Whole-transcript majority amplification with one total-variation error
ProvedInfoLossLAR.WorstCase.BDD.amplification_joint_tvLet independent ideal trials return complete candidate secrets, with each trial correct with probability at least . Let be any failure event such that failure implies at least of the complete candidates are incorrect. Suppose the law of the actual full -trial transcript is within total variation of the full independent ideal product law. Then
The actual transcript may contain arbitrary correlations. The bound charges total variation once for the whole joint transcript; it assumes neither independence of actual trials nor conditional independence given a shared sample pool. The theorem is useful for amplifying a whole-secret decoder while controlling joint source replacement.
Formalization Note. Finite outcome laws are AFT.FiniteDist; readout returns complete secrets rather than coordinatewise votes. The predicate F may be any decidable event satisfying the stated failure condition.
import Definitions.Def_InfoLossLAR_LoadedBatchBundle_05 set_option autoImplicit false open AFT open scoped BigOperators
theorem InfoLossLAR.WorstCase.BDD.amplification_joint_tv {Outcome Secret : Type*} [Fintype Outcome] [DecidableEq Secret]
(trials : ℕ) (law : Fin trials → FiniteDist Outcome)
(readout : Fin trials → Outcome → Secret) (secret : Secret)
(fail : (Fin trials → Outcome) → Prop) [DecidablePred fail]
(hsuccess : ∀ i, (3/4 : ℝ) ≤ ∑ x, if readout i x = secret then (law i).weight x else 0)
(hfail : ∀ outcomes, fail outcomes → trials ≤ 2 *
(∑ i, if readout i (outcomes i) = secret then 0 else 1))
(actual : FiniteDist (Fin trials → Outcome)) (jointError : ℝ)
(hjoint : FiniteDist.TVBound actual (FiniteDist.independent law) jointError) :
(∑ outcomes, if fail outcomes then actual.weight outcomes else 0) ≤
(7/8 : ℝ)^trials + jointError := by sorry