{"id":"794c7175-89bf-4f12-a0b8-9a4e0c755760","arxiv_id":"2412.19772","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Direct estimates of trajectory irreversibility can be made by extrapolating plug-in D_KL estimates to infinite sample size, and single retinal neurons show nonzero irreversibility over 200 ms windows.","lead":"This paper introduces a way to measure the arrow of time directly from recorded time series, without fitting a model, by correcting the bias that small samples create. It applies the method to retinal neurons and finds that single neurons carry detectable signatures of irreversibility, implying their activity is not Markovian.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bias expansion assumes independent word samples; retina data are correlated and the shuffle control does not validate the extrapolation, so the nonzero single-neuron DKL may be an artifact of an invalid 1/N correction.","rationale":"The reader's conditional verdict is well-founded. The paper's strongest assets are clean synthetic demonstrations on Markov systems and a controlled shuffle test, but the synthetic data do not probe the non-Markovian, correlated regime that motivates the retina application. The bias correction is the crux of the central claim; without verifying it on correlated time-reversible data, the single-neuron result cannot be distinguished from a sampling artifact. The missing zero-count rule is a further reproducibility gap but is secondary to the independence assumption. A simulation with a known time-reversible non-Markovian process is the decisive experiment: if the method returns zero there, the retina claim gains support; if not, the verdict should move toward rejection. Since that test has not been done, the conditional verdict is appropriate and no change is needed.","tokens_in":7803,"tokens_out":7862,"duration_ms":74852,"concrete_test":"Generate a long stationary time series from a time-reversible, non-Markovian process with known DKL=0 (e.g., a stationary Gaussian process with a 1/f-like spectrum, thresholded to binary) using the same word length, binning, and overlap as the retina analysis; apply the estimator and extrapolation of Eq (10) exactly as described. If the extrapolated DKL is significantly nonzero, the independence assumption behind the bias expansion is violated in a way that produces false positives for irreversibility, undermining the retina claim. If it is zero within error bars, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim is that the plug-in estimator in Eq (9) has the bias expansion in Eq (10) with the leading coefficient in Eq (11). This expansion is derived from Eq (6), which states that the covariance of the plug-in probability estimates is δ_{W,W'} P(W)/N, a result that holds only for N independent draws from the true distribution. For the retina experiment, the samples are 200 ms words cut from a continuous spike train, and consecutive words overlap and are strongly correlated; spike trains exhibit correlations on time scales comparable to or longer than the window. No test of independence, nor any check that the 1/N + 1/N^2 extrapolation is valid for such correlated data, is reported. If the effective sample size is much smaller than N, the leading bias is not A/N with A from Eq (11), and extrapolating with that form can leave a residual bias that masquerades as genuine irreversibility. The shuffle control in Fig 3 breaks temporal correlations and only demonstrates that the estimator is unbiased for independent samples; it does not test the null hypothesis of a time-reversible but correlated process. Additionally, Eq (9) is undefined whenever a word W is observed but its time reverse ~W is not; the text never states how zero reverse counts are handled. Any regularization (pseudocount, omission) alters the bias and is not accounted for in Eq (10). Together these gaps mean the method is not established in the regime where it is applied to real neural data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a model-free estimator of the Kullback-Leibler divergence between the distributions of forward and time-reversed trajectories, based on plug-in frequency estimates and a finite-sample bias correction. The authors derive a bias expansion (Eqs. 10-11), extrapolate to infinite sample size, validate the method on a two-state Markov chain (recovering D_KL = 0) and a three-state cycle (recovering the analytic entropy-production rate), and apply it to spike trains from salamander retinal ganglion cells, reporting nonzero D_KL for single neurons and supralinear growth with window length.","tokens_in":8087,"tokens_out":9734,"duration_ms":96587,"significance":"If the estimator is valid in the regimes where it is applied, this is a significant methodological contribution: it provides a direct, dynamics-agnostic route to quantifying irreversibility from time series, with clean synthetic validation and a novel neural-data application. The paper also gives a concrete formula for the leading finite-sample bias, which is useful beyond the specific extrapolation procedure. However, the strength of the retina claim depends on assumptions about sample independence and the handling of unobserved words, and these assumptions are not checked; the method as presented is therefore not yet established in the regime of its headline application.","major_comments":[{"comment":"The plug-in estimator in Eq. (9) is