Nyquist-Shannon sampling theorem

The Nyquist-Shannon sampling theorem is the foundational bridge between the continuous and the discrete — the theorem that guarantees a bandlimited continuous signal can be perfectly reconstructed from its samples, provided the sampling rate exceeds twice the maximum frequency present in the signal. Named for Harry Nyquist, who first articulated the bandwidth criterion in 1928, and Claude Shannon, who proved the theorem rigorously in 1949, it is the mathematical justification for the entire digital world: every digital audio recording, every JPEG image, every video stream, every telephone conversation depends on this theorem.

The theorem states that if a signal is bandlimited — that is, its Fourier transform contains no frequency components above some maximum frequency B — then sampling the signal at a rate greater than 2B (the Nyquist rate) captures all the information in the original signal. The samples are sufficient to reconstruct the original continuous signal exactly, via convolution with a sinc function — the ideal low-pass filter. This is not an approximation. It is an exact equality, provided the conditions are met.

The Mathematics of Perfect Reconstruction

The theorem rests on the convolution structure of the Fourier transform. In the frequency domain, sampling a continuous signal at uniform intervals corresponds to convolving its Fourier transform with a periodic train of Dirac delta functions — a process that replicates the spectrum at integer multiples of the sampling frequency. If the sampling rate exceeds 2B, these replicated spectra do not overlap; the original spectrum can be isolated by an ideal low-pass filter and the original signal recovered. If the sampling rate is below 2B, the spectra overlap — a phenomenon called aliasing — and the original signal is lost irrecoverably. The high-frequency components fold into the low-frequency band, producing artifacts that no post-processing can remove.

The reconstruction formula is a direct consequence of the Fourier series representation of the sampled spectrum. The sinc function — sin(πx)/(πx) — emerges as the inverse Fourier transform of the ideal rectangular low-pass filter. Each sample is multiplied by a shifted sinc function, and the sum of these sinc interpolants reconstructs the original signal at every point. The sinc is the reconstruction filter implied by the theorem, and its infinite support means that perfect reconstruction requires knowing the entire signal — past, present, and future — which is why real-time digital systems use approximate reconstruction filters with finite impulse responses.

Why the Theorem Matters Beyond Engineering

The Nyquist-Shannon theorem is not merely a result in signal processing or digital signal processing. It is a structural claim about the relationship between continuous and discrete representations of information. It says that continuity is not an irreducible property of the world; it is a property that can be captured, preserved, and transmitted through discrete means. The theorem is the reason we can believe that a digital recording of a violin concerto 'is' the concerto — not a representation of it, but a complete information-theoretic equivalent.

The theorem also underlies the information theory framework that Shannon built: information is what remains when redundancy is removed, and the sampling theorem tells us exactly how much information a continuous signal contains. A bandlimited signal with bandwidth B contains exactly 2B independent samples per second — no more, no less. This is the information-theoretic compression of infinity into finitude. The continuous signal has infinitely many points, but only 2B degrees of freedom per second. The theorem is a statement about the effective dimensionality of function spaces, not merely about engineering practice.

The theorem's dependence on the Fourier basis is also philosophically significant. The reconstruction works because the Fourier transform diagonalizes the translation operator, and the sinc function is the dual basis in the time domain. The theorem assumes that the natural decomposition of the signal is into eternal sinusoids — the same assumption that underlies all of Fourier analysis. A signal that is not bandlimited in the Fourier sense may still be perfectly reconstructible from samples under a different basis, but the Nyquist-Shannon theorem does not apply. The theorem is frame-dependent: it is true for Fourier-bandlimited signals, not for all signals.

The Nyquist-Shannon sampling theorem is often taught as a guarantee: sample fast enough, and you lose nothing. This is the pedagogy of engineering confidence. But the theorem's hidden ontological commitments are more radical than its practical utility. It assumes that the signal is bandlimited in the Fourier basis, that the sampling is uniform, that the reconstruction is ideal, and that the signal extends infinitely in time. None of these conditions is met in any real system. The theorem is not a guarantee of perfection; it is a guarantee of what perfection would look like under a specific set of representational assumptions. The digital world is not a faithful copy of the analog world. It is a world built on the assumption that the analog world was Fourier-bandlimited all along — an assumption that is mathematically convenient, physically approximate, and ontologically unproven. The digital revolution did not digitize reality. It digitized a model of reality, and we have mistaken the model for the thing.