Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

The compensated per-action variable satisfies the scalar interface while the remaining-budget field has not crossed

Proved
G26G.Sharp.runF_SI

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

bandit-algorithmslinear-banditsonline-learning

Setting. A⊆Rd\mathcal A\subseteq\mathbb R^dA⊆Rd is a compact action set spanning Rd\mathbb R^dRd, B⊆A\mathcal B\subseteq\mathcal AB⊆A is a compact generator (A⊆K:=Conv⁡B\mathcal A\subseteq\mathcal K:=\operatorname{Conv}\mathcal BA⊆K:=ConvB), e:N→Qde:\mathbb N\to\mathbb Q^de:N→Qd is the direction enumeration, and RRR is a GIA-FTRL run of TTT rounds (IsGIAFTRLRun) with cumulative loss estimate Sn=∑j<nθ^jS_n=\sum_{j<n}\hat\theta_jSn​=∑j<n​θ^j​, run filtration (Gn)(\mathcal G_n)(Gn​), learning rate βn=16d+n\beta_n=16d+\sqrt nβn​=16d+n​ (gbeta) and magnitude constant B=236(d+1)B=2^{36}(d+1)B=236(d+1) (gB). The filtration used below is Fn=Gn\mathcal F_n=\mathcal G_nFn​=Gn​ for n<Tn<Tn<T and FT=F\mathcal F_T=\mathcal FFT​=F (Fx). The standing hypotheses Sharp hold: A\mathcal AA lies in the unit ball, k=dim⁡K≥1k=\dim\mathcal K\ge1k=dimK≥1, ∥θ∥A≤1\|\theta\|_{\mathcal A}\le1∥θ∥A​≤1, a∗∈Aa^*\in\mathcal Aa∗∈A is the unique minimizer of a↦a⊤θa\mapsto a^\top\thetaa↦a⊤θ on A\mathcal AA, and the instance is sharp with modulus ς>0\varsigma>0ς>0 (H.sig): Δ(a)=(a−a∗)⊤θ≥ς∥a−a∗∥S\Delta(a)=(a-a^*)^\top\theta\ge\varsigma\|a-a^*\|_{\mathcal S}Δ(a)=(a−a∗)⊤θ≥ς∥a−a∗∥S​ on K\mathcal KK. The corruption is bounded by CCC: ∑t<T∥θt−θ∥A≤C\sum_{t<T}\|\theta_t-\theta\|_{\mathcal A}\le C∑t<T​∥θt​−θ∥A​≤C a.s. (IsCorruptedG). The projective body is Uθ={Δ(a)−1(a−a∗):a∈K∖{a∗}}‾U_\theta=\overline{\{\Delta(a)^{-1}(a-a^*):a\in\mathcal K\setminus\{a^*\}\}}Uθ​={Δ(a)−1(a−a∗):a∈K∖{a∗}}​ and the cone body is Dθ={λu:λ∈[0,1],u∈Uθ}D_\theta=\{\lambda u:\lambda\in[0,1],u\in U_\theta\}Dθ​={λu:λ∈[0,1],u∈Uθ​}; Λ=ς−1\Lambda=\varsigma^{-1}Λ=ς−1, An=∑j<n∥θj−θ∥AA_n=\sum_{j<n}\|\theta_j-\theta\|_{\mathcal A}An​=∑j<n​∥θj​−θ∥A​, and j∗j^*j∗ is the index of a∗a^*a∗ among the reference atoms (js).

Fix a∈Ka\in\mathcal Ka∈K and a buffer sequence τ\tauτ with τn>0\tau_n>0τn​>0, τn≤τn+1≤τn+1\tau_n\le\tau_{n+1}\le\tau_n+1τn​≤τn+1​≤τn​+1, an interval end N≤TN\le TN≤T, a start m0m_0m0​, and assume B(βn+1+3)≤τnB(\beta_{n+1}+3)\le\tau_nB(βn+1​+3)≤τn​ for n≥m0n\ge m_0n≥m0​. Consider the compensated variable

YnF(a)=(a−a∗)⊤Sn−En(a−a∗)+2C+2τn,En(w)=∑j<nw⊤(θj−θ),Y^F_n(a)=(a-a^*)^\top S_n-E_n(a-a^*)+2C+2\tau_n,\qquad E_n(w)=\sum_{j<n}w^\top(\theta_j-\theta),YnF​(a)=(a−a∗)⊤Sn​−En​(a−a∗)+2C+2τn​,En​(w)=j<n∑​w⊤(θj​−θ),

and the alive events InW={max⁡K(−Wj)≤τj for all m0≤j≤n}I^W_n=\{\max_{\mathcal K}(-W_j)\le\tau_j\text{ for all }m_0\le j\le n\}InW​={maxK​(−Wj​)≤τj​ for all m0​≤j≤n} (IW), where WWW is the remaining-budget field of Lemma G.30. Then SI(m0,N)\mathrm{SI}(m_0,N)SI(m0​,N) holds for (YF(a),IW)(Y^F(a),I^W)(YF(a),IW) with thresholds τn\tau_nτn​, budgets βn+1\beta_{n+1}βn+1​ and B=236(d+1)B=2^{36}(d+1)B=236(d+1).

On InWI^W_nInW​ one has YnF(a)≥τnY^F_n(a)\ge\tau_nYnF​(a)≥τn​, because Wn(a)≥−τnW_n(a)\ge-\tau_nWn​(a)≥−τn​ and the compensator differs from the shift by at most 2C2C2C. The relative increment and conditional variance bounds come from Lemma G.10. These variables carry the far-reference estimate of Lemma G.29.

Formalization Note Rounds are 0-based.

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 IAFTRL Matrix
open G26G
open G26 G26R
variable {d : ℕ} {𝒜 ℬ : Set (Fin d → ℝ)} {e : ℕ → Fin d → ℚ} {θ astar : Fin d → ℝ}
variable {Ω : Type*} [mΩ : MeasurableSpace Ω] {P : Measure Ω} [IsProbabilityMeasure P]
  {T : ℕ} {R : GRun d Ω}
open G26G.Sharp
variable (H : Sharp 𝒜 ℬ θ astar)
include H
Formal statement
theorem G26G.Sharp.runF_SI (hR : IsGIAFTRLRun P 𝒜 ℬ e T R) {C : ℝ} (hcor : IsCorruptedG P 𝒜 θ C T R)
    {a : Fin d → ℝ} (ha : a ∈ Kset ℬ) {τ : ℕ → ℝ} (hτpos : ∀ n, 0 < τ n)
    (hτmono : ∀ n, τ n ≤ τ (n + 1)) (hδ : ∀ n, τ (n + 1) - τ n ≤ 1) {m0 N : ℕ} (hN : N ≤ T)
    (hτB : ∀ n, m0 ≤ n → gB d * (gbeta d (n + 1) + 3) ≤ τ n) :
    SI P (Fx R T) (YF R θ astar C τ a) (IW 𝒜 ℬ R θ astar C τ m0) m0 N τ
      (fun n => gbeta d (n + 1)) (gB d) := 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.29, (G.635)–(G.641) (p.184)

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