Professional Documents
Culture Documents
R.K.R. Yarlagadda, Analog and Digital Signals and Systems, DOI 10.1007/978-1-4419-0034-0_2, 39
Springer ScienceBusiness Media, LLC 2010
40 2 Convolution and Correlation
Note the symbol (**) for correlation. Correlation isIf xt and yt are orthogonal, the energy and power
related to the convolution. As in autocorrelation, contained in the energy or power signal
the cross-correlation in (2.1.3) is a function of t, the zt xt yt are respectively given by
time shift between the function xt, and the shifted
version of the function ht.
Ez Ex Ey or Pz Px Py : (2:1:8)
2.1.1 Scalar Product and Norm Some of the important properties of the norm are
stated as follows:
The scalar valued function hxt; yti of two signals
xt and yt of the same class of signals, i.e., either 1: kxtk 0 if and only if xt 0; (2:1:9a)
energy or power signals, is defined by
8 R1
>
> xty tdt; 2: kxt ytk kxtk kytk;
>
>
>
> 1 triangular inequality (2:1:9b)
>
>
< energysignals: 2:1:4a
hxt;yti T=2 3: kaxtk jajkxtk: (2:1:9c)
>
> R
>
> lim T1 xty tdt;
>
> T!1
>
> T=2
:
powersignals: 2:1:4b
In (2.1.9c), a is some constant. One measure of
Superscript (*) indicates complex conjugation. Our distance, or dissimilarity, between xt and yt is
discussion will be limited to a subclass of power kxt ytk. A useful inequality is the Schwarzs
signals, namely periodic signals. In that case, inequality given by
assuming that both the time functions have the
same period (2.1.4b) can be written in the symbolic jhxt; ytij kxtkkytk: (2:1:9d)
form as follows:
Z
1
hxT tyT ti xty tdt: (2:1:4c) The two sides are equal when xt or yt is zero or if
T yt axt where a is a scalar to be determined. This
T
can be seen by noting that
Even though our interest is in real functions, for gen-
erality, we have used complex conjugates in the above kxt aytk2 hxt ayt; xt ayti
equations. The norm of the function is defined by hxtxti a hxt; yti
Ex ; energy signals ahxt; yti jaj2 hyt; yti
kxtk hxt; xti1=2 :
Px ; power signals kxtk2 a hxt; yti
(2:1:5) ahxt; yti jaj2 kytk2 :
It gives the energy or power in the given energy or (2:1:10)
the power signal. The two functions, xt and yt,
are orthogonal if
Since a is arbitrary, select
hxt; yti 0: (2:1:6)
2 This definition describes a higher algebra and
hxt; yti
2
kxt aytk kxtk 2 allows us to study the response of a linear time-
kytk2 invariant system in terms of a signal and a system
response to be discussed in Chapter 6. It should be
) kxtk2 kytk2 hxt; yti2 0: emphasized that the end result of the convolution
(2:1:12) operation is a function of time. Coming back to the
sifting property of the impulse functions, consider
Equality exists in (2.1.9d) only if xt ayt 0. the equation given in (2.1.1). Two special cases are
Another possibility is the trivial case being either one of interest.
of the functions or both are equal to zero. Ziemer
Z1
and Tranter (2002) provide important applications
on this important topic. ft dt fadt ada
Correlations in terms of time averages: Cross- 1
(2:2:2a)
correlation and autocorrelation functions can be Z1
expressed in terms of the time average symbols and ft bdbdb dt ft;
1
Z1
Rxh t xtht tdt hxtht ti; (2:1:13a) Z1
dt dt dadt ada dt: (2:2:2b)
1
1
ZT=2
1
RT;xh t xT thT t tdt
T 2.2.1 Properties of the Convolution
T=2 (2:1:13b)
Z Integral
1
xT thT t tdt hxT thT t ti:
T
T 1. Convolution of two functions, x1 t and x2 t,
satisfies the commutative property,
In the early part of this chapter we will deal with
convolution and correlation associated with aper- yt x1 t x2 t x2 t x1 t: (2:2:3)
iodic signals. In the later part we will concentrate on
convolution and correlation with respect to both This equality can be shown by defining a new
periodic and aperiodic signals. Most of the material variable, b t a, in the first integral in (2.2.1)
in this chapter is fairly standard and can be seen in and simplifying the equation.
