{"id":"b442a27e-e670-459b-a148-81e6de6957f6","arxiv_id":"2510.07140","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A review of the author's parameter-free spike-train synchrony and directionality measures, including two published latency-correction algorithms, with simulated examples.","lead":"This paper reviews a family of time-resolved, parameter-free measures for spike-train synchrony and directionality, including ISI- and SPIKE-Distance, SPIKE-Synchronization, and SPIKE-Order, plus two previously published latency-correction algorithms. It is a synthesis aimed at neuroscientists and data analysts who need practical tools for comparing spike trains.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 33's transitivity extrapolation is arithmetically inconsistent as written: the unnormalized sum over intermediate k inflates δ by n−m−1, so the §4.2 latency-correction algorithm is mis-specified until corrected.","rationale":"The reader accepted the manuscript as a sound review, with the weakest assumption being the transitivity premise of Eq. 33. I agree that transitivity is an idealization, but the more immediate problem is that Eq. 33 is arithmetically inconsistent even under that idealization: the unnormalized sum over intermediate k multiplies the target latency by n−m−1. This is not merely a limitation in real, jittered data—it is an internal inconsistency in the algorithm as specified. Because Section 4.2's central claim (Extrapolation is almost as good as reduced matrix simulated annealing but faster) depends on this formula, the manuscript cannot be accepted without correction or clarification of Eq. 33. I emphasize that this is likely a repairable typo—perhaps a missing normalization factor in the transcription—and the broader review content may well be sound. That is why I recommend CONDITIONAL rather than REJECT: the article should be accepted only after the formula is corrected and the algorithm re-verified. I am not flagging any concern about authorial intent or the quality of the review as a whole; the issue is scoped to a specific equation whose correctness is load-bearing for the latency-correction section.","tokens_in":18011,"tokens_out":12330,"duration_ms":101892,"concrete_test":"Construct a noiseless synfire chain with N=5 spike trains and known additive latencies L_n = n·dt, so the exact pairwise differences are δ(n,m) = L_n − L_m. Set stop diagonal d=1 and evaluate Eq. 33 for (n,m)=(5,1): the printed formula yields 3δ(5,1) rather than δ(5,1). Then check the original published version (Kreuz et al., J Neurosci Methods 416, 2025, Eq. for Extrapolation) for a 1/(n−m−1) prefactor. If the prefactor is present, the manuscript has a typographical error that should be corrected; if it is absent, the algorithm as described cannot recover the true latencies and the §4.2 performance claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central methodological novelty in Section 4.2 is the Extrapolation direct shift, which uses Eq. 33 to replace overlap-affected outer entries of the spike time difference matrix. But Eq. 33 as printed cannot be correct under the additive-latency model it invokes. If δ(a,b) = L_a − L_b, then each term δ(n,k) + δ(k,m) equals L_n − L_k + L_k − L_m = L_n − L_m = δ(n,m). Summing k = m+1 … n−1 therefore gives (n−m−1)·δ(n,m), not δ(n,m). For example, with n=5, m=1, the formula returns 3δ(5,1). This factor grows with the distance in spike-train index and would inflate exactly the outer matrix entries that the extrapolation is meant to repair. Since the subsequent shift computation in Eq. 35 averages over the full STDM, these inflated entries would corrupt the inferred shifts and undermine the claim that the Extrapolation direct shift 'achieves almost the same performance' as reduced matrix simulated annealing. The reader's concern about transitivity is real but secondary: even granting transitivity, the printed formula needs a normalization factor of 1/(n−m−1) (or a different summation convention). The condition 'm−n>d' also appears inconsistent with the n>m ordering implied by the summation index; this may be a typographical issue, but it compounds the ambiguity. As written, Section 4.2's algorithm is not fully specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This article is a review of the author's SPIKE-family measures of spike train synchrony and directionality. The first part introduces the ISI-Distance and SPIKE-Distance, their time-resolved profiles, and multivariate extensions. The second part describes SPIKE-Synchronization and the Spike Train Order/SPIKE-Order framework, including the Synfire Indicator and the leader-to-follower sorting