Epistemic Support-Point Filter (ESPF)

• ESPF is a non-Bayesian recursive filter representing uncertainty through normalized possibility distributions and bounded support regions.
• It employs evidential mechanisms such as compatibility weighting, surprisal-aware pruning, and adaptive dispersion to update support points.
• The framework minimizes epistemic entropy via a falsification approach, ensuring robust state estimation in sparse or adversarial environments.

The Epistemic Support-Point Filter (ESPF) is a non-Bayesian recursive state-estimation framework in which uncertainty is represented by possibility distributions and bounded supports rather than posterior probabilities. In the 2025 formulation, ESPF is presented as a “bounded possibilistic framework for ordinal state estimation” that maintains a structured region of plausibility or non-rejection through compatibility-weighted support updates, surprisal-aware pruning, adaptive dispersion, and sparse-grid quadrature. In the 2026 formulation, ESPF is a recursive filter for normalized possibility distributions whose propagation phase enacts Jaynesian maximum entropy and whose measurement update enacts Popperian falsification, with a uniqueness theorem inside the class of epistemically admissible evidence-only filters (Jah et al., 28 Aug 2025, Jah, 10 Mar 2026).

1. Epistemological basis and intended problem class

ESPF is motivated by the claim that traditional state-estimation methods rely on probabilistic assumptions that can collapse epistemic uncertainty into scalar beliefs and thereby risk overconfidence in sparse, biased, discontinuous, adversarial, or otherwise poorly calibrated sensing environments. The framework is explicitly possibilistic rather than probabilistic: admissibility, non-rejection, compatibility, necessity, and surprisal replace posterior densities, expected-value criteria, and likelihood-weighted mass transport. The 2025 paper formulates this as a shift from posterior inference to maintaining a structured region of plausibility, while the 2026 paper formulates it as a commitment to “be quick to embrace ignorance and slow to assert certainty” (Jah et al., 28 Aug 2025, Jah, 10 Mar 2026).

This orientation leads to a sharp distinction between propagation and measurement. In propagation, ESPF spreads support as widely as the dynamics and process noise allow, under a maximum-ignorance principle. In the measurement update, it eliminates only those hypotheses falsified by evidence alone. The 2026 analysis states that any rule incorporating prior possibility into survivor ranking is strictly suboptimal and risks race-to-bottom bias; the admissible ranking must depend only on evidence, operationalized through whitened innovation scores $q_k^{(i)}$ (Jah, 10 Mar 2026).

The framework is therefore positioned neither as a Bayesian posterior filter nor as a purely set-membership method. The 2025 paper contrasts ESPF with Kalman, EKF, UKF, and particle filtering by emphasizing support geometry, falsifiability-based pruning, and ordinal plausibility, while also distinguishing it from bounded filters that lack compatibility, surprisal, and Choquet-based multi-model fusion (Jah et al., 28 Aug 2025).

2. Possibilistic state representation and support geometry

In the 2026 formalization, the filter state is a normalized possibility distribution

$$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$

defined over a finite support $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$. For each level $\alpha$, the $\alpha$-cut is

$C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$

Its geometry is summarized by a minimum-volume enclosing ellipsoid (MVEE), with volume

$V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$

The 2025 paper uses an equivalent bounded-support viewpoint: ESPF maintains a possibility distribution $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$ with support

$$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$

and, in its basic version, assumes a uniform possibility distribution over a compact convex support so that every point inside $\mathcal{S}_k$ is equally plausible and every point outside is impossible (Jah, 10 Mar 2026, Jah et al., 28 Aug 2025).

Support-points are deterministic epistemic hypotheses rather than weighted Monte Carlo particles. For a hyperrectangular support,

$$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$0

the 2025 paper defines a $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$1 pattern consisting of a center point

$$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$2

and symmetric axis points

$$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$3

with $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$4. It also generalizes support generation using Smolyak sparse grids and Clenshaw–Curtis nodes, mapping normalized sparse-grid points $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$5 into the current support (Jah et al., 28 Aug 2025).

Possibility theory supplies the algebraic backbone. The 2025 paper states normalization as $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$6, joint possibility as $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$7, marginalization as $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$8, and conditioning through Zadeh-style min-based fusion. Necessity for a set $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$9 is

$\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$0

and logarithmic surprisal is

$\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$1

These definitions are not auxiliary notation; they govern the update logic of the filter (Jah et al., 28 Aug 2025).

3. Recursive mechanics: propagation, update, pruning, and regeneration

The recursive flow given in the 2025 paper is: initialize the support from an admissible region and necessity; generate support-points using either a $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$2 deterministic pattern or sparse-grid quadrature; predict each point through $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$3; form predicted support via possibilistic propagation or Minkowski expansion; map to measurement space through $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$4; compute compatibility and residual admissibility; prune support-points using surprisal thresholding; fuse prior support and measurement compatibility with sup–min logic; extract a mode or representative estimate; estimate dispersion from the surviving support-points; adapt radius and spread using dispersion and surprisal; regenerate a new support set around the updated epistemic center; and repeat (Jah et al., 28 Aug 2025).

