Recursive least squares with stabilized inverse factorization (Q2639594)

A sequence of linear least squares problems \(\| A^{(k)}x- b^{(k)}\| \to \min.\) with \[ A^{(k)}=\left( \begin{matrix} \lambda A^{(k-1)}\\ a^{(k)T}\end{matrix} \right),\quad b^{(k)}=\left( \begin{matrix} \lambda b^{(k-1)}\\ \beta^{(k)}\end{matrix} \right) \] is considered. Such problems arise e.g. in signal processing since new observations have to be added continuously and older observations have to be deemphasized. The usual recursive-least-squares scheme is not suited for parallelization, while a parallel variant involving inverse factorization is shown in the paper to be numerically unstable. The authors develop a stabilized version using a Jacobi correction algorithm for inverting triangular matrices that is numerically stable and can be efficiently implemented on a systolic array; this implementation is also described in detail.