Proposition 4.1, (4.3a), p. 7 — vec(C^(q)E)ᵀ = vec(C^(q))ᵀF^(q) and ∂/∂C^(q) = ∂/∂A^(P+q)·Eᵀ
ProvedTensorBTD.Cogradient.proposition_4_1Let , the Kronecker product of the matrix with the identity of order . Then:
- for every ,
- (chain rule (4.3a)) let be a complex-valued function of the factor matrices of the unstructured CPD, and let be the structured unknowns. If is real-differentiable at the factor matrices (3.6) of , then the cogradient of with respect to is
the right-hand side being the cogradient of with respect to evaluated at .
The proposition lets one first differentiate with respect to as if it were unstructured and then transfer to ; it gives (4.5b) from (4.5a).
Formalization Note. vec is column-major (index (column, row), Mathlib's Matrix.vec) and is Mathlib's Kronecker product, whose row and column pairs are ordered as in Definition 2.9. The hypothesis that is real-differentiable (Fréchet, over ) at the point is the presupposition of "applying the chain rule"; without it the identity can fail because deriv returns at non-differentiable points. Every function the mission applies it to (in particular ) is a polynomial in the real and imaginary parts and satisfies it. The Jacobian form (4.3b) is not stated here.
import Mathlib import Definitions.Def_TensorBTD_Cogradient_Setting
namespace TensorBTD.Cogradient
open Matrix
/-- Proposition 4.1, p. 7, with (4.3a): `vec(C^(q) E)ᵀ = vec(C^(q))ᵀ F^(q)` for
`F^(q) = E ⊗ 𝕀_{I_(P+q)}`, and the chain rule `∂/∂C^(q) = ∂/∂A^(P+q) · Eᵀ` for the cogradient of
a function `g` of the unstructured factor matrices, real-differentiable at the point, composed with
`A^(P+q) = C^(q) E`. -/
theorem proposition_4_1 {P Q R : ℕ} {I : Fin P ⊕ Fin Q → ℕ} {L : Fin R → ℕ} :
(∀ (q : Fin Q) (C : Matrix (Fin (I (.inr q))) (Fin R) ℂ),
vec (C * TensorBTD.Gramian.E L) = vecMul (vec C) (kronecker (TensorBTD.Gramian.E L) (1 : Matrix (Fin (I (.inr q))) (Fin (I (.inr q))) ℂ))) ∧
∀ (g : (TensorBTD.Gramian.GIdx I L → ℂ) → ℂ) (z : TensorBTD.Gramian.Unk I L → ℂ), DifferentiableAt ℝ g (TensorBTD.Gramian.fullVec z) →
∀ q : Fin Q, cogradC (g ∘ TensorBTD.Gramian.fullVec) z q = cogradMatG g (TensorBTD.Gramian.fullVec z) (.inr q) * (TensorBTD.Gramian.E L)ᵀ := by sorry
end TensorBTD.Cogradient
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.