A parallel projection method for solving generalized linear least-squares problems (Q1112561)

Let \(A\in {\mathbb{R}}^{m,n}\), \(b\in {\mathbb{R}}^ m\) and K be a convex polyhedral cone in \({\mathbb{R}}^ m\). To find a vector \(x\in {\mathbb{R}}^ n\) minimizing the 2-norm distance from Ax-B to K, the authors propose a parallel projection method based on the dual problem: \((\min (1/2)<y,y>+<b,y>\) such that \(A^ Ty=0\) and \(y\in K^ 0\), where \(K^ 0\) denotes the polar of the cone K. Using a column partioning technique to decompose A, the dual problem is split into some independent smaller subproblems which consist in finding projections onto subspaces. These subproblems can be solved in parallel. The convergence analysis presented in the paper shows that the authors' method converges to the solution of the dual problem.