Control Theory Basics

18 May 2019 - 37 mins read time
Tags: control theory system modelling feedback control transfer function stability root locus frequency response PID state-space

This note is adapted from course materials for MNE3049 Control Principles at City University of Hong Kong. Instructor: Dr. Hon Yuen Tam.

Central Question

How can we use measured error and limited actuation to reshape a dynamical system so that it tracks, rejects disturbances, stays stable, and remains robust when the model is imperfect?

Control theory starts with a plant, not a gain. A heater stores thermal energy before its temperature changes. A motor-load system stores kinetic energy before speed changes. A robot joint may obey a commanded torque but still saturate, vibrate, delay, or amplify sensor noise. The controller is therefore not an add-on formula; it is a device for changing the closed-loop dynamics of an imperfect physical system.

The causal spine is:

\[\text{system boundary} \rightarrow \text{states, inputs, outputs, disturbances} \rightarrow \text{dynamic model} \rightarrow \text{open-loop behavior} \rightarrow \text{stability} \rightarrow \text{feedback} \rightarrow \text{time and frequency tradeoffs} \rightarrow \text{controller structure} \rightarrow \text{implementation limits}.\]

If Signals & Systems gives the language of linear time-invariant models, control theory gives that language a purpose: shaping motion, regulation, and robustness under uncertainty.

1. Define the System Boundary First

A control problem begins by deciding what is inside the plant boundary and what is outside it.

Control loop diagram showing reference error controller actuator plant disturbance sensor and measured output
Figure 1. Standard feedback loop with plant boundary, actuator input, measured output, disturbance, sensor, controller, and reference. The controller only acts through available input channels.

The usual roles are:

This boundary matters because control is not “make the system do anything.” It is “use limited actuation and imperfect measurement to influence selected variables.” A controller cannot directly regulate an unmeasured variable unless it can infer it; it cannot move an unactuated mode unless the plant physics couples that mode to an input.

2. States, Parameters, and Operating Regimes

A state is the smallest set of variables that, together with future inputs, determines future behavior. For a heated chamber, temperature may be enough. For a mass-spring-damper, position and velocity are needed. For a circuit, capacitor voltages and inductor currents are natural states.

Parameters such as mass (m), damping (b), stiffness (k), thermal resistance (R), and thermal capacitance (C) are not states. They describe the plant. If they drift with temperature, wear, or payload, the model must either include that variation or treat it as uncertainty.

An operating point is a nominal condition around which we analyze behavior. Many plants are nonlinear, but local linear models can be accurate near a chosen operating regime. The model is valid only where its assumptions remain true:

The assumptions are part of the model. They are not paperwork after the mathematics.

3. From Physics to Control Models

Two simple plants are enough to introduce most of classical control: a first-order thermal process and a second-order mechanical servo.

3.1 Thermal Process: One Dominant Stored-Energy State

Let (\theta(t)) be temperature rise above ambient. A simple energy balance gives

\[C\dot{\theta} = -\frac{1}{R}\theta + K_h u + d,\]

where (C) is thermal capacitance ([\mathrm{J/K}]), (R) is thermal resistance ([\mathrm{K/W}]), (K_h u) is heater power ([\mathrm{W}]), and (d) is an unknown heat disturbance ([\mathrm{W}]).

With (\tau=RC), the model can be written as

\[\tau\dot{\theta}+\theta = K u + d_\theta.\]

The physical meaning is direct: the chamber has one dominant memory state and one dominant time constant. If forcing stops, it forgets exponentially rather than instantly.

3.2 Mechanical Servo: Inertia, Dissipation, and Stiffness

For a mass-spring-damper driven by force,

\[m\ddot{x}+b\dot{x}+kx = u+d,\]

where (m) is mass ([\mathrm{kg}]), (b) is damping ([\mathrm{N\,s/m}]), (k) is stiffness ([\mathrm{N/m}]), (u) is actuator force ([\mathrm{N}]), and (d) is disturbance force ([\mathrm{N}]).

This plant can overshoot and oscillate because it stores both kinetic and potential energy. The state-space form makes that memory explicit:

\[\dot{x}_1=x_2, \qquad \dot{x}_2=-\frac{k}{m}x_1-\frac{b}{m}x_2+\frac{1}{m}u+\frac{1}{m}d, \qquad y=x_1.\]

In matrix form,

\[\dot{x}=Ax+Bu+Ed, \qquad y=Cx.\]

3.3 Linearization

A nonlinear plant may be written as

\[\dot{x}=f(x,u,d), \qquad y=h(x,u,d).\]

Around an equilibrium ((x_0,u_0,d_0)) satisfying (f(x_0,u_0,d_0)=0), small deviations satisfy

\[\delta\dot{x}=A\delta x+B\delta u+E\delta d, \qquad \delta y=C\delta x+D\delta u+F\delta d,\]

where the matrices are Jacobians at the operating point. This is why classical control often works on nonlinear hardware: the loop is designed around the regime where the plant behaves approximately linearly.

4. Open-Loop Dynamics Before Control

Before adding feedback, ask how the plant moves by itself.

