(3.13), p. 12 — I_d(α) ≥ (1/(2κ²))|α − ᾱ_d|²
ProvedSAARate.SharpLD.display_3_13convex-analysislarge-deviationsp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1rate-function
Let be a real random variable on a probability space , let , and suppose almost surely. Let and let
be the rate function (Cramér transform) of , as in (3.5). With ,
In the paper for a direction , , and . The quadratic lower bound is what turns the sharpness constant into the exponential rate .
Formalization Note is the published extended-real Legendre transform legendre of the extended-real log-MGF logMGF; for one has and the inequality is the page's statement there, not a junk case. The statement is made for an arbitrary bounded measurable .
Preamble
import Mathlib import Definitions.Def_BellWilliams2001_ThresholdPolicy_LargeDeviation open MeasureTheory ProbabilityTheory Filter Topology BellWilliams2001.ThresholdPolicy
Formal statement
namespace SAARate.SharpLD
theorem display_3_13 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω)
[IsProbabilityMeasure P] (X : Ω → ℝ) (hX : Measurable X) (κ : ℝ) (hκ : 0 < κ)
(hXκ : ∀ᵐ ω ∂P, |X ω| ≤ κ) :
∀ α : ℝ, (((1 / (2 * κ ^ 2)) * |α - ∫ ω, X ω ∂P| ^ 2 : ℝ) : EReal) ≤
legendre (logMGF P X) α := by sorry
end SAARate.SharpLD
Source
Shapiro & Homem-de-Mello, On Rate of Convergence of Optimal Solutions of Monte Carlo Approximations of Stochastic Programs, preprint (SPEPS copy, edoc.hu-berlin.de), p. 12, proof of Theorem 3.1, (3.5), (3.13)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.