Appendix B.2, p. 51 — chain value is a resolvent entry
ProvedPolicyGradTheory.ChainLB.chain_value_eq_resolventchain-mdpmatrix-resolventp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1policy-value
Fix the Figure 2 chain with and a direct parameter whose three coordinates at each interior state lie strictly between zero and one, with every forward coordinate less than . Let and . Then
This identifies the discounted return with the matrix entry differentiated in Appendix B.2.
Formalization Note The parameter assumptions are those of Proposition 4.1. The value depends only on the forward coordinates; the other coordinates need not sum to at most one for the algebraic expression to be defined.
Preamble
import Mathlib import Definitions.Def_PolicyGradTheory_ChainLB_Chain
Formal statement
namespace PolicyGradTheory.ChainLB
/-- Appendix B.2, p. 51: the return from s₀ is the last-column resolvent entry. -/
theorem chain_value_eq_resolvent (H : ℕ) (hH : 1 ≤ H)
(θ : EuclideanSpace ℝ (Fin H × Fin 3))
(hθ : ∀ i j, 0 < θ (i, j) ∧ θ (i, j) < 1)
(hθ1 : ∀ i, θ (i, (0 : Fin 3)) < 1 / 4) :
chainValue H θ = chainResolvent H (forwardProb H θ) 0 (Fin.last (H + 1)) := by sorry
end PolicyGradTheory.ChainLB
Source
arXiv:1908.00261v5, App. B.2, p. 51, paragraph before Lemma B.1
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.