For the thermal plant with (u=d_\theta=0),

\[\theta(t)=\theta(0)e^{-t/\tau}.\]

For a standard underdamped second-order mode,

\[x_n(t)=e^{-\zeta\omega_n t} \left(A\cos\omega_d t+B\sin\omega_d t\right), \qquad \omega_d=\omega_n\sqrt{1-\zeta^2}.\]

Poles and eigenvalues summarize this natural motion:

Complex plane showing stable real poles, stable complex poles, marginal imaginary poles, and unstable right half plane poles
Figure 2. Pole geometry as motion geometry. Pole location encodes decay, oscillation, settling speed, and instability.

Control does not create dynamics from nothing. It moves or reshapes modes that already exist in the plant-controller interconnection.

5. Stability: Necessary but Not Sufficient

Stability is the first pass/fail test.

Internal stability asks whether internal states return or remain bounded after perturbation. BIBO stability asks whether bounded inputs produce bounded outputs. For minimal LTI systems these align closely, but they are conceptually different: hidden unstable modes can be dangerous even if a particular input-output transfer function hides them.

For continuous-time LTI systems, a common pole test is:

Stability alone is not enough. A stable controller may still be too slow, too oscillatory, too noisy, too fragile, or impossible for the actuator to implement. The design problem is:

stabilize the plant with acceptable speed, damping, accuracy, noise sensitivity, saturation behavior, and robustness.

6. Transfer Functions and the Characteristic Equation

For an LTI model under zero initial conditions, the transfer function is

\[G(s)=\frac{Y(s)}{U(s)}.\]

The zero-initial-condition phrase matters. A transfer function describes input-output behavior, not arbitrary stored initial energy.

For the running examples,

\[G_\theta(s)=\frac{\Theta(s)}{U(s)}=\frac{K}{\tau s+1}, \qquad G_x(s)=\frac{X(s)}{U(s)}=\frac{1}{ms^2+bs+k}.\]

Poles are denominator roots and govern natural motion. Zeros are numerator roots and shape how inputs appear at the output. Non-minimum-phase zeros, delays, and hidden lightly damped modes can make control much harder.

For standard negative feedback with controller (D(s)), plant (G(s)), and sensor path (H(s)),

\[E(s)=R(s)-H(s)Y(s), \qquad U(s)=D(s)E(s), \qquad Y(s)=G(s)U(s).\]

Thus,

\[\frac{Y(s)}{R(s)} = \frac{D(s)G(s)}{1+D(s)G(s)H(s)}.\]

The characteristic equation is

\[1+D(s)G(s)H(s)=0.\]

This is the core mathematical fact of feedback: the controller changes the denominator of the closed-loop system. It reshapes the poles, and therefore the motion.

7. Time-Domain Performance: Speed, Damping, Accuracy

Step-response metrics are compact descriptions of transient and steady behavior:

Step response curve showing rise time overshoot settling band settling time and steady state error
Figure 3. Step-response metrics: rise time, overshoot, settling band, settling time, and steady-state error.

The standard second-order closed-loop template is

\[G_{cl}(s)=\frac{\omega_n^2}{s^2+2\zeta\omega_n s+\omega_n^2}.\]

For (0<\zeta<1),

\[M_p\approx e^{-\pi\zeta/\sqrt{1-\zeta^2}}\times 100\%, \qquad t_s\approx\frac{4}{\zeta\omega_n}\]

for the common 2% settling-time approximation. These are design heuristics for dominant second-order behavior, not universal laws for every plant.

When the closed loop is stable, the final-value theorem gives

\[\lim_{t\to\infty}y(t)=\lim_{s\to 0}sY(s).\]

Classical system type counts the number of open-loop integrators. More type improves polynomial tracking in ideal models, but integrators also add phase lag and can worsen robustness. Integral action is powerful because it removes constant steady-state error; it is dangerous because it adds controller memory that can wind up during saturation.

8. Frequency-Domain Design: Bandwidth, Margins, and Noise

Frequency response asks how the system reacts to sinusoidal inputs of different frequencies. For

\[G(s)=\frac{1}{\tau s+1},\]

we evaluate (G(j\omega)):

\[|G(j\omega)|=\frac{1}{\sqrt{1+(\omega\tau)^2}}, \qquad \angle G(j\omega)=-\tan^{-1}(\omega\tau).\]

A first-order plant follows slow commands well but attenuates and lags fast commands. That lag becomes dangerous when feedback tries to push bandwidth too high.

Schematic Bode magnitude showing high low frequency gain crossover phase margin region and high frequency roll off
Figure 4. Loop-shaping tradeoff. High low-frequency gain improves tracking and disturbance rejection, while high-frequency roll-off limits noise and unmodelled dynamics.

Important frequency-domain ideas:

These tools are not separate worlds. They are different projections of the same robustness problem.

