Spectral Estimation Notes
Spectral Estimation Notes
Spectral Estimation Notes
f X=
( x1 : t1 ) f X ( x1 : t1 + δ ) .
If we put δ = −t1 , we can conclude that the second order density function is
a function of time difference.
Wide sense stationary process:
A random process is said to be wide sense stationary if the mean is constant
and the autocorrelation of a process is a function of time difference. The
process is said to be an ergodic process if the time average is equal to the
ensemble average and time average auto-correlation is equal to the
ensemble autocorrelation.
1
minimum the N -point DFT. For example, the computation yields samples
of the periodogram
2
1 N −1
x(n)e− j 2π kn
k
=
Pxx ( )
N
∑
N n =0
= /N
k 0,1, 2,..., N − 1
At the frequencies f k = k / N .
0.9
FFT length= M =
∆f
N
Number of FFTs=
= 1.11N ∆f
M
2
Quality factor=
= Q 1.11N ∆f
N M N 0.9
Number of computations= log 2 M = log 2
M 2 2 ∆f
1.28
FFT length= M =
∆f
2N
Number of FFTs== 1.56 N ∆f
M
Quality factor=
= Q 1.39 N ∆f
2N M 1.28
Number of computations= log 2 M = N log 2
M 2 ∆f
2N
For windowing × M multiplications are needed.
M
1.28 5.12
Total computations = 2 N + N log 2 =N log 2
∆f ∆f
1.28
FFT length= 2M =
∆f
N N N
Number of FFTs= 2k + 1= 2 + 1= +1 ≅
2M M M
Quality factor=
= Q 2.34 N ∆f
N 1.28
Number of computations= ( M log 2 2M ) = N log 2
M ∆f
3
Limitations of Non-parametric Methods for Power Spectrum
Estimation:
These methods have the following limitations:
i) It needs long data sequences to obtain the necessary frequency
resolution.
ii) Spectral leakge effects due to windowing.
iii) The assumption of the autocorrelation estimate rxx ( m) to be zero
for m ≥ N . This assumption limits the frequency resolution and
quality of the power spectrum estimate.
iv) Assumption that the data are periodic with period N . These
assumptions may not be totally realistic.
Example:
1. Determine the frequency resolution of the Bartlett, Welch and
Blackman-Tukey methods of power spectrum estimation for quality
factor Q = 10. Assume that overlap in Welch’s method is 50%. Given
the length of the sample sequence is 1000.
Solution: Data given:
Quality factor Q = 10.
4
Overlap in Welch’s method=50%
Bartlett Method:
Quality factor==
QB 1.11N ∆f
QB 10
=⇒ ∆f frequency
= resolution= = 0.009 .
1.11N 1.11× 1000
Welch Method:
Quality factor==
QW 1.39 N ∆f
QW 10
=⇒ ∆f frequency
= resolution= = 0.0072 .
1.39 N 1.30 × 1000
Blackman-Tukey Method:
Quality factor==
QBT 2.34 N ∆f
QBT 10
=⇒ ∆f frequency
= resolution= = 0.0042 .
2.34 N 2.34 × 1000
2. A continuous-time signal xa (t ) is band limited to 5 kHz, i.e., xa (t ) has
a spectrum X a ( f ) that is zero for f > 5 kHz. Only 10 seconds of the
signal has been recorded and is available for processing. We would
like to estimate the power spectrum of xa (t ) using the available data
in a radix-2 FFT algorithm, and it is required that the estimate have
a resolution of at least 10 Hz. Suppose that we use Bartlett’s method
of periodogram averaging.
(a) If the data is sampled at the Nyquist rate, what is the minimum
section length that you may use to get the desired resolution?
(b) Using the minimum section length determined in part (a), with
10 seconds of data, how many sections are available for
averaging?
(c) How does your choice of sampling rate affect the resolution and
variance of your estimate? Are there any benefits to sampling
above the Nyquist rate?
Solution:
(a) If we sample at Nyquist rate, f s = 10 kHz, then a resolution of
∆f =10 Hz (in analog frequency) implies that we want a resolution
(in radian) of
5
∆f
∆ω = 2π = 2π × 10−3 .
fs
0.9
Now, for Bartlett’s method, ∆ω =2π .
L
0.9
So, we want to use a section length of L ≥ 2π
900 samples.
=
∆ω
(b) Sampling at 10 kHz, 10 seconds of data corresponds to
N= (10)(10 × 103 )= 105 samples.
Therefore, with a 1024-point DFT the number of sections we may
N
have in Bartlett’s method = is k = 98 .
1024
(c) If the sampling rate is increased then ∆ω decreases which, in turn,
requires a longer section length for a given resolution. However,
an increase in the sampling rate produces a corresponding
increase in the total number of samples within a T second interval.
Therefor, since the variance (normalized) is
L
V=
N
increasing the sampling rate has no effect. Thus, provided that the
sampling rate is not less than the Nyquist frequency, the
resolution and the variance do not depend on the sampling rate.