Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

BanditAlgorithm.partial_monitoring_hard_game_lower_bound

Proved

by Shuze Chen · Jul 30, 2026 · Mathlib c5ea003 (Lean v4.30.0)

lower-boundsonline-learningpartial-monitoring

(Hard games, Theorem 37.12) If GGG is globally observable but not locally observable, then there exists a game-dependent constant cG>0c_G > 0cG​>0 such that for every horizon nnn,

Rn∗(G)≥cG n2/3.R_n^*(G) \ge c_G\, n^{2/3}.Rn∗​(G)≥cG​n2/3.
Preamble
import Definitions.Def_PartialMonitoringGame
import Mathlib.Analysis.SpecialFunctions.Pow.Real


open MeasureTheory ProbabilityTheory
Formal statement
theorem BanditAlgorithm.partial_monitoring_hard_game_lower_bound {k d : ℕ} {𝕊 : Type*}
    [Fintype 𝕊] [MeasurableSpace 𝕊] (G : PartialMonitoringGame k d 𝕊)
    (hglob : GloballyObservable G) (hloc : ¬ LocallyObservable G) :
    ∃ c : ℝ, 0 < c ∧
      ∀ n : ℕ, c * (n : ℝ) ^ ((2 : ℝ) / 3) ≤ pmMinimaxRegret G n := by
  sorry
Source
L&S Theorem 37.12, p.488
Read-back

What the Lean code literally says, in plain math · claude-fable-5

Setup. GGG is a partial-monitoring game: losses L(a,i)∈RL(a,i)\in\mathbb RL(a,i)∈R and signals Φ(a,i)∈Σ\Phi(a,i)\in\SigmaΦ(a,i)∈Σ over kkk actions and ddd outcomes. Rn∗\mathcal R^*_nRn∗​ is its minimax regret at horizon nnn: the infimum over signal-only policies π\piπ of the supremum over deterministic outcome sequences j∈{1,…,d}nj\in\{1,\dots,d\}^nj∈{1,…,d}n of max⁡aE[∑t<n(L(At,jt)−L(a,jt))]\max_a\mathbb E\big[\sum_{t<n}(L(A_t,j_t)-L(a,j_t))\big]maxa​E[∑t<n​(L(At​,jt​)−L(a,jt​))], where the expectation is over the interaction law in which the round-ttt action comes from π\piπ applied to the past (action, signal) pairs and the observed signal is Φ(At,jt)\Phi(A_t,j_t)Φ(At​,jt​). A pair of actions is neighbouring when both cells (sets of outcome distributions making the action weakly optimal) are nonempty of affine dimension d−1d-1d−1 and their intersection is nonempty of affine dimension d−2d-2d−2; a global estimator for such a pair (a,b)(a,b)(a,b) is a function fff of (action, signal) with ∑cf(c,Φ(c,i))=L(a,i)−L(b,i)\sum_c f(c,\Phi(c,i))=L(a,i)-L(b,i)∑c​f(c,Φ(c,i))=L(a,i)−L(b,i) at every outcome iii; a local estimator additionally vanishes on actions whose cell does not contain the intersection.

Assertion.

∃ c>0:∀ n∈N,c n2/3  ≤  Rn∗.\exists\,c>0:\quad\forall\,n\in\mathbb N,\quad c\,n^{2/3}\;\le\;\mathcal R^*_n.∃c>0:∀n∈N,cn2/3≤Rn∗​.

Hypotheses.

  • Σ\SigmaΣ is finite and carries a σ\sigmaσ-algebra; kkk, ddd, and the losses are arbitrary (no boundedness of LLL).
  • GGG is globally observable: every neighbouring pair admits a global estimator.
  • GGG is not locally observable — unfolding the negation: some neighbouring pair admits no local estimator (in particular, at least one neighbouring pair exists).

Edge cases.

  • The inequality is claimed for all nnn, including n=0n=0n=0, where it degenerates to 0≤R0∗0\le\mathcal R^*_00≤R0∗​ (n2/3n^{2/3}n2/3 is the real power, 02/3=00^{2/3}=002/3=0).
  • Rn∗\mathcal R^*_nRn∗​ inherits the real inf/sup junk conventions (empty or unbounded sets give the value 000); the hypotheses do not exclude these degenerate regimes explicitly.
Human review
  • Endorsed by Community (Bot) · Jul 30, 2026

  • Endorsed by Shuze Chen · Jul 30, 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