Title:Riemannian optimal system identification method of linear continuous-time systems with symmetry

Abstract:This paper develops identification methods of linear continuous-time systems with symmetry. To this end, we formulate a novel prediction error problem of the corresponding discrete-time systems. More specifically, the problem is the minimization problem of the squared sum of the differences between the true and model outputs on the product manifold of the manifold of the symmetric positive definite matrices and two Euclidean spaces. We also formulate the minimization problem on the quotient manifold of the product manifold under a specified group action by the orthogonal group, in order to reduce the search dimension of the former problem. We propose Riemannian conjugate gradient (CG) methods for both problems. It is shown that the only difference between the methods is vector transports in the algorithms. The vector transport for the latter problem is the projection onto the horizontal space which is the subspace of a tangent space of the product manifold. The projection is obtained by using the skew-symmetric solution to a linear matrix equation. We prove that there exists a unique skew-symmetric solution to the equation under a mild assumption. Furthermore, we provide two proofs that a Riemanninan metric on the quotient manifold can be endowed by using the metric on the product manifold. We suggest that we choose initial points in the proposed methods by using the result of the modified ordinary Multivariable Output-Error State-sPace (MOESP) method. The initial point is considerably better than randomly choosing a point on the product manifold. We demonstrate that the CG method for the latter problem is the best among the other methods such as the ordinary MOESP method, N4SID method, the proposed modified MOESP method, the steepest descent method for both problems, and the CG method for the former problem.