CHAPTER 7

Random Processes-Spectral Characteristics

7.0
INTRODUCTION

All of the foregoing discussions concerning random process€s have involved

the time domain. That is, we have characterized processes by means of auto-
correlation, cross-correlation. and covariance functions without any consider-
tion of spectral properties. As is well known, both time domain ond frequency
domain analysis methods exist for analyzing linear systems and deterministic
waveforms. But what about randorn waveforms? Is there some way to
describe random processes in the frequency domain? The answer is yes. and
it is the purpose of this chapter to introduce the most important concepts that
apply to characterizing random processes in the frequency domain.
The spectral description of a deterministic waveform is obtained by
Fourier transforming the waveform. and the reader would be correct in con-
cluding that Fourier transforms play an important role in the spectral char-
acterization o[ random u'aveforms. However. the direct transformation
approach is not attractive for random waveforms because the transform
may not exist. Thus. spectral analysis of random processes requires a bit
more subtlety than do deterministic signals.
An appropriate spectrum to be associated with a random process is intro-
duced in the following section. The concepts rely heavily on theorS of Fourier
transforms. Readers wishing to refresh their background on Fourier theory
are referred to Appendix D. u'here a short revieu' is given.

7.1
POWER DENSIT}'SPECTRU}I AND ITS PROPERTIES

The spectral properties of a deterntinistic signal r(l) are ctrntained in its

Fourier transform X(at\ given by

220
O
?|i-.
 22t
x(a) -J]*n,r" -i")t dt (7.r-r)
csrprrn 7:
Random
The function X(rrr), sometimes called simply the spectrum of x(l), has the unit
Processes-
o[volts per hertz when ,r(l) is a voltage and describes the way in which relative Spectral
signal voltage is distributed with frequency. The Fourier transform can, there- Characteristics
fore, be considered to be a voltage densit't' spectrum applicable to x(0. Both the
amplitudes and the phases of the frequencies present in x(t) are described by
X(co).For this reason, if X(a) is known then x(r) can be recovered by means of
the rnlerse Fourier transform

.r(,) - *f **go\ei'' da (7.t-21

In other words. X(o) lbrms a complete description- of .x(t) and vice versa.
In attempting to apply (7.1-l) to a random process, we immediately
encounter problems. The principal problem is the fact that X(ut) may not
exist for most sample functions of the process. Thus, we conclude that a
spectral description of a random process utilizing a voltage density spectrum
(Fourier transform) is not feasible because such a spectrum may not exist.
Other problems arise if Laplace transforms are considered (Cooper and
McGillem , l97l , p. 132).
On the other hand, if we turn our attention to the description of the pow'er
in the random process as a function of frequency. instead of voltage, it results
that such a function does exist. We next proceed to develop this function,
called the pou,er density spectrumt of the random process.

The Power Density Spectrum

For'a random process X(r). let.rr(r) be defined as that portion of a sample

runction .r(l) that exists betu'een - f and f; that is
[.r(r)
:lO" -T<t<T (7. r -3)
-t1(I)
elsewhere
Now so long as I is finite. \r'e presume that xr(r) has bounded variation. u'iil
satisly
1r
I lxr(t)l dt < a (7.1-4)
J-r
and will have a Fourier transform (see Appendix D for conditions sufficient
for the existence of Fourier transforms). which u'e denote X7\a\. given by

(7.1-s)
XTQo) - f_rr7(t)e-t'' o, - I'-rx(r)e-i''
ctr

tMany books call this function a Fo$'er spectral dcnsit1,. We shall occasionalll use also the names
or pot{er speilrum.
pov'er densily
- 222 The energy contained in x(l) in the interval (-i", I) ist
Probability,-
Random Variables, E(D _ [, xT.o a, = l'_rte) dt (7.1-6)
and Random J-r
Signal Principles Sinc xr(l) is Fourier transformable, its energy must also be related lo X7@t)
by Parseval's theorem. Thus, from (7.1-6) and (D-21) of Appendix D

E(D- [' *21t1dt: +r{* xrtr)|2 d, (7.t-7)

J_r
By dividing the expressions in (7.1-7)by 2T,we obtain the average power
P(1') in r(t) over the interval (- T, T):

.p/7'\
\r ,, --
t[r xz(t),tt=*f
2r J-, _ryt- o, (T:r-8)

At this point we observe that lxr(ail2 /2i" is a power density spectrum because
power results through its integration. However, it is not the function that we
scek,for two reasons..One is the fact that (7.1'8) does not represent the power
in an entire sample fuirction. There remains the step of tetting I become
arbitrarily large so as to include all power in the ensemble member. The
second reason is that.(7.1-8) is only the power in one sample function and
does not represent the process. In other words, P(JI) is actually a random
variable with respect to the random process. By taking the expected value in
(7.1-E), we can obtain an average power Pyy for the random process.I
From the above discussion it is clear that we must still form the limit as
I + oo and take the expected value of (7.1-8) to obtain a suitable power
density spectrum for the random process. It is important that the limiting
operation be done last (Thomas, 1969, p. 98, or Cooper hnd McGillem,
l9?1, p. 134, and 1986, p. 233). After ther operations are performed,
(7.1-8) can be written

txx-
D
1T-+ !'-,rtx'tda'
Equation (7.1-9) establishes two important facts. First. average power
Pyy in a random process X(l) is given by the time average of its second
moment:

txx-
Jg+J:. Elx2(t\ldt - Alqx2Q)ll
D (7. r -10)

trlre assume a real process X(l) and interpret .r( l) as eithe r thc voltage across a I -Q impedance or the
current through I Q. In other words, we shall assume a l-Q real impedance whenever we discuss
eocrgt or power in subsequent work. unless specificalll" statcd otherwise.
fln uking the expected value we replace.r(r) by Xltlia{7.1-8) since the integral of -t'(r) is an
operation performed on all sample functions of X(r).
For a procegg..lhat is at least wide-sense stationary , Hx2(il -P, a constant, 223
and Pyy : X2. Second, Pyy an be obtained by a friqueniy domain integra- cxerrr,n 7:
tion. If we dcfine the power density spectrum for the random process by Random
Processcs-
exx@i:Jgry (7.r-r r) Spectral
Charactcristics
the applicable integral, which we call the power formula, is

Pxx : *f yy(a)do (7.t-tz)

*,
from (7.1-9). Two examples will illustrate the above concepts. '

EXAMPLE z.t-r; Consider the random process

X(t) : 16 cos(a.5, + @)

where As and a.b are real constants and @ is a random variable uniformty
distribu,ted on the interval (0,r12). we shall find the average powcr Pxx
in X(l) by use of (7.1-10): Mean-squared value is

elx'(t)l: El,Aicos21o4r + @)l : Ee . +cos(2ar6r + 2@)]

.t? r/2 2
_Ai
T + 7iJ, -cos(2ant +20\d0
=4t7t -4sin(2aot')
This process is not even wide-sense stationary, since the above function is
time-dependent. The time average of the above expression is

Aqrx2u)r):
lgl *r ,l+ 4,in(?-or)a,

which easily evaluates to

Pxx:AlE[x2(t)D-AAp
EXAMPLE r.t-2. we reconsider the process of the above example to find
9xtkt) and average power Pxx by use of (7.t-tl) and (7.1-12), respec-
tively. First u,e find X7(r..r):
1r
. Xr(r):.J_rnocos(arsr + O) exp(-jat\dt

:+exp(7@) o)tlctt
J'..ro,r(oo -

!exp1-i@) J'..*r,-r(at + oitldt

+
224
sl{(r'r - r'1*r1
Probability. - Asr exp( l@, \@ - a1ll t
Random Variables.
and Random
Signal Principles
* Aorerp(-.io,iffif
Next we determine lXTl.l)l: : .Y7!o)Xi(r,r) and find its expected value.
After some simple algebraic reduction we obtain
EllX rk t)l2l i* f sinr[t<., - r.,.,0)fl * f sinr[(t,,, + r,,,r,)7'] ]
-= 2 f
2T 1; 11,,: ,,rlrit' , Xr;;7p I
In obtaining this rcsult we have neglected cross terms each having factors
at widely separate frequencies rrts and -@11 SUCh that their product is
small. Now it is knorvn that

ri* I[sintlr)l
I*:':IL o/ J -6(a)
(Lathi. 1968. p. l4). so (7.1-ll) and the above result give

'? xtko\ : +[6(r'r - aro) * 6(a + a1,)]

Finally. we use this result to obtain average po\\'er from (7.1-12):

r)^.
D :+ [- *t^. - tool * 6(a * ototld, - 4
!. J -oc

Thus, P1.; found here agrees with that of the earlier Example 7.1-1.

Properties of the Porer Density Spectrum

The power density' spectrum possesses a number of important properties:

(l) 9s.y@\20 (7.1-13)
(2f 9 xxGo\: .f s.xko) X(r) real (7.1-14)
(3) I xxko) is real (7.1- 15)

(r) xskt)da --l{EU'r(r)l} (7.1-16)

*,f *"
Property from the definition (7.1-ll)and the fact that the expected
I follou's
value of a nonnegative function is nonnegative. Similarll'. property 3 is true
from (7.1-ll) sin-ce lXr(ar)|2 is real. Some reflecrion on ih. properties of
Fourier transforms of real functions u'ill verify propert)' 2 (see Problem
7.1-9). Property' 4 is just another statement of (7.1-9).
Sometimes another property is included in a list of properties:
(5) 9 rr@) - ot29 tt\a) (7.1- 17)
It says that the power density spectrum of the derivative *11 dX(t)ldt is o? : 225
times the power, spectrum of X(t). Proof of this property is left as a.reader cxrpten 7:
exercisc (Problem 7. l-
l0). Random
A final property we list is Processcs-
Spectral
(6) *f xx(a)ei* dut: AlRyy(t.r + r)I (7. r-r 8) Characteristics
_,
9 xx(r\ = *xx(t, t + t)le--i-' dt (7.1-t9)
f*_*n
It states that the power density spectrum and the time average of the auto-
correlation function form a Fourier transform pair. We prove this very impor-
tant property in Section 7 .2. Of qurse, if X(l) is at least wide-sense stationary,
AlRyy(t,, + r)J = R;a(r), and property 6 indicates that the power spectrurn
and the autocorrelation function form a Fourier transform pair. Thus
ro
9 yyfui) - | Ryy()e-r" dt (7.t-z0)
J_x
Rxxk) = rrrcoyei-' da (7.1-2t)
*l*_*,
for a wide-sense stationary prodrss.

Bandwidth of the Power Density Spectrum

Assume that X(t) is a lowpass process; that is, its spectral components are
clustered near @: 0 and have decreasing magnitudes at higher frequencies.
Such processes are also called baseband. Except for the fact that the area of
9'xx@\ is not necessarily unity, I ys,fu) has characteristics similar to a prob-
ability density function (it is nonnegative and real). Indeed,.by dividing
9 xxki by its area, a new function is formed with area of unity that is
analogous to a density function.
Recall that standard deviation is a measure of the spread in a densitl'
.function. The analogous quantity for the normalized power spectrum is a
measure of its spread that we call rnrs bandv,idth,t which we denote W'r^,
(rad's). Now since 9xx@) is an even function for a real process. its "mean
value" is zero and its "standard deviation" is the square root of its second
moment. Thus, upon normalization. the rms ban{width is given by

,29 r^.1r1d,
w'l,n : I- 0 t-221
,, ,,, a\ cto
Jl-

tTbc ortation rms bandwidth stands fs root-mtdtt-.squared handu,idth.

