Existence of the hyperbolic singular value decomposition (Q2365712)

Let \(A_ 1\) and \(A_ 2\) be rectangular complex matrices. In signal processing and other applications, it is desirable to find the eigendecomposition of \(X=A_ 1A_ 1^*-A_ 2A_ 2^*\) without explicitly forming the matrix \(X\). The authors observe that \(X\) can be written as \(APA^*\), where \(A=[A_ 1A_ 2]\) and \(P\) is a signature matrix (diagonal with diagonal entries \(\pm 1)\). They prove here that when \(A\) is an \(n\times m\) complex matrix and \(P\) is a signature matrix, there exists a factorization of \(A\) which easily gives the eigendecomposition of \(APA^*\). In an earlier paper they proved this under the assumption that \(APA^*\) has full rank. Specifically, they show here that there exist an \(n\times n\) unitary matrix \(U\), an \(m\times m\) matrix \(V\), an \(n\times m\) diagonal matrix \(D\), and another signature matrix \(\hat P\) such that \(A=UDV^*\) and \(V^*PV=\hat P\). If \(\text{rank}(APA^*)=\text{rank}(A)\) then \(D\) is real, otherwise some entries of \(D\) are not determined. The diagonal entries of \(D\) which are determined are real and nonnegative and the authors call these the hyperbolic singular values of \(A\).