Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2.1 (2.13) — unsuccessful iterations bounded by successful ones

Proved
AdaptiveCubic.SecondOrder.theorem_2_1

by mikedeng1 · Oct 8, 2026 · Mathlib 0df444a (Lean v4.33.1)

cubic-regularisationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1worst-case-complexity

Consider a run of the ARC algorithm and fix j≥0j\ge0j≥0. Let Sj={k≤j:ρk≥η1}\mathcal S_j=\{k\le j: \rho_k\ge\eta_1\}Sj​={k≤j:ρk​≥η1​} and Uj={k≤j:ρk<η1}\mathcal U_j=\{k\le j:\rho_k<\eta_1\}Uj​={k≤j:ρk​<η1​} be the successful and unsuccessful iterations up to jjj (2.9). Assume (2.10): for some γ3∈(0,1]\gamma_3\in(0,1]γ3​∈(0,1], σk+1≥γ3σk\sigma_{k+1}\ge\gamma_3\sigma_kσk+1​≥γ3​σk​ on every very successful iteration k≤jk\le jk≤j. Let σˉ>0\bar\sigma>0σˉ>0 satisfy σk≤σˉ\sigma_k\le\bar\sigmaσk​≤σˉ for all k≤j+1k\le j+1k≤j+1. Then

∣Uj∣≤⌈−log⁡γ3log⁡γ1 ∣Sj∣+1log⁡γ1log⁡σˉσ0⌉.|\mathcal U_j| \le \left\lceil -\frac{\log\gamma_3}{\log\gamma_1}\,|\mathcal S_j| + \frac{1}{\log\gamma_1}\log\frac{\bar\sigma}{\sigma_0}\right\rceil .∣Uj​∣≤⌈−logγ1​logγ3​​∣Sj​∣+logγ1​1​logσ0​σˉ​⌉.

The weight grows by at least γ1\gamma_1γ1​ on every unsuccessful iteration and cannot shrink too fast on successful ones, so a bound on σk\sigma_kσk​ caps the number of unsuccessful iterations in terms of the successful ones. This converts every bound on successful iterations into a bound on all iterations.

Formalization Note The page assumes σk≤σˉ\sigma_k\le\bar\sigmaσk​≤σˉ only for k≤jk\le jk≤j (2.12). Its proof bounds σj+1\sigma_{j+1}σj+1​ (the product in (2.15) covers iterations 0,…,j0,\dots,j0,…,j), and the printed statement fails for j=0j=0j=0 with iteration 000 unsuccessful and σˉ=σ0\bar\sigma=\sigma_0σˉ=σ0​ (∣U0∣=1|\mathcal U_0|=1∣U0​∣=1 but the right-hand side is 000). The hypothesis is therefore posed for k≤j+1k\le j+1k≤j+1.

Preamble
import Mathlib
import Definitions.Def_AdaptiveCubic_Cauchy_IsARCRun

open scoped RealInnerProductSpace
Formal statement
namespace AdaptiveCubic.SecondOrder

/-- Theorem 2.1, (2.13), p. 5 (Cartis, Gould & Toint, ARC Part II, preprint rev. 15 Sep 2009).
For any fixed `j ≥ 0`, with `S_j` the successful and `U_j` the unsuccessful iterations `k ≤ j`
(2.9), assume (2.10) (`σ_{k+1} ≥ γ₃ σ_k` on every very successful `k ≤ j`, `γ₃ ∈ (0, 1]`) and let
`σ̄ > 0` bound `σ_k`. Then `|U_j| ≤ ⌈−(log γ₃ / log γ₁)|S_j| + (1/log γ₁) log(σ̄/σ₀)⌉`.

Correction: the page assumes (2.12) `σ_k ≤ σ̄` only for `k ≤ j`; the proof uses `σ_{j+1} ≤ σ̄`
((2.15) bounds `σ_{j+1}`), and the printed statement fails for `j = 0`, iteration `0` unsuccessful,
`σ̄ = σ₀`. The hypothesis is posed for `k ≤ j + 1`. -/
theorem theorem_2_1 {n : ℕ} (f : EuclideanSpace ℝ (Fin n) → ℝ) (γ₁ γ₂ η₁ η₂ : ℝ)
    (x s : ℕ → EuclideanSpace ℝ (Fin n)) (σ : ℕ → ℝ)
    (B : ℕ → EuclideanSpace ℝ (Fin n) →L[ℝ] EuclideanSpace ℝ (Fin n))
    (hrun : AdaptiveCubic.Cauchy.IsARCRun f γ₁ γ₂ η₁ η₂ x s σ B)
    (j : ℕ) (γ₃ : ℝ) (hγ₃ : 0 < γ₃) (hγ₃' : γ₃ ≤ 1)
    (h210 : ∀ k ≤ j, η₂ < AdaptiveCubic.Cauchy.rho f (B k) (σ k) (x k) (s k) → γ₃ * σ k ≤ σ (k + 1))
    (σbar : ℝ) (hσbar : 0 < σbar) (h212 : ∀ k ≤ j + 1, σ k ≤ σbar) :
    (((Finset.range (j + 1)).filter
        (fun k => AdaptiveCubic.Cauchy.rho f (B k) (σ k) (x k) (s k) < η₁)).card : ℤ) ≤
      ⌈-(Real.log γ₃ / Real.log γ₁) *
          (((Finset.range (j + 1)).filter
            (fun k => η₁ ≤ AdaptiveCubic.Cauchy.rho f (B k) (σ k) (x k) (s k))).card : ℝ) +
        1 / Real.log γ₁ * Real.log (σbar / σ 0)⌉ := by sorry

end AdaptiveCubic.SecondOrder
Source
Cartis, Gould & Toint, Adaptive cubic regularisation methods for unconstrained optimization. Part II, preprint rev. 15 Sep 2009 (UNamur), p. 5, Theorem 2.1, (2.13)
Human review
  • Endorsed by Shuze Chen · Oct 8, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 8, 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