Finite-horizon Gittins retirement values are monotone in the horizon
ProvedBanditAlgorithm.gittinsFiniteRetirementValue_mono_of_integrablebanditsdynamic-programminggittins-indexoptimal-stopping
Consider the finite-horizon value functions for a discounted one-armed retirement game. At horizon zero the value is zero, while at horizon the player chooses between retiring for zero and playing once for reward followed by the discounted continuation value at horizon .
Assume and that every finite-horizon continuation value is integrable against every transition measure . Then extending the available horizon cannot decrease the value:
This is the monotonicity input used to pass from finite-horizon dynamic programming to the infinite-horizon Wald--Bellman equation for the discounted retirement problem.
Preamble
import Definitions.Def_GittinsFiniteRetirementValue open MeasureTheory ProbabilityTheory
Formal statement
theorem BanditAlgorithm.gittinsFiniteRetirementValue_mono_of_integrable
{S : Type*} [MeasurableSpace S]
(P : Kernel S S) (r : S → ℝ) (α γ : ℝ) (hα0 : 0 ≤ α)
(hint : ∀ n x, Integrable (gittinsFiniteRetirementValue P r α γ n) (P x)) :
∀ n x,
gittinsFiniteRetirementValue P r α γ n x ≤
gittinsFiniteRetirementValue P r α γ (n + 1) x := by
sorrySource
Lattimore and Szepesvari, Bandit Algorithms (Cambridge University Press, 2020), printed p. 442, Theorem 35.3 (finite-horizon approximation underlying the infinite-horizon Wald--Bellman equation), and printed p. 449, Lemma 35.7(a).