Statistical and Distributional Descriptors

Statistical and Distributional Descriptors explicitly characterize the empirical values observed within a declared Behavioral Signal Descriptor support. These descriptors summarize properties such as location, magnitude, dispersion, range, quantiles, empirical distribution, distribution shape, tails, occupancy, and robust summaries. This family of descriptors primarily captures how values are distributed over eligible support and usually does not preserve temporal, spatial, symbolic, or relational ordering of those values unless such ordering is explicitly part of the descriptor definition. It is important to recognize that a statistically named quantity is not automatically behaviorally meaningful solely because it is mathematically well defined; behavioral interpretation requires contextual grounding beyond mathematical existence.

Meaning and Scope of Statistical and Distributional Descriptors

A Statistical or Distributional Descriptor is a scalar, vector, distributional, or structured characterization of the empirical value population associated with a declared signal object and valid support. The definition of such a descriptor must explicitly identify:

• The input-value semantics (what the values represent)
• The support population (which observations are included)
• The weighting or membership rule (how observations contribute)
• The estimator or functional (the mathematical operation applied)
• Parameterization (e.g., quantile probabilities, bin edges)
• Output units or scale
• Interpretation within the behavioral context

Distributional order invariance means that descriptors like mean, variance, quantiles, histograms, and many moment-based summaries remain unchanged when the same multiset of values is permuted. Thus, even if temporal behavior, transitions, autocorrelation, recurrence, or waveform morphology change completely, these descriptors do not reflect those changes. Consequently, identical statistical descriptors do not establish identical behavioral dynamics.

It is necessary to distinguish Statistical and Distributional Descriptor computation from Descriptor Aggregation. Computing the mean, variance, quantiles, or empirical distribution of signal values within one declared descriptor support produces a descriptor of that support. In contrast, computing a mean, median, distribution, or other functional across already constructed descriptor instances constitutes an aggregation of descriptor instances. Although the same mathematical operator can serve both roles, scientific identity derives from the input objects and semantics rather than the operator name alone.

Empirical Observation Population and Weighting

The empirical observation population contributing to a statistical descriptor depends on the declared signal object and its support. Observations may be samples, frames, amplitudes, vector magnitudes, event-associated values, segment-level values, spatial samples, or other explicitly declared observational units. The population must adhere to Descriptor Support semantics rather than being inferred merely from the length of an array or raw data structure.

Equal observation weighting applies when each retained value is intended to contribute equally to the descriptor. However, irregular sampling, unequal dwell times, variable spatial areas, event exposure, confidence weighting, or other scientifically justified designs require unequal weighting. Weight semantics must be declared explicitly, and numerical normalization of weights alone does not guarantee scientific appropriateness.

The weighted empirical cumulative distribution function (CDF) is rendered as:

Here, E is the set of eligible observations within the declared descriptor support, i indexes an eligible observation, x_i is observation value i, w_i is the declared nonnegative weight of observation i, W is the total eligible weight, I is an indicator function (1 if the condition is true, 0 otherwise), x is the evaluation value, and F̂_w(x) is the weighted empirical cumulative fraction of values at or below x. Equal weights recover the ordinary empirical CDF. The empirical CDF describes the observed value distribution rather than a behavioral-truth probability distribution.

Missing, invalid, censored, saturated, clipped, out-of-range, and structurally unavailable observations represent distinct states that alter the empirical population. Such observations may be excluded, retained with explicit semantics, censored, imputed, or otherwise treated depending on the declared protocol. The resulting population and denominator change accordingly. Missing or undefined values should not be replaced with zero unless zero explicitly carries the intended signal meaning.

Location, Magnitude, and Scale

The arithmetic mean is a location descriptor appropriate when arithmetic averaging is meaningful for the input scale and observation population. It is sensitive to extreme values and depends on observation weighting. The mean is interpretable in the original input units and represents the center of mass under equal weighting. The mean of signed values differs fundamentally from mean absolute magnitude, root-mean-square magnitude, and other nonnegative magnitude summaries.

The median and trimmed or winsorized location summaries provide alternatives with different robustness and information-loss trade-offs. The median identifies a central order statistic, robust to extreme values but discarding some distributional detail. Trimmed means remove or attenuate extremes according to declared rules, while winsorized means replace extremes with nearer values. Robustness to extremes should not be interpreted as automatic robustness to systematic bias, multimodality, support mismatch, or invalid observations.

Magnitude descriptors such as mean absolute value, root-mean-square (RMS) value, norms, or related nonnegative summaries are meaningful only when their input and scale semantics justify them. Signed location is distinct from signal magnitude, and magnitude or squared magnitude is distinct from physical energy or power unless supported by signal definition and units.

