Filter (signal processing)

A filter, in the signal-processing sense, is an LTI system whose purpose is to emphasize signal power in one frequency range while deemphasizing power in another. Every LTI system is technically a filter — its frequency response $H(f)$ determines how it weights different frequencies — but we use the word specifically when this frequency-selective behavior is the intended purpose.

Ideal vs practical

An ideal filter passes signal power in its passband without distortion and completely blocks signal power outside. “Without distortion” means $|H(f)|$ is constant in the passband and the phase $\angle H(f)$ is linear in $f$ (so all frequencies are delayed by the same amount — multiplying by a constant or shifting in time isn’t distortion; changing the shape is).

Real filters approximate the ideal but never reach it:

• The transition between passband and stopband is gradual, not sharp.
• The passband isn’t perfectly flat.
• The phase isn’t perfectly linear.

Filter design is the engineering art of choosing among these compromises.

Bandwidth definitions

A filter’s bandwidth is the width of its passband. Three definitions are in common use:

• Absolute bandwidth: the range where $|H(f)|$ is exactly nonzero. Only meaningful for ideal filters; real filters have nonzero tails everywhere.
• Half-power bandwidth (or 3 dB bandwidth): the range where $|H(f){|}^2$ is at least half its peak value. Equivalently, $|H(f)|\ge (1/\sqrt{2})|H{|}_{\max }$, a magnitude drop of $3\text{dB}$. The most common practical definition.
• Null bandwidth: the range between the first nulls (zeros of $|H(f)|$) on either side of the peak. Useful for sinc-shaped responses.

Bandwidth is traditionally measured on the positive frequency axis only — even though the spectrum extends to negative frequencies. A lowpass filter with cutoff at $1\text{kHz}$ has a $1\text{kHz}$ bandwidth (from 0 to 1 kHz), not 2 kHz.

The four canonical filter types

• Lowpass (LPF): passes near DC, blocks high frequencies.
• Highpass (HPF): blocks DC, passes high frequencies.
• Bandpass (BPF): passes a band around some center frequency.
• Bandstop (BSF) / notch: blocks a narrow band, passes everything else.

These four cover almost every filter in practical use. The transfer functions are on the formula sheet; the pattern is consistent — poles in the left half-plane attenuate frequencies near their imaginary parts, zeros in the left half-plane (or at the origin) kill response at their imaginary parts.

Ideal filter impulse response

The impulse response of an ideal lowpass filter is a sinc:

$h(t)=2A{f}_m\mathrm{sinc}(2f_m(t-t_0)),$

(with the time shift $t_0$ added for causality, but the sinc extends to $t=-\infty$, so an ideal filter is fundamentally non-causal). The slow $1/t$ decay of the sinc means an ideal filter has long ringing in time — oscillations that don’t die out fast. This is why filters with sharp frequency cutoffs tend to ring around sharp signal transitions.

Practical first-order filters (like an RC lowpass) don’t have these problems — they’re causal and don’t ring — but they don’t have brick-wall responses either; the transition is gradual.

The trade-off between time-domain niceness and frequency-domain sharpness is fundamental — it’s the time-bandwidth tradeoff in another guise.

Where filters appear

• Audio amplifiers: lowpass to remove hiss; highpass to remove rumble.
• Radio receivers: bandpass to select one station.
• Power supplies: notch at 60 Hz to remove line-frequency hum.
• Digital cameras: anti-aliasing lowpass to bandlimit before sampling.
• Control systems: lead/lag compensators that shape phase and amplitude.

The Bode plot is the standard graphical representation of a filter’s frequency response and the tool used to design and analyze them.

Filter families

For sharper rolloffs than first-order filters give, you go to higher-order filters. The main families:

• Butterworth: maximally flat passband, gentle rolloff.
• Chebyshev type I: sharper rolloff at the cost of passband ripple.
• Chebyshev type II: sharp rolloff with stopband ripple.
• Elliptic (Cauer): sharpest possible rolloff, with both passband and stopband ripple.
• Bessel: maximally linear phase, gentle rolloff.

Filter design then becomes choosing a family, choosing an order, and choosing where to place the poles and zeros to meet specifications. Done with software in practice (MATLAB butter, cheby1, etc.); the engineer specifies type and specs, the software produces $H(s)$.

Simplest realisations (Electronics I)

At the circuit level, a filter is any network that responds differently to different frequencies. The four canonical responses — lowpass, highpass, bandpass, bandstop — and the very simplest physical realisations of the first two are a single resistor and capacitor: the RC lowpass filter (series $R$, shunt $C$) and the RC highpass filter (series $C$, shunt $R$). These first-order RC networks recur constantly inside amplifiers — as coupling capacitors, bypass paths, and parasitic bandwidth limits — which is why every microelectronics circuit is, in part, a collection of implicit filters.