circuits and systems books. For example, see 2. Convolution operation satisfies the distributive
Ambardar (1995), Carlson (1975), Ziemer and property, i.e.,
Tranter (2002), Simpson and Houts (1971), Peebles
(1980), and others. x1 t x2 t x3 t x1 t x2 t
x1 t x3 t:
2.2 Convolution (2:2:4)
2 1 3 2 3
Z Z1 Zt
dyt d
y 0 t 4 x1 ax2 t ada5 4 x2 a bda5x1 bdb
dt dt
1 1 1
Z1 2 3
dx2 t a Z1 Ztb
x1 a db x1 t x02 t; 4
dt x2 ldl5x1 bdb;
1
1 1
2 3 2 3
dyt d Zt Zt
) x1 t x2 t 4 x2 ldl5x1 t4 x1 bdb5x2 t:
dt dt
dx1 t dx2 t 1 1
x2 t x1 t : (2:2:8)
dt dt
(2:2:6a)
Example 2.2.1 Find the convolution of a function
xt and the unit step function ut and show it is a
Equation (2.2.6a) can be generalized for higher
running integral of xt:
order derivatives. We can then write
Solution: This can be seen from
m m Z1
m d x1 t d yt
x1 t x2 t m
x2 t xt ut xbut bdb
dt dtm
(2:2:6b) 1
Zt
m m d i xi t 1; b5t
y t Note xi t ; xbdb; ut b :
dti 0; b4t
1
(2:2:9) &
m n d m x1 t dn x2 t
x1 t x2 t
dtm dtn
(2:2:6c) 6. Convolution of two delayed functions x1 t t1
d mn yt and x2 t t2 are related to the convolution of
ymn t:
dtmn x1 t and x2 t.
Solution: Convolution of these two functions is properties to simplify the evaluations are illu-
Z1 strated. A few comments are in order before the
yt x1 ax2 t ada examples. First, the convolution yt x1 t
1 x2 t is an integral operation and can use either
Z1 one of the integrals in (2.2.1). Note that yt; 1 <
0:5da 1 0:5da 2 t < 1 is a time function. The expression for
1 the convolution, say at t t0 , will yield a zero
0:3dt a 1 0:7dt a 3da value for those values of t0 over which
x1 b and x2 t0 b do not overlap. The area
Z1 under the product x1 bx2 t0 b, i.e., the integral
0:50:3da 1dt a 1da of this product gives the value of the convolution at
1 t t0 . Sketches of the function x1 b and the time
Z1
reversed and delayed function x2 t0 b on the
0:50:7da 1dt a 3da same figure would be helpful in identifying the lim-
1 its of integration of the product x1 bx2 t0 b.
Z1 As a check, the value of the convolution at end
0:50:3da 2dt a 1da points of each range must match, except in the
1 case of impulses and/or their derivatives in the inte-
Z1 grand of the convolution integral. This is referred to
0:50:7da 2dt a 3da as the consistency check. The following steps can be
1 used to compute the convolution of two functions
0:15dt 0:35dt 4 x1 t and x2 t:
0:15dt 1 0:35dt 5:
New variable Reverse Shift
x2 t!x2 b!x2 b!x2 t b
Notes: If an impulse function is in the integrand of
New variable Multiply the two functions
the form dat b, then use (see (1.4.35), which is x1 t!x1 b !x1 bx2 t b:
Integrate R1
! x1 bx2 t bdb yt:
dat b 1=jajdt b=a: 1
Convolution of two functions exists if the convolu- 1 t a=2
tion integral exists. Existence can be given only in x1 t P and
a a
terms of sufficient conditions. These are related to
1 t b=2
signal energy, area, and one sidedness. It is simple to x2 t P ; b a > 0: (2:3:1)
b b
give examples, where the convolution does not exist.
