Dynamic Feedback Linearization

• Dynamic Feedback Linearization is a control methodology that augments nonlinear systems with compensator dynamics, enabling exact linearization via static feedback.
• It extends static feedback methods by introducing virtual inputs and additional states, allowing systems that fail classical conditions to achieve flatness and controllability.
• Algorithmic and geometric techniques, including categorical search and symmetry-based methods, offer practical strategies for constructing minimal dynamic extensions in control applications.

Dynamic Feedback Linearization is a geometric control methodology that generalizes static feedback linearization by allowing the introduction of a dynamic precompensator—effectively augmenting a nonlinear plant with auxiliary states and controls in order to render the overall closed-loop system exactly linearizable via static feedback. This approach encompasses the construction, analysis, and synthesis of dynamic extensions for nonlinear systems; it provides a universal framework for flatness, enables the linearization of systems failing classical static conditions, and admits both constructive and categorical optimality perspectives.

1. Foundations and Definitions

Dynamic feedback linearization seeks to realize an exact linear map from new “virtual” control inputs to system outputs by augmenting the original control system with additional compensator dynamics. For a general nonlinear system

$\dot x = f(x, u), \qquad y = h(x, u),$

static feedback linearization is possible if there exist a diffeomorphic state change and a static feedback law such that the closed-loop dynamics can be expressed in Brunovsky canonical form

$\dot \xi = A\xi + B\nu,$

where $\xi$ and $\nu$ are new state and input variables. However, static feedback linearizability is topologically restrictive and characterized by involutive flag and controllability rank conditions (Hunt–Su–Meyer, Jakubczyk–Respondek).

Dynamic feedback linearization relaxes these requirements by introducing an auxiliary (precompensator) system

$$\begin{aligned} \dot z &= a(x, z, v), \ u &= b(x, z, v), \end{aligned}$$

with additional compensator state $z$ and new input $v$, such that the extended state $(x, z)$ admits static feedback linearization to the canonical linear form for some static feedback law and coordinate transformation. The goal is to find—where possible—the minimal order of such dynamic compensation (minimal dynamic extension) and associated coordinate/feedback mappings (D'Souza, 1 Feb 2026).

2. Geometric and Algorithmic Characterizations

Dynamic feedback linearizability is directly tied to geometric structures built from the control vector fields and their Lie brackets. The key invariants are:

• Linearizability distributions: A sequence

$$\mathcal D^0 = \mathrm{span}\{g_1, ..., g_m\},\quad \mathcal D^{j+1} = \mathcal D^j + [f, \mathcal D^j]$$

for $j \ge 0$, where $g_i$ are the input vector fields.

• Involutivity and controllability criteria: Static feedback linearization requires all $\mathcal D^j$ to be involutive of constant rank and $\mathcal D^{n-1} = TX$ on an $n$-dimensional manifold.
• Dynamic extension: When static criteria fail, dynamic feedback linearizability seeks a compensator and mapping that ensure the corresponding lifted distributions become involutive, possibly via a sequence of input prolongations.

The categorical framework models the class of all possible dynamic extensions as a category whose objects are control systems and whose morphisms are extension maps. The search for the minimal dynamic extension is cast as finding a shortest “arrow” in this category, admitting dynamic programming and $A^*$-style constructive algorithms (D'Souza, 1 Feb 2026). The “leading integrability defect” provides an admissible heuristic for guiding search in infinite prolongation spaces. Existence conditions are precisely those of the extended system's involutivity and controllability.

3. Constructive Procedures and Minimal Dynamic Extension

Several explicit and algorithmic procedures have been developed:

• Successive one-fold prolongations: For multi-input systems (notably two-input), the algorithm proceeds by selecting a control to prolong (i.e., introducing $\dot u_j = v_j$), thereby augmenting the state with $u_j$, and searching for an involutive corank-one subdistribution in the derived chain. This process repeats, with each prolongation strictly increasing the minimal involutive index, until the enlarged system becomes static-feedback linearizable or further extension is not possible (Nicolau et al., 10 May 2025).
• Constructive synthesis: Given suitable vector field structures, the algorithm replaces non-involutive distributions by their involutive subdistributions, ensuring at each prolongation an “involutivity gain.” The process is guaranteed to terminate in at most $n-2$ steps for flat two-input systems (Nicolau et al., 10 May 2025).
• Symmetry-based methods: For systems admitting Lie group actions, constructing static-feedback-linearizable quotients and contact sub-connections yields criteria and procedures for dynamic feedback linearization. This perspective employs geometric reductions, partial prolongations along group orbits, and explicit construction of the compensator from bundle mappings and derived flag analysis (Clelland et al., 2021).
• Endogenous and exogenous feedbacks: Canonical constructions distinguish between endogenous (output-derivative based) and exogenous (arbitrary compensator) cases (Jindal et al., 2024), with the lift of discretizations preserving dynamic feedback linearizability in sampled-data systems.

