The Fourier transform decomposes into constant-amplitude sinusoids. But bells fade, circuits settle, and filters decay — oscillations with an envelope. The Laplace transform generalises by letting the exponent be a full complex number: the imaginary part is frequency, the real part is the decay rate. The Fourier transform is then just the slice where the decay rate is zero.
Sampling bends the frequency axis into a circle
Figure
Left, the s-plane: frequency runs up the vertical axis as a straight line, and the stable region is the entire left half. Right, the z-plane: that straight frequency axis has become the unit circle, and the stable left half-plane has become the inside of it. Every geometric statement carries across with the line bent into a loop — which is why discrete stability means poles inside the circle. It also explains aliasing: walk far enough around a circle and you arrive back where you started, so the discrete spectrum must repeat.
The geometric picture
A transfer function is a ratio of polynomials, and a polynomial is characterised by its roots. Numerator roots are zeros; denominator roots are poles, so named because the magnitude surface shoots upward there like a tent pole. Their positions determine the system completely, up to gain — a handful of points replacing a differential equation.
Poles, zeros, and why stability is a circle
Interactive
Peak at 0.085 · fsPoles × · Zeros ○Stable — poles inside the unit circle
The response on the right is not drawn — it is computed by the distance rule: walk around the unit circle, and at each frequency take the product of distances to the zeros divided by the product of distances to the poles. Poles pull the response up, zeros push it down. Slide the radius toward 1 and the peak sharpens, because the denominator distance is shrinking. Push past 1 and the filter is unstable — which is exactly why sharp recursive filters are fragile, and why high orders are built as cascaded biquads.
A notch filter is a zero placed on the frequency axis at the unwanted frequency. A resonator is a pole placed close to it. And the whole apparatus explains something we asserted two episodes ago: a lap around the circle returns to the start, so the discrete spectrum must repeat. Aliasing is what happens when sampling bends the frequency axis into a closed loop.
The best and worst of DSP
The best is that it does things not possible any other way — responses no component can produce, perfect reproducibility across every unit and temperature, systems that reconfigure themselves, images computed from measurements that never formed a picture.
The worst is that it is easy to get wrong in ways that look right. Alias a signal and you get a plausible spectrum that is a lie. Deconvolve carelessly and you amplify nothing into something. Use a sharp filter on shape-carrying data and you add features that were never there. In every case the output looks fine, nothing raises an error, and the conclusion is wrong.
Which is why the same question keeps returning: where does the information live, what is this tool assuming that I have not checked, and what would it look like if this were going wrong? The mathematics is the easy part. The judgement is the part that is yours.
Key points
The Laplace transform generalizes the Fourier transform by decomposing into sinusoids that may grow or decay, rather than only constant amplitude ones.
Writing the exponent as s equals sigma plus j omega splits e to the s t into an envelope and a rotation. Sigma is the decay rate, omega is the frequency. The Fourier transform is the Laplace transform restricted to the imaginary axis, where sigma is zero.
For ordinary systems, the transfer function is a ratio of polynomials. The numerator's roots are the zeros; the denominator's roots are the poles, so named because the magnitude surface shoots up there like a tent pole.
Pole and zero positions completely determine the system up to an overall gain. A handful of points replaces a differential equation or an infinite impulse response.
The magnitude of the frequency response is the product of distances to the zeros divided by the product of distances to the poles. Poles pull the response up, zeros push it down, and closeness to the frequency axis determines sharpness.
A notch filter is a zero on the frequency axis at the unwanted frequency. A resonator is a pole placed close to it.
A continuous system is stable if and only if all poles lie in the left half of the s plane. One pole on the wrong side means exponential growth, and it is fatal rather than merely undesirable.
Butterworth filters place poles evenly on a circle, producing maximum flatness. Chebyshev filters place them on an ellipse squashed toward the axis, which sharpens each peak — giving both the faster roll-off and the passband ripple from one geometric decision.
The z transform is the discrete equivalent, related by z equals e to the s T. The imaginary axis becomes the unit circle, the stable left half plane becomes the inside of that circle, and the unstable right half becomes the outside.
Discrete stability therefore means all poles strictly inside the unit circle — the claim from episode fourteen, now explained.
The frequency response is found by walking around the unit circle: zero frequency at z equals one, Nyquist at z equals minus one, negative frequencies on the way back round.
Because a lap around the circle returns to the start, the discrete frequency response must repeat. Aliasing is what happens when sampling bends the frequency axis into a closed loop.
Sharp recursive filters need poles very close to the unit circle, so small coefficient rounding can push one outside and destabilize the filter. Cascading biquads helps because each section's pole positions depend on only a couple of coefficients.
Converting a recursion to a transfer function is mechanical: each one sample delay becomes a factor of z to the minus one. Cascaded systems multiply transfer functions; parallel systems add them.
Digital filter families are not derived from scratch. Known analog designs — Butterworth on a circle, Chebyshev on an ellipse — are mapped into the z domain to produce the recursion coefficients.
The failures in this field are dangerous because they look like successes. Aliasing, careless deconvolution, ringing filters on shape-carrying signals, and marginally stable recursions all produce plausible output with no error raised. The mathematics is the easy part; the judgement is yours.