Eq. (3.18), p. 291 — Φ_{s+1}(x) ≤ f(x) + (1 − 1/√κ)^s (Φ₁(x) − f(x))
OpenConvexOptAlg.NesterovStrong.eq_3_18accelerated-gradientconvex-optimizationestimate-sequencep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
Let be -strongly convex and -smooth with , and put . Let be a run of Nesterov's accelerated gradient descent and let be the functions defined from the points by (3.17). Then for every and every ,
The functions thus approach from below at the geometric rate ; this is one half of the estimate-sequence argument for Theorem 3.18.
Formalization Note The inequality is stated for every ; at it is an equality. The positivity is implied by the other hypotheses whenever and is stated for definiteness of .
Preamble
import Mathlib import Definitions.Def_OnlineConvexOpt_ConvexBasics_StronglyConvexOn import Definitions.Def_ConvexOptAlg_NesterovStrong_Defs open scoped InnerProductSpace
Formal statement
namespace ConvexOptAlg.NesterovStrong
/-- Bubeck, proof of Theorem 3.18, Eq. (3.18), p. 291: for a `β`-smooth, `α`-strongly convex
`f` on `ℝⁿ` (gradient map `g`, `κ = β/α`) and a run `(x, y)` of Nesterov's accelerated gradient
descent, the functions `Φ_s` of (3.17) satisfy
`Φ_{s+1}(z) ≤ f(z) + (1 − 1/√κ)^s (Φ₁(z) − f(z))` for every `z` and every `s`. -/
theorem eq_3_18 {n : ℕ} (f : EuclideanSpace ℝ (Fin n) → ℝ)
(g : EuclideanSpace ℝ (Fin n) → EuclideanSpace ℝ (Fin n)) (α β : ℝ)
(hα : 0 < α) (hβ : 0 < β)
(hsc : OnlineConvexOpt.ConvexBasics.StronglyConvexOn Set.univ f g α)
(hsm : IsBetaSmooth f g β)
(x y : ℕ → EuclideanSpace ℝ (Fin n)) (hrun : IsNesterovSCRun g α β x y)
(s : ℕ) (z : EuclideanSpace ℝ (Fin n)) :
Phi f g α β x (s + 1) z ≤
f z + (1 - 1 / Real.sqrt (kappa α β)) ^ s * (Phi f g α β x 1 z - f z) := by sorry
end ConvexOptAlg.NesterovStrong
Source
Bubeck, arXiv:1405.4980v2, proof of Theorem 3.18, Eq. (3.18), p. 291