procedure. The third part turns to latency correction, presenting direct-shift and simulated-annealing algorithms for non-overlapping events and the reduced-matrix Extrapolation and reduced-matrix simulated annealing for overlapping events. The presentation is unified and illustrated on synthetic data, with examples of real-data applications. The central technical claims are the definitions and normalization properties of the measures and the effectiveness of the two latency-correction algorithms.","tokens_in":18431,"tokens_out":6952,"duration_ms":57630,"significance":"SPIKE-Synchronization, SPIKE-Distance, and related measures are widely used, and a single self-contained account of the definitions, profiles, and algorithmic extensions has practical value. The manuscript benefits from explicit formulas for each measure, references to open-source implementations (SPIKY, PySpike, cSPIKE), and a clear discussion of leader-follower sorting. The latency-correction section is the least standard part, but its presentation is undermined by the error in Eq. (33); until that is corrected, the algorithmic contribution cannot be evaluated from the manuscript alone. If the formula is repaired, the paper would be a useful reference for practitioners.","major_comments":[{"comment":"The extrapolation formula is not correct as printed. Under the additive-latency model δ(a,b)=L_a−L_b, each summand δ(n,k)+δ(k,m) equals δ(n,m), so the sum over k=m+1,...,n−1 returns (n−m−1)·δ(n,m), not δ(n,m). The condition 'm−n>d' is also inconsistent with the summation bounds: for n>m one would need n−m>d; for n<m the sum over k=m+1,...,n−1 is empty. The formula should be normalized, e.g. δ(n,m)= [1/(n−m−1)] Σ_k [δ(n,k)+δ(k,m)] with n−m>d, or the intended convention should be stated explicitly. As written, the Extrapolation direct shift is not fully specified, and the claim in §4.2 that it achieves almost the same performance as reduced-matrix simulated annealing cannot be checked.","section":"§4.2, Eq. (33)"},{"comment":"The extrapolation relies on transitivity of pairwise spike time differences, which is exact only for global additive latencies. Real data with jitter, missing or extra spikes, or nonstationary propagation violate transitivity, and the outer entries being extrapolated are precisely the ones most affected by overlap. The manuscript should state this limitation explicitly and should provide evidence on how the extrapolation degrades when the assumption is violated, e.g., by comparing extrapolated outer entries with direct overlap-free estimates in the simulations of [14]. This is not a reason to abandon the method, but it is necessary for judging the 'almost the same performance' claim.","section":"§4.2, Eq. (33)"}],"minor_comments":[{"comment":"The summation in Eq. (17) is written as Σ_{m,n}, which is ambiguous and would include both n and m. It should be Σ_{m≠n} or Σ_{m=1,m≠n}^N to match the text: averaging over all N−1 bivariate coincidence indicators involving spike i.","section":"§3.2, Eq. (17)"},{"comment":"The equivalence of the two definitions of the Synfire Indicator is asserted but not shown. A one-line rearrangement of the double sum over spike trains would make the equality transparent.","section":"§3.3, Eqs. (24) and (27)"},{"comment":"Several typos should be corrected: 'propogates' in the abstract, 'reduxed matrix' in §4.2, 'electrophyiological' in §2.3, and 'next on' in the description of Fig. 8b.","section":"Throughout"},{"comment":"The iterative scheme in Fig. 9 is described qualitatively. Specifying the stopping criterion and the parameter choices (stop diagonal d, annealing temperature/cooling schedule) would improve reproducibility, although the reader is referred to [14] for details.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a self-review of the author's own SPIKE-family methods. This is appropriate for the special issue, but the editor may wish to ensure that the review clearly acknowledges that the latency-correction algorithms in Section 4 are previously published in [13,14] and that the present article's contribution is their unified presentation. The most important fix is Eq. (33), which as printed is mathematically inconsistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know first: this is a review article, not a research paper. Thomas Kreuz is reviewing his own family of measures — ISI-Distance, SPIKE-Distance, SPIKE-Synchronization, SPIKE-Order, and Spike Train Order — along with two latency-correction algorithms from his own earlier work. There is no new method and no new experimental result. That is fine for a special-issue review, and the paper does a genuinely useful job of putting the definitions, formulas, and illustrative examples in one place. I checked several formulas against the cited sources and they are consistent. The paper also points to open-source implementations (SPIKY, PySpike, cSPIKE), which is a real plus for reproducibility. The application anecdotes and citation trail show that these methods are in use well beyond the author's own group, so the review is not pure self-promotion.