Minimax State Estimation for a Dynamic System Described by a Differential-Algebraic Equation
In this report we address the linear state estimation problem: to estimate a linear transformation $\ell(φ)$ of the state $φ$ through an algorithm $\widehat{\ell(φ)}$ operating on measurements $y$, where $Lφ=f,y=Hφ+η$. We study the estimation problem in terms of the minimax estimation framework: to find a linear algorithm $\widehat{\widehat{\ell(φ)}}$ that minimizes the worst case error $\sup_{φ,η}d(\ell(φ),\widehat{\ell(φ)}) $. A key feature of the presented estimation approach is to fix a class of linear operators $L$, $H$; given any pair $L,H$ from that class we describe a class $\mathcal L$ of all solution operators $\ell$ such that the worst case error is finite. We formulate a duality theorem (like Kalman duality principle) that is the estimation problem is equal to the optimal control problem if $G$ is convex bounded subset of the corresponding Hilbert space, $L$ is a closed linear mapping. We obtain optimal estimations as solutions of the linear operator equations if $G$ is an ellipsoid. Then we apply this to the state estimation for the linear differential-algebraic equations (DAE). The minimax observer for DAE is represented in the form of the minimax filter. For discrete time DAEs we present the online minimax estimator.