Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Equation (2.2) — the interchange inequality E[(1 − e^{−βτ})∫₀^{T(s*)} r(s*)e^{−βt}dt] ≥ E[(1 − e^{−βT(s*)})∫₀^τ r̄(t)e^{−βt}dt]

Proved
TsitsiklisGittins.IndexTheorem.eq2_2_interchange_inequality

by mikedeng1 · Oct 6, 2026 · Mathlib 0df444a (Lean v4.33.1)

bandit-algorithmsinterchange-argumentp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Let (Ω,F,P)(\Omega,\mathcal F,\mathbb P)(Ω,F,P) be a probability space and β>0\beta>0β>0. Let T≥0T\ge 0T≥0 be a random variable (the duration T(s∗)T(s^*)T(s∗) of a play at the top state s∗s^*s∗), let τ∈[0,∞]\tau\in[0,\infty]τ∈[0,∞] be a random time (the first time bandit i∗i^*i∗ is played, with τ=∞\tau=\inftyτ=∞ if it is never played), and let rˉ(t)\bar r(t)rˉ(t), t≥0t\ge 0t≥0, be a jointly measurable bounded random reward rate with rˉ(t)≤r(s∗)\bar r(t)\le r(s^*)rˉ(t)≤r(s∗) for all ttt. Then

E[(1−e−βτ)∫0T(s∗)r(s∗)e−βt dt] ≥ E[(1−e−βT(s∗))∫0τrˉ(t)e−βt dt].\mathbb E\Big[(1-e^{-\beta\tau})\int_0^{T(s^*)}r(s^*)e^{-\beta t}\,dt\Big]\ \ge\ \mathbb E\Big[(1-e^{-\beta T(s^*)})\int_0^{\tau}\bar r(t)e^{-\beta t}\,dt\Big].E[(1−e−βτ)∫0T(s∗)​r(s∗)e−βtdt] ≥ E[(1−e−βT(s∗))∫0τ​rˉ(t)e−βtdt].

In the proof of Lemma 2.1 this inequality is equivalent to J(π′)≥J(π)J(\pi')\ge J(\pi)J(π′)≥J(π), where π′\pi'π′ plays bandit i∗i^*i∗ once and then mimics the optimal policy π\piπ; it is the interchange step that makes s∗s^*s∗ a top-priority state.

Formalization Note e−β⋅∞=0e^{-\beta\cdot\infty}=0e−β⋅∞=0, and ∫0∞\int_0^{\infty}∫0∞​ is the integral over (0,∞)(0,\infty)(0,∞); τ\tauτ takes values in [0,∞][0,\infty][0,∞]. The statement uses only rˉ(t)≤r(s∗)\bar r(t)\le r(s^*)rˉ(t)≤r(s∗), which is all the page's argument uses; no independence between τ\tauτ and T(s∗)T(s^*)T(s∗) is assumed. Boundedness and measurability of rˉ\bar rrˉ make both expectations finite.

Preamble
import Mathlib
Formal statement
namespace TsitsiklisGittins.IndexTheorem

/-- Equation (2.2) of Tsitsiklis (1994), p. 196: with `τ` the (possibly infinite) first time bandit
`i*` is played, `T` the duration `T(s*)` of a play at `s*`, `r̄(t)` the reward rate under the
policy at time `t` and `r̄(t) ≤ r(s*)` for all `t`,
`E[(1 − e^{−βτ}) ∫₀^{T(s*)} r(s*) e^{−βt} dt] ≥ E[(1 − e^{−βT(s*)}) ∫₀^τ r̄(t) e^{−βt} dt]`.
Here `e^{−β·∞} = 0` and `∫₀^∞` is the integral over `(0, ∞)`. -/
theorem eq2_2_interchange_inequality {Ω : Type*} [MeasurableSpace Ω]
    (P : MeasureTheory.Measure Ω) [MeasureTheory.IsProbabilityMeasure P]
    (β : ℝ) (hβ : 0 < β)
    (T : Ω → ℝ) (hT : Measurable T) (hT0 : ∀ ω, 0 ≤ T ω)
    (τ : Ω → ENNReal) (hτ : Measurable τ)
    (rbar : Ω → ℝ → ℝ) (hrbar : Measurable (Function.uncurry rbar))
    (C : ℝ) (hC : ∀ ω t, |rbar ω t| ≤ C)
    (rstar : ℝ) (hle : ∀ ω t, rbar ω t ≤ rstar) :
    ∫ ω, (1 - (if τ ω = ⊤ then 0 else Real.exp (-(β * (τ ω).toReal)))) *
          (∫ t in (0 : ℝ)..T ω, rstar * Real.exp (-(β * t))) ∂P ≥
      ∫ ω, (1 - Real.exp (-(β * T ω))) *
          (∫ t in {t : ℝ | 0 < t ∧ ENNReal.ofReal t < τ ω}, rbar ω t * Real.exp (-(β * t))) ∂P := by sorry

end TsitsiklisGittins.IndexTheorem
Source
Tsitsiklis, A Short Proof of the Gittins Index Theorem, Ann. Appl. Probab. 4 (1994), p. 196 (PDF p. 3), Section 2, proof of Lemma 2.1, equation (2.2)
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 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