Kendall's renewal theorem: geometric convergence of forces
ProvedMarkovChainCLT.kendall_summable_geometric_of_renewal_tendsto_geometricKendall's renewal theorem, the geometric-convergence direction. Let be a probability distribution on the positive integers (, , ), and let be its renewal sequence: and
so that is the probability that a renewal occurs at time when inter-renewal times are i.i.d. with law . Suppose converges geometrically to a positive limit: there are , and with
Then the inter-renewal law has a geometric tail: there is with
This is the implication (i) (ii) of Kendall's Theorem as stated in Meyn and Tweedie (Theorem 15.1.1), and it is the analytic engine behind the equivalence "geometric ergodicity geometric drift" (their Theorem 15.0.1) that the source invokes in Remark 1. In the present mission it is what the remaining open leaf geoDriftCondition_of_geometricallyErgodic needs: applied to the renewal sequence of a small set with minorization on (and concentrated where the geometric rate constant is bounded), it turns the geometric total-variation convergence at the points of into geometric tails for the regeneration times, which is condition (ii) of Theorem 15.0.1 and yields the drift function .
Proof idea. Write and for ; the renewal equation is . The hypothesis says is analytic on , so extends analytically to and satisfies on the open unit disc. has no zero on the closed unit disc: , and at a point of the unit circle a zero of would force as , contradicting . By compactness is zero-free on a disc with , so is analytic there and coincides with on the unit disc; uniqueness of power series coefficients and Cauchy's estimates give for every . No aperiodicity hypothesis is needed: it is implied by the convergence of to a positive limit.
Formalization Note The renewal recursion is written with the index shift to keep natural-number subtraction honest. The distribution is normalised by HasSum a 1; the limit must be positive (for a defective or periodic law the renewal sequence does not converge geometrically to a positive limit, so nothing is lost). The conclusion is stated as summability of for some , the form in which it is consumed (a finite geometric moment of the inter-renewal time).
import Mathlib.Analysis.SpecificLimits.Basic import Mathlib.Topology.Algebra.InfiniteSum.Real open Filter open scoped Topology BigOperators
theorem MarkovChainCLT.kendall_summable_geometric_of_renewal_tendsto_geometric
(a : ℕ → ℝ) (ha0 : ∀ k, 0 ≤ a k) (ha_zero : a 0 = 0) (hsum : HasSum a 1)
(v : ℕ → ℝ) (hv0 : v 0 = 1)
(hv : ∀ n : ℕ, v (n + 1) = ∑ j ∈ Finset.range (n + 1), a (j + 1) * v (n - j))
(vlim M ρ : ℝ) (hvlim : 0 < vlim) (hρ0 : 0 ≤ ρ) (hρ1 : ρ < 1)
(hconv : ∀ n : ℕ, |v n - vlim| ≤ M * ρ ^ n) :
∃ κ : ℝ, 1 < κ ∧ Summable (fun k : ℕ => a k * κ ^ k) := by sorry