Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Killed inverse moment of a scalar interface process with positive drift

Proved
G26.SI.inv_moment

by Canzhe Zhao · Nov 18, 2026 · Mathlib 0df444a (Lean v4.33.1)

bandit-algorithmslinear-banditsonline-learning

Throughout, (Ω,F,P)(\Omega,\mathcal F,\mathbb P)(Ω,F,P) is a probability space, (Fn)n≤N(\mathcal F_n)_{n\le N}(Fn​)n≤N​ are sub-σ\sigmaσ-algebras, Yn:Ω→RY_n:\Omega\to\mathbb RYn​:Ω→R is a process, InI_nIn​ are alive events with In+1⊆InI_{n+1}\subseteq I_nIn+1​⊆In​, and τn>0\tau_n>0τn​>0, βn\beta_nβn​, BBB are reals. The scalar interface SI(m,N)\mathrm{SI}(m,N)SI(m,N) asks that YnY_nYn​ and InI_nIn​ be Fn\mathcal F_nFn​-measurable, that YnY_nYn​ be a.s. bounded, and that for m≤n≤Nm\le n\le Nm≤n≤N (resp. m≤n<Nm\le n<Nm≤n<N), almost surely on InI_nIn​,

Yn≥τn,∣Yn+1−Yn∣≤3YnB,E[(Yn+1−Yn)2 ∣ Fn]≤10βn+2Yn2B(n+1).Y_n\ge\tau_n,\qquad |Y_{n+1}-Y_n|\le\frac{3Y_n}{B},\qquad \mathbb E\big[(Y_{n+1}-Y_n)^2\,\big|\,\mathcal F_n\big]\le 10\beta_n+\frac{2Y_n^2}{B(n+1)} .Yn​≥τn​,∣Yn+1​−Yn​∣≤B3Yn​​,E[(Yn+1​−Yn​)2​Fn​]≤10βn​+B(n+1)2Yn2​​.

Write κn−=E[1InYn−P]\kappa^-_n=\mathbb E[\mathbf 1_{I_n}Y_n^{-P}]κn−​=E[1In​​Yn−P​] (kinv). Assume SI(m,N)\mathrm{SI}(m,N)SI(m,N) holds, P≥2P\ge2P≥2, 3P/B≤1/23P/B\le1/23P/B≤1/2, 8(P+1)≤B8(P+1)\le B8(P+1)≤B, and Δ>0\Delta>0Δ>0. Let m≤t≤Nm\le t\le Nm≤t≤N with 2(m+1)≤t+12(m+1)\le t+12(m+1)≤t+1, (P+1)Δ≤τm(P+1)\Delta\le\tau_m(P+1)Δ≤τm​ and P+1≤2(m+1)P+1\le2(m+1)P+1≤2(m+1). Assume that for m≤n<tm\le n<tm≤n<t the conditional drift is at least Δ\DeltaΔ a.s. on InI_nIn​, E[Yn+1−Yn∣Fn]≥Δ\mathbb E[Y_{n+1}-Y_n\mid\mathcal F_n]\ge\DeltaE[Yn+1​−Yn​∣Fn​]≥Δ, and that βn≥0\beta_n\ge0βn​≥0 with 40(P+1)βn(n+1)≤τn240(P+1)\beta_n(n+1)\le\tau_n^240(P+1)βn​(n+1)≤τn2​. Then

E[1It Yt−P]≤(4(t+1)Δ)P.\mathbb E\big[\mathbf 1_{I_t}\,Y_t^{-P}\big]\le\Big(\frac{4}{(t+1)\Delta}\Big)^P .E[1It​​Yt−P​]≤((t+1)Δ4​)P.

This is the inverse-moment estimate of Lemma G.22, isolated from the bandit run. Lemma G.29 applies it to the compensated per-action variables (with Δ\DeltaΔ a positive gap), and Lemma G.28 applies it to the normalized per-direction variables (with Δ=1\Delta=1Δ=1).

Formalization Note Rounds are 0-based, so t+1t+1t+1 plays the role of the paper's ttt. An inverse multiplied by a zero indicator is zero.

Preamble
import Mathlib
import Definitions.Def_IAFTRL_LinEnv
import Definitions.Def_IAFTRL_GIAFTRL
import Definitions.Def_IAFTRL_G26_body
import Definitions.Def_IAFTRL_G26_run
import Definitions.Def_IAFTRL_G26_window

open MeasureTheory ProbabilityTheory Set
open G26
variable {Ω : Type*}
variable [mΩ : MeasurableSpace Ω] {μ : Measure Ω} {F : ℕ → MeasurableSpace Ω}
  {Y : ℕ → Ω → ℝ} {I : ℕ → Set Ω} {m N : ℕ} {τ β : ℕ → ℝ} {B : ℝ}
Formal statement
theorem G26.SI.inv_moment [IsProbabilityMeasure μ] (S : SI μ F Y I m N τ β B) {P : ℕ} (hP : 2 ≤ P)
    (hPB : (P : ℝ) * (3 / B) ≤ 1 / 2) (hB8 : 8 * ((P : ℝ) + 1) ≤ B) {Δ : ℝ} (hΔ : 0 < Δ)
    {t : ℕ} (hmt : m ≤ t) (htN : t ≤ N) (hmhalf : 2 * ((m : ℝ) + 1) ≤ (t : ℝ) + 1)
    (hτm : ((P : ℝ) + 1) * Δ ≤ τ m) (hmP : (P : ℝ) + 1 ≤ 2 * ((m : ℝ) + 1))
    (hdr : ∀ n, m ≤ n → n < t → ∀ᵐ ω ∂μ, ω ∈ I n → Δ ≤ μ[fun ω => Y (n + 1) ω - Y n ω | F n] ω)
    (hτβ : ∀ n, m ≤ n → n < t → 40 * ((P : ℝ) + 1) * β n * ((n : ℝ) + 1) ≤ τ n ^ 2)
    (hβ : ∀ n, 0 ≤ β n) :
    kinv μ Y I P t ≤ (4 / (((t : ℝ) + 1) * Δ)) ^ P := by sorry
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, Lemma G.22 (p.165), second display, and (G.568) of Lemma G.28 (p.178), abstracted

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