Digital Signal Processing

(EDSP632)
6th Sem B. Tech. (ECE)

Text book and Grading

Digital Signal Processing, John G Proakis, PE publication

Grading will be done out of 150 Marks:

Midsem (30)
Endsem (75)
Internal Assessment (45)
2 or 3 class tests (25)
Interaction during Class/Viva (20)
Registration in this subject will be automatically cancelled
for those who fail to attend at least 75% classes.

Prerequisite and the Syllabus

Prerequisite: ESAS432 Discrete time signal
and system theory, Z-Transform, Frequency Analysis
of discrete time signals and systems

The syllabus of EDSP632:

Review of fundamentals, DFT and its applications,
FFT Algorithms, Implementation structures and
analysis of DTLTI systems, Digital Filter Design,
Multirate Systems, Linear Prediction and Optimum
Filters, Power spectrum estimation techniques

Signal
The signal represents a physical quantity that varies
with time, space or any other independent variable(s).
Dimensionality (no. of variables in the function):
Sound: 1-Dimension
Greyscale image i(x,y) : 2-D
Video: 3 x 3-D: {r(x,y,t) g(x,y,t) b(x,y,t)}

Signal processing
The Signal Processing involves in modifying a
signal with one or more of the following
objectives
Information Enhancement (equalization, diversity etc.)
Information Extraction (radar signal analysis)
Information Rearrangement (speech coding in mobile)

Digital Signal Processing

Digital Signal processing using
Microprocessors
General Purpose Digital Signal Processors
ASIC

A/D & signal processors speed

Wide-band signals still difficult to treat
(real-time systems)

Finite word-length effect

(Round-off: error caused by rounding math calculation result
to nearest quantization level.)

Application areas
Space (Photograph enhancement, Intelligent sensory analysis etc)
Medical (CT, MRI, ultrasound etc.)
Commercial (Image and sound compression etc.)
Telephone (Echo reduction, data compression, Filtering etc.)
Military (Radar, Sonar, secure communication etc.)
Industrial (Process Monitoring and control etc.)

Review of Fundamentals

The Concept of frequency

For Continuous - time sinusoidal signal x(t)
x(t+T) = x(t); T is time period = 1/F (freq.)
Increasing F results in an increase in rate of
oscillations (more periods are included in unit time)
Two sinusoids x1(t) and x2(t) with distinct freq F1
and F2 are themselves distinct.

For Discrete time sinusoids

x(n) = Acos(n+); = 2f;
f (cycles per sample) = F/Fs; Fs is the sampling freq

Periodic, if f is a rational number.

for periodicity, x(n+N) = x(n) for all n i.e.
Acos(2f(n+N)+) = Acos(2fn+)
This relation is true if and only if there exists an
integer K such that 2fN = 2k i.e. f=k/N
If k and N are relatively prime then N is called the
fundamental period of x[n].
A small change in freq can results in large
change in period i.e. f1= 31/60 implies N1=60 but
f2=30/60 implies N2= 2.

 Two sinusoids x1[n] and x2[n] whose

frequencies are separated by an integer
multiple of 2 are identical (more precisely,
indistinguishable)
Acos(( +2)n+) = Acos(2n+ (n+))
= Acos(n+)
Sinusoids with <= <= i.e. -1/2<=f<=1/2
are distinct. If it appears that a sinusoid has f
outside this range then definitely in the above
range its identical sinusoid does exists.

 The highest rate of oscillation in a discrete

time sinusoid is attained when = (or - )
or equivalently f=1/2 or -1/2.
As varies from 0, /8, /4, /2, then f =
/2 will increases as 0, 1/16, 1/8, , and
N = , 16, 8, 4, 2
When > e.g. 3/2 then f= > which
should be represented by its identical
sinusoid cos(2- )n.

Classification of systems

Static (memoryless) Vs Dynamic (with

memory)
(static, if o/p depends upon present i/p only
e.g. y(n)= nx(n)+bx3(n). Dynamic, if depends
upon past or future I/P (may include present I/P
as well) e.g. y(n)=x(n-1)+3x(n))
Time-invariant Vs Time Variant systems
(if i/p, x(n) results in y(n) then x(n-k) must
results in y(n-k), otherwise system is time
variant. Ex: TI, y(n)=x(n)-x(n-1); TV, y(n)=
x(n)*cos(wn))

 Linear Vs Nonlinear
(linear, must satisfy superposition principle, otherwise
nonlinear.)

