Performance bounds for recognition of jump-linear systems

Multiple model algorithms have been used extensively to model jump-linear systems. Although the optimal solution to systems subject to abrupt parameter changes is well known, its calculation is impractical and many suboptimal approaches have been proposed. In this paper, the authors investigate the performance of a broad class of systems approximated by generalized pseudo-Bayesian (GPB) algorithms and the interacting multiple model (IMM) algorithm by bounding the L₁ distance of a suboptimal prediction density from the truth. The relation of this L₁ bound to the probability of error in the system identification problem is discussed