Lemma A.9 — conditional squared Hellinger distance is at most four times joint distance
ProvedStatComplexityDM.TabularPS.hellinger_cond_le_four_jointchange-of-measurehellinger-distancep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
Let and be two joint probability laws on finite alphabets , with -marginals and conditional kernels , . Then
This compares the conditional discrepancies used layer by layer with the discrepancy of the joint observation.
Formalization Note All probability laws are vectors on finite alphabets. A conditional kernel is supplied even at an to which its marginal assigns zero mass; those values do not affect the left side.
Preamble
import Mathlib import Definitions.Def_StatComplexityDM_TabularPS_MDP
Formal statement
namespace StatComplexityDM.TabularPS
/-- Lemma A.9, p. 71, on finite alphabets. -/
theorem hellinger_cond_le_four_joint {X Y : Type*} [Fintype X] [Fintype Y]
(PX QX : X → ℝ) (K L : X → Y → ℝ)
(hPX : StatComplexityDM.LowerBound.IsDist PX) (hQX : StatComplexityDM.LowerBound.IsDist QX)
(hK : ∀ x, StatComplexityDM.LowerBound.IsDist (K x)) (hL : ∀ x, StatComplexityDM.LowerBound.IsDist (L x)) :
∑ x, PX x * FoundationsRL.GeneralDM.hellingerSq (K x) (L x) ≤
4 * FoundationsRL.GeneralDM.hellingerSq
(fun xy : X × Y => PX xy.1 * K xy.1 xy.2)
(fun xy : X × Y => QX xy.1 * L xy.1 xy.2) := by sorry
end StatComplexityDM.TabularPS
Source
arXiv:2112.13487v3, Lemma A.9, p. 71
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.