Column-wise symmetric block partitioned tensor decomposition
Christopher Mueller-Smith, Predrag Spasojevic
Abstract
Symmetric block partitioned tensors (SBPT) are a useful structure in signal processing applications, often generated from computing higher-order statistics on observed data. Such tensors often follow the rank (R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</sub> , R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</sub> , 1) SBPT structure, but in some applications the partitioning of the factor matrices is not known a priori. We propose a both blind and non-blind column-wise SBPT decomposition algorithms that are better scalable to high-dimensional tensors because they avoid large matrix inversions. We apply the algorithms to simulated SBPTs and demonstrate that they estimate factor matrices having high congruence with the originals across a range of collinearity values for the columns of the original factor matrices.
BibTeX
@inproceedings{icassp2016_columnwisesymmet,
title = {Column-wise symmetric block partitioned tensor decomposition},
author = {Christopher Mueller-Smith and Predrag Spasojevic},
booktitle = {ICASSP 2016},
year = {2016}
}