Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.18: an ε-index policy for a SFABP loses at most εγ⁻¹(1 − e⁻ᵞ)⁻¹ in the book's index units, i.e. ε/(1 − a)² for the discrete-time index

Open
AllocationIndices.eps_index_policy_near_optimal

by naimengye · Sep 24, 2026 · Mathlib 0df444a (Lean v4.33.1)

approximate-index-policygittins-indexsuboptimality-bound

Theorem 4.18. If hhh is an ε\varepsilonε-index policy for an SFABP FFF with the discount parameter γ\gammaγ, then Rh(F)≥R(F)−εγ−1(1−e−γ)−1R_h(F) \ge R(F) - \varepsilon\gamma^{-1}(1 - e^{-\gamma})^{-1}Rh​(F)≥R(F)−εγ−1(1−e−γ)−1.

Formally: for a SFABP in the Bandit Algorithms model (a common countable state space, bounded reward, discount factor a=e−γ∈(0,1)a = e^{-\gamma} \in (0,1)a=e−γ∈(0,1)), ε≥0\varepsilon \ge 0ε≥0, and a policy π\piπ that at every decision time almost surely continues a bandit whose Gittins index is within ε\varepsilonε of the maximal index among the current states, for every initial state-vector xxx,

sup⁡π′Rπ′(x)−ε(1−a)2≤Rπ(x),\sup_{\pi'} R_{\pi'}(x) - \frac{\varepsilon}{(1 - a)^2} \le R_\pi(x),π′sup​Rπ′​(x)−(1−a)2ε​≤Rπ​(x),

where Rπ(x)R_\pi(x)Rπ​(x) is the discounted value of π\piπ from xxx. At ε=0\varepsilon = 0ε=0 this is the index theorem.

The units of ε\varepsilonε. The book's ε\varepsilonε-bounds of §4.10 are in the continuous-time units of the index, reward per unit time with Wτ=E∫0τe−γt dtW_\tau = \mathbb{E}\int_0^\tau e^{-\gamma t}\,dtWτ​=E∫0τ​e−γtdt. Corollary 4.16's bound εγ−1=ε∫0∞e−γtdt\varepsilon\gamma^{-1} = \varepsilon\int_0^\infty e^{-\gamma t}dtεγ−1=ε∫0∞​e−γtdt shows it. For a Markov family with decision times 0,1,2,…0, 1, 2, \dots0,1,2,…, that index is γ/(1−a)\gamma/(1 - a)γ/(1−a) times the discrete-time index (2.6), whose denominator is E∑t<τat\mathbb{E}\sum_{t<\tau} a^tE∑t<τ​at. An ε\varepsilonε-index policy for the discrete index is therefore an εγ/(1−a)\varepsilon\gamma/(1-a)εγ/(1−a)-index policy in the book's units, and the book's bound becomes εγ1−a⋅γ−1(1−e−γ)−1=ε/(1−a)2\frac{\varepsilon\gamma}{1-a}\cdot\gamma^{-1}(1 - e^{-\gamma})^{-1} = \varepsilon/(1-a)^21−aεγ​⋅γ−1(1−e−γ)−1=ε/(1−a)2. Reading the book's constant with the discrete index, i.e. ε/((1−a)ln⁡(1/a))\varepsilon/((1-a)\ln(1/a))ε/((1−a)ln(1/a)), gives a false statement for a<1/ea < 1/ea<1/e. Two one-state bandits paying ν\nuν and ν−ε\nu - \varepsilonν−ε show it: the ε\varepsilonε-index policy that always plays the second loses ε/(1−a)\varepsilon/(1-a)ε/(1−a), which at a=0.1a = 0.1a=0.1 is 1.11ε>0.48ε1.11\varepsilon > 0.48\varepsilon1.11ε>0.48ε. Exact value iteration on 120 random families (up to three bandits with up to three states each) found the worst ε\varepsilonε-index policy within 0.840.840.84 of the bound ε/(1−a)2\varepsilon/(1-a)^2ε/(1−a)2 and never above it, while the other reading was exceeded in 5 families.

Preamble
import Definitions.Def_AllocationIndices_Superprocess

open MeasureTheory ProbabilityTheory BanditAlgorithm
Formal statement
namespace AllocationIndices

theorem eps_index_policy_near_optimal {T : Type*} [MeasurableSpace T] [Countable T]
    [MeasurableSingletonClass T] (P : Kernel T T) [IsMarkovKernel P] {r : T → ℝ}
    (hr : BoundedReward r) {a : ℝ} (ha0 : 0 < a) (ha1 : a < 1) {ε : ℝ} (hε : 0 ≤ ε) {k : ℕ}
    {π : MarkovBanditPolicy k T} (hπ : IsEpsIndexPolicy P r a ε π) (x : Fin k → T) :
    (⨆ π' : MarkovBanditPolicy k T, markovBanditDiscountedValue P r a π' x) -
        ε / (1 - a) ^ 2 ≤
      markovBanditDiscountedValue P r a π x := by sorry

end AllocationIndices
Source
Gittins, Glazebrook and Weber, Multi-armed Bandit Allocation Indices, 2nd ed., Wiley 2011, doi:10.1002/9780470980033, §4.10 p. 112, Theorem 4.18 (Glazebrook 1982c); ε converted from the book's continuous-time index units to the discrete-time index (2.6)
Human review
  • Endorsed by Shuze Chen · Sep 25, 2026

    Confirmed by the moderator at approval.

  • Endorsed by naimengye · Sep 25, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me