Brunovsky Normal Form in Control Systems

• Brunovsky Normal Form is a canonical coordinate representation for controllable LTI systems that transforms matrices into a chain-of-integrators structure.
• It exposes controllability indices, facilitating feedback linearization and optimal actuator design in practical control applications.
• Modern numerical methods optimize transformation parameters to enhance robustness and mitigate floating-point sensitivity in system computations.

Brunovsky normal form is a canonical coordinate representation for finite-dimensional controllable linear systems, central to system-theoretic classification, feedback linearization, and various control design methodologies. For a controllable pair (A,B) in continuous-time LTI systems, there exists a similarity transformation that brings the system matrices (A,B) to a block-diagonal companion, or “chain-of-integrators,” structure. This form exposes the controllability indices and enables both theoretical and computational advances in optimal control, actuator design, and the generalization to nonlinear and multi-input systems (Geshkovski et al., 2021, Yang et al., 5 Dec 2025, Nicolau et al., 2023).

1. Mathematical Definition and Structural Features

Given a continuous-time LTI system $\dot{x} = A x + B u$, $x\in\mathbb{R}^n$, $u\in\mathbb{R}^m$, the pair (A,B) is controllable if

$$\mathrm{rank}[\,B,\,A B,\,\ldots,\,A^{n-1}B\,] = n.$$

Under this condition, there exists an invertible matrix $P \in \mathrm{GL}_n(\mathbb{R})$ such that under the coordinate transformation $z = P^{-1} x$, the system assumes the Brunovsky, or block-companion, form: $\dot{z} = J z + B' u,$ where $J = \mathrm{diag}(J_1,\dots,J_m)$, each $J_i$ is a companion matrix of size $\sigma_i \times \sigma_i$, and $B' = (1,0,\dots,0)^\top$ in each block ($e_1 \in \mathbb{R}^{\sigma_i}$). The block sizes $\sigma_1,\dots,\sigma_m$ — the controllability indices — summing to $n$, characterize the minimal orders of integrator chains each input controls.

The single-input ($m=1$) case corresponds to a single $n \times n$ companion matrix; for multi-input cases, the block structure reflects the partition induced by the respective controllability indices (Geshkovski et al., 2021, Yang et al., 5 Dec 2025).

2. Computation and Parametrization of the Brunovsky Form

A constructive route to the Brunovsky form begins with reduction to staircase form using an orthogonal similarity transformation, ensuring numerical stability. For a controllable pair $(A,B)$, the process yields block–upper–triangular $(A_s, B_s)$, from which the family of similarity transformations $T$ mapping $(A_s, B_s)$ to the unique Brunovsky form $(A_b, B_b)$ can be linearly parametrized.

For each controllability index $\mu_i$, the blocks in $A_b$ are

$$A^b_i = \begin{bmatrix} 0 & I_{\mu_i-1} \ 0 & 0 \end{bmatrix}, \qquad B^b_i = \begin{bmatrix} 0 \ \vdots \ 1 \end{bmatrix}.$$

The full transformation is parametrized through observation matrices $C_j$ and parameter matrices $S^r_{k_j}, S^f_{k_j}$, subject to algebraic invertibility constraints ensuring that the transformation and decoupling matrices are nonsingular. For numerical conditioning, a "deadbeat" gain is applied to make the closed-loop system nilpotent, circumventing the risk of unbounded matrix powers in the transformation algorithm. Optimization over the parameter set is performed to minimize the condition number of the transformation matrices, yielding transformations with several orders of magnitude improved reliability relative to classical approaches (Yang et al., 5 Dec 2025).

3. Controllability Indices and System Decomposition

The controllability indices $(\sigma_1,\dots,\sigma_m)$ or $(\mu_1,\dots,\mu_m)$ are central invariants, encoding the minimal number of derivatives on each input needed to affect all states. Given the controllability matrix

$$C = [B, AB, \ldots, A^{n-1}B] \in \mathbb{R}^{n \times nm},$$

the first $n$ independent columns are grouped according to their originating input; the counts define the indices, and their sum gives the state dimension. The Weyr characteristic, its conjugate partition, organizes block sizes for the staircase form. This blockwise decomposition underpins both the existence and the explicit construction of the Brunovsky representation (Geshkovski et al., 2021, Yang et al., 5 Dec 2025).

