Term 2, proof of Theorem 9, p. 26 — (√µ/(2s) + µ/√s)X² + (β/4)Y² − √µXY ≥ 0 when √s ≤ β
ProvedHessianDamping.IPAHDSC.term2_nonnegconvex-optimizationhessian-dampinginertial-algorithmp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1proximal-algorithmstrong-convexity
Let , and , and put . Then:
- if , then for all real numbers ,
- if and only if ;
- if , then .
In the proof of Theorem 9 this is applied with and to show that the second group of remainder terms is nonnegative under the step-size condition .
Formalization Note is assumed so that is the page's quantity; under the theorem's hypotheses it follows from .
Preamble
import Mathlib import Definitions.Def_HessianDamping_IPAHDSC_Setting
Formal statement
namespace HessianDamping.IPAHDSC
/-- Term 2, proof of Theorem 9, p. 26: if `√μ/(2s) + μ/√s ≥ μ/β` then
`(√μ/(2s) + μ/√s) X² + (β/4) Y² - √μ X Y ≥ 0` for all reals `X`, `Y`; the condition is
equivalent to `√s ≤ (β/2)(1 + √(1 + 2/(β√μ)))`; and `√s ≤ β` implies that condition. -/
theorem term2_nonneg (μ β s : ℝ) (hμ : 0 < μ) (hs : 0 < s) (hβ : 0 < β) :
(Real.sqrt μ / (2 * s) + μ / Real.sqrt s ≥ μ / β →
∀ X Y : ℝ,
0 ≤ (Real.sqrt μ / (2 * s) + μ / Real.sqrt s) * X ^ 2 + β / 4 * Y ^ 2 -
Real.sqrt μ * X * Y) ∧
(Real.sqrt μ / (2 * s) + μ / Real.sqrt s ≥ μ / β ↔
Real.sqrt s ≤ β / 2 * (1 + Real.sqrt (1 + 2 / (β * Real.sqrt μ)))) ∧
(Real.sqrt s ≤ β → Real.sqrt μ / (2 * s) + μ / Real.sqrt s ≥ μ / β) := by sorry
end HessianDamping.IPAHDSC
Source
Attouch, Chbani, Fadili, Riahi, First-order optimization algorithms via inertial systems with Hessian driven damping, arXiv:1907.10536v2, p. 26, proof of Theorem 9, Term 2
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.