Smoothed integer counts certify entropy-penalty bounds
Provedmme_entropy_penalty_le_smoothed_countsmatrix-multiplicationregional-extraction
Adding one to every entry of each of two nonnegative integer count vectors and normalizing by the total plus the alphabet size gives a valid entropy-penalty bound. Zero entries and zero total are allowed. Only the two exact count-total identities are required; numerical logarithm bounds remain separate. The full matrix exponent bound remains a separate obligation.
Preamble
import Theorems.Thm_mme_entropy_penalty_le_pair_reference open scoped BigOperators open MME.RegionRate MME.RecursiveThinSplit
Formal statement
theorem mme_entropy_penalty_le_smoothed_counts
{half : ℕ} {parent : Fin 3 → ℕ} (alpha : Split half parent → ℝ)
(ha : ∀ c, 0 ≤ alpha c) (hprob : ∑ c, alpha c = 1)
(i j : Fin 3) (hij : i ≠ j) (D : ℕ)
(a b : Fin (half + 1) → ℕ) (haD : ∑ v, a v = D) (hbD : ∑ v, b v = D) :
Real.log 2 * entropyPenalty alpha ≤
-(∑ v, mme_modern_marginal (fun c : Split half parent => c.val i) alpha v *
Real.log (((a v : ℝ) + 1) / ((D : ℝ) + (half + 1)))) -
(∑ v, mme_modern_marginal (fun c : Split half parent => c.val j) alpha v *
Real.log (((b v : ℝ) + 1) / ((D : ℝ) + (half + 1)))) - entropy alpha := by sorrySource
Kernel-checked regional count identities and exact extraction with empty regions.