{"id":"24b372d5-c980-4807-96b9-b09dc54350da","arxiv_id":"2411.14297","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GPD-based local dimension estimates need a measure property called regular variation, which singular measures on fractal attractors typically lack, making the estimates depend on resolution.","lead":"This paper tests a popular statistical method for estimating how 'fractal' a chaotic system is, and finds the method's core assumption often fails. The result matters because the method is widely used on climate and weather data, so many published numbers may depend on arbitrary analysis choices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Cantor-set proof of non-regular variation asserts an exact b=1/3 ratio for every point and radius; that identity is not established and, as stated, it is the regular-variation limit, so the analytical core of the central claim is unsupported as written.","rationale":"The reader's weakest_assumption concerns finite-sample bias in the numerical R(r) estimates; the reader's rationale also notes that the Cantor derivation does not cleanly prove the failure. I single out the Cantor derivation because it is the only analytical support for the paper's central mathematical claim that singular measures on non-integer-dimensional sets are typically not regularly varying. The printed b=1/3 identity is not true for arbitrary points and radii, and taken literally it states a regular-variation limit, making the proof internally inconsistent as written. This concern is load-bearing because if the Cantor counterexample is removed, the remaining evidence is purely numerical and inherits the finite-orbit caveat already flagged by the reader. A targeted exact computation of µ(B_r(ζ)) would settle whether the counterexample survives with corrected limiting sequences. If it does, the paper's conclusion can remain conditional with a repaired derivation; if it does not, the central claim loses its only rigorous anchor. Therefore no change to the reader's CONDITIONAL verdict is needed.","tokens_in":16047,"tokens_out":10251,"duration_ms":102202,"concrete_test":"Implement an exact recursive computation of µ(B_r(ζ)) for the (1/2,1/2) Bernoulli measure on the middle-third Cantor set, evaluating R_b(r) = µ(B_{br}(ζ))/µ(B_r(ζ)) for b=1/2 and b=1/3 at radii r=3^{-N}, r=2·3^{-N}, and r=3^{-N+1}/2 for N=1,...,40, using both ζ=0 (an endpoint) and ζ drawn from the invariant measure. If the b=1/2 ratios along r=2·3^{-N} tend to 1 while the b=1/3 ratios along a nested sequence tend to 1/2, the non-regularity conclusion survives and the derivation can be repaired by stating the sequences explicitly; if both ratios tend to b^{log2/log3} for generic ζ, the paper's central counterexample fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mathematical demonstration in Section III.A that the (1/2,1/2) Bernoulli measure on the middle-third Cantor set is not regularly varying rests on two ratio computations. The b=1/2 endpoint construction is only an obstruction if it is combined with a second value of b giving a different index; the text supplies b=1/3 and asserts, due to self-similarity, that µ(B_{r/3}(ζ))/µ(B_r(ζ)) = 1/2 = (1/3)^{log2/log3} for any point ζ and any radius. This identity is not established: for a generic point ζ, the ball B_r(ζ) eventually crosses gaps or neighbouring cylinders, so the ratio depends on ζ and on r modulo the ternary grid; the equality holds exactly only for special points such as endpoints and selected radii, not for 'any point, any radius'. Moreover, the b=1/3 equality, if valid, states exactly the regular-variation limit with index log2/log3, so by itself it supports rather than refutes Eq. (5). The contradiction only emerges if the b=1/2 construction produces ratio tending to 1 along a decreasing sequence of radii (forcing γ=0) while b=1/3 produces ratio tending to 1/2 along another sequence (forcing γ=log2/log3). The manuscript does not specify such sequences or prove convergence of the b=1/2 ratio; it only exhibits one radius r=2/3^{N+1} for each N. Since Eq. (5) is a limit as r→0 for each fixed b, one must show the ratio along the chosen sequence converges. As written, therefore, the analytic centerpiece does not cleanly prove non-regular variation. The numerical R(r) oscillations in Figures 3-5 could fill the gap only if the finite-orbit recurrence proxy is shown to be unbiased; the paper itself notes that the method may be sensitive to data order. Hence the claim that singular measures on non-integer-dimensional sets are typically not regularly varying lacks a rigorous demonstration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates the mathematical assumptions behind the Generalized Pareto Distribution-based (GPD) estimator of the local dimension, which the authors call the Exceedance-Based Dimension (EBD) algorithm. It argues that the EBD algorithm requires, in addition to ergodicity and stationarity, the existence of the local dimension almost everywhere and regular variation of the invariant measure. The central claim is that singular measures supported on sets of non-integer dimension are typically not regularly varying, so that EBD estimates become resolution dependent. Section III.A attempts to prove non-regular variation for the (1/2,1/2) Bernoulli measure on the middle-third Cantor set; Sections III.B–III.D present numerical evidence for the fat Cantor set, the Hénon map, and the solenoid, and Section IV discusses the ambiguity of index-based extremal index estimators for continuous-time processes sampled at fixed time steps. The paper concludes that the EBD algorithm should not be applied without additional checks and that the extremal