Identifiability Issues for Parameter-Varying and Multidimensional Linear Systems
Abstract This paper considers the identifiability of state space models for a system that is expressed as a linear fractional transformation (LFT): a constant matrix (containing identified parameters) in feedback with a finite-dimensional, block-diagonal (“structured”) linear operator. This model structure can represent linear time-invariant, linear parameter-varying, uncertain, and multidimensional systems. Families of input-output equivalent realizations are characterized as manifolds in the parameter space whose tangent spaces — and orthogonal complements — can be obtained via singular value decomposition. As illustrated by a numerical example, restricting iterative parameter estimation algorithms (e.g., maximum-likelihood with nonlinear programming) to the orthogonal directions offers significant computational advantages.