REVIEW 3 major objections 5 minor 4 references
Global Permutation Entropy
T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read 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
desk verdict A genuinely new index with a working implementation, but the evidence for its advantages over permutation entropy is thinner than the paper claims. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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
What would settle it
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.
Extended reading notes
Core claim
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
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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).
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 (3)
- [Section 3.1, Eq. (1), Section 4] 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 3.2, Figure 5] 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 3.3, Figure 10] 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.
minor comments (5)
- [Section 4] 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.
- [Figure 3] Add confidence intervals or error bars; the text mentions 100 realizations but the plot shows only means.
- [Section 3.2] 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.
- [Eq. (2)] 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 2.1] Minor typo: 'explicitely' should be 'explicitly'.
Circularity Check
No significant circularity: GPE is a direct empirical entropy with no fitted parameters; computational feasibility rests on independent algorithmic results.
full rationale
GPE is defined in Eq. (1) as the Shannon entropy of the empirical distribution over all C(n,k) ordinal patterns. No parameter is fitted to a target, and no prediction is derived from a fit; the measure is a straightforward aggregation of permutation-pattern frequencies. The faster-convergence claim in Section 3.1 is an empirical observation supported by Monte Carlo means, and the statement that GPE has a larger sample size is an informal interpretation, not a mathematical identity that forces the result. The computational feasibility claims cite prior algorithmic work; although DV24 is co-authored by two of the present authors, the paper also cites independent prior work EL21 and BL24, and the dominant O(n^2 log^4 n) term for the 6-profile is attributed to BL24. Thus the self-citation is not load-bearing. The paper's central measure and experiments are defined without importing its conclusions as assumptions, so no circular step is present.
Assumptions & free parameters
free parameters (3)
- GPE order k =
2, 3, 4, 5, 6 in experiments
- window size w =
Range from about half period to 1.5P+1; 30-60 for P=60
- PE delay tau and delay averaging set =
tau = 1, 10, 20 and averages over 1-10
assumptions (4)
- domain assumption For i.i.d. continuous random variables, all k-permutations are equally likely, so the true normalized GPE and PE are 1.
- domain assumption The counting algorithms of [EL21], [BL24], and [DV24] correctly and efficiently count all permutation patterns of order up to 6, together with the tie-breaking rule 'values are ordered according to their time of appearance'.
- ad hoc to paper The average sliding-window entropy curve for a periodic signal has a minimum near half the period, which is used to choose GPE's window size.
- domain assumption The Gaussian-noise corruption model in the synthetic experiments is a relevant proxy for structural information in time series.
Cite this review
Pith. "Pith review of Global Permutation Entropy." pith.science (2026). https://pith.science/paper/6UNOPF6F
@misc{pith2026250819955,
author = {Pith},
title = {Pith review of: Global Permutation Entropy},
year = {2026},
howpublished = {\url{https://pith.science/paper/6UNOPF6F}},
note = {Machine review of arXiv:2508.19955}
}
read the original abstract
Permutation Entropy, introduced by Bandt and Pompe, is a widely used complexity measure for real-valued time series that is based on the relative order of values within consecutive segments of fixed length. After standardizing each segment to a permutation and computing the frequency distribution of these permutations, Shannon Entropy is then applied to quantify the series' complexity. We introduce Global Permutation Entropy (GPE), a novel index that considers all possible patterns of a given length, including non-consecutive ones. Its computation relies on recently developed algorithms that enable the efficient extraction of full permutation profiles. We illustrate some properties of GPE and demonstrate its effectiveness through experiments on synthetic datasets, showing that it reveals structural information not accessible through standard permutation entropy. We provide a Julia package for the calculation of GPE at `https://github.com/AThreeH1/Global-Permutation-Entropy'.
Reference graph
Works this paper leans on
-
[74]
Conditional entropy of ordinal patterns
[UK14] Anton M Unakafov and Karsten Keller. “Conditional entropy of ordinal patterns”. In: Physica D: Nonlinear Phenomena 269 (2014), pp. 94–102. [ZSW13] Xiaojun Zhao, Pengjian Shang, and Jing Wang. “Measuring information interactions on the ordinal pattern of stock time series”. In: Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 87.2 (...
work page 2014
-
[197]
Counting Permutation Patterns with Multidimensional Trees
11 [BL24] Gal Beniamini and Nir Lavee. “Counting Permutation Patterns with Multidimen- sional Trees”. In: arXiv preprint arXiv:2407.04971 (2024). [BP02] Christoph Bandt and Bernd Pompe. “Permutation entropy: a natural complexity measure for time series”. In: Physical review letters 88.17 (2002), p. 174102. [DV24] Joscha Diehl and Emanuele Verri. “Efficien...
work page Pith review arXiv 2024
-
[2025]
A new kind of permutation entropy used to classify sleep stages from invisible EEG microstructure
[Ban17] Christoph Bandt. “A new kind of permutation entropy used to classify sleep stages from invisible EEG microstructure”. In: Entropy 19.5 (2017), p
work page 2017
-
[2668]
Scinet: Time series modeling and forecasting with sample con- volution and interaction
[Liu+22] Minhao Liu et al. “Scinet: Time series modeling and forecasting with sample con- volution and interaction”. In: Advances in Neural Information Processing Systems 35 (2022), pp. 5816–5828. [LK17] Douglas J Little and Deb M Kane. “Variance of permutation entropy and the influ- ence of ordinal pattern selection”. In: Physical Review E 95.5 (2017), p...
work page 2022
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.