How to manipulate Frequencies in Discrete-time Domain?

Two Main Approaches

• Difference Equations (an • Discrete Fourier Transform

LTI system) (in short, DFT)
• Remember we have introduced
three kinds of Fourier transforms.
• However, what we are able to
deal with in the discrete-time
x[n]: input, y[n]: output domain is usually a finite-
duration signal.
• That is, building a system that • DFT is the final (fourth) Fourier
makes use of the current and transform, where its input is a
previous inputs x[n], x[n-1], x[n- discrete-time finite-duration
2] …. signal.
as well as the previous outputs
y[n-1], y[n-2], …

Details will be introduced in later courses

DTFT Theorems Reviews (From Oppenheim)

 Linearity
x1[n]  X1(ejw), x2[n]  X2(ejw)
implies that
a1x1[n] + a2x2[n]  a1X1(ejw) + a2X2(ejw)

 Time shifting
x[n]  X(ejw)
implies that
x n  n d   e
 jwn d
 
X e
jw
 DTFT Theorems (continue)

 Frequency shifting
x[n]  X(ejw)
implies that
e
jw 0 n

x n   X e
j w  w0 

 Time reversal
x[n]  X(ejw)
If the sequence is time reversed, then x[n]  X(ejw)
 DTFT Theorems (continue)

 Differentiation in frequency
x[n]  X(ejw)
implies that
nx n   j
dX e  
jw

dw
 Parseval’s theorem
x[n]  X(ejw)
implies that  

 X e 
2
 x n 
2 1 jw
E   dw
n  
2

 DTFT Theorems (continue)

 The convolution theorem

x[n]  X(ejw) and h[n]  H(ejw),
and if y[n] = x[n]  h[n], then
Y(ejw) = X(ejw)H(ejw)

 The modulation or windowing theorem

x[n]  X(ejw) and w[n]  W(ejw),
and if y[n] = x[n]w[n], then

 
Y e
jw 1
  W e
X e
j j  w  
d  Periodic convolution,
2 Y(ejw)= Y(ej(w+2))

 DTFT Theorems (continue)
• Hence, unlike continuous Fourier transform, DTFT’s
time and frequency domains are not that
analogous in DTFT.
• For example, convolution in the time domain
implies multiplication in the frequency domain.
• But multiplication in the time domain implies
periodic-convolution in the frequency domain.
 DTFT Pairs

 n   1
  n  n0   e
 jw n 0

1 (    n  )   2   w  2 k 
k  
 DTFT Pairs (continue)
Real Exponentials

1
a u  n  ( a  1) 
n
 jw
1  ae

1
u n    jw
     w  2 k 
1 e k  

1
 n  1 a u  n 
n
( a  1) 
1  ae 
2
 jw
 DTFT Pairs (continue)
sin(ax)/x : called sync function
sin(ax)/sin(x) : called Dirichlet Kernel

r sin w p  n  1 
n
1
u  n  ( r  1)   jw  j2w
1  2 r cos w p e r e
2
sin w p

sin w c n 1 w  wc
 X e   
jw

n 0 wc  w  

1 0n M sin  w  M  1  / 2   jw M / 2
x n     e
0 otherw ise sin  w / 2 
 DTFT Pairs (finally)


  2   w  w 0  2 k 
jw 0 n
e
k  


cos  w 0 n       e   w  w 0  2 k    e   w  w 0  2 k 
j  j

k  
Example: Determining a Fourier Transform Pair-
wise Formulas

 Suppose we wish to find the Fourier transform of

x[n] = anu[n-5].

x1 n   a u n  
n 1
 jw
 X1 e  
jw

1  ae
x 2 n   x1 n  5 
 j5w
Thus, X 2 e   e
jw  j5w
X1 e   jw e
 jw
1  ae
 j5w
 
5
x n   a x 2 n , so X e
5 jw a e
 jw
1  ae
 Another Example

 Determining the impulse response for a difference

equation
y[n](1/2) y[n1] = x[n] (1/4)x[n1]
 To find the impulse response, we set x[n] = [n].
Then the above equation becomes
h[n](1/2) h[n1] = [n]  (1/4)[n1]
 Applying DTFT, we obtain
H(ejw)  (1/2)e-jwH(ejw) = 1  (1/4) e-jw
 So H(ejw) = (1  (1/4) e-jw) / (1  (1/2) e-jw)
continue
 Then, from the pairwise table, we know

a u n  ( a  1) 
n 1
 jw
1  ae
thus, (1/2)nu[n]  1 / (1  (1/2) e-jw)

