[Intermediate] — Bode plots, what “bandwidth” actually means, and why frequency-domain thinking is how working engineers reason about loops.
The idea
Feed a linear system a sinusoid and, after transients die, the output is a sinusoid at the same frequency — only the amplitude and phase differ.
That’s it. Substitute into the transfer function and you get a complex number whose magnitude is the gain and whose angle is the phase shift, at that frequency.
Sweep across the range of interest and you have a complete description of the system — one that you can measure directly on real hardware without knowing the model. That last point is why the frequency domain dominates practical control: the Bode plot of a plant is something you can obtain experimentally on a plant nobody has ever modelled.
Bode plots
Two plots against :
- Magnitude in decibels:
- Phase in degrees:
Why log scales. They turn multiplication into addition. Since cascaded systems multiply, on a Bode plot they add — so you can sketch the plot of a complicated transfer function by adding the plots of its factors. That’s the whole reason for the format.
Decibels worth memorising:
| Gain | dB |
|---|---|
| 0.01 | −40 |
| 0.1 | −20 |
| 0.5 | −6 |
| 1 | 0 |
| 2 | +6 |
| 10 | +20 |
| 100 | +40 |
0 dB means gain of 1 — the input passes through unchanged. That frequency is where everything interesting happens.
The building blocks
Every rational transfer function is a product of these, so its Bode plot is their sum:
| Factor | Magnitude | Phase |
|---|---|---|
| Gain | flat at | 0° |
| Integrator | −20 dB/decade, all | −90°, all |
| Differentiator | +20 dB/decade | +90° |
| Pole | flat, then −20 dB/dec after | 0° → −90° |
| Zero | flat, then +20 dB/dec after | 0° → +90° |
| Delay | flat at 0 dB | , unbounded |
Each pole costs you 20 dB/decade of rolloff and 90° of phase lag. Each zero gives them back.
A first-order pole in detail:
dB
0 ┼────────────╮
│ ╲ −20 dB/decade
−20 ┤ ╲
└──────┬────────╲──────→ log ω
ωc
deg
0 ┼──────╮
│ ╲
−45 ┤ ● ← at ωc, phase is exactly −45°
│ ╲
−90 ┤ ╰────────
At the corner frequency the magnitude is −3 dB and the phase is −45°. The asymptotes miss by 3 dB there, which is close enough for design work.
The delay row is the one to stare at. A pure delay costs you nothing in magnitude and unbounded phase. It gives you no benefit and takes away the thing keeping you stable. → Time delay
Sketching by hand
Still worth being able to do, because it builds the intuition that the software plot doesn’t:
- Start at low frequency with the gain and any integrators (slope = −20 dB/dec per integrator)
- At each pole corner, subtract 20 dB/decade from the slope
- At each zero corner, add 20 dB/decade
- Phase: each pole contributes −90°, spread over roughly a decade either side of its corner; each zero +90°
The final slope is dB/decade — the pole excess sets the high-frequency rolloff.
Second-order resonance
Rolls off at −40 dB/decade past with 180° of total phase lag. But for there’s a resonant peak:
dB
│ ╱╲ ← peak grows as ζ → 0
0 ┼─────╯ ╲
│ ╲╲ −40 dB/dec
└──────┬────╲╲──→ log ω
ωn
Lightly damped modes show up as sharp peaks, and they’re the practical danger in mechanical systems. Structural resonances, drive-belt compliance, unmodelled flexible modes — a peak that crosses 0 dB destabilises the loop even if the rigid-body design was fine. It’s why you notch-filter a known resonance, and why a controller tuned on a rigid model can go unstable on the real machine.
Loop shaping
The central design idea in the frequency domain, and the thing most worth taking away from this note.
You don’t design the closed-loop response directly. You shape the open-loop gain , because it’s what you can see and what the controller directly modifies.
Recall from note 02:
is the complementary sensitivity (reference to output), the sensitivity (disturbance to output).
