Purdue University: ECE438 - Digital Signal Processing with Applications 1

ECE438 - Laboratory 7:
Discrete-Time Random Processes (Week 2)
October 6, 2010

1 Bivariate Distributions
In this section, we will study the concept of a bivariate distribution. We will see that bivariate
distributions characterize how two random variables are related to each other. We will also
see that correlation and covariance are two simple measures of the dependencies between
random variables, which can be very useful for analyzing both random variables and random
processes.

1.1 Background on Bivariate Distributions

Sometimes we need to account for not just one random variable, but several. In this section,
we will examine the case of two random variables–the so called bivariate case–but the theory
is easily generalized to accommodate more than two.
The random variables X and Y have cumulative distribution functions (CDFs) FX (x)
and FY (y), also known as marginal CDFs. Since there may be an interaction between X
and Y , the marginal statistics may not fully describe their behavior. Therefore we define a
bivariate, or joint CDF as

FX,Y (x, y) = P (X ≤ x, Y ≤ y). (1)

If the joint CDF is sufficiently “smooth”, we can define a joint probability density func-
tion,
∂2
fX,Y (x, y) = FX,Y (x, y). (2)
∂x∂y
Conversely, the joint probability density function may be used to calculate the joint CDF:
Z y Z x
FX,Y (x, y) = fX,Y (s, t)ds dt. (3)
−∞ −∞

Questions or comments concerning this laboratory should be directed to Prof. Charles A. Bouman,
School of Electrical and Computer Engineering, Purdue University, West Lafayette IN 47907; (765) 494-
0340; bouman@ecn.purdue.edu
Purdue University: ECE438 - Digital Signal Processing with Applications 2

The random variables X and Y are said to be independent if and only if their joint CDF
(or PDF) is a separable function, which means

fX,Y (x, y) = fX (x)fY (y) . (4)

Informally, independence between random variables means that one random variable does
not tell you anything about the other. As a consequence of the definition, if X and Y are
independent, then the product of their expectations is the expectation of their product.

E[XY ] = E[X]E[Y ] (5)

While the joint distribution contains all the information about X and Y , it can be very
complex and is often difficult to calculate. In many applications, a simple measure of the
dependencies of X and Y can be very useful. Three such measures are the correlation,
covariance, and the correlation coefficient.

• Correlation Z ∞ Z ∞
E[XY ] = xyfX,Y (x, y)dx dy (6)
−∞ −∞

• Covariance
Z ∞ Z ∞
E[(X − µX )(Y − µY )] = (x − µX )(y − µY )fX,Y (x, y)dxdy (7)
−∞ −∞

• Correlation coefficient
E[(X − µX )(Y − µY )] E[XY ] − µX µY
ρXY = = (8)
σX σY σX σY

If the correlation coefficient is 0, then X and Y are said to be uncorrelated. Notice that
independence implies uncorrelatedness, however the converse is not true.

1.2 Samples of Two Random Variables

In the following experiment, we will examine the relationship between the scatter plots for
pairs of random samples (Xi , Zi ) and their correlation coefficient. We will see that the
correlation coefficient determines the shape of the scatter plot.
Let X and Y be independent Gaussian random variables, each with mean 0 and variance
1. We will consider the correlation between X and Z, where Z is equal to the following:

1. Z = Y

2. Z = (X + Y )/2

3. Z = (4 ∗ X + Y )/5

4. Z = (99 ∗ X + Y )/100
Purdue University: ECE438 - Digital Signal Processing with Applications 3

Notice that since Z is a linear combination of two Gaussian random variables, Z will also
be Gaussian.
Use Matlab to generate 1000 i.i.d. samples of X, denoted as X1 , X2 , . . . , X1000 . Next,
generate 1000 i.i.d. samples of Y , denoted as Y1 , Y2 , . . . , Y1000 . For each of the four choices
of Z, perform the following tasks:

1. Use equation (8) to analytically calculate the correlation coefficient ρXZ between X and
Z. Show all of your work. Remember that independence between X and Y implies that
E[XY ] = E[X]E[Y ]. Also remember that X and Y are zero-mean and unit variance.

2. Create samples of Z using your generated samples of X and Y .

3. Generate a scatter plot of the ordered pair of samples (Xi , Zi ). Do this by plotting
points (X1 , Z1 ), (X2 , Z2 ), . . . , (X1000 , Z1000 ). In order to plot points without connecting
them with lines, use the plot command with the ’.’ format.
plot(X,Z,’.’)
Use the command subplot(2,2,n) (n=1,2,3,4) to plot the four cases for Z in the same
figure. Be sure to label each plot using the title command.

