Estimation of parameters for Hilbert space-valued partially observable stochastic processes (Q1111295)

The author considers a discrete-time, partially observed linear system in which the state and observation processes take values in real separable Hilbert spaces, and the system and observation noises are sequences of mutually independent i.i.d. random elements with zero means. The drift coefficient is assumed to be of the form \(A(\theta)=A_ 0+\sum^{m}_{1}\theta_ iA_ i\), where the \(A_ i\) are given bounded linear operators and \(\theta =(\theta_ 1,...,\theta_ m)\) is an unknown parameter vector. The problem considered in the paper is to estimate \(\theta\) in terms of the observation process. Under suitable stability and identifiability assumptions, similar to those for finite-dimensional systems [see, e.g., \textit{L. Ljung}, Stoch. Syst.: Model., Identif., Optim. I; Math. Program. Stud. No.5, 169-190 (1976; Zbl 0369.93038)], the author establishes the strong consistency and asymptotic normality of a class of modified least squares estimators, which, in the finite-dimensional case, is a particular class of the so- called prediction error estimators [cf. Ljung, op. cit.]. The proofs use the strong law of large numbers for martingales and ergodic sequences, and a central limit theorem for q-dependent stationary processes. The author argues for the importance of considering Hilbert-space-valued processes, which allows one to treat, e.g., distributed parameter systems.