Product weights bound entropy with an exact partition function
Provedmme_entropy_penalty_le_product_partitionmatrix-multiplicationtensor-extraction
Arbitrary positive coordinate weights give an entropy-penalty upper bound with the logarithm of their exact partition function. Normalizing one coordinate proves the result without approximate normalization assumptions. 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_entropy_penalty_le_product_partition
{half : ℕ} {parent : Fin 3 → ℕ} (alpha : Split half parent → ℝ)
(ha : ∀ c, 0 ≤ alpha c) (hprob : ∑ c, alpha c = 1)
(weights : Fin 3 → Fin (half + 1) → ℝ)
(hpos : ∀ i j, 0 < weights i j)
(Z : ℝ) (hZ : 0 < Z)
(hpartition : ∑ c : Split half parent, ∏ i, weights i (c.val i) = Z) :
Real.log 2 * entropyPenalty alpha ≤
Real.log Z -
(∑ i, ∑ j, mme_modern_marginal (fun c : Split half parent => c.val i) alpha j *
Real.log (weights i j)) - entropy alpha := by sorrySource
Checked tensor restrictions and released global histogram extraction.