grokkingstuff Home Blog Projects Wiki Calculators About

Nonlinear & Adaptive Control

date2026-07-24tags:ele: :control:

The whole of classical control — root locus, Bode plots, Nyquist, LQR, the Routh-Hurwitz criterion — rests on a single fiction: that the plant is linear time-invariant, $\dot{x} = Ax + Bu$, $y = Cx + Du$. This fiction is extraordinarily productive precisely because linear systems admit a complete, constructive theory: superposition holds, the Laplace transform diagonalizes convolution, and stability reduces to the location of eigenvalues. But it is a fiction. Real plants have saturating actuators, Coulomb friction, dead zones, backlash, exponential growth that clips, and state-dependent dynamics. The question nonlinear control answers is not "is the linear theory wrong?" — it obviously is — but "how much of the linear theory survives, and what must be rebuilt from scratch?"

The local-global gap

Hartman and Grobman proved, independently and nearly simultaneously in 1959–60, that near a hyperbolic equilibrium $x^*$ of $\dot{x} = f(x)$ (an equilibrium where $Df(x^*)$ has no eigenvalue with zero real part), the nonlinear flow is topologically conjugate to its linearization. That is, there exists a homeomorphism $h$ such that $h(\phi_t(x)) = e^{At} h(x)$, with $A = Df(x^*)$. The phase portrait near $x^*$ is, up to continuous deformation, the phase portrait of the linearized system. This is a remarkable theorem: it says linearization is not merely an approximation but a qualitatively exact description in a neighborhood of any hyperbolic fixed point.

[fn:: Hartman-Grobman is local in the strongest sense: the conjugacy holds on some neighborhood of $x^*$ whose size is not specified by the theorem and is often impractically small. Grobman's proof preceded Hartman's by a year but circulated less widely; the theorem is sometimes credited to Poincaré, who had the right intuition decades earlier but lacked the topological machinery. The requirement of hyperbolicity is essential — at a non-hyperbolic equilibrium (e.g. a center), the linearization and the nonlinear flow can differ qualitatively, which is exactly why one cannot casually linearize through a Hopf bifurcation.]

The catch is "neighborhood." A satellite's attitude dynamics are locally linearizable around any operating point, but the operating point moves across the full attitude manifold. A power amplifier's gain is linear for small signals and compresses for large ones. The plant may be stable at every frozen operating point and yet exhibit pathological behavior under transitions. The gap between local theory and global operation — between "there exists a neighborhood" and "the plant operates over the full state space" — is the entire subject matter of nonlinear control.

Lyapunov stability

Aleksandr Lyapunov's 1892 doctoral dissertation introduced two methods for analyzing the stability of $\dot{x} = f(x)$ at an equilibrium $x^* = 0$. The indirect method linearizes: compute $A = Df(0)$, check eigenvalues. If all eigenvalues have strictly negative real part, $x^*$ is asymptotically stable; if any has strictly positive real part, it is unstable. The indeterminate case (any eigenvalue with zero real part) is genuinely nonlinear and the linearization is inconclusive. The indirect method is, in effect, Hartman-Grobman turned into a computational recipe — and it shares the same locality caveat.

The direct method is more powerful and more frustrating. Find a continuously differentiable function $V: \mathbb{R}^n \to \mathbb{R}$ such that:

negative definite for asymptotic stability).

Then $x^* = 0$ is stable (or asymptotically stable if $\dot{V} < 0$). The theorem is beautiful because it makes no reference to linearization — it is a genuinely nonlinear result. It is also, in a precise sense, not constructive: Lyapunov's theorem says that if such a $V$ exists, then stability follows, but it provides no procedure for finding one. For physical systems with an energy function (mechanical, electrical, Hamiltonian), the energy itself is a natural candidate, and $\dot{V}$ is then dissipation. For arbitrary nonlinear systems, the search for $V$ is the hard part.

