Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2(a), corrected — ∥dI(xk)∥≤max⁡{1,sup⁡jαj} C rk\|d_I(x^k)\| \le \max\{1,\sup_j\alpha^j\}\,C\,r^k∥dI​(xk)∥≤max{1,supj​αj}Crk for k∈Tk \in \mathcal Tk∈T under the restricted Gauss–Seidel rule

Proved
TsengCGD.Linear.theorem2_a

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

coordinate-descentgauss-seidelnonsmooth-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1

Fix nnn and constants L≥0L\ge0L≥0, 0<λ‾≤λˉ0<\underline\lambda\le\bar\lambda0<λ​≤λˉ. There is a constant C>0C>0C>0, depending only on nnn, LLL, λ‾\underline\lambdaλ​, λˉ\bar\lambdaλˉ, with the following property. Consider any instance of problem (1) (with its standing assumptions) in which ∇f\nabla f∇f is LLL-Lipschitz on dom⁡P\operatorname{dom}PdomP (22), and any run {xk},{Hk},{dk},{αk}\{x^k\},\{H^k\},\{d^k\},\{\alpha^k\}{xk},{Hk},{dk},{αk} of the CGD method with iterates in dom⁡P\operatorname{dom}PdomP, satisfying Assumption 1 with λ‾,λˉ\underline\lambda,\bar\lambdaλ​,λˉ, with {Jk}\{\mathcal J^k\}{Jk} chosen by the restricted Gauss–Seidel rule (12) along T⊆{0,1,… }\mathcal T\subseteq\{0,1,\dots\}T⊆{0,1,…}, and with PPP block-separable with respect to every Jk\mathcal J^kJk. If the stepsizes are bounded, αj≤αˉ\alpha^j\le\bar\alphaαj≤αˉ for all jjj, then for every k∈Tk\in\mathcal Tk∈T

∥dI(xk)∥ ≤ max⁡{1,αˉ}  C  rk,rk=∑ℓ=kτ(k)−1∥dℓ∥.\|d_I(x^k)\|\ \le\ \max\{1,\bar\alpha\}\;C\;r^k,\qquad r^k=\sum_{\ell=k}^{\tau(k)-1}\|d^\ell\|.∥dI​(xk)∥ ≤ max{1,αˉ}Crk,rk=ℓ=k∑τ(k)−1​∥dℓ∥.

The paper states ∥dI(xk)∥≤sup⁡jαj C rk\|d_I(x^k)\|\le\sup_j\alpha^j\,C\,r^k∥dI​(xk)∥≤supj​αjCrk without the block-separability hypothesis; we state the corrected form above. As printed, the bound fails twice. (1) With n=1n=1n=1, P=0P=0P=0, f(x)=x2/2f(x)=x^2/2f(x)=x2/2, Hk=1H^k=1Hk=1, T=N\mathcal T=\mathbb NT=N and constant stepsizes αk=ε\alpha^k=\varepsilonαk=ε, it reads ∣xk∣≤εC∣xk∣|x^k|\le\varepsilon C|x^k|∣xk∣≤εC∣xk∣, false for ε<1/C\varepsilon<1/Cε<1/C; the paper's proof gives the factor max⁡{1,sup⁡jαj}\max\{1,\sup_j\alpha^j\}max{1,supj​αj}. (2) With n=2n=2n=2, dom⁡P={x1+x2≤0}\operatorname{dom}P=\{x_1+x_2\le0\}domP={x1​+x2​≤0}, P=0P=0P=0 there, f(x)=−x1f(x)=-x_1f(x)=−x1​, Hk=IH^k=IHk=I, alternating blocks {1},{2}\{1\},\{2\}{1},{2} and x0=0x^0=0x0=0, every coordinate step is blocked, so r0=0r^0=0r0=0 while dI(x0)=(12,−12)d_I(x^0)=(\tfrac12,-\tfrac12)dI​(x0)=(21​,−21​); the proof splits dI(xk)d_I(x^k)dI​(xk) over the blocks, which needs block-separability, as Theorem 2(b) assumes. Where the theorem is used, in (b), sup⁡jαj≤1\sup_j\alpha^j\le1supj​αj≤1 and PPP is block-separable, so (b) is unaffected.

The bound controls the optimality residual at the start of each Gauss–Seidel cycle by the length of the steps taken during the cycle; combined with the error bound (Assumption 2(a)) it drives the linear rate of Theorem 2(b).

Formalization Note The constant is quantified before the problem data, the run and T\mathcal TT, so it is uniform (it may depend on nnn, which is fixed). sup⁡jαj\sup_j\alpha^jsupj​αj is replaced by an explicit upper bound αˉ\bar\alphaαˉ. The hypothesis xk∈dom⁡Px^k\in\operatorname{dom}Pxk∈domP for all kkk is the domain of the method (the subproblem (6) is posed at x∈dom⁡Px\in\operatorname{dom}Px∈domP, and the proof uses xℓ,xk∈dom⁡Px^\ell,x^k\in\operatorname{dom}Pxℓ,xk∈domP); with αk≤1\alpha^k\le1αk≤1 it follows from convexity. T={t0<t1<⋯ }\mathcal T=\{t_0<t_1<\cdots\}T={t0​<t1​<⋯} is a strictly increasing map with t0=0t_0=0t0​=0, k=tik=t_ik=ti​ and τ(k)=ti+1\tau(k)=t_{i+1}τ(k)=ti+1​. dId_IdI​ is the full direction (13) with H=IH=IH=I.

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

theorem theorem2_a {n : ℕ} (L lam lamBar : ℝ) :
    ∃ C > 0, ∀ (f : TsengCGD.Global.Vec n → ℝ) (D : Set (TsengCGD.Global.Vec n)) (P : TsengCGD.Global.Vec n → ℝ) (c : ℝ)
      (J : ℕ → Finset (Fin n)) (H : ℕ → Matrix (Fin n) (Fin n) ℝ)
      (x d : ℕ → TsengCGD.Global.Vec n) (α : ℕ → ℝ) (t : ℕ → ℕ),
      TsengCGD.Global.Standing f D P c → GradLipOn f D L → TsengCGD.Global.IsCGDRun f D P c J H x d α →
      (∀ k, x k ∈ D) → TsengCGD.Global.Assumption1 H lam lamBar → RestrictedGaussSeidel J t →
      (∀ k, TsengCGD.Global.BlockSeparable D P (J k)) →
      ∀ abar : ℝ, (∀ j, α j ≤ abar) →
      ∀ i, ‖TsengCGD.Global.dI f D P c (x (t i))‖ ≤ max 1 abar * C * r t d i := by sorry

end TsengCGD.Linear
Source
Tseng and Yun, A coordinate gradient descent method for nonsmooth separable minimization, Math. Program. Ser. B 117 (2009), pp. 405–406, Theorem 2(a) (corrected; see the statement)
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