A system identification framework for globally optimal models.

System identification constructs dynamical models from measured data. In a new unifying misfit versus latency (MvL) framework, we deal with inexact data by modifying the data (in least- squares sense) with misfits and adding unobservable, latent inputs. The MvL identification problem can be solved without any heuristics by means of a multiparameter eigenvalue problem (MEVP), resulting in globally optimal models. The main research objectives of this proposal are: deepening our mathematical understanding of the MvL framework and studying how the problem is related to (and may benefit from) operator theory, algebraic geometry and (behavioral) systems theory, extending the MvL approach to the multi-input multi-output (MIMO) case, and developing tailored numerical methods to perform the involved structured matrix computations.