Behavioral Signal Descriptors

Behavioral Signal Descriptors are explicitly defined characterizations of properties, structures, events, dynamics, or relationships found in identified behavioral-signal evidence over declared supports. A descriptor is produced by applying a specified descriptor definition to eligible signal evidence under declared parameters, assumptions, and output semantics. It is important to establish that a descriptor is not the raw signal itself, nor is it a behavioral annotation or reference; it is not automatically a measurement in the metrological sense, nor is it synonymous with a feature or the same object as a representation or model output.

Meaning of Behavioral Signal Descriptors

A Behavioral Signal Descriptor is an identifiable and reproducible characterization obtained from one or more behavioral-signal inputs over a declared support, with explicit signal-property semantics, computation, parameterization, assumptions, output form, and interpretive limits. Descriptor identity is determined by what is characterized and how it is defined, not merely by a column name, software function, numerical value, or eventual predictive usefulness.

The scientific purpose of descriptors is to make selected properties of behavioral-signal evidence explicit in forms that can be compared, summarized, related, monitored, or used in further analysis. Descriptors can characterize aspects such as level, distribution, variability, rate, duration, temporal organization, spectral organization, shape, geometry, kinematics, regularity, symbolic structure, or cross-signal relationships without asserting a behavioral category or reference truth merely by being computed.

It is crucial to distinguish characterization from transformation and labeling. A transformation produces another signal-like or coefficient representation—examples include a filtered signal, a spectrum, or transform coefficients—whereas a descriptor states a defined property of eligible evidence or of a declared transformed object. An annotation or reference assigns a behavioral assertion or evidential role to a target. For example, a Fourier spectrum is a transformed signal, whereas the spectral centroid is a descriptor summarizing that spectrum. A trajectory is a signal or geometric path, whereas trajectory curvature is a descriptor of its shape. A behavioral label assigns a category, while an event rate is a descriptor of frequency.

Descriptors extend beyond simple scalar features. A descriptor definition can yield scalar, vector, structured values, distributions, relational values, or time-indexed descriptor contours when such outputs themselves constitute the defined characterization. For example, a vector-valued descriptor whose components jointly define one characterization differs from an arbitrary vector assembled from several independent descriptors; the latter is an organizational representation rather than a new descriptor merely because it is stored as one vector.

Descriptor Definition, Instance, and Identity

Here,

• i indexes the descriptor instance,
• D denotes the descriptor definition,
• F_D is the declared mapping specified by the definition,
• X_i is the identified eligible signal evidence for instance i,
• S_i is the declared support over which the characterization is formed,
• θ_D are the resolved descriptor parameters,
• A_D are the declared assumptions or contextual conditions required by the definition,
• d_i is the resulting descriptor instance.

The output need not be scalar; it can be vector or structured. This equation expresses descriptor identity conceptually rather than mandating every implementation to follow this exact computational form.

Descriptor definition, descriptor instance, descriptor value, and descriptor interpretation are distinct. The definition identifies the characterized property and its mapping. An instance applies this definition to identified evidence and support. The value is the resulting numerical or structured output. Interpretation states what this output means under the declared signal semantics, units, assumptions, and intended use. Equal numerical values from different definitions or supports need not be scientifically equivalent.

Signal-input semantics are integral to descriptor identity. Applying the same descriptor definition to different signals—such as acceleration magnitude, joint angle, gaze direction, speech activity, inter-event intervals, or symbolic sequences—can produce different meanings even if the mathematical operation is identical. Input semantics include channel, modality, coordinate frame, units, calibration state, preprocessing state, entity identity, and any other property materially affecting interpretation.

Parameter and assumption identity also matter. Parameters such as window length, frequency band, event threshold, embedding dimension, spatial reference frame, smoothing level, normalization convention, symbolic vocabulary, or lag range can materially change what a descriptor measures or how comparable its instances are. Mathematically equivalent parameterizations can represent scientifically different descriptor variants. Defaults must be treated as explicit parameter choices rather than invisible constants.

Output semantics, dimensionality, and units vary by descriptor. Outputs can retain physical units, combine multiple units, be dimensionless, express rates, angles, ratios, counts, probability-like quantities under justified semantics, distributions, vectors, or other structured forms. Numerical storage type or decimal precision alone do not establish measurement scale, physical meaning, or behavioral resolution.

Descriptor Support, Contours, and Aggregation

Descriptor support is the declared region of evidence—temporal, spatial, event-based, entity-based, relational, or otherwise—over which a descriptor instance is defined. Relevant supports can include samples, frames, windows, segments, events, event-centered neighborhoods, episodes, sessions, trajectories, spatial regions, entity sets, or relations. The same descriptor definition applied over different supports can produce scientifically different interpretations.

Local descriptors and descriptor contours arise when a descriptor is applied repeatedly over successive or irregular supports, producing a time- or support-indexed contour of descriptor instances. This contour differs from the underlying raw signal and from continuous behavioral annotations or reference trajectories. It records repeated characterizations whose temporal resolution depends on support length, overlap, step size, sampling, and the descriptor itself.

