Theorem 1, p. 13 — for f ∈ S²_{μ,L} and 0 < s ≤ 1/L, the NAG-SC high-resolution ODE satisfies f(X(t)) − f(x⋆) ≤ (2‖x₀ − x⋆‖²/s) e^{−√μ t/4}
ProvedHighResODE.NAGSCODE.theorem_1Let , that is, is -smooth, -strongly convex with , of class with a Lipschitz-continuous Hessian, and let be its minimizer. Let be a step size and . Let be the solution of the high-resolution ODE of NAG-SC,
Then for every ,
The function value along the high-resolution ODE of Nesterov's accelerated gradient method for strongly convex functions converges linearly, at the accelerated rate rather than ; with the bound reads . It is the continuous-time counterpart of the discrete rate for NAG-SC in Theorem 3 of the same paper.
Formalization Note. A solution is any pair of curves with the one-sided derivative of on , having the derivative the ODE prescribes, and the initial conditions above; existence and uniqueness (Proposition 2.1) are not assumed, so "the solution" is read as "every solution". The minimizer is a point with for all . is the Fréchet derivative of at applied to . The Hessian's Lipschitz constant is not fixed.
import Mathlib import Definitions.Def_HighResODE_NAGSCODE_Setting open scoped InnerProductSpace
namespace HighResODE.NAGSCODE
theorem theorem_1 {n : ℕ} (f : EuclideanSpace ℝ (Fin n) → ℝ) (μ L s : ℝ) (hf : IsS2 f μ L)
(xs : EuclideanSpace ℝ (Fin n)) (hmin : ∀ z, f xs ≤ f z) (hs : 0 < s) (hsL : s ≤ 1 / L)
(x0 : EuclideanSpace ℝ (Fin n)) (X V : ℝ → EuclideanSpace ℝ (Fin n))
(hsol : IsNAGSCODE f μ s x0 X V) :
∀ t, 0 ≤ t →
f (X t) - f xs ≤ 2 * ‖x0 - xs‖ ^ 2 / s * Real.exp (-(Real.sqrt μ * t / 4)) := by sorry
end HighResODE.NAGSCODE
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.