{"id":"2c595fe3-e446-4c20-bc31-8fea0409b8dd","arxiv_id":"2508.19955","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Global Permutation Entropy is a new, efficiently computable entropy index over all ordinal patterns of a time series and is claimed to converge faster and detect noise changes more robustly than standard permutation entropy in synthetic tests.","lead":"This paper introduces Global Permutation Entropy, a time-series complexity measure that counts the relative order of all groups of k points in a window, not just nearby points, using fast combinatorial counting. In synthetic experiments, it claims faster convergence to maximum entropy for random data and earlier detection of noise changes than standard permutation entropy, with less parameter tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GPE's faster-convergence claim rests on an unproven 'larger sample size' intuition; effective sample size is O(n), not C(n,k), and may be smaller than PE's.","rationale":"The central claim is that GPE is practically computable and reveals structural information not accessible via PE. The mathematical definition is sound and the computational complexity relies on published algorithms. However, the three empirical advantages depend on a statistical intuition that the paper does not justify. The most load-bearing weakness is the 'larger effective sample size' claim: overlapping k-subsets are heavily dependent, and the effective sample size of a U-statistic is O(w), not C(w,k). For k=2 the exact calculation shows PE has a larger effective sample size, and the paper excludes that case; for k>2 no theory or error-barred experiments support the claim. A simple simulation can settle it. Additionally, Section 3.2's AUC comparison selects the best order/delay on the test data before reporting, which inflates GPE's apparent advantage; this is a secondary but serious methodological issue. Because the core definition and implementation are contributions, the appropriate verdict is CONDITIONAL: the paper should either prove or empirically demonstrate the convergence and noise-detection advantages with proper validation. The reader's weakest_assumption (dependence breaking the sample-size intuition) is exactly the same concern, so I agree.","tokens_in":8426,"tokens_out":17812,"duration_ms":195157,"concrete_test":"Simulate 10,000 i.i.d. standard normal series for lengths w = 20, 50, 100 and orders k = 3, 4, 5. Compute GPE(k) and PE(k; τ=1) for each series and report mean, variance, and MSE relative to 1. If for any k>2 GPE's MSE is not below PE's (or its mean is not closer to 1), the faster-convergence advantage in Sec. 3.1 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 and Section 4 claim GPE converges faster to 1 than PE for i.i.d. data because it uses all C(w,k) subsets, 'effectively increasing the sample size.' This is not established and is likely false. For i.i.d. continuous data, GPE pattern counts are (non-symmetric) U-statistics over all k-subsets; their variance decays as O(1/w), not O(1/C(w,k)), so the effective sample size is O(w), with a constant depending on k. For k=2, an exact covariance calculation gives m_eff(GPE) ≈ 0.75w vs m_eff(PE) ≈ 3w, so PE converges faster—consistent with the paper's 'except for order 2' caveat. For k=3, the asymptotic leading constant is not obviously favorable to GPE; the paper provides no variance/MSE analysis. Figure 3 shows only means over 100 realizations, with no error bars, so the claim of faster convergence is unsupported. The same dependence issue undermines the 'finer granularity' argument in Section 4.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Global Permutation Entropy (GPE), defined as the Shannon entropy of the empirical distribution of ordinal patterns over all strictly increasing index subsets of size k (Eq. 1). It uses recent permutation-pattern counting algorithms to make computation feasible for k up to 6 and provides a Julia package. The authors claim that GPE offers advantages over classical permutation entropy (PE): faster convergence to 1 for i.i.d. data (k>2), quicker detection of a sudden noise increase in periodic signals, and greater robustness to window-size choice with less parameter tuning. These claims are supported by synthetic experiments on convergence (Sec. 3.1), noise detection (Sec. 3.2), and linearly increasing noise (Sec. 3.3).","tokens_in":8722,"tokens_out":8369,"duration_ms":94987,"significance":"The definition of GPE is conceptually clean and parameter-free beyond the order k and window size w. The computational contribution is real: the paper correctly identifies that recent algorithmic advances make full permutation profiles up to order 6 practical, and the Julia package is a concrete, reusable artifact. However, the claimed empirical advantages