Model Order Reduction (MOR)

• Model order reduction is a technique that approximates high-dimensional systems with low-dimensional models while preserving essential input/output behavior and error bounds.
• Projection-based methods such as POD, Balanced Truncation, and Krylov methods form the core of MOR by constructing reduced-order models that capture dominant system characteristics.
• MOR is applied in power systems, integrated circuits, and PDE simulations to enable large-scale, real-time control and efficient uncertainty quantification.

Model order reduction (MOR) comprises a class of mathematical and computational techniques that approximate complex, high-dimensional dynamical systems by models of much lower dimension, yet with preserved essential system behavior. In control, simulation, and optimization for physical, engineering, and industrial systems, MOR enables large-scale simulation, uncertainty quantification, and real-time control by constructing reduced-order models (ROMs) that faithfully replicate input/output and state trajectories of the original full-order model (FOM), often with certified error bounds or preserved system structure.

1. Mathematical Foundations and Classical Approaches

Model order reduction typically begins with a high-dimensional system—often the result of spatial discretization of partial differential equations (PDEs), or from discretized circuits or multi-physics models—represented in state-space or descriptor (generalized state-space) form: $E \dot{x}(t) = A x(t) + B u(t),\quad y(t) = C x(t)$ with $E, A \in \mathbb{R}^{n \times n}$, $B \in \mathbb{R}^{n \times m}$, $C \in \mathbb{R}^{p \times n}$, and $n \gg 1$.

MOR seeks a mapping onto a reduced-order system: $$\tilde{E} \dot{\tilde{x}}(t) = \tilde{A} \tilde{x}(t) + \tilde{B} u(t),\quad \tilde{y}(t) = \tilde{C} \tilde{x}(t),\;\; \tilde{x} \in \mathbb{R}^r,\, r \ll n$$ such that the transfer function $H(s) = C (sE - A)^{-1} B$ is well approximated by $$\tilde{H}(s) = \tilde{C} (s\tilde{E} - \tilde{A})^{-1} \tilde{B}$$ with error $\| H - \tilde{H} \|$ small in a suitable norm.

Projection-based methods are central. In Galerkin or Petrov-Galerkin projection, one chooses trial and test subspaces (matrices $V, W$), typically via: $$x \approx V y_r,\quad W^T(E V \dot{y}_r - A V y_r - B u) = 0$$ yielding $\tilde{E} = W^T E V$, etc. (Melendez et al., 2022)

Key methodologies include:

• Proper Orthogonal Decomposition (POD): SVD-based extraction of modes that capture dominant energy directions of solution snapshots, optimal in the sense of minimizing the projection error for the sampled data (Binder et al., 2020, Fang et al., 26 Jul 2025).
• Balanced Truncation (BT): Simultaneous diagonalization of reachability and observability Gramians, discarding states with low joint controllability/observability. The method provides a priori error bounds: $$\| H - H_r \|_\infty \leq 2 \sum_{i=r+1}^n \sigma_i$$ with $\sigma_i$ the Hankel singular values (Giamouzis et al., 2023, Giamouzis et al., 2024).
• Moment-Matching (Krylov Methods): Construction of reduced spaces that interpolate moments of the transfer function at selected expansion points. Generalized by extended Krylov subspaces which match both low- and high-frequency moments for improved accuracy (Chatzigeorgiou et al., 2020, Oyaro et al., 2015).
• Reduced Basis (RB) Methods: Greedy or adaptive selection of parameter samples combined with local optimal subspaces, tailored for parametric PDEs (Melendez et al., 2022).

2. Advanced Reduction Frameworks: Beyond Linear Subspaces

Classical subspace-based MOR is fundamentally limited by the Kolmogorov $n$-width of the solution manifold, $$d_n(S) = \inf_{\dim W = n} \sup_{x \in S} \| x - P_W x \|$$, which may decay only algebraically for transport- or advection-dominated problems (Wotte et al., 5 Nov 2025). To overcome this, several advanced techniques are now prominent:

Nonlinear and Manifold-based Reduction:

• Quadratic and Higher-Order Manifolds: The quadratic-manifold approach parameterizes solutions on nonlinear, data-learned manifolds of the form $x \approx V a + \frac12 H(a \otimes a)$, capturing geometric curvature in the solution set and yielding substantially improved accuracy for strongly nonlinear power system transients—with >100× speedup and error reductions of order 5–10 compared to linear POD (Farooq et al., 5 Dec 2025).
• Lie Group–Based MOR (MORLie): Embeds the reduction process on group orbits, i.e., approximating solution trajectories as $x(t) \approx g(t) \cdot x_0$ for $g(t)$ in a low-dimensional Lie group acting on the state manifold. This group Kolmogorov width can decay exponentially, breaking the algebraic barrier for transport problems and yielding extremely compact ROMs for large deformation or geometric motion (Wotte et al., 5 Nov 2025).

