Control Theory

Every system that moves, heats, flies, or processes has dynamics — it responds to inputs over time according to physical laws. Control theory is the mathematics of making those responses do exactly what we want. Want a cruise control to hold exactly 65 mph uphill and downhill? Control theory. Want a quadcopter to hover perfectly in gusty wind? Control theory. Want a chemical reactor to stay at precisely 350°C? Control theory.

The core loop is deceptively simple: measure the output of a system, compare it to what we want (the reference), compute the error, and apply a correction based on that error. This is feedback — the most powerful idea in engineering. An open-loop system just hopes for the best (like a toaster that runs for a fixed time). A closed-loop system constantly adjusts based on what's actually happening. Almost every engineered system you interact with daily runs a closed-loop controller.

Before designing a controller, you must model the system you are controlling (the plant). This means writing differential equations that describe how the plant responds to inputs over time. Three archetypal physical systems illustrate that different physics share the same mathematics:

Solving differential equations directly is tedious for anything beyond simple cases. The Laplace transform converts the problem from the time domain (where things change) to the frequency domain (where things are algebraic). The magic: differentiation becomes multiplication by the complex variable s. An ODE transforms into an algebraic equation you can solve with high school algebra, then transform back to the time domain using tables or partial fractions.

The Laplace transform integrates the function from 0 to infinity against an exponentially decaying oscillation e^{-st}. The real part of s controls decay; the imaginary part controls oscillation. Every stable system has its poles (denominator roots) in the left half of the complex s-plane.

A transfer function G(s) is the single most important concept in classical control. It is defined as the ratio of the Laplace transform of the output to the Laplace transform of the input, assuming zero initial conditions. G(s) is always a rational function: a ratio of two polynomials in s. The roots of the numerator are zeros (frequencies the system blocks). The roots of the denominator are poles (frequencies where the system resonates). The positions of the poles in the complex plane reveal EVERYTHING about stability, speed, and oscillation. If all poles have negative real parts, the system is stable. A pole at s=0 means an integrator (it never forgets). A complex conjugate pole pair means oscillation.

A closed-loop system feeds the output back to the input through a sensor, comparing it to the reference. The resulting block diagram has a forward path (controller + plant) and a feedback path (sensor). ${ref('Block diagram algebra', 'foundations/elementary_algebra', 'Algebraic manipulation of block diagrams reduces to simple fraction arithmetic')} yields the closed-loop transfer function: forward path divided by (1 plus the loop gain). The denominator 1 + G(s)H(s) = 0 is called the characteristic equation — its roots are the closed-loop poles, which determine the system's behavior.

The step response is the system's reaction to an instantaneous change from 0 to 1 in the reference. It is the single most informative test in control engineering. From the step response, we extract: rise time (how fast it responds), peak time (when the maximum occurs), percent overshoot (how much it exceeds the target), settling time (when it stays within 2% of final value), and steady-state error (permanent offset from target). For a ${ref('second-order system', 'differential_equations/ode', 'Second-order ODEs: the three damping regimes and their step responses')}, all these metrics can be calculated analytically from just two parameters: the damping ratio ζ (zeta) and the natural frequency ωₙ (omega-n).

A system is stable if ALL its closed-loop poles have negative real parts. The Routh-Hurwitz criterion tests this WITHOUT solving the polynomial — a lifesaver for high-order systems where finding polynomial roots analytically is impossible. You construct the Routh array from the characteristic polynomial's coefficients. If all entries in the first column have the same sign, all poles are in the left half-plane and the system is stable. Each sign change in the first column indicates a pole in the right half-plane (unstable). The criterion also reveals the critical value of parameters at the boundary of stability.

The steady-state error is the permanent difference between reference and output as time → ∞. It depends on the system type — the number of pure integrators (poles at s=0) in the open-loop transfer function. A Type 0 system has constant error to a step input. A Type 1 system (one integrator) has ZERO error to a step (it can perfectly track a constant reference). A Type 2 system has zero error to both steps and ramps. The constants K_p (position), K_v (velocity), and K_a (acceleration) quantify these errors. Adding integrators improves accuracy but destabilizes — the fundamental tradeoff in control between precision and stability.

Classical control (1930s-1960s) developed three powerful graphical techniques that let engineers DESIGN controllers visually rather than solving equations analytically. All three analyze the open-loop system to predict closed-loop behavior.