 ExAMPLE z.t-r. Given the power spectrum
l0
Probability,
Random Variables,
I xx(r') =
11..,' 1r71gfii
and Random
Signal Principles where the 6-dB bandwidth is l0 radians per socond. we find [7r-r' First,
using (C-28) from APPendix C,

[- do _ ,.rs [-
lO da
J-*il + Grllofi' - '" J --(loo +,2)2
:rostmrmL.fr* 'EJL}
: 50n

Next, from (C-30) of APPendix C:

[* l}r;,z do [- ,2 d,
- ,^s
'" J -o" (loo +'2\2
J -"" tl ; (rflofP

=ro5tmL.fi,"n-'ffi)Ll
= 5000r
Thus

w^': {@-
so"
torad/s

Although W,^,and the 6-dB bandwidth of 9 yy(a) are equal in this case,

they are not equal in general

The above concept is readily extended to a process

that has a bandpass
components cluster
form of power .p..tiu-; that is, its significant spectral
wi process ,Y(t) is real'
ncar some frequencies r.rs and -a.rs. If that
assume the
i rr<rlwill bi real and have even symmetry about' ot:0' With this assump'
tion we define a mean frequencY at bY

eyy(c,t)do
f,
@o: -;6---- (7. r-23)

a'
Le *r@)
and rms bandwidth bY
ql* (, - h)'s *r@) a,
wk: (7.1-24)
o rr1'1a'
l*
The reader is encouraged to sketch a few lowpass
and bandpass power spec-
i*rnr .rA justify fiims.lf why
for the factor of 4 appears in (7 'l-24)'
7.2 227
RELATIOITISHIP BETWEEN POWER, SPECTRUM
rt\Lrrvr AND
AUTocoRnf,LATIoN FLTNCTIoN ^r\r'
ffi:''
In Section ?.1 it was stated that the inverse Fourier transform of the power 3ffi:f-
density spectrum is the time average of the autocorrelation function; that is Characteristics

*f--, xx(.V^ do = AlRyy(tl + r)l (7.2-t)

This expression will now be proved.
If we ur(7.1-5), which is the definition of X7Qg, in the defining equation
(7.1-l l) for the power spectrum we havet

.? xxd,t=
;gg ,lr+['_rru,rei'', ilif'_r*r,r',o- -', a,r)

= J55#
f ,f ,Elx(t)x(t2)ls-i.<tz-',t
dt2 dr1 (7 .z-2)

The expectation in the integrand of (7.2-?) is identified as the autocorrelation

function of .Y(0:
$X(t)X(rz)l: Ryy(t1,t2\ -T < (11 and tz)<T (7.2-3)
Thus, (7.2-z1bccomes

' g [' Rxx(tt, r2)s-io/tz-',t dr1dt2 (7.2-4)

t+@Zt+ J_T
[' J_T
xxko): -lim

Next, ut seek the inverse Fourier transform of (7.2-a\.I The transform

will exist since the power spectrum is real, nonnegative, and is absolutely
integrable from (7.1-12) because we presume Pxx < oo. The inverse transform
is

[- gxxko)ei-'d'
:lrJ-*

- I2nl-n
[* ll]L+
f-rf-,
Rxx(r, t1)s-Htt-t) dt dtt d" do) (7.2-s)

- ;g+ f ,f-rRxx(t'"'*!* *'* "+t)

dadtldt
From (A-23) and the event symmetry of impulses, we recognize the inner
integral of (;.2-5) as the impulse 2n6(t - rr * t) - t - r), so -2r6(t1

fWe use .f(0 i! 17.1-5), rather than x(l), to imply that the operations performed take place on the
process as oppoced to one sample function.
fOur develqpoent follows that of Middlcton (1960), pp. 142, 194.
228
Probability, l,f -'xx@\'i" 't'
Random Variables,
and Random : ;*# t1)6(r1 - t - rdtl ttt (7.2-61
Signal Principles f ,f ,Rrr(t'
The definition of the impulse given in (A-16) allou's the immediate evaluation
of the integral over ,r:
I [- Ixxkiej" d,
2,. J -."

_ T_x.l.l llr
'li*^-l J_r Rrr(r,t+r)(h -T<t*r<T 17.2-7\

In the limit as ln + m the condition -T < r * r < f has little meaning. The
quantity on the right side of (7.2-11 is recognized as the time average of the
process' autocorrelation function

(r. t + r) (tt (7.2-8)

A[Ryy(t,/ + r)] - ll1
,R.xr +f
Thus. \7.2-7) says that the inverse Fourier transform of the power density
spectrum is the time average of the process' autocorrelation function. which
proues (7.2-l). The dircct transform must also hold. giving
fr
9 xxkol - , * t\le-t'' dr (7.2-e)
J__r[^r..r'(r,
which shows that I y7{ro) and AlRysQ, t + r)] form a Fourier transform pair:

AIR^.y(t, r * r)l <-> 9s.s.Qoi) (7.2- l0)

For the important case where X(t) is at least wide-sense stationary,

,{[R1'1.(t, I + r)] : Rxx(r) and we get

9 xx@t= 'v(r|e-i" dt (7.2-l I )

Jl-^.
Ryy(r) - + ttlei-' cta (7.2- I2)
J1,",,r,

R.r.r.(r) <+ I yy(u.t) (7.2-t3)

The expressions ('7.2-11) and (1.2-12) are usually called the lViener-Kltirttchine
relations after the great American mathematician Norbert Wiener (1894-1964)
and the Russian mathematician A. I. Khintchine (1894-1959). Thq'form the
basic link between the time domain description (correlation functions) of
processes and their description in the frequency domain (poiier spectrurn).
From (7.2-13), it is clear lhat knowledge of the po\\'er spectrum ol a
process a.llows complete recovery of the autocorrelation funciion uhen X(l)
is at least wide-sense stationary; for a nonstationary process. trnll' thE time
average of the autocorrelation function is recoverable from (;.1-10).
EXAIIPLE z.z-t. The power spectrum will bc founJ fior the random procesc 229
of Example 6.2-l ttrat has the autocorrelation function .*G z,

RxxG) - tAi/2)cos(a.5r) lH:L

where ls and ab are constants. This equation can be written in the forrn Spectral
Characteristics

RxxG) =+(eju+' + e-lq')

Now we note that the inverse transform of a frequency domain impulse
function is

[" 6(a)d-'d. - 2n
:ZtrJ-*' )
from (A-2) of Appendix A. Thus \

| <+ 2n6(a)
and, from the frequency-shifting property of Fourier transforms given by
(D7) of Appendix D, we get
eJ%' ,- 2'r.6(o - aro)

By using this last result, the Fourier transform of Rxx(r) becomes

xx(a) :+[(a, - aro) * d(ar * a.rs)]

This function and RxxG) are illustrated in Figure 7.2-1.

EXAMPLE 7.2-2. As another example, assume a wide-sense stationrSt

process X(r) has an autocorrelation function
- (l:|lr)
Rxx(,):
lloll ;rffif '
where f > 0 and ,{s are constants. R.y;(r) is sketched in Figure 7 .2-2a for
Ao, 0.The power spectrum is found from the Fourier transform

e xx@): dt
I]_^, 7.(t1e-i-'
PlT
I tt +(rlT)Js-i"dt*AolJo U-G/T'1fe-i"dt
-As J-r
These integrals evaluate using (C-45) and (C.46) to give
9 xxl,rrl - .4of Sa2gof p)
which is sketched in Figure i.2-2b.
230
Probability,
Random Variebhs,
and Random
Signal Principlcs

-@6 0 otg 0)

(D)

FIGURE 7.2.7
The autocorrelation function (a) and power density spectrum (D) of the wide-sense
stationary random process of Example 7.2-1.

7.3
CROSS.POWER DENSITY SPECTRUM AND ITS PROPERTIES

Consider a real random process W(t) given by the sum of t$'o other real
processesx(l) and Y(l):
tt'(t1 (7.3-1)
-X(t)+Y(t)
The autocorrelation function of IV(t) is
Rww(t, r + r) - E[W(t)W(t + t)l
- E{[x(0 + r(r)][x(r + r) + Y(t + r)]]
: Rxx(t, r + r) I Ryy(t, t + r)
*R1.y(t, t+t)* Ryy(t,t+r) (7.3-2)

Now if we take the time average of both sides of (7 .3-2) and Fourier transform
the resulring expression by applying (7 .2-9't, we have
9ww@): I xx(a)* 9yy(w) + gVlnxy(t, t + r)l) + glAlRtr'(t, I + r)J)
(7.3-3)
 231
CHAPTEN, T:
Random
Proccsscs-
Spcctral
Characteristics

0
(a)

aao
fr--
-4' -2rc 0A$to)
T T
(D)
TT

FIGURE 7.2.2
The autocorrelation function (a), and power spectrum (D), for the wide-sense
stationary process of Exanple 7.2-2.

where 9l-l reptesents the Fourier transform. It is clear that the left side of
(7.3-3) is just the power spectrum of W(t). Similarly, the first two'right-side
terms are the power spectrums of X(r) and f(r), respectively. The second two
right-side terms are new quantities that are the subjects of this section. It will
be shown that they are cross-power densily spectrums defrned by (7.3-12) and
(7.3-14) below.

The Cross-Power Density Spectrum

For two real random processes X(r) and Y(l), we define xr(r) and y7(l) as
truncated ensemble members; that is

xr(r) : -T<t<T (7.v4)

{;t" elsewhere

and

{d,', ;rh:;'
,yr(r): (7.3-s)

Both xr[) and y/l) are assumed to have bounded variation and to be mag-
nitude integrabhoverthe interval (-T,I) as indicated by (7.1-4). As a con-
232 sequence. they will possess Fourier transforms that we denote by X1(ar) and
Probability, Yrko), respectively:
Random Variables.
xr(t) <+ X1(ar) (7.3-6)
and Random
Signal Principles -r'r(r) <+ Y7(ar) (7.3-7)
We next define the cro.ss pov'er P.u.(f) in the two prmsses within the
interval (- f, I) by

pxy(r) =
+ll, 'r,,,., 'r(t) ttr = +J]r,,,, 1,(ttdt (7 3-8)
Since .xr(r) and -r'1(l) are Fourier transformable, Parseval's theorem (D-20)
applies: its left side is the same as (7.3-8). Thus, we may write

Pxv(r) =+l'rr,,, 1'(t)dr = *l-_-Ury9 o, (7.3-e)

This cross poweris a random quantity since its value will vary depending on
which ensemble member is considered. We form the average cross power,
denoted Pxv(T), by taking the expected value in (7.3-9). The result is

Fr,.( D: +l'_,*rru. t)dt : +-l* *%@ o, (7.3-10)

Finally, we form the total average cross power P xv by letting 7n + oo:

Pxv : ),*+l'_,^r,,,, t) dr : *J* ;* lW o, (7.3-r r)

It'is clear that the integrand involving @ can be defined as a cross-power

density spectrum; it is a function of ar which we denote :

rxx(,,:)- T+*
lim
ElxiklYrko)l (7.3-12)
2T
Thus, we obtain what we call the cross-pov'er formula

Pxy : )rf (7.3-r 3)

*" ^.yb\da
By repeating the above procedure, we can also define ancther cross-power
density spectrum b1'
ElYlkt)Xr(ar)l
evx@) _ (7.3- l 4)
lT* 2T
Cross power is given by

Pyx : *l* yx@)rta - pit (7.3- I 5)

*n
Total cross po\\erP.r.r. * P yy can be interpreted as the additional power tw6
processes are capable of generating, over and above their individual powers,
due to the fact that they are not orthogonal.
Properties of the Cross-Power fD,ensity Spectrum 233
r Nerren 7:
Sonre properties of the cross-power spectrum of real random processes X(l) Random
and Y(l) are listod below without formal proofs. Processes-
Spectral
ll) :/'
^t.(a)
- .{ 1r.(-ar) - ./'\ s.Qo\ (7.3-16) Characteristics

(2) Re l./'Tylo,lland Re 1./ t,t@\l are even functions of ro (see Problem 7.3-4)
(7.3- r 7)

