Nonlinear Control: Feedback Linearization

Designing controllers for a system that behaves in a straight line is a well-solved problem with powerful tools like PID and state-space methods. But what do you do when the system is fundamentally curved, where its dynamics twist and turn in complex ways? Feedback linearization offers an elegant, algebraic approach. It applies nonlinear coordinate transformations and state feedback to convert a nonlinear system into an equivalent linear system, effectively "canceling" the troublesome nonlinearities and allowing you to apply the full arsenal of linear control design to certain classes of nonlinear systems.

The Challenge of Nonlinear Systems and the Linearization Idea

A nonlinear dynamic system can be described in state-space form as $\dot{x}=f(x)+g(x)u$, $y=h(x)$, where $x$ is the state vector, $u$ is the control input, and $y$ is the output. The functions $f(x)$ and $g(x)$ are nonlinear, making stability analysis and controller design profoundly more difficult than for linear systems. Traditional Jacobian linearization works by approximating the system's behavior around a single operating point (like an equilibrium), but this approximation degrades as the system moves away from that point.

Feedback linearization takes a radically different, geometric approach. Instead of approximating the nonlinearity, it seeks to cancel it exactly through a clever combination of a change of variables and a specially crafted control law. The goal is to make the transformed system behave exactly like a linear one, not just approximately, over a potentially large region of operation. This is achieved by finding a nonlinear state transformation $z=T(x)$ and a control law of the form $u=\alpha (x)+\beta (x)v$, where $v$ is a new, external control input. When applied, the system dynamics in the new $z$-coordinates become $\dot{z}=Az+Bv$, a completely linear system.

Input-Output Linearization

The most direct application of this idea is input-output linearization. Here, the objective is to linearize the relationship between the system output $y$ and the new input $v$, even if the internal state dynamics are not fully linearized. The key is differentiating the output until the input $u$ explicitly appears.

The number of times you must differentiate $y$ to have $u$ appear is called the relative degree $r$. For a system with relative degree $r$, the derivatives are: $$\dot{y}=L_{f}h(x)\text{(does not contain u)}$$ $$y^{(r)}=L_f^{r}h(x)+[L_g{L}_f^{r-1}h(x)]u$$ Here, $L_{f}h(x)$ is the Lie derivative of $h$ along $f$, a directional derivative that measures how $h$ changes along the flow of $f$. The term $L_g{L}_f^{r-1}h(x)$ is crucial—if it is non-zero, we can design a control law.

We then choose our control input $u$ to cancel the nonlinearities and impose linear dynamics: $$u=\frac{1}{L_g{L}_f^{r-1}h(x)}[-L_f^{r}h(x)+v]$$ Substituting this into the equation for $y^{(r)}$ yields the simple, linear relationship: $y^{(r)}=v$. You can now set $v$ to achieve desired linear behavior, such as $v=-K_{0}y-K_1\dot{y}-...-K_{r-1}{y}^{(r-1)}$, to place poles for the output dynamics. For example, controlling a single-link robotic arm involves differentiating the joint angle (output) until the motor torque (input) appears, then applying this exact cancellation law.

State-Space Exact Linearization and Coordinate Transformation

A stronger result is full state-space exact linearization. The goal here is to find a state transformation $z=T(x)$ that, when combined with state feedback $u=\alpha (x)+\beta (x)v$, renders the entire state equation $\dot{z}=Az+Bv$ linear and controllable. This is possible if and only if the system satisfies certain geometric conditions: the vector fields $g,a{d}_{f}g,...,a{d}_f^{n-1}g$ must be linearly independent and involutive (a technical condition related to being "curl-free") in a region of interest. These are known as the controllability and integrability conditions.

When satisfied, the process involves solving a set of partial differential equations to find the transformation $T(x)$. The new coordinates, $z$, are often chosen such that $z_1=h(x)$, $z_2=L_{f}h(x)$, ..., up to the relative degree. The remaining coordinates (for a system where $r<n$) complete the transformation but are not directly influenced by the new control input $v$. This leads us to the critical concept of internal dynamics.

Zero Dynamics and Internal Stability

After applying input-output linearization, the output $y$ and its derivatives (up to $r-1$) become part of a linear subsystem. However, if the relative degree $r$ is less than the system order $n$, there remain $n-r$ internal states that have been rendered unobservable from the output. These are the internal dynamics. Their stability is paramount.

To analyze them, we study the zero dynamics: the internal dynamics when the initial conditions and the control law are constrained to keep the output identically zero ($y(t)=0$). If the zero dynamics are asymptotically stable, the overall system is said to be minimum phase, and the input-output linearization controller will be successful. If the zero dynamics are unstable (a non-minimum phase system), the cancellation can cause hidden states to blow up, even while the output tracks perfectly. For instance, an aircraft's pitch rate might be linearized, but if the associated internal altitude dynamics are unstable, the plane could dive or climb catastrophically while the pitch appears controlled.

Control Design After Linearization

Once the system is transformed into a linear equivalent (either fully or just the input-output map), you can apply any linear control technique. For the linear system $\dot{z}=Az+Bv$, you can design a state feedback law $v=-Kz$ to achieve desired pole placement for stability and performance. Tracking control is also straightforward: for the linear input-output system $y^{(r)}=v$, you can design $v$ to make $y$ track a reference signal $y_d(t)$ by incorporating feedforward of $y_d^{(r)}$.

It is critical to remember that the final control law $u=\alpha (x)+\beta (x)v$ is implemented using the original states $x$, not the transformed states $z$. The transformation $T(x)$ is used in the design process, but the controller itself is a static nonlinear function of the measurable states $x$. This synthesized $u$ is what enacts the cancellation and imposes the linear behavior.

Common Pitfalls

Over-reliance on Exact Cancellation: The technique depends on perfect knowledge of the nonlinear functions $f(x)$ and $g(x)$. In reality, model inaccuracies and disturbances mean cancellation is never perfect. The residual "un-canceled" nonlinearities can degrade performance or cause instability. The correction is to design robust linear controllers (e.g., with high gain or sliding mode characteristics) for the linearized system or to incorporate adaptive elements to handle uncertainty.

Ignoring Zero Dynamics: Focusing solely on the linearized input-output behavior while neglecting the internal dynamics is a recipe for hidden failure. Always check if the system is minimum phase. For non-minimum phase systems, you must use alternative methods like output redefinition or approximate linearization that don't excite the unstable internal modes.

Misunderstanding the Relative Degree: The relative degree must be well-defined and constant in the region of operation. If the term $L_g{L}_f^{r-1}h(x)$ becomes zero at some point (a singularity), the control law blows up. Furthermore, some systems may not have a finite relative degree. The correction involves carefully analyzing the system's structure across the entire intended operating envelope before committing to this design path.

Summary

• Feedback linearization is a geometric control method that uses exact nonlinear state feedback and coordinate changes to transform a nonlinear system into an equivalent linear system for controller design.
• Input-output linearization involves repeatedly differentiating the output until the input appears, then designing a control law to cancel nonlinearities and impose linear dynamics on the input-output map.
• The relative degree determines how many derivatives are needed, and the success of the method hinges on the stability of the resulting internal dynamics, analyzed via the zero dynamics.
• A system with stable zero dynamics is minimum phase and is a good candidate for this method.
• After linearization, standard linear control techniques (pole placement, tracking) are applied to the transformed system, and the resulting nonlinear control law is implemented using the original system states.
• The approach is sensitive to model errors and requires careful attention to singularities and the definition of the relative degree across the operating region.