Some of these are a*a, a*u(t), cos(t)*u(t), eat *eat,
a > 0. Convolution of energy signals and the same- Solution: First
sided signals always exist. In Chapter 4 we will be
discussing Fourier transforms and the transforms Z1
make it convenient to find the convolution. yt x1 t x2 t x1 bx2 t bdb: (2:3:2)
1
x1(t) x2(t)
(a) (b)
x1() x2() x2()
(f) (g)
t > 0 : x2 (t ) and x1 () t > a : x2 (t ) and x1 ()
(h) (i)
t b < a : x2 (t ) and x1 () t b > a : x2 (t ) and x1 ()
(j) (k)
b > a : y (t) = x1 (t)x2 (t) b = a : y (t) = x1 (t)x2 (t)
(l) (m)
Case 3: a5t b: This corresponds to the complete integral operation, which is a smoothing operation.
overlap of the two functions and the functions are Convolution values at end points of each range
shown in Fig. 2.3.1i. The convolution integral and must match (consistency check) as we do not have
the area is any impulse functions or their derivatives in the func-
tions that are convolved. Some of these are dis-
Za cussed below.
1 1 The areas of the two pulses are each equal to 1
yt db ; a5t b: (2:3:5)
ab b and the area of the trapezoid is given by
0
8
>0; t0 tx1 a; tx2 b ) ty tx1 tx2 a b: (2:3:10)
>
>
>
>t
>
> ; 05 t 5 a
>
>
>
ab
> A special case is when a b and the function yt
>
<1 given in Fig. 2.3.1m, a triangle, is
yt ; a t 5b : (2:3:8)
>
> b
>
>
>
> abt
>
> ; b t 5a b t a=2 t a=2 ht ai
>
> ab
>
> P P L : (2:3:11) &
: a a a
0; tab
(a) (b)
x() h()
(c) (d)
h() h ( t ), t 0
(e) (f)
h ( t ) , t < T and x ( ) h ( t ) , t < T and x ( )
(g) (h)
y(t)
(i)
dy1 t 1 d
b: y2 t 1 eat ut
dt a dt
1 d 1 d1 eat
1 eat ut ut
(a) a dt a dt
at at y (t ) = x(t ) * x(t ) 1=a1 eat dt eat ut
x(t ) = e u (t ) x(t ) = e u (t )
1=adt dt eat ut
1=aeat ut: (2:3:25) &
Solution:
tT=2 tT=2
It follows that y2 t 1; 15t51. In this case,
Z Z convolution does not exist. &
yt xada xada
1 1
xt ut T=2 xt ut T=2
xt ut T=2 ut T=2 2.4 Convolution and Moments
hti
xt P : (2:3:27)
T In the examples considered so far, except in the
cases of impulses, convolution is found to be a
The output is the convolution of xt with a pulse smoothing operation. We like to quantify and com-
width of T with unit amplitude and the process is a pare the results of the convolution of nonimpulse
running average. & functions to the Gaussian function. In Section 1.7.1,
the moments associated with probability density
Example 2.3.8 Find the derivative of the running functions were considered.
average of the function in (2.3.27) and express the A useful result can be determined by consid-
function xt in terms of the derivative of yt. ering the center of gravity convolution in terms
Solution: McGillem and Cooper (1991) give an of the centers of gravity of the factors in the
interesting solution for this problem. convolution. First, the moments mn x of a wave-
form xt and its center of gravity Z are, respec-
dut T=2 dut T=2 tively, defined as
y0 t xt
dt dt
T T Z1
xt d t xt d t
2 2 mn x tn xtdt; (2:4:1)
xt T=2 xt T=2 1
2 3
Z1 Z1 Z1 Average signal power
4t Signal-to-noise ratio :
m1 y tytdt glht ldl5dt Noise power; s2n
1 1 1 (2:4:8)
2 3
Z1 Z1
gl4 tht ldt5dl: Example 2.4.1 Verify the result is true in (2.4.7)
using the functions
1 1
Z1 m1 g m2 g
gldl m1 gm0 h m1 hm0 g: (2:4:4) Zg 1; s2g Z2g 1;
m0 g m0 g
1
s2h 1 note gt ht;
From the area property, it follows that m0 y
yt gt ht tet utsee Example 2:3:4:
m0 gm0 h. The center of gravity is
Z1 Z1
t
m1 y m1 g m1 h m0 y te dt 1; m1 y t2 et dt 2;
) Zy Zg Zh : (2:4:5)
m0 y m0 g m0 h 0 0
Z1
Consider the expression for the squares of the m2 y t3 et dt 6;
spread of yt in terms of the squares of the spreads 0
of gt and ht. The derivation is rather long and m1 y m2 y
Zy 2; s2y Z2y 2 )
only results are presented. m0 y m0 y
That is, the variance of y is equal to the sum of the In Chapter 10, we will make use of this in quantiza-
variances of the two factors. It also verifies that tion methods, wherein A and SNR are given and
convolution is a broadening operation for pulses. determine s2n . This, in turn, provides the informa-
Noting that if gt and ht are probability density tion on the size of the error that can be tolerated.