Causal Vs Noncausal
(o/p should depend upon present and past i/p but not on
future i/p {x(n+1)..}, otherwise noncausal. If signal is first
recorded then offline processing is done then only non
causal systems are possible to implement)

Stable Vs Unstable
(BIBO stable if every bounded i/p produces a bounded o/p,
otherwise unstable.)

FIR systems
This classification depends upon system
characteristics; the impulse response.
If h(n)=0 for n<0 and n>M then causal FIR
system (finite number of symbols in h(n))
For a causal FIR system
y(n) = (k=0 to M-1) h(k)x(n-k) = h(0)x(n)
+h(1)x(n-1) + ----- h(M)x(n-M+1)

 To form the o/p at any instant n=n0 the most

recent M input samples are required.
System must have a memory size M to
remember past M Input symbols.
This type of realization is called non recursive
realization of systems.
The FIR system acts as a window of size M
thus the impulse response of FIR system is
also called a window

IIR systems
The h(n) has infinite number of symbols e.g.
h(n)= anu(n).
y(n) = (k=0 to ) h(k)x(n-k) = h(0)x(n)+h(1)x(n1) + h(2)x(n-2) ----- up to
To form the o/p at any instant of time n=n 0 the
system must remember all the previous input
samples.

 To store all previous values of I/P, it will

require infinite memory space, which is
practically impossible.
Is it possible to realize an IIR system?
(fortunately , Yes.)

Recursive/Non Recursive systems

The classification recursive and non recursive
systems is based on method of implementation.
An FIR system can be implemented by either
method but an IIR system can only be
implemented by recursive method.
In recursive systems the output at any instant
of time n=n0 depends upon one or more
previous values of O/Ps. e.g.
y(n) = ay(n-1) + bx(n)

The LTI System Characteristics

Obeys Convolution Principle

y(n)= T[x(n)]
T represents the operation performed by the system

If x(n) is an impulse then y(n)= h(n)= T[(n)]

y(n)= T[(k= - to ) x(k) (n-k)]

= T[..+x(-1)(n+1)+x(0) (n)+x(1)(n-1)+..]
= +x(-1)T[(n+1)]+x(0)T[(n)]+x(1)T[(n-1)]+..
By applying superposition property of linear systems
= (k= - to ) x(k)T[(n-k)]
Apply time-invariance property
= (k= - to ) x(k)h(n-k)]

= x(n)*h(n)

Condition for Causality: h(n)=0 for n<0

Using convolution sum,
y(n)= (k= - to ) h(k)x(n-k)
= (k= 0 to ) h(k)x(n-k) + (k= - to -1) h(k)x(n-k)
= [h(0)x(n)+h(1)x(n-1)+..] + [h(-1)x(n+1)+ ..]
The first sum involves present and past values of
x(n).
The second sum involves future values of inputs
thus all terms of this sum should be zero for
causality.
It is guaranteed if, h(n)=0 for n<0
It is necessary and sufficient condition for causality.

Condition for BIBO Stability: Sh= (k= - to ) |h(k)|<

|y(n)| = |(k= - to ) h(k)x(n-k)|

<= (k= - to ) |h(k)||x(n-k)|
(absolute value of sum is always less than or equal to sum of
absolute values)

x(n) is called bounded when there exists a

finite constant Mx such that |x(n)|<=Mx<
If i/p is bounded then,
|y(n)|<=(k= - to )|h(k)|Mx =Mx(k= - to ) |h(k)|

 Applying same definition, y(n) is bounded

if |y(n)|<.
Thus, Mx(k= - to ) |h(k)|<
As Mx is constant thus sufficient condition
for BIBO stability is Sh= (k= - to ) |h(k)|<

Z-Domain conditions for Causality and stability

System function is defined as

H(z)= Z[h(n)]= (n= - to ) h(n)z-n
ROC of H(z) must be outside of a circle
then causal and vice versa.
For stability unit circle must be included in
the ROC of H(z). (prove it?)

