9.

6 Correlation of Discrete-Time Signals

A signal operation similar to signal convolution, but with completely different

physical meaning, is signal correlation. The signal correlation operation can be

performed either with one signal (autocorrelation) or between two different signals

(crosscorrelation). Physically, signal autocorrelation indicates how the signal energy

(power) is distributed within the signal, and as such is used to measure the signal

power. Typical applications of signal autocorrelation are in radar, sonar, satellite,

and wireless communications systems. Devices that measure signal power using

signal correlation are known as signal correlators. There are also many applications

of signal crosscorrelation in signal processing systems, especially when the signal

is corrupted by another undesirable signal (noise) so that the signal estimation

(detection) from a noisy signal has to be performed. Signal crosscorrelation can be

also considered as a measure of similarity of two signals.

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–90
 Definition 9.3: Discrete-Time Autocorrelation and Crosscorrelation

Given two discrete-time real signals (sequences) and . The autocorre-

lation and croosscorrelation functions are respectively defined by

 


   
 


    

where the parameter is any integer, .

Using the definition for the total discrete-time signal energy, we see that for

, the autocorrelation function represents the total signal energy, that is

  

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–91
 Naturally, the autocorrelation and crosscorrelation sums are convergent under

assumptions that the signals and have finite total energy. It can be

observed that  
  . In addition, it is easy to show that the

autocorrelation function is an even function, that is


 


Hence, the autocorrelation function is symmetric with respect to the vertical axis.

Also, it can shown that


 

(see Problem, 9.29).

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–92
Problem 9.29

Using the change of variables as in the definition formula for the

auto-correlation function, we obtain the required result as follows

 

  



 

Using the change of variables as in the definition formula for the

cross-correlation function, we have

 


    

 
 

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–93
 Note that the above defined autocorrelation and crosscorrelation functions are

applicable to the energy signals (they have finite total energy). The autocorrelation

and crosscorrelation functions can be also defined for power signals (they have

infinite energy, but finite power). In such a case, we have to redefine these sums

along the definition formula of the discrete-time signal average power.

It is interesting to observe that the autocorrelation and cross correlation functions

can be evaluated using the discrete-time convolution as follows


 


It is left to students as an exercise to establish these results, Problem 9.30.

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–94
Problem 9.30

Using the change of variables as in the definition formula for the

auto-correlation function, we have


! "# $"#

$" #

Using the change of variables as in the definition formula for the

cross-correlation function, we obtain


%
! " # $" #

$"#

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–95
 As a measure of similarity of two signals, we can use the correlation coefficient

defined by
&
'
&
'
&
& ''

Note the correlation coefficient satisfies &
' . This can be established

by observing the fact that &' is the inner product of the vectors

that contain respectively the samples of and . Similarly, the relations

( (
&
& and ' ' represent the square

of the corresponding vector Euclidean norm so that the correlation coefficient

geometrically represents the angle between the vectors and , that is

&
'
&' ( (
&
& ''

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–96
 When the correlation coefficient is close to one then the signals are similar (they

almost overlap) and when the correlation coefficient is close to zero the signals are

very different (orthogonal as the matter of fact). When the correlation coefficient

is close to minus one the signals are asimilar (opposite direction, but almost the

same sample values). Note that the correlation coefficient can be also defined in

terms of parameter , that is

)
*
)
*
)
) **

in which case the same lower and upper bounds hold due to the fact that
)
* )
) * * , see Problem 9.31.

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–97
Problem 9.31

Using the fact that the sum a nonnegative numbers is nonnegative, that is, staring

with the following sum

+
0
/ 0
, - .+

valid for any real , we have

+ + +
0 0 0
/ 0 / 0
, -.+ , -.+ , - .+
0
132 1
4 13251
4

Taking 1 21 4 1 4 , we obtain
0
1 21 4 1 21 4
162 0 1
4 13251
4
1 4 1 4
0
1 2 1 4 1 21 4

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–98
 The obtained result indicates that

78 7
9 738:79 738 79

which establishes the required result.

The crosscorrelation operation is used for detection (estimation) of signals from

measured signals that contain the original signal corrupted by an additive noise, that

is , where is the original signal and is noise. The signal

that has the highest correlation with the signal is considered as the

best estimate of the signal and denoted by .

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–99
 DTFT of Autocorrelation and Crosscorrelation Functions

The DTFT of the auto- and crosscorrelation functions can be found similarly

to the DTFT of the convolution function. Define the following DTFT pairs
; ; and < < . Then, the auto- and crosscorrelation

functions of these two signals satisfy

; ;? ; < ;
=3>5=3> =3>:=@ <?

The proof of the property follows the convolution property proof. The quantity
; <
is called the energy spectral density of the signal ; . Hence, the

discrete-time signal energy spectral density is the DTFT of the signal autocorrelation

function.

The slides contain the copyrighted material from Linear Dynamic Systems and Signals, Prentice Hall, 2003. Prepared by Professor Zoran Gajic 9–100