4. Empirically compute an estimate of the correlation coefficient using your samples Xi

and Zi and the following formula.
PN
i=1 (Xi − µ̂X )(Zi − µ̂Z )
ρ̂XZ = qP
N PN
i=1 (Xi − µ̂X )2 i=1 (Zi − µ̂Z )2

INLAB REPORT:

1. Hand in your derivations of the correlation coefficient ρXZ along with your numerical
estimates of the correlation coefficient ρ̂XZ .

2. Why are ρXZ and ρ̂XZ not exactly equal?

3. Hand in your scatter plots of (Xi , Zi ) for the four cases. Note the theoretical correlation
coefficient ρXZ on each plot.

4. Explain how the scatter plots are related to ρXZ .

2 Autocorrelation for Filtered Random Processes

In this section, we will generate discrete-time random processes and then analyze their be-
havior using the correlation measure introduced in the previous section.
Purdue University: ECE438 - Digital Signal Processing with Applications 4

2.1 Background
A discrete-time random process Xn is simply a sequence of random variables. So for each n,
Xn is a random variable.
The autocorrelation is an important function for characterizing the behavior of random
processes. If X is a wide-sense stationary (WSS) random process, the autocorrelation is
defined by
rXX (m) = E[Xn Xn+m ] m = . . . , −1, 0, 1, . . . . (9)
Note that for a WSS random process, the autocorrelation does not vary with n. Also, since
E[Xn Xn+m ] = E[Xn+m Xn ], the autocorrelation is an even function of the “lag” value m.
Intuitively, the autocorrelation determines how strong a relation there is between samples
separated by a lag value of m. For example, if X is a sequence of independent identically
2
distributed (i.i.d.) random variables each with zero mean and variance σX , then the auto-
correlation is given by

rXX (m) = E[Xn Xn+m ]

(
E[Xn ]E[Xn+m ] if m 6= 0
=
E[Xn2 ] if m = 0
2
= σX δ(m) . (10)

We use the term white or white noise to describe this type of random process. More precisely,
a random process is called white if its values Xn and Xn+m are uncorrelated for every m 6= 0.

x(n) y(n)
H(ejω )

Figure 1: A new LTI system diagram

If we run a white random process Xn through an LTI filter as in figure 1, the output
random variables Yn may become correlated. In fact, it can be shown that the output
autocorrelation rY Y (m) is related to the input autocorrelation rXX (m) through the filter’s
impulse response h(m).
rY Y (m) = h(m) ∗ h(−m) ∗ rXX (m) (11)

2.2 Experiment
Consider a white Gaussian random process Xn with mean 0 and variance 1 as input to the
following filter.
y(n) = x(n) − x(n − 1) + x(n − 2) (12)
Calculate the theoretical autocorrelation of Yn using (10) and (11). Show all of your work.
Purdue University: ECE438 - Digital Signal Processing with Applications 5

Generate 1000 independent samples of a Gaussian random variable X with mean 0

and variance 1. Filter the samples using (12). We will denote the filtered signal Yi ,
i = 1, 2, . . . , 1000.
Draw 4 scatter plots using the form subplot(2,2,n), (n = 1, 2, 3, 4). The first scatter plot
should consist of points, (Yi , Yi+1 ), (i = 1, 2, . . . , 900). Notice that this correlates samples
that are separated by a lag of “1”. The other 3 scatter plots should consist of the points
(Yi , Yi+2 ), (Yi , Yi+3 ), (Yi , Yi+4 ), (i = 1, 2, . . . , 900), respectively. What can you deduce about
the random process from these scatter plots?
For real applications, the theoretical autocorrelation may be unknown. Therefore, rY Y (m)
may be estimated by the sample autocorrelation, rY′ Y (m) defined by
N −|m|−1
1
rY′ Y (m)
X
= Y (n)Y (n + |m|) − (N − 1) ≤ m ≤ N − 1 (13)
N − |m| n=0

where N is the number of samples of Y .

Use Matlab to calculate the sample autocorrelation of Yn using (13). Plot both the
theoretical autocorrelation rY Y (m), and the sample autocorrelation rY′ Y (m) versus m for
−20 ≤ m ≤ 20. Use subplot to place them in the same figure. Does equation (13) produce
a reasonable approximation of the true autocorrelation?

INLAB REPORT:
For the filter in equation (12),

1. Show your derivation of the theoretical output autocorrelation, rY Y (m).

2. Hand in the four scatter plots. Label each plot with the corresponding theoretical
correlation, using rY Y (m). What can you conclude about the output random process
from these plots?