index should not be included in the GPD exceedance distribution.","tokens_in":16442,"tokens_out":29045,"duration_ms":237297,"significance":"The manuscript addresses a practically important and timely question: the GPD-based local dimension estimator is widely used in climate and dynamical-systems applications, and a demonstration that its output can be resolution dependent would be a significant cautionary result. The paper has several concrete strengths: the exact dimension of the solenoid is known and the approximation in Appendix B is validated against it over 16 orders of magnitude; the code and data are publicly available; and the discussion of the extremal index in the peaks-over-threshold framework, including the worked example in Section II, is instructive. If the analytic and numerical evidence is brought to the required standard, the paper would be a valuable contribution. However, the central analytic proof in Section III.A is incomplete as written, and the numerical evidence for the non-regularity of the Hénon, Lorenz, and solenoid measures rests on an uncontrolled sampling proxy.","major_comments":[{"comment":"The proof that the (1/2,1/2) Bernoulli measure on the middle-third Cantor set is not regularly varying is incomplete as written. The b=1/2 construction with r=2/3^{N+1} yields ratio 1 only when ζ is an endpoint in the specific configuration described; the assertion that such radii “can be found for any point ζ ∈ C∞” is not proved and is not a consequence of the displayed ball-measure formula. The b=1/3 calculation, if the self-similarity ratio µ(B_{r/3}(ζ))/µ(B_r(ζ)) = 1/2 is accepted, is exactly the regular-variation value b^{log2/log3} and therefore by itself supports Eq. (5) with γ = log2/log3 rather than refuting it; the contradiction requires a second sequence of radii with a different limiting ratio for the same ζ, and the text does not supply such a sequence for a fixed generic ζ. Since this is the only analytic demonstration of non-regular variation, the central claim needs a corrected proof (for example, using the ternary structure to show that for a typical ζ the ratio for b=1/2 along r=3^{-n} does not converge to 2^{-γ}).","section":"Section III.A, Eq. (5)"},{"comment":"The statement that the solenoid provides a “counterexample” to the claim in Ref. 23 is an overstatement. The evidence is the numerical evaluation of the approximate formula (B3), which rests on three simplifying assumptions (uniform measure along branches, straight branches, and total length 2π) whose error is not quantified. The fact that the slope of the approximate log-measure plot matches the known Hausdorff dimension over 16 orders of magnitude is a useful validation of the approximation, but it does not prove that the true invariant measure of the solenoid is not regularly varying. The authors should either prove a rigorous lower bound on the oscillations of µ_S(B_{r/2}(ζ))/µ_S(B_r(ζ)) using the exact solenoid dynamics, or describe the result as numerical evidence that challenges the applicability of Ref. 23.","section":"Section III.D and Appendix B"},{"comment":"The numerical computation of R(r) = µ(B_{r/2}(ζ))/µ(B_r(ζ)) by retaining the 5000 closest recurrences of a single finite orbit and treating them as samples of the invariant measure is an uncontrolled approximation; the paper itself notes that the method “might be sensitive to the order in which the data is distributed.” No convergence check with respect to trajectory length is reported, and the 1000-point averages in Figures 5 and 8 do not by themselves rule out finite-sample bias. Consequently, the observed oscillations for Hénon, Lorenz 63, and Lorenz 96, and the claim in Section V that no singular measure on a non-integer-dimensional attractor was regularly varying, are not fully established. The authors should add a finite-sample robustness test (for example, varying the number of recurrences and comparing with direct Monte Carlo estimates of µ(B_r)) or temper the conclusions.","section":"Section III.B and Figures 3–5"}],"minor_comments":[{"comment":"The displayed ratio formula has a missing closing parenthesis in µ(B_r(ζ)), and the preceding sentence contains a grammatical typo: “The measure of ball a radius r.”","section":"Section III.A"},{"comment":"The Cantor ternary function C(·) is used in the ball-measure formula but is never defined; please define it.","section":"Section III.A"},{"comment":"The phrase “a nowhere dense set of isolated points” is contradictory for a Cantor set; the intended meaning is presumably “a nowhere dense set with no isolated points.”","section":"Section III.B"},{"comment":"The figure caption contains the informal instruction “(Change x for ζ in figure)”; please remove it and make the notation consistent.","section":"Figure 1"},{"comment":"The statement that for Axiom A systems the local dimension is constant µ-a.e. and equal to the Hausdorff dimension needs a qualifier (for example, for the SRB measure); as written it is too broad.","section":"Section III.D"},{"comment":"The notation P_{φ_k} and the definition of the Poincaré section would benefit from a short explanatory paragraph or a diagram; the current derivation is dense.