The NMF problem and lattice-subspaces (Q2185834)

The nonnegative matrix factorization (NMF) problem for a nonnegative \(n\times m\) real matrix \(A\) is the problem of determining two nonnegative real \(n\times p\) matrix \(F\) and \(p\times m\) matrix \(V\) such that \(A =FV\) with \(p<\min \{n, m\}\). When \(p\) is the rank of \(A\), the problem is known as nonnegative rank factorization (NRF) of \(A\) and some results have existed in the literature. The problem does not have always a solution. In this paper, an algorithmic method is given for finding \(F, V\) such that the ranks of \(F\) and \(A\) are the same, under some condition on \(p\). The method is based on the theory of lattice-subspaces and positive bases, namely, it involves a positive basis of a minimal lattice-subspace which contains a maximal set of linearly independent rows of \(A\). The author notes that when \(A\) is symmetric and nonnegative, the method does not warrant a symmetric factorization \(A = WW^T\). Examples are provided to illustrate the algorithm which has different cases. A MATLAB code is used in some of the examples. An appendix is given on lattice-subspaces and positive bases.