Numerically Reliable Brunovsky Transformations

Shaohui Yang, Colin N. JonesarXiv 2025 · DOI · arXiv

Abstract

The Brunovsky canonical form is a fundamental representation in control theory, enabling efficient analysis and high-performance control design. However, its computation has historically been plagued by numerical instability, particularly for high-dimensional or poorly conditioned systems. This paper introduces a novel optimization-based framework that leverages staircase reduction and deadbeat pre-processing to reliably compute Brunovsky transformations. By parameterizing the family of valid transformations and optimizing for numerical stability, the proposed method achieves construction errors significantly lower than classical algebraic approaches.

Introduction

In control theory, the Brunovsky canonical form represents a multi-variable linear system as a set of independent integrator chains. This structure is highly desirable because it simplifies feedback linearization and enhances the computational efficiency of Model Predictive Control (MPC) through matrix sparsity.

The block-diagonal structure of the Brunovsky canonical form.

Despite its theoretical elegance, calculating the transformation matrix $T$ is notoriously difficult. Standard algorithms developed in the 1960s, such as Luenberger’s method, often involve computing high powers of the system matrix and inverting nearly singular matrices. For systems with large eigenvalues or those near the boundary of controllability, these methods produce ill-conditioned transformations that are practically unusable.

Methodology

The proposed approach by Yang and Jones (EPFL) addresses these challenges by combining robust numerical linear algebra with optimization. The method follows three primary stages:

  1. Staircase Reduction: The system is first transformed into a staircase form using orthogonal transformations. This step is numerically stable and reveals the system's structural parameters (Kronecker indices).
  2. Deadbeat Pre-processing: To prevent numerical explosion when dealing with high powers of the system matrix, a "deadbeat" control gain is applied. This shifts all system eigenvalues to zero, making the matrix nilpotent. This ensures that matrix powers eventually vanish rather than grow uncontrollably.
  3. Optimization over Transformation Families: The core innovation is the recognition that the Brunovsky transformation is not unique. The authors parameterize the entire family of valid transformations and use a quasi-Newton solver to find the specific transformation that minimizes a condition-number-based cost function. This "hunts" for the most numerically stable version of the transformation.

Experimental Results

The method was evaluated using 100 randomly generated controllable systems (15 states, 4 inputs) to compare its performance against the classical Luenberger method.

Numerical performance comparison: The proposed method maintains stability and low error even as system conditioning worsens.

The results demonstrate a clear advantage for the optimization-based approach. While the Luenberger method's error increases exponentially with system ill-conditioning, the proposed method maintains construction errors between $10^{-5}$ and $10^{-10}$. Furthermore, the resulting system matrices exhibit the desired nilpotent structure with minimal numerical noise, and the transformation matrices themselves are significantly better conditioned.

Conclusion

This work challenges the long-standing assumption that canonical forms involving nilpotent structures cannot be computed reliably. By shifting from a rigid algebraic problem to a flexible optimization framework, the authors provide a robust tool for modern control engineering. Future research aims to provide a formal proof for the completeness of the transformation parametrization and explore the global optimality properties of the stability cost function.

Summary updated

Abstract

The Brunovsky canonical form provides sparse structural representations that are beneficial for computational optimal control, yet existing methods fail to compute it reliably. We propose a technique that produces Brunovsky transformations with substantially lower construction errors and improved conditioning. A controllable linear system is first reduced to the staircase form via an orthogonal similarity transformation. We then derive a simple linear parametrization of the transformations yielding the unique Brunovsky form. Numerical stability is further enhanced by applying a deadbeat gain before computing system matrix powers and by optimizing the linear parameters to minimize condition numbers.
BibTeX
@misc{yang2025numerically,
  author = {Shaohui Yang and Colin N. Jones},
  title = {{Numerically Reliable Brunovsky Transformations}},
  year = {2025},
  eprint = {2512.05910},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.05910},
}
Figure