(G.6), p. 46 — for the optimal rotation O, ‖Û − U⁰O‖²_F ≤ 8‖X̂ − X⁰‖₁/n²_min
ProvedRobustSpecClust.Round.eq_G6davis-kahaneigenvector-perturbationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1spectral-clustering
Let be a labelling with every cluster nonempty (), the true clustering matrix and . Let be symmetric with , and let consist of orthonormal top- eigenvectors of . Let be an optimal rotation, i.e. an orthogonal matrix minimising over all orthogonal . Then
This is the eigenvector-perturbation step of the proof of Theorem 2: it combines the Davis–Kahan theorem, the eigengap of and the box constraint on .
Formalization Note The page writes without naming ; §4.2 (p. 17) calls it "the optimal rotation matrix". The statement is made for every minimiser (minimisers exist since the orthogonal group is compact), which is stronger than the existence of one good .
Preamble
import Mathlib import Definitions.Def_RobustSpecClust_Round_Setting
Formal statement
namespace RobustSpecClust.Round
open Matrix
theorem eq_G6 {n r : ℕ} [NeZero r] (φ : Fin n → Fin r) (hφ : ∀ k, 0 < clusterSize φ k)
(Xh : Matrix (Fin n) (Fin n) ℝ) (hsym : Xhᵀ = Xh) (hbox : ∀ i j, 0 ≤ Xh i j ∧ Xh i j ≤ 1)
(Uh : Matrix (Fin n) (Fin r) ℝ) (hU : IsTopEigvecs Xh Uh)
(O : Matrix (Fin r) (Fin r) ℝ) (hO : O ∈ Matrix.orthogonalGroup (Fin r) ℝ)
(hOopt : ∀ O' ∈ Matrix.orthogonalGroup (Fin r) ℝ,
frobSq (Uh - U0 φ * O) ≤ frobSq (Uh - U0 φ * O')) :
frobSq (Uh - U0 φ * O) ≤ 8 * l1 (Xh - trueClust φ) / nmin φ ^ 2 := by sorry
end RobustSpecClust.Round
Source
arXiv:1912.07546v3, Appendix G, proof of Theorem 2, p. 46, (G.6); O as in §4.2, p. 17 ("the optimal rotation matrix")
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.