Weighted covariance matrix

Another interesting approach was developed in [19], in which ICA estimation was performed by ordinary eigenvalue decomposition of a quadratically weighted covariance matrix, i.e., $E\{\Vert{\bf x}\Vert^2{\bf x}{\bf x}^T\}$. This approach needs the assumption that all the independent components have different distributions (in particular, different kurtoses). It was realized later [20] that this approach is in fact a special case of the cumulant tensor approach of Section 4.3.5.