 By the shifting property, in the time domain, the

second term of H(ejw) is
(1/4)(1/2)n1u[n1]  (1/4) e-jw / (1  (1/2) e-jw)
 Hence,
h[n] = (1/2)nu[n]  (1/4)(1/2)n1u[n1]
 Symmetry Property of DTFT

 Conjugate-symmetric sequence: xe[n] = xe*[ n]

 If a real sequence is conjugate symmetric, then it is
called an even sequence satisfying xe[n] = xe[ n].
 Conjugate-asymmetric sequence: xo[n] = xo*[ n]
 If a real sequence is conjugate antisymmetric, then it is
called an odd sequence satisfying x0[n] =  x0[n].
 Any sequence can be represented as a sum of a
conjugate-symmetric and asymmetric sequences,
x[n] = xe[n] + xo[n], where xe[n] = (1/2)(x[n]+ x*[n])
and xo[n] = (1/2)(x[n]  x*[n]).
 Symmetry Property of DTFT
(continue)
 Similarly, a Fourier transform can be decomposed
into a sum of conjugate-symmetric and anti-
symmetric parts:
X(ejw) = Xe(ejw) + Xo(ejw) ,
where
Xe(ejw) = (1/2)[X(ejw) + X*(ejw)]
and
Xo(ejw) = (1/2)[X(ejw)  X*(ejw)]
 Symmetry Property of DTFT
(continue)

 Fourier Transform Pairs (if x[n]  X(ejw))

 x*[n]  X*(e jw)
 x*[n]  X*(ejw)
 Re{x[n]}  Xe(ejw) (conjugate-symmetry part of X(ejw))
 jIm{x[n]}  Xo(ejw) (conjugate anti-symmetry part of X(ejw))
 xe[n] (conjugate-symmetry part of x[n])  XR(ejw) = Re{X(ejw)}
 xo[n] (conjugate anti-symmetry part of x[n])
 jXI(ejw) = jIm{X(ejw)}
 Symmetry Property of DTFT
(continue)
 Fourier Transform Pairs (if x[n]  X(ejw))
 Any real xe[n]  X(ejw) = X*(ejw) (Fourier transform is
conjugate symmetric)
 Any real xe[n]  XR(ejw) = XR(ejw) (real part is even)
 Any real xe[n]  XI(ejw) = XI(ejw) (imaginary part is odd)
 Any real xe[n]  |XR(ejw)| = |XR(ejw)| (magnitude is even)
 Any real xe[n]  XR(ejw)= XR(ejw) (phase is odd)
 xo[n] (even part of real x[n])  XR(ejw)
 xo[n] (odd part of real x[n])  jXI(ejw)
 Example of Symmetry Properties

 The Fourier transform of the real sequence x[n] = anu[n]

for a < 1 is jw
 
X e
1
 jw
1  ae
 Its magnitude is an even function, and phase is odd.
 Sinusoidal Responses of LTI System
(application of symmetric property)
• We have seen that ejwn is an eigen function of any
LTI system, with the corresponding ‘eigenvalue’ (or
frequency response) H(ejw).
• What happens when feeding a real sinusoid of
frequency w0 into the LTI system with the frequency
response H(ejw)?