\n\nNow the soft spot. The stress-test note is right, and it is more concrete than the reader's transitivity concern. Equation 33, the extrapolation rule for the SPIKE time difference matrix, is arithmetically wrong as written. Under the additive-latency model, δ(n,k) + δ(k,m) = δ(n,m), so summing over k from m+1 to n−1 gives (n−m−1)·δ(n,m), not δ(n,m). The printed formula would inflate exactly the outer matrix entries it is meant to repair. The condition \"m−n>d\" also looks backwards given the summation index (which implies n>m). A reader trying to implement the Extrapolation direct shift from this paper alone will get corrupt shifts. Since the original algorithm is in ref [14], this is not fatal for the field, but it is a genuine error in the exposition that a referee should catch. The transitivity assumption itself is a simplification — real data with jitter and missing spikes does not have exact additive latencies — but the algorithm is explicitly scoped to sparse spike trains with well-defined global events, so that is a reasonable limitation rather than a hidden flaw.\n\nBottom line: this is a solid review for newcomers and for researchers wanting a single reference to these measures. Use it for orientation, but for any actual implementation go back to the primary sources, especially for Section 4.2. It deserves a serious referee, not a desk reject, because the core content is sound and the error is local and fixable. I would not cite it in my own methods work — I would cite the original method papers — but I might assign it to a student as an entry point.","headline":"Useful consolidation of the author's own spike-train synchrony measures, but the exposition of the latency-correcting extrapolation in Eq. 33 has a real arithmetic error that needs fixing.","tokens_in":18851,"tokens_out":2200,"would_cite":false,"duration_ms":20653,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A review of parameter-free, time-resolved spike train synchrony measures and two latency-correction algorithms that handle overlapping global events, with the fast extrapolation shift nearly matching simulated annealing.","keywords":["spike train synchrony","ISI-Distance","SPIKE-Distance","SPIKE-Synchronization","SPIKE-Order","latency correction","synfire pattern","spike time difference matrix"],"falsifier":"Generate a set of spike trains with overlapping global events in which latencies are deliberately non-additive, e.g., propagation delays that differ between events or depend on event index; run the Extrapolation direct shift and reduced-matrix simulated annealing. If the extrapolation end cost is substantially worse than simulated annealing's on a dataset where the two were reported to tie, the transitivity assumption is falsified in that regime.","tokens_in":17920,"feed_emoji":"🧠","tokens_out":3563,"duration_ms":30902,"temperature":0.7,"pith_summary":"This review makes the case that a single family of parameter-free, time-resolved measures—the ISI-Distance, SPIKE-Distance, SPIKE-Synchronization, and SPIKE-Order—can capture both the degree and the direction of synchrony in sets of spike trains, and that these measures can serve as the basis for correcting systematic latency differences between spike trains. The central new claim is that two recently proposed latency-correction algorithms succeed even when global spiking events overlap, a regime where naive spike matching produces spurious coincidences. The faster algorithm, which extrapolates the spike time difference matrix from its unaffected inner diagonals, is reported to reach almost the same alignment quality as a simulated-annealing approach at a fraction of the computational cost. If correct, these tools give experimental neuroscientists a way to separate true synchrony from propagation delays and to rank neurons from leader to follower.","feed_headline":"New latency-correction algorithm nearly matches simulated annealing","feed_subtitle":"A transitivity-based extrapolation fixes overlapping neural events at a fraction of the cost, review shows.","key_machinery":"The central object is the spike time difference matrix (STDM), built from pairwise averaged time differences between matched spikes after adaptive coincidence