Complexity and Regularity Descriptors

Complexity and Regularity Descriptors are explicit characterizations of sequential regularity, pattern persistence, ordinal or symbolic organization, compressibility, recurrence of states, multiscale structure, scaling behavior, fractal-like organization, local divergence, and intrinsic predictability in declared Behavioral Signal evidence. These descriptors are mathematical tools designed to quantify specific features of temporal or sequential data reflecting behavior, yet the terms regularity, irregularity, entropy, randomness, complexity, nonlinearity, fractal scaling, predictability, and chaos are not synonyms. Different families of descriptors operationalize different mathematical properties and can rank the same signals differently. A larger value in one descriptor does not universally mean that the signal is more complex, less regular, healthier, more adaptive, more random, or more behaviorally sophisticated. Each descriptor must be interpreted in the context of its formal definition, assumptions, parameters, and intended operational meaning.

Meaning and Conceptual Boundaries

A Complexity or Regularity Descriptor is a reproducible characterization whose meaning depends on ordered signal structure, pattern construction, state recurrence, scaling across resolutions, symbolic parsing, reconstructed-state geometry, divergence, or prediction behavior under declared parameters and assumptions. Many such descriptors are useful finite-data statistics that do not imply that the underlying behavioral process is nonlinear, deterministic, low-dimensional, fractal, or chaotic. Rather, they quantify operational aspects of the signal or its transformations as defined by the user’s analytical choices.

Regularity concerns recurrence, repeatability, predictability, or restricted pattern variation under a declared definition. Complexity can refer to richer organization across patterns, scales, states, or predictive structure and has no single universally accepted scalar definition. For example, highly periodic signals can be very regular but structurally simple, while white-noise-like signals can be highly irregular yet lack the organized multiscale structure intended by some complexity concepts.

Randomness and stochasticity concern generative uncertainty or inherent unpredictability in signal generation. Nonlinearity indicates the failure of an adopted linear relationship or null model. Deterministic irregularity can arise without stochastic randomness, producing signals that appear irregular but are generated by deterministic rules. Chaos requires a deterministic dynamical interpretation with sensitive dependence on initial conditions and additional evidential conditions. Entropy-like, recurrence, fractal, or divergence descriptors can inform these questions but do not establish them by name alone.

Order sensitivity and domain boundaries are critical. A histogram entropy describes the value distribution without temporal order; spectral entropy describes the spectral energy distribution; time-frequency entropy describes localized spectral-temporal representations; Sample Entropy and Permutation Entropy use sequence-pattern structure; event recurrence describes repeated event occurrence times; state-space recurrence describes returns to neighborhoods in a reconstructed or observed state space. The shared words entropy or recurrence do not make these quantities interchangeable.

Template-Matching Regularity and Entropy-Like Descriptors

A generic delayed template vector is constructed as

where x_i is observation i of the declared ordered scalar series, i is the template-start index, m is the positive integer template dimension, τ is the positive integer sample delay or another explicitly mapped delay, and v_i^(m,τ) is the resulting template vector. The same mathematical vector construction can serve pattern matching or state-space reconstruction but does not automatically represent the complete physical or behavioral state of a person.

Approximate Entropy is a template-matching regularity statistic based on logarithmic match frequencies under declared template dimension, delay, distance metric, tolerance, self-match convention, normalization, and finite-sample rules. It includes self matches in its classical construction and can depend materially on record length and parameter choices, often producing biased estimates if parameters or data length are not carefully controlled.

A representative Sample Entropy abstraction is

where m is the template dimension, r is the declared matching tolerance under a declared distance and scaling convention, B is the eligible count or consistently normalized frequency of distinct template pairs that match at dimension m, and A is the corresponding eligible count or frequency of those pair relationships that remain matched when extended to dimension m+1. Exact counting, delay, edge eligibility, self-match exclusion, and normalization conventions must be documented. Values of A = 0, B = 0, or too few matches can produce infinite, undefined, or unstable results rather than a valid zero.

Sample Entropy is a self-match-excluding template-continuation statistic related to regularity, not an exact universal entropy rate. It is sensitive to parameter choices including m, r, delay, distance metric, support length, amplitude scaling, noise, missingness, and preprocessing. Larger m or smaller r can drastically reduce match counts, and canonical-looking parameter values should not be transferred across modalities without justification.

Fuzzy or soft-matching entropy variants conceptually replace hard binary matches with graded similarity, reducing discontinuity with respect to small distance changes. However, membership function choice, exponent, local/global mean removal, tolerance scaling, and normalization are integral to descriptor identity. Fuzzy Entropy is not one universally standardized formula and requires explicit definition.

Ordinal, Symbolic, and Compression-Based Descriptors

Normalized Permutation Entropy is defined as

where M ≥ 2 is the ordinal-pattern order, Π is the set of admissible ordinal patterns under the declared construction and tie policy, π is one ordinal pattern, p(π) is its empirical relative frequency among eligible delayed windows, and H_PE is Shannon permutation entropy normalized by ln(M!) under the conventional full-pattern normalization. Delay, ties, missing patterns, finite support, and alternative normalizations can change the result.

