Zero Bayes Residual Criterion
Abstract
Equal-prior statistical residual vanishes exactly when the transcript laws are mutually singular.
Definition 1.1 (Statistical residual is half the common mass).
Formalization. D5/S3/Observer/MeasureSeparation/ZeroBayesResidualCriterion.statisticalResidual (✓ std3).
Source. Repository-derived.
Commentary.
The infimum of two measures is their canonical common mass. For two probability laws, half of its mass on the full transcript space is the equal-prior optimal binary error, equivalently the Le Cam one-minus-total-variation formula.
Theorem 1.2 (Zero residual is equivalent to mutual singularity).
Proof. Machine-checked in Lean as D5/S3/Observer/MeasureSeparation/ZeroBayesResidualCriterion.statistical_residual_eq_zero_iff_mutually_singular (✓ std3). ∎
Source. Repository-derived.
Commentary.
The transcript carrier is arbitrary and both state-indexed laws are probability measures on its measurable structure.
Zero residual is zero total common mass. That is exactly a zero measure infimum, hence lattice disjointness; the pinned Mathlib equivalence identifies disjoint measures with mutually singular measures.
References
- Truth anchor:
D5/S3/Observer/MeasureSeparation/ZeroBayesResidualCriterion.statisticalResidual - Truth anchor:
D5/S3/Observer/MeasureSeparation/ZeroBayesResidualCriterion.statistical_residual_eq_zero_iff_mutually_singular