Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Reference identification for GIA-FTRL on Laplace histories

Proved
IAFTRL.lemma_G_ref_id

by Canzhe Zhao · Nov 17, 2026 · Mathlib 0df444a (Lean v4.33.1)

bandit-algorithmslinear-banditsonline-learning

Setting. The sharp instance of Proposition 3.1. Here A\mathcal AA is a compact action set in the unit ball spanning Rd\mathbb R^dRd, and B\mathcal BB is a compact generator. The enumeration eee is surjective, a∗a^*a∗ is the unique optimal action, the instance is sharp, and Δ(a)=(a−a∗)⊤θ\Delta(a)=(a-a^*)^\top\thetaΔ(a)=(a−a∗)⊤θ. GIA-FTRL chooses at round nnn the reference action brnb_{r_n}brn​​ of (3.4), where rnr_nrn​ is the least index j∈Jj\in Jj∈J minimizing bj⊤S+βnhjb_j^\top S+\beta_nh_jbj⊤​S+βn​hj​ over the candidate atoms bjb_jbj​ with reinforced priorities hjh_jhj​ of (3.2).

Claim. There is a round TrT_rTr​, depending only on the instance, such that the following holds for every n≥Trn\ge T_rn≥Tr​ and every S∈RdS\in\mathbb R^dS∈Rd. If

(a−a∗)⊤S ≥ 4n7/8Δ(a)for all a∈K,(a-a^*)^\top S\ \ge\ 4n^{7/8}\Delta(a)\qquad\text{for all }a\in\mathcal K,(a−a∗)⊤S ≥ 4n7/8Δ(a)for all a∈K,

then the reference action is brn=a∗b_{r_n}=a^*brn​​=a∗.

Formalization Note. This is the deterministic part of the statement "on GtcorG_t^{\rm cor}Gtcor​ the reference is a∗a^*a∗" in Lemma G.26. The Laplace lower bound of (G.554) holds on GtcorG_t^{\rm cor}Gtcor​, and the claim shows that this bound alone forces the reference to be a∗a^*a∗ for large nnn.

Preamble
import Mathlib
import Definitions.Def_IAFTRL_LinEnv
import Definitions.Def_IAFTRL_GIAFTRL
Formal statement
namespace IAFTRL

open MeasureTheory

/-- **Reference identification on Laplace histories.** Zhao–Tan, *Best-of-Both-Worlds FTRL for Linear Bandits* (STOC submission version), Lemma G.26, p.174: "On `G_t^cor` the reference is
`a*`", with the reference index (3.4) `r_t = argmin_{j∈J} {b_jᵀ L̂_{t−1} + β_t h_j}` and the
reinforced priorities (3.2).

Deterministic form: in the sharp setting there is a round `T_r` such that for every round
`n ≥ T_r` and every cumulative estimate `S` with `(a − a*)ᵀS ≥ 4 n^{7/8} Δ(a)` on `𝒦`, the
reference action of program (3.6) is `b_{r_n} = a*`. -/
theorem lemma_G_ref_id {d : ℕ} (hd : 1 ≤ d) (𝒜 ℬ : Set (Fin d → ℝ)) (e : ℕ → Fin d → ℚ)
    (h𝒜 : IsCompact 𝒜) (hℬ : IsCompact ℬ) (hgen : IsGenerator 𝒜 ℬ)
    (hball : ∀ a ∈ 𝒜, a ⬝ᵥ a ≤ 1) (hspan : Submodule.span ℝ 𝒜 = ⊤)
    (he : Function.Surjective e) (θ astar : Fin d → ℝ)
    (hastar : astar ∈ 𝒜) (huniq : ∀ a ∈ 𝒜, a ≠ astar → astar ⬝ᵥ θ < a ⬝ᵥ θ)
    (hsharp : IsSharp 𝒜 θ astar) :
    ∃ Tr : ℕ, ∀ (n : ℕ) (S : Fin d → ℝ), Tr ≤ n →
      (∀ a ∈ Kset ℬ, 4 * (n : ℝ) ^ ((7 : ℝ) / 8) * ((a - astar) ⬝ᵥ θ) ≤ (a - astar) ⬝ᵥ S) →
      gref ℬ e n S = astar := by sorry

end IAFTRL
Source
Zhao–Tan, Best-of-Both-Worlds FTRL for Linear Bandits: The Interplay of Information Acquisition, Corruption, and Geometry (STOC submission version; arXiv:2611.01299), Appendix G.6, Lemma G.26 ("On G_t^cor the reference is a*"), with the reference rule (3.4) and priorities (3.2)

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