§3.2, p. 16 — increment of v_k
ProvedHessianDamping.IGAHD.vK_succ_subconvex-optimizationhessian-dampinginertial-algorithmp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
For an IGAHD run and , the auxiliary vector of (15) satisfies
The identity aligns the energy's vector term with the gradient step in the algorithm and is used in the one-step energy estimate.
Preamble
import Mathlib import Definitions.Def_HessianDamping_IGAHD_Setting
Formal statement
namespace HessianDamping.IGAHD
/-- The v-increment identity in the proof of Theorem 6, p. 16. -/
theorem vK_succ_sub {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
(f : H → ℝ) (hfconv : ConvexOn ℝ Set.univ f) (hfdiff : Differentiable ℝ f)
(L : ℝ) (hL : 0 < L)
(hLip : ∀ u v : H, ‖gradient f u - gradient f v‖ ≤ L * ‖u - v‖)
(xstar : H) (hxstar : ∀ z : H, f xstar ≤ f z)
(α β s : ℝ) (hα : 3 ≤ α) (hs : 0 < s) (hsL : s ≤ 1 / L)
(hβ0 : 0 ≤ β) (hβs : β < 2 * Real.sqrt s)
(x y : ℕ → H) (hrun : IsIGAHDRun f α β s x y)
(k : ℕ) (hk : 1 ≤ k) :
vK f α β s x xstar (k + 1) - vK f α β s x xstar k =
-(s * tK α (k + 1)) • gradient f (y k) := by sorry
end HessianDamping.IGAHD
Source
Attouch, Chbani, Fadili, Riahi, First-order optimization algorithms via inertial systems with Hessian driven damping, arXiv:1907.10536v2, p. 16, proof of Theorem 6, v-increment
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.