Lemma 1 — strong law of large numbers along a predictably selected subsequence
ProvedKMNewsvendor.optional_skipping_sllnLet be independent real random variables, all distributed like , and let ( trivial). Let be selection indicators such that each is -measurable, i.e. a deterministic measurable function of , and such that almost surely for infinitely many . Let be measurable with integrable. Then, almost surely,
The subsequence of selected demands is chosen adaptively, after seeing the earlier demands, and the strong law still holds along it. This is the probabilistic engine of the paper: the censoring decisions of an inventory policy are exactly such predictable selections.
Formalization Note Lean's index is the source's ; ξ t is a Boolean given by a measurable function of D 0, …, D (t-1) (for , a constant). "" is "almost surely, frequently ξ t = true". The source's hypothesis "" is formalized as integrability of , which makes a finite real number. For the finitely many with an empty selection the Lean quotient is , which does not affect the limit.
import Mathlib open MeasureTheory ProbabilityTheory Filter Finset Topology
namespace KMNewsvendor
/-- Lemma 1 (p. 934): the strong law of large numbers along a predictably selected subsequence.
`ξ t` (Lean period `t`, the paper's `ξ_{t+1}`) is a deterministic measurable function of
the demands `D i`, `i < t`; infinitely many periods are selected almost surely; `f (D)` is integrable. -/
theorem optional_skipping_slln {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω)
[IsProbabilityMeasure μ]
(D : ℕ → Ω → ℝ) (hmeas : ∀ t, Measurable (D t)) (hindep : iIndepFun D μ)
(hident : ∀ t, IdentDistrib (D t) (D 0) μ μ)
(ξ : ℕ → Ω → Bool)
(hξ : ∀ t, ∃ g : (Fin t → ℝ) → Bool, Measurable g ∧ ξ t = fun ω => g (fun i => D i ω))
(hinf : ∀ᵐ ω ∂μ, ∃ᶠ t in atTop, ξ t ω = true)
(f : ℝ → ℝ) (hf : Measurable f) (hint : Integrable (fun ω => f (D 0 ω)) μ) :
∀ᵐ ω ∂μ, Tendsto (fun T : ℕ =>
(∑ t ∈ range T, if ξ t ω = true then f (D t ω) else 0) /
(((range T).filter (fun t => ξ t ω = true)).card : ℝ))
atTop (𝓝 (∫ ω, f (D 0 ω) ∂μ)) := by sorry
end KMNewsvendor
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Setting.
- is a probability space.
- are measurable, mutually independent, and each has the same distribution as under .
- are selection indicators.
Hypotheses.
- For every there is a measurable with for all . So is determined by the strictly earlier values. In particular, is a constant, since is a single point.
- For -almost every , for infinitely many .
- is measurable, and is -integrable.
Conclusion. For -almost every ,
Degenerate cases. When no is selected (for example ), the quotient has denominator and takes the value . By hypothesis 2, this happens for only finitely many almost surely, so it does not affect the limit.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.