Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 3.2, p. 13 — for f ∈ S^{1,1}_{μ,L} the NAG flow (57) has a unique C² solution, ℒ′ ≤ −ℒ − μ/2‖x′‖² (59), and (60)

Proved
NAGFlow.Flow.lemma_3_2

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

accelerated-gradientexponential-decaylyapunov-functionnag-flowp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Let VVV be a real Hilbert space and f∈Sμ,L1,1f\in\mathcal S^{1,1}_{\mu,L}f∈Sμ,L1,1​ with μ≥0\mu\ge0μ≥0: fff is continuously differentiable, μ\muμ-convex,

f(x)−f(y)−⟨∇f(y),x−y⟩≥μ2∥x−y∥2,f(x)-f(y)-\langle\nabla f(y),x-y\rangle\ge\frac{\mu}{2}\|x-y\|^2,f(x)−f(y)−⟨∇f(y),x−y⟩≥2μ​∥x−y∥2,

and ∇f\nabla f∇f is LLL-Lipschitz with 0<L<∞0<L<\infty0<L<∞. Let x∗x^*x∗ be a global minimizer of fff, γ0>0\gamma_0>0γ0​>0, γ(t)=μ+(γ0−μ)e−t\gamma(t)=\mu+(\gamma_0-\mu)e^{-t}γ(t)=μ+(γ0​−μ)e−t the solution of (54) γ′=μ−γ\gamma'=\mu-\gammaγ′=μ−γ, γ(0)=γ0\gamma(0)=\gamma_0γ(0)=γ0​, and x0,v0∈Vx_0,v_0\in Vx0​,v0​∈V.

Then the NAG flow

γx′′+(μ+γ)x′+∇f(x)=0,x(0)=x0, x′(0)=v0−x0,(57)\gamma x''+(\mu+\gamma)x'+\nabla f(x)=0,\qquad x(0)=x_0,\ x'(0)=v_0-x_0, \tag{57}γx′′+(μ+γ)x′+∇f(x)=0,x(0)=x0​, x′(0)=v0​−x0​,(57)

admits a unique solution x∈C2([0,∞);V)x\in C^2([0,\infty);V)x∈C2([0,∞);V). Moreover, with v=x+x′v=x+x'v=x+x′ and the Lyapunov function

L(t)=f(x(t))−f(x∗)+γ(t)2∥v(t)−x∗∥2,(58)\mathcal L(t)=f(x(t))-f(x^*)+\frac{\gamma(t)}{2}\|v(t)-x^*\|^2, \tag{58}L(t)=f(x(t))−f(x∗)+2γ(t)​∥v(t)−x∗∥2,(58)

L\mathcal LL is differentiable on [0,∞)[0,\infty)[0,∞) and

L′(t)≤−L(t)−μ2∥x′(t)∥2,(59)\mathcal L'(t)\le-\mathcal L(t)-\frac{\mu}{2}\|x'(t)\|^2, \tag{59}L′(t)≤−L(t)−2μ​∥x′(t)∥2,(59)

which implies

L(t)+μ2∫0tes−t∥x′(s)∥2 ds≤e−tL(0),t≥0.(60)\mathcal L(t)+\frac{\mu}{2}\int_0^te^{s-t}\|x'(s)\|^2\,ds\le e^{-t}\mathcal L(0),\qquad t\ge0. \tag{60}L(t)+2μ​∫0t​es−t∥x′(s)∥2ds≤e−tL(0),t≥0.(60)

The lemma is the continuous model behind the accelerated methods of the paper: the NAG flow decays the Lyapunov function at the rate e−te^{-t}e−t for every μ≥0\mu\ge0μ≥0, and the discrete schemes are designed so that one step contracts Lk\mathcal L_kLk​ by a factor 1/(1+αk)1/(1+\alpha_k)1/(1+αk​), the discrete analogue of e−te^{-t}e−t.

