Comparison of order-recursive adaptive RLS algorithms

The finite word-length effects of the class of order-recursive least-squares algorithms are studied. The order-recursive algorithms can be modeled as arrays of interconnected cells or building blocks, which makes it possible to evaluate their numerical accuracy by studying the cell and then the combined roundoff noise effects in the array. The roundoff error of the local variable in each cell is given by a state-space model. The states of this error model are decoupled from each other: the roundoff error of a lower order residual (input) does not contribute significantly to the local state error, but only propagates through each building block to the higher order residual (output). This analysis is confirmed by numerical simulation.<>