Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2.1 a), p. 6 — for 0 < γ_k < 2/L and P_R of (2.3), (1/L)E‖∇f(x_R)‖² ≤ (D_f² + σ²Σγ_k²)/Σ(2γ_k − Lγ_k²)

Proved
GhadimiLan.RSG.theorem_2_1_a

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

complexitynonconvex-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 (L>0L>0L>0), bounded below, with f∗=inf⁡x∈Rnf(x)f^*=\inf_{x\in\mathbb R^n}f(x)f∗=infx∈Rn​f(x). Let GGG be a Borel stochastic first-order oracle and run the randomized stochastic gradient (RSG) method

xk+1=xk−γk G(xk,ξk),k≥1,x_{k+1}=x_k-\gamma_k\,G(x_k,\xi_k),\qquad k\ge1,xk+1​=xk​−γk​G(xk​,ξk​),k≥1,

from x1x_1x1​, under Assumption A1 with noise level σ≥0\sigma\ge0σ≥0: conditionally on the past, G(xk,ξk)G(x_k,\xi_k)G(xk​,ξk​) is an unbiased estimate of ∇f(xk)\nabla f(x_k)∇f(xk​), and E∥G(xk,ξk)−∇f(xk)∥2≤σ2\mathbb E\|G(x_k,\xi_k)-\nabla f(x_k)\|^2\le\sigma^2E∥G(xk​,ξk​)−∇f(xk​)∥2≤σ2. Let N≥1N\ge1N≥1, let the stepsizes satisfy 0<γk<2/L0<\gamma_k<2/L0<γk​<2/L for k=1,…,Nk=1,\dots,Nk=1,…,N, and let the output index RRR, independent of the noise, have the mass function

PR(k)=Pr⁡{R=k}=2γk−Lγk2∑j=1N(2γj−Lγj2),k=1,…,N.P_R(k)=\Pr\{R=k\}=\frac{2\gamma_k-L\gamma_k^2}{\sum_{j=1}^N(2\gamma_j-L\gamma_j^2)},\qquad k=1,\dots,N.PR​(k)=Pr{R=k}=∑j=1N​(2γj​−Lγj2​)2γk​−Lγk2​​,k=1,…,N.

Then ∥∇f(xR)∥2\|\nabla f(x_R)\|^2∥∇f(xR​)∥2 is integrable and

1L E[∥∇f(xR)∥2]≤Df2+σ2∑k=1Nγk2∑k=1N(2γk−Lγk2),Df=[2(f(x1)−f∗)L]1/2,\frac1L\,\mathbb E\big[\|\nabla f(x_R)\|^2\big]\le\frac{D_f^2+\sigma^2\sum_{k=1}^N\gamma_k^2}{\sum_{k=1}^N(2\gamma_k-L\gamma_k^2)},\qquad D_f=\Big[\frac{2(f(x_1)-f^*)}{L}\Big]^{1/2},L1​E[∥∇f(xR​)∥2]≤∑k=1N​(2γk​−Lγk2​)Df2​+σ2∑k=1N​γk2​​,Df​=[L2(f(x1​)−f∗)​]1/2,

where the expectation is taken with respect to both RRR and the noise ξ1,…,ξN\xi_1,\dots,\xi_Nξ1​,…,ξN​.

This is the basic complexity bound for stochastic gradient descent on smooth nonconvex objectives: with the stepsize policy of Corollary 2.2 it gives E∥∇f(xR)∥2=O(1/N+σ/N)\mathbb E\|\nabla f(x_R)\|^2=O(1/N+\sigma/\sqrt N)E∥∇f(xR​)∥2=O(1/N+σ/N​), and it is the result on which the two-phase RSG method and the zeroth-order RSGF method of the same paper are built.

Formalization Note f∈CL1,1f\in\mathcal C^{1,1}_Lf∈CL1,1​ is the published predicate IsBetaSmooth f g L with g=∇fg=\nabla fg=∇f; f∗f^*f∗ is the greatest lower bound of the range of fff. Three hypotheses implicit in the paper are explicit: L>0L>0L>0 (the paper divides by LLL), γk>0\gamma_k>0γk​>0 (otherwise PRP_RPR​ is not a probability mass function), and the independence of RRR from the whole noise sequence (Step 0 draws RRR first). Assumption A1 is stated in the conditional form the proof uses, with respect to any filtration to which the noise is adapted; the noise need not be independent. Integrability of ∥∇f(xR)∥2\|\nabla f(x_R)\|^2∥∇f(xR​)∥2 is part of the conclusion, since Lean's integral of a non-integrable function is 000.

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

/-- Theorem 2.1 a), Eq. (2.4) (Ghadimi & Lan, arXiv:1309.5549v1, p. 6). Let `f ∈ C^{1,1}_L(ℝⁿ)`
(`L > 0`) have gradient map `g = ∇f` and be bounded below with `f* = inf f`. Run the RSG method
(2.2) from `x1` with a Borel oracle `G` satisfying Assumption A1 (noise level `σ`), stepsizes
`0 < γ_k < 2/L` for `k = 1, …, N` (`N ≥ 1`), and an output index `R` with mass function (2.3)
drawn independently of the noise. Then `‖∇f(x_R)‖²` is integrable and
`(1/L) E‖∇f(x_R)‖² ≤ (D_f² + σ² Σ_{k=1}^N γ_k²) / Σ_{k=1}^N (2γ_k − Lγ_k²)`, where
`D_f = [2(f(x_1) − f*)/L]^{1/2}` and the expectation is over both `R` and the noise. -/
theorem theorem_2_1_a {n : ℕ} (f : E n → ℝ) (g : E n → E n) (L : ℝ)
    (hf : ConvexOptAlg.SmoothGD.IsBetaSmooth f g L) (hL : 0 < L)
    (fstar : ℝ) (hfstar : IsGLB (Set.range f) fstar)
    {Ξ : Type*} [MeasurableSpace Ξ] (G : E n → Ξ → E n) (hG : Measurable (Function.uncurry G))
    {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
    (ℱ : Filtration ℕ ‹MeasurableSpace Ω›) (ξ : ℕ → Ω → Ξ) (σ : ℝ)
    (N : ℕ) (hN : 1 ≤ N) (γ : ℕ → ℝ) (hγ : ∀ k ∈ Finset.Icc 1 N, 0 < γ k ∧ γ k < 2 / L)
    (x1 : E n) (x : ℕ → Ω → E n) (hx : IsRSGRun G γ x1 ξ x)
    (hA1 : AssumptionA1 μ ℱ g G ξ x σ)
    (R : Ω → ℕ) (hR : IsRandomOutputIndex μ R L γ N)
    (hRind : IndepFun R (fun ω k => ξ k ω) μ) :
    Integrable (fun ω => ‖g (x (R ω) ω)‖ ^ 2) μ ∧
    1 / L * ∫ ω, ‖g (x (R ω) ω)‖ ^ 2 ∂μ ≤
      (Df f x1 fstar L ^ 2 + σ ^ 2 * ∑ k ∈ Finset.Icc 1 N, γ k ^ 2) /
        ∑ k ∈ Finset.Icc 1 N, (2 * γ k - L * γ k ^ 2) := by sorry

end GhadimiLan.RSG
Source
Ghadimi & Lan, arXiv:1309.5549v1, Theorem 2.1 a), Eqs. (2.3)–(2.5), 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