Certified logarithms bound product-dual entropy penalties
Provedmme_rational_product_dual_penalty_certificatematrix-multiplicationtensor-extraction
Positive normalized rational coordinate weights and certified logarithm bounds give a rational upper bound for the maximum-entropy penalty. Zero distribution atoms are allowed. The root exponent bound remains a separate obligation.
Preamble
import Theorems.Thm_mme_entropy_penalty_le_product_dual open scoped BigOperators open MME.RegionRate MME.RecursiveThinSplit
Formal statement
theorem mme_rational_product_dual_penalty_certificate
{half : ℕ} {parent : Fin 3 → ℕ}
(alpha : Split half parent → ℚ)
(ha : ∀ c, 0 ≤ alpha c) (hprob : ∑ c, alpha c = 1)
(weights : Fin 3 → Fin (half + 1) → ℚ)
(hpos : ∀ i j, 0 < weights i j)
(hmass : ∑ c : Split half parent, ∏ i, weights i (c.val i) ≤ 1)
(weightLower : Fin 3 → Fin (half + 1) → ℚ)
(alphaUpper : Split half parent → ℚ)
(hwlog : ∀ i j, (weightLower i j : ℝ) ≤ Real.log (weights i j : ℝ))
(halog : ∀ c, 0 < alpha c → Real.log (alpha c : ℝ) ≤ (alphaUpper c : ℝ))
(bound : ℚ) :
let m := fun i j => ∑ c : {c : Split half parent // c.val i = j}, alpha c.val;
-(∑ i, ∑ j, m i j * weightLower i j) + ∑ c, alpha c * alphaUpper c ≤ bound →
Real.log 2 * entropyPenalty (fun c => (alpha c : ℝ)) ≤ (bound : ℝ) := by sorrySource
Checked tensor restrictions and released global histogram extraction.