Professional Documents
Culture Documents
If is very long (e.g. an hour of music...) it does not make sense to compute
the
Fourier transform—the signal is nonstationary.
Typically we break up the signal into shorter segments and compute the Fourier
transform of each segment (time-frequency analysis).
otherwise
How does this relate to the Fourier transform of the original signal?
where
Rectangular window
N=19
0.8
0.6
0.4
W(ω
)
0.2
−0.2
−0.4
−2pi −pi 0 pi 2pi
ω
The window gives distortions: smearing (resulting in loss of resolution) and side-
lobe interference (resulting in confusion)
4 8a non-parametric spectrum est. October 1, 2015
Spectral analysis
Example:
10 cos(ω0 n)
N=19 w[n]
5
R e ( X (ω))
^
−
5−p −pi/ 0 pi/2 pi
i 2 ω
, with
corresponding with
Example
cos(ω0 n)
0
x[n]
−1
0 5 10 15 20
n
10 N=19
L=19
|X(ω)|
0
−pi −pi/2 0 pi/2 pi
ω
0.5 cos(ω 0 n)
x [n] 0
^
−0.5
−1
0 5 10 15 20 25 30 35 40 45 50
n
10 N=19
L=50
|X(ω)|
0
−pi −pi/2 0 pi/2 pi
ω
1 cos(ω 1 n) + cos(ω 2 n)
x [n]
0
^
−1
−2
0 10 20 30 40 50 60
n
20 N=20 omega1=0.2 ⋅ 2π
L=500 omega2=0.22 ⋅ 2π
15 1/N=0.05
|X(ω)|
10
0
−pi −pi/2 0 pi/2 pi
ω
2
cos(ω1 n) + cos(ω 2 n)
1
x^[n]
0
−1
−2
0 10 20 30 40 50 60
n
20 N=40 Rectangular windowed
omega1=0.2 ⋅ 2π
L=500 omega2=0.22 ⋅ 2π
15 1/N=0.025
|X (ω)|
10
^
0
−pi −pi/2 0 pi/2 pi
ω
The resolution is further limited by the height of the side lobes (confusion).
Other windows can be designed to obtain lower side lobes, at the expense of a
wider main lobe. E.g., Bartlett, Hann, Hamming, Blackman, Kaiser, etc.
Bartlett window
0.5
0
0 2 4 6 8 10 12 14 16 18
n
1 N=19
Bartlett window
|W(ω)|
0
−pi −pi/2 0 pi/2 pi
ω
Bartlett window
2
1 cos(ω1 n) + cos(ω 2 n)
x^[n]
0
−1
−2
0 20 40 60 80 100 120
n
8 N=40 Bartlett windowed
omega1=0.2 ⋅ 2π
L=500 omega2=0.22 ⋅ 2π
6 1/N=0.025
|X (ω)|
4
^
0
−pi −pi/2 0 pi/2 pi
ω
Hamming window
Hamming window
0.5
0
0 2 4 6 8 10 12 14 16 18
n
1 N=19
Hamming window
|W(ω)|
0
−pi −pi/2 0 pi/2 pi
ω
Hamming window
1 cos(ω1 n) + cos(ω 2 n)
x^[n]
0
−1
−2
0 20 40 60 80 100 120
n
8 N=40 Hamming windowed
omega1=0.2 ⋅ 2π
L=500 omega2=0.22 ⋅ 2π
6 1/N=0.025
|X (ω)|
4
^
0
−pi −pi/2 0 pi/2 pi
ω
Example: communication signal (binary phase shift keying, with random bits )
bpsk signaal
2
x[n] −1
−2
0 10 20 30 40 50 60 70 80 90 100
n
40
N=100
30
|X(ω)|
20
10
0
−pi −pi/2 0 pi/2 pi
ω
Applying the DFT directly on the signal gives a chaotic picture. Do the
peaks correspond to certain frequencies?
x[n]
−1
−2
0 100 200 300 400 500 600 700 800 900 1000
n
200
N=1000
150
|X(ω)|
100
50
0
−pi −pi/2 0 pi/2 pi
ω
A larger gives points in the frequency domain, but the spectrum does not
converge. We need to average.
1 1 1 1 1
0 0 0 0 0
−1 −1 −1 −1 −1
−2 −2 −2 −2 −2
0 50 100 100 150 200 200 250 300 300 350 400 400 450 500
n n n n n
20 40 30 30 40
15 30 30
20 20
10 20 20
10 10
5 10 10
0 0 0 0 0
−pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi
ω ω ω ω ω
Averaged periodogram
1
N=10000 bpsk spectrum
L=100 single pulse (dirichlet)
0.8 Ts=6
0.4
0.2
0
−pi −pi/2 0 pi/2 pi
Definition
determine
where
asymptotically unbiased:
and for .
so that
This limits the resolution: ability to resolve closely spaced sinusoids. The
resolution
is defined by the half-power width of the main lobe,
Res
Var
bpsk signaal
2
0
x[n]
−1
−2
0 100 200 300 400 500 600 700 800 900 1000
n
200
N=1000
150
100
|X(ω)|
50
0
−pi −pi/2 0 pi/2 pi
ω
where
Bias
it follows that
Variance
Var
The window gives a way to trade-off the spectral resolution (width of main lobe)
with
the confusion (sidelobe amplitude).
Bias
Variance
Var Var
This goes to as .
Bartlett’s method
2 2 2 2 2
1 1 1 1 1
0 0 0 0 0
−1 −1 −1 −1 −1
−2 −2 −2 −2 −2
0 50 100 100 150 200 200 250 300 300 350 400 400 450 500
n n n n n
20 40 30 30 40
15 30 30
20 20
10 20 20
10 10
5 10 10
0 0 0 0 0
−pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi −pi −pi/2 0 pi/2 pi
ω ω ω ω ω
Bias
Precisely as before,
Res
Variance
Var Var
Bias
Variance
One shows that, for 50% overlap (some correlation among the sequences )
Var
This is a factor worse than for Bartlett’s estimate, for the same . But, twice
as many segments can be averaged ( doubles), so in fact the variance is a
factor
better:
Var