Proposition 4.2, p. 7 — V = A^(p_1) ⊙ ⋯ ⊙ A^(p_N) satisfies VᴴV = ∗ₙ A^(n)ᴴA^(n)
ProvedTensorBTD.Cogradient.proposition_4_2gramianhadamard-productkhatri-rao-productp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let for , let be any permutation of , and let
be the Khatri–Rao product of the matrices in the order . Let be the Hadamard (entrywise) product of all the Gramians . Then
This is the standard identity that turns the Gramian of a Khatri–Rao product into a Hadamard product of small Gramians; it is the reason the cogradient of Theorem 4.4 can be evaluated without forming .
Formalization Note. The rows of are indexed by the tuple rather than by an integer; the entry in row and column is . Columns are 0-based. For both sides are the all-ones matrix (empty products).
Preamble
import Mathlib import Definitions.Def_TensorBTD_Cogradient_Setting
Formal statement
namespace TensorBTD.Cogradient
open Matrix
/-- Proposition 4.2, p. 7: for matrices `A^(n) ∈ ℂ^{I_n × J}`, `n = 1, …, N`, and any permutation
`p` of the modes, the Khatri–Rao string `V = A^(p_1) ⊙ ⋯ ⊙ A^(p_N)` satisfies `Vᴴ V = W`, where
`W = ∗_n A^(n)ᴴ A^(n)` is the Hadamard product of the Gramians. -/
theorem proposition_4_2 {N J : ℕ} {I : Fin N → ℕ} (A : (n : Fin N) → Matrix (Fin (I n)) (Fin J) ℂ)
(p : Equiv.Perm (Fin N)) :
(krString A p)ᴴ * krString A p = Matrix.of fun j j' => ∏ n, ((A n)ᴴ * A n) j j' := 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. 7, Proposition 4.2
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.