(B.1), p. 21 — every eigenvalue ω of B_π solves ω² − 2λµω/(1 + √(1 − λ²)) + (1 − √(1 − λ²))/(1 + √(1 − λ²)) = 0 for an eigenvalue µ of L_π
ProvedAccelVI.MVC.eq_B_1block-matrixmomentump2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1spectrum
Let be a valid finite MDP and a policy, and let be the matrix of App. B built with the step sizes (4.2). For every (complex) eigenvalue of there is an eigenvalue of such that
So the spectrum of is determined by that of , each contributing at most two eigenvalues of .
Formalization Note. The paper states that to every eigenvalue of correspond at most two eigenvalues of satisfying (B.1); the formal statement records the direction the proof of Lemma B.2 uses: every eigenvalue of arises in this way. Spectra are of the complexified matrices.
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]
/-- (B.1), p. 21: every eigenvalue ω of B_π (step sizes (4.2)) is a root of
ω² − 2λμ/(1 + √(1 − λ²)) ω + (1 − √(1 − λ²))/(1 + √(1 − λ²)) = 0 for some eigenvalue μ of L_π. -/
theorem eq_B_1 (M : MDP S A) (hM : M.Valid) (pol : S → A → ℝ) (hpol : IsPolicy pol) :
∀ ω ∈ spectrum ℂ ((Bpol M pol).map (algebraMap ℝ ℂ)),
∃ μ ∈ spectrum ℂ ((LPol M pol).map (algebraMap ℝ ℂ)),
ω ^ 2 - (2 * (M.lam : ℂ) * μ / (1 + (Real.sqrt (1 - M.lam ^ 2) : ℂ))) * ω
+ ((1 - Real.sqrt (1 - M.lam ^ 2)) / (1 + Real.sqrt (1 - M.lam ^ 2)) : ℝ) = 0 := by sorry
end AccelVI.MVC
Source
Goyal & Grand-Clément, A First-Order Approach to Accelerated Value Iteration, arXiv:1905.09963v7, p. 21, proof of Lemma B.2, display (B.1)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.