[fn:: The analogy to program verification is instructive and not merely decorative. Hoare-logic correctness of a loop requires an invariant $I$ such that $I \land G \land \text{body} \Rightarrow I'$ and $I \land \neg G \Rightarrow \text{postcondition}$. The invariant is the Lyapunov function: it is the object whose existence makes the proof go through, and discovering it is the irreducibly creative step. Both fields have the same structure — a well-specified verification condition whose witnesses must be conjured into existence by the engineer. This is why "just find a Lyapunov function" is as unhelpful as "just find the invariant": the theorem is sound, the bottleneck is synthesis. Automation (sum-of-squares programming, SMT-based invariant generation) has helped at the margins but not closed the gap.]

A worked example: the scalar system $\dot{x} = -x^3$. The linearization is $\dot{x} = 0$, which is inconclusive (zero eigenvalue, non-hyperbolic). Try $V(x) = \tfrac{1}{2} x^2$. Then $\dot{V} = x \dot{x} = -x^4 \leq 0$, and $\dot{V} < 0$ for $x \neq 0$. So $x = 0$ is asymptotically stable — a conclusion the linearization could not deliver. The convergence is not exponential ($\dot{x} \sim -x^3$ gives $x(t) \sim t^{-1/2}$), which linear theory would have flagged immediately had it applied; this is a genuinely nonlinear decay rate.

LaSalle's invariance principle

The requirement $\dot{V} < 0$ is often too strong. Consider the undamped pendulum $\ddot{\theta} + \sin\theta = 0$ with total energy $V = \tfrac{1}{2}\dot{\theta}^2 - \cos\theta$: $\dot{V} = 0$ identically (energy is conserved), so the direct method with strict inequality gives nothing — yet we know the pendulum orbits the equilibrium. LaSalle's invariance principle (1960, contemporaneous with but independent of Barbălat's lemma) relaxes the requirement: if $\dot{V} \leq 0$ and the largest invariant set within $\{x : \dot{V}(x) = 0\}$ is $\{0\}$, then $x = 0$ is asymptotically stable.

This is the practical version of Lyapunov theory. Most physical Lyapunov candidates are energies or pseudo-energies whose derivative vanishes on some manifold (the zero-velocity set, the no-slip set), and LaSalle lets us conclude asymptotic stability by checking that the only trajectory which stays on that manifold is the equilibrium itself. Without it, a great many systems — including damped-but-not-collocated mechanical systems — would be unanalyzable by direct Lyapunov methods.

Describing functions

The describing-function method is the harmonic linearization: given a nonlinear element $N(\cdot)$ in a feedback loop with a linear plant $G(s)$, approximate $N$ by its response to a sinusoidal input. If the input is $A \sin(\omega t)$ and the nonlinearity produces a periodic output, one retains only the fundamental harmonic and treats the nonlinearity as an amplitude-dependent gain $N(A)$, the sinusoidal-input describing function (SIDF). A limit cycle is then predicted to exist where

$$ N(A) G(j\omega) = -1, $$

the nonlinear analog of the Nyquist crossing of $-1$. Solving for $A$ and $\omega$ gives the predicted amplitude and frequency of the limit cycle.

[fn:: Why does this approximation work as well as it does in practice? The reason is filtering. Most physical plants have strictly proper transfer functions — they are low-pass. If the nonlinearity generates harmonics at $2\omega$, $3\omega, \ldots$, the plant attenuates them more than the fundamental, so the signal fed back to the nonlinearity's input is approximately sinusoidal at $\omega$, which is exactly the assumption the SIDF was built on. This is a self-consistency argument: the approximation is good because the system makes it good. When the plant is not low-pass (e.g. a pure integrator chain, or a plant with a resonance near a harmonic), the method degrades unpredictably. The honest summary: describing functions predict limit cycles that, when simulated, often exist at approximately the predicted amplitude and frequency — but they can also predict cycles that never appear (the true system has none) and miss cycles that do. Treat results as engineering estimates, not theorems.]

Sliding mode control

The idea, due to Utkin (1977), is to define a switching surface $s(x) = 0$ in state space — typically a linear combination $s = c^\top e$ where $e$ is the tracking error — and enforce a discontinuous control that drives the state onto the surface and holds it there. The control law

$$ u = -\eta \, \mathrm{sign}(s) $$

guarantees $\dot{s} = -\eta |s| < 0$ for $s \neq 0$, so the surface is reached in finite time. Once on the surface, the system is constrained to $s = 0$, and the equivalent control — the continuous value $u_{eq}$ that would maintain $s \equiv 0$ — governs the reduced-order dynamics. The appeal is that the closed-loop behavior on the surface depends only on $s = c^\top e$, not on the plant parameters: this is the famous "invariance" of sliding mode to matched uncertainties.

The cost is chattering: the sign function switches infinitely fast across the surface, and real actuators cannot, so the system oscillates at high frequency around $s = 0$ with amplitude set by the actuator's bandwidth and delay. The standard fix is the boundary layer: replace $\mathrm{sign}(s)$ with $\mathrm{sat}(s/\phi)$, a saturation with width $\phi$. Inside the layer the control is continuous (proportional, gain $\eta/\phi$); outside it switches as before. This trades chattering for a steady-state error of order $\phi$.

[fn:: There is a deeper tension here that the boundary layer papers over. The paradigm is discontinuous control for continuous plants — a mathematical device (the sign function) applied to a physical system whose actuators are themselves continuous and rate-limited. The discontinuity is not in the plant but in the engineer's idealization of the controller. Second-order sliding modes (the "super-twisting" algorithm of Levant) achieve finite-time convergence with a continuous control signal by driving not just $s$ but $\dot{s}$ to zero, removing chattering at the source at the cost of greater analysis complexity. Whether this is "more elegant" or "more fragile" depends on one's taste and on how well one knows the plant — which, in the regime where sliding mode is needed at all, is usually poorly.]

Feedback linearization

If the plant has the affine form $\dot{x} = f(x) + g(x)u$, $y = h(x)$, and $f, g, h$ are smooth, one can in principle find a diffeomorphism $z = T(x)$ and an input transformation $u = \alpha(x) + \beta(x) v$ that render the system in the chain-of-integrators form $\dot{z}_1 = z_2, \dot{z}_2 = \cdots = v$ (full-state feedback linearization) or $y^{(r)} = v$ for the output (input-output linearization, $r$ being the relative degree). The nonlinearity is cancelled by the control: $\alpha$ subtracts $f(x)$, $\beta$ divides by the input gain. Linear design — pole placement, LQR, whatever — is then performed on the synthetic linear system.

Two subtleties dominate practice. First, the transformation $T$ must be a diffeomorphism — smooth, invertible, with smooth inverse — over the operating region, which restricts the class of admissible plants and the region of validity. Second, if $r < n$ (the relative degree is less than the system order), there are internal dynamics $\dot{\eta} = \phi(\eta, \xi)$ invisible from the input-output map. The zero dynamics — the internal dynamics restricted to the manifold where the output is identically zero — must be asymptotically stable for the linearized design to be safe, exactly analogous to the minimum-phase requirement in linear nonminimum-phase systems. If the zero dynamics are unstable, the internal states blow up while the output tracks perfectly: the system is non-minimum phase and feedback linearization is unsafe.

[fn:: The robustness cost is severe and is the reason feedback linearization is more a conceptual tool than a deployed one. Exact cancellation requires exact knowledge of $f$ and $g$. If the true system is $f + \Delta f$, $g + \Delta g$, the residual $\Delta f + \Delta g \, u$ enters the transformed dynamics unattenuated and destroys the linear structure. Robust and adaptive extensions exist (adaptive feedback linearization, sliding-mode observers for the uncertainty) but they reintroduce the nonlinearities the method was designed to remove. The honest reading: feedback linearization is the cleanest illustration of what "cancelling the nonlinearity" means, and a useful design approach when the model is accurate and the zero dynamics are benign — chemical process control, some aerospace applications — but it is not a general-purpose solution. The contrast with sliding mode is instructive: sliding mode does not cancel, it dominates, and domination is robust where cancellation is not.]

