Model order reduction via Lie groups
Yannik Wotte, Patrick Buchfink, Silke Glas · 2025
Read Model order reduction via Lie groups by Yannik Wotte, Patrick Buchfink, Silke Glas (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-2511-03520.txt.
Introduction
Yannik P Wotte and colleagues build MORLie, a framework that shrinks large physical simulations by respecting their Lie group geometry. Many systems live on curved manifolds where rotations and motions act naturally, yet classic reduction flattens that structure into linear subspaces. MORLie instead approximates the full system with a smaller system that still lives on a Lie group.
Unlike older group methods, the framework handles dynamics that lack symmetry, a common practical case. Non-intrusive variants learn the small model from data alone. A key bound shows the method can beat the classic width limit that caps all linear approaches. Tests include a deforming body tracked through noisy point clouds.
Catalog record
| CREATOR | Yannik Wotte, Patrick Buchfink, Silke Glas |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 80 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03520
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Model order reduction via Lie groups Yannik Wotte, Patrick Buchfink, Silke Glas Abstract Lie groups and their actions are ubiquitous in the description of physical systems, and we explore implications in the setting of model order reduction (MOR). We present a novel framework of MOR via Lie groups, called MORLie, in which high-dimensional dynamical systems on manifolds are approximated by low-dimensional dynamical systems on Lie groups. In comparison to other Lie group methods we are able to attack non-equivariant dynamics, which are frequent in practical applications, and we provide new non-intrusive MOR methods based on the presented geometric formulation. We also highlight numerically that MORLie has a lower error bound than the Kolmogorov N N -width, which limits linear-subspace methods. The method is applied to various examples: 1. MOR of a simplified deforming body modeled by a noisy point cloud data following a sheering motion, where MORLie outperforms a naive POD approach in terms of accuracy and dimensionality reduction. 2. Reconstructing liver motion during respiration with data from edge detection in ultrasound scans, where MORLie reaches performance approaching the state of the art, while reducing the training time from hours on a computing cluster to minutes on a mobile workstation. 3. An analytic example showing that the method of freezing (a previous Lie group method for MOR introduced in [ 1 ] ) is analytically recovered as a special case, showing the generality of the geometric framework. 1 Introduction Figure 1 : Summary of MORLie: Given a full order model (FOM) based on a vector field X\in\Gamma(T\mathcal{M}) , a reduced order model (ROM) on a lower-dimensional Lie group G is expressed using an action \Phi:G\times\mathcal{M}\rightarrow\mathcal{M} and a reduction map \rho:\mathcal{M}\rightarrow\mathfrak{g} . The goal is to choose (G,\Phi,\rho) such that \bar{x}(t)=\Phi(g(t),x_{0}) approximately follows the FOM dynamics \dot{\bar{x}}\approx X(\bar{x}) . Model order reduction (MOR) is an essential step for the repeated simulation and optimization of distributed and multi-scale engineering systems, such as soft robotic manipulators [ 2 , 3 , 4 ] , hybrid reaction-diffusion systems [ 5 , 6 ] , fluids [ 7 ] and systems with fluid-structure interactions [ 8 ] . These systems are frequently modeled by (parameterized) partial differential equations (PDEs), which yield high dimensional ordinary differential equations (ODEs) upon spatial discretization – often considered as full order models (FOMs). The core idea of MOR is to approximate these FOMs by low-dimensional surrogate models, called reduced order models (ROMs), that can be evaluated with far lower computational complexity. A big class of MOR techniques begins with large datasets of snapshots, which could be solutions from high-fidelity simulations of FOMs or state measurements from real world experiments, for an interesting range of system parameters and initial conditions. In the following, we denote the high-dimensional state-manifold of the FOM as \mathcal{M} , a set of solution snapshots (classically, the solution manifold) as S\subseteq\mathcal{M} , and we lump parameters and initial conditions as \mu\in\mathcal{P} , with \mathcal{P} a compact set. In linear-subspace methods [ 9 , 10 ] , \mathcal{M} is a Hilbert space, and