Lemma 1, p. 20 — vectors orthogonal to a_1, …, a_d and with positive inner product with a_{d+1}, …, a_{L_1}
ProvedBadLocalMin.Smooth.lemma_1linear-algebraneural-networksp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and let be linearly independent. Then for every there are linearly independent vectors such that, for every ,
In the proof of Theorem 1 this lemma is applied to the vectors of Assumption 1(b), with the squared rows last, to build the output data.
Formalization Note. Vectors are indexed from 0: is a ⟨j - 1, _⟩, so "" becomes j.val < d and "" becomes d ≤ j.val. The inner product is dotProduct.
Preamble
import Mathlib
Formal statement
namespace BadLocalMin.Smooth
/-- Lemma 1 (p. 20). Vectors are indexed from 0: the paper's a₁, …, a_{L₁} are `a 0`, …,
`a (L1 - 1)`, so "1 ≤ j ≤ d" is `j.val < d` and "d < j ≤ L₁" is `d ≤ j.val`. -/
theorem lemma_1 (L1 L2 d : ℕ) (hL : L1 ≤ L2) (a : Fin L1 → Fin L2 → ℝ)
(ha : LinearIndependent ℝ a) (hd1 : 1 ≤ d) (hdL : d < L1) :
∃ u : Fin (L2 - d) → Fin L2 → ℝ, LinearIndependent ℝ u ∧
∀ l j, (j.val < d → u l ⬝ᵥ a j = 0) ∧ (d ≤ j.val → 0 < u l ⬝ᵥ a j) := by sorry
end BadLocalMin.Smooth
Source
Ding, Li & Sun, arXiv:1911.01413v3, p. 20, Lemma 1 (eqs. (24a)–(24b)); proof Appendix F, pp. 46–47
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.