Module 4

Signal Representation and Baseband Processing

Version 2 ECE IIT, Kharagpur

Lesson 18
Response of Linear System to Random Processes

Version 2 ECE IIT, Kharagpur

After reading this lesson, you will learn about

Modeling of thermal noise and power spectral density; Time domain analysis of a linear filter for random input; Representation of narrow-band Gaussian noise; Low-pass equivalent components of narrow-band noise; Band-pass Gaussian noise and its spectral density;

A noise waveform is a sample function of a random process. Thermal noise is expected to manifest in a communication receiver for an infinite time and hence theoretically noise may have infinite energy. Thermal noise is typically modeled as a power signal. Usually, some statistical properties of thermal noise, such as its mean, variance, auto correlation function and power spectrum are of interest. Thermal noise is further modeled as a wide-sense-stationary (WSS) stochastic process. That is, if n(t) is a sample function of noise, a) the sample mean of n(t1), i.e. n(t) at t = t1, is independent of the choice of sampling instant t1 and b) the correlation of two random samples, n(t1) and n(t2) depends only on the interval / delay (t2-t1), i.e., E[n(t1 ) n(t2 )] = Rn (t2 t1 ) = Rn ( ) The auto-correlation function (ACF), Rx() of a WSS process, x(t) is defined as: ACF = Rx ( ) = E[ x(t ) x(t + )] . Rx() indicates the extent to which two random variables separated in time by vary with each other. Note that,

Rx (0) = E[ x(t ) x(t )] = x 2 , the mean square of x (t).

Power Spectral Density (psd)

a. b. Specifies distribution of power of the random process over frequency f. If Sx(f) is the two-sided psd of x(t), the power in a small frequency band f at f1 is [Sx(f1). f]; psd Sx(f) of thermal noise is a real, positive even function of frequency. The power in a band f1 to f2 is:
f1

f2

S x ( f )df +

f2 f1

S x ( f )df

= 2 S x ( f )df ;
f1

f2

in Volt2/Hz

For a deterministic waveform, the psd and ACF form a Fourier transform pair. The concept is extended to random processes and we may write for thermal noise process,

S x ( f ) = F [ Rx ( ) ] =
and

Rx ( )e

4.18.1

Rx ( ) = F

[ S x ( f )] = S x ( )e j df
Version 2 ECE IIT, Kharagpur

Now, as noted in Lesson #17, the psd for white noise is constant:

Sn ( f ) =

N0 2

4.18.2

Hence, the ACF for such noise process is,

N N Rn ( ) = F 1 0 = 0 . ( ) 2 2

4.18.3

As we know, the signal, carrying information, occupies a specific frequency band and it is sufficient to consider the effect of noise, which manifests within this frequency band. So, it is useful to study the features of band-limited noise. For a base band additive white Gaussian noise (AWGN) channel of bandwidth W Hz,

Sn ( f ) =

N0 , f <W 2 = 0 , Elsewhere

4.18.4

Simple calculation now shows that, the auto-correlation function for this base band noise is: Rn ( ) = WN 0 sin c(2W ) 4.18.5 For pass-band thermal noise of bandwidth B around a centre frequency fc, the results can be extended:

Sn ( f ) =

N0 B , f fc < 2 2 = 0 , Otherwise

4.18.6 4.18.7

The ACF now is given by, Rn ( ) = BN 0 (sin cB ).cos 2 f c

In many situations, it is necessary to analyze the characteristics of a noise process at the output of a linear system, which transforms some excitation given at its input. This is important because, the system being linear in nature, obeys the principle of superposition and if we excite the system with a noise process and analyze the response noise process, we can use this knowledge for multiple situations. For example, a specific case of interest may be to analyze the output of a linear filter when a noisy received signal is fed to it. For simplicity, we discuss about response of linear systems which are time-invariant. Though such analysis is more elegant when carried out in the frequency domain, we start with a time-domain analysis to provide some insight.

Time-domain analysis for random input to a linear filter

Let us consider a linear lowpass filter whose impulse response is h(t) and let us excite the filter with white Gaussian noise. The input being a random process, it is not so important to get only an expression for the filter output y(t). It is statistically more significant to obtain expressions for the mean, variance, ACF and other parameters of the output signal. Now, in general, if x(t) indicates the input to a linear system, the mean of the output y(t)

Version 2 ECE IIT, Kharagpur

is, y = x h(t )dt , where x is the mean of the input process. The mean square value of
0

the output is: y =

Rx (2 1)h(1)h(2 )d 2d 1
00

When the input is white noise, we know

Rn ( ) =

N0 ( ) , 2

and y = 0

4.18.8

So, the mean square of the output noise process is:

y =
2

N0 (2 1 )h(1 )h(2 )d 2 d 1 2 00

N = 0 h 2 ( ) d 2 0

4.18.9

Ex 4.18.1: Let us consider a single-stage passive R-C lowpass filter whose impulse response is well known: 1 h(t ) = e t u (t ) , where = . The 3 dB cutoff frequency of the filter is fcutoff = R.C 1 . It is straight forward to see that the average of noise at the output of this low2RC pass filter is zero: Further, y 2 =

y = x e t d = n = 0

N0 N0 N 2e2 d = = 0 = N 0 f cutoff 4 RC 2 0 4 2

We observe that y 2 , the noise power at the output of the filter, is proportional to the filter BW.

