Theorem 2 — a.s. for class [G]
ProvedAzumaWeightedSums.IteratedLog.theorem2_limsup_le_oneLet be a sequence of random variables for which [G] holds with for all : martingale differences with respect to a filtration with a.s. for every real . Let be a real sequence, and put and . If
then
This is the upper half of the law of the iterated logarithm for weighted sums of conditionally sub-Gaussian martingale differences, including bounded martingale differences with . The constant matches the classical Hartman–Wintner law for independent standard Gaussian increments, so it cannot be improved in general.
Formalization Note The lim sup is stated in its equivalent eventual form, avoiding Lean's real-valued limsup (which is on unbounded sequences): for every , almost surely, for all sufficiently large . Since , for all large , so only meaningful values of the iterated logarithm enter. Both parts of (4.1) are hypotheses. Indices start at ; the filtration's initial -field is arbitrary.
import Mathlib import Definitions.Def_AzumaWeightedSums_IteratedLog_ClassG import Definitions.Def_AzumaWeightedSums_IteratedLog_WeightedSums open MeasureTheory ProbabilityTheory Filter Topology
namespace AzumaWeightedSums.IteratedLog
/-- Azuma 1967, Theorem 2, display (4.2), p. 362: if `(x_n)` satisfies [G] with
`τ(x_n) ≤ 1` for all `n`, and the real weights `(a_n)` satisfy (4.1),
`a_n²/D_n² → 0` and `D_n² → ∞`, then
`limsup_{n→∞} |S_n| / √(2 D_n² log log D_n²) ≤ 1` almost surely.
The lim sup is stated in its equivalent eventual form: for every `ε > 0`, almost surely,
`|S_n| ≤ (1 + ε) √(2 D_n² log log D_n²)` for all sufficiently large `n`. -/
theorem theorem2_limsup_le_one {Ω : Type*} {mΩ : MeasurableSpace Ω} {μ : Measure Ω}
[IsProbabilityMeasure μ] {ℱ : Filtration ℕ mΩ} {x : ℕ → Ω → ℝ}
(hx : IsCondSubgaussianOne μ ℱ x) (a : ℕ → ℝ)
(ha_ratio : Tendsto (fun n => a n ^ 2 / sqWeightSum a n) atTop (𝓝 0))
(ha_div : Tendsto (fun n => sqWeightSum a n) atTop atTop) :
∀ ε : ℝ, 0 < ε → ∀ᵐ ω ∂μ, ∀ᶠ n in atTop,
|weightedSum a x n ω|
≤ (1 + ε) * Real.sqrt (2 * sqWeightSum a n
* Real.log (Real.log (sqWeightSum a n))) := by sorry
end AzumaWeightedSums.IteratedLog
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.