Consistent Nonparametric Spectrum Estimation Via Cepstrum Thresholding

M.Venkatanarayana & Dr.T.Jayachandra Prasad
Signal Processing : An International Journal (SPIJ), Volume (4): Issue (5) 292
Consistent Nonparametric Spectrum Estimation Via
Cepstrum Thresholding
M. Venkatanarayana narayanamoram@yahoo.co.in
Associate Professor/ECE/
K.S.R.M.College of Engg.
Kadapa-516003, India
Dr. T.Jayachandra Prasad jp.talaril@gmail.com
Principal, Professor in ECE
RGMCET
Nandyal-518 501, India
Abstract
For stationary signals, there are number of power spectral density estimation
techniques. The main problem of power spectral density (PSD) estimation
methods is high variance. Consistent estimates may be obtained by suitable
processing of the empirical spectrum estimates (periodogram). This may be
done using window functions. These methods all require the choice of a certain
resolution parameters called bandwidth. Various techniques produce estimates
that have a good overall bias Vs variance tradeoff. In contrast, smooth
components of this spectral required a wide bandwidth in order to achieve a
significant noise reduction. In this paper, we explore the concept of cepstrum for
non parametric spectral estimation. The method developed here is based on
cepstrum thresholding for smoothed non parametric spectral estimation. The
algorithm for Consistent Minimum Variance Unbiased Spectral estimator is
developed and implemented, which produces good results for Broadband and
Narrowband signals.
Keywords: Cepstrum, Consistency, Cramer Rao Lower Bound, Unbiasedness.
1. INTRODUCTION
The main objective of spectrum estimation is the determination of the Power Spectral density
(PSD) of a random process. The estimated PSD provides information about the structure of the
random process, which can be used for modeling, prediction, or filtering of the deserved process.
Digital Signal Processing (DSP) Techniques have been widely used in estimation of power
spectrum. Many of the phenomena that occur in nature are best characterized statistically in
terms of averages [20].
Power spectrum estimation methods are classified as parametric and non-parametric. Former
one a model for the signal generation may be constructed with a number of parameters that can
be estimated from the observed data. From the model and the estimated parameters, we can
compute the power density spectrum implied by the model. On the other hand, do not assume
any specific parametric model of the PSD. They are based on the estimate of autocorrelation
sequence of random process from the observed data. The PSD estimation is based on the
assumption that the observed samples are wide sense stationary with zero mean. Traditionally

four techniques are used to estimate non parametric spectrum such as Periodogram, Bartlett
method (Averaging periodogram), Welch method (Averaging modified periodogram) and
Blackman-Tukey method (smoothing periodogram) [18] and [19].
2. CEPSTRUM ANALYSIS
The cepstrum of a signal is defined as the Inverse Fourier Transform of the logarithm of the
Periodogram. The cepstrum of })({ 1
0
−=
=
Nt
tty can be defined as [7],[8] and [13]
1,......0;)ln(
1 1
0
−== ∑
−
=
Nke
N
c pj
N
p
pk
kω
φ (1)
Consider a stationary, discrete-time, real valued signal })({ 1
0
−=
=
Nt
tty , the Periodogram estimate is
given by
( )
21
0
21ˆ
∑
−
=
−
=
N
t
ftj
p ety
N
π
φ (2)
A commonly used cepstrum estimate is obtained by replacing pφ with the periodogram pφˆ .
1,.....,0
;)ˆln(
1
ˆ
1
0
−=
= ∑
−
=
Nk
e
N
c pj
N
p
pk
kω
φ
(3)
to make unbiased estimate the cepstrum coefficients only at origin is modified, remaining are
unchanged.



==
+=
2/,......1ˆ
577126.0ˆ00
Nkcc
cc
kk
(4)
In this approach, we smooth






p
^
lnφ by thresholding the estimated cepstrum }{ kc , not by
direct averaging of the values of






p
^
lnφ . The following test can be used to infer whether kc is
likely to be equal or close to zero and, there fore, whether kc should be truncated to zero [9]-
[12].
( )




≤
=
elsec
Nd
cif
c
k
k
k
k
2/1
0~
µπ
(5)
The spectral estimate corresponding to { }kc~ is given by
1,........0;~exp
~ 1
0
−=





= ∑
−
=
−
Npec
N
k
kj
kp
pω
φ (6)
The proposed non parametric spectral estimate is obtained from pφ
~
by a simple scaling
1,.....0,
~
ˆ
ˆˆ −== Nppp φαφ (7)

where torscalingfacaisN
p
p
N
p
pp
α
φ
φφ
α ˆ;
~
~ˆ
ˆ 1
0
2
1
0
∑
∑
−
=
−
=
=
Statistics of log periodogram
The mean and variance of the k th component of the log periodogram of the signal,
2
log kY ,
assuming that the spectral component kY is Gaussian, are, respectively, given by [1]-[6],