In the prediction step, the 2025 formulation uses bounded epistemic process noise $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$5 and pointwise propagation

$\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$6

The predicted support is then

$\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$7

equivalently $\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$8. The 2026 formulation describes the same phase as regeneration of support over the reachable state region by a max–min kernel rule,

$\{\chi^{(i)}\}_{i=1}^M \subset \mathbb{R}^n$9

and interprets this as the Jaynesian maximum-entropy phase in which support spreads as widely as the dynamics allow (Jah et al., 28 Aug 2025, Jah, 10 Mar 2026).

The measurement update admits two related parameterizations in the literature. The 2026 optimality paper computes, for each surviving hypothesis, a whitened squared innovation

$\alpha$0

a compatibility score

$\alpha$1

and a conjunctive update

$\alpha$2

The 2025 bounded-support paper defines residuals $\alpha$3, constructs a residual-compatible ellipsoid from $\alpha$4 and $\alpha$5, and then declares a point compatible when

$\alpha$6

Surprisal-aware pruning removes support-points whose surprisal exceeds a threshold, and posterior fusion uses

$\alpha$7

In both formulations, measurement never increases possibility; it only preserves or cuts down support (Jah, 10 Mar 2026, Jah et al., 28 Aug 2025).

After update, the 2025 paper estimates a geometric spread tensor

$\alpha$8

defines dispersion

$\alpha$9

and adapts the support radius and spread using $\alpha$0, a reference surprisal, and temporal decay. The regenerated kernel is

$\alpha$1

with new support-points generated from a Cholesky factorization of $\alpha$2 (Jah et al., 28 Aug 2025).

4. Optimality theory, entropy, and the Gaussian limit

The 2026 paper provides the strongest formal characterization of ESPF. It defines the admissible class $\alpha$3 of epistemically admissible recursive estimators by five conditions: possibilistic representation over a finite support; conjunctive updates satisfying $\alpha$4; geometric non-degeneracy through bounded MVEE eigenvalues; commitment point admissibility, meaning the estimate $\alpha$5 must be one of the support points; and evidence referencing, meaning survivor ranking must depend only on innovations $\alpha$6, not on prior possibility values (Jah, 10 Mar 2026).

Its optimality criterion is possibilistic minimax entropy. For a possibility distribution $\alpha$7,

$\alpha$8

This decomposes as

$\alpha$9

with the integral term non-positive. The paper interprets $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$0 as support entropy and the integral term as gradient entropy, so the ignorance functional depends on both the size of the admissible region and the flatness of the possibility field inside it. In the worst case, minimizing ignorance is equivalent to minimizing $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$1 of the surviving support (Jah, 10 Mar 2026).

Three lemmas structure the proof. The Possibilistic Entropy Lemma identifies $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$2 as the correct ignorance functional and states, among other properties, that uniform possibility yields $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$3, that concentration lowers entropy, and that entropy tends to $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$4 as possibility collapses to a singleton. The Possibilistic Cramér–Rao Lemma gives a lower bound on entropy reduction per measurement: $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$5 The Evidence-Optimality Lemma then proves that, among all evidence-only selection and assignment pairs with fixed survivor count, the unique minimizer selects the hypotheses with the smallest $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$6 and assigns $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$7. This uniquely minimizes both integrated cut-volume entropy and $C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$8 within the evidence-only class (Jah, 10 Mar 2026).

The main theorem states that, in the falsification regime

$C_\alpha=\{i:\pi^{(i)}\ge \alpha\}.$9

ESPF is the unique filter in $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$0 that minimizes $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$1 subject to the PCRB constraint. The theorem further states that ESPF achieves the bound with equality and that any alternative ranking that incorporates prior possibility fails to minimize the cut volumes and is strictly worse. The paper distinguishes this falsification regime from a diffusion regime, in which propagation should maximize spread and the minimax update criterion is not operative (Jah, 10 Mar 2026).

The Gaussian-limit relation is formulated differently in the two ESPF papers. The 2026 paper states that the Kalman filter is recovered in the Gaussian epistemic limit, where $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$2, $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$3, and minimizing possibilistic entropy becomes equivalent to minimizing $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$4. The 2025 paper states that if possibility distributions are interpreted as Gaussian densities, noise is additive Gaussian, fusion is done probabilistically, and sigma-points are chosen via the unscented transform, then ESPF reduces exactly to the UKF in both prediction and measurement update (Jah, 10 Mar 2026, Jah et al., 28 Aug 2025).