Gain scheduling

The oldest adaptive method — older than the word "adaptive" in this context — is gain scheduling. Linearize the plant at a grid of operating points $\rho_i$, design a local LTI controller $K_i$ for each, and interpolate (usually linearly) the controller gains as a function of the measured scheduling variable $\rho$. Flight control has used this since the 1950s: the aircraft's dynamics change with Mach number and dynamic pressure, so one designs controllers across the flight envelope and switches or blends between them.

It is a heuristic, and a remarkably effective one. But the stability argument is not obvious. Each frozen controller $K_i$ stabilizes the frozen plant $P(\rho_i)$ — but the closed-loop under time-varying $\rho(t)$ is a switched system, and switched systems can be unstable even when every constituent subsystem is stable.

[fn:: The counterexample is easy to construct: two stable linear systems $A_1, A_2$ whose convex combination $\lambda A_1 + (1-\lambda) A_2$ has an eigenvalue in the right half-plane for some $\lambda$, or whose time-varying composition $\dot{x} = A_{\sigma(t)} x$ has an unstable induced flow. The phenomenon is real — a slow sweep through a region where the frozen controllers are each fine but the transitions excite unstable modes. Sufficient conditions exist (dwell-time bounds, common Lyapunov functions, the "frozen-property + slow variation" results of Lawrence Rifford and others), but in practice gain schedules are tuned by simulation and flight test, not certified a priori. This is arguably the honest state of affairs for most adaptive methods: the engineering works, the proofs lag, and the gap is bridged by Monte Carlo and experience.]

Model Reference Adaptive Control (MRAC)

MRAC is the paradigmatic adaptive scheme. A reference model $W_m(s)$ specifies the desired closed-loop behavior; the adaptive controller adjusts its parameters so the plant output tracks the reference model output. The error $e = y - y_m$ drives the adaptation. The MIT rule (developed at MIT in the 1950s) sets $\dot{\theta} = -\gamma e \, \partial e / \partial \theta$, a gradient descent on $e^2$ — intuitive, but not derived from any stability argument, and it can destabilize. The Lyapunov-based adaptation law, due to Narendra and others, chooses $\dot{\theta}$ so that a Lyapunov function of $e$ and the parameter error $\tilde{\theta}$ has negative derivative, guaranteeing boundedness and (with persistent excitation) convergence.

The robustness problems are well known and serious. Parameter drift: in the absence of persistent excitation, the adaptation can wander into regions where the parameters are meaningless (e.g. parameter values that cancel unmodeled high-frequency dynamics), producing a closed loop that is technically stable under the model but fragile to disturbances the model ignored. Lack of persistent excitation: the reference input must be "rich enough" — containing enough frequency content — to identify the parameters, but a rich reference is exactly what the operator does not want (it is a disturbance from the production standpoint). The fixes (σ-modification, normalization, projection, leakage terms) are robustness patches that sacrifice the ideal convergence guarantees for boundedness under realistic conditions.

[fn:: MRAC and its descendants (L1 adaptive control, which inserts a low-pass filter between the adaptation and the control signal to decouple robustness from adaptation rate) represent the modern engineering face of adaptive control. The theoretical community remains divided on how much these methods add over a well-tuned gain schedule or a robust $H_\infty$ controller — the steelman for MRAC is the regime where the plant parameters drift during operation in a way no fixed schedule can anticipate, which is rarer than the literature implies. The steelman against: most "adaptive" successes are gains scheduled on a variable that happens to be hard to measure, and once you measure it, a schedule suffices. The honest summary: adaptive control is the right answer to a real but narrow class of problems, and its history is largely the history of discovering how narrow that class is.]

Related