Efficient Identification of Butterfly Sparse Matrix Factorizations - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2022

Efficient Identification of Butterfly Sparse Matrix Factorizations

Résumé

Many well-known matrices $\mathbf{Z}$ are associated to fast transforms corresponding to factorizations of the form $\mathbf{Z} = \mathbf{X}^{(J)} \ldots \mathbf{X}^{(1)} $, where each factor $ \mathbf{X}^{(\ell)} $ is sparse and possibly structured. This paper investigates essential uniqueness of such factorizations. Our first main contribution is to prove that any $N \times N$ matrix having the so-called butterfly structure admits a unique factorization into $J$ butterfly factors (where $N = 2^J$), and that the factors can be recovered by a hierarchical factorization method. This contrasts with existing approaches which fit the product of the butterfly factors to a given matrix via gradient descent. The proposed method can be applied in particular to retrieve the factorizations of the Hadamard or the Discrete Fourier Transform matrices of size $2^J$. Computing such factorizations costs $\mathcal{O}(N^2)$, which is of the order of dense matrix-vector multiplication, while the obtained factorizations enable fast $\mathcal{O}(N \log N)$ matrix-vector multiplications. This hierarchical identifiability property relies on a simple identifiability condition in the two-layer and fixed-support setting that was recently established. While the butterfly structure corresponds to a fixed prescribed support for each factor, our second contribution is to obtain identifiability results with more general families of allowed sparsity patterns, taking into account permutation ambiguities when they are unavoidable. Typically, we show through the hierarchical paradigm that the butterfly factorization of the Discrete Fourier Transform matrix of size $2^J$ admits a unique sparse factorization into $J$ factors, when enforcing only $2$-sparsity by column and a block-diagonal structure on each factor.
Fichier principal
Vignette du fichier
main.pdf (605.02 Ko) Télécharger le fichier
S_1.pdf (2.25 Ko) Télécharger le fichier
S_2.pdf (2.29 Ko) Télécharger le fichier
S_3.pdf (2.27 Ko) Télécharger le fichier
S_4.pdf (2.31 Ko) Télécharger le fichier
dst_dct_supp_N_16.pdf (2.56 Ko) Télécharger le fichier
dst_dct_supp_N_4.pdf (1.3 Ko) Télécharger le fichier
dst_dct_supp_N_8.pdf (1.6 Ko) Télécharger le fichier
partitioning-binary-tree.pdf (44.18 Ko) Télécharger le fichier
proof-hierarchical.pdf (35.11 Ko) Télécharger le fichier
w_2_1.pdf (2.31 Ko) Télécharger le fichier
w_3_1.pdf (1.31 Ko) Télécharger le fichier
w_3_2.pdf (2.35 Ko) Télécharger le fichier
w_4_2.pdf (2.32 Ko) Télécharger le fichier
w_4_3.pdf (2.33 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03362626 , version 1 (01-10-2021)
hal-03362626 , version 2 (11-11-2021)
hal-03362626 , version 3 (15-02-2022)
hal-03362626 , version 4 (04-04-2022)
hal-03362626 , version 5 (02-08-2022)
hal-03362626 , version 6 (07-10-2022)

Identifiants

Citer

Léon Zheng, Elisa Riccietti, Rémi Gribonval. Efficient Identification of Butterfly Sparse Matrix Factorizations. 2022. ⟨hal-03362626v3⟩
438 Consultations
629 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More