(22), proof of Lemma 4, p. 14 — ρ_r(z) ≥ ‖v‖ with v = ((−c + A⊤y)⁺, b − Ax)
ProvedRestartPD.LPSharp.eq_22linear-programmingnormalized-duality-gapp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , , and consider the LP Lagrangian on , with the Euclidean norm on . For set
Then for every the normalized duality gap satisfies
This is the half of Lemma 4 that controls primal infeasibility and dual infeasibility by the normalized duality gap, uniformly in the radius .
Formalization Note The page's assumption that (19) has a solution and the radius are not needed here and are omitted. When the claim reads .
Preamble
import Mathlib import Definitions.Def_RestartPD_LPSharp_PrimalDual import Definitions.Def_RestartPD_LPSharp_LP open scoped InnerProductSpace Matrix
Formal statement
namespace RestartPD.LPSharp
theorem eq_22 {n m : ℕ} (A : Matrix (Fin m) (Fin n) ℝ) (b : EuclideanSpace ℝ (Fin m))
(c : EuclideanSpace ℝ (Fin n)) (z : PDSpace n m) (hz : z ∈ (lpX : Set (EuclideanSpace ℝ (Fin n))) ×ˢ Set.univ)
(r : ℝ) (hr : 0 < r) :
((eucl n m (vvec A b c z) : ℝ) : EReal) ≤ rho (lpL A b c) lpX Set.univ r z := by sorry
end RestartPD.LPSharp
Source
Applegate, Hinder, Lu & Lubin, Faster First-Order Primal-Dual Methods for Linear Programming using Restarts and Sharpness, arXiv:2105.12715v4, p. 14, (22) in the proof of Lemma 4
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.