4. Role in Optimal Actuator Design and Control

The Brunovsky form enables reframing otherwise intractable optimal actuator placement problems for finite-dimensional linear systems. The core observation is that the minimum controllability cost—quantified via the worst-case maximized inverse-controllability Gramian—is directly related to the norm of the inverse of the change-of-basis matrix $P$. Reformulating: $\mathcal{C}^*(B,T) = \kappa(T) \|P^{-1}\|$ with $\kappa(T)$ a function of the block-companion structure, the actuator optimization problem becomes a spectral minimization over $P(B)$: $$\min_{\|B\|=1} \|P(B)^{-1}\| \equiv \max_{\|B\|=1} \lambda_{\min}(P(B)P(B)^\top)$$ where existence of minimizers and symmetry properties are transparent in this form. This reformulation bypasses the need for diagonalizability or randomization methods and highlights the emergence of continuous symmetries and non-uniqueness phenomena in the optimal design, as well as non-convexity that complicates gradient-based algorithms. Global search methods (e.g., differential evolution) have been required in high-dimensional settings (Geshkovski et al., 2021).

5. Extensions to Nonlinear Systems and Generalizations

The Brunovsky normal form generalizes to nonlinear control-affine systems via feedback linearization and flatness theory. For such systems, classical Brunovsky coordinates correspond to static-feedback-linearizable cases (i.e., zero differential degree). For flat, non-static cases, the system admits a finite chain of "prolongations" (differentiations of controls) before linearizability is achieved, with the exact structure and number of integrators determined by the "differential weight." In systems with two inputs and five states that are $x$-flat (flat outputs depend on state only), any such system is brought to a normal form generalizing Brunovsky’s: either a triangular structure comprising chains of integrators followed by nonlinear equations, or, for maximal differential degree, non-triangular forms exhibiting novel singularity phenomena.

These canonical forms make explicit the minimal required integration chains, nonlinearities, and define algebraic/differential regularity conditions under which $x$-flatness holds. Notably, the Brunovsky form is maximally $x$-flat and free of nonlinearities and control-space singularities, properties generically lost in generalized normal forms (Nicolau et al., 2023).

6. Numerical Conditioning and Computational Considerations

Traditional algorithms for computing Brunovsky transformations have suffered from poor numerical reliability, particularly in high-dimensional or ill-conditioned systems. Modern algorithms employ numerically stable steps: first, reduction to staircase form via orthogonal similarity; next, parameterization and optimization of transformations; and further, deadbeat gain application for nilpotency. Efficient optimization over the transformation parameters, with explicit cost functions based on the $\omega$-condition number or standard $2$-norm condition, yields transformations robust to floating-point error and scale.

Empirical evaluations demonstrate 5–10 orders of magnitude improvements in accuracy compared to standard Luenberger-style procedures, and the ability to exploit the full geometric freedom of Brunovsky transformations within the algebraic constraints of controllability (Yang et al., 5 Dec 2025).

7. Significance and Theoretical Impact

The Brunovsky normal form is foundational in the theory and computation of controllable systems. It exposes the structure of linear dynamics most amenable to feedback, clarifies invariant properties under feedback and similarity, and underpins both actuator and observer design. Its generalizations to nonlinear systems demarcate the boundary between true feedback linearizability and settings requiring dynamic compensators, prolongations, or encountering unavoidable singularities. As an organizing principle, it highlights the role of block-companion structure, intrinsic symmetries, and the computational landscape of system transformation—informing design, analysis, and robust computation in modern control theory (Geshkovski et al., 2021, Yang et al., 5 Dec 2025, Nicolau et al., 2023).