Permutation Entropy measures the entropy of ordinal-pattern frequencies and thus characterizes local ordering structure rather than amplitude distribution. Embedding order, delay, normalization, overlapping patterns, finite-data coverage of the M! pattern space, and invariance to strictly monotonic amplitude transformations under standard ordinal construction are important considerations. High Permutation Entropy does not prove stochastic randomness or chaos.

Ordinal ties and quantization arise from true plateaus, sensor quantization, rounding, clipping, or repeated states. Tie policies such as order of occurrence, equal-rank patterns, exclusion, deterministic perturbation, or another declared rule produce different ordinal distributions. Random jitter should not be inserted invisibly when exact reproducibility matters.

Symbolic-sequence descriptors arise after explicit symbolization of continuous or discrete evidence. Symbolization can use thresholds, quantiles, ordinal patterns, domain-defined states, or adaptive rules and support word frequencies, block entropy, transition diversity, grammar-like restrictions, and parsing statistics. Symbolization is lossy; its alphabet, thresholds, mapping, missing-symbol rule, and temporal spacing are part of descriptor identity.

Lempel–Ziv-like and compression-related descriptors characterize novelty, phrase growth, or compressibility in a declared symbolic sequence under a specified parsing algorithm and normalization. Raw phrase count, length-normalized variants, alphabet-normalized variants, and actual compressor ratios are not interchangeable. Finite sequence length or symbolization can dominate apparent complexity.

Multiscale Complexity and Scale Profiles

Multiscale complexity analysis evaluates a regularity or complexity statistic across a family of explicitly constructed temporal scales rather than treating one single-scale value as the complete system complexity. Conventional coarse-graining, filtering/downsampling, composite or refined-composite constructions, scale-dependent ordinal methods, and other variants exist. The scale-construction rule materially defines the descriptor output and interpretation.

Increasing scale changes temporal grain, effective sampling interval, frequency content, variance, series length, match counts, ordinal-pattern counts, and estimator stability. Higher scales usually contain fewer effective observations, so an apparently smooth complexity profile can become invalid or unstable where support is insufficient.

Retention of the full complexity-versus-scale profile is recommended, with optional summaries such as mean over a declared scale range, area or sum across valid scales, slope, crossover scale, selected-scale vector, or another explicit functional. A scalar complexity index discards scale-specific structure and should not replace the profile when differences across scales are scientifically important.

Multiscale interpretation requires caution. Uncorrelated random noise can be highly irregular at one scale yet lose structure rapidly with coarse-graining, while correlated or structured processes can retain nontrivial organization across scales. Do not interpret monotonic increase or decrease of an entropy profile as universally better or more complex without a declared theoretical meaning.

State-Space Reconstruction and Recurrence Descriptors

A generic recurrence matrix is defined as

where i and j are eligible state-vector indices, z_i and z_j are observed or reconstructed state vectors under a declared state construction, d is the declared state-space distance or metric, ε is the declared recurrence threshold or radius, I is an indicator function, and R_{i,j} is binary recurrence status for that eligible pair. Temporally adjacent pairs can require exclusion through a declared Theiler or temporal-exclusion rule. Fixed threshold and fixed recurrence-rate constructions define different recurrence analyses.

Delay-coordinate state-space reconstruction from a scalar series is a mathematical construction distinct from direct observation of a complete physical state. Embedding dimension, delay, coordinate scaling, metric, temporal exclusion, missingness, and preprocessing determine neighborhood geometry. A selected embedding dimension should not be interpreted directly as the number of physiological or behavioral degrees of freedom.

Recurrence plots and recurrence quantification characterize returns of observed or reconstructed states to declared neighborhoods. Recurrence Rate quantifies density of recurrence points; Determinism quantifies diagonal-line structure; Average/Max Diagonal Length measure time of predictable evolution; Laminarity quantifies vertical-line structure; Trapping Time estimates average time trapped in states; Recurrence-Time Descriptors summarize intervals between recurrences; Recurrence-Line Entropy quantifies diversity of line lengths. These descriptors quantify different recurrence structures and should not be collapsed into one generic recurrence complexity value.

Threshold, metric, embedding, normalization, and recurrence-density dependence are critical. A fixed distance threshold compares recurrence under one geometric radius, whereas choosing thresholds to force equal recurrence rates removes recurrence density as a between-case difference and changes the interpretation of line-based RQA measures. The construction policy must remain explicit.

State-space recurrence differs from repeated event occurrence. An event-return interval is based on recurrence of declared events in time; an RQA recurrence point is based on closeness of state vectors under a metric and threshold. Neither form should substitute for the other because they answer different scientific questions.