A A
x  n   A cos  w 0 n    
j  j  jw 0 n

jw 0 n
e e e e
2 2
 Sinusoidal Responses of LTI System
(application of symmetric property)
• If the impulse response h[n] is real, by the
symmetry property, H(ejw ) = H*(ejw ) 0 0

• Hence, y  n   A H  e jw  e j e jw n  A H  e  jw  e  j e  jw n
0 0 0 0

2 2

H e e e  e  (e
A j A  j 
 
jw 0 jw 0 n jw 0 jw 0 n
H e )
2 2

H e e e [H e e e
A j A j jw 0 n 
 
jw 0 jw 0 n jw 0
]
2 2
 A R e{ H  e e e
jw 0 j jw 0 n
}

 A H e  R e{e e }, w here    H  e 
jw 0 j jw 0 n j jw 0
e

 A H e  cos  w n     
jw 0
0
 Efficient Representation of
Difference Equations
• We have mentioned that difference equation
is a common way for realizing an LTI system
(although not all LTI system is able to be
implemented by difference equations)
N M

 a k y n  k    b m x n  m 
k 0 m 0

• To represent a difference equation efficiently,

z-transform introduced below is widely used.
 Z-transform
• Representing a discrete-time signal (or a sequence)
as a polynomial.

X z    x n z
n

n  

• For CS students: It has the same form of the

generating function for combinations in
combinatorics.
– But are developed from different domains.
– In fact, the difference equation is the same as the
“recurrence relation” in combinatorics.
 Z-transform
• For a right-sided sequence (x[n]=0 for n<0), the
z-transform is


X z   x n  z
n

n0

• Example: what is the z-transform of delta-

function? 

X z    n  z
n
1
n0
 Example of z-transform for a finite
sequence

n n1 0 1 2 3 4 5 N>5

x[n] 0 2 4 6 4 2 1 0

X z   2  4 z
1 2 3 4 5
 6z  4z  2z z
 Z-Transform vs. DTFT
• Discrete-time Fourier Transform

    x n e
X e
jw  jwn

n  
• Hence, DTFT is equivalent to substituting z=ejw into
the z-transform


X z   x n  z
n
|z  e jw
n  

• More specifically, the z -transform is a generalization

of the DTFT, where the DTFT evaluates the z-transform
only on the complex unit circle (|z| = 1):
Z-Transform vs. DTFT

The unit circle in the complex z plane.

DTFT is only evaluated on this circle.
 Z-transform vs. Convolution
• Remember that the convolution results can be seen as
the coefficients of polynomial products.
• Since z-transform represents a sequence as a polynomial,
it has the property that

• Time domain convolution implies Z-domain multiplication

 Proof
• Convolution

• Take the z-transform on both sides:

 Proof (con’t)

Time domain convolution implies Z-domain multiplication

 Time Delay Property
• It can be easily shown that time delay of n0
samples is equivalent to multiplying z-n0 in the
z-domain.
 Z-transform Applying to Systems
• In the above, z-transform is applied to a signal.
• Now, we apply it to the LTI system realized by a
difference equation:
N M

 a k y n  k    b m x n  m 
k 0 m 0

• Taking z-transforms for both sides, we have

N M

 a k Y  z z  b m X  z z
k m

k 0 m 0
 Z-transform Applying to Systems
M


m
• Hence bm z
Y z  m 0

X z  N


k
ak z
k 0
• We call
Y z
H (z)  A fractional form
X z
the system function of this LTI system
• General Definition: the system function of an LTI
system is the z-transform of the output signal
divided by the z-transform of the input signal.
 System Function vs. Frequency
Response
• Hence, when feeding an input signal X(z) to an
LTI system with the system function H(z), the
output is the product Y(z)=X(z)H(z)
• When z=ejw, we obtain the output spectrum
Y(ejw)=X(ejw)H(ejw).

X(z) Y(z) (= H(z)X(z)) Z-transform

H(z)/H(ejw)
X(ejw) Fourier transform
Y(ejw) (= H(ejw)X(ejw))
Z-transform and Impulse Response
• In addition, the convolution in time domain
implies multiplication in the z-domain (and
frequency domain).
• So,
Time domain convolution

x[n] h[n] y[n] = x[n]h[n]

X(z) Y(z) (= H(z)X(z)) Z-transform

H(z)/H(ejw)
X(ejw) Fourier transform
Y(ejw) (= H(ejw)X(ejw))
 Z-transform
• Hence, we can represent an LTI system by
either h[n] (impulse response), H(ejw)
(frequency response), or H(z) (z-transform)
• In particular, we usually use H(z) as a fractional
form: M

b
m
m
z
m 0
H (z)  N


k
ak z
k 0

to represent a difference-equation LTI system.