(3) Im llf sykolland Im l,f yr.(o\l are odd functions ol'rr,, (see Problem 7.3-4).
(7.3- r 8)

(4\ :/'\y(al -- 0 and ,/ t tko\ 0 if X(t) and l'(r) are orthogonal. (7.3-19)
-
(5) ii X(r)'and lr{r)'are'uncorrel':.r'ted ancl hare constant lneans i- and i-'
./'tt.ko\ : ,(/) t.x(al - 2nf iO(r) (7.3-20)

(6) l[R;y1r. , + r)] ,+ .l y1.Qtt'1 (7.3-21)

l[R1.q.1t., + r)] *+ ./ yllot)
' (7.3-22\
In the abore properties. Re [.] and lm [.] represent the real and inraginary
parts. respecti\e11,. and .4[.] represents the time average. as usual. defined by
(6.2-2t).
Propertl lfollows tiom (7.3-12) and (7.3-14). Properties 2 and 3 are
proved b1 considering the symmetry that X7@) and Yr@) must pos'sess for
real processes. Properties 4 and 5 may be proved by substituting the integral
(Fourier transform) forms for X7(a) and Yr{:.o\ into E[.['](co\Yrko)l and
shorving that the function has the necessary behal'ior under the stated assump-
tions.
Propertl 6 states that the cross-power density spectrum and the time
average of the cross-correlation function are a Fourier transform pair: its
dcvelopment is given in Section 7.4. For the case of jointll' wide-sense sta-
tionary processs. (7.3-21) and (7.3-22) reduce to the especially useful forms

f yy(a\- [* Ryyft\c-i-' dt (7.3-23)

J-*
e,r(.): Jl.-^
,^.1r\e-i-' ctt (7.3-24\

:
R4 (r)
*J]_ r'., 1ot\ei'u' cta
(7.3-2s)

Ryxk) -- r.tt')siu)t da (7.3-26)

+rJ^"
i.l ExArtPLt'+r. Suppose we are gt\/en a cross-po\\'er spectrum defined by
:.1
{. { 1 a * iholtl/ -ll'.a<ll:
1,,

1fi
r'r'(a) - lo elsewhere
>:
2v where W > O,a and b arereal constants. We use (7.3-25) to find the cross-
Probabilily, correlation function. It is
Random Variables.
and Random
Signal Principles
Rxy(r) =*f ,Q+ifla"da
L]* o)t"t o b 1t4'

- Ztt!-w- -"a+i ffil-*'*" "

on using (c-45) and (c-46) this expression will readily reduce to

Rxy(r) : +,V;l:,] *i
**[.'" V: -# I ;]
I
: ;fu\@wt - b) sin( wt) * blut cos( llzr)l

EXAMpLE z.lz. We determine the cross-correlation function corresponding

to the cross-power density spectrum
8 -
I xv(a):
(o */ar)3

where o > 0 is a constant' we write I xt @) : 4G(a) where G(u) --

2l@+ jat)3. From pair 17 of Appendix E, we have

g(r) - u(t)ie-"'<+ G(ar)

From the linearity property of Fourier transforms and pair 17, u'e get

Rxy(z) - 4u(t)ie-"'
lgxy@)l and Ryy(r) are sketched in Figure 7.3-1. :

*7.4
RELATIONSHIP BETWEEN CROSS-POWER SPECTRUM AND
CROSS-CORRELATION FUN CTION

In the following discussion we show that

exv@):
J:{;15,.Jl.^,,,,,
r* r) atl,-i" a, (74-r)

as indicated in (7.3-21).
The development consists of using the transforms of the truncated
processes given bY

Xr@):
. ['
J_r
X(t)e-)-!dt (1.4-2)

Yr@): Y(tr)e-i-t' dt, (7.4-3)

I:,
 235
a2f,r"(r)
3 cxerrr,n 7:
Random
Processcs-
Spcctral
Characteristics

$vnott

0
otla
(D)

spectrum (b) for the

[i:::3r];il,1"" function (a) and magnitude of cross-power
process of ExamPle 7.1'2'

(7'4-3) to form
in (7.3-12). First, we use (7'4'2) and
i't'
xi@)Yr@): I'-rx(id'' *!'-rY(t)e dtl 0.4-4)

cross-power density spectrum'

This result is used in (7.3-t2) to form the

exvko)=;'yyff@
: J;+'[J:, xl)dd o,I'.,Y(t)g-t'" '"1
-tt drdtt Q'4-s)
= Jg *!' ,Jlrn't,, t1;e:jox"
we
developed in Section 7'2' First'
Next. we us€ a procedure similar to that as an
inr.erse Fourier transform both siOes
oi t7'a-5) and identify one integral
impulse function

i
l*
236
:!fi [* ,'/'s'v(oleiu" rla
Probability. J ---
Random Variables.
: +,J dt ttt /'r'| -"
and Random
Signal Princrples * ;* *l: ,.[1 , ^ ' ' tr' 11)1' 1ei"''
't'
- trl +r ,11, ^,,,, ,, + .[l_""".,- "+'t , dut tlr 1 dt

(7.-1-6)

Further reduction is possible on usc of (A-16). Thc definition of the

impulse allos's the irnmediate solution for the intcgral over I1.

|rl* ..''' 1v'(aY""' do : llr..L +[, R \ ).(r. t -f t) dt (7 4-7)

which is valid for -T < r * r < f. Thiscondition arises from therequirement

that the irnpulse in (7.4-61bc within the range of integration. The condition
can be ignt-rred as f --, x, in the limit. Equation (7.4-7) indicates that tlre
cross-power spectrum and the tinre-aVerage cross-correlation function form
a Fourier transform pair. The result proves (7.4-l ) because it is the direct
transform pert of thc pair. while 17.2-7) is the inverse part.
It should be noted from (7.4-7) that siven thc cross-power spectrum. the
crr'rss-correlation function cannot in seneral be recovered: onlf its time aver-
age can. Ft",r i61n,ly uide-sense stationar)' processes. honever. the cross-
correlation function R;.1(r; t'anbe found fronr :/ y.yko) since its time average
is just Ryr (r).
Although u,e shall not give the proof. a development sirrrilar to the above

EXA\IpLE z.r-r. Let the cross-correlation function of tu'o procbsses X(l)

and )'( r) be
.48
R1y (l.l:- t) : {sin(r,,,11r) -F cos[r,,,1,(2l + r)]i
i .

u'here A. B. and rr.i, are constants. We find the cross-po\\er spectrum by

use of Q.4-l ). First. the tirne average is formed

lim (t. t + t)
7'*x +r ,
R.r.r clt

AB ' AB
: 2
stn(ar'r)* ''
: [' cos[ar6(2r
z lT-zrJ_r L \,\ + rl)dt
:
R The integral is readill- evaluated and is found to be zero. Finally we
f
l: Fourier transform the time-averaged cross-correlation function q'ith the
aid ol pair l2 of Appendix E:
 237
q ,/ r.ytat = <'unprrn 7:
..,
:r
.:;
"l+sin(atr)) Random
-jnAB _ a4) Proc'esses-
l,:
,r.
= -a_ [6(a., - 6(ar + @o)J
Syrectral
Characteristics

7.5
POWER SPECTRUMS FOR DISCRETE.TIME PROCESSES AND
SEQUENCES

As with continuous time processes. power spectrums for discrete-time (DT)

processes and sequences. as defined in Sectit-ln 6.1. result from Fourier trans-
.formation of :.rppropriatc sorrclation functions. To f.illi' appreciate thcse
topics. we need materral to be developed in Chapter 8. The reason is that
the developments depend, in part. on sampling theory and the response of a
netu,ork to a random signal (the material of Chapter 8). Hou'ever. for com-
pleteness irer'e: it is reasonable ro just summarize the imporf ant results, with-
out proofs. We shall consider only baseband processes and sequences that are
at least jointly wide-sense stationary.

Discrete-Time Processes

Let I'(l) be a band-limited+ random process for which samples are taken at
tinres nT,. n 0. tl. +2. .. .. to form a discrete-titne process X(rrl). When
-
samplin-e is the result of multipl),ing X(l) by' a periodic "sampling" pluse train
consisting of rectangular pulses of short duration 7,, and amplitude l/Tp
occdrring each r. seconds. the sampling theory of Chapter 8 shou's that the
Drocess' autocorrelation function R.vr(r) has a sampled representation.
denoted by Rr r (r). that is eiren b1'

Rr r (r) - R11(r,;,,1*., (+)

Rrr'(r)'i
',t -
= ttT,)

:i )6('l - nr') (7.5- l )

^,;t;^
The second sum is an approrimation to the first sum that results liom use of
narrou sample pulses ( I, small). The third sum derives from t.he second bv use

! tA bandlimited process. frrr our presRnt purposes, is one with a power density sPeclrum that is zero
at all frequencies except o\er a frnitc band where larl < @,/2. u'ith ar. - 2tr1T, and r. is the constant
tt time betwcen any adjacenr pairs ol samples. I is called the sampling iilterwl or sanrpling perittd.
I
 of (A-29). Sampling with very narrow pulses (impulses in the limit) is called
Probahlity, ideal sampling.
Random Variables, Direct Fourier transformation of (7.5- l ) defines the power spectrum of the
and Random discrete-time (DT) random process, which we denote by ,/.1,7.,@)
Signal Principles
g x,.x.@\- RyrlnT,le-i'-r, (7.5-z)
,,1
It is to be noted that (7.5-2) is periodic in the variable a., with period
@s :2nlTr'
It is of interest to show how the power spectrum of the DT randorn
process is related to the power spectrum of the process X(l). First. we sub-
stitute the relationship (Problem 7.5-2)

,L6(r-
nr,) -l,L__*n*. 17.s-ry

into the middle form of (7.5-l) and then Fourier transform the result to get
(Problem 7.5-3)
rN
lr
9 x,x,ko) -+ D Y rr(a- nu-t,) (7.s-4)
' I ,=-o.
Now I xx@)is the central term of (7.5-4) where n : 0. If we write a rectan-
gular llnction recl(ala,\ on both sides to select out 9'y;.fut), u'e have
9 xx@) - r. rect(af o4).f x.t @) (i. s-s)
Our result, (7.5-5), shows that the process'power spectrum is f. times the
central period portion of the periodic power spectrum of the DT process.
Thus, (7.5-Z) in its central period yields the process' power spectrum within
a constant factor.
Alternatively, (7.5-5) can be obtained another way. In Chapter 8 it is
shown that the autocorrelation lunction has the following valid representa-
tion:

Ryy(t) - I R1-1.(nl) Sa[ar,(r - nT,\12) (7.s-6)

based on sampling theory. The direct Fourier transformation ol (7.5-6)

produces

9 xx@)- I, rect( ot/a\) i tT,1g-t"'r'

^.rrt,
- r* rect( -t.,l5 rl,ol ( 7.5-7 )

Which is the same as (7.5-5).

The above procedures and results apply to an)' other wide-sense station-
ary, band-limited. baseband process. say, )'(l). One only needs to revise sub.
scripts. Similarll. for X(t) and Y1t) jointly u'ide-sense stationary. appropriate
cross-correlation functions and cross-power spectrums result from subscripts
XY and YX. cxrprrn 7:
It is emphasized that the above results require processes to be barrd- Random
limited and the sample rate alr to be large enough.t For processes that are not Processes-
perfectly band-limited, there is usually a practical value of ar above which the Spectral
power spectrum can be considered negligible. If the sample rate exceeds at Characteristics
least twice this practical value, the power density expressions can still be uscd,
but give the desired density with a small (negligible) error. The error is due to
an effect called aliasing, and it can be reduced by faster sampling.

