Lecture 4 Slides DFT Sampling Theorem
Lecture 4 Slides DFT Sampling Theorem
1
Fourier Analysis
a0
x(t ) = + ( an cos n0t +bn sin n0t )
2 n =1
Fourier Series
1 T2
cn = T x(t )e jn0t dt ( cn is complex generally.)
T 2
x(t ) = ce
n =
n
jn0 t
,(n = 0, 1, 2,...)
X ( ) = x(t )e jt dt (Fourier transform)
1
x(t ) =
2
X ( )e jt d (Inverse Fourier transform)
2
Discrete Fourier Series
2 2 90 1
j ( k + mN ) n j kn 120 60
ek + mN (n) = e N
=e N
= ek (n) 0.5
150 30
2 n
N 1
1 e j 2
e
j
N
= =0 180 0 Re
j 2 / N
n=0 1 e
210 330
Discrete Fourier series is finite because 240 300
only N of the harmonics are independent. 270
1 N 1 j 2 kn / N
ck = xn e ; k = 0,1,..., N 1 Example with N = 8 .
N n =0
N 1
xn = c e
k =0
k
j 2 kn / N
3
Fourier Analysis Summary
The basic premise of Fourier analysis is that any signal can be expressed as a linear
superposition, that is, a sum or integral of sinusoidal signals. The presence of an infinite
sum or integral prevents exact numerical computation of the corresponding transform.
4
5
2. The Discrete Fourier Transform (DFT)
N 1
X k = xn e j 2 kn / N
n =0
The Inverse DFT
1 N 1
xn = X k e j 2 kn / N ; n = 0,1,..., N 1
N k =0
6
The Discrete Fourier Transform (DFT)
(W ) = ( e )
N N
j 2 k / N
k
N = e j 2 k = 1
7
The Discrete Fourier Transform (DFT)
There are two important properties that make the DFT so eminently useful in signal
processing.
First, the N-point DFT provides a unique representation of the N-samples of a
finite duration sequence.
Second, the DFT provides samples of the DTFT of the sequence at a set of equally
spaced frequencies. This sampling process results in the inherent periodicity of the DFT.
Understanding the underlying periodicity of DFT is absolutely critical for the correct
application of DFT and meaningful interpretation of the results obtained.
We note that sampling in one domain is equivalent to periodization in the other domain.
Periodic replication may cause frequency-domain or time-domain aliasing.
8
9
DFT input sequence x(n) x(2)
x(3) x(1)
n n
N
x(n)
0 1 2 3 4 5 6 7 8
x(4) x(0)
NT
x(5) x(7)
x(6)
DFT output sequence X(k) X(2)
X(3) X(1)
k k
N
X(k)
0 1 2 3 4 5 6 7 8
X(4) X(0)
0=2 /NT
N0 X(5) X(7)
s
X(6)
N, k, n: N: points of input; k: the sequence number of spectrum X(k), n: the sequence
number of input x(n).
For N=8, s sampling frequency, 0 fundamental frequency X(7)= s 0
10
Examples:
Determine the N-point DFTs of the following sequences defined over 0 n < N.
(a) x[n] = 4 n, N = 8.
(b) x[n] = 4 sin(0.2n), N = 10.
(c) x[n] = 5(0.8)n, N = 16.
11
Example: The DFT of a Rectangular Pulse
x[n] is of length 5
We can consider x[n] of any length greater than 5
Lets pick N=5
Calculate the DFS of the periodic form of x[n]
j ( 2 k /5) n
4
X [k ] = e
n =0
1 e j 2 k
=
1 e ( )
j 2 k /5
5 k = 0, 5, 10,...
=
0 else
2
k
5
12
If we consider x[n] of length 10
We get a different set of DFT coefficients
Still samples of the DTFT but in different places
13
Amplitude and Phase Spectra of DFT
Phase spectrum = ( X ) = [ X 0 X 1 X N 1 ]
0.2
xk
-0.2
0 50 100 150 200
k
14
Relative Power Spectrum (2)
xk 0.2
-0.2
0 50 100 150 200
k
1 0
2
|Xm|
dB
0.5 -50
0 -100
0 50 100 0 50 100
m m
15
Linear Phase Shift of DFT
If a signal vector, x, is delayed (shifted to the right) k samples, then the phase of DFT{x}
is decreased by 2 mk / N rad.
Proof: Let x = [ x0 , x1 ,..., xN 1 ] . Let y be the delayed version of x so that
[ yk , yk +1 ,..., yN + k 1 ] = [ x0 , x1 ,..., xN 1 ] . The DFT of y is
N + k 1 2 mn N 1 2 m ( k + i ) 2 mk
j j j
Ym =
n=k
yn e N
= xi e
i =0
N
= X me N
;
m = 0,1,..., N 1.
(To get the second sum, we let i = n k .)
Thus, the phase angle of X m is decreased by 2 mk / N rad.
This is called linear phase shift because the phase shift is a linear function of m.
