{"id":"d0913c9e-85c3-4821-a0da-acaaa8dc16c2","arxiv_id":"2412.03734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A bisecting k-means partition with up to about one million cells yields transfer-operator approximations for the Lorenz system that converge for invariant statistics and autocorrelations, while the quasi-invariant Koopman eigenfunction grows increasingly intermittent.","lead":"The authors used a modified bisecting k-means algorithm to partition the Lorenz attractor into over a million cells and built a transfer operator that reproduces the system's statistics and autocorrelations with increasing accuracy as the partition is refined. The work is a test of how far data-driven, partition-based Koopman methods can be pushed in chaotic systems, with potential applications in climate and turbulence modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed convergence and saturation near 10^5 cells are measured against a single short reference trajectory with only 10^3–10^4 independent samples; without reference error bars, the saturation may be an artifact of reference sampling noise.","rationale":"The paper's strongest and most central assertions are empirical convergence statements, and their only ground truth is temporal statistics from the same T=10^5 trajectory used to construct the operator. The uncertainty in that ground truth is admitted in the text but never quantified. Since the claimed first-order convergence and saturation are precisely the features that would be affected by a noise floor, this is the load-bearing weak point. A longer or independent reference, or at least bootstrap error bars, would discriminate between operator approximation error and reference sampling error. The eigenfunction intermittency claim is honestly presented as hard to assess and is less central. I agree with the reader's identification of this weakest assumption and with the conditional verdict; my concern does not move the verdict but makes the required condition concrete: quantify reference uncertainty before asserting the saturation behavior.","tokens_in":11088,"tokens_out":5644,"duration_ms":61942,"concrete_test":"Generate a new, substantially longer independent reference trajectory (e.g., T=10^6 or 10^7 at the same Δt=10^-3), compute the temporal cumulants and z-autocorrelation from it, and recompute the error curves in Figs. 3.3–3.5 with this improved reference. If the saturation near 10^5 cells persists and the first-order slope is unchanged, the claim stands; if the errors continue to decrease past 10^5 cells or the measured slopes change, the saturation and convergence are artifacts of the original reference noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claims—first-order convergence of cumulants and saturation near 10^5 cells (Figs. 3.3–3.4), and convergence of the z-autocorrelation to the temporal average (Fig. 3.5)—all compare the data-driven operator against temporal statistics computed from the same T=10^5 trajectory that generated the operator (Eqs. 3.1–3.4, §3). The authors acknowledge this limitation: “we have between 10^3 and 10^4 decorrelated in time samples… empirical error… should begin dominating the inaccuracy at around three digits” (§3.1). However, they do not quantify the reference uncertainty, add error bars, or use an independent or longer reference. With only 10^3–10^4 independent samples, the standard error of a correlation or cumulant estimate is roughly 1–3%, so a plateau at the 10^-3 level is exactly the expected noise floor of the reference, not necessarily a property of the operator approximation. If the reference is noisier than assumed, the reported first-order slope and saturation point in Figs. 3.3–3.5 could be dominated by reference error. Since every error measurement is in-sample (same data for operator and reference), the headline convergence claim rests on an unquantified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modified bisecting k-means algorithm for generating hierarchical, nearly uniform partitions of state space, and uses these partitions as piecewise-constant dictionaries in an extended dynamic mode decomposition (EDMD) of the transfer operator. The method is applied to the Lorenz equations with up to O(10^6) cells. The central numerical claims are: (i) invariant-measure cumulants converge at roughly first order in the number of cells and saturate near 10^5 cells (Figs. 3.3 and 3.4); (ii) the z-autocorrelation computed from the operator converges to the temporal autocorrelation as resolution increases, with the Perron-Frobenius operator at larger timescales converging fastest (Fig. 3.5); and (iii) a quasi-invariant-set Koopman eigenfunction becomes increasingly intermittent as resolution increases (Figs. 3.6 and 3.7). The authors are explicit that the eigenfunction convergence itself is difficult to assess and that the temporal reference is limited to 10^3–10^4 decorrelated samples.","tokens_in":11227,"tokens_out":3933,"duration_ms":41237,"significance":"If the convergence claims hold, the paper demonstrates a practical, single-CPU method for constructing transfer-operator approximations with dictionary sizes that are unusually large for EDMD-type methods, while keeping matrix entries well-sampled through nearly uniform hierarchical partitions. The paper is also honest about the limitations of the eigenfunction interpretation. Strengths include: a clear, reproducible experimental setup (initial condition, timestep, threshold parameter p = 1.4e-6 are specified); a parameter-free construction in the sense that no constants are fitted to the reported statistics; and consistent