Skip to content

Fourier-Basis Expansion

Transform method — instantiates Independent Generating Set Design

Represents any signal in a fixed, universal orthonormal basis of sinusoids, turning it into frequency coordinates that reconstruct it exactly.

Fourier-Basis Expansion represents a signal in a basis you do not have to discover: the sinusoids of every frequency. Its defining move is that the basis is prescribed and universal — the same complex exponentials work for every signal on the interval, so there is nothing to fit, extract, or prune. Because the sinusoids are mutually orthogonal, each coordinate is just the signal's inner product with one frequency, the whole set of coordinates is the spectrum, and summing the sinusoids back weighted by those coordinates reconstructs the original exactly. Where most mechanisms in this archetype build a basis from vectors in hand, Fourier expansion adopts one whose completeness and independence are guaranteed in advance — its job is to encode and reconstruct, not to select.

Example

A tuner app listens to a sustained guitar note that sounds slightly off. The waveform in time is a nondescript wiggle. Fourier-Basis Expansion projects that wiggle onto the sinusoid basis: the coordinates come back as a sharp spike at the fundamental frequency plus a comb of smaller spikes at its harmonics. Reading the fundamental's coordinate against the twelve-tone standard shows it sitting a few cents flat.

The point is that the same fixed basis handled a signal it had never seen. Nobody fit a model to this note; the sinusoids were waiting, orthonormal, and the note's identity fell out as a handful of coordinates. Inverting the transform — resumming the sinusoids weighted by those coordinates — rebuilds the original waveform to the sample, which is what makes the spectrum a faithful re-coordinatization rather than a lossy summary.

How it works

  • Adopt the prescribed basis. The generators are the complex exponentials (equivalently sines and cosines) at each admissible frequency; no candidate pool is searched.
  • Analyze into coordinates. Each coordinate is the signal's inner product with one sinusoid — cheap and exact precisely because the basis is orthonormal (a fast transform computes them all at once).
  • Synthesize back. The inverse sum rebuilds the signal from its coordinates; orthonormality makes analysis and synthesis mutual inverses.

What distinguishes it from its neighbours is that completeness and independence are theorems here, not steps — the method spends its effort on encoding and reconstruction, not on proving it has a basis.

Tuning parameters

  • Frequency resolution (record length) — a longer analysis window resolves closely-spaced frequencies but assumes the signal's content holds still across it.
  • Windowing / apodization — tapering the record before transforming suppresses spectral leakage from the finite window, at the cost of blurring resolution.
  • Real vs. complex form — cosine/sine pairs versus complex exponentials; the complex form carries phase cleanly, the real form is easier to read.
  • Basis truncation — how many harmonics to retain; dropping high frequencies band-limits and smooths, but erases sharp edges.
  • Sampling rate — sets the highest representable frequency; too low and high content folds back as aliasing.

When it helps, and when it misleads

Its strength is that the basis is free, exact, and physically meaningful: for stationary and periodic signals the coordinates are the frequency content, energy is preserved across the transform (Parseval), and the fast transform makes it cheap. No other mechanism here gives a complete orthonormal basis with zero fitting.

It misleads whenever the signal is not globally stationary. A transient or a jump discontinuity has no compact frequency description, so its energy smears across the whole spectrum and a truncated expansion overshoots the edge — the Gibbs phenomenon[1] — leaving ripples that never fully vanish. The classic misuse is reading a leakage artifact of the finite window as a real spectral line, or forcing the global basis on a signal whose content moves in time. The discipline is to check stationarity, window deliberately, and reach for a localized basis — Wavelet Multiresolution Analysis — when the interesting structure is transient.

How it implements the components

Fourier-Basis Expansion fills the prescribed-basis, encode-and-reconstruct side of the archetype:

  • admissible_generation_operations — the combination rule is harmonic superposition: every admissible signal is a weighted sum of the basis sinusoids.
  • coordinate_encoding_contract — the coordinates are the inner products of the signal with each sinusoid; that vector of coefficients is the spectrum.
  • synthesis_and_reconstruction_map — the inverse transform resums the sinusoids to rebuild the signal exactly.
  • orthonormality_or_normalization_constraint — the sinusoids are mutually orthonormal, so coordinates are independent and the transform preserves energy.

It does not extract or prune a basis from candidate vectors — that is Basis Extraction from a Spanning Set — nor build a localized, multiscale basis, which is Wavelet Multiresolution Analysis; and it needs no biorthogonal partner, unlike the Dual-Basis Transform.

Notes

Fourier is the special self-dual case of a basis: because the sinusoids are orthonormal, the analysis basis and the synthesis basis coincide, so the biorthogonal machinery a Dual-Basis Transform needs for a non-orthogonal basis collapses to nothing here. That collapse is exactly why the expansion is so cheap — and why abandoning orthonormality (for a wavelet packet, say, or an oblique basis) reintroduces a distinct dual basis you then have to track.

References

[1] Gibbs phenomenon — near a jump discontinuity, a truncated Fourier series overshoots the true value by a fixed proportion that narrows but does not shrink as more terms are added. It is the standard illustration of why a global sinusoid basis represents localized features poorly. ↩