Where is matrix multiplication locally open? (Q503425): Difference between revisions

A map \(f\) from a metric space \(M_1\) into a metric space \(M_2\) is locally open at \(x\in M_1\) if for every \(\varepsilon>0\) there is a \(\delta>0\) such that \(B(f(x),\delta)\subset f(B(x,\varepsilon))\). In particular, one can study the problem of when composition of linear operators (``matrix multiplication'' in the finite-dimensional case) is locally open. More precisely, let \(X, Y, Z\) be normed spaces and \(L(X,Y), L(Y,Z)\) and \(L(X,Z)\) be the corresponding spaces of operators, equipped with their operator norms. Let \(\Phi\) be defined on \(L(Y,Z)\times L(X,Y)\to L(X,Z)\) by \(\Phi(S,T)=S\circ T\). The problem the author studies is: At what points \((S_0,T_0)\) is \(\Phi\) locally open? The author succeeds in giving a complete answer in the finite-dimensional case (both real and complex) in terms of dimensions related to the operators (matrices) and spaces involved. The characterization is a bit involved, but goes like this: Matrix multiplication is open at \((S_0,T_0)\) if and only if \[ \dim(\operatorname{range}T_0\cap\operatorname{ker}S_0)\leq \dim Y - \dim Z \] or \[ \dim X - \dim (\operatorname{range}T_0\cap(\operatorname{ker}S_0)^\perp)\leq \dim Y - \dim \operatorname{range} S_0. \] Despite this positive result, it is interesting to read the authors reflections at the end of the paper: ``The formulation and the proof of our main theorem was strongly dependent on the fact that we are dealing with finite-dimensional spaces\(\dots\) Part of our results have a counterpart for continuous operators on separable Hilbert spaces, but we are still far from an understanding how an interplay of the ranges, the kernels and the spectra gives rise to local openness.''