On the nonnegativity of \(XX^+\) and its relevance in econometrics (Q762872)

Let X be an \((n\times p)\) matrix of rank \(m>0\) with \(r\leq n\) ''non- proportional'' rows and let the (r\(\times p)\) matrix B contain these rows. Then \(X=AB\) with A having exactly one nonzero element in each row. A necessary and sufficient condition that \(XX^+\) has only nonnegative elements, is that \(m=r\) and that each column of A is either nonnegative or nonpositive. \((X^+\) denotes the Moore-Penrose inverse of the matrix X.) The relevance of this result regarding least squares estimation of the systematic part of linear models is discussed. Extensions are given for weighted least squares estimation in case of heteroskedasticity.