The ?-EM algorithm and its basic properties

Abstract

A new class of statistical algorithms is presented and examined. The method is called the D-EM algorithm. This novel algorithm contains the traditional EM algorithm as a special case of D = 1. The choice of the design parameter D affects the eigenvalues of Hessian matrices for likelihood maximization. This causes much faster convergence than the traditional EM algorithm. Convergence theorems are given for the basic D-EM algorithm and its practical variants. Numerical evaluation shows fast convergence at nearly one-third the iteration counts and one-half the CPU time relative to the traditional method.