Pathwise Gittins terminal-potential envelope
ProvedBanditAlgorithm.gittins_current_terminal_potential_le_stack_envelopebanditsgittins-indexprobability
On a common stack realization of the arms, consider any length- interleaving. The sum of the arms' current Gittins retirement values after this history is bounded above by an explicit nonnegative envelope. For each arm, the envelope consists of the absolute discounted-reward values at its first stack states, together with a geometric factor multiplying its initial absolute value and the absolute rewards observed along the first subsequent stack states.
This pathwise bound is uniform over the action interleaving and supplies the domination needed to remove the terminal retirement term in the finite-horizon Gittins comparison.
Preamble
import Definitions.Def_GittinsTerminalPotential
Formal statement
open MeasureTheory ProbabilityTheory ENNReal
theorem BanditAlgorithm.gittins_current_terminal_potential_le_stack_envelope
{k N : ℕ} {S : Type*} [MeasurableSpace S]
(P : Kernel S S) [IsMarkovKernel P]
{r : S → ℝ} (hr : Measurable r)
{α : ℝ} (hα0 : 0 < α) (hα1 : α < 1)
(hint : DiscountedRewardIntegrable P r α)
(ω : Fin k → ℕ → S) (a : Fin N → Fin k) :
let count := fun (i : Fin k) (t : ℕ) ↦
∑ s : Fin N, if (s : ℕ) < t ∧ a s = i then 1 else 0
let stackHistory : MarkovBanditHistory k S N :=
((fun t ↦ ((fun i ↦ ω i (count i t)), a t)),
fun i ↦ ω i (count i N))
let absoluteValue := fun y : S ↦
∫ path, ∑' t : ℕ, α ^ t * |r (path t)|
∂markovChainMeasure P y
let envelope := ∑ i : Fin k,
((∑ u ∈ Finset.range (N + 1), absoluteValue (ω i u)) +
(∑' t : ℕ, α ^ t) *
(absoluteValue (ω i 0) +
∑ v ∈ Finset.range N, |r (ω i (v + 1))|))
currentGittinsRetirementPotential P r α stackHistory ≤ envelope := by
sorry
Source
Tor Lattimore and Csaba Szepesvári, Bandit Algorithms, §35.4, proof of Lemma 35.10, printed pp. 452–453 (free PDF pp. 460–461), https://tor-lattimore.com/downloads/book/book.pdf