Lemma 7 — for identically distributed with , almost surely
ProvedGenEmpLik.Expansion.lemma_7borel-cantellip2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1probability
Let be identically distributed real random variables, not necessarily independent, with for some . Then
- for every ,
- almost surely,
With this shows that a finite-variance stationary sequence has no single observation of order , which makes the event of Lemma 6 hold eventually.
Formalization Note Lean indexes the sequence from : Z m is the paper's , so the event of clause 1 is , and the maximum in clause 2 is over the first variables Z 0, …, Z (n-1). "Infinitely often" is Filter.limsup of the events along atTop. Identical distribution is Mathlib's IdentDistrib; no independence is assumed. Powers with real exponent are Real.rpow.
Preamble
import Mathlib open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
namespace GenEmpLik.Expansion
/-- Lemma 7 (Duchi, Glynn & Namkoong, arXiv:1610.03425v3, p. 31). Let `Z 0, Z 1, …` (the paper's
`Z₁, Z₂, …`) be identically distributed, possibly dependent, with `E[|Z₁|^k] < ∞` for some `k > 0`.
Then for every `ε > 0`, `P(|Z_n| ≥ ε n^{1/k} i.o.) = 0` (Lean's `Z m` is the paper's `Z_{m+1}`,
so the threshold is `ε (m+1)^{1/k}`), and `max_{i ≤ n} |Z_i| / n^{1/k} → 0` almost surely. -/
theorem lemma_7 {Ω : Type*} [MeasurableSpace Ω] {P : Measure Ω} [IsProbabilityMeasure P]
(Z : ℕ → Ω → ℝ) (hident : ∀ i, IdentDistrib (Z i) (Z 0) P P) {k : ℝ} (hk : 0 < k)
(hmom : Integrable (fun ω => |Z 0 ω| ^ k) P) :
(∀ ε : ℝ, 0 < ε →
P (limsup (fun m : ℕ => {ω | ε * ((m : ℝ) + 1) ^ (1 / k) ≤ |Z m ω|}) atTop) = 0) ∧
(∀ᵐ ω ∂P, Tendsto (fun n : ℕ => (⨆ i : Fin n, |Z i ω|) / (n : ℝ) ^ (1 / k))
atTop (𝓝 0)) := by sorry
end GenEmpLik.Expansion
Source
Duchi, Glynn & Namkoong, Statistics of Robust Optimization: A Generalized Empirical Likelihood Approach, arXiv:1610.03425v3, p. 31, Lemma 7
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.