undefined whenever a word W is observed but its time reverse ~W is not: the term ln[PN(W)/PN(~W)] diverges. The manuscript never states how zero reverse counts are handled (pseudocounts, omission, or otherwise), and any such regularization changes the bias and invalidates the unmodified expansion in Eqs. (10)-(11). This is not a technicality for the retina application, where rare words are common; the authors must specify the procedure used and account for its effect on the bias.","section":"Eq. (9) and Section 'We start by discretizing'"},{"comment":"The bias expansion in Eq. (10) is derived from Eq. (6), which assumes that the N word samples are independent draws from the true distribution. In the retina experiment, the 200 ms words are cut from a continuous spike train, and consecutive words are likely overlapping and correlated; the paper reports no test of independence and no check that the 1/N + 1/N^2 extrapolation is valid for such data. The shuffle control in Fig. 3 breaks temporal correlations, so it only demonstrates unbiasedness for independent samples; it does not test the null hypothesis of a time-reversible but correlated process. If the effective sample size is much smaller than N, the extrapolation can leave a residual bias that masquerades as genuine irreversibility, directly undermining the central claim about single-neuron D_KL.","section":"Eq. (10) and Fig. 3"},{"comment":"The derivation of the bias expansion is only sketched ('following the same path as for the entropy'), and the coefficient A in Eq. (11) is asserted without a full derivation or a reference. Since the entire extrapolation procedure rests on this expansion, the authors should provide a complete derivation (or a supplementary appendix) that covers the treatment of palindromic words, the definition of Omega', and the conditions under which the 1/N^2 term has the stated form.","section":"Eqs. (10)-(11)"}],"minor_comments":[{"comment":"The standard Miller-Madow correction for the plug-in entropy estimator is (Omega - 1)/(2N), not Omega/(2N); if the authors intend the large-Omega approximation, they should state this explicitly.","section":"Eq. (8)"},{"comment":"The term 'model-free' is used for a method that assumes a particular parametric bias expansion and sample independence; consider softening this wording to 'assumption-light' or 'dynamics-agnostic'.","section":"Abstract and Section 'In summary'"},{"comment":"The caption contains a typo: '1=N' should be '1/N'.","section":"Fig. 1 caption"},{"comment":"The inset shows D_KL(T) normalized by D_KL(T=10 Delta_tau) and averaged over ten neurons; please clarify how the normalization is performed and whether the error bars include the across-cell variance.","section":"Fig. 3 inset"},{"comment":"The text does not report the number of spike trains, the total recording duration, or the number of words used per cell; these details are necessary to assess the reliability of the extrapolation in Fig. 3.","section":"Retina application"}],"recommendation":"major_revision","confidential_remarks":"The synthetic validation is genuinely clean, and the bias formula in Eq. (11) appears to be correct for independent multinomial samples (I checked it via a second-order Taylor expansion). The retina claim, however, is the most novel and also the most fragile part of the paper: the lack of any independence check or a correlated-reversible null control means the headline result could be an artifact of the extrapolation. The issues are fixable with additional analysis (e.g., using non-overlapping windows, estimating an effective sample size, or testing on a reversible correlated surrogate), so I would not reject the manuscript, but the revision needs to be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful takeaway: this is a short Letter that imports the well-known entropy-estimation trick—expand the finite-sample bias in powers of 1/N and extrapolate—to the DKL between forward and reverse trajectories. The new piece is the explicit bias formula, Eqs. (10)-(11), and the application to single retinal ganglion cells. The synthetic checks are genuinely clean: the estimator returns zero for a two-state equilibrium system and matches the analytic entropy-production rate for a three-state cycle. That part is credible.\n\nThe problem is in the step from synthetic to real data. Eq. (9) is undefined whenever a word is observed but its time reverse is not; the paper never says how zero counts are handled. Any fix—pseudocount, dropping those words—changes the bias, and the 1/N expansion has to be reworked accordingly. That is a concrete omission.\n\nThe larger issue is the independence assumption behind Eq. (6) and hence Eq. (10). For the retina data, the 200 ms words are cut from a continuous spike train; they are not independent samples, and the paper does not test whether the 1/N + 1/N^2 form actually holds in that regime. The shuffle control in Fig. 3 only checks that the estimator is unbiased when the data have been decorrelated; it says nothing about a time-reversible but correlated process. So the nonzero single-neuron DKL could in principle be an artifact of the extrapolation, and the paper doesn't rule that out.\n\nI think the authors know the limitations; the text says 'there is no magic' and admits the extrapolation relies on the same logic used for entropy estimation. But for a method paper, the missing validation on a time-reversible correlated null is a real gap, not a nitpick.