Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proposition 15 — Riemannian Fletcher–Reeves CG with strong Wolfe steps, c₂ < 1/2, non-expansive transport: liminf ‖Df(x_k)‖ = 0

Proved
RiemOpt.FR.proposition_15

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

conjugate-gradientfletcher-reevesglobal-convergencep2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1riemannian-optimization

Let M\mathcal MM be a smooth manifold modelled on a real Hilbert space with a Riemannian metric, f:M→Rf:\mathcal M\to\mathbb Rf:M→R differentiable with gradient ∇f\nabla f∇f, and RRR a retraction family with transport Tx,Rx(v)Rx=DRx(v)T^{R_x}_{x,R_x(v)}=\mathrm DR_x(v)Tx,Rx​(v)Rx​​=DRx​(v). Consider the Fletcher–Reeves nonlinear conjugate gradient iteration

xk+1=Rxk(αkpk),p0=−∇f(x0),pk+1=−∇f(xk+1)+∥Df(xk+1)∥xk+12∥Df(xk)∥xk2 Txk,xk+1Rxkpk,x_{k+1}=R_{x_k}(\alpha_kp_k),\qquad p_0=-\nabla f(x_0),\qquad p_{k+1}=-\nabla f(x_{k+1})+\frac{\|\mathrm Df(x_{k+1})\|_{x_{k+1}}^2}{\|\mathrm Df(x_k)\|_{x_k}^2}\,T^{R_{x_k}}_{x_k,x_{k+1}}p_k,xk+1​=Rxk​​(αk​pk​),p0​=−∇f(x0​),pk+1​=−∇f(xk+1​)+∥Df(xk​)∥xk​2​∥Df(xk+1​)∥xk+1​2​​Txk​,xk+1​Rxk​​​pk​,

with step lengths αk>0\alpha_k>0αk​>0 satisfying the strong Wolfe conditions (1a), (2) with 0<c1<c2<120<c_1<c_2<\tfrac120<c1​<c2​<21​, and assume that it never stops (Df(xk)≠0\mathrm Df(x_k)\neq0Df(xk​)=0 for all kkk). Suppose that

  1. the transport does not increase the norm of the search direction: ∥Txk,xk+1Rxkpk∥xk+1≤∥pk∥xk\|T^{R_{x_k}}_{x_k,x_{k+1}}p_k\|_{x_{k+1}}\le\|p_k\|_{x_k}∥Txk​,xk+1​Rxk​​​pk​∥xk+1​​≤∥pk​∥xk​​ for all kkk;
  2. fff is bounded below;
  3. the pull-backs f∘Rxkf\circ R_{x_k}f∘Rxk​​ are Lipschitz continuously differentiable on span{pk}\mathrm{span}\{p_k\}span{pk​} with a uniform constant L>0L>0L>0: hk(t)=f(Rxk(tpk))h_k(t)=f(R_{x_k}(tp_k))hk​(t)=f(Rxk​​(tpk​)) is differentiable and ∣hk′(s)−hk′(t)∣≤L∣s−t∣ ∥pk∥xk2|h_k'(s)-h_k'(t)|\le L|s-t|\,\|p_k\|_{x_k}^2∣hk′​(s)−hk′​(t)∣≤L∣s−t∣∥pk​∥xk​2​.

Then

lim inf⁡k→∞∥Df(xk)∥xk=0.\liminf_{k\to\infty}\|\mathrm Df(x_k)\|_{x_k}=0.k→∞liminf​∥Df(xk​)∥xk​​=0.

This is the Riemannian counterpart of the classical global convergence result of Al-Baali for the Fletcher–Reeves method: on a manifold the only extra ingredient is the non-expansiveness of the transport attached to the retraction.

Formalization Note Hypotheses 2 and 3 are those of Zoutendijk's theorem (Theorem 2), which the paper's proof invokes; they are not among the conditions of Lemma 14 and are stated explicitly. Descent of pkp_kpk​ is not assumed (Lemma 14 proves it). The run never stops, since βk+1\beta_{k+1}βk+1​ divides by ∥Df(xk)∥2\|\mathrm Df(x_k)\|^2∥Df(xk​)∥2. The lim inf⁡\liminfliminf is stated explicitly: for every ε>0\varepsilon>0ε>0 and every NNN there is k≥Nk\ge Nk≥N with ∥Df(xk)∥xk<ε\|\mathrm Df(x_k)\|_{x_k}<\varepsilon∥Df(xk​)∥xk​​<ε. One retraction family is used for all kkk; the step equation identifies TRxk(αkpk)MT_{R_{x_k}(\alpha_kp_k)}\mathcal MTRxk​​(αk​pk​)​M with Txk+1MT_{x_{k+1}}\mathcal MTxk+1​​M, whose norm is used in hypothesis 1.

Preamble
import Mathlib
import Definitions.Def_RiemOpt_FR_Setting

open Bundle Manifold
open scoped ContDiff
Formal statement
namespace RiemOpt.FR

variable {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
  {M : Type*} [TopologicalSpace M] [ChartedSpace E M] [IsManifold 𝓘(ℝ, E) ∞ M]
  [RiemannianBundle (fun x : M ↦ TangentSpace 𝓘(ℝ, E) x)]
  [IsContMDiffRiemannianBundle 𝓘(ℝ, E) ∞ E (fun x : M ↦ TangentSpace 𝓘(ℝ, E) x)]

theorem proposition_15
    (f : M → ℝ) (R : (x : M) → TangentSpace 𝓘(ℝ, E) x → M)
    (grad : (x : M) → TangentSpace 𝓘(ℝ, E) x) (c₁ c₂ : ℝ)
    (x : ℕ → M) (p : (k : ℕ) → TangentSpace 𝓘(ℝ, E) (x k)) (α : ℕ → ℝ)
    (hrun : IsFRRun f R grad c₁ c₂ x p α) (hc₂ : c₂ < 1 / 2)
    (hT : ∀ k : ℕ, ‖transportedDir R x p α k‖ ≤ ‖p k‖)
    (hbdd : BddBelow (Set.range f)) (L : ℝ) (hL : 0 < L) (hlip : LineLipschitz f R x p L) :
    ∀ ε > 0, ∀ N : ℕ, ∃ k ≥ N, ‖RiemOpt.BFGS.Df (E := E) f (x k)‖ < ε := by sorry

end RiemOpt.FR
Source
Ring, Wirth, Optimization Methods on Riemannian Manifolds and Their Application to Shape Space, SIAM J. Optim. 22 (2012), p. 611, Proposition 15
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