AGC

DSP Power Spectral Estimation

The purpose of these methods is to

obtain an approximate estimation of the
j
power spectral density yy (e ) of a
given real random process { yn }

Professor A G Constantinides 1
AGC
DSP Autocorrelation
The autocorrelation sequence is
yy [m] E{ y[m n] y[n]}
The pivot of estimation is the Wiener-
Khintchine formula (which is also known
as the Einstein or the Rayleigh formula)

j jm
yy (e ) yy [m]e
m

Professor A G Constantinides 2
AGC
DSP Classification
The crux of PSD estimation is the determination of
the autocorrelation sequence from a given process.
Methods that rely on the direct use of the given
finite duration signal to compute the autocorrelation
to the maximum allowable length (beyond which it
is assumed zero), are called Non-parametric
methods
Methods that rely on a model for the signal
generation are called Modern or Parametric
methods.
Personally I prefer the names Direct and Indirect
Methods

Professor A G Constantinides 3
AGC
DSP Classification & Choice

The choice between the two options is

made on a balance between simple and
fast computations but inaccurate PSD
estimates Vs computationally involved
procedures but enhanced PSD estimates

Professor A G Constantinides 4
AGC
DSP Direct Methods & Limitations
Apart from the adverse effects of noise,
there are two limitations in practice
Only one manifestation { y [ n]} , known

as a realisation in stochastic processes, is
available
Only a finite number of terms, say 2 N 1 ,
is available

Professor A G Constantinides 5
AGC
DSP Assumptions
Assume { y[n]} to be
Ergodic so that statistical expectations can
be replaced by summation averages
Stationary so that infinite averages can be
estimated from finite averages
Both of these averages are to be derived
from
{ y [n]}

Professor A G Constantinides 6
AGC
DSP Windowing
Thus an approximation is necessary. In
effect we have a new signal {x[n]} given
by
x[n] w[n] y [n]

where w[n] is a window of finite

duration selecting a segment of signal
from { y [n]}.

Professor A G Constantinides 7
AGC
DSP The Periodogram
The Periodogram is defined as
N 1 2
1
I N (e j ) x[ n ]e jn
N n 0
j1 N 1 jn
N 1
jr
I N (e ) x[n]e x[r ]e
N n 0 r 0
Clearly evaluations at
2
k k
N
are efficiently computable via the FFT.
Professor A G Constantinides 8
AGC
DSP Limited autocorrelations
Let
N 1 m
xx [m] x[n m ]x[n]
n 0

which we shall call the autocorrelation

sequence of this shorter signal.
These are the parameters to be used
for the PSD estimation.

Professor A G Constantinides 9
AGC
DSP PSD Estimator
It can be shown that

j 1 N 1 jm
I N (e ) xx [ m ]e
N m ( N 1)
The above and the limited autocorrelation
expression, are similar expressions to the
PSD. However, the PSD estimates, as we
shall see, can be bad.
Measures of goodness are the bias and
the variance of the estimates?
Professor A G Constantinides 10
AGC
DSP The Bias

The Bias pertains to the question:

Does the estimate tend to the correct
value as the number of terms taken
tends to infinity?
If yes, then it is unbiased, else it is biased.

Professor A G Constantinides 11
AGC
DSP Analysis on Bias
For the unspecified window case
considered thus far, the expected value
of the autocorrelation sequence of the
truncated signal is
N 1 m
E{ xx [m]} E{ x[n m ]x[n]}
n 0
N 1 m
E{w[n m ] y[n m ]w[n] y[n]}
n 0

Professor A G Constantinides 12
AGC
DSP Analysis on Bias
N 1 m
or
E{ xx [m]} ww [m] yy [m]
n 0

Thus
j 1 N 1 jm
E{I N (e )} E { xx [ m ]}e
N m ( N 1)
1 j j 2
yy (e ) * W (e )
2N

Professor A G Constantinides 13
AGC
DSP Analysis on Bias
The asterisk denotes convolution.
The bias is then given as the difference
between the expected mean and the
true mean PSDs at some frequency.
j k j k
B E{I N (e ) yy (e )
1
yy (e ) * W (e ) yy (e jk )
j j 2
2N
Professor A G Constantinides 14
AGC
DSP Example
For example take a rectangular window
then ,
2
j 2 sin( N / 2)
W (e )
sin( / 2)
which, when convolved with the true
PSD, gives the mean periodogram, ie a
smoothed version of the true PSD.
Professor A G Constantinides 15
AGC
DSP Example

Note that the main lobe of the window has a

width of 2 / N
and hence as N

j j
we have lim E{I N (e )} yy (e )
N

at every point of continuity of the PSD.