\n\nStill, the paper deserves a serious referee. The bias expansion is a useful concrete result, the synthetic validation is honest, and the application is interesting. I'd send it out with a request that the authors specify the regularization, and that they add a surrogate test that preserves correlations while enforcing time-reversal symmetry.\n\nRecommendation: conditional accept, with those two additions. It's a solid Letter, not a game-changer.","headline":"A clean bias-correction recipe for plug-in irreversibility estimates, with a provocative retina result that outruns the paper's own validation.","tokens_in":8628,"tokens_out":3867,"would_cite":false,"duration_ms":41280,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.70.Ln"],"model":"deepseek-v4-flash","headline":"A model-free estimator of forward-reverse trajectory divergence, corrected for finite-sample bias, recovers zero irreversibility in equilibrium and detects a nonzero arrow of time in single retinal neurons.","keywords":["Kullback-Leibler divergence","arrow of time","irreversibility","finite-sample bias","retinal ganglion cells","non-Markovian dynamics","entropy production","model-free estimation"],"falsifier":"Generate a long stationary, time-reversible but non-Markovian time series with known zero irreversibility, for example a Gaussian process with symmetric autocorrelation, and apply the estimator to 200 ms windows: if the extrapolated $D_{KL}$ is systematically nonzero across many realizations the bias expansion is missing a term, while if it is zero within error the correction works as claimed.","tokens_in":7594,"feed_emoji":"⏳","tokens_out":12680,"duration_ms":124592,"temperature":0.7,"pith_summary":"The paper tries to establish that the arrow of time can be read directly from observed time series, with no model of the underlying dynamics. It quantifies irreversibility as the Kullback-Leibler divergence ($D_{KL}$), a measure of how different two probability distributions are, between the distribution of trajectory windows and their time reverses. The central technical point is that the naive plug-in estimate of this divergence is systematically biased upward by finite sample size, with a computable $1/N$ plus $1/N^2$ correction, so extrapolating to infinite data removes the bias. Calibrated on simulations, the corrected estimator returns $D_{KL}=0$ for a detailed-balance system and the right entropy-production rate for a driven cycle; applied to salamander retinal neurons it finds nonzero irreversibility in single cells over 200 ms windows, implying strongly non-Markovian dynamics. If this works, it gives a reliable way to detect time-asymmetry in systems where no heat flow or thermodynamic flux is directly measurable.","feed_headline":"Single neurons show the arrow of time, model-free estimate finds","feed_subtitle":"Corrected for finite samples, the estimator sees irreversibility in 200-millisecond windows of retinal spike trains.","key_machinery":"The machinery is the plug-in estimator of Eq. (9) joined to the bias expansion of Eq. (10). The central object is $D_{KL}$, the Kullback-Leibler divergence between the probability of a discretized trajectory word $W$ and the probability of its time-reversed partner $\\tilde W$; the estimator counts how much more often forward words appear than reversed words. Because empirical counts fluctuate, the plug-in estimate is biased upward, with a leading term $A/N$ whose coefficient is $A=\\frac{1}{2}[\\Omega'+\\sum_W P(W)/P(\\tilde W)]$. Extrapolating across sample sizes removes the leading bias, and using the asymptotic form $D_{KL}(T)\\to\\sigma T+\\sigma_1$ extracts a rate $\\sigma$ from finite windows.","core_discovery":"The central claim is that a model-free, direct estimator of the forward-reverse trajectory divergence can be made quantitatively reliable by correcting its finite-sample bias. The plug-in estimator is $\\hat{D}_N(T)=\\sum_W P_N(W)\\ln[P_N(W)/P_N(\\tilde W)]$, and the paper derives the bias expansion $\\langle \\hat{D}_N(T)\\rangle = D_{KL}(T)+A/N+B/N^2+\\cdots$ with $A=\\frac{1}{2}[\\Omega'+\\sum_W P(W)/P(\\tilde W)]$, then extrapolates to $N\\to\\infty$. On a simulated two-state Markov chain obeying detailed balance this recovers $D_{KL}=0$ within a standard deviation, and on a driven three-state cycle the extrapolated $D_{KL}(T)/T$ converges to the known entropy-production rate $\\sigma\\approx20.7\\mathrm{s}^{-1}$ when the time bin is small enough. On salamander retinal data, the extrapolated $D_{KL}(T=200\\mathrm{ms})$ is nonzero for almost every single neuron, vanishes after shuffling the spike trains, and grows supralinearly with $T$ on the 40-200 ms scale.","pith_inferences":["A natural next test is to push the same estimator to longer windows, where the supralinear growth should cross over to a linear $D_{KL}(T)\\sim\\sigma T$ if a finite entropy-production rate exists; locating that crossover would give a rate for retinal spike trains.","Because the $1/N$ coefficient involves ratios $P(W)/P(\\tilde W)$, data sets in which many time-reversed words are never observed may need a regularized estimator, and sensitivity of the extrapolation to that regulator would reveal when the bias expansion has broken down.","Applied to pairs or populations of simultaneously recorded neurons, the same method could test whether the arrow-of-time evidence is enhanced by correlations between