Computing symmetric nonnegative rank factorizations (Q651217)

The paper concerns the nonnegative rank factorization (NRF): given \(A\in \mathbf{R}_+^{m\times n}\) of rank \(r\) find nonnegative factors \(W\in \mathbf{R}_+^{m\times k}\) of full column rank and \(H\in \mathbf{R}_+^{k\times n}\) of full rank such that \(k=r\) and \(A=WH.\) The authors present an algorithm for NRF of some completely positive (CP) matrices whose rank is equal to their CP-rank. This one can compute the symmetric NRF of any nonnegative symmetric rank-\(r\) matrix that contains a diagonal principal submatrix of that rank and size. The first part is an introduction in nature. The second part reviews the previous efforts in the literature, concerning the computation of symmetric NRF. In the third part, knowing that the NRF factors might not be unique, causing a form of ill-conditioning, the authors recall an important property of symmetric positive semidefinite matrices that is then used to bind all possible solutions of the symmetric NRF. In the fourth part, using the above property, the authors propose an algorithm that is applicable to any symmetric, rank-2 matrix that is doubly nonegative. The fifth part provides properties of the sought factors and uniqueness results. The sixth section describes the algorithm, called IREVA, for the aforementioned types of matrices. The numerical validation of the IREVA is presented in the seventh part for the case of a symmetric matrix \(A\in \mathbf{R}_+^{6\times 6}\) of rank-3 that contains a diagonal principal submatrix of equal rank.