Identification of independent components using cumulants and coherences
P. Ruiz · 2002
The characterization of independent stationary stochastic components (sources), can theoretically be achieved by using the 2/sup nd/ and 4/sup th/ order cumulants of partially correlated measurements, which are linearly related to the components of interest. The general model supposes as many sources as measurements. The authors recall first why 4/sup th/ order cumulants achieve a solution, while 2/sup nd/ order cumulants fail. However, the main algorithms (Comon, Cardoso, Gaeta-Lacoume) are generally used with a 2 observations -2 independent components models, to avoid complexity of calculation. The authors first examine what is achieved when the number of independent components is not well estimated. They suggest a new algorithm for independent components identification, which enables more than two sources, and based on both 2/sup nd/ and 4/sup th/ order cumulants.>