detection, together with its symmetric cost matrix. The adaptive coincidence window, half the minimum of the four neighboring interspike intervals, is what makes the whole framework parameter-free: each spike is matched to at most one spike in another train, and coincidences are defined locally. For overlap, the extrapolation formula δ^{(n,m)} = Σ_k [δ^{(n,k)} + δ^{(k,m)}] replaces overlap-corrupted outer matrix entries by sums of inner entries, assuming that latencies are additive.","core_discovery":"On the paper's own terms, the discovery is that synchrony and directionality in spike trains can be quantified without any time-scale parameter and in a time-resolved way, and that the same framework extends naturally from measurement to correction. The paper argues that SPIKE-Synchronization, built on adaptive coincidence detection, yields an unambiguous 0-or-1 spike matching that is local and scale-free; SPIKE-Order and Spike Train Order turn this matching into leader-follower assignments and a Synfire Indicator. The newer latency-correction algorithms then use the spike time difference matrix derived from matched spikes, and the paper's key technical claim is that when global events overl","pith_inferences":["The transitivity assumption behind the extrapolation is likely to break down when different events have different propagation speeds or when spikes are missing; a natural extension would be to test the algorithm against ground-truth non-additive latency models and measure how much the end cost degrades.","The paper describes the framework as universal for discrete datasets, so a testable extension is to apply the same latency-correction machinery to climate, communication, or social-event time series and check whether alignment improves known event correspondences.","The paper notes an open interplay between spike train sorting and latency correction; an implication is that the two operations should perhaps be iterated jointly rather than sequentially, since changing the order changes which spike pairs are used to estimate latencies."],"forward_implications":["If the measures behave as stated, researchers can compare synchrony across datasets without choosing a time scale, and can localize synchrony changes in time.","If the extrapolation direct shift matches simulated annealing, latency correction becomes feasible for very large spike train sets where annealing's high computational cost is prohibitive.","If the iterative rematching scheme works, then overlapping global events, previously a failure mode, can be disentangled and aligned close to the true latencies.","The Synfire Indicator and the sorted leader-follower order provide a quantitative test for synfire-like propagation in any discrete event data."],"fun_headline_variants":["Time-scale-free spike synchrony and directionality measures","Latency correction algorithm nearly matches annealing at lower cost","SPIKE-Synchronization: scale-free, time-resolved neural spike matching","Transitivity-based extrapolation fixes overlapping neural events","New algorithm optimizes spike alignment without time-scale"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The extrapolation step assumes that the time difference between any two spike trains is the sum of the differences through an intermediate spike train, which is exact only when every spike train differs from every other by a single global constant delay.","fun_headline_variants_meta":{"raw":{"variants":["Time-scale-free spike synchrony and directionality measures","Latency correction algorithm nearly matches annealing at lower cost","SPIKE-Synchronization: scale-free, time-resolved neural spike matching","Transitivity-based extrapolation fixes overlapping neural events","New algorithm optimizes spike alignment without time-scale"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3235,"prompt_tokens":719,"completion_tokens":2516,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":2436}},"tokens_in":463,"tokens_out":2516,"duration_ms":10894,"temperature":1.0,"reasoning_tokens":2436,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:59:18.908937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a set of spike trains with overlapping global events in which latencies are deliberately non-additive, e.g., propagation delays that differ between events or depend on event index; run the Extrapolation direct shift and reduced-matrix simulated annealing. If the extrapolation end cost is substantially worse than simulated annealing's on a dataset where the two were reported to tie, the transitivity assumption is falsified in that regime.","supporting_citations":[],"review_version":1}