438book2012 Vol1 PDF

Digital Signal Processing in a Nutshell (Volume I)
Jan Allebach, Krithika Chandrasekar
ID: 12395631
www.lulu.com
5 800077 300189
Digital Signal Processing in a

Nutshell (Volume I)
Jan P Allebach
Krithika Chandrasekar
Contents
1 Signals
1.1 Chapter Outline . . . . . . . . . .
1.2 Signal Types . . . . . . . . . . . .
1.3 Signal Characteristics . . . . . . .
1.4 Signal Transformations . . . . . . .
1.5 Special Signals . . . . . . . . . . .
1.6 Complex Variables . . . . . . . . .
1.7 Singularity functions . . . . . . . .
1.8 Comb and Replication Operations
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
7
10
14
17
23
27
2 Systems
2.1 Chapter Outline . .
2.2 Systems . . . . . . .
2.3 System Properties .
2.4 Convolution . . . . .
2.5 Frequency Response
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
29
29
29
32
36
44
3 Fourier Analysis
3.1 Chapter Outline . . . . . . . . . . . . .
3.2 Continuous-Time Fourier Series (CTFS)
3.3 Continuous-Time Fourier Transform . .
3.4 Discrete-Time Fourier Transform . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
51
51
51
57
70
4 Sampling
4.1 Chapter Outline . . . . . . . . . . .
4.2 Analysis of Sampling . . . . . . . . .
4.3 Relation between CTFT and DTFT
4.4 Sampling Rate Conversion . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
79
79
79
87
90
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
5 Z Transform (ZT)
5.1 Chapter Outline . . . . . . . . . . . . . . . . . . . .
5.2 Derivation of the Z Transform . . . . . . . . . . . . .
5.3 Convergence of the ZT . . . . . . . . . . . . . . . . .
5.4 ZT Properties and Pairs . . . . . . . . . . . . . . . .
5.5 ZT,Linear Constant Coefficient Difference Equations
5.6 Inverse Z Transform . . . . . . . . . . . . . . . . . .
5.7 General Form of Response of LTI Systems . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
99
99
99
103
107
108
114
124
6 Discrete Fourier Transform (DFT)

6.1 Chapter Outline . . . . . . . . . . . . . .
6.2 Derivation of the DFT . . . . . . . . . . .
6.3 DFT Properties and Pairs . . . . . . . . .
6.4 Spectral Analysis Via the DFT . . . . . .
6.5 Fast Fourier Transform (FFT) Algorithm
6.6 Periodic Convolution . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
129
129
129
132
134
137
142
7 Digital Filter Design

7.1 Chapter Outline . . . . . . . . . . . . .
7.2 Overview . . . . . . . . . . . . . . . . .
7.3 Finite Impulse Response Filter Design .
7.4 Infinite Impulse Response Filter Design
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
147
147
147
154
165
8 Random Signals
8.1 Chapter Outline . . . . . . . .
8.2 Single Random Variable . . . .
8.3 Two Random Variables . . . .
8.4 Sequences of Random Variables
8.5 Estimating Distributions . . . .
8.6 Filtering of Random Sequences
8.7 Estimating Correlation . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
177
177
177
181
184
191
195
203
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Chapter 1
Signals
1.1
Chapter Outline
In this chapter, we will discuss:
1. Signal Types
2. Signal Characteristics
3. Signal Transformations
4. Special Signals
5. Complex Variables
6. Singularity Functions
7. Comb and Replication Operators
CHAPTER 1. SIGNALS
1.2
Signal Types
There are three different types of signals:

1. Continuous-time (CT)
2. Discrete-time (DT)
3. Digital (discrete-time and discrete-amplitude)
The different types of signals are shown in Figure 1.2(a) below.
Figure 1.2(a) Different types of signals.
Comments
The CT signal need not be continuous in amplitude
The DT signal is undefined between sampling instances
Levels need not be uniformly spaced in a digital signal
We make no distinction between discrete-time and digital signals in
theory
The independent variable need not be time
1.3. SIGNAL CHARACTERISTICS
x(n) may or may not be sampled from an analog waveform, i.e.

xd (n) = xa (nT ) T- sampling interval
Throughout this book, we will use the subscripts d and a only
when necessary for clarity.
Equivalent Representations for DT signals
Figure 1.2(b) Equivalent representations for DT signals.
1.3
Signal Characteristics
Signals can be:

1. Deterministic or random
2. Periodic or aperiodic
For periodic signals, x(t) = x(t + nT ), for some fixed T and for all
integers n and times t.
3. Right-sided, left-sided, or two-sided
For right-sided signals, x(t) 6= 0 only when t t0 .
If t0 0, the signal is causal.
For left-sided signals, x(t) 6= 0 only when t t0 .
If t0 0, the signal is anticausal.
For 2-sided or mixed causal signals, x(t) 6= 0 for some t < 0, and
x(t) 6= 0 for some t > 0.
4. Finite or infinite duration
For finite duration signals, x(t) 6= 0 only for < t1 t t2 < .
CHAPTER 1. SIGNALS
Metrics
Metrics
Energy
CT signal
R
Ex = |x(t)|2 dt
Power
1
Px = lim
T 2T
DT signal
X
|x(n)|2
Ex =
n
|x(t)|2 dt
Magnitude
Area
Mx = maxt |x(t)|
R
Ax = |x(t)|dt
Average value
1
xavg = lim
T 2T
N
X
1
|x(n)|2
N 2N + 1
Px = lim
n=N
Mx = maxn |x(n)|
X
Ax =
|x(n)|
n
x(t)dt
T
N
X
1
x(n)
N 2N + 1
xavg = lim
n=N
Comments
If signal is periodic, we need only average over one period, e.g. if
x(n)=x(n+N) for all n
Px =
1
N
n0 +N
X1
|x(n)|2 , for any n0
n=n0
root-mean-square (rms) value

p
xrms = Px
Examples
Example 1
Figure 1.3(a) Signal for Example 1.

(t+1)/2
e
, t > 1
x(t) =
0,
t < 1
1.3. SIGNAL CHARACTERISTICS
The signal is:

1. deterministic
2. aperiodic
3. right-sided
4. mixed causal
Metrics
1. Ex =
Ex =
1
0
|e(t+1)/2 |2 dt,
es ds =
let s = t + 1
es |0
=1
2. Px = 0
3. Mx = 1
4. Ax = 2
5. xavg = 0
Example 2
Figure 1.3(b) Signal for Example 2.

x(n) = A cos(n/4)
The signal is:
1. deterministic
2. periodic (N = 8)
10
CHAPTER 1. SIGNALS
3. 2-sided
4. mixed causal
Metrics
1. Ex =
2. Px =
7
1X
|A cos(n/4)|2
8 n=0
7
A2 X
[1 + cos(n/2)]
16 n=0
A2
Px =
2
Px =
A
3. xrms =
2
4. Mx = A
5. Ax is undefined
6. xavg = 0
1.4
Signal Transformations
Continous-Time Case
Consider
1. Reflection
y(t) = x(t)
1.4. SIGNAL TRANSFORMATIONS
2. Shifting
y(t) = x(t t0 )
3. Scaling
y(t) = x( at )
4. Scaling and Shifting

Example 1
y(t) = x( at t0 )
What does this signal look like?
center of pulse:
t
a
t0 = 0 t = at0
11
12
CHAPTER 1. SIGNALS
left edge of pulse:
t
a
right edge of pulse:
1
2
t0 = 21
t0 =
t = a(t0 12 )
t
a
t = a(t0 + 21 )
Example 2
y(t) = x
tt0
a
What does this signal look like?

center: (t t0 )/a = 0 t = t0
left edge of pulse: (t t0 )/a = 1/2 t = t0 a/2
right edge of pulse: (t t0 )/a = 1/2 t = t0 + a/2
Discrete-Time Case
Consider
1.4. SIGNAL TRANSFORMATIONS

1. Reflection
y(n) = x(n)
2. Shifting
y(n) = x(n n0 )
Note that n0 must be an integer. Non-integer delays can only be

defined in the context of a CT signal, and are implemented by
interpolation.
3. Scaling
a. downsampler
y(n) = x(Dn), D Z
b. upsampler

x(n/D), n/D Z, D Z
y(n) =
0,
else
13
14
1.5
CHAPTER 1. SIGNALS
Special Signals
Unit step
a. CT case

1,
u(t) =
0,
t>0
t<0
b. DT case

1, n 0
u(n) =
0, else
Rectangle

1, |t| < 1/2
rect(t) =
0, |t| > 1/2
1.5. SPECIAL SIGNALS
Triangle

1 |t|, |t| 1
(t) =
0,
else
Sinusoids
a. CT
xa (t) = A sin(a t + )
where A is the amplitude
a is the frequency
is the phase
Analog frequency

radians
cycles
a
= 2fa
sec
sec
15
16
CHAPTER 1. SIGNALS
Period

sec
1
T
=
cycle
fa
b. DT
xd (n) = A sin(d n + )
digital
frequency

radians
cycles
d
= 2
sample
sample
Comments
1. Depending on , xd (n) may not look like a sinusoid
2. Digital frequencies 1 = 0 and 2 = 0 + 2k are equivalent

x2 (n) = A sin(2 n + )
= A sin[(0 + 2k)n + ]
= x1 (n), for all n
3. xd (n) will be periodic if and only if d = 2(p/q) where p and q are
integers. In this case, the period is q.
1.6. COMPLEX VARIABLES

Sinc
sinc(t) =
1.6
sin(t)
(t)
Complex Variables
Definition
Cartesian coordinates
z = (x, y)
Polar coordinates
z = R
p
R = x2 + y 2
x = R cos ()
17
18
CHAPTER 1. SIGNALS
= arctan
y
x
y = R sin ()
Alternate representation
define j = (0, 1) = 1 (/2)
then z = x + jy
Additional notation
Re {z} = x
Im {z} = y
|z| = R
Algebraic Operations
Addition (Cartesian Coordinates)
z1 = x1 + jy1
z2 = x2 + jy2
z1 + z2 = (x1 + x2 ) + j(y1 + y2 )

Multiplication (Polar Coordinates)
z1 = R1 1
z2 = R2 2
z1 z2 = R1 R2 (1 + 2 )
Special Examples
1. 1 = (1, 0) = 10
2. 1 = (1, 0) = 1180
3. j 2 = (190)2 = 1180 = 1
19
20
CHAPTER 1. SIGNALS
Complex Conjugate
z = x jy = R ()
Some useful identities

1
Re {z} = [z + z ]
2
1
Im {z} = [z z ]
j2
|z| = zz
Complex Exponential
Taylor series for exponential function of real variable x
ex =
k=0
X
xk
k!
Replace x by j
j2
j3
j4
j5
+
+
+
2!
3!
4!
5!
2
4
3
5
= (1
+
...) + j(
+
...)
2!
4!
3!
5!
= cos + j sin
ej = 1 + j +
ej
cos2 + sin2 = 1
sin
= arctan
=
cos
|ej | =
Alternate form for polar coordinate representation of complex number

z = R = Rej
Multiplication of two complex numbers can be done using rules for

multiplication of exponentials:
z1 z2 = (R1 ej1 )(R2 ej2 ) = R1 R2 ej(1 +2 )
Complex exponential signal

CT x(t) = Aej(t)
(t) = a t + instantaneous
phase

rad
d(t)
= a
instantaneous phasor velocity
dt
sec
21
22
CHAPTER 1. SIGNALS
DT x(n) = Aej(d n+)

d is the phasor velocity in

radians .
sample
Example
d = /3 = /6
Aliasing
1. It refers to the ability of one frequency to mimic another
2. For CT case, ej1 t = ej2 t for all t 1 = 2
3. In DT, ej1 n = ej2 n if 2 = 1 + 2k
1.7. SINGULARITY FUNCTIONS

Consider
7
5
1 =
2 =
6
6
x1 (n) = ej1 n x2 (n) = ej2 n
In general, if 1 = + and 2 = ( ), then

x1 (n) = ej1 n
= ej(+)n
= ej[(+)n2n]
= ej[(+)n]
= ej()n
= x2 (n)
1.7
Singularity functions
CT Impulse Function
1
t
Let (t) = rect( )
Z
Note that
(t)dt = 1.
What happens as 0?
1 > 2 > 3
23
24
CHAPTER 1. SIGNALS
In the limit, we obtain

(t) = lim (t)
0

0, t 6= 0
(t) =
, t = 0
and
Z
(t)dt = 1
We will use impulses in three ways:

1. to sample signals
2. as a way to decompose signals into elementary components
3. as a way to characterize the response of a class of systems to an
arbitrary input signal
Sifting property
Consider a signal x(t) multiplied by (t t0 )for some fixed t0 :
1.7. SINGULARITY FUNCTIONS

From the mean value theorem of calculus,
Z
x(t) (t t0 )dt = x()
for some which satisfies

t0
t0 +
As 0, t0 and x() x(t0 )

So we have
Z
x(t)(t t0 )dt = x(t0 )
provided x(t) is continuous at t0 .
Equivalence
Based on sifting property,
x(t)(t t0 ) x(t0 )(t t0 )
Indefinite integral of (t)

Z
( )d
Let u (t) =
25
26
CHAPTER 1. SIGNALS
Let 0,
Z t
u(t) =
( )d
When we define the unit step function, we did not specify its value at
t=0. In this case, u(0) = 0.5
More general form of sifting property

Z

x( )( t0 )d =
x(t0 ), a < t0 < b

0,
t0 < a or b < t0
DT Impulse Function (unit sample function)

(n) =
1, n = 0
0, else
1.8. COMB AND REPLICATION OPERATIONS
27
Sifting Property
n2
X

x(n)(n n0 ) =
n=n1
x(n0 ), n1 n0 n2
0,
else
Equivalence
x(n)(n n0 ) = x(n0 )(n n0 )
Indefinite sum of (n)
u(n) =
n
X
(m)
m=
1.8
Comb and Replication Operations
Sifting property yields a single sample at t0 . Consider multiplying x(t) by

an entire train of impulses:
Define
xs (t) = combT [x(t)]
X
= x(t)
(t nT )
n
x(t)(t nT )
X
n
x(nT )(t nT )
28
CHAPTER 1. SIGNALS
Area of each impulse = value of sample there.

