Proof of Theorem 4, p. 939 — in program (3), the optimal a = min{d, v₀ − v₁}
ProvedZipkinLostSales.LNatural.optimal_fillConsider the lost-sales model with lead time , nonnegative independent demands with common law of finite mean, nonnegative unit costs and discount factor . Let be the optimal cost with periods to go in the transformed state (with ).
After demand is observed in a period whose continuation cost is , the paper's program (3) chooses the amount of demand to fill and the remaining inventory :
The claim is that the optimal is : for every , and ,
where .
The paper notes this "is not hard, but also not essential" for Theorem 4; it is what makes the two-step program (4)–(5) agree with the recursion (1). It depends on the model: it uses , and the specific , and fails for an arbitrary continuation function.
Formalization Note. The program is stated in the form (4), after the substitution (so is ); is the infimum over the interval , and the equality says the infimum equals the value at the left endpoint. Time is the number of periods to go (fbar … k is the paper's , for every period of every horizon). Vectors are indexed with . The cost hypotheses (, ) read the paper's "unit cost" and "discount rate"; nonnegative demand is the paper's; finite mean demand is added so that is finite.
import Mathlib import Definitions.Def_ZipkinLostSales_LNatural_Model open MeasureTheory
namespace ZipkinLostSales.LNatural
/-- Proof of Theorem 4 (Zipkin 2008, p. 939): in program (3)/(4) with continuation `f̄_{t+1}`
(here `fbar … k`), filling as much demand as possible (`a = min{d, v₀ − v₁}`, i.e.
`v₊ = max(v₀ − d, v₁)`) is optimal, so `κ̄(v, ζ | d) = q̂(v₀ − v₁ − d) + γ f̄_{t+1}(v₊)` with `v₊`
the transition of p. 938. -/
theorem optimal_fill (L : ℕ) (hL : 0 < L) (μ : Measure ℝ) [IsProbabilityMeasure μ]
(hμ : μ (Set.Iio 0) = 0) (hμint : Integrable id μ)
(c hh p γ : ℝ) (hc : 0 ≤ c) (hhh : 0 ≤ hh) (hp : 0 ≤ p) (hγ0 : 0 < γ) (hγ1 : γ ≤ 1)
(k : ℕ) (v : Fin L → ℝ) (hv : v ∈ V L) (ζ : ℝ) (hζ : ζ ≤ 0) (d : ℝ) (hd : 0 ≤ d) :
kappa hh p γ (fbar c hh p γ μ k) d v ζ =
qhat hh p (vext v 0 - vext v 1 - d) + γ * fbar c hh p γ μ k (next v d ζ) := by sorry
end ZipkinLostSales.LNatural
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.