over PE are not yet established. The 'larger sample size' rationale in Secs. 3.1 and 4 is statistically inaccurate as stated, and the experimental evaluations in Secs. 3.2 and 3.3 use post hoc parameter selection that can inflate apparent performance. If the authors supply a proper variance analysis and prespecified evaluation protocols, GPE could be a useful new tool for time-series analysis.","major_comments":[{"comment":"The claim that GPE converges faster because it uses all C(w,k) subsets, 'effectively increasing the sample size,' is not supported by statistical reasoning. For i.i.d. continuous data, the pattern counts are U-statistics of degree k; the variance of p(σ) decays as O(1/w), not O(1/C(w,k)). For k=2, an exact calculation gives Var(p̂_GPE)≈1/(3w) and Var(p̂_PE)≈1/(12w), so the effective sample sizes are ≈3w/4 and ≈3w; PE converges faster, consistent with the paper's exception. For k≥3 no variance/MSE analysis is provided, and Figure 3 reports only means without error bars or confidence intervals. The faster-convergence advantage therefore remains unverified, and the 'finer granularity' argument in Section 4 inherits the same problem.","section":"Section 3.1, Eq. (1), Section 4"},{"comment":"The caption states that for each window size 'the highest average AUC among orders k=2,3,4 is selected' for GPE, and for PE the highest among orders, delays, and the delay-average. Selecting the best-performing configuration on the test data and reporting that AUC is an optimistic estimate of discriminative performance; it does not provide a valid comparison between methods. Because the PE search space is larger, the selection bias may also differ across methods. Please report results with parameters fixed a priori or use a separate validation split / cross-validation.","section":"Section 3.2, Figure 5"},{"comment":"The conclusion that GPE 'requires significantly less parameter tuning' is based on a post hoc search over PE delays and window sizes ('Through extensive experimentation and fine-tuning, we found...'), with no prespecified grid, selection rule, or correction for multiple comparisons. As reported, the comparison is not reproducible and does not establish robustness. Provide the full parameter grid and a defined selection criterion (e.g., best on a validation set) for both methods.","section":"Section 3.3, Figure 10"}],"minor_comments":[{"comment":"The statement that 'at most one histogram element changes for PE' should be 'at most two' when the removed and added ordinal patterns are of different types.","section":"Section 4"},{"comment":"Add confidence intervals or error bars; the text mentions 100 realizations but the plot shows only means.","section":"Figure 3"},{"comment":"The methodology states that 95% confidence intervals are averaged over 100 Monte Carlo runs, but the AUC figures do not display them; include them in the plots.","section":"Section 3.2"},{"comment":"The condition C(w,k) ≫ k! is necessary but, because of strong dependence among the subsets, not sufficient for reliable entropy estimation; the text should acknowledge this.","section":"Eq. (2)"},{"comment":"Minor typo: 'explicitely' should be 'explicitly'.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The core definition and computational contribution are solid, but the empirical claims need substantial strengthening. The convergence claim appears to rest on a misconception about effective sample size; a careful U-statistic variance analysis or suitably designed simulations with error bars could settle it. The noise-detection and robustness comparisons should use prespecified parameter choices or cross-validation. The paper may be publishable after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GPE is worth knowing: it replaces consecutive/equally spaced ordinal patterns with all increasing index subsets, giving a clean definition, and the authors show it can be computed in near-quadratic time using recent permutation-pattern counting algorithms, with a Julia package. That is a real contribution, and the outlook sketch of a corner-tree entropy is a nice idea.\n\nThe problems are in the empirical claims. The \"larger sample size\" argument in Sections 3.1 and 4 is not established. The C(w,k) subsets are heavily dependent; variance of a U-statistic over all k-subsets decays like 1/w, not 1/C(w,k), so the effective sample size is O(w), not C(w,k). The paper even admits order 2 behaves differently, which is a hint that the intuition is unreliable. Figure 3 plots means only, no error bars or variance analysis, so \"converges faster\" is unsupported for k>2.\n\nThe noise detection comparison (Sec. 3.2) selects the highest AUC over orders and delays on the same test data before reporting, which inflates the apparent advantage. No multiplicity correction is mentioned. That makes \"GPE clearly outperforms PE\" not trustworthy as stated.