Corollary 3, p. 67 — a σ-sub-Gaussian variable satisfies (209) and hence (210)
ProvedInfoRelaxTS.Gap.corollary_3concentrationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1probabilitysub-gaussian
Let be a real random variable on a probability space and . Suppose is -sub-Gaussian with variance proxy :
Then satisfies the tail condition of Lemma 4,
and hence
This is the form in which Lemma 4 is used for the sub-Gaussian mean-reward metrics of Lemma 5.
Formalization Note Definition 3 (p. 66) prints the bound ; every use in the paper ((213), (214) with , (236)) has , and with the printed form the corollary would fail for . The statement uses Mathlib's HasSubgaussianMGF for with parameter , which also records that is integrable for every .
Preamble
import Mathlib open MeasureTheory ProbabilityTheory
Formal statement
namespace InfoRelaxTS.Gap
theorem corollary_3 {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
(X : Ω → ℝ) (σ : ℝ) (hσ : 0 < σ)
(hX : HasSubgaussianMGF (fun ω => X ω - ∫ x, X x ∂μ) ⟨σ ^ 2, sq_nonneg σ⟩ μ) :
(∀ z : ℝ, 0 ≤ z →
(μ {ω | (∫ x, X x ∂μ) + z * σ ≤ X ω}).toReal ≤ Real.exp (-z ^ 2 / 2)) ∧
(∀ z : ℝ, 0 < z →
∫ ω, max (X ω - ((∫ x, X x ∂μ) + z * σ)) 0 ∂μ ≤ σ / z * Real.exp (-z ^ 2 / 2)) := by sorry
end InfoRelaxTS.Gap
Source
Min, Maglaras, Moallemi, Thompson Sampling with Information Relaxation Penalties, arXiv:1902.04251v2, pp. 66–67, Definition 3 (208), Corollary 3
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.