Differential Flatness & Geometric Parametrization

• Differential flatness is a structural property that enables a finite collection of flat outputs and their shifts to parameterize all system trajectories.
• Geometric parametrization constructs explicit algebraic maps ensuring that every trajectory in flat-output space corresponds to a dynamically feasible state-input sequence.
• This approach underpins practical design methods for motion planning, trajectory optimization, and feedback control in nonlinear, underactuated, and discrete-time systems.

Differential flatness is a structural property of control systems—both in continuous and discrete time—which guarantees the existence of a finite-dimensional geometric parametrization of all system trajectories via a collection of so-called flat outputs and a finite number of their (time-)shifts or derivatives. Geometric parametrization refers to the explicit construction of algebraic maps (typically surjective submersions) from sequences or jets of the flat output to the original state and input variables, ensuring that arbitrary reference trajectories in flat-output space correspond to dynamically feasible trajectories of the full nonlinear system. The theory underlies powerful design methodologies in motion planning, trajectory generation, and feedback control for a wide range of nonlinear and underactuated systems.

1. Foundational Definitions: Flatness, Submersions, and Shift Systems

For a discrete-time control system

$$x_{+} = f(x, u), \quad x \in \mathbb{R}^n,\, u \in \mathbb{R}^m$$

flatness is characterized by the existence of a finite collection of sequences $(y_0,\sigma y,\ldots,\sigma^r y)$, called flat coordinates (or flat outputs and their forward shifts), and a smooth submersion (map with full-rank Jacobian) $\varphi$ such that

$$\varphi : (y_0, \sigma y, \ldots, \sigma^r y) \mapsto (x, u)$$

and, for all trajectories of the trivial shift system

$$y^i_{j+} = y^i_{j+1}, \quad j=0, \ldots, r_i-1,\, i=1, \ldots, m$$

the dynamics are respected: $$\varphi_x(\sigma(y_0,\ldots,\sigma^r y)) = f(\varphi_x(y_0,\ldots,\sigma^r y), \varphi_u(y_0,\ldots,\sigma^r y))$$ The requirement that $\varphi$ be a submersion translates to the algebraic rank condition

$$\operatorname{rank}\, \frac{\partial (\varphi_x, \varphi_u)}{\partial (y_0, \sigma y, \ldots, \sigma^r y)} = n + m$$

Thus, the nonlinear system can be "flattened": all its trajectories are parametrized by a finite tuple of freely chosen sequences (the flat outputs and their shifts) (Kurt et al., 2023).

2. Geometric Parametrization and Role of Shifts

Geometric parametrization in discrete time leverages the forward shift operator ($\sigma$), which replaces the derivative in continuous-time theory. The trivial shift system of the flat outputs—sequences ($y, \sigma y, \ldots, \sigma^r y$)—plays the role of an iterated chain of unit-delay blocks, analogous to integrator chains in the continuous case. The geometric role is that the map $\varphi$ must push forward the canonical shift vector fields on y-space into the vector field described by the system dynamics in (x, u)-space.

The geometric structure is formalized by constructing towers of codistributions and their involutive distributions on the extended input space, mirroring the construction of Cartan distributions and exterior differential systems in the jet bundle formalism for ODEs. The main geometric objects are:

• Submersion $\varphi$: surjective map, full rank Jacobian, encoding the local diffeomorphism between flat coordinates and the state-input space.
• Codistributions and Cauchy-Characteristic Distributions: iteratively built to identify and split off "trivial shift parts" of the system.
• Flat Parametrization: The explicit formulas

$$\begin{align*} x &= \varphi_x(y_0,\sigma y, \ldots, \sigma^r y) \ u &= \varphi_u(y_0,\sigma y, \ldots, \sigma^r y) \end{align*}$$

provide the complete geometric (algebraic) parametrization of all trajectories (Kurt et al., 2023).

3. Necessary and Sufficient Flatness Conditions and Algorithmic Decomposition

Flatness is characterized via a normal-form decomposition. A key result is that, for a system to be flat, it must be possible to perform a coordinate transformation (a bundle-preserving diffeomorphism) to bring the system into a recursively "split" form: $$\begin{align*} \tilde{x}^+ &= G(\tilde{x}, v) \ w^+ &= \text{trivial shift of }w \end{align*}$$ Here $w$ are the variables associated with the projectable, involutive subbundles in the input space which transform as flat "directions" under shift; the remaining system in $(\tilde{x}, v)$ is of lower dimension. The process is recursively applied; at each stage, one searches for nontrivial involutive subbundles ($W$) of input directions whose pushforward is projectable by $f_*$.

