Theorem 2.16 — for and , every solution of converges weakly to a minimizer
ProvedInertialAVD.Traj.theorem_2_16convex-optimizationinertial-dynamicsp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1weak-convergence
Let be a real Hilbert space and let be a continuously differentiable convex function such that . Let , , and let be a solution of
Then converges weakly, as , to a point in : there is with for all and
This is the convergence of the trajectories of the continuous-time model of Nesterov's accelerated gradient method, for damping parameters ; the case is open.
Formalization Note The conclusion is weak convergence, tested against every ; in an infinite-dimensional it does not imply norm convergence. Solutions are pairs satisfying (1) on as a first-order system, with .
Preamble
import Mathlib import Definitions.Def_InertialAVD_Traj_Setting
Formal statement
namespace InertialAVD.Traj
theorem theorem_2_16 {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
(Φ : H → ℝ) (hΦconv : ConvexOn ℝ Set.univ Φ) (hΦC1 : ContDiff ℝ 1 Φ)
(hargmin : ∃ z : H, ∀ y : H, Φ z ≤ Φ y)
(α t₀ : ℝ) (hα : 3 < α) (x v : ℝ → H) (hsol : IsSolution Φ α t₀ x v) :
∃ xbar : H, (∀ y : H, Φ xbar ≤ Φ y) ∧ WeakTendstoAtTop x xbar := by sorry
end InertialAVD.Traj
Source
Attouch, Chbani, Peypouquet, Redont, Fast convergence of inertial dynamics and algorithms with asymptotic vanishing damping, Optimization Online preprint 5179 (Oct. 2015), p. 7, Theorem 2.16
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.