 Cascading of LTI Systems
• Cascading of LTI systems can be described in
the z-domain too.
Z-transform vs. Laplace transform
Remark
• Z-transform transfers a discrete-time signal to the
z-domain.
• It is analogous to the Laplace transform for the
continuous-time signal that is transformed to the s-
domain.
• Z-transform: for solving difference equations
• Laplace transform: for solving differential equations
 Time Delay System
• Recall that time delay of n0 samples is
equivalent to convolving with (n-n0) in time
domain, or multiplying z-n0 in the z-domain.

• We call z-1 the unit-delay system

System Diagram of A Causal FIR System
• The signal-flow graph of a causal FIR system can be
represented by z-transforms:

x[n] b0 +
y[n]
M z 1
y n    b m x n  m  x[n-1] b1 +

m 0 z1
x[n-2] b2 +

z1
x[n-M] bM +
 Deconvolution
• Deconvolution or inverse filtering
 Example
• Deconvolution or inverse filtering
Poles and Zeros of Fractional Form
• Let H(z) = P(z)/Q(z) be the system function of
a rational form.
• Let us represent both P(z) and Q(z) as
polynomials of z (not z-1)

Then
• Poles: the roots of Q(z)
• Zeros: the roots of P(z)
Example of Poles and Zeros
 Example and Physical Meanings
• Three zeros:

• Three poles: 0, 0, 0 (or called a three-order pole at 0)

• Physical meanings
• Pole:
– The pole of a z-transform H(z) are the values of z for
which H(z)= .
• Zero:
– The zero of a z-transform H(z) are the values of z for
which H(z)=0.
Poles and zeros of negative powers of z
 Pole-Zero Plot
• It is useful to display the zeros and poles of H(z)
on the complex z-plane.
• Knowing the poles and zeros is equivalent to
knowing the LTI system.
zeros
 poles
 Significance of the Zeros of H(z) for
FIR System
• The poles a FIR system
are always in the origin
(or a multi-order zero).
• The zeros of a FIR filter
thus play a major role in
the frequency response.
• Recall that DTFT is
evaluated on the unit-
circle on the z-plane.
• If the zeros are placed
close to the unit circle,
we can image that the
nearby frequencies will
be lowered.
 Example: The L-point Running-
Sum Filter

• The system function is

• We have to represent it as rational form in z.

• From geometric series,
 Example: The L-point Running-
Sum Filter
• The zeros of H(z) will be determined by the roots of the
numerator polynomial,

• The poles are determined by the denominator. There

are (L-1)-th roots at z=0 and a root at z=1.
• So, the is a common location at z=1 for both pole and
zero; they are canceled.
• Hence, finally, the L-point runing-sum filter has L-1
zeros and an (L-1)th-order pole at the origin.
 Example: The L-point Running-
Sum Filter
• For instance, L=10
 Example: The L-point Running-
Sum Filter
• The magnitude response: note that there are
a total of 9 zero responses from [-, )
Recall: Ideal Low-pass Filter’s
frequency response
 Passband
• In the above example of 10-point running sum
filter, the passband is roughly centered at w=0.
It acts like a low-pass filter.
• Can we move the passband to other locations
by re-placing zeros, so that the filter acts like
some other frequency-selection filters?
Frequency-selection Filters: Ideal
High-pass Filter
Ideal Band-pass Filter
Ideal Band-stop Filter
 Bandpass Filter by Placing the Zeros
of an FIR system
• We can move the passband to a frequency away from
w=0 by changing the zero locations:
• The obvious way is to use all but one of the roots of
unity as the zeros of an FIR filter.

