From Retirable Recovery State to Finite-Suffix Certificates: Bounded-Memory Assurance and Model-Uncertainty Boundaries for Finite-State Bayesian Filters
Year: 2026
Type: preprint
Abstract: Finite-state Bayesian filters can provide different bounded-memory guarantees depending on the model and the assurance target. For a known post-restart model with positive length-L observation products, block Hilbert contraction ρ < 1, local implementation error q, and bank envelope E = Lq/(1 − ρ), deleting an all-prior vertex bank once D_t* + E + (L − 1)q ≤ r preserves posterior total-variation error at most τ = r/4 for every future time and forgotten prior. A phase-reset scalar handoff retains adaptive bounds; q(L/(1 − ρ) + L − 1) ≤ r is a necessary feasibility condition for these sufficient guards. Across 32 matched episodes (8,192 predictions) of the four-state original4 filter, the retained hybrid matches all 7,946 permanent-bank current-action certificates (97.00%) with 65.54% fewer kernel calls and a 16 B final per-stream serialized payload; full deletion yields 7,938 action certificates in 10 B. These action counts include the active-bank phase; the persistent posterior guarantee applies after the retirement gate. A recorded-likelihood HMM-MOS replay yields 313 certificates, alongside conditional dynamic-set synchronization and sufficient Q_n continuation contracts. Under parameter uncertainty around η_0 = 1/5, the selected bounded-drift retirement guard is infeasible above ε* ≈ 8.9031 × 10⁻⁶; this is not a general impossibility threshold. Across the same episodes, 33 nominal certificates on [0.19, 0.21] and 763 on [0.10, 0.30] admit opposite-sign endpoint/prior witnesses, including 26 and 573 post-deletion positions. As an alternative assurance target, exact scaled-Bernstein suffix tables through depth K = 12 classify all 24,573 contexts with zero unresolved slack. Complement symmetry halves the raw output table from 2,048 to 1,024 B ROM with 2 B suffix RAM, smaller than the 2,011–2,220 B bit-packed minimal Moore layouts. Multiaffine endpoint reduction shows that fixed-parameter certificates do not automatically remain sound under within-suffix drift. On the frozen trajectories, fixed-interval K = 12 action coverage is 87.79% narrow and 71.85% wide; drift-qualified K = 8 coverage is 66.53% and 26.71%. These exact finite-trace counts compare current-action availability, not posterior-recovery rates or native application performance.
Authors: Cao, Xu, Xiao, Wenxuan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233587053
Cites:
Cited by:
Related to:
