Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 4 — Hölder-type Lipschitz dependence of the coordinate subproblem's minimizer on the linear term

Proved
TsengCGD.Linear.lemma4

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

coordinate-descentnonsmooth-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1sensitivity-analysis

Let c>0c>0c>0 and let P:ℜn→(−∞,∞]P:\Re^n\to(-\infty,\infty]P:ℜn→(−∞,∞] be proper, convex and lower semicontinuous. Let h:ℜn→ℜh:\Re^n\to\Reh:ℜn→ℜ be continuously differentiable and satisfy, for some ρ>0\rho>0ρ>0 and p>1p>1p>1,

(∇h(u)−∇h(v))T(u−v) ≥ ρ ∥u−v∥ppfor all u,v∈ℜn,(\nabla h(u)-\nabla h(v))^T(u-v)\ \ge\ \rho\,\|u-v\|_p^p\qquad\text{for all }u,v\in\Re^n,(∇h(u)−∇h(v))T(u−v) ≥ ρ∥u−v∥pp​for all u,v∈ℜn,

and let qqq be the conjugate exponent, 1p+1q=1\tfrac1p+\tfrac1q=1p1​+q1​=1. Fix x∈dom⁡Px\in\operatorname{dom}Px∈domP, a nonempty index set J⊆N\mathcal J\subseteq\mathcal NJ⊆N, and two vectors gˉ,g~∈ℜn\bar g,\tilde g\in\Re^ngˉ​,g~​∈ℜn. Let dˉ\bar ddˉ minimize gˉTd+h(d)+cP(x+d)\bar g^Td+h(d)+cP(x+d)gˉ​Td+h(d)+cP(x+d) and d~\tilde dd~ minimize g~Td+h(d)+cP(x+d)\tilde g^Td+h(d)+cP(x+d)g~​Td+h(d)+cP(x+d), both over the vectors ddd with dj=0d_j=0dj​=0 for j∉Jj\notin\mathcal Jj∈/J. Then

∥dˉ−d~∥p ≤ ρ−q/p ∥gˉJ−g~J∥qq/p.\|\bar d-\tilde d\|_p\ \le\ \rho^{-q/p}\,\|\bar g_{\mathcal J}-\tilde g_{\mathcal J}\|_q^{q/p}.∥dˉ−d~∥p​ ≤ ρ−q/p∥gˉ​J​−g~​J​∥qq/p​.

The lemma says that the coordinate subproblem's solution moves Hölder-continuously with its linear coefficients; with h(u)=∥u∥2/2h(u)=\|u\|^2/2h(u)=∥u∥2/2, p=q=2p=q=2p=q=2, ρ=1\rho=1ρ=1 it is the 1-Lipschitz dependence of dI(x;J)d_I(x;\mathcal J)dI​(x;J) on ∇f(x)J\nabla f(x)_{\mathcal J}∇f(x)J​ that the proof of Theorem 2(a) uses.

Formalization Note PPP is the pair (D,P)(D,P)(D,P) with D=dom⁡PD=\operatorname{dom}PD=domP. A minimizer is a vector supported on J\mathcal JJ with x+d∈Dx+d\in Dx+d∈D and objective value no larger than that of any other such vector (the others have value +∞+\infty+∞); the bound is asserted for every pair of minimizers, which is the paper's statement since the minimizers are unique. Coordinates are 0-based (Fin n), and ∥⋅∥p\|\cdot\|_p∥⋅∥p​, ∥⋅∥q\|\cdot\|_q∥⋅∥q​ are the explicit sums of rpows, of nonnegative bases.

Preamble
import Mathlib
import Definitions.Def_TsengCGD_Linear_Basic
open Filter Topology Finset Matrix
open scoped RealInnerProductSpace
Formal statement
namespace TsengCGD.Linear

theorem lemma4 {n : ℕ} (D : Set (TsengCGD.Global.Vec n)) (P : TsengCGD.Global.Vec n → ℝ) (c : ℝ)
    (hP : ProxNewton.Inexact.IsProperClosedConvex D P) (hc : 0 < c)
    (h : TsengCGD.Global.Vec n → ℝ) (hh : ContDiff ℝ 1 h) (ρ p q : ℝ) (hρ : 0 < ρ) (hp : 1 < p)
    (hq : 1 / p + 1 / q = 1)
    (hmono : ∀ u v : TsengCGD.Global.Vec n, ρ * pNorm p (u - v) ^ p ≤ ⟪gradient h u - gradient h v, u - v⟫)
    (x : TsengCGD.Global.Vec n) (hx : x ∈ D) (J : Finset (Fin n)) (hJ : J.Nonempty) (gb gt db dt : TsengCGD.Global.Vec n)
    (hdb : TsengCGD.Global.SupportedOn J db ∧ x + db ∈ D ∧ ∀ d', TsengCGD.Global.SupportedOn J d' → x + d' ∈ D →
      ⟪gb, db⟫ + h db + c * P (x + db) ≤ ⟪gb, d'⟫ + h d' + c * P (x + d'))
    (hdt : TsengCGD.Global.SupportedOn J dt ∧ x + dt ∈ D ∧ ∀ d', TsengCGD.Global.SupportedOn J d' → x + d' ∈ D →
      ⟪gt, dt⟫ + h dt + c * P (x + dt) ≤ ⟪gt, d'⟫ + h d' + c * P (x + d')) :
    pNorm p (db - dt) ≤ ρ ^ (-q / p) * pNormOn q J (gb - gt) ^ (q / p) := by sorry

end TsengCGD.Linear
Source
Tseng and Yun, A coordinate gradient descent method for nonsmooth separable minimization, Math. Program. Ser. B 117 (2009), p. 396, Lemma 4
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