Professional Documents
Culture Documents
2 2
2 2
References
Introduction to
Time-Frequency Analysis and Wavelet Transforms
Arun K. Tangirala
Lecture 2.2: Basic definitions and concepts - Part II
December 3, 2014
Lecture 2.2
December 3, 2014
December 3, 2014
References
Objectives
Lecture 2.2
References
Energy signal
Energy
The energy of a continuous-time signal x(t) and a discrete-time signal x[k] are, respectively, defined as
(Cohen, 1994),
Z 1
1
X
2
Exx =
|x(t)| dt ;
Exx =
|x[k]|2
(1)
1
A signal with finite energy, i.e., 0 < Exx < 1 is said to be an energy signala .
a The
Examples: exponentially decaying signals, all finite-length (duration) bounded amplitude signals
Lecture 2.2
December 3, 2014
References
Power signal
Power
The average power of a continuous-time signal x(t) and a discrete-time signal x[k] are, respectively,
defined as (Cohen, 1994)
Pxx
1
= lim
T !1 2T
k=N
X
1
|x[k]|2
N !1 2N + 1
T
T
|x(t)|2 dt ;
Pxx = lim
(2)
k= N
A signal with finite power, i.e., 0 < Pxx < 1 is said to be a power signal.
Examples: periodic signals, random signals
I All finite-duration (and amplitude) signals have Pxx = 0. In general, any energy signal is not a power
signal and vice versa. However, it is possible that a signal is neither an energy nor a power signal.
Arun K. Tangirala, IIT Madras
December 3, 2014
Lecture 2.2
References
Energy Density
Equation (1) gives rise to the idea of an energy density in time. Drawing analogies with probability density
function and densities in mechanics, the quantity
Sxx (t) = |x(t)|2
(3)
is termed as the energy density per unit time. It can also be thought of as an instantaneous power.
1
0.8
Energy density
Signal
0.5
0.6
0.4
0.5
0.2
1
0
50
100
150
200
250
300
0
0
50
Time
Lecture 2.2
100
150
200
250
300
Time
December 3, 2014
References
Power Density
Similarly, the power density in time can be defined as
xx (t) =
|x(t)|2
T
(4)
I For the discrete-time case, the energy and power density in time are not defined since the time domain
is not a continuum. The distribution functions exist nevertheless.
I On the other hand, we can think of energy and power densities of d.t. signals in a transform domain,
provided that the new domain is continuous and that the transform is energy / power preserving. This is the
basis for defining spectral densities of c.t. and d.t. signals in the Fourier (frequency) domain.
The energy / power densities in frequency domain share a strong connection with the time-domain
characteristics (properties) of the signal, specifically the covariance functions.
December 3, 2014
Lecture 2.2
References
Cross-covariance function
The cross-covariance function (CCVF) is a measure of the linear dependence between two timelagged (random or deterministic) signals.
I
It is based on the notion of covariance, a quantity that measures the linear dependence between
two zero-lagged deterministic signals (or two random variables).
A normalized version known as, cross-correlation function (CCF), is more suitable for analysis
since it is invariant to the choice of units (for signals).
Caution: It is a common practice in signal processing to use the alternative terms cross-correlation and
normalized cross-correlation, for CCVF and CCF, respectively.
Lecture 2.2
December 3, 2014
References
(5)
l]
k=0
I
I
xp yp [l]
xp xp [0] yp yp [0]
Observe that by setting xp = yp and l = 0 in (5), we obtain the average power of the periodic
signal.
December 3, 2014
Lecture 2.2
References
1
X
x[k]y[k
(6)
l]
k= 1
As before,
I
I
xx [0] yy [0]
Observe that by setting x = y and l = 0 in (6), we obtain the energy of the aperiodic signal.
Lecture 2.2
xy [l]
December 3, 2014
References
The CCVF measures the linear dependence between x[k] and time-shifted y[k] (by l samples). This
property is used in testing linear relationships between two signals.
The CCVF specializes to auto-covariance function (ACVF) for univariate signals, which is a widely
used tool for periodicity detection and echo cancellation.
MATLAB: xcov, xcorr
December 3, 2014
10
Lecture 2.2
References
Auto-covariance functions
The ACVFs of periodic and (finite-energy) aperiodic deterministic signals are, respectively,
Np 1
1 X
xp [k]xp [k
xp xp [l] =
Np
l];
xx [l]
1
X
x[k]x[k
(7)
l]
k= 1
k=0
Lecture 2.2
+ Np ] =
xp xp [l]
December 3, 2014
11
References
Examples
1
0.4
0.6
0.2
0.4
0.2
ACVF
Amplitude
xp [k] = cos(2f k)
1
=) xp xp [l] = cos(2f l)
2
0.6
0.8
0
0.2
0.4
0.4
0.6
0.6
0.8
1
0
10
15
20
Samples
25
30
35
40
=)
xx [l] =
0 , < 0
25
30
35
40
0.6
0.5
0.4
4
0.1
0
0
20
40
60
80
Samples
20
Lags
10
0.2
15
12
0.8
0.3
e2
10
0.9
0.7
k<0
ACVF
:0,
0.8
0
Amplitude
x[k] =
8
<ek ,
0
0.2
0.05.
100
0
0
20
40
60
80
100
Lags
12
Lecture 2.2
References
Concluding remarks
The properties of covariance functions have been listed for deterministic signals. However, they
also apply to signals corrupted with noise as well.
Definitions of covariance functions for stochastic signals involve ensemble averages (in place of
time averages), also known as expectations.
Although ACVF is theoretically suited for detecting periodicities embedded in noise, a significantly
better tool for periodicity detection is the spectral density (obtained by Fourier analysis).
Lecture 2.2
December 3, 2014
13
References
Bibliography I
Cohen, L. (1994). Time Frequency Analysis: Theory and Applications. Upper Saddle River, New Jersey, USA:
Prentice Hall.
Proakis, J. and D. Manolakis (2005). Digital Signal Processing - Principles, Algorithms and Applications. New
Jersey, USA: Prentice-Hall.
Tangirala, A. K. (2014). Principles of System Identification: Theory and Practice. CRC Press, Taylor & Francis
Group.
December 3, 2014
14