Deterministic Kalman filters for uncertain dynamical systems

Karl Kunisch ⋆,† and Jesper Schröder † · 2025

Read Deterministic Kalman filters for uncertain dynamical systems by Karl Kunisch ⋆,† and Jesper Schröder † (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-00463.txt.

Introduction

Kunisch and Schröder rebuild the Kalman filter on deterministic ground. The filter is the standard tool for tracking a noisy linear system from flawed measurements, usually justified by probability. Their version drops the statistics: dynamics and noise levels are uncertain but bounded in a deterministic sense, and the question becomes how wrong the estimate can be.

They derive error bounds tied to the size of those uncertainties, showing exactly how model doubt propagates into tracking error. Two worked examples with numerical implementations let readers compare the deterministic estimator against classical ones. For engineers who distrust tidy noise assumptions, the paper offers guarantees that do not lean on them.

Catalog record

CREATORKarl Kunisch ⋆,† and Jesper Schröder †
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE86 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-00463

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deterministic Kalman filters for uncertain dynamical systems Karl Kunisch ⋆,† and Jesper Schröder † Abstract . The Kalman(-Bucy) filter is the natural choice for the state reconstruction of disturbed, linear dynamical systems based on flawed and incomplete measurements. Taking a deterministic viewpoint this work investigates possible extensions of the concept to systems with uncertain dynamics and noise covariances. In a theoretical analysis error bounds in terms of the variance of the uncertainties are derived. The article concludes with a numerical implementation of two example systems allowing for a comparison of the estimators. Keywords: state estimation, filtering, parameter dependent systems AMS subject classification: 34F05, 34H05, 49N10, 93B53, 93C15 1. Introduction In this work we aim at reconstructing the state of a linear perturbed dynamical system of the form \displaystyle\dot{x}(t) \displaystyle=A_{\sigma}x(t)+Bv(t)\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), \displaystyle x(0) \displaystyle=x_{0}+\eta, where the disturbances in the dynamics and the initial state are given by v\in L^{2}(0,T;\mathbb{R}^{m}) and \eta\in\mathbb{R}^{n} , respectively. Additionally the system matrix A_{\sigma} depends on a parameter \sigma\in\Sigma representing uncertainties of the system dynamics. We assume access to a measurement of the system given as y(t)=Cx(t)+\mu(t)\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), affected by the disturbance \mu\in L^{2}(0,T;\mathbb{R}^{r}) . The disturbances are assumed to be unknown and of deterministic nature. They are associated with parameter dependent positive definite matrices \Gamma_{\sigma}\in\mathbb{R}^{n,n} , R_{\sigma}\in\mathbb{R}^{m,m} , and Q_{\sigma}\in\mathbb{R}^{r,r} representing their magnitudes. Reconstructing a system’s state based on incomplete and disturbed measurements has a rich history going back at least to Wiener’s seminal work [ 22 ] , where disturbances are modeled as random processes. The pioneering works of Kalman and Bucy introduced the highly celebrated Kalman(-Bucy) filter for time discrete [ 9 ] and for time continuous [ 10 ] linear systems with Gaussian noise. Due to its effectiveness it is widely applicable in practice and it inspired numerous variations to address particularly challenging systems such as the extended Kalman filter for nonlinear dynamical systems and the ensemble Kalman filter [ 6 ] for systems of very large dimension. The assumption of normally distributed noise, however, implies state independent disturbances. Hence exact knowledge of the system dynamics is required and modeling uncertainties and errors introduced by, e.g., material dependent parameters, moving sensors, or model reduction, are not addressed explicitly. Similarly, the Kalman filter relies on exact knowledge on the distribution of the errors, i.e., their respective covariances which, in practice, are known only approximately. In [ 17 ] , see also the monograph [ 18 ] , the authors consider time continuous, quadratically stable dynamics subject to bounded uncertainties. They design an estimator based on solutions of associated Riccati equations and analyze the asymptotic behavior with respect to time. It is shown that the expected difference of estimator and disturbed state converges to a neighborhood of zero. For a discussion of time discrete systems affected by noise with uncertain covariances we refer to [ 14 ] and the references therein, see also [ 20 ] for a more recent analysis. The remainder of the article is organized as follows. Section 2 gives an overview on the Kalman filter in the stochastic as well as in the deterministic setting. We compare the two formulations, point out connections, and set up the framework for our investigation. In Section 3 the uncertainties are formulated