Nonnegative matrix factorization of a correlation matrix (Q1025855)

A correlation matrix is a symmetric positive-semidefinite \(n \times n\) matrix \(C\) which has a unit main diagonal. The basic question is essentially concerned with finding a factorization of the type \(C \approx AA^T\) such that \(AA^T\) should be a low-rank approximation of \(C\) together with an additional nonnegativity constraint. First, in section 2, an application from financial engineering in which the non-negative matrix factorization arises is discussed. In section 3, a counterexample for the strict equality \(C= AA^T\) is given. Then, in section 4, an algorithm -- based on a two step procedure -- is developed for obtaining a non-negative matrix \(A\) which, moreover, aims at minimizing the Frobenius norm of the matrix \(C-AA^T\) subject to \(A_{ij} \geq 0\) and \((AA^T)_{ii} = 1\). Finally, two alternative norms, the weighted Frobenius norm or the Kullback-Leibler divergence are considered and numerical experiments based on the Frobenius norm are also presented in section 5.