Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Adaptive history KL is bounded by expected outside-action count

Proved
BanditAlgorithm.pmStochMeasure_kl_toReal_le_expected_outside_count

by Harry_Xu · Aug 5, 2026 · Mathlib c5ea003 (Lean v4.30.0)

banditsinformation-theorypartial-monitoringtextbook

Consider two stochastic partial-monitoring environments uuu and vvv used with the same adaptive policy π\piπ. Let NNN be a set of actions whose feedback laws agree under the two environments. Suppose every action outside NNN has one-round Kullback–Leibler divergence at most a finite constant BBB, and all one-round divergences are finite. If T~N(n)\widetilde T_N(n)TN​(n) denotes the number of rounds up to time nnn on which the learner selects an action outside NNN, then

D(Pun∥Pvn)≤B Eu[T~N(n)].D(P_u^n\|P_v^n) \le B\,\mathbb E_u[\widetilde T_N(n)].D(Pun​∥Pvn​)≤BEu​[TN​(n)].

Here PunP_u^nPun​ and PvnP_v^nPvn​ are the complete feedback-history laws induced by uuu and vvv. This is the real-valued expected-count form of the adaptive KL chain rule used in information-theoretic lower bounds for partial monitoring.

Formalization Note The platform chain-rule theorem is naturally stated in extended nonnegative reals. This theorem records the finiteness argument and converts its predictable selection-mass bound into the expected count on complete histories.

Preamble
import Theorems.Thm_BanditAlgorithm_pmStochMeasure_kl_le_outside_selection
import Theorems.Thm_BanditAlgorithm_pmStoch_expected_actionSetCount_eq_selection_mass

open MeasureTheory ProbabilityTheory InformationTheory
open scoped BigOperators ENNReal
Formal statement
theorem BanditAlgorithm.pmStochMeasure_kl_toReal_le_expected_outside_count
    {k d : ℕ} {𝕊 : Type*} [Fintype 𝕊] [MeasurableSpace 𝕊]
    (G : PartialMonitoringGame k d 𝕊) (π : PMPolicy k 𝕊)
    (u v : Fin d → ℝ) (hu : u ∈ stdSimplex ℝ (Fin d))
    (hv : v ∈ stdSimplex ℝ (Fin d)) (N : Finset (Fin k)) (B : ℝ≥0∞)
    (hB : B ≠ ⊤)
    (hfin : ∀ c, klDiv (pmSignalMeasure G u c) (pmSignalMeasure G v c) ≠ ⊤)
    (heq : ∀ c, c ∈ N → pmSignalMeasure G u c = pmSignalMeasure G v c)
    (hle : ∀ c, c ∉ N → klDiv (pmSignalMeasure G u c)
      (pmSignalMeasure G v c) ≤ B) : ∀ n : ℕ,
    (klDiv (pmStochMeasure G π u hu n)
      (pmStochMeasure G π v hv n)).toReal ≤
      B.toReal *
        ∫ h, ∑ t : Fin n,
          (if (h t).1 ∉ N then (1 : ℝ) else 0)
          ∂pmStochMeasure G π u hu n := by
  sorry
Source
Lattimore and Szepesvári, Bandit Algorithms (Cambridge University Press, 2020), Theorem 37.12 proof, printed p. 490, Eqs. (37.8)–(37.9), with the outside-neighbourhood count defined on printed p. 491.

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