Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

(3.10), sign corrected — one-step bound on Ψ(x^ag_k) for convex Ψ (p. 17)

Proved
NonconvexAG.Stoch.eq_3_10

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

accelerated-gradientconvex-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1stochastic-approximation

Let Ψ:Rn→R\Psi:\mathbb R^n\to\mathbb RΨ:Rn→R be convex and differentiable with LΨL_\PsiLΨ​-Lipschitz gradient. Run the RSAG method (Algorithm 3) from x0x_0x0​ with step sizes α1=1\alpha_1=1α1​=1, αk∈(0,1)\alpha_k\in(0,1)αk​∈(0,1) (k≥2k\ge2k≥2), βk,λk>0\beta_k,\lambda_k>0βk​,λk​>0, any oracle and any noise sequence, with δk=G(xkmd,ξk)−∇Ψ(xkmd)\delta_k=G(x^{md}_k,\xi_k)-\nabla\Psi(x^{md}_k)δk​=G(xkmd​,ξk​)−∇Ψ(xkmd​). Then for every sample point, every k≥1k\ge1k≥1 and every x∈Rnx\in\mathbb R^nx∈Rn,

Ψ(xkag)≤(1−αk)Ψ(xk−1ag)+αkΨ(x)+αk⟨∇Ψ(xkmd),xk−1−x⟩−βk∥∇Ψ(xkmd)∥2−βk⟨∇Ψ(xkmd),δk⟩+LΨβk22∥∇Ψ(xkmd)+δk∥2.\Psi(x^{ag}_k)\le(1-\alpha_k)\Psi(x^{ag}_{k-1})+\alpha_k\Psi(x)+\alpha_k\langle\nabla\Psi(x^{md}_k),x_{k-1}-x\rangle-\beta_k\|\nabla\Psi(x^{md}_k)\|^2-\beta_k\langle\nabla\Psi(x^{md}_k),\delta_k\rangle+\frac{L_\Psi\beta_k^2}{2}\|\nabla\Psi(x^{md}_k)+\delta_k\|^2 .Ψ(xkag​)≤(1−αk​)Ψ(xk−1ag​)+αk​Ψ(x)+αk​⟨∇Ψ(xkmd​),xk−1​−x⟩−βk​∥∇Ψ(xkmd​)∥2−βk​⟨∇Ψ(xkmd​),δk​⟩+2LΨ​βk2​​∥∇Ψ(xkmd​)+δk​∥2.

This is the one-step inequality for the aggregated sequence in the convex case, from which (3.11) is obtained by combining with the stochastic analogue of (2.22) and taking expectations.

Formalization Note The paper prints +βk⟨∇Ψ(xkmd),δk⟩+\beta_k\langle\nabla\Psi(x^{md}_k),\delta_k\rangle+βk​⟨∇Ψ(xkmd​),δk​⟩ in (3.10) and in the line above it. By (3.3), xkag−xkmd=−βk(∇Ψ(xkmd)+δk)x^{ag}_k-x^{md}_k=-\beta_k(\nabla\Psi(x^{md}_k)+\delta_k)xkag​−xkmd​=−βk​(∇Ψ(xkmd​)+δk​), so ⟨∇Ψ(xkmd),xkag−xkmd⟩=−βk∥∇Ψ(xkmd)∥2−βk⟨∇Ψ(xkmd),δk⟩\langle\nabla\Psi(x^{md}_k),x^{ag}_k-x^{md}_k\rangle=-\beta_k\|\nabla\Psi(x^{md}_k)\|^2-\beta_k\langle\nabla\Psi(x^{md}_k),\delta_k\rangle⟨∇Ψ(xkmd​),xkag​−xkmd​⟩=−βk​∥∇Ψ(xkmd​)∥2−βk​⟨∇Ψ(xkmd​),δk​⟩ and the printed sign is wrong pathwise; the Lean states the corrected sign. The error has mean zero and does not affect (3.11). The statement is pathwise.

Preamble
import Mathlib
import Definitions.Def_ConvexOptAlg_SmoothGD_Defs
import Definitions.Def_GhadimiLan_RSG_Model
import Definitions.Def_NonconvexAG_Stoch_RSAG
open MeasureTheory ProbabilityTheory
open scoped InnerProductSpace
Formal statement
namespace NonconvexAG.Stoch

open GhadimiLan.RSG (E)

/-- (3.10), p. 17, with the sign of the inner-product term corrected: for convex `Ψ`, every
sample point `ω`, every `k ≥ 1` and every `x ∈ ℝⁿ`,
`Ψ(x^ag_k) ≤ (1 − αₖ)Ψ(x^ag_{k−1}) + αₖΨ(x) + αₖ⟨∇Ψ(x^md_k), x_{k−1} − x⟩
  − βₖ‖∇Ψ(x^md_k)‖² − βₖ⟨∇Ψ(x^md_k), δₖ⟩ + (L_Ψβₖ²/2)‖∇Ψ(x^md_k) + δₖ‖²`.
(The paper prints `+ βₖ⟨∇Ψ(x^md_k), δₖ⟩`; by (3.3) the sign is `−`.) -/
theorem eq_3_10 {n : ℕ} (Ψ : E n → ℝ) (g : E n → E n) (LΨ : ℝ)
    (hΨ : ConvexOptAlg.SmoothGD.IsBetaSmooth Ψ g LΨ) (hconv : ConvexOn ℝ Set.univ Ψ)
    {Ω Ξ : Type*} (G : E n → Ξ → E n) (ξ : ℕ → Ω → Ξ)
    (α β lam : ℕ → ℝ) (hstep : RSAGStepsizes α β lam) (x0 : E n) :
    ∀ k, 1 ≤ k → ∀ ω : Ω, ∀ x : E n,
      Ψ (xagSeq G α β lam x0 ξ k ω) ≤
        (1 - α k) * Ψ (xagSeq G α β lam x0 ξ (k - 1) ω) + α k * Ψ x
          + α k * ⟪g (xmdSeq G α β lam x0 ξ k ω), xSeq G α β lam x0 ξ (k - 1) ω - x⟫_ℝ
          - β k * ‖g (xmdSeq G α β lam x0 ξ k ω)‖ ^ 2
          - β k * ⟪g (xmdSeq G α β lam x0 ξ k ω), delta G g α β lam x0 ξ k ω⟫_ℝ
          + LΨ * β k ^ 2 / 2 *
              ‖g (xmdSeq G α β lam x0 ξ k ω) + delta G g α β lam x0 ξ k ω‖ ^ 2 := by sorry

end NonconvexAG.Stoch
Source
Ghadimi & Lan, Accelerated Gradient Methods for Nonconvex Nonlinear and Stochastic Programming, arXiv:1310.3787v1, p. 17, (3.10), sign of the ⟨∇Ψ(x^md_k), δ_k⟩ term corrected
Human review
  • Endorsed by Shuze Chen · Oct 9, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 9, 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