The replication operator similarly provides a compact way to express
periodic signals:
xp (t) = repT [x0 (t)] =
x0 (t nT )
Note that x0 (t) = xp (t)rect

t
T
Chapter 2
Systems
2.1
Chapter Outline

1. System Properties
2. Convolution
3. Frequency Response
2.2
Systems
A system is a mapping which assigns to each signal x(t) in the input

domain a unique signal y(t) in the output range.
29
30
CHAPTER 2. SYSTEMS
Examples
Z
t+1/2
1. CT y(t) =
x( )d
t1/2
let x(t) = sin(2f t)

Z t+1/2
y(t) =
sin(2f t)dt
t1/2
1
t+1/2
cos(2f t)|t1/2
=
2f
1
=
{cos[2f (t + 1/2)] cos[2f (t 1/2)]}
2f
use cos( ) = cos cos sin sin
1
{[cos(2f t) cos(f ) sin(2f t) sin(f )]
2f
[cos(2f t) cos(f ) + sin(2f t)sin(f )]}
y(t) =
sin(f )
sin(2f t)
f
= sinc(f )sin(2f t)
y(t) =
Look at particular values of f

f = 0 : x(t) 0
y(t) 0
f = 1 : x(t) = sin(2t)
sinc(1) = 0
y(t) 0
2.2. SYSTEMS
2. DT y(n) =
31
1
[x(n) + x(n 1) + x(n 2)]
3
Effect of filter on output:

1. widened,smoothed, or smeared each pulse
2. delayed pulse train by one sample time
32
CHAPTER 2. SYSTEMS
2.3
System Properties
Notation:
y(t) = S[x(t)]
A. Linearity
def. A system S is linear (L) if for any two inputs x1 (t) and x2 (t) and any
two constants a1 and a2 , it satisfies:
S[a1 x1 (t) + a2 x2 (t)] = a1 S[x1 (t)] + a2 S[x2 (t)]
Special cases:
1. homogeneity (let a2 = 0)
S[a1 x1 (t)] = a1 S[x1 (t)]
2. superposition (let a1 = a2 = 1)
S[x1 (t) + x2 (t)] = S[x1 (t)] + S[x2 (t)]
Examples
Z
t+1/2
y(t) =
x( )d
t1/2
let x3 (t) = a1 x1 (t) + a2 x2 (t)

Z
t+1/2
y3 (t) =
x3 ( )d
t1/2
t+1/2
[a1 x1 ( ) + a2 x2 ( )]d
t1/2
t+1/2
= a1
t+1/2
x1 ( )d + a2
t1/2
x2 ( )d
t1/2
= a1 y1 (t) + a2 y2 (t)
system is linear
We can similarly show that
y(n) = 13 [x(n) + x(n 1) + x(n 2)] is linear.
2.3. SYSTEM PROPERTIES
33
Consider two additional examples:

3. y(n) = nx(n)
let x3 (n) = a1 x1 (n) + a2 x2 (n)
y3 (n) = nx3 (n)
= (a1 nx1 (n) + a2 nx2 (n))
= a1 y1 (n) + a2 y2 (n)
system is linear
1, x(t) < 1
x(t), 1 x(t) 1
y(t) =
1,
1 < x(t)
Suspect that system is nonlinear - find a counterexample.

x1 (t) 1 y1 (t) 1
x2 (t)
1
2
y2 (t)
1
2
x3 (t) = x1 (t) + x2 (t) =
3
2
y3 (t) 1 6=
3
2
system is not linear

Because it is memoryless, this system is completely described by a curve
relating input to output at each time t:
34
CHAPTER 2. SYSTEMS
Importance of linearity: We can represent response to a complex input in

terms of responses to very simple inputs.
B. Time Invariance
def. A system is time-invariant if delaying the input results in only an
identical delay in the output, i.e.
if y1 (t) = S[x1 (t)]
and y2 (t) = S[x1 (t t0 )]
then y2 (t) = y1 (t t0 )
Examples
1. y(n) =
1
[x(n) + x(n 1) + x(n 2)]
3
assume
1
[x1 (n) + x1 (n 1) + x1 (n 2)]
3
let x2 (n) = x1 (n n0 )
y1 (n) =
1
[x2 (n) + x2 (n 1) + x3 (n 2)]
3
1
= [x1 (n n0 ) + x1 (n 1 n0 ) + x1 (n 2 n0 )]
3
1
= [x1 (n n0 ) + x1 (n n0 1) + x1 (n n0 2)]
3
= y1 (n n0 )
y2 (n) =
system is TI.
We can similarly show that
2.3. SYSTEM PROPERTIES

Z
35
t+1/2
2. y(t) =
x( )d is TI
t1/2
3. y(n) = nx(n)
assume y1 (n) = nx1 (n)
let x2 (n) = x1 (n n0 )
y2 (n) = nx2 (n)
= nx1 (n n0 )
6= (n n0 )x1 (n n0 )
system is not TI.
C. Causality
def. A system S is causal if the output at time t depends only on x( ) for
t
Casuality is equivalent to the following property:
If x1 (t) = x2 (t), t t0 then y1 (t) = y2 (t), t t0
D. Stability
def. A system is said to be bounded-input bounded-output (BIBO) stable
if every bounded input produces a bounded output,i.e.
Mx < My < .
Example
1
[x(n) + x(n 1) + x(n 2)]
3
Assume |x(n)| Mx for all n
1. y(n) =
1
|x(n) + x(n 1) + x(n 2)|
3
1
[|x(n)| + |x(n 1)| + |x(n 2)|]
3
Mx
|y(n)| =
36
2.4
CHAPTER 2. SYSTEMS
Convolution
Characterization of behavior of an LTI system

1. Decompose input into sum of simple basis functions xi (n)
x(n) =
N
1
X
ci xi (n)
i=0
y(n) = S[x(n)]
=
N
1
X
ci S[xi (n)]
i=0
2. Choose basis functions which are all shifted versions of a single basis
function x0 (n)
i.e. xi (n) = x0 (n ni )
Let yi (n) = S[xi (n)], i=0,...,N-1
then yi (n) = y0 (n ni )
and y(n) =
N
1
X
ci y0 (n ni )
i=0
3. Choose impulse as basis function

Denote impulse response by h(n).
Now consider an arbitrary input x(n).
2.4. CONVOLUTION
Sum over both the set of inputs and the set of outputs.
x(n) = x(0)(n) + x(1)(n 1) + x(2)(n 2) + x(3)(n 3)

y(n) = x(0)h(n) + x(1)h(n 1) + x(2)h(n 2) + x(3)h(n 3)
37
38
CHAPTER 2. SYSTEMS
Convolution sum
x(n) =
x(k)(n k)
k=
y(n) =
x(k)h(n k)
k=
let ` = n k k = n `
X
X
y(n) =
x(n `)h(`) =
x(n `)h(`)
`=
`=
Notation and identity

For any signals x1 (n) and x2 (n), we use an asterisk to denote their
convolution; and we have the following identity.
x1 (n) x2 (n) =
=
X
k=
x1 (n k)x2 (k)
x1 (k)x2 (n k)
k=
Characterization of CT LTI systems

Approximate x(t) by a superposition of rectangular pulses:
2.4. CONVOLUTION
x (t) =
x(k) (t k)
k=
Find response to a single pulse:
Determine response to x (t):
X
y (t) =
x(k)h (t k)
k=
Let 0 (d ) k
(t) (t) x (t) x(t)
h (t) h(t) y (t) y(t)
Convolution Integral
Z
x( )(t )d
x(t) =
Z
x( )h(t )d
y(t) =
39
40
CHAPTER 2. SYSTEMS
Notation and Identity

For any signals x1 (t) and x2 (t), we use an asterisk to denote their
convolution; and we have the following identity.
Z
x1 (n) x2 (n) =
x1 ( )x2 (t )d
Z
=
x1 (t )x2 ( )d
Example
DT system y(n) =
W 1
1 X
x(n k) W - integer
W
k=0
Find response to x(n) = en/D u(n).

W - width of averaging window
D - duration of input
To find impulse response, let x(n) = (n) h(n) = y(n).

W 1
1 X
1/W, 0 n W 1
h(n) =
(n k) =
0,
else
W
k=0
Now use convolution to find response to x(n) = en/D u(n).
X
y(n) =
x(n k)h(k)
k=
Case 1: n < 0
y(n) = 0
2.4. CONVOLUTION
41
Case 2: 0 n W 1
y(n) =
n
X
x(n k)h(k)
k=0
n
1 X (nk)/D
e
W
k=0
1 n/D X k/D
e
e
W
k=0
Geometrical Series
N
1
X
zk =
k=0
1 zN
, for any complex number z
1z
1
, |z| < 1
1z
k=0

1 n/D 1 e(n+1)/D
y(n) =
e
W
1 e1/D
zk =
Case 3: W n
42
CHAPTER 2. SYSTEMS
y(n) =
W
1
X
x(n k)h(k)
k=0
W 1
1 X (nk)/D
e
W
k=0
W 1
1 n/D X k/D
=
e
e
W
k=0

1 n/D 1 eW/D
e
W
1 e1/D

1 1 e(W/D [n(W 1)]/D
=
e
W 1 e1/D
y(n) =
Puttingeverything together
0,
n<0

1 1 e(n+1)/D
,
0nW 1
y(n) =
W 1 e1/D
(W/D
1 1e
e[n(W 1)]/D , W n
W 1 e1/D
Causality for LTI systems

y(n) =
n
X
k=
x(k)h(n k) +
x(k)h(n k)
k=n+1
System will be causal second sum is zero for any input x(k).
This will be true h(n k) = 0, k = n + 1, ...,
2.4. CONVOLUTION
43
h(k) = 0, k < 0
An LTI system is causal h(k) = 0, k < 0, i.e. the impulse response is
a causal signal.
Stability of LTI systems

Suppose the input is bounded, i.e Mx <
X
x(k)h(n k)
y(n) =
k=
|y(n)| = |
x(k)h(n k)|
k=
|x(k)||h(n k)|
k=
Mx
|h(k)|
k=
Therefore, it is sufficient for BIBO stability that the impulse response be

absolutely summable.
X
Suppose
|h(k)| <
6 .
k
Consider y(0) =
x(k)h(k).

Assuming h(k) to be real-valued, let x(k) =
then y(0) =
1,
1,
h(k) > 0
h(k) < 0
|h(k)| <
6
For an LTI system to be BIBO stable,X

it is necessary that the impulse
response is absolutely summable, i.e.
|h(k)| < .
k
Examples
y(n) = x(n) + y(n 1)
Find the impulse response.
Let x(n) = (n), then h(n) = y(n).
Need to find solution to
44
CHAPTER 2. SYSTEMS
y(n) = (n) = 2y(n 1)

This example differs from earlier ones because the system is recursive, i.e
the current output depends on previous output values as well as the
current and previous inputs.
1. must specify initial conditions for the system (assume y(-1)=0)
2. cannot directly write a closed form expression for y(n)
Find output sequence term by term.
y(0) = (0) + 2y(1) = 1 + 2(0) = 1
y(1) = (1) + 2y(0) = 0 + 2(1) = 2
y(2) = (2) + 2y(1) = 0 + 2(2) = 4
Recognize general form.
h(n) = y(n) = 2n u(n)
1. Assuming system is initially at rest, it is causal
X
2.
|h(n)| <
6 system is not BIBO stable
n
2.5
Frequency Response
We have seen that the impulse signal provides a simple way to

characterize the response of an LTI system to any input. Sinusoids play a
similar role
Consider x(n) = ejn , is fixed
Denote response by y (n)
Now consider
2.5. FREQUENCY RESPONSE
45
Since ejm ejn = ej(n+m)

y (n + m) = y (n)ejm
Let n = 0,
y (m) = y (0)ejm for all m
For a system which is homogeneous and time-invariant, the response to

a complex exponential input x(n) = ejn is y(n) = y (0)x(n), a
frequency-dependent constant times the input x(n). Thus, complex
exponential signals are eigenfunctions of homogeneous, time-invariant
systems.
Comments
We refer to the constant of proportionality as the frequency
response of the system and denote it by
H(ej ) = y (0)
We write it as a function of ej rather than for two reasons:
1. digital frequencies are only unique modulo 2
2. there is a relation between frequency response and the Z
transform that this representation captures
Note that we did not use superposition. We will need it later when
we express the response to arbitrary signals in terms of the
frequency response.
46
CHAPTER 2. SYSTEMS
Magnitude and Phase of Frequency Response

In general, H(ej ) is complex-valued,
i.e. H(ej ) = A()ej()
Thus for x(n) = ejn
y(n) = A()ej[n+()]
Suppose response to any real-valued input x(n) is real-valued.
1
Let x(n) = cos(n) = [ejn + ejn ].
2
Assuming superposition holds,
y(n) =

1
H(ej )ejn + H(ej )ejn
2
[y(n)] =

1
[H(ej )] ejn + [H(ej )] ejn
2
y(n) = [y(n)] H(ej ) = [H(ej ]

Expressed in polar coordinates
A()ej() = A()ej()
A() = A() even
() = () odd
Examples
1
[x(n) + x(n 1)]
2
Let x(n) = ejn
1
y(n) = [ejn + ej(n1) ]
2
1
= [1 + ej ]ejn
2
1
H(ej ) = [1 + ej ]
2
Factoring out the half-angle:
1. y(n) =
1 j/2 j/2
e
[e
+ ej/2 ]
2
= ej/2 cos(/2)
H(ej ) =

|H(ej )| = |ej/2 || cos(/2)|
= | cos(/2)|
H(ej ) = ej/2 + cos(/2)

/2,
cos(/2) 0
H(ej ) =
/2 , cos(/2) < 0
Note:
even symmetry of |H(ej )|
odd symmetry of H(ej )|
periodicity of H(ej ) with period 2
low pass characteristic
1
[x(n) x(n 2)]
2
Let x(n) = ejn
2. y(n) =
1 jn
[e
ej(n2) ]
2
1
= [1 ej2 ]ejn
2
y(n) =
1
[1 ej2 ]
2
1
= jej [ (ej ej )]
j2
H(ej ) =
47
48
CHAPTER 2. SYSTEMS
H(ej ) = jej sin()
|H(ej )| = |j||ej || sin()|

= | sin()|
H(ej ) = j + ej + sin()

/2 ,
sin() 0
H(ej ) =
/2 , sin() < 0
The filter has a bandpass characteristic.

3. y(n) = x(n) x(n 1) y(n 1)
Let x(n) = ejn , how do we find y(n)?
Assume desired form of output,
i.e. y(n) = H(ej )ejn
H(ej )ejn = ejn ej(n1) H(ej )ej(n1)

H(ej )[1 + ej ]ejn = [1 ej ]ejn

1 ej
j
H(e ) =
1 + ej
=
1
jej/2 [ j2
(ej/2 ej/2 )]
ej/2 [ 21 (ej/2 + ej/2 )]

sin(/2)
=j
cos(/2)
= j tan(/2)
|H(ej )| = | tan(/2)|

/2,
tan(/2) 0
H(ej ) =
/2 , tan(/2) < 0
Comments
What happens at = /2?
Factoring out the half-angle is possible only for a relatively
restricted class of filters
49
50
CHAPTER 2. SYSTEMS
Chapter 3
Fourier Analysis
3.1
Chapter Outline

1. Continuous-Time Fourier Series (CTFS)
2. Continuous-Time Fourier Transform (CTFT)
3. Discrete-Time Fourier Transorm (DTFT)
3.2
Continuous-Time Fourier Series (CTFS)
Spectral representation for periodic CT signals

Assume x(t) = x(t + nT )
We seek a representation for x(t) of the form
x(t) =
N 1
1 X
Ak xk (t) (1a)
T
k=0
xk (t) = cos(2fk t + k ) (1b)

Since x(t) is periodic with period T, we only use the frequency
components that are periodic with this period:
51
52
CHAPTER 3. FOURIER ANALYSIS
How do we choose amplitude Ak and the phase k of each component?

k 6= 0:
Ak xk (t) = Ak cos(2kt/T + k )
Ak jk j2kt/T Ak jk j2kt/T
=
e e
+
e
e
2
2
= Xk ej2kt/T + Xk ej2kt/T
k = 0:
A0 x0 (t) = A0 cos(0 )
= X0
3.2. CONTINUOUS-TIME FOURIER SERIES (CTFS)
53
Given a fixed signal x(t), we dont know at the outset whether an exact
representation of the form given by Eq.(1) even exists.
However, we can always approximate x(t) by an expression of this form.
So we consider
N
1
X
1
Xk ej2kt/T
x
b(t) =
T
k=(N 1)
We want to choose the coefficients Xk so that x

b(t) is a good
approximation to x(t).
Define e(t) = x
b(t) x(t).
Recall
1
Pe =
T
1
=
T
T /2
|e(t)|2 dt
T /2
T /2
|e
x(t) x(t)|2 dt
T /2
2

N
1
X

1
j2kt/T

Xk e
x(t) dt
T
T /2

k=(N 1)
Z
N
1
X
1 T /2 1
=
Xk ej2/pikt/T x(t)
T T /2 T
k=(N 1)
N
1
X
1
X` ej2`t/T x (t) dt
T
1
=
T
T /2
`=(N 1)
("
#
Z
N 1
1 T /2
1 X
j2kt/T
Pe =
Xk e
T T /2
T
k=N +1
"
#
N 1
1 X
j2`t/T
X` e
T
`=N +1
"
#
N 1
1 X
j2/pikt/T
Xk e
x (t)
T
k=N +1
"
#
)
1 X
j2`t/T
x(t)
X` e
+ x(t)x (t) dt
T
`=N +1
54
1
Pe = 3
T
1
2
T
N
1
X
N
1
X
Xk X`
T /2
ej2(kl)t/T dt
T /2
k=N +1 `=N +1
N
1
X
T /2
Xk
x (t)ej2kt/T dt
T /2
k=N +1
Z T /2
N
1
X
1
2
X`
x(t)ej2lt/T dt
T
T
/2
`=N +1
Z T /2
1
+
|x(t)|2 dt
T T /2
T /2
j2(kl)t/T
T /2
T /2

T
j2(kl)t/T
e
dt =

j2(k l)
T /2
h
i
T
=
e[j(k`)] e[j(k`)]
j2(k `)
= T sinc(k `)

T, k = l
=
0, k 6= l
T /2
x(t)ej2lt/T dt.
el =
Let X
T /2
Pe =
1
T2
N
1
X
n
o
e X X
e k + Px
|x(k)|2 Xk X
k
k
N +1
Xk -unknown coefficients in the Fourier Series approximation

ek -fixed numbers that depend on x(t)
X
We want to choose values for the coefficients Xk ,k = N + 1, ..., N 1
which will minimize Pe .
Fix l between N + 1 and N 1.
Let Xl = Al + jBl .
and consider
Pe
Al
and
Pe
Bl
3.2. CONTINUOUS-TIME FOURIER SERIES (CTFS)
Pe =
1
T2
N
1
X
55
n
o
ek Xk X
e k + Px
|X(k)|2 Xk X
k=N +1
o
1 Pe n
Pe
el (Al jBl )X
el
= 2
(Al + jBl )(Al jBl ) (Al + jBl )X
Al
T Al
o
1 n
el X
el
= 2 Xl + Xl X
T
n
o
2
el
= 2 Re X l X
T
n o
Pe
el
= 0 Re {Xl } = Re X
Al
Similarly
Pe
1
el + j X
el
= 2 jXl jXl j X
Bl
T
n
o
2
el Xl
= j 2 Im X
T
and
n o
Pe
el
= 0 Im {Xl } = Im X
Bl
To summarize
N 1
1 X
For x
b(t) =
Xk ej2kt/T
T
k=N +1
to be a minimum mean-squared error approximation to the signal x(t)

over the interval T /2 t T /2, the coefficients Xk must satisfy
Z T /2
Xk =
x(t)ej2kt/T dt
T /2
Example
56
T /2
x(t)ej2kt/T dt
Xk =
T /2
/2
=A
ej2kt/T dt
/2
A
ej2kt/T
j2kt/T
A
=
[ejk /T ejk /T ]
j2k/T
sin(k /T )
= A
k /T
=
Line spectrum
Xk = A sinc(k /T )

|Xk | = |A sinc(k /T )| Xk =
0,
sinc(k /T ) > 0
, sinc(k /T ) < 0
3.3. CONTINUOUS-TIME FOURIER TRANSFORM

Comments
Z
57
T /2
x(t)ej2kt/T dt
Recall Xk =
T /2
T /2
1. X0 =
x(t)dt = T xavg
T /2
2. Xk = 0, k 6= 0
3. lim |Xk | = 0
T
What happens as N increases?

