LOGSPLINE INDEPENDENT COMPONENT ANALYSIS

Most recent maximum likelihood approaches to independent component analysis (ICA) are based on nonparametric density estimation. In this paper, we present an algorithm by using the logsplines approach to density estimation. The logarithmic source density functions are modeled by polynomial splines or a linear combination of B-splines with (a) parameters or coefficients of the B-splines estimated by maximizing the log-likelihood function, and (b) knots of the B-splines determined by a stepwise procedure so as to minimize the approximation errors in modeling the log-density functions. We showed in a comparative study that our new algorithm has performed very favorably when compared to several popular density estimation based procedures