use of the hierarchical partition to define a meaningful refinement path from four cells to over one million cells. The main weakness is that all error measurements compare the operator against temporal statistics computed from the same trajectory that generated the operator, and the reference uncertainty is acknowledged but never quantified.","major_comments":[{"comment":"The central convergence and saturation claims are measured entirely against temporal averages from the same T = 10^5 trajectory that was used to construct the operator (Eqs. 3.1–3.4). As the authors state, this trajectory contains only 10^3–10^4 decorrelated samples, which implies a standard error on the order of 1–3% for the reference statistics. The saturation seen near 10^5 cells at relative errors around 10^-3 is therefore exactly what one would expect from reference sampling noise. The paper should quantify this uncertainty directly, for example by showing bootstrap confidence intervals for the temporal averages, or by validating against an independently generated longer trajectory, before asserting that the saturation and the first-order slope are properties of the operator approximation.","section":"§3.1, Figs. 3.3–3.5"},{"comment":"The claimed 'first-order convergence rate' is supported only by a dashed guide line drawn through the points; no regression slope, confidence interval, or quantitative fit is reported. Given that the saturation may be a reference-noise artifact, the paper should either provide a formal fit with uncertainty or soften the first-order claim to a qualitative observation. This is load-bearing because the first-order convergence is a headline result in the abstract and conclusions.","section":"§3.1, Figs. 3.3–3.4"},{"comment":"The claim that the Koopman eigenfunction 'becomes increasingly intermittent as resolution increases' is based on visual inspection of three resolutions. No quantitative intermittency metric, norm, or convergence diagnostic is provided, and the authors themselves note that it is unclear in what sense the eigenfunctions converge. If this eigenfunction is a central object of the paper, the claim should be supported by a reproducible measure, such as a scale-dependent variance or a Sobolev-type norm; otherwise, the statement should be presented as a qualitative observation rather than a finding.","section":"§3.1, Fig. 3.7"}],"minor_comments":[{"comment":"The captions and axis labels of Figs. 3.3–3.5 contain corrupted glyphs such as '/uni2212', '/g22', and '/g24', which make the figures unintelligible in the current version. These need to be repaired before publication.","section":"§3.1, figure captions"},{"comment":"The definitions of Nmin, Nthreshold, and pmin are confusingly written. In particular, pmin is introduced as a 'minimum probability' but then used in expressions where it must be interpreted as a percentage or a scaled quantity (e.g., Nmin = floor(100/pmin) and Nthreshold = ceil(pmin Nmin)). The units and scaling should be made explicit.","section":"§2, clustering threshold"},{"comment":"The caption says the hierarchy starts 'with four cells (top right)' while the main text says 'four partitions (top left)'; these should be reconciled.","section":"Figure 3.1 caption"},{"comment":"References [1] and [2] are the same paper, and references [8] and [9] are also duplicates. These should be consolidated.","section":"References"},{"comment":"The paper would benefit from a data/code availability statement, since the numerical experiments are a central contribution and the exact algorithm parameters are only partially specified.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the author's own prior work (refs. [34,35]) for the definitions of the generator and Perron-Frobenius operator estimators. Editors may wish to confirm that the current paper's contribution—the modified bisecting k-means hierarchy and its scaling demonstration—is sufficiently distinct and that the heavy self-citation is justified. The main technical risk is the unquantified in-sample reference error; even if the authors choose not to add an independent reference trajectory, a thorough bootstrap analysis would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a computational study of a modified bisecting k-means for building piecewise-constant transfer operators for Lorenz, and it takes the partition size up to about 10^6 cells. The new bit is the probability-threshold splitting rule and the scale of the convergence study. It does that well, and the authors are candid about the parts that don't converge cleanly, especially the Koopman eigenfunction.\n\nWhat I think is actually solid: the algorithm is simple, runs on a single core, produces nearly uniform entropy partitions, and the cumulant convergence plots show first-order behavior before a plateau. The autocorrelation comparisons across generator and Perron-Frobenius at different timescales are informative. The eigenfunction \"intermittency\" observation is interesting and honestly flagged as not properly converged.\n\nThe soft spots are the ones you'd guess. Every \"ground truth\" statistic is computed from the same T=10^5 trajectory that generates the operator. They estimate 10^3–10^4 independent samples, which puts the reference error around 1–3%. That is exactly the level at which the cumulant curves plateau (three digits). So the saturation at ~10^5 cells could easily be the noise floor of the reference, not a property of the operator. The paper acknowledges this in one sentence but doesn't quantify it with error bars or cross-check with a longer or independent trajectory. That weakens the central \"convergence\" claim and the reported first-order slope. The slope itself is probably real, but the plateau is unproven.