Theorem 2, p. 14 — for f ∈ S²_{μ,L} and 0 < s ≤ 1/L, the high-resolution heavy-ball ODE satisfies f(X(t)) − f(x⋆) ≤ 7‖x₀ − x⋆‖² e^{−√μ t/4}/(2s)
ProvedHighResODE.HeavyBallODE.theorem_2Theorem 2 (Convergence of the heavy-ball ODE). Let , that is, is twice continuously differentiable with Lipschitz-continuous Hessian, -smooth, and -strongly convex with , and let be its minimizer. Fix a step size and an initial point , and let be the solution of the high-resolution ODE of the heavy-ball method,
Then for every ,
The high-resolution ODE keeps the terms that the classical low-resolution limit discards. The theorem shows that this continuous-time model of Polyak's heavy-ball method converges linearly at the accelerated rate ; in the paper it is contrasted with the high-resolution ODE of NAG-SC (Theorem 1), which carries an additional gradient-correction term.
Formalization Note. is EuclideanSpace ℝ (Fin n). The solution is any pair of curves with the derivative of and the derivative of on (one-sided at ), satisfying the two initial conditions; since the solution is unique (Proposition 2.1 of the paper), quantifying over all solutions is the paper's statement. The minimizer is a point with for all , which exists for strongly convex .
import Mathlib import Definitions.Def_HighResODE_HeavyBallODE_Setting
namespace HighResODE.HeavyBallODE
theorem theorem_2 {n : ℕ} (f : HighResODE.NAGSC.E n → ℝ) (μ L s : ℝ) (hf : IsS2 f μ L) (xs : HighResODE.NAGSC.E n)
(hmin : ∀ z : HighResODE.NAGSC.E n, f xs ≤ f z) (hs : 0 < s) (hsL : s ≤ 1 / L) (x0 : HighResODE.NAGSC.E n) (X V : ℝ → HighResODE.NAGSC.E n)
(hsol : IsHeavyBallODE f μ s x0 X V) :
∀ t : ℝ, 0 ≤ t →
f (X t) - f xs ≤ 7 * ‖x0 - xs‖ ^ 2 / (2 * s) * Real.exp (-(Real.sqrt μ * t / 4)) := by sorry
end HighResODE.HeavyBallODE
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.