Definition of Non-uniform Discrete Fourier Transform

(NDFT)
Now we want to generalize the definition and the computation of the Fourier transform
from the regular sampling to the irregular sampling domain. In the general case, the
definition of the Nonuniform Discrete Fourier Transform (NDFT) is the same as the one
given by Equation 3, taking into consideration that the samples can the taken at irregular
intervals both in time ( ) and/or in frequency ( ).

However, in practice, we want to take into consideration a more restricted case, which is
the case where the samples are irregularly taken in the time domain t but regularly taken
in the frequency domain. That is to say that the samples P(m) of the irregular Fourier
transform are taken at multiples of a quantity , which is a fixed quantity in the Fourier

domain. The fixed quantity in the regular case corresponds to . The extension from
regular to irregular sampling, therefore, depends on the duration of the signal p(t) and not
on the fact that the samples are taken at regular or irregular intervals.

The definition of the nonuniform discrete Fourier transform (NDFT) is as follows:

It is common practice to set where T is the range of extension for the samples
. In that case the formulation of the NDFT is very similar to the one of the DFT except
of the presence of the spatial coordinates instead of the index n. In this case, the NDFT
is defined as:

From a computational point of view, two differences have to be noticed between DFT and

NDFT. The first difference is that samples in frequency are taken at intervals in the

irregular case instead of in the regular case (T being the duration of the signal p(t), with

, and N is the number of samples of the signal p(t)). The second difference is
that, instead of the integer index n in the regular case, in the irregular case the irregular
sampling coordinate appears in the exponent.
Signal Reconstruction by using the Fourier transform
The irregular sampling problem is concerned with the problem of dealing with signals
and images which may be represented by samples on an irregular grid. One of the most
frequent problems to solve is the one of reconstruction of the signal from its samples. The
NDFT can be used directly for reconstruction. By computing the Fourier coefficients at
all required discrete regular frequencies and then Fourier transforming back with the use
of the inverse DFT, a kind of interpolation of the irregularly sampled signal can be
obtained.

In order to explain this application, let us first examine the problem of reconstruction of a
signal from its regular samples with the use of the inverse Fourier transform.

Signal reconstruction from regularly sampled data

Let us consider the following one-dimensional signal . It assumes values

, the plot of which is shown in

Figure 1 a. The corresponding Fourier transform is function P(f) visualized in Figure 1 b.
The total duration of the signal is T=17.

Figure 1: Original signal p(t) (a) and its Fourier transform P(f) (b).

Let us consider a regular subsampling of the original signal p(t), . It assumes values

. The regular sampling pattern used is , the values of

which are .

The sampled values are shown as magenta circles in Figure 2 a. The total number of
regular samples is N=9. The DFT of the regularly sampled signal is plotted in Figure 2 b
in magenta superimposed to the Fourier transform of the original signal in blue. It is
possible to notice that the Fourier transform of the regularly sampled signal has the same
shape of the original Fourier transform, the peaks are slightly lower (the signal loses
some energy) and the higher frequencies are missing, in fact only half of the Fourier
coefficients are present.

Figure 2: Regularly sampled signal superimposed on the original signal p(t) (a)

and its Discrete Fourier transform (DFT) (in magenta) superimposed on the
Fourier transform of the original signal (in blue)(b).

If we Fourier transform the back, which is to say we use the inverse DFT, we
obtain the original signal p(t) if the regular sampling frequency is above the Nyquist limit
of twice the highest frequency present in the original signal. A smoothed (low-pass)
approximation of the original signal is obtained, if the Nyquist condition is not met. In
the case of N=9 the Nyquist condition is respected and the signal is reconstructed
perfectly as shown in Figure 3. In the case the number of samples is lowered and
therefore the sampling interval increases, the reconstruction is imperfect as shown from
the examples in Figure 4 a for N=8 and Figure 4 b for N=6.
Figure 3: Reconstructed signal using the inverse Fourier Transform (IDFT) using N=9
regularly spaced samples.
Figure 4: Reconstruction of the original signal (in blue) using N=8 regularly spaced
samples (a) and N=6 regularly spaced samples (b). The reconstruction is not perfect
because the number of samples is reduced.

Signal reconstruction from irregularly sampled data

Now we want to reconstruct the signal p(t) from a collection of samples taken irregularly
using the inverse Fourier transform. Let us take into consideration an irregular sampling

of the original signal p(t). The number of samples

taken into consideration is the same as in the case of the regular sampling, N=9.
The irregular sampling sequence is , so these
are the irregular sampling coordinates that have to be used in the computation of the
NDFT. The irregular samples are shown as green circles in Figure 5 a and the
corresponding NFT is shown as the green curve in Figure 5 b.

Figure 5: Irregularly sampled signal superimposed on the original signal p(t) (a)

and its Nonuniform Discrete Fourier transform (NDFT) (in green) superimposed
to the Fourier transform of the original signal, P(f) (in blue) and the DFT of the regularly

sampled signal, (in magenta) (b).

Now it is possible to calculate the regular inverse DFT of the Fourier transform ,

which provides a reconstruction of the original signal p(t). This reconstruction is

shown in green in Figure 6, superimposed to the original signal in blue.
Figure 6: Reconstructed signal using the inverse Fourier Transform (IDFT) of the NDFT

The reconstructed signal has the same shape of the original signal and provides a good
approximation, however it is not an interpolation of the original signal, as the values of
the original signal at the sampling points are not recovered.

Let us make a comparison now between the computation of the DFT and the computation
of the NDFT. For the computation of the DFT it is necessary to calculate the matrix:

where is the transpose of vector which

contains the N=9 indices of the Fourier coefficients. The coefficients obtained by the

multiplication are shown in Figure 7 a.

Using a matrix formulation the calculation of the DFT can be expressed as:

In order to compute the NDFT it is necessary to calculate the matrix:

where is the transpose of vector which contains the N=9 coordinates of the irregular
sampling sequence. Once again, we remind that T=17 is the total length of the signal p(t).

The coefficients obtained by the multiplication are shown in Figure 7 b.

Using a matrix formulation the calculation of the NDFT can be expressed as:

Inspection of Figure 7 helps visualize the difference in the two matrices used for the
regular and irregular calculation of the Fourier transform. Each line in the diagrams on
Figure 7 corresponds to a matrix row. The regular circles in Figure 7 a correspond to the
column values as column index assumes the N=9 values from -4 to 4. In Figure 7 b the
N=9 cross values are placed irregularly and their position is dependent on the sampling
sequence represented by vector .

By using these coefficients as exponents of the exponential Fourier kernel, the complex
values shown in Figure 8 are obtained. The values of the Fourier exponential functions in
the regular matrix are shown as magenta circles on the complex plane, while the
corresponding functions for the irregular matrix are shown as green crosses in the
complex plane. Since matrices and are square matrices of dimension N, one
would expect to see values. Since the values shown are much fewer, this means
that some values are assumed by elements of the matrices more than once.
Figure 7: Regular matrix coefficients obtained from the vector multiplication

(a) and irregular matrix coefficients obtained from the vector

multiplication (b) as row and colum values of a matrix.

Figure 8: Values assumed by the elements of the regular matrix in magenta and of the
irregular matrix in green, in the complex plane.