Auto-correlation function (ACF)

In general, the autocorrelation of a random process at the output of a linear two-port network is:

R y ( ) = Rx (2 1 )h(1 ).h(2 )d 2 d 1
00

4.18.10

Specifically, for white noise,

R y ( ) =
00

N0 (2 1 )h(1 ).h(2 )d 2 d 1 2

Version 2 ECE IIT, Kharagpur

N = 0 h(1 ).h(1 + )d 1 2 0
Considering the RC lowpass filter of Ex #4.18.1, we see,

4.18.11

N R y ( ) = 0 e e ( + ) d 2 0
=

N0
4

e , 0
R y ( ) =

As, R() is an even function , R() = R(- ) and hence,

N0
4

4.18.12

This is an exponentially decaying function of with a peak value of

at = 0. 4 As Ry() and the power spectrum Sy() are Fourier transform pair, we see that the power spectrum of the output noise is:

No

S y ( ) =

N0 2 2 2 2 +

4.18.13

Representation of Narrow-band Gaussian Noise

Representation and analysis of narrow pass band noise is of fundamental importance in developing insight into various carrier-modulated digital modulation schemes which are discussed in Module #5. The following discussion is specifically relevant for narrowband digital transmission schemes. Let, x(t) denote a zero-mean Gaussian noise process band-limited to B/2 around centre frequency fc. There are several ways of analyzing such narrowband noise process and we choose an easy-to-visualize approach, which somewhat approximate. To start with, we consider a sample of a noise process over a finite time interval and apply Fourier series expansion while stretching the time interval to . If x(t) is observed over an interval

T T t , we may write 2 2
where, w0 =

x(t ) = ( xcn cos nw0t + xsn sin w0t )

n=1 T /2

2 T

and xcn

2 = x(t ) cos nw0tdt , n = 1,2, T T/ 2

T /2

2 x(t )sin nw0tdt , n = 1,2,.. and xsn = T T/ 2

4.18.14 Version 2 ECE IIT, Kharagpur

It can be shown that xcn and xsn are Gaussian random variables. The centre frequency fc can now be brought in by the following substitution:

nw0 = (nw0 wc ) + wc

n =1

x(t ) = {xcn cos[(nw0 wc )t + wct ] + xsn sin[(nw0 wc )t + wct ]}

xc (t ) cos wct xs (t )sin wct ;
4.18.15 4.18.16 4.18.17

where, xc (t ) = and xs (t ) =

xcn cos(nw0 wc )t + xsn sin(nw0 wc )t

n=1

xcn cos(nw0 wc )t + xsn sin(nw0 wc )t

n=1

Another elegant and equivalent expression for x(t) is:

x(t ) = xc (t ) cos wct xs (t )sin wct = r (t ) cos[ wct + (t )]

4.18.18

where, and

r (t ) = xc 2 (t ) + xs 2 (t ) ;

x (t ) (t ) = tan 1 s ; 0 (t ) < 2 xc (t ) It is easy to recognize that, xc (t ) = r (t ) cos (t ) and xs (t ) = r (t )sin (t )

Low pass equivalent components of narrow band noise

Let, xct and xst represent samples of xc(t) and xs(t). These are Gaussian distributed random variables with zero mean as the original noise process has zero mean. E [ xct ] = E [ xst ] = 0 4.18.19 Now, an expression for variance of xct, by definition, looks like the following: [ xcn cos(nw0 wc )t + xsn sin(nw0 wc )t ] 2 E 4.18.20 E x = ct [ x cos(mw w )t + x sin(mw w )t ] 0 c sm 0 c n =1 m =1 cm However, after some manipulation, the above expression can be put in the following form: 2 E xct

[ xcn .xcm cos(nw0 wc )t.cos(mw0 wc )t + xcn .xsm cos(nw0 wc )t. = sin(mw0 wc )t + ... + xsn .xcm sin(nw0 wc )t.co s(mw0 wc )t n =1 m =1 + xsn .xsm sin(nw0 wc )t.sin(mw0 wc )t ] 4.18.21

Version 2 ECE IIT, Kharagpur

Here,
2 T / 2 2 T /2 ( ).cos . ( ).cos xcn xcm = E x t nw tdt x t mw tdt 0 0 T T / 2 T T / 2

