Theorem 3.9, p. 14 — near a locally symmetric point,
ProvedProjLikeRetr.Spectral.projection_spectral_manifoldAssume the hypotheses of Theorem 3.5: is a submanifold of of some dimension ; is the associated spectral set; and ; and there is such that is strongly locally symmetric (3.15).
Then, after restricting if necessary, there is such that for every symmetric matrix with :
- the projection of onto is unique, ; and
- for every with , the projection of onto is unique and equals
Norms are Frobenius on matrices and Euclidean on vectors, and is the set of nearest points of to . The theorem gives the projective retraction on a spectral manifold in closed form: an eigendecomposition of plus a projection in onto the much smaller manifold .
Formalization Note The page writes and as functions; here both are sets (nearest points, IsMetricProjection) and their being singletons is part of the conclusion. The radius is a because the proof "restrict[s] if necessary" so that Lemma 3.1 (local uniqueness of ) and Lemma 3.8 apply; shrinking keeps (3.15) true. The page states Theorem 3.5 for or ; covers both, since a submanifold is . The proof's "Let " means any symmetric in the ball, as stated here. Theorem 3.5's conclusion ( is a manifold) is not assumed.
import Mathlib import Definitions.Def_RandomGradFree_Nonsmooth_IsMetricProjection import Definitions.Def_ProjLikeRetr_Spectral_SpectralSet import Definitions.Def_ProjLikeRetr_Spectral_IsSubmanifold open scoped Matrix Matrix.Norms.Frobenius open RandomGradFree.Nonsmooth
namespace ProjLikeRetr.Spectral
/-- Theorem 3.9, p. 14 (projection onto spectral manifolds), under the hypotheses of
Theorem 3.5, p. 12. -/
theorem projection_spectral_manifold {n d : ℕ} (M : Set (EuclideanSpace ℝ (Fin n)))
(hM : IsSubmanifold 2 d M) (Xbar : Matrix (Fin n) (Fin n) ℝ)
(hXbar : Xbar ∈ specSet (M ∩ sortedDesc n)) (δ : ℝ) (hδ : 0 < δ)
(hsym : IsStronglyLocallySymmetric M (eig Xbar) δ) :
∃ δ₀ : ℝ, 0 < δ₀ ∧ δ₀ ≤ δ ∧ ∀ X : Matrix (Fin n) (Fin n) ℝ, X.IsHermitian →
‖X - Xbar‖ ≤ δ₀ / 2 →
∃ p : EuclideanSpace ℝ (Fin n), {z | IsMetricProjection M (eig X) z} = {p} ∧
∀ U ∈ Matrix.orthogonalGroup (Fin n) ℝ,
X = U * Matrix.diagonal (WithLp.ofLp (eig X)) * Uᵀ →
{Y | IsMetricProjection (specSet (M ∩ sortedDesc n)) X Y} =
{U * Matrix.diagonal (WithLp.ofLp p) * Uᵀ} := by sorry
end ProjLikeRetr.Spectral
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.