The expectation maximization (EM) algorithm is an iterative procedure used to determine maximum likelihood estimators in situations of incomplete data. In the case of independent Poisson variables it converges to a solution of a problem of the form min ∑[〈ai,x〉 − bi log 〈ai, x〉] such that x ⩾0. Thus, it can be used to solve systems of the form Ax = b, x⩾0 (with A stochastic and b positive.) It converges to a feasible solution if it exists and to an approximate one otherwise (the one that minimizes d (b, Ax), where d is the Kullback–Leibler information divergence). We study the convergence of the multiplicatively relaxed version proposed by Tanaka for use in positron emission tomography. We prove global convergence in the underrelaxed and unrelaxed cases. In the overrelaxed case we present a local convergence theorem together with two partial global results: the sequence generated by the algorithm is bounded and, if it converges, its limit point is a solution of the aforementioned problem.
[1]
L. Shepp,et al.
A Statistical Model for Positron Emission Tomography
,
1985
.
[2]
W. Cheney,et al.
Proximity maps for convex sets
,
1959
.
[3]
Eiichi Tanaka.
A Fast Reconstruction Algorthm for Stationary Positron Emission Tomography Based on a Modified EM Algorithm
,
1987,
IEEE Transactions on Medical Imaging.
[4]
K. Lange,et al.
EM reconstruction algorithms for emission and transmission tomography.
,
1984,
Journal of computer assisted tomography.
[5]
G. McCormick.
Second Order Conditions for Constrained Minima
,
1967
.