Predictive State Filter¶
Recursive state estimator — instantiates Entity Persistence Across Observation Gaps
Carries an entity's state forward through an observation gap as a probability distribution anchored on the last confirmed sighting, widening the uncertainty envelope as time passes so the estimate never masquerades as an observation.
When an entity slips out of view, most of what you knew about it a moment ago is still roughly true — and getting less true by the second. Predictive State Filter is the machinery that carries that knowledge across the gap: it takes the last confirmed observation, applies a motion or dynamics model to predict where the entity is now, and represents the answer as a probability distribution whose spread grows the longer the entity stays unseen. Its one defining commitment is that a prediction is not an observation — the filter never collapses its estimate into a hard fact, because the whole point of the widening envelope is to say, in numbers, "I believe it is somewhere in here, and my confidence is decaying." It is the active, state-propagating answer to a gap, and it is precisely the mechanism a sibling registry refuses to be when detailed prediction isn't warranted.
Example¶
A self-driving car is tracking a cyclist in the next lane when a parked delivery truck slides between them and the cyclist vanishes from every camera. The naïve move — drop the cyclist and re-detect whoever appears next — would be dangerous: for a second and a half the car is blind to a road user it knows is there. The Predictive State Filter refuses to forget. It takes the cyclist's last confirmed position and velocity, propagates them forward under a plausible motion model, and maintains a growing ellipse of "the cyclist is probably here" that drifts down the road at roughly the last-seen speed.
Crucially the ellipse spreads: after one second the cyclist could have braked, accelerated, or begun a turn, so the region of plausible positions balloons, and the filter's confidence in any single point drops accordingly. The planner treats that whole uncertain region as occupied and keeps clearance from all of it. When the cyclist emerges past the truck's front bumper, the fresh detection lands inside the predicted ellipse, the filter snaps its uncertainty back down, and tracking resumes — no new entity minted, no history lost, no phantom certainty about where the cyclist "must" be.
How it works¶
- Anchor on the last confirmed observation. Freeze the last real measurement — position, velocity, appearance, timestamp — as the initial condition; everything downstream is propagated from it, never from a guess.
- Propagate through a dynamics model. Push the state forward with whatever motion model the domain justifies: constant-velocity, a physical envelope of reachable positions, or a learned trajectory prior.
- Grow the uncertainty on a clock. Inflate the covariance (or bound) with elapsed time and process noise, so the estimate degrades gracefully rather than staying falsely sharp.
- Correct on re-detection. When a real observation returns, fuse it with the prediction — the classic predict-then-update loop — and contract the uncertainty back toward observation-grade.
Tuning parameters¶
- Process-noise / dynamics aggressiveness — how fast the envelope grows per second. Too tight and the filter stays overconfident and loses the entity when it maneuvers; too loose and the prediction becomes so vague it's useless for planning.
- Motion-model fidelity — constant-velocity versus a maneuver-aware or multi-model bank. Richer models track erratic entities better but cost computation and can overfit noise.
- Gap horizon — how long the filter keeps propagating before it gives up and hands the entity to a dormant record. Long horizons bridge real occlusions; over-long ones let a stale track compete with genuine new observations.
- Re-detection gate width — how close a new observation must fall to the prediction to be fused. Wide gates recover from drift but invite mis-association; narrow gates stay clean but drop entities that moved unexpectedly.
When it helps, and when it misleads¶
Its strength is short, physical gaps — occlusions, sensor dropouts, momentary loss — where the entity genuinely persists and its near-future state is constrained by dynamics. Here the filter is the difference between smooth continuity and a jarring forget-and-rediscover, and its honest uncertainty band lets everything downstream reason about where it might be instead of pretending to know.
Its failure mode is the overconfident or stale track. Because the filter emits a tidy distribution, it invites treating prediction as fact — the archetype's invented certainty trap — and a mis-tuned model can stay too sharp, causing track switching when two entities cross or letting a ghost track linger and steal the next real detection. The recursive machinery it descends from, the Kalman filter,[1] is optimal only when its noise and linearity assumptions hold; push it past a long gap or a hard maneuver and the pretty ellipse is fiction. The guarding discipline is to keep uncertainty growing honestly with time, cap the propagation horizon, and hand the entity to a dormant record rather than let the filter narrate a present it can no longer see.
How it implements the components¶
observation_gap_state_model— it is the gap model: a dynamics-driven propagation of state through the unobserved interval, producing an estimate the archetype demands be distinguishable from an observation.last_confirmed_observation— it pins the propagation to the last real measurement and its timestamp, so every predicted value is traceable to genuine evidence.confidence_and_uncertainty_state— the covariance/bound it carries and inflates over time is the explicit, decaying confidence the archetype requires.
It does not decide whether a returning presentation is the same entity — the identity match on reappearance is reappearance_association_gate / identity_persistence_criterion, held by Reappearance Association Protocol — and it does not judge when non-detection means the entity is gone; that calibrated absence_evidence_calibration and its retirement_rule trigger belong to Absence-Evidence Calibration Test. The filter propagates state; it neither matches returns nor retires.
Related¶
- Instantiates: Entity Persistence Across Observation Gaps — supplies the active, state-propagating continuity hypothesis for short physical gaps.
- Sibling mechanisms: Dormant Entity Registry · Absence-Evidence Calibration Test · Reappearance Association Protocol · Multi-Observer Sighting Reconciliation · Persistent Identifier Resolver
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Predictive State Filter operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it carries an entity's state forward through an observation gap as a probability distribution anchored on the last confirmed sighting, widening the uncertainty envelope as time passes so the estimate never masquerades as an observation.
Independent corroboration: The frozen evidence defines Predictive State Filter as 'Carries an entity's state forward through an observation gap as a probability distribution anchored on the last confirmed sighting, widening the uncertainty envelope as time passes so the estimate never masquerades as an observation', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Propagating a state estimate and widening uncertainty between observations descends from control and navigation engineering, especially Kalman filtering.
Related originating lineages:
- Robotics & Automation — Tracking and localization materially developed practical use across sensor gaps.
- Statistics & Experimental Design — Statistics supplies Bayesian state-space inference and uncertainty propagation.
Review resolution: Both blind reviewers agree that engineering design is the primary origin. Reconciliation resolves alternate origin disagreement. Formative alternate lineages are retained as statistics_experimental_design, robotics_automation; later breadth of use is recorded separately as domain_reach=multi_domain, while origin_mode=cross_disciplinary_synthesis describes the relationship among origin lineages.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The filter and the Dormant Entity Registry are two answers to the same gap, chosen by gap length and dynamics: the filter predicts a current state when the entity's near-future is physically constrained and the gap is short; the registry declines to predict when it isn't. A well-built system hands an entity from one to the other as a propagated track ages past the horizon where its dynamics still constrain anything.
References¶
[1] The Kalman filter is the classic recursive predict-then-update estimator (R. E. Kálmán, 1960) that maintains a mean and covariance for a linear-Gaussian system; extended and unscented variants handle nonlinear dynamics. It is used here as the established archetype of state-propagating estimators, not as a cited result — and its optimality guarantees hold only while its modeling assumptions do. withdrawn registry ↩