[Intermediate] — Turning physics into a transfer function, why the Laplace transform is worth the trouble, and where poles and zeros come from.

From physics to a differential equation

Every plant model starts with a balance law. Three examples that cover most of what you’ll meet:

Mass–spring–damper:

RLC circuit (voltage across the capacitor):

DC motor (speed from voltage, neglecting inductance):

Notice the first two are the same equation. Mass ↔ inductance, damping ↔ resistance, spring ↔ 1/capacitance. That analogy is why control theory transfers across domains — the mathematics doesn’t know whether it’s describing a car suspension or a filter. → Hardware: Electricity

Why Laplace

Differential equations are awkward to compose. If you have a controller feeding a plant feeding a sensor, combining three ODEs means substitution and a mess.

The Laplace transform converts differentiation into multiplication:

With zero initial conditions, becomes .

So an ODE becomes an algebraic equation, and cascading systems becomes multiplying their transfer functions. That’s the entire payoff: composition becomes arithmetic.

The transforms worth knowing:

(step)
(ramp)
(delay)

That last one matters — a pure time delay is , which is not a rational function, and it’s why delay is so destructive to stability. → Frequency Response

The transfer function

For the mass–spring–damper:

A transfer function is a complete description of a linear time-invariant system’s input–output behaviour. Everything about how it responds — speed, damping, oscillation, stability — is encoded in that ratio of polynomials.

Written generally:

Poles () — roots of the denominator. Zeros () — roots of the numerator.

What poles mean

The single most important idea in classical control:

Poles determine the shape of the response. Their locations are the system’s natural modes.

Each pole contributes a term to the response:

Pole locationContributesBehaviour
Real, negative ()decaying exponential — stable
Real, positive ()growing — unstable
Complex pair ()damped oscillation
Pure imaginary ()sustained oscillation — marginally stable
At originconstantintegrator
              Im
               │      × ← unstable (right half plane)
     stable    │
        ×      │
   ────────────┼────────────  Re
        ×      │      ×
               │
    LEFT half plane = STABLE

That’s the criterion the entire field rests on. → Stability

Distance from the imaginary axis sets speed — a pole at decays ten times faster than one at . The pole closest to the axis dominates, which is what makes second-order approximation of high-order systems work in practice.

What zeros mean

Less intuitive, and they matter more than people expect.

Zeros don’t affect stability — they don’t appear in the characteristic equation. They shape the transient: they can add overshoot, speed up the initial response, or cancel a pole.

Right-half-plane zeros are the ones to watch. A system with an RHP zero is non-minimum phase, and it exhibits initial undershoot — it goes the wrong way first.

Real examples:

  • Backing a car with a trailer — to move the trailer left, you must first steer right
  • A bicycle — countersteering: to turn left at speed, push the bars right first
  • Boiler drum level — adding cold feedwater makes the level drop momentarily as steam bubbles collapse, before rising
  • Aircraft altitude via elevator — pitching up initially loses altitude

RHP zeros impose a hard limit on achievable bandwidth. You cannot control faster than an RHP zero allows, no matter how clever the controller — pushing gain up makes the loop unstable. That’s a fundamental constraint, not a tuning problem, and recognising a non-minimum-phase plant early saves a lot of wasted effort.

Block diagram algebra

The reason transfer functions are worth it — composition becomes arithmetic.

Series:
Parallel:

Negative feedback — the one to memorise:

with forward path and the feedback path. For unity feedback ():

is the loop gain, and it’s the central object of classical design. Note the closed-loop poles are the roots of — the characteristic equation. Everything in root locus and Nyquist is about the roots of that one equation.

And the insensitivity result from note 01 falls straight out: when , regardless of .

Linearisation

Transfer functions require linearity, and real systems aren’t. The standard move is to linearise about an operating point.

For , expand in a Taylor series about and keep the first-order terms:

The pendulum is the standard example:

Valid to about 20°, and the linear model is where the whole classical toolkit applies.

Three things to keep in mind:

A linearised model is local. Valid near the operating point and nowhere else. A model linearised at hover doesn’t describe aggressive flight.

Gain scheduling is the practical response: linearise at several operating points, design a controller for each, and interpolate. Standard in aerospace and engine control.

Some nonlinearities cannot be linearised away — saturation, backlash, friction, hysteresis. Those need nonlinear methods.

System identification

Often you can’t derive a model — the plant is too complex or poorly characterised. So you measure one.

Step response — apply a step, record the output, fit a model. The workhorse for process control, and the basis of Ziegler–Nichols tuning. A first-order-plus-dead-time fit

describes a surprising number of industrial processes with three parameters.

Frequency sweep — inject sinusoids across a range and measure gain and phase. Gives the Bode plot directly and is the most informative method, at the cost of time. → Frequency Response

Least-squares / ARX fitting — apply a rich excitation (PRBS is standard) and fit a discrete model numerically. What System Identification Toolbox and its open-source equivalents do.

Practical warnings:

  • Excite the frequencies you care about. A slow ramp tells you nothing about high-frequency behaviour
  • Stay in the linear region. Large inputs hit saturation and you’ll fit a model of the saturation
  • Validate on different data than you fitted to — the same discipline as any model fitting
  • An identified model is only valid over the range you excited. Extrapolation is guessing

Standard forms

Two you should recognise instantly, because most plants approximate one of them.

First order:

is the DC gain, the time constant — the response reaches 63% in , and 98% in . Thermal systems, simple RC circuits, motor speed.

Second order:

is the natural frequency, the damping ratio. Everything mechanical with mass and compliance. These two parameters determine the entire character of the response — the subject of the next note.


↑ Back to Control Theory