Descriptor Support

Descriptor Support is the explicitly declared evidence scope to which one Behavioral Signal Descriptor instance applies and from which its scientific meaning is determined. It defines the precise subset of behavioral-signal evidence that contributes to a given descriptor value, thereby grounding the descriptor’s interpretation in its evidence base. Descriptor Support can be temporal, spatial, sample-based, event-based, entity-based, channel-based, relational, trajectory-based, or otherwise structured to reflect the scientific context. Importantly, Descriptor Support is not merely window length, nor is it synonymous with signal sampling, segmentation, a window weighting function, descriptor aggregation, or model receptive field. It is not an incidental implementation detail: changing descriptor support changes which evidence contributes, which behavioral scale is characterized, and ultimately what a descriptor value means.

Meaning and Boundaries of Descriptor Support

Descriptor Support is defined as an identifiable subset, indexed region, membership set, or relational scope of eligible behavioral-signal evidence over which a descriptor definition is instantiated. In other words, Descriptor Support answers the question: Which evidence belongs to this descriptor instance? This subset can include contiguous or noncontiguous samples, time intervals, frames, windows, segments, event neighborhoods, episodes, trajectories, spatial regions, channels, entities, entity pairs, or multisignal correspondences, depending on the descriptor’s scientific purpose and operational definition.

Descriptor Support differs from the mathematical notion of support of a signal or function. In mathematics, support typically means the domain subset on which a function is nonzero. Descriptor Support, by contrast, identifies the evidence scope admitted to a descriptor instance under its scientific definition. While these two notions can coincide in special cases, they must not be conflated or treated as synonyms.

To clarify further, Descriptor Support is distinct from related but separate concepts:

• Observation Interval: States the overall available observation extent but not which evidence contributes to a descriptor.
• Sampling Grid: Specifies represented coordinate points but not membership in a descriptor instance.
• Frame: A declared unit with modality-dependent meaning, often a basic temporal or spatial unit.
• Window Support: Selects a local evidence region relevant for descriptor computation.
• Window Function/Taper: Weights values within or around the signal-processing operation without changing membership.
• Segment: Scientifically defined boundaries or anchors that can supply descriptor membership rules.
• Event: Point or interval anchors around which supports can be defined.
• Receptive Field: Describes computational dependence in models or filters.
• Aggregation Support: The set or scope over which several descriptor instances are summarized.

Support Identity and Coordinate Semantics

Support identity is more than start time and duration. A reproducible Descriptor Support requires a support identifier, coordinate basis, explicit start and end or explicit members, source-signal identity and version, channel or entity scope, inclusion rule, validity state, anchor, and any transformation needed to relate source coordinates to support coordinates. Two supports with identical numerical bounds can remain scientifically distinct when they refer to different entities, signal versions, coordinate systems, modalities, or membership rules.

Coordinate bases for Descriptor Support can include sample index, frame index, clock time, experiment-relative time, event-relative time, spatial coordinate, trajectory parameter, symbolic position, or another explicitly defined index. Conversions among coordinate systems must preserve units, origin, direction, synchronization or mapping uncertainty, and rounding conventions where they can change membership.

Boundary semantics are crucial. For interval-like supports, boundaries can be closed, open, or half-open. Mapping continuous-time boundaries to discrete samples or frames requires declared rules. Exact-boundary samples, duplicate timestamps, irregular timestamps, rounding, quantization, and tie handling affect membership. For example, intervals [a,b], [a,b), and sample-index intervals can contain different evidence even if their displayed start and end values look identical.

A generic valid-support relation can be expressed as follows:

Here, i indexes the descriptor instance, u is an elemental coordinate or member of the declared support domain, S_i^n is the nominal support intended for descriptor instance i, v_i(u) is a declared binary validity or eligibility indicator for member u, and S_i^v is the valid subset that actually contributes under that validity rule. Validity can also involve richer states than a binary mask, but this relation makes the distinction between nominal and contributing support explicit.

Forms and Generation of Descriptor Support

Fixed local supports include sample blocks, frames, and fixed-duration or fixed-count temporal windows. Support length measured in samples can differ from support duration measured in time when sampling is irregular or differs across signals. Fixed supports simplify comparability but may mix distinct behavioral states if the support is poorly matched to the phenomenon.

Sliding and successive support generation uses parameters such as support length, hop or step size, and overlap. Overlap is a consequence of the relationship between support length and hop for ordinary fixed sliding windows, not an independent source of new evidence. Repeated overlapping supports generate distinct descriptor instances that share input evidence and thus should not be treated automatically as independent observations.

