(2.46), p. 767 — Ee^{hX_j} ≤ ∫_{u<0} e^{hu} dF_j(u) + b_gj sup_{u≥0} e^{hu−g(u)}
ProvedNagaevLD.GenMoment.eq_2_46Let be a real random variable with distribution function , let , and let be finite. For , let be any upper bound of over , for example
when this supremum is finite. Then
The bound comes from splitting the expectation at and writing on . In the proof of Theorem 2.5 it is used with , for which the supremum is computed in (2.47).
Formalization Note The paper's supremum is a real supremum, which Lean would evaluate to when it is infinite; the statement is therefore quantified over every upper bound , which is equivalent when the supremum is finite. The integrability of on (finiteness of ) is assumed explicitly. The proof uses ; allowing keeps the negative part integrable, so both the displayed equality and inequality represent finite expectations. The statement does not need independence or any other property of .
import Mathlib import Definitions.Def_NagaevLD_GenMoment_Setting open MeasureTheory ProbabilityTheory
namespace NagaevLD.GenMoment
/-- (2.46), p. 767: splitting `E e^{h X_j}` at `0` and bounding the part on `X_j ≥ 0` by
`b_gj` times any upper bound `M` of `e^{hu - g(u)}` over `u ≥ 0` (the supremum of the page).
The nonnegative tilt is the one used in the proof of Theorem 2.5. -/
theorem eq_2_46 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
(n : ℕ) (X : Fin n → Ω → ℝ) (hXm : ∀ j, Measurable (X j)) (g : ℝ → ℝ) (j : Fin n)
(hbg : IntegrableOn (fun ω => Real.exp (g (X j ω))) {ω | 0 ≤ X j ω} P)
(h M : ℝ) (hh : 0 ≤ h) (hM : ∀ u : ℝ, 0 ≤ u → Real.exp (h * u - g u) ≤ M) :
(∫ ω, Real.exp (h * X j ω) ∂P) =
(∫ ω in {ω | X j ω < 0}, Real.exp (h * X j ω) ∂P) +
(∫ ω in {ω | 0 ≤ X j ω},
Real.exp (h * X j ω - g (X j ω)) * Real.exp (g (X j ω)) ∂P) ∧
(∫ ω, Real.exp (h * X j ω) ∂P)
≤ (∫ ω in {ω | X j ω < 0}, Real.exp (h * X j ω) ∂P) + bg P X g j * M := by sorry
end NagaevLD.GenMoment
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.