Section 5 — (7) holds iff and ; the points of are the costs of the strategies
ProvedDEpenoux.LinearProgram.bellman_fixed_iff_A_and_BIn d'Epenoux's production–inventory model with stock levels , a stochastic kernel , real costs and discount , let be the Bellman operator
and let . Then:
- a vector solves (7), , if and only if and , that is, for every and every admissible , and for every ;
- a vector lies in if and only if it is the cost of some strategy: there is a map with for all such that
i.e. , the system (2).
This is the decomposition with which Section 5 turns the dynamic program (7) into a linear program: (A) are the linear constraints (9), and (B) selects the strategy costs.
Formalization Note The page writes "" and ""; here ranges over the admissible potentials , as in the paper's count of constraints. "The points of the set correspond to the costs associated with every possible strategy" is formalized as the equivalence in item 2, with the cost of described as a solution of (2) through the policy operator BertsekasDiscountedPolicyOp.
import Mathlib import Definitions.Def_DEpenoux_LinearProgram_Model
namespace DEpenoux.LinearProgram
theorem bellman_fixed_iff_A_and_B (σ : ℕ) (p d : Fin (σ + 1) → Fin (σ + 1) → ℝ)
(hp0 : ∀ j s, 0 ≤ p j s) (hp1 : ∀ j, ∑ s, p j s = 1)
(lam : ℝ) (hlam0 : 0 < lam) (hlam1 : lam < 1) :
(∀ u : Fin (σ + 1) → ℝ,
BertsekasDiscountedBellmanOp (model σ p d hp0 hp1) lam u = u ↔
u ∈ setA p d lam ∧ u ∈ setB p d lam) ∧
(∀ u : Fin (σ + 1) → ℝ,
u ∈ setB p d lam ↔
∃ J : Fin (σ + 1) → Fin (σ + 1), IsStrategy J ∧
BertsekasDiscountedPolicyOp (model σ p d hp0 hp1) lam J u = u) := by sorry
end DEpenoux.LinearProgram
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.