Discrete-Time Sequences

As noted in Section 6.1. the DT sequence is essentially a DT process, but some

notation is changed to be consistent with that in the DSP (digital signal
processing) literature. Since computers treat samples as simply a sequence
of numbers, they care not about the time separation (I.) used between sam-
ples. In recognition of this fact, the explicit dependence on I is dropped and
X(nT,\ is written as X[nl, a function of index n denoted by brackets rather
than parentheses. Two other changes must also be recognized.
First, a new "variable" O, called the discrete frequencl', is defined, even
though it is actually an angle, as
Q : <oI, (7.s-8)
Second, it results that :-transforms are extensively used in DSP. For our
purposes it can be recognized that the right side of (7.5-2) is a z-transform
of the sequence of samples of the autocorrelation function. For example, the
bilateral (or ru'o-sdedl :-transforur, denoted by 9x,x,G) (boldface type used
for the :-transform). of the sequence Ryy[n] is

9 t',x,G)- Rvyln):-" (7.s-e)

,,8""
By. comparing (7.5-2) u'ith (7.5-9) we have

' .v x.t.,(a, Ry',.lnle-i't,,r' -,,L Rxr.[r] s-ina

-,i*
- -?t',x,G)l -I
I

t'.t',@i!) (7.5- r0)

l---.'ro

In general, -- can be any- complex number. However, (7.5-10) shows that the
power density speL-trum is the :-transform evaluated on the unit circle [where
: - exp(./Q). l:l - l. and the angle of : is Al. The second right-side sum of
(7.5-10) is called the,ii-rcr.'ie-rirne Fourier transforrrr (DTFT) of the sequence
R;,'[n].

tlfthe frequencl of rhe highesr nonzero spectral term in .? rt@) is ll'^ . called the spc('trul c.rrerrr- Ihe
sampling rate must hrr.r. > lll',. uhere 2lt'1 is known as the.\1'tTrrist r.tte (scc Section 8.7r.
2N EXAMPLE z.$t. Suppose the autocorrelation sequence of a DT sequence
Probability: X[n] is
.Random Variabhr
and Random
R;r [r] = ,fi lal < I

Signal Principles We use (7.5-10) to find the DTFT of X[r].

gx,x,ko\: s-ine --,,L int"-ina

,l,R.rxlrl
-,,i- o-n'-i,,. ,,ttu-/ndl (l)
[
We use the sum of (C-62) by identifying ,,'- [aexp(lQ)]-r and u':
aexp(-iQ). respectively. for the tro sums in the last form of (l). We have
i.- I
{
tq,
,{.v,v.(o) -r1Q
t_
t t
t-ut' --.-rS)

* (l - ar)
{2 : oT,
tq: (l + a') - 2acos(Q)
t,
t:
Note ihiit this power spectrum i; periodic in the yariable ft. aq are all
power ipectrurns of all f)T sequences.
Developments similar to those leading to (7.-5-10) produce expressions for
the power spectrurn of Y[n] for a process ]'(r) and for the applicable cross-
power spectrums for ,\'[r] and I[n]. These are lelt as reader exercises.
In (7.-5-10) the DTFT was defined as the second right-side form u,here we
say .?1...1("in) is the DTFT of Ry.1[n]. There is also an inyerse DTFT lor
IDTFT) u,here R1.y[rr] can be recorrred from 9t,,^..(rin). The two form a
DTFT pair:

'f t',.t,(.ia ) - Ryylnle-t"a (DTFT) (7.5-lr)

,,L
lf"
R1 1 [rr]
/"Q
- ,"J- _.9 r' r',(c'aii' r/ S) (I DTFT) (7.5- r 2)

ExAltpLE 7.s-2. We shou' that (?-5-12) is valid by provin-e that the right
side results in R1.i [n].

+f ,'? Y,.Y'1eia1't"erlQ - *1. ,,,i- R s .1 fnfie-in'e e

tna
cl {z

- iL R.vvfnfl: [" ,itn-nttQ 7n

/7t J-n
f*
= iJ Rvltnr-grrtr]
' (tt-ttt\tr
Since sin[(n - nt)n]ll(n - ni.vl=;;;, all nt t' n and equals unitl for
nt = n, the sum reduces to R5.1[nl. This result equals the left side' of
(7.5-12), proving the IDTFT is valid.
 EX^M'LE 7S3. Let X(l) be a (whitet) gaussian.noise random pro
xilHffi:t"tfi ]!ru*^1*irxl;':.; t'itxTJl?.;.ilrfi .ry[
other value;
of k- yfnl-
tf a new sequence x[nl+osx[n- rJ+ 0.25x
[n-21is ftrrmed, rhen ir can be shown-tirrout.nr 7.5-r3) that
I tzt tt6to2_ k=0
Rrr,rAr : A:*land -l
li:*2and -2 (l)
lil,,r_ilj, all other k
It can also hc shou'n (probrem 7.5-14) rhat the power specrrum
of r[rJ is
2
tl y,1..(a) - gy,r.(de): f Ryr.[AJ e-jke
A':-2
_ o', ,",] + 20 cos(e) cos(2A)]
+
,

- 16
[2 8 e)
from (7.5- l l).
_.- -orr
erlmple consists of using MATLAB for simulating the sequence
Y[a], computing estimates of its autocorrelation function
un-d po*.i,p."-
I trum. and comparing rhe resurts to the true quantities. The
MATLAB
fr cod_eused is given in Figure 7.5-1. Specifically, we first use
Tf the procedures
;:, of Exampl.' 5.6- l to generate r002 values of zero-mean, gaussian
*. random
ti variable.l'with a variance of 4. These values are used to generate
I000
t:' values of )'[r] as defined abore. Next, we estimate
dP
the autocorrelation
Iri sequence Rr r [kJ using rhe formula (Childers. 1997, p.297)
sl
I v-l-lA'
H kl:i
Rr'r'I Y[n)Y[n+lkt] lkt <N
rC
E
I
ilit u'here A = 1000. Results are shown as the stem plot
of Figure 7.5-2,
where the exact function of ( r ) is shown as the piecewise
fi continuous
c{.\'e for comparison. \\'e see rhat some significani .r.o6 remain. eyen
P, u'ith N
:.1 - 1000 data -points. Finaily, we compute an estimate of the po*er
spectrum using the formula (Childers, I997, p. 303)
.\'- |
g'r.r {-'s
- l=-(.\'-lr
I R'r U.l e-ike {2 - atT,

Direct results sheru'n in Figure -.5-3 for N : 1000 prove to be verl,noisy,

so thecodt'that rr'as used has averaged adjacent intervals
of 100 frequen-
cies to Srrooth the plot to thal shou'n in Figure 7.s-4
[the lower cur'e is
the exact result of tl)].Clearl1. e\.en lor 1000 d.t" values our esrimare
Shr)\\'S sotltc significant error.

t\\'hitc noise is properly dcfined in Secrion 7.t,

24t
242 tttfttttttft$t lre*r1o 7.5-3 ttttlttt$ffitfrffi,
Probability, cloG
Random Variabhs, B r 1000r
and Random lrlr - !l t vlrl'arco of r
Signal Principles lig ' 19,
raada('3tat.',110),
r r .grt (rrver) rrradn( 1, U+2 ) r t t . t . d. Orumlaa raado wrrLrblo
n.*(3:!l+2)ltr(a)
n1 - x(2rf+l)l tr(a-l)
xa2 r xn1(1:N); tx(u-2)
y - r + 0.5rna1 + 0.25rxa2l
Ryfr r zcro. ( 1, 2rtr-1) I t lnltlall,zc
for L r -lt+1:Il-1
nd:<1 r na,x( [1 1+hI ],:ul"n( [t{+k !tl.) ; .1,-,,1..,1.:
s

naxZ r Erx( tt'1-LI ; :aj.n( :,!l-k Nl ) ; .' ; -;i

Rlrfr (lt+h) . nrn(y(adrrl) . ry(u&t2 ) ) . l[i I rutocorrclatl,on
.nd
Rtnrc r r(rrarr [0 0 1/a 518 ?r.lLG 518 Lll0 0l , t tnrc valuc of R1ryr
t-2r![-Lt -
e - -5;t +pLltis2r9i lllzP.Li t f rcqucncy vcctor
Stmc . xrrar/16c (11+20rcoe (w) + 8rcog (2rr1 ) ; t tnre valuc of Stry
Sftrl-noLgy - abs (fftrhLft (f ft (Ryy) ) ) I t Dor'lr tgcctral crtLntt.
ltp r 1001 t anrnbcr of poLats to awcragrc (rtD Erst b. cvcaly divLsLble
t btr li+1)
Syy r zcro.(1, (t+1)/etp1;
tor I r 1: ltD:ll t motb tbc Dorsr sDcctnrr cetl.natc
lf I < (ri+1)/2
sy:r( (t-1) /rt9+t1 . Ecau(stnr_nol.ty(i: i+stp-l) ) ;
clse
lryy( (l-1) /gtp+l) - mean(stry-noiry(l-1:i+etp-2) );
rnd
snd
clf
glot ( [ -Iag -3: 3 lag] , Rtnrc, 'k, )
ho1d
rtcu( -lag: Iag, Rtry(ll-Iag:N+1a9),, k, )

rIabcI ('Lag')
ItH:tjiHn::xi"l,oo,,
f Lgrurc
glot (r, SYY-aokY(1:lr) ,'k'I
xlabcl ('Normallzcd Frcqu€acy (rad) ,)
ylatrcl (,uaguitude, )
tLtlo ( 'Pw6r SDoctnD, )
f Lgrurc
plot (w(round lotglz) : etp: Ietrgtb(w) ),Sy1t,, --k, erre Strnr€,,h, )
xlebol ('trorualizcd Frequoncf? (rad), )
ylabeI ('MagmLtuds')
tltlc ('Poreer St,€ctnln' )

. FIGURE 7.5.I
The MATLAB code used in Example 7.5-3.
 Autmrrclrtim 243
cxrpren 7:
Random
Proccsscs-
5
Spectral
I
Characteristics
ie
C
to
i1

_l r : I I I I r I I r r

-r0 -8 -6 -4 -2 A 2 I 6 a r0
L-ag

FIGURE 7.5-2
True autocorrelation function and estimated autocorrelation sequence applicable to
Example 7.5-3.

Discrete Fourier Transform

Equations (7.5-ll) and (7.5-12) apply to a band-limited signal, be it determi-

nistic or iandom. The principal points to be made are: (l) there is theoretically
an infinite number of values of autocorrelation function needed to complete
the sampled signal's power spectrum, and (2) this computed power spectrum is
a continuous function of Q that consists of nonoverlapping replicas of the
power spectrum of the continuous signal that was sampled. When a digital

Powcr spectnrm

10

50
€
Io
y,

l0

0
-4 -3 ;2 -l 0 I I 3 4
NormaIized frequency (rad)

FIGIIRE 7.5.3
Unsmoothed pouer spectrum estimate applicable to Example 7.5-3.
 Powcr rpoctnrm
l6
Probability,
Random Variablcs.
and Random
Signal Principles
l4

t2 in\
i\
o
l0 i'/ \.,
it\\
!
it \\
tl
E
a8
\\
a!