The PID (Proportional-Integral-Derivative) controller runs over 90% of all industrial control loops. Its three terms each address a different temporal aspect of the error. P (Proportional): responds to the PRESENT error. Larger Kp means faster response but more overshoot and possible instability. I (Integral): accumulates PAST error. This eliminates steady-state error because any persistent error keeps accumulating until the control action cancels it. But too much integral causes overshoot and slow settling. D (Derivative): predicts FUTURE error from the current rate of change. This adds damping — it resists rapid changes and reduces overshoot. But derivative amplifies noise, so real implementations use a filtered derivative or omit it entirely (PI control).

Tuning a PID means finding the right Kp, Ki, Kd values. Methods like Ziegler-Nichols provide starting values from step response experiments. Modern practice uses optimization to automatically find optimal gains.

Tune a PID controller in real time! Adjust Kp, Ki, and Kd sliders and watch the closed-loop step response update instantly. See how P speeds up response, I eliminates steady-state error, and D adds damping. Find the sweet spot where rise time, overshoot, and settling time are all balanced.

Classical control (transfer functions, Bode, root locus) works beautifully for single-input single-output (SISO) systems. But a modern fighter jet has multiple control surfaces and sensors — a multi-input multi-output (MIMO) system. The state-space approach, pioneered by Rudolf Kalman in the 1960s, represents ANY dynamic system (linear or nonlinear, SISO or MIMO, continuous or discrete) as a unified set of first-order matrix differential equations. The system is described by four matrices — A (dynamics), B (input), C (output), D (feedthrough) — and the entire behavior is captured in the matrix exponential.

Two fundamental questions define the limits of what control can achieve. Controllability: can we drive the system from any initial state to any desired state in finite time using the available inputs? The answer is YES if the controllability matrix [B, AB, A²B, ..., A^{n-1}B] has full row rank. This is a linear algebra rank condition. Observability: can we reconstruct the full internal state from output measurements over a finite time interval? YES if the observability matrix [C; CA; CA²; ...; CA^{n-1}] has full column rank. If both hold, we can design an observer (Kalman filter) that estimates the state and a controller that uses those estimates — the celebrated separation principle.

If the system is controllable, we can use state feedback (u = -Kx) to place the closed-loop poles ANYWHERE we want. The closed-loop dynamics become (A - BK), and we choose K so that the eigenvalues of (A - BK) match our desired pole locations. For SISO systems, Ackermann's formula gives K directly. For MIMO systems, numerical methods compute K. Combined with an observer (estimated state x-hat instead of true state x), the complete controller is: u = -K·x_hat, where x_hat comes from a Luenberger observer or Kalman filter. The separation principle guarantees that controller and observer designs are independent.

Almost all modern controllers run on microprocessors — from the tiny chip in your car's ABS to industrial PLCs controlling chemical plants. This means continuous signals must be sampled (measured at discrete instants) and converted to digital numbers. The ${ref('sampling theorem', 'differential_equations/fourier_laplace', 'Fourier & Laplace: the Nyquist-Shannon theorem requires sampling at least twice the highest frequency')} guarantees perfect reconstruction if the sampling rate exceeds twice the signal bandwidth.

Just as the Laplace transform converts continuous differential equations to algebraic equations in s, the Z-transform converts discrete difference equations to algebraic equations in z. A delay of one sample becomes multiplication by z^{-1}. The mapping from s-plane to z-plane is z = e^{sT} where T is the sample period. The stability boundary shifts: in continuous time, poles must be in the left half s-plane. In discrete time, poles must be inside the unit circle in the z-plane. A discrete PID controller is implemented as a difference equation: the integral becomes a running sum and the derivative becomes a finite difference. These numerical approximations introduce subtle errors that must be accounted for.

Modern control engineering follows a model-based design workflow. 1. Model the plant physics (ODEs from first principles). 2. Design the controller in simulation (MATLAB/Simulink, Python Control Systems Library). 3. Validate against a high-fidelity plant model. 4. Auto-generate C code from the controller block diagram. 5. Deploy to the target hardware (PLC, microcontroller, FPGA). 6. Hardware-in-the-loop testing with real actuators and sensors. This workflow has revolutionized industries from automotive (engine control units) to aerospace (flight control computers) to consumer electronics (camera image stabilization).

Adjust damping zeta and natural frequency omega-n to see how second-order systems respond. Watch the animated curve build: underdamped oscillates (ζ<1), critically damped settles fastest without overshoot (ζ=1), overdamped is sluggish (ζ>1). Metrics update live: overshoot, peak time, settling time.