Maximum Likelihood Estimates and a Kernel k-Means Iterative Algorithm for Normal Mixtures

Duda et al. established a connection between maximum likelihood estimates and a k-Means algorithm approximating the Mahalanobis distance by the Euclidean distance for Normal Mixtures. They suggested that a more accurate result might be possible if identical covariance matrices were assumed. In this paper that is shown to be true by using a kernel K-means algorithm that does not rely on approximating the Mahalanobis distance.