\n\nAlso missing: any comparison with a standard Ulam grid at comparable resolution, and no code or data release, so the reproducibility is limited to trusting the figure curation. The novelty relative to the authors' own JFM papers [34,35] and the clustering paper [16] is incremental, but the probability-threshold criterion is distinct and the 10^6-cell test is new.\n\nThe stress-test note you passed along lands: the concern about the reference trajectory is the main real limitation. But I don't think it sinks the paper; it just means the convergence claims need more careful statistical treatment before they're stated as firmly as they are.\n\nI'd send this to a good referee. The method is useful, the scaling claim is worth checking, and the flaws are addressable—add error bars, run a longer reference, and do a Ulam-grid baseline for Lorenz. For me, this is a conditional accept rather than a reject.","headline":"Useful scaling study of a k-means partition-based transfer operator, but the convergence claims are measured against a reference trajectory short enough to produce the observed saturation itself.","tokens_in":11852,"tokens_out":2505,"would_cite":true,"duration_ms":25091,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q25","68R10","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a piecewise-constant dictionary built by modified bisecting k-means, scaled to over one million cells, makes data-driven transfer operators converge on the Lorenz attractor's invariant statistics, autocorrelations…","keywords":["modified bisecting k-means","transfer operator","Koopman eigenfunction","Perron-Frobenius operator","extended dynamic mode decomposition","Lorenz equations","invariant measure","piecewise-constant dictionary"],"falsifier":"Recompute the same cumulant and autocorrelation errors against a second, much longer trajectory, say $T=10^6$ or $10^7$, generated independently; if the saturation near $10^5$ cells persists, the paper's convergence picture holds, and if the error drops substantially, the earlier saturation was reference-sampling noise. A simpler check is to plot log error versus log cell count and test that the slope is approximately one only below the resolution limit set by the decorrelated sample count.","tokens_in":10773,"feed_emoji":"🌀","tokens_out":6139,"duration_ms":57650,"temperature":0.7,"pith_summary":"The paper is trying to establish that the extended dynamic mode decomposition, supplied with a state-space partition built by a modified bisecting k-means algorithm, can approximate the transfer operator of a chaotic system well enough to recover its steady-state statistics, temporal autocorrelations, and a distinctive Koopman eigenfunction. The test case is the Lorenz equations, with dictionaries of piecewise-constant functions up to roughly one million cells constructed from a single trajectory of length $T=10^5$. The authors report first-order convergence of invariant cumulants that saturates near $10^5$ cells, convergence of the $z$-autocorrelation toward the time-series value, and a quasi-invariant Koopman eigenfunction that becomes increasingly intermittent as resolution increases. The practical stake is that this partition-based approach can be pushed to very large dictionaries on a single CPU core, so the limits appear to be set by available data rather than by the cost of the dictionary.","feed_headline":"Million-cell dictionaries converge on Lorenz chaos","feed_subtitle":"At a million cells, data-driven transfer operators match Lorenz time-series statistics and reveal an intermittent eigenfunction.","key_machinery":"The central object is a partition of state space into leaf cells of nearly equal probability, produced by a modified bisecting k-means algorithm with a splitting threshold $p_{\\min}$: a cell is split only if its probability exceeds the threshold, yielding near-uniform entropy and no user-specified cluster count. Each cell is an element of the piecewise-constant dictionary for the extended dynamic mode decomposition, and the classifier maps each state to an integer cell index. From the resulting cell-index time series, the paper builds sparse matrix representations of the infinitesimal generator $Q$ and the Perron-Frobenius operator, then computes eigenvalues and eigenvectors by inverse iteration. The near-uniform entropy property ensures every column of the operator is estimated from sufficiently many samples, while the hierarchical tree provides consistent coarse-to-fine partitions and makes classification of new data cheap.","core_discovery":"On the paper's own terms, the central claim is that a data-driven discretization of the continuity equation for a noiseless chaotic system can be refined to over a million piecewise-constant cells and, in that limit, reproduces the invariant measure, the autocorrelation of the $z$ variable, and the quasi-invariant Koopman eigenfunction. The cumulants converge at first order in cell count until roughly $10^5$ cells, after which the error saturates at about three digits, consistent with the finite number of decorrelated samples in the reference trajectory. The $z$-autocorrelation computed from the operator approaches the temporal average as resolution increases, and this holds both for the infinitesimal generator and for Perron-Frobenius operators built at several timescales. The associated Koopman eigenfunction develops finer-scale, increasingly intermittent structure with resolution, and the authors conclude that while statistics converge, the