Theorem 4.4 — , and if the trajectory ends inside or outside
ProvedTikhonovHDD.Strong.theorem_4_4Let be a real Hilbert space, , , and . Let be convex and twice Fréchet differentiable, with gradient Lipschitz continuous on bounded sets and , and let be nonincreasing, of class , with . Let be a global -solution of
Assume that
that there exist and such that for every , and in addition that
- in case : ;
- in case : there exists such that for large enough.
If is the element of minimum norm of the nonempty convex closed set , then
In addition,
if there exists such that the trajectory stays either in the open ball , or in its complement.
This is the strong-convergence result for which the Tikhonov term is added to the Hessian-damped inertial dynamics: without it, trajectories are only known to converge weakly to some minimizer, while here they approach the distinguished minimizer of minimum norm.
Formalization Note Stated for every global -solution of (5) (existence and uniqueness is Theorem 2.1 of the paper). is written as: for every , for arbitrarily large , which avoids the junk value of Lean's Filter.liminf. The alternative "in the ball, or in its complement" uses one for both branches: either for all , or for all . is the real power. Positivity of is not assumed; it follows from the hypotheses for and .
import Mathlib import Definitions.Def_TikhonovHDD_Strong_Setting open Set Filter Topology
namespace TikhonovHDD.Strong
/-- Theorem 4.4 (p. 21): `liminf_{t→+∞} ‖x(t) − x⋆‖ = 0` (for every `δ > 0`,
`‖x(t) − x⋆‖ < δ` for arbitrarily large `t`), and `‖x(t) − x⋆‖ → 0` if there is `T ≥ t₀`
such that `{x(t) : t ≥ T}` stays in the open ball `B(0, ‖x⋆‖)` or stays in its complement.
Standing hypotheses (those of Theorem 4.4, arXiv:1911.12845v2, p. 21): the General
assumption, `t₀ > 0`, `α ≥ 3`, `β ≥ 0`, `x` a global `C²`-solution of (5),
`∫_{t₀}^{+∞} ε(t)/t dt < +∞`, `β/(ε(t) t^{α/3+1}) ∫_{t₀}^t ε²(s) s^{α/3+1} ds → 0`,
`ε̇(t) ≤ −(aβ/2) ε²(t)` for `t ≥ t₁` (some `a > 1`, `t₁ ≥ t₀`), for `α = 3`: `t² ε(t) → +∞`,
for `α > 3`: `t² ε(t) ≥ (2/3)α((1/3)α − 1 + βc²)` eventually (some `c > 0`), and `x⋆` the
minimum-norm element of `argmin g`. -/
theorem theorem_4_4
{H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
(g : H → ℝ) (α β t₀ : ℝ) (ε ε' : ℝ → ℝ) (u₀ v₀ : H) (x xd xdd : ℝ → H) (xstar : H)
(ht₀ : 0 < t₀) (hα : 3 ≤ α) (hβ : 0 ≤ β)
(hGA : GeneralAssumption g t₀ ε ε')
(hx : IsSolution g α β t₀ ε u₀ v₀ x xd xdd)
(hε_int : MeasureTheory.IntegrableOn (fun t => ε t / t) (Ici t₀))
(hε_lim : Tendsto
(fun t => β / (ε t * t ^ (α / 3 + 1)) * ∫ s in t₀..t, ε s ^ 2 * s ^ (α / 3 + 1))
atTop (𝓝 0))
(hε_deriv : ∃ a > (1 : ℝ), ∃ t₁ ≥ t₀, ∀ t ≥ t₁, ε' t ≤ -(a * β / 2) * ε t ^ 2)
(hα3 : α = 3 → Tendsto (fun t => t ^ 2 * ε t) atTop atTop)
(hα3lt : 3 < α → ∃ c > (0 : ℝ),
∀ᶠ t in atTop, 2 / 3 * α * (1 / 3 * α - 1 + β * c ^ 2) ≤ t ^ 2 * ε t)
(hxstar : IsMinNormMinimizer g xstar) :
(∀ δ > (0 : ℝ), ∃ᶠ t in atTop, ‖x t - xstar‖ < δ) ∧
((∃ T ≥ t₀, (∀ t ≥ T, ‖x t‖ < ‖xstar‖) ∨ (∀ t ≥ T, ‖xstar‖ ≤ ‖x t‖)) →
Tendsto (fun t => ‖x t - xstar‖) atTop (𝓝 0)) := by sorry
end TikhonovHDD.Strong
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.