On best rank \(n\) matrix approximations (Q448365): Difference between revisions

Let \(A=(a_{ij})\) be a real \(N\times M\) matrix. The author studies its approximation by rank~\(n\) matrices in the norm \[ |A|_{p,q}=\Big(\sum_{i=1}^N\Big(\sum_{j=1}^M|a_{ij}|^q\Big)^\frac{p}{q}\Big)^\frac{1}{p}, \] where \(1\leq p,q\leq\infty\). A similar question regarding best rank~\(n\) approximation to a totally positive kernel on \([a,b]\times [c,d]\) has been studied by \textit{C.~A.~Micchelli} and the author [J. Approx. Theory~24, 51--77 (1978; Zbl 0396.41014)]. Full analogy between these results and the present study holds only for \(p=q=1\). The author proves that if \(\mathrm{rank}\,A=M\), then there exists a best rank \(M-1\) approximation to~\(A\) in \(|.|_{1,1}\) which agrees with~\(A\) on \(M-1\) rows and \(M-1\) columns. He also proves that if \(A\) is strictly totally positive (i.e., all its minors are positive) and \(1\leq n<N,M\), then a best rank~\(n\) approximation to~\(A\) in \(|.|_{1,1}\) is given by a matrix which agrees with~\(A\) on \(n\) rows and \(n\) columns.