9. Sensitivity and the Waterbed Tradeoff

Let the loop gain be

\[L(s)=D(s)G(s)H(s).\]

The sensitivity and complementary sensitivity functions are

\[S(s)=\frac{1}{1+L(s)}, \qquad T(s)=\frac{L(s)}{1+L(s)}.\]

They satisfy

\[S(s)+T(s)=1.\]

Small (S) at low frequency means good tracking and disturbance rejection. Small (T) at high frequency means less sensor-noise transmission and less excitation of unmodelled fast dynamics. Since they sum to one, the controller cannot make both small everywhere. Control design is a redistribution of error, noise, and uncertainty across frequency.

This is the “waterbed” intuition: pushing sensitivity down in one region tends to push it up somewhere else, especially for plants with delays, right-half-plane zeros, or bandwidth limits.

10. Classical Design Views

Map of classical control tools linking characteristic polynomial root locus Bode Nyquist Nichols and closed loop design
Figure 5. Classical design views. Routh-Hurwitz, root locus, Bode, Nyquist, and Nichols answer the same closed-loop question from different angles.

10.1 Routh-Hurwitz

Routh-Hurwitz checks whether a characteristic polynomial has right-half-plane roots without solving for all roots. For

\[s^3+a_2s^2+a_1s+a_0,\]

a continuous-time cubic is stable if

\[a_2>0,\qquad a_1>0,\qquad a_0>0,\qquad a_2a_1>a_0.\]

This is an algebraic sanity check for a controller-dependent denominator.

10.2 Root Locus

Root locus asks:

As a gain changes, where do the closed-loop poles move?

Branches start at open-loop poles and end at open-loop zeros or infinity. The plot makes visible whether increasing gain speeds the response, improves damping, or drives the loop unstable.

10.3 Bode, Nyquist, and Nichols

Bode plots expose bandwidth and margins. Nyquist connects open-loop frequency response to exact closed-loop stability. Nichols charts make loop-shaping tradeoffs visible in a gain-phase plane. They are complementary tools, not competing methods.

11. Controllers as Model-Shaping Tools

The standard PID form is

\[D(s)=K_P+\frac{K_I}{s}+K_Ds.\]
PID term diagram showing proportional present error integral accumulated error derivative slope and implementation caveats
Figure 6. PID terms as controller memory and prediction. Proportional action reacts to present error, integral action remembers accumulated error, and derivative action reacts to trend.

The terms have different physical meanings:

Real derivative action is filtered; ideal differentiation is not implementable over infinite bandwidth. Real integral action needs anti-windup because actuators saturate. When saturation occurs, the actuator cannot realize the requested input, but the integrator may keep accumulating error. Anti-windup keeps controller memory consistent with actuator reality.

Lead and lag compensators express similar loop-shaping goals:

\[D_{lead}(s)=K\frac{1+\tau s}{1+\alpha\tau s},\qquad 0<\alpha<1,\] \[D_{lag}(s)=K\frac{1+\tau s}{1+\beta\tau s},\qquad \beta>1.\]

Lead compensation adds phase near crossover and can improve damping or bandwidth. Lag compensation increases low-frequency gain while trying not to push crossover too high. Feedforward, reference shaping, and two-degree-of-freedom structures are added when tracking and disturbance rejection should not be forced through exactly the same path.

12. Implementation Limits: What Fails First

Control designs fail in hardware for reasons that may not appear in the first transfer function:

A practical validation workflow is:

  1. identify the plant boundary and safety limits;
  2. estimate or identify the model;
  3. design with margin rather than exact cancellation;
  4. simulate nominal, disturbed, saturated, and parameter-shifted cases;
  5. test at low gain and low authority first;
  6. log reference, measurement, input, saturation, and fault flags;
  7. increase bandwidth only after the limiting mechanism is understood.

Control is therefore both mathematics and commissioning discipline.

13. State-Space Bridge

Transfer functions are excellent for single-input/single-output loop design, but they hide internal structure. State-space models expose it:

\[\dot{x}=Ax+Bu, \qquad y=Cx+Du.\]

The eigenvalues of (A) are internal modes. For minimal realizations, they correspond to transfer-function poles.

State space diagram showing input through B to state A dynamics output through C and controllability observability questions
Figure 7. State-space bridge. Controllability asks whether inputs can move each mode; observability asks whether outputs reveal each mode.

The controllability matrix is

\[\mathcal{C}=\begin{bmatrix}B & AB & \cdots & A^{n-1}B\end{bmatrix}.\]

The observability matrix is

\[\mathcal{O}= \begin{bmatrix} C\\ CA\\ \vdots\\ CA^{n-1} \end{bmatrix}.\]

Full rank means every mode can be influenced or seen through the chosen channels. No controller can move a mode it cannot actuate; no observer can reconstruct a mode that leaves no measurement signature.

With full-state feedback,

\[u=-Kx+Nr, \qquad A_{cl}=A-BK.\]

This is the modern version of the same idea: choose feedback to move closed-loop modes.

For the full multivariable story, continue to Modern, Multivariable, and Networked Control.

14. Bridge to Nonlinear Control

Classical control is usually local. It assumes the model is approximately linear near the operating point and that the plant stays in the region where the design assumptions hold.

