Proposition 4.1, p. 7 — vec(C^(q)E)ᵀ = vec(C^(q))ᵀF^(q) and ∂/∂vec(C^(q))ᵀ = ∂/∂vec(A^(P+q))ᵀ·F^(q)ᵀ
ProvedTensorBTD.Gramian.proposition_4_1Let be the block diagonal matrix and . Then:
- for every ,
- (chain rule (4.3b)) for every tensor and every vector of unknowns of the (rank- rank-1) BTD, the columns of the BTD Jacobian that belong to are
where the right-hand Jacobian is that of the unstructured CPD residual (all factor matrices free) evaluated at the factor matrices of , i.e. with .
The proposition lets one ignore the structure when differentiating, and then correct by the fixed matrix .
Formalization Note. Part (4.3a) of the proposition (the matrix form of the cogradient) is not formalized here; this mission formalizes the vec identity and the Jacobian form (4.3b), which is what Theorem 4.5 uses. Row vectors times matrices are Mathlib's vecMul; vec is column-major, indexed by (column, row). The left-hand side of part 2 is the submatrix of at the columns , the right-hand side the submatrix of the unstructured CPD Jacobian at the columns , times . Since is not an unknown of the BTD, is read as the derivative of the unstructured CPD residual, as in the proof of Theorem 4.5.
import Mathlib import Definitions.Def_TensorBTD_Gramian_Setting
namespace TensorBTD.Gramian
open Matrix
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,
Matrix.vec (C * E L) = Matrix.vecMul (Matrix.vec C) (F I L q)) ∧
(∀ (T : Tensor I) (z : Unk I L → ℂ),
(jac (residual T) z).submatrix id
(fun y : Fin R × Fin (I (.inr q)) => (⟨.inr q, y⟩ : Unk I L)) =
(jac (cpdResidual T) (fullVec z)).submatrix id
(fun y : Col L × Fin (I (.inr q)) => (⟨.inr q, y⟩ : GIdx I L)) *
(F I L q)ᵀ) := by sorry
end TensorBTD.Gramian
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.