Fourier Transform · aiwiki.page

The Fourier transform is a mathematical operation that converts a function of time or position into a function of frequency. It represents the original function through complex exponential components, which encode sinusoidal oscillations with different amplitudes and phases. Together with its inverse, it connects two descriptions of the same information: one in the original domain and another in the frequency domain. It is a central tool in harmonic analysis, signal processing, and mathematical physics. (numpy.org)

Historical development

The transform is named after Joseph Fourier, whose Théorie analytique de la chaleur, published in 1822, developed methods for studying heat conduction using trigonometric expansions. His work helped establish the connection between physical processes and the decomposition of functions into oscillatory components. (people.math.harvard.edu)

A Fourier series represents a periodic function using discrete harmonics. The Fourier transform extends this framework to functions on an unbounded domain, with frequency generally varying continuously. The distinction is important: a periodic function has a discrete harmonic description, whereas a nonperiodic function can require a continuous spectrum. (math.utah.edu)

Definition and interpretation

One common convention defines the transform of an absolutely integrable function $f(t)$ by

\widehat f(\nu)=\int_{-\infty}^{\infty} f(t)e^{-2\pi i\nu t}\,dt.

Here $i^2=-1$ , and $\nu$ denotes frequency in cycles per unit of $t$ . The integral measures the contribution associated with each oscillatory component. Under suitable hypotheses, the inverse formula is

f(t)=\int_{-\infty}^{\infty} \widehat f(\nu)e^{2\pi i\nu t}\,d\nu.

For example, if both $f$ and $\widehat f$ are absolutely integrable, inversion holds almost everywhere, and pointwise wherever $f$ is continuous. (math.utah.edu)

Other conventions use angular frequency $\omega=2\pi\nu$ , placing a factor $1/(2\pi)$ in the inverse formula, or distribute normalization symmetrically between the two transforms. These conventions describe the same operation but produce different constants in formulas. (math.utah.edu)

The transformed function usually takes complex values. Its magnitude describes spectral strength, while its argument describes phase. Phase cannot generally be discarded without losing information needed for reconstruction. For a real-valued input, positive and negative frequencies satisfy conjugate symmetry: $\widehat f(-\nu)=\overline{\widehat f(\nu)}$ . Negative frequencies are therefore part of the mathematical representation, not additional independent components of a real signal. (numpy.org)

Mathematical properties

The transform is a linear map: transforming a linear combination gives the same linear combination of the transforms. A translation $f(t-t_0)$ multiplies the spectrum by $e^{-2\pi i\nu t_0}$ ; multiplying $f(t)$ by $e^{2\pi i\nu_0t}$ shifts its spectrum by $\nu_0$ . Scaling obeys

\mathcal F\{f(at)\}(\nu) =\frac{1}{|a|}\widehat f(\nu/a), \qquad a\ne0.

Thus compression in one domain corresponds to expansion in the other. (math.stanford.edu)

Under suitable regularity and decay assumptions, a derivative becomes multiplication by frequency:

\mathcal F\{f'\}(\nu)=2\pi i\nu\,\widehat f(\nu).

The convolution of two functions becomes the pointwise product of their transforms. These properties simplify the analysis of filters and constant-coefficient differential equations. (math.stanford.edu)

Plancherel’s theorem extends the transform to square-integrable functions and, with the convention above, gives

\int_{\mathbb R}|f(t)|^2\,dt = \int_{\mathbb R}|\widehat f(\nu)|^2\,d\nu.

It makes the transform a unitary operator on the Hilbert space $L^2(\mathbb R)$ . For signals, this expresses preservation of integrated squared magnitude, often interpreted as signal energy. (math.utah.edu)

Discrete computation

For a finite sequence $x_0,\ldots,x_{N-1}$ , the discrete Fourier transform (DFT) is commonly defined by

X_k=\sum_{n=0}^{N-1}x_n e^{-2\pi i kn/N}, \qquad k=0,\ldots,N-1.

Its inverse uses the opposite exponential sign and a factor $1/N$ . The DFT produces a finite set of frequency coefficients and is naturally associated with a periodic extension of the sequence. (numpy.org)

A fast Fourier transform (FFT) is an efficient algorithm for calculating the DFT, not a different mathematical transform. Direct evaluation requires $O(N^2)$ arithmetic operations; standard FFT methods reduce this to $O(N\log N)$ . This reduction in computational complexity enables large-scale spectral calculations and convolution. (fftw.org)

Sampling and spectral limitations

A computed spectrum reflects both the underlying signal and how it was measured. Sampling can cause aliasing, in which distinct continuous frequencies become indistinguishable. The Nyquist–Shannon sampling theorem establishes reconstruction conditions for band-limited signals; spectral content must remain below half the sampling rate in the usual strictly band-limited formulation. (docs.scipy.org)

Finite observation introduces spectral leakage: restricting a signal to a time interval spreads its spectral components. A window function can reduce leakage away from peaks, but typically broadens those peaks. Zero-padding gives a denser frequency grid without creating additional measured information or improving the intrinsic ability to separate nearby components. (docs.scipy.org)

Applications

In signal processing, Fourier methods reveal periodic components and implement frequency-selective filtering. In image processing, multidimensional transforms describe spatial frequencies and support filtering and reconstruction. In optics, they describe diffraction patterns; in X-ray crystallography, Fourier relationships connect crystal structure with diffraction measurements. (numpy.org)

A global transform describes frequency content without directly showing when that content occurs. The short-time Fourier transform addresses this limitation by transforming successive windowed portions of a signal, producing a time–frequency representation whose temporal and frequency resolution depend on the window. (docs.scipy.org)