Real plants may include saturation, switching, friction, impacts, geometric nonlinearities, dead zones, gain scheduling, and constraints. Linear tools still help locally, but global behavior needs phase-plane reasoning, Lyapunov analysis, invariant sets, nonlinear compensation, or hybrid analysis.

For those topics, continue to Topics in Nonlinear Systems. For adaptive, optimal, robust, and learning-based extensions, continue to Adaptive, Optimal, Robust, and Learning Control.

15. Practitioner’s Perspective: Why the Textbook Toolbox Is Not the Job

The sections above are the clean framework. Experienced practitioners (see the Zhihu discussion by Fang Chenxi in the sources) add a blunt corrective: control is an engineering discipline that happens to use mathematics, not a catalogue of named algorithms. The points below are my distillation of that view, checked against the framework above.

15.1 Twelve Working Rules

1. Control is about uncertainty, and tuning is the real work. No named method (PID, LQR, MPC, $H_\infty$, sliding mode) is applied off the shelf. A method that is easy to tune and diagnose usually beats one with better theoretical performance, because performance is set by the parameters, not the controller structure. Models with many parameters are a liability.

2. The Bode plot is the working tool. For practical SISO design, shaping the magnitude curve in different frequency bands, matching it to identification data and to observed noise and jitter, is most of the job. Root locus mainly tunes one gain; Nyquist is excellent for deciding stability but gives little design guidance on high-order loops, where software already reports margins.

3. PID is a low-order special case of a general rational compensator. Any realizable SISO controller is a ratio of polynomials, which factors into lead, lag, and notch sections. PID has four free parameters; a real motor or chassis loop may need a 6th to 12th order compensator. PID works when blindly applied because the integral term is kept below crossover, and a filtered derivative acts as a lead.

\[D(s)=\underbrace{\frac{b_ms^m+\cdots+b_0}{a_ns^n+\cdots+a_0}}_{\text{lead}\,\times\,\text{lag}\,\times\,\text{notch}} \quad\supset\quad K_P+\frac{K_I}{s}+\frac{K_Ds}{\tau_fs+1}.\]

4. Observers are the dual of feedback, not a truth machine. An observer is judged by how it shapes the whole loop (bandwidth, margins), not by how close $\hat x$ is to $x$. Designing controller and observer separately is safe in theory but can fail in practice: an LQR and a Kalman filter each have good margins alone (under the classical assumptions: continuous time, infinite horizon, exact full-state feedback, loop broken at the plant input; see section 7.3 of the adaptive-control note), yet the LQG combination has no guaranteed margin. A Kalman filter and an MPC tuned by different people can show similar fragility. A correct closed-loop simulation of the whole loop can reveal the sensitivity peak; what simulation often misses is model uncertainty and neglected dynamics, so check the combined loop in the frequency domain on perturbed models.

5. Pole placement hurts when poles sit near zeros. Moving a pole close to a zero is fragile, and large pole moves demand large control effort. In high-order plants most lightly damped poles are near zeros, so aggressive placement is rarely a good idea.

6. Most “noise” at standstill is a limit cycle. Jitter when a steering wheel or lane-keeping loop is at rest is often a nonlinear limit cycle rather than sensor noise passed through the closed loop, though this is a hypothesis to test with frequency checks and gain sweeps, which give diagnostic evidence, not a definitive classification. Sensor noise is typically far above the mechanical bandwidth, and a linear time-invariant loop does not create new frequencies. Typical causes are dead zones, Coulomb friction with gain, Stribeck friction, and backlash. Practical rules: keep gain curves monotonic (a non-monotone gain map invites limit cycles), and be cautious about adding switches, since unanalysed switching nearly always creates chattering.

7. Strong nonlinearities are traded, not solved. Gain scheduling handles mild nonlinearity (engines, chassis, aircraft). For strong ones, the trade is explicit: zero gain inside an error band costs steady-state accuracy; higher gain in a stick-slip speed band costs noise; freezing the integrator in a dead zone costs response speed. If accuracy and quietness both fail, improve the plant (less friction, backlash, dead zone) instead of the controller.

8. Identify linear models; do not expect exact parameters. Fitting a linear model to a sine sweep of a strongly nonlinear steering system gives coefficients that depend on the excitation amplitude and have no direct physical meaning. Nonlinear parameters (friction, backlash) are not inherently unidentifiable, but identifying them needs an assumed model structure, structural identifiability, excitation that exercises the nonlinearity, suitable measurements, and validation on separate data. Parameters also map nonlinearly from the spectrum, so a good open-loop frequency response is often the realistic goal for the linear part. Keep excitation matched to the operating regime: violent excitation softens structures and excites modes absent in normal operation.

9. Adaptive control is not a free lunch. Gain scheduling uses an independently measured variable; adaptive control adapts from input-output signals. Classical self-tuning regulators and MRAS assume certainty equivalence, require persistent excitation, and in closed loop identification and control are coupled. Convergence of estimates plus stability of the controller does not guarantee stability of the whole, since finite-escape-time systems can diverge before parameters converge. Treat adaptation as a high-gain nonlinear controller with extra states, not as a way to avoid knowing the plant.

