8. Nonlinear Systems and Control

Introduction and Context

Linear control theory provides powerful tools for modelling, analysing, and designing feedback systems. Superposition, transfer functions, eigenvalue methods, frequency response, and state-space techniques make many engineering problems tractable. However, physical systems are rarely linear over their complete operating range. Actuator saturation, friction, backlash, dead zones, switching devices, aerodynamic forces, magnetic saturation, chemical reaction kinetics, and geometric constraints all introduce nonlinear behaviour.

A nonlinear system is one in which the relationship between inputs, states, and outputs does not satisfy the superposition principle. This is not merely a mathematical inconvenience. Nonlinearity can produce multiple equilibria, limit cycles, finite escape, bifurcations, hysteresis, subharmonic oscillations, and responses that depend strongly on initial conditions. A linear approximation may accurately describe behaviour near one operating point while failing completely during large disturbances or near constraints.

Nonlinear control therefore requires both analytical insight and engineering judgement. This chapter develops three classical analysis tools—phase-plane analysis, describing functions, and Lyapunov methods—and then introduces important design approaches: feedback linearization, sliding-mode control, and nonlinear stabilization. The emphasis is on understanding what each method can establish, what assumptions it requires, and where its conclusions should not be overextended.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish linear and nonlinear system behaviour and identify common sources of nonlinearity.

2. Analyse second-order systems using phase portraits, nullclines, and equilibrium classification.

3. Explain the purpose and limitations of describing-function analysis.

4. Apply Lyapunov’s direct method and LaSalle’s invariance principle to stability questions.

5. Formulate feedback-linearizing controllers and recognize the conditions for exact linearization.

6. Explain sliding-mode control, reaching conditions, robustness, and chattering.

7. Compare nonlinear stabilization strategies in terms of performance, robustness, measurement, and implementation.

8. Evaluate nonlinear control designs in practical and safety-critical settings.

Where Linear Models Cease to Be Reliable

A nonlinear state model is commonly written as

\[

\dot{x}=f(x,u,t), \qquad y=h(x,u,t),

\]

where the state vector is \(x\), the control input is \(u\), and the output is \(y\). In an autonomous system, explicit dependence on time is absent. An equilibrium \((x_e,u_e)\) satisfies

\[

f(x_e,u_e)=0.

\]

If the functions are differentiable, a local linear model about an equilibrium is obtained from the Jacobians:

\[

\delta\dot{x}=A\delta x+B\delta u,

\]

where \(A=\partial f/\partial x\) and \(B=\partial f/\partial u\) evaluated at the equilibrium. This model describes infinitesimal or sufficiently small deviations, not necessarily the global behaviour.

For example, a pendulum satisfies

\[

\ddot{\theta}+\frac{g}{l}\sin\theta=0.

\]

Replacing \(\sin\theta\) by \(\theta\) gives a useful small-angle model near the downward equilibrium. It cannot describe the upright equilibrium, full rotations, or the energy variation associated with large angles. Similarly, a motor model with saturated voltage has dynamics that change when the actuator reaches its limit. A controller designed using an unconstrained linear model may demand an impossible input and create integrator windup or sluggish recovery.

Important nonlinear phenomena include:

• Multiple equilibria: different operating states may coexist, with stability depending on the state.
• Limit cycles: isolated closed trajectories may represent sustained oscillations.
• Bifurcation: a qualitative change in equilibria or periodic solutions occurs as a parameter varies.
• Hysteresis: the output depends on the direction or history of input variation.
• Finite escape: a state may become unbounded in finite time in an ideal mathematical model.
• State-dependent controllability: an input may influence the system effectively in one region but weakly or not at all in another.

These effects explain why local pole placement or frequency-domain margins alone cannot guarantee satisfactory nonlinear operation.

Phase-Plane Analysis: Geometry Before Algebra

For a second-order autonomous system, define \(x_1=x\) and \(x_2=\dot{x}\). The model becomes

\[

\dot{x}_1=f_1(x_1,x_2),\qquad \dot{x}_2=f_2(x_1,x_2).

