Orthogonality of matrices (Q5915532): Difference between revisions

Let \(\|\;\|\) denote any norm on the space of \(m\times n\) matrices over \(\mathbb{R}\) or \(\mathbb{C}\) (where for convenience \(m\leq n\)). Two \(m\times n\) matrices \(A\) and \(B\) are called orthogonal (in the Birkhoff-James sense) if \(\|A+\mu B\|\geq\|A\|\) holds for every scalar \(\mu\); in other words, \(0\) is the best approximation to \(A\) from the space of matrices spanned by \(B\). If \(s_{1}(A)\geq s_{2} (A)\geq\dots\geq s_{m}(A)\) are the singular values of \(A\), then the Schatten \(p\)-norm \(S_{p}(A)\) is defined to be \(\{ \sum_{i}s_{i}(A)^{p}\} ^{1/p}\) if \(1\leq p<\infty\) and \(s_{1}(A)\) if \(p=\infty\). The authors characterize the pairs of matrices which are orthogonal under the various Schatten \(p\)-norms. For example, for \(p=1\), the condition (i) \(A\) is orthogonal to \(B\) is equivalent to each of the conditions: (ii) for some \(m\times n\) matrix \(F\) with \(S_{\infty}(F)=1\) we have \(\text{tr}(AF^{\ast})=S_{1}(A)\) and \(\text{tr}(BF^{\ast})=0\); and (iii) if \(A\) has rank \(k\) and the singular value decomposition \(A=U^{\ast}SV\), then \[ UBV^{\ast}= \begin{pmatrix} B_{11} & B_{12}\cr B_{21} & B_{22} \end{pmatrix} \text{ with }|\text{tr}(B_{11})|\leq S_{1}(B_{22}) \] where \(B_{11}\) is a \(k\times k\) submatrix. This extends earlier results of \textit{R. Bhatia} and \textit{P. Semrl} [ibid. 287, No. 1-3, 77-85 (1999; Zbl 0937.15023)] and of \textit{F. Kittaneh} [ibid. 151, 119-124 (1991; Zbl 0724.15015)]. The authors also give a counterexample to a conjecture in the former paper concerning orthogonality under a matrix norm \(\|.\|\) induced from a vector norm.