A truncated projected SVD method for linear discrete ill-posed problems (Q870762)

The authors consider ill-posed problems of the form \(Ax=b\), where \(A\in{\mathbb R}^{m\times n}\) is of ill-determined rank and \(m\geq n\). They choose \(W\in{\mathbb R}^{n\times \ell}\) with orthonormal columns spanning \(\mathcal{W}\) such that \(AW\) has full rank. They form the QR factorization \(AW=QR\) where \(Q\in{\mathbb R}^{m\times \ell}\) has orthonormal columns and \(R\) is upper triangular. They introduce the orthogonal projectors \(P_W=WW^T\), \(P_W^\perp=I-P_W\), \(P_Q=QQ^T\) and \(P_Q^\perp=I-P_Q\), and split the solution \(x\) according to \(x=x'+x''\), where \(x'=P_W x\) and \(x''=P_W^\perp\), thus splitting the system as \(P_QAx'+P_QAx''=P_Qb\) and \(P_Q^\perp Az''=P_Q^\perp b\). The truncated singular value decomposition (SVD) is applied to the last equation to compute \(x_k^{\prime\prime}\in{\mathbb R}^n\setminus{\mathcal W}\), where the number \(k\) of terms is determined by the discrepancy principle. Then an approximate solution \(x_k'\) of the second last equation is computed. Numerical examples show the efficiency of the method.