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is a useful cautionary study, but the central analytic proof needs repair and the numerical evidence should be framed more carefully. I do not see circularity or a citation-pattern problem; the self-citation to the code repository is appropriate. The paper’s fit with a math.DS journal depends on whether the authors can provide a rigorous proof of non-regular variation for the Cantor measure and a rigorous or clearly qualified treatment of the solenoid claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nTwo things you should know. First, the central warning is sound: the GPD-based (EBD) local dimension estimator requires regular variation of the invariant measure, and that property is rarely checked and often fails for singular measures on fractal attractors. If true, many published local dimensions and extremal indices are threshold- or resolution-dependent. Second, the paper's flagship analytic demonstration—the Cantor set—is not written carefully enough to prove the point as it stands, though the gap is repairable.\n\nWhat's genuinely new is the systematic focus on regular variation as the unexamined assumption, and the across-system survey (Cantor, fat Cantor, Hénon, solenoid, Lorenz 63/96, Hénon-Heiles). The solenoid approximate measure in Appendix B is a nice touch: it reproduces the oscillation pattern and matches the known dimension independently, which bolsters the claim beyond the finite-orbit recurrence proxy. The fat Cantor case is a useful control, showing the property is not just about fractality. The extremal index section on sampling-frequency ambiguity is a practical point that deserves attention from climate scientists. The paper ships code and data.\n\nThe soft spots are real but not fatal. The Cantor proof in Section III.A is confusing: the assertion that the b=1/3 ratio equals 1/2 for any point and any radius is not true as stated—it holds for some radii but not for radii that straddle gaps. The b=1/2 construction only displays a single radius per level, not a sequence, and the text's claim that it forces gamma=0 works only for a particular endpoint choice. The contradiction actually emerges by taking r_n=1/3^n for b=1/3 (ratio 1/2) and r_n=2/3^{n+1} at left endpoints for b=1/2 (ratio 1); the authors should write it that way. As written, a sharp reader cannot reconstruct the proof.\n\nSecond, the numerical R(r) oscillations are computed from the 5000 closest recurrences of a finite orbit, and the paper itself concedes the method may be sensitive to data order. So the Hénon and Lorenz evidence is suggestive, not conclusive, on its own. The solenoid approximation mitigates this for that system, but the approximation is unquantified.\n\nThird, the conclusion that 'we could not find any regularly varying measure in systems having a dimension different from an integer' is stronger than the experiments support; it's a reasonable conjecture, not an established fact.\n\nWho is this for? Anyone using EBD/GPD local dimensions or extremal index estimators on real data—especially climate and atmospheric science. They need to read this before trusting published numbers.\n\nBottom line: it deserves a serious referee. The core message is likely correct and practically important. The Cantor derivation needs rewriting, the numerical proxy caveat needs honest treatment, and the generality claims should be scaled back. I would accept with major revisions if the authors fix the proof exposition.","headline":"A likely correct and practically important caution about GPD-based local dimensions, but the Cantor proof is not clean and the numerical evidence has a sampling caveat.","tokens_in":17002,"tokens_out":12170,"would_cite":true,"duration_ms":93631,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37C45","60G70","37A50"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper demonstrates that the standard GPD-based local dimension estimator fails when the invariant measure is not regularly varying, which is the norm for singular measures on non-integer-dimensional sets.","keywords":["extreme value theory","local dimension","Generalized Pareto Distribution","regular variation","fractal dimension","extremal index","dynamical systems","attractor geometry"],"falsifier":"Run the EBD algorithm on the Hénon attractor with trajectory lengths spanning at least three orders of magnitude and thresholds $q=0.98$, $0.99$, and $0.995$ at a fixed set of reference points; if the estimated local dimension is the same across all these choices at small radii, the central claim that the lack of regular variation makes estimates resolution dependent is wrong.","tokens_in":15813,"feed_emoji":"📉","tokens_out":13136,"duration_ms":109382,"temperature":0.7,"pith_summary":"The paper tries to establish when the standard exceedance-based dimension (EBD) estimator actually works. That estimator assumes ergodicity, stationarity, existence of the local dimension almost everywhere, and regular variation; the paper argues the last condition is the least justified. Regular variation means the scaling ratio $\\mu(B_{br}(\\zeta))/\\mu(B_r(\\zeta))$ settles to a pure power law $b^\\gamma$ as the radius shrinks, and when it fails the excess distribution does not converge to the exponential law that the estimator relies on. For singular measures supported on non-integer-dimensional sets, the normal situation for the Hénon, solenoid, Lorenz 63, and Lorenz 96 attractors, the paper shows this condition typically fails, making the estimated local dimension oscillate with ball radius and depend on threshold and trajectory length. If true, this matters because the GPD approach is widely applied to real data, where the same output could change with arbitrary analysis choices.","feed_headline":"GPD dimension estimates shift with scale on fractal attractors","feed_subtitle":"The standard estimator assumes a smooth scaling that fractal measures usually lack, so outputs depend on radius and data length.","key_machinery":"The load-bearing object is the regular variation property of the invariant measure, written $\\mu(B_{br}(\\zeta))/\\mu(B_r(\\zeta)) \\to b^\\gamma$ as $r \\to 0$ for every $0