Consider a function expanded in terms of orthonormal basis functions of a given vector space
One can change the representation by using other set of basis functions. For example, the function could be expanded in terms of (Legendre series), or (Laguerre series) or (Fourier series).
The coefficients are computed through
Now imagine instead of discrete functions , , we had a continuous set of functions (replace ). Then instead of , we would have a function that depends on the continuous parameter , which we can call .
Then, we could represent a function through the "coefficients"
defined as
where is called the kernel of the integral transform .
Note that the "coefficients" are not constants anymore (as it is the case in a series expansion), but a function of the continuous parameter .
Different kernels produce different transforms. Some examples follow below.
Laplace transform
Fourier transform
Mellin transform
All those transforms are linear because they are inside integrals, which are linear operators. A given integral transform has an inverse transform defined by
Fourier transform
Simplified derivation
The expansion of a function in a Fourier series requires the function to be defined within an interval and be periodic. As a result, we obtain a sum (series) of harmonic functions with discrete amplitudes. In order to remove these constraints (namely, to use a non-periodic function and continuous amplitudes), we need to use a Fourier transform.
Suppose we take a periodic function . We build a Fourier series, but this time, let's define the angular frequency (which implies , since because ). Substituting, we have
(the term inside the brackets are the coefficients ). Now, we do , which implies that the sum becomes an integral and the angular frequency becomes continuous:
The innermost integral (the one inside the brackets) is called the Fourier transform of , and the outermost integral is called the inverse Fourier transform. There are different conventions on how to deal with the factor of : some authors distribute it equally between the two transforms as factors of , others use the factor of only on the inverse Fourier transform. In this document, we take the factor only on the inverse transform.
Formal definition
For a function , the function is called the Fourier transform of , and one can use for it the notation .
The interpretation is as follows. The Fourier transform gives out a complex function of the angular frequency. It originated from the coefficients of the Fourier series, so it can be interpreted as the (complex) amplitudes of the sum of (complex) harmonic functions that describe . The fact that those functions are complex means that the Fourier transform also encodes information on the phase of those harmonic functions (think of Euler's formula ).
Trivial example: the Fourier transform of is
This means that in the case of a harmonic function in (usually time space) we get a spike at in (usually the frequency space). Using exponential functions, we can build the Fourier transform of sines and cosines (sums of Dirac delta functions, some multiplied by ).
Inverse Fourier transform
Given an image function , we define the function with
as long as it exists for each , as the inverse Fourier transform of . In the case that is Fourier-transformable and differentiable in portions and its transform is , then the inverse transform of is
In practice, one does not compute directly the inverse transformation because the integrals typically become difficult, but instead one uses the properties of the Fourier transform described below to express in terms of known transformations.
Fourier transform: properties
(Exercise: show all of them using the definition of Fourier transform)
Linearity
For all , if and , then
Application: consider a real signal (for example, the sound of a typical musical instrument, where the base frequency is "louder" than the first harmonic ). Given that , one can see that . This means that for a given signal, the Fourier transform selects all the frequencies present in it (it gives out a complex spectrum).
Conjugation
Scaling
for .
Example: Note that because of the properties of the Dirac delta,
Displacement in time
Displacement in frequency
Derivative with respect to time
Derivative with respect to frequency
Convolution theorem („Faltungsprodukt")
We define a convolution of as
The Fourier transform of a convolution is .
Another way to state this is using the inverse Fourier transform:
Let's derive this, starting by multiplying the Fourier transforms of and :
Let . Then, we obtain
The quantity inside brackets is , and the outer integral is the Fourier transform, exactly as stated in our definition.
Duality
If is continuous in , .
Example: we saw above that . Then, it's not surprising that .
Symmetry
If the function is even or odd, then is also even or odd.
Inversion
Parseval's identity
Exercise: prove this relation (hint: substitute one of the by the inverse Fourier transform definition).
General Parseval's relation
Fourier transform in several dimensions
For dimensions, we have
The integral is performed in all space, and this doesn't necessarily mean from to in each coordinate. For example, an azimuthal angular integral covering all space means .
References
Lang & Pucker (2016) Mathematische Methode in der Physik (3rd edition). Springer.
Karpfinger, C. (2022) Höhere Mathematik in Rezepte (4th edition). Springer Spektrum.
Hassani, S. (1999) Mathematical Physics. Springer.
Korn, G.A.; Korn, T.M. (1968) Mathematical handbook for scientists and engineers. Dover publications.
Boas, M. (1983) Mathematical Methods in the Physical Sciences. Wiley.