.I -l
t/\
i/
-2 -t 0 I 2 3 4
Nom'ralized frequencr' (rad)

FIGURE 7.Y
Smoothed estimated and exact power spectrums applicable to Example 7.5-3.

computer is uscd to compute either (7.5-l I) or (7.5-12) some problems imme-

diately arise.
In regard to point ( I ). the computer cannot use an infinite number of .

values R.r.r[n]. even if they were available, since we cannot wait forever for
the final result, Furthermore, such a procqdure could involve an infinite
amount of memory and/or an infinite number of calculations. Thus, from a
practical standpoint, the computer typically works with a finite number of
values of R1.x.[n], say, N. For truly band-limited signals this limitation requires
truncation of R1.y[n] to A'ralues, and some error will result in computing the
power spectrum. Truncation results in some spectral distoriion and aliasing
(the overlap of spectral replicas due to distortion-caused spectral spread-see
Section 8.7 lor more detail). Because even truncated waveforms haye spec-
trums that decrease with increasing frequency, there is always some practical
frequency beyond which the spectrum's magnitude becomes negligible: this
fact allows truncation errors to be reduced by sampling at a rate,higher than
twice the practical frequencl'.
ln regard to point (2). even if a computer could determine the continuous
power spectrum (the DTFT). it must still produce results for discrete values of
Q (or ar, since g - ,I,). Thus, it is common practice to discretize the variable
Q (or ar) to a finite number of values. Usually the number is chosen the same
as the number of sample ralues (denoted by N above) to facilitate practical
computations (by using the fast Fourier transform or FFT).
From the above comments it is clear that the practical use of the DTFT of
(7.5-ll) and the IDTFT of (7.5-12) involves sequences of finite length N. Let
us define the extent of the samples as N[, and the discrete values {r{ &nd Q1. of
at and O, respectivelv. by
Zrfr/(Nfr) k : 0, l, ... , (N - l)
e)k: (7.s- r 3)
S2r: @Trl,o-t-2trklN k:0, 1,',..,(N - l) (7.s- r 4)
Next, rve define the discrete values of the power spec.trum by 245

9 x,x,ku*) = I x,x,(e;o^ ) - I r,x,(d''*T') (7.s-r s) cxepren 7:

Random
9 x,x,lkl : I x,x,(eio^ ) - I x,x,(d2rk I N 7 (7.5-r6) Processes-
Spectral
These definitions allow us to write the truncated version of (7.5-l l) and the Characteristics
discretized version of (7.5-12), respectively, as
N-t
g x,x,lkl: In:0 Rxxlnlr-t2nnklN k : 0, l, . . ., (N - l) (7.s-t7)

I N-l
Rxxlnl=;I '?x,x,lk142n/-tN n:0, 1, " ', (N - l) (7.s-18)
ft=0

Equation (7.5-17) is called the discrete Fourier transform (DFT) of the

finite autocorrelation sequence, while (7.5-18) is known as the inverse discrete
Fourier transform (IDFT) of the discretized power spectrum. Together the two
results form a DFT pair.
An alternative version of the DFT pair results from use of the DT
sequence Ryy(nTr'1:
N-t
9t,t',(r,,x\ - I Rxx@T)r-ina*r' k:0. 1,...,(N- l)
n=O
(7.5-19)

. rl-l
xxr.(rrr): gy,y,(o*);1na1r' n 0. ..., (N - l) (7.s-a0)
- 1,

"E
EXAMPLE 7.H. To set the staBe for the following example we determine
the autocorrelation function RxxG) for a continuous time process X(l)
that we assume has the power spectrum

exxkt)= Kcos (#;) r*(h)

Here K > 0 is a real constant and Wy is the spectral extent of the process.
On inverse Fourier transformation we get

RxxG) : ry lSafw s.t - (tt /2)l* Sa[ w yt * (tr l2)ll

(KWx12)cos(Wyt)
(1)
(nl2): - (tvs.t)z
In computing (l)
we have used (C-7) and (C-45) to easil,v evaluate the
inverse transform's integral.

EXAMrLE ?.s-s. We demonstrate the solution of (7.5-19) by use of

MATLAB softu'are. Fint. we assume values of the power spectrum are EE
l--*fr
to be found at N - 5 frequencies. Ne:<t, we choose to calculate the power
spectrum of Example 7.5{ by using five true values of the autocorrelation
 function of (l) when the sampling rate is ,, -2r/7,:21421,. Thesc
Probability, values are found to be
Random Variables.
Rxxt0l : KZlQtT,) Rxxlll = K2/(3nT,\
and Random
Signal Principles Rxx[2] - -K2l05rL) Rxx[3] : K2/(35nT,')
Rxr-[4] -K2l63nT,)
-
The MATLAB code to solve (7.5-19) when these values are used is given
in Figure 7.5-5.
In Figure 7.5-6 the solid line curve is the amplitude-normalized plot
of the true power spectrum of (l) in Example 7.5-4 versus normalized
frequency [-n . (Q : ,Tr). r]. The computed frequency values are
shown as the stem plots at five points. Clearly. the program is producing
the correct values of the power spectrum at the limited number of points
used. This example is revisited in Problem 7.5-15. where the number of
frequencies is increased to show that accuracy is preserved while calculat-
ing the function at more discrete points.

7.6
SOME NOISE DEFINITIONS AND OTHER TOPICS

In many practical problems it is helpful to sometimes characterize noise

through its power density spectrum. Indeed, in the following discussions we
defin, two forms of noise on the basis of their power spectrums. We also
consider the response of a product device when one of its input waveforms
is a random signal or noise.

White rnd Colored Noise

A sample function n(l) of a wide-sense stationary noise random process f(t) is

called n'hite noise if the power density spectrum of N(t) is a constant at all
frequencies. Thus, we define
9 Nx@) - -t"ol2 (7.6-l)
for white noise, where .'l"s is a real positive constant. By inverse Fourier
transformation of (7.6-l), the autocorrelation function of //(l) is found to be
R,v,,s(r)
- 1. t"s/2)6(r) \1.6-2)
Thc above two functions are illustrated in Figure 7.6-1. White noise derises its
name by analogy with "white" light. u'hich contains all visible light frequen-
cies in its spectrum.
White noise is unrealizable. as can be seen by the fact that it possesses
infinite average power:

*l* *' ,'r ('o) d<o :a (7.6-3)

ttttttttttttttt Exeqrlc 7 . 5-5 ttttttttttBtttttt 247

clcar cHePrrn 7:
Random
trr 5; Processcs-
t . 1l Spectral
Ta r 1i Characteristics

163 r pi/Tel

lf rsm(Nr2) rr g
vr - -ur:rt : 2rwxlH 3 w:ri - 2rrxlN; t frequcncy vector
elee
r r -ur:N + w:r/l{ z Z)rrl[ 3 r:ri
cad
tau - -0.5tN: ( 0.5rf,-1) ;
giacl r zerog(lrlcngrtb(tau) );
ginc2 = zerog ( 1, lengrth (tau) ) ;
f . fl.ad,(Dl.*tau - ell? -. 0); t avol.d aLtr(O) /0
eincl(f ) - sia(Dlrtau(fl -Oil2)./(pl*tau(fl -eLl2lt
f = f ind(pi*tau - ell2 == 0); t eia(0) /O - 1
eincl (f ) - oDea (1, lengtb(f ) ) ;
f = f ind(Di*t,au + 9Ll2 -= 0);
sinc2(f ) = sin(Dirtau(fl +9L12)./(pi*tau(f y +9il2li
f = find(Di*tau + 9Ll2 '.= 0) i
sine2(f ) = oneB(1,lcngth(f));
Rtnre = fs*13*s1l (2.OL) *(sincl + sinc2) ; t tnre
autocorrelation
12.= -wx z 2*wxl1,28 : wrr - Z*vxlL28i t frequetreyvector
Stnre = k*cos (OL*wZl (2rvx) ) ; % tnre Doner spectnrm
Sxx = abs(f ftshift(fft(Rtnre,N) ) ); % estimatedgower spectrun

clf
Dlot (w2 , Stru?, 'k' I
hold
gtem(w, Sxx, 'k' )

xlabel ('Normrlized Frequency (rad) ')

ylabe1 ( 'Magmitude' )
title ( , Power SDectnur' )
FIGURE 7.5-5
MATLAB code used in Example 7.5-5.

However, one type of real-iiorld noise closely approximates *'hite noise-

Thernral noisc generated by thermal agitation of electrons in an1' electrical
conductor has a power spectrum that is constant up to very high lrequencies
and then decreases. For example, a resistor at temperature f in kelvin
248 I
Probability. 0.9
Random Variables.
0.8
and Random
Signal Principles 0.7 l-
l) o.u l-
!
a
.=

:- ::f
c
a

rI -1 -l 0 r 2
N ormahzcd frequsry (rrd)

FIGURE 7.tr
True power spectrum (solid curve) and MATlAB-calculated pou'er spectrum (stem
plot) applicable to Exampie 7.5-5.

produces a noise voltage across its open-circuited terminals having the power
spectrumt (Carlson. 1975, p. I l8)
l.t^ol2)(alr,ullT\
I n^1(a\ - (7.6-4)
eb"T - |
where a-7.64(10-12) kelvin-seconds is a constant. At a temperature of
T - 290 K lusually called room temproture although it corresponds to a rather
cool room at 63"F), this function remains above 0.9 (",1'o/2) foi frequencies up

0 0
(a) (D)

FIGURE 7.6I
(a) The autocorreldtion function and (b) the po\\'er densitl' spectrum of white noise .

lAdaptedliont Peehles ( 19761 v'ith pcrmission t,.! publishers Addison-Weslcv.

Adranced Bwk Program.f

tThe unir of 9r-r.(ro) is actually volts squared per heru. According to our convention. we obtain
watts per hertz by presuming the voltage exists across a l-Q resistor.
-
to l0r2 Hz or 1000GHz. Thus, thermal noise has a neaily flat spectrum at all 249
frequencies that are likely to ever be used in radio, microwave, or millimeter- cunrrsn 7:
wave systems.t Random
Noise having a nonzero and constant power spectrum over a finite fre- Processes-
qumcy band and zero everywhere else is called band-limited white noise- Spectral
Figure 7.6-2a depicts such a power spectrum that is lowpass. Here Characteristics

lPn
t
?pp(o) - | w -W<ot<W (7.6-5t
Io elsewhere

Inverse transforrnation of (7.6-5) gives the autocorrelation function shown in

Figure 7.6-2b:

Rrvrv(r):rW (7.6-6)

FIGURE 7.6.2
Power density spectrum (a) and autocorrelation function (D) of lowpass band'
limited u'hite noise.

tThis statement must be rccramined for f < 290K, such as in some superconducting s1'stem or
otfcr low-temperature deviccs (masers).
250 The constant P equals the power in the noise.
Probability, Band-limited white noise can also be bandpass as illustrated in Figure
Random Variables, 7.G3. The applicable power spectrum and autocorrelation function are:
and Random
Signal Principles gxx(a)=frrp ab - (w/2\ < la{ <ab+ W/2)
Io elsewhere
(7.6-7)

and

R,v,v (r) -- rtfficos(a.5 r) (7.6-8)

where a.5 and w are constants and P is the power in the noise.

.c- 1w uo* w
1

li
I
i
tIv

FIGURE 7.G3
Power density spectrurn (a) and autocorrelation function (D) for bandpass band-
limited white noise. fAdapred from Peebles ( 1976) *ith permission of pubtishers
Addison-Wesley, Adrunced Book Program.j
 Again, by analogy with colored light that has only a portion of the visible 25r
light frequencies in its spectrum, we define colored noise as any noise that is cnr,rren 7:
not white. An example serves to illustrate colored noise. Random
Pr<icesscs-
EXAMPLE 7.6t. A wide-sense stationary noise process N(l) has an auto- Spcctral
correlation function Characteristics

Rrurv(r) - Pe 3l'l
where P is a constant. We find its power spectrum. It is

I xn@l= -3it"-i'<ot 4,
I]_*
: dt *" ,G-ia), 6,
"J;r-(3+ia,)r J:-
These integrals easily evaluate using (C-45) to give
PP6P
9xx@)-3qi.,+3_io:g*;l
This power spectrum is sketched in Figure 7.6-4 along with the preceding
autocorrelation function.

