CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Asymptotic comparison of tensorial and vectorial ICA methods Authors:  Joni Virta - Aalto University (Finland) [presenting]
Bing Li - The Pennsylvania State University (United States)
Klaus Nordhausen - Vienna University of Technology (Austria)
Hannu Oja - University of Turku (Finland)
Abstract: Naturally tensor-valued data is more and more common nowadays and various different probabilistic tensorial models have been developed in response. The tensor independent component model and various methods for estimating its parameters are discussed. The methods can be divided into two classes, tensorial and those requiring vectorization of the data prior to use, a procedure that causes loss of structural information. To see how severe the loss is the different methods are compared based on their asymptotic efficiencies, yielding the conclusion that under the model vectorization is practically always inferior to the purely tensorial ICA methods.