10. MPC is truncated LQR. For an LTI model, a finite-horizon predictor loses the closed-form gain $u=-Kx$, the optimality guarantee unless a terminal condition holds, and guaranteed feasibility once constraints are added. MPC earns its cost when constraints are active and model-based prediction really matters, not as a default choice for lateral or longitudinal vehicle control. Kalman’s key theoretical contribution was the duality between estimation and optimal control; Kalman filter formulas are one numerical realization of that duality, not the theory itself.

11. Stability certificates are conservative. Lyapunov, passivity, and small-gain results are sufficient conditions. Small-gain says the loop gain should stay below one everywhere, which is rarely a useful operating condition. A system certified stable may still lack robustness. Use these tools for qualitative reasoning about known systems, and for the dynamics that remain: unmodelled fast modes, whose singular-perturbation analysis is where nonlinear robust design is most convincing.

12. Practice shapes understanding. A good lab project is a high-bandwidth (over about 10 Hz), high-order, uncertain plant with a tracking objective, not an inverted pendulum stabilized at a point. The mathematics that pays off most is calculus/analysis, linear algebra, and ordinary differential equations; Laplace transforms are shorthand for the Fourier picture in which differentiation becomes multiplication, poles make a frequency component grow without bound, and zeros annihilate it.

15.2 Laplace Transforms Are Fourier Analysis in Disguise

Most of the Laplace-domain machinery is a bookkeeping convenience for one fact: complex exponentials are eigenfunctions of linear time-invariant (LTI) systems. Take a plant described by

\[\sum_{k=0}^{n}a_k\,y^{(k)}(t)=\sum_{k=0}^{m}b_k\,u^{(k)}(t).\]

For an input $u=e^{st}$ the response is $y=G(s)e^{st}$, because $\frac{d}{dt}e^{st}=s\,e^{st}$ turns differentiation into multiplication, and

\[G(s)=\frac{\sum_k b_ks^k}{\sum_k a_ks^k}.\]

On the imaginary axis $s=j\omega$ this is the frequency response. Three consequences follow directly:

Writing $s$ instead of $d/dt$ saves symbols, but it hides the time-domain meaning. Keeping the sinusoidal picture in mind also explains why frequency-domain plots (Bode plots) match laboratory tests so directly: a sweep measures $G(j\omega)$ point by point.

Worked example: an armature circuit. Let $r$ be a commanded current, $u$ the voltage the controller applies, and $i$ the current in a motor winding with inductance $L$ and resistance $R$. Kirchhoff’s voltage law gives the plant ODE, and the sinusoidal ansatz turns it into algebra in one line:

 time-domain ODE        sinusoidal input         algebra (no calculus)      transfer function
 L di/dt + R i = u  -->  u = U e^{jωt},          (jωL + R) I = U      -->   H(jω) = I/U = 1/(R + jωL)
                         i = I e^{jωt}

With $L=1$ mH and $R=2\ \Omega$: the DC gain is $1/R=0.5$ A/V, the pole is at $s=-R/L=-2000$ rad/s (time constant $0.5$ ms), and at the corner frequency $\omega=2000$ rad/s the gain falls by $3$ dB and the phase is $-45^\circ$. The same fact in the time domain is convolution, $i(t)=\int_0^t h(t-\tau)u(\tau)\,d\tau$ with impulse response $h(t)=\tfrac1L e^{-Rt/L}$; the Fourier or Laplace transform turns the convolution into the product $I=HU$, which is why cascaded blocks multiply. In the feedback loop below, the same algebra gives the closed-loop transfer function directly:

Simple feedback control loop with controller C, plant P, and sensor F Figure 8. A single-loop feedback system with controller $C$, plant $P$, and sensor $F$. Source: File:Simple_feedback_control_loop2.svg. License: CC BY-SA 3.0.

\[\frac{Y}{R}=\frac{CP}{1+CPF},\qquad S=\frac{1}{1+CPF},\qquad T=\frac{CPF}{1+CPF},\qquad S+T=1 .\]

A zero works the same way. The compensator $1+\tau s$ has a zero at $s=-1/\tau$: if the input contains the component $e^{-t/\tau}$, the compensator maps it to zero. A pole is the opposite: the denominator $1+\alpha\tau s$ vanishes at $s=-1/(\alpha\tau)$, so a forcing term near that exponent is amplified.

15.3 Why PID Works, and What the General Compensator Is

Any realizable SISO LTI controller is a ratio of real polynomials. By the fundamental theorem of algebra each polynomial factors into real linear factors $(s+a)$ and real irreducible quadratics $(s^2+bs+c)$. The factors have familiar names:

\[N(s)=\frac{s^2+2\zeta_z\omega_0s+\omega_0^2}{s^2+2\zeta_p\omega_0s+\omega_0^2},\qquad \zeta_z<\zeta_p,\]

whose gain at $\omega_0$ is exactly $\zeta_z/\zeta_p$, so the depth and width of the notch are set by the two damping ratios.

