2.2.4. Discrete Fourier Transform and FFT http://s-mat-pcs.oulu.fi/~ssa/ESignals/sig2_2_4.

htm

Previous section

2.2.4. Discrete Fourier Transform and FFT

Often we have only (equally spaced) samples f(tk), tk = a + k(b-a)/N, k = 0, 1, ..., N of the signal f(t).
Then f may be considered to be restricted to the measuring interval [a,b] so that outside of the interval f(t)
0. Hence

F(v) = f(t) e-j2 vtdt

= f(t) e-j2 vtdt

f(tk ) e-j2 vtk (tk+1-tk)

= [(b-a)/N] f(tk ) e-j2 vtk

= (v), v R.

If a = 0, b = 1, then tk = k/N and we usually denote f(tk) = f(k) and (v) is evaluated at points n = 0, 1, ...,
N-1.

Then (v) gets the form

1
(n) = f(k) e-j2 kn/N, n = 0, 1, ..., N-1.
N

(n) is called an N-point discrete Fourier transform of f. (even in the general case (a,b))

Obviously | (n)| = | (N-n)| .

The inverse discrete Fourier transform (IDTF)

(k) e j2 kn/N = f(n)

gives the original values f(n), n = 0, 1, ..., N - 1, when (k) are the values of the discrete Fourier
transform.

Frequency Resolution

Assume that the signal is sampled (at least) by the Nyquist frequency, the number of samples is N and that
the sampling time is T0 seconds ( (a,b)=(0,T) ). Then the sampling interval is T = T0 /N, sampling
frequency f s = 1/T and the highest frequency of the signal is (at most)

1 N
f max = = .
2T 2T0

The discrete Fourier transform (DFT) gives the values of the amplitude spectrum at the frequencies 1/T0

1 of 3 12/14/2009 2:23 AM
2.2.4. Discrete Fourier Transform and FFT http://s-mat-pcs.oulu.fi/~ssa/ESignals/sig2_2_4.htm

,2/T0 , ..., N / 2T0 - 1/T0 but also at N / 2T0 , N / 2T0 + 1/T0 , ..., N/T0 which, by the symmetry, can be
obtained from the the first N values.

The frequency resolution of the amplitude spectrum, obtained by DFT, is

1 1 fs
f= = = .
T0 NT N

In practise DFT often is evaluated at M points, M > N, which means addition of frequency points ( f = 1
/ MT) i.e. interpolation or equivalently in time domain adding zeros after observations (zero padding)
(increase of sampling time).
Usually, the spectrum programs give the spectrum at points n = 0, 1, ..., N-1 instead of the true analog
frequencies. The real analog frequencies can be found if the sampling interval T = T0 /N (or the sampling
frequency 1/T = N/T0 ) is known. Sometimes the spectrum is given on [ -N / 2T0 , N / 2T0 ] or on [-½,½].

Example 2.2.3: Finding the analog frequencies

Fast Fourier Transform, FFT

Denote W = e-j2 /N and choose N = 2L. Then

1
(n) = f(k) Wkn, n = 0, 1, ... , N-1,
N
1
= f(k) Wkn + f(k + N/2) W (k + N/2)n
N
1
= f(k) + f(k + N/2) W nN/2 Wkn
N

Since

1, when n = 2l
W nN/2 = e-j n = cos n = (-1)n =
-1, when n = 2l + 1

is

1
(2l) = f(k) + f(k + N/2) W k2l = ½ (l)
N

and

1
(2l+1) = f(k) - f(k + N/2) W k W k2l = ½ (l)
N

where

1
(l) = f(k) - f(k + N/2) e-j2 /(N/2) kl
N/2

2 of 3 12/14/2009 2:23 AM
2.2.4. Discrete Fourier Transform and FFT http://s-mat-pcs.oulu.fi/~ssa/ESignals/sig2_2_4.htm

is the N/2-point DFT of the signal g(k) = f(k) + f(k + N/2) and

1
(l) = f(k) - f(k + N/2) e-j2 /(N/2) kl
N/2

is the N/2-point DFT of the signal h(k) = Wk [f(k) - f(k + N/2)]. Since N/4 is an integer, ja can further
be divided into two N/4-point DFT etc. until in the last stage N 1-point transforms are calculated (N = 2L).
The number of arithmetic operations LN = N log2 N.

Problems: P18, P19, P20, P21, P24

3 of 3 12/14/2009 2:23 AM