Symplectic Model Reduction

• Symplectic model reduction is a structure-preserving technique for high-dimensional Hamiltonian systems that maintains canonical symplectic forms for long-time stability.
• It employs specialized basis construction algorithms such as cotangent lift, complex SVD, and greedy methods to preserve energy and key invariants during projection.
• The approach enhances computational efficiency, facilitates hyperreduction via methods like SDEIM, and enables robust simulation of nonlinear and quantum systems.

Symplectic model reduction refers to a family of structure-preserving model order reduction (MOR) techniques for high-dimensional Hamiltonian systems. Unlike classical methods, which risk destroying the crucial symplectic and energy-conservation properties of Hamiltonian dynamics, symplectic model reduction constructs reduced-order models (ROMs) that retain the canonical symplectic structure, ensuring long-time stability, boundedness, and fidelity of physical invariants under projection. This approach underpins both traditional linear-subspace reduction (e.g., Proper Symplectic Decomposition, PSD), manifold-based extensions, and modern machine learning methods that enforce symplecticity in nonlinear settings.

1. Symplectic Projection for Hamiltonian Systems

Let the full-order system have state $z\in\mathbb{R}^{2n}$ and admit the Hamiltonian form

$$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$

where $H$ is the Hamiltonian and $J_{2n}$ the standard symplectic matrix. A symplectic subspace $W\subset\mathbb{R}^{2n}$ admits a basis $A\in\mathbb{R}^{2n\times2k}$ satisfying $A^T J_{2n} A = J_{2k}$.

The symplectic projection is defined by the symplectic inverse $A^+ = J_{2k}^T A^T J_{2n}$, yielding $A^+A=I_{2k}$. Reducing via $z\approx Ay$ produces the reduced system

$$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$0

which is itself Hamiltonian with inherited structure and stability. This guarantees exact preservation of the reduced Hamiltonian $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$1 and of the symplectic form $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$2 for any pairs $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$3 (Peng et al., 2014, Tyranowski et al., 2019, Afkham et al., 2017).

2. Symplectic Basis Construction Algorithms

Central to symplectic model reduction is the construction of a reduced-order basis (ROB) $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$4 satisfying the symplecticity constraint. Several principal methods are:

• Cotangent Lift (Block-Diagonal Ansatz): Restricts $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$5 to block-diagonal form via SVD of concatenated position-momentum snapshots. This yields $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$6, where $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$7 (Peng et al., 2014, Tyranowski et al., 2019).
• Complex SVD (cSVD): Forms complex-valued snapshots $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$8, computes the truncated SVD, and constructs $$\dot z = J_{2n} \nabla_z H(z), \qquad J_{2n} = \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix},$$9 via

$H$0

which yields a basis that is both symplectic and orthonormal (Peng et al., 2014, Herkert et al., 2024).

• Greedy Algorithms: Iteratively enrich the symplectic basis by maximizing errors (e.g., Hamiltonian error $H$1) on a snapshot set. At each step, symplectic Gram–Schmidt orthogonalization maintains $H$2. Exponential convergence is proven under Kolmogorov $H$3-width decay assumptions (Afkham et al., 2017).
• Non-Orthonormal SVD-like Decomposition: An SVD-like factorization produces non-orthonormal but symplectic ROBs with lower projection error than cSVD in some cases (Buchfink et al., 2019).
• Randomized Symplectic Decompositions: Algorithms such as randomized complex SVD (rcSVD) utilize sketching and power iteration to efficiently compute symplectic bases for large-scale problems, yielding projection errors within a small constant factor of the optimal cSVD bases (Herkert et al., 2023, Herkert et al., 2024).

The following table summarizes basis construction approaches:

Method	Structure	Symplecticity	Orthonormality	Key Reference
Cotangent Lift	Linear	Yes	Yes	(Peng et al., 2014)
Complex SVD	Linear	Yes	Yes	(Herkert et al., 2024)
SVD-like	Linear	Yes	No	(Buchfink et al., 2019)
Greedy	Linear	Yes	Yes (option.)	(Afkham et al., 2017)
Randomized cSVD	Linear	Yes	Yes	(Herkert et al., 2023)
Autoencoder/ML	Nonlinear	Yes	N/A	(Niggl, 2024)

3. Extensions: Weighted Norms, Forced/Dissipative Systems, and Nonlinear Manifolds

Symplectic model reduction generalizes in several important directions:

• Weighted Inner Products: For discretizations yielding non-Euclidean norms (e.g., energy norm from finite elements), all symplectic projections and basis-generation steps can be defined with respect to a weighted inner product. The weighted symplectic form and basis $H$4 must satisfy $H$5 for a suitable weight matrix $H$6 (Afkham et al., 2018).
• Forced and Dissipative Hamiltonian Systems: With Rayleigh dissipation or external forces, reduction proceeds via extension of PSD with additional constraints to preserve the dissipativity (e.g., block structure in $H$7). This ensures the reduced system maintains energy decay and Lyapunov stability properties (Peng et al., 2016).
• Manifold-Based and Nonlinear Model Reduction: For systems where the Kolmogorov $H$8-width decays slowly (e.g., transport-dominated dynamics), linear-subspace ROMs become inefficient. Symplectic model reduction is extended to nonlinear trial manifolds $H$9, where $J_{2n}$0 satisfies a weak symplectic condition. The Symplectic Manifold Galerkin (SMG) projection reduces Hamiltonian systems onto such manifolds, guaranteeing a reduced Hamiltonian structure and energy conservation (Buchfink et al., 2021, Sharma et al., 2023). Data-driven quadratic manifolds and weakly symplectic convolutional autoencoders have demonstrated superior accuracy for both linear and nonlinear systems (Sharma et al., 2023).

