Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2 — nonconvex SGD on the robust surrogate: (1/T)Σ E‖∇F(θᵗ)‖² − 4L_θz²ε/(γ − L_zz) ≤ 4σ√(L_φΔ_F/T)

Proved
CertDRO.SGD.theorem_2

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

convergence-ratedistributionally-robust-optimizationnonconvex-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1stochastic-gradient

Let Z⊆RmZ\subseteq\mathbb R^mZ⊆Rm be nonempty, convex and closed; let the cost ccc satisfy Assumption A and the loss ℓ:Rd×Rm→R\ell:\mathbb R^d\times\mathbb R^m\to\mathbb Rℓ:Rd×Rm→R satisfy Assumption B, both with the Euclidean norm, and let Θ=Rd\Theta=\mathbb R^dΘ=Rd. Let γ>Lzz\gamma>L_{zz}γ>Lzz​ and

φγ(θ;z0)=sup⁡z∈Z{ℓ(θ;z)−γc(z,z0)},F(θ)=EP0[φγ(θ;Z)],Lφ=Lθθ+LθzLzθγ−Lzz,\varphi_\gamma(\theta;z_0)=\sup_{z\in Z}\{\ell(\theta;z)-\gamma c(z,z_0)\},\qquad F(\theta)=\mathbb E_{P_0}[\varphi_\gamma(\theta;Z)],\qquad L_\varphi=L_{\theta\theta}+\frac{L_{\theta z}L_{z\theta}}{\gamma-L_{zz}},φγ​(θ;z0​)=z∈Zsup​{ℓ(θ;z)−γc(z,z0​)},F(θ)=EP0​​[φγ​(θ;Z)],Lφ​=Lθθ​+γ−Lzz​Lθz​Lzθ​​,

where P0P_0P0​ is a probability distribution with P0(Z)=1P_0(Z)=1P0​(Z)=1 and each φγ(θ;⋅)\varphi_\gamma(\theta;\cdot)φγ​(θ;⋅) is P0P_0P0​-integrable. Suppose:

  1. FFF is bounded below and ΔF>0\Delta_F>0ΔF​>0 satisfies ΔF≥F(θ0)−inf⁡θF(θ)\Delta_F\ge F(\theta^0)-\inf_\theta F(\theta)ΔF​≥F(θ0)−infθ​F(θ);
  2. σ>0\sigma>0σ>0 and EP0∥∇F(θ)−∇θφγ(θ;Z)∥22≤σ2\mathbb E_{P_0}\|\nabla F(\theta)-\nabla_\theta\varphi_\gamma(\theta;Z)\|_2^2\le\sigma^2EP0​​∥∇F(θ)−∇θ​φγ​(θ;Z)∥22​≤σ2 for every θ\thetaθ;
  3. ϵ≥0\epsilon\ge0ϵ≥0 and the inner maximization is solved by a jointly measurable ϵ\epsilonϵ-approximate oracle z^(θ,z0)∈Z\hat z(\theta,z_0)\in Zz^(θ,z0​)∈Z;
  4. T≥1T\ge1T≥1 and T≥LφΔF/σ2T\ge L_\varphi\Delta_F/\sigma^2T≥Lφ​ΔF​/σ2.

Run Algorithm 1 from θ0\theta^0θ0 for TTT steps with i.i.d. samples zt∼P0z^t\sim P_0zt∼P0​ and constant stepsize α=ΔF/(LφTσ2)\alpha=\sqrt{\Delta_F/(L_\varphi T\sigma^2)}α=ΔF​/(Lφ​Tσ2)​, i.e. θt+1=θt−α∇θℓ(θt;z^(θt,zt))\theta^{t+1}=\theta^t-\alpha\nabla_\theta\ell(\theta^t;\hat z(\theta^t,z^t))θt+1=θt−α∇θ​ℓ(θt;z^(θt,zt)). Then

1T∑t=0T−1E[∥∇F(θt)∥22]−4Lθz2γ−Lzzϵ≤4σLφΔFT.\frac1T\sum_{t=0}^{T-1}\mathbb E\big[\|\nabla F(\theta^t)\|_2^2\big]-\frac{4L_{\theta z}^2}{\gamma-L_{zz}}\epsilon\le4\sigma\sqrt{\frac{L_\varphi\Delta_F}{T}} .T1​t=0∑T−1​E[∥∇F(θt)∥22​]−γ−Lzz​4Lθz2​​ϵ≤4σTLφ​ΔF​​​.

This is the paper's computational guarantee: stochastic gradient descent with an approximate adversary reaches an approximate stationary point of the Wasserstein penalty objective at the rate T−1/2T^{-1/2}T−1/2 of smooth nonconvex optimization, up to an additive bias proportional to the accuracy ϵ\epsilonϵ of the inner maximization.

