The discrete Fourier transform in Mathematica

The Fourier transform is a very useful and widely used tool in many branches of maths. The continuous Fourier transform, available in Mathematica as FourierTransform, tries to perform the Fourier transform of a symbolic expression. This is useful for solving many problems if the Fourier transform can also be expressed in a nice symbolic form, but such a form is not necessarily available (and indeed the Fourier transform might not exist at all).

In practice, the discrete Fourier transform, which works with a finite set of data points and approximates the continuous Fourier transform, is more important. The DFT is available in Mathematica as Fourier, and the inverse DFT as InverseFourier. Note that Mathematica uses the 1/\sqrt{n} normalisation by default for both the DFT and the IDFT.


Here is an example usage of the DFT in Mathematica:

xmax = 1;
nx = 128;
dx = xmax / nx;
xs = Range[0, xmax - dx, dx];
f[x_] := Exp[Sin[6 Pi x]] + Sin[40 Pi x];
data = Map[f, xs];
ListLinePlot[Transpose[{xs, data}]]

dft = Fourier[data];
numax = 1/(2*dx);
dnu = 1/xmax;
nus = Join[Range[0, numax - dnu, dnu], Range[-numax, -dnu, dnu]];
Transpose[{nus, Abs[dft]}],
PlotRange -> All]

The first block generates a list data which consists of sampled values of the function f at the values xs, which are 128 values between 0 and 1 (not including the endpoint at 1). These sampled points are plotted.

The second plot calculates the discrete Fourier transform of data (according to the default normalisation). Then the list nus is constructed: these are the wavenumbers that correspond to each element of dft. (With the default normalisation, this is not the angular wavenumber: the wavenumber of \sin(40\pi x) is said to be 20, while the angular wavenumber is 40\pi. The symbol k is often used for the angular wavenumber; I am using \nu for the wavenumber.) We then plot the absolute values of dft against the wavenumbers nus.


Note how the values of \nu_{\max} and \Delta\nu are related to the values of \Delta x and x_{\max}:

\nu_{\max} = \frac{1}{2\Delta x} \text{ and } \Delta\nu = \frac{1}{x_{\max}}.

What this means is the following:

  • If you sample more densely (so \Delta x is smaller), then you can pick up higher frequencies.
  • If you sample for a longer domain (so x_{\max} is larger), then you can pick up lower frequencies.

The highest frequency that you can pick up, $1/2\Delta x$, is called the Nyquist frequency and is equal to half the sampling rate of the data. If your function in fact oscillates at a higher frequency than this, then your sampled data will not pick up that component properly, and many of your results will be inaccurate. This effect is known as aliasing.


The Fourier transform can be used to solve many different problems, and is particularly useful for solving linear partial differential equations. Numerical PDE solvers that make use of the DFT are part of a class of numerical schemes known as spectral methods. For linear problems where the solution is oscillatory or where a ‘conservation law’ (usually analogous to conservation of mass or momentum) applies, spectral methods tend to be better than finite differences. The DFT can also be useful for quasilinear PDEs, using a family of methods known as pseudospectral methods.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.