functions then (2.4.7) is valid. In communications Notes: For readers interested in independent random
theory we are faced with a signal, say gt is cor- variables, the probability density function of a sum of
rupted by a noise nt with the variance, s2n . The two independent random variables is the convolution
signal-to-noise ratio (SNR) is given by of the density functions of the two factors of the
52 2 Convolution and Correlation
convolution, and the variance of the sum of the two Solution: yt is a triangular function (see Example
random variables equals the sum of their variances. 2.3.1) given by
For a detailed discussion on this, see Peebles (2001).& 1 ht ai
yt L : (2:4:12)
a a
2.4.1 Repeated Convolution and the The mean values of the two rectangular pulses are
Central Limit Theorem a/2 (see Section 1.7). The mean value of yt is
2a=2 a. The variance of each of the rectangular
Convolution operation is an integral operation, which pulses is
is a smoothing operation. In Example 2.3.1, we have
s2i m2 m21 a2 =12; i 1; 2: (2:4:13a)
considered the special case of the convolution of two
identical rectangular pulses and the convolution of The variance is given by s2y s21 s22 a2 =6. The
these two pulses resulted in a triangular pulse (see Gaussian approximation is
Fig. 2.3.1m). The discontinuities in the functions
being convolved are not there in the convolved signal. 1 2 2
ytjN2 p eta =a =3 : (2:4:13b)
As more and more pulse functions convolve, the resul- pa2 =3
tant functions become smoother and smoother.
Repeated convolution begins to take on the bell- This Gaussian and the triangle functions are sym-
shaped Gaussian function. The generalized version metric around a. They are sketched in Fig. 2.4.1.
of this phenomenon is called the central limit theorem. Even with N 2, we have a good approximation. &
It is commonly presented in terms of probability den-
sity functions. In simple terms, it states that if we
convolve N functions and one function does not dom-
inate the others, then the convolution of the N func-
tions approaches a Gaussian function as N ! 1. In
the general form of the central limit theorem, the
means and variances of the individual functions that
are convolved are related to the mean and the variance
of the Gaussian function (see Peebles (2001)).
Given xi t; i 1; 2; :::; N; the convolution of
Fig. 2.4.1 Triangle function yt in (2.4.12) and the Gaus-
these functions is
sian function in (2.4.13b)
yt x1 t x2 t ::: xN t: (2:4:9)
Example 2.4.3 In Example 2.4.1 we considered two
The function yt can be approximated using m0 N , identically exponentially decaying functions:
the sum of the individual means of the functions, x1 t et ut x2 t. The convolution of these
and s2N the sum of the individual variances by two functions is given by y2 t tet ut. Approx-
imate this function using the Gaussian function.
1 2 2
yt q etm0 N =2sN : (2:4:10) Solution: The Gaussian function approximations
2ps2N of yn t, considering n 2 and for n large, are,
respectively, given below. Note that m0 y 2.