A lead $\frac{1+\tau s}{1+\alpha\tau s}$ with $0<\alpha<1$ gives its largest phase boost, $\arcsin\frac{1-\alpha}{1+\alpha}$, at $\omega=1/(\tau\sqrt{\alpha})$. So the complete family of controllers is a chain of leads, lags, notches, and resonances, and design is choosing how many of each and where to put them.

PID is a four-parameter slice of that family. With a realistic filtered derivative,

\[D(s)=K_P+\frac{K_I}{s}+\frac{K_Ds}{\tau_fs+1},\]

the parameters are $K_P,K_I,K_D,\tau_f$. An order-$n$ compensator has $2n+1$ free coefficients: 13 for sixth order and 17 for eighth order. High-order plants (a motor drive with a flexible load, or a chassis with several modes) can need that freedom. PID works when applied blindly because $K_I$ is kept small enough to act only below the loop crossover, so it does little to the phase near crossover, and a filtered $K_D$ term behaves like a lead. It is weak when the plant needs a notch, a second lag, or shaping across several decades, and when parameters have to be found by trial because no model is available.

PID controller block diagram Figure 9. PID structure: the error $e(t)=r(t)-y(t)$ feeds three parallel terms, proportional $K_Pe$, integral $K_I!\int e\,d\tau$, and derivative $K_D\,de/dt$, which are summed into the plant input $u$. Source: File:PID_en.svg. License: CC BY-SA 3.0.

The general compensator is the same loop with the three parallel boxes replaced by a chain of sections, each chosen for a purpose:

              gain      lead (phase boost)     lag / integrator       notch (kills a mode)
   e --->[   K   ]--->[ (1+τs)/(1+ατs) ]--->[ (1+τ_l s)/(1+βτ_l s) ]--->[ N(s) ]---> u
                          around ωc               below ωc               at ω0

Designing a lead section. The lead $\dfrac{1+\tau s}{1+\alpha\tau s}$ with $0<\alpha<1$ raises the phase near the crossover frequency $\omega_c$. Its maximum phase lead $\phi_{max}$ occurs at $\omega_{max}=1/(\tau\sqrt\alpha)$, with gain $1/\sqrt\alpha$ there. To get a required boost $\phi_{max}$:

\[\alpha=\frac{1-\sin\phi_{max}}{1+\sin\phi_{max}},\qquad \tau=\frac{1}{\omega_c\sqrt\alpha}.\]

Worked example: a $45^\circ$ boost at $\omega_c=10$ rad/s gives $\alpha=0.172$ and $\tau=0.241$ s. At $10$ rad/s the section has magnitude $2.41$ ($+7.7$ dB) and phase exactly $45^\circ$, so the open-loop gain must be reduced by that factor if crossover is to stay at $10$ rad/s. A lead cannot be made arbitrarily strong: smaller $\alpha$ raises high-frequency gain by $1/\alpha$ and amplifies sensor noise, which is why leads are limited to roughly $60^\circ$ per section and why two leads in series are used for more.

Bode diagram of a lead compensator Figure 10. Bode diagram of a lead compensator: the magnitude rises between the zero and the pole, and the phase bump peaks at $\omega_m$ with value $\phi_m$. Source: File:Bodelead.png. License: CC BY-SA 4.0.

Designing a notch. For $N(s)=\dfrac{s^2+2\zeta_z\omega_0s+\omega_0^2}{s^2+2\zeta_p\omega_0s+\omega_0^2}$ the gain at $\omega_0$ equals $\zeta_z/\zeta_p$ exactly. With $\zeta_z=0.05$ and $\zeta_p=0.7$ the notch is $0.071$ deep ($-22.9$ dB) at $\omega_0$, and a larger $\zeta_p$ widens it. A wide notch costs phase lag below $\omega_0$ and phase lead above it, so it should be only as wide as the uncertainty in the resonance frequency requires.

15.4 How Useful Are the Classical Pictures in Practice?

One practitioner’s ranking, with my caveats added:

Tool Good for Limits in practice
Root locus seeing how closed-loop poles move with one gain; teaching why large gain destabilizes a minimum-phase loop and small gain destabilizes a non-minimum-phase one it varies a single parameter, while real designs shape many; a numerical root finder gives the same facts
Nyquist plot exact closed-loop stability from encirclements, including open-loop unstable plants and delays; margin intuition for low-order loops on high-order loops the curve is hard to read by eye; software reports the margins directly
Bode plot loop shaping: slopes, crossover, margins, and noise, all additive on a log scale; matching to identification data and to observed jitter gives margins, not a complete robustness guarantee

The Bode plot is the working tool because design, identification, noise, and test data all live naturally in the frequency domain. Nyquist remains the right argument when the plant is open-loop unstable or the loop is conditionally stable, so it is better to treat it as the theory behind the margin numbers than as a hand-drawing exercise.

