The transformed discounted value solves the original average cost equation
ProvedArapostathisAC.RossReduction.rewritten_acoeaverage-cost-optimality-equationdiscounted-costp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let have the same admissible actions and bounded nonnegative costs as , and let its transition law be the transformation in Theorem 5.6 for . Write for its discounted value. Every value is finite, and the value function is bounded. Moreover, with and ,
This is the equation at the center of the reduction from average cost for to discounting for .
Formalization Note The minimum is attained, and the real expectation series is summable. The explicit finiteness conclusion prevents conversion of an infinite discounted value to Lean's default real value zero. The cost bound comes from §5.1.
Preamble
import Mathlib import Definitions.Def_ArapostathisAC_RossReduction_CMP
Formal statement
namespace ArapostathisAC.RossReduction
/-- The discounted optimality equation for the transformed process,
rewritten as the original process's average cost optimality equation. -/
theorem rewritten_acoe {A : Type*} [MetricSpace A] [MeasurableSpace A] [BorelSpace A]
(M M' : CMP A) (α : ℝ) (hα : 0 < α) (hα1 : α < 1)
(hP : ∀ i, ∀ a ∈ M.U i, α ≤ prob M i a 0)
(hmod : ModifiedLaw M M' α)
(C : ℝ) (hc : ∀ i, ∀ a ∈ M.U i, M.c i a ≤ C) :
(∀ i, discValue M' (1 - α) i ≠ ⊤) ∧
(∃ B : ℝ, ∀ i, |(discValue M' (1 - α) i).toReal| ≤ B) ∧
ACOE M (α * (discValue M' (1 - α) 0).toReal)
(fun i => (discValue M' (1 - α) i).toReal) := by sorry
end ArapostathisAC.RossReduction
Source
Arapostathis et al., Discrete-time controlled Markov processes with average cost criterion: a survey, SIAM J. Control Optim. 31(2) (1993), p. 304, displayed ACOE in proof of Theorem 5.6; https://doi.org/10.1137/0331018
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.