0
(a)

-303a)
(D)

FIGURE 7.G4
The autocorrelation funoion (a) and power spectrum (b) of the colored noise of
Example 7.6-1. lAdaptedfrom Peebles ( 1976) with permission of publishers Addisotr
Wesley, Advanced Book Program.l
 Product lhvice Response to a Random Signal
Probabili$.
Random Variablcs, Product devices are frequently encountered in electrical systems. Often they
and Random involve the product of a random waveform X(r) (either signal or noise or the
Signal Principlcs sum of signal and noise) with a cosine (or sine) "carrier" wave as illustrated in
Figure 7.G5. The response is the new process
Y(tl - X(tlAocos(a5r) (7.6-e)

where le and {rs are constants. We seek to find the power spectrum I vv@) of
Y(t) in terms of the power spectrum tl y;.Qo\ of X1r).
The autocorrelation function of Y(t) is
Rry(r. r + r) - EIY(t)Y(r + r)l
- EIA2yXQ)X(t * rlcos(r,.r6l) cos(a4t * a41r)]
:-Ai, tt rt +t r)[cos(aror)
-\tn..^/-. -\ +r cos(2a.5r
^^^11., r +r atr)]
.. -\l (7.6-10)
Tl(r..r(t.
Even if X(r) is wide-sense stationary Y(r) is not since Rl'r'(1. t + r) depends on
r, Thus, we apply (7. I - l9) to obtain I y,1.(a) after we take the time average cf
RyvQ,l*r): Let X(t) be assumed wide-scnse stationary. Then (7.6-10)
becomes

AlRyy(t,r + rrl :+Rr.1 (r) cos(ahr) (7.6-r l )

On Fourier transforming (7.6-l I I u,e have

I vv@) - $V ryko -r,ro) + I xs.(.+ alo)l (7.6-t2)

A possible power density spectrum of X(t) and that givbn by (7.6-12) are
illustrated in Figure 7.6-6.It presumes that X(t) is a lowpass process, although
this is not a constraint in applf ing (7.6-12).

E One important use of the product device is in recovery

ExAr\rpLE ?.6-2.
i.. (demodulation) of the information signal (music, speech, etc.) conveyed
F in the wave transmitted from a conventional broadcast radio station that
V, uses AM {antplirude nrodularion ). The wave received by a receiver tuned to
i{ a station with frequency osl2lr is one input to the product device. The

i( cos (oor)

FIGURE 7.G5
A product of interest in electrical $stems. lAdapted from Peebles ( 1976 ) n'ith
permission of publishers Addison-\l:esley, Advanced $ook Progran.j
 2s3
CHAPTER. T:
Random
Processes-
Spectral
Characteristics

0
(D)

FIGURE 7.6.6
Power density spectrums applicable to Figure 7.6-5; (a) at the input and (D) at the
output. lAdapted from Peebles (1976) with perntission of publishers Addison-Wesley,
Advanced Book Program.l

other is a "local oscillator" signal ls cos(@sl) generated within the recei-

ver. The product device output passes through a lowpass filter rvhich has
as its output the desired information signal. Unfortunately, this signal
also contains noise because noise is also present at the input to the pro-
duct device; the input noise is added to the received radio wave. We shall
calculate the power in the output noise of the product demodulator.
Let the power spectrum of the input noise, denoted X(t). be approxi-
mated by an idealized (rectangular) function with bandwidth W pp cen-
tered at *a.6. Thus,
I.t"ol2 -@x - (W nr /2) < e) < -rrt * (W nr l2)
e xx@) ao - (War/2) < o < cro * (W'ar 12)
'
t or"o2 elsewhere

where -4'ol2 is the po\\'er density within the noise band. 81' applying
(7.6-12) the power densitl' spectrum of the output noise )'(r) ol the pro-
duct device is readill' lound (by sketch) to be
Probability;-'
l rroAzl8 -2@s - (W ar 12) < ot < -2c,4 + (W nrl2)
Random Variables, I yy(o) -
| "ro,qip -War/2<@<WnrlT
and Random 2q - (WnrlL\ < @ < Zao * (WnrlT)
Signal Principles
ld'n*'' elsewhere
Now only the noise in the band -Warl2 < o) < Wnr/2 cannot be
removed by a lowpass filter (which usually follows the product device
to reinove unwanted noise and other undesired outputs) bccause the
desired signal is in the same band. This remaining component of I yy
(ar) gives rise to the final output noise power, denoted Nn,

N' : zoI ftt'a1'12 "a'oAi acu

' --
-l"oAiW^'
g"
J -rr.^,,, 4

*7.7
POWER SPECTRUMS OF COMPLEX PROCESSES

Power spectrums may readily be defined for complex processes. We consider

only those processes that are at least wide-sense stationary. In terms of the
autocorrelation function RzzG) of a complex random process Z(t), the power
density spectrum is defined as its Fourier transform

g zz@) : dr (7 .7 -t\
I*_**r7k)e-t"
The inverse transform applies, so

RzzG)
- ;J _*rrrro)cl"
do (7.7-2)

For two jointly wide-sense stationary complex processes Zr,(t) and 2,,(t).
their cross-power ciensity spectrum and cross-correlation function are a
Fourier transform pair:

g z^2,(r) : dt (7.7 3)
I*_*^r^z,lrle-i"
lf*
:
Rz^2,(r)
*J _*rr,r,(otler" da (7 .7 -4)

An equivalent statement is:

Rz,,z,(t) * I z*z,ko) (7.6-s)

ExAt\tpLEz.z-t. We reconsider the complex process V(t) of Example 6.7-l

E and find its pou,er spectrum. From the previous example
H
Rl'l'(r) : s!('\tjA
,l=l
H
 On Fourier transforming this autocorrelation function we obtain
cnrrren 7:

evvko)- s[r*'F""1 Random

Processes-
Spectral'
N

- I7,91ri^'7
Characteristics
/.'
n:l
/V

:2n6(co - aro)
D7,
n=l

after using pair 9 of Appendix E.

7.8
SUMMARY

The material of Chapter 6 developed random processes through the use of

time domain concepts. This chapter developed processes from the standpoint
of the frequency domain. Central to the development is the concept of a
spectrum describing the manner in which pov:er is distributed with frequency
in a random process. This power density spectrum defines the process'spectral
propertieg, and is somewhat analogous to the voltage density with frequency
for deterministic waveforms. However, the analogy is principally due to the
fact that in both cases, the time domain characteristics are related to those of
the frequency domain through Fourier transforms. Important topics discussed
herein were:
. The power density spectrum and its properties were developed for any
(stationary or nonstationary) random process.
. It was proven that the time average of a process' autocorrelation function
and the power density spectrpm form a Fourier transform pair.
. Similarly. a cross-power density spectrum and its ptoperties were presented
for two random processes.
. The power densitl' spectrums of discret"-time processes and sequences were
developed and related to the power spectrums of the continuous-time pro-
cess from which they were derived via sampling. The relationship of the
discrete-time Founer transform (DTFT) and the inverse DTFT (IDTFT)
were related to the bilateral z-transform and methods used in digital signal
processing (DSP).
. Computer examples (and chapter-end problems) were included to support
(using MATLABI the materials on discrete-time processes and sequences.
. Some practical ntrise definitions of rvhite and colored noise were included.
The respo.nse of a product device was analyzed: it is valuable as a model for
some mlxers and phase detectors used in the practical world.
. The chapter closcd with some more advanced material on power spectrums
of complex random processes
2s6 PROBLEMS
Probability.
Random veriables. 7.1-1.' We are given the random process
and Randsr x1t; : ls cos(osl * @)
Signat Pririphs
where ls and o\ are constants and @ is a random variable uniformly distrib-
uted on thc interval (0, zr).
1a) Is X(t) wide-sense stationarl'?
(b) Find the power in X(t) by using (7.1-10).
(c) Find the power spectrum of X(t) by using (7.1-ll) and calculate power
from (7.1-12). Do your two powers agree?

7,1-2. Work Problem 7.1-l if the process is defined by

X(r) = u(t)Ascos(a.rsr * O)

where u(r) is the unit-step function

*7.1-3. Work Problem 7.1-2 assuming @ is uniform on the interval (0,r12).

7.14. Work Problem 7.1-l if the random process is given by X(t) = lo sin(c.b, + O).
*7.1-5. Work Problem 7.1-l if the random process is

X(r) : lfr cos2larel + O)

7.1-6. Let A6 and 86 be random variables. We form the random process

X(t): lq cos(@ol) + 8o sin(a.rel)

where a.6 is a real constant.
(a) Show that if As and 8s are uncorrelated with zero means and equal
variances. then X(l) is wide-sense stationary. :

(D) Find the autocorrelation function of X(t\.

(c) Find the power density spectrum.

7.1-7. A limiting form for the impulse function was given in Example 7.1-2. Give
arguments to show that the following are also true:

to) /T- T expl-ra2r'l : 6(o)

T
(D) - d(a)
mrexp[-lcul4
7.f-8. Work Problem 7.1-7 for the following cases:

(a)
JgI#2:6(o)
t" j:+ ru - lalrl : 6(o)
7.1-9. Show that (7.1-14) is true.

7.1-f0. Prove (7.1-7). lHint: Use (D-6) of Appendix D and the definition of the deri-
vative.]
7.1-ll. A random pioot. is dcfined by 257

l,(r) - X(,)cos(.ql, + O) ffiz'

Random
where X(l) is a lowpass wide-sense starionary process. ob is a reat constant. processcs-
and (t is a random variable uniformly distributed on the rnterval (0,2r). Find Spectral
and sketch the power density spectrum of I(r) in terms of that of X(t). Characreristics
Assume G) is independenr of X(r).

7.1-12. fhtermine.which of the following functions can and cannot be valid power
density spectrums. For those that are not. explain why.

-7 (D) exP[-1c.r-l)rJ
tat
ffi
I

#-6(,.,) (d) ;#lW

7.1-13. Work Problem 7.1-12 for the following functions.
cos(3co) (D)
. tut
T#
I
(Iffi
kt#: (r/)#
7.1-14. Given that X(r): IiIr a;X;(tl where {o;} is a set of real constants and the
. processes X;(l) are-stationary and orthogonal. show that
,\.

!/' 7.1.(al:
It=l a):/'1,1.,1to)

7.1-15. A random proc€ss is given by

X{rl=l11cos(Qr*69)
nhere 16 is a real constant. Q is a random variable with density function .fa().
and O is a random variable uniformlv distributed on the interval (0.22r)
independent of Q. Show that rhe poo.ispectrum of X(r) is

g;r(ro) =+Llnto) + la.-utll

7.1-16. If X(r) is a stationary process. find the power spectrum of
l'(t)-.4,, +86X(l)
in terms of the power spectrum of Xtt) if lo and Bs are real constants.

?.1-17. Find the rms bandwidth of the pou'er spectrum

P
9xx@):
I l*(a.,7ll'yj r,,r<KW
{ -
[o lu,l> KW
u'here P,W, and K are real positiveconstants. If r( -.+ oo, what happens?
258 7.1-l& Find the rms bandwidth of the power spectrum
Probability,
exx@\:
Random Variables.
and Random If 'o""ol2rtt) il,:Y,
Signal Principles wherr W >0 and'P> 0 areconstants.

7.1-19. Determine the rms bandwidths of the power spcctrums given by:

(a) I yy(c,t) =
P lalcw
0 lcol>W
(b) I yy(a\ :
P{l - lcolwll kol. W
0 lcol> W

- wherc P and W are real positive constants.

*7.1-m. Given the power spectrum

P P
"7'',kol:-
AA\-' - '+ -

['* (T)')' ['*(#)']'

where P, o. and W are real positive constants, find the mean frequency and
rms bandwidth.

7.1-21. Show that the rms bandwidth of the power spectrum of a real bandpass
process X(r) is given by
w?^,:4w-n - dll
wherea.6 is given by (7.1-23) and W'tsgiven bythe right side of (7.1-22).
:
7.L-tL The autocorrelation function of a random process X(l) is
R.rx(r) : 3 * 2exP(-4r2)

(a) Find the power spectrum of X(r).

