Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Note 4.2 (corrected): under Condition D, S is selected in {S, Λ(λ)} iff ν(S, x) ≥ λ, and at λ = ν(S, x) a control u is optimal iff ν(S, x, u) = ν(S, x)

Proved
AllocationIndices.conditionD_optimal_control_iff

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

condition-dgittins-indexsuperprocesses

Note 4.2. If a superprocess SSS satisfies Condition D, a control u∈Γ(x)u \in \Gamma(x)u∈Γ(x) is optimal for {S,Λ}\{S, \Lambda\}{S,Λ} if and only if ν(S,x,u)=ν(S,x)≥λ\nu(S, x, u) = \nu(S, x) \ge \lambdaν(S,x,u)=ν(S,x)≥λ. This follows by supposing that Λ\LambdaΛ in the statement of Condition D has the parameter ν(S,x)\nu(S, x)ν(S,x).

Stated here, corrected. For a decision process on a countable state space with bounded rewards, a∈(0,1)a \in (0,1)a∈(0,1), satisfying Condition D, and every state xxx:

  1. for every λ\lambdaλ, it is optimal to select SSS in state xxx in the SFAS {S,Λ(λ)}\{S, \Lambda(\lambda)\}{S,Λ(λ)} (some optimal policy continues SSS at time 000 from (x,Λ)(x, \Lambda)(x,Λ)) if and only if λ≤ν(S,x)=max⁡v∈Γ(x)ν(S,x,v)\lambda \le \nu(S, x) = \max_{v \in \Gamma(x)} \nu(S, x, v)λ≤ν(S,x)=maxv∈Γ(x)​ν(S,x,v);
  2. for λ=ν(S,x)\lambda = \nu(S, x)λ=ν(S,x), the parameter the note's own justification uses, a control u∈Γ(x)u \in \Gamma(x)u∈Γ(x) is optimal to apply (some optimal policy continues SSS with uuu at time 000 from (x,Λ)(x, \Lambda)(x,Λ)) if and only if ν(S,x,u)=ν(S,x)\nu(S, x, u) = \nu(S, x)ν(S,x,u)=ν(S,x).

Why not every λ\lambdaλ. The printed equivalence fails for λ<ν(S,x)\lambda < \nu(S, x)λ<ν(S,x), in both directions. Take a=1/2a = 1/2a=1/2 and a state xxx with two controls. uuu pays 111 and leads to a state paying 111 forever. u′u'u′ pays 111 and leads to a state paying 000 forever. Both have index 1=ν(S,x)1 = \nu(S, x)1=ν(S,x), and Condition D holds with g(x)=ug(x) = ug(x)=u. At λ=1/2\lambda = 1/2λ=1/2, applying u′u'u′ earns 1+1/2=3/21 + 1/2 = 3/21+1/2=3/2 against 222, so u′u'u′ is not optimal, although ν(S,x,u′)=ν(S,x)≥λ\nu(S, x, u') = \nu(S, x) \ge \lambdaν(S,x,u′)=ν(S,x)≥λ. Conversely, let uuu pay 111 and then 000 forever and u′u'u′ pay 0.90.90.9 and then 0.10.10.1 forever. Then ν(S,x)=1>ν(S,x,u′)=0.9\nu(S, x) = 1 > \nu(S, x, u') = 0.9ν(S,x)=1>ν(S,x,u′)=0.9, Condition D holds with g(x)=ug(x) = ug(x)=u, and at λ=0\lambda = 0λ=0 both controls earn 111, so u′u'u′ is optimal. At λ=ν(S,x)\lambda = \nu(S, x)λ=ν(S,x) the note is right: selecting SSS is exactly as good as retiring, and a control is optimal exactly when some stationary policy starting with it has index ν(S,x)\nu(S, x)ν(S,x) (the constrained stopping problem has an optimal stationary policy, so the supremum in (4.1) is attained where it matters).

Preamble
import Definitions.Def_AllocationIndices_Superprocess

open MeasureTheory ProbabilityTheory BanditAlgorithm
Formal statement
namespace AllocationIndices

theorem conditionD_optimal_control_iff {S U : Type*} [MeasurableSpace S] [Countable S]
    [MeasurableSingletonClass S] [MeasurableSpace U] [Fintype U] [Nonempty U]
    [MeasurableSingletonClass U] (D : DecisionProcess S U) (hD : D.BoundedRewards)
    {a : ℝ} (ha0 : 0 < a) (ha1 : a < 1) (hcond : ConditionD D a) (x : S) :
    (∀ lam : ℝ, OptimalToSelect D a lam x ↔ lam ≤ superIndexMax D a x) ∧
    (∀ u ∈ D.avail x, OptimalToApply D a (superIndexMax D a x) x u ↔
      superIndex D a x u = superIndexMax D a x) := by sorry

end AllocationIndices
Source
Gittins, Glazebrook and Weber, Multi-armed Bandit Allocation Indices, 2nd ed., Wiley 2011, doi:10.1002/9780470980033, §4.2 p. 83, Note 4.2; stated at the parameter λ = ν(S, x) the note itself uses, since the printed claim fails for λ < ν(S, x)
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