Professional Documents
Culture Documents
Signals
∗
James McNames
Biomedical Signal Processing Laboratory (http://bsp.pdx.edu)
February 7, 2002
Technical Report 02-1
Abstract
Over the years I have encountered many occasions where I need to know how the fast fourier
transform (FFT) of a sampled signal (discrete-time) precisely relates to the Fourier transform of
a continuous-time signal. This short technical article is intended to serve as a common reference
to answering this question so that I can stop deriving the answer each time I encounter it.
1 Derivation
Let the Fourier transform of a continuous-time signal x(t) be defined as follows:
∞
F{x(t)} ≡ x(t)e−jωt dt. (1)
−∞
where T is the sampling period. The sum is multiplied by T to ensure that the sampled signal,
p(t)x(t), has approximately the same area over any given interval,
t1 t1
x(t) · p(t) dt ≈ x(t) dt, (3)
t0 t0
for any t. An alternate justification for this impulse train could be derived by taking the limit as
the duty cycle approaches zero of a pulse train with fundamental period T . If the height of each
pulse is scaled so that the area under each pulse is equal to one, then the sampled signal, x(t)p(t),
will only retain the same power as the original signal if the height is increased by a factor of T 1
The sampled, discrete-time signal, x[n] is defined as
in [1].
1
The Fourier transform of xs (t) = x(t)p(t) is therefore
This proof used the following properties of the unit impulse (Dirac delta) function:
and
∞ ∞
x(t) · δ(t − t0 ) dt = x(t0 ) · δ(t − t0 ) dt, (15)
−∞ −∞
∞
= x(t0 ) δ(t − t0 ) dt, (16)
−∞
= x(t0 ). (17)
Now suppose that our signal is either finite duration or has been multiplied by a window function
to make it finite duration. Further, suppose that the window lasts exactly N samples. Then
Equation 13 reduces to
N
−1
Xs (ω) = T x[n]e−jωnT . (18)
n=0
2
1
Signal x(t)
0.5
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (seconds)
0.2
Real X(ω)
0.15
0.1
0.05
0
0 5 10 15 20 25 30 35 40 45 50
Frequency (rad/sec)
Imaginary X(ω)
−0.05
−0.1
0 5 10 15 20 25 30 35 40 45 50
Frequency (rad/sec)
or, conversely,
Xs (ω)|ω= k 2π = T X[k]. (21)
N T
2 Example
Figure 2 shows the signal
x(t) = e−5t u(t) (23)
which has a well known Fourier transform,
1
X(ω) = . (24)
a + jω
The signal and its Fourier transform are shown in blue. The sampled signal is shown in red and the
FFT approximation of the Fourier transform is also shown in red. For this example the sampling rate
was chosen to be 100 Hz and the number of samples was 26 = 64. Note that the frequency resolution
can be increased by zero-padding, but this does not affect the precision of the FFT estimate of the
continuous-time Fourier transform.
3
References
[1] Alan V. Oppenheim and Alan S. Willsky. Signals & Systems. Prentice Hall, second edition,
1997.