Functionals and aggregation summarize descriptor contours or collections of instances. Means, quantiles, extrema, slopes, distributions, counts, or other summaries characterize the contour but define a new characterization that may discard temporal or distributional structure. Aggregation of descriptor values differs from computing the same descriptor directly on a larger pooled support because nonlinear operations need not commute with aggregation.

Weighting and variable-length support alter aggregate interpretations. Equal-instance, time-weighted, duration-weighted, event-weighted, or other aggregation schemes answer different scientific questions when segments, events, intervals, or sampling densities differ. The denominator and weighting semantics must be stated explicitly to avoid misleading dominance by long episodes or densely sampled periods.

Multiscale descriptor characterization evaluates the same property over multiple window lengths, event scales, temporal resolutions, spatial scales, frequency bands, or organizational supports to reveal scale dependence. Values at distinct scales must remain identifiable unless an explicit multiscale definition combines them. Smaller support does not automatically mean better behavioral resolution.

Families of Behavioral Signal Descriptors

Statistical and Distributional Descriptors characterize value level, spread, quantiles, distribution shape, concentration, extremes, robust location, or other distributional properties over declared support. These depend on sampling density, support composition, missing data, outliers, and whether observations are legitimately comparable. Examples include mean, variance, median, skewness, kurtosis, and quantile ranges.

Temporal Descriptors characterize timing, duration, rate, rhythm, interval structure, transitions, trends, temporal dependence, or event organization. Descriptors computed directly from sampled signals differ from those computed from detected events or states, the latter inheriting thresholding, timing uncertainty, and extraction errors. Examples include event rate, inter-event interval distribution, temporal autocorrelation, and trend slope.

Spectral and Time-Frequency Descriptors characterize how signal variation is organized by frequency and, when localized analysis is used, by both frequency and time or scale. Representative properties include dominant frequency, band energy, spectral shape, concentration, or localized energy patterns. These depend on sampling rate, support length, stationarity assumptions, transform definition, spectral leakage, frequency resolution, and preprocessing. Examples include spectral centroid, band power, spectrogram peaks.

Morphological and Shape Descriptors characterize waveform, contour, event, trajectory, or geometric form—such as peak shape, width, asymmetry, curvature, path shape, contour geometry, or other declared structural properties. Shape characterization differs from semantic classification of the behavior that produced the shape. Examples include peak width, curvature, contour eccentricity.

Spatial and Kinematic Descriptors characterize position, displacement, distance, orientation, velocity, acceleration, angular motion, path length, joint configuration, spatial extent, or relations among body parts, objects, or tracked entities. Interpretation requires explicitly stated coordinate frame, units, reference origin, dimensionality, calibration, and entity identity. Kinematics describe motion without implying kinetic forces or causal explanations unless supported by evidence. Examples include Euclidean distance traveled, angular velocity, joint angle variance.

Complexity and Regularity Descriptors characterize predictability, irregularity, recurrence, scaling, compressibility, entropy-like structure, fractal behavior, or dynamical organization under explicitly stated definitions. Different complexity measures operationalize different mathematical notions; a larger value does not universally mean more complex, less regular, healthier, or more behaviorally sophisticated. Examples include approximate entropy, sample entropy, fractal dimension, recurrence rate.

Linguistic and Symbolic Descriptors characterize properties computed from declared symbolic, token, transcript, event-symbol, action-code, or other discrete behavioral sequences when those objects are legitimate signal representations for the analysis. Properties include frequency, diversity, sequence structure, transition organization, repetition, duration, or symbolic regularity. These differ from behavioral annotations or labels whose symbols carry semantic assertions. Examples include symbol entropy, n-gram frequency, transition probabilities.

Cross-Signal and Relational Descriptors characterize correspondence, synchrony, relative timing, dependence, coupling, similarity, coordination, coherence, or structured relations among two or more signals, entities, modalities, or event streams. They require compatible supports and explicit correspondence semantics. Correlation, synchrony, coherence, or other dependence measures do not by themselves establish causality, shared mechanism, or behavioral agreement. Examples include cross-correlation, phase-locking value, coherence.

Descriptor Interpretation, Sensitivity, and Comparability

Mathematical meaning and behavioral meaning must be distinguished. A variance, entropy, centroid, curvature, rate, distance, or coupling value has a precise mathematical definition, but its behavioral interpretation depends on the characterized signal, how it was obtained, the support used, any preprocessing, and the scientific claim made. One must not infer a behavioral construct directly from a mathematically named descriptor without context.

Descriptor invariance, equivariance, and sensitivity relate to how values change under transformations. Some descriptors are intentionally insensitive to translation, rotation, amplitude offset, temporal shift, scale, or other nuisance transformations; others are defined to change under these transformations. Desired behavior should follow the target property; greater invariance is not always preferable.