N 1
1 X
Let x
bN (t) =
Xk ej2kt/T
T
k=N +1
T /2
|x(t)|2 dt < ,
1. If
T /2
T /2
|b
xN (t) x(t)|2 dt 0
T /2
T /2
|x(t)|dt < and other Dirichilet conditions are met,
2. If
T /2
x
bN (t) x(t) for all t where x(t) is continuous
3. In the neighborhood of discontinuities, x
bN (t) exhibits an overshoot
or undershoot with maximum amplitude equal to 9 percent of the
step size no matter how large N is (Gibbs phenomena)
3.3
Continuous-Time Fourier Transform
Spectral representation for aperiodic CT signals

Consider a fixed signal x(t) and let xp (t) = repT [x(t)].
58
What happens to the Fourier series as T increases?

= T /2
= T /4
Fourier coefficients
Z
T /2
Xk =
T /2
Let T .
x(t)ej2kt/T dt

k/T f
Xk X(f )
Z
X(f ) =
x(t)ej2f t dt
Fourier series expansion

xp (t) =
1X
Xk ej2kt/T
T
Let T .
xp (t) x(t)
k/T f Xk X(f )
Z
1 X
df
T
k=
Z
x(t) =
X(f )ej2f t df
Fourier Transform pairs

Forward transform
Z
X(f ) =
x(t)ej2f t dt
Inverse Transform
Z
x(t) =
X(f )ej2f t df
Sufficient conditions for Existence of CTFT

1. x(t)
Z has finite energy
|x(t)|2 dt <
2. x(t)
Z is absolutely integrable
|x(t)|dt <
and it satisfies Dirichlet conditions
59
60
Transform relations
Linearity
CT F T
a1 x1 (t) + a2 x2 (t) a1 X1 (f ) + a2 X2 (f )
Scaling and shifting

t t0 CT F T
x
|a|X(af )ej2f t0
a
Modulation
x(t)ej2f0 t
CT F T
X(f f0 )
Reciprocity
CT F T
X(t) x(f )
Parsevals relation
Z
Z
|x(t)|2 dt =
|X(f )|2 df
Initial value
Z
x(t)dt = X(0)
Comments
1. Reflection is a special case of scaling and shifting with a = 1 and
t0 = 0, i.e.
CT F T
x(t) X(f )
2. The scaling relation exhibits reciprocal spreading
3. Uniqueness of the CTFT follows from Parsevals relation
CTFT for real signals

If x(t) is real, X(f ) = [X(f )] .
|X(f )| = |X(f )| and Xf = X(f )
In this case, the inverse transform may be written as:
Z
|X(f )| cos[2f t + X(f )]df
x(t) = 2
0

Additional symmetry relations:
x(t) is real and even X(f ) is real and even.
x(t) is real and odd X(f ) is real and odd.
Important transform pairs:

CT F T
rect(t) sinc(f )
CT F T
(t) 1
Proof
(by sifting property)
F {(t)} =
(t)ej2f t dt = 1
CT F T
1 (f ) (by reciprocity)
ej2f0 t
CT F T
(f f0 ) (by modulation property)
61
62
CT F T
cos(2f0 t)
1
[(f f0 ) + (f + f0 )]
2
Generalized Fourier transform

Note that (t) is absolutely integrable but not square integrable.

t
1
Consider (t) = rect
.
T
Z
| (t)|dt = 1
| (t)|2 dt =
| (t)|2 dt =
lim
The function x(t) 1 is neither absolutely nor square integrable; and the
integral
Z
1ej2f t dt is undefined.
Even when neither condition for existence of the CTFT is satisfied, we

may still be able to define a Fourier transform through a limiting process.
Let xn (t), n = 0, 1, 2, ... denote a sequence of functions each of which has
a valid CTFT Xn (f ).
Suppose lim xn (t) = x(t), that function does not have a valid transform.
n
If X(f ) = lim Xn (f ) exists, we call it the generalized Fourier transform

n

of x(t) i.e.
CT F T
x0 (t) X0 (f )
CT F T
x1 (t) X1 (f )
CT F T
x2 (t) X2 (f )
x(t)
GCT F T
X(f )
Example
Let xn (t) = rect(t/n).
Xn (f ) = nsinc(nf ) (by scaling)
lim xn (t) = 1
What is lim Xn (f )?
n
Xn (f ) 0, f 6= 0
Xn (0)
Z
What is
Xn (f )df ?
By the initial value relation

Z
Xn (f )df = xn (0) = 1
lim Xn (f ) = (f )
n
and we have
1
GCT F T
(f )
63
64
Efficient calculation of Fourier transforms

Suppose we wish to determine the CTFT of the following signal.
Brute force approach:

1. EvaluateZtransform integral directly
Z 2
0
j2f t
X(f ) =
(1)e
dt +
(1)ej2f t dt
2
2. Collect terms and simplify
Faster approach
1. Write x(t) in terms of functions whose transforms are known

x(t) = rect
t+1
2

+ rect
t1
2
2. Use transform relations to determine Xf

X(f ) = 2 sinc(2f )[ej2f ej2f ]
X(f ) = j4 sinc(2f )sin(2f )
65
Comments
1. Ax = 0 and X(0) = 0
2. x(t) is real and odd and X(f ) is imaginary and odd
CTFT and CT LTI systems

The key factor that made it possible to express the response y(t) of an
LTI system to an arbitrary input x(t) in terms of the impulse response
h(t) was the fact that we could write x(t) as the superposition of
impulses (t).
We can similarly express x(t) as a superposition of complex exponential
signals:
Z
x(t) =
X(f )ej2f t df
e ) denote the frequency response of the system, i.e. for a fixed

Let H(f
frequency f.
66
then by homogeneity
and by superposition
Thus, the response to x(t) is

Z
e )X(f )ej2f t dt
H(f
y(t) =
But also
Z
y(t) =
Y (f )ej2f t dt
e )X(f ) (1)
Y (f ) = H(f
We alsoZ know that
y(t) =
h(t )x( )d
e )?
What is the relation between h(t) and H(f
Let x(t) = (t) y(t) = h(t)
then X(f ) = 1 and Y (f ) = H(f ).
e ) = H(f )
From Eq(1), we conclude that H(f
Since the frequency response is the CTFT of the impulse response, we
will drop the tilde.
Summarizing, we have two equivalent characterizations for CT LTI

systems.
Z
y(t) =
h(t )x( )d
Y (f ) = X(f )H(f )
Convolution Theorem
Since x(t) and h(t) are arbitrary signals, we also have the following
Fourier transform relation
Z
CT F T
x1 ( )x2 (t )d X1 (f )X2 (f )
CT F T
x1 (t) x2 (t) X1 (f )X2 (f )
Product Theorem
By reciprocity, we also have the following result:
CT F T
x1 (t)x2 (t) X1 (f ) X2 (f )
This can be very (
useful for calculating transforms of certain functions.
1
[1 + cos(2t)], |t| 1/2
Example x(t) =
2
0,
|t| > 1/2
Find X(f )
1
[1 + cos(2t)]rect(t)
2

1
1
X(f ) =
(f ) + [(f 1) + (f + 1)] sinc(f )
2
2
Since convolution obeys linearity, we can write this as
x(t) =
67
68
1
X(f ) =
2

1
(f ) sinc(f ) + [(f 1) sinc(f ) + (f + 1) sinc(f )]
2
All three convolutions here are of the same general form.
Identity
For any signal w(t),
w(t) (t t0 ) = w(t t0 )
Proof:
w(t) (t t0 ) =
w( )(t t0 )d = w(t t0 ) (by sifting property)
Using the identity,

1
1
X(f ) =
(f ) sinc(f ) + [(f 1) sinc(f ) + (f + 1) sinc(f )]
2
2

1
1
=
sinc(f ) + [sinc(f 1) + sinc(f + 1)]
2
2
Fourier transform of Periodic Signals

We previously developed the Fourier series as a spectral
representation for periodic CT signals
Such signals are neither square integrable nor absolutely integrable,
and hence do not satisfy the conditions for existence of the CTFT
By applying the concept of the generalized Fourier transform, we
can obtain a Fourier transform for periodic signals
69
This allows us to treat the spectral analysis of all CT signals within

a single framework
We can also obtain the same result directly from the Fourier series
Let x0 (t) denote one period of a signal that is periodic with period T, i.e.
x0 (t) = 0, |t| > T /2.
Define x(t) = repT [x0 (t)].
The fourier series representation for x(t) is
1X
x(t) =
Xk ej2kt/T
T
k
Taking the CTFT directly, we obtain

(
)
1X
j2kt/T
X(f ) = F
Xk e
T
k
n
o
1X
=
Xk F ej2kt/T (by linearity)
T
k
1X
Xk (f k/T )
=
T
k
Also
Z
T /2
x(t)ej2kt/T dt
Xk =
T /2
Z
x0 (t)ej2kt/T dt
= X0 (k/T )
Thus
X(f ) =
1X
X0 (k/T )(f k/T )
T
k
1
= comb T1 [X0 (f )]
T
Dropping the subscript 0, we may state this result in the form of a
transform relation:
CT F T
repT [x(t)]
1
comb T1 [X(f )]
T
70
For our derivation, we required that x(t) = 0, |t| > t/2. However, when
the generalized transform is used to derive the result, this restriction is
not needed.
Example
3.4
Discrete-Time Fourier Transform
Spectral representation for aperiodic DT signals

As in the CT case, we may derive the DTFT by starting with a
spectral representation (the discrete-time Fourier series) for
periodic DT signals and letting the period become infinitely long
Instead, we will take a shorter but less direct approach
Recall the continuous-time fourier series:
1 X
x(t) =
Xk ej2kt/T (1)
T
k=
3.4. DISCRETE-TIME FOURIER TRANSFORM

Z
71
T /2
x(t)ej2kt/T dt (2)
Xk =
T /2
Before, we interpreted (1) as a spectral representation for the CT

periodic signal x(t).
Now lets consider (2) to be a spectral representation for the sequence
Xk , < k < .
We are effectively interchanging the time and frequency domains.
We want to express an arbitrary signal x(n) in terms of complex
exponential signals of the form ejn .
Recall that is only unique modulo 2.
To obtain this, we make the following substitutions in (2).
Z T /2
Xk =
x(t)ej2kt/T dt
T /2
kn
T x(t) X(ej )
2t/T
Xk x(n)
T
d
dt 2
x(n) =
1
2
X(ej )ejn d
This is the inverse transform.

To obtain the forward transform, we make the same substitution in (1).
x(t) =
1 X
Xk ej2kt/T
T
k=
T x(t) X(ej )
kn
Xk x(n)
2t/T
X(ej ) =
X
n=
x(n)ejn
72
Putting everything together
X(ej ) =
x(n)ejn
n=
1
2
x(n) =
X(ej )ejn d
Sufficient conditions for existence

X
|x(n)|2 < or
n
|x(n)| < plus Dirichlet conditions.
Transform Relations
1. linearity
a1 x1 (n) + a2 x2 (n)
DT F T
a1 X1 (ej ) + a2 X2 (ej )
2. shifting
x(n n0 )
DT F T
X(ej )ejn0
DT F T
X(ej(0 ) )
3. modulation
x(n)ej0 n
4. Parsevals relation
|x(n)|2 =
n=
1
2
5. Initial value
X
n=
x(n) = X(ej0 )
|X(ej )|2 d
73
Comments
1. As discussed earlier, scaling in DT involves sampling rate changes;
thus, the transform relations are more complicated than in CT case
2. There is no reciprocity relation, because time domain is a
discrete-parameter whereas frequency domain is a
continuous-parameter
3. As in CT case, Parsevals relation guarantees uniqueness of the
DTFT
Some transform pairs

1. x(n) = (n)
X(ej ) =
(n)ejn
= 1 (by sifting property)
2. x(n) = 1
does not satisfy existence conditions
X
ejn does not converge in ordinary sense

n
cannot use reciprocity as we did for CT case

Consider inverse transform
Z
1
x(n) =
X(ej )ejn d
2
What function X(ej ) would yield x(n) 1?
74

Let X(ej ) = 2(), ,
Z
1
2()ejn d = 1 (again by sifting property).
then x(n) =
2
Note that X(ej ) must be periodic with period 2, so we have:
X
X(ej ) = 2
( 2k)
k
3. ej0 n
DT F T
4. cos(0 n)

5. x(n) =
rep2 [2( 0 )] (by modulation property)
DT F T
rep2 [( 0 ) + ( + 0 )]
1, 0 n N 1
0, else
X(ej ) =
N
1
X
ejn
n=0
1 ejN
1 ej
= ej(N 1)/2

6. y(n) =
sin(N/2)
sin(/2)
1, (N 1)/2 n (N 1)/2
0, else
75
N is odd
y(n) = x(n + (N 1)/2) where x(n) is signal from Example 5
Y (ej ) = X(ej )ej(N 1)/2 (by sifting property)

j(N 1)/2 sin(N/2)
= e
ej(N 1)/2
sin(/2)
Y (ej ) =
sin(N/2)
sin(/2)
What happens as N ?
y(n) 1
< n <
2/N 0
Y (ej ) rep2 [2()]
Z
1
Y (ej )d = y(0)
2
DTFT and DT LTI systems

Recall that in CT case we obtained a general characterization in terms of
CTFT by expressing x(t) as a superposition of complex exponential
signals and then using frequency response H(f) to determine response to
each such signal.
Here we take a different approach.
For any DT LTI system, we know that input and output are related by
X
y(n) =
h(n k)x(k)
k
76
Thus
Y (ej ) =
y(n)ejn
XX
n
(
X X
X
)
h(n k)e
jn
x(k)
h(n k)x(k)ejn
H(ej )ejk x(k)
(by sifting property)
where H(ej ) is the DTFT of the impulse response h(n).

Rearranging,
Y (ej ) = H(ej )
x(k)ejk
k
j
= H(e )X(ej )
How is H(ej ) related to the frequency response?
Consider
x(n) = ej0 n
X(ej ) = rep2 [2( 0 )]
X(ej ) + 2
( 0 2k)
Y (ej ) = H(ej )X(ej )

X
= 2
H(ej0 =2k )( 0 2k)
k
= H(e
j0
)2
( 0 2k)
Y (ej ) = H(ej0 )X(ej )

y(n) = H(ej0 )x(n)
77
so H(ej ) is also the frequency response of the DT system.