\]

The phase plane is the \((x_1,x_2)\) plane. Each point has a vector \((\dot{x}_1,\dot{x}_2)\), and a trajectory is the path followed by the state. Unlike a time plot, a phase portrait reveals the geometry of motion and the influence of initial conditions.

An equilibrium occurs where both derivatives vanish. The curves \(\dot{x}_1=0\) and \(\dot{x}_2=0\) are called nullclines. Their intersections identify equilibria; their positions divide the plane into regions in which the direction of motion can be inferred. Trajectories may approach a stable node or focus, leave an unstable equilibrium, or follow a saddle separatrix.

For a linearized system \(\dot{x}=Ax\), eigenvalues provide a local classification. Negative real eigenvalues generally indicate a stable node, complex eigenvalues with negative real parts indicate a stable focus, and eigenvalues of opposite sign indicate a saddle. Purely imaginary eigenvalues require nonlinear analysis because linearization may be inconclusive. In nonlinear systems, these classifications are local unless stronger arguments establish a region of attraction.

Consider the damped pendulum

\[

\dot{\theta}=\omega,

\qquad

\dot{\omega}=-\frac{g}{l}\sin\theta-c\omega.

\]

The downward positions \(\theta=2k\pi,\omega=0\) are stable equilibria, while \(\theta=(2k+1)\pi,\omega=0\) are unstable in the absence of special control. The energy

\[

V(\theta,\omega)=\frac{1}{2}ml^2\omega^2+mgl(1-\cos\theta)

\]

satisfies \(\dot V=-ml^2c\omega^2\leq0\). Thus energy decreases through damping. The closed energy contours of the undamped pendulum become spiralling trajectories when damping is introduced. The boundary separating oscillations from complete rotations is associated with the separatrix of the undamped system.

Phase-plane methods are especially useful for visualizing attraction regions, saturation-induced trajectories, and oscillations. Their principal limitation is dimensionality: direct visualization becomes difficult beyond two states, although projection, simulation, and invariant-set methods remain useful.

Describing Functions and Nonlinear Oscillation

The describing-function method approximates a static nonlinear element by its first harmonic response to a sinusoidal input. Suppose a nonlinearity receives

\[

v(t)=A\sin\omega t.

\]

Its output is generally nonsinusoidal. The fundamental component can be represented as

\[

y_1(t)=N(A)A\sin(\omega t+\phi(A)),

\]

where \(N(A)\) is the amplitude-dependent describing function. For an odd symmetric nonlinearity, the phase is often zero and \(N(A)\) is real.

When a nonlinear element is placed in a linear feedback loop, a predicted limit cycle can be found approximately from the harmonic-balance equation

\[

1+G(j\omega)N(A)=0.

\]

Graphically, the Nyquist locus of the linear part is compared with \(-1/N(A)\). An intersection predicts an oscillation amplitude and frequency. For a relay with output levels \(\pm M\), an ideal symmetric describing function is approximately

\[

N(A)=\frac{4M}{\pi A}.

\]

The method is valuable in control systems containing relays, saturation, dead zones, and other static nonlinearities. It can explain why a feedback loop that is stable under linear analysis develops a sustained oscillation after a relay or quantizer is introduced.

However, a describing function is not an exact stability theorem. It neglects higher harmonics and assumes that the rest of the loop filters them sufficiently. It may miss subharmonic oscillations, non-sinusoidal periodic orbits, quasiperiodic motion, and interactions between several nonlinearities. Predictions should therefore be verified through simulation, phase-plane analysis, experiments, or more rigorous tools.

Lyapunov Stability: An Energy-Like View

Lyapunov methods assess stability without explicitly solving the differential equations. For an equilibrium shifted to the origin, consider a scalar function \(V(x)\). A positive-definite function satisfies \(V(0)=0\) and \(V(x)>0\) for all nonzero \(x\) in a specified region. Along system trajectories,

\[

\dot V(x)=\nabla V(x)^T f(x).

\]

If \(V\) is positive definite and \(\dot V\leq0\), the equilibrium is stable in the sense of Lyapunov. If \(\dot V<0\) away from the origin, asymptotic stability follows under standard regularity assumptions. If both \(V\) and \(-\dot V\) are radially unbounded, global asymptotic stability may be established.

