Proof of Theorem 3.18, p. 291 — f(y_t) − f(x*) ≤ ((α + β)/2)‖x₁ − x*‖²(1 − 1/√κ)^{t−1}
OpenConvexOptAlg.NesterovStrong.thm_3_18_rateaccelerated-gradientconvergence-rateconvex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
Let be -strongly convex and -smooth with , , and let be a minimizer of on . Let be a run of Nesterov's accelerated gradient descent. Then for every ,
This is the rate obtained by combining (3.18) at with (3.19); the exponential form of Theorem 3.18 follows from .
Formalization Note The existence of the minimizer is the book's standing assumption (p. 242). The exponent is a natural-number subtraction, guarded by .
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, p. 291, the display combining (3.18) and (3.19) (first and
last members): for a `β`-smooth, `α`-strongly convex `f` on `ℝⁿ` with minimizer `x*` and a run
`(x, y)` of Nesterov's accelerated gradient descent, for every `t ≥ 1`,
`f(y_t) − f(x*) ≤ ((α + β)/2)‖x₁ − x*‖² (1 − 1/√κ)^{t−1}`. -/
theorem thm_3_18_rate {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)
(xstar : EuclideanSpace ℝ (Fin n)) (hmin : ∀ z, f xstar ≤ f z)
(x y : ℕ → EuclideanSpace ℝ (Fin n)) (hrun : IsNesterovSCRun g α β x y)
(t : ℕ) (ht : 1 ≤ t) :
f (y t) - f xstar ≤
(α + β) / 2 * ‖x 1 - xstar‖ ^ 2 * (1 - 1 / Real.sqrt (kappa α β)) ^ (t - 1) := by sorry
end ConvexOptAlg.NesterovStrong
Source
Bubeck, arXiv:1405.4980v2, proof of Theorem 3.18, p. 291, display after (3.19) (first and last members)