We thus have two equivalent characterizations for the response y(n) of a
DT LTI system to any input x(n).
y(n) =
h(n k)x(k)
Convolution Theorem
Since x(n) and h(n) are arbitrary signals, we also have the following
transform relation:
X
DT F T
x1 (k)x2 (n k) X1 (ej )X2 (ej )
k
or
x1 (n) x2 (n)
DT F T
X1 (ej )X2 (ej )
Product theorem
As mentioned earlier, we do not have a reciprocity relation for the DT
case.
However, by direct evaluation of the DTFT, we also have the following
result:
Z
DT F T 1
x1 (n)x2 (n)
X1 (ej() )X2 (ej )d
2
Note that this is periodic convolution.
78
Chapter 4
Sampling
4.1
Chapter Outline

1. Analysis of Sampling
2. Relation between CTFT and DTFT
3. Sampling Rate Conversion (Scaling in DT)
4.2
Analysis of Sampling
A simple scheme for sampling a waveform is to gate it.
T - period
- interval for which switch is closed
/T - duty cycle
xs (t) = s(t)x(t)
79
80
CHAPTER 4. SAMPLING
Fourier Analysis
Xs (f ) = S(f ) X(f )

s(t) = repT 1 rect t
1
comb T1 [sinc( f )]
T
1X
=
sinc( k/T )(f k/T )
T
S(f ) =
Xs (f ) =
1
T
sinc( k/T )X(f k/T )
How do we reconstruct x(t) from xs (t)?
4.2. ANALYSIS OF SAMPLING
Nyquist condition
Perfect reconstruction of x(t) from xs (t) is possible if
X(f ) = 0, |f | 1/(2T )
Nyquist sampling rate
1
fs =
= 2W
T
Ideal Sampling
What happens as 0?
X
s(t)
(t mT )
m
xs (t)
x(mT )(t mT ) = combT [x(t)]
Xs (f )
1X
1
X(f k/T ) = rep T1 [X(f )]
T
T
k
An ideal lowpass filter will again reconstruct x(t)

In the sequel, we assume ideal sampling
81
82
CHAPTER 4. SAMPLING
Transform Relations
1
comb T1 [X(f )]
T
CT F T 1
combT [x(t)]
rep T1 [X(f )]
T
CT F T
repT [x(t)]
Given one relation, the other follows by reciprocity.
Whittaker-Kotelnikov-Shannon Sampling Expansion

Xr (f ) = Hr (f )Xs (f )
Hr (f ) = T rect(T f )
xr (t) = hr (t) xs (t)
hr (t) = X
sinc(t/T )
xs (t) =
x(mT )(t mT )
m

X
t mT
xr (t) =
x(mT )sinc
T
m
xr (nT ) =
x(mT )sinc(n m)
= x(nT )
83
If Nyquist condition is satisfied, xr (t) x(t).
Zero Order Hold Reconstruction
xr (t) =

x(mT )rect
t T /2 mT
T
= hZO (t) xs (t)

hr (t) = rect
t T /2
T
Xr (f ) = HZO (f )Xs (f )
HZO (f ) = T sinc(T f )ej2f (T /2)
84
CHAPTER 4. SAMPLING
One way to overcome the limitations of zero order hold

reconstruction is to oversample, i.e. choose fs = 1/T >> W
This moves spectral replications farther out to where they are
better attenuated by sinc(Tf)
What happens if we inadvertently undersample, and then
reconstruct with an ideal lowpass filter?
Effect of undersampling
frequency truncation error
Xr (f ) = 0,
|f | fs /2
aliasing error
Frequency components at f1 = fs /2 + fold down to and mimic
frequencies f2 = fs /2 .
Example
x(t) = cos[2(5000)t]
sample at fs = 8 kHz
reconstruct with ideal lowpass filter having cutoff at 4 kHz
Sampling
xr (t) = cos[2(3000)t]
Note that x(t) and xr (t) will have the same sample values at times
t = nT (T=1/(8000)).
x(nT ) = cos {2(5000)n/8000}
= cos {2(5000)n/8000}
= cos {2[(5000)n0(8000)n]/8000}
= cos {2(3000)n/8000}
= xr (nT )
85
86
Reconstruction
CHAPTER 4. SAMPLING
4.3. RELATION BETWEEN CTFT AND DTFT
4.3
Relation between CTFT and DTFT
We have already shown that

Xs (fa ) =
1
rep T1 [X(fa )]
T
Suppose we evaluate the CTFT of xs (t) directly

(
)
X
Xs (fa ) = F
xa (nT )(t nT )
n
xa (nT )F {(t nT )}
Xs (fa ) =
xa (nT )ej2fa nT
Recall
Xd (ejd )
X
n
xd (n)ejd n
87
88
CHAPTER 4. SAMPLING
But xd (n) = xa (nT ).

Let d = 2fa T = 2(fa /fs ).

d
Rearranging as fa =
fs , we obtain
2
h i
d
Xd (ejd ) = Xs
fs
2
Example
CT Analysis
xa (t) = cos(2fa0 t)
1
Xa (fa ) = [(fa fa0 ) + (fa + fa0 )]
2
Xs (fa ) = fs repfs [Xa (fa )]
fs X
=
[(fa fa0 kfs ) + (fa + fa0 kfs )]
2
k
4.3. RELATION BETWEEN CTFT AND DTFT
89
DT Analysis
xd (n) = xa (nT )
= cos(2fa0 nT )
= cos(d0 n)
d0 = 2fa0 T = 2(fa0 /fs )
X
X(ejd ) =
[(d d0 2k) + (d + d0 2k)]
k
Relation between CT and DT Analyses

h i
d
Xd (ejd ) = Xs
fs
2
fs X
Xs (fa ) =
[(fa fa0 kfs ) + (fa + fa0 kfs )]
2
k

fs X
d fs
d fs
jd
Xd (e ) =
fa0 kfs +
+ fa0 kfs
2
2
2
k
Recall (ax + b) =
1
(x + b/a).
|a|

2fa0
k2fs
fs X 2
d
2
fs
fs
fs
k

2
2fa0
k2fs
+
d +
fs
fs
fs
X
=
[(d d0 2k) + (d + d0 2k)]
Xd (ejd ) =
Aliasing
CT
fa1 = fs /2 + a folds down to fa2 = fs /2 a .
DT

Let d = 2
fa
fs

d = 2
a
fs
d1 = + d is identical to d2 = d .
90
4.4
CHAPTER 4. SAMPLING
Sampling Rate Conversion
Recall definitions for DT scaling.
Downsampling
y(n) = x(Dn)
Upsampling

y(n) =
x(n/D), if n/D is an integer

0,
else
4.4. SAMPLING RATE CONVERSION
91
In CT, we have avery

simple transform relation for scaling:
f
CT F T 1
x(at)
X
|a|
a
What is the transform relation for DT?

1. As in definitions for downsampling and upsampling, we must treat
these cases separately
2. Relations will combine elements from both scaling and sampling
transform relations for CT
Downsampling
y(n) = x(Dn)
X
Y (ej ) =
x(Dn)ejn
n
let m = Dn n = m/D
X
Y (ej ) =
x(m)ejm/D (m/D is an integer)
m
To removerestrictions on m, define a sequence:

1, m/D is an integer
sD (m) =
0, else
then
92
CHAPTER 4. SAMPLING
Y (ej ) =
sD (m)x(m)ejm/D
Alternate expression for sD (m):

D=2
1
[1 + (1)m ]
2
1
= [1 + ej2m/2 ]
2
s2 (m) =
s2 (0) = 1
s2 (1) = 0
s2 (m + 2k) = s(m)
D=3
s3 (m) = 31 [1 + ej2m/3 + ej2(2)m/3 ]
s3 (0) = 31 [1 + 1 + 1] = 1
s3 (1) = 13 [1 + ej2/3 + ej2(2)/3 ] = 0
s3 (2) = 13 [1 + ej2(2)/3 + ej2(4)/3 ] = 0

s3 (m + 3k) = s3 (m)
In general,
sD (m) =
D1
1 X j2km/D
e
D
k=0
1 1 ej2m
=
D 1 ej2m/D

sD (m) =
1,
0,
Y (ej ) =
m/D is an integer
else
X
sD (m)x(m)ejm/D
X
X 1 D1
ej2km/D x(m)ejm/D
D
m
k=0
1
D
1
D
D1
XX
x(m)ej[(+2k)/D]m
k=0 m
D1
X
k=0
X(ej(+2k)/D )
93
94
CHAPTER 4. SAMPLING
Decimator
To prevent aliasing due to downsampling, we prefilter the signal to
bandlimit it to highest frequency d = /D.
The combination of a low pass filter followed by a downsampler is
referred to as a decimator.
With prefilter
Without prefilter
Upsampling

y(n) =
x(n/D), n/D is an integer

0,
else
Y (ej ) =
x(n/D)ejn
(n/D) is an integer
sD (n)x(n/D)ejn
Let m = n/D n = mD
X
Y (ej ) =
sD (mD)x(m)ejmD but sD (mD) 1
m
Y (ej ) = X(ejD )
95
96
CHAPTER 4. SAMPLING
Interpolator
To interpolate between the nonzero samples generated by upsampling, we
use a low pass filter with cutoff at d = /D.
The combination of an upsampler followed by a low pass filter is referred
to as an interpolator.

Time Domain Analysis of an Interpolator
y(n) =
v(k)h(n k)
v(k) = sD (k)x(k/D)
let ` = k/D k = `D
X
y(n) =
x(`)h(n `D)
`
1
2
1
=
2
h(n) =
H(ej )ejn d
/D
ejn d
/D

1

CTFT1 rect

2
2/D
t=n/2

2 n
1 2
sinc
=
2 D
D 2

1
n
= sinc
D
D
=
y(n) =
x(`)h(n `D)
X
`

x(`)sinc
n `D
D
97
98
CHAPTER 4. SAMPLING
Chapter 5
Z Transform (ZT)
5.1
Chapter Outline

1. Derivation of the Z transform
2. Convergence of the ZT
3. ZT properties and pairs
4. ZT and linear, constant coefficient difference equations
5. Inverse Z transform
6. General Form of Response of LTI Systems
5.2
Derivation of the Z Transform
1. Extension of DTFT
2. Transform exists for a larger class of signals than does DTFT
3. Provides an important tool for characterizing behavior of LTI
systems with rational transfer functions
Recall sufficient conditions for existence of the DTFT
99
100
CHAPTER 5. Z TRANSFORM (ZT)

1.
|x(n)|2 <
2.
or
X
|x(n)| <
By using impulses we were able to define the DTFT for periodic

signals which satisfy neither of the above conditions
Example 1
Consider x(n) = 2n u(n)
It also satisfies neither condition for existence of the DTFT.

Define x
e(n) = rn x(n),
r>2
5.2. DERIVATION OF THE Z TRANSFORM
101
Now
|e
x(n)| =
n=0
|rn x(n)|
n=0
(2/r)n
n=0
X
n=0
|e
x(n)| =
1
, (2/r) = 1 or r > 2
1 2/r
DTFT of x
e(n) exists
X
e j ) =
X(e
x
e(n)ejn
n=0
e j ) =
X(e
[2(rej )1 ]n
n=0
1
, |2(rej )1 | < 1 or r > 2
1 2(rej )1
e j ) in terms of x(n)
Now lets express X(e
e j ) =
X(e
=
X
n=0
rn x(n)ejn
x(n)(rej )n
n=0
Let z = rej and define the Z transform (ZT) of x(n) to be the DTFT of
x(n) after multiplication by the convergence factor rn u(n).
e j ) =
X(z) = X(e
x(n)z n
For the example x(n) = 2n u(n).

X(z) =
1
,
1 2z 1
|z| > 2
It is important to specify the region of convergence since the transform is

not uniquely defined without it.
102
Example 2
Let y(n) = 2n u(n 1)
Y (z) =
y(n)z n
1
X
2n z n
n=
Y (z) =
=
=
1
X
(z/2)n
n=
(z/2)n
n=1
(z/2)n + 1
n=0
Y (z) = 1
1
, |z/2| < 1 or |z| < 2
1 z/2
z/2
1 z/2
1
=
, |z| < 2
1 2z 1
Y (z) =
So we have
ZT
x(n) = 2n u(n) X(z) =
1
, |z| > 2
1 2z 1
1
ZT
y(n) = 2n u(n 1) Y (z) =
, |z| < 2
1 2z 1
5.3. CONVERGENCE OF THE ZT

1. The two transforms have the same functional form
2. They differ only in their regions of convergence
Example 3
w(n) = 2n , < n <
w(n) = x(n) y(n)

By linearity of the ZT,
W (z) = X(z) Y (z)
1
1
=
=0
1 2z 1
1 2z 1
But note that X(z) and Y (z) have no common region of convergence.
Therefore, there is no ZT for
w(n) = 2n , < n <
5.3
Convergence of the ZT
1. A series
X
n=0
un
103
104
is said to converge to U if given any real > 0, there exists an

M such
that
integer
1

NX

un U < for all N > M

n=0
2. Here the sequence un and the limit U may be either real or complex
3. A sufficient condition for convergence of the series
un is that of
n=0
absolute convergence i.e.
|un | <
n=0
4. Note that absolute convergence is not necessary for ordinary

convergence as illustrated by the following example
The series
(1)n
n=0
1
= ln(2)
n
is convergent, whereas the series
X
1
is not.
n
n=1
5. However, when we discuss convergence of the Z transform, we

consider only absolute convergence Thus, we say that
X
X(z) =
x(n)z n
n
converges at z = z0 if
X
X
x(n)z n =
|x(n)|r0n <
0
n
where r0 = |z0 |
Region of Convergence for ZT

As a consequence of restricting ourselves to absolute convergence, we
have the following properties:
5.3. CONVERGENCE OF THE ZT
105
P1. If X(z) converges at z = z0 , then it converges for all z for which

|z| = r0 , where r0 = |z0 |.
Proof:
X
X
x(n)z n =
|x(n)|rn
0
< by hypothesis
P2. If x(n) is a causal sequence, i.e. x(n) = 0, n < 0, and X(z) converges
for |z| = r1 , then it converges for all Z such that |z| = r > r1 .
Proof:
X
X

x(n)z n =
|x(n)|rn
n=0
<
n=0
|x(n)|r1n
n=0
< by hypothesis
P3. If x(n) is an anticausal sequence, i.e. x(n) = 0, n > 0 and X(z)
converges for |z| = r2 , then it converges for all z such that |z| = r < r2 .
Proof:
0
X

X
x(n)z n =
|x(n)|rn
n=
<
n=0
|x(n)|r2n
n=0
< by hypothesis
P4. If x(n) is a mixed causal sequence, i.e. x(n) 6= 0 for some n < 0 and
x(n) 6= 0 for some n > 0, and X(z) converges for some |z| = r0 , then
there exists two positive reals r1 and r2 with r1 < r0 < r2 such that X(z)
converges for all z satisfying r1 < |z| < r2 .
Proof:
Let x(n) = x (n) + x+ (n) where x (n) is anticausal and x+ (n) is causal.
Since X(z) converges for |z| = r0 , X (z) and X+ (z) must also both
106
converge for |z| = r0 .

