Fourier Transform: Complex Rotations, Not Automatic Spirals

The Fourier transform decomposes a signal into oscillatory components. Its complex representation is often drawn as rotating vectors and then as a chain of epicycles. The picture is powerful, but it invites a confusion: complex rotation is not automatically a spiral. Radius may remain constant while phase advances.

The central question is therefore twofold: how does projection onto rotations measure frequency, and under what conditions does the resulting trajectory become genuinely spiral? The distinction supports the site’s thesis: visual resemblance alone does not identify a mechanism.

A real sinusoid oscillates on a line. Euler’s formula lets us see it as the projection of a point moving around the unit circle: the horizontal coordinate gives cosine and the vertical coordinate gives sine. A complex number stores both quadratures at once. MIT multivariable-calculus notes present this correspondence among the complex exponential, sine, cosine, and polar form. ^[1]

If θ = ωt, the point rotates at angular speed ω. Its modulus is exactly 1. The curve is closed and periodic: it is a circle. “Winding” correctly describes the phase; “spiral” is incorrect until radial evolution is introduced.

For an integrable function f(t), the continuous transform measures its correlation with each rotation of frequency ω. The negative sign rotates the kernel opposite to the tested signal. If f contains a coherent component at that frequency, products stop cancelling and their integral forms a non-negligible complex vector.

The magnitude |F(ω)| describes component strength; the argument arg F(ω) describes relative phase. No material point necessarily follows this curve. F is a representation: it converts time dependence into a frequency distribution.

3.1 Orthogonality and cancellation

Over an interval containing an integer number of periods, two exponentials of distinct frequencies cancel on average. This orthogonality explains selectivity. On a finite window that does not respect the periods, cancellation is incomplete: energy from one line spreads into neighboring frequencies.

A computer observes N samples $x_{n}$ and computes N coefficients. The discrete Fourier transform is not a vague approximation to the continuous transform: it is an exact algebraic operation on a finite vector, with precise assumptions about periodicity and sampling.

Direct evaluation takes a number of operations proportional to $N^{2}$. The Cooley–Tukey algorithm factors the DFT into subproblems and brings the usual cost down to order N log N for favorable sizes. Their 1965 paper did not invent Fourier analysis, but it made large-scale numerical computation far more accessible. ^[2]

Truncating a signal means multiplying it by a window. In frequency, this multiplication becomes convolution: the ideal spectrum is spread by the window spectrum. Leakage is therefore not a mysterious FFT defect; it follows from finite observation and edge discontinuities.

Hann, Hamming, Blackman, and Kaiser windows trade main-lobe width against side-lobe height. Harris systematized these compromises for harmonic analysis with the DFT. There is no universally best window: the choice depends on desired frequency separation and amplitude dynamic range. ^[3]

A damped complex exponential has rotating phase and decreasing magnitude. Its trajectory in the complex plane is a converging logarithmic spiral. With positive growth, the same equation produces a diverging spiral.

This form appears in eigenmodes of dissipative linear systems. It belongs naturally to the Laplace transform, whose complex variable s = σ + iω combines radial growth or decay with phase rotation. Fourier corresponds to the axis σ = 0: constant kernel magnitude, hence circular rotation.

A global spectrum says which frequencies occur over the full duration, but not when they occur. A short-time Fourier transform slides a window through time and builds a spectrogram. A short window localizes events better but separates frequencies less finely; a long window does the reverse.

This time–frequency limit is not merely an interface problem. It follows from the mathematics of the analysis: asking simultaneously for a very precise time and a very precise frequency requires incompatible information within a finite window.

• Audio: pitch, timbre, filtering, and perceptual compression
• Images: spatial filtering, diffraction, and reconstruction
• Telecommunications: modulation and band sharing
• Physics: normal modes, propagation, and diffraction
• Data: periodicity detection and noise reduction

Frequency decomposition alone does not reveal the physical cause of a signal. Two different mechanisms can produce similar spectra; conversely, one nonstationary mechanism can produce different spectra under different windows. Fourier organizes components, but causal interpretation requires a model of the system.

The Fourier transform measures coherence with complex rotations. The circle belongs to the kernel; a spiral appears when amplitude evolves with phase. This separation avoids a misleading metaphor while preserving geometric intuition: analyzing a signal means testing how much rotations at different speeds add instead of cancelling.