−=−
=−−
=
12/,.........1)log(
2/,02log)log(
}{log
2
Kk
Kk
YE
k
k
Y
Y
k
γλ
γλ (8)
where 05772156649.0=γ is the Euler constant, and






−=
=
=
∑
∑
∞
=
∞
=
1
2
1
22
12/,....1
1
2/,0
1
)5.0(
!
)var(log
n
n n
k
Kk
n
Kk
n
n
Y (9)
where )1)........(2).(1.(.1)( −+++≅ naaaaa n . Furthermore,
;
2
1
)5.0(
!
1
2
2∑
∞
=
=
n n n
n π
;
6
1
1
2
2∑
∞
=
=
n n
π
Note from (8) that the expected value of the k th component of the log-periodogram equals the
logarithm of the expected value of the periodogram plus some constant. This surprising linear
property of the expected value operator is of course a result of the Gaussian model assumed
here. From (9) the variance of the k th log-periodogram component of the signal is given by the
constant.
Statistics of Cepstrum
The mean of the cepstral component of the signal is obtained from (8) and is given by [1], [2] and
[7]
n
K
k
kYy
K
kn
K
j
K
ncE ξ
π
λ
1
}
2
exp{)log(
1
)}({
1
0
−= ∑
−
=
(10)
where


 =
=
oddnif
evennornif
n
,0
0,2log2
ξ
the variance of the cepstral components is obtained from (9) and given by for 2/..,.........0 Kn =





<<−+
=−+
=
=
2
0),2(
21
2
,0),2(
22
))(),(cov())(var(
1021
1021
K
nifkk
K
k
K
K
nifkk
K
k
K
ncncnc yyy
(11)
and for mnKmn ≠= ,2/....,,.........1,0,





±±±=−−
=
otherwise
K
mnifkk
K
mcnc yy
,0
2
,.........4,2)2(
2
))(),(cov(
102
(12)
where
2
2
0
π
=k ;
6
2
1
π
=k
The covariance matrix of cepstral components of the signal, assuming the spectral components
of the signal are statistically independent complex Gaussian random variables. The covariance
matrix of cepstral components given by (11) and (12) is independent of the underlying power
spectral density which characterizes the signal under the Gaussian assumption. The covariance
of cepstral components under the Gaussian assumption is a fixed signal independent matrix that
approaches, for large K a diagonal matrix given by







<=<
==
=
otherwise
K
mnif
K
K
mnif
K
mcnc yy
,0
2
0,
6
1
2
,0,
3
1
))(),(cov(
2
2
π
π
(13)
Cepstrum algorithm
1. Let a stationary, discrete-time, real valued signal })({ 1
0
−=
=
Nt
tty
2. Compute the periodogram estimate of pφ using FFT.
)(ˆ ωφp =
N
1
| ∑
−
=
−
1
0
)(
N
t
tj
ety ω
|
2
3. First apply natural logarithm and take IFFT to compute the cepstrum estimate.
1,.....,0
;)ˆln(
1
ˆ
1
0
−=
= ∑
−
=
Nk
e
N
c pj
N
p
pk
kω
φ
4. Compute the threshold by choosing the appropriate value of µ depending on the type of
signal and determine the cepstral coefficients
( )




≤
=
elsec
Nd
cif
c
k
k
k
k
2/1
0~
µπ
5. Compute the spectral estimate corresponding to { }kc~ is given by
1,........0;~exp
~ 1
0
−=





= ∑
−
=
−
Npec
N
k
kj
kp
pω
φ
6. Obtain the proposed non parametric spectral estimate by a simple scaling
1,.....0,
~
ˆ
ˆˆ −== Nppp φαφ

Simulation Results
In this section, we present experimental results on the proposed algorithm for simulated data to
estimate the power spectrum. The performance of proposed method is verified for simulated data,
generated by applying Gaussian random input to a system, which is either broad band or narrow
band.The MA broad band signal is generated by using the difference equation [18]
1,....1,0
),4(2401.0)3(1736.0
)2(3508.0)1(3544.0)()4(4096.0
)3(8843.0)2(5632.1)1(3817.1)(
−=
−+−
+−+−+=−
+−−−+−−
Nt
tete
tetetety
tytytyty
(14)
where )(te is a normal white noise with mean zero and unit variance. The ARMA narrow band
signal is generated by using the difference equation
1,....1,0
),2(25.0)1(21.0)()4(8556.0
)3(19.0)2(61.1)1(2.0)(
−=
−+−−=−
+−−−+−−
Nt
tetetety
tytytyty
(15)
The number of samples in each realization is assumes as N=256.
After performing 1000 Monte Carlo Simulations, the comparison of the mean Power Spectrum,
Variance and Mean Square Error for the broad band signal and narrow band signals, obtained
using periodogram and cepstrum approach along with the true power spectrum are shown in
Figure 1 (a) , (b) and (c) and Figure 2 (a), (b) and (c) respectively.
0 0.5 1 1.5 2 2.5 3
5
10
15
20
ensemblepowerspectrum,db
frequency,w
True
Periodogram
Cesptrum
FIGURE 1: (a) PSD vs frequency for broadband signal