\n\nSection 3.3 is exploratory and visual; the claim of robustness across window sizes is based on plots, not a metric. The window-size heuristic for GPE (minimum around half period) is reasonable but unproven.\n\nOn the positive side, the definition and normalization are sound, the computational complexity statements are specific, and the code is public. I could not execute it, so reproducibility is unverified. Citation of [DV24] is a preprint by two of the authors but it's a separate algorithmic result, not circular. Minor: the package link appears not archived/versioned.\n\nBottom line: this is a methods paper with a useful new measure and a working tool, but the comparative claims need stronger statistical backing. It deserves a serious referee; I would not accept it as is. The referee should ask for error bars or MSE curves, a principled treatment of AUC selection, and ideally a dependence-aware analysis of variance.","headline":"A genuinely new index with a working implementation, but the evidence for its advantages over permutation entropy is thinner than the paper claims.","tokens_in":9197,"tokens_out":2554,"would_cite":true,"duration_ms":26778,"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":"Global Permutation Entropy extends permutation entropy to all ordinal patterns of a given length, and the paper shows it converges faster to the maximal value, detects sudden noise increases sooner, and needs less parameter tuning than stan","keywords":["global permutation entropy","permutation entropy","ordinal patterns","permutation pattern counting","corner trees","time series complexity","noise detection","Shannon entropy"],"falsifier":"Take i.i.d. standard-normal windows of sizes w=20,50,100 and Monte Carlo 10^4 replicates; compare the empirical bias and variance of GPE(k) and PE(k;1) for k=3,4,5 at matched window lengths. If GPE(k) is not consistently closer to 1 or lower-variance than PE(k), the central faster-convergence claim fails. Separately, on the two-segment noisy periodic signal, if GPE's AUC advantage over PE disappears when PE's delay is oracle-selected per window, the claimed robustness is a tuning artifact.","tokens_in":8364,"feed_emoji":"📈","tokens_out":7608,"duration_ms":77964,"temperature":0.7,"pith_summary":"Global Permutation Entropy (GPE) is a proposal to measure time-series complexity by the Shannon entropy of the empirical distribution over every ordinal pattern of length k that appears in any increasing index subset of the data, not just the consecutive blocks used by permutation entropy (PE). The paper argues that this change is computationally feasible up to order 6 thanks to recent permutation-pattern-counting algorithms based on corner trees, and that it yields concrete benefits on synthetic data: GPE approaches the maximal entropy value 1 faster than PE for i.i.d. sequences of order k>2, detects a sudden noise increase in short periodic signals more quickly, and remains informative across window sizes and orders where PE saturates or needs careful delay tuning. If these findings hold, GPE offers a largely parameter-free alternative to PE for exploratory analysis of short or structured signals. The authors provide an implementation of the 2- through 6-profiles.","feed_headline":"Global permutation entropy spots noise changes faster than PE","feed_subtitle":"Counting all ordinal patterns, not just consecutive ones, makes the entropy measure converge faster and need less tuning.","key_machinery":"The load-bearing object is the full permutation profile of order k: for each permutation σ in S_k, the number of strictly increasing index tuples (i_1<...<i_k) whose values have relative order σ, divided by C(n,k). GPE is the Shannon entropy of that profile. The algorithmic enabler is the corner tree—a recursive combinatorial structure that encodes occurrence counts of permutation patterns—together with its generalizations (pattern trees and tree double posets), which reduce naive O(n^k) pattern counting to O(n log n) for orders 2–3, O(n^{7/4} log^2 n) for order 5, and O(n^2 log^4 n) for order 6. In a sliding window, PE changes at most one histogram bin per step, while GPE can change many bi","core_discovery":"The central claim is that the global permutation profile—counts of all k! permutations over all C(n,k) increasing index tuples, normalized by C(n,k)—carries structural information that the consecutive-pattern PE profile misses, and that the entropy of this profile is now practical to compute. Concretely, the paper demonstrates three advantages. For pure i.i.d. data with k>2, GPE(k) converges to 1 faster than PE(k) on the same window, because its histogram has C(window,k) entries rather than window−k+1. For a periodic signal that abruptly