= =

4 4

T /2 T /2

T 2 T / 2 T / 2
T /2 T /2

x(t1 ) x(t2 ) cos nw0t1.cos mw0t2 dt1dt2

4.18.22

T 2 T / 2 T / 2

Rx (t2 t1 )cos nw0t1.cos mw0t2 dt1dt2

Rx(t2-t1) in the above expression is the auto-correlation of the noise process x(t). Now,

t putting t2 t1 = u and 1 = v , we get, T

xcn xcm

T ( 1 +v ) 2 4 = + cos ( ).cos ( . ) nw Tv R u mw u v T du dv x 0 0 T 1/ 2 T ( 1 +v ) 2
1/ 2
1 T ( +v ) 2 4 2 mu cos 2 nv cos 2 mv Rx (u ).cos du = T 1/ 2 T 1 T ( + v ) 2 1/ 2

2 mu du dv sin 2 mv Rx (u ).sin T 1 T ( + v ) 2 4.18.23

1 T ( +v ) 2

Now for T and putting f m = are:

1 T ( +v ) 2

m , where fm tends to 0 for all m, the inner integrands T

lim T
and

Rx (u ).cos 2 f mudu = S x ( f m ) , say

lim T

1 T ( +v ) 2 1 T ( +v ) 2

Rx (u ).sin 2 f mudu = 0

1 T ( +v ) 2

Let us choose to write,

f lim
T

m = = f 0 T

Now from Eq. 4.18.23, we get a cleaner expression in the limit:

Version 2 ECE IIT, Kharagpur

limT xcn xcm = 4S x ( f )

T

1/ 2

cos 2 nv cos 2 mvdv

= 2S x ( f ) , m = n = 0, mn

1/ 2

4.18.24

Following similar procedure as outlined above, it can be shown that,

= 0, and lim T xcn xsm = 0 ,

T

T xsn xsm = 2S x ( f ) , lim T

m=n mn all m, n 4.18.25 4.18.26

Eq. 4.18.24 26 establish that the coefficients xc-s and xs-s are uncorrelated as T approaches . Now, referring back to Eq. 4.18.21, we can see that,
2 2 2 2 E xct = lim xcn cos ( nw0 wc )t + sin ( nw0 wc )t T n =1

4.18.27 4.18.28

2 = lim S x ( f m ) = 2 S x ( f )df = xt 2 T T n=1 0

Here xt 2 denotes the mean square value of x(t). Similarly, it can be shown that, E xst 2 = E xct 2 = xt 2

[ ]

[ ]

4.18.29

Since, xst = xct = 0 , we finally get,

st 2 = ct 2 = x 2 , the variance of x(t).

It may also be shown that the covariance of xct and xst approach 0 as T approaches . Therefore, ultimately it can be shown that xct and xst are statistically independent. So, xct and xst are uncorrelated Gaussian distributed random variables and they are statistically independent. They have zero mean and a variance equal to the variance of the original bandpass noise process. This is an important observation. xct is called the in-phase component and xst is called the quadrature component of the noise process.

Spectral Density of In-phase and Quadrature Component of Bandpass Gaussian Noise

Following similar procedures as adopted for determining mean square values of xc(t) and xs(t), we can compute their auto-correlation and cross-correlation functions as below:

Version 2 ECE IIT, Kharagpur

ACF of xct = Rxc ( ) = 2 S x ( f ).cos 2 ( f f c ) df

Rxs ( ) = ACF of xst = 2 S x ( f ).cos 2 ( f fc ) df = Rxc ( )

0

4.18.30 4.18.31

Cross Correlation Function (CCF) between xct and xst is:

Rxc xs ( ) = 2 S x ( f ).sin 2 ( f fc ) df
and

4.18.32 4.18.33

Rxc xs ( ) = Rxc xs ( )

0 fc

Eq.4.18.30 can be expressed in the following convenient manner:

Rxc ( ) = S x ( f ) cos 2 ( f f c ) df +
0

S x ( f )cos 2 ( f fc ) (df )

= =

fc fc

S x ( f + fc )cos 2 f df + S x ( f f c ) cos 2 f df
0

fc

[ S x ( f + fc ) + S x ( f fc )] cos 2 f df

4.18.34

From this, using inverse Fourier Transform, one gets,

S xc ( f ) = S x ( f + f c ) + S x ( f fc )

4.18.35 4.18.36

Moreover, S xc ( f ) = S xs ( f ) = S x ( f + f c ) + S x ( f f c )

Note that the power spectral density of xc(t) is the sum of the negative and positive frequency components of Sx(f) after their translation to the origin. The following steps summarize the method to construct psd of S xc ( f ) or S xs ( f ) from Sx(f): a. Displace the +ve frequency portion of the plots of Sx(f) to the left by fc. b. Displace the ve frequency portion of Sx(f) to the right by fc. c. Add the two displaced plots. If fc is not the centre frequency, the psd of xc(t) or xs(t) may be significantly different from what may be guessed intuitively.

Problems
Q4.18.1) Consider a pass band thermal noise of bandwidth 10 MHz around a center frequency of 900 MHz. Sketch the auto co-relation function of this pass band thermal noise normalized to its PSD. Q4.18.2) Sketch the pdf of typical narrow band thermal noise. Version 2 ECE IIT, Kharagpur