Proposition C.2(iii), p. 34 — D_h(p, x_j) → D_h(p, x) whenever x_j → x in Δ_p
ProvedReinfRegGames.StrictStable.proposition_C_2_iiibregman-divergenceconvex-analysisp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1penalty-function
Let be a -strongly convex penalty function on and . is the one-sided derivative (C.3) and is the Bregman divergence (C.4), which may equal when . is the union of the relative interiors of the faces of that contain (C.5).
If with all and in , then
Continuity of on is what makes the set in the proof of Theorem 5.2 a neighbourhood of the vertex (the case ).
Formalization Note The one-sided derivative is encoded as the predicate "the difference quotient tends to as " with real, and as ; no real-valued function with a junk value stands for . By part (i) the derivatives at and at exist; the statement takes them as data with the defining limit as hypothesis and concludes .
Preamble
import Mathlib import Definitions.Def_agt_games import Definitions.Def_ReinfRegGames_Extinction_Model
Formal statement
namespace ReinfRegGames.StrictStable
theorem proposition_C_2_iii {B : Type*} [Fintype B] (h : (B → ℝ) → ℝ) (K : ℝ) (hpen : ReinfRegGames.Extinction.IsPenalty h K)
(p : B → ℝ) (hp : p ∈ stdSimplex ℝ B)
(xs : ℕ → B → ℝ) (ds : ℕ → ℝ) (x : B → ℝ) (d : ℝ)
(hxs : ∀ j, xs j ∈ stdSimplex ℝ B ∧ ∀ α, 0 < p α → 0 < xs j α)
(hx : x ∈ stdSimplex ℝ B ∧ ∀ α, 0 < p α → 0 < x α)
(hds : ∀ j, Filter.Tendsto (fun t : ℝ => (h (xs j + t • (p - xs j)) - h (xs j)) / t) (nhdsWithin 0 (Set.Ioi 0)) (nhds (ds j)))
(hd : Filter.Tendsto (fun t : ℝ => (h (x + t • (p - x)) - h x) / t) (nhdsWithin 0 (Set.Ioi 0)) (nhds d))
(hlim : Filter.Tendsto xs Filter.atTop (nhds x)) :
Filter.Tendsto (fun j => h p - h (xs j) - ds j) Filter.atTop (nhds (h p - h x - d)) := by sorry
end ReinfRegGames.StrictStable
Source
Mertikopoulos & Sandholm, Learning in games via reinforcement and regularization, arXiv:1407.6267v2, pp. 33–34, Proposition C.2(iii), with (C.3)–(C.5)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.