Remarks on completey positive matrices (Q4253152)

A real \(n\times n\) symmetric matrix \(A\) which can be written in the form \(A=BB^{\top}=\sum b_{i}b_{i}^{\top}\) for some \(n\times k\) matrix \(B\) with non-negative entries and columns \(b_{1},b_{2},\dots,b_{k}\) is called a completely positive matrix. Clearly \(B\) is not uniquely determined by \(A\), and the question arises as to how sparse \(B\) may be chosen for a given completely positive matrix. If each column \(b_{i}\) of \(B\) has at most \(t\) nonzero entries, then the representation \(A=BB^{\top}\) is called a \(k\)-support \(t\) rank \(1\) representation. For example, it is known that an \(n\times n\) diagonally dominant symmetric matrix with non-negative entries is completely positive and has a \(k\)-support \(2\) rank \(1\) representation for some \(k\) [cf. \textit{J. H. Drew, C. R. Johnson} and \textit{R. Loewy}, Linear Multilinear Algebra 37, No. 4, 303-310 (1994; Zbl 0815.15019)] The main theorem proved here is the following (Theorem 2.1). If \(A\) is an \(n\times n\) diagonally dominant symmetric matrix with non-negative entries\ and \(n\geq 4,\) then \(A\) has a \(\lfloor n^{2}/4\rfloor\)-support \(3\) rank \(1\) representation. The paper also includes some related results.