- Where : , — good tracking, good disturbance rejection
- Where : , — no tracking, and noise is rejected
So:
Want high gain at low frequency and low gain at high frequency. High gain where the references and disturbances live; low gain where the sensor noise and unmodelled dynamics live.
|L| dB
│────╲ ← high gain: tracking, disturbance rejection
│ ╲
0 ┼──────●────────── ← crossover ωc: sets bandwidth,
│ ╲ and phase margin is measured HERE
│ ╲────── ← low gain: noise rejection, robustness
└──────────────→ log ω
The crossover frequency — where — is the single most important number in a design:
- It sets the closed-loop bandwidth, hence the speed: , and rise time
- The slope through crossover should be about −20 dB/decade. Steeper means more phase lag at crossover, hence less margin. This is Bode’s gain–phase relationship doing its work
- Phase margin is measured here, and it’s what determines the damping
And is the algebraic statement of the constraint you can’t escape: you cannot make both small at the same frequency. Rejecting a disturbance at means tracking noise at .
Bandwidth
is the frequency where the closed-loop gain drops to −3 dB. The system tracks references below it and ignores those above it.
Higher bandwidth means faster response. The connections:
And higher bandwidth costs you:
- More noise passes through — you’re tracking at frequencies where the signal is mostly noise
- More actuator effort, so saturation becomes likely
- Less robustness, because high-frequency model error is exactly where your model is worst
The hard limits on bandwidth are worth knowing, because they’re the ones you can’t tune your way past:
| Limit | Rough cap on |
|---|---|
| RHP zero at | |
| Time delay | |
| Unstable pole at | (a lower bound — you must be fast enough) |
| Actuator | can’t demand what it can’t deliver |
| Sensor noise | can’t act on what you can’t measure |
Note the third: an unstable plant forces bandwidth up, while an RHP zero or delay forces it down. A plant with both an unstable pole and an RHP zero close together may be impossible to control at all — the required bandwidth exceeds the permitted one. That’s a fundamental result, not a design failure, and it’s the sort of thing worth checking before spending a month tuning.
Phase and gain margin
Defined here, used properly in the next note.
Gain margin — how much you can multiply the gain before instability. Read at the frequency where phase = −180°.
Phase margin — how much extra phase lag you can tolerate. Read at crossover, where .
dB
│────╲
0 ┼─────╲●──────────
│ ╲ ↕ gain margin
└───────╲──┬──→
ω(−180°)
deg
│──╲
│ ╲ ↕ phase margin
−180┼────╲──●────────
└────┬────────→
ωc
Targets: PM ≥ 45°, GM ≥ 6 dB. Below ~30° phase margin a system is oscillatory and fragile.
The rule of thumb worth memorising:
So 45° phase margin gives roughly — noticeable overshoot but usable. 70° gives , the classic target. This is the bridge between frequency-domain design and time-domain specifications, and it’s why an engineer can tune in the frequency domain while thinking about overshoot.
Why work here at all
Given root locus exists, the case for the frequency domain:
You can measure it. A frequency sweep gives you the plot with no model at all. You cannot measure a root locus.
Delays are exact. is just a phase shift. In root locus, delay requires a rational approximation.
Uncertainty is naturally expressed here. “My model is good to 100 rad/s and I don’t trust it above that” is a frequency-domain statement, and robust control is built on it. → Nyquist and Robustness
Margins are meaningful. “45° of phase margin” is a real, defensible robustness claim. “Poles at ” tells you nothing about how much the plant can change.
It composes. Cascaded systems add on a Bode plot — which makes it possible to see, at a glance, what a compensator does to a loop.
Related
- Stability and Root Locus — the time-domain view
- Nyquist and Robustness — margins done properly
- PID Control — what P, I and D do to these plots
- Control theory map
Software engineering
Infrastructure & security
Computer science
Hardware & engineering