4. Symplectic Model Reduction via Machine Learning

Recent advances leverage neural-network architectures to learn nonlinear, structure-preserving reductions:

• Symplectic Autoencoders (AE): Parameterize encoder and decoder as symplectic maps. Architectures stack exact-symplectic nonlinear layers (e.g., generating functions, gradient-based updates) and linear symplectic blocks (PSDLayer). Training minimizes reconstruction loss, potentially with symplectic penalty terms or direct manifold optimization (Niggl, 2024, Brantner et al., 2023).
• Manifold Optimization: The symplectic Stiefel manifold constraints are enforced via Riemannian optimization, tangent-space projection, and special retraction maps. Manifold-adapted Adam variants accelerate large-scale training (Niggl, 2024).
• Universal Approximation Achievability: There exists a universal approximation theorem for symplectic embeddings, showing that any smooth symplectic embedding can be realized to arbitrary accuracy via compositions of symplectic shear layers (generalized MLPs acting as generating functions) (Feng et al., 3 Jun 2026). Consequently, symplecticity-preserving autoencoders (SpAE) constructed as compositions of such layers form highly expressive, exactly structure-preserving reductions.

Empirical studies universally report that symplectic autoencoder reductions maintain Hamiltonian invariants and boundedness over long-time horizon, outperforming both classical PSD and non-symplectic autoencoder baselines (Niggl, 2024, Brantner et al., 2023, Feng et al., 3 Jun 2026).

5. Efficient Hyperreduction and Nonlinear Term Approximation

For nonlinear Hamiltonian systems, the online evaluation of nonlinear terms can become a computational bottleneck. Symplectic DEIM (SDEIM) methods are introduced:

• Symplectic DEIM (SDEIM): By constructing a DEIM basis $J_{2n}$1 and appropriate interpolation schemes, the nonlinear vector field is approximated in such a way that the resulting reduced system remains Hamiltonian. This enables efficient $J_{2n}$2 online evaluation (where $J_{2n}$3 is the number of interpolation points) without sacrificing structural fidelity (Afkham et al., 2017, Peng et al., 2014).
• Adaptive and Randomized Algorithms: Randomized sketching (e.g., rcSVD) and greedy selection for the DEIM basis significantly reduce offline computational cost while preserving symplectic optimality within a controlled error (Herkert et al., 2024, Herkert et al., 2023).

6. Applications, Benchmarks, and Robustness

Extensive evidence supports the effectiveness of symplectic model reduction:

• Long-time Integration: All symplectic ROMs exhibit bounded error in energy invariants and stable trajectories over hundreds to thousands of system periods, even for highly oscillatory or nonlinear problems (Tyranowski et al., 2019, Afkham et al., 2017).
• Parameter Generalization: Greedy and data-driven manifold methods generalize to unseen parameter values and new regimes; the Kolmogorov $J_{2n}$4-width rate is a key predictor of achievable accuracy (Afkham et al., 2017, Sharma et al., 2023).
• Computational Efficiency: Randomized and greedy basis construction reduces offline setup costs by up to an order of magnitude, and SDEIM drastically lowers online operation counts (Herkert et al., 2024, Afkham et al., 2017).
• Quantum Systems: For quantum and physically realizable (PR) systems, symplectic Petrov-Galerkin projections guarantee exact preservation of canonical commutation relations and other algebraic structures, supporting high-fidelity H₂ optimal model reduction in quantum networks (Borzi et al., 8 May 2026).

7. Limitations, Best Practices, and Research Directions

Symplectic model reduction is indispensable whenever long-time energy preservation, structural stability, or geometric fidelity of Hamiltonian systems are critical. Key recommendations are:

• Always enforce symplecticity in basis construction and projection if system invariants are essential for application correctness and reliability (Tyranowski et al., 2019, Peng et al., 2014).
• For marked non-Euclidean spatial discretizations, adopt weighted or problem-adapted symplectic bases (Afkham et al., 2018).
• For high-dimensional, nonlinear, or weakly compressible systems (slow $J_{2n}$5-width decay), leverage quadratic/symplectic manifold generalizations or learning-based approaches (Buchfink et al., 2021, Feng et al., 3 Jun 2026).
• Symplectic Greedy and Randomized basis generation are preferred for large-scale or many-query scenarios (Afkham et al., 2017, Herkert et al., 2023, Herkert et al., 2024).
• Use symplectic integrators for time-stepping of reduced models (Tyranowski et al., 2019).
• Research continues on global optimality of non-orthonormal symplectic bases, extensions to broader geometric structures (e.g., Poisson, contact), global topology constraints in deep symplectic networks, and adaptive regularization for hyperreduction (Buchfink et al., 2019, Feng et al., 3 Jun 2026).

Symplectic model reduction therefore provides a rigorously justified, computationally efficient, and physically faithful foundation for advanced surrogate modeling, uncertainty quantification, and real-time simulation of Hamiltonian and quantum systems.