Structure-Preserving and DAE Reduction:

• Joint Differential/Algebraic Reduction: For nonlinear DAEs in power networks, block-diagonal projection with structure-preserving POD or BPOD ensures that both dynamic (differential) and network (algebraic) variables are reduced, yielding a reduced NDAE with preserved index and constraints, and orders-of-magnitude speedup on large power systems (Nadeem et al., 2024).
• Optimization-Based Structured Reduction (SOBMOR): Direct optimization of low-order ROM parameters within a parametrization that enforces port-Hamiltonian or second-order symmetry, using only frequency response samples, achieves $\mathcal{H}_\infty$-level error competitive with unconstrained methods for large pH/SSO systems (Schwerdtner et al., 2020).

Time-Limited and Frequency-Aware Optimality:

• Finite-Horizon $\mathcal{H}_2$–Optimal Reduction: Extension of $\mathcal{H}_2$-optimality conditions to finite time intervals, along with iterative projection algorithms satisfying first-order optimality for the time-limited norm, can reduce transient simulation errors compared to classical (infinite-horizon) reduction, particularly relevant for control with time-windowed specifications (Goyal et al., 2017).

Non-Intrusive, Data-Driven, and Certified Reduction:

• Data-Driven, Two-Sided Moment Matching: Identification of ROMs from time-domain input/output data, matching both left and right moments at selected interpolation points, enables high-fidelity, model-free reduction of wind-farm models with 200+ turbines, with total extraction/simulation speedups exceeding 500× vs classical methods (Gong et al., 2024).
• Certified ROM from Data: The use of simulation functions and interface controllers gives formal, trajectory-wise error bounds for ROMs identified solely from two input-state sequences, enabling controller synthesis and formal verification for unknown linear dynamical systems (Samari et al., 3 Feb 2025).
• Active Learning and Neural Surrogates: Greedy error estimator–guided sampling, PAC-validation, and neural-network integrators facilitate non-intrusive construction of robust digital twins, achieving validated prediction error within 1–2% with dramatic speed and sample-efficiency gains compared to random-based snapshot strategies (Zhuang et al., 2022).

3. Algorithmic Realizations and Computational Complexity

The practical design of MOR hinges on scalable algorithms and efficient offline–online decomposition:

• Balanced Truncation via Krylov Subspaces: Classical BT has cubic complexity and is infeasible for $n > 10^4$. Low-rank BT leverages extended Krylov subspace (EKS) methods to project large Lyapunov equations onto small rational subspaces, enabling BT with rigorous error control up to $n \approx 10^5$–$1.6 \times 10^5$ with orders-of-magnitude ROM compression over industry tools (Giamouzis et al., 2023, Garyfallou et al., 2023, Giamouzis et al., 2024).
• Moment-Matching and Householder Techniques: TurboMOR iteratively applies block-congruence with Householder transformations, producing block-diagonal, passivity-preserving ROMs in RC networks with many ports and arbitrary moment matching, lowering simulation time and memory requirements relative to Arnoldi-based approaches (Oyaro et al., 2015, Chatzigeorgiou et al., 2020).
• Efficient Nonlinear Manifold Reductions: Quadratic-manifold methods require SVD and least-squares fitting for the manifold coefficients ($H$), but remain feasible for offline basis computation and admit fast online evaluation due to the small projected dimension ($r \leq 20$) (Farooq et al., 5 Dec 2025).
• Reduced Basis and Hyperreduction: For non-affine or nonlinear parameter dependence, efficient offline/online decomposition is achieved via Empirical Interpolation Method (EIM), Discrete Empirical Interpolation Method (DEIM), or gappy POD (Key et al., 2023, Melendez et al., 2022).
• Error Estimation and Validation: Many modern pipelines accompany a posteriori error indicators (POD residuals, simulation functions, PAC confidence certificates) to ensure and control the fidelity of the ROM in both training and application regimes (Samari et al., 3 Feb 2025, Zhuang et al., 2022).

