Modified Fisher scoring algorithms using Jacobi or Gauss-Seidel subiterations (Q1389253)

The paper deals with the large dimensional problem of estimating an unknown \(p\)-dimensional vector parameter via maximum likelihood or maximum penalized likelihood methods. When \(p\) is large, the traditional algorithms such as Fisher scoring and Newton-Raphson algorithm become impractical because each iteration requires the exact solution of large linear systems. The authors propose to incorporate Fisher scoring method with approximate solution of linear systems instead of the exact one. They focus on two well-known approximate methods, namely Jacoby and Gauss-Seidel, which can be realized as subiteration procedures within each iteration of Fisher scoring. The main advantage of such approach is that it does not demand inversion of large matrices. Simulation studies show that this modified algorithm exhibits much faster convergence than the EM algorithm. Applications of the modified Fisher scoring algorithm in emission and transmission tomography are discussed.