Note that by specifying poles and zeros,

we completely describe the LTI system.
L=10
 Bandpass Filter by Placing the Zeros
of an FIR system
• Frequency magnitude response:
 Bandpass Filter with Real Coefficients
• However, the above filter have complex coefficients.
• By the symmetry property of DTFT, we know that the
magnitude response should be an even function if we
hope the coefficients are all real.
 Practical Band Pass Filtering
• The above examples enable us a process of
manipulating (adding, moving, removing) the zeros
interactively by using a simulation system (eg. Matlab)
for filter design.
• We show a filter obtained by such an “interactive
design” below, although much better filters can be
deigned by more sophisticated methods.
• It is a useful illustration of the power of the z-
transform to simplify the analysis of such problems.
• The underlying reason is that the z-transform coverts
difficult problems involving convolution and frequency
response into simple algebraic ideas based multiplying and
factoring polynomials.
Practical Band Pass Filtering
 Digital Filter Structures

• We know that there could be multiple ways to

realize the system function of a difference
equation.
• Z-transform is also useful in help characterize
these solutions, and find an efficient
implementation structure.
 Digital Filter Structures (for IIR Filter)
N M

y n    a y n  k    b x n  k 
k k
k 1 k 0

v[ n ]   b  x n  k 
k
k 0

y n    a y n  k   v [ n ]
k
k 1

Direct Form I implementation

On the z-domain

1  M k 
H (z)  H 2 (z)H1(z)  ( N
)  bk z 
1   ak z
k  k 0 
k 1

or equivalently
 M k 
V ( z )  H 1 ( z ) X ( z )    bk z  X ( z )
 k 0 

1
Y ( z )  H 2 ( z )V ( z )  ( N
)V ( z )
1   ak z
k

k 1
By changing the order of H1 and H2, onsider the equivalence on
the z-domain:
H (z)  H1(z)H 2 (z)

where  M k  H 2 (z)  (
1
H 1 ( z )    bk z  N
)
1   ak z
k
 k 0 
k 1
Let
1
W (z)  H 2 (z) X (z)  ( N
) X (z)
1   ak z
k

Then
k 1

 M 
Y ( z )  H 1 ( z )W ( z )    b k z
k
W ( z )
 k 0 
 N

In the time domain, w[ n ]  a k

w[ n  k ]  x[ n ]
k 1
M

y[ n ]  b k
w[ n  k ]
k 0

We have the following equivalence for implementation:

We assume M=N
here
Note that the exactly the same signal, w[k], is stored in the
two chains of delay elements in the block diagram. The
implementation can be further simplified as follows:

Direct Form II (or

Canonic Direct Form)
implementation
By using the direct form II implementation, the number of
delay elements is reduced from (M+N) to max(M,N).

Example: 1 2z
1

H (z)  1 2
1  1 .5 z  0 .9 z

Direct form I implementation Direct form II implementation

Representing by signal-flow graph

Example: the signal-flow graph of direct form II.

Example:

1 2 1 1
1 2z z (1  z )(1  z )
H (z)  1 2
 1 1
1  0 . 75 z  0 . 125 z (1  0 . 5 z )( 1  0 . 25 z )

Cascade structure: direct form I implementation

Cascade structure: direct form II implementation

 Cascade Form

From the fundamental theorem of algebra, we know that

an n-th order real-coefficient polynomial has n roots that
are real or complex-conjugate pairs.

From this theorem, it ensures that we can factor the

numerator and denominator polynomials. We can express H(z) in
the form:
M1 M 2
1 1  1
 (1  f k z )  (1  g k z )( 1  g k z )
k 1 k 1
H (z)  N1 N2
1 1  1
 (1  c k z )  (1  d k z )( 1  d k z )
k 1 k 1

where M=M1+2M2 and N=N1+2N2, gk and gk* are a complex

conjugate pair of zeros, and dk and dk* are a complex conjugate
pair of poles.
• A general form is

1 2
b 0 k  b1 k z  b2 k z
Ns

H (z)   1  a1 k z
1
 a2k z
2
k 1

where N s  ( N  1) / 2 
and we assume that MN. The real poles and
zeros have been combined in pairs.