0 0.5 1 1.5 2 2.5 3
0
5
10
15
20
25
30
variance,db
frequency,w
Periodogram
Cesptrum
FIGURE 1: (b) Variance vs frequency for broadband signal
0 0.5 1 1.5 2 2.5 3
0
5
10
15
20
25
30
meansquareerror
frequency,w
Periodogram
Cesptrum
FIGURE 1: (c) Mean Square Error vs frequency for broadband signal

0 0.5 1 1.5 2 2.5 3
-10
0
10
20
30
40
50
frequency,w
True
Periodogram
Cesptrum
FIGURE 2: (a) PSD vs frequency for narrowband signal
0 0.5 1 1.5 2 2.5 3
0
5
10
15
20
25
30
35
40
45
50
variance,db
frequency,w
Periodogram
Cesptrum
FIGURE 2: (b) Variance vs frequency for narrowband signal

0 0.5 1 1.5 2 2.5 3
0.5
1
1.5
2
2.5
x 10
4
meansquareerror
frequency,w
Periodogram
Cesptrum
FIGURE 2: (c) Mean Square Error vs frequency for narrowband signal
From the above results we can say that
1. In the case of broad band signal the spectral estimates through cepstrum approach has
very smooth response compared to the periodogram approach. However it can be
observed that the mean square error is more in the case of periodogram and least with
cepstrum thresholding approach.
2. In the case of broad band signals, variance obtained through cepstrum thresholding
approach is very small as compared to the periodogram approach.
3. It is also observed that the mean square error estimated through cepstrum approach for
narrowband signals is less compared to broadband signals.
Comparison among the traditional methods and the cepstrum method
In order to evaluate the performance of the cepstrum technique, which is compared with the
traditional methods such basic Peridogram, Bartlett method, Welch method and Blackman and
Tukey [21] for simulated ARMA narrow band signal, which is generated by using equation (15).
TABLE 1: Comparison table for the parameters mean and variance (Record length N=128).
From the comparison table 1, for short record length, with respect to mean and variance, the
cepstrum technique produces better results in comparison with the traditional methods. For
longer record length, with reduced computational complexity, the cepstrum method produces the
The various PSD
techniques
Mean Variance
Cepstrum 0.0090 2.4023e-004
Periodogram 0.0092 4.8587e-004
Black-man and Tukey 0.0521 0.0047
Welch 0.0138 8.9491e-004
Bartlett 0.2474 0.0637

values of mean and variance as same as that of the Welch method, but these methods are better
than the remaining techniques. For 1000 Monte carlo simulations, the ensemble power spectrum
for various techniques is shown in figure 3.
0 0.5 1 1.5 2 2.5 3
-40
-35
-30
-25
-20
-15
-10
-5
0
frequency,w
Periodogram
Cesptrum
blackman
welch
barlett
FIGURE 3: an ensemble power spectrum of an ARMA narrowband signal by using the traditional methods
and the cepstrum method
Results for MST Radar data
The concept of cepstrum is applied to atmospheric data collected from the MST Radar on 10
th
August 2008 at Gadhanki, Tirupati, India. 150 sample functions, each having 256 samples are
used to know the performance of cepstrum in comparison with the standard periodogram. The
better results are obtained through the cepstrum than the periodogram. The comparison of the
mean Power Spectrum, Variance for Radar data, obtained using periodogram and cepstrum
approach are shown in Figure 4 (a) and (b) respectively. It is observed that the smooth power
spectra and less variance in cepstrum than that of the periodogram.

0 0.5 1 1.5 2 2.5 3
10
30
10
31
10
32
10
33
10
34
PSD,dB
Frequency
peridogram and cepstrum
periodogram
cepstrum
FIGURE 4: (a) Mean Power Spectra Vs Frequency for MST Radar data
0 0.5 1 1.5 2 2.5 3
10
60
10
62
10
64
10
66
10
68
PSD,dB
Frequency
variance of both peridogram and cepstrum
varper
varcep
FIGURE 4: (b) Variance Vs Frequency for MST Radar data