The distinction between stability and asymptotic stability is essential. Stability means that sufficiently small initial errors remain small. Asymptotic stability additionally requires convergence to the equilibrium. If \(\dot V\) is only negative semidefinite, convergence cannot automatically be claimed. LaSalle’s invariance principle resolves many such cases: trajectories approach the largest invariant set contained in \(\{x:\dot V=0\}\).

For a mechanical system, kinetic plus potential energy often provides a natural candidate. For a linear system, a quadratic candidate \(V=x^TPx\), with \(P=P^T>0\), leads to the Lyapunov equation

\[

A^TP+PA=-Q,

\]

where \(Q>0\). The existence of a positive-definite solution for every \(Q>0\) is equivalent to Hurwitz stability of \(A\). Nonlinear analysis extends this idea, but constructing a useful function is often the principal challenge.

Control Lyapunov functions (CLFs) generalize this reasoning to design. If, for every nonzero state, at least one admissible input makes \(\dot V<0\), a feedback law may be selected to stabilize the system. Constraints, input saturation, uncertainty, and actuator limits must be included; otherwise a mathematically valid CLF design may be physically unrealizable.

Nonlinear Stabilization Strategies

Feedback Linearization

Feedback linearization cancels selected nonlinear terms through state feedback and, where necessary, a change of coordinates. For a single-input system in normal form,

\[

y^{(r)}=a(x)+b(x)u,

\]

with \(b(x)\neq0\), choose

\[

u=\frac{-a(x)+v}{b(x)}.

\]

Then \(y^{(r)}=v\). A linear design such as

\[

v=y_d^{(r)}-k_{r-1}e^{(r-1)}-\cdots-k_1\dot e-k_0e

\]

can produce a desired error polynomial. The method is conceptually direct and can achieve excellent tracking when the model is accurate.

Its risks are equally important. Exact cancellation fails when parameters are uncertain, states are poorly measured, or \(b(x)\) approaches zero. Internal or zero dynamics must be stable; otherwise input-output linearization can hide unstable state behaviour. High-gain cancellation may amplify noise and require excessive control effort. Practical designs often combine nominal cancellation with robust or adaptive terms rather than relying on exact knowledge.

Sliding-Mode Control

Sliding-mode control defines a switching surface, commonly

\[

s(x)=\dot e+\lambda e,

\qquad \lambda>0.

\]

The control is selected so that trajectories reach \(s=0\) and remain there. A representative law is

\[

u=u_{eq}-K\operatorname{sgn}(s),

\]

where \(u_{eq}\) maintains ideal motion on the surface and the discontinuous term overcomes bounded uncertainty. With \(V=\tfrac12s^2\), a reaching condition such as \(s\dot s\leq-\eta|s|\) establishes finite-time approach in the ideal model.

Sliding control is robust to matched uncertainties—uncertainties entering through the same channel as the control. It is not automatically robust to unmatched disturbances. The discontinuity also causes chattering when implemented with finite sampling, actuator dynamics, sensor noise, and unmodelled high-frequency modes. Boundary layers, saturation approximations, higher-order sliding modes, and observer-based implementations reduce chattering but may weaken exact robustness.

Backstepping and Nonlinear Damping

Backstepping constructs stabilizing virtual controls recursively for strict-feedback systems. At each stage, a Lyapunov function is extended to include the next tracking error. The approach handles certain nonlinearities without cancellation and can accommodate adaptive parameter estimates. Its drawbacks include algebraic complexity, possible explosion of terms, and sensitivity to unmodelled dynamics.

A simpler strategy is nonlinear damping injection. If the system has a passive or mechanical structure, a controller can add damping while shaping the potential energy. This approach is often less aggressive than exact cancellation and may provide superior robustness, although it may not achieve the fastest tracking response.

An Illustrative Applied Case: Stabilization of an Inverted Pendulum

Background

An inverted pendulum mounted on a cart is a standard model for balancing, robotics, and transportation systems. The upright position is unstable, and the cart actuator is subject to force limits. The system is nonlinear because of trigonometric geometry and because the actuator cannot provide unlimited force.

