Convergence analysis for a multiplicatively relaxed EM algorithm

From MaRDI portal





This paper studies the convergence of the multiplicatively relaxed EM (expectation maximization) algorithm, which is an iterative procedure to determine maximum likelihood estimators in situations of incomplete data. Let \(b\in R^ m_{>0}\) and \(A\in R^{m\times n}_{\geq 0}\) with rows \(a^ i\neq 0\) and element \(a_{ij}\) satisfying \(\sum^ m_{i=1} a_{ij}=1\), \(1\leq j\leq n\). Define \(\Gamma=\{x\in R^ n_{\geq 0}: Ax>0\}\). Consider \(f: \Gamma\to R\) defined by \(f(x)=\sum^ m_{i=1}(\langle a^ i,x\rangle -b_ i\log \langle a^ i,x\rangle)\). The problem is \(\min f(x)\hbox{ s.t. }x\in\Gamma\). Define \(P_ \alpha(x)_ j=x_ jc_ j(x)^ \alpha\), where \(c_ j(x)=\sum^ m_{i=1}b_ ia_{ij}/\langle a^ i,x\rangle\) and \(\alpha \in (0,2)\). The relaxed \(EM\) algorithm is defined as \(x^ 0\in R^ n_{>0}\), \(x^{k+1}=P_ \alpha(x^ k)\). The author proves global convergence of \(x^ k\) to a solution of the problem in the underrelaxed case \(\alpha\in (0,1)\) and unrelaxed case \(\alpha=1\). A local convergence theorem is also presented for the overrelaxed case \(\alpha > 1\). Moreover, it is shown that the sequence is bounded and that the limit, if it exists, is a solution of the problem. The problem considered here is originated in the analysis of positron emission tomography, i.e., to obtain a maximum likelihood estimator in the case of independent Poisson random variables.











This page was built for publication: Convergence analysis for a multiplicatively relaxed EM algorithm

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3986078)