Professor A G Constantinides 16
AGC
DSP Asymptotically unbiased

j
Thus NI ( e ) is an asymptotically
unbiased estimator of the true PSD.
The result can be generalised as
follows.

Professor A G Constantinides 17
AGC
DSP Windows & Estimators
For the window to yield an unbiased
estimator it must satisfy the following:
1) Normalisation condition

N 1
w [ n] N
2
n 0

2) The main lobe width must decrease

as 1/N

Professor A G Constantinides 18
AGC
DSP The Variance
The Variance refers to the question on
the goodness of the estimate:
Does its variance of the estimate
decrease with N? ie does the
expression below tend to zero as N
tends to infinity?
j j j
var{I N (e )} E{( I N (e )) } ( E{I N (e )})
2 2

Professor A G Constantinides 19
AGC
DSP Analysis on Variance
If the process is Gaussian then (after
very long and tedious algebra) it can be
shown that
var{I N (e j )} ( E{I N (e j )}) 2 A
where
2
1 j j ( ) j ( )
A
yy ( e )W ( e )W *
( e )d
2N
Professor A G Constantinides 20
AGC
DSP Analysis

Hence it is evident that as the length of

data tends to infinity the first term
remains unaffected, and thus the
periodogram is an inconsistent
estimator of the PSD.

Professor A G Constantinides 21
AGC
DSP Example
For example for the rectangular window
taken earlier we have
j j
var{I N (e )} ( E{I N (e )}) C
2

where

sin ( N m )
2
N 1
C yy [m]
m ( N 1) N sin

Professor A G Constantinides 22
AGC
DSP Decaying Correlations
If yn has yy 0 for m m0
then for N m0 we can write above

j j sin N
2
var{I N (e )} yy (e ) 1
2

N sin
From which it is apparent that
j
j
yy (e ) 0,
2

var{I N (e )} j
2 yy (e ) elsewhere
2

Professor A G Constantinides 23
AGC
DSP Variance is large

Thus even for very large windows the

variance of the estimate is as large
as the quantity to be estimated!

Professor A G Constantinides 24
AGC
DSP Smoothed Periodograms
Periodograms are therefore inadequate for
precise estimation of a PSD.
To reduce variance while keeping estimation
simplicity and efficiency, several modifications
can be implemented
a) Averaging over a set of periodograms of
(nearly) independent segments
b) Windowing applied to segments
c) Overlapping the windowed segments for
additional averaging
Professor A G Constantinides 25
AGC
DSP Welch-Bartlett Procedure
Typical is the Welch-Bartlett procedure as
follows.
Let yn be an ergodic process from which we
are given M data points for the signal y[n] .
1) Divide the given signal into K M / N
blocks each of length N .
2) Estimate the PSD of each block
3) Take the average of these estimates

Professor A G Constantinides 26
AGC
DSP Welch-Bartlett Procedure
Step 2 can take different forms for
different authors.
For the Welch-Bartlett case the
periodogram is suggested as
N 1 2
j 1 jn
r
IN (e ) xr [m]e
N n 0

Professor A G Constantinides 27
AGC
DSP Welch-Bartlett Procedure
where the segment xr [n] is a
windowed portion of yn
xr [n] w[n] y[n r ( N N0 )]

And N0 is the overlap.

(Strictly the Bartlett case has a
rectangular window and no overlap).

Professor A G Constantinides 28
AGC
DSP Comments
FFT-based Spectral estimation is limited by
a) the correlation assumed to be zero beyond the
measurement length and
b) the resolution attributes of the DFT.
Thus if two frequencies are separated by
then a data record of length
N 2 /
is required.
(Uncertainty Principle)

Professor A G Constantinides 29
AGC
DSP Narrowband Signals
The spectrum to be estimated is some cases
may contain narrow peaks (high Q
resonances) as in speech formants or
passive sonar.
The limit on resolution imposed by window
length is problematic in that it causes bias.
The derived variance formulae are not
accurate

Professor A G Constantinides 30