Situation and Analysis

A linear model around the upright position can support local pole placement, but its validity decreases as the pendulum angle grows. A controller designed only for that local model may recover small disturbances yet fail after a large push. Moreover, the cart may reach its travel boundary while the controller continues demanding corrective motion.

A nonlinear model can be expressed schematically as

\[

M(q)\ddot q+C(q,\dot q)\dot q+g(q)=Bu,

\]

where \(q\) contains cart position and pendulum angle. The energy associated with the pendulum provides a natural variable for a swing-up controller, while a separate stabilizer is used near the upright equilibrium.

Strategy

A practical hybrid design has two modes. During swing-up, an energy-based controller injects or removes energy to move the pendulum toward the desired energy level. Near the upright position, a local stabilizer—possibly feedback linearization, linear quadratic control, or a Lyapunov-based nonlinear law—takes over. Hysteresis or a carefully designed switching region prevents rapid mode changes.

The controller must include cart-force saturation, state estimation, and anti-windup measures. A safety supervisor can stop the experiment if the cart approaches its travel limit or if the estimated angle becomes unreliable. Simulation should test parameter variation, sensor noise, delayed sampling, friction, and disturbances rather than only the nominal model.

Outcome and Learning Points

The central outcome is not merely upright regulation but reliable operation across the complete manoeuvre. The case demonstrates that no single local linear controller is sufficient for a large-region task. It also illustrates the value of combining energy methods, local stabilization, switching logic, constraints, and supervisory protection.

The main lessons are:

• local stability is not equivalent to global recovery;
• actuator and workspace limits must enter the design from the beginning;
• switching surfaces require hysteresis and verification;
• simulations should include uncertainty and implementation effects;
• safety constraints are part of control design, not an afterthought.

Practical Applications and Engineering Responsibilities

Nonlinear control is used in robotic manipulators, aircraft and spacecraft, electric drives, autonomous vehicles, process plants, power converters, magnetic bearings, biomedical devices, and renewable-energy systems. In robotics, gravity compensation and feedback linearization can improve trajectory tracking. In power electronics, switching nonlinearities and saturation make averaging, hybrid modelling, and robust control important. In aerospace systems, aerodynamic coefficients vary with speed and angle, while actuator authority may be limited.

Industrial implementation requires more than proving a theorem. Engineers must examine sensor resolution, sampling time, computational delay, quantization, actuator bandwidth, thermal limits, fault modes, and maintainability. A discontinuous controller that is elegant in continuous time may excite structural modes in hardware. A nonlinear observer may become unreliable when measurements are temporarily lost. Certification-oriented industries require traceability of assumptions, test coverage, fail-safe behaviour, and clearly defined operating envelopes.

Digital twins, real-time simulation, automatic differentiation, and data-driven model identification increasingly support nonlinear control development. Machine-learning models can approximate unknown dynamics, but they should normally be combined with physical constraints, uncertainty bounds, monitoring, and fallback controllers. Black-box prediction alone does not provide a stability guarantee. Control barrier functions are also being used to enforce state safety, while control Lyapunov functions promote performance; their combination leads to optimization-based controllers that balance safety and regulation.

Limitations, Risks, and Best Practices

Common mistakes include treating a local linear model as a global model, ignoring unstable internal dynamics, cancelling uncertain nonlinearities exactly, assuming negative-semidefinite Lyapunov derivatives imply asymptotic convergence, and interpreting a describing-function prediction as proof of a limit cycle. Another frequent error is omitting actuator saturation from both analysis and simulation.

A disciplined nonlinear-control workflow is as follows:

1. Define the operating region, constraints, disturbances, and performance requirements.

2. Derive a physically meaningful nonlinear model and identify uncertain parameters.

3. Locate equilibria and examine local linearizations, but do not stop there.

4. Select an analysis tool appropriate to the question: phase plane for geometry, describing functions for approximate harmonic oscillations, and Lyapunov methods for rigorous stability claims.

