Symplectic eigenvalue problem via trace minimization and Riemannian optimization

Nguyen Thanh Son, P.-a. Absil, Bin Gao, Tatjana Stykel

Abstract

We address the problem of computing the smallest symplectic eigenvalues and the corresponding eigenvectors of symmetric positive-definite (SPD) matrices in the sense of Williamson's theorem. It is formulated as minimizing a trace cost function over the symplectic Stiefel manifold. We first investigate various theoretical aspects of this optimization problem, such as characterizing the sets of critical points, saddle points, and global minimizers, as well as proving that non-global local minimizers do not exist. Based on our recent results on constructing Riemannian structures on the symplectic Stiefel manifold and the associated optimization algorithms, we then propose solving the symplectic eigenvalue problem in the framework of Riemannian optimization. Moreover, a connection of the sought solution with the eigenvalues of a special class of Hamiltonian matrices is discussed. Numerical examples are presented.

Introduction

In standard linear algebra, symmetric matrices are typically diagonalized via orthogonal transformations. However, in fields such as quantum mechanics, Hamiltonian dynamics, and control theory, we frequently encounter Williamson’s Theorem. This theorem states that any symmetric positive-definite (SPD) matrix $M$ can be diagonalized using a symplectic transformation. The resulting diagonal entries, known as symplectic eigenvalues, characterize the fundamental properties of the underlying physical or mathematical system.

While numerical solvers for standard eigenvalue problems are highly mature, the symplectic counterpart presents unique challenges. Symplectic matrices must preserve a specific skew-symmetric structure ($X^T J X = J$). Unlike the set of orthogonal matrices, the set of symplectic matrices is non-compact and unbounded, which complicates traditional optimization and iterative methods. This paper proposes a rigorous framework to compute the $k$ smallest symplectic eigenvalues by treating the task as a trace minimization problem on the symplectic Stiefel manifold.

Theoretical Framework and Optimization Landscape

The core of our approach involves formulating the search for the $k$ smallest symplectic eigenvalues as a minimization problem over the symplectic Stiefel manifold $Sp(2k, 2n)$. We define a trace cost function and analyze its geometric properties.

A significant theoretical contribution of this work is the characterization of the optimization landscape. We mathematically define the sets of critical points, saddle points, and global minimizers. Crucially, we provide a proof that the cost function is devoid of non-global local minimizers. This "well-behaved" landscape ensures that gradient-based solvers are not trapped in suboptimal local minima, providing a theoretical guarantee that the global solution is reachable from any suitable starting point.

Riemannian Optimization Approach

To solve the minimization problem, we leverage the Riemannian structure of the symplectic Stiefel manifold. Rather than using standard Euclidean optimization with external constraints, we treat the search space as a curved manifold.

Algorithm and Retraction

We utilize a Riemannian gradient descent algorithm that incorporates a Cayley retraction. This retraction serves as a specific update rule that maps a tangent vector back onto the manifold, ensuring that every iteration of the algorithm stays exactly within the symplectic constraints. The process is summarized in a two-step algorithm:

  1. Subspace Optimization: Solve the Riemannian trace minimization problem to identify a basis for the symplectic subspace.
  2. Diagonalization: Perform a small-scale diagonalization to extract the exact symplectic eigenvalues and eigenvectors.

This method also reinforces the link between the symplectic eigenvalues of an SPD matrix $M$ and the standard eigenvalues of the Hamiltonian matrix $J_{2n}M$, offering a unified perspective on these two mathematical structures.

Numerical Experiments

The proposed algorithm was evaluated using three primary numerical benchmarks to test precision, scalability, and robustness.

Synthetic and Physics-Based Models

  1. Synthetic SPD Matrices: Used to verify the algorithm's precision and convergence rates against known ground truths.
  2. Wire Saw Model: A gyroscopic system (size $n=2000$) from mechanical engineering. In this test, the Riemannian approach achieved a normalized residual of approximately $1.3 \times 10^{-14}$. This outperformed the standard eigs function ($1.4 \times 10^{-12}$) and the specialized symplLanczos algorithm ($1.7 \times 10^{-10}$).
  3. Piston Rod Model: A large-scale discretized model of a combustion engine component (size $n=8053$). The algorithm successfully handled this high-dimensional problem, maintaining a residual of $9.9 \times 10^{-12}$.

Discussion and Future Work

Our results demonstrate that the Riemannian optimization framework is highly competitive with established methods. Its primary advantage lies in its robust geometric foundation and the theoretical guarantee of global convergence.

However, some limitations remain. Because the symplectic manifold is unbounded, the cost function is only bounded from below, meaning the algorithm is naturally suited for finding the smallest eigenvalues. Finding the largest eigenvalues currently requires applying the algorithm to the matrix inverse. Additionally, as a first-order method, the gradient descent can exhibit slower convergence when very close to the minimizer. Future research will focus on developing second-order Riemannian methods, such as Riemannian Newton or Trust-Region algorithms, to accelerate final convergence and improve computational efficiency for ultra-large-scale systems.

Summary updated

Abstract

We address the problem of computing the smallest symplectic eigenvalues and the corresponding eigenvectors of symmetric positive-definite matrices in the sense of Williamson's theorem. It is formulated as minimizing a trace cost function over the symplectic Stiefel manifold. We first investigate various theoretical aspects of this optimization problem such as characterizing the sets of critical points, saddle points, and global minimizers as well as proving that non-global local minimizers do not exist. Based on our recent results on constructing Riemannian structures on the symplectic Stiefel manifold and the associated optimization algorithms, we then propose solving the symplectic eigenvalue problem in the framework of Riemannian optimization. Moreover, a connection of the sought solution with the eigenvalues of a special class of Hamiltonian matrices is discussed. Numerical examples are presented.
BibTeX
@article{son2021symplectic,
  author = {Nguyen Thanh Son and P.-A. Absil and Bin Gao and Tatjana Stykel},
  title = {{Symplectic eigenvalue problem via trace minimization and Riemannian optimization}},
  journal = {SIAM J. Matrix Anal. Appl.},
  year = {2021},
  url = {https://openreview.net/forum?id=N83rUOkRgC},
}
Figure