S is approximated by a subspace W\subseteq\mathcal{M} , e.g., spanned by suitable reduced basis elements \{\varphi_{1},\cdots,\varphi_{N}\},\;\varphi_{i}\in W [ 11 ] . Introducing time-dependent coefficients c_{\mu}^{i}(t)\in\mathbb{R} indexed by i , real solutions u_{\mu}(t)\in S are approximated as u_{\mu}(t)\approx\sum_{i=1}^{N}c_{\mu}^{i}(t)\varphi_{i}\,. (1) Linear-subspace methods scale to high-dimensional systems, can preserve physical structure and can capture input-output behavior for control applications [ 12 , 13 ] . Theoretical error bounds are well-known: the Kolmogorov N -width [ 14 ] gives, depending on S , a lower bound to the approximation error that can be achieved by an N -dimensional subspace W of \mathcal{M} . However, linear-subspace methods are ill-suited for S with a slowly decaying Kolmogorov N -width, e.g., in certain transport dominated problems from fluid dynamics or heat transport, where solutions do not evolve on fixed subspaces [ 1 ] . This slow decaying Kolmogorov N -width is often referred to as the Kolmogorov barrier, and to break it, nonlinear MOR methods are required [ 15 ] , for example by allowing also the basis elements to change over time: u_{\mu}(t)\approx\sum_{i=1}^{N}c_{\mu}^{i}(t)\varrho_{\mu,i}(t)\,. (2) Research on MOR methods going beyond the Kolmogorov barrier include: adaptive basis methods [ 16 , 17 ] , which recompute the reduced basis over time, polynomial approximation methods [ 18 , 19 , 20 ] , and machine learning approaches [ 21 , 22 , 23 , 19 , 24 ] , which are sufficiently complex to express relations of the form ( 2 ) but typically incur large computational cost. A recent differential geometric description of MOR encompasses many of the aforementioned methods in a common description as submanifold methods [ 25 ] : given the FOM as a dynamical system on a \mathcal{M} , the ROM is a dynamical system on a submanifold (rather than a subspace) \mathcal{N} of \mathcal{M} . In the present article we use a new geometric approach, utilizing the action of Lie groups. Lie groups act on the underlying configuration and state manifolds for a wide variety of physical systems, including rigid bodies, flexible bodies and fluids, where they allow to describe and exploit symmetries of the systems’ kinematics and dynamics [ 26 ] . With reference to Figure 1 our approach reduces a dynamical system on a high-dimensional state-manifold \mathcal{M} to a lower-dimensional dynamical system on a Lie group G . Let \Phi:G\times\mathcal{M}\rightarrow\mathcal{M} denote an action of G on \mathcal{M} , and u_{\mu,0}\in\mathcal{M} the initial condition of an FOM trajectory, then we work towards approximations of the type u_{\mu}(t)\approx\Phi(g_{\mu}(t),u_{\mu,0})\,. (3) We show that our method induces a distribution on the state-space manifold, rather than picking a submanifold of it. Thus initial states u_{0}\in\mathcal{M} are not projected to a submanifold, and approximations can reproduce initial states exactly. Early Lie group methods for MOR [ 1 , 27 , 28 ] split dynamics into group and manifold parts (vertical and horizontal components on a principal bundle, respectively) and applied submanifold methods to reduce the latter. The splitting explicitly required equivariance of the underlying dynamical systems, and MOR methods were intrusive, i.e., based on explicit knowledge of the FOM dynamics. Our method does not require equivariance, and allows a novel non-intrusive (i.e., data-driven) approach that exploits the presence of approximate group motions. Recently [ 29 ] used a learning approach to find a ROM on a subgroup of G=\mathrm{Diff}(\mathbb{R}^{d}) to reduce Burgers equation with discontinuous shocks. Our method works towards a general geometric approach enabling other choices of G , in a fashion that is compatible with machine-learning, but does not rely on it. Concretely, our contributions are: 1. Development of a new MOR procedure via Lie groups, which we name MORLie (Section 4.1 ). 2. Formalization of a group Kolmogorov N -width , providing a lower bound for the approximation error of MORLie (Section 4.2 ). 3. Development of optimization problems to compute ROMs, both intrusive and non-intrusive, and numerical methods to solve them (Section 5 ). 4. Multiple example applications: point clouds undergoing sheering motions, liver tracking during respiration, and analytically recovering the method of freezing [ 1 ] (Section 6 ). The article is ordered as follows. We conclude the introduction with related literature and notation. Section 2 provides