(3), p. 1257 — f* satisfies f(x) = min_{z≥0}{q(x,z) + γE[f(x_+)]}, the minimum is attained, and f* is the least solution
ProvedLostSalesOldNew.LeadTime.optimality_equationConsider the lost-sales system of Zipkin (2008), §2, with lead time , cost data , discount factor , and i.i.d. demand with law on with finite mean. Let be the infinite-horizon optimal cost, the limit of the finite-horizon optimal costs of recursion (2) as the horizon grows, let be the transformed one-period cost and the next state. Then:
- for every state ,
- if or , the infimum is a minimum: for every there is an order with , the next state taken with ;
- is the least solution: every function with values in that satisfies the equation of item 1 at every satisfies for every .
This is the functional equation (3) of the paper: the optimal cost of the infinite-horizon problem satisfies the Bellman equation with a minimizing order, and it is singled out among the solutions. The lead-time bound (6) is obtained by substituting into it.
Formalization Note. The paper says the optimal cost "uniquely satisfies" (3), calls "the solution" and "an optimal policy", without naming a function class for uniqueness. Uniqueness in an unspecified class is not formalized; item 3 states the class-free part of it, that is the least nonnegative solution. Attainment (item 2) is stated under or : when and demand has unbounded support, but every order leaves a positive expected penalty, so the paper's is not attained in that degenerate case. The optimal cost is defined as (value iteration from the terminal condition of (2)), with values in . The equation is stated at nonnegative states, the states the system visits.
import Mathlib import Definitions.Def_LostSalesOldNew_LeadTime_Model
namespace LostSalesOldNew.LeadTime
open MeasureTheory
theorem optimality_equation (K : Costs) (D : Measure ℝ) [IsProbabilityMeasure D]
(hD : LostSalesOldNew.StateReduction.IsDemand D) (hmean : Integrable (fun d : ℝ => d) D)
(L : ℕ) (hL : 1 ≤ L) :
(∀ x : Fin L → ℝ, (∀ i, 0 ≤ x i) → fStar K D L x = bellman K D L (fStar K D L) x) ∧
((0 < K.c ∨ 0 < K.hhat) → ∀ x : Fin L → ℝ, (∀ i, 0 ≤ x i) → ∃ z : ℝ, 0 ≤ z ∧
fStar K D L x = ENNReal.ofReal (qL K D L (Fin.snoc (α := fun _ => ℝ) x z))
+ ENNReal.ofReal K.γ * ∫⁻ d, fStar K D L (LostSalesOldNew.StateReduction.next x z d) ∂D) ∧
(∀ g : (Fin L → ℝ) → ENNReal,
(∀ x : Fin L → ℝ, (∀ i, 0 ≤ x i) → g x = bellman K D L g x) →
∀ x : Fin L → ℝ, (∀ i, 0 ≤ x i) → fStar K D L x ≤ g x) := by sorry
end LostSalesOldNew.LeadTime
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.