The difference between \(5\times 5\) doubly nonnegative and completely positive matrices (Q840651): Difference between revisions

A matrix \(X\) is called \textit{completely positive} if it is a Hermitian square of a nonnegative matrix, i.e., if \(X=N^tN\) for some (entrywise) nonnegative matrix \(N\). Let \(C_n\) denote the convex cone of all \(n\times n\) completely positive matrices. Likewise, let \(D_n\) be the convex cone of all nonnegative and positive semidefinite \(n\times n\) matrices. Obviously, \(C_n\subseteq D_n\), and it is well-known that equality holds iff \(n\leq 4\). In a dual viewpoint, \(C_n^*\) is the convex cone of all \(n\times n\) copositive matrices, and \(D_n^*\) is the convex cone of all symmetric \(n\times n\) matrices that can be written as a sum of a nonnegative and a positive semidefinite matrix. Horn was the first to construct a matrix \(H\in C_n^*\setminus D_n^*\), cf.~\textit{M.~Hall} and \textit{M.~Newman} [Proc. Camb. Philos. Soc. 59, 329--339 (1963; Zbl 0124.25302)]. Let us call a matrix \(X\in D_n\setminus C_n\) exceptional. The authors of the paper under review study these with an emphasis on the \(5\times 5\) case. An extreme ray \(X\in D_5\) can have rank \(1\) (in which case \(X\in C_5\)) or \(3\) (in which case \(X\) is exceptional). A characterization of all matrices of the later type is given in Theorem 3. In Section 5 a separating hyperplane is constructed for every matrix of this type. Also, the authors present an algorithm to decompose \(X\in D_n\) into \(X=Y+Z\) for \(Y\in C_n\) and \(Z\in D_n\), where \(Z\) is CP-irreducible (that is, it does not admit a nontrivial decomposition of this form). As a final result let us mention that \(X\in D_5\) is CP-irreducible iff it is exceptional and an extreme ray of \(D_5\).