Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 3.3 — expected regret of Exp3.P against adaptive adversaries

Proved
RegretBandits.Adversarial.exp3P_expected_regret

by mikedeng1 · Oct 4, 2026 · Mathlib 0df444a (Lean v4.33.1)

adversarial-banditsbanditsp2o-batch-pfp1bp2o-gran-per-chapterp2o-plan-bookp2o-v1regret-bounds

Let K≥2K \ge 2K≥2 arms and a horizon n≥1n \ge 1n≥1 be given, and run Exp3.P with

β=ln⁡KnK,η=0.95ln⁡KnK,γ=1.05Kln⁡Kn\beta = \sqrt{\frac{\ln K}{nK}}, \qquad \eta = 0.95\sqrt{\frac{\ln K}{nK}}, \qquad \gamma = 1.05\sqrt{\frac{K \ln K}{n}}β=nKlnK​​,η=0.95nKlnK​​,γ=1.05nKlnK​​

against any adaptive adversary, whose gains gi,t∈[0,1]g_{i,t} \in [0,1]gi,t​∈[0,1] may depend on the forecaster's past actions. Then the expected regret satisfies

E Rn≤5.15nKln⁡K+nKln⁡K,\mathbb E\, R_n \le 5.15\sqrt{nK\ln K} + \sqrt{\frac{nK}{\ln K}},ERn​≤5.15nKlnK​+lnKnK​​,

where Rn=max⁡i∑t=1ngi,t−∑t=1ngIt,tR_n = \max_{i} \sum_{t=1}^n g_{i,t} - \sum_{t=1}^n g_{I_t,t}Rn​=maxi​∑t=1n​gi,t​−∑t=1n​gIt​,t​ is the regret, with the maximum inside the expectation.

Against a non-oblivious adversary, a bound on the pseudo-regret does not bound the expected regret; this theorem gives an O(nKln⁡K)O(\sqrt{nK\ln K})O(nKlnK​) bound on the expected regret itself, matching the lower bound nK/20\sqrt{nK}/20nK​/20 of Theorem 3.4 up to a logarithmic factor.

Formalization Note The expectation is the integral of the regret under the law of the run; the regret is bounded by nnn and measurable, so the integral is not a default value. The statement is asserted for every n≥1n \ge 1n≥1. When 1.05Kln⁡K/n>11.05\sqrt{K\ln K/n} > 11.05KlnK/n​>1 the rule can have negative entries (and then no run exists), and in that range the bound already follows from Rn≤nR_n \le nRn​≤n.

Preamble
import Mathlib
import Definitions.Def_RegretBandits_Adversarial_Protocol
import Definitions.Def_RegretBandits_Adversarial_Exp3P

open MeasureTheory
Formal statement
namespace RegretBandits.Adversarial

/-- Theorem 3.3 (Bubeck–Cesa-Bianchi, arXiv:1204.5721v2, p. 32), expected regret of Exp3.P. If
Exp3.P is run with `β = √(ln K/(nK))`, `η = 0.95 √(ln K/(nK))`, `γ = 1.05 √(K ln K/n)` against any
adaptive adversary with gains in `[0,1]`, then
`E R_n ≤ 5.15 √(n K ln K) + √(nK/ln K)`, eq. (3.17), where `R_n` is the regret (maximum inside
the expectation). -/
theorem exp3P_expected_regret {K : ℕ} (hK : 2 ≤ K) (n : ℕ) (hn : 1 ≤ n) (g : Adversary K)
    {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
    (I : ℕ → Ω → Fin K)
    (hI : IsRun P
      (exp3PProb (Real.sqrt (Real.log K / (n * K)))
        (1.05 * Real.sqrt (K * Real.log K / n))
        (0.95 * Real.sqrt (Real.log K / (n * K))) g) I) :
    ∫ ω, regret g n I ω ∂P ≤
      5.15 * Real.sqrt (n * K * Real.log K) + Real.sqrt (n * K / Real.log K) := by sorry

end RegretBandits.Adversarial
Source
Bubeck, Cesa-Bianchi, Regret Analysis of Stochastic and Nonstochastic Multi-armed Bandit Problems, arXiv:1204.5721v2, p. 32, Theorem 3.3, Eq. (3.17)
Human review
  • Endorsed by Shuze Chen · Oct 4, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 4, 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