(D) What is the average power in X(r)?
(c) What fraction of the power lies in the frequency Uaria -l l0 :
@ s I lJ1"!

7.1-23. State whether or not each of the following functions can be a valid power
density spectrum. For those that cannot, explain why.

l'le-*p(l+'2) (b) +jht)

1o;
I * iut \- ---\'-l -"r\
' cos(3r,r)exp{-ai
(c) @) 6tanlt2otl(t
+<.rr1l
(12+o-t
-jl=e
(e) cosr(r,.,)exp(-8a;2) (/) (-ja)(ja)t1-t -;ro)r(3 * j.)2

7.1:l,t lf 92iy@) is a valid power spectrum.of-a random process X(l), discuss

whether the functions dgyy(a\lda and d!grr.1u11dut2 can be valid power
spectrums.
7.1-25. (a) Rework Problan 7.1-15 and show that eveir if @ is a constant (not 259
random) the power spectrum is still given by ilffi Z,

9 xxki: @A]ol2ya(ar) +/n(-ar)l Random

Processis-
lHint: Timc-average autocorrelation function before Fourier transform- Spgptnf
ing to obtain I xx@\.| Characteristics
(6) Find the totalpowcr in x(t) and show that it is independent of the forrn of
the density function fa@).

7.1-26. Find the rms bandwidth of the power spectrum

9 xx@\: l /[ + (al W)2J3
where W >0 isaconstant.
7.1-27. Work Problem 7.1-26 for the power spectrum

I xxko) : ,'lll + @/ W)2f

7.l-LS. Work Problem 7.1J.6 for the power spectr'-:rn
f xx@\ - l/[t + (a/W)2f
1.1-29. Work Problem 7.1-26 for the power spectrum

r7.l-30. Generalize Problems 7.1-26 and 7.1-28 by finding the rms bandwidth of the
power spectrum

fxxki-tl\+kt/W)IJN
where N Z2 is an integer.

*7.1-31. Generalize Problems 7.1-27 and 7.1-29 by finding the rms bandwidth of the
power spectrum

I x.x@i : @2 /lt + ko/W)21'

:hereN>3isaninteger.
7.1-32, Assume a random process has a power spectrum

e xr@) : 11- @2 /g) la,,l ' 6

[0 elsewhere

Find (a) the average power, (b) the rms bandwidth, and (c) the autocorrelation
function of the process.

7.1-33. Show that rms bands'idth of a lowpass random process l'(r). as gin'en by
(7.1-22), can also be obtained from

,, ,2
: -l dzRrrk)l
wrm:
' R;rreJ dr'
l,=o
where Rr.x(r) is the autocorrelation function'of X(r).
- 260 7.1-y. Assume a random process X(0 has. a power spectrum
Probability.
Random Variablcs. erxtot)=#
and Random
SignalPrinciplcs where W >0 isaconstant.
(a) Sketch ,{ xt'(.').
(t?) At what positive value of ar. denoted by o4n:*. does ./'.1.1.(ro) reach a
maximum value?

7.1-35. Treat the pou'er spectrum of Problem 7.1-34 as bandpass and find its mean
frequency ri4y and rms bandwidth I/,n.,..

7.t-36. Work Problem 7.1-35. except assume the power spectrum

6lo'rl
xx(,,l:
all: + ,,rT
7.1-37. Determine which of the foltowing funciions can be a valid power density of
some random process. For those that cannot. give at least one reason why.
3

(a) f t-zf a(ar - 3r)

n=-3
(b) exp[-4 sin(6r,.r)]

(c) 6(r..' - 4) + 6(a.r * 4\ + e-i-' cos2{ l0r;

-5o-t
td)
,* *;af
7.1-3t. Find ars and I7r., for the power spectrum shown in Figure. P7 .6-6.

7.2-1. Find the power density spectrum of the random process for which
Rr'.r(r) = P cosa(c'ror)

if P and @s ate- constants. Determine the power in the process b1 use of

(7.r-12). ;

7.2-2. A random process has the po\r'er density spectrum

.rslyeo)=#
Find the average power in rhe process.

7,2-3. Work Problem 7.2-2 for the power speclrum

txx@l:r

7.24. Work Problem 7.2-2 for the pou'er spectrum

6,,2
t '.t'r-(rr.,) = (,
*rpf
7.2'5. Assume ,Y(l) is a wide-scnse stationary proccss with nonzero mean valuc 261
X +0. Showthat
ffi2,
9 xxki = 2r*?6(a) * [* Cyy(r)e-t-, dt ]rT::l_
J -a,
Spectral
where Cy/tl is the autocovariance function of X(t). Characrerisrics

7.2-6. For a randoa process X(r), assume that

R.rx(r) - Pe-'2/2ot

where P > 0 and a > 0 are constants. Find the power density spectrum of
X(t).lHint: Use Appendix E to evaluate the Fourier transform of nr.r.(.).J

7.2-7. A random process has an autocorrelation function

[ 4t-Qtlr\ o<r.Tl2
Rys.ft)
- lPll +(2r/T\l -r12 < r<0
[0 r<-Tl2 and t>T/Z
Find an<t sketch its power density spectrum. (Hinr: Use Appendix E.)

*7.2-8. A random process X(l) has a periodic autocorrelation function where the
function of Problem 7.2-7 forms the central period of duration 7'. Find and
sketch the power spectrum.

7.2-9. If the random processes of Problem 7.1-14 are stationary, zero-mean. statis-
tically indepcndent processes, show that the power spectrum of the sum is the
rr;lxs'
Fo r sta ri o na rv i ndepcnden r processes wi th
:lT:.il *LT:::f,:lil
7,2-10- Given that a process X(t) has the autocorrelation function
Rxr'(r) : lg-oltl cos(ro6r)

where A > 0- cr > 0. and rr.rs are real constants. find the power spectrum of
X(t).

7.?-ll. A random process l'(r) having the pou'er spectrum of Problem 7.2-3 is applied
to an ideal differentiator.
(a) Find tk power spectrum of the differentiator's output.
(6) What is the power in the derivative?

7.2-12. Work Problem 7.1-l I for the power spectrum of Problem 7.2-4.

7.2-13. A wide-sense stationary random process X(t) is used to define another process
.by

l'(r) : [* ,,{,.*',, - () r/€

J--
u'here i(l)is some real function having a Fourier trandorm H(al').Sho*' rhat
the pouer s.[Ectrum of Y(t) is given by

9vt'@)=9xx(a)lH(r\|2
262 7.2-14. A deterministic signal I cos(a6t). where A and ar1 are real constants. is added
Probability.
to a noise procrss tV(t) for which
Random Variabhs.
w1
and Random J/rr,(-) :
Signal Principles
lt,r*..I
andW>0isaconstant.
(a) Find the ratio of average signal power to average noise power.
(h) What value of llz maximizes the signal-to-noise ratio? What is the con-
sequence of choosing this value of If "?

7.1-15. A random process has the autocorrelation function

Rx.r(r) - 8 cos21,r4,r) exp( - l/lrl )

u'here 8. rral. and W are postiive constants.

(c) Find and sketch the power spectrum of X(l) when a4 is at least several
times larger lhan W.
(b) Compute the average po\\'er in the lowpass part of the power spectrum.
Repeat for the bandpass part. In each case assume a6> W.

*1.2-16. Generalize Problem '7 .2-15 b1 replacing cosl(r,.,11r; with cos^'(a4yr) where N > 0
is art integer. What is the resulting power spectrum when N is (a) odd. and (b)
even?

*7.2-17. The product of a wide-sense stationarl gaUssian random process .\'(I) with
itself delayed by I seconds forms a new process Y(t!= X(tlA'(t - f).
Determine (a) the autocorrelation function. and (h) the power spectrum of
)'(t). t,Hint: Use the fact that ElXt^'rX\X4'l : f[,Yr X]ElX\Xi+ E[XrX]l
E|X2X41+ 4,Xrx4l ElXzXr] - 2E[Xr jElX/1E[Xr]E[X.] for gaussian random
variablesXt.Xz,.Yl.andXq.(Thonras.l969.p.64.)i

1.2-18. For a random process X(l). assume its autocorrelation function is

Rr'.r'(t., + r) : 11,'-{r- cosl(l4t)

(a) Is X(t) wide-sense stationary'l

(6) Find Ryl(r).
(c) Find the power spectrum of X(r).

7.2-19. A random process is defined by I(r) : X(tl - X(t - a). where X(r) is a wide-
sense stationarv process and d > 0 is a constant. Find the autocorrelation
function and power density spectrum of I(l) in terms of the corresponding
quantities for l'(r).

1.2-2O. Find the autocorrelation function corresponding to the power density spec-
trum

I 57 + l2or
./'7tko)=(16+r,rffi:foI

lHint: Use a partial fraction expansion (Peebles and Giuma. 1991. pp.
149- 156) and Table E- I .l
7.2-21. Find the autocorrelation function corresponding to'the power spectrum 263

exxko)=# ffJil''
Proccsscs-
lHint: Use the convolution property of Fourier transforms given by (D-16).; Spectral
Characteristics
*7.2-22. Find the power spertrum corresponding to the autocorrelation function

Rxx(r) - [cos(or) + sin(clrl)]e-"|'l

wherea>0isaconstant.
*7.I1. Joint wide-sense stationary random processes X(l) and y(ridefine a pr(rcess
w'(tl bY
' W(t) : X(t)cos(rr4,r) + I(r) sin(a.5r)

where a.5 is a real positive constant

(a) Develop some conditions on the mean values and correlation functions of
X(t) and Y(t) such that W(t) is wide-sense stationary.
(b) With the conditions of part (a) applied to W(t), find its power spectrum in
terrns of power spectrums of X(t) and I(r).
(r ) If X(t) and I/(t) are also uncorrelated. what is the power spectrum of
w(t)?

732. A random process is given by

W(t): AX(t)+ By(t\

u'here A and B are real constants and X(r) and I(r) are jointly wide-senr
stationary processes.
(a) Find the power spectrum I ww(a) of W().
(b) Find I n.w(-l if X(t) and Y(t) are uncorrelated.
(c) Find the cross-power spectrums 9 t'y,(a) and .7 y,s,(a).

'7.33. Define two random processes by

X(t) : z4 cos(a4t * O)
I(r): tlz(r)cos(a4, + O)

u'hire A and @0 are real positive constanrs. O is a random variable indepen-

dent of l&'(r). and WQ) is a random process u,ith a constant mean value W.By
using (7.3-12), show that

I xyko\ : +[d(, - ab) + 6(ro * <o6)]

regardless of the form of the probabilitl densitl, function of O.

7.!4. Decompose the cross-power spectrunrs intrr rc-al and imaginary parts acco'rd-
ine to

,/'t'yko): Rr r (ur) + j/y1'(r..r)

./tx@l: Rlr(ar) + jl12;@\
2g and provc that
Probability. : Ryr(-<.r\ = - R yxk t\
R;y(rrr)
Random Variables,
Itvko\: Iyr(-ar) : -lvx(a)
and Random
Signal Principles
7.$,5. From the rmrlts of Problem 7.3-4, prove (7.3-16).

7.3{. Show that (7.3-19) and (7.3-20) are true.

7.!7. Find thc cross-correlation function Rr.r.(l,l + r) and cros-power spectrum

9xvfu,\ for the delay-and-multiply device of Problem 7.2-17. {Hint: Use
the fact that f[Xs XzXi - ElXr]flX:Xrl + ElX2lEIXtXrl+ ElX3JElXrXrl-
zElXiqX2lE[Xl] for three gaussian random variables X1. X2, and Xt.
(Thomas, l%9. p. 64.))

