One-step descent of the RSAG method (proof of Theorem 3 a), p. 16)
ProvedNonconvexAG.Stoch.one_stepaccelerated-gradientnonconvex-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1stochastic-approximation
Let be differentiable with -Lipschitz gradient . Run the RSAG method (Algorithm 3) from with step sizes , (), , any oracle and any noise sequence, and write . Then for every sample point and every ,
This is the stochastic counterpart of (2.15): the descent of the long-step sequence in one iteration, with the oracle error appearing through a quadratic term and a linear term that vanishes in conditional expectation.
Formalization Note The inequality is pathwise: it holds for every realization of the noise, so no probabilistic assumption is made. It is the outer ends of the paper's chain (the second display of the proof).
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)
/-- One-step descent of the RSAG method (proof of Theorem 3 a), p. 16, second display, outer
ends), pathwise for every sample point `ω` and every `k ≥ 1`, with
`δₖ = G(x^md_k, ξₖ) − ∇Ψ(x^md_k)`:
`Ψ(xₖ) ≤ Ψ(xₖ₋₁) − λₖ(1 − L_Ψλₖ)‖∇Ψ(x^md_k)‖² + (L_Ψ(1 − αₖ)²/2)‖x^ag_{k−1} − xₖ₋₁‖²
+ (L_Ψλₖ²/2)‖δₖ‖² − λₖ⟨∇Ψ(xₖ₋₁) − L_Ψλₖ∇Ψ(x^md_k), δₖ⟩`. -/
theorem one_step {n : ℕ} (Ψ : E n → ℝ) (g : E n → E n) (LΨ : ℝ)
(hΨ : ConvexOptAlg.SmoothGD.IsBetaSmooth Ψ g LΨ)
{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,
Ψ (xSeq G α β lam x0 ξ k ω) ≤
Ψ (xSeq G α β lam x0 ξ (k - 1) ω)
- lam k * (1 - LΨ * lam k) * ‖g (xmdSeq G α β lam x0 ξ k ω)‖ ^ 2
+ LΨ * (1 - α k) ^ 2 / 2 *
‖xagSeq G α β lam x0 ξ (k - 1) ω - xSeq G α β lam x0 ξ (k - 1) ω‖ ^ 2
+ LΨ * lam k ^ 2 / 2 * ‖delta G g α β lam x0 ξ k ω‖ ^ 2
- lam k * ⟪g (xSeq G α β lam x0 ξ (k - 1) ω)
- (LΨ * lam k) • g (xmdSeq G α β lam x0 ξ k ω),
delta G g α β lam x0 ξ k ω⟫_ℝ := by sorry
end NonconvexAG.Stoch
Source
Ghadimi & Lan, Accelerated Gradient Methods for Nonconvex Nonlinear and Stochastic Programming, arXiv:1310.3787v1, p. 16, §3.1, proof of Theorem 3 a), second display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.