Geometry-preserving Lie Group Integrators For Differential Equations On The Manifold Of Symmetric Positive Definite Matrices

Lucas Drumetz, Alexandre Reiffers-Masson, Naoufal El Bekri, Franck VermetGSI 2021

Abstract

Differential equations on the manifold of Symmetric Positive Definite (SPD) matrices are central to various fields, including finance, signal processing, and control theory. Standard numerical integrators often fail to preserve the positive definiteness of solutions, leading to mathematically invalid results. This paper presents a framework utilizing Lie group integrators to solve these equations while intrinsically maintaining the manifold structure. By employing the Runge-Kutta-Munthe-Kaas (RKMK) algorithm and specific Lie group actions, the proposed methods demonstrate superior accuracy, stability, and robustness to step size compared to traditional Euclidean and projection-based approaches.

Introduction

Symmetric Positive Definite (SPD) matrices, denoted as $Sym_n^+$, are fundamental in representing covariance, diffusion tensors, and kernels. In many applications, these matrices evolve over time according to Ordinary Differential Equations (ODEs). However, the space of SPD matrices is not a flat Euclidean space but a curved manifold.

Traditional numerical solvers, such as the Euler method or the fourth-order Runge-Kutta (RK4), operate under the assumption of a flat vector space. When applied to ODEs on $Sym_n^+$, these "flat-space" methods can produce updates that "shoot off" the manifold. This results in matrices with non-positive eigenvalues, which are physically and mathematically meaningless in the context of the problem.

The Problem of Manifold Departure

The authors formalize the failure of classical methods through a boundary analysis. Specifically, Theorem 1 in the paper demonstrates that if an update vector possesses even a single negative eigenvalue, a sufficiently large integration step $h$ will inevitably push the solution outside the SPD manifold.

To mitigate this, researchers often resort to extremely small time steps, which increases computational overhead without addressing the underlying geometric mismatch. Alternatively, "extrinsic" methods attempt to project the Euclidean result back onto the manifold (Riemannian RK4). However, these projections can introduce inaccuracies and fail to capture the true dynamics of the system.

Lie Group Integration Framework

To address these limitations, the paper proposes the use of Lie group integrators. The fundamental shift is moving from additive updates ($P + \Delta P$) to multiplicative group actions.

Transitive Lie Group Actions

A Lie group $G$ acts on the manifold $Sym_n^+$ such that for any $P \in Sym_n^+$ and $g \in G$, the action $\phi(g, P)$ remains within $Sym_n^+$. The paper focuses on two primary actions:

  1. General Linear Group ($GL_n(\mathbb{R})$): Utilizes congruence transformations ($MPM^T$). This is particularly effective for linear dynamical systems.
  2. Symplectic Group ($SP(2n)$): Acts on the Siegel upper half-space, providing a robust framework for solving Riccati differential equations, which are essential in Kalman filtering and Linear Quadratic Regulator (LQR) problems.

The RKMK Algorithm

The authors adapt the Runge-Kutta-Munthe-Kaas (RKMK) algorithm. The process involves lifting the ODE from the manifold to the associated Lie algebra (a flat vector space), performing numerical integration there, and then mapping the result back to the manifold using the exponential map. This "intrinsic" approach ensures the solution never leaves the manifold, regardless of the step size.

Case Study: Multivariate Geometric Brownian Motion

The framework was evaluated using a synthetic case study of Multivariate Geometric Brownian Motion (GBM). This choice allows for a direct comparison between numerical approximations and a known closed-form solution. The experiments compared three approaches:

  • Euclidean RK4: Standard integration without manifold constraints.
  • Riemannian RK4: Euclidean integration followed by a projection onto the SPD manifold.
  • Lie-RK4: The proposed geometry-preserving integrator.

Results and Discussion

The results highlight several key advantages of the Lie group approach:

  • Manifold Preservation: Visualizations of $2 \times 2$ matrices show that while Euclidean RK4 drifts into invalid territory, Lie-RK4 remains perfectly on the manifold surface.
  • Accuracy: Lie-RK4 significantly outperforms both Euclidean and Riemannian methods. It maintains lower error rates in both standard Euclidean distance and affine-invariant metrics.
  • Robustness to Step Size: A critical finding is that Lie group integrators remain stable and accurate even with large integration steps ($h$). In contrast, Euclidean methods quickly diverge or produce invalid matrices as $h$ increases.
  • Stability in Control: The method successfully handles Riccati equations, which are notoriously sensitive to numerical errors that break symmetry or positive definiteness.

Conclusion

The introduction of Lie group integrators for SPD manifolds provides a mathematically rigorous and numerically stable framework for solving complex differential equations. While the computational cost per step is higher due to matrix exponentials and inversions, the ability to use larger time steps and the guarantee of manifold preservation offer significant advantages. Future work points toward the integration of these geometric constraints into machine learning, potentially using neural networks to learn dynamics within the Lie algebra to ensure physically valid outputs.

Summary updated

Abstract

In many applications, one encounters signals that lie on manifolds rather than a Euclidean space. In particular, covariance matrices are examples of ubiquitous mathematical objects that have a non Euclidean structure. The application of Euclidean methods to integrate differential equations lying on such objects does not respect the geometry of the manifold, which can cause many numerical issues. In this paper, we propose to use Lie group methods to define geometry-preserving numerical integration schemes on the manifold of symmetric positive definite matrices. These can be applied to a number of differential equations on covariance matrices of practical interest. We show that they are more stable and robust than other classical or naive integration schemes on an example.
BibTeX
@inproceedings{drumetz2021geometry,
  author = {Lucas Drumetz and Alexandre Reiffers-Masson and Naoufal El Bekri and Franck Vermet},
  title = {{Geometry-preserving Lie Group Integrators For Differential Equations On The Manifold Of Symmetric Positive Definite Matrices}},
  booktitle = {GSI},
  year = {2021},
}
Figure