Abstract :
This paper compares the use of the kurtosis function and higher-order moment functions for the extraction of independent components from a hybrid mixture. The kurtosis function, based on the 4-th order moments, is a theoretically attractive contrast function since it possess a traceable extrema-correspondence. Numerical motivations have lead many researchers to look into contrast functions involving higher-order moments. To facility the study, we define super-ness and sub-ness of non-Gaussian processes as terms that are dependent not only only on the nature of the signals themselves, but also on the the contrast functions used. In terms of the extrema-correspondence, we report some noteworthy discrepancies between the kurtosis and higher-order moment functions. Finally, we verify that for even-positive-integer and equal-moment components, both the necessary and sufficient conditions are valid
Keywords :
higher order statistics; method of moments; optimisation; signal detection; extrema-correspondence; higher-order moment functions; independent component analysis; kurtosis function; necessary conditions; signal detection; source separation; subness; sufficient conditions; superness; Array signal processing; Deconvolution; Entropy coding; Feature extraction; Flip-flops; Independent component analysis; Noise reduction; Optimized production technology; Sensor arrays; Sensor phenomena and characterization;