Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.3: under Condition D, the index policy that continues a superprocess of maximal ν(S_i, x_i) with its Condition-D control is optimal for a SFAS

Proved
AllocationIndices.superprocess_index_theorem

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

condition-dgittins-indexindex-theoremsuperprocesses

Theorem 4.3 (Index theorem for a SFAS). A policy for a simple family of alternative superprocesses comprised of S1,…,SnS_1, \dots, S_nS1​,…,Sn​, which all satisfy Condition D, is optimal if it is an index policy with respect to ν(S1,⋅,⋅),…,ν(Sn,⋅,⋅)\nu(S_1, \cdot, \cdot), \dots, \nu(S_n, \cdot, \cdot)ν(S1​,⋅,⋅),…,ν(Sn​,⋅,⋅).

Formally, in discrete time: DDD a decision process on a countable state space SSS with finite control type UUU and bounded rewards, a∈(0,1)a \in (0,1)a∈(0,1), and ggg a Condition-D control for DDD (ggg feasible, and whenever it is optimal to select SSS in state xxx in {S,Λ(λ)}\{S, \Lambda(\lambda)\}{S,Λ(λ)} it is optimal to apply g(x)g(x)g(x); the nnn superprocesses of the family are copies of DDD, a family of distinct superprocesses being the disjoint-union case). If a policy π\piπ for the family of nnn superprocesses at every decision time almost surely continues a superprocess iii of maximal index ν(D,xi)=max⁡u∈Γ(xi)ν(D,xi,u)\nu(D, x_i) = \max_{u \in \Gamma(x_i)} \nu(D, x_i, u)ν(D,xi​)=maxu∈Γ(xi​)​ν(D,xi​,u) and applies the control g(xi)g(x_i)g(xi​), then π\piπ is optimal: it is feasible and its payoff equals the supremum of the payoffs of all feasible policies from every initial state-vector. Since ν(D,x,g(x))=ν(D,x)\nu(D, x, g(x)) = \nu(D, x)ν(D,x,g(x))=ν(D,x) (Note 4.2 at λ=ν(D,x)\lambda = \nu(D, x)λ=ν(D,x)), π\piπ is an index policy with respect to ν(D,⋅,⋅)\nu(D, \cdot, \cdot)ν(D,⋅,⋅).

Why the control is pinned to ggg. The book's proof runs the index policy with the controls gi(xi)g_i(x_i)gi​(xi​) of Condition D. An index policy that breaks a tie among controls differently need not be optimal. Take n=1n = 1n=1, 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, and u′u'u′ pays 111 and leads to a state paying 000 forever. Both have index 111, Condition D holds with g(x)=ug(x) = ug(x)=u, and the index policy that applies u′u'u′ earns 111 against 222.

With singleton control sets this is Theorem 2.1 (the proved platform theorem). The book's proof reduces the semi-Markov case to stochastic discounting and runs a prevailing-stake argument; here the family is Markov with unit decision intervals.

Preamble
import Definitions.Def_AllocationIndices_Superprocess

open MeasureTheory ProbabilityTheory BanditAlgorithm
Formal statement
namespace AllocationIndices

theorem superprocess_index_theorem {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) {g : S → U} (hg : IsConditionDControl D a g)
    {n : ℕ} {π : SFASPolicy n S U} (hπ : IsSuperIndexPolicy D a g π) :
    IsOptimalSFASPolicy D a π := by sorry

end AllocationIndices
Source
Gittins, Glazebrook and Weber, Multi-armed Bandit Allocation Indices, 2nd ed., Wiley 2011, doi:10.1002/9780470980033, §4.3 pp. 83-85, Theorem 4.3 (index theorem for a SFAS, Whittle 1980) with its proof; the control applied is the Condition-D control g, as in the proof, since an index policy breaking ties among controls otherwise need not be optimal
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