4. Representative Applications and Numerical Performance

Power Systems:

• Quadratic-manifold ROM reduces simulation time by >100× and error by 5–10× over linear POD, even under severe grid faults. RMS state errors drop to $10^{-3}$ ($r = 3$–6) (Farooq et al., 5 Dec 2025).
• Structure-preserving DAE reduction yields a 40-state ROM for a 2000-bus system with RMS trajectory errors below $10^{-3}$, 24–28× speedup (Nadeem et al., 2024).

Integrated Circuits and Electrical Networks:

• Low-rank BT via EKS for RLCk models reduces ROM order by up to ×22 relative to ANSYS RaptorX, at S-parameter errors <1.4×$10^{-3}$ for $N \approx 10^5$–$1.6 \times 10^5$, with scalable memory and runtime (Giamouzis et al., 2023, Giamouzis et al., 2024, Garyfallou et al., 2023).
• TurboMOR matches arbitrary even moments using block-diagonal models, attaining 3–10× faster reduction and lower simulation cost for many-port RC networks (Oyaro et al., 2015).

Parametric PDEs and Physics:

• Laplace-based contour-integral MOR for parametric PDEs yields exponential decay of Kolmogorov width and snapshot singular values, achieves almost immediate all-at-once evaluation, and 20–40× online acceleration vs time-marching POD for diffusion, advection, and transport problems (Guglielmi et al., 2022).
• In nuclear physics, RB methods and eigenvector continuation uniformly achieve sub-percent errors and speedup of 100×–10,000× in emulators of large many-body Hamiltonians (Melendez et al., 2022).

5. Extensions, Limitations, and Future Directions

Several ongoing challenges and extensions include:

• Kolmogorov Width Limitations and Nonlinear Geometry: Linear projection methods face a barrier for transport- or convection-dominated problems (polynomial decay), which is exceeded only by nonlinear/traveling-manifold, group, or kernel-based MOR (Wotte et al., 5 Nov 2025, Farooq et al., 5 Dec 2025).
• Certification and Verification: Especially in data-driven/self-supervised ROMs, providing trajectory-wise certification under minimal data remains active; ongoing advances include SF-based guarantees and PAC certificates (Samari et al., 3 Feb 2025, Zhuang et al., 2022).
• Structure-Preservation: For physical models (port-Hamiltonian, pH, or second-order systems), maintaining energy, passivity, and symmetry often requires sophisticated parametrization or constrained optimization to match the accuracy of unconstrained ROMs (Schwerdtner et al., 2020).
• Hyperreduction and High-Dimensional Parameter Domains: Efficient nonlinear reduction with fast online cost (e.g., EIM, DEIM, gappy-POD) and sampling in high-dimensional parameter spaces (active subspaces, machine-learning–driven selection) are actively investigated (Melendez et al., 2022, Zhuang et al., 2022).
• Non-Intrusive and Data-Driven MOR: Techniques extracting ROMs solely from time-domain data or measurement—bypassing explicit system matrices—now match or surpass classical methods in speed and robustness for very large, complex systems (Gong et al., 2024, Samari et al., 3 Feb 2025).

6. Software Ecosystem and Community Practices

Modern MOR research benefits from robust software infrastructure, including open-source projects (pyMOR, libROM, MORLAB, pressio) supporting projection-based, balanced truncation, and data-driven approaches. Seamless integration of these MOR pipelines into large simulation toolchains or EM-circuit post-layout flows is now standard for industrial-scale design (Giamouzis et al., 2023, Garyfallou et al., 2023, Melendez et al., 2022).

7. Summary Table: Canonical MOR Techniques and Domains

Method	Class/Scope	Typical Problem Domains
Balanced Truncation	LTI, Descriptor	Circuit, EM, Water Networks
POD-Galerkin	Nonlinear/PDE	CFD, Structural, Financial PDEs
Extended Krylov	Descriptor, MIMO	IC Circuit, Power grid simulation
Quadratic/Kernal Manifolds	Nonlinear	Power system dynamics
Lie group/MORLie	Transport/Deform	Shape deformation, Biomechanics
Data-driven Moment Matching	Black box	Wind farms, Unknown systems
Structure-Preserving Optimization	PH/SSO, Certified	Mechanical, Hamiltonian

In summary, MOR provides a deeply developed and diversified arsenal of techniques for high-fidelity, scalable modeling and simulation in scientific computing and engineering. The ongoing evolution of the field is characterized by a convergence of algebraic, geometric, statistical, and data-driven methods, with increased emphasis on structure preservation, error certification, and scalability to truly large-scale, real-world systems.