A geometric perspective of state estimation using Kalman filters
Mateusz Baran, Ronny Bergmann · 2025
Read A geometric perspective of state estimation using Kalman filters by Mateusz Baran, Ronny Bergmann (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-2506-01086.txt.
Introduction
Mateusz Baran and Ronny Bergmann give Kalman filters a new geometric foundation in this 2025 arXiv paper. Kalman filters estimate a moving system's state from noisy readings, and the shape of the state space matters greatly in robotics and motion tracking. Two schools handle that shape differently, one through Lie groups and one through curved-space geometry.
The authors show how the two schools relate, then build a description based on affine connections that covers both. That wider view gives users more freedom in choosing the state space shape, and it still works with standard tools like parameter fitting. They test the idea on a rigid-motion problem and on tracking a point on a sphere, using a software library they built for the purpose.
Catalog record
| CREATOR | Mateusz Baran, Ronny Bergmann |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-01086
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A geometric perspective of state estimation using Kalman filters Mateusz Baran, Ronny Bergmann Abstract Geometry of the state space is known to play a crucial role in many applications of Kalman filters, especially robotics and motion tracking. The Lie group-centric approach is currently very common, although a Riemannian approach has also been developed. In this work we explore the relationship between these two approaches and develop a novel description of Kalman filters based on affine connections that generalizes both commonly encountered descriptions. We illustrate the results on two test problems involving the special Euclidean group and the tangent bundle of a sphere in which the state is tracked by geometric variants of the extended Kalman filter and the unscented Kalman filter. The examples use a newly developed library GeometricKalman.jl. The new approach provides a greater freedom in selecting the structure of the state space for state estimation and can be easily integrated with standard techniques such as parameter estimation or covariance matrix estimation. I Introduction Over the decades numerous different variants of the Kalman filter have been developed [ 1 , 2 , 3 , 4 ] . Many of them find applications in robotics or navigation, where the geometry of the state space is known to play a crucial role [ 5 ] . The core subject of our study are dynamical systems affected by noise. We would like to estimate parameters of state of such systems based on our knowledge of its dynamics and some measurements. In the continuous and deterministic setting it can be formalized as tracking the state of a system p(t) which evolves in the configuration space \mathcal{M} . The initial state is p_{0}=p(0)\in\mathcal{M} . This evolution is described by a differential equation \frac{\mathrm{d}}{\mathrm{d}t}{p(t)}=f(p(t),q(t),t) (1) for some function f . Its first argument, p(t) , is the state at time t\in[0,T] , q(t) describe control parameters and T is the upper bound on simulation time. An example of such system can be found in [ 6 ] . Kalman filters allow us to estimate the value of p in situations where there is significant noise, f might be a very rough approximation to the true dynamics and we have some observations to help us track the state. In this work we describe a novel and generalized presentation of Kalman filters. In this framework we explain how similarities of multiple existing approaches like error-state Kalman filters, invariant filter and optimization-based filters can be emphasized and used to develop new variants. We explain theoretical foundations in a way that should be approachable by practitioners without omitting important parts of the theory. We attempted to collect the most prominent variants of Kalman filters and present them in a unified geometric framework. Lie groups are commonly used in robotics and aeronautics [ 1 ] as well as in control theory [ 7 ] to represent configuration spaces. A recent literature review can be found in [ 8 ] . Riemannian manifolds have been used for example for articulated tracking [ 9 ] and more generally in computer vision [ 10 ] . We generalize both approaches to arbitrary manifolds with an affine connection. We demonstrate our approach by generic implementation of selected variants GeometricKalman.jl (available at https://github.com/JuliaManifolds/GeometricKalman.jl ) based on one of the Julia packages for numerical differential geometry, Manifolds.jl [ 11 ] and Riemannian optimization, Manopt.jl [ 12 ] . This work was enabled by recent developments in applications of differential geometry, such as computational anatomy [ 13 , 14 ] , shape analysis [ 15 , 16 , 17 ] or medical image analysis [ 18 , 19 ] . Other notable achievements include the Geomstats library [ 20 ] , which implements selected variants of Kalman filters on manifolds. As a guiding example we consider a segmented rigid body. Its state is described by the position and orientation of the first segment as well as pairs of angles between two consecutive segments. Such model is widely considered in robotics [ 21 ] and modelling of human and animal skeletons, especially the hip and shoulder joints [ 22 ] . The natural geometry of those pairs of angles is that of a sphere [ 23 ] , which cannot be handled using the standard approach of Lie group representation [ 24 ] . Following contributions are made in this work. First, we demonstrate that the theory of connection manifolds is more fundamental to Kalman filters than Lie theory. Next, we provide a systematic review of different variants of Kalman filters in the developed framework. A new library using this approach was written. Finally, we also show some paths for future developments this approach opens. II Theoretical foundations II-A Basic concepts in differential geometry and Lie groups Manifolds with an affine connection [ 13 , 25 ] are a commonly encountered basic differential geometry objects useful in applications. An d -dimensional connection manifold \mathcal{M} is characterized by an atlas of charts \varphi_{i}\colon U_{i}\to\mathbb{R}^{d} , i\in\mathcal{I} , where \mathcal{I} is some index set, U_{i}\subseteq\mathcal{M} is open, and \varphi_{i} is a smooth bijection. The union of all U_{i} , i\in\mathcal{I} is \mathcal{M} . At each point we have a tangent space T_{p}\mathcal{M} , which is a vector space defined as follows: In a chart (\varphi,U) around p\in U consider curve \gamma\colon I\to U\subset\mathcal{M} , where I an interval containing 0 , such that \gamma(0)=p . We call the equivalence class X=[\frac{\mathrm{d}}{\mathrm{d}t}{\varphi\circ\gamma}]\in T_{p}\mathcal{M} of curves \delta through \delta(0) having the same derivative with respect to \varphi as \gamma a tangent vector . This definition is indeed independent of the chart \varphi , see also [ 26 , p. 33] . We represent X also in coordinates with respect to the chart \varphi , by writing the curve \gamma as \gamma(t)=\varphi^{-1}\bigl(\varphi(p)+tc\bigr),\qquad\text{ for some }c\in\mathbb{R}^{d} Since \varphi is a chart, we can identify X with these coordinates c\in\mathbb{R}^{d} and even more also p with its coordinates \varphi(p) . For both coordinate representations we write X_{c} and p_{c} , respectively, for short. Similarly we can represent X also with respect to a basis B_{p}=\{e_{1},e_{2},\dots,e_{d}\} of by coefficients c_{B_{p}}(X) , that is X=\sum_{i=1}^{d}c_{B_{p}}(X)_{i}e_{i} . We denote by T\mathcal{M} the disjoint union of all tangent spaces T\mathcal{M}=\{(p,T_{p}\mathcal{M})\mid p\in\mathcal{M}\} , called the tangent bundle . Similar to a point in coordinates, we can also consider the curve \gamma_{c}(t)\coloneqq\varphi(\gamma(t)) in coordinates of the chart as well as its derivative \frac{\mathrm{d}}{\mathrm{d}t}{\gamma_{c}} in the same way as a tangent vector. More generally we consider vector fields along a curve \Xi(t)\in T_{\gamma(t)}\mathcal{M} , which we can again also write in coordinates denoted by \Xi_{c}(t)\in\mathbb{R}^{d} . Additionally we assume that \mathcal{M} is equipped with an affine connection, which is a method of relating tangent spaces at different points. The relation is described by the Christoffel symbol \Gamma_{i}\colon\mathbb{R}^{d}\times\mathbb{R}^{d}\times\mathbb{R}^{d}\to\mathbb{R}^{d} [ 27 , Def. 15.2] which depends on the chart \varphi as well. A vector Y\in T_{p}\mathcal{M} can undergo parallel transport in direction X\in T_{p}\mathcal{M} by solving the following ODE in a chart \varphi around p : Find the curve \gamma and the vector fields \Xi,\Psi along \gamma such that \frac{\mathrm{d}}{\mathrm{d}t}{\begin{pmatrix}\gamma_{c}(t)\\ \Xi_{c}(t)\\ \Psi_{c}(t)\end{pmatrix}}=\begin{pmatrix}\Xi_{c}(t)\\ -\Gamma(\gamma_{c}(t),\Xi_{c}(t),\Psi_{c}(t))\\ -\Gamma(\gamma_{c}(t),\Xi_{c}(t),\Psi_{c}(t)),\end{pmatrix}, (2) With the initial conditions \gamma_{c}(0)=p_{c} , \Xi_{c}(0)=X_{c} , and \Psi_{c}(0)=Y_{c} . Then, the parallel transport of Y in direction