Kernel representation of Kalman observer and associated H-matrix based discretization

¹ Ecole Polytechnique (CMAP), Inria (Team Idefix), Institut Polytechnique, Paris Palaiseau 91128, France
² Inria – LMS, Ecole Polytechnique, CNRS, Institut Polytechnique de Paris Team MΞDISIM, Inria Saclay - Ile-de-France, Turing Building, Palaiseau 91128, France

^* Corresponding author: philippe.moireau@inria.fr

Received: 17 December 2021
Accepted: 30 October 2022

Abstract

In deterministic estimation, applying a Kalman filter to a dynamical model based on partial differential equations is theoretically seducing but solving the associated Riccati equation leads to a so-called curse of dimensionality for its numerical implementation. In this work, we propose to entirely revisit the theory of Kalman filters for parabolic problems where additional regularity results proves that the Riccati equation solution belongs to the class of Hilbert-Schmidt operators. The regularity of the associated kernel then allows to proceed to the numerical analysis of the Kalman full space-time discretization in adapted norms, hence justifying the implementation of the related Kalman filter numerical algorithm with H-matrices typically developed for integral equations discretization.