cells, rather than only by temporal correlations within each cell.","The method should transfer to other high-dimensional time series with a well-defined forward-reverse pairing, such as gene expression or weather records, where model-based entropy-production estimates are unavailable."],"forward_implications":["A nonzero extrapolated $D_{KL}$ from a stationary time series is evidence that the underlying dynamics are not time-reversal invariant, without needing a physical model of the system.","For Markovian dynamics the method recovers the thermodynamic entropy-production rate from trajectory data alone, matching the analytic rate when the discretization bin is small enough.","Because any binary Markov model obeys detailed balance, the observed nonzero $D_{KL}$ in single retinal neurons identifies the non-Markovian character of the spike train.","The supralinear growth of $D_{KL}(T)$ means that evidence for the arrow of time accumulates synergistically across successive brief windows rather than additively."],"supporting_citations":[{"why":"Supplies the classic statement that plug-in information estimates carry a systematic 1/N bias, the starting point for the correction.","marker":"[33]"},{"why":"Derives the upward bias of information measures under limited sampling, giving the entropy-parallel form used here.","marker":"[36]"},{"why":"Introduces the strategy of extrapolating across sample sizes for entropy and information in neural spike trains, which this paper adapts to the Kullback-Leibler divergence.","marker":"[34]"},{"why":"Links trajectory-level time reversal to the large-deviation structure of stochastic dynamics, grounding the use of DKL as irreversibility evidence.","marker":"[13]"},{"why":"Defines entropy production along a stochastic trajectory, connecting DKL to a thermodynamic rate.","marker":"[14]"},{"why":"Provides the network formula for the entropy-production rate used to check the three-state cycle result.","marker":"[21]"},{"why":"Supplies the salamander retina experiment with naturalistic movies whose responses provide the neural data analyzed here.","marker":"[37]"}],"fun_headline_variants":["Model-free estimator recovers arrow of time in single neurons","Bias-corrected irreversibility measure works on single neurons","Direct estimator of time irreversibility fixed for finite samples","Arrow of time seen in retinal neurons via model-free estimate","Finite-sample-corrected irreversibility estimator passes tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the observed trajectory windows are independent draws from a single fixed distribution, so that the finite-sample bias has exactly the $1/N$ plus $1/N^2$ form that the extrapolation removes; no test of this assumption is reported for the retina spike trains.","fun_headline_variants_meta":{"raw":{"variants":["Model-free estimator recovers arrow of time in single neurons","Bias-corrected irreversibility measure works on single neurons","Direct estimator of time irreversibility fixed for finite samples","Arrow of time seen in retinal neurons via model-free estimate","Finite-sample-corrected irreversibility estimator passes tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3850,"prompt_tokens":973,"completion_tokens":2877,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2793}},"tokens_in":589,"tokens_out":2877,"duration_ms":20096,"temperature":1.0,"reasoning_tokens":2793,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:52:43.299811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long stationary, time-reversible but non-Markovian time series with known zero irreversibility, for example a Gaussian process with symmetric autocorrelation, and apply the estimator to 200 ms windows: if the extrapolated $D_{KL}$ is systematically nonzero across many realizations the bias expansion is missing a term, while if it is zero within error the correction works as claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classic statement that plug-in information estimates carry a systematic 1/N bias, the starting point for the correction."},{"cited_title":"Panzeri and A","cited_arxiv_id":null,"evidence_quote":"Derives the upward bias of information measures under limited sampling, giving the entropy-parallel form used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the strategy of extrapolating across sample sizes for entropy and information in neural spike trains, which this paper adapts to the Kullback-Leibler divergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Links trajectory-level time reversal to the large-deviation structure of stochastic dynamics, grounding the use of DKL as irreversibility evidence."},{"cited_title":"Seifert, Entropy production along a stochastic trajec- tory and an integral fluctuation theorem.Physical Review Letters 95(4), 040602 (2005)","cited_arxiv_id":null,"evidence_quote":"Defines entropy production along a stochastic trajectory, connecting DKL to a thermodynamic rate."},{"cited_title":"Schnakenberg, Network theory of microscopic and macroscopic behavior of master equation systems","cited_arxiv_id":null,"evidence_quote":"Provides the network formula for the entropy-production rate used to check the three-state cycle result."},{"cited_title":"Tkaˇ cik, O","cited_arxiv_id":null,"evidence_quote":"Supplies the salamander retina experiment with naturalistic movies whose responses provide the neural data analyzed here."}],"review_version":1}