EXTENDED QUASINEWTON METHOD FOR THE ICA
Toshinao Akuzawa
As extensions to the strict method developed in [1], varia
tions of the Newton method for the independent component
analysis(ICA) are proposed. Our method presented here is
highly practical and simple. Concrete merits of our algo
rithm are as follows. i) Robust under gaussian noises. In
the presence of strong gaussian noises it outperforms the ex
isting methods like the JADE[2] and the FICA[3]. ii) By
two deformations it becomes considerably stable globally.
Although the first deformation is apparently unnatural, a
justification is given based on the random matrix arguments.
iii) Each step of the methods proposed here resolves itself
into the determination of the inverse of 2 2 matrices or
generalized inverse matrices of 3 2 matrices. There is no
need to deal with gigantic matrices and little computational
resources are required.