Introduction to Nonlinear Control

Core Idea

Nonlinear control addresses systems where superposition does not hold — the response to a sum of inputs is not the sum of individual responses — which means linear analysis tools (transfer functions, root locus, Bode plots) are insufficient for capturing the full range of system behavior. Phase plane analysis plots state trajectories (x₁ vs x₂) for second-order autonomous systems, revealing equilibrium points, their stability character (stable/unstable nodes, foci, saddle points, centers), and global behavior features such as limit cycles — isolated closed orbits that attract or repel nearby trajectories. Unlike linear systems, nonlinear systems can exhibit multiple equilibrium points with different stability properties, amplitude-dependent frequency and damping, jump phenomena, and sustained oscillations (limit cycles) that have no linear counterpart. The describing function method extends frequency-domain analysis to systems with a single nonlinear element (such as saturation, dead zone, or backlash) in an otherwise linear loop by approximating the nonlinearity's response to a sinusoidal input as an amplitude-dependent complex gain N(A). Limit cycles are predicted where the Nyquist plot of the linear part G(jω) intersects the curve −1/N(A). Linearization around an equilibrium point (Jacobian linearization) recovers a locally valid linear model, but its predictions apply only in a neighborhood of the operating point, and it cannot predict global phenomena like limit cycles or multiple equilibria.

How It's Best Learned

Simulate the Van der Pol oscillator and the pendulum with friction as canonical nonlinear systems, plotting phase portraits and identifying equilibrium points, limit cycles, and regions of attraction. Then add a saturation nonlinearity to a linear feedback system and use the describing function method to predict the amplitude and frequency of any resulting limit cycle, verifying against simulation. Compare the linearized model's predictions with the actual nonlinear response at small and large amplitudes to understand the limits of linearization.

Common Misconceptions

• Linearization is not wrong for nonlinear systems — it is locally valid near an equilibrium point, and most practical control design uses linearized models. The mistake is assuming the linearized model captures global behavior, limit cycles, or behavior far from the operating point.
• Limit cycles are not the same as underdamped oscillations in linear systems — linear oscillations decay to zero or grow without bound, while a limit cycle is a self-sustaining oscillation at a fixed amplitude that persists indefinitely and attracts neighboring trajectories.
• The describing function method is an approximation that assumes the nonlinearity's higher harmonics are filtered out by the linear plant (the "filtering hypothesis") — it can miss limit cycles or predict spurious ones when the plant does not sufficiently attenuate harmonics, so its predictions should always be verified by simulation.

Explainer

Every linear analysis tool you have learned — transfer functions, root locus, Bode plots — rests on superposition: the response to a scaled or summed input is the same scaling or sum of the individual responses. This property makes linear systems analytically tractable; the entire frequency-domain machinery exists because of it. Nonlinear systems break superposition. A pendulum with large-angle swings, a motor that saturates, a pneumatic valve with backlash — none of these can be captured by a single transfer function valid across all operating conditions. The same system can have qualitatively different behaviors at different amplitudes, making linear predictions misleading or flat-out wrong.

The phase plane is the primary visualization tool for second-order nonlinear systems. Instead of plotting a signal versus time, you plot the state variables against each other — typically position on one axis and velocity on the other. Each initial condition generates a trajectory, and the collection of all trajectories forms the phase portrait. Equilibrium points appear where trajectories converge (stable nodes and foci), diverge (unstable nodes and foci), or pass through a saddle point with mixed behavior. Reading a phase portrait tells you the global behavior of the system without solving the differential equation: you can see which initial conditions lead to stable operation, which lead to divergence, and whether there are self-sustaining oscillations. Unlike linear systems, which have at most one equilibrium, nonlinear systems can have multiple equilibria — each potentially with different stability character.

Limit cycles are isolated closed orbits in the phase plane — trajectories that form a closed loop and attract (stable limit cycle) or repel (unstable limit cycle) nearby trajectories. The Van der Pol oscillator is the canonical example: a system with a nonlinear damping term that acts as negative damping at small amplitudes (driving the oscillation up) and positive damping at large amplitudes (pulling it back), settling to a stable limit cycle at a fixed amplitude regardless of initial conditions. Linear systems cannot produce this behavior: a linear underdamped system oscillates at constant amplitude if started exactly on the orbit, but does not attract or repel nearby trajectories. Limit cycles explain why many physical systems — heart rhythms, relaxation oscillators, and wind-induced structural vibrations — sustain oscillations at a fixed amplitude.

The describing function method extends frequency-domain thinking to a single-nonlinearity system. If the linear part of the loop attenuates harmonics well (the "filtering hypothesis"), the nonlinearity's response to a sinusoidal input A·sin(ωt) can be approximated by its fundamental harmonic component, characterized by an amplitude-dependent complex gain N(A). Limit cycles are predicted where the Nyquist curve of the linear part G(jω) intersects the curve −1/N(A) in the complex plane — directly analogous to the Nyquist stability criterion. The intersection point predicts the frequency and amplitude of the oscillation. The method is an approximation, but it gives a tractable analytical answer for systems where simulation alone would require sweeping many initial conditions.

Finally, Jacobian linearization provides the bridge back to your existing toolkit. Near any equilibrium point, a differentiable nonlinear system can be approximated by a linear system whose matrix is the Jacobian of the vector field evaluated at that point. If the Jacobian's eigenvalues have negative real parts, the equilibrium is locally asymptotically stable — and your linear control design methods apply in a neighborhood around that operating point. The critical word is "locally": the linearized model may predict stability near the equilibrium while the nonlinear system has unstable trajectories further away (a saddle), or it may miss a limit cycle that exists at larger amplitudes. Jacobian linearization is indispensable for practical design, but the phase plane and describing function methods remain necessary to understand what happens beyond the small-signal regime.