Tensor techniques in signal processing: algorithms for the canonical polyadic decomposition (PARAFAC)

Abstract

Low rank tensor decomposition has been playing for the last years an important role in many applications such as blind source separation, telecommunications, sensor array processing, neuroscience, chemometrics, and data mining. The Canonical Polyadic tensor decomposition is very attractive when compared to standard matrix-based tools, manly on system identification. In this thesis, we propose: (i) several algorithms to compute specific low rank-approximations: finite/iterative rank-1 approximations, iterative deflation approximations, and orthogonal tensor decompositions. (ii) A new strategy to solve multivariate quadratic systems, where this problem is reduced to a best rank-1 tensor approximation problem. (iii) Theoretical results to study and proof the performance or the convergence of some algorithms. All performances are supported by numerical experiments