Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 11.15: RepeatedHE with ε>0\varepsilon > 0ε>0 and N0N_0N0​ initial samples of arm 1 is BIC as long as ε<13E[G⋅1{G>0}]\varepsilon < \frac13\mathbb{E}[G \cdot \mathbf 1\{G > 0\}]ε<31​E[G⋅1{G>0}], G=GN0+1G = G_{N_0+1}G=GN0​+1​

Proved
IntroBandits.repeatedHE_bic

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

bayesian-incentive-compatibilityhidden-explorationincentivized-exploration

Theorem 11.15. RepeatedHE with exploration probability ε>0\varepsilon > 0ε>0 and N0N_0N0​ initial samples of arm 1 is BIC as long as ε<13 E[G⋅1{G>0}]\varepsilon < \frac13\,\mathbb{E}[G \cdot \mathbf 1\{G > 0\}]ε<31​E[G⋅1{G>0}], where G=GN0+1G = G_{N_0+1}G=GN0​+1​.

Formally: PPP a prior (a probability measure on mean vectors) supported on the finite F⊆[0,1]2F \subseteq [0,1]^2F⊆[0,1]2 with μ10≥μ20\mu^0_1 \ge \mu^0_2μ10​≥μ20​; any reward family with finitely many reward values; any bandit algorithm ALG\mathrm{ALG}ALG; N0∈NN_0 \in \mathbb{N}N0​∈N; ε>0\varepsilon > 0ε>0 with

ε<13∑s∈valuesN0max⁡(0, ∑μ∈FP(μ)∏i≤N0Dμ1({si}) (μ2−μ1))=13 E[G⋅1{G>0}],\varepsilon < \frac13\sum_{s \in \mathrm{values}^{N_0}} \max\Big(0,\ \sum_{\mu \in F} P(\mu)\prod_{i \le N_0} D_{\mu_1}(\{s_i\})\,(\mu_2 - \mu_1)\Big) = \frac13\,\mathbb{E}[G \cdot \mathbf 1\{G > 0\}],ε<31​s∈valuesN0​∑​max(0, μ∈F∑​P(μ)i≤N0​∏​Dμ1​​({si​})(μ2​−μ1​))=31​E[G⋅1{G>0}],

G=E[μ2−μ1∣S1,N0]G = \mathbb{E}[\mu_2 - \mu_1 \mid S_{1,N_0}]G=E[μ2​−μ1​∣S1,N0​​] the posterior gap given the N0N_0N0​ initial samples of arm 1. Then, for every horizon TTT, RepeatedHE (Algorithm 11.2, run with ALG\mathrm{ALG}ALG, N0N_0N0​, ε\varepsilonε) is Bayesian incentive-compatible in the sense of Definition 11.4: for every round t≤Tt \le Tt≤T and every two distinct arms a,a′a, a'a,a′ such that Pr⁡[rect=a]>0\Pr[\mathrm{rec}_t = a] > 0Pr[rect​=a]>0 under the law of the compliant run,

E[(μa−μa′) 1{rect=a}]≥0,i.e.E[μa−μa′∣rect=a]≥0.\mathbb{E}\big[(\mu_a - \mu_{a'})\,\mathbf 1\{\mathrm{rec}_t = a\}\big] \ge 0, \quad\text{i.e.}\quad \mathbb{E}[\mu_a - \mu_{a'} \mid \mathrm{rec}_t = a] \ge 0 .E[(μa​−μa′​)1{rect​=a}]≥0,i.e.E[μa​−μa′​∣rect​=a]≥0.

Reading. The event Et−1E_{t-1}Et−1​ of (11.1) is the sure event of the compliant law. The initial rounds recommend arm 1, so for them the condition reads μ10≥μ20\mu^0_1 \ge \mu^0_2μ10​≥μ20​; every later round is a HiddenExploration with signal StS_tSt​ (Lemma 11.10), and E[Gt1{Gt>0}]\mathbb{E}[G_t \mathbf 1\{G_t > 0\}]E[Gt​1{Gt​>0}] is nondecreasing in ttt because St+1S_{t+1}St+1​ determines StS_tSt​. Nothing is assumed about ALG\mathrm{ALG}ALG; when E[G1{G>0}]=0\mathbb{E}[G \mathbf 1\{G > 0\}] = 0E[G1{G>0}]=0 no ε\varepsilonε qualifies (Remark 11.16).

Preamble
import Definitions.Def_IntroBandits_Agents

open MeasureTheory ProbabilityTheory BanditAlgorithm
Formal statement
namespace IntroBandits

theorem repeatedHE_bic (P : Measure (Fin 2 → ℝ)) [IsProbabilityMeasure P]
    (F : Finset (Fin 2 → ℝ)) (hF : P (↑F)ᶜ = 0)
    (hunit : ∀ μ ∈ F, ∀ a, μ a ∈ Set.Icc (0 : ℝ) 1) (fam : RewardFamily)
    (hprior : priorMean P F 1 ≤ priorMean P F 0) (A : BanditPolicy 2) (N₀ : ℕ) {ε : ℝ}
    (hε : 0 < ε) (hε3 : ε < initialGapPlus P F fam N₀ / 3) (T : ℕ) :
    IsBIC F (repeatedHELaw P F fam A N₀ ε T) recHE := by sorry

end IntroBandits
Source
Slivkins, Introduction to Multi-Armed Bandits, arXiv:1904.07272 (FnT ML 12, 2019), §11.4 p. 151, Theorem 11.15 with its proof (monotonicity of E[G_t 1{G_t > 0}])
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