Accelerating the EM algorithm using rescaled block-iterative methods

Block-iterative methods, in which only part of the data is used at each step, can converge significantly faster than simultaneous methods, such as EMML or SMART, in which all the data is employed at each step. The authors discuss the rescaled block-iterative (RBI) approach to both algorithms. When a nonnegative solution exists, these RBI algorithms converge to a solution for any configuration of subsets. The RBI-EMML reduces to the "ordered subset" method when "subset balance" holds. When there is no nonnegative solution block-iterative methods produce limit cycles, from which an approximate solution can be obtained using a "feedback" approach.