(3.11), p. 11 — |Λ″(t)| = |E[X²e^{tX}]/E[e^{tX}] − Λ′(t)²| ≤ |κ² − Λ′(t)²| ≤ κ²
ProvedSAARate.SharpLD.display_3_11large-deviationsmoment-generating-functionp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1probability
Let be a real random variable on a probability space with almost surely, and let . Then for every
The second derivative of the cumulant generating function is the variance of under the exponentially tilted law, and (3.11) bounds it by . Through (3.12) this makes Lipschitz with constant , which is what gives the conjugate strong convexity modulus .
Formalization Note The statement is the conjunction of the identity and the two printed inequalities; is Mathlib's cgf and , are iterated deriv. It is stated for an arbitrary bounded measurable .
Preamble
import Mathlib open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
namespace SAARate.SharpLD
theorem display_3_11 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω)
[IsProbabilityMeasure P] (X : Ω → ℝ) (hX : Measurable X) (κ : ℝ)
(hXκ : ∀ᵐ ω ∂P, |X ω| ≤ κ) :
∀ t : ℝ,
deriv (deriv (cgf X P)) t =
(∫ ω, X ω ^ 2 * Real.exp (t * X ω) ∂P) / mgf X P t - (deriv (cgf X P) t) ^ 2 ∧
|(∫ ω, X ω ^ 2 * Real.exp (t * X ω) ∂P) / mgf X P t - (deriv (cgf X P) t) ^ 2| ≤
|κ ^ 2 - (deriv (cgf X P) t) ^ 2| ∧
|κ ^ 2 - (deriv (cgf X P) t) ^ 2| ≤ κ ^ 2 := 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. 11, proof of Theorem 3.1, (3.11)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.