Pairwise-Agreement Symbolic Entropy Saturates by Construction: Implications for Delay Selection

A family of methods selects the phase-space reconstruction delay τ as the point where a pairwise-agreement symbolic entropy H(m,τ) stops growing, on the premise that this onset marks the decorrelation delay. We show that the saturation is built into the construction rather than discovered in the data, so that the premise fails for the pairwise-agreement symbol. The agreement symbol is invariant under complementation of the thresholded sign vector, so the symbol space has exactly 2m−1 elements, the map from complement orbits to attainable agreement patterns is a bijection, and H(m,τ)≤m−1 for every signal. The plateau is fixed by the symbolization and the threshold-crossing probability alone: for independent Bernoulli(p) signs it equals H∞(m,p)≤m−1, with equality if and only if p=1/2, and an independent control confirms this to 0.005 bits for m=2,…,8. Saturation therefore occurs for any input, including i.i.d. noise, and the plateau carries no dynamical information; what remains is the rate of approach, which depends on the marginal and not only on the dependence structure. Translating a record by half its standard deviation leaves the mutual-information curve and its first minimum unchanged to machine precision while moving the criterion’s delay by 27% of its own value. We test the consequences on five chaotic systems at six noise levels, each verified to have a positive largest Lyapunov exponent under two integration steps and two accumulation lengths, and on two machinery datasets. On clean data our low-pass-filtered variant returns delays 1.08–4.69 times the average-mutual-information reference and, under the as-used protocol, carries the largest bias of the six; the filter raises the delay bias by 71% (31.6→54.1 samples) rather than reducing it; and the threshold that reproduces the reference is absent for one of five systems on clean data and for two at 10 dB. Under a harmonized protocol that gives all six estimators one search window, threshold and dimension range, the ordering changes and the variant is no longer the worst, so the ranking is protocol-dependent. On MFPT bearing data the comparison inverts: the most biased estimator under the as-used protocol classifies best (0.917 from repeated grouped cross-validation, 95% cluster interval over records [0.798,0.987]) and the one closest to the reference worst (0.552, [0.408,0.697]), while replacing the per-record (τ,m) by a single pooled pair drops the winner to 0.743, so about a fifth of its score is the record-specific calibration acting as a record fingerprint rather than any property of the delay. All conclusions are bounded to the five systems, six noise levels and two datasets examined. We advise against reading a decorrelation delay off the saturation onset of a pairwise-agreement symbolic entropy; where only classification accuracy is sought it may serve as a black-box feature heuristic, but its output should not be reported as a delay.

Authors

Publication Details

Journal
Entropy
Published
2026-11-18
DOI
https://doi.org/10.3390/e28111272
Primary Topic
Chaos control and synchronization
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pairwise-Agreement Symbolic Entropy Saturates by Construction: Implications for Delay Selection

Jiahang Cui, Zhenmin Zhao, Jianghong Li
1 links
Entropy
Chaos control and synchronization
article

Pairwise-Agreement Symbolic Entropy Saturates by Construction: Implications for Delay Selection

Jiahang Cui, Zhenmin Zhao, Jianghong Li
article en

Abstract

A family of methods selects the phase-space reconstruction delay τ as the point where a pairwise-agreement symbolic entropy H(m,τ) stops growing, on the premise that this onset marks the decorrelation delay. We show that the saturation is built into the construction rather than discovered in the data, so that the premise fails for the pairwise-agreement symbol. The agreement symbol is invariant under complementation of the thresholded sign vector, so the symbol space has exactly 2m−1 elements, the map from complement orbits to attainable agreement patterns is a bijection, and H(m,τ)≤m−1 for every signal. The plateau is fixed by the symbolization and the threshold-crossing probability alone: for independent Bernoulli(p) signs it equals H∞(m,p)≤m−1, with equality if and only if p=1/2, and an independent control confirms this to 0.005 bits for m=2,…,8. Saturation therefore occurs for any input, including i.i.d. noise, and the plateau carries no dynamical information; what remains is the rate of approach, which depends on the marginal and not only on the dependence structure. Translating a record by half its standard deviation leaves the mutual-information curve and its first minimum unchanged to machine precision while moving the criterion’s delay by 27% of its own value. We test the consequences on five chaotic systems at six noise levels, each verified to have a positive largest Lyapunov exponent under two integration steps and two accumulation lengths, and on two machinery datasets. On clean data our low-pass-filtered variant returns delays 1.08–4.69 times the average-mutual-information reference and, under the as-used protocol, carries the largest bias of the six; the filter raises the delay bias by 71% (31.6→54.1 samples) rather than reducing it; and the threshold that reproduces the reference is absent for one of five systems on clean data and for two at 10 dB. Under a harmonized protocol that gives all six estimators one search window, threshold and dimension range, the ordering changes and the variant is no longer the worst, so the ranking is protocol-dependent. On MFPT bearing data the comparison inverts: the most biased estimator under the as-used protocol classifies best (0.917 from repeated grouped cross-validation, 95% cluster interval over records [0.798,0.987]) and the one closest to the reference worst (0.552, [0.408,0.697]), while replacing the per-record (τ,m) by a single pooled pair drops the winner to 0.743, so about a fifth of its score is the record-specific calibration acting as a record fingerprint rather than any property of the delay. All conclusions are bounded to the five systems, six noise levels and two datasets examined. We advise against reading a decorrelation delay off the saturation onset of a pairwise-agreement symbolic entropy; where only classification accuracy is sought it may serve as a black-box feature heuristic, but its output should not be reported as a delay.

EntropyVol. 28(11)
Key Research and Development Projects of Shaanxi Province
Chaos control and synchronization
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon