Discrete-Time Signals and Systems

Seoul National University 1.
Discrete time signals and systems

School of Electrical Engineering
1. Discrete-time signals and systems

• Discrete - time signals: sequences
– Discrete-time signals are represented as sequences of numbers

– A sequence is a function whose domain is the set of integers
– Define the nth number in the sequence by x [ n ], a set of numbers x can be represented
by
x = { x [ n ]} , −∞ < n < ∞
– Note that
x [ n ] = xa ( nT )
xn
x0
x −1
x1
discrete-time signal analog signal

x2
where T denote the sampling period.
– A delayed or shifted version of x [ n ] : n
y [ n ] = x [ n − n0 ]
where n0 is an integer
420.461: Digital Signal Processing copyright@Yong-Hwan Lee

Seoul National University 1. Discrete time signals and systems
– An unit sample sequence is defined by

⎧0; n = 0
δ [ n] = ⎨
⎩1; n ≠ 0
which is also referred to as a discrete-time impulse or simply as an impulse
– An arbitrary sequence can be represented as a sum of weighted and delayed

impulses ∞
x [ n] = ∑ x [ k ]δ [ n − k ]
k =−∞
Ex) x [ n ] = m−3δ [ n + 3] + m1δ [ n − 1] + m4δ [ m − 4] m1

m−3
4
-3 -2 -1 0 1 2 3 5 6
– The unit step sequence is defined by
⎧1, n ≥ 0 m4
u [ n] = ⎨
⎩0, n < 0
∞
⇒ u [ n ] = δ [ n ] + δ [ n − 1] + ... = ∑ δ [ n − k ]
k =0
-1 0 1 2 3
⇒ δ [ n ] = u [ n ] − u [ n − 1]

• Let x [ n ] be an exponential sequence given by

x [ n ] = Aα nu [ n ]
jω
– For complex α = α e 0
and A = A e jφ ,
x [ n ] = A e jφ α e jω0n
n
= A α ⎡⎣cos (ω o n + φ ) + j sin (ω 0 n + φ ) ⎤⎦
n
– Complex exponential sequence ( α = 1)

x [ n ] = A cos (ω 0 n + φ ) + j A sin (ω 0 n + φ )
where ω 0 is called the frequency of the complex sinusoid and φ is called the phase
x [ n ] = Ae
j (ω o + 2π )n
– Since = Ae jω0n e j 2π n = Ae jω0n ,
complex exponential sequences with frequencies (ω 0 + 2π k ) , where k is an integer,
are indistinguishable from one another
– A sinusoidal sequence can be represented by x [ n ] = A cos(ω 0 n + φ ), ∀n
– Since x [ n ] = A cos ⎡⎣(ω 0 + 2π k ) n + φ ⎤⎦ = A cos (ω 0 n + φ ) ,
we consider ω in a frequency interval of length 2π ,
−π < ω 0 ≤ π or 0 ≤ ω 0 < 2π .

– A sequence is called periodic with an integer period of N, if

x [ x ] = x [ n + N ] , ∀n
– In order for a sinusoidal sequence to be
A cos (ω 0 n + φ ) = A cos ⎡⎣ω 0 ( n + N ) + φ ⎤⎦
it requires for ω 0 N = 2π k .
2π
– A sinusoidal sequence is not necessarily periodic with a period of and may not
ω0
be periodic at all, depending on the value of ω 0
1 2π k
Ex ω0 = π ⇒ N = ⋅ 3 = 6k
3 π
ω 0 = 1 ⇒ N = 2π k ⇒ No integer N
• High and low frequency concept: A continuous time sinusoidal signal x [ n ] = A cos ( Ω0t + φ )
oscillates more rapidly as Ω0 increases. However, x [ n ] = A cos (ω 0 n + φ ) oscillates
more rapidly as ω 0 increases from 0 toward π .
But, as ω 0 increases from π to 2π , the oscillation becomes slower due to symmetry
with respect to π.
⎛π ⎞ ⎛ 15 ⎞
Ex : A cos ⎜ n + φ ⎟ = A cos ⎜ π n + φ ⎟
⎝8 ⎠ ⎝8 ⎠

• Discrete-time systems
xn yn
T {}
⋅ y [ n ] = T { x [ n ]}
– Ideal delay system

y [ n ] = x [ n − nd ] , nd is a positive integer
– Moving averager
1 N −1
y [ n ] = { x [ n ] + x [ n − 1] + ... + x [ n − N + 1]} = ∑ x [ n − k ]
1
N N k =0
– Memoryless system
A system is referred to as memoryless if the output y [ n ] depends only on the input x [ n ]
Ex : y [ n ] = x 2 [ n ] , ∀n
– Linear system : principle of superposition
T {ax1 [ n ] + bx2 [ n ]} = aT { x1 [ n ]} + bT { x2 [ n ]}
Ex: accumulator n
y [n] = ∑ x[k ]
k = −∞

– Time-invariant system (TIS): A time shift or delay of the input sequence results in a
corresponding shift in the output sequence
y1 [ n ] = T { x [ n − nd ]} ⇒ y [ n − nd ] = y1 [ n ]
Ex : y [ n ] = x 2 [ n ] → TIS
n
y [ n] = ∑ x[k ] → TIS
k =−∞
y [ n ] = x [ Mn ] : compressor
y1 = x1 [ Mn ] = x [ Mn − no ]
y [ n − n0 ] = x ⎡⎣ m ( n − n0 ) ⎤⎦ ≠ y1 [ n ] → not TIS
– Causality: A system is causal if the output sequence at n = no depends only on the
input sequences for n ≤ n0, ∀n0
Ex: y [ n ] = x [ n + 1] − x [ n ] ⇒ non causal
y [ n ] = x [ n ] − x [ n − 1] ⇒ causal
– Stability: A system is stable in the BIBO sense iff every bounded input sequence
produces a bounded output sequence
x [ n ] ≤ B X < ∞ ⇒ y [ n ] ≤ B y < ∞ , ∀n
n
⎧0 n<0
y [ n] = ∑ u [ k ] = ⎨n + 1
k =−∞ ⎩ n≥0
• Linear time-invariant systems
– Let nk [ n ] be the response of the system to the input δ [ n − k ] , i.e., hk [ n ] = T {δ [ n − k ]}

y [ n ] = T { x [ n ]}
⎧ ∞ ⎫ ∞
= T ⎨ ∑ x [ k ]δ [ n − k ]⎬ = ∑ x [ k ]T {δ [ n − k ]}
⎩ k =−∞ ⎭ k =−∞
∞
= ∑ x [ k ] h [ n]
k =−∞
k
– Time invariance implies

δ [ n ] → h [ n ] ⇒ δ [ n − k ] ⇒ h [ n − k ] = hk [ n ]
– Discrete-time convolution
∞
y [ n] = ∑ x [ k ]h [ n − k ] Δ x [ n] ∗ h [ n]
k =−∞
– To compute h [ n − k ] , first reflect h [ k ] about the origin to obtain h [ −k ] and then shift
the origin to k = n ;
hk h −k h 6− k
0 1 -2 -1 0 0 1 2 3 4 5
Ex h [ n ] = u [ n ] − u [ n − N ]
⎧1, 0 ≤ n ≤ N − 1
=⎨
⎩0, otherwise
x [ n ] = a nu [ n ]
n < 0; y [ n ] = 0
1 − a n+1
n
n ≤ 0 ≤ N − 1; y [ n ] = ∑ a = k
k =0 1− a
n
a n− N +1 − a n+1
n > N − 1; y [ n ] = ∑ ak =
k = n − N +1 1− a
⎛1− aN ⎞
=a n − N +1
⎜ ⎟
⎝ 1− a ⎠
n − N −1 b g n 0 k
yn
n
0

• Properties of LTI system
– Commutative x [ n] ∗ h [ n] = h [ n] ∗ x [ n]
∞
⇒ y [ n] = ∑ x [ k ]h [ n − k ]
k =−∞
∞
∑ x [ n − k ]h [ k ]
=
k =−∞
x [ n ] ∗ ( h [ n ] + h [ n ]) = x [ n ] ∗ h [ n ] + x [ n ] ∗ h [ n ]
1 2 1 2
– Cascaded connection
x [n] v [n] y [n]
h1 h2
Letting h [ n ] Δ h1 [ n ] ∗ h2 [ n ] , we have
v [ n ] = x [ n ] ∗ h1 [ n ]
y [ n ] = v [ n ] ∗ h2 [ n ]
= x [ n ] ∗ h1 [ n ] ∗ h2 [ n ]
= x [ n] ∗ h [ n]

– A Linear system is stable iff the impulse response is absolutely summable, i,e.,
∞
S Δ ∑ h(k ) < ∞
k =−∞
Proof
Sufficient condition: ∞ ∞
y [ n] = ∑ h[k ] x[n − k ] ≤ ∑ h[k ] x[n − k ]
k =−∞ k =−∞
If x [ n ] is bounded, i.e., x n ≤ BX [ ]
∞
then, y [ n ] ≤ BX ∑ h[k ]
k =−∞
∞
∑ h[k ] < ∞
k =−∞
⇒ y [ n] < ∞
Necessary condition:
Show that if S = ∞ , a bounded input can cause an unbounded output.
Consider a bounded sequence given by
⎧ h * [ −n]
, h [ n] ≠ 0 [ ]
2
⎪ ∞ ∞ h k
x [ n] = ⎨ h [ −n] ⇒ y [ 0] = ∑ x [ − k ]h [ k ] = ∑ =S =∞
⎪ 0 h [ k ]
, h [ n] = 0
k =−∞ k =−∞
⎩

Ex h [ n ] = a nu [ n ]
∞
1
⇒S =∑a =
n
n =0 1− a
a <1⇒ S < ∞
a ≥ 1 ⇒ S = ∞ ⇒ unstable
xn yn Backward xn
Accumulator
difference
h [ n ] = u [ n ] ∗ (δ [ n ] − δ [ n − 1])
= u [ n ] − u [ n − 1]
= δ [ n]
Note: The property of convolution can be used to analyze the time-invariant system
xn one sample Forward yn Backward

⇒
delay difference difference
δ n −1 δ n +1 −δ n
h [ n ] = δ [ n − 1] ∗ (δ [ n + 1] − δ [ n ])
= δ [ n ] − δ [ n − 1]

• Linear constant coefficients difference equations
– An Nth order linear const. difference eq. is represented by

N M
∑ ak y [ n − k ] = ∑ bk x [ n − k ]
k =0 k =0
When the right side term is equal to zero, it is called the homogeneous difference eqn.
x [n] y [n]
Ex: Accumulator
n n −1
y [ n] = ∑ x [ n] = x [ n] + ∑ x [ k ] = x [ n] + y [ n − 1] +
k =−∞ k =−∞
⇒ y [ n ] − y [ n − 1] = x [ n ] D
y [ n − 1]
Ex: Moving average
h [ n] =
1
M
( u [ n ] − u [ n − M ])
1 M −1
⇒ y [ n] = ∑ x[n − k ] xn 1 yn
M k =0
M _ ∑
h [ n ] = (δ [ n ] − δ [ n − M ]) ∗ u [ n ]
1
M M-sample
delay
y [ n ] − y [ n − 1] =
1
M
( x [ n ] − x [ n − M ])

Ex
y [ n ] = ay [ n − 1] + x [ n ]
Consider x [ n ] = K δ [ n ] and y [ -1] = c ← auxiliary cond.
For n ≥ 0;
⇒ y [ 0] = ac + K
y [1] = ay [ 0] + 0 = a 2c + aK
:
y [ n ] = a n+1 ⋅ c + a n K , n ≥ 0 ⇒ y [ n ] = a n+1 ⋅ c + Ka nu [ n ]
For n < −1;
y [ −2] = a −1 ⎡⎣ y [ −1] − x [ −1]⎤⎦ = a −1 ⋅ c
y [ −3] = a −1 ⋅ a −1c = a −2c
:
y [ n ] = a n+1 ⋅ c, n ≤ −1

• Frequency-domain representation of DTS

x [ n ] = e jω n
∞
⇒ y [ n] = ∑ h [ k ]e ω[ j n−k ]
k =−∞
∞
=e jω n
∑ h [ k ]e
k =−∞
− jω k
∞
Let H ( e jω ) Δ ∑ h [ k ]e − jω k
k =−∞
= H R ( e jω ) + jH ( e jω )
( )
= H ( e jω ) e
jθ H e jω
y [ n ] = H ( e jω ) e jω n
eigen-value eigen-function of the system

Ex: Ideal delay y [ n ] = x [ n − n0 ]

For x [ n ] = e jω n ,
y [ n] = e = e − jω n0 e jω n ⇒ H ( e jω ) = e − jω n0
jω [ n − n0 ]
Since h [ n ] = δ [ n − n0 ] for an ideal delay system,

∞
H ( e jω ) = ∑ δ [ n − n ]e
0
− jω n
= e − jω n0 ⇒ H ( e jω ) = 1
n =−∞
{
Arg H ( e jω ) = −ω n0 }
– For x [ n ] = ∑α k e jω n , k
k
the output of the system is
(
y [ n ] = ∑α k H e jωk e jωk n ) ← principle of superposition
k
Ex: x [ n ] = A cos (ω 0 n + φ ) =
2
(
A jφ jω0n
e e + e − jφ e − jω0n )
⇒ y [ n] =
A⎡
2 ⎣ ( ) ( )
H e jω0 e jφ e jω n + H e − jω0 e − jφ e − jω0n ⎤⎦
If h [ n ] is real, H ( e − jω ) = H ∗ ( e − jω ) = H ( e jω ) e − jθ
⇒ y [ n ] = A H ( e jω ) cos (ω 0 n + φ + θ ) , where θ =arg ⎡⎣ H ( e jω ) ⎤⎦

jω
Ex: Delay H e 0 = 1;( ) θ = −ω 0 n0
⇒ y [ n ] = A cos (ω 0 ( n − n0 ) + φ )
– Discrete-time LTI → periodicity of 2π
( ) ∑ h[n]e
∞
j (ω + 2 kπ ) − j (ω + 2 kπ )n
⇒H e =
k =−∞
∞
= ∑ h [ n]e
k =−∞
− jω n
= H ⎡ e(
⎣
jω )
⎤
⎦
⇒ H (e jω
) is periodic with a period of 2π
Ex: Ideal frequency selective filters (LPF)
c h
H e jω
ω
−π − ω 0 −π ω 0 π 2πrω 0 2π 2π + ω 0

⎧ 1
Ex: ⎪ − M1 ≤ n ≤ M 2
h [ n] = ⎨ M1 + M 2 + 1
⎪
⎩ 0 otherwise
⇒ H (e ) =
M2
1
jω
∑
M 1 + M 2 + 1 n=− M1
e − jω n
e + jω M1 − e (
− jω M 2 +1)
1
=
M1 + M 2 + 1 1 − e − jω
e ( 2 1) e
− jω M +1− M / 2 jω ( M1 + M 2 +1) / 2 − jω ( M1 + M 2 +1) / 2
−e
=
M1 + M 2 + 1 1 − e − jω
−π −π
ω ( M 1 + M 2 + 1)
1 sin
− jω ( M 2 − M1 ) 2
= e θ
M1 + M 2 _1 ω
sin
2 π
−π

• Representation of sequences by Fourier transform

– A stable sequence can be represented by a Fourier integral of the form
∫ π X ( e ) e dω
1 π
x [ n] = ω ω j j n
- - - (A )
2π 2
∞
where X ( e jω ) =
∑ x [ n]e
n =−∞
− jω n
- - - (B)
( )
= X ( e jω ) e
jθ X e jω
← polar cord. form
= X R ( e jω ) + jX 1 ( e jω ) ← rectangular cord. form
Since X ( e jω ) is periodic with period 2π , it is of the form of Fourier series for the
continuous-variable periodic function X ( e ).
jω
– Note that Eq (A) is the inverse of Eq. (B)

⎛ ∞ ⎞
∫ π X (e )e
1 π 1 π
x [ n] = jω jω n
= ∫−π ⎜⎝ n∑ x [ m ]e − jω n ⎟ e jω n dω
2π − 2π =−∞ ⎠
∞
1 π jω n−m
= ∑ x [ m ] ∫ e ( ) dω = x [ n ]
n =−∞ 2π 2π
⎛ 1 π sin π ( n − m ) ⎞
= δ [ n − m] ⎟
jω ( n − m )
⎜∵ ∫π e dω =
π ( n − m)
⎝ 2π ⎠
2

If x [ n ] is absolutely summable (i.e., stable sequence), X ( e ) exists.

jω
–
∞
X (e jω
) = ∑ x [ n]e − jω n
n =−∞
∞ ∞
≤ ∑ x [ n ] e − jω ≤ ∑ x [ n] < ∞
n =−∞ n =−∞
– Furthermore, x [ n ] can be showen to converge uniformly to a continuous function of ω .

⇒ Any stable sequence (or system) have a finite and continuous frequency response:
(Sufficient condition for existence of H ( e jω ) )
– A finite-length sequence is absolutely summable
Ex: x [ n ] = a nu [ n ]
∞ ∞
X ( e jω ) = ∑ a nu [ n] e− jωn = ∑a e n − jω n
n =−∞ n =−∞
1
=− jω
, if ae− jω = a < 1
1 − ae
– If a sequence is not absolutely summable, but square summable ; i.e.,
∞
∑ x [ n] < ∞
2
n =−∞
it can be represented by a Fourier transform with mean-square convergence, i.e.,
2
X (e ) − ∑ x [ n] e
π M
lim ∫ jω − jω n
dω = 0
M →∞ −π
n =− M

Ex: Ideal LPF

⎧⎪1 ω < ω c
H LP ( e jω ) = ⎨
⎪⎩0 ω c < ω < π
1 ωc jω n sin πω c n
⇒ hLP ( n ) =
2π ∫−ωc
e d ω = , ∀n
πn
⇒ non causal and not absolutely summable
H Lp ( e jω ) is discontinuous at ω = ω c
But is mean-square summable
Ex: x [ n ] = 1, ∀n : neither absolutely nor square summable
∞
⇒ X ( e jω ) = ∑ 2πδ (ω + 2kπ )
k∞=−∞
Since X (e jω
) = ∑ 2πδ (ω − ω 0 + 2kπ )
k =−∞
1 π
⇒ x [ n] = ∫ 2πδ (ω − ω 0 ) e jω n dω
2π −π
= e jω0n ⇒ When ω = 0, x [ n ] = 1
0
X (e jω
) is not finite for all ϖ

– Theory of generalized functions

∞ ∞
∑e
n =−∞
j (ω −ω 0 )n
= ∑ 2πδ (ω − ω
k =−∞
0 + 2kπ )
∞
⇒ x [ n ] = ∑ ame jω m n
⇒ X (e jω
) = ∑ ∑ 2π a δ (ω − ω
m m + 2kπ )
m k =−∞ m
– Conjugate symmetric sequence: xe [ n ] = xe [ −n ]

*
– Conjugate anti-symmetric sequence: xo [ n ] = − xo [ −n ]

*
⇒ x [ n ] = xe [ n ] + xo [ n ]
xe [ n ] =
1
2
( x [ n ] + x∗ [ − n ]) even sequence : xe [ n ] = xe [ − n ]
odd sequence: xo [ n ] = − xo [ − n ]
xo [ n ] = ( x [ n ] − x∗ [ − n ])
1
2
⇒ X ( e jω ) = X e ( e jω ) + X o ( e jω ) = X R ( e jω ) + jX I ( e jω )
where
X e ( e jω ) =
1
2
(
X ( e jω ) + X ∗ ( e jω ) )
(
X o ( e jω ) = X ( e jω ) − X ∗ ( e − jω )
1
2
)
Note that X e ( e ) = X ( e ) and X o ( e ) = − X ( e )

jω ∗ − jω jω ∗ − jω
–
Re{ x [ n ]} ↔ X e ( e jω ) Re { x [ n ]} ↔ X ( e jω ) = X ∗ ( e jω )
j Im { x [ n ]} ↔ X o ( e jω )
xe [ n ] ↔ X R ( e jω )
xo [ n ] ↔ jX I ( e jω )
X ( e jω ) =
1
Ex: a <1
1 − ae − jω
X ∗ ( e jω ) = = X ( e jω )
1
− jω
1 − ae
X R ( e jω ) = (1 − a cosω ) = X R ( e jω )
1
1 + a − 2a cos ω
2
X I ( e jω ) = − X I ( e − jω )
− a sin ω
θ X ( e jω ) = tan −1 = −θ X ( e − jω )
1 − a cos ω

• Fourier transform theorems

F { x [ n ]} = X ( e jω )
{
x [ n ] = F −1 x ( e jω ) }
⇒ X [ n ] ←⎯
F
→ X ( e jω )
– Linearity: ax1 [ n ] + bx2 [ n ] ←⎯→ aX 1 ( e ) + bX 2 ( e )

F jω jω
– Delay: x [ n − n0 ] ←⎯
F
→ X ( e jω ) e − jω n 0
e jω0n ←⎯
F
(
→ X e j (ω −ω0 ) n )
– Time reversal: x [ −n ] ←⎯→ X ( e )
F − jω
If x [ n ] is real, x [ − n ] ←⎯→ X ( e )
F ∗ + jω
– Differentiation:
X ( e jω )
d
nx [ n ] ←⎯
F
→j
dω
– Parseval’s theorem:
∞ 2
X ( e jω ) dω
1 π
∑ x [ n] =
2
E Δ
n =−∞ 2π ∫π
−

– Convolution:
x [ n ] ←⎯
F
→ X ( e jω ) , h [ n ] ←⎯
F
→ H ( e jω )
∞
⇒ y [ n] = ∑ x [ k ]h [ n − k ] = x [ n] ∗ h [ n] ←⎯→Y ( e ω ) = X ( e ω ) H ( e ω )
k =−∞
F j j j
→ Y ( e jω ) = ∫ π X ( e )H ( e ) dξ
1 π
y [ n ] = x [ n ] h [ n ] ←⎯ (ω ξ )
F j ξ j −
2π −
periodic convolution
⇒ Multiplication of two sequences is equivalent to periodic convolution of

corresponding Fourier Transforms.
Ex: y [ n ] − 1 y [ n − 1] = x [ n ] − 1 x [ n − 1]
2 4
1 1 − jω
1 − e − jω e
⇒ H (e ) =
1 1 1
⇒ δ [ n ] − δ [ n − 1] = h [ n ] − h [ n − 1] jω 4 = − 4
4 2 1 − jω 1 − jω 1
1− e 1− e 1 − e − jω
2 2 2
n n −1
1 ⎛1⎞ 1⎛1⎞
a u [ n ] ←⎯→
n F
; α < 1 ⇒ h [ n ] = ⎜ ⎟ u [ n ] − ⎜ ⎟ u [ n − 1]
1 − ae − jω ⎝2⎠ 4⎝ 2⎠

• Discrete-time Random Signals
– Consider an LTI system with impulse response h [ n ]

∞
y [ n] = ∑ h[n − k ] x[k ]
k =−∞
When x [ n ] is a WSS discrete-time random process,
μ X [ n ] Δ E { X [ n ]} , μY [ n ] Δ E { y [ n ]}
– If stationary, μ X [ n ] = μ X
⎧ ∞
⎫
μY [ n ] = E ⎨ ∑ h [ n − k ] x [ k ]⎬
⎩ k =−∞ ⎭
∞ ∞
= ∑ h [ n − k ] E { x [ k ]} = ∑ h [ k ] E { x [ n − k ]}
k =−∞ k =−∞
∞
= μX ∑ h[k ]
k =−∞
= μX H (e j0 )

– Autocorrelation of y n [ ]
φ yy [ n, n + m ] = E { y [ n ] y [ n + m]}
⎧ ∞ ∞ ⎫
= E ⎨ ∑ ∑ h [ k ] h [ l ] x [ n − k ] x [ n − l + m ]⎬
⎩ k =−∞ l =−∞ ⎭
= ∑∑ h [ k ]h [l ] E { x [ n − k ] x [ m + m − l ]}
k l
= ∑∑ h [ k ]h [l ]φ xx [ m + k − l ] = φ yy [ m ]
k l
Let l − k Δ i. Then,
∞ ∞
φ yy [ m ] = ∑ φ [ m − i ] ∑ h [ k ] h [i + k ]
xx
i =−∞ k =−∞
∞
Autocorrelation sequence of h [ n ]: φh [i ] Δ ∑ h [ k ] h [i + k ] ( )
H e jω
k =−∞
∞
⇒ φ yy [ m ] = ∑ φ [i ]φ [ m − i ] ωa ωb
h xx
i =−∞
⇒ Φ yy ( e jω ) = Φ h ( e jω ) Φ xx ( e jω ) = H ( e jω ) Φ xx ( e jω )
2
where Φ h ( e jω ) = H ( e jω ) H ∗ ( e jω ) = H ( e jω )
2
Note : E { y 2 [ n ]} = φ yy ( 0 ) = Φ yy ( e jω ) dω = H ( e jω ) Φ xx ( e jω ) dω ≥ 0
1 π 1 π 2
2π ∫π
− 2π ∫π
−

Seoul National University 2. Periodic sampling of continuous-time signals
2. Periodic Sampling of Continuous-time Signals

• Impulse sampling: ideal continuous-to–discrete (C/D) converter
xc ( t ) xs ( t )
x [ n ] = xc ( nT ) , − ∞ < n < ∞
s ( t ) = ∑ δ ( t − nT )
⎛ 1⎞
where T ⎜ = ⎟ is the sampling period and f s is called the sampling frequency
⎝ fs ⎠
∞
xs ( t ) = xc ( t ) ∑ δ ( t − nT ) ⇐ shifting
n =−∞
property of the impulse function
∞
= ∑ xc ( nT )δ ( t − nT )
n =−∞

– Fourier transform of xs ( t )
∞
X s ( f ) = F { xs ( t )} = ∫ ∑ xc ( nT )δ ( t − nT ) e − j 2π ft dt
n =∞
∞
= ∑ xc ( nT ) e − j 2π fnT
n =∞
⎧ ∞ ⎫
X s ( f ) = X c ( f ) ∗ F ⎨ ∑ δ ( t − nT ) ⎬
⎩ n=−∞ ⎭
∞
⎛ n⎞
Ω s = 2π / T = 2π f s
1
= Xc ( f ) ∗ ∑ δ ⎜ f − ⎟
T n =−∞ ⎝ T⎠
1 ∞ ⎛ n⎞
= ∑ X c ⎜ f − ⎟ ⇐ periodic function
T n=∞ ⎝ T⎠
Xs ( f )
f s − W ≤ W ⇔ f s ≤ 2W
− fs −W 0 W fs − W f s
⇒ Aliasing
Xc ( f )
−W W
• Natural Sampling
xc ( t ) xs ( t )
p (t )
0 τ T T +τ
∞
1 nτ
p (t ) = ∑ pe
n =−∞
n
j 2π f s nt
; pn =
T
sin c
T
xs ( t ) = xc ( t ) p ( t )
⎧ ∞
⎫ Xs ( f )
⇒ X s ( f ) = F ⎨ xc ( t ) ∑ pn e j 2π f s nt ⎬
⎩ n =−∞ ⎭
∞
⎛ n⎞
= ∑ pn X c ⎜ f − ⎟
n =−∞ ⎝ T⎠
− fs 0 fs

• Sample and hold (S/H)
xc ( t ) xs ( t )
0τ
1 ⎛τ ⎞
p (t ) P( f ) = sinc ⎜ f ⎟
s (t ) T ⎝T ⎠
⎡ ∞
⎤
xs ( t ) = ⎢ xc ( t ) ∑ δ ( t − nT ) ⎥ ∗ p ( t )
⎣ n =−∞ ⎦
⎡ 1 ∞
⎛ n ⎞⎤
⇒ X s ( f ) = ⎢ X c ( f ) ∗ ∑ δ ⎜ f − ⎟⎥ p ( f )
⎣ T n=−∞ ⎝ T ⎠⎦
1 ∞ ⎛ n⎞
= p( f ) ∑ X ⎜ f − ⎟ Xs ( f )
T n=−∞ ⎝ T⎠
− fs 0 fs

• Nyquist Sampling Theorem
– If xc ( t ) is strictly band limited, i.e., X ( f ) = 0 for

f >W
1
and the sampling frequency is chosen such that f s = = 2W ,
T
then
∞
Xs ( f ) = ∑ x ( nT ) e
n =−∞
c
− j 2π fnT
Since
1
Xc ( f ) =
X s ( f ), f ≤ W ,
2W
if the sample values xc ( nT ) are specified for all time, X c ( f ) is uniquely determined
by using the Fourier series.
⇒ xc ( t ) is uniquely determined by xc ( nT ) , n = 0, ±1, ±2, ⋅ ⋅ ⋅.

• Reconstruction of a band limited signal from its samples: Interpolation formula

∞
xc ( t ) = ∫ X c ( f ) e j 2π ft df
−∞
πf
1 ∞ −j
∑ x [ n ] e W e j 2π ft df
W
=∫
−W 2W
n =−∞
∞
1
∑ x [ n]
W j 2π f ( t − nT )
=
n =−∞ 2W ∫−W
e df
∞
sin π ( t − nT ) / T
= ∑ x [ n] π ( t − nT ) / T
n =−∞
∞
sin π t / T
= ∑ x [ n] h(t − nT );
n =−∞
h(t ) =
πt /T
⇒ A band-limited signal of finite energy can be completely recovered by its samples
taken at a rate of 2W/sec
H( f )
– Ideal reconstruction filter
sin π t / T
h(t ) =
πt /T
f
1 1
−
2T 2T
• Poisson sum formula
A periodic signal xs ( t ) with period T is represented by the Fourier series,

∞
1
xs ( t ) = ∑xe
n =−∞
n
j 2π f s nt
; fs =
T
1 T2
where xn = ∫ T xs ( t ) e − j 2π f s nt dt
T −2
∞
1 ∞ j 2π f s nt ⎛ 1 T2 1⎞
Ex : ∑ δ ( t − nT ) = ∑ e ⎜∵ xn = ∫− T δ (τ )e
− j 2π f s nt
dt = ⎟
n =−∞ T n=−∞ ⎝ T 2 T⎠
x(t ) = ∑δ ( t − nT ) y (t )
ht bg
∞ ∞
y (t ) = h (t ) ∗ ∑ δ ( t − nT ) =
n =−∞
∑ h ( t − nT )
n =−∞
⎡ 1 ∞ j 2π fs nt ⎤
y (t ) = h (t ) ∗ ⎢ ∑ e ⎥
⎣ T n=−∞ ⎦
∞
1 1 ∞
= ∑ ∫ h (τ )e dτ = ∑ H ( nf s ) e j 2π f s nt
j 2π f s n( t −τ )
T n=−∞ T n=−∞
∞
1 ∞
⇒ ∑ h ( t − nT ) =
n =−∞
∑
T n=−∞
H ( nf s ) e j 2π f s nt

• Discrete-time processing of continuous-time signals
bg
xc t
C/D
xn
hn
yn bg
yr t
D/C
discrete - time
T system T
∞
⎛ ω 2π k ⎞
x [ n ] = xc ( nT ) ⇒ X ( e jω ) =
1
∑ Xc ⎜ −
T k =−∞ ⎝ T T ⎠
⎟;
π ( t − nT )
∞ sin
yr ( t ) = ∑ y [ n ] T ;
n =−∞ π ( t − nT )
T
⇒ Yr ( Ω ) = H r ( Ω ) Y ( e jΩT )
⎧ π
⎪TY ( e ) , Ω <
jΩT
low pass
=⎨ T
reconstruction ⎪⎩ 0 , otherwise
filter
– The freq-domain representation is easier than the time-domain representation

• LTI discrete-time systems

Y ( e jω ) = H ( e jω ) X ( e jω )
⇒ Yr ( jΩ ) = H r ( Ω ) H ( e jΩT ) X ( e jΩT ) ; ω = ΩT
1 ∞ ⎛ 2π k ⎞
= H r ( Ω ) H ( e jΩT ) ∑ Xc ⎜Ω − ⎟
T k =−∞ ⎝ T ⎠
⎧ π
⎪H (e ) X c (Ω), Ω <
jΩT
=⎨ T
⎪⎩ 0 , otherwise
– If X c ( Ω ) is band limited and the sampling rate is larger than the Nyquist rate,
Yr ( Ω ) = H eff ( Ω ) X c ( Ω )
the effective frequency response is
⎧ π
⎪H (e ), Ω <
jΩT
H eff ( Ω ) = ⎨ T

• Example of signal reconstruction

Ex: Ideal LPF

⎧1, ΩT < ωc
H eff ( Ω ) = ⎨
⎩0, otherwise
Ex: Ideal band-limited differentiator
d
yc ( t ) = xc ( t ) ⇒ H c ( Ω ) = jΩ
dt
⎧ π
⎪ j Ω, Ω <
⇒ H eff ( Ω ) = ⎨ T
⎪⎩0 , otherwise
jω
H ( e jω ) = , ω <π
T
⎧ cos π n
π n cos π n − sin π n ⎪ , n≠0
⇒ h [ n] = = ⎨ nT
π n 2T ⎪⎩ 0 , n = 0

– Impulse invariance
⎛ω ⎞
( )
H e jω = H c ⎜ ⎟ = H ( Ω); ω < π
⎝T ⎠
⇒ When h [ n ] = Thc ( nT ) , the discrete-time system is said to be an impulse-invariance
version of the continuous-time system.
1 ∞ ⎛ ω 2π k ⎞
∵ H (e ) = ∑ Hc ⎜ −
jω
⎟ ⇐ X s (Ω) = X ( e ) ω =ΩT = X ( e )
jω jΩT
T k =−∞ ⎝ T T ⎠
⎛ω ⎞
⇒ H ( e jω ) = H c ⎜ ⎟ , ω < π
1
T ⎝T ⎠
A
Ex: hc (t ) = Ae u (t ) ⇒ H c ( s ) =
s0t
s − s0
h [ n ] = Thc ( nT ) = Ae s0nT u [ n ]
⎛ω ⎞
or H ( e jω ) =
AT AT
H ( z) = s0T −1 s0T − jω
≠ Hc ⎜ ⎟
1− e z 1− e e ⎝T ⎠
∵ hc (t ) is not strictly band - limited
The resulting discrete-time frequency response is aliased

• Continuous-time processing of discrete-time signals
xn bg
xc t
hc ( t )
yc ( t ) yn
D/C C/D
T π ( t − nT ) T
∞ sin π
xc ( t ) = ∑ x [ n] T ; X c (Ω) = T X ( e jΩT ) , Ω ≤
n =−∞ π ( t − nT ) T
T
π ( t − nT )
∞ sin π
yc ( t ) = ∑ y [ n ] T ; Yc ( Ω ) = H c ( Ω ) X c ( Ω ) ; Ω <
n =−∞ π ( t − nT ) T
T
1 ⎛ω ⎞
Y ( e jω ) = Yc ⎜ ⎟ , ω <π
T ⎝T ⎠
= H c ( Ω ) X c ( Ω ) = H ( e jω ) X ( e jω ) ⇒ H ( e jΩT ) = H c ( Ω ) = H eff ( Ω ) ; Ω ≤ π /T

Ex: Non-integer delay H ( e jω ) = e − jωΔ , ω <π

If Δ is an integer , y [ n ] = x [ n − Δ ] ⇒ h [ n] = δ [ n − Δ ]
If Δ is not an integer , H c ( f ) = e − j 2π f ΔT
yc ( t ) = xc ( t − ΔT )
⇒ y [ n ] = xc ( nT − ΔT )
π ( t − ΔT − kT )
∞ sin
= ∑ x[k ]
k =−∞
T
π ( t − ΔT − kT )
T t = nT
∞
sin π ( n − k − Δ ) sin π ( n − Δ )
= ∑ x[k ] ⇒ h [ n] =
k =−∞ π (n − k − Δ) π (n − Δ)
1
e.g., Δ = ;
2
⎛ T⎞
yc ( t ) = xc ⎜ t − ⎟
⎝ 2⎠
−T 0 T 2T 3T

⎡ M⎤
– Ex: Moving average with non-integer delay y [ n] = w ⎢n − ⎥
⎣ 2⎦
⎧ 1
⎪ , 0≤n≤M
h [ n] = ⎨ M + 1
⎪⎩ 0, otherwise
ω ( M + 1)
sin ωM
⇒ H (e ) =
1 −j
jω 2 e 2
, ω <π
M +1 ω
sin
2
When M=5,
for x[n] = cos ( 0.25π n )
⇒ y[n] = 0.308cos ⎡⎣ 0.25π ( n − 2.5 ) ⎤⎦

• Changing the sampling rate

– Decimation by M (sampling rate compressor)
xn xd [ n] = x [ nM ]
↓M
x [ n ] = xc (nT ) ⇒ xd [ n ] Δ x [ nM ] = xc ( nMT )
1 ∞ ⎛ ω 2π k ⎞
X ( e jω ) = ∑ Xc ⎜ − ⎟
T k =−∞ ⎝ T T ⎠
Let T ′ = MT . Then,
1 ∞ ⎛ ω 2π k ⎞
X d ( e jω ) = ∑ Xc ⎜ ' − ' ⎟
T ' =−∞ ⎝ T T ⎠
1 ∞
⎛ ω 2π ⎞
=
MT
∑X
=−∞
c ⎜ −
⎝ MT MT ⎠
⎟ ⇐ = i + kM
1 ⎡1 ∞
M −1
⎛ ω − 2π i 2π k ⎞ ⎤
= ∑ ⎢ T ∑ X c ⎜ MT − T ⎟ ⎥
i =0 ⎣
M k =−∞ ⎝ ⎠⎦
1 M −1 ⎛ j ω −M2π i ⎞
= ∑ X ⎜e ⎟
M i =0 ⎝ ⎠

– Decimation by M (cont.)
To avoid aliasing, a LPF is
required before down-sampling
xn x [ n] xd [ n ]
LPF ↓M
T T ' = MT
hd [ n ]
Example: When M=3:

– Interpolation by L (sampling rate expander)

c h
Hi e jω
xe [ n] xi [ n]
L
xn
↑L LPF
T T '=T /L
hi [ n] −2π −π π π π 2π
−
T L L
xi [ n ] = xc ( nT ′ ) ; T ′ =
L X c (Ω)
⎡n⎤
= x⎢ ⎥
⎣L⎦ Ω
−W W
X (e jω )
⎧ ⎡n⎤
⎪x , n = 0, ± L, ±2 L,...
xe [ n ] = ⎨ ⎢⎣ L ⎥⎦ ω = ΩT
⎪ 0 , otherwise −2π −π π 2π
⎩
∞
X e (e jω )
= ∑ x [ k ]δ [ n − kL]
k =−∞ ω = ΩT ′
−3π / L −π / L 0 π /L 3π / L
⎛ ∞ ∞
⎞
⇒ X e ( e ) = ∑ ⎜ ∑ x [ k ]δ [ n − kL ] ⎟ e − jω n
jω
k =−∞ ⎝ k =−∞ ⎠ X i (e jω )
∞
= ∑ x[k ]e − jω Lk
= X ( e jω L )
ω = ΩT ′
k =−∞
−2π −π / L π /L 2π

– D/C conversion
πn
sin
hi [ n ] = L ⇒ ⎧1 n = 0
πn ⎨
⎩0 n = ± L, ±2 L...
L
⎡n⎤ ⎛ nT ⎞
xi [ n ] = ∑ x [ k ] hi [ n − k ] = x ⎢ ⎥ = xc (nT ' ) = xc ⎜ ⎟
k ⎣L⎦ ⎝ L ⎠
In practice, we use an approximate LPF.
– Linear interpolation
⎧ n
⎪1 − , n < L
h [ n] = ⎨ L
⎪ 0,
⎩ otherwise
x [ n ] = ∑ xe [ k ] h [ n − k ]
k c h
Hi e jω
= ∑ x [ m ] h [ n − mL ] H ce h: L = 5
jω
m
⎡ ωL ⎤
2
1 ⎢ sin 2 ⎥
H (e ) = ⎢
jω
⎥
π 2π 4π
L ⎢ sin ω ⎥ 5 5 5
⎣ 2 ⎦

L
– Changing the sampling rate by a non-integer ratio R =
M
xn xe [ n ] xi [ n ] xd [ n ]
↑L LPF LPF ↓M
T T MT
T L T L
L L
xn xe [ n ] xi [ n] xd [ n ]
↑L LPF ↓M
G=L
⎛π π ⎞
ω = min ⎜ , ⎟
c ⎝L M ⎠

L
– Changing the sampling rate by a non-integer ratio R = (cont.)
M

• Multirate signal processing
– Interchange the filtering and the down sampling process
x [n ] x a [n ] y [n ] x [n ] xb [n ] y [n ]
↓ M H (z ) H (z )M
↓ M
ω − 2π i
X b ( e jω ) = H ( e jω M ) X ( e jω )
M −1
X a (e ) =
1
∑ X (e
j
jω M
)
M i =0
⎛ j ω −M2π i ⎞
M −1
Y (e ) =
1
jω
∑ Xb ⎜ e ⎟
M i =0 ⎝ ⎠
1 M −1 ⎛ j ω −M2π i ⎞
= ∑ X ⎜e
M i =0 ⎝
⎟H e
⎠
(
j (ω − 2π i )
)
M −1
⎛ j ω −M2π i ⎞
= H (e )
1 jω
∑ X ⎜e ⎟
M i =0 ⎝ ⎠
= H ( e jω ) X a ( e jω )

• Multirate signal processing
– Interchange the filtering and the down sampling process
x [ n] xa [ n] y [ n] x [ n] xb [ n] y [ n]
↓M H (z) H ( zM ) ↓M
ω − 2π i
X b ( e jω ) = H ( e jω M ) X ( e jω )
M −1
X a (e ) =
1
∑ X (e
j
jω M
)
M i =0
⎛ j ω −M2π i ⎞ 1 ⎛ j ω −M2π i ⎞
( )
M −1 M −1
Y (e ) =
1
jω
∑ Xb ⎜ e ⎟= ∑ X ⎜e ⎟H e
j (ω − 2π i )
M i =0 ⎝ ⎠ M i =0 ⎝ ⎠
M −1
⎛ j ω −M2π i ⎞
= H (e ) ⎟ = H (e ) X a (e )
1 jω
∑ X ⎜e jω jω
M i =0 ⎝ ⎠
– Interchange the filtering and the interpolation process
x [ n] xa [ n] y [ n] x [ n] xb [ n] y [ n]
H (z) ↑L ↑L H ( zL )
Y ( e jω ) = X a ( e jω L ) = X ( e jω L ) H ( e jω L )
X b ( e jω ) = X ( e jω L ) ⇒ Y ( e jω ) = H ( e jω L ) X b ( e jω ) = X ( e jω L ) H ( e jω L )

– Poly phase implementation of interpolation filters
x [ n] u [ n] y [ n]
↑L H (z)
Note that only every Lth sample of u [ n ] is nonzero

Consider decomposition of an impulse response h [ n ] by
⎧h [ n + k ] , n = mL, m = 0, ±1, ±2, ⋅ ⋅⋅
hk [ n ] = ⎨
⎩ 0 , otherwise
By successively delaying these subsequences, we can reconstruct h [ n ] by
L −1
h [ n] h [ n] e0 [ n ] h0 [ n ] h [ n ] = ∑ hk [ n − k ]
↓L ↑L k =0
ek [ n ] = h [ nL + k ] = hk [ nL ]
z
h [ n + 1] e1 [ n ]
↓L ↑L z −1 +
z
h [ n + 2] e2 [ n ]
↓L ↑L z −2
...
...
...
z eM −1 [ n]
−( M −1)
↓L ↑L z
h[ n + M −1]
h [ n] e0 [ n ] x [ n]
↓L ↑L +
( )
E0 z L
z z −1
z −1
h [ n + 1] e1 [ n ] y [ n]
↓L ↑L + ( )
E1 z L
+
:
⇒
...
...
+
...
z −1
z eM −1 [ n] z −1
h[ n + M −1]
↓L ↑L ( )
EM −1 z L
x [n ] E0 ( z ) ↑L + y [n ]
⇒ z −1
E1 ( z ) ↑L +
z −1
E L −1 ( z ) ↑L
• Digital processing of analog signals
xc (t) Anti- x (t) x ⎡⎣n⎤⎦ x̂ ⎡⎣n⎤⎦ Discrete-time ŷ ⎡⎣n⎤⎦ Compensated y(t)

a
aliasing S&H A/D D/A reconstructure
system
filter filter
Δ Δ
– Quantization error: − < q [ n ] = x [ n ] − xˆ [ n ] ≤
2 2
x [n ] x̂ [ n ] x [n ] x̂ [ n ]
Q ⇔ +
↑
q [n ]
– Assuming that the error sequence q [ n ] is uncorrelated with x [ n ] and it is uniformly

⎡ Δ Δ⎤
distributed over ⎢ − , ⎥ ,
⎣ 2 2⎦ Δ
Δ
2
1
σ =∫
2 2
Δ q ⋅ dq =
2
Δ
q
− 12
2
2−2 k X m2
For a (K+1)- bit quantizer with full-scale value X m , σ =
2
– q
12
σ X2 X
SQNR = 10 log 2 = 6.02 K + 10.8 − 20log m
σq σX
⇒ SQNR is increases approximately 6dB for each bit added to the quantizer.

Seoul National University
Transform analysis of L.T.I. systems

• Frequency response of LTI systems
xn yn
hn
y [ n] = ∑ x [ k ] h [ n − k ] ⇔ Y ( z ) = H ( z ) X ( z )
k
⇒ Y (e ) = H (e ) X (e ) = H (e ) X (e ) e ( ( ) ( ))
jω jω jω jω jω j ArgH e jω + ArgX e jω
↑ eigenfunction x [ n ] = e jω n
complex gain (or eigenvalue)
– Ideal frequency-selective filters

⎧⎪1, ω ≤ ω c
H lp ( e
jω
) = ⎨0,
⎪⎩ otherwise ω c < ω ≤ π
⇒ H hp ( e jω ) = 1 − H lp ( e jω ) : ideal high - pass filter
⇒ hhp [ n ] = δ [ n ] − hlp [ n ]
sin ω c n
= δ [ n] − ⇐ not computationally realizable
Digital Signal Processing
πn Yong-Hwan Lee
– Ideal delay hid [ n ] = δ [ n − nd ]

⇒ H id ( e jω ) = e − jω nd = 1 e − jω nd , ω < π
– Low-pass filter with linear phase (or delay)

⎧⎪e − jω nd , ω ≤ ω c
H p (e ) = ⎨
jω
⎪⎩ 0 ,ω c ≤ ω ≤ π
sin ω c ( n − nd )
⇒ hhp [ n ] = ← Regardless of nd , it is still noncausal
π ( n − nd )
– Group delay: A measure of the nonlinearity of the phase
I (ω ) = −
d
dω
(
Arg H ( e jω ) )
(
Example Arg H ( e ) ≈ −φ0 − ω nd
jω
)
For an input x [ n ] = s [ n ] cos ω 0 n,
y [ n ] ≈ s [ n − nd ] cos (ω 0 ( n − nd ) − φ0 )
⇒ The time delay of the envelope s [ n ] of narrowband signal centered at ω 0 is

given by the negative of the slope of the phase at ω 0.
Digital Signal Processing Yong-Hwan Lee

• Linear constant coefficient difference equation

N M
∑ ak y [ n − k ] = ∑ bk x [ n − k ]
k =0 k =0
N M
⇒ ∑ ak z − kY ( z ) = ∑ bk z − k X ( z )
k =0 k =0
M
Y ( z) ∑
−k
bz k
It does not uniquely specify
⇒ H (z) Δ = k =0
⇐ the impulse response of an
X ( z) N
∑a z k =0
k
−k
LTI system
b0 ∏ (1 − ck z −1 )
M
= k =1
⇒ zeros: z = ck ; poles: z = d k
a0 ∏ (1 − d k z )
N
−1
k =1
Example
H (z) =
(1 + z )
=
−1 2
1 + 2 z −1 + z −2
=
Y ( z) 1 3
⇒ zeros: z = −1, −1; poles: z = , −
⎛ 1 −1 ⎞⎛ 3 −1 ⎞ 1 + 1 z −1 − 3 z −2 X ( z ) 2 4
⎜1 − z ⎟⎜ 1 + z ⎟
⎝ 2 ⎠⎝ 4 ⎠ 4 8
1 3
⇒ y [ n] + y [ n − 1] − y [ n − 2] = x [ n ] + 2 x [ n − 1] + x [ n − 2]
4 8
– Causal system
h [ n ] is a right-sided sequence
⇒ The region of convergence of H c ( z ) should be outside the outermost pole
– Stable system: absolutely summable
∞
∑ h [ n] < ∞ ⇔ ∑ h [ n] z
n =−∞ n
−n
< ∞ for z =1
⇒ The ROC of H(z) includes the unit circle

– Example 5
y [ n ] − y [ n − 1] + y [ n − 2] = x [ n ]
2
1 1
⇒ H (z) = = 1 2
1 − z −1 + z −2 ⎛⎜1 − z −1 ⎞⎟ (1 − 2 z −1 )
5 1 2
2 ⎝ 2 ⎠
To be causal, ROC ⇒ z >2
1
To be stable, ROC ⇒ 2 < z < 2
– In order for an LTI system to be both stable and causal, the ROC must be outside the
outermost pole and include the unit circle, i.e., all poles inside the unit circle
– Depending upon the choice of ROC, the same difference equation results in a different
impulse response

– Inverse system
1
Let H ( z ) be H ( z ) =
i i H (z)
G( z) H ( z ) H i ( z ) = 1 ⇒ g [ n ] = h ( n ) ∗ hi [ n ] = δ [ n ]
To hold this equation, the ROC of Hi(z) and H(z) must overlap
If H(z) is causal, the ROC is z > max
k
dk
⇒ The ROC of H i ( z ) should overlap with z > max d k
k
Example 1:
1 − 0.5 z −1
H (z) = with ROC: z > 0.9
1 − 0.9 z −1
1 − 0.9 z −1
⇒ Hi ( z ) = ⇒ ROC: z > 0.5 (to be overlapped with z > 0.9)
1 − 0.5 z −1
⇒ hi [ n ] = 0.5n u [ n ] − 0.9 ⋅ 0.5n−1 u [ n − 1]
Example 2:
z −1 − 0.5
H ( z) = with ROC : z > 0.9
1 − 0.9 z −1
−2 + 1.82 z −1
⇒ Hi ( z ) =
1 − 2 z −1
For an ROC z < 2; hi1 [ n ] = 2 ⋅ 2n u [ − n − 1] − (1 ⋅ 8 ) ⋅ 2n−1 u [ − n ] ⇒ stable & noncausal
For an ROC z > 2; hi 2 [ n ] = −2 ⋅ 2n u [ n ] + 1 ⋅ 8 ⋅ 2n−1 u [ n − 1] ⇒ unstable & causal
If H(z) is causal with zeros at ck , H i ( z ) will be causal iff the Roc of H i ( z ) is

z > max ck
k
If H i ( z ) is stable, the RoC of H i ( z ) must include the unit circle, max

k
ck < 1
H(z) and its inverse are stable and causal iff both poles and zeros of H(z) are
inside the limit circle. ⇒ minimum-phase system
– Impulse response of for rational system functions

M −N N
Ak
H (z) = ∑ B z
−
+ ∑1− d −1
,M ≥ N
=0 k =1 k
z
FIR term IIR term
If causal,
M −N N
h [ n] = ∑ B δ [ n − ] + ∑ A d u [ n], k
n
k M ≥N
=0 k =1
M −N N
⇒ y [ n] = ∑ B x[n − ] + ∑ Ak d knu [ n] ∗ x [ n]
=0 k =1
Example: A first-order IIR filter

1
y [ n ] − ay [ n − 1] = x [ n ] ⇒ H ( z ) = ; ROC : z > a
1 − az −1
For stability, a < 1 ⇒ h [ n ] = a nu [ n ]
FIR system
M
H ( z ) = ∑ bk z − k
k =0
M ⎧b , 0 ≤ n ≤ M
⇒ h ⎡⎣ n ⎤⎦ = ∑ bkδ ⎡⎣ n − k ⎤⎦ = ⎪⎨ n
k =0 ⎪⎩0, otherwise
M
⇒ y ⎡⎣ n ⎤⎦ = ∑ bk x ⎡⎣ n − k ⎤⎦
k =0
Example
⎧a n , 0 ≤ n ≤ M
h [ n] = ⎨
⎩0 , otherwise M =7
1 − a M +1 z (
M − M +1)
⇒ H (z) = ∑a z n −n
=
n =0 1 − az −1
2π k
j
Zeros : zk = ae M +1
, k = 0,1, 2,..., M
M
⇒ y [ n] = ∑ a k x [ n − k ]
k =0
or y [ n ] − ay [ n − 1] = x [ n ] − a M +1 x [ n − M − 1]

• Frequency response for rational systems
∏ (1 − c e ω )
M M
∑ bk e− jωk k
−j
H ( e jω ) = H ( z ) z =e jω =
b0
k =0
= k =1
∏ (1 − d e ω )
N N
∑a e − jω k a0 −j
k k
k =0 k =1
M
∏ 1− c e − jω
∏ (1 − c e ω )(1 − c e ω )
M
k −j * j
⇒ H (e )
2
jω b ⎛b ⎞ k k
= 0 H ( e jω )
k =1 2
N =⎜ 0 ⎟ k =1
∏ 1− d e
a0 − jω
∏ (1 − d e ω )(1 − d e ω )
N
k ⎝ a0 ⎠ −j * j
k k
k =1
k =1
– Gain (or attenuation)
Gain = 20log10 H ( e jω )
M M
b0
= 20log10 + 20∑ log10 1 − ck e − jω
− 20∑ log10 1 − d k e− jω (dB )
a0 k =1 k =1
– Phase response
⎛ b0 ⎞ M
Arg ( H (e ) ) = Arg ⎜ ⎟ + ∑ Arg (1 − ck e ) − ∑ Arg (1 − d k e − jω ) ; − π < Arg ( x ) ≤ π
M
jω − jω
⎝ a0 ⎠ k =1 k =1
– Group delay
grd ( H (e ) ) = −∑ Arg (1 − ck e ) + ∑ Arg (1 − d k e − jω )

M M
jω d − jω d
k =1 dω k =1 dω

• Frequency response of a single pole or zero
– Single zero
H ( z ) = 1 − az −1 ⇒ H ( e jω ) = (1 − ae − jω ) ; a = re jθ
⇒ H ( e jω ) = 1 − re jθ e− jω = (1 − re jθ e − jω )(1 − re − jθ e jω )
2
= 1 + r 2 − 2r cos (ω − θ )
r sin (θ − ω )
( )
⇒ A rg H ( e jω ) = tan −1
1 − r cos (θ − ω )
(
⇒ grd H ( e jω ) = ) d ⎡
dω ⎣
(
A rg H ( e jω ) ⎤
⎦ ) r = 0.9
r 2 − r cos (θ − ω ) θ =0
φ3 =
1 + r 2 − 2r cos (θ − ω ) θ = 0.5π
` θ =π
v2 v
z−a
H ( z) =
3
θ
ω v1 z
e jω − re jθ
⇒ H (e )
jω v3
= = = v3
e jω v1
( )
A rg H ( e jω ) = φ3 − φ1 = φ3 − ω
– Example: Real single zero
H ( z ) = 1 − az −1
⇒ H ( e jω ) = (1 − ae − jω ) ; a = re jθ
If θ =π , a = − r
⇒ H (e jω
) = 1 + re − jω 2
= 1 + r 2 − 2r cos ω
r sin ω
(
⇒ A rg H ( e jω ) = tan −1 )
1 + r cos ω
(
⇒ grd H ( e jω ) =
d ⎡
dω ⎣
)
A rg H ( e jω ) ⎤
⎦( )
r 2 + r cos ω
=
1 + r 2 + 2r cos ω

– Second-order IIR system

1 1
H ( z) = =
(1 − re jθ
z −1 )(1 − re jθ z −1 ) 1 − 2r cosθ z −1 + r 2 z −2
⇔ y [ n ] − 2r cosθ y [ n − 1] + r 2 y [ n − 2] = x [ n ]
r n sin ( n + 1)θ
⇒ h [ n] = u [ n]
sin θ
v1
ω v3
−θ
v2
H (e )
jω v 1
= 3 =
v1 ⋅ v2 v1 v2
r sin (θ − ω ) r sin (θ + ω )
( )
A rg H ( e jω ) = − tan −1
1 − r cos (θ − ω )
− tan −1
1 − r cos (θ + ω )
1
H (z) =
1 − 0.9 2 z −1 + 0.81z −2
– Third-order IIR system
0.05634 (1 + z −1 )(1 − 1.0166 z −1 + z −2 )

H ( z) =
(1 − 0.683z )(1 − 1.4461z
−1 −1
+ 0.7957 z −2 )
⇒ z z = −1, e ± j1.0376 (59.45 )
⇒ z p = 0.683, 0.892e ± j 0.6257 (35.85 )

• Relationship between magnitude and phase
– In general, magnitude information ⎯ ⎯→

No
←⎯⎯ phase information
No
– In cases of rational system functions, there is some constraint between magnitude and
phase
⎛1⎞
H ( e jω ) = H ( e jω ) H ∗ ( e jω ) = H ( z ) H ∗ ⎜ * ⎟
2
⎝ z ⎠ z =e jω
∏ (1 − c z ) ∏ (1 − c z )
M M
−1 ∗
⎛1⎞ b
k k
b0
H ( z) = k =1
; H∗⎜ ∗ ⎟ = 0 k =1
∏ (1 − d z ) ⎝ z ⎠ a0
∏ (1 − d z )
N N
a0 −1 ∗
k k
k =1 k =1
– The square of the magnitude frequency response is the evaluation of z-transform on

the unit circle
C (z) Δ H (z)H * ( )
1
z
*
– Question: Can we know of H(z) from C(z) ?
H ( z ): pole d k ; zero ck
⎛1⎞ conjugate reciprocal pairs
H ∗ ⎜ ∗ ⎟ : pole ( d k∗ ) ; zero (c )
−1 ∗ −1
k
⎝z ⎠
– If H(z) is assumed to be causal & stable, all its poles are in the unit circle
⇒ Poles of H(z) can be identified uniquely
Example System with the same C(z)
2 (1 − z −1 )(1 + 0.5 z −1 )
H1 ( z ) = π π
⎛ j
4 −1
⎞⎛ −j
4 −1
⎞
⎜ 1 − 0.8e z ⎟⎜ 1 − 0.8e z ⎟
⎝ ⎠⎝ ⎠
H2 ( z ) =
(1 − z )(1 + 2 z )
−1 −1 −0.5 −2
π π
⎛ j
4 −1
⎞⎛ −j
4 −1
⎞
⎜1 − 0.8e z ⎟⎜ 1 − 0.8e z ⎟
⎝ ⎠⎝ ⎠
⇒ H1 ( e jω ) = H 2 ( e jω )
⎛1⎞
C1 ( z ) = H1 ( z ) H1* ⎜ * ⎟
⎝z ⎠
⎛1⎞
C2 ( z ) = H 2 ( z ) H 2* ⎜ * ⎟ ⇒ C1 ( z ) = C2 ( z )
⎝z ⎠

• All pass system
– A stable system function of the form

z −1 − a ∗
H ap ( z ) =
1 − az1
has a frequency-response magnitude independent of ω.
e − jω − a∗ 1 − a∗e jω − jω
H ap ( e ) =
jω
− jω
= − jω
e ⇒ H ap ( e jω ) = 1
1 − ae 1 − ae
r sin (ω − θ )
( )
A rg H ap ( e jω ) = −ω − 2 tan −1
1 − r cos (ω − θ )
– In general, the transfer function of an all-pass system is given by
H ap ( z ) = A ∏
z −1 − d k
Mr Mc
(z −1
− ek
∗
)( z −1
− ek )
k =1 1 − d k z
−1 ∏ (1 − e z )(1 − e z ) ;
k =1
−1 ∗ −1
A > 0, d k < 1, ek < 1
k k
– All-pass system can be used for compensating group delay distortion.

– Causal all-pass system has positive group delay (but not continuous phase) property

– Example
z −1 ∓ 0.9
H (z) =
1 ∓ 0.9 z1
z p = 0.9
z p = −0.9

– Example
0.8
π
4
4 1 2
− − .75 0.5
3

• Minimum-phase system
– It has its poles and zeros all inside the unit circle.
⇒ Any minimum-phase system H ( z ) is causal and stable.
⇒H
−1
(z) is a causal, stable, and min-phase system
– Cascade of both systems will not introduce any delay
H ( z ) H −1 ( z ) = 1 ⋅ e jo
– Energy concentration
For an arbitrary stable and causal system F ( e jω ) , let G ( e jω ) be the min-phase
function s.t., G ( e jω ) = F ( e jω ) , ∀ω .
Then, for any sequence x [ n ] and K,

K K
∑ ( x [ n ] ∗ g [ n ])k ≥ ∑ ( x [ n ] ∗ f [ n ])k
2 2
k =0 k =0
⇒ The signal energy is transported with the smallest possible delay using a
minimum-phase system
K K
Example If x [ n ] = δ [ n ] , ∑ g 2 [k ] ≥ ∑ f 2 [k ]
k =0 k =0

– The min-phase function has its energy maximally concentrated near the time origin.
proof
F ( e jω )
Let H ( e jω
)Δ , ∀ω be stable, causal and all pass
G ( e jω )
H ( e jω ) = 1
xn yn zn
c h
G e jω c h
H e jω
min-phase all-pass
phase shifter
The energy of y [ n ]
ω0
2π
E y Δ E { y [ n ]} = Y ( e jω ) dω , ω 0 =
1 2
∫ω
2 2
ω0 −
2
0
T
ω0
Ez Δ E { z 2 [ n ]} = H ( e jω ) Y ( e jω ) dω = E y
1 2
ω 0 ∫− 2
2
ω0
K
Assume y [ n ] vanishes for n < 0 a n d n < K , i .e ., Ey = ∑
n=0
y 2 [n ]
But z [ n ] is not necessarily zero for n > K
K
Ez = Ey ≥ ∑
n=0
z 2 [n ]
– H (z) cannot be uniquely determined from G ( z ) , because cascading all-pass filters
does not affect the magnitude of the frequency response.

– Any rational rational transfer function can be expressed as

H (z ) = H m in ( z ) H a p ( z )
Assume that H ( z ) has all zeros and poles inside the unit circle except one
zero outside the unit circle at z =
1
*
, c <1
c
H ( z ) = H1 ( z )( z −1 − c* )
−1 z −1 − c*
= H1 ( z ) ⋅ (1 − cz ) ⋅
1 − cz −1
minimum phase
all pass
where H1 ( z ) is minimum phase
H min ( z ) has poles and zeros of H ( z ) inside the unit circle plus zeros that are
conjugate reciprocals of the zeros of H ( z ) outside the unit circle
H ap ( z ) contains all the zeros of H ( z ) outside the unit circle

Example
1 −1 1
−1 1+ z z −1 +
1 + 3z
H (z ) = 1
= 3⋅ 3
1
⋅
1
3
1 + z −1 1 + z −1 1 + z −1
2 2 3
H m in ( z ) H ap ( z )

– The amplitude characteristics of a minimum-phase filter is fully determined by the

phase characteristics and vice versa ⇒ We can compute G ( e jω ) from G ( e jω )
Proof
Since G ( z ) has its poles and zeros inside the unit circle, ln G ( e jω ) is analytic for z > 1
⇒ ln G ( e jω ) has a causal and well behaved inverse Fourier transform, i.e.,
∞
ln G ( e jω
) = ∑ g [ n] e − jω n
= X ( e jω ) + jφG ( e jω )
n =0
∞ ∞
where X ( e jω ) = ln G ( e jω ) = ∑ g [ n ] cos nω ; φG ( e jω ) = −∑ g [ n ] sin nω
n =0 n=0
We can see that X ( e ) φG ( e )

jω jω
and form a Hilbert transform pair.
Since X ( e jω ) is the logarithm of a real-valued symmetric function of frequency, it
is real and symmetric ⇒ x [ n ] = x [ −n ] , ∀n
∞
⇒ X ( e jω ) = x [ 0] + 2∑ x [ n ] cos nω
n =0
⇒ g [ 0] = x [ 0] and g [ n ] = 2 x [ n ] , n ≥ 1
Thus we can compute φG ( e )

jω

– Frequency-response compensation
If H d ( z ) is a min-phase system, perfect compensation is generally possible
sn sc n
Hd z bg Hc zbg
H d ( z ) = H min ( z ) H ap ( z )
1
Let H c ( z ) = ⇒ G ( z ) = H d ( z ) H c ( z ) = H ap ( z )
H min ( z )
The frequency response magnitude is exactly compensated with some delay

Example
H ( z ) = (1 − 0.9e j 0.6π z −1 )(1 − 0.9e − j 0.6π z −1 )(1 − 1.25e j 0.8π z −1 )(1 − 1.25e− j 0.8π z −1 )
⇒ H min ( z ) = 1.252 (1 − 0.9e j 0.6π z −1 )(1 − 0.9e − j 0.6π z −1 )
⋅ (1 − 0.8e j 0.8π z −1 )(1 − 0.8e− j 0.8π z −1 )
⇒ H ap ( z) =
(1 − 1.25e j 0.8π
z −1 )(1 − 1.25e − j 0.8π z −1 )
(1 − 0.8e j 0.8π
z −1 )(1 − 0.8e − j 0.8π z −1 )

H ( z) H min ( z ) H ap ( z )

In general, min-energy delay property h [ 0 ] ≤ h m in [ 0 ]

z −1 − a
H ( z ) = H m in ( z ) , a <1
1 − a z −1
1 − a z −1
lim H m in ( z ) = lim − 1 H (z )
z→ ∞ z→ ∞ z −a
1
⇒ h m in [ 0 ] = − h [ 0 ] ⇒ h [ 0 ] = − a h m in [ 0 ] a <1
a
⇒ h [ 0 ] ≤ h m in [ 0 ]
– Maximum phase system is a stable system whose poles and zeros are all outside the
unit circle M
∏ (z − c ) k
H (z ) = k =1
N
c k > 1, d k > 1
∏ (z − d )
k =1
k
stable ⇒ R oc : z < smallest dk

⇒ all poles contribute to h [ n ] with terms of the form − ( d k ) u [− n − 1]
n
⇒ anti causal!

Note that FIR max-phase sequence can be made causal by introducing a finite
delay
z −1 − c k
H m a x ( z ) = h m in [ 0 ] ∏ (1 − c ∗
)∏ 1 − c k∗ z
k
= H m in ( z ) H a p ( z )
– Property of min-phase systems
minimum phase-lag
arg ⎡⎣ H ( e jω ) ⎤⎦ = arg ⎡⎣ H min ( e jω ) ⎤⎦ + arg ⎡⎣ H ap ( e jω ) ⎤⎦
negative for 0 ≤ ω ≤ π
minimum group-delay
minimum energy delay property
h [ 0 ] ≤ h m in [ 0 ]

Seoul National University 4. Structure for discrete-time systems
Structure for discrete-time systems

• Characterization of an LTI system can be represented by difference equation, impulse
response or system function
b0 + b1 z −1
Example: H (z) = , z >a
1 − az −1
⇒ Impulse response : h [ n ] = b0 a nu [ n ] + b1a n −1u [ n − 1]
⇒ Infinite duration impulse response
⇒ Not possible to implement the system by discrete time convolution
⇒ The output can be calculated by a recursive computation algorithm
y [ n ] = ay [ n − 1] + b0 x [ n ] + b1 x [ n − 1]
• For description of systems for implementation, use the block diagram and/or signal flow
graph method

• Basic block diagram symbols

yn
zn =xn +yn
xn
xn a y n = ax n
y n = x n −1
xn −1
z
Example y [ n ] = a1 y [ n − 1] + a2 y [ n − 2] + bx [ n ]
xn b yn
z −1
a1
y n −1
z −1
a2

y [ n ] = ∑ ak y [ n − k ] + ∑ bk x [ n − k ]
N M
–
k =1 k =0
bg
H1 z H2 z bg
xn b0
yn
z −1 z −1
b1 a1
x n −1 y n −1
z −1 z −1
b2 a2
x n−2
: :
:
bM −1 a N −1
x n − M +1
y n − N +1
z −1 z −1
x n− M bM aN
y n− N
< direct form I>

– By interchanging the cascaded blocks,

Y (z)
= H (z)
X (z)
= H1 ( z ) H 2 ( z )
= H 2 ( z ) H1 ( z )
x n ω [n ] b0 y n x [n ] b0 y [n ]
z −1 z −1 z −1
ω [n − 1]
a1 b1 a1 b1
z −1 z −1 :
a2 : z −1
ω [n − 2 ] ⇒ a2
: b2
a N −1 b N −1 :
:
a N −1
−1 −1
z z b N −1
aN ω [n − N ] bN
z −1
aN
bN
< direct form II or canonic direct form >

• Signal flow graph representation
node j branch j , k
ω j ⎣⎡ n ⎦⎤ ω k ⎡⎣ n ⎤⎦
node k
Example
d
a
b e
x n ω1 n c ω2 n y n
ω1 [ n ] = x [ n ] + aω 2 [ n ] + bω 2 [ n ]
ω 2 [ n ] = cω1 [ n ]
y [ n ] = dx [ n ] + eω 2 [ n ]

Example
xn ωn xn ωn b0 yn
yn
b0
z −1
−1
z
a b1
ω n −1
a b1
ω [ n ] = x [ n ] + aω [ n − 1]
y [ n ] = b0ω [ n ] + b1ω [ n − 1]
⇒ W ( z ) = X ( z ) + aW ( z ) z −1 ⇒ X ( z ) = (1 + az −1 )W ( z )
Y ( z ) = b0W ( z ) + bW
1 ( z ) z −1 ⇒ Y ( z ) = ( b0 + b1z −1 )W ( z )
Y (z) X (z)
⇒W (z) = =
b0 + b1 z −1 1 + az −1
Y ( z ) b0 + b1 z −1
⇒ H (z) = =
X ( z ) 1 + az −1

• Basic structures for IIR systems
– Direct form I
1 + 2 z −1 + z −2
H (z) =
1 − 0.75 z −1 + 0.125 z −2
xn 1 yn
z −1 z −1
2 0.75
z −1 z −1
1
−0125
.
– Direct form II
xn yn
z −1
2
0.75
z −1
−0125
.

– Cascade form
∏ (1 − g ) ∏ (1 − h z )(1 − h z )
M1 M2
−1 −1 * −1
k
z k k
H ( z ) = A kN=1 k =1
∏ (1 − c z )∏ (1 − d z −1 )(1 − d k* z −1 )
1 N2
−1
k k
k =1 k =1
1 − ( d k + d k* ) z −1 + d k z −2
2
z −1 z −1 z −1
cd + d h
1
*
1
z −1
c
− h1 + h *
1 h z −1 − ( h2 + h2* ) c1 −g1
− d1
2
h1
2
cd 2 + d 2* h − d2
2
h2
2
Example H (z) =
(1 + z )(1 + z ) −1 −1
(1 − 0.5 z )(1 − 0.25 z )−1 −1
xn yn
z −1 z −1
0.25 0.5

– Parallel form
Np N1
Ak N Bk (1 − ek z −1 )
H ( z ) = ∑ gk z + ∑ +∑
2
−k
k =1 1 − ck z k =1 (1 − d k z )(1 − d k z )
−1 −1 * -1
k =0
N = N1 + 2 N 2 N2
e0 k + e1k z −1
M ≥ N, Np = M − N
∑
k =1 1 − a1k z
−1
− a2 k z −2
g0
z −1
g1
z −1
g2
: : :
gNp
x n e01 yn
z −1
a 11
e11
z −1
a 21
A1
z −1
c1

1 + 2 z −1 + z − 2
Example H (z) =
1 − 0.75 z −1 + 0.125 z −1
−7 + 8 z −1
=8+
1 − 0.75 z −1 + 0.125 z −1
18 25
=8+ −
1 − 0.5 z −1 1 − 0.25 z −1
xn −7 yn
z −1
0.75 8
z −1
−01125
.
xn 8 18 yn
0.5 z −1
−0.25
0.25 z −1

• Transposed forms
– Signal flow graph methods provide procedures for transforming graphs into different forms
– Transposition of a flow graph (flow graph reversal) is obtained by reversing the directions
of all branches, while keeping the branch transnittances and reversing the roles of the
input and output
1
Example H (z) =
1 − az −1 x n yn
z −1
a
Transposed form
yn x n
The difference is
z −1
the change in ordering
a
xn yn
z −1
a
Example Transposed form of a second-order structure

xn ωn b0 yn
z −1
ω [ n ] = a1ω [ n − 1] + a2ω [ n − 2] + x [ n ]
a1 b1
z −1 y [ n ] = b0ω [ n ] + b1ω [ n − 1] + b2ω [ n − 2]

a2 b2
Transposed direct form II

yn xn
b0
z −1
a1 b1
−1
z
a2 b2
xn b0 v0 n yn
b1 z −1 a1
b2 z −1 a
2

• Basic network structure for FIR systems
– Direct form
⎧bn , 0 ≤ n ≤ M
Causal FIR systems y [ n ] = ∑ bk x [ n − k ] ⇒ h [ n ] = ⎨
M
k =0 ⎩0 , otherwise
Transversal filter or
xn z −1 z −1 z −1
" tapped delay line
h0 h1 h2 h[M ]
yn
In the transposed form,
yn z −1 z −1
"
h0 h1 h2 h M
xn
"
z −1 z −1 z −1 z −1 yn
⋅ ⋅ ⋅
h M h M −1 h [ M − 2] h0
xn
"

– Cascade form
H ( z ) = ∑ h [ n ] z − n = ∏ ( b0 k + b1k z −1 + b2 k z −2 ); M s = ⎢⎣( M + 1) / 2 ⎥⎦
M Ms
k =0 k =1
x [ n] b01 b02 b0Ms y [ n]

"
z−1 z−1 z−1
b11 b12 b1Ms

−1
−1 −1
z z z
b21 b22 b2Ms

– Structure for linear phase FIR systems

Causal linear phase FIR systems
h [ M − n ] = h [ n ] or h [ M − n ] = − h [ n ] for 0 ≤ n ≤ M
⇒ y [ n] = ∑ h [ k ] x [ n − k ]
M
k =0
M
−1
⎡M ⎤ ⎡ M⎤
∑ h [ k ] x [ n − k ] + h ⎢ 2 ⎥x ⎢ n − h[k ] x[n − k ]
2 M
= + ∑
2 ⎥⎦ k = M / 2+1
k =0 ⎣ ⎦ ⎣
M M
−1 −1
⎡M ⎤ ⎡ M⎤ 2
= ∑ h[k ] x [n − k ] + h ⎢ ⎥ x ⎢n − ⎥ + ∑ h[ M − k ] x [n − M + k ]
2
k =0 ⎣2⎦ ⎣ 2 ⎦ k =0
When h [ M − n ] = h [ n ] (type - I system),
M
−1
⎡M ⎤ ⎡ M⎤
y [ n] = ∑ h [ k ] ( x [ n − k ] + x [ n − M + k ]) + h ⎢ 2 ⎥ x ⎢ n −
2
k =0 ⎣ ⎦ ⎣ 2 ⎥⎦
z −1 z −1 z −1
z −1 z −1 z −1
"
h0 h1 h2 LM M − 1OP LM M OP
h
N2 Q h
N2Q
yn
Symmetry condition
⇒ zeros of H ( z ) occur mirror image pairs
1
z1*
z3
z1
z4 z2 1
z* z2
1
*
z 3
1
z1
If h [ n ] is real, zeros of H ( z ) occur in complex conjugate pairs.
H ( z ) = h [ 0] (1 + z −1 )(1 + az −1 + z −2 )(1 + bz −1 + z −2 )(1 + cz −1 + dz −2 + cz −3 + z −4 )

2
1 ⎧ 1⎫ 1
where a = z2 + , b = 2 Re{ z3 } , c = −2Re ⎨ z1 + ⎬ , d = 2 + z1 +
z2 ⎩ z1 ⎭ z1

Seoul National University 5. Filter design techniques
Filter Design Techniques

• Design procedure
– Specify the desired frequency Response

– Approximate the specifications
⎛ω ⎞
– Realize the filter H ( e jω ) = H eff ⎜ ⎟ , ω <π
⎝T ⎠

• Design of discrete-time IIR filters from continuous-time filters
– Filter design by impulse invariance

h [ n ] = Td hc ( nTd )
bg
xa τ xn yn bg
ya τ
C/D hn D/C
T Td T
⎛ ω 2π ⎞ ω
⇔ 2π f ( = Ω )
∞
H ( e jω ) =
∑ H c⎜ + k ⎟
k =−∞ ⎝ Td Td ⎠ T d
The discrete-time filter specifications are transformed to continuous-time filter

⎛ω ⎞
specifications by H ( e jω ) = H c ⎜ ⎟ , ω ≤ π , if H c ( f ) = 0, f ≥
1
⎝ Td ⎠ 2Td
Design a suitable continuous-time filter H c ( s)
Transform to the desired discrete-time filter H ( z )

⎧ N A es t , t ≥ 0
For H c ( s ) = ∑ k , hc ( t ) = ⎪⎨∑
k
N
A k
k =1
k =1 s − sk
⎪⎩ 0 , t >0
By sampling hc ( t ) ,
h [ n ] = Td hc [ nTd ] = ∑ Td Ak e s nT u [ n ] = ∑ Td Ak ( e s T ) u [ n]
N N n
k d k d
k =1 k =1
N
Td Ak
⇒ H (z) = ∑
k =1 1 − e z −1
sT
k d
A pole at s = sk in the s-plane is transformed to a pole at z = e s T in the z-plane. k d
If H c ( s ) is stable, the real part of sk < 0

⇒ es T < 1
k d
⇒ pole is inside the unit circle

⇒ H ( z ) is also stable
Although the poles in the s-plane are mapped in the z-plane in a simple manner,
the zeros are not.

– Butterworth (BW) filter
The magnitude response of a BW filter is
1 1
Hc ( f ) ⇐ H ( s ) H ( −s ) =
2
= 2N 2N
⎛ f ⎞ ⎛ s ⎞
1+ ⎜ ⎟ 1+ ⎜ ⎟
⎝ fc ⎠ ⎝ 2π f c ⎠
where f c is the 3dB frequency
⎧ j
π
( 2 k + N −1) ⎫
⇒ poles are ⎨ sk = Ω c e 2N
; k = 0,1, 2, , 2 N − 1⎬
⎩ ⎭

Example Design a continuous-time Butterworth (BW) filter s.t.
1
0.89125 ≤ H c ( f ) ≤ 1, 0 ≤ f ≤ 0.1
1− δ1
1
H c ( f ) ≤ 0.17783 0.15 ≤ f ≤
2 δ2
Let Td = 1 so that ω = Ω
ωp ωs
The magnitude response of a BW filter is
1 1 δ 1 = 0.10875
Hc ( f ) ⇐ H ( s ) H ( −s ) =
2
= 2N 2N
⎛ f ⎞ ⎛ s ⎞ δ 2 = 0.17783
1+ ⎜ ⎟ 1+ ⎜ ⎟
⎝ fc ⎠ ⎝ 2π f c ⎠ ω p = 0.2π
where f c is the 3dB frequency ω s = 0.3π
By solving
⎧ ⎛ 0.1 ⎞ 2 N ⎛ 1 ⎞
2
⎪1 + ⎜ ⎟ =⎜ ⎟ ⇒ N = 5.8858 ⇒ N = 6
⎪ ⎝ c ⎠f ⎝ 0.89125 ⎠
⎨ 2N f c = 0.11216 0.7032
⎪ ⎛ 0.15 ⎞
2
⎛ 1 ⎞
⎪1 + ⎜ f ⎟ = ⎜ 0.17783 ⎟
⎩ ⎝ c ⎠ ⎝ ⎠
π
⎛ −0.182 ± j 0.679 ⎞
⇒ poles are ⎜⎜ −0.497 ± j 0.497 ⎟⎟
12
sk = 2π f c e
j ( 2 k + N −1)
⎛ 0.1 ⎞
1+ ⎜ ⎟ = 1.258928
2N
⎜ −0.679 ± j 0.182 ⎟ ⎝ fc ⎠
⎝ ⎠
⇒ f c = 0.111918

0.12093
Hc ( s ) =
( s + 0.3645 + 0.4945)( s + 0.99455 + 0.4945) ⋅ ( s 2 + 1.35855 + 0.4945)
2 2
A1 A2
= +
( s + 0.182 + j 0.679 ) ( s + 0.182 − j 0.679 )
A3 A4
+ +
( s + 0.497 + j 0.497 ) ( s + 0.497 − j 0.497 )
A5 A6
+ +
( s + 0.679 + j 0.182 ) ( s + 0.679 − j 0.182 )
A1 A2
⇒ H (z) = + + ⋅⋅⋅
(1 − e−0.182e− j 0.679 z −1 ) (1 − e0.182e+ j 0.679 z −1 )
z1 = e −0.182− j 0.679 = 0.8336 ( 0.7782 − j 0.6280 )
= 0.6487 − j 0.5235 ⇒ z1 = 0.69489
2
0.2871 − 0.4466 z −1
H (z) =
1 − 1.2974 z −1 + 0.6949 z −2
−2.1428 + 1.1455 z −1
+
1 − 1.0691z −1 + 0.3699 z −2
1.8557 − 0.63032 z −1
+
1 − 0.9972 z −1 + 0.2570 z −2
This method is useful only for band-limited filters

– Bilinear transformation
To avoid aliasing, we employ the bilinear transform s = 2 ⎛⎜ 1 − z −1 ⎞⎟
−1
Td ⎝ 1 + z ⎠
2π f
ω
∞ −∞ −π π
⎡ 2 ⎛ 1 − z −1 ⎞ ⎤
H ( z ) = Hc ⎢ ⎜ −1 ⎟ ⎥
⎣ Td ⎝ 1 + z ⎠ ⎦
π
2 ⎛ 1 − z −1 ⎞ 1+ s
s= ⎜ −1 ⎟
⇒z= 2 ⇐ s = σ + j 2π f
Td ⎝ 1 + z ⎠ 1− s
Td
2
σT
1 + d + jπ fTd
= 2
σ Td
1− − jπ fTd
2
⇒ z < 1, if σ < 0

The j 2π f axis of the s-plane maps onto the unit circle of the z-plane
For s = j 2π f ,
1 + jπ fTd 1 + jπ fTd
z= ⇒ z =1 ⇒ e jω =
1 − jπ fTd 1 − jπ fTd
or 2 ⎛ 1 − e − jω ⎞
⎟ ( = σ + j 2π f )
s= ⎜ jΩ Im
Td ⎝ 1 + e− jω ⎠
⎛ω ⎞
2e − jω 2 j sin ⎜ ⎟
2 ⎝2⎠ σ
= Re
Td ⎛ω ⎞
2e − j 2 ω cos ⎜ ⎟ −π
⎝2⎠
2 ω
= j tan ω
Td 2
π
1 ω ⎛ 2 ω⎞
⇒ σ = 0 and f = tan ⎜ or Ω = tan ⎟
π Td 2 ⎝ Td 2⎠
⎛T Ω⎞
or ω =2tan -1 ⎜ d ⎟ b
Ω or 2πf g
⎝ 2 ⎠
−π
ω = 2 tan −1
FG ΩT IJ
H2K
d
= 2 tan −1 bπT f g
d

Ω
Ωs
Ωp
Hc (Ω)
ω
ωp ωs π ω

Example Bilinear transformation of a BW filter
0.89125 ≤ H ( e jω ) ≤ 1, 0 ≤ ω ≤ 0.2π
H ( e jω ) ≤ 0.17783, 0.3π ≤ ω ≤ π
2 0.2π
0 ≤ ω ≤ 0.2π ⇒ 0 ≤ Ω ≤ tan = 0.6498; Td = 1
Td 2
2 0.3π
0.3π ≤ ω ⇒ tan ≤ Ω ⇒ 1.01905 ≤ Ω
Td 2
2N
⎛ 0.6498 ⎞
2
1 ⎛ 1 ⎞
Hc (Ω) =
2
⇒1+ ⎜ ⎟ =⎜ ⎟ = 1.2589
⎝ Ωc ⎠ ⎝ 0.89125 ⎠
2N
⎛Ω⎞
1+ ⎜ ⎟
⎝ Ωc ⎠
2N
⎛ 1.01905 ⎞
2
⎛ 1 ⎞
1+ ⎜ ⎟ =⎜ ⎟ = 31.6220
⎝ Ωc ⎠ ⎝ 0.17783 ⎠
⎧⎪ ⎡⎛ 1 ⎞
2
⎤ ⎡⎛ 1 ⎞
2
⎤ ⎫⎪
⎨ ⎢⎜ ⎟ − 1⎥ ⎢ ⎜ ⎟ − 1⎥⎬
⎪⎣⎢ ⎝ 0.17783 ⎠ ⎥
⎦ ⎣ ⎢ ⎝ 1.89125 ⎠ ⎥⎦ ⎭⎪ 2.0729
⇒N=⎩ ≅ ≅ 5.30267
⎛ 1.01905 ⎞ 0.3909
2log ⎜ ⎟
⎝ 0.6498 ⎠
⇒ N = 6 ⇒ Ω c = 0.76622

π
⎛ ±j
⎜ 0.76622 ⋅ e , = 0.7401 ± j 0.1983
12
⎜ ±j π
3
⎜
⇒ Poles are 0.76622 ⋅ e 12
= 0.5418 ± j 0.5418
⎜
⎜
5
±j π
⎜ 0.76622 ⋅ e = 0.1983 ± j 0.7401

12
0.20238
⇒ Hc ( s) =
( s + 0.39965 + 0.5871)( s 2 + 1.08365 + 0.5871)
2
1
×
( s + 1.48025 + 0.5871)
2
0.0007378 (1 + z −1 )
6
⇒ H (z) =
(1 − 1.2686 z −1
+ 0.7051z −2 )(1 − 1.0106 z −1 + 0.3583 z −2 )
1
×
(1 − 0.9044 z −1
+ 0.2155 z −2 )


The magnitude of N-th order Butterworth filter can be represented by

ωc ΩcTD ω
H ( e jω ) =
2 1 2
; tan = ⇐ Ω = tan
ω
2N
⎛ ⎞ 2 2 Td 2
⎜ tan 2 ⎟
1+ ⎜ ⎟
ω
⎜ tan c ⎟
⎝ 2 ⎠
Note that this BW filter is maximally flat, H ( e jω c

) = 12 , and periodic with a period
of 2π.
Maximally flat ⇔ The first (2N-1) derivatives of are zero at ω = 0

H ( e jω )
2
It is not easy to directly find the poles and zeros in the z-plane from
since the magnitude-squared function should be factorized into H ( z ) H ( z −1 ) to
determine H ( z ) .

• Design of type-I Chebyshev filters

1
1
H (Ω) =
2
1− ε
⎛Ω⎞
1 + ε 2VN2 ⎜ ⎟
⎝ Ωc ⎠
– VN ( x ) is the N-th order Chebyshev polynomial defined by Ωp Ωs
VN ( x ) = cos ( N cos −1 x )
V0 ( x ) = cos 0 = 1
V1 ( x ) = cos ( cos −1 x ) = x
V2 ( x ) = cos ( 2cos −1 x ) = 2 x 2 − 1
V3 ( x ) = 2 x ( x 2 − 1) − x = 4 x 3 − 3 x
VN +1 ( x ) = 2 xVN ( x ) − VN −1 ( x )
⇒ 0 ≤ x ≤ 1 ⇒ 0 ≤ VN2 ( x ) ≤ 1
x > 1 ⇒ cos −1 x is imaginary ⇒ VN ( x ) ∼ hyperbolic cosine
⇒ VN ( x ) increases monotonically

– Property of VN ( x )
{VN ( x )} forms an orthogonal set over −1 ≤ x ≤ 1
⎧ 0, M ≠N
1 ⎪
VN ( x ) VM ( x ) dx = ⎨π / 2, M = N ≠ 0
1
∫ −1
1 − x2 ⎪π , M = N =0
⎩
For each N, VN (1) = 1
For even N, VN ( −1) = 1; VN ( 0 ) = ( −1)
N /2
VN ( x ) = VN ( − x )
For odd N, VN ( −1) = −1; VN ( 0 ) = 0
VN ( x ) = −VN ( − x )
The zeros of {VN ( x )} lie in −1 ≤ x ≤ 1
The range of {VN ( x )} is between –1 and 1 for −1 ≤ x ≤ 1

1
Ω 1
⇒ For 0 ≤ ≤ 1, 1 ≤ H ( Ω ) ≤
2
1− ε
Ωc 1+ ε 2
Ω
> 1, H ( Ω ) decreases monotonically
Ωc
Ωp Ωs
Ripple factor: ε 2 ⇒ ripple amplitude δ =1- 1
1+ε 2
Since N ( )
V 1 = 1
H e ( jω pT
) = 1 +1ε 2
⇐ H ( e jωT ) =
1
1 + ε 2VN ⎡⎣ k tan (ωT / 2 )⎤⎦
Rule: choose ε , N and Ωc

– Consider the first zeros of 1 + ε VN ( x ) = 0

2 2
Letting cos r ≡ cos(u + jv) = x ⇒ cos −1 x = r

VN ( x ) = cos ( N cos −1 x ) = cos Nr
= cos Nu cosh Nv − j sin Nu sinh Nv
= ± j /ε
1
⇒ cos Nu cosh Nv = 0 & sin Nu sinh Nv = ±
ε
⇒ u = ( 2 M + 1)π / 2 N , M = 0,1,2, ,2 N − 1
1 1
v= sinh −1
N ε
The poles of the Chebyshev filter lie on an ellipse in the s-plane
σ2 Ω2
+ = 1; s = σ + jΩ
sinh 2 v cosh 2 v
σ p = ( sinh v ) cosθ = aΩc cosθ

Ω p = ( cosh v ) sin θ = bΩc sin θ
1⎛ ⎞ 1⎛ N − ⎞
1 1 1 1
−
a = ⎜α N − α N ⎟ , b = ⎜α + α ⎟ , α = ε + 1 + ε
N −1 −2
2⎝ ⎠ 2⎝ ⎠
N =3
• Design of Type II Chebyshev Filters
1
Hc (Ω) =
2
−1
⎡ ⎛ Ω ⎞⎤
1 + ⎢ε 2VN ⎜ c ⎟ ⎥
⎣ ⎝ Ω ⎠⎦
• Elliptic filter
– The error is distributed over the entire frequency band ⇒ Equal ripple in all band
1
Hc (Ω) =
2
⎛Ω⎞ ⇒ min order of the filter