• If there are an odd number of zeros, one of

the coefficients b2k will be zero.
It suggests that a difference equation can be
implemented via the following structure consisting
of a cascade of second-order and first-order
systems:

cascade form of implementation (with a direct form II realization of

each second-order subsystems)
 Parallel Form

• We also know that a rational function can be

represented as sum of partial fractions.

• If we represent H(z) by additions of low-order rational

systems:

N 1
B k (1  e k z )
p N1 N2
Ak
  1 c  (1  d
k
H (z)  Ck z  1
 1  1
k 0 k 1 k
z k 1 k
z )( 1  d k z )

where N=N1+2N2. If MN, then Np = MN.

 Parallel Form

• Alternatively, we may group the real poles in

pairs, so that

N 1
e 0 k  e1 k z
p Ns

  1 a
k
H (z)  Ck z  1 2
k 0 k 1 1k
z  a2k z
Illustration of parallel-form
structure for six-order
system (M=N=6) with the
real and complex poles
grouped in pairs.
Example: consider still the same system

1 2 1
1 2z z  7  8z
H (z)  1 2
8 1 2
1  0 . 75 z  0 . 125 z 1  0 . 75 z  0 . 125 z
another alternation of the same system

1 2
1 2z z 18 25
H (z)  1 2
8 1
 1
1  0 . 75 z  0 . 125 z 1  0 .5 z 1  0 . 25 z

Hence, given a system

function, there are many
ways to implement it.

There are equivalent when

infinite-precision arithmetic
is used. However, their
behavior with finite-
precision arithmetic could
be quite different.
Remark: While the signal flow graph is an efficient way
to represent a difference equation, not all of its
instances are realizable:
If a system function has poles, a corresponding signal
flow graph will have feedback loops.
A signal flow graph is computable if all loops contain at
least one unit delay element. Eg.

A non-computable system

Computable systems
 Discrete Fourier Transform (DFT)

Currently, we have investigated three cases of Fourier transform,

 Fourier series (for continuous periodic signal)
 Continuous Fourier transform (for continuous signal)
 Discrete-time Fourier transform (for discrete-time signal)

All of them have infinite integral or summation in either time or

frequency domains.
There is still another type of Fourier transform

Consider a discrete sequence that is periodic in the time

domain.

Eg., it can be obtained by a periodic expansion of a

finite-duration sequence, ie., we image that a finite-
length sequence repeats, over and over again, in the
time domain).

Then, in the frequency domain, the spectrum shall be

both periodic and discrete, ie, the frequency sequence
is also made up of a finite-length sequence, which
repeats over and over again in the frequency domain.
 Discrete Fourier Transform
• Considering both the finite-length (or finite-duration)
sequences in one period of the time and frequency
domains, leads to a transform called discrete Fourier
transform (DFT).
• In principle, DFT acts like performing multiple (a bank
of) FIR filters simultaneously at the same time.
 DFT Definition
• If x[n] is a finite-length sequence (n0 only when
|n|<N) , its DTFT X(ejw) is still a periodic continuous
function with period 2.

• The DFT of x[n], denoted by X(k) is as follows:

where W  e  j  2  / N  , and Wn are the roots of WN = 1.

 Relation to DTFT for Finite-length
sequence
• X(k) is the uniform samples of the DTFT X(ejw) at the
discrete frequency wk = (2/N)k, when the frequency
range [0, 2] is divided into N equally spaced points.
• Note that the above explains only the relation
between DFT and DTFT for the simple case that x[n]
is a finite-length sequence.
• We will investigate more in-depth relationships
between DFT continuous FT (and also between DFT
and DTFT) for length-unlimited sequences in the
future.
From Kuhn 2005
From Kuhn 2005
 Four types of Fourier Transforms
Time domain non- Time domain periodic
periodic
Frequency Continuous Fourier Fourier series
domain non- transform (both (time domain
periodic domains are continuous and
continuous) periodic, frequency
domain discrete)
Frequency DTFT DFT/DTFS
domain periodic (time domain (time domain discrete,
discrete, frequency frequency domain
domain continuous discrete, and both
and band-limited or finite-duration or
periodic) periodic)