eigenfunctions themselves are harder to declare convergent. The paper also finds that the generator is more dissipative than Perron-Frobenius operators constructed at larger timescales, but that Perron-Frobenius operators at timescales much longer than the target eigenvalue lose meaning.","pith_inferences":["I infer that the near-uniform-entropy splitting rule effectively makes the hierarchy an adaptive quadrature for state space, and the same splitting criterion could be replaced by local error indicators if a different observable family were the target.","I infer that if a much longer reference trajectory were used, the saturation near $10^5$ cells would likely move outward, which would confirm that the accuracy ceiling is data-limited rather than dictionary-limited.","I infer that the increasingly intermittent Koopman eigenfunction means early-warning indicators built from such eigenfunctions should be reported together with the partition resolution, since their spike structure changes qualitatively as the cell count grows."],"forward_implications":["For the Lorenz test case, invariant cumulants converge at first order as the partition is refined, saturating near $10^5$ cells for a trajectory of length $T=10^5$.","The $z$-autocorrelation recovered from the operator converges to the temporal average as resolution increases, for both the generator and Perron-Frobenius constructions.","The quasi-invariant Koopman eigenfunction develops finer and more intermittent structure with increasing resolution, so individual eigenfunctions are harder to call converged even while statistical quantities converge.","The infinitesimal generator is more dissipative than Perron-Frobenius operators built at larger timescales, but very large timescales skip the target eigenvalue; the paper recommends an iterative strategy of estimating timescales from the generator first and then constructing a Perron-Frobenius operator at an appropriate timescale.","The full computation runs on a single CPU core, showing that the modified bisecting k-means dictionary scales to over a million terms without specialized hardware."],"supporting_citations":[{"why":"Supplies the Lorenz equations as the chaotic test system whose statistics are approximated.","marker":"[25]"},{"why":"Supplies the Runge-Kutta 4 scheme used to generate the reference trajectory of length $T=10^5$.","marker":"[4]"},{"why":"Provides the original bisecting k-means algorithm that the paper modifies with a splitting criterion.","marker":"[37]"},{"why":"Provides the partition-based construction of the sparse generator and Perron-Frobenius operator, along with the ensemble-average formulas used for comparison.","marker":"[34]"},{"why":"Supplies the same data-driven operator construction and the assumptions underlying the formulas in the companion application.","marker":"[35]"},{"why":"Supplies the inverse iteration and sparse linear algebra methods used to extract eigenvalues and eigenvectors of the large operators.","marker":"[17]"}],"fun_headline_variants":["Bisecting k-means builds million-cell Lorenz operators","Million-cell transfer operators match Lorenz statistics","Data-driven operators expose intermittent Lorenz eigenfunctions","Generator vs Perron-Frobenius: Lorenz operator timescales"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reference statistics come from a single trajectory of length $T=10^5$, which the authors estimate contains only $10^3$ to $10^4$ effectively independent samples; if those reference values are noisier than the reported three digits, the observed saturation near $10^5$ cells could be an artifact of the reference rather than a property of the operator approximation.","fun_headline_variants_meta":{"raw":{"variants":["Bisecting k-means builds million-cell Lorenz operators","Million-cell transfer operators match Lorenz statistics","Data-driven operators expose intermittent Lorenz eigenfunctions","Generator vs Perron-Frobenius: Lorenz operator timescales"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1445,"prompt_tokens":883,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":498}},"tokens_in":499,"tokens_out":562,"duration_ms":6334,"temperature":1.0,"reasoning_tokens":498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:08:51.860966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the same cumulant and autocorrelation errors against a second, much longer trajectory, say $T=10^6$ or $10^7$, generated independently; if the saturation near $10^5$ cells persists, the paper's convergence picture holds, and if the error drops substantially, the earlier saturation was reference-sampling noise. A simpler check is to plot log error versus log cell count and test that the slope is approximately one only below the resolution limit set by the decorrelated sample count.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Lorenz equations as the chaotic test system whose statistics are approximated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Runge-Kutta 4 scheme used to generate the reference trajectory of length $T=10^5$."},{"cited_title":"Steinbach, G","cited_arxiv_id":null,"evidence_quote":"Provides the original bisecting k-means algorithm that the paper modifies with a splitting criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the partition-based construction of the sparse generator and Perron-Frobenius operator, along with the ensemble-average formulas used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the inverse iteration and sparse linear algebra methods used to extract eigenvalues and eigenvectors of the large operators."}],"review_version":1}