Eq. (3.19) — suboptimality bounded by the squared gradient norm
ProvedStochQuasiNewton.SQN.suboptimality_le_gradient_sqconvex-optimizationp2o-batch-p200bp2o-gran-per-chapterp2o-plan-paperp2o-v1strong-convexity
Let be and suppose for every , with (the lower bound in (3.4)). Then for all ,
and if minimizes , then for every
This inequality converts the gradient term of the descent inequality (3.18) into a suboptimality term, yielding the recursion (3.22).
Formalization Note Only the lower Hessian bound is assumed, as in the paper's derivation. A published Boyd–Vandenberghe statement of the same inequality exists on the platform in the older Mathlib environment c5ea003; it cannot be imported from this mission's environment, so the inequality is restated here.
Preamble
import Mathlib open scoped RealInnerProductSpace
Formal statement
namespace StochQuasiNewton.SQN
/-- Eq. (3.19) and the line after it: if `F` is `C²` with `∇²F(w) ≻ λ I` for every `w` (the
lower bound of (3.4)), then `F(v) ≥ F(w) − ‖∇F(w)‖² / (2λ)` for all `v, w`, and for a minimizer
`w*` of `F`, `2λ [F(w) − F(w*)] ≤ ‖∇F(w)‖²` for every `w`. -/
theorem suboptimality_le_gradient_sq {n : ℕ} (F : EuclideanSpace ℝ (Fin n) → ℝ)
(hF : ContDiff ℝ 2 F) (lam : ℝ) (hlam : 0 < lam)
(hHess : ∀ w v : EuclideanSpace ℝ (Fin n), v ≠ 0 →
lam * ‖v‖ ^ 2 < ⟪fderiv ℝ (gradient F) w v, v⟫)
(wstar : EuclideanSpace ℝ (Fin n)) (hwstar : IsMinOn F Set.univ wstar) :
(∀ w v : EuclideanSpace ℝ (Fin n), F w - 1 / (2 * lam) * ‖gradient F w‖ ^ 2 ≤ F v) ∧
∀ w : EuclideanSpace ℝ (Fin n), 2 * lam * (F w - F wstar) ≤ ‖gradient F w‖ ^ 2 := by sorry
end StochQuasiNewton.SQN
Source
Byrd, Hansen, Nocedal, Singer, A Stochastic Quasi-Newton Method for Large-Scale Optimization, SIAM J. Optim. 26(2) (2016), p. 1017, Eq. (3.19) and the displayed inequality after it
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.