Incomplete alternating projection method for large inconsistent linear systems (Q2472379)

This paper is concerned with iterative methods for solving the linear least-squares problem \(\min \| A x - b \| _2^2\), where \(A\) is an \(m\times n\) matrix and \(b\) is a given vector of length \(m\). The methods under consideration are based on a reformulation in terms of the minimal distance between two affine subspaces. This allows the application of von Neumann's alternating projection method to linear least-squares problems [cf. \textit{H. D. Scolnik, N. Echebest, M. T. Guardarucci} and \textit{M. C. Vacchino}, Math. Program. 111, No. 1--2 (B), 273--300 (2008; Zbl 1344.65046)]. However, one of these projections becomes rather expensive as \(m,n\) increase. The authors propose to replace this projection by a finite number of steps of an inexpensive iteration. It is shown how the Fejér monotonicity of the alternating projection method can be preserved despite the additional approxmation error imposed by the inner iteration. An academic numerical example demonstrates the viability of the proposed modifications.