5. Multi-model inference, diagnostics, and empirical evaluations

The 2025 paper extends the framework to multi-model inference through a maxitive capacity and the Choquet integral. For a finite hypothesis set $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$5 and possibility distribution $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$6,

$V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$7

and the inverse surprisal relation is written as

$V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$8

For a real-valued function $V_\alpha = c_n\bigl(\det \Pi_\alpha\bigr)^{1/2}.$9, the Choquet integral is

$\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$0

which, in the special case $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$1 and $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$2, simplifies to $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$3. The paper uses this construction to describe non-additive aggregation across competing hypotheses or experts without collapsing ordinal structure into linear weights (Jah et al., 28 Aug 2025).

Empirical reports are given in two distinct forms. The 2025 bounded-support paper compares ESPF against the Unscented Kalman Filter in orbital estimation scenarios deliberately chosen to violate Gaussian assumptions. It considers a LEO satellite of 2000 kg and $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$4 area observed over 6 days from Arecibo, Kwajalein, and Diego Garcia, and a GEO satellite of 1000 kg whose area changes from $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$5 to $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$6 mid-track over 180 minutes with 15-second measurements. Reported results are: LEO nominal, UKF RMS 1.05 km versus ESPF RMS 0.97 km with average surprisal 0.01 and necessity retention 84%; LEO with bias, UKF RMS 1.44 km versus ESPF RMS 0.73 km with average surprisal 0.10 and necessity retention 88%; and GEO with area change, UKF RMS 4.30 km versus ESPF RMS 1.09 km with average surprisal 0.098 and necessity retention 91% (Jah et al., 28 Aug 2025).

The 2026 paper reports a separate numerical validation over a 2-day, 877-step Smolyak Level-3 orbital tracking run with state dimension $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$7, $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$8 support points, process noise $\pi_{x_k}(x):\mathbb{R}^n\to[0,1]$9, measurement noise $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$0 and $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$1, and the same three ground stations. In the nominal run, $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$2 stayed positive throughout, around $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$3–$$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$4 nats, so the filter remained in the diffusion regime; the Epistemic Width Monitor reported $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$5–$$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$6, prune count around $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$7–$$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$8, steady-band $$\mathcal{S}_k = \{ x \in \mathbb{R}^n \mid \pi_{x_k}(x) > 0 \},$$9, necessity near zero, low stationary surprisal, and $\mathcal{S}_k$0 near $\mathcal{S}_k$1–$\mathcal{S}_k$2. In the stress run, with a $\mathcal{S}_k$3 m/s crosstrack maneuver at day 1.0 and a $\mathcal{S}_k$4 m Arecibo range bias active from day 0, $\mathcal{S}_k$5 remained positive, roughly $\mathcal{S}_k$6–$\mathcal{S}_k$7 nats, but prune count rose to $\mathcal{S}_k$8–$\mathcal{S}_k$9, necessity to $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$00–$$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$01, and surprisal produced peaks around $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$02 and $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$03. The paper’s interpretation is that the MVEE sign alone is too coarse and that necessity and surprisal are earlier-warning signals of model–reality divergence (Jah, 10 Mar 2026).

6. Related frameworks, boundaries, and recurrent misconceptions

ESPF is sometimes conflated with other uses of “epistemic” in logic, but the term is not used uniformly across arXiv. The 2023 paper on Depth-Bounded Epistemic Logic does not define ESPF; it defines DBEL and DPAL, where agent knowledge is gated by modal depth through atoms $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$04 and $$\pi:\{1,\ldots,M\}\to(0,1],\qquad \max_i \pi^{(i)}=1,$$05, and public announcements consume depth. That work can be read as a bounded epistemic filtering or approximation framework, but its subject is modal-depth-limited reasoning, not recursive possibilistic state estimation (Arthaud et al., 2023).

A second source of confusion comes from epistemic answer set programming. The 2021 paper “Refining the Semantics of Epistemic Specifications” does not mention “Epistemic Support-Point Filter” directly. Its nearest mechanisms are foundedness, epistemic splitting, subjective constraint monotonicity, knowledge-minimization via maximality among epistemic models, and the selection of reflexive autoepistemic equilibrium models. The 2019 paper on Founded (Auto)Epistemic Equilibrium Logic likewise does not define ESPF; it studies foundedness, epistemic splitting, FEEL, and FAEEL, proving that FAEEL is the first semantics shown to satisfy both foundedness and epistemic splitting (Su, 2021, Fandinno, 2019).

This suggests a broader terminological pattern, but only as an interpretation. A plausible implication is that “epistemic” names a family of mechanisms that restrict admissible conclusions or hypotheses by explicit criteria of support, evidence, depth, or foundedness. Within that broader family resemblance, however, ESPF has a specific meaning: a support-point-based possibilistic estimator whose uncertainty object is a bounded region of non-rejection and whose update logic is based on compatibility, min/sup fusion, pruning, and, in the 2026 theory, evidence-only minimax entropy (Jah et al., 28 Aug 2025, Jah, 10 Mar 2026).