What the plots actually measure. For the open-loop transfer function $L(s)=C(s)P(s)F(s)$ the two classical robustness numbers are

\[GM=\frac{1}{|L(j\omega_{180})|}\ \ \text{where}\ \angle L(j\omega_{180})=-180^\circ,\qquad PM=180^\circ+\angle L(j\omega_c)\ \ \text{where}\ |L(j\omega_c)|=1 .\]
Worked example: $L(s)=\dfrac{1}{s(s+2)(s+3)}$. The phase reaches $-180^\circ$ when the imaginary part of the denominator, $\omega(6-\omega^2)$, vanishes, i.e. at $\omega_{180}=\sqrt6=2.449$ rad/s, where $ L =1/30$, so $GM=30$ ($29.5$ dB). The gain crosses $0$ dB at $\omega_c=0.166$ rad/s with phase $-97.9^\circ$, so $PM=82.1^\circ$. These match the MATLAB screenshot below, and they illustrate why margins alone are not the whole story: the loop has large margins but crosses over very slowly, so it would track poorly. Margins measure distance to the critical point, not performance.

Bode plot with gain and phase margins annotated for 1/(s(s+2)(s+3)) Figure 11. Gain margin $29.5$ dB at $2.45$ rad/s and phase margin $82.1^\circ$ at $0.166$ rad/s for $L(s)=1/\big(s(s+2)(s+3)\big)$. Source: File:Bode-margin 1 over s(s+2)(s+3).png. License: Public domain.

The Nyquist plot expresses the same stability test exactly: the closed loop is stable if and only if $Z=N+P=0$, where $P$ is the number of open-loop poles in the right half-plane and $N$ is the number of clockwise encirclements of $-1$ by $L(j\omega)$ for $-\infty<\omega<\infty$. For a plot like the one below, $P=0$ and the curve does not encircle $-1$, so the loop is stable; the distance from the curve to $-1$ is the stability margin in a single number (the minimum of $ 1+L(j\omega) $, which equals $1/|S|_\infty$).

Nyquist plot of G(s)=1/(s^2+s+1) Figure 12. Nyquist plot of $G(s)=1/(s^2+s+1)$ for $\omega\ge0$ (blue) and $\omega\le0$ (green). Source: File:Nyquist example.svg. License: Public domain.

15.5 Identification: Three Questions Before Any Fit

Model-based design needs a model, and textbook transfer functions are not the plant. Three questions come before any parameter estimate:

  1. Is the system close enough to linear? Identify a linear model only over a regime where the response scales with the input. A linear fit to a strongly nonlinear plant returns coefficients that depend on the excitation amplitude and are not parameters. Nonlinear parameters can be identified, but only with an assumed structure, excitation that exercises the nonlinearity, and suitable measurements.
  2. What model structure and order? Compare candidate orders and pole-zero counts against data, and prefer the smallest model whose residuals look like noise. The mapping from a measured spectrum to parameters is nonlinear and can be ill-conditioned, so a good open-loop frequency response is often a more reliable deliverable than precise parameter values.
  3. Is the data good enough? An input with energy at the frequencies that matter (a chirp or multisine, not a step alone), a good signal-to-noise ratio across the band, and coherence checks. Poor data typically produces a model with a gain and little else.

Beyond the fit, unmodelled dynamics near the stability margin matter most. Small flexible modes, delays, and sampling effects can be invisible at low amplitude and become dominant when the loop is pushed hard, so identify at the amplitudes the controller will see, and keep margin for what the model leaves out.

A concrete workflow. A frequency-domain identification run follows these steps:

 design input ──► apply to plant ──► record u(t), y(t) ──► estimate H(ω) and coherence ──► fit rational model ──► validate
 (multisine/chirp,   (amplitude in      (anti-aliasing,       H1 = Suy/Suu, γ² = |Suy|²/(Suu Syy)     (order, poles,     (new data,
  covers band)        linear regime)     long enough)                                                  zeros, delay)      Bode overlay)
Here $S_{uy}$ is the cross-spectrum of input and output and $S_{uu}$ the input auto-spectrum. The estimate $\hat H_1(\omega)=S_{uy}/S_{uu}$ is unbiased for output noise, and the coherence $\gamma^2(\omega)= S_{uy} ^2/(S_{uu}S_{yy})$ lies in $[0,1]$: values near 1 mean the output is a linear function of the input at that frequency, and low values flag measurement noise, an unmodelled input, or nonlinearity. A sweep that gives high coherence at low amplitude and low coherence at high amplitude is direct evidence of a nonlinearity that a single linear model hides.

If a discrete-time model is fitted in the time domain instead, the standard ARX form is

\[y_k+a_1y_{k-1}+\cdots+a_ny_{k-n}=b_1u_{k-1}+\cdots+b_mu_{k-m}+e_k,\]

and least squares gives $\hat\theta=(\Phi^\top\Phi)^{-1}\Phi^\top y$ with $\theta=[a_1\ \cdots\ a_n\ b_1\ \cdots\ b_m]^\top$. When the input does not excite some frequency band, $\Phi^\top\Phi$ is ill-conditioned and the corresponding parameters are poorly determined, which is the quantitative form of “the data lack richness”; and because poles are roots of the fitted polynomial, a small change in the coefficients of a high-order polynomial can move them a long way. That is why a good frequency-response fit is often more trustworthy than the individual parameter values.

