Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proposition 4.1, pp. 425–426 — P_ε → max-entropy optimal plan and L^ε_C → L_C as ε → 0; P_ε → a ⊗ b as ε → ∞

Proved
CompOT.EntropicLimit.proposition_4_1

by mikedeng1 · Oct 8, 2026 · Mathlib 0df444a (Lean v4.33.1)

convergenceentropic-regularizationoptimal-transportp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1

Let a∈Σna \in \Sigma_na∈Σn​, b∈Σmb \in \Sigma_mb∈Σm​ be histograms and C∈Rn×mC \in \mathbb R^{n\times m}C∈Rn×m a cost matrix. For each ε>0\varepsilon > 0ε>0 let PεP_\varepsilonPε​ be the solution of the entropic problem (4.2), and let P0⋆P^\star_0P0⋆​ be the optimal coupling of the Kantorovich problem with maximal entropy (4.3). Then

  1. Pε→P0⋆P_\varepsilon \to P^\star_0Pε​→P0⋆​ as ε→0+\varepsilon \to 0^+ε→0+;
  2. LCε(a,b)=⟨C,Pε⟩−εH(Pε)→LC(a,b)=⟨C,P0⋆⟩L^\varepsilon_C(a,b) = \langle C, P_\varepsilon\rangle - \varepsilon\mathbf H(P_\varepsilon) \to L_C(a,b) = \langle C, P^\star_0\rangleLCε​(a,b)=⟨C,Pε​⟩−εH(Pε​)→LC​(a,b)=⟨C,P0⋆​⟩ as ε→0+\varepsilon \to 0^+ε→0+;
  3. Pε→a⊗b=ab⊤=(aibj)i,jP_\varepsilon \to a\otimes b = ab^\top = (a_ib_j)_{i,j}Pε​→a⊗b=ab⊤=(ai​bj​)i,j​ as ε→+∞\varepsilon \to +\inftyε→+∞ (4.4).
Pε→ε→0argmin⁡P{−H(P):P∈U(a,b), ⟨P,C⟩=LC(a,b)},Pε→ε→∞a⊗b.P_\varepsilon \xrightarrow{\varepsilon\to 0} \operatorname*{argmin}_P\{-\mathbf H(P) : P \in U(a,b),\ \langle P,C\rangle = L_C(a,b)\},\qquad P_\varepsilon \xrightarrow{\varepsilon\to\infty} a\otimes b.Pε​ε→0​Pargmin​{−H(P):P∈U(a,b), ⟨P,C⟩=LC​(a,b)},Pε​ε→∞​a⊗b.

Small regularization recovers the most diffuse exact optimal transport plan, while large regularization yields the independent coupling.

Formalization Note PεP_\varepsilonPε​ is given as any family satisfying the optimality predicate of (4.2) for each ε>0\varepsilon > 0ε>0 (its values at ε≤0\varepsilon \le 0ε≤0 are irrelevant), and P0⋆P^\star_0P0⋆​ as any matrix satisfying the predicate of (4.3); existence and uniqueness of both are separate items of this mission, so the hypotheses are satisfiable and determine the objects. Convergence of matrices is entrywise. H\mathbf HH uses 0log⁡0=00\log 0 = 00log0=0 (the book's −∞-\infty−∞ convention for zero entries is not used). Indices are 0-based.

Preamble
import Mathlib
import Definitions.Def_CompOT_EntropicLimit_Defs
Formal statement
namespace CompOT.EntropicLimit

open Filter Topology

/-- Proposition 4.1, pp. 425–426. Let `a ∈ Σ_n`, `b ∈ Σ_m`, `P_ε` the solution of (4.2) for
each `ε > 0`, and `P₀` the maximal-entropy optimal coupling (4.3). Then `P_ε → P₀` as
`ε → 0⁺`, `L^ε_C(a,b) = ⟨C,P_ε⟩ - ε H(P_ε) → L_C(a,b) = ⟨C,P₀⟩` as `ε → 0⁺`, and
`P_ε → a ⊗ b` as `ε → +∞` (4.4). -/
theorem proposition_4_1 {n m : ℕ} (C : Matrix (Fin n) (Fin m) ℝ)
    (a : Fin n → ℝ) (b : Fin m → ℝ)
    (ha : a ∈ stdSimplex ℝ (Fin n)) (hb : b ∈ stdSimplex ℝ (Fin m))
    (Pe : ℝ → Matrix (Fin n) (Fin m) ℝ)
    (hPe : ∀ ε, 0 < ε → IsEntropicOptimal C a b ε (Pe ε))
    (P₀ : Matrix (Fin n) (Fin m) ℝ) (hP₀ : IsMaxEntropyOptimal C a b P₀) :
    Tendsto Pe (𝓝[>] 0) (𝓝 P₀) ∧
      Tendsto (fun ε => entObjective C ε (Pe ε)) (𝓝[>] 0) (𝓝 (CompOT.Assignment.frob C P₀)) ∧
      Tendsto Pe atTop (𝓝 (Matrix.of fun i j => a i * b j)) := by sorry

end CompOT.EntropicLimit
Source
Peyré & Cuturi, Computational Optimal Transport (FnT ML 2019), Proposition 4.1, pp. 425–426, eqs. (4.3)–(4.4)
Human review
  • Endorsed by Shuze Chen · Oct 8, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 8, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me