Abstract
Numerous problems in optics, quantum physics, stability analysis, and control of dynamical systems can be brought to an optimization problem with matrix variable subjected to the symplecticity constraint. As this constraint nicely forms a so-called symplectic Stiefel manifold, Riemannian optimization is preferred, because one can borrow ideas from unconstrained optimization methods after preparing necessary geometric tools. Retraction is arguably the most important one which decides the way iterates are updated given a search direction. Two retractions have been constructed so far: one relies on the Cayley transform and the other is designed using quasi-geodesic curves. In this paper, we propose a new retraction which is based on an SR matrix decomposition. We prove that its domain contains the open unit ball which is essential in proving the global convergence of the associated gradient-based optimization algorithm. Moreover, we consider three applications—symplectic target matrix problem, symplectic eigenvalue computation, and symplectic model reduction of Hamiltonian systems—with various examples. The extensive numerical comparisons reveal the strengths of the proposed optimization algorithm.
Introduction
The symplectic Stiefel manifold, denoted as $\mathrm{Sp}(2k, 2n)$, consists of matrices $X \in \mathbb{R}^{2n \times 2k}$ that satisfy the symplecticity constraint $X^T J_{2n} X = J_{2k}$, where $J$ is the standard skew-symmetric Poisson matrix. This manifold appears naturally in Hamiltonian mechanics, quantum physics, and control systems. Optimizing over this manifold is non-trivial due to its non-convexity and unboundedness. Riemannian optimization offers a principled framework by treating the manifold as a curved space, but it requires efficient "retractions"—mappings from the tangent space back to the manifold—to update iterates while maintaining the constraint.
Riemannian Framework and Metrics
To perform optimization, one must define a Riemannian metric on the tangent space. This paper explores two primary choices: the standard Euclidean metric and a Canonical-like metric. The choice of metric influences the definition of the gradient and the projection operators, which are essential for navigating the manifold's geometry.
The Euclidean metric is often computationally simpler, while the Canonical-like metric can offer better stability in specific applications like model reduction. The geometric tools derived for these metrics, such as the orthogonal projections onto the tangent and normal spaces, are summarized in the table above.
The SR Decomposition-Based Retraction
The core contribution of this work is a new retraction based on the SR decomposition. Analogous to how the QR decomposition is used for the standard Stiefel manifold (orthogonal matrices), the SR decomposition—where $S$ is a symplectic matrix and $R$ is a specialized upper triangular-like matrix—is the natural fit for the symplectic Stiefel manifold.
The authors adapt a symplectic Gram-Schmidt process to compute the SR factor of a rectangular matrix. A critical theoretical result established is that the SR retraction is well-defined for any tangent vector within the open unit ball (spectral norm less than one). This "unit ball" property is a prerequisite for ensuring the global convergence of Riemannian gradient-based algorithms.
Applications and Numerical Results
The proposed SR-based optimization was evaluated across several domains using high-dimensional mathematical models:
- Symplectic Target Matrix Problem: Finding the nearest symplectic matrix to a given target (e.g., approximating quantum gates). Results show that the Euclidean metric is significantly faster for this class of problems.
- Symplectic Eigenvalue Computation: Tested on synthetic matrices and a real-world Wire Saw Model. The SR optimization achieves high accuracy while ensuring eigenvalues remain purely real, satisfying physical requirements.
- Symplectic Model Reduction: Applied to Hamiltonian systems including the Linear Wave, Sine-Gordon, Schrödinger, and Vlasov equations.
Numerical experiments demonstrate that the SR retraction maintains the symplecticity constraint more robustly than existing quasi-geodesic methods, which often suffer from numerical "drift." In the case of the Vlasov equation, the optimization-based approach reduced state errors by orders of magnitude compared to traditional SVD-based methods.
Conclusion
The SR decomposition provides a theoretically sound and numerically stable retraction for optimization on the symplectic Stiefel manifold. While the current implementation focuses on first-order gradient methods, the robust constraint preservation and global convergence properties make it a superior choice for complex physical simulations. Future work will explore second-order Riemannian methods, such as Newton or Conjugate Gradient variants, to accelerate convergence near optimal solutions.
Summary updated