4. Theory to Algorithm: Categorical Search and Heuristics

The search for a minimal dynamic precompensator can be organized as a category-theoretic problem. The universe of all attainable extensions is a category; each object is a control system presented as an exterior differential system, and each morphism is a smooth, covering submersion corresponding to a dynamic extension. Primitive arrows correspond to one-step prolongations (integrator chains) along directions dictated by the regular zero-dynamics foliation. The partial order is by compensator order (i.e., the number of new states introduced).

• Dynamic programming: The cost function is additive under composition and enables Bellman-type equations for optimal extension (D'Souza, 1 Feb 2026).
• Heuristics: The “leading integrability defect” is a lower bound on the required number of one-step corrections and is used to efficiently prune the search space, dramatically accelerating heuristic $A^*$-style algorithms for infinite extension domains (D'Souza, 1 Feb 2026).
• Block diagram representation: The dynamic compensator manifests as a cascade of integrators with nullifier feedback, followed by algebraic mappings to physical controls.

5. Applications, Case Studies, and Examples

Dynamic feedback linearization extends the applicability of feedback-linearizing control to systems not amenable to static techniques. Illustrative instances include:

• Chained form and flat systems: Successive one-fold prolongations render the 4-state chained form system fully static-feedback linearizable with precisely two prolongations (i.e., compensator order two) (Nicolau et al., 10 May 2025).
• Symmetric mechanical systems: For the vertical take-off and landing (VTOL) system, dynamic feedback linearizability is revealed via symmetry-based reduction and suitable prolongation of input derivatives (Clelland et al., 2021).
• Block diagram example: Each primitive arrow in the categorical approach can be mapped to a block-chain realization: the virtual input enters a chain of integrators, whose output, via algebraic transformation, becomes the plant input.

A representative table of dynamic feedback linearization strategies:

Strategy	Required Structure	Output
One-fold Prolongation (Nicolau et al., 10 May 2025)	Flat two-input systems	Static-linearizable extension in ≤n–2 steps
Categorical Search (D'Souza, 1 Feb 2026)	General smooth nonlinear systems	Minimal-compensator via optimal arrow
Symmetry-Based (Clelland et al., 2021)	Lie symmetry present	Compensator from reduction/contact construction

6. Theoretical Implications and Limitations

Dynamic feedback linearization provides a unifying geometric and algorithmic framework:

• Flatness equivalence: For two-input systems, dynamic feedback linearizability via successive prolongation is equivalent to flatness (Nicolau et al., 10 May 2025).
• Constructive completeness: The presented approaches guarantee the construction of a compensator when possible, under verifiable geometric conditions (involutivity, corank, compatibility).
• Computational tractability: Symbolic methods (e.g., PDE reduction, Lie algebra computations) rapidly become intractable at high dimensions, motivating categorical/algorithmic approaches. These provide both optimality guarantees in finite prolongation space and heuristic efficiency in infinite domains (D'Souza, 1 Feb 2026).
• Nonuniqueness and extension space: There may exist multiple, non-minimal dynamic extensions for the same system; the categorical viewpoint clarifies the landscape of such possibilities, introducing optimality and pruning.
• Discretization preservation: First-order discretization schemes can be constructed to maintain dynamic feedback linearizability, provided the compensator and mappings are sampled/lifted in alignment (Jindal et al., 2024).

7. Broader Perspectives and Connections

Dynamic feedback linearization integrates with several core concepts in nonlinear systems and control:

• Flatness and endogenous feedback: The theory underpins differential flatness and the flat-output parameterization via endogenous dynamic feedback (Nicolau et al., 10 May 2025, Clelland et al., 2021).
• Data-driven and approximate approaches: When dynamic feedback linearizability is unobtainable or model parameters are unknown, data-driven or learning-based feedback design may seek approximate linearization but do not guarantee the exact dynamic canonical form (K. et al., 18 Aug 2025, Floren et al., 2022).
• Sampled-data and digital control: Recent developments provide methods for constructing discretizations that are dynamically feedback linearizable, extending the applicability to digital and hybrid implementations (Jindal et al., 2024).
• Limitations: Constructive dynamic feedback linearization remains intractable in general—for high-dimensional or large-input systems, the exponential growth of prolongation paths necessitates heuristic or problem-specific simplifications (D'Souza, 1 Feb 2026).

Dynamic feedback linearization constitutes a central methodology for extending the power of linear control to broad classes of nonlinear, multivariable, and high-relative-degree systems, furnishing both foundational theory and actionable synthesis algorithms grounded in geometric control, category theory, and algorithmic optimization.