3. Hand in your plots of rY Y (m) and rY′ Y (m) versus m. Does equation (13) produce
a reasonable approximation of the true autocorrelation? For what value of m does
rY Y (m) reach its maximum? For what value of m does rY′ Y (m) reach its maximum?

4. Hand in your Matlab code.

3 Correlation of Two Random Processes

3.1 Background
The cross-correlation is a function used to describe the correlation between two separate
random processes. If X and Y are jointly WSS random processes, the cross-correlation is
defined by
cXY (m) = E[Xn Yn+m ] m = . . . , −1, 0, 1, . . . . (14)
Purdue University: ECE438 - Digital Signal Processing with Applications 6

Similar to the definition of the sample autocorrelation introduced in the previous section, we
can define the sample cross-correlation for a pair of data sets. The sample cross-correlation
between two finite random sequences Xn and Yn is defined as
N −m−1
1
c′XY (m)
X
= X(n)Y (n + m) 0 ≤ m ≤ N − 1 (15)
N − m n=0

N −1
1
c′XY (m) =
X
X(n)Y (n + m) 1 − N ≤ m < 0 (16)
N − |m| n=|m|
where N is the number of samples in each sequence. Notice that the cross-correlation is not
an even function of m. Hence a two-sided definition is required.
Cross-correlation of signals is often used in applications of sonar and radar, for example to
estimate the distance to a target. In a basic radar set-up, a zero-mean signal X(n) is trans-
mitted, which then reflects off a target after traveling for D/2 seconds. The reflected signal
is received, amplified, and then digitized to form Y (n). If we summarize the attenuation and
amplification of the received signal by the constant α, then

Y (n) = αX(n − D) + W (n) (17)

where W (n) is additive noise from the environment and receiver electronics.
In order to compute the distance to the target, we must estimate the delay D. We can do
this using the cross-correlation. The cross-correlation cXY can be calculated by substituting
(17) into (14).

cXY (m) = E[X(n)Y (n + m)]

= E[X(n)(αX(n − D + m) + W (n + m))]
= αE[X(n)X(n − D + m)] + E[X(n)]E[W (n + m)]
= αE[X(n)X(n − D + m)]

Here we have used the assumptions that X(n) and W (n+m) are uncorrelated and zero-mean.
By applying the definition of autocorrelation, we see that

cXY (m) = αrXX (m − D) (18)

Because rXX (m − D) reaches its maximum when m = D, we can find the delay D by
searching for a peak in the cross correlation cXY (m). Usually the transmitted signal X(n)
is designed so that rXX (m) has a large peak at m = 0.

3.2 Experiment

Down load radar.mat

Purdue University: ECE438 - Digital Signal Processing with Applications 7

Using (15) and (16), write a Matlab function C=CorR(X,Y,m) to compute the sample
cross-correlation between two discrete-time random processes, X and Y , for a single lag
value m.
To test your function, generate two length 1000 sequences of zero-mean Gaussian random
variables, denoted as Xn and Zn . Then compute the new sequence Yn = Xn + Zn . Use CorR
to calculate the sample cross-correlation between X and Y for lags −10 ≤ m ≤ 10. Plot
your cross-correlation function.

INLAB REPORT:

1. Submit your plot for the cross-correlation between X and Y . Label the m-axis with
the corresponding lag values.

2. Which value of m produces the largest cross-correlation? Why?

3. Is the cross-correlation function an even function of m? Why or why not?

4. Hand in the code for your CorR function.

Next we will do an experiment to illustrate how cross-correlation can be used to measure

time delay in radar applications. Down load the MAT file radar.mat and load it using the
command load radar. The vectors trans and received contain two signals corresponding to
the transmitted and received signals for a radar system. First compute the autocorrelation
of the signal trans for the lags −100 ≤ m ≤ 100. (Hint: Use your CorR function.)
Next, compute the sample cross-correlation between the signal trans and received for the
range of lag values −100 ≤ m ≤ 100, using your CorR function. Determine the delay D.

INLAB REPORT:

1. Plot the transmitted signal and the received signal on a single figure using subplot.
Can you estimate the delay D by a visual inspection of the received signal?
′
2. Plot the sample autocorrelation of the transmitted signal, rXX (m) vs. m for −100 ≤
m ≤ 100.

3. Plot the sample cross-correlation of the transmitted signal and the received signal,
c′XY (m) vs. m for −100 ≤ m ≤ 100.

4. Determine the delay D from the sample correlation. How did you determine this?

References
[1] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed., McGraw-
Hill, New York, 1991.