Lemma 13 — projection correction: and
ProvedIntervalToDynamic.lemma_13dynamic-regretinterval-regretonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be closed and convex, and let satisfy for some . For every and every , put , , and let be the corrected gradient of at , so that , with when . Then
The lemma lets a linearized loss at the projected point be charged to the unprojected decision without growing the gradient, which is the basis of domain conversion.
Differences from the source: none.
Preamble
import Mathlib import Definitions.Def_IntervalToDynamic_proj import Definitions.Def_IntervalToDynamic_correctedGrad
Formal statement
namespace IntervalToDynamic
/-- Lemma 13 (projection correction). Wang–Zhao–Zhou, *On the Relation Between Interval Regret and Dynamic Regret*, arXiv:2609.34423, Lemma 13, p.30.
Let `X ⊆ ℝ^d` be closed and convex and `‖g‖₂ ≤ G`, `G ≥ 0`. For every `u ∈ X` and every `y ∈ ℝ^d`,
with `x = Π_X[y]`, `n = y − x` and `d = correctedGrad g n`, one has `‖d‖₂ ≤ ‖g‖₂ ≤ G` and
`gᵀ(x − u) ≤ dᵀ(y − u)`.
**Formalization Note.** `X` is nonempty because `u ∈ X`, so `proj X y` is the projection. -/
theorem lemma_13 {d : ℕ} (hd : 0 < d) (X : Set (EuclideanSpace ℝ (Fin d))) (hXc : IsClosed X) (hXv : Convex ℝ X)
(g : EuclideanSpace ℝ (Fin d)) (G : ℝ) (hG : 0 ≤ G) (hg : ‖g‖ ≤ G) (u y : EuclideanSpace ℝ (Fin d)) (hu : u ∈ X) :
‖correctedGrad g (y - proj X y)‖ ≤ ‖g‖ ∧ ‖g‖ ≤ G ∧
inner ℝ g (proj X y - u) ≤ inner ℝ (correctedGrad g (y - proj X y)) (y - u) := by sorry
end IntervalToDynamic
Source
Wang–Zhao–Zhou, On the Relation Between Interval Regret and Dynamic Regret, arXiv:2609.34423, Lemma 13, p.30
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.