Remark 4.21, (4.33)–(4.34) — coordinate gradients of the entropic dual
ProvedCompOT.EntropicDual.remark_4_21_gradientsentropygradientoptimal-transportp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
For probability histograms , , a cost matrix , and , let be the entropic dual objective. At any potentials and , its partial derivatives are
These formulas identify the row and column marginal errors of the scaled Gibbs matrix and underlie the log-domain updates.
Formalization Note Each partial derivative is a one-variable HasDerivAt statement obtained by changing one coordinate with Function.update.
Preamble
import Mathlib import Definitions.Def_CompOT_EntropicDual_Defs
Formal statement
namespace CompOT.EntropicDual
/-- Remark 4.21, (4.33)–(4.34), p. 449: the two partial gradients of
the entropic dual objective, stated coordinatewise. -/
theorem remark_4_21_gradients {n m : ℕ} (C : Matrix (Fin n) (Fin m) ℝ)
(a : Fin n → ℝ) (b : Fin m → ℝ) (ε : ℝ)
(hn : 0 < n) (hm : 0 < m) (hε : 0 < ε)
(ha : a ∈ stdSimplex ℝ (Fin n)) (hb : b ∈ stdSimplex ℝ (Fin m))
(f : Fin n → ℝ) (g : Fin m → ℝ) :
(∀ i, HasDerivAt
(fun t : ℝ => dualObjEnt C a b ε (Function.update f i t) g)
(a i - Real.exp (f i / ε) *
∑ j, gibbs C ε i j * Real.exp (g j / ε)) (f i)) ∧
(∀ j, HasDerivAt
(fun t : ℝ => dualObjEnt C a b ε f (Function.update g j t))
(b j - Real.exp (g j / ε) *
∑ i, gibbs C ε i j * Real.exp (f i / ε)) (g j)) := by sorry
end CompOT.EntropicDual
Source
Peyré & Cuturi, Computational Optimal Transport (FnT ML 2019), Remark 4.21, (4.33)–(4.34), p. 449. https://doi.org/10.1561/2200000073
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.