Example 2.4.2 Illustrate the effects of convolution
1 2
and compare yt to a Gaussian function by con- y2 t p et2 =22 ;
sidering the convolution 2p2
(2:4:14)
1 2
yt x1 t x2 t; yn t p etn =2n :
2pn
(2:4:11)
1 t a=2 For sketches of these functions for various values of
xi t P ; i 1; 2:
a a n, see Ambardar (1995). &
2.4 Convolution and Moments 53
Amplitude scaling:
Commutative:
x1 t x2 t x2 t x1 t:
Distributive:
x1 t x2 t x3 t x1 t x2 t x1 t x3 t:
Associative:
x1 t x2 t x3 t x1 t x2 t x3 t:
Delay:
x1 t t1 x2 t t2 x1 t t2 x2 t t1 yt t1 t2 :
Impulse response:
xt dt xt:
Derivatives:
m n
x1 t x02 t x01 t x2 t y0 t; x1 t x2 t ymn t:
Step response:
Rt
yt xt ut 1 xada; y0 t xt dt xt:
Area:
R1
Ax1 t x2 t Ayt; where Axt 1 xtdt:
Duration:
tx1 tx2 ty :
Symmetry:
Time scaling:
R1
X
1 Solution: yt ht xT t eab coso0 t b
yt ht dT t ht dt kT ydb 0
k1
X
1 Z1
ht dt kT: (2:5:2) eab coso0 t y coso0 b
k1 0
sino0 t y sino0 bdb
Noting that ht dt kT ht kT, it follows 2 1 3
Z
that 4 eab coso0 bdb5 coso0 t y
0
X
1 2 3
Z1
yt ht kT yT t: (2:5:3)
4 eab sino0 bdb5 sino0 t y: (2:5:5)
k1
0
Figure 2.5.1a,b gives the sketches of the functions Using the identities given below (see (2.5.7 a, b, and
dT t and ht. The sketches for the convolution are c.)), (2.5.5) can be simplified.
shown in Fig. 2.5.1c,d. In the first case, there were
no overlaps, whereas in the second case there are yt a=a2 o20 coso0 t y
overlaps. & o0 =a2 o20 sino0 t y
Example 2.5.2 Derive an expression for the convo- 1
lution yt ht xT t, q coso0 t y tan1 o0 =a
a2 o20 (2:5:6)
k0
xT t xt nT and hT t ht nT;
n1 n1
yk ht; o0 2p=T: (2:5:8b)
(2:5:10a)
xT t; t0 t < t0 T
2.5.2 Convolution of Two Periodic xt ;
0; otherwise
Functions
hT t; t0 t < t0 T
ht (2:5:10b)
0; otherwise:
In Section 1.5 energy and power signals were con-
sidered. The energy in a periodic function is infinity Note that the time-limited functions, xt and ht,
and its average power is finite. One period of a are defined from the periodic functions xT t and
periodic function has all its information. In the hT t. Using (2.5.10b) the periodic convolution is
same vein, the average convolution is a useful mea-
Z
sure of periodic convolution. Such averaging pro- 1
cess is called periodic or cyclic convolution. The con- yT t xT ahT t ada
T
volution of two periodic functions with different T
1 X 1
ZT=2
xt ht nT; 1
T n1 yt lim xaht ada: (2:5:14)
T!1 T
(2:5:11a) T=2
yT t xT t
hT t
1 X 1 This reduces to the periodic convolution if
yt nT;yt xt ht: (2:5:11b) xt and ht are periodic with the same period.
T n1
yT (t ) (a) yT (t )
Z1
Rxh t xtht tdt (2:6:1a)
1
2.6.2 Cross-Correlation and Convolution
Cross-correlation: Periodic:
Z The cross-correlation function is related to the con-
1
RT; xht xT thT t tdt (2:6:1b) volution. From (2.6.3) we have
T
T
Rxh t xt ht xt ht; (2:6:4a)
Autocorrelation: Aperiodic:
Rhx t ht xt ht xt: (2:6:4b)
Z1
Rx t xtxt tdt (2:6:1c) Equation (2.6.4a) can be seen by first rewriting the
1 first integral in (2.6.3) using a new variable t a,
and then simplifying it. That is,
Autocorrelation: Periodic:
Z Z1 Z1
1 Rxh t xtht tdt xaht ada
RT;x t xT txT t tdt: (2:6:1d)
T 1 1
T
xt ht: (2:6:4c)
Notes: Cross- and autocorrelations of periodic
functions and random signals are referred to as Equation (2.6.4b) can be similarly shown. Noting
average periodic cross- and autocorrelation functions. the explicit relation between correlation and convo-
In the case of random or noise signals, the average lution, many of the convolution properties are
cross-correlation function is defined by applicable to the correlation. To compute the cross
58 2 Convolution and Correlation
8
> 0; t T The significance of rxh t can be seen by consider-
>
>
<1h
>
atT
i ing some extreme cases. When xt aht; a > 0,
Rxh t a 1 e ; T < t T we have the correlation coefficient rxh t 1. In the
>
> case of xt aht; a < 0 and rxh t 1. In
>
> 1
: eaT eaT eat ; t > T
a communication theory, we will be interested in
(2:6:11c) & signals that are corrupted by noise, usually identi-
fied by nt, which can be defined only in statis-
tical terms. In the following, we will consider the
analysis without going through statistical analysis.