Formalization Note The inequality is stated in [0,∞][0,\infty][0,∞] with the ϵ\epsilonϵ-term moved to the right-hand side, which is equivalent because all terms are finite and nonnegative; the expectations are lower Lebesgue integrals over P0⊗TP_0^{\otimes T}P0⊗T​, so a non-integrable integrand is never read as 000. Hypotheses implicit in the paper and made explicit: γ>Lzz\gamma>L_{zz}γ>Lzz​ (the paper writes γ≥Lzz\gamma\ge L_{zz}γ≥Lzz​ but divides by γ−Lzz\gamma-L_{zz}γ−Lzz​), ZZZ convex and closed, σ>0\sigma>0σ>0 and ΔF>0\Delta_F>0ΔF​>0 (needed for a positive stepsize), T≥1T\ge1T≥1, ϵ≥0\epsilon\ge0ϵ≥0, FFF bounded below, integrability of φγ(θ;⋅)\varphi_\gamma(\theta;\cdot)φγ​(θ;⋅), and measurability of the oracle. The cost is not assumed differentiable.


Retired 2026-10-27: a Lean-vs-paper fidelity audit found this statement weaker than, or narrower than, the cited result. A corrected statement will be resubmitted.

Preamble
import Definitions.Def_CertDRO_SGD_model
Formal statement
namespace CertDRO.SGD

open MeasureTheory

/-- Theorem 2 (Convergence of Nonconvex SGD), p. 6. Let Assumptions A and B hold with the ℓ²-norm,
`Θ = ℝ^d`, `γ > Lzz`, and let `F(θ) = E_{P₀}[φ_γ(θ; Z)]`. Let `Δ_F ≥ F(θ⁰) − inf_θ F(θ)`, assume
`E‖∇F(θ) − ∇_θ φ_γ(θ, Z)‖² ≤ σ²` for all `θ`, and run Algorithm 1 with an `ε`-approximate maximization
oracle and constant stepsize `α = √(Δ_F / (L_φ T σ²))`, `L_φ = Lθθ + Lθz Lzθ / (γ − Lzz)`. If
`T ≥ L_φ Δ_F / σ²`, then
`(1/T) Σ_{t=0}^{T−1} E‖∇F(θ^t)‖² − (4 Lθz² / (γ − Lzz)) ε ≤ 4σ √(L_φ Δ_F / T)`
(stated with the `ε`-term moved to the right-hand side, in `ℝ≥0∞`). -/
theorem theorem_2 {d m : ℕ} (Z : Set (Data m)) (c : Data m → Data m → ℝ)
    (ℓ : Param d → Data m → ℝ) (gθ : Param d → Data m → Param d) (gz : Param d → Data m → Data m)
    (Lθθ Lzz Lθz Lzθ γ : ℝ)
    (hS : Standing Z c ℓ gθ gz Lθθ Lzz Lθz Lzθ γ)
    (P₀ : Measure (Data m)) [IsProbabilityMeasure P₀] (hP₀ : P₀ Zᶜ = 0)
    (hint : ∀ θ, Integrable (fun z => robustSurrogate Z ℓ c γ θ z) P₀)
    (ε : ℝ) (hε : 0 ≤ ε) (zhat : Param d → Data m → Data m)
    (hzhat : IsOracle Z ℓ c γ ε zhat) (hzhat_meas : Measurable (Function.uncurry zhat))
    (σ : ℝ) (hσ : 0 < σ)
    (hvar : ∀ θ, ∫⁻ z, ENNReal.ofReal (‖gradient (robustObjective Z ℓ c γ P₀) θ -
        gradient (fun θ' => robustSurrogate Z ℓ c γ θ' z) θ‖ ^ 2) ∂P₀ ≤ ENNReal.ofReal (σ ^ 2))
    (θ₀ : Param d) (ΔF : ℝ) (hΔF_pos : 0 < ΔF)
    (hbdd : BddBelow (Set.range (robustObjective Z ℓ c γ P₀)))
    (hΔF : robustObjective Z ℓ c γ P₀ θ₀ - ⨅ θ, robustObjective Z ℓ c γ P₀ θ ≤ ΔF)
    (T : ℕ) (hT : 0 < T) (hTL : Lphi Lθθ Lzz Lθz Lzθ γ * ΔF / σ ^ 2 ≤ T) :
    (T : ENNReal)⁻¹ * ∑ t ∈ Finset.range T,
        ∫⁻ ω, ENNReal.ofReal (‖gradient (robustObjective Z ℓ c γ P₀)
            (run gθ zhat (Real.sqrt (ΔF / (Lphi Lθθ Lzz Lθz Lzθ γ * T * σ ^ 2))) θ₀ ω t)‖ ^ 2)
          ∂(Measure.pi fun _ : Fin T => P₀) ≤
      ENNReal.ofReal (4 * Lθz ^ 2 / (γ - Lzz) * ε
        + 4 * σ * Real.sqrt (Lphi Lθθ Lzz Lθz Lzθ γ * ΔF / T)) := by sorry

end CertDRO.SGD
Source
Sinha, Namkoong, Volpi, Duchi, Certifying Some Distributional Robustness with Principled Adversarial Training, arXiv:1710.10571v5, p. 6, Theorem 2 (proof in §B.3, pp. 40–41)
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