Controlling noise error in block iterative methods (Q501961): Difference between revisions

The authors consider the iterative solution of large, sparse and ill-conditioned systems of linear equations like they arise from the computerized tomography problem. When solved by an iterative method and if, as usual, there is noise present in the right hand data, these systems show the phenomenon of semi-convergence, see [\textit{F. Natterer}, The mathematics of computerized tomography. Stuttgart: B. G. Teubner; Chichester etc.: John Wiley \& Sons (1986; Zbl 0617.92001)], i.e., the iterative solution first approaches a regularized solution and then starts to be increasingly corrupted by noise. The solution method considered here is a kind of block Kaczmarz iteration with iteration parameter and a weighting matrix and, possibly, project onto the null space, and is called simultaneous iterative reconstruction technique (SIRT). Relying on their earlier work [``Convergence studies on iterative algorithms for image reconstruction'', IEEE Trans. Med. Imaging 22, No. 5, 569--579 (2003; \url{doi:10.1109/TMI.2003.812253})] and on the work of \textit{T. Elfving} et al. [SIAM J. Sci. Comput. 34, No. 4, A2000-A2017 (2012; Zbl 1254.65044)], they consider a regularized problem containing 2 parameters and are able to estimate the noise error using certain singular values and to propose 3 strategies for choosing the iteration parameters. When cycling through all blocks of the decomposed system and using a constant relaxation parameter per cycle, they prove that their method converges to the least squares solution. In their numerical experiments this cycle-wise method shows better for noiseless data whereas the not fully proven case of iteration-wise changing parameters is better for noisy data.