From properties 2 and 3, there exist two positive reals r1 and r2 with
r1 < r0 < r2 such that X (z) converges for |z| < r2 and X+ (z) converges
for |z| > r1 .
The ROC for X(z) is just the intersection of these two ROCs.
Example 3
x(n) = 2|n| u(n)
X(z) =
2|n| z n
=
=
1
X
2n z n +
n=
X
1
(2
2n z n
n=0
z)n 1 +
n=0
(21 z 1 )n
n=0
1
1
=
1+
1 21 z
1 21 z 1
|21 z| <
|21 z 1 | <
|z| < 2
Combining everything
3z 1 /2
X(z) =
,
1 5z 1 /2 + z 2
|z| > 1/2
1/2 < |z| < 2
5.4. ZT PROPERTIES AND PAIRS
5.4
107
ZT Properties and Pairs
Transform Relations
1. Linearity
ZT
a1 x1 (n) + a2 x2 (n) a1 X1 (z) + a2 X2 (z)

2. Shifting
ZT
x(n n0 ) z n0 X(z)
3. Modulation
ZT
z0n x(n) X(z/z0 )

4. Multiplication by time index
ZT
nx(n) z
d
[X(z)]
dz
5. Convolution
ZT
x(n) y(n) X(z)Y (z)

6. Relation to DTFT
If XZT (z) converges for |z| = 1, then the DTFT of x(n) exists and
XDT F T (ej ) = XZT (z)|z=ej
Important Transform Pairs

ZT
1. (n) 1,
ZT
2. an u(n)
all z
1
,
1 az 1
ZT
3. an u(n 1)
ZT
4. nan u(n)
|z| > a
1
,
1 az 1
az 1
2,
(1 az 1 )
ZT
5. nan u(n 1)
az 1
|z| < a
|z| > a
2,
(1 az 1 )
|z| < a
108
5.5
ZT,Linear Constant Coefficient

Difference Equations
1. From the convolution property, we obtain a characterization for all

LTI systems
impulse response: y(n) = h(n) x(n)

transfer function: Y (z) = H(z)X(z)
2. An important class of LTI systems are those characterized by
linear, constant coefficient difference equations
M
N
X
X
y(n) =
ak x(n k)
h` y(n `)
k=0
`=1
nonrecursive
N =0
always finite impulse response (FIR)
recursive
N >0
usually infinite impulse response (IIR)
3. Take ZT of both sides of the equation
M
N
X
X
y(n) =
ak x(n k)
b` y(n `)
Y (z) =
k=0
M
X
`=1
N
X
ak z k X(z)
k=0
`=1
M
X
H(z) =
Y (z)
=
X(z)
b` z ` Y (z)
ak z k
k=0
N
X
1+
`=1
b` z `
5.5. ZT,LINEAR CONSTANT COEFFICIENT DIFFERENCE EQUATIONS109

4. M N
Multiply numerator and denominator by z M .
M
X
ak z M k
"k=0
H(z) =
z
M N
N
X
b` z
N `
`=1
5. M < N
Multiply numerator and denominator by z N .
M
X
z N M
ak z M k
H(z) =
zN +
k=0
N
X
b` z N `
`=1
6. By the fundamental theorem of algebra, the numerator and

denominator polynomials may always be factored
M N
M
Y
H(z) =
M <N
H(z) =
(z zK )
k=1
z M N
QN
`=1 (z
p` )
QM
z N M k=1 (z zk )
N
Y
(z p` )
`=1
7. Roots of the numerator and denominator polynomials:

zeros z1 , ..., zM
poles p1 , ..., pN
If M N , have M N additional poles at |z| =
If M < N , have N M additional zeros at z = 0
8. The poles and zeros play an important role in determining system
behavior
Example
1
y(n) = x(n) + x(n 1) y(n 2)
2
110
1
Y (z) = X(z) + z 1 X(z) z 2 Y (z)
2
1 + z 1
Y (z)
=
H(z) =
X(z)
1 + 21 z 2
zeros : z1 = 0 z2 = 1
poles : p1 = j/ 2 p2 = j/ 2
z(z + 1)
z 2 + 1/2
z(z + 1)
=
(z j/ 2)(z + j/ 2)
H(z) =
Effect of Poles and Zeros on Frequency Response

Frequency response H(ej )
CT F T
h(n) HDT F T (ej ) = H(ej )

ZT
h(n) HZT (z)

HDT F T (ej ) = HZT (ej )
Let z = ej .

H(ej ) = HZT (ej )
Assume M < N .
M
Y
z N M
(z zk )
k=1
H(z) =
N
Y
(z p` )
`=1
ej(N M )
M
Y
(ej zk )
k=1
H(ej ) =
N
Y
(ej p` )
`=1
M
Y
|H(ej )| =
k=1
N
Y
|ej zk |
|ej p` |
`=1
H(ej ) = (N M ) +
M
X
k=1
(ej zk )
N
X
(ej p` )
`=1
Example
z(z + 1)
(z j/ 2)(z + j/ 2)
|ej + 1|
|H(ej )| =
j
|e j/ 2||ej + j/ 2|
H(ej ) = + (ej + 1) (ej j/ 2) (ej + j/ 2)

Contribution from a single pole
H(z) =
112
=0
|H(ej0 )| =
|~a| = 2
|~b| = 1q
|~c| =
~ =
|d|
1+
q
|~
a||~b|
~
|~
c||d|
1
2
3
2
3
2
|H(ej0 )| = 43 = 1.33
H(ej0 ) = ~a + ~b ~c d~
~a = 0
~b = 0

~c = arctan 12
d~ = ~c
H(ej0 ) = 0
= /4
|H(ej/4 )| =
r
|~a| =
1+
|~b| = 1
|~
a||~b|
~
|~
c||d|
1
2
2
1
2
= 1.85

|~c| =
~ =
|d|
r2
2
2
2
1
2
2
5
2
|H(ej/4 )| = 1.65
H(ej/4 ) = ~a + ~b ~c d~
~a = 22.5
~b = 45
~c = 0
d~ = 63.4
H(ej/4 ) = 4.1
= /2
|H(ej/2 )| =
|~a| = 2
|~b| = 1
|~c| = 1 12
~ = 1 + 1
|d|
|~
a||~b|
~
|~
c||d|
|H(ej/2 )| = 2.83
H(ej/4 ) = ~a + ~b ~c d~
~a = 45
~b = 90
~c = 90
d~ = 90
H(ej/2 ) = 45
General Rules
1. A pole near the unit circle will cause the frequency response to
increase in the neighborhood of that pole
114
2. A zero near the unit circle will cause the frequency response to
decrease in the neighborhood of that zero
5.6
Inverse Z Transform
1. Formally, the
I inverse ZT may be written as
1
x(n) = j2
X(z)z n1 dz
c
where the integration is performed in a counter-clockwise direction

around a closed contour in the region of convergence of X(z) and
encircling the origin.
2. If X(z) is a rational function of z, i.e. a ratio of polynomials, it is
not necessary to evaluate the integral
3. Instead, we use a partial fraction expansion to express X(z) as a
sum of simple terms for which the inverse transform may be
recognized by inspection
4. The ROC plays a critical role in this process
5. We will illustrate the method via a series of examples
Example 1
The signal x(n) = (1/3)n u(n) is input to a DT LTI system described by
y(n) = x(n) x(n 1) + (1/2)y(n 1)
Find the output y(n).

What are our options?
direct substitution
Assume y(1) = 0.
y(0) = x(0) x(1) + (1/2)y(1) = 1 0 + (1/2)(0) = 1.000
y(1) = x(1) x(0) + (1/2)y(0) = 1/3 1 + (1/2)(1.0) = 0.167
y(2) = x(2) x(1) + (1/2)y(1) = 1/9 1/3 + (1/2)(0.167) = 0.306
y(3) = x(3)x(2)+(1/2)y(2) = 1/271/9+(1/2)(0.306) = 0.227
5.6. INVERSE Z TRANSFORM
115
convolution
y(n) =
h(n k)x(k)
k=
Find impulse response by direct substitution.

h(n) = (n) (n 1) + (1/2)h(n 1)
h(0) = 1 0 + (1/2)(0) = 1
h(1) = 0 1 + (1/2)(1) = 1/2
h(2) = 0 0 + (1/2)(1/2) = 1/4
h(3) = 0 0 + (1/2)(1/4) = 1/8
Recognize h(n) = (n) (1/2)n u(n 1).
y(n) =
h
X
i
(n k) (1/2)(nk) u(n k 1) (1/3)k u(k)
k=
= (1/3)n u(n) (1/2)n
n1
X
(2/3)k u(n 1)
k=0
(2/3)n
u(n)
1 2/3
= 4(1/3)n u(n) 3(1/2)n u(n)
n
= (1/3) u(n) (1/2)
n1
DTFT
x(n) = (1/3)n u(n)
X(ej ) =
1
1 13 ej
h(n) = (n) (1/2)n u(n 1)

1
H(e ) = 1
1
1 12 ej
j
1 ej
1 12 ej
116

=
=
1
y(n) =
2
1 ej
(1
1 j
)(1 13 ej )
2e
j
1e
1 56 ej + 61 ej2
1 ej
ejn d
1 56 ej + 61 ej2
ZT
x(n) = (1/3)n u(n)
X(z) =
1
,
1 13 z 1
|z| >
1
3
y(n) = x(n) x(n 1) + (1/2)y(n 1)

Y (z) = X(z) z 1 X(z) + (1/2)z 1 Y (z)
H(z) =
1 z 1
Y (z)
=
X(z)
1 12 z 1
What is the region of convergence?

Since system is causal, h(n) is a causal signal.
H(z) converges for |z| > 1/2.
Y (z) = H(z)X(z)
=
(1
1 z 1
13 z 1 )
1 1
)(1
2z
ROC[Y (z)] = ROC[H(z)] ROC[X(z)]

= z : |z| > 1/2
117
Partial Fraction Expansion (PFE)

(Two distinct poles)
1
1 z 1
A1
A2

=
+
1 12 z 1
1 13 z 1
1 13 z 1
1 1
2z
To solve for A1 and A2 , we multiply both sides by 1 21 z 1

to obtain

1 z 1 = A1 1 31 z 1 + A2 1 12 z 1
A1 = 3
1 = A1 + A2
13 A1
1 =
12 A2
A2 = 4
so
(1
3
4
1 z 1
=
+
check
13 z 1 )
1 12 z 1
1 13 z 1
1 1
)(1
2z

1 z 1 = 3 1 31 z 1 + 4 1 12 z 1
Now
y(n) = y1 (n) + y2 (n)
Possible ROCs
where
3
1
1 1 , |z| < 2 or |z| >
1 2z
4
1
Y2 (z) =
1 1 , |z| < 3 or |z| >
1 3z
Y1 (z) =
1
2
1
3
1 13 z 1
118
and
ROC[Y (z)] = ROC[Y1 (z)] ROC[Y2 (z)]

1
= z : |z| >
2
Recall transform pairs:
1
|z| > |a|
1 az 1
ZT
an u(n)
ZT
an u(n 1)
ZT 1
3
1 1
1 2 z
ZT
4
1 31 z 1

1 n
2

1 n
3
1
|z| < |a|
1 az 1
u(n)
u(n)
and
n
n
1
1
u(n) + 4
u(n)
y(n) = 3
2
3
Consider again
1 z 1

1
1 13 z 1
3
4
=
+
1 12 z 1
1 13 z 1
Y (z) =
1 1
2z
There are 3 possible ROCs for a signal with this ZT.

Case II:
1
1
< |z| <
3
2
Possible ROCs
3
1
, |z| < or |z| >
Y1 (z) =
2
1 12 z 1
4
1
Y2 (z) =
, |z| < or |z| >
3
1 13 z 1
119
1
2
1
3
and
ROC[Y (z)] = ROC[Y1 (z)] ROC[Y2 (z)]

ROC[Y1 (z)] = z : |z| < 12

ROC[Y2 (z)] = z : |z| > 13
n
3
1
ZT 1
3
u(n 1)
2
1 12 z 1
n
1
4
ZT 1
u(n)
1 1 4 3
1 3z
n
n
1
1
y(n) = 3
u(n 1) + 4
u(n)
2
3
Case III:
1
|z| <
3

1
ROC[Y1 (z)] = z : |z| <
2

1
ROC[Y2 (z)] = z : |z| <
3
n
1
1
3
ZT
3
u(n 1)
2
1 12 z 1
n
4
1
ZT 1
4
u(n 1)
3
1 13 z 1
n
n
1
1
y(n) = 3
u(n 1) 4
u(n 1)
2
3
120
Summary of Possible ROCs and Signals

1 z 1

Y (z) =
1
1
1 z 1
1 z 1
2
3
1
3
ROC
< |z|
< |z| <
|z| < 13
1
2
1
2
n Signal n
3 12 u(n) + 4 31 u(n)
n
n
3 12 u(n 1) + 4 13 u(n)
n
n
3 21 u(n 1) 4 13 u(n 1)
Residue Method for Evaluating Coefficients of PFE

1 z 1
A1
A2

=
+
1 12 z 1
1 13 z 1
1
1 13 z 1

1 1
Y (z)
A1 = 1 z
2
z= 12

1 z 1
=
1 13 z 1 z= 1
Y (z) =
1 1
2z
12
=
1 2/3
= 3

1
1 z 1 Y (z)
3
z= 13

1 z 1
=
1 12 z 1 z= 1
A2 =
13
=
1 3/2
=4
121
Example 2 (complex conjugate poles)
1 + 12 z 1
(1 jz 1 ) (1 + jz 1 )
A1
A2
X(z) =
+
1 jz 1
1 + jz 1
Y (z) =

1 + 12 z 1
A1 =
1 + jz 1 z=j

1
1
=
1 j
2
2

1 + 12 z 1
A2 =
1 jz 1 z=j

1
1
1+ j
=
2
2
check

j
1
1 12 j
2 1+ 2
+
=
1 jz 1
1 + jz 1
1
2

1 + jz 1 + 12 1 + 2j 1 jz 1
(1 jz 1 ) (1 + jz 1 )
1 1
1 + 2z
=
(1 jz 1 )(1 + jz 1 )
1
2
1 12 j
Note that A1 = A2
This is necessary for y(n) to be real-valued
Use it to save computation
122
Example 3 (poles with multiplicity greater than 1)

Y (z) =
1 + z 1
2
1 12 z 1 (1 z 1 )
1
< |z| < 1
2
recall
az 1
ZT
nan u(n)
(1 az 1 )
2,
|z| > |a|
Try
A1
A2
1 + z 1
=

+
1 1 2
1 1 2
1
1
z 1
1 2z
1 2z
(1 z )

1 + z 1
A1 =
= 3
1 z 1 z= 1
2

1 + z 1
A2 =
=8
2
1 12 z 1 z=1
check
3
1 12 z 1
2 +
3(1 z 1 ) + 8(1 z 1 + 41 z 2 )
8
=
2
1 z 1
1 12 z 1 (1 z 1 )
=
5 5z 1 + 2z 2

1 12 z 1 (1 z 1 )
6=
1 + z 1
2
1 12 z 1 (1 z 1 )
General form of terms in PFE for pole p` with multiplicity m`

A` , m` 1
A` , m`
A` , 1
m +
m` 1 + ... + (1 p z 1 )
1
(1 p` z 1 ) `
`
(1 p` z )
Now we have
1 + z 1
A1,2
A1,1
A2
+
=
+

2
2
1
1
1 z 1
1 2z
1 12 z 1 (1 z 1 )
1 12 z 1
A1,2 = 3, A2 = 8
2
To solve for A1,1 , multiply both sides by 1 21 z 1 .
2

1 12 z 1
1 + z 1
1 1
=
A
+
A
z
+
A
1
1,2
1,1
2
1 z 1
2
1 z 1
differentiate with respect to z 1

2
1
1 1
+
A
z
P (z)
1
=
0
+
A
1,1
2
2
2
2
(1 z 1 )
1
let z =
2
2
A1,1 = 2
2 = 4
(1 2)
Example 4
(numerator degree denominator degree)
Y (z) =
1 + z 2
1 12 z 1 12 z 2
reduce degree of numerator by long division so

3 z 1
1 + z 2
=
2
+
1 12 z 1 21 z 2
1 12 z 1 12 z 2
In general, we will have a polynomial of degree M N .
M
N
X
k=0
ZT 1
Bk z k
M
N
X
k=0
Bk (n k)
123
124
5.7
General Form of Response of LTI

Systems
If both X(z) and H(z) are rational

Y (z) = H(z)X(z)

PH (z)
PX (z)
=
QH (z) QX (z)
PH (z)
PX (z)
= N
N
H
X
Y
Y

H 1
X 1
1 p` z
1 p` z
`=1
Y (z) =
`=1
PY (z)
Ny
1 pY` z 1
`=1
PY (z) = PH (z)PX (z)

NY = NH + NX
X
pY` is a combined set of poles pH
` and p` .
Drop superscript/subscript Y.
Accounting for poles with multiplicity > 1
Y (z) =
P (z)
D
Y
1 p` z 1
m `
`=1
D - number of distinct poles

N=
D
X
m`
`=1
For M < N
Y (z) =
m`
D X
X
`=1 k=1
A`k
(1 p` z 1 )
5.7. GENERAL FORM OF RESPONSE OF LTI SYSTEMS
125
Each term under summation will give rise to a term in the output
y(n)
It will be causal or anticausal depending on location of pole relative
to ROC
Poles between origin and ROC result in causal terms
Poles separated from origin by ROC result in anticausal terms
For simplicity, consider only causal terms in what follows
Real pole with multiplicity 1

