Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

(2.8), p. 6 — one RSG step: f(x_{k+1}) ≤ f(x_k) − (γ_k − Lγ_k²/2)‖∇f(x_k)‖² − (γ_k − Lγ_k²)⟨∇f(x_k), δ_k⟩ + (L/2)γ_k²‖δ_k‖²

Proved
GhadimiLan.RSG.eq_2_8

by mikedeng1 · Oct 8, 2026 · Mathlib 0df444a (Lean v4.33.1)

nonconvex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1sgdstochastic-approximation

Let f:Rn→Rf:\mathbb R^n\to\mathbb Rf:Rn→R be differentiable with LLL-Lipschitz gradient ∇f\nabla f∇f, i.e. f∈CL1,1(Rn)f\in\mathcal C^{1,1}_L(\mathbb R^n)f∈CL1,1​(Rn). Let x1,x2,…x_1,x_2,\dotsx1​,x2​,… be generated by the RSG recursion xk+1=xk−γkG(xk,ξk)x_{k+1}=x_k-\gamma_k G(x_k,\xi_k)xk+1​=xk​−γk​G(xk​,ξk​) from x1x_1x1​, with arbitrary oracle GGG, noise ξk\xi_kξk​ and real stepsizes γk\gamma_kγk​, and write δk=G(xk,ξk)−∇f(xk)\delta_k=G(x_k,\xi_k)-\nabla f(x_k)δk​=G(xk​,ξk​)−∇f(xk​) for the oracle error. Then for every k≥1k\ge1k≥1 and every outcome,

f(xk+1)≤f(xk)−(γk−L2γk2)∥∇f(xk)∥2−(γk−Lγk2)⟨∇f(xk),δk⟩+L2γk2∥δk∥2.f(x_{k+1})\le f(x_k)-\Big(\gamma_k-\frac L2\gamma_k^2\Big)\|\nabla f(x_k)\|^2-\big(\gamma_k-L\gamma_k^2\big)\langle\nabla f(x_k),\delta_k\rangle+\frac L2\gamma_k^2\|\delta_k\|^2.f(xk+1​)≤f(xk​)−(γk​−2L​γk2​)∥∇f(xk​)∥2−(γk​−Lγk2​)⟨∇f(xk​),δk​⟩+2L​γk2​∥δk​∥2.

This is the one-step descent inequality from which the analysis of the RSG method starts. It is purely deterministic: it holds along every realization of the noise, with no probabilistic assumption.

Formalization Note CL1,1\mathcal C^{1,1}_LCL1,1​ is the published predicate IsBetaSmooth f g L (with g=∇fg=\nabla fg=∇f). The statement is pathwise in ω\omegaω.

Preamble
import Mathlib
import Definitions.Def_ConvexOptAlg_SmoothGD_Defs
import Definitions.Def_GhadimiLan_RSG_Model
open MeasureTheory ProbabilityTheory
open scoped InnerProductSpace
Formal statement
namespace GhadimiLan.RSG

/-- Eq. (2.8) (Ghadimi & Lan, arXiv:1309.5549v1, proof of Theorem 2.1, p. 6): for
`f ∈ C^{1,1}_L(ℝⁿ)` with gradient map `g = ∇f` and an RSG run `x` (recursion (2.2)), for every
`k ≥ 1` and every outcome `ω`, with `δ_k = G(x_k, ξ_k) − ∇f(x_k)`,
`f(x_{k+1}) ≤ f(x_k) − (γ_k − (L/2)γ_k²)‖∇f(x_k)‖² − (γ_k − Lγ_k²)⟨∇f(x_k), δ_k⟩ + (L/2)γ_k²‖δ_k‖²`.
This is a pathwise (deterministic) inequality. -/
theorem eq_2_8 {n : ℕ} (f : E n → ℝ) (g : E n → E n) (L : ℝ)
    (hf : ConvexOptAlg.SmoothGD.IsBetaSmooth f g L)
    {Ω Ξ : Type*} (G : E n → Ξ → E n) (γ : ℕ → ℝ) (x1 : E n) (ξ : ℕ → Ω → Ξ)
    (x : ℕ → Ω → E n) (hx : IsRSGRun G γ x1 ξ x) (k : ℕ) (hk : 1 ≤ k) (ω : Ω) :
    f (x (k + 1) ω) ≤
      f (x k ω) - (γ k - L / 2 * γ k ^ 2) * ‖g (x k ω)‖ ^ 2
        - (γ k - L * γ k ^ 2) * ⟪g (x k ω), rsgNoise G g ξ x k ω⟫_ℝ
        + L / 2 * γ k ^ 2 * ‖rsgNoise G g ξ x k ω‖ ^ 2 := by sorry

end GhadimiLan.RSG
Source
Ghadimi & Lan, arXiv:1309.5549v1, proof of Theorem 2.1, Eq. (2.8), p. 6
Human review
  • Endorsed by Shuze Chen · Oct 8, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 8, 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