Ranked descent for a duplicate-free Pareto-cell cover
OpenBanditAlgorithm.partial_monitoring_duplicate_free_pareto_cover_ranked_descentbanditsconvex-geometrygraph-theorypartial-monitoring
Let be a nonempty, duplicate-free collection of Pareto-optimal actions in a finite partial-monitoring game with at least one outcome. Then, for every mixed outcome in the probability simplex, there is a root action and a natural-number rank such that every non-root action has a neighbouring successor satisfying
Thus the Pareto-cell adjacency graph admits an acyclic in-tree oriented toward a -optimal root, including boundary mixtures where several actions tie. This is the geometric connectivity and monotone-tree component used by partial-monitoring estimators.
Formalization Note Duplicate-freeness is expressed extensionally by equality of all loss coordinates implying equality of actions; the decreasing rank makes termination explicit.
Preamble
import Definitions.Def_PartialMonitoringGame open scoped BigOperators
Formal statement
theorem BanditAlgorithm.partial_monitoring_duplicate_free_pareto_cover_ranked_descent
{k d : ℕ} {𝕊 : Type*}
(G : PartialMonitoringGame k d 𝕊) (hd : 0 < d)
(S : Finset (Fin k)) (hSne : S.Nonempty)
(hpareto : ∀ a ∈ S, ParetoOptimalAction G a)
(hunique : ∀ a ∈ S, ∀ b ∈ S,
(∀ i, G.L a i = G.L b i) → a = b) :
∀ lam : Fin d → ℝ, lam ∈ stdSimplex ℝ (Fin d) →
∃ root ∈ S, ∃ rank : Fin k → ℕ,
∀ b ∈ S, b ≠ root →
∃ c ∈ S, NeighbouringActions G b c ∧
∑ i : Fin d, G.L c i * lam i ≤
∑ i : Fin d, G.L b i * lam i ∧
rank c < rank b := by
sorrySource
Tor Lattimore and Csaba Szepesvári, Bandit Algorithms, Cambridge University Press (2020), Lemma 37.7 p. 484, Exercise 37.10 p. 509, and Lemma 37.21 pp. 501–502, especially the finite-subsequence boundary argument. https://tor-lattimore.com/downloads/book/book.pdf