Adaptive BDD transcript failure under one whole-law replacement
ProvedInfoLossLAR.WorstCase.BDD.adaptive_trace_tv_boundLet and be probability distributions on the same finite space of complete transcripts, with . Let be a failure event, and let be events for finitely many stages such that every transcript in lies in at least one . If
then
The events may be the first incorrect stage after a correct adaptive prefix. The full actual and reference transcript laws may each have arbitrary dependencies between stages; the total variation hypothesis is on the entire transcript and is charged once. This isolates the probability step used to combine a correct BDD trajectory with a correlated source pool.
Formalization Note. The finite laws and eventwise total-variation bound are the existing platform AFT.FiniteDist and TVBound; no algorithmic BDD implementation is asserted.
import Definitions.Def_InfoLossLAR_LoadedBatchBundle_05 set_option autoImplicit false open AFT open scoped BigOperators
theorem InfoLossLAR.WorstCase.BDD.adaptive_trace_tv_bound {Ω : Type*} [Fintype Ω] (steps : ℕ)
(actual ideal : FiniteDist Ω) (err : ℝ)
(htv : FiniteDist.TVBound actual ideal err)
(bad : Fin steps → Ω → Prop) [∀ i, DecidablePred (bad i)]
(fail : Ω → Prop) [DecidablePred fail]
(hfail : ∀ ω, fail ω → ∃ i, bad i ω)
(bound : Fin steps → ℝ)
(hstage : ∀ i, (∑ ω, if bad i ω then ideal.weight ω else 0) ≤ bound i) :
(∑ ω, if fail ω then actual.weight ω else 0) ≤ err + ∑ i, bound i := by sorry