Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.3, p. 13 — DRS residuals vanish at rate o(1/√k); u^k, v^k share stationary cluster points with φ = lim φ^DR_γ(s^k); level-bounded φ gives bounded iterates

Proved
NonconvexDRS.DRS.theorem_4_3

by mikedeng1 · Oct 9, 2026 · Mathlib 0df444a (Lean v4.33.1)

douglas-rachfordnonconvex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1stationaritysubsequential-convergence

Suppose Assumption I holds: φ1:Rp→R\varphi_1:\mathbb R^p\to\mathbb Rφ1​:Rp→R is LLL-smooth (L>0L>0L>0) and σ\sigmaσ-hypoconvex with σ∈[−L,L]\sigma\in[-L,L]σ∈[−L,L], φ2:Rp→R‾\varphi_2:\mathbb R^p\to\overline{\mathbb R}φ2​:Rp→R is proper and lower semicontinuous, and φ=φ1+φ2\varphi=\varphi_1+\varphi_2φ=φ1​+φ2​ has a minimizer. Let the stepsize γ>0\gamma>0γ>0 and relaxation λ\lambdaλ lie in the ranges of Theorem 4.1: either λ∈(0,2)\lambda\in(0,2)λ∈(0,2) and γ<min⁡{2−λ2[σ]−,1L}\gamma<\min\{\frac{2-\lambda}{2[\sigma]_-},\frac1L\}γ<min{2[σ]−​2−λ​,L1​}, or σ>0\sigma>0σ>0, 2≤λ<41+1−p2\le\lambda<\frac4{1+\sqrt{1-p}}2≤λ<1+1−p​4​ and pλ−δ4σ<γ<pλ+δ4σ\frac{p\lambda-\delta}{4\sigma}<\gamma<\frac{p\lambda+\delta}{4\sigma}4σpλ−δ​<γ<4σpλ+δ​, with p=σ/Lp=\sigma/Lp=σ/L and δ=(pλ)2−8p(λ−2)\delta=\sqrt{(p\lambda)^2-8p(\lambda-2)}δ=(pλ)2−8p(λ−2)​. Let (sk,uk,vk)k∈N(s^k,u^k,v^k)_{k\in\mathbb N}(sk,uk,vk)k∈N​ be any sequence generated by DRS,

uk∈prox⁡γφ1(sk),vk∈prox⁡γφ2(2uk−sk),sk+1=sk+λ(vk−uk),u^k\in\operatorname{prox}_{\gamma\varphi_1}(s^k),\qquad v^k\in\operatorname{prox}_{\gamma\varphi_2}(2u^k-s^k),\qquad s^{k+1}=s^k+\lambda(v^k-u^k),uk∈proxγφ1​​(sk),vk∈proxγφ2​​(2uk−sk),sk+1=sk+λ(vk−uk),

from an arbitrary s0∈Rps^0\in\mathbb R^ps0∈Rp. Then:

  1. the residual vanishes with rate min⁡i≤k∥ui−vi∥=o(1/k)\min_{i\le k}\|u^i-v^i\|=o(1/\sqrt k)mini≤k​∥ui−vi∥=o(1/k​);
  2. (uk)(u^k)(uk) and (vk)(v^k)(vk) have the same cluster points; every such cluster point u′u'u′ is stationary for φ\varphiφ, i.e. 0∈∂^φ(u′)0\in\hat\partial\varphi(u')0∈∂^φ(u′); and the limit φ⋆=lim⁡kφγDR(sk)\varphi_\star=\lim_k\varphi^{\mathrm{DR}}_\gamma(s^k)φ⋆​=limk​φγDR​(sk) exists, is finite, and equals φ(u′)\varphi(u')φ(u′) for every cluster point u′u'u′;
  3. if φ\varphiφ is level bounded, then (sk)(s^k)(sk), (uk)(u^k)(uk) and (vk)(v^k)(vk) are bounded.

This is the paper's convergence result for nonconvex DRS: it holds for every relaxation λ∈(0,2)\lambda\in(0,2)λ∈(0,2) and, when φ1\varphi_1φ1​ is strongly convex, beyond λ=2\lambda=2λ=2, under stepsize bounds that Theorem 4.7 shows to be tight.

