Algorithms for inversion of diagonal plus semiseparable operator matrices (Q1865921)

Consider a Hilbert space \(H=H_1\oplus\cdots\oplus H_N\), and an operator equation in \(H\): \(Rx=y\). Conformal the direct sum, suppose \(R\) is represented by a matrix with entries \(R_{ij}\) with \(R_{ij}=g_ih_j\) for \(j>i\) and \(R_{ij}=p_iq_j\) for \(i>j\), where \(p_i,g_i:{\mathbb C}\to H_i\) and \(q_j,h_j: H_j\to{\mathbb C}\) are linear bounded operators. Three fast algorithms are proposed for inverting \(R\), and hence for solving \(Rx=y\). The first algorithm assumes strong regularity and inverts successive leading submatrices. The second is a doubling algorithm in the style of \textit{L. Greengard} and \textit{V. Rokhlin} [Commun. Pure Appl. Math. 44, 419-452 (1991; Zbl 0727.65068)], based on a divide and conquer technique using \(H=\oplus_i H_i^{(1)}\) and subsequently \(H_i^{(p)}=\oplus_j H_j^{(p+1)}\). The third method reduces the problem to a system of linear difference equations. These algorithms are then applied to integral equations (differential equations are reduced to integral equations first). The \(H_i\) result from splitting up the integration interval into subintervals (see Greengard and Rokhlin, loc. cit.) For the numerical computation of the integrals, Chebyshev quadrature formulas are used. Numerical results illustrate the performance of the methods.