Baseline, centering, and origin dependence affect these descriptors. Mean, absolute magnitude, RMS, and related statistics can change after offset removal, baseline correction, normalization, or coordinate transformation. A descriptor computed after centering characterizes the transformed signal, not the original absolute level, so preprocessing state must be part of the interpretation.

Dispersion, Range, and Relative Variation

Weighted descriptive mean and variance are defined as:

where E is the eligible observation set, i indexes observations, x_i is observation value i, w_i is the declared nonnegative weight of observation i, W is the total eligible weight defined by W = Σ_{i∈E} w_i, μ_w is the weighted descriptive mean, and σ_w² is the weighted descriptive second central moment. This convention is descriptive and empirical; it is not a universal unbiased sample-variance estimator. Finite-sample corrections and weighted-sample variance conventions differ and must be declared when inferential estimation is intended.

Variance and standard deviation characterize dispersion around a declared center under a specified estimator convention. Population-style denominators differ from sample estimation denominators, and weighting conventions, degrees-of-freedom adjustments, and software defaults vary. Standard deviation shares the input unit, while variance has squared units. Neither reveals skewness, multimodality, tails, or temporal organization alone.

Range, interquartile range (IQR), median absolute deviation (MAD), and related robust spread descriptors differ in properties. Range depends strongly on extremes and sample size; IQR describes central spread; MAD requires a declared center and scaling convention if interpreted as a standard-deviation analogue. These summaries have distinct roles and should not be treated interchangeably.

Relative dispersion descriptors such as the coefficient of variation (CV) are meaningful only when division by a reference level is justified, typically on a ratio scale. CV is unstable when the mean is near zero, undefined or misleading if the mean changes sign or the scale is interval or bipolar. A dimensionless output does not guarantee comparability across signals.

Dispersion depends on support composition, weighting, multimodality, preprocessing, and genuine state mixture. High variance can reflect behavioral heterogeneity, transitions, multiple regimes, artifacts, or scale changes; low variance can indicate stable behavior, clipping, saturation, excessive smoothing, or insufficient dynamic range. Therefore, dispersion requires signal-context interpretation.

Quantiles, Empirical Distributions, and Occupancy

Quantiles, percentiles, median, interquantile ranges, and quantile vectors are order-statistic descriptors of the empirical value distribution. The quantile probability and interpolation or selection convention must be specified when finite samples admit multiple definitions. Similar quantiles do not imply identical full distributions.

The empirical cumulative distribution function (CDF) is a distributional descriptor preserving the cumulative weighted or unweighted value distribution over support. Quantiles can be conceptually derived from the empirical CDF. Comparing empirical CDFs can reveal distributional differences hidden by mean and variance comparisons alone.

Histograms and empirical probability-mass representations depend on bin definitions, edges, normalization, weights, and support population. Choices of bin width, origin, closed/open edge conventions, and number of bins affect apparent shape. A normalized histogram represents relative empirical occupancy of bins and is not automatically a calibrated probability model for future or latent behavior.

Density estimates provide smoothed representations of an empirical value distribution. Bandwidth, kernel choice, boundary treatment, weighting, and finite-sample size influence apparent modes and tails. Density estimates differ from histograms and from parametric probability distributions fitted under additional assumptions.

Threshold occupancy, exceedance fraction, below-threshold fraction, central-band occupancy, and related distributional descriptors require declaration of threshold units, direction, inclusivity, reference baseline, and denominator. Thresholds may arise from physical meaning, calibration, population criteria, or other justified rules. Choosing thresholds solely to maximize downstream discrimination alters scientific interpretation.

Moments, Distribution Shape, Tails, and Special Forms

Raw moments summarize value distributions mathematically; central moments center values before powering. Standardized moments normalize scale to characterize shape. Different moment orders emphasize different distribution aspects and increasingly weight extreme deviations. Some finite moments may not exist or be stably estimable for every empirical or modeled distribution.

Skewness describes asymmetric distributional shape under a declared estimator convention. It differs from mean–median difference, directional trend, temporal asymmetry, and waveform asymmetry. Sample skewness conventions and finite-sample corrections vary across software, so the estimator must be identified rather than relying on the label alone.

Kurtosis and excess kurtosis are convention-dependent fourth-moment shape descriptors associated with tail weight and concentration relative to a reference distribution. It must be explicitly stated whether ordinary kurtosis or excess kurtosis is reported, which estimator or bias correction was used, and why the common informal description "peakedness" is incomplete. Kurtosis values should not be compared across implementations without confirming conventions.

