Lemma B.2, p. 21 — for a reversible policy chain, ρ(B_π) = (1 − √(1/κ))/(1 + √(1/κ)) (printed (1 − √κ)/(1 + √κ))
ProvedAccelVI.MVC.lemma_B_2momentump2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1reversible-markov-chainspectral-radius
Let be a valid finite MDP and a policy whose transition matrix defines an irreducible, reversible Markov chain, and let be the matrix of App. B built with the constant step sizes (4.2). With , the spectral radius of is
Together with Lemma B.1 this gives the linear convergence rate of Momentum Value Computation in Theorem 5.5.
Formalization Note. The lemma prints , which is negative since and so cannot be a spectral radius; the proof's final display computes , the rate of Theorem 5.5, which is the value stated. The spectral radius is that of the complexified matrix over , an extended nonnegative real compared with the (positive) rate.
Preamble
import Mathlib import Definitions.Def_mm_basic import Definitions.Def_AccelVI_MVC_Setting
Formal statement
namespace AccelVI.MVC
variable {S A : Type*} [Fintype S] [DecidableEq S] [Fintype A]
/-- Lemma B.2, p. 21 (with the rate (1 − √(1/κ))/(1 + √(1/κ)) of its proof for the printed
(1 − √κ)/(1 + √κ)): if π defines an irreducible, reversible Markov chain, then for step sizes (4.2)
ρ(B_π) = (1 − √(1/κ))/(1 + √(1/κ)). -/
theorem lemma_B_2 (M : MDP S A) (hM : M.Valid) (pol : S → A → ℝ) (hpol : IsPolicy pol)
(hrev : AccelVI.AVC.IsReversible (LPol M pol)) :
spectralRadius ℂ ((Bpol M pol).map (algebraMap ℝ ℂ)) = ENNReal.ofReal (rateMVC M.lam) := by sorry
end AccelVI.MVC
Source
Goyal & Grand-Clément, A First-Order Approach to Accelerated Value Iteration, arXiv:1905.09963v7, p. 21, Lemma B.2 (constant corrected per its proof, p. 22)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.