Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

(6.4) — the doubling loop of RACDM terminates with L^i≤2Li\hat L_i\le 2L_iL^i​≤2Li​

Proved
NesterovRCD.Adaptive.eq_6_4

by mikedeng1 · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

backtrackingconvex-optimizationcoordinate-descentp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1

Let f:Rn→Rf:\mathbb R^n\to\mathbb Rf:Rn→R be differentiable with coordinate-wise Lipschitz partial derivatives (2.2): for every iii, xxx and u∈Ru\in\mathbb Ru∈R,

∣∇if(x+uei)−∇if(x)∣≤Li∣u∣.|\nabla_i f(x+ue_i)-\nabla_i f(x)|\le L_i|u| .∣∇i​f(x+uei​)−∇i​f(x)∣≤Li​∣u∣.

Fix a point xxx, a coordinate iii and an estimate ℓ\ellℓ with 0<ℓ≤Li0<\ell\le L_i0<ℓ≤Li​. Then the doubling loop of RACDM (6.1) started at xxx, iii, ℓ\ellℓ terminates: some d≥0d\ge0d≥0 satisfies ∇if(x)⋅∇if(x−(2dℓ)−1∇if(x)ei)≥0\nabla_i f(x)\cdot\nabla_i f(x-(2^d\ell)^{-1}\nabla_if(x)e_i)\ge0∇i​f(x)⋅∇i​f(x−(2dℓ)−1∇i​f(x)ei​)≥0. Moreover the accepted estimate satisfies

2d(x,i,ℓ) ℓ ≤ 2Li,2^{d(x,i,\ell)}\,\ell\ \le\ 2L_i ,2d(x,i,ℓ)ℓ ≤ 2Li​,

where d(x,i,ℓ)d(x,i,\ell)d(x,i,ℓ) is the number of doublings (the least such ddd).

This is the bound (6.4) of the proof of Theorem 7: once the estimate exceeds LiL_iLi​ the sign test is passed, so the loop never overshoots LiL_iLi​ by more than a factor 222. It gives both that the method is well defined and the factor that enters the decrease bound 3/(8Li)3/(8L_i)3/(8Li​).

Formalization Note Rn\mathbb R^nRn is EuclideanSpace ℝ (Fin n) (scalar coordinates: the paper sets N=nN=nN=n in §6.1), coordinates are indexed 0,…,n−10,\dots,n-10,…,n−1 for the paper's 1,…,n1,\dots,n1,…,n, and the gradient is an explicit map ggg with g(x)=∇f(x)g(x)=\nabla f(x)g(x)=∇f(x), as in the published definition ConvexOptAlg_CoordDescent_Defs that this mission reuses (its IsCoordSmooth f g L is (2.2) with one-dimensional blocks and ∣⋅∣|\cdot|∣⋅∣ as block norm; it includes differentiability). The paper states (6.4) for "the internal cycle" with the entry assumption L^ik≤Lik\hat L_{i_k}\le L_{i_k}L^ik​​≤Lik​​; we state it for an arbitrary entry point xxx, coordinate iii and entry estimate 0<ℓ≤Li0<\ell\le L_i0<ℓ≤Li​ (positivity is implicit in the paper: the method divides by L^i\hat L_iL^i​). Every intermediate estimate 2d′ℓ2^{d'}\ell2d′ℓ, d′≤dd'\le dd′≤d, is at most the accepted one, so the bound on the accepted one is the bound "during the internal cycle". Convexity is not needed and not assumed.

Preamble
import Mathlib
import Definitions.Def_ConvexOptAlg_CoordDescent_Defs
import Definitions.Def_NesterovRCD_Adaptive_RACDM
Formal statement
namespace NesterovRCD.Adaptive

open ConvexOptAlg.CoordDescent

/-- (6.4), p. 18: if the inner cycle of RACDM (6.1) is entered on coordinate `i` at `x` with an
estimate `0 < ℓ ≤ L_i`, the cycle terminates, and the accepted estimate `2^d ℓ` satisfies
`L̂_i ≤ 2 L_i`. -/
theorem eq_6_4 {n : ℕ} (f : Vec n → ℝ) (g : Vec n → Vec n) (L : Fin n → ℝ)
    (hL : IsCoordSmooth f g L) (x : Vec n) (i : Fin n) (ℓ : ℝ) (hℓ : 0 < ℓ) (hℓL : ℓ ≤ L i) :
    (∃ d : ℕ, ¬ (g x i * g (trial g x i (2 ^ d * ℓ)) i < 0)) ∧
      2 ^ doublings g x i ℓ * ℓ ≤ 2 * L i := by sorry

end NesterovRCD.Adaptive
Source
Nesterov, Efficiency of coordinate descent methods on huge-scale optimization problems, CORE Discussion Paper 2010/2, p. 18, proof of Theorem 7, (6.4)
Human review
  • Endorsed by Shuze Chen · Oct 7, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 7, 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