Proof of Theorem 4.4, p. 8 — f_BTD = ½(f^(1) − f^(2) − f̄^(2) + f^(3)) for every mode n
ProvedTensorBTD.Cogradient.fBTD_decompositioncanonical-polyadic-decompositionleast-squaresp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1tensor-decomposition
Let be an th-order tensor, . First, is the objective of the unstructured CPD evaluated at the structured factor matrices (3.6): for all unknowns ,
Second, for all factor matrices of the unstructured CPD and every mode , with
one has
where is the objective of the unstructured CPD.
This splits the objective into a part depending only on the factors, a part linear in the data, and a constant; it opens the proof of Theorem 4.4.
Formalization Note. The inner product is Definition 2.1's, conjugate on the first argument. The identity holds for every mode simultaneously, because the mode- unfolding of the residual is (3.8). All quantities are cast to .
Preamble
import Mathlib import Definitions.Def_TensorBTD_Cogradient_Setting
Formal statement
namespace TensorBTD.Cogradient
open Matrix
/-- Proof of Theorem 4.4, p. 8, first sentence: `f_BTD` is the objective of the unstructured CPD
at the factor matrices `A^(n)` of (3.6), and for every mode `n`,
`f = ½ (f^(1) − f^(2) − conj f^(2) + f^(3))` with `f^(1) = ‖A^(n) V^{n}ᵀ‖²`,
`f^(2) = ⟨T_(n), A^(n) V^{n}ᵀ⟩`, `f^(3) = ‖T_(n)‖²`. -/
theorem fBTD_decomposition {P Q R : ℕ} {I : Fin P ⊕ Fin Q → ℕ} {L : Fin R → ℕ} (T : TensorBTD.Gramian.Tensor I) :
(∀ z : TensorBTD.Gramian.Unk I L → ℂ, fBTD T z = fCPD T (TensorBTD.Gramian.fullVec z)) ∧
∀ (x : TensorBTD.Gramian.GIdx I L → ℂ) (n : Fin P ⊕ Fin Q),
fCPD T x = (1 / 2 : ℂ) *
(fOne n x - fTwo T n x - starRingEnd ℂ (fTwo T n x) + fThree T n) := by sorry
end TensorBTD.Cogradient
Source
Sorber, Van Barel & De Lathauwer, Optimization-Based Algorithms for Tensor Decompositions, SIAM J. Optim. 23(2) (2013), doi:10.1137/120868323, authors' manuscript (Lirias), p. 8, proof of Theorem 4.4, first sentence; (3.6)–(3.8), p. 6
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.