Equation (32), p. 54 — derivative of the chain resolvent
ProvedPolicyGradTheory.ChainLB.resolvent_derivative_identitychain-mdpderivativematrix-resolventp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Fix a chain with , forward probabilities , and . For states and an interior state , the derivative in its forward probability is
This is the entrywise derivative identity used to bound all higher order partial derivatives in Appendix B.2.
Formalization Note The displayed identity is the sign-rearranged form of the paper's equation (32), which writes its negative on both sides. The input is restricted to the open probability cube, where the resolvent is nonsingular.
Preamble
import Mathlib import Definitions.Def_PolicyGradTheory_ChainLB_Chain
Formal statement
namespace PolicyGradTheory.ChainLB
/-- Appendix B.2, (32), p. 54: derivative of one resolvent entry in an interior forward probability. -/
theorem resolvent_derivative_identity (H : ℕ) (hH : 1 ≤ H)
(p : Fin H → ℝ) (hp : ∀ i, 0 < p i ∧ p i < 1)
(a b : Fin (H + 2)) (i : Fin H) :
fderiv ℝ (fun q : Fin H → ℝ => chainResolvent H q a b) p (Pi.single i (1 : ℝ)) =
chainGamma H * chainResolvent H p a ⟨i.val + 1, by omega⟩ *
(chainResolvent H p ⟨i.val + 2, by omega⟩ b -
chainResolvent H p ⟨i.val, by omega⟩ b) := by sorry
end PolicyGradTheory.ChainLB
Source
arXiv:1908.00261v5, App. B.2, (32), p. 54
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.