Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proof of Theorem 5.2, Part IV, pp. 23–24 — h*(y) − y_{α*} = max_x {Σ_{µ≠α*} x_µ z_µ − h(x)} is nondecreasing in the relative scores z_µ

Proved
ReinfRegGames.StrictStable.monotone_step

by mikedeng1 · Oct 9, 2026 · Mathlib 0df444a (Lean v4.33.1)

convex-analysisfenchel-couplinglearning-in-gamesp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Let hhh be a penalty function on Δ=Δ(B)\Delta=\Delta(B)Δ=Δ(B) with conjugate h∗(y)=max⁡x∈Δ{⟨y∣x⟩−h(x)}h^*(y)=\max_{x\in\Delta}\{\langle y|x\rangle-h(x)\}h∗(y)=maxx∈Δ​{⟨y∣x⟩−h(x)}, and fix a vertex α∗∈B\alpha^*\in Bα∗∈B. For a score vector yyy write zμ=yμ−yα∗z_\mu=y_\mu-y_{\alpha^*}zμ​=yμ​−yα∗​ (μ≠α∗\mu\neq\alpha^*μ=α∗) for the relative scores. Then

h∗(y)−yα∗=max⁡x∈Δ{∑μ≠α∗xμzμ−h(x)},h^*(y)-y_{\alpha^*}=\max_{x\in\Delta}\Big\{\sum_{\mu\neq\alpha^*}x_\mu z_\mu-h(x)\Big\},h∗(y)−yα∗​=x∈Δmax​{μ=α∗∑​xμ​zμ​−h(x)},

and consequently, for two score vectors y,y′y,y'y,y′ with zμ′<zμz'_\mu<z_\muzμ′​<zμ​ for all μ≠α∗\mu\neq\alpha^*μ=α∗,

h∗(y′)−yα∗′ ≤ h∗(y)−yα∗.h^*(y')-y'_{\alpha^*}\ \le\ h^*(y)-y_{\alpha^*}.h∗(y′)−yα∗′​ ≤ h∗(y)−yα∗​.

Since Fh(eα∗,y)=h(eα∗)+h∗(y)−yα∗F_h(e_{\alpha^*},y)=h(e_{\alpha^*})+h^*(y)-y_{\alpha^*}Fh​(eα∗​,y)=h(eα∗​)+h∗(y)−yα∗​, the Fenchel coupling to the vertex does not increase when every relative score decreases. In the proof of Theorem 5.2 this shows that an orbit started in Uε∗U^*_\varepsilonUε∗​ cannot leave it.

Formalization Note The page writes a strict inequality h∗(y′)−yα∗′<h∗(y)−yα∗h^*(y')-y'_{\alpha^*}<h^*(y)-y_{\alpha^*}h∗(y′)−yα∗′​<h∗(y)−yα∗​. For a nonsteep penalty this is false: when the maximizer is the vertex eα∗e_{\alpha^*}eα∗​ itself (e.g. the quadratic penalty with every zμ≤−1z_\mu\le-1zμ​≤−1), both sides equal −h(eα∗)-h(e_{\alpha^*})−h(eα∗​). The proof of Theorem 5.2 only needs the weak inequality (the first exit point has coupling exactly εKmin⁡/2\varepsilon K_{\min}/2εKmin​/2 by continuity, while the initial coupling is strictly smaller), so the weak inequality is stated. The maximum is written as a real sSup over the image of Δ\DeltaΔ, attained for a penalty function.

Preamble
import Mathlib
import Definitions.Def_agt_games
import Definitions.Def_ReinfRegGames_Extinction_Model
Formal statement
namespace ReinfRegGames.StrictStable

theorem monotone_step {B : Type*} [Fintype B] [DecidableEq B] (h : (B → ℝ) → ℝ) (K : ℝ) (hpen : ReinfRegGames.Extinction.IsPenalty h K)
    (αstar : B) (y y' : B → ℝ) :
    ReinfRegGames.Extinction.conj h y - y αstar =
        sSup ((fun x => (∑ μ ∈ Finset.univ.erase αstar, x μ * (y μ - y αstar)) - h x) ''
          stdSimplex ℝ B) ∧
      ((∀ μ, μ ≠ αstar → y' μ - y' αstar < y μ - y αstar) →
        ReinfRegGames.Extinction.conj h y' - y' αstar ≤ ReinfRegGames.Extinction.conj h y - y αstar) := by sorry

end ReinfRegGames.StrictStable
Source
Mertikopoulos & Sandholm, Learning in games via reinforcement and regularization, arXiv:1407.6267v2, pp. 23–24, proof of Theorem 5.2, Part IV (the sentence "More rigorously, note that F_k(x*_k, y_k) = …")
Human review
  • Endorsed by Shuze Chen · Oct 9, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 9, 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