2.6.4 Quantitative Measures Noise signal nt is assumed to have a zero average
of Cross-Correlation value. That is,
Rxh t Rxh t
rxh t s
1 1 p
; Cross-correlation function can be used to compare
R R Ex Eh
2
x tdt 2
h tdt two signals. The signals xt and ht are uncorre-
1 1 lated if the average cross-correlation satisfies the
(2:6:12a) relation
(2:6:15)
It should be noted that the case of xt ht, the
correlation coefficient reduces to
Example 2.6.2 If the signals x(t) and a zero average
Rx t noise signal n(t) are uncorrelated, then show
rxx t : (2:6:13)
Rx 0
ZT=2
Correlation measures are very useful in statistical 1
lim xtnt tdt 0 for all t: (2:6:16)
analysis. See Yates and Goodman (1999), Cooper T!1 T
T=2
and McGillem (1999) and others.
60 2 Convolution and Correlation
Solution: Using (2.6.14) and (2.6.15), we have correlation detector (or receiver) shown in
Fig. 2.6.2. The received signals yi t are assumed
ZT=2
1 to be of the form in (2.6.19). Decide which signal
lim xtnt tdt
T!1 T has been transmitted using the cross-correlation
T=2 function.
2 32 3
ZT=2 ZT=2
6 1 76 1 7 yi t xi t noise; i 1 or 2: (2:6:19)
4 lim xtdt54 lim nt tdt5 0:
T!1 T T!1 T
T=2 T=2
Solution: Let the transmitted signal be x1 t. Using
(2:6:17) the top path in Fig. 2.6.2, we have
Example 2.6.4 Derive the expressions for the cross- t1
correlation Rxh t and Rhx t assuming xt Pt :5; ht tP ; Rxh t
2
xt et ut; ht e2t ut: Z1
Solution: Using the expression in (2.6.3), we have xtht tdt: (2:6:23)
1
Z1
Rxh t xa thada Solution: See Fig. 2.6.4c for ht t for an arbitrary
1 t. The function ht t starts at t t and ends at
Z1 t 2 t. As t varies from 1 to 1, there are five
eat e2a ua tuada: (2:6:22a) possible regions we need to consider. These are
1 sketched in Fig. 2.6.4 d,e,f,g,h. In each of these cases
both the functions are sketched in the same figure,
Consider the following and then the corresponding which allows us to find the regions of overlap. The
correlations: regions of overlap are listed in Table 2.6.1.
Z2t
t2
Rxh t t tdt ttt2t
t0
2
0
Example 2.6.5 Derive the crosscorrelation Rxh t Case 5: 2 < t: See Fig. 2.6.4 h. There is no overlap
for the following functions: and
62 2 Convolution and Correlation
(c) (d)
(e) (f)
(h)
(g)
Rxh ( )
(i)
n h t io hti
AC P TL : (2:7:11b)
T T
1 jtj; 0 jtj 1 Example 2.7.3 Find the autocorrelation of the func-
Rx t Rx t Lt:
0; Otherwise tion yt coso0 tPt=T.
