Theorem 2.5.5 — Kolmogorov's maximal inequality
ProvedDurrettProbability.kolmogorov_maximal_inequalityLet be independent, square-integrable, with , and write . Then for every ,
Compare Chebyshev's inequality, which under the same hypotheses bounds only by the same quantity. The maximal inequality controls the entire path up to time at no extra cost, and that is what makes it useful: to show a random series converges one needs the partial sums to be uniformly close to each other over a whole tail, not merely close at one time.
The mechanism is a first-passage decomposition. Let be the event that while for every ; the are disjoint, depends only on , and depends only on , so independence kills the cross terms in and leaves .
Formalization Note The maximum over is expressed as the existence of such a with , which is the same event and avoids introducing a maximum over an index range that is empty when ; in that case both sides are the trivial bound . Square-integrability is membership in , which is what makes the variances finite.
import Mathlib import Definitions.Def_DurrettProbability_Series open MeasureTheory ProbabilityTheory Filter
namespace DurrettProbability
theorem kolmogorov_maximal_inequality {Ω : Type*} [MeasurableSpace Ω] {μ : Measure Ω}
[IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (hmeas : ∀ i, Measurable (X i))
(hindep : iIndepFun X μ) (hL2 : ∀ i, MemLp (X i) 2 μ) (hmean : ∀ i, μ[X i] = 0)
(n : ℕ) (x : ℝ) (hx : 0 < x) :
(μ {ω | ∃ k ∈ Finset.Icc 1 n, x ≤ |partialSum X k ω|}).toReal
≤ Var[partialSum X n; μ] / x ^ 2 := by sorry
end DurrettProbabilityConfirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.