A
1 pz 1
ZT 1
Apn u(n)
126
Complex conjugate pair of poles with multiplicity 1

A
A
+
1
1 pz
1 p z 1
ZT 1
Apn u(n) + A (p )n u(n)

[Apn + A (p )n ] u[n] = |A|ejA (|p|ejp )n + |A|ejA (|p|ejp )n u(n)
= 2|A||p|n cos(pn + A)u(n)
sinusoid with amplitude 2|A||p|n
1. grows exponentially if |p| > 1
2. constant if p = 1
3. decays exponentially if |p| < 1
digital frequency d = p radians/sample
phase A
Examples
5.7. GENERAL FORM OF RESPONSE OF LTI SYSTEMS
127
Real Pole with multiplicity 2

A
(1
2
pz 1 )
ZT 1
?
ZT
Recall nan u(n)
az 1
2 , |z| > |a|

(1 az 1 )
"
#
A
pz 1
A
n+1
u(n + 1)
2 = (A/p)z
2 p (n + 1)p
1
1
(1 pz )
(1 pz )
A
(n + 1)pn+1 u(n + 1) = A(n + 1)pn u(n)
p
Example (A=1)
Similar results are obtained with complex conjugate poles that have
multiplicity 2
In general, repeating a pole with multiplicity m results in
multiplication of the signal obtained with multiplicity 1 by a
polynomial in n with degree m 1
128
Example
A
(1
ZT 1
3
pz 1 )
A(n + 1)(n + 2)pn u(n)
Stability Considerations
A system is BIBO stable if every bounded input produces a
bounded output
X
A DT LTI system is BIBO stable
|h(n)| <
n
|h(n)| < H(z) converges on the unit circle
A causal DT LTI system is BIBO stable all poles of H(z) are

strictly inside the unit circle
Stability and the general form of the response

1. real pole with multiplicity 1
Apn u(n) is bounded if |p| 1
2. complex conjugate pair of poles with multiplicity 1
2|A||p|n cos(pn + A)u(n) is bounded if p| 1
3. real pole with multiplicity 2
A(n + 1)pn u(n) is bounded if |p| < 1
Chapter 6
Discrete Fourier
Transform (DFT)
6.1
Chapter Outline

1. Derivation of the DFT
2. DTFT Properties and Pairs
3. Spectral Analysis via the DFT
4. Fast Fourier Transform (FFT) Algorithm
5. Periodic Convolution
6.2
Derivation of the DFT
Summary of Spectral Representations

Signal Type
CT, Periodic
CT, Aperiodic
DT, Aperiodic
DT, Periodic
Transform
CT Fourier Series
CT Fourier Transform
DT Fourier Transform
Discrete Fourier Transform
129
Frequency Domain
Discrete
Continuous
Continuous
Discrete
130
CHAPTER 6. DISCRETE FOURIER TRANSFORM (DFT)
Computation of the DTFT

X(ej ) =
x(n)ejn
n=
In order to compute X(ej ), we must do two things:

1. Truncate summation so that it ranges over finite limits
2. Discretize = k
Suppose we choose

x(n), 0 n N 1
1. x
e(n) =
0,
else
2. k = 2k/N, k = 0, ..., N 1
then we obtain
N
1
X
e
X(k)
x
e(n)ej2kn/N
n=0
This is the forward DFT. To obtain the inverse DFT, we could discretize
the inverseZDTFT:
2
1
x(n) = 2
X(ej )ejn d
0
2k
N
k =
Z 2
N 1
2 X
d
N
0
k=0
e
X(ej ) X(k)
ejn ej2k/N
This results in
x(n)
N 1
1 X e
X(k)ej2kn/N
N
k=0
Note approximation sign

e
1. truncated x(n) so X(k)
X(ej2k/N )
2. approximated integral by a sum
6.2. DERIVATION OF THE DFT
131
Alternate approach
Use orthogonality of complex exponential signals
Consider again
e
X(k)
=
N
1
X
x
e(n)ej2kn/N
n=0
For fixed 0 m N 1, multiply both sides by ej2km/N and sum

over k
N
1
X
j2km/N
e
X(k)e
=
k=0
N
1
X
"N 1
X
k=0
n=0
N
1
X
x
e(n)
j2km/N
e
=
X(k)e
k=0

x
e(n)
N
1
X
N 1
1 X e
X(k)ej2km/N
N
k=0
Summarizing
X(k) =
N
1
X
x(n)ej2kn/N
n=0
x(n) =
N 1
1 X
X(k)ej2kn/N
N
k=0
where we have dropped the tildes
ej2km/N
ej2k(nm)/N
1 ej2(nm)
1 ej2(nm)/N
x
e(n)N (n m)
= Nx
e(m)
x
e(m) =
N
1
X
n=0
n=0
k=0
x
e(n)e
N
1
X
n=0
N
1
X
#
j2kn/N
132
6.3
DFT Properties and Pairs
1. Linearity
DF T
a1 x1 (n) + a2 x2 (n) a1 X1 (k) + a2 X2 (k)

2. Shifting
DF T
x(n n0 ) X(k)ej2kn0 /N
3. Modulation
DF T
x(n)ej2k0 n/N X(k k0 )

4. Reciprocity
DF T
X(n) N x(k)
5. Parsevals relation
N
1
X
|x(n)|2 =
n=0
N 1
1 X
|X(k)|2
N
k=0
6. Initial value
N
1
X
x(n) = X(0)
n=0
7. Periodicity
x(n + mN ) = x(n)for all integers m
X(k + `N ) = X(k) for all integers `
8. Relation to DTFT of a finite length sequence
Let x0 (n) 6= 0 only for 0 n N 1
x0 (n)
Define x(n) =
DT F T
X0 (ej )
x0 (n + mN )
m
DF T
x(n) X(k)
Then X(k) = X0 (ej2k/N )
6.3. DFT PROPERTIES AND PAIRS
133
Transform Pairs
1. x(n) = (n),
X(k) = 1,
0nN 1
0 k N 1 (by relation to DTFT)
2. x(n) = 1, 0 n N 1
X(k) = N k, 0 k N 1 (by reciprocity)
3. x(n) = ej2k0 n , 0 n N 1
X(k) = N (k k0 ), 0 k N 1 (by modulation property)
4. x(n) = cos(2k0 n/N ), 0 n N 1
N
X(k) =
[(k k0 ) + (k (N k0 ))] ,
2

5. x(n) =
X(k) = ej
0k N 1
1, 0 n M 1, 0 n N 1
0, else
2k
N (M 1)/2
sin[2kM/(2N )]
(by relation to DTFT)
sin[2k/(2N )]
134
6.4
Spectral Analysis Via the DFT
An important application of the DFT is to numerically determine the

spectral content of signals. However, the extent to which this is possible
is limited by two factors:
1. Truncation of the signal - causes leakage
2. Frequency domain sampling - causes picket fence effect
Truncation of the Signal

Let x(n) = cos(0 n)
Recall that
X(ej ) = rep2 [( 0 ) + ( + 0 )]
1, 0 n N 1
0, else
sin(N/2)
Recall W (ej ) = ej(N 1)/2
sin(/2)
Note that W (ej ) = W (ej(+2m) ) for all integers m.
Also, let w(n) =
Now let xt (n) = x(n)w(n)
t-truncated
6.4. SPECTRAL ANALYSIS VIA THE DFT
135
By the product theorem

Z
1
Xt (ej ) =
W (ej() )X(ej )d
2
Z
1
=
W (ej() )rep2 [( 0 ) + ( + 0 )]d
2
Z
1
W (ej() )[( 0 ) + ( + 0 )]d
Xt (e ) =
2
Z
1
W (ej() )( 0 )d
=
2
Z
+
W (ej() )( + 0 )d
j
Xt (ej ) =
1
2

W (ej(0 ) ) + W (ej(+0 ) )
Frequency Domain Sampling

Let
xr (n) =
xt (n + mN )
r-replicated
Recall from relation between DFT and DTFT of finite length sequence
that
Xr (k) = Xt (ej2k/N )
i
1h
W (ej(2k/N 0 ) ) + W (ej(2k/N +0 ) )
=
2
Consider two cases:
136
1. 0 = 2k0 /N for some integer k0

Example: k0 = 1, N = 8, 0 = /4
i
1 h
Xr (k) =
W (ej2(kk0 )/N ) + W (ej2(k+k0 )/N )
2
Example: k0 = 1, N = 8
In this case:
N
Xr (k) =
[(k k0 ) + (k (N k0 ))] , 0 k N 1
2
as shown before. There is no picket fence effect.
2. 0 = 2k0 /N + /N for some integer k0
Example: k0 = 1, N = 8, 0 = 3/8
6.5. FAST FOURIER TRANSFORM (FFT) ALGORITHM
137
Picket fence effect:

Spectral peak is midway between sample locations
Each sidelobe peak occurs at a sample location
1. Both truncation and picket fence effects are reduced by increasing N
2. Sidelobes may be supressed at the expense of a wider mainlobe by
using a smoothly tapering window
6.5
Fast Fourier Transform (FFT)

Algorithm
The FFT is an algorithm for efficient computation of the DFT. It is not a

new transform.
Recall
X (N ) (k) =
N
1
X
x(n)ej2kn/N , k = 0, 1, ..., N 1
n=0
Superscript (N) is used to show the length of the DFT. For each value of
k, computation of X(k) requires N complex multiplications and N-1
additions.
Define a complex operation (CO) as 1 complex multiplication and 1
complex addition.
Computation of length N DFT then requires approximately N 2 COs.
To derive an efficient algorithm for computation of the DFT, we employ a
divide-and-conquer strategy.
138
Assume N is even.
X (N ) (k) =
N
1
X
x(n)ej2kn/N
n=0
N
1
X
X (N ) (k) =
N
1
X
x(n)ej2kn/N +
n=0,even
x(n)ej2kn/N
n=0,odd
N/21
x(2m)ej2k(2m)/N
m=0
N/21
x(2m + 1)ej2k(2m+1)/N
m=0
X (N ) (k) =
N/21
X
X
x(2m)ej2km/(N/2) + ej2k/N
x(2m + 1)ej2km/(N/2)
N/21
m=0
m=0
m = 0, ..., N/2 1
Let x0 (n) = x(2m),

x1 (n) = x(2m + 1),
m = 0, ..., N/2 1
Now have
(N/2)
X (N ) (k) = X0
(N/2)
(N/2)
(k) + ej2k/N X1
(k), k = 0, ..., N 1
(N/2)
(k) and X1
(k) are both periodic with period N/2,
Note that X0
while ej2k/N is periodic with period N.
Computation
Direct
X (N ) (k), k = 0, ..., N 1
N 2 COs
Decimation by a factor of 2
(N/2)
(k), k = 0, ..., N/2 1
N 2 /4 COs
(N/2)
(k), k = 0, ..., N/2 1
N 2 /4 COs
X0
X1
(N/2)
X (N ) (k) = X0
COs
(N/2)
(k) + ej2k/N X1
(k), k = 0, ..., N 1

Total
N 2 /2 + N COs
For large N, we have nearly halved computation.

Consider a signal flow diagram of what we have done so far:
If N is even,we can repeat the idea with each N/2 pt. DFT:
If N = 2M , we repeat the process M times reulting in M stages.
139
140
The first stage consists of 2 point DFTs:

X (2) (k) =
1
X
x(n)ej2kn/2
n=0
X (2) (0) = x(0) + x(1)

X (2) (1) = x(0) x(1)
Flow diagram of 2 pt. DFT Full example for N=8(M=3)
141
Ordering of Input Data
Computation (N = 2M )
M = log2 N stages
N COs/stage
Total: Nlog2 N COs
Comments
The algorithm we derived is the decimation-in-time radix 2 FFT
1. input in bit-reversed order
142

2. in-place computation
3. output in normal order
The dual of the above algorithm is the decimation-in-frequency
radix 2 FFT
1. input in normal order
2. in-place computation
3. output in normal order
The same approaches may be used to derive either
decimation-in-time or decimation-in-frequency mixed radix FFT
algorithms for any N which is a composite number
All FFT algorithms are based on composite N and require O(N log
N) computation
The DFT of a length N real signal can be expressed in terms of a
length N/2 DFT
6.6
Periodic Convolution
We developed the DFT as a computable spectral representation for DT

signals. However, the existence of a very efficient algorithm for calculating
it suggests another application. Consider the filtering of a length N signal
x(n) with an FIR filter containing M coefficients, where M << N
y(n) =
M
1
X
h(m)x(n m)
m=0
Computation of each output point requires M multiplications and M-1

additions.
Based on the notion that convolution in the time domain corresponds to
multiplication in the frequency domain, we perform the following
computations:
Compute X (n) (k) N log2 N COs
Extend h(n) with zeros to length N and compute
H (N ) (k) N log2 N COs
Multiply the DFTs
6.6. PERIODIC CONVOLUTION

Y (N ) (k) = H (N ) (k)X (N ) (k) N complex mutliplications
Compute inverse DFT

ye(n) = DF T 1 Y (N ) (k) N log2 N COs
Total: 3N log2 N COs + N complex multiplications
The computation/output point is 3 log2 N COs+ 1 complex
multiplication.
How Periodic Convolution Arises

Consider inverse DFT of product Y [k] = H[k]X[k]
y(n) =
N 1
1 X
H[k]X[k]ej2kn/N
N
k=0
Substitute for X[k] in terms of x[n]

(N 1
)
N 1
X
1 X
j2km/N
H[k]
x[m]e
ej2kn/N
y(n) =
N
m=0
k=0
Interchange order of summations

)
( N 1
N
1
X
1 X
j2k(nm)/N
H[k]e
y(n) =
x[m]
N
m=0
k=0
N
1
X
x[m]h[n m]
m=0
Note periodicity of sequence h[n]

h[n m] = h[(n m) mod N ]
143
144
Periodic vs. Aperiodic Convolution

Aperiodic convolution
y[n] = h[n] x[n] =
h[n m]x[m]
m=
Periodic convolution (N=9)
y[n] = h[n] x[n] =
N
1
X
m=0
h[(n m)modN ]x[m]
6.6. PERIODIC CONVOLUTION
Comparison
Aperiodic convolution
145
146
Periodic convolution (N = 9)
Zero Padding to Match Aperiodic Convolution

Suppose x[n] has length N and h[n] has length M. Periodic convolution
with length M+N-1 will match aperiodic result
Chapter 7
Digital Filter Design

7.1
Chapter Outline

1. Overview of filter design problem
2. Finite impulse response filter design
3. Infinite impulse response filter design
7.2
Overview
Filter design problem consists of three tasks

1. Specification - What frequency response or other
characteristics of the filter are desired?
2. Approximation - What are the coefficients or, equivalently,
poles and zeros of the filter that will approximate the desired
characteristics?
3. Realization - How will the filter be implemented?
Consider a simple example to illustrate these tasks.
147
148
CHAPTER 7. DIGITAL FILTER DESIGN
Filter Design Example

An analog signal contains frequency components ranging from DC
(0 Hz) to 5 kHz
Design a digital system to remove all but the frequencies in the
range 2-4 kHz
Assume signal will be sampled at 10 kHz rate
Specification
Ideal analog filter
Ideal digital filter
7.2. OVERVIEW
Filter Tolerance Specification
pl , pu - lower and upper passband edges

sl , su - lower and upper stopband edges
l , u - lower and upper stopband ripple
- passband ripple
Approximation
Design by positioning poles and zeros (PZ plot design)
149
150

Transfer function for PZ plot filter

z ej/10 z ej3/10 z ej9/10

H(z) = K
z 0.8ej5/10 z 0.8ej7/10

z ej/10 z ej3/10 z ej9/10

z 0.8ej5/10 z 0.8ej7/10
Choose constant K to yield unity magnitude response at center of
passband
Frequency response of PZ plot filter
Comments on PZ plot design

