Measurability of finite-horizon Gittins Bellman values
ProvedBanditAlgorithm.measurable_gittinsFiniteRetirementValuebanditsmeasurabilityoptimal-stopping
Every finite-horizon Bellman value function for the discounted retirement game is measurable. The proof inducts on the horizon, using measurability of integration against a probability kernel.
Preamble
import Definitions.Def_GittinsFiniteRetirementValue open MeasureTheory ProbabilityTheory
Formal statement
theorem BanditAlgorithm.measurable_gittinsFiniteRetirementValue
{S : Type*} [MeasurableSpace S]
(P : Kernel S S) {r : S → ℝ} (hr : Measurable r)
(α γ : ℝ) :
∀ n, Measurable (gittinsFiniteRetirementValue P r α γ n) := by
sorrySource
Lattimore and Szepesvári, Bandit Algorithms (CUP 2020), Theorem 35.3 (Wald--Bellman equation), printed pp.442--443, specialized to the discounted retirement game used in Lemma 35.7 on printed p.449.