Balanced Truncation for Model Reduction

• Balanced truncation is a model reduction technique that removes states with low energy contributions by balancing controllability and observability Gramians.
• It produces reduced-order models with a quantifiable H-infinity error bound, ensuring stability and preserving core system behavior.
• The method extends to unstable, bilinear, stochastic, and data-driven systems, offering structure-preserving reductions for diverse applications.

Balanced truncation is a foundational model order reduction technique for dynamical systems, yielding truncated low-order models that retain essential input-output properties of the high-dimensional original system. The core idea is to identify and remove states that are simultaneously hard to control and hard to observe, based on an energy-theoretic analysis via controllability and observability Gramians. Balanced truncation is rigorously characterized for asymptotically stable linear time-invariant (LTI) systems, but its principles extend to unstable systems, bilinear and quadratic-bilinear systems, switched and networked systems, infinite-dimensional and stochastic systems, and various structure-preserving or non-intrusive settings. This article develops the theoretical basis, extensions, and modern algorithmic workflow of balanced truncation, delineating key error bounds, generalizations, and applications.

1. Classical Balanced Truncation: Theory and Algorithm

Balanced truncation for stable continuous-time LTI systems starts from the minimal realization

$\dot x = A x + B u, \quad y = C x + D u,$

with $A \in \mathbb{R}^{n \times n}$ Hurwitz, $B \in \mathbb{R}^{n \times m}$, $C \in \mathbb{R}^{p \times n}$. The reachability (controllability) Gramian $P > 0$ and observability Gramian $Q > 0$ are the unique positive definite solutions to

$$A P + P A^T + B B^T = 0, \qquad A^T Q + Q A + C^T C = 0.$$

The balancing transformation $T$ is constructed such that, in coordinates $\bar x = T x$, the transformed Gramians are equal and diagonal,

$$\bar P = T P T^T = \bar Q = T^{-T} Q T^{-1} = \Sigma = \mathrm{diag}(\sigma_1, \dots, \sigma_n), \quad \sigma_1 \geq \ldots \geq \sigma_n > 0,$$

where the $A \in \mathbb{R}^{n \times n}$0 are the Hankel singular values and quantify joint controllability and observability (Paré et al., 2019).

After partitioning and projecting onto the $A \in \mathbb{R}^{n \times n}$1 dominant balanced directions, the reduced-order system

$A \in \mathbb{R}^{n \times n}$2

satisfies the celebrated $A \in \mathbb{R}^{n \times n}$3 error bound

$A \in \mathbb{R}^{n \times n}$4

where $A \in \mathbb{R}^{n \times n}$5 are the original and reduced transfer functions (Paré et al., 2019). Balanced truncation thus provides a structure-preserving, stability-preserving scheme equipped with a closed-form worst-case error bound.

2. Generalizations: Unstable, Bilinear, Stochastic, and Nonlinear Systems

Balanced truncation extends beyond the stable LTI case to a variety of structured and nonlinear systems:

• Unstable Systems: For systems with unstable $A \in \mathbb{R}^{n \times n}$6, finite-horizon snapshot-based methods yield converged balancing transformations on the optimal time window, balancing even unstable subspaces without projection. Singular directions associated with unstable modes converge as the time horizon increases, ensuring a well-posed basis for model reduction (Flinois et al., 2015).
• Bilinear Systems: Two main frameworks exist. Type I BT solves linear Lyapunov equations for Gramians but only guarantees local energy bounds; Type II BT introduces a control magnitude bound $A \in \mathbb{R}^{n \times n}$7 and solves Riccati-type inequalities for the controllability Gramian, yielding global energy bounds and a true $A \in \mathbb{R}^{n \times n}$8 a priori error estimate,

$A \in \mathbb{R}^{n \times n}$9

for admissible controls $B \in \mathbb{R}^{n \times m}$0 (Redmann, 2017). Infinite-dimensional generalizations and associated error bounds are established in a functional-analytic setting (Becker et al., 2018).