1. Passband asymmetry is due to extra zeros in lower stopband
2. Phase is highly nonlinear
3. Zeros on unit circle drive magnitude response to zero and
result in phase discontinuities
4. Pole-zero plot yields intuition about filter behavior, but does
not suggest how to meet design specifications:
7.2. OVERVIEW
151
Realization
Cascade form of transfer function

z ej/10 z ej3/10 z ej9/10

H(z) = K
z 0.8ej5/10 z 0.8ej7/10

z ej/10 z ej3/10 z ej9/10

z 0.8ej5/10 z 0.8ej7/10
Cascade Form Realization

Cascade form in second order sections with real-valued coefficients
H(z) = K[z 2 2 cos(/10)z + 1]

z 2 2 cos(3/10)z + 1
2
z 1.6 cos(5/10)z + 0.64

z 2 2 cos(9/10)z + 1
2
z 1.6 cos(7/10)z + 0.64
Convert to negative powers of z
H(z) = K[1 2 cos(/10)z 1 + z 2 ]

1 2 cos(3/10)z 1 + z 2
1 1.6 cos(5/10)z 1 + 0.64z 2

1 2 cos(9/10)z 1 + z 2
1 1.6 cos(7/10)z 1 + 0.64z 2
(Ignoring overall time advance of 2 sample units)
Overall System
152
Second order stages

Hi (z) =
1 + a1i z 1 + a2i z 2
, i = 1, 2, 3
1 b1i z 1 b2i z 2
yi [n] = xi [n] + a1i xi [n 1] + a2i xi [n 2] + b1i yi [n 1] + b2i yi [n 2]
Parallel Form Realization

Expand transfer function in partial fractions
H(z) =
a01 + a11 z 1
a02 + a12 z 1
d0 + d1 z 1 + d2 z 2 +
+
1
2
1 b11 z b21 z
1 b12 z 1 b22 z 2
7.2. OVERVIEW
Direct Form Realization

Multiply out all factors in numerator and denominator of transfer
function
a0 + a1 z 1 + ... + a6 z 6
H(z) =
1 b1 z 1 ... b4 z 4
153
154
7.3
Finite Impulse Response Filter Design
Ideal Impulse Response

Consider same ideal frequency response as before
Inverse DTFT
Z
1
h[n] =
H()ejn d
2 (
)
Z 2/5
Z 4/5
1
jn
jn
h[n] =
e d +
e d
2
4/5
2/5
h[n] =
i h
io
1 nh j2n/5
e
ej4n/5 + ej4n/5 ej2n/5
j2n
h[n] =
i
h
io
1 n j3n/5 h jn/5
e
e
ejn/5 + ej3n/5 ejn/5 ejn/5
j2n

n
n
1
1
+ ej3n/5 sinc
h[n] = ej3n/5 sinc
5
5
5
5
h[n] = 0.4 cos(3n/5)sinc(n/5)
7.3. FINITE IMPULSE RESPONSE FILTER DESIGN
155
Approximation of Desired Filter Impulse Response by

Truncation
FIR filter equation
y[n] =
M
1
X
am x[n m]
m=0
Filter coefficients (assuming M is odd)

am = hideal [m (M 1)/2], m = 0, ..., M 1
Truncation of Impulse Response
156
Frequency Response of Truncated Filter

Analysis of Truncation
ideal infinite impulse response - hideal [n]
Actual, finite impulse response - hactual [n]
Windowsequence - w[n]
1 , |n| (M 1)/2
w[n] =
0 , else
Relation between ideal and actual impulse responses
hactual [n] = hidealZ[n]w[n]
1
Hactual () =
Hideal ()W ( )d
2
Effect of Filter Length
DTFT of Rectangular Window

n=(M 1)/2
W () =
ejn
n=(M 1)/2
let m=n+(M-1)/2
=
M
1
X
ej[m(M 1)/2]
m=0
= ej(M 1)/2
=
sin(M/2)
sin(/2)
1 ejM
1 ej
157
158
Graphical View of Convolution

1
Hactual () =
2
Hideal ()W ( )d
Relation between Window Attributes and Filter

Frequency Response
Parameter
W ()
Mainlobe width
Area of first sidelobe
Hactual ()
Transition bandwidth
Passband and stopband ripple
159
Design of FIR Filters by Windowing

Relation between ideal and actual impulse responses
hactual [n] = hideal[n] w[n]
Choose window sequence w[n] for which DTFT has
1. minimum mainlobe width
2. minimum sidelobe area
Kaiser window is the best choice
1. based on optimal prolate spheroidal wavefunctions
2. contains a parameter that permits tradeoff between mainlobe
width and sidelobe area
Kaiser Window
(
I0 [(1[(n)/]2 )1/2 ]
,
I0 ()
0N M 1
0,
else
= (M 1)/2
I0 (.) zeroth-order modified Bessel function of the first kind
w[n] =
160
Effect of Parameter
Filter Design Procedure for Kaiser Window

1. Choose stopband and passband ripple and transition bandwidth
2. Determine paramter
A =
20 log10
0.112(A 8.7)
0.5842(A 21)0.4 + 0.07886(A 21)
=
0.0
, A > 50
, 21 A 50
, A < 21
3. Determine filter length M

A8
M = 2.285
+1
4. Example
= 0.1 radians/sample= 0.05 cycles/sample
= 0.01
= 3.4
M = 45.6
Filter Designed with Kaiser Window
161
162
Kaiser Filter Frequency Response
M = 47
= 3.4
= 0.01 = 40 dB
= 0.05
sl = 0.175
pl = 0.225
pu = 0.375
su = 0.425
Optimal FIR Filter Design

Although the Kaiser window has certain optimal properties, filters
designed using this window are not optimal
Consider the design of filters to minimize integral mean-squared
frequency error
Problem:
Choose hZactual [n] = (M 1)/2, ..., (M 1)/2 to minimize
1
E=
|Hactual () Hideal ()|2 d
2
163
Minimum Mean-Squared Error Design

Using Parsevals relation, we obtain a direct solution to this problem
Z
1
E=
|Hactual () Hideal ()|2 d
2
X
=
|hactual [n] hideal [n]|2
n=
(M 1)/2
|hactual [n] hideal [n]|2
n=(M 1)/2
(M +1)/2
n=
|hideal [n]|2 +
|hideal [n]|2
n=(M +1)/2
Solution
Set hactual [n] = hideal [n], n = (M 1)/2, ..., (M 1)/2
Minimum error is given by
(M +1)/2
X
X
E=
|hideal [n]|2 +
|hideal [n]|2
n=
n=(M +1)/2
This is the truncation of the ideal impulse response

We conclude that original criteria of minimizing mean-squared error
was not a good choice
What was the problem with truncating the ideal impulse response?
(See figure on following page)
Minimax (Equi-ripple) Filter Design

No matter how large we pick M for the truncated impulse response,
we cannot reduce the peak ripple
Also, the ripple is concentrated near the band edges
Rather than focussing on the integral of the error, lets consider the
maximum error
Parks and McClellan developed a method for design of FIR filters
based on minimizing the maximum of the weighted frequency
domain error
164
Effect of Filter Length: Parks-McClellan Equiripple

Filters
7.4. INFINITE IMPULSE RESPONSE FILTER DESIGN
7.4
165
Infinite Impulse Response Filter Design
IIR vs. FIR Filters

FIR Filter Equation
y[n] = a0 x[n] + a1 x[n 1] + ... + aM 1 x[n (M 1)]
IIR Filter Equation
y[n] = a0 x[n] + a1 x[n 1] + ...aM 1 x[n (M 1)] b1 y[n 1]
b2 y[n 2] ... bN y[n N ]
IIR filters can generally achieve given desired response with less
computation than FIR filters
It is easier to approximate arbritrary frequency response
characteristics with FIR filters, including exactly linear phase
Infinite Impulse Response (IIR) Filter Design:

Overview
IIR digital filter designs are based on established methods for
designing analog filters
Approach is generally imited to frequency selective filters with ideal
passband/stopband characteristics
Basic filter type is low pass
166

Achieve highpass or bandpass via transformations
Achieve multiple stop/pass bands by combining multiple filters with
single pass band
IIR Filter Design Steps

Choose prototype analog filter family
1. Butterworth
2. Chebychev Type I or II
3. Elliptic
Choose analog-digital transformation method
1. Impulse Invariance
2. Bilinear Transformation
Transform digital filter specifications to equivalent analog filter
specifications
Design analog filter
Transform analog filter to digital filter
Perform frequency transformation to achieve highpass or bandpass
filter, if desired
Prototype Filter Types
Transformation via Impulse Invariance

Sample impulse response of analog filter
hd [n] = T ha (nT )
Hd () =
X
k

Ha (f k/T )

f=
Note that aliasing may occur
Impulse Invariance
Implementation of digital filter
2T
167
168

1. Partial fraction expansion of analog transfer function
(assuming all poles have multiplicity 1)
Ha (s) =
N
X
Ak
s sk
k=1
2. Inverse Laplace transform

ha (t) =
N
X
Ak esk t u(t)
k=1
3. Sample Impulse response

hd [n] = T ha (nT ) = T
N
X
Ak esk nT u(nT )
k=1
=T
N
X
Ak pnk u[n], pk = esk T
k=1
4. Take Z transform
Hd (z) =
N
X
k=1
T Ak
1 pk z 1
5. Combine terms
a0 + a1 z 1 + ...aM 1 z (M 1)
1 + b1 z 1 + b2 z 2 + ...bN z N
6. Corresponding filter equation
Hd (z) =
y[n] = ao x[n] + a1 x[n 1] + ...aM 1 x[n (M 1)] b1 y[n

1] b2 y[n 2] ... bN y[n N ]
Impulse Invariance Example

Design a second order ideal low pass digital filter with cutoff at
d = /5 radians/sample (Assume T=1)
1. Analog cutoff frequency
fc =
d
2T
= 0.1Hz
c = 2fc = /5 rad/sec
Use Butterworth analog prototype
Ha (s) =
0.3948
[s (0.4443 + j0.4443)][s (0.4443 j0.4443)]
169
Apply partial fraction expansion

Ha (s) =
j0.4443
j0.4443
+
[s (0.4443 + j0.4443)] [s (0.4443 j0.4443)]
Compute inverse Laplace Transform

ha (t) = [j0.4443e(0.4443+j0.4443)t + j0.4443e(0.4443j0.4443)t ]u(t)
Sample impulse response
h
hd [n] = j0.4443e(0.4443+j0.4443)n
i
+j0.4443e(0.4443j0.4443)n u[n]

= j0.4443(0.641ej0.4443 )n

+j0.4443(0.641ej0.4443 )n u[n]
Compute Z transform
j0.4443
j0.4443
+
1 0.6413ej0.4443 z 1
1 0.6413ej0.4443 z 1
0.2450z 1
=
1 1.1581z 1 + 0.4113z 2
Hd (z) =
Combine terms
Hd (z) =
(1
0.2450z 1
0.6413ej0.4443 z 1 )
0.6413ej0.4443 z 1 )(1
Find difference equation

y[n] = 0.2450x[n 1] + 1.1581y[n 1] 0.4113y[n 2]
Impulse Invariance Method-Summary

Preserves impulse response and shape of frequency response, if
there is no aliasing
Desired transition bandwidths map directly between digital and
analog frequnecy domains
Passband and stopband ripple specifications are identical for both
digital and analog filter, assuming that there is no aliasing
170

The final digital filter design is independent of the sampling interval
parameter T
Poles in analog filter map directly to poles in digital filter via
transformation pk = esk T
There is no such relation between the zeros in the two filters
Gain at DC in a digital filter may not equal unity, since sampled
impulse response may only approximately sum to 1
Bilinear Transformation Method

Mapping between s and z

2 1 z 1
s=
T 1 + z 1

2 1 z 1
Hd (z) = Ha
T 1 + z 1
z=
1 + (T /2)s
1 (T /2)s
Mapping between s and z planes

Mapping between analog and digital frequencies
Bilinear Transformation-Example No. 1

Design low-pass filter
1. cutoff frequency dc = /5 rad/sample
2. transition bandwidth and ripple
d = 0.2 radians/sample= 0.1 cycles/sample
= 0.1
3. use Butterworth analog prototype
4. set T = 1
|Hd (d 1)| = 1
|Hd (d 2)| =
Map digital to analog frequencies
171
172

a =
2
T
tan(d /2)
d1 = /5 0.1
a1 = 0.3168 rad/sec
d2 = /5 + 0.1
a2 = 1.0191 rad/sec
Solve for filter order and cutoff frequency
1
1+(a /c )2N
|Ha (a )|2 =
|Ha (a1 )| = 1
|Ha (a2 )| =

1 log[(|Ha (a2 )|2 1)/(|Ha (a1 )|2 1)]
N=
2
log(a2 /a1 )
a2
ac =
(|Ha (a2 )|2 1)1/(2N )
Result
N =3
ac = 0.4738
Determine transfer function of analog filter
Ha (s)Ha (s) =
1
1+(s/jac )2N
Poles are given by
sk = ac e
2k
+
+
2
2N
2N
, k = 0, 1, ..., 2N 1
Take N poles with negative real parts for H(s)

s0 = 0.2369 + j0.4103 = 0.4738ej2/3
s1 = 0.4738 + j0.0000 = 0.4738ej
s2 = 0.2369 j0.4103 = 0.4738ej2/3
Transfer function of anlaog filter
3
ac
Ha (s) =
(s s0 )(s s1 )(s s2 )

Transform to discrete-time
filter

2 1 z 1
Hd (z) = Ha
T 1 + z 1
3
ac
(1 + z 1 )3
8(1 z 1 s0 )(1 z 1 s1 )(1 z 1 s2 )
0.0083 + 0.0249z 1 + 0.0249z 2 + 0.0083z 3
=
1.0000 2.0769z 1 + 1.5343z 2 0.3909z 3
Hd (z) =
Bilinear Example No.2

Design low-pass filter dc = /5 rad/sample
173
174
1. Cutoff frequency
2. Transition bandwidth and ripple
d = 0.1 radians/sample = 0.05 cycles/sample
= 0.01
3. Use Butterworth analog prototype
4. set T = 1
Result:N = 13
175
Frequency Transformations of Lowpass IIR Filters

Given a prototype lowpass digital filter design with cutoff frequency
c , we can transform directly to:
1. a lowpass digital filter with a different cutoff frequency
2. a highpass digital filter
3. a bandpass digital filter
4. a bandstop digital filter
1
General form of transformation H(z) = Hlp (Z)}Z = G(z 1 )

Example - lowpass to lowpass
1. Transformation
z 1
Z 1 =
1 z 1
2. Associated design
formula
c
sin c
2

=
c
sin c +
2
176
Chapter 8
Random Signals
8.1
Chapter Outline

1. Single Random Variable
2. Two Random Variables
3. Sequences of Random Variables
4. Estimating Distributions
5. Filtering of Random Sequences
6. Estimating Correlation
8.2
Single Random Variable
Random Signals
Random signals are denoted by upper case X and are completely
characterized by density function fX (x)
Z b
P (a X b) =
fX (x)dx
a
fX (x)x = P (x X x + x)
177
178
CHAPTER 8. RANDOM SIGNALS
The cumulative
distribution function is FX (x) = P (X x)
Z x
fX (x)dx
FX (x) =
dFX (x)
= fX (x)
dx
Properties of density function
1. fX (x) 0
Z
2.
fX (x)dx = 1
Example 1
fX (x) = ex u(x)
Z
fX (x)dx =
0
ex dx
0
= ex
=1
What is the probability X 1?