5. Design with measurement, actuation, sampling, and computation limits included.

6. Establish regions of attraction, robustness margins, and failure responses.

7. Validate using nonlinear simulation, worst-case tests, hardware-in-the-loop, and controlled experiments.

8. Document assumptions and define the conditions under which the guarantee applies.

Ethically and professionally, engineers must not report local or nominal performance as universal safety. A controller should transition to a safe state when sensors, actuators, or computational resources fail. Cybersecurity is also relevant: malicious manipulation of state estimates or control commands can drive a nonlinear system outside its verified region. Governance therefore includes access control, logging, independent testing, change management, and clear responsibility for deployment decisions.

Current Developments

Research and practice increasingly combine classical nonlinear control with optimization, adaptive control, learning, and formal verification. Model predictive control handles constraints explicitly and can use nonlinear prediction models, although online computation and recursive feasibility remain concerns. Differential-geometric methods extend feedback linearization to more complex systems. High-order sliding modes reduce chattering, while finite-time and fixed-time methods provide stronger convergence specifications.

Data-driven and reinforcement-learning controllers are being investigated for systems whose dynamics are difficult to model. The most credible engineering direction is often a hybrid one: physical models provide structure and safety arguments, while data improve uncertain terms or operating efficiency. Stability certificates, uncertainty quantification, runtime monitoring, and interpretable fallback policies are essential when learning components influence safety-critical systems.

Chapter Summary

Nonlinear systems violate superposition and can display multiple equilibria, limit cycles, bifurcations, hysteresis, and state-dependent behaviour. Linearization remains useful, but it is local and must not be mistaken for a global description. Phase-plane analysis reveals trajectories and attraction regions for second-order systems. Describing functions provide approximate predictions of harmonic oscillations but are not rigorous stability proofs. Lyapunov methods establish stability through energy-like functions, with LaSalle’s principle handling negative-semidefinite derivatives.

Feedback linearization cancels nonlinearities under restrictive modelling and invertibility conditions. Sliding-mode control offers robustness to matched uncertainty but introduces chattering and implementation concerns. Backstepping, damping injection, energy shaping, and hybrid control provide alternative strategies. Successful nonlinear control integrates mathematical analysis with constraints, uncertainty, sensing, computation, safety, and validation.

Key Terms

Nonlinear system: A system that does not satisfy the superposition principle.

Equilibrium: A state and input pair for which the state derivative is zero.

Phase portrait: A geometric representation of state trajectories in state space.

Nullcline: A locus on which one component of the state derivative is zero.

Limit cycle: An isolated periodic trajectory of a dynamical system.

Bifurcation: A qualitative change in system behaviour caused by parameter variation.

Linearization: Approximation of nonlinear dynamics by first-order Taylor expansion near an operating point.

Describing function: An amplitude-dependent first-harmonic approximation of a nonlinear element.

Harmonic balance: An approximate method for determining periodic responses by equating fundamental harmonics.

Lyapunov function: A scalar function used to assess stability through its value and derivative.

Positive definite: A function that is zero at the equilibrium and positive elsewhere in a specified region.

LaSalle’s invariance principle: A result establishing convergence to the largest invariant subset where a nonpositive Lyapunov derivative is zero.

Control Lyapunov function: A function for which an admissible control can produce stabilizing decrease.

Feedback linearization: Nonlinear feedback and coordinate transformation used to obtain linear closed-loop dynamics.

Relative degree: The number of differentiations of an output required before the input appears explicitly.

Zero dynamics: Internal dynamics that remain when the controlled output is constrained to zero.

Sliding surface: A state-space manifold toward which sliding-mode control drives the system.

Chattering: High-frequency switching caused by practical implementation of discontinuous control.

Region of attraction: The set of initial states that converge to a specified equilibrium.

Matched uncertainty: Uncertainty entering through the same input distribution as the control.

Review Questions

Short Answer Questions

1. Why can a linearized model fail to predict the behaviour of a nonlinear system after a large disturbance?

2. What information is obtained from the intersections of nullclines?

3. State the harmonic-balance equation used with a describing function.

4. Distinguish stability from asymptotic stability in Lyapunov analysis.

5. What is chattering, and why does it occur in sliding-mode control?

Long Answer Questions

1. Explain phase-plane analysis for a second-order nonlinear system, including equilibria, nullclines, trajectories, and separatrices.

