Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 4.1 — with β_k ≥ 0, Zoutendijk and sufficient descent, gradients bounded away from 0 force Σ‖u_k − u_{k−1}‖² < ∞

Proved
NonlinCG.PropStar.lemma_4_1

by mikedeng1 · Oct 6, 2026 · Mathlib 0df444a (Lean v4.33.1)

conjugate-gradientnonlinear-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1

Let fff satisfy Assumptions 2.1 at the starting point x1x_1x1​ and let (xk,dk,βk,αk)(x_k, d_k, \beta_k, \alpha_k)(xk​,dk​,βk​,αk​) be a run of the conjugate gradient iteration (1.2)–(1.3) with βk≥0\beta_k \ge 0βk​≥0 for all k≥2k \ge 2k≥2. Suppose the line search satisfies the Zoutendijk condition ∑k≥1⟨gk,dk⟩2/∥dk∥2<∞\sum_{k\ge1} \langle g_k, d_k\rangle^2/\|d_k\|^2 < \infty∑k≥1​⟨gk​,dk​⟩2/∥dk​∥2<∞ (2.7) and the sufficient descent condition ⟨gk,dk⟩≤−σ3∥gk∥2\langle g_k, d_k\rangle \le -\sigma_3\|g_k\|^2⟨gk​,dk​⟩≤−σ3​∥gk​∥2 for all k≥1k \ge 1k≥1 (4.1), with 0<σ3≤10 < \sigma_3 \le 10<σ3​≤1. If the gradients are bounded away from zero, i.e. there is γ>0\gamma > 0γ>0 with ∥gk∥≥γ\|g_k\| \ge \gamma∥gk​∥≥γ for all k≥1k \ge 1k≥1 (4.3), then dk≠0d_k \ne 0dk​=0 for all k≥1k \ge 1k≥1 and, with uk:=dk/∥dk∥u_k := d_k/\|d_k\|uk​:=dk​/∥dk​∥,

∑k≥2∥uk−uk−1∥2<∞.(4.5)\sum_{k \ge 2} \|u_k - u_{k-1}\|^2 < \infty. \qquad (4.5)k≥2∑​∥uk​−uk−1​∥2<∞.(4.5)

The lemma says that, as long as the gradients stay away from zero, the normalized search directions change more and more slowly. It is the first of the two ingredients of the proof by contradiction of Theorem 4.3. Its hypotheses are expected to be contradictory on every actual run (that is what Theorem 4.3 shows); the lemma is a step of that argument.

Formalization Note The paper's standing smoothness assumption "fff is smooth" (1.1) is stated as global continuous differentiability, in addition to Assumptions 2.1. The run has positive steplengths and is indexed from 111. The Zoutendijk sum is written as ∑k⟨gk,dk⟩2/∥dk∥2\sum_k \langle g_k,d_k\rangle^2/\|d_k\|^2∑k​⟨gk​,dk​⟩2/∥dk​∥2, which equals ∑kcos⁡2θk∥gk∥2\sum_k \cos^2\theta_k\|g_k\|^2∑k​cos2θk​∥gk​∥2 where the latter is defined. The series (4.5) is stated as the summability of k↦∥uk+2−uk+1∥2k \mapsto \|u_{k+2} - u_{k+1}\|^2k↦∥uk+2​−uk+1​∥2 over k≥0k \ge 0k≥0. As in the paper, neither (4.4) nor Property (*) is assumed.

Preamble
import Mathlib
import Definitions.Def_NonlinCG_PropStar_Setting
Formal statement
namespace NonlinCG.PropStar

/-- Lemma 4.1 (p. 12). Under Assumptions 2.1, for a run of (1.2)–(1.3) with `β_k ≥ 0`, a line
search satisfying the Zoutendijk condition (2.7) and the sufficient descent condition (4.1), if
the gradients are bounded away from zero (4.3), then `d_k ≠ 0` and
`Σ_{k≥2} ‖u_k − u_{k−1}‖² < ∞` (4.5), where `u_k := d_k/‖d_k‖`. -/
theorem lemma_4_1 {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
    [FiniteDimensional ℝ E]
    (f : E → ℝ) (hf : ContDiff ℝ 1 f) (β α : ℕ → ℝ) (x d : ℕ → E) (σ₃ : ℝ)
    (hA : NonlinCG.FRBound.Assumptions21 f (x 1)) (hrun : NonlinCG.FRBound.IsCGRun f β α x d)
    (hβ : ∀ k ≥ 2, 0 ≤ β k)
    (hZ : NonlinCG.FRBound.ZoutendijkCondition f x d)
    (hσ₃ : 0 < σ₃ ∧ σ₃ ≤ 1) (hD : SufficientDescent f σ₃ x d)
    (h43 : ∃ γ > 0, ∀ k ≥ 1, γ ≤ ‖NonlinCG.FRBound.g f x k‖) :
    (∀ k ≥ 1, d k ≠ 0) ∧
      Summable (fun k : ℕ => ‖u d (k + 2) - u d (k + 1)‖ ^ 2) := by sorry

end NonlinCG.PropStar
Source
Gilbert & Nocedal, Global convergence properties of conjugate gradient methods for optimization, INRIA Rapport de Recherche 1268 (June 1990), HAL inria-00075291v1, p. 12, Lemma 4.1 (4.5); (4.1), (4.3) on p. 11
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 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