Scaling, Fractal, and Long-Range Structure

The core Detrended Fluctuation Analysis (DFA) scaling relation is

where s is the detrending window or scale size under the declared DFA construction, F(s) is the resulting root-mean-square detrended fluctuation at scale s, C is a positive fitted scale factor, and α is the fitted scaling exponent over an explicitly declared log–log scaling range. Detrending order, scale range, crossover structure, weighting, fit diagnostics, and finite record length are part of the estimate. α should not automatically be relabeled a Hurst exponent outside assumptions that justify that relationship.

Hurst-like, rescaled-range, DFA, and related long-range scaling descriptors are distinct estimator families. Persistent, uncorrelated-like, and antipersistent interpretations depend on the process class and estimator assumptions. Trends, periodicities, crossovers, nonstationarity, filtering, and finite support can distort fitted exponents. The actual estimator name and fit range must be preserved rather than reporting a generic fractal exponent.

Fractal-dimension descriptors conceptually differ. Graph-based estimators such as Higuchi-like fractal dimension characterize the signal graph's geometric complexity, distinct from state-space correlation dimension or geometric fractal dimension of a spatial object. Graph-based estimators depend on scale range and sample geometry; correlation dimension depends on reconstructed state space, neighborhood scales, embedding, and evidence of a scaling region. Similar numerical dimensions from different constructions do not imply the same dynamical object.

Multifractal descriptors require sufficient data length and scaling evidence. They distinguish a single scaling exponent from a spectrum of local scaling exponents or generalized dimensions and require scale range, moment orders, regression or partition construction, and uncertainty quantification. A wide estimated multifractal spectrum can be produced or distorted by finite samples, trends, heavy-tailed distributions, noise, or estimator bias and is not automatic evidence of rich behavioral organization.

Divergence, Predictability, and Evidence for Nonlinear Structure

Lyapunov-like and local-divergence descriptors estimate how nearby reconstructed trajectories separate over time under declared state-space reconstruction, neighbor-selection rule, temporal exclusion, divergence statistic, time units, and fitted region. Variants include global largest Lyapunov exponent estimates, finite-time or local divergence, mean divergence curves, and inverse-longest-recurrence-line proxies. A positive finite-data estimate is not sufficient proof of deterministic chaos.

Intrinsic or local predictability descriptors such as nearest-neighbor forecast error, horizon-dependent prediction error, local analog forecasting, or improvement over a declared baseline model characterize predictability under a specific prediction rule and state construction. These descriptors differ from supervised prediction of behavioral labels and do not establish causality.

Surrogate-data-supported nonlinear evidence tests whether a chosen complexity or dynamical statistic is inconsistent with a declared null process while preserving selected properties. The null hypothesis, surrogate-generation method, preserved statistics, number of surrogates, randomization seed (when relevant), discriminating statistic, and decision rule must be specified. Rejection of a linear stochastic null is evidence against that null, not proof of deterministic chaos or identification of the true generating mechanism.

Entropy-rate, conditional-entropy, excess-entropy, statistical-complexity, and predictive-information-like quantities have theoretical or estimator-specific meanings different from finite-template SampEn, Permutation Entropy, recurrence-line entropy, spectral entropy, or histogram entropy. Do not rename a convenient finite-data statistic as one of these information-theoretic quantities without the estimator and assumptions required by that definition.

Adequacy, Uncertainty, Sensitivity, and Provenance

Data adequacy and invalid states are critical in complexity and regularity estimation. Estimates can fail because support is too short, sampling is inappropriate, too few template matches or ordinal patterns occur, state space is too sparse, a recurrence threshold yields no useful recurrence structure, a scaling region is absent, a DFA fit shows crossover instead of one slope, neighbor divergence lacks a stable fit region, or missingness/interpolation dominates the result. Undefined, infinite, unstable, unidentifiable, out-of-domain, and valid-but-uncertain values must be distinguished from zero.

A worked example using synthetic signals—periodic, noisy periodic, correlated stochastic, white-noise-like, and structured irregular sequences—demonstrates why one universal complexity ranking fails. Compare Sample Entropy, Permutation Entropy, a multiscale profile, one RQA measure, one scaling/fractal descriptor, and one divergence or predictability descriptor where assumptions allow. Show sensitivity to record length, sampling, filtering or detrending, normalization, tolerance or ordinal order, recurrence threshold, embedding, scale range, and fit region. Preserve descriptor definition/version, source signal and preprocessing state, support, sampling/timing, missingness, template/embedding parameters, distance and tolerance, self-match rule, ordinal tie rule, symbolization, multiscale construction, recurrence metric/threshold/Theiler rule, scaling estimator/range, fractal estimator, neighbor and divergence settings, surrogate/null specification when used, invalid-state rules, implementation/version, uncertainty, and sensitivity findings. Conclude that reproducible computation establishes conformance to the descriptor definition, not behavioral complexity, nonlinearity, or chaos.