A Benchmark for Entropy Estimators

TL;DR

This paper evaluates how well different entropy estimation methods work on complex systems with known entropy values, revealing which methods are most accurate.

Contribution

The study introduces a benchmark using certified entropy values from dynamical systems to evaluate entropy estimators.

Findings

Approximate Entropy and symbolic methods provided accurate estimates within error bounds across all systems.

Sample Entropy systematically underestimated entropy, while Permutation Entropy showed large biases for certain maps.

Abstract

This study assessed the performance of several entropy estimators for numerical time series and symbolic data on non-trivial one-dimensional dynamical systems whose Kolmogorov–Sinai entropy is known with certified accuracy: recent computer-assisted proof techniques provide rigorous values together with explicit error bounds. We considered four classes of interval maps, including piecewise expanding maps with and without a Markov partition and an intermittent Pomeau–Manneville map, and generated long orbits for each system. We then compared the certified entropy with the output of widely used estimators: Approximate Entropy, Sample Entropy, Permutation Entropy, a symbolic Plug-In estimator of the entropy rate, and the Non-Sequential Recursive Pair Substitution (NSRPS) method (the latter two with Grassberger-type bias correction). Our experiments reveal substantial, dynamics-dependent differences in accuracy and robustness. In particular, Approximate Entropy and the symbolic methods (Plug-In and NSRPS) consistently yielded estimates within the rigorous error bars across all systems, whereas Sample Entropy showed a marked systematic underestimation, and Permutation Entropy exhibited large biases, especially for expanding maps without a Markov partition. The resulting benchmark provides a quantitative testbed for evaluating entropy estimation techniques in deterministic dynamical systems.

Figures