Formalization Note Rate (i) is k⋅min⁡i≤k∥ui−vi∥→0\sqrt k\cdot\min_{i\le k}\|u^i-v^i\|\to0k​⋅mini≤k​∥ui−vi∥→0. "Stationary" is 0∈∂^φ(u′)0\in\hat\partial\varphi(u')0∈∂^φ(u′) in the regular (Fréchet) subdifferential of φ\varphiφ as an R‾\overline{\mathbb R}R-valued function, which is what the proof establishes and is stronger than limiting stationarity. Cluster points are MapClusterPt points. The page's closing sentence of (ii), "In fact, for each kkk one has dist⁡(0,∂^φ(vk))≤1−γσ2γ∥uk−vk∥\operatorname{dist}(0,\hat\partial\varphi(v^k))\le\frac{1-\gamma\sigma}{2\gamma}\|u^k-v^k\|dist(0,∂^φ(vk))≤2γ1−γσ​∥uk−vk∥", is false as printed (for φ1≡0\varphi_1\equiv0φ1​≡0, φ2=a2x2\varphi_2=\frac a2x^2φ2​=2a​x2 the distance is ∥uk−vk∥/γ\|u^k-v^k\|/\gamma∥uk−vk∥/γ) and is not part of this statement; the separate item theorem_4_3_ii_dist states it with the constant 1−γσγ\frac{1-\gamma\sigma}{\gamma}γ1−γσ​. Reciprocal bounds are multiplied out. The added hypothesis L>0L>0L>0 is needed for p=σ/Lp=\sigma/Lp=σ/L; L=0L=0L=0 makes φ1\varphi_1φ1​ affine. No a-priori decrease, summability or boundedness is assumed.

Preamble
import Mathlib
import Definitions.Def_NonconvexSplitting_Shared_LimitingSubdiff
import Definitions.Def_NonconvexDRS_DRS_Setting

open NonconvexSplitting.Shared
open Filter Topology
open scoped InnerProductSpace
Formal statement
namespace NonconvexDRS.DRS

/-- Theorem 4.3 (Subsequential convergence), p. 13, without the "In fact" sentence of (ii)
(see `theorem_4_3_ii_dist`). -/
theorem theorem_4_3 {p : ℕ} (φ₁ : EuclideanSpace ℝ (Fin p) → ℝ) (φ₂ : EuclideanSpace ℝ (Fin p) → EReal)
    (L σ : ℝ) (hL : 0 < L) (hσ : -L ≤ σ ∧ σ ≤ L)
    (hsmooth : IsLSmooth φ₁ L) (hhypo : IsHypoconvex φ₁ σ)
    (hprop : IsProper φ₂) (hlsc : LowerSemicontinuous φ₂)
    (hsol : ∃ xs, ∀ x, (φ₁ xs : EReal) + φ₂ xs ≤ (φ₁ x : EReal) + φ₂ x)
    (γ lam : ℝ) (hγ : 0 < γ) (hrange : StepRange L σ γ lam)
    (s u v : ℕ → EuclideanSpace ℝ (Fin p)) (hrun : IsDRSRun φ₁ φ₂ γ lam s u v) :
    Tendsto (fun k : ℕ => Real.sqrt (k : ℝ) *
        (Finset.range (k + 1)).inf' Finset.nonempty_range_add_one (fun i => ‖u i - v i‖))
      atTop (𝓝 0) ∧
    ({y | MapClusterPt y atTop u} = {y | MapClusterPt y atTop v} ∧
      (∀ y, MapClusterPt y atTop u → IsRegularSubgrad (phi φ₁ φ₂) y 0) ∧
      ∃ φstar : ℝ, Tendsto (fun k => dre φ₁ φ₂ γ (s k)) atTop (𝓝 (φstar : EReal)) ∧
        ∀ y, MapClusterPt y atTop u → phi φ₁ φ₂ y = φstar) ∧
    (LevelBounded (phi φ₁ φ₂) →
      Bornology.IsBounded (Set.range s) ∧ Bornology.IsBounded (Set.range u) ∧
        Bornology.IsBounded (Set.range v)) := by sorry

end NonconvexDRS.DRS
Source
Themelis & Patrinos, Douglas-Rachford splitting and ADMM for nonconvex optimization: tight convergence results, arXiv:1709.05747v4, p. 13, Theorem 4.3 (without the "In fact" sentence of (ii))
Human review
  • Endorsed by Shuze Chen · Oct 10, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 10, 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