Tail and extreme-value descriptors include upper or lower quantiles, tail occupancy, maximum and minimum, trimmed extremes, or exceedance magnitudes. These are sensitive to support size, outliers, clipping, censoring, and rare-event sampling. An observed maximum is a sample extreme, not necessarily a physical maximum of the behavioral process.

Static value-distribution entropy is meaningful only when a discrete, quantized, histogram, or otherwise explicit probability-mass representation has been defined. The binning or symbolization and logarithm convention must be stated. Entropy of the empirical amplitude distribution discards ordering and must not be conflated with sample entropy, permutation entropy, entropy rate, multiscale entropy, or other order-sensitive complexity descriptors.

Special empirical distribution forms require modified interpretation: discrete or quantized values, sparse or zero-inflated signals, bounded scales, saturated or clipped signals, multimodal distributions, and circular or angular values. Ordinary linear mean, variance, quantiles, and histogram choices may be inappropriate or incomplete for some of these forms, especially circular variables whose topology differs from a line.

Robustness, Finite-Sample Behavior, and Multivariate Summaries

Robustness and influence describe how descriptors react to outliers or contamination. Mean and variance are sensitive to outliers, while extrema and quantiles vary in sensitivity. Median, trimmed statistics, and median absolute deviation (MAD) intentionally reduce sensitivity to selected departures but may suppress genuine rare behavior and do not correct systematic measurement bias, clipping, or inappropriate support.

Finite-sample behavior influences estimator bias, variance, instability, and definedness. Small valid sample counts, limited dynamic range, repeated identical values, sparse tails, few occupied histogram bins, or excessive missingness can make higher moments, quantiles, density estimates, entropy, or relative-spread descriptors unstable or undefined. Effective observation count or total weight should be reported when it materially affects interpretation.

Weighted and irregularly sampled statistical descriptors arise when samples represent unequal time spans or observation opportunities. Ordinary equal-sample statistics may characterize sampling patterns rather than the intended time-weighted distribution. Time weighting, event weighting, spatial weighting, or other designs change the estimand and must remain explicit instead of being treated as mere technical corrections.

Multivariate statistical summaries for declared vector-valued signals or jointly defined signal objects include component-wise location and spread, covariance or correlation structure, and robust multivariate summaries conceptually. Component units, coordinate frame, scaling, and joint observation membership must be preserved. Covariance within one declared multivariate signal differs from broader cross-signal synchrony, lagged dependence, coherence, or causal relations.

Descriptor redundancy and complementarity exist within statistical summaries. Mean, median, quantiles, variance, interquartile range, skewness, and histogram-derived descriptors may be correlated or mathematically related without being semantically identical. Avoid selecting or discarding descriptors solely based on pairwise correlation without considering the distinct properties each descriptor preserves.

Interpretation, Verification, and Provenance

Consider two short signal supports, each containing the same set of values but arranged in different temporal orders, and a third support containing an extreme value:

• Support A: Values [1, 3, 2, 4]
• Support B: Values [4, 2, 3, 1] (same multiset, different order)
• Support C: Values [1, 3, 2, 40] (includes an extreme outlier)

Computing mean, median, variance, selected quantiles, histograms or empirical CDFs, skewness, tail behavior, and robust summaries shows:

• Supports A and B share identical statistical and distributional descriptors despite differing temporal order.
• Support C’s extreme value strongly affects mean and variance but has less influence on median and trimmed mean, demonstrating robustness.
• Histograms and empirical CDFs reveal concentration differences driven by the extreme value.
• Skewness and tail occupancy descriptors highlight asymmetry introduced by the outlier.
• Robust summaries such as median absolute deviation remain stable across supports.

This example highlights that statistical descriptors do not encode temporal organization and that extreme values can disproportionately influence some but not all descriptors.

Verification, sensitivity, uncertainty, and provenance form a coherent responsibility. Verification involves confirming observation membership, valid count or total weight, weighting, units, center convention, denominator, bias correction, quantile algorithm, threshold semantics, histogram bins, density bandwidth, logarithm base for entropy, circular treatment, missingness policy, and deterministic recomputability on test data when useful. Sensitivity analysis examines plausible support boundaries, preprocessing, outlier policies, weights, quantile conventions, binning, estimator definitions, and finite-sample composition.

Provenance documentation must preserve descriptor identity and version, input signal and preprocessing state, support, empirical population, weights, estimator and convention, parameters and defaults, units, missing/invalid states, effective sample size or total weight when material, implementation/version, uncertainty or stability diagnostics, and sensitivity results. Arithmetic reproducibility establishes computational conformance but not behavioral validity.