Theorem 4.2, p. 541 — |Eh(W) − 𝒫_λh| ≤ 2λ^{−1}[8m+5+4(nφ(m+1))^{1/2}][Var W − λ + 2(2m+1)Σp_i²] + 32[6m+3+(nφ(m+1))^{1/2}]nφ(m+1)
ProvedPoissonDepTrials.MixInv.theorem_4_2Let be Bernoulli random variables on a probability space , with , padded by for and . Write and . Assume Ibragimov's mixing condition (4.1): there is a non-increasing sequence such that for all and every event , almost surely, where . Then for every and every real function on the nonnegative integers with ,
where is the expectation of under the Poisson law with mean .
Since is twice the total variation distance between the law of and the Poisson law, this is an explicit Poisson approximation for sums of dependent trials. Its first term has order , which improves on Theorem 4.1's when is large.
Formalization Note The trials are with IsBernoulliTrials (measurable, zero outside , at most almost surely); is the definition prob, so it cannot disagree with . Constant padding variables generate the trivial -algebra, so (4.1) for the padded sequence is (4.1) for . Double sums over are computed in . is min (1 / Real.sqrt λ) 1 and is Real.sqrt (n * φ (m + 1)). No hypothesis is added: at every , almost surely, and both sides are with Lean's , as in the paper. The factor is the bare of the paper, not .
import Mathlib import Definitions.Def_PoissonDepTrials_MixInv_Setting open MeasureTheory ProbabilityTheory
namespace PoissonDepTrials.MixInv
theorem theorem_4_2 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
(n : ℕ) (X : ℕ → Ω → ℕ) (hX : IsBernoulliTrials P n X) (φ : ℕ → ℝ) (hφ : IbragimovMixing P X φ) (m : ℕ)
(h : ℕ → ℝ) (hh : ∀ k, |h k| ≤ 1) :
|∫ ω, h (W n X ω) ∂P - poissonExp (lam P n X) h| ≤
2 * (lam P n X)⁻¹ * (8 * m + 5 + 4 * Real.sqrt ((n : ℝ) * φ (m + 1)))
* (Var[fun ω => (W n X ω : ℝ); P] - lam P n X
+ 2 * (2 * m + 1) * ∑ i ∈ Finset.Icc 1 n, prob P X i ^ 2)
+ 32 * (6 * m + 3 + Real.sqrt ((n : ℝ) * φ (m + 1))) * n * φ (m + 1) := by sorry
end PoissonDepTrials.MixInv
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.