Abstract
This note illustrates some shortcomings of the criterion proposed in "Blind source separation based on endpoint estimation with applications to the MLSP 2006 data competition" when the number of samples is finite. This algorithm considers mutually independent sources with semibounded support, however, even for a sufficient sample size for which the finite bound of the support of the density of the output can be estimated accurately, the endpoint estimate nearest to the mean might be in the unbounded side of this density. In that case, the superadditivity of the least absolute endpoint estimate is usually violated, causing the loss of the contrast function property and of the capability of discriminating hidden sources, for practical versions of this criterion.
Original language  English 

Pages (fromto)  863865 
Number of pages  3 
Journal  Neurocomputing 
Volume  74 
Issue number  5 
DOIs 

Publication status  Published  Feb 2011 
Externally published  Yes 
Keywords
 Blind source separation
 Independent component analysis
 Least absolute endpoint