• Quadratic-Bilinear Systems: Algebraic Gramians are derived from Volterra series representations, leading to generalized quadratic Lyapunov equations. Balancing and truncation protocols extend accordingly, with local Lyapunov stability for the reduced model (Benner et al., 2017).
• Stochastic Systems: For SPDEs or bilinear systems perturbed by Wiener noise, balanced truncation is defined via mean-square energy Gramians. The error in the output process is quantifiable in terms of the trace-norm difference in Hankel singular values, e.g.,

$B \in \mathbb{R}^{n \times m}$1

(Becker et al., 2018).

• Nonlinear and Quadratic Output Systems: For LTI systems with quadratic observables, the system is recast into quadratic-bilinear (or high-dimensional MIMO) form, and balanced truncation proceeds via transformed Gramians. Efficient algorithms exploit structural properties to reduce computational complexity, with favorable performance for large-scale outputs (Pulch et al., 2017).

3. Structure-Preserving and Specialized Extensions

Balanced truncation has been adapted to preserve valuable system structures:

• Descriptor and Passive Systems: Reciprocal Positive-Real Balanced Truncation uses Riccati or Lur’e equations for the positive-real Gramians, ensuring preservation of reciprocity and passivity in the reduced model. A key step is the factorization of the signature matrix associated with the system’s symmetry (Tanji, 2018).
• Second-Order Mechanical Systems: Positive-real balanced truncation for second-order systems ensures that reduced models retain symmetric positive-definite mass and stiffness matrices, and preserve overdamped or passive structure. The gap-metric error of the reduction satisfies

$B \in \mathbb{R}^{n \times m}$2

where $B \in \mathbb{R}^{n \times m}$3 are positive-real singular values (Dorschky et al., 2020).

• Networked and Graph-Laplacian Systems: For interconnected passive agent networks, Kronecker-structured Gramians decouple node and network topologies, enabling simultaneous reduction of agent dynamics and topology while guaranteeing passivity and synchronization in the reduced-order model (Cheng et al., 2017).
• Switched Systems: For linear switched systems, mode-coupled Lyapunov equations yield per-mode Gramians, and balanced truncation applies mode-wise. The $B \in \mathbb{R}^{n \times m}$4-error for arbitrary (sufficiently slow) switching signals is proportional to the sum of truncated singular values across all modes (Gosea et al., 2017, Petreczky et al., 2013).
• $B \in \mathbb{R}^{n \times m}$5-Positive Systems: For discrete-time Hankel $B \in \mathbb{R}^{n \times m}$6-positive systems, balanced truncation up to order $B \in \mathbb{R}^{n \times m}$7 yields a reduced model that is Hankel totally positive (i.e., a sum of first-order lags with internal positivity), preserving minimality and state-space symmetry (Grussler et al., 2020).
• Frequency- and Time-Limited BT: Extensions to focus accuracy on specified frequency bands or time intervals involve modifying Lyapunov and balancing procedures, yielding tighter in-band or in-window a priori error bounds (Du et al., 2016, Kürschner, 2017).
• Conformal Mapping: By composing with conformal maps, BT can be generalized to systems with poles in arbitrary open domains. Modified Gramians, computed via quadrature or Lyapunov solvers, preserve spectral constraints, and the BT a priori bound holds in the $B \in \mathbb{R}^{n \times m}$8 norm of the mapped domain (Borghi et al., 2024).

4. Data-Driven and Non-intrusive Balanced Truncation

Recent advances enable balanced truncation and its generalizations to be implemented non-intrusively, relying solely on transfer function samples:

• Projection-Based Approaches: Constructs Krylov or Loewner-type subspaces using frequency-domain data $B \in \mathbb{R}^{n \times m}$9, forms "projected" Gramians whose diagonal-dominant structure admits closed-form Gramian factorizations, and recovers the dominant Hankel singular values directly from data. The classical error bound remains valid (Zulfiqar, 13 Feb 2026).
• Quadrature-Based and ADI-Based Methods: Approximates Gramians by frequency quadrature, constructing low-rank Gramian factors from frequency response samples. For the ADI-based (Petrov-Galerkin) framework, Krylov subspaces are built using transfer function samples at mirror-image shifts, and small Lyapunov or Riccati systems are solved entirely in the sample space (Zulfiqar, 29 Jun 2025). These approaches are practical for very large-scale or proprietary systems where internal state-space matrices are unavailable.
• Extensions to Generalized BT: Data-driven frameworks have been successfully applied to positive-real BT, bounded-real BT, stochastic BT, frequency-limited, and time-limited variants, using only frequency response samples. The same data-driven SVD and projection procedures yield reduced models that are passivity-preserving, bounded-real, or minimum-phase, matching the a priori error guarantees of their intrusive counterparts (Zulfiqar, 13 Feb 2026, Zulfiqar, 29 Jun 2025).
• Second-Order and Bilinear Systems: Nonintrusive BT for systems with proportional damping and K-power bilinear systems builds Gramian factors from tensorized quadrature of transfer function samples, supports efficient (real-valued) SVD/projection, and recovers structure-preserving ROMs with performance comparable to intrusive algorithms (Wang et al., 4 Jun 2025, Wang et al., 13 Apr 2026).

5. Balanced Truncation for Nonstandard and Infinite-Dimensional Systems

Balanced truncation is rigorously established for infinite-dimensional control systems (e.g., PDEs, SPDEs):

• Infinite-dimensional LTI, Bilinear, and Stochastic Systems: Functional-analytic constructions define Gramians as convergent infinite sums over input/output functionals, and the abstract Hankel operator has compactness properties ensuring existence of singular value decomposition and convergence of reduced-order models. Error bounds in trace and Hardy-mixed norms are analogous to the finite-dimensional case (Becker et al., 2018).
• Snapshot-Based Balanced POD for Large/Unstable Flows: For numerical applications in fluid dynamics or high-dimensional unstable PDEs, empirical/snapshot-based balanced POD methods apply balanced truncation directly from impulse or frequency response data, matching analytic performance while avoiding projection onto stable subspaces (Flinois et al., 2015).

6. Model Manifold Interpretation, Boundary Approximations, and Interpolation

Balanced truncation, singular-perturbation approximation, and their hybrids can be understood as "boundary approximations" on a "model manifold" parameterizing all system behaviors under coordinate and Gramian transformations (Paré et al., 2019):

• MBAM Unification: Balanced truncation corresponds to a limit in which specific balanced-realization parameters (Hankel singular values) are sent to zero, truncating unimportant directions. Singular perturbation approximation arises from sending state scaling parameters to infinity, imposing fast-subsystem constraints. Intermediate approximations interpolate these behaviors, providing a unified geometric understanding of model reduction (Paré et al., 2019).
• Implication: This viewpoint inspires new algorithms that interpolate between BT and singular perturbation, with potential application to nonlinear systems where analogous parameterizations may be achieved.

7. Algorithmic and Practical Considerations

Balanced truncation is widely used due to its clear workflow and scalability, especially when combined with sparse or low-rank computational solvers:

1. Compute or approximate the controllability and observability Gramians (or their generalizations).
2. Compute Cholesky or square-root Gramian factors.
3. Perform an SVD of the cross Gramian product to obtain balanced coordinates.
4. Partition and truncate balanced system matrices according to Hankel singular value decay.
5. Assemble the reduced-order model, with structure and stability as required.

For large-scale applications, rational Krylov subspace, low-rank ADI iterations, and purely data-driven SVD/projection methods are prevalent (Kürschner, 2017, Zulfiqar, 13 Feb 2026, Zulfiqar, 29 Jun 2025). All such methodologies preserve computable a priori error certificates and system-theoretic properties (e.g., passivity, internal positivity, reciprocity, network topology) under suitable conditions.

---

In conclusion, balanced truncation has evolved from a classical LTI analysis tool to a unifying, structure-preserving, and data-enabled workflow for broad classes of dynamical systems, including unstable, bilinear, networked, switched, and infinite-dimensional regimes, as well as for non-intrusive, application-driven contexts (Redmann, 2017, Paré et al., 2019, Borghi et al., 2024, Zulfiqar, 29 Jun 2025, Zulfiqar, 13 Feb 2026).