2. Derive the basic Lyapunov stability conditions and explain how LaSalle’s invariance principle extends them.

3. Explain feedback linearization, relative degree, zero dynamics, and the practical consequences of model uncertainty.

4. Discuss the advantages and limitations of describing-function analysis for predicting limit cycles.

5. Compare feedback linearization, sliding-mode control, backstepping, and energy-based stabilization for an uncertain mechanical system.

Analytical / Application Questions

1. For a damped pendulum, propose an energy-based Lyapunov function, calculate its derivative, and state what conclusion can be drawn about stability and convergence.

2. A relay is inserted into a feedback loop and sustained oscillation is observed. Describe how a describing-function calculation could estimate its amplitude and frequency, and identify two reasons the estimate may be inaccurate.

3. Design a control architecture for an inverted pendulum that must recover from large angles while respecting actuator and cart-position limits. Justify the use of multiple control modes, safety supervision, and nonlinear analysis.

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes the principal limitation of a local linearization?

A. It cannot be used to analyse any equilibrium point.

B. It describes only behaviour sufficiently near the selected operating point.

C. It always predicts a larger stability region than the nonlinear model.

D. It removes actuator saturation from the physical system.

Correct Answer: B. It describes only behaviour sufficiently near the selected operating point.

Question 2

In a phase plane, an equilibrium point is located where:

A. The trajectory has maximum curvature.

B. The output reaches its largest value.

C. All components of the state derivative are zero.

D. The input changes sign.

Correct Answer: C. All components of the state derivative are zero.

Question 3

The describing-function method primarily approximates a nonlinear element by its:

A. DC component only.

B. First harmonic response to a sinusoidal input.

C. Exact state-space realization.

D. Worst-case transient response.

Correct Answer: B. First harmonic response to a sinusoidal input.

Question 4

Which equation represents the usual harmonic-balance condition for a nonlinear feedback loop?

A. G(jω)+N(A)=0

B. G(jω)N(A)=1

C. 1+G(jω)N(A)=0

D. G(jω)/N(A)=0

Correct Answer: C. 1+G(jω)N(A)=0

Question 5

If a positive-definite Lyapunov function has a negative-definite derivative, the equilibrium is generally:

A. Asymptotically stable locally.

B. Necessarily unstable globally.

C. A limit cycle.

D. Unobservable.

Correct Answer: A. Asymptotically stable locally.

Question 6

Why does a negative-semidefinite Lyapunov derivative require additional reasoning?

A. It proves finite escape immediately.

B. It does not by itself prove convergence to the equilibrium.

C. It is incompatible with positive-definite functions.

D. It applies only to linear systems.

Correct Answer: B. It does not by itself prove convergence to the equilibrium.

Question 7

Feedback linearization of y^(r)=a(x)+b(x)u requires, in the basic form, that:

A. b(x) be nonzero in the operating region.

B. a(x) be identically zero.

C. the system have no internal dynamics.

D. the input be discontinuous.

Correct Answer: A. b(x) be nonzero in the operating region.

Question 8

Sliding-mode robustness is most directly associated with uncertainties that are:

A. Matched with the control input channel.

B. Unmeasurable under every condition.

C. Located only in stable zero dynamics.

D. Introduced by numerical integration.

Correct Answer: A. Matched with the control input channel.

Question 9

A major practical disadvantage of an ideal discontinuous sliding-mode law is:

A. It cannot reject any disturbance.

B. It produces chattering in the presence of implementation imperfections.

C. It requires the plant to be linear.

D. It eliminates all measurement noise.

Correct Answer: B. It produces chattering in the presence of implementation imperfections.

Question 10

Why is a hybrid strategy appropriate for an inverted pendulum recovering from large angles?

A. A single local linear controller is always globally valid.

B. Energy-based swing-up and local upright stabilization address different operating regions.

C. Hybrid control avoids all actuator constraints.

D. Phase-plane analysis is unnecessary for hybrid systems.

Correct Answer: B. Energy-based swing-up and local upright stabilization address different operating regions.