mathematically and three designs for Kalman filters for uncertain systems are proposed and characterized. In part we shall focus on the estimator which arises as the weighted average of the individual Kalman filters and minimizes the average energy. Section 4 contains the technical proof that the distances of the proposed estimators to the Kalman filter associated with the hidden parameter can be estimated from above in terms of the level of variance in the uncertain matrices. The article concludes with a numerical investigation of the designs at the hand of two examples in Section 5 . Throughout the article we use the following notation. We say that a multivariate random variable X is distributed according to \mathcal{N}(m,\Sigma) , where m\in\mathbb{R}^{d} and \Sigma\in\mathbb{R}^{d,d} is symmetric and positive definite, if X is normally distributed with mean m and covariance matrix \Sigma . In short we write X\sim\mathcal{N}(m,\Sigma) . For a symmetric, positive definite matrix P\in\mathbb{R}^{d,d} the associated norm of x\in\mathbb{R}^{d} is defined via \|x\|_{M}^{2}=x^{\top}Mx . The cardinality of a finite set \mathcal{M} is denoted by |\mathcal{M}| . Finally, we abbreviate some common L^{p} spaces, Sobolev spaces, and spaces of continuous functions equipped with their default norms as follows. We set L^{2}(t_{1},t_{2};\mathbb{R}^{d})=\mathcal{L}_{t_{1}}^{t_{2}} , H^{1}(t_{1},t_{2};\mathbb{R}^{d})=\mathcal{H}_{t_{1}}^{t_{2}} , C([t_{1};t_{2}];\mathbb{R}^{d})=\mathcal{C}_{t_{1}}^{t_{2}} , and C([0;t];\mathbb{R}^{d,d})=\mathcal{C}_{t}^{d,d} . For the spaces of vector valued function the dimension d is suppressed in the notation and becomes clear from context. 2. The Kalman filter We commence with a brief presentation of two different formulations of the Kalman filter. Even though our proposed Kalman filter for uncertain systems relies only on the deterministic formulation we include a recap of the stochastic formulation enriching the interpretation of our approach. For ease of presentation we restrict ourselves to time-invariant systems. While the concepts and results presented in this section are well known in the literature we include them for the purpose of a self-contained work. 2.1. The stochastic formulation We illustrate the stochastic filter based on the original work [ 10 ] . Even though this formulation of the model lacks some mathematical rigor, the formal representation suffices to illustrate the underlying concept. For a more thorough treatment based on Itô integrals we refer to [ 16 , 24 ] . Consider the disturbed system with system state X_{t} and measured output Y_{t} modeled by (2.1) \displaystyle\tfrac{\mathrm{d}}{\mathrm{d}t}X_{t} \displaystyle=AX_{t}+BV_{t}\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), \displaystyle Y_{t} \displaystyle=CX_{t}+W_{t},\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), where A\in\mathbb{R}^{n,n} describes the linear dynamics of the system, C\in\mathbb{R}^{r,n} describes the incomplete measurements taken from the state and B\in\mathbb{R}^{n,m} encodes how the noise enters the dynamics. The white noise affecting the dynamics and the measurement is modeled by the multivariate Gaussian random variables V_{t}\in\mathbb{R}^{m} and W_{t}\in\mathbb{R}^{r} , respectively. They have zero mean and covariances given by \displaystyle\text{cov}[V_{t},V_{s}] \displaystyle=R\,\delta(t-s), \displaystyle\text{cov}[W_{t},W_{s}] \displaystyle=Q\,\delta(t-s), with given symmetric, positive definite matrices R\in\mathbb{R}^{m,m} and Q\in\mathbb{R}^{r,r} . The Gaussian distribution of the state at time t is described by X_{t} with initial Gaussian distribution X_{0} with mean x_{0}\in\mathbb{R}^{n} and covariance \Gamma\in\mathbb{R}^{n,n} , symmetric, positive definite. Finally, V_{t} , W_{t} , and X_{0} are assumed to be mutually independent and Y_{t} encodes the distribution of the output of the system at time t . The objective, as stated in [ 10 ] , is to optimally estimate the realized state x(t) , of the system based on realized, measured outputs y(s) , 0\leq s\leq t . Kalman and Bucy find that the optimal estimate \widehat{x}(t) is characterized via (2.2) \displaystyle\dot{\widehat{x}}(t) \displaystyle=A\widehat{x}(t)+\Pi(t)C^{\top}Q^{-1}(y(t)-C\widehat{x}(t))\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), \displaystyle\widehat{x}(0) \displaystyle=x_{0}, where \Pi is the unique solution to the differential Riccati equation (2.3) \displaystyle\dot{\Pi}(t) \displaystyle=A\Pi(t)+\Pi(t)A^{\top}-\Pi(t)C^{\top}Q^{-1}C\Pi(t)+BRB^{\top}\penalty\ \penalty\ \penalty\ \penalty\ \penalty\ t\in(0,T), \displaystyle\Pi(0) \displaystyle=\Gamma. These formulas have since been confirmed by the stochastic community using rigorous mathematical concepts, see e.g. [ 16 , 24 ] . Interpreting ( 2.1 ) as stochastic differential equations with noise terms defined by Brownian motions and solutions defined via the Itô integral one characterizes the conditional expectation and covariance. More precisely, given a measured output the

Read the complete text