(2.10), p. 7 — under A1, E[⟨∇f(x_k), δ_k⟩ | ξ_[k−1]] = 0
ProvedGhadimiLan.RSG.eq_2_10martingalesnonconvex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1sgdstochastic-approximation
Let , let be a Borel oracle, and let be an RSG run from satisfying Assumption A1 with respect to a filtration (noise adapted; conditional unbiasedness ; ). Write . Then for every the random variable is integrable and
The cross term in the summed descent inequality (2.9) therefore has mean zero, which is what lets the expectation of (2.9) be bounded without assuming the noise is independent.
Formalization Note In the paper the conditioning is on the history ; any filtration to which the noise is adapted is allowed here, the history filtration being one of them. No moment assumption on is made: its integrability is part of the claim.
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
/-- Eq. (2.10) (Ghadimi & Lan, arXiv:1309.5549v1, proof of Theorem 2.1, p. 7): under Assumption
A1 along an RSG run, for every `k ≥ 1` the scalar `⟨∇f(x_k), δ_k⟩` (with
`δ_k = G(x_k, ξ_k) − ∇f(x_k)`) is integrable and `E[⟨∇f(x_k), δ_k⟩ | ℱ_{k-1}] = 0` almost surely.
No integrability of `∇f(x_k)` is assumed: it is part of the claim. -/
theorem eq_2_10 {n : ℕ} (f : E n → ℝ) (g : E n → E n) (L : ℝ)
(hf : ConvexOptAlg.SmoothGD.IsBetaSmooth f g 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]
(ℱ : Filtration ℕ ‹MeasurableSpace Ω›) (ξ : ℕ → Ω → Ξ) (σ : ℝ)
(γ : ℕ → ℝ) (x1 : E n) (x : ℕ → Ω → E n) (hx : IsRSGRun G γ x1 ξ x)
(hA1 : AssumptionA1 μ ℱ g G ξ x σ) (k : ℕ) (hk : 1 ≤ k) :
Integrable (fun ω => ⟪g (x k ω), rsgNoise G g ξ x k ω⟫_ℝ) μ ∧
μ[fun ω => ⟪g (x k ω), rsgNoise G g ξ x k ω⟫_ℝ | ℱ (k - 1)] =ᵐ[μ] 0 := by sorry
end GhadimiLan.RSG
Source
Ghadimi & Lan, arXiv:1309.5549v1, proof of Theorem 2.1, Eq. (2.10), p. 7
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.