Averaging operators for exponential splittings (Q2373397)

Given a matrix \(M\in{\mathbb C}^{n\times n}\), the author considers the averaging of \(M\) defined as \[ A=(1/k)\sum_{j=0}^{k-1} J^jMJ^{-j}, \] where \(J\in{\mathbb C}^{n\times n}\) is a unipotent matrix, that is, \(J^k=I\) for some smallest \(k\in{\mathbb N}\). The two most central examples of unitary unipotents with \(k=n\) are (1) if \(J_c\) is a unitary circulant matrix with ones on the first sub-diagonal and at position \((1,n)\), then \(A\) is circulant, and (2) if \(J_d=e^D\), where \(D=(2\pi i)/n\, {\text{diag}}\,(0,1,2,\ldots,n-1)\), then \(A={\text{diag}}\,M\). With appropriate splittings, a resplitting of Toeplitz-plus-Hankel (T+H) matrices is constructed. This approach supports a general hierarchy of structured matrices which is then described and used to apply exponential splitting methods with T+H matrices. In applications, small rank perturbations, which typically can be expected to result from boundary conditions, are allowed. Diagonal perturbations can also be handled. Based on a formula due to \textit{A. Lenard} [J. Funct. Anal. 10, 410--423 (1972; Zbl 0237.47002)], a very fast iterative method for finding the nearest T+H matrix in the Frobenius norm is devised.