An explicit algorithm is provided:

• Compute the codistribution $P_0 = \operatorname{span}\{ d f^i \}$ on $X \times U$.
• Identify involutive, projectable subbundles $W_0$, their annihilators $P_1=P_0 \cap \text{Ann}(W_0)$, and check integrability and characteristic conditions.
• Rectify distributions to put the system in "split" (normal) form; iterate the process until all inputs have been exhausted or no projectable direction can be found, in which case the system is not flat.
• At each step, closed-form coordinate transformations can be constructed by integrating the 1-forms in the codistributions or by solving small ODEs.

This algorithmic approach yields not only a flatness test but also an explicit construction of the geometric parametrization (the map $\varphi$), thus providing the full flat geometric coordinates required for trajectory planning and control (Kurt et al., 2023).

4. Comparative Analysis: Discrete-Time vs. Continuous-Time Flatness

Continuous-time flatness theory is expressed in the language of jet bundles and submersions from the flat output jet (variables $y, \dot y, \ldots, y^{(r)}$) to the original system variables, with the system ODEs being identically satisfied under arbitrary trajectories of the flat outputs and their derivatives. The critical mathematical constructs—submersions, involutive distributions, and integrability via Frobenius—are paralleled in discrete time by submersions involving shifts, shift-distribution analogues of the Cartan distribution, and towers of codistributions. The recursive splitting strategy in discrete time, with shifts replacing derivations, provides full analogues of the continuous-time geometric machinery (Kurt et al., 2023).

5. Parametrization via B-Splines, Bézier Curves, and Trajectory Optimization

The geometric parametrization inherent in flatness theory admits efficient finite-dimensional representations for trajectory design. Two primary methods are prominent:

• B-Splines: Flat outputs are parametrized as vector-valued B-splines whose control points directly encode the reference trajectory. Waypoint and smoothness constraints are linear in control points; the resultant optimization problem is a convex quadratic program with additional constraints imposed as needed (Nguyen et al., 2016).
• Bézier Curves: Bézier representations exploit convex-hull properties and their closed algebra under differentiation, addition, and multiplication. For flat systems, all system variables (states and inputs) can be written as explicit combinations of Bézier control points of the flat output and its derivatives. Systematic pulling-back of state and input constraints yields semi-algebraic feasibility sets in the flat output control point space, tractable via symbolic or sum-of-squares relaxations (Bekcheva, 2020).

These representations tightly couple the geometric parametrization enabled by flatness with tractable optimization-based planning pipelines, supporting satisfaction of stringent constraints and offering interpretability and computational tractability for high-dimensional systems.

6. Extensions: Flatness in Discrete-Time Systems Beyond Pure Forward Shifts

Recent developments generalize discrete-time flatness to allow the flat output's dependence on both forward and backward shifts, not just forward shifts. This "δ-flatness" property guarantees a one-to-one correspondence between arbitrary sequences of the flat output (including past values) and system trajectories, fully analogous to the classical continuous-time case. The extension preserves the key advantages (local controllability, geometric trajectory planning, and feedback linearizability), with geometric criteria naturally formulated in terms of shift-invariant codistributions/distributions and their involutivity (Diwold et al., 2020).

7. Broader Implications and Connections

Geometric parametrization provided by flatness unifies structural system-theoretic properties with tractable synthesis methods. It enables the decoupling of motion planning from feedback control, facilitates the embedding of complex constraints in trajectory generation, and applies to a range of model classes—continuous/discrete, nonlinear, underactuated, multi-agent. The approach provides both the theoretical underpinnings and practical algorithms for high-performance, constraint-aware, and feedback-stabilizable planning in advanced robotics and control (Kurt et al., 2023, Nguyen et al., 2016, Bekcheva, 2020).

---

References:

• "Flatness of Discrete-time Systems, a simple Approach" (Kurt et al., 2023)
• "Flatness-based nonlinear control strategies for trajectory tracking of quadcopter systems" (Nguyen et al., 2016)
• "Constraints on Nonlinear Finite Dimensional Flat Systems" (Bekcheva, 2020)
• "A Trajectory-Based Approach to Discrete-Time Flatness" (Diwold et al., 2020)