Proposition 4.2 [Grincevičius (1980)] — Lévy-type maximal inequality for the partial sums of a perpetuity
ProvedGoldieRenewal.Kesten.proposition_4_2Let , , be independent pairs of real random variables, each with the law of , and let , , , be the partial products and sums
so that . Then for every and all ,
where is a median of the law of , and the inequality holds for every choice of medians.
This extends Lévy's symmetrization inequality from sums of independent variables to the affine recursion; in the proof of Theorem 4.1 it bounds the probability that some partial sum is large by the tail of the limit.
Formalization Note The sequence is stored 0-based in Lean (the -th Lean variable is the paper's ). "The maximum exceeds " is written as "some has the term exceeding ". No condition on beyond measurability is assumed, as in the paper, which notes that the assumption of Grincevičius (1981) is not needed.
import Mathlib import Definitions.Def_GoldieRenewal_Kesten_Perpetuity
namespace GoldieRenewal.Kesten
open MeasureTheory ProbabilityTheory
open scoped ENNReal
/-- **Proposition 4.2** [Grincevičius (1980)] (Goldie, *Implicit renewal theory and tails of
solutions of random equations*, Ann. Appl. Probab. 1(1) (1991), p. 136). Let `(Q_n, M_n)`,
`n = 1, 2, …`, be independent, each with the law `μ` of `(Q, M)`. With
`Π_j = M₁ ⋯ M_j`, `R_n = Σ_{k=1}^n Π_{k−1} Q_k`, `Π_{j,n} = Π_{k=j+1}^n M_k`,
`R_{j,n} = Σ_{k=j+1}^n Π_{j,k−1} Q_k`, for all `x, y ∈ ℝ`,
`P(max_{j=1,…,n} (R_j + Π_j med(R_{j,n} + Π_{j,n} y)) > x) ≤ 2 P(R_n + Π_n y > x)`.
**Formalization Note** The sequence is stored 0-based: `Q i`, `M i` (`i = 0, 1, …`) are the
paper's `Q_{i+1}`, `M_{i+1}` (see `piProd`, `partialSum`, `piProdFrom`, `partialSumFrom`).
`med X` is any median of the law of `X`; the paper fixes none, so the inequality is asserted for
every choice `med j` of medians, `j = 1, …, n`. "max > x" is "some `j ∈ {1, …, n}` has `… > x`".
No moment or Cramér condition is assumed, as in the paper. -/
theorem proposition_4_2 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
(μ : Measure (ℝ × ℝ)) (Q M : ℕ → Ω → ℝ) (hQ : ∀ i, Measurable (Q i))
(hM : ∀ i, Measurable (M i))
(hindep : iIndepFun (fun i ω => (Q i ω, M i ω)) P)
(hlaw : ∀ i, P.map (fun ω => (Q i ω, M i ω)) = μ)
(n : ℕ) (x y : ℝ) (med : ℕ → ℝ)
(hmed : ∀ j ∈ Finset.Icc 1 n,
IsMedian (P.map (fun ω => partialSumFrom Q M j n ω + piProdFrom M j n ω * y)) (med j)) :
P {ω | ∃ j ∈ Finset.Icc 1 n, x < partialSum Q M j ω + piProd M j ω * med j}
≤ 2 * P {ω | x < partialSum Q M n ω + piProd M n ω * y} := by sorry
end GoldieRenewal.Kesten
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.