(9.22), p. 530 — matrix KL conjugate
ProvedCompOT.Barycenter.kl_matrix_conjugateconvex-conjugateoptimal-transportp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1relative-entropy
Let be a finite matrix with strictly positive entries, and let be any real matrix of the same shape. For nonnegative matrices , define by summing the scalar generalized KL terms. Its convex conjugate is
The maximizing matrix has entries .
This is the separable transformation that turns the KL projection problem into the exponential dual objective.
Formalization Note The theorem states both the universal upper bound and attainment, so its use of “max” is literal even when an index type is empty.
Preamble
import Mathlib import Definitions.Def_CompOT_Barycenter_Defs
Formal statement
namespace CompOT.Barycenter
/-- The separable matrix KL conjugate (9.22), p. 530. -/
theorem kl_matrix_conjugate {n m : ℕ}
(hn : 0 < n) (hm : 0 < m)
(U K : Matrix (Fin n) (Fin m) ℝ) (hK : ∀ i j, 0 < K i j) :
(∀ P : Matrix (Fin n) (Fin m) ℝ, (∀ i j, 0 ≤ P i j) →
(∑ i, ∑ j, U i j * P i j) - klMatrix P K ≤
∑ i, ∑ j, K i j * (Real.exp (U i j) - 1)) ∧
(∑ i, ∑ j, U i j * (K i j * Real.exp (U i j))) -
klMatrix (fun i j => K i j * Real.exp (U i j)) K =
∑ i, ∑ j, K i j * (Real.exp (U i j) - 1) := by sorry
end CompOT.Barycenter
Source
Peyré & Cuturi, Computational Optimal Transport (FnT ML 2019), (9.22), proof of Proposition 9.1, p. 530
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.