Measurability of the infinite-horizon retirement value from finite Bellman convergence
ProvedBanditAlgorithm.measurable_gittinsRetirementValue_of_finite_tendstobanditsgittins-indexmeasurabilityoptimal-stopping
Let be the measurable finite-horizon Bellman values of a discounted one-armed retirement game, and suppose that they converge pointwise to the infinite-horizon retirement value :
Then the map is measurable.
This is the measurability step needed to define threshold stopping rules from sublevel sets of the retirement value and, through the fair-charge characterization, of the Gittins index.
Preamble
import Definitions.Def_GittinsFiniteRetirementValue import Mathlib.MeasureTheory.Constructions.BorelSpace.Metrizable open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
theorem BanditAlgorithm.measurable_gittinsRetirementValue_of_finite_tendsto
{S : Type*} [MeasurableSpace S]
(P : Kernel S S) [IsMarkovKernel P]
{r : S → ℝ} (hr : Measurable r)
(α γ : ℝ)
(hconv : ∀ x,
Tendsto (fun n ↦ gittinsFiniteRetirementValue P r α γ n x)
atTop (𝓝 (gittinsRetirementValue P r α γ x))) :
Measurable (gittinsRetirementValue P r α γ) := by
sorrySource
Lattimore and Szepesvari, Bandit Algorithms (Cambridge University Press, 2020), printed pp. 442-443, Theorem 35.3 and the following note that measurability is the technical issue; specialized to printed p. 449, Lemma 35.7.