7.3{. If X(t) and Y(t1 are real random processes, determine whkh of tlre iollowing
functions can be valid. For those that are not, state at least one reason why.
(a') Rxxftl = exp(-lrl) (b) I Rx r(rlt s r',Fqyrtol
(c) R;;(r) = 2sin(3r) (d) f txko\-61rc*7lity
(e).?y(af =tffi (f) 9xv@):3+jo2
lg\ ,? yyQo) = 186(ar)

7.3-9. Form the product of two statistically independent jointly wide-sense station-
ary random processes X(t) and Y(l) as
w,(t\ - X(r)I(r)
Find general expressions for the following correlation functions and power
spectrums in terms of those of X(t) and Y(l): (a) Rs,y(t,'t + r) and y.a'(a\. I
(D) Rxr.(t.l r) and xwk t'), and (c) RwxQ,l + r) and s,,y(co). (d) If
* I I
Rx.r'(r) = (W t / r)Sa(W t)
and
R r.r.(r) : (W 2 / n)Sa(W2t)

with constants W2 > Wr, find explicit functions for Rs.s.(t, I + r) and
I wn'(r\.

7.$10. An engineer is working with the function

Rxy(r) - P(l * r)exp(- W'r')
where P > 0 and W > 0 are constants. He suspects that the function may not
be a valid cross-correlation for two jointly stationary processes X(l) and Y(l),
as he has been told. Determine if his suspicions are true. p/l)rl: Find the cross-
power spectrum and see if it satisfies properties (7.3-16) through (7.3-18).1

' A
7.3-fl. wide-sense stationary process X(t) is applied to aD ideal differentiator
having the response )/(l): dX(t'1/dt. The cross-correlation of the input.
output processes is known to be
Rxy(r) - dRyyft)/dt
 (a) Determine ?yy(a) and ?yy(al in terms of the pow€r spectrum gxxkt\ 265
of x(0. IF*2,
(b) Since I xx@t\ must be real, nonnegative, and have even symmetry, what Random
are the properties of 91sy@)? prooess*_

Tjll?- The cross+orrelation of jointly wide-scnse stationary prooesscs X(t) and Y(t\ lff.IlLnr,,*
is assumed to be

Rxv(r) = Bu(r)exP(-Wt)
where I > 0 and W > 0 are constants.
(a'l Find Ry;.(r).
(D) Find I xvki and I yy(a\.

7.I13. Work Problem 7.3-12 for the function

Rxvt) = Bu(r)rexP(-Wr)

7.y14. The cross-power spectrum for random processes X(t) and Y(t) can be written
as

9xv@l= 9xx(a\H(a)
where I xx@\ is the power spectrum of X(l) and f/(c.r) is a function with an
inverse Fourier transform /r(r). Denve expressions for Ryy(r) and Ryy(r) in
terms of Rxx(r) and i(r).

7.It5. Determine the cross-power density spectrum to the croes-

correlation function "orr.rpording

Rxy(r) - u(-r) *.#b +b- 26r-1a-h)'l1

where a > 0 and D > 0 are constants.

7.$16. Find the cross-correlation function corresponding to the cross-power qrec-

trum

etykut=#fr
*7.+1. Again consider the random procbsses of Problem 7.3-3.
(a) Use (6.3-l l) to show that thc cross-correlation function is given by
Ati,
Rxy(t, t + t): + {cos(r,.5r) * E[cos(2O)] cos(2a.tt * aror)

- ltrin(zo)l sin(2aal + c,tr))

where the expectation is with respect to O only.

(D) Find the time average of R1 y(t.t* z) and determine the cross-p+:'u'er
density spectrum I n(a).
7.$1. Fourier transform the first right-side form of (7.5-l), assuming Tris srnall
enough that Rs(r)'does not change appreciably over any time inten'al of
length 7;. Does your result boconre (7.5'2) when Tn + 0?
- 266 1.*2. Show that (7.5-3) is true.
ProU"t ititV.
7.$3. Substitute (7.5-3) into the middle form of (7.5-l) and prove that (7.5-4) is true.
Random Variables,
and Random
7.94. Fourier transform (7.5-6) and show that (7.5-7) is true. (Hint: Make use of the
Signal Principles
table of transforms in Appendix E.)

7.'5. Periodic samples of the autocorrelation function of u'hite norsc N(l) with
period T, are defined by

Rrv,v(k?i,):
l;, f ;3
Find the power spectrum of the DT random process.

7.5{. A discrete-time random sequence X[nl has a DTFT given by

I r,r',(.ia\: lo cos(Qr .."(Y)

rL
where ,46 is a real positive constant. Find the sequence R;.;[nJ by use of the
IDTFT of (7.5-12).

7.57. A random sequence Y[n] is formed by adding the white noise sequence of
Problem 7.5-5 to a one-unit delayed white noise sequence according to

Yln)- il[r]+b1N[n- l]
where b1 is a real constant. Show that

Itt+tllir
: I btd
k=o
RyytkJ k: *l and = I
Io all other k

7.fi. Find the power density spectrum of the sequence I[n] of Problem 7.5-7 by use
of (7.5-l l).

7.$9. Work Problem 7.5-7 except for the sequence Ylnl = N[n] + b-Nln - ml. where
b, is a real constant and z is a positive integer. Use (7.5-l l) to find the power
spectrum of )/[r].
7.$10. Extend Problem 7.5-7 to consider the sequence

Y[n] : N[n]+ bs.\'[n - lJ+ fuNln - 2]

where D2 is a real constant. Find Rr r [k] and the applicable power spectrum
from (7.5-l l).

7.5-ll. The white noise of Problem 7.5-5 is added to a "signal" X that is a random
variable statistically independent of the noise and having a mean value of zero
and a variance oi.. Assume X is a constant u'ith time. Denote the DT sequence
of samples of signal-plus-noise by ll'lnl: X[n] + N[n] where. of course."
Xlnl - X, a constant with n. Find the autocorrelation function of this
sequence and then determine the DTFT of the sequence. lHint: Use the
known sum
 f cx(-jnQ) : 2n D ,(t - nzi).| 267
n=-e cxerrrn T:
Random
7.*12. A DT random scquence X[nJ has an autocorrelation function defined by tbe Processcs-
sequence Rxxlnl-- (l - lnllN) for Inl < N and is zero for lzl > H, with iV a Spectral
positive integer. Find the DTFT of the sequence. Characteristics

7.$13. For the sequenoc Y[n] defined in Example 7.5-3, show that its autocorrelation
sequencre is given by (l).

7.$14. Show that the autocorrelation sequenct defined in Equation (l) of ExampJe
7.5-3 has the power spectrum of equarion (2).

?.315: ..n_eX91k1Eii-*pf9 uT,,ry.: l0 and

'/\f.
''
-""'' ij''!!1-j.5r{
exeept obser.re the effect cf increasing
ffi
rEf.
7.5-16. Extend Examplc 7.5a by showing that the autocorrelation function for the
continuous-time process having the power spectrum

ey11(a)= Kcos2(#) ,*(h)

is
' Rr.x( n:#l[Sa(wyfl+ Sa( w,tt -r) + Sa( wyt + n)l
_(KWxr12)sin(Wx!')
wr4r2 - (wxi2l

7.5-17. Rework Example 7.5-5 except assume the random process defined in Problem
7.5-16. Rework your results for N : l0 and compare to the results for N : 5.
&*
ET
-==t
7.5-18. Show that (7.5.12) can be written as

-q lT

[r] : *f e d?'r, cta

: , .r,x,(-')
R1.1

7.G1. (a) Sketch the power spectrum of (7.6-a) as a function of q@lT.

(r) For what values of ar will ?xy(.) remain above 0.5(-l"ol2) when
T :4.2K (the value of liquid helium at one atmosphere of pressure)?
These values form the region where th0rmal noise is approximately
white in some amplifiers operated at very lou' temperatures, such as a
maser.

7.6-2. For the po\\er spectrum given in Figure 7"'.6-2a. shos,(hat(i.6-6) definas the
correspondir:g band-limited noise autocorrelation function.

7.G3, Show that (7.68) gives the autocorrelation function of the bandpass b,and-
Iimited noir defined bl' Figure 7.6-1a.
268 7.H. A lowpass random X(l) has a continuous power spoctrum g yy@i1and
process

Probability,
J/xl(0) 10. Find the bandwidth W of a lowpass band-limited white-noise
power spectrum having a density /.rl (0) and the same total power as in X(r).
Random Vanables.
and Random
Signal Principlcs 7.G5. Work Problem 7.64 for a bandpass process assunring I xx(on) f 0, where <r.rs

is some convenient frequency about which the spectral components of X(r)

cluster.

-rou - 2W -@o + \tr rrfO - W @o'2W

FTGURE YI.6-6

1.ffi. The power spectrum of a bandpass process l'(t) is shown in Figure P7.6-6.
X(l) is applied to a product device where the second multiplying input
is 3cos(a.ret). Plot.the pou'er spectrum of the device's output 3X(r)cos(a,,ul).

7.G7. Let the "carrier" z{1,cos(abt) in Figure 7.6-5 be modified to add a phase
random variable O so that Y(t): AuX(t)cos(ar6l + O). If (-) is uniformly dis-
tributed on (0, 2n) and is independent of X(r), find Ryr(t, t -i r) and ,/ y.t'(-\
when .t'(l) is wide-sense stationary. :

7.Gt. Assume a stationary bandpass process X(t) is adequately approximated by the

power spectrum

9 tt@l -- Pu(o - crtXro - arg) exp[-(4., - ,r,\2 lbl

+ Pu(-o - rr.tX-rr.r - @o) exp[-(r..r + ol,f,lbl
u'here as, P > 0. and b > 0 are constants. The product y(r) = X(t)cos(arst) is
formed.
(a) Find and sketch the pon'er spectrum of )'(r).
(b) Determine the average power in X(t) and )/(t).
7.Glg. Approximate (7.6-4) as a rectangular lowpass power spectrum with a constant
amplitude -a-ol2 for lrrri < Il' where W is the angular frequencl at which
(7.6-4\ drops to -{-0/4 when I:2K. What average-noise power exists in
the approximate poo'er spectrum if .l"o/2: 5.5(t0-le;? (Hint: Assume the
exponential is adequatell' approximated by the first three terms in its series
representation. )

7.G10. A signal si(r) - 2.3cos(1000r) plus an input noise process .Y,(t) having the
power spectrum shoun in Figure P7.6-l0a are applied to the product device
shown in (D). The ideal lowpass filter (LPF) acts only to remove all spectral
 269
CHA?TEN,7:
Random
Processes-
Spoctral
-rrq) -rmo -500 0 500 l00O 1500 at Characteristics
(c)

r,(4.4(r) /o + Alo(r)

3 coa (lfiDt)

nGUnE P7.5-r0

and noise) that are outside the iraird lc.rl - 500 ratl:s and
components (signal
affect components inside the band.
docs not
(a) Find the output dc level Iz6.
(D) Stetch.the power density spectrum of No(r).
(c) What signai power to .u.rlg. noise power ratio, vi tfi!i,tt)], occurs at
. the output? Note that this circuit acts as a detector of the signal's ampli-
. trde in the presence of noise.
*7.7-1. A complex random process is given by

Z(t) - Ada'
whcrc Q is a random variable with probability density function/o(.) and A is a
omplex constant. Show that the power spectrum of Z(t\ is

,/ zzko') - 274Aiz .fa@)\

"7.7-2. Compute the power spectrum of the complex process of Probtem 6.7'4.
*7.7-3. Ler X(r) and Y(l) be statistically independent p! ocesses with power spectrurrs
I xx@\: 26(ar) + l/U + (arl10)21
and

9vv@\:4/U+ktl2)21
A complex process
Z(t) - [x(0 +jY(t)]exp(jaar)
is formed where a4 is a constant much larger than 10.
(a) Determine the autocorrelation function of Z(t).
(61 Find and sketch the power spectrum of Z(t).