An alternating maximization method for approximating the hump of the matrix exponential (Q2411640)

If \(A\in\mathbb{C}^{n\times n}\) is stable (eigenvalues in the left-half plane), then the curve \(\Gamma(t)=\|e^{tA}\|_2\), \(t\geq0\) has humps where local maxima \(\Gamma_h=\Gamma(t_h)\) are attained. Define \(\gamma(t,v)=\|e^{tA}v\|_2\), with \(t\geq0\) and \(v\in\mathbb{C}^n\), \(\|v\|_2=1\), then \(\max_{\|v\|_2=1}\gamma(t_h,v)=\gamma(t_h,v_h)=\Gamma_h=\max_{t\in I}\gamma(t,v_h)\) where \(I\) is an interval containing \(t_h\). Note that \(v_h\) is a right singular vector of \(e^{t_hA}\) with singular value \(\Gamma_h\). This is used to compute a (local) hump iteratively by alternatingly computing the maximum of \(\gamma(t,v)\) over \(v\) for \(t\) fixed, and use the result to fix \(v\) and compute a maximum over \(t\). A convergence analysis is given and it is proved that the method converges to the local maximum or alternates between accumulation points where \(\gamma(t,v)\) is not maximal. Implementation aspects (e.g. finding a starting vector for \(v\) and a stopping criterion) are discussed and the method is extensively tested numerically.