Finite-horizon upper bound from a random start, for measurable locally bounded costs
ProvedOWMR.ss_poisson_upper_measapproximation-algorithmsinventoryinventory-theoryoperations-researchrenewal-theorystochastic-control
Let be a measurable space with a point , let be i.i.d. with law on , and let be measurable, so that the demands have law with . Let be measurable and bounded above on every bounded interval, and let , and be reals. Let (an event) and (a real number) be measurable functions of the path that depend only on . Suppose that once active the process stays active and follows the rule , with if and otherwise, and that it starts at most . Then there is a finite such that, for all , with set to ,
where (ssAvgCost) is the long-run average cost of the policy started at .
This extends the bound for continuous to the costs used in Appendix C: the stockout probability and the width loss .
Preamble
import Definitions.Def_owmr_single_location open MeasureTheory open scoped ENNReal NNReal
Formal statement
namespace OWMR
theorem ss_poisson_upper_meas {α : Type} [MeasurableSpace α] [Zero α]
(μ : Measure α) [IsProbabilityMeasure μ] (g : α → ℝ≥0) (hg : Measurable g)
(ρ : Measure ℝ≥0) [IsProbabilityMeasure ρ] (hρg : μ.map g = ρ)
(hρ : 0 < ∫⁻ x, (x : ℝ≥0∞) ∂ρ) (f : ℝ → ℝ) (hfm : Measurable f)
(hfb : ∀ a b : ℝ, ∃ C, ∀ y ∈ Set.Icc a b, f y ≤ C) (hf0 : ∀ y, 0 ≤ f y)
(K s S M : ℝ) (hK : 0 ≤ K) (hsS : s ≤ S)
(act : ℕ → (ℕ → α) → Prop) [∀ t, DecidablePred (act t)] (X : ℕ → (ℕ → α) → ℝ)
(hactm : ∀ t, MeasurableSet {d | act t d}) (hXm : ∀ t, Measurable (X t))
(hactd : ∀ (t : ℕ) (d d' : ℕ → α), (∀ r < t, d r = d' r) → (act t d ↔ act t d'))
(hXd : ∀ (t : ℕ) (d d' : ℕ → α), (∀ r < t, d r = d' r) → X t d = X t d')
(hmono : ∀ t d, act t d → act (t + 1) d)
(hstep : ∀ t d, act t d → X (t + 1) d = (if X t d ≤ s then S else X t d) - (g (d t) : ℝ))
(hstart0 : ∀ d, act 0 d → X 0 d ≤ M)
(hstart : ∀ t d, ¬ act t d → act (t + 1) d → X (t + 1) d ≤ M) :
∃ C : ℝ≥0∞, C ≠ ∞ ∧ ∀ T N : ℕ, T ≤ N →
∫⁻ z, ∑ t ∈ Finset.range T, (if act t (padHist N z) then ENNReal.ofReal
(K * (if X t (padHist N z) ≤ s ∧ X t (padHist N z) < S then 1 else 0)
+ f (if X t (padHist N z) ≤ s then S else X t (padHist N z))) else 0)
∂(Measure.pi fun _ : Fin N => μ)
≤ T * ssAvgCost ρ f K s S + C := by sorry
end OWMRSource
M. Hu, "A 5/2 Guarantee for Echelon Stock (s,S) Policies in One-Warehouse Multi-Retailer Systems", working paper (manuscript of November 2026), Appendix B (proof of Proposition 2), Step 4, p. E.4, and its use in Appendix C, Steps 4 and 8, pp. E.14-E.16 and E.21-E.22; renewal-reward identities (10) and (11), pp. 13-14.