Abstract We study the influence of noise in the EM algorithm and ISRA. Both methods have been used for emission computed tomography. In the underlying linear model Ax b A x ∼ b the noise can be in the data vector b and/or in the matrix A. Under certain conditions it is shown that the noise error asymptotically grows like O (k) O (k), with k the iteration index. For ISRA we also introduce a relaxation parameter and give conditions under which convergence is maintained.
Tommy Elfving (Mon,) studied this question.