Aliasing

Aliasing is the phenomenon that occurs when a signal is sampled at a rate below its Nyquist rate, causing high-frequency components to fold into the low-frequency band and become indistinguishable from genuine low-frequency signals. The term derives from the Latin alias ('otherwise'), because the aliased frequency appears to be a different frequency than it actually is — a false identity imposed by the sampling process.

In the frequency domain, aliasing is the overlap of spectral replicas. When a continuous signal is sampled at rate $, its spectrum is replicated at integer multiples of $. If the original signal contains frequency components above /2$ — the Nyquist frequency — the replicas overlap, and the energy from the high-frequency region folds into the baseband. The aliased frequency $ of a component at frequency $ is given by:

1021714f_a = |f - k f_s|1021714

where $ is the integer that minimizes the absolute value. A signal at 7 kHz sampled at 10 kHz appears at 3 kHz — an alias that cannot be separated from a genuine 3 kHz component by any subsequent processing.

Aliasing is not merely an engineering inconvenience. It is an information-theoretic catastrophe: the aliased components are not merely distorted; they are indistinguishable from legitimate signals. Once sampling has occurred below the Nyquist rate, the original signal cannot be recovered. The information is lost not through noise but through structural ambiguity — the mapping from continuous to discrete becomes many-to-one, and the inverse does not exist.

The standard remedy is anti-aliasing filtering: a low-pass filter applied before sampling to ensure that no frequency components above /2$ reach the sampler. But this introduces a different compromise: the anti-aliasing filter itself has a transition band, and sharp cutoffs require high-order filters with long impulse responses, introducing latency and phase distortion. The choice of sampling rate is therefore not merely a matter of applying the Nyquist theorem but of trading off aliasing against filter complexity, latency, and phase fidelity.

Aliasing appears beyond signal processing. In computer graphics, spatial aliasing produces jagged edges and moiré patterns. In numerical analysis, aliasing of high-wavenumber modes into low-wavenumber modes is the mechanism behind the instability of under-resolved simulations of turbulent flows. In any domain where a continuous structure is represented by discrete samples, the same hazard exists: frequencies too high for the sampling grid will masquerade as frequencies that fit.

Aliasing is the revenge of the continuous on the discrete. We sample to make the world manageable, and the world responds by wearing masks. The high-frequency detail that we could not afford to capture does not disappear politely; it reappears in disguise, corrupting what we thought we had preserved. Aliasing is the proof that compression is not free, that the decision to discretize is a decision about what to destroy, and that the destruction, if careless, will haunt the reconstruction.