Lemma 4.4.1 — the Palm expansion of a functional of the count
ProvedPalmQueueing.Ordering.palm_count_expansionLemma 4.4.1. Let be a stationary point process with finite intensity and Palm probability , with points (). For all functions with , and for all ,
whenever the series converges absolutely.
A stationary expectation of a functional of the count in a window is expanded into a series of second differences of against Palm expectations of the residual window . For convex every coefficient is non-negative, so the expansion is monotone in the quantities ; this is what turns a comparison of the Palm distributions into a comparison of the stationary counts (Property 4.4.1).
Formalization Note. The book states (4.4.1) "for all functions " without a convergence qualification. Read literally this fails: the partial sums differ from by the boundary term , with , which need not vanish for heavy-tailed counts and sparse even when . The statement therefore assumes , which covers the book's use (convex with finite expectation, all terms non-negative).
import Mathlib import Definitions.Def_PalmQueueing_Palm_PointProcess
/-!
# Lemma 4.4.1: the Palm expansion of a functional of the count (§4.4, p.295)
-/
namespace PalmQueueing.Ordering
open MeasureTheory
open PalmQueueing.Palm
variable {Ω : Type*} [MeasurableSpace Ω]
/-- **Lemma 4.4.1** (p.295). Let `N` be a stationary point process with finite intensity. For all
functions `f : ℝ₊ → ℝ` with `f(0) = 0`, and for all `x > 0`,
`(4.4.1) E_P[f(N[0,x))] = Σ_{n=0}^{∞} ( f(n+1) − 2f(n) + f((n−1)⁺) ) λ E_{P⁰}(x − T_n)⁺`.
A stationary expectation of a functional of the count in a window is expanded into a series of
**second differences** of `f` against Palm expectations of the residual window `(x − T_n)⁺`. The
second difference is what makes the lemma useful: `f` convex means every coefficient
`f(n+1) − 2f(n) + f((n−1)⁺)` is non-negative, so the whole expansion is monotone in the quantities
`E_{P⁰}(x − T_n)⁺` — which is exactly what Property 4.4.1 needs to turn a `≤_cx` comparison of the
`T_n[P⁰]` into a `≤_cx` comparison of the counts `N[0,x)[P]`.
That equivalence, Property 4.4.1, is the point of §4.4:
> (i) `T_n[P⁰] ≤_cx T̃_n[P̃⁰]` for all `n ≥ 1`; (ii) `N[0,x)[P] ≤_cx Ñ[0,x)[P̃]` for all
> `x ∈ ℝ₊`
are equivalent. The section's Example 4.4.1, "Feller's paradox revisited", shows that the
corresponding statement for `≤_i` is **false** — the Palm order does not pass to the stationary
one — so the restriction to `≤_cx` is essential and this lemma is where it enters.
`f(0) = 0` is a hypothesis, not a normalisation: the coefficient at `n = 0` is
`f(1) − 2f(0) + f(0) = f(1) − f(0)`, and the series is arranged around it.
**Correction of the page (recorded in `HARD.md`).** The page states (4.4.1) "for all functions
`f`" without saying in what sense the series converges, and read literally it is false: the
second differences telescope, `Σ_{n ≤ M} Δ²f(n) a_n = Σ_{k ≤ M} f(k) P(N[0,x) = k)
+ f(M+1) a_M − f(M) a_{M+1}` with `a_n = λE_{P⁰}(x − T_n)⁺ = E_P(N[0,x) − n)⁺`, and for a mixed
Poisson `N` with heavy-tailed intensity and `f` supported on a sparse set the boundary term does
not vanish although `E_P|f(N[0,x))| < ∞`. The theorem is therefore stated under the hypothesis
that the series of (4.4.1) converges absolutely, `Σ_n |Δ²f(n)| λE_{P⁰}(x − T_n)⁺ < ∞`, which
covers the use the book makes of it (convex `f`, where every term is non-negative) and makes the
right-hand side an honest sum rather than Lean's `tsum` of a non-summable family (`= 0`). -/
theorem palm_count_expansion (S : PalmSetting Ω) (f : ℝ → ℝ) (hf0 : f 0 = 0) (x : ℝ) (hx : 0 < x)
(hsum : Summable fun n : ℕ =>
|f ((n : ℝ) + 1) - 2 * f (n : ℝ) + f (max ((n : ℝ) - 1) 0)| *
(S.lam * ∫ ω, max (x - S.N.T (n : ℤ) ω) 0 ∂S.P0)) :
∫ ω, f ((S.N.count ω (Set.Ico (0 : ℝ) x)).toReal) ∂S.P
= ∑' n : ℕ,
(f ((n : ℝ) + 1) - 2 * f (n : ℝ) + f (max ((n : ℝ) - 1) 0)) *
(S.lam * ∫ ω, max (x - S.N.T (n : ℤ) ω) 0 ∂S.P0) := by sorry
end PalmQueueing.Ordering
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.