Variable-length and adaptive supports have extents that follow observed segment length, event duration, signal content, change points, quality conditions, behavioral scale, or another justified rule. Adaptive supports can improve alignment with heterogeneous phenomena but make support length data-dependent and can alter descriptor comparability, uncertainty, and distribution.

Event-centered and anchor-relative supports are defined over pre-event, post-event, symmetric, asymmetric, or phase-relative neighborhoods around a declared event or landmark. Anchor identity, anchor uncertainty, inclusion of the anchor itself, pre/post extents, and overlapping event neighborhoods should be preserved. Event-centered support does not imply error-free event definitions or timestamps.

Segment-, episode-, and trajectory-derived supports use membership defined by independently established behavioral or signal units. The descriptor characterizes evidence within these boundaries without redefining why the boundaries exist. Variable support duration, heterogeneous internal composition, and uncertain boundaries remain part of descriptor interpretation.

Irregular and noncontiguous supports have membership as an explicit set of valid samples, frames, events, spatial points, trajectory portions, or masked intervals rather than a single contiguous interval. Gaps may be excluded because evidence is invalid or because they are scientifically outside the target support. Descriptor definitions must state whether their mathematics remain meaningful on irregular or disconnected evidence.

Windows, Boundaries, Anchors, and Edge Policies

A window’s support is distinct from its weighting or taper. A temporal window selects which samples or times belong to a descriptor instance, while a Hann, Hamming, rectangular, Gaussian, or other weighting function changes how included evidence contributes. Descriptor definitions must preserve both concepts when both matter. A tapered computation is not the same as changing support selection.

Support anchors and descriptor timestamps identify the descriptor instance but do not imply that the descriptor characterizes only that instant. An anchor can be the support start, center, end, event anchor, centroid of valid support, or another declared coordinate. The chosen timestamp identifies the descriptor instance; descriptor timestamp density is not equal to effective temporal extent.

Partial-support policies near observation, segment, session, or data-quality boundaries include dropping the partial instance, truncating support, padding, reflection or extension, minimum valid fraction requirements, or returning invalid/undefined results. These policies change evidence membership or computation assumptions and should be part of descriptor identity or provenance when they can affect interpretation.

Padding and extrapolation produce generated evidence rather than observed members. Zero padding, reflection, replication, model-based extrapolation, or other edge extensions can be legitimate computational devices but may alter statistics, spectra, derivatives, morphology, and other descriptors. Padding rules must be preserved, and nominal observed support must be distinguished from synthetic extensions.

Supports can be causal (past-only), centered (symmetrical around anchor), or future-inclusive. Centered supports can use evidence after the nominal descriptor timestamp, while causal supports restrict membership or computational dependence to evidence available up to the relevant time. This distinction is critical for real-time interpretation, forecasting, and leakage-sensitive evaluation; however, causal support is not universally preferable.

This timeline shows one larger segment support (blue), three overlapping sliding window supports (green) with declared hop, an event-centered support (orange) around an event anchor (red line), an invalid gap excluded from valid support (gray thick line), and a partial support at the right edge (purple). Each local support produces a descriptor instance indexed by its chosen anchor (circles with labels). Overlapping descriptor instances share evidence, and descriptor timestamp density is not equal to support duration.

Valid Support, Missingness, and Sufficiency

Nominal support is the scope the descriptor is intended to characterize. Observed support is the evidence actually available. Valid support excludes evidence declared unusable or ineligible. Effective contributing support can additionally reflect weighting, transform requirements, masks, or estimator-specific exclusions. These notions must remain distinguishable when they differ.

Validity masks exclude artifact-contaminated, missing, saturated, occluded, unsynchronized, low-confidence, or otherwise unusable evidence according to declared criteria. Exclusion is distinct from imputation or correction: a corrected value can re-enter support under a declared preprocessing state, whereas a masked value contributes no observed evidence under the mask.

Support coverage quantifies the proportion of valid support relative to nominal support:

where i is the descriptor-instance index, S_i^v is the declared valid support, S_i^n is the nominal support, μ is the measure appropriate to the support semantics—such as valid sample count, duration, spatial measure, or another justified measure—and C_i is the valid-support coverage. Coverage quantifies represented support under the declared measure; it does not establish signal quality, descriptor correctness, behavioral validity, or support sufficiency by itself.