16
Linear Phase Shift Example
X phase (rad)
1
xn
0.5 -10
0
-20
0 20 40 0 10 20
Sample (n) Index (m)
0
-2mk/N
Y phase (rad)
1
yn = xn-k
0.5 -10
0
-20
0 20 40 0 10 20
Sample (n) Index (m)
17
DTF Properties
18
3. The Fast Fourier Transform (FFT)
W6=W14=W22=...
90
1
120 60
0.8
W5=W13=W21=... 0.6
W7=W15=W23=...
150 30
0.4
0.2
W4=W12=W20=...180 0 W0=W8=W16=...
210 330
W3=W11=W19=... W1=W9=W17=...
240 300
270
W2=W10=W18=...
19
The Fast Fourier Transform (FFT)
If N = power of 2,
N
# complex products = log 2 N instead of N 2 .
2
10
10
flops
5
10
0
10
0 500 1000 1500 2000 2500
FFT size (N)
20
4. Fourier analysis of signals using the DFT
The application of DFT requires three steps: (a) sample the continuous-time signal,
(b) select a finite number of samples for analysis, and (c) compute the spectrum at a
finite number of frequencies.
21
The effects of rectangular windowing (truncation) on the spectrum of a sinusoidal signal. In this case,
windowing can be interpreted as modulation of a sinusoidal carrier by the window function.
22
Peak merging (loss of spectral resolution) when two spectral lines are closer than the
width of the mainlobe of the window.
23
2). Effects of time-windowing on signals with continuous spectra
Thus, the Fourier transform of the windowed signal is obtained by convolving the Fourier
transform of the original signal with the Fourier transform of the window.
24
The effects of
windowing on the
spectrum of an ideal
bandpass signal using a
sum of equally spaced
sinusoidal components:
25
Good windows and the uncertainty principle
Smearing The predominant effect of the main lobe is to smear or spread the original
spectrum.
The result is loss of resolution. An ideal spectral line in the original spectrum
will have a width of about 2/T0 after windowing. Two equal amplitude sinusoids with
frequencies less than 2/T0 apart will blend with each other and may appear as a single
sinusoid.
Leakage The major effect of the side lobes is to transfer power from frequency bands
that contain large amounts of signal power into bands that contain little or no power. This
transfer of power, which is called leakage, may create false peaks (that is, peaks at
wrong frequencies), nonexisting peaks, or change the amplitude of existing peaks.
A good window should have a narrow main lobe (to minimize spectral spreading) and
low side lobes (to minimize spectral leakage). Unfortunately, as we show below, it is
impossible to satisfy both of these requirements simultaneously. This is a consequence
of the uncertainty principle of Fourier transforms.
26
27
5. Sampling theorem
Three different continuous-time signals with the same set of sample values, that is, x[n] =
xc1(nT) = xc2(nT) = xc3(nT).
If T were decreased to 1, the sample vectors would differ.
In general, how small must the time step be in order to convey all the information in x(t ) ,
that is, in order to be able to recover x(t ) from its sample vector, x?
28
The Sampling Theorem
The proof of the sampling theorem lies in the following relationship. X is the Fourier
transform before sampling.
N 1
1 2 m
DFT { x} = xn e jnT
= X j
n =0 T m = T
The terms in the sum are disjoint (do not overlap) provided X ( j ) = 0 at and above
half the sampling rate, that is, for / T .
If this condition holds, then (with m = 0 ) X ( j ) = T * DFT( x) , and so
x(t ) = FT 1 { X ( j} is recoverable from the samples.
29
Terminology in sampling operation
The highest frequency FH, in Hz, present in a bandlimited signal xc(t) is called the Nyquist
frequency. The minimum sampling frequency required to avoid overlapping bands is 2FH,
which is called the Nyquist rate. The actual highest frequency that the sampled signal x[n]
contains is Fs/2, in Hz, and is termed as the folding frequency.
30
6. Aliasing
Frequency-domain
interpretation of uniform
sampling.
31
Lecture 4: Summary
The Discrete Fourier Transform (DFT) is a finite orthogonal transform which provides
a unique representation of N consecutive samples x[n], 0 n N 1 of a sequence
through a set of N DFT coefficients X[k].
The DFT does not provide any information about the unavailable samples of the
sequence, that is, the samples not used in the computation. The interpretation or physical
meaning of the DFT coefficients depends upon the assumptions we make about the
unavailable samples of the sequence.
The DFT is widely used in practical applications to determine the frequency content
of continuous-time signals (spectral analysis). The basic steps are: (a) sampling
the continuous-time signal, (b) multiplication with a finite-length window (Hann or
Hamming) to reduce leakage, (c) computing the DFT of the windowed segment.
The value of the DFT stems from its relation to the DTFT, its relation to convolution
operations, and the existence of very efficient algorithms for its computation.
These algorithms are collectively known as Fast Fourier Transform (FFT) algorithms.
The FFT is not a new transform; it is simply an efficient algorithm for computing
the DFT.
32