Cyclic Seesaw Process for Optimization and Identification

摘要

A known approach to optimization is the cyclic (or alternating or block coordinate) method, where the full parameter vector is divided into two or more subvectors and the process proceeds by sequentially optimizing each of the subvectors, while holding the remaining parameters at their most recent values. One advantage of such a scheme is the preservation of potentially large investments in software, while allowing for an extension of capability to include new parameters for estimation. A specific case of interest involves cross-sectional data that is modeled in state–space form, where there is interest in estimating the mean vector and covariance matrix of the initial state vector as well as certain parameters associated with the dynamics of the underlying differential equations (e.g., power spectral density parameters). This paper shows that, under reasonable conditions, the cyclic scheme leads to parameter estimates that converge to the optimal joint value for the full vector of unknown parameters. Convergence conditions here differ from others in the literature. Further, relative to standard search methods on the full vector, numerical results here suggest a more general property of faster convergence for seesaw as a consequence of the more “aggressive” (larger) gain coefficient (step size) possible.