Descriptors are sensitive to preprocessing, sampling, and support definitions. Filtering, denoising, artifact handling, normalization, resampling, coordinate transformations, event detection, segmentation, support boundaries, overlap, and missing-data treatments can alter descriptor values. Such dependencies may be scientifically legitimate but must be traceable and not mistaken for behavioral changes without examination of processing and support conditions.

Finite-support and resolution effects impact descriptor stability. Descriptors estimated from short, sparse, irregular, or low-resolution evidence can be unstable, biased, undefined, or unable to represent slower or finer structures. Conversely, longer supports can mix behavioral states or obscure local dynamics. Support sufficiency is descriptor- and property-specific; no universal minimum duration or sample count applies.

Descriptor comparability requires compatibility in descriptor definition, input semantics, support semantics, parameters, units, preprocessing state, implementation behavior, and relevant context, or an explicit transformation that justifies comparison. Identical descriptor names in different software libraries or datasets do not establish definition equivalence.

Descriptor Sets, Features, and Representations

Descriptor sets are collections of explicitly identified descriptor instances or definitions assembled for joint characterization. Breadth of descriptors does not guarantee independence: several descriptors can share the same signal, support, transformation, event detector, mathematical ingredients, or upstream descriptor and thus be redundant or strongly dependent. More descriptor columns do not automatically represent more independent behavioral information.

A feature is a role that a quantity assumes in a specific analytical, statistical, or predictive task, whereas descriptor identity is grounded in the signal property and descriptor definition. A descriptor can be used as a feature; a feature can also be derived from transformed signals, metadata, learned representations, annotations, or other quantities. Predictive usefulness neither defines nor validates descriptor status.

A representation organizes or encodes evidence into a structured object such as a vector, sequence, matrix, tensor, graph, symbolic structure, or embedding. Descriptors can become coordinates or components of such an object. Assembling descriptors into a representation does not create new descriptor identities unless an additional declared descriptor definition characterizes the assembled evidence.

Explicit descriptors have defined property semantics. Learned or model-assisted quantities can estimate explicitly defined properties, in which case the resulting descriptor must preserve the target property, training-state dependence, model version, calibration, uncertainty, and validation necessary for that interpretation. Latent embedding coordinates or hidden activations without stable standalone property semantics are better treated as representation components rather than descriptors merely because they are numeric.

Descriptor Quality, Uncertainty, and Provenance

Descriptor quality is multidimensional and property-specific. Relevant considerations include definitional correctness, computational correctness, semantic validity, repeatability, reproducibility, robustness, sensitivity to intended versus nuisance changes, numerical stability, missing or invalid-state behavior, resolution, and fitness for the intended characterization. No single downstream model score, correlation, reliability coefficient, or software test establishes all dimensions of descriptor quality.

Descriptor uncertainty and invalidity states must be explained to support interpretation. Uncertainty arises from finite support, input noise, uncertain event or segment boundaries, calibration, preprocessing, sampling, parameter choices, stochastic estimation, model-assisted computation, implementation differences, or uncertain upstream quantities. It is important to distinguish uncertain descriptor values from undefined descriptors, unavailable input, out-of-domain computations, saturated outputs, or mathematically valid values whose behavioral interpretation is unsupported.

Consider a short multimodal behavioral episode involving a motion signal (e.g., hand acceleration), an event stream (e.g., button presses), and a symbolic stream (e.g., spoken keywords). A statistical descriptor such as mean acceleration magnitude characterizes average movement level over a session (input semantics: calibrated accelerometer; support: session duration; property: mean level; output: scalar in m/s²; assumption: sensor calibration). A temporal descriptor such as event rate characterizes button press frequency (input semantics: detected button-press event timestamps; support: session duration; output: count rate in Hz; assumption: accurate event detection). A spectral descriptor such as dominant frequency characterizes rhythmic movement components (input: acceleration; support: sliding window; output: scalar frequency in Hz; assumption: stationarity within window). A symbolic descriptor such as keyword diversity characterizes the variety of spoken words (input: recognized tokens; support: session; output: count of unique tokens; assumption: correct transcription). A cross-signal descriptor such as temporal correlation quantifies synchrony between movement bursts and word onsets (input: acceleration bursts and word timestamps; support: overlapping windows; output: correlation coefficient; assumption: aligned signals). None of these descriptor values is automatically a behavioral label, reference, causal explanation, or predictive feature without further contextualization.

Descriptor provenance encompasses all information required to reproduce and scientifically interpret a descriptor instance. This includes descriptor identity and version, characterized property, input signal identity and semantics, entity or modality, support, preprocessing state, sampling conditions, parameters and defaults, assumptions, units and output structure, event or transformation dependencies, implementation and version, fitted or learned state when applicable, invalid or missing status, uncertainty or sensitivity findings, computation run identity, and derivation from upstream evidence or descriptors. A defensible Behavioral Signal Descriptor states what property was characterized, from which evidence, over what support, by which definition and parameters, with what meaning and limitations, and under which conditions its values are comparable.