Lemma 3.1, p. 263 — ‖Π_X(y) − x‖² + ‖y − Π_X(y)‖² ≤ ‖y − x‖² for x ∈ X
ProvedConvexOptAlg.Subgradient.lemma_3_1convex-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1projection
Let be convex, let and , and let be a point of nearest to in the Euclidean norm. Then
Projecting onto a convex set therefore never moves a point farther from any point of the set; this is the fact that lets the projected subgradient method ignore the projection step in its distance bookkeeping.
Formalization Note This is the second claim of Lemma 3.1. Its first claim, , is the published theorem ConvexOptimization.projection_iff_obtuse_angle (forward direction), which is a separate item of this mission. The projection is given as a relation (IsMetricProjection), so no closedness of is needed: the statement is about any nearest point.
Preamble
import Mathlib import Definitions.Def_OnlineConvexOpt_FirstOrder_Protocol import Definitions.Def_ConvexOptAlg_Subgradient_Defs
Formal statement
namespace ConvexOptAlg.Subgradient
/-- Bubeck, Lemma 3.1, p. 263, second claim: for a convex set `X`, a point `x ∈ X` and any
`y`, the projection `p = Π_X(y)` satisfies `‖p - x‖² + ‖y - p‖² ≤ ‖y - x‖²`. The projection
is given as a relation (`IsMetricProjection X y p`: `p ∈ X` is a nearest point of `X` to `y`).
The first claim `(Π_X(y) - x)ᵀ(Π_X(y) - y) ≤ 0` is the published
`ConvexOptimization.projection_iff_obtuse_angle`. -/
theorem lemma_3_1 {n : ℕ} (X : Set (EuclideanSpace ℝ (Fin n))) (hXconv : Convex ℝ X)
(x y p : EuclideanSpace ℝ (Fin n)) (hx : x ∈ X)
(hp : OnlineConvexOpt.FirstOrder.IsMetricProjection X y p) :
‖p - x‖ ^ 2 + ‖y - p‖ ^ 2 ≤ ‖y - x‖ ^ 2 := by sorry
end ConvexOptAlg.Subgradient
Source
Bubeck, arXiv:1405.4980v2, Lemma 3.1, p. 263
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.