Proposition A.1(iii), p. 31 — Q_α(y) → 0 whenever y_α − y_β → −∞ for some β ≠ α
ProvedReinfRegGames.TimeAvg.proposition_A_1_iiichoice-mapp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1penalty-function
Let be a finite set, a penalty function on (Definition 2.1) and its choice map, . Fix two distinct indices in . Then the -component of the choice map vanishes when the score of falls infinitely far behind the score of :
in the uniform sense: for every there is such that for every score vector with .
This is the property that lets a large score gap be read off as a vanishing choice probability; Proposition C.5 uses it.
Formalization Note The page's limit statement ("whenever ", i.e. along every sequence) is equivalent to the uniform – form stated here. "" is the predicate IsChoice h y x, so the bound holds for every maximizer.
Preamble
import Mathlib import Definitions.Def_agt_games import Definitions.Def_ReinfRegGames_Extinction_Model
Formal statement
namespace ReinfRegGames.TimeAvg
theorem proposition_A_1_iii {B : Type*} [Fintype B] [DecidableEq B]
(h : (B → ℝ) → ℝ) (K : ℝ) (hpen : ReinfRegGames.Extinction.IsPenalty h K) (α β : B) (hαβ : β ≠ α) :
∀ ε : ℝ, 0 < ε → ∃ M : ℝ, ∀ y x : B → ℝ, ReinfRegGames.Extinction.IsChoice h y x → y α - y β < -M → x α < ε := by sorry
end ReinfRegGames.TimeAvg
Source
Mertikopoulos & Sandholm, Learning in games via reinforcement and regularization, arXiv:1407.6267v2, p. 31, Proposition A.1(iii)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.