 We have derived the condition for stability as;

Sh= (k= - to ) |h(k)|<
We also know thatH(z)= (n= - to ) h(n)z-n
|H(z)| = |(n= - to ) h(n)z-n| <= (n= - to ) |h(n)||z-n|

At unit circle i.e. z = 1

|H(z=1)| <= Sh
For stability Sh< i.e. |H(z=1)| <
According to the definition of ROC above condition is satisfied
if unit circle falls in it

This is necessary and sufficient condition for stability and its

converse is also true.
For a causal linear time invariant system the ROC is outside
of its highest pole thus condition for BIBO stability reduces to
the condition that all poles must lie inside unit circle.

Natural and forced response

Let H(z)= B(z)/A(z) has N simple distinct poles and X(z) = N(z)/Q(z) has
L simple distinct poles.

Using convolution theorem and assuming initial conditions zero (two

sided z-transform can be used)
Y(z) = H(z)X(z)= B(z)N(z) / A(z)Q(z)

 If X(z) = 0 then Y(z) = 0 i.e. the above

expression does not include the zero input
response
On partial fraction expansion
Y(z) = (k=1toN)Ak/(1-pkz-1) +
(k=1toL) Qk/(1-qkz-1)
Each coefficient Ak and Qk depends upon
both H(z) and X(z).

 Taking inverse z-transform

y(n) = (k=1toN)Ak(pk)nu(n)+
(k=1toL)Qk(qk)nu(n)
First sum depends upon system poles thus called Natural
response ynr(n).
If all pk fall inside the unit circle and n the ynr(n) reduces
to zero thus it is also called transient response ytr(n).
(influence of I/P is reflected in Ak). If system poles are close
to unit circle it will have a longer transient response.
Second sum depends upon poles of input signal thus
called forced response yfr(n)

 If all qk fall inside unit circle and n then the

yfr(n) will also reduce to zero. (it is just the
response to a limited duration i/p)
If I/P is a sinusoid then the poles will be
complex conjugate and fall on the unit circle.
Then O/P will also be a sinusoid of different
amplitude and phase.
In this case it is also called the steady state
response of the system (it persists as long as
input is ON).

Direct form realizations

Transfer function can be written as

M

H z

b z

k 0

1 ak z

vn

k 1

Direct Form I Represents

H z H2 z H1 z

1
N

bk z k

k k 0
1 ak z
k 1

V z H1 z X z bk z k X z
k 0

1
V z
Y z H2 z V z
N

k
1 ak z
k 1

yn

b xn k

k 0

a yn k vn

k 1

 Replace order of cascade LTI systems

H z H1 z H2 z bk z
N
k 0
1 a z k

k
k 1

X z
W z H2 z X z
N

k
1 ak z
k 1

k
b
z
W z

k
k 0

Y z H1 z W z
wn
yn

a wn k xn

k 1

b wn k

k 0

wn
yn

a wn k xn

k 1

b wn k

k 0

 No need to store the same

data twice in previous
system
So we can collapse the
delay elements into one
chain
This is called Direct Form II
or the Canonical Form
Theoretically no difference
between Direct Form I and II
Implementation wise
Less memory in Direct II
Difference when using
finite-precision arithmetic

Fourier analysis - tools

Input Time Signal

2.5

Frequency spectrum

2
1.5
1

Periodic

0.5
0
0

time, t

Continuous

2.5
2
1.5
1

(period T)

FS

Discrete

Aperiodic

FT

Continuous

0.5

T
1
c k s(t) e j k t dt
T
0
j2 f t

S(f) s(t) e
dt

0
0

time, t

10

12

2.5
2

Periodic

1.5
1
0.5

(period T)

0
0

time, tk

Discrete
2.5

Aperiodic

2
1.5
1
0.5
0
0

time, tk
6

10

12

2kn
N

j
1
~
N
ck s[n] e
N
n0

DFS** Discrete
DTFT

Continuous

DFT** Discrete

Note: j = -1, = 2 /T, s[n]=s(tn), N = No. of samples

**

S(f) s[n] e j 2 f n
n
2kn
j
1 N1
~
N
ck s[n] e
N
n 0

Calculated via FFT