Figure 6 also suggests that the estimated Gaussian subspace [W.sub.g-Ed] of the SSTA is spanned by the last 6 SVecs (with a confidence level [p.sub.sig] = 95%) and so [N.sub.g-Ed] = 6 and, [N.sub.ng-Ed] = 5 for the non-Gaussian subspace [W.sub.ng-Ed].
Using the estimated non-Gaussian subspace (Section 3.2), we apply here ICA to the leading [N.sub.ICA] = 5 SVecs of the SSTA dataset (see Figure 6).
Further experiments show that IC1 is quite robust with the use of [N.sub.ICA] < 5 leading SVecs with the remaining ICs being slight mixtures of those obtained with [N.sub.ICA] = 5.
R: Orthogonal rotation matrix filled by singular vectors of M(y) (SVecs)
[R.sub.ICA]: Orthogonal matrix relating the SVecs and ICA vectors
[R.sub.ISA]: Orthogonal matrix relating the SVecs and ISA sources