Support sufficiency is descriptor-specific adequacy of available evidence for producing an interpretable instance. Sufficiency can depend on valid duration, sample count, number of events, frequency cycles, spatial points, entity observations, dynamic range, or another property required by the descriptor definition. A support can have high percentage coverage yet remain insufficient for a descriptor whose estimator requires more absolute evidence or structural diversity.

Undefined, invalid, unstable, and unavailable descriptor outcomes may arise from support issues. A descriptor can be mathematically undefined because required variation or events are absent, invalid because support violates admissibility criteria, unstable because evidence is insufficient for reliable estimation, or unavailable because required source data do not exist. These states should not be converted silently into zero, default values, or behavioral absence.

Multichannel, Entity, Spatial, and Relational Support

Descriptor Support can apply to one channel, a declared channel set, vector-valued samples, modality-specific evidence, or synchronized evidence from multiple modalities. Channel identities, coordinate frames, units, sampling differences, and transformations needed to make joint support meaningful must be preserved. Simultaneous availability alone does not guarantee commensurability.

Entity and spatial support can characterize one participant, body part, tracked object, spatial region, trajectory subset, dyad, group, or other entity set. Entity identity, role, spatial reference frame, inclusion rule, and membership changes must be explicit. Group-level or relational descriptors should not be interpreted as properties of each individual entity merely because all contributed evidence.

Relational and paired supports compare two or more signals or entities. They define which observations correspond, temporal or event overlap requirements, and whether support uses intersection, union, matched events, nearest-neighbor correspondence, lagged pairing, or another declared relation. Relational support is more than two separate univariate supports because the correspondence relation determines which paired evidence contributes.

Changing availability in multisignal support can occur when one modality, channel, or entity becomes unavailable within otherwise valid support, producing time-varying membership. Distinguish complete-case intersection, partial multimodal support, modality-specific descriptor instances, and explicit missing-modality states. Apparent cross-signal evidence should not be created by pairing values outside scientifically defensible correspondence regions.

Support Scale, Descriptor Contours, and Dependence

Support scale relates to the property being characterized. Supports that are too short produce unstable estimates or miss slower structure; supports too long mix states, average transitions, violate local assumptions, or obscure brief behavior. Mixed-scale supports combine processes with different characteristic durations. No universally optimal support size exists independent of descriptor definition, signal semantics, and scientific question.

Multiscale support involves deliberate evaluation of related descriptor definitions or instances across several temporal lengths, spatial scales, event neighborhoods, frequency-related supports, or organizational scopes. Scale-indexed descriptor values should remain distinguishable unless explicitly combined. Agreement across scales does not prove one scale is intrinsically correct.

Descriptor contours are ordered collections of local descriptor instances indexed by support anchor or another support coordinate. Support length, hop or contour step, overlap, timestamp rule, and effective temporal resolution differ from contour sampling density. For example, a descriptor contour sampled every 100 ms can consist of values each summarizing several seconds of shared evidence.

Dependence among descriptor instances arises from overlapping supports, shared preprocessing, nested supports, common events, repeated entities, or shared source evidence. Dense contours contain many values but not the same number of independent behavioral observations. This dependence matters when descriptor instances are compared, summarized, modeled, or interpreted statistically, without requiring a general statistical-dependence tutorial.

Support Appropriateness, Sensitivity, and Provenance

Support appropriateness differs from mechanical support verification. Verification confirms that coordinates, membership, boundaries, masks, channel/entity identities, anchors, and edge policies were implemented as specified. Scientific appropriateness asks whether the support meaningfully characterizes the intended signal property. A perfectly reproduced support may still be scientifically mismatched to the behavior or descriptor.

Support sensitivity analysis compares plausible support lengths, hops, overlap, boundary conventions, event extents, anchor rules, edge policies, validity masks, sufficiency thresholds, multichannel correspondence rules, and multiscale choices. It examines whether descriptor values, contours, distributions, rankings, or scientific conclusions materially change. Support should not be chosen solely because one configuration maximizes downstream predictive performance or agreement with a desired result.

Descriptor Support provenance includes all information needed to reproduce and interpret evidence membership. This includes: support identity and version; source signal and preprocessing version; coordinate basis; bounds or explicit members; boundary convention; support-generation rule; length/hop/overlap; anchor and descriptor timestamp; segment or event identity when used scientifically; channel/entity/spatial scope; validity masks; nominal and valid coverage; sufficiency state; edge and padding policy; causal or future-inclusive status; correspondence rules for relational supports; effective computational dependencies; multiscale identity; sensitivity results; and implementation details needed to reproduce membership. A defensible Descriptor Support makes clear exactly which evidence a descriptor characterizes and which assumptions determine that membership.