Procrustes problems in finite dimensional indefinite scalar product spaces (Q556872): Difference between revisions

Let \({\mathbb F}^n\) be an \(n\)-dimensional vector space, where \(\mathbb{F}\) is either \(\mathbb R\) or \(\mathbb C\). For \(x,y\in{\mathbb F}^n\), let \((x,y)\) be the usual inner product. If \(H\) is a nonsingular Hermitian matrix in \({\mathbb F}^{n\times n}\), then \([x,y]=(Hx,y)\) defines an indefinite scalar product for \({\mathbb F}^n\). A factorization \(A=UM\) is called an \(H\)-polar decomposition, if \(U\) is invertible, \(U^\star HU=H\), and \(M^\star H=HM\), where \(U^\star\) and \(M^\star\) are the adjoints of \(U\) and \(M\), respectively. If a decomposition \(A=UM\) is \(H\)-polar and \(G\)-polar, then it is called a \((G,H)\)-polar decomposition. Let \(X\) and \(Y\) be two vectors constructed from two measurements of a certain quadratic form. The task of the Procrustes problem is to determine some \(U\in{\mathbb F}^{n\times n}\) with \(U^\star HU=H\) that optimizes \(f(U)=\text{ tr}\,[(UX-Y)^\star H(UX-Y)]\). The author shows that there is such a \(U\) if and only if the matrix \(A=YX^\star H\) admits a semidefinite \(H\)-polar decomposition, \(A=UM\). The author also finds conditions for the solution of a similar problem, namely the \((G,H)\)-Procrustes problem. Here \(U^\star HU=H\), \(U^\star GU=G\), and \(H^{-1}G=\mu^2G^{-1}H\) for a certain \(\mu\in{\mathbb R}\) is required. The last section of the paper is devoted to the numerical computation of \(H\)-polar and \((G,H)\)-polar decompositions.