Formalization Note. VVV is a real Hilbert space; the duality pairing is the inner product (Riesz). Uniqueness is equality on [0,∞)[0,\infty)[0,∞), because a function R→V\mathbb R\to VR→V is unconstrained at negative times. Derivatives are relative to [0,∞)[0,\infty)[0,∞), so one-sided at t=0t=0t=0; "L′(t)≤…\mathcal L'(t)\le\dotsL′(t)≤…" means that L\mathcal LL has a one-sided derivative ddd at ttt with d≤−L(t)−μ2∥x′(t)∥2d\le-\mathcal L(t)-\frac{\mu}{2}\|x'(t)\|^2d≤−L(t)−2μ​∥x′(t)∥2. The integral in (60) is an interval integral of a continuous integrand (the solution is C2C^2C2), so no junk value for non-integrable functions enters. The hypothesis that x∗x^*x∗ is a minimizer is the paper's standing assumption that argmin⁡f\operatorname{argmin}fargminf is nonempty.

Preamble
import Mathlib
import Definitions.Def_NAGFlow_Flow_Setting
Formal statement
namespace NAGFlow.Flow

open Set

/-- Lemma 3.2, p. 13. Let `f ∈ S^{1,1}_{μ,L}` with `μ ≥ 0`, `γ₀ > 0`, `γ(t) = μ + (γ₀ − μ)e^{−t}`
(the solution of (54)), `x₀, v₀ ∈ V`, and `x*` a global minimizer of `f`. Then the NAG flow (57)
`γx″ + (μ + γ)x′ + ∇f(x) = 0`, `x(0) = x₀`, `x′(0) = v₀ − x₀`, admits a solution
`x ∈ C²([0, ∞); V)`, unique on `[0, ∞)`; and, with `v = x + x′` and the Lyapunov function (58)
`ℒ(t) = f(x(t)) − f(x*) + γ(t)/2 ‖v(t) − x*‖²`, `ℒ` is differentiable on `[0, ∞)` (right derivative
at `0`) with
`ℒ′(t) ≤ −ℒ(t) − μ/2 ‖x′(t)‖²` (59), and
`ℒ(t) + μ/2 ∫₀ᵗ e^{s−t} ‖x′(s)‖² ds ≤ e^{−t} ℒ(0)` for all `t ≥ 0` (60). -/
theorem lemma_3_2 {V : Type*} [NormedAddCommGroup V] [InnerProductSpace ℝ V] [CompleteSpace V]
    (f : V → ℝ) (gradf : V → V) (μ L γ₀ : ℝ) (hf : NAGFlow.PredCorr.IsS11 f gradf μ L) (hγ₀ : 0 < γ₀)
    (x₀ v₀ xstar : V) (hxstar : ∀ y, f xstar ≤ f y) :
    ∃ x : ℝ → V, IsNAGSolution2 f gradf μ γ₀ x₀ v₀ x ∧
      (∀ y : ℝ → V, IsNAGSolution2 f gradf μ γ₀ x₀ v₀ y → EqOn y x (Ici 0)) ∧
      (∀ t : ℝ, 0 ≤ t → ∃ dL : ℝ,
        HasDerivWithinAt (lyapFlow f xstar (gammaFn μ γ₀) x (fun s => x s + dI x s)) dL
          (Ici 0) t ∧
        dL ≤ -lyapFlow f xstar (gammaFn μ γ₀) x (fun s => x s + dI x s) t
          - μ / 2 * ‖dI x t‖ ^ 2) ∧
      (∀ t : ℝ, 0 ≤ t →
        lyapFlow f xstar (gammaFn μ γ₀) x (fun s => x s + dI x s) t
          + μ / 2 * ∫ s in (0 : ℝ)..t, Real.exp (s - t) * ‖dI x s‖ ^ 2
        ≤ Real.exp (-t) * lyapFlow f xstar (gammaFn μ γ₀) x (fun s => x s + dI x s) 0) := by sorry

end NAGFlow.Flow
Source
Luo & Chen, arXiv:1909.03145v4, Lemma 3.2 and Eqs. (59)–(60), p. 13
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