Moriba Jah, University of Texas Austin; Van Haslett, University of Texas Austin
https://doi.org/10.64861/ONYH2737
Keywords: Space Domain Awareness, Space Situational Awareness, State Estimation, Orbit Determination, Non-Bayesian Filtering, Possibility Theory, Epistemic Uncertainty, Hypothesis Testing, Model Mismatch, Sparse Observations, GEO Tracking, Angles-Only Measurements,
Abstract:
A probabilistic filter can become increasingly confident while becoming increasingly wrong. When model assumptions are violated by an unmodeled maneuver, an undisclosed sensor bias, or dynamics that are bounded rather than stochastic, the covariance continues to shrink. The filter interprets incoming observations as confirmation of its current trajectory because that is what its algebra demands. It has no mechanism for representing contradiction; normalization ensures that evidence always increases the weight of some hypothesis rather than eliminating hypotheses outright. The confidence it reports is therefore an artifact of its assumptions, not a property of the data. It fails silently.
We introduce the Epistemic Support-Point Filter (ESPF), a non-Bayesian recursive estimator built on a different principle: confidence should emerge only from survival under contradiction, not from the algebra of normalization. The ESPF maintains a finite set of explicit hypotheses spanning an admissible region of state space. Evidence acts by falsifying hypotheses whose predicted observations are geometrically incompatible with what was observed. Crucially, the ESPF admits the existence of mutually incompatible evidence without forcing reconciliation: a falsified hypothesis is eliminated, and nothing is renormalized to compensate. What survives is what has not yet been contradicted. Compatibility encodes non-contradiction, not belief, frequency, or calibration, and therefore cannot be reinterpreted probabilistically. When the model is violated, the ESPF fails informatively: surprisal spikes, necessity surges, and the Epistemic Width Monitor signals that the filter is operating outside its warrant, diagnostics with no probabilistic analog.
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness
Best Paper Award Winner 2026