Nonlinear Control and Lyapunov Methods

Core Idea

Linear control (pole placement, LQR, H-infinity) is inadequate for systems with significant nonlinearities: saturation, friction, unmodeled stiffness, or dynamics where equilibrium motion and linearization break down. Lyapunov stability theory provides a nonlinear framework: you design a Lyapunov function (energy-like measure) and a feedback law that drives the function to decrease along all trajectories, guaranteeing asymptotic stability without linearization. Feedback linearization cancels nonlinearities by algebraic (input-output) or (differential-geometric) transformations; backstepping recursively stabilizes subsystems; passivity-based control exploits the structure of conservative systems. These methods guarantee stability for the true nonlinear system, not just a linearization.

How It's Best Learned

Design a stabilizing controller for a simple nonlinear system (pendulum, magnetic levitation, or a nonlinear spring-mass) using Lyapunov direct method: propose a Lyapunov function V (often kinetic + potential energy or a quadratic form), compute dV/dt in terms of state and control input, choose u to make dV/dt negative definite, and verify that trajectories from any initial condition converge to the origin. Compare with linearized (LQR) control: observe that LQR works only near equilibrium, whereas the Lyapunov-designed controller stabilizes from arbitrary initial conditions.

Common Misconceptions

• A Lyapunov function always exists if the system is stabilizable; existence is not guaranteed — finding one requires insight, and for general nonlinear systems it is computationally hard (Sum-of-Squares methods attempt it but scale poorly).
• If a Lyapunov function V exists with dV/dt < 0 everywhere, the system is globally exponentially stable; you get asymptotic stability at best, not exponential unless special conditions hold.
• Feedback linearization is always better than linear control because it handles nonlinearity exactly; feedback linearization requires knowing the system model precisely and can amplify measurement noise due to differentiation in the control law.

Explainer

From linear control, you know how to place poles, design lead-lag compensators, and ensure stability margins. But these tools rely on a linear model — once you linearize around an operating point, they lose validity beyond a small neighborhood. Real systems have saturation (actuators can only push so hard), friction (often proportional to sign(velocity) rather than velocity), gravity (affecting equilibrium), and unmodeled stiffness or hysteresis. For these systems, linear control can fail: an LQR controller tuned for small disturbances may oscillate wildly or saturate under large ones.

Lyapunov stability theory provides a nonlinear alternative. Rather than analyzing transfer functions and Bode plots, you work directly with the state equations ẋ = f(x, u). The central idea is Lyapunov's second method: a function V(x) (analogous to energy) that is positive definite at the origin and decreases along all system trajectories guarantees that trajectories converge to the origin. If you can design a feedback law u = k(x) such that dV/dt < 0 everywhere, you've proven the system is asymptotically stable without ever computing eigenvalues or frequency responses.

Finding the right Lyapunov function is the art: for a mechanical system, use kinetic + potential energy; for an electrical system, use magnetic energy stored in inductors; for a general nonlinear system, guess a quadratic form and verify. Once you have a candidate V, compute dV/dt = ∇V ⊤ f(x, k(x)) in terms of the control input, and choose k(x) to make dV/dt negative. This direct method is elegant: stability is guaranteed by construction, and the feedback law is often nonlinear in structure (e.g., proportional to state or nonlinear functions of state) in ways that linear pole placement cannot express.

Feedback linearization is a more ambitious approach: choose the control law to *cancel* the nonlinearities and make the closed-loop system linear. For a system with input affine structure (ẋ = f(x) + g(x)u), if the system has "relative degree" (loosely, how many times you must differentiate the output to see the input directly), you can choose u to achieve exact linearization. The control law will be nonlinear, but the closed-loop system behaves as a linear transfer function. The catch: this requires knowing f and g accurately, and if the model is wrong or measurement is noisy (especially if the control law involves derivatives), performance degrades sharply.

Backstepping recursively builds up a stabilizing controller for cascade-like systems. Suppose you have ż = f(z) + g(z)ζ and ζ̇ = u. You first stabilize the z-subsystem by treating ζ as a "virtual control" and designing ζ = α(z). Then, you augment the Lyapunov function to include a penalty for |ζ − α(z)| and design the real input u to drive this error to zero while maintaining z-stability. This recursive approach scales to higher-order systems and is particularly effective for systems with a cascade or hierarchical structure (e.g., attitude control for aircraft, where you first stabilize roll and pitch, then impose commanded yaw).

Passivity-based control leverages energy structure. If a system is passive (energy non-increasing), it has a natural tendency toward stable equilibrium. Connecting a passive controller (one that only dissipates energy) to a passive system guarantees stability by the passivity theorem. This is why simple proportional or damping-like feedback stabilizes mechanical systems: the system and controller are both passive, energy always flows from input to dissipation, and no instability loops can form. Modern examples include impedance control in robotics (designing the robot to behave like a passive mechanical impedance) and power systems (ensuring power converters are passive to avoid islanding instability).

The trade-off between nonlinear methods is sophistication vs. robustness. Feedback linearization gives exact linearization but is fragile to model error. Lyapunov direct method proves stability but requires finding the right V (hard for high-dimensional systems) and the resulting controller may not optimize a performance objective — you get stability, not optimality. Backstepping is systematic but requires the system to have a specific structure. Passivity methods are robust to modeling error but apply mainly to conservative systems. Modern practice combines them: use passivity if available, augment with backstepping for additional variables, validate with Lyapunov analysis, and finally test robustness against model uncertainty. For systems with no clear structure or high dimensionality, learning-based approaches (neural network controllers trained in simulation) are increasingly complementing analytical design.