§3.3, pp. 359–360 — the representative (x(π), y(π)) of a stationary policy is a feasible solution of the dual LP
ProvedHordijkKallenbergLP.ExtremePoint.representative_feasiblelinear-programmingmarkov-chainmarkov-decision-processp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let with , and let be a stationary rule: for and for every state . Then the representative defined by (6) is a feasible solution of the dual linear program:
Feasibility of the representative is the first half of the correspondence between stationary policies and dual solutions of §3.3, and it is part of the conclusion of Theorem 10.
Formalization Note. The vector in (6) uses the denominator ; with the printed the representative is in general not nonnegative.
Preamble
import Mathlib import Definitions.Def_MarkovDecisionProcesses_AverageReward import Definitions.Def_BlackwellDiscreteDP_NearOne_LimitMatrix import Definitions.Def_HordijkKallenbergLP_ExtremePoint_Model open Matrix MarkovDecisionProcesses BlackwellDiscreteDP.NearOne
Formal statement
namespace HordijkKallenbergLP.ExtremePoint
/-- **§3.3, pp. 359–360 (unnumbered): the representative is a feasible solution.** For every
stationary rule `π` (randomized, supported on the admissible actions) and every `β` with
`β_j > 0`, `∑_j β_j = 1`, the representative `(x(π), y(π))` of (6) satisfies the constraints
(3)–(5) of the dual linear program.
Hordijk and Kallenberg, Linear Programming and Markov Decision Chains, Management Sci.
25(4):352–362 (1979), DOI 10.1287/mnsc.25.4.352, pp. 359–360, §3.3
("First, we show that (x(π), y(π)) is a feasible solution.").
**Formalization Note.** `γ` in `repY` uses the denominator `∑_{k ∈ E_j} p*_ki` (see `gamma`):
with the printed `∑_k p*_ki` the claim fails when transient states feed an ergodic set. -/
theorem representative_feasible {S A : Type*} [Fintype S] [Fintype A] [DecidableEq S]
[DecidableEq A] (M : StationaryMDP S A) (β : S → ℝ) (hβpos : ∀ j, 0 < β j)
(hβsum : ∑ j, β j = 1) (π : S → A → ℝ) (hπ : IsStationaryRule M π) :
(repX M β π, repY M β π) ∈ dualFeasible M β := by sorry
end HordijkKallenbergLP.ExtremePoint
Source
Hordijk and Kallenberg, Linear Programming and Markov Decision Chains, Management Sci. 25(4):352–362 (1979), DOI 10.1287/mnsc.25.4.352, pp. 359–360, §3.3, feasibility of (x(π), y(π))
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.