Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proof of Theorem 2.1, p. 7 — E‖∇f(x_R)‖² = Σ(2γ_k − Lγ_k²)E‖∇f(x_k)‖² / Σ(2γ_k − Lγ_k²)

Proved
GhadimiLan.RSG.weighting_identity

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

nonconvex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1randomizationsgdstochastic-approximation

Let f∈CL1,1(Rn)f\in\mathcal C^{1,1}_L(\mathbb R^n)f∈CL1,1​(Rn) with L>0L>0L>0, let GGG be a Borel oracle, let the noise variables ξk\xi_kξk​ be measurable, and let x1,x2,…x_1,x_2,\dotsx1​,x2​,… be the RSG run from x1x_1x1​ with stepsizes 0<γk<2/L0<\gamma_k<2/L0<γk​<2/L for k=1,…,Nk=1,\dots,Nk=1,…,N, N≥1N\ge1N≥1. Let the output index RRR take values in {1,…,N}\{1,\dots,N\}{1,…,N} with Pr⁡{R=k}=PR(k)=(2γk−Lγk2)/∑j=1N(2γj−Lγj2)\Pr\{R=k\}=P_R(k)=(2\gamma_k-L\gamma_k^2)/\sum_{j=1}^N(2\gamma_j-L\gamma_j^2)Pr{R=k}=PR​(k)=(2γk​−Lγk2​)/∑j=1N​(2γj​−Lγj2​), and let RRR be independent of the noise sequence (ξk)k≥1(\xi_k)_{k\ge1}(ξk​)k≥1​. If ∥∇f(xk)∥2\|\nabla f(x_k)\|^2∥∇f(xk​)∥2 is integrable for every k=1,…,Nk=1,\dots,Nk=1,…,N, then ∥∇f(xR)∥2\|\nabla f(x_R)\|^2∥∇f(xR​)∥2 is integrable and

E∥∇f(xR)∥2=∑k=1N(2γk−Lγk2) E∥∇f(xk)∥2∑k=1N(2γk−Lγk2).\mathbb E\|\nabla f(x_R)\|^2=\frac{\sum_{k=1}^N\big(2\gamma_k-L\gamma_k^2\big)\,\mathbb E\|\nabla f(x_k)\|^2}{\sum_{k=1}^N\big(2\gamma_k-L\gamma_k^2\big)}.E∥∇f(xR​)∥2=∑k=1N​(2γk​−Lγk2​)∑k=1N​(2γk​−Lγk2​)E∥∇f(xk​)∥2​.

This is the step where the randomized output enters: the expected squared gradient at the random iterate is the PRP_RPR​-weighted average of the expected squared gradients along the trajectory.

Formalization Note The independence of RRR from the noise is implicit in the paper (Step 0 draws RRR before any oracle call, and the expectation is "with respect to RRR and ξ[N]\xi_{[N]}ξ[N]​"); it is stated explicitly, and without it the identity fails. The left side is the integral of ∥∇f(xR(ω)(ω))∥2\|\nabla f(x_{R(\omega)}(\omega))\|^2∥∇f(xR(ω)​(ω))∥2, not the weighted sum.

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

/-- The weighting identity in the proof of Theorem 2.1 (Ghadimi & Lan, arXiv:1309.5549v1, p. 7,
display after "Dividing both sides"): if the output index `R` has the mass function (2.3) on
`{1, …, N}` and is independent of the noise sequence `(ξ_k)`, and each `‖∇f(x_k)‖²`
(`k = 1, …, N`) is integrable, then `‖∇f(x_R)‖²` is integrable and
`E‖∇f(x_R)‖² = Σ_{k=1}^N (2γ_k − Lγ_k²) E‖∇f(x_k)‖² / Σ_{k=1}^N (2γ_k − Lγ_k²)`. -/
theorem weighting_identity {n : ℕ} (f : E n → ℝ) (g : E n → E n) (L : ℝ)
    (hf : ConvexOptAlg.SmoothGD.IsBetaSmooth f g L) (hL : 0 < L)
    {Ξ : Type*} [MeasurableSpace Ξ] (G : E n → Ξ → E n) (hG : Measurable (Function.uncurry G))
    {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
    (ξ : ℕ → Ω → Ξ) (hξ : ∀ k, Measurable (ξ k))
    (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)
    (R : Ω → ℕ) (hR : IsRandomOutputIndex μ R L γ N)
    (hRind : IndepFun R (fun ω k => ξ k ω) μ)
    (hint : ∀ k ∈ Finset.Icc 1 N, Integrable (fun ω => ‖g (x k ω)‖ ^ 2) μ) :
    Integrable (fun ω => ‖g (x (R ω) ω)‖ ^ 2) μ ∧
    ∫ ω, ‖g (x (R ω) ω)‖ ^ 2 ∂μ =
      (∑ k ∈ Finset.Icc 1 N, (2 * γ k - L * γ k ^ 2) * ∫ ω, ‖g (x k ω)‖ ^ 2 ∂μ) /
        ∑ k ∈ Finset.Icc 1 N, (2 * γ k - L * γ k ^ 2) := by sorry

end GhadimiLan.RSG
Source
Ghadimi & Lan, arXiv:1309.5549v1, proof of Theorem 2.1, p. 7, display after "Dividing both sides"
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