15.6 Lessons From Failures, With Care

Practitioner accounts often cite well-known cases where a textbook method failed once its hidden assumptions were broken: LQG regulators that lost robustness once an observer was added (Doyle’s 1978 counterexample), adaptive flight control that behaved well in simulation and poorly under real unmodelled dynamics (the X-15 MH-96 adaptive system, often discussed in connection with the 1967 loss), and MPC or Kalman filter stacks tuned separately and combined late. The common thread is not that these methods are wrong, but that

The same sources argue that control as a research field has stagnated, citing Y.-C. Ho’s remark that “control is dead.” That is a rhetorical claim, and the evidence it rests on (little change in core theory for decades, many papers revisiting old results) is contestable. A reasonable reading is that classical theory is mature, while problems at the interface with learning, networks, and hardware remain open.

15.7 A Careful Reading List

One highly opinionated answer ranks textbooks sharply. I keep the recommendations and drop the star ratings, because taste differs, and because some books it dismisses are widely used and perfectly good for a first pass.

Area Frequently recommended Why
Classical, first pass Franklin, Powell and Emami-Naeini; Goodwin, Graebe and Salgado correct, practical, with design examples
Linear systems Rugh clear, standard treatment
Nonlinear Khalil; Feng Chunbo (in Chinese) standard theory; the latter is compact with engineering examples
Optimal Bryson and Ho; Kirk or Lewis for variety the older text still reads well; the others are conventional coursework texts
Adaptive Ioannou and Sun (Robust Adaptive Control); Krstic, Kanellakopoulos and Kokotovic rigorous and complete for linear, then nonlinear, adaptive control
Robust Doyle, Francis and Tannenbaum; Skogestad and Postlethwaite; Zhou (Essentials of Robust Control) feedback limits, loop shaping, and $H_\infty$

Two cautions about reading lists. A book that makes everything look easy may be hiding the hard cases, and a book that seems impenetrable may be rewarding once its notation is mastered. A good test is whether, after reading a chapter, you can state what the method cannot do.

15.8 A Second Route to Learning

Alongside textbooks, organize study in three layers: modules (matrix theory, linear systems, nonlinear systems; read one book at a time, even a little), theorems (one result per week, implemented and simulated so that its assumptions are visible), and papers (skim background, problem, assumptions, and method; promote a paper to a theorem when it is relevant). Keep three notes for each paper: the assumptions, the guarantee, and what breaks if an assumption fails. Progress rests on a broad base, not a narrow stack.

Reading note. The source is an opinionated forum answer. Strong claims such as “simulation is invalid for control”, “MPC is misused in autonomous driving”, or “adaptive control does not exist” are best read as warnings against uncritical use, not as theorems. Simulation, MPC, and adaptive laws remain valid tools when their assumptions are checked on hardware.

What This Framework Lets Us Do

This framework lets us reason through control as an engineering workflow:

Where the Framework Stops Being Reliable

The framework becomes unreliable when its assumptions are violated: large nonlinear excursions, strong delays, saturating actuators, hidden flexible modes, changing parameters, poor sensing, unmodelled constraints, or unsafe test conditions. In those regimes, the local LTI design may still be a useful inner loop, but it must be checked with nonlinear, robust, sampled-data, or experimental methods.

Where the Subject Leads Next

Basic control leads naturally to:

Compact Recall Map

  1. Control starts by defining the plant boundary.
  2. A useful model names states, inputs, outputs, disturbances, and assumptions.
  3. Open-loop poles describe the plant’s natural motion.
  4. Feedback changes the characteristic equation and reshapes closed-loop poles.
  5. Stability is necessary but not enough.
  6. Time-domain tools describe speed, damping, and accuracy.
  7. Frequency-domain tools describe bandwidth, margins, noise, and robustness.
  8. PID, lead, lag, feedforward, and 2-DOF designs are loop-shaping tools.
  9. State-space methods expose controllability, observability, and internal modes.
  10. Nonlinear and implementation limits decide where the basic model stops being safe.

Technical and Editorial Audit

Area Correction or decision
Central question Reframed basic control around limited actuation, measurement, stability, robustness, and implementation.
Preserved material Retained thermal and mass-spring-damper examples, transfer functions, second-order metrics, Routh-Hurwitz cubic condition, sensitivity functions, PID/lead/lag, and internal links.
Units Added units for the thermal and mechanical model variables and parameters.
Assumptions Made linearization, zero initial conditions, dominant second-order formulas, and local LTI validity explicit.
Sign conventions Used standard negative feedback with (e=r-y_m) and characteristic equation (1+DGH=0).
Figures Added seven original SVG diagrams for loop boundary, pole geometry, step metrics, Bode tradeoff, classical design views, PID terms, and state-space structure.
Safety Added implementation limits and a cautious validation workflow before increasing controller authority.
Internal links Preserved links to Signals & Systems, Modern Control, Nonlinear Systems, and Adaptive/Optimal/Robust/Learning Control.

Sources and Further Reading