Skip to Main Content
Subset models of autoregressive reflection coefficients have a number of non-zero reflections in the subset while the other reflections are zero. Speech analysis and seismic exploration are areas of applications where true subset processes may be found. The estimation of subsets of reflection coefficients differs completely from any linear subset selection procedure. This paper presents three selection strategies with an increasing computational effort. An efficient algorithm can use all three strategies, but picks out the cheapest solution for the occasion. All strategies use the Weak Parameter Criterion WPC for the selection. One reason for the selection of subset models can be an adequate representation of true processes. On the other hand, selected subsets below the WPC-order give often a smaller prediction error and a closer approximation of the power spectral density than selected complete models.