Aliasing occurs when sampling violates the Nyquist criterion, causing high-frequency components to 'fold back' into the passband as spurious low-frequency signals. Anti-aliasing filters remove high frequencies before sampling; reconstruction filters (interpolation) convert discrete signals back to continuous form while suppressing alias images.
From the Nyquist theorem, you know that a signal sampled at rate f_s can faithfully represent frequencies up to f_s/2. But the theorem's statement is also a warning: what happens when a component above f_s/2 is present at sampling time? It does not disappear. It reappears at a different, lower frequency — a ghost signal that was never in the original content. A 1100 Hz tone sampled at 2000 Hz does not cause a blank; it appears as a 900 Hz tone. That spurious tone is an alias, and once it is sampled in, it is indistinguishable from a genuine 900 Hz signal. The damage is irreversible.
The geometry of aliasing is a folding operation around the Nyquist frequency f_N = f_s/2. Think of the frequency axis folded in half at f_N. Any component at f_N + Δ folds back to f_N − Δ. A component at f_N + 500 Hz aliases to f_N − 500 Hz. If multiple aliases fold onto the same frequency, their amplitudes and phases combine — the resulting corruption is not even a simple impostor but a mixture. In the spectrum, you can visualize this as copies of the signal's baseband spectrum appearing at multiples of f_s, each reflected alternately. The baseband spectrum and all its alias copies sum into the sampled signal. The Nyquist criterion is satisfied when all these reflected copies are separated by adequate gaps, so only the original baseband is nonzero in the range [0, f_N].
The solution before sampling is an anti-aliasing filter — a lowpass filter placed ahead of the analog-to-digital converter that attenuates everything above f_N before the sampler can see it. Ideally this would be a brick-wall cutoff at exactly f_N, but real analog filters have gradual rolloff. Engineers handle this by choosing a sampling rate somewhat higher than 2f_max, leaving a transition band between f_max and f_N where the filter can roll off. Audio CD uses f_s = 44,100 Hz for content up to ~20,000 Hz, leaving a 2,050 Hz transition band for the anti-aliasing filter. Oversampled systems (high-definition audio at 192,000 Hz) gain a huge transition band, allowing simple, gentle anti-aliasing filters rather than steep analog filters — this simplification is often worth the extra storage cost.
Going the other direction — converting a sampled sequence back to a continuous signal — requires a reconstruction filter. The digital-to-analog converter produces a staircase or impulse train, whose spectrum contains the desired baseband plus alias images at f_s, 2f_s, 3f_s, and so on. A lowpass reconstruction filter passes the baseband and rejects all images, recovering the smooth analog waveform. Mathematically, perfect reconstruction uses a sinc interpolation kernel (the inverse Fourier transform of a rectangular spectrum), which reconstructs the signal exactly at the Nyquist rate. Practical filters approximate this with finite-impulse-response (FIR) or IIR designs that trade off sharpness for computational cost and phase linearity.
The anti-aliasing and reconstruction filters are bookends of every digital signal processing chain. Every audio interface, digital camera, and data-acquisition system includes them, often invisibly integrated into dedicated ICs. Understanding aliasing connects directly to practical measurement design: when you acquire vibration data at 10,000 samples/second and see an unexpected 1,200 Hz component, the first question is whether anything in your system generates a signal at 8,800 Hz (= f_s − 1,200 Hz) that the anti-aliasing filter failed to reject. Aliasing is the most common source of spurious spectral content in digital measurements, and recognizing its signature — a component that appears stronger or shifts in frequency as f_s is changed — is a fundamental diagnostic skill.