3. CONSLUSION & FUTURE WORK
The problem in traditional methods is that the variance becomes proportional to square of power
spectrum instead of converging into zero, thus the estimated spectrum is an inconsistent. In this
paper the new technique has been proposed, called cepstrum, which gives reduce variance while
evaluating the smoothed nonparametric power spectrum estimation. The expression for mean
and variance of the cepstrum has been presented. The total variance reduction is more through
broadband signals when compared to narrowband signals. All results are verified by using MAT
lab 7.0.1. The concept of Cepstrum can be also extended for higher order spectral estimations.
4. REFERENCES
1. Y.Ephraim and M.Rahim, “On second-order statistics and linear estimation of cepstral
Coefficiets”, IEEE Trans.Speech Audio Processing, vol.7, no.2, pp.162-176, 1999.
2 A.H.Gray Jr., “Log spectra of Gaussian signals”, Journal Acoustical Society of America,
Vol.55, No.5, May 1974.
3 Grace Wahba, “Automatic Smoothing of the Log Periodogram”, JASA, Vol.75, Issue 369,
Pages: 122-132.
4. Herbert T. Davis and Richard H. Jones, “Estimation of the Innovation Variance of a Stationary
Time Series”: Journal of the American Statistical Association, Vol. 63, No. 321 (Mar., 1968), pp.
141- 149.
5. Masanobu Taniguchi, “On selection of the order of the spectral density model for a stationary
Process”, Ann.Inst.Statist.Math.32 (1980), part-A, 401-419.
6. Yariv Ephraim and David Malah, “Speech Enhancement Using a Minimum Mean Square Error
Short-Time Spectral Amplitude Estimator”, IEEE trans., on ASSP, vol.32, No.6, December,
1984.
7. P.Stoica and N. Sandgren, “Smoothed nonparametric spectral estimation via cepstrum
Thresholding” IEEE Signal Processing Magazine, November, 2006, pp. 34-45.
8. D. G. Childers D. P. SLciAner and R. C. Kernemit, “The Cepstrum: A Guide to Processing”,
Proceedings of the IEEE, Vol. 65, no. 10, October 1977.
9. P. Moulin, “Wavelet thresholding techniques for power spectrum estimation” IEEE Trans.
Signal Processing, vol. 42, no. 11, pp. 3126–3136, 1994.
10. H.-Y. Gao, “Choice of thresholds for wavelet shrinkage estimate of the spectrum” J. Time
Series Anal., vol. 18, no. 3, pp. 231–251, 1997.
11. A.T. Walden, D.B. Percival, and E.J. McCoy, “Spectrum estimation by wavelet thresholding of
multitaper estimators,” IEEE Trans. Signal Processing, vol. 46, no. 12, pp. 3153–3165, 1998.
12. A.R. Ferreira da Silva, “Wavelet denoising with evolutionary algorithms” in Proc. Digital Sign.,
2005, vol. 15, pp. 382–399.
13. B.P. Bogert, M.J.R. Healy, and J.W. Tukey, “The quefrency analysis of time series for echoes:
Cepstrum, pseudo-autocovariance, cross-cepstrum and saphe cracking,” in Time Series
Analysis, M. Rosenblatt, Ed. Ch. 15, 1963, pp. 209–243.
14. E.J. Hannan and D.F. Nicholls, “The estimation of the prediction error variance,”J. Amer.

Statistic. Assoc., vol. 72, no. 360, pp. 834–840, 1977.
15. M. Taniguchi, “On estimation of parameters of Gaussian stationary processes,” J. Appl. Prob.,
vol. 16, pp. 575–591, 1979.
16. PETR SYSEL JIRI MISUREC, “Estimation of Power Spectral Density using Wavelet
Thresholding”, Proceedings of the 7th WSEAS International Conference on circuits,
systems, electronics, control and signal processing (CSECS'08).
17. Petre Stoica and Niclas Sandgren, “Cepstrum Thresholding Scheme for Nonparametric
Estimation of Smooth Spectra”, IMTC 2006 - Instrumentation and Measurement Technology
Conference Sorrento, Italy 24-27 April 2006.
18. P. Stoica and R.Moses, “Spectral Analysis of Signals”, Englewood Cliffs, NJ: Prentice Hall,
2005.
19. John G. Proakis and Dimitris G. Manolakis, “Digital Signal Processing: Principles and
Applications”, PHI publications, 2nd
edition, Oct 1987.
20. M.B.Priestley, “Spectral Analysis and Time series” Volume-1, Academic Press, 1981.
21. Alexander D.Poularikas and Zayed M.Ramadan, “Adaptive Filtering Primer with Matlab”,
CRC Press, 2006

Consistent Nonparametric Spectrum Estimation Via Cepstrum Thresholding

Recommended

More Related Content

What's hot (19)

Similar to Consistent Nonparametric Spectrum Estimation Via Cepstrum Thresholding (20)

Recently uploaded (20)

Consistent Nonparametric Spectrum Estimation Via Cepstrum Thresholding