(2:7:11a) Solution:
Z1 h t i htti
This is sketched in Fig. 2.7.2b indicating that there is
Ry t P P coso0 tcoso0 ttdt
correlation for jtj < 1 and no correlation for T T
jtj 1. The peak value of the autocorrelation is 1
Rx()
(a) (b)
If o0 is large, Ry t in (2.7.13) can be approximated Note that the periods of the functions,
by the first term and xT t and hT t, are assumed to be the same and
the constant (1/T) before the integrals in (2.8.1a
hti and b). If they have different periods, computation
1
RY t TL coso0 t: (2:7:15) of (2.8.1a) is difficult and these cases will not be
2 T
discussed here. Many of the cross-correlation and
AC function properties derived earlier for the aper-
The envelope of the autocorrelation function in iodic case apply for the periodic functions with
(2.7.16) is a triangular function, which follows some modifications. Note that
since the correlation of the two identical rectangular
functions is a triangular function. Noting that the
cosine function oscillates between 1, the envelope RT;xh t RT;hx t;
of the autocorrelation function in (2.7.15) is shown RT;x 0 RT;h 0 2RT;xh t: (2:8:2)
in Fig. 2.7.3. &
tZ0 T Solution: Z
1 1
RT;xh t xT thT t tdt; hT t t a: RT;xT;1 t X2s 0dt X2s 0; (2:8:5a)
T T
t0 T
X
1
Z
ht t nT: (2:8:3b) 1
n1
RT;xT;2 t xT;2 txT;2 t tdt
T
T
The expression for periodic convolution is given in Z
c2 k
terms of aperiodic convolution and cosko0 tdt
2T
T
Z
1 X 1 c2 k
RT;xh t Rxh t nT; cosko0 2t t 2ykdt
T n1 2T
T
tZ0 T
ZT
Rxh t nT xtht t nT: (2:8:3c) c2 k cosko0 t c2 k cosko0 t
dt :
2T 2
t0 0
(2:8:5b)
The details of the derivation are left as an exercise.
Copies of Rxh t will overlap if the width of Rxh t is Note that the integral of a cosine function over any
wider than T. integer number of periods is zero.
Example 2.8.1 Give the lower bound on the period T b. The cross-correlation of a constant and a cosine
so that there are no overlaps in the cross-correlation function over one period is zero. Also note that
of the functions xT t and hT t given below. See the two functions are orthogonal. That is
Example 2.6.5. hxT1 t; xT2 ti 0. &
Note 1X 1
P X2s 0 c2 k: (2:8:10)
2 k1
Z
1
xT;1 txT;2 t tdt The difference between the total power and the dc
T
T power is the variance and is given by
Z
1
ckcm cosk mo0 t mo0 t yk
2T 1X 1
T
Z Variance c2 k: (2:8:11) &
1 2 k1
ymdt hkhm cosk mo0 t mo0 t
2T
T Example 2.8.4 Consider the corrupted signal
yk ymdt 0: yt xt nt, where nt is assumed to be
noise. Assuming the signal xt and noise nt are
Similarly the fourth term in (2.8.6) goes to zero. uncorrelated, derive an expression for the autocor-
From the last example, relation function of yt.
c2 k c2 m Solution:
RT;x t cosko0 t cosmo0 t; k 6 m:
2 2
(2:8:7) & ZT=2
1
Ryy t lim ytyt tdt
T!1 T
T=2
These results can be generalized using the last two ZT=2
examples and the autocorrelation of a periodic 1
lim xt ntxt tnt tdt;
function xT t is given as follows: T!1 T
T=2
X
1 ZT=2 ZT=2
xT t Xs 0 ck cosko0 t yk; (2:8:8) lim xtxt tdt lim xtnt tdt
k1 T!1 T!1
T=2 T=2
T=2
R T=2
R
1X 1
lim ntxt tdt lim xtxt tdt:
) RT;x t X2s 0 c2 kcosko0 t;o0 2p=T: T!1 T!1
2 k1 T=2 T=2
(2:8:9) (2:8:12)
2.7.3 Show the identity these constants. Suppose we are interested in deter-
mining the period T from these two sketches, which
ACxt t0 ACxt: function is better, the given function or its autocor-
relation? Why?
2.7.4 Derive the AC function step by step for the
2.8.2 Let yT t A xT t; A constant. Repeat
function xt coso0 tPt=T. the last problem, except for the plots.
Use the integral formula by assuming
o0 p and T 4. Verify the results in Example 2.8.3 a. Show that the following functions are
2.7.3 using the information provided in this pro- orthogonal over a period:
blem. Give the appropriate bounds. xT t coso0 t y; yt A b. Show the func-
tions xt Pt; yt t are orthogonal.
2.7.5 Show that the autocorrelations of the function
x2 t eat ut for a 0 do not exist. 2.8.4 Consider the signal zt xt yt. Show
2.8.1 a. Derive the time-average periodic autocorre- that the AC of this function is given by
lation function Rx; T t for the following periodic
function using the integral formula. Rz t Rx t Ry t Rxy t Ryx t: