(2.4) — for class [G]
ProvedAzumaWeightedSums.IteratedLog.mgf_weightedSum_leLet be a sequence of random variables on a probability space with a filtration , belonging to class [G] with for all : martingale differences with a.s. for every real . Then for every , every real sequence and every real ,
This is the moment-generating-function form of the Azuma–Hoeffding inequality: a weighted sum of conditionally sub-Gaussian martingale differences is sub-Gaussian with variance proxy . It is the first step of the proof of Lemma 2.
Formalization Note The expectation is a lower Lebesgue integral of with values in , so the inequality also asserts that the left side is finite; a Bochner integral would be for a non-integrable integrand and make the bound trivially true.
import Mathlib import Definitions.Def_AzumaWeightedSums_IteratedLog_ClassG import Definitions.Def_AzumaWeightedSums_IteratedLog_WeightedSums open MeasureTheory ProbabilityTheory Filter Topology
namespace AzumaWeightedSums.IteratedLog
/-- Azuma 1967, display (2.4) (proof of Lemma 2, pp. 358–359): under [G] with `τ(x_n) ≤ 1`,
`E exp(t Σ_{k=1}^n b_k x_k) ≤ exp((t²/2) Σ_{k=1}^n b_k²)` for every `n`, every real
sequence `(b_k)` and every real `t`. -/
theorem mgf_weightedSum_le {Ω : Type*} {mΩ : MeasurableSpace Ω} {μ : Measure Ω}
[IsProbabilityMeasure μ] {ℱ : Filtration ℕ mΩ} {x : ℕ → Ω → ℝ}
(hx : IsCondSubgaussianOne μ ℱ x) (b : ℕ → ℝ) (n : ℕ) (t : ℝ) :
∫⁻ ω, ENNReal.ofReal (Real.exp (t * weightedSum b x n ω)) ∂μ
≤ ENNReal.ofReal (Real.exp (t ^ 2 / 2 * sqWeightSum b n)) := by sorry
end AzumaWeightedSums.IteratedLog
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.