Z
P (X 1) =
ex dx
0
= ex 1
= 1 e
8.2. SINGLE RANDOM VARIABLE

Expectation
Z
E {g(X)} =
g(x)fX (x)dx where g is a function of X
Special Cases:
1. Mean value
g(x) = x
Z
X = E {X} =
xfX (x)dx
2. Second moment
g(x) = x2
Z

X 2 = E X 2 = x2 fX (x)dx
3. Variance
g(x) = (X X)2
Expectation is a linear operator
E {ag(x) + bh(x)} = aE {g(x)} + bE {h(x)}

2
X
= E (X X)2
n
o
2
= E X 2 2XX + X

2
= E X2 X
Example 2
Z
E {X} =
xex dx
Integrate
by parts
Z
Z
udv = uv vdu
u=x
dv = ex
Z
Z

ex dx
xex dx = xex 0 +
0
0

1
= xex 0 ex
0
1
=
du = 1
v = ex
179
180
Example 3

E X2 =
x2 ex dx
Integrate by parts
u = x2 du = 2x v = ex
Z
0
dv = ex

x2 ex dx = x2 ex 0 + 2
=
xex dx
2
1
2
2
= 1 X = 1 X
= 1 X = 1
2
=
X
X and Y have the same mean and same standard deviation
Transformations of a Random Variable

Y = aX + b
Y = aX + b
8.3. TWO RANDOM VARIABLES
181

Y2 = E (Y Y )2
n
2 o
= E aX + b (aX + b)

= a2 E (X X)2
2
= a2 X
FY (y) = P {Y y}
= P {aX + b y} assume a> 0

yb
=P X
a

yb
= FX
a
1
fY (y) = fX
a
yb
a

1
yb

In general fY (y) = fX
a
a
8.3
Two Random Variables
1. Joint density function

Z
P {a X b c Y d} =
fXY (x, y)dxdy

a
2. Joint distribution
Z x Zfunction
y
FXY (x, y) =
fXY (x, y)dxdy
3. MarginalZdensities
fX (x) = fXY (x, y)dy
Z
fY (y) = fXY (x, y)dx
4. Expectation
Z Z
E {g(X, Y )} =
g(x, y)fXY (x, y)dxdy
182
5. Linearity
E {ag(x, y) + bh(x, y)} = aE {g(x, y)} + bE {h(x, y)}
6. Correlation
g(x, y) = xy
7. Covariance
g(x, y) = (x X)(y Y ) XY = E {g(x, y)}
8. Correlation coefficient
2
XY
= E {g(x, y)}

1
=
E {XY } E XY E XY + X Y
X Y

1
=
XY X Y
X Y
9. Independence
fXY (x, y) = fX (x)fY (y)
2
XY
= 0 and XY = X Y
Example 4

2,
0,
fXY (x, y) =
fX (x) =
fY (y) =
0yx1
else
0
0,
Z 1
2dy = 2x, 0 x 1
else
2dx = 2(1 y), 0 y 1
0,
else
8.3. TWO RANDOM VARIABLES
XY =
2xydxdy
x=0
Z 1
y=0
Z
x
x=0
1

2ydy dx
y=0
x3 dx
=
x=0

4 1
1
x
=
4 0
4
X=
x(2x)dx =
0
2
2 31
x | =
3 0
3
1
Y =
3
X2
Z
=
x2 (2x)dx
1
2 4
x
4 0
1
=
2
183
184

1 4
2 9
1
=
18
= y2
2
X
=
XY X Y
X Y
1/4 2/9
=
1/18
1
=
2
2
XY
=
8.4
Sequences of Random Variables
Characteristics
X1 , X2 , X3 , ..., XN denote a sequence of random variables
Their
behavior is completely characterized by

P xl1 X1 xu1 , xl2 X2 xu2 , ..., xlN XN xuN
Z
xu
1
xu
2
xu
N
...
xl1
xl2
xlN
fX1 ,X2 ,...,XN (x1 , x2 , ..., xN )dx1 , dx2 , ..., dxN
It is apparent that this can get complicated very quickly

In practice, we consider only 2 cases:
1. X1 , X2 , ..., XN are mutually independent
N
Y
fX1 ,X2 ,...,XN (x1 , x2 , ..., xN ) =
fXi (xi )
i=1
2. X1 , X2 , ..., XN are jointly Gaussian

Example 5
2
X1 , X2 , ..., Xn are i.i.d with mean X and std. deviation X
N1 , N2 , ..., Nn are i.i.d with zero mean
Yi = aXi + Ni
8.4. SEQUENCES OF RANDOM VARIABLES

E {Yi } = aE {Xi } + E {Ni }
Y = aX
n
o
2
Y 2 = E (aXi + Ni )
= a2 X 2 + N 2
Y2 = Y 2 Y
2
2
= a2 X 2 + N
a2 X
2
2
2
= a2 X
+ N
XY = E {Xi (aXi + Ni )}
= aX 2
2
XY
aX 2 a X
2
1/2
2 + 2 )
X (a2 X
N
X

=a p
2 + 2
a 2 X
N
=r
1+
1
2
N
2
2
(a X
)
185
186
187
188
189
190
8.5. ESTIMATING DISTRIBUTIONS
8.5
191
Estimating Distributions
So far, we have assumed that we have an analytical model for the

probability densities of interest.
This was reasonable for quantization. However, suppose we do not know
the underlying densities. Can we estimate it from observations of the
signal?
Suppose we observe a sequence of i.i.d rvs
X1 , X2 , ..., XN
let E(X) = X
How do we estimate X ?
X =
N
1 X
Xn
N n=1
How good is the estimate?

X is actually a random variable, lets call it Y
)
(
N
1 X
Xn
E {Y } = E
N n=1
= E {X}
= X
!2
N
1 X
o
n
2
|Xn X |
E |Y X | = E
N
n=1
=
N
N
1 X X
N
E {[Xm X ][Xn X ]}
N 2 m=1 n=1
N
1 X
2
N X
N 2 m=1
1 2
N X
192
How do we estimate variance?

N
X
2
2 = 1
[Xn X ] [X X ]]
X
N n=1
N
N
1 X
1 X
2
|Xn X | 2[X X ]
[Xn X ]
N n=1
N n=1
N
1 X
[X X ]2
N n=1
N
1 X
2
|Xn X | [X X ]2
N n=1
call this r.v Z

N
n
o
1 X 2
2
X E |X X |
E {Z} =
N n=1
!2
N
1 X
n
o
2
E |X X | = E
[Xn X ]
N
n=1
=
N
N
1 XX
E {[Xm X ][Xn X ]}
N 2 m=1 n=1
1
2
N X
N2
1 2
= X
N
=
1 2
2
E {Z} = X
X
N

N 1
2
= X
N
So the estimate is not unbiased
We define unbiased estimate as:
N
2
X
=
2
unb
N 1 X
N
1 X
2
=
(Xn c
X)
N 1 n=1
8.5. ESTIMATING DISTRIBUTIONS
193
N
1 X
2
|Xn c
X|
N n=1
(N
)
N
X
1 X
2
2
|Xn | 2c
Xn + c
=
X
X
N n=1
n=1
2
X
=
2
X
=
unb
N
1 X
|Xn |2 2X
N n=1
N
2
N 1 X
So far we have seen how to estimate statistics such as mean and variance
of a random variable.
These ideas can be extended to estimation of any moments of the random
variable. But suppose we want to know the actual density itself.
How do we do this?
We need to compute the histogram.
Consider a sequence of i.i.d r.vs with density fX (x)
We divide the domain of X into intervals

Ik = x : k 21 x k + 12
194
We then define the r.v.s

1, X Ik
Uk =
0, else
We observe a sequence of i.i.d r.v.s Xn n = 1, ..., N and compute the
corresponding r.v.s Unk
Finally we let
PN
Z k = N1 n=1 Unk
Then since this r.v has the same form as the estimator of mean which we
analyzed earlier, we know that

E Z k = E Unk
1 2
2
Z
k
k =
N U

Now what is E Unk

E Unk = 1.P {Xn Ik } + 0.P {Xn
/ Ik }
= P {Xn Ik }
Z (k+ 12 )
fX (x)dx
=
(k 12 )
So we see that we are estimating the
integral of the density function over
the range k 12 x k + 12
According to the fundamental theorem of calculus, there exists

k 21 k + 12
such that
Z (k+ 12 )
fX ( ) =
fX (x)dx
(k 12 )

= E Zk
Assuming fX (x) is continuous,

As 0

E Z k = fX ( ) fX (k)
defined as
2
Uk
= 2
= Uk
U k
8.6. FILTERING OF RANDOM SEQUENCES
195
For fixed fX (x), U k = P {X Ik } will increase with increasing so there

is a tradeoff between measurement noise and measurement resolution.
Of course, for fixed and fX (x), we can always improve our result by
increasing N.
What is the variance of Unk ?
n
2 o
E Unk
= 12 P {X Ik } + 02 P {X
/ Ik }
= P {X Ik }
= Uk
2
2
so U
k = Uk (Uk ) = Uk (1 Uk )
We would expect to operate in range where is small so

2
k
U
k U
8.6
Filtering of Random Sequences
Let us define X (n) = E {Xn }

rXX (m, n) = E {Xm Xn }
196
Special case
A process is w.s.s. if
1. X (n) X
2. rXX (m, n) = rXX (n m, 0) rXX (n m)
Convention:
rAB [a, b] = rAB [b a]
Consider
X
Y [n] =
h[n m]X[m]
m
E {Yn } =
h[n m]E {Xm }
If Xn is w.s.s
E {Yn } =
h[n m]X
h[n]X
= X [x]is a constant
To find rY Y , we first define cross-correlation

rXY [m, n] = E {Xm , Yn }
(
)
X
= E Xm
h[n k]Xk
k
h[n k]E {Xm Xk }
h[n k]rXX [k m]
let l=k-m k = m + l
rXY [m, n] =
h[n (m + l)]rXX (l)
h[n m l]rXX (l)
rXY [m, n] = rXY [n m]

then
rY Y [m, n] = E {Ym Yn }
(
)
X
h[n k]Xk
= E Ym
k
h[n k]rXY [m k]
let l = m k k = m l
X
=
h[n (m l)]rXY (l)
k
h[n m + l]rXY (l)
h[n m l]rXY (l)
so rY Y [m, n] = rY Y [n m]
summarizing:
If X is w.s.s, Y is also w.s.s
197
198
Interpretation of Correlation
Recall that for two different random variables X and Y, we defined
covariance to be
2
XY
= E {(X X ) (Y Y )}
and correlation coefficient

2
XY = XY
X Y
as measures of how related X and Y are,i.e whether or not they tend to
vary together?
Since X[m] and X[m + n] or X[m] and Y [m + n] are just two different
pairs of r.v.s, we can see that
rXX [n] = E {X[m]X[m + n]}
rXY [n] = E {X[m]Y [m + n]}
are closely related to covariance and correlation coefficient.
Suppose X and Y have zero mean, these r.vs have the same
interpretation. How related are Xm and Xm+n or Ym and Ym+n ?
Lets consider some examples:
Suppose X[n] is i.i.d with zero mean.
Since mean of a w.s.s process is just a constant offset or shift, we can let
it equal zero without loss of generality.
Then rXX [n] = E {X[m]X[m + n]}
2
X , n = 0
rXX [n] =
0,
else
2
rXX [n] = X
[n]
Now suppose
Y [n] =
7
X
2m X[n m]
m=0
This is a filtering operation with

h[n] = {2n [u[n] u[n 8]}
We already know that
rXY [n] = h[n] rXX [n]
2
= X
h[n]
199
rXY [n] = E {X[m]Y [m + n]}

What about the sequence Y [n]?
Is it i.i.d?
Recall
rY Y [n] = h[n] rXY [n]
2
= h[n] X
h[n]
X
2
= X
h[n m]h[m]
m
2
X
h[n + m]h[m]
Cases
1. n 7,
rY Y [n] = 0
2. 7 n 0
2
rY Y [n] = X
n+7
X
2 n
2(n+m) 2m = X
2
m=0
n+7
X
m=0
2(n+8)
2 n 1 2
= X
2
1 22
h
i
4
2
2n 2(n16)
= X
3
3. 0 n 7
rY Y [n] =
2 n
X
2
7
X
22m
l =m+n
m=n
2 n
= X
2
n+7
X
22(ln)
l=0
2 n
= X
2
n+7
X
2 n
22l = X
2
l=0
4 2 n
= X
[2 2n16 ]
3
1 22(n+8)
1 22
22m
200
4. n > 7
rY Y [n] = 0
Summarizing
0,
2
4 X
[2n 2(n16) ],
3
rY Y [n] =
4 2 [2n 2(n16) ],
3
0,
n < 7
7 n 0
0n7
n>7
201
202
8.7. ESTIMATING CORRELATION
8.7
203
Estimating Correlation
Assume X and Y are jointly w.s.s processes

E[Xn ] = X
E[Yn ] = X
rXY [m, m + n] = E {X[m]Y [m + n]}
= rXY [n]
Suppose we form the sum
Z[n] =
M 1
1 X
X[m]Y [m + n]
M m=0
Consider
E {Z[n]} =
M 1
1 X
E {X[m]Y [m + n]}
M m=0
E {Z[n]} =
M 1
1 X
rXY [n] = rXY [n]
M m=0
So we have an unbiased estimator of correlation.

Lets look a little more closely at what the computation involves:
Suppose we computed
rXY
[n] = Z[n] for the interval N n N
=
M 1
1 X
X[m]Y [m + n]
M m=0
For n = N
X[0]X[1]...X[M 1]
Y [N ]Y [N + 1]...Y [N + M 1]
For n = 0
X[0]X[1]...X[M 1]
Y [0]Y [1]...Y [M 1]
For n = N
X[0]X[1]...X[M 1]
Y [N ]Y [N + 1]...Y [N + M 1]
In each case, we used same data set for X, bu the data used for Y varied.
204
Suppose instead we collect Z sets of data

X[0]...X[M 1]
Y [0]...Y [M 1]
and use this data to compute estimate for every n
let Xtr [n] = X[n] {u[n] u[n m]}
let Ytr [n] = Y [n] {u[n] u[n m]}
Then let
v[n] =
Xtr [m]Ytr [m + n]
M
1
X
X[m]Y [m + n] {u[m + n] u[m + n M ]}
m=0
E {V [n]} =
M
1
X
E {X[m]Y [m + n]} {u[m + n] u[m + n M ]}
m=0
M
1
X
rXY [n] {u[m + n] u[m + n M ]}
m=0
= rXY [n]
M
1
X
{u[m + n] u[m + n M ]}
m=0
So an unbiased estimate of correlation is given by

X
1
1 M

rXY
[n] =
X[m]Ytr [m + n]
M n m=0
Cover image taken from website

http://www.fmwconcepts.com/imagemagick

438book2012 Vol1 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

438book2012 Vol1 PDF

Uploaded by

Copyright:

Available Formats

Digital Signal Processing in a Nutshell (Volume I)

Jan Allebach, Krithika Chandrasekar

Digital Signal Processing in a

6 Discrete Fourier Transform (DFT)

7 Digital Filter Design

There are three different types of signals:

Figure 1.2(a) Different types of signals.

1.3. SIGNAL CHARACTERISTICS

x(n) may or may not be sampled from an analog waveform, i.e.

Equivalent Representations for DT signals

Figure 1.2(b) Equivalent representations for DT signals.

Signals can be:

|x(n)|2 , for any n0

root-mean-square (rms) value

Figure 1.3(a) Signal for Example 1.

1.3. SIGNAL CHARACTERISTICS

The signal is:

Figure 1.3(b) Signal for Example 2.

1.4. SIGNAL TRANSFORMATIONS

4. Scaling and Shifting

right edge of pulse:

What does this signal look like?

1.4. SIGNAL TRANSFORMATIONS

Note that n0 must be an integer. Non-integer delays can only be

1.5. SPECIAL SIGNALS

2. Digital frequencies 1 = 0 and 2 = 0 + 2k are equivalent

1.6. COMPLEX VARIABLES

1.6. COMPLEX VARIABLES

Some useful identities

1.6. COMPLEX VARIABLES

Alternate form for polar coordinate representation of complex number

Multiplication of two complex numbers can be done using rules for

Complex exponential signal

DT x(n) = Aej(d n+)

1.7. SINGULARITY FUNCTIONS

In general, if 1 = + and 2 = ( ), then

In the limit, we obtain

We will use impulses in three ways:

1.7. SINGULARITY FUNCTIONS

x(t) (t t0 )dt = x()

for some which satisfies

As 0, t0 and x() x(t0 )

provided x(t) is continuous at t0 .

Indefinite integral of (t)

More general form of sifting property

x(t0 ), a < t0 < b

DT Impulse Function (unit sample function)

1.8. COMB AND REPLICATION OPERATIONS

Comb and Replication Operations

Sifting property yields a single sample at t0 . Consider multiplying x(t) by

Area of each impulse = value of sample there.

xp (t) = repT [x0 (t)] =

Note that x0 (t) = xp (t)rect

In this chapter, we will discuss:

A system is a mapping which assigns to each signal x(t) in the input

let x(t) = sin(2f t)

Look at particular values of f

Effect of filter on output:

let x3 (t) = a1 x1 (t) + a2 x2 (t)

2.3. SYSTEM PROPERTIES

Consider two additional examples:

Suspect that system is nonlinear - find a counterexample.

x3 (t) = x1 (t) + x2 (t) =