmonic and multi-well potentials, convergence with N, and comparisons to Masson et al.'s binning and residence-time methods. The comparisons show the expected advantage of bin-free inference, especially in undersampled regions.\n\nSoft spots, in proportion. The derivation of Eq. (10) as printed is internally inconsistent: adding εI to the test-point blocks of the augmented likelihood and taking ε→0 would pin f* to zero, not produce the stated posterior. The correct limit is zero precision on the test points (equivalently ε→∞ in their notation), and then the standard GP formulas follow. So the final expressions are right, but the presentation needs fixing.\n\nThe larger assumption is the Euler-discretized, noiseless likelihood. The paper explicitly states that it does not include observation noise and proposes an HMM extension in the Discussion, but it never tests how the method degrades when localization noise is comparable to thermal kicks or when τ is too large for Euler-Maruyama. For real single-molecule data this matters. It does not sink the paper, but it should be flagged more prominently and tested.\n\nTwo smaller issues: hyperparameters are hand-set (and in the multi-well example, the length scale was chosen with knowledge of the target's fine detail), and no code or data are provided, which limits reproducibility. The citation pattern is mostly fair—Masson et al. and residence time are the right baselines—though ref 37 is cited only for an analogy and never compared, so the 'first time' claim is not fully checked against it.\n\nVerdict: the central claim—bin-free force inference with credible intervals from a Langevin trace—holds up under the stated assumptions. The paper is a solid methods contribution for single-molecule and coarse-grained MD analysis, not a conceptual breakthrough. A serious referee should look at it; I'd accept it into review and ask for a corrected derivation, a robustness check on observation noise, and preferably code. For a reading group, it's a decent example of GP inference in physics, but not a must-read.","headline":"Clean, correct GP mapping for Langevin force inference; the derivation typo and untested observation-noise assumption are addressable, not fatal.","tokens_in":14905,"tokens_out":5943,"would_cite":true,"duration_ms":53855,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that a zero-mean Gaussian process prior on the force in an overdamped Langevin model yields a closed-form Gaussian posterior predictive at test points, enabling continuous effective-force inference with credible…","keywords":["Gaussian process","Langevin dynamics","effective force","Bayesian nonparametrics","Gibbs sampling","time trace analysis","friction coefficient","single-molecule"],"falsifier":"Generate a long synthetic trajectory from the same overdamped Langevin equation with a known force, then add Gaussian observation noise of increasing variance to the recorded positions before running the method; if the inferred effective force deviates systematically from the ground truth as the noise grows, the noiseless-assumption breakdown is demonstrated. Alternatively, hold the physical parameters fixed and increase the time step $\\tau$ until the Euler-Maruyama scheme is no longer accurate; a growing bias in the recovered force would falsify the method's claims for coarse sampling.","tokens_in":13904,"feed_emoji":"⚛️","tokens_out":5890,"duration_ms":52829,"temperature":0.7,"pith_summary":"The paper proposes a Bayesian method to infer the effective force acting on a particle from a single time trace of its position, without binning or other preprocessing. The key claim is that a zero-mean Gaussian process prior on the force, combined with the Euler-discretized overdamped Langevin likelihood, makes the posterior predictive at arbitrary test points exactly Gaussian, with means and covariances given by closed-form matrix expressions. A Gibbs sampler alternates between sampling the force and the friction coefficient, so both can be learned simultaneously from the same data. If correct, the method delivers continuous force profiles with full credible intervals even in undersampled regions, which existing histogram-based and binned Bayesian methods cannot do.","feed_headline":"No binning: GP inference yields continuous force curves","feed_subtitle":"Closed-form Gaussian updates deliver credible intervals and joint friction inference from raw positions.","key_machinery":"The central object is the conjugate pair formed by a zero-mean Gaussian process prior on the force and the Gaussian likelihood coming from Euler-discretized overdamped Langevin dynamics. The GP kernel (squared exponential) and the thermal-noise covariance combine so that the posterior predictive at test points is Gaussian with the closed-form mean and covariance of Eqs. (12)–(13); a Gibbs sampler using a Gamma prior on the friction coefficient exploits the same conjugacy to sample $\\zeta$ and $f$ alternately.","core_discovery":"The central discovery is that the Gaussian prior is conjugate to the Gaussian thermal kicks in the overdamped Langevin equation, so after marginalizing the force values at the data points, the predictive distribution for the force at any chosen test points is a multivariate Gaussian. Its mean is $\\tilde{\\mu} = \\zeta K_*^T (K + (2\\zeta kT/\\tau) I)^{-1} v$ and its covariance is $\\tilde{K} = K_{**} - K_*^T (K + (2\\zeta kT/\\tau) I)^{-1} K_*$ (Eqs. 12–13). This turns force inference into a matrix-algebraic operation with no arbitrary discretization choice, and the same machinery supplies the conditional needed to Gibbs-sample the friction coefficient. The paper demonstrates on synthetic data from harmonic, quartic, and multi-well potentials that the method recovers the ground truth force with fewer data than the binned Masson et al. method or the residence-time histogram analysis do.","pith_inferences":["Because the likelihood assumes noiseless observations, the largest practical risk is measurement noise; explicitly modeling a Gaussian observation error by learning hidden positions, as the paper suggests, would extend the method to experimental single-particle tracking where localization error is significant.","The closed-form predictive should make the method amenable to sparse or inducing-point GP approximations, which would address the cubic scaling that currently limits the method to about 50,000 points.","The same conjugacy structure could be extended to memory kernels or position-dependent friction by choosing appropriate kernels or by Gibbs-sampling additional hyperparameters, though these go beyond what the paper tests.","A direct comparison of posterior predictive error against the true force on simulated data with added measurement noise would quantify the breakdown of the noiseless assumption; the paper does not report such a test."],"forward_implications":["For a single unbinned time trace of positions, the method yields a continuous effective force estimate with pointwise credible intervals, including in regions where data are sparse.","The accuracy of the inferred force improves as the number of data points grows, and the method converges to the ground truth force with fewer data points than the binned comparison methods tested.","When the friction coefficient is unknown, the Gibbs sampler still recovers both $\\zeta$ and the force, though the force estimate is less accurate than when $\\zeta$ is fixed.","The stiffness parameter $h = \\zeta/\\tau$ controls the signal-to-noise ratio; the method performs better for small stiffness (large thermal kicks relative to friction), consistent with the Langevin equation's structure.","The approach avoids binning artifacts such as the underestimation of forces in steep potential regions that arises when averaging within spatial bins."],"supporting_citations":[{"why":"Supplies the Gaussian process prior and posterior predictive formulas that become the closed-form mean and covariance of Eqs. (12)–(13).","marker":"26"},{"why":"Provides the overdamped Langevin dynamics equation that defines the model and the thermal-force statistics.","marker":"7"},{"why":"The finite-difference (Euler) scheme used to discretize the Langevin equation and build the likelihood in Eqs. (2)–(3).","marker":"31"},{"why":"The Gaussian conditioning derivation used to combine the prior and likelihood and arrive at the predictive distribution.","marker":"34"},{"why":"Provides the Gibbs sampling and Metropolis-Hastings framework used to sample the friction coefficient.","marker":"35"},{"why":"The binned Bayesian method used as the primary comparison baseline; the GP approach aims to remove its binning artifacts.","marker":"24"},{"why":"The residence-time histogram method used as the baseline for potential-energy comparison.","marker":"38"},{"why":"The signal-to-noise discussion that frames the stiffness dependence of the inference quality.","marker":"37"}],"fun_headline_variants":["Gaussian process force inference skips binning","Closed-form GP yields force curves with error bars","No binning: GP learns forces from raw trajectories","Bayesian GP extracts forces from Langevin data","Continuous force curves without binning via GP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inference assumes the recorded positions are exact samples of an overdamped Langevin process with a single constant friction coefficient, and that the Euler time step is small enough that the discretized likelihood is accurate; if measurement noise is sizable or the sampling interval is too coarse, the likelihood is misspecified and the inferred force and friction become biased.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian process force inference skips binning","Closed-form GP yields force curves with error bars","No binning: GP learns forces from raw trajectories","Bayesian GP extracts forces from Langevin data","Continuous force curves without binning via GP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1191,"prompt_tokens":886,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":234}},"tokens_in":502,"tokens_out":305,"duration_ms":3391,"temperature":1.0,"reasoning_tokens":234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:44:06.411283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long synthetic trajectory from the same overdamped Langevin equation with a known force, then add Gaussian observation noise of increasing variance to the recorded positions before running the method; if the inferred effective force deviates systematically from the ground truth as the noise grows, the noiseless-assumption breakdown is demonstrated. Alternatively, hold the physical parameters fixed and increase the time step $\\tau$ until the Euler-Maruyama scheme is no longer accurate; a growing bias in the recovered force would falsify the method's claims for coarse sampling.","supporting_citations":[{"cited_title":"E.; Williams, C","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process prior and posterior predictive formulas that become the closed-form mean and covariance of Eqs. (12)–(13)."},{"cited_title":"Nonequilibrium Statistical Mechanics; Oxford University Press, 2001","cited_arxiv_id":null,"evidence_quote":"Provides the overdamped Langevin dynamics equation that defines the model and the thermal-force statistics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The finite-difference (Euler) scheme used to discretize the Langevin equation and build the likelihood in Eqs. (2)–(3)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Gaussian conditioning derivation used to combine the prior and likelihood and arrive at the predictive distribution."},{"cited_title":"R.; Best, N","cited_arxiv_id":null,"evidence_quote":"Provides the Gibbs sampling and Metropolis-Hastings framework used to sample the friction coefficient."},{"cited_title":"B.; Casanova, D.; T \\\" u rkcan, S.; Voisinne, G.; Popoff, M","cited_arxiv_id":null,"evidence_quote":"The binned Bayesian method used as the primary comparison baseline; the GP approach aims to remove its binning artifacts."},{"cited_title":"L.; Pralle, A.; Stelzer, E","cited_arxiv_id":null,"evidence_quote":"The residence-time histogram method used as the baseline for potential-energy comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The signal-to-noise discussion that frames the stiffness dependence of the inference quality."}],"review_version":1}