becomes noisier, sliding-window GPE separates the two regimes with higher AUC than PE for short windows and moderate noise levels. For a periodic signal who","pith_inferences":["A step the paper leaves open is theoretical: modeling GPE(k) on i.i.d. data as a functional of a U-statistic over dependent k-subsets could yield explicit bias and variance bounds, converting the empirically observed faster convergence into a theorem.","For k=2, GPE is determined entirely by the relative order of all pairs—essentially an entropy over ascents and inversions—so its exact behavior on monotone, periodic, and random signals can likely be characterized in closed form, offering a clean test of the general claims.","The corner-tree entropy proposed in the conclusion could be computed in O(n log n); testing whether it preserves GPE's noise-detection gains would show whether the full profile is necessary or a cheaper projection suffices.","Whether GPE's synthetic advantages transfer to real EEG, finance, or forecasting data remains open; the authors explicitly list finding such real datasets as future work, and the released implementation makes that test immediate."],"forward_implications":["GPE(k) for k up to 6 can be computed in near-quadratic time, so the global ordinal profile becomes a practical tool on time series of the lengths that PE already handles.","For noisy periodic signals, sliding-window GPE requires no delay parameter; a window roughly half a period to one period long gives informative entropy curves, simplifying exploratory analysis.","Averaging sliding-window GPE over window sizes yields a curve whose minimum sits near half the period, giving a data-driven period estimate.","Because GPE's entropy values move by many small steps rather than by coarse 1/window increments, entropy traces of short signals can resolve structure at a finer scale than PE."],"supporting_citations":[{"why":"Defines permutation entropy and the sliding-window setup, the baseline GPE is compared against throughout.","marker":"[BP02]"},{"why":"Introduces corner trees and shows certain permutation patterns can be counted in O(n log n), the algorithmic basis for the 2-, 3-, and 4-profiles.","marker":"[EL21]"},{"why":"Adds double-poset counting that extends the corner-tree method and is used for the 5- and 6-profiles.","marker":"[DV24]"},{"why":"Provides pattern-tree techniques that make full 5-, 6-, and 7-profiles computable in near-quadratic time, the key feasibility result for GPE.","marker":"[BL24]"},{"why":"The authors' Julia implementation of the 2- through 6-profiles used in the experiments.","marker":"[Avh+25]"},{"why":"Supplies the recommendation to average PE over delays, which the paper adopts as the best-tuned PE baseline in Section 3.3.","marker":"[Ban17]"}],"fun_headline_variants":["GPE beats PE at spotting noise shifts in short windows","All-pattern entropy sees noise changes that consecutive PE misses","Count every pattern, not just runs: entropy sharpens change detection","Global permutation entropy: faster convergence, sharper noise detection","Non-consecutive patterns boost entropy's noise sensitivity"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The convergence and granularity advantages assume that the C(w,k) overlapping index subsets inside a window act as a larger, effectively usable sample for entropy estimation, even though these subsets are strongly dependent and the paper supplies no theory or resampling evidence for that assumption.","fun_headline_variants_meta":{"raw":{"variants":["GPE beats PE at spotting noise shifts in short windows","All-pattern entropy sees noise changes that consecutive PE misses","Count every pattern, not just runs: entropy sharpens change detection","Global permutation entropy: faster convergence, sharper noise detection","Non-consecutive patterns boost entropy's noise sensitivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001073,"raw_usage":{"total_tokens":4295,"prompt_tokens":677,"completion_tokens":3618,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":3537}},"tokens_in":421,"tokens_out":3618,"duration_ms":31565,"temperature":1.0,"reasoning_tokens":3537,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:20:21.430996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take i.i.d. standard-normal windows of sizes w=20,50,100 and Monte Carlo 10^4 replicates; compare the empirical bias and variance of GPE(k) and PE(k;1) for k=3,4,5 at matched window lengths. If GPE(k) is not consistently closer to 1 or lower-variance than PE(k), the central faster-convergence claim fails. Separately, on the two-segment noisy periodic signal, if GPE's AUC advantage over PE disappears when PE's delay is oracle-selected per window, the claimed robustness is a tuning artifact.","supporting_citations":[],"review_version":1}