Proof of Theorem 4.4, p. 8 — ∂f^(1)/∂A^(n) = Ā^(n)·W^{n}
ProvedTensorBTD.Cogradient.cograd_f1gramiankhatri-rao-productp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1wirtinger-derivative
Consider the unstructured CPD with factor matrices and fix a mode . Let
where is the Khatri–Rao string of all factor matrices except . Then the cogradient of with respect to the entries of is
where is the entrywise complex conjugate of and is the Hadamard product of the Gramians , .
This is the factor-dependent part of the cogradient of .
Formalization Note. The cogradient is coordinatewise, with all other factor matrices held fixed; the mode may be any of the modes, including , where is a free matrix of the unstructured CPD.
Preamble
import Mathlib import Definitions.Def_TensorBTD_Cogradient_Setting
Formal statement
namespace TensorBTD.Cogradient
open Matrix
/-- Proof of Theorem 4.4, p. 8, first display: for every mode `n` of the unstructured CPD,
`∂f^(1)/∂A^(n) = Ā^(n) · W^{n}` with `f^(1) = ‖A^(n) V^{n}ᵀ‖²`. -/
theorem cograd_f1 {P Q R : ℕ} {I : Fin P ⊕ Fin Q → ℕ} {L : Fin R → ℕ} (x : TensorBTD.Gramian.GIdx I L → ℂ)
(n : Fin P ⊕ Fin Q) :
cogradMatG (fOne n) x n = (TensorBTD.Gramian.cpdFactor x n).map (starRingEnd ℂ) * W {n} (TensorBTD.Gramian.cpdFactor x) := 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 display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.