Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2.2 — at most ⌈κpϵ−p⌉\lceil\kappa_p\epsilon^{-p}\rceil⌈κp​ϵ−p⌉ successful iterations decrease the model by αϵp\alpha\epsilon^pαϵp

Proved
AdaptiveCubic.Cauchy.theorem_2_2

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 with parameters γ2≥γ1>1\gamma_2\ge\gamma_1>1γ2​≥γ1​>1, 1>η2≥η1>01>\eta_2\ge\eta_1>01>η2​≥η1​>0, σ0>0\sigma_0>0σ0​>0, and suppose f(xk)≥flowf(x_k)\ge f_{\rm low}f(xk​)≥flow​ for all kkk. Let ϵ>0\epsilon>0ϵ>0, α>0\alpha>0α>0 and p>0p>0p>0. Fix a horizon jjj such that the model decreases on every iteration up to it,

mk(sk)<f(xk)for all k≤j,(2.6)m_k(s_k)<f(x_k)\quad\text{for all }k\le j, \qquad (2.6)mk​(sk​)<f(xk​)for all k≤j,(2.6)

and let So\mathcal S_oSo​ be any set of successful iterations k≤jk\le jk≤j with

f(xk)−mk(sk)≥αϵpfor all k∈So.(2.18)f(x_k)-m_k(s_k)\ge\alpha\epsilon^p\quad\text{for all }k\in\mathcal S_o. \qquad (2.18)f(xk​)−mk​(sk​)≥αϵpfor all k∈So​.(2.18)

Then

∣So∣≤⌈κp ϵ−p⌉,κp=f(x0)−flowη1α.|\mathcal S_o|\le\left\lceil\kappa_p\,\epsilon^{-p}\right\rceil,\qquad \kappa_p=\frac{f(x_0)-f_{\rm low}}{\eta_1\alpha}.∣So​∣≤⌈κp​ϵ−p⌉,κp​=η1​αf(x0​)−flow​​.

Each such iteration lowers fff by at least η1αϵp\eta_1\alpha\epsilon^pη1​αϵp, and fff never increases along the run, so the total available decrease f(x0)−flowf(x_0)-f_{\rm low}f(x0​)−flow​ caps their number. This is the generic counting step behind every complexity bound of the paper.

Formalization Note The paper assumes (2.6) on every iteration (p. 4) and sums over all of So\mathcal S_oSo​ "with say jm≤∞j_m\le\inftyjm​≤∞ as the largest index". Here the statement is posed for each finite horizon jjj, with (2.6) assumed up to jjj; the infinite case follows by letting jjj grow. The optimality measure Fk≥0F_k\ge0Fk​≥0 and the set SFϵ={k∈S:Fk>ϵ}\mathcal S^\epsilon_F=\{k\in\mathcal S: F_k>\epsilon\}SFϵ​={k∈S:Fk​>ϵ} of (2.16)–(2.17) only restrict which So\mathcal S_oSo​ are allowed; since FkF_kFk​ is arbitrary, allowing every set of successful iterations is the same theorem.

Preamble
import Mathlib
import Definitions.Def_AdaptiveCubic_Cauchy_IsARCRun

open scoped RealInnerProductSpace
Formal statement
namespace AdaptiveCubic.Cauchy

/-- Theorem 2.2, (2.19), p. 6 (Cartis, Gould & Toint, ARC Part II, preprint rev. 15 Sep 2009),
finite-horizon form. Let `f(x_k) ≥ f_low` for all `k`, `ϵ > 0`, `α > 0`, `p > 0`. Fix a horizon `j`
on which (2.6) `m_k(s_k) < f(x_k)` holds for every `k ≤ j`, and let `S_o` be any set of successful
iterations `k ≤ j` with (2.18) `f(x_k) − m_k(s_k) ≥ α ϵ^p` for `k ∈ S_o`. Then
`|S_o| ≤ ⌈κ_p ϵ^{−p}⌉` with `κ_p = (f(x_0) − f_low)/(η₁ α)`.

The paper assumes (2.6) on every iteration (p. 4, last paragraph); here it is assumed up to the
horizon, which is the "largest index `j_m`" of the page's proof. The optimality measure `F_k` and
`S^ϵ_F` of (2.16)–(2.17) only constrain which `S_o` are allowed; since `F_k ≥ 0` is arbitrary, the
statement for all successful `S_o` is the same theorem. -/
theorem theorem_2_2 {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 : IsARCRun f γ₁ γ₂ η₁ η₂ x s σ B)
    (flow : ℝ) (hlow : ∀ k, flow ≤ f (x k))
    (ε : ℝ) (hε : 0 < ε) (α p : ℝ) (hα : 0 < α) (hp : 0 < p) (j : ℕ)
    (h26 : ∀ k ≤ j, model f (B k) (σ k) (x k) (s k) < f (x k))
    (So : Finset ℕ)
    (hSo : So ⊆ (Finset.range (j + 1)).filter (fun k => η₁ ≤ rho f (B k) (σ k) (x k) (s k)))
    (h218 : ∀ k ∈ So, α * ε ^ p ≤ f (x k) - model f (B k) (σ k) (x k) (s k)) :
    (So.card : ℤ) ≤ ⌈(f (x 0) - flow) / (η₁ * α) * ε ^ (-p)⌉ := by sorry

end AdaptiveCubic.Cauchy
Source
Cartis, Gould & Toint, Adaptive cubic regularisation methods for unconstrained optimization. Part II, preprint rev. 15 Sep 2009 (UNamur), p. 6, Theorem 2.2, (2.16)–(2.19)
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