1 + ε 2U N2 ⎜ ⎟
Ω
⎝ c⎠
U N (Ω) : Chebyshev rational function
of degree N

• Example
– Filter specifications
Passband
0.99 ≤ H ( e jω ) ≤ 1.01; ω ≤ 0.4π
Stopband
H ( e jω ) ≤ 0.001; 0.6π ≤ ω ≤ π
This implies that δ1 = 0.01, δ 2 = 0.001, ω p = 0.4π and ωs = 0.6π

Butterworth filter with N=14 Chev-type I filter with N=8

2 ⎛ 1 − z −1 ⎞
s= 5. Filter design techniques
Td ⎜⎝ 1 + z −1 ⎟⎠
Chev-type II filter with N=8 Elliptic filter with N=6

– Poles and zeros of the designed filters

• Design of FIR filters by windowing
– FIR filter design directly approximates the desired frequency response of the discrete-
time system
– When the desired frequency response is

∞
H d ( e jω ) = ∫ π H ( e ) dω
1
∑ h [ n]e hd [ n ] =
− jω n π jω
;
2π
d − d
h =−∞
consider the design of a causal FIR filter that approximates the ideal response
– The truncation of the desired impulse response results in the Gibbs phenomenon
– Rectangular Windowing
⎧h [ n], 0 ≤ n ≤ M
h [ n] = ⎨ d ⇒ h [ n ] = hd [ n ]ω r [ n ]
⎩ 0 , otherwise
where ω r [ n ]is the windowing function given by

⎧1, 0 ≤ n ≤ M
ω r [ n] = ⎨
⎩0, otherwise
⇒ H ( e jω ) =
1
∫ π H ( e )ω ( e ( ) ) dθ
π θ j j ω −θ
2π − d r
Note that H ( e jω ) is a smeared version of H d ( e jω )

– Effect of windowing
c h
Hd e jω
Wr ( e jω )
ω ω
0
c h
ω r e jω
Wr ( e ) = ∑e
M
jω − jω n
peak side lobe
n =0
1 − e − jω ( M +1)
= 2π
1 − e − jω M +1 design paramters
4π
ω ( M + 1) ΔWm =
sin ωM M +1
−j
= 2 e 2
ω
sin
2
• Common windowing functions

– Bartlett (triangular) ⎧ 2n M
⎪ M, 0≤n≤
2
⎪
⎪ 2n M
wB [ n ] = ⎨2 − , ≤n≤M
⎪ M 2
⎪ 0, otherwise
⎪⎩
– Hanning
⎧ ⎛ 2π n ⎞
⎪0.5 ⎜1 − cos ⎟, 0 ≤ n ≤ M
whan [ n ] = ⎨ ⎝ M ⎠
⎪ 0
⎩ , otherwise
– Hamming
⎧ 2π n
⎪0.54 − 0.46cos ,0≤n≤M
wham [ n ] = ⎨ M
– Blackman
⎧ ⎛ 2π n ⎞ 4π n
⎪0.42 − 0.5cos ⎜ ⎟ + 0.08cos , 0≤n≤M
wBL [ n ] = ⎨ ⎝ M ⎠ M
⎪ 0
⎩

Trade off between the main lobe width and the peak of the side lobe
Peak side lobe Main-lobe Peak approx.
width error
Rectangular − 13 4π /( M + 1) − 21
Bartlett − 25 8π / M − 25
Hanning − 31 8π / M − 44
Hamming − 41 8π / M − 53
Blackman − 57 12π / M − 74
Rectangular window
Bartlett window Hamming window
Hanning window Blackman window

– Kaiser window
⎧ ⎡ n −α ⎤ ⎤
2
⎪ I0 ⎢ β 1 − ⎡
⎢ ⎥ ⎥
⎪ ⎢ ⎣ α ⎦ ⎥⎦
w[ n ] = ⎨ ⎣ ,0 ≤ n ≤ M , α =
M
⎪ I0 ( β ) 2
⎪
⎩ 0 ,otherwise
where I 0 ( ⋅ ) is the zeroth-order modified Bessel function of the first kind defined by
1 π jx sinθ 2 ⎛ π⎞
I0 ( x ) =
∫ e dθ ≈ cos ⎜ x − ⎟ ( for large x )
2π −π πx ⎝ 4⎠
As β = 0 → rectangular window case
If the window is tapered more (i.e., β is increased), the main lobe width becomes
wider and the peak of the side lobe becomes lower

Increasing M while holding β constant

results in the main lobe width decreased
but does not affect the peak of the side
lobe.
Given δ is fixed, ω p is defined to be

the highest frequency such that
H ( e jω ) ≥ 1 − δ
and ω s to be the lowest frequency such that

H ( e jω ) ≤ δ
Then β can be empirically determined by

⎧0.1102 ( A − 8.7 ) , A > 50
⎪⎪
β = ⎨0.5842 ( A − 21) + 0.07886 ( A − 21) 21 ≤ A ≤ 50
0.4
⎪ 0 , A < 21
⎪⎩
A−8
M= ;
2.285Δω
where Δω = ω s − ω p , A = −20log δ

– Comparison of commonly used windows

Design example
ω p + ωs
ω p = 0.4π ; ω s = 0.6π ⇒ ω c = = 0.5π
2
Δω = ω s − ω p = 0.2π
δ 1 = 0.01; δ 2 = 0.001 ⇒ δ = 0.001
⇒ A = −20log δ = 60dB
⇒ β = 0.1102 ( A − 8.7 ) = 5.65326
A−8 60 − 8
M= = = 36.219 ⇒ M = 37; Type - II FIR
2.285Δω 2.285 × 0.2π
⇒ h [ n ] = hd [ n ] ⋅ ω [ n ]
⎧ ⎡ 2 2⎤
1
⎪ ⎛ ⎛ n − 18.5 ⎞ ⎞ ⎥
I 0 ⎢5.65326 ⎜1 − ⎜ ⎟ ⎟ ⎥
⎪⎪ ⎢ ⎝ 18.5 ⎠ ⎠
= ⎨ sin 0.5π ( n − 18.5 ) ⎢ ⎝
⋅ ⎣ ⎦⎥ , 0 ≤ n ≤ M
⎪ π ( n − 18.5 ) I 0 ( 5.65326 )
⎪
⎪⎩0, otherwise


• Linear system with generalized linear phase
– For casual system, zero phase is not achievable

Phase distortion exists
Linear phase results in a simple time shift
– System with linear phase
Ideal delay
H id ( e jω ) = e − jωα , ω <π
⇒ H id ( e jω ) = 1
A rg ⎡⎣ H id ( e jω ) ⎤⎦ = −αω
∂
grd ⎣⎡ H id ( e jω ) ⎦⎤ = − αω = α
∂ω
sin π ( n − α )
From the inverse Fourier transform, hid [ n ] = , −∞<n<∞
π (n −α )
For an input x [ n ] ,
sin π ( n − α )
y [ n] = x [ n] ∗
π (n −α )
∞
sin π ( n − k − α )
= ∑
k =−∞
x[k ]
π (n − k −α )
In particular, if α = nd , hid [ n ] = δ [ n − nd ]
⇒ y [ n ] = x [ n ] ∗ hid [ n ] = x [ n − nd ]
Ideal LPF with linear phase

⎧⎪e− jωα , ω < ω c sin ω c ( n − α )
H p (e ) = ⎨
jω
⇒ h p [ n] =
⎪⎩0, ωc < ω ≤ π π (n −α )
If α = nd , the impulse response is symmetric about n = α , i.e.,

h p [ 2α − n ] = h p [ n ]

– Generalized linear phase
H ( e jω ) = A ( e jω ) e
− j (αω − β )
= A ( e jω ) cos ( β − ωα ) + jA ( e jω ) sin ( β − ωα ) ⇒ τ (ω ) = grd ⎡⎣ H ( e jω ) ⎤⎦
H (e jω
∞
) = ∑ h [ n]e − jω n =−
d
dω
{ }
A rg ⎡⎣ H ( e jω ) ⎤⎦
n =−∞
∞ ∞
=α
= ∑ h [ n] cos ωn − j ∑ h [ n]sin ωn
n =−∞ n =−∞
=−∑
sin ( β − ωα ) h [ n ] sin ω n
tan ( β − ωα ) = ⇒ ∑ h [ n ] sin ⎡⎣ω ( n − α ) + β ⎤⎦ = 0, ∀ω
cos ( β − ωα ) ∑ []
h n cos ω n n
This is satisfied when β = 0 or π , 2α is an integer, and h [ 2α − n ] = h [ n ]

Alternatively, if β =π /2 or 3π / 2, ∑ h [ n] cos ⎡⎣ω ( n − α )⎤⎦ = 0
⇒ 2α = M = an integer and h [ 2α − n ] = − h [ n ]
This is a necessary condition for h [ n ] to have a constant group delay

There are other systems having linear phase characteristics without this
symmetry condition

– Causal generalized linear phase systems

∞
If causal, ∑ h [ n]sin ⎡⎣ω ( n − α ) + β ⎤⎦ = 0, ∀ω
n =0
Causal FIR systems of length (M+1) have generalized linear phase, if

h [ n ] = h [ 2α − n ] or h [ n ] = − h [ 2α − n ]
If
⎧h [ M − n ] , 0 ≤ n ≤ M
h [ n] = ⎨
⎩ 0 , otherwise
jω M
then H ( e ) = A (e )e
−
jω jω 2
e
where A ( e ) is a real, even periodic function of ω

e
jω
If
⎧−h [ M − n ] , 0 ≤ n ≤ M
h [ n] = ⎨
⎩ 0 , otherwise
jω M jω M +π
then H ( e ) = jA ( e ) e = A ( e ) e
− −
jω jω 2 jω 2
o o
where A ( e ) is a real, odd periodic function of ω

o
jω

Type I FIR linear phase systems

h [ n] = h [ M − n] 0≤n≤M
where M is an ever integer

M
−1 ωM
⎡M ⎤ −j
H ( e jω ) = ∑ h [ n ] e − jω n = ∑ h [ n ] e − jω n + h ⎢ ⎥ e
M 2 M
n =0 n =0 ⎣2⎦
2
+ ∑ h [ n] e
M
− jω n
n= +1
2
M M
−1 ωM −1
2
⎡M ⎤ −j 2 2
= ∑ h [ n] e + ∑ h [ M − n] e ( )
− jω n − jω M − n
+ h⎢ ⎥e
n =0 ⎣2⎦ n =0
M
−1
( ) ⎡ M ⎤ − jω
2 M
= ∑ h [ n] e − jω n − jω ( M − n )
+e + h⎢ ⎥e 2
n =0 ⎣2⎦
M M
−1 ωM
⎡ M⎤ ⎡M ⎤
M
2 − jω 2 −j
= 2 ∑ h [ n ] cos ⎢ n − ⎥ ω +h ⎢ ⎥ e 2
= ∑ a [ k ] cos ω k ⋅ e 2
n =0 ⎣ 2⎦ ⎣2⎦ k =0
where
⎡M ⎤ ⎡M ⎤ M
a [ 0] = h ⎢ ⎥ , a [ k ] = 2h ⎢ − k ⎥ , k = 1, 2,.... .
⎣2⎦ ⎣2 ⎦ 2
ωM
⇒ H (e ) = A (e )e
−j
jω jω 2
e
where M
Ae ( e jω ) = ∑ a [ k ] cos ω k
2
k =0

Example
⎧1, 0 ≤ n ≤ 4
h [ n] = ⎨
⎩0, otherwise
5ω
sin
⇒ H ( e jω ) = ∑ e − jω n
4
= 2 e − j 2ω
n =0
ω
sin
2

Type II linear phase systems

h [ n] = h [ M − n], 0 ≤ n ≤ M
where M is an odd integer
⇒ H ( e jω ) = ∑ h [ n ] e − jω n
M
n =0
M +1
2
⎡M +1 ⎤ ⎡ ⎛ 1 ⎞ ⎤ − j ω2M
= ∑ 2h ⎢ − k ⎥ cos ⎢ω ⎜ k − ⎟ ⎥ ⋅ e
k =1 ⎣ 2 ⎦ ⎣ ⎝ 2 ⎠⎦
Example
⎧1, 0 ≤ n ≤ 5
h [ n] = ⎨
⎩0, otherwise
sin 3ω 5
⇒ H (e )=
−j ω
jω 2
e
ω
sin
2
Type III FIR linear- phase systems

h [ n ] = −h [ M − n ] , 0 ≤ n ≤ M
where M is an even integer
M
ωM
⎡M ⎤
⇒ H ( e ) = j 2∑ h ⎢ − k ⎥ ⋅ sin ω k ⋅ e
2 −j
jω 2
k =1 ⎣ 2 ⎦
Example
h [ n ] = δ [ n ] − δ [ n − 2]
H ( e jω ) = 1 − e − j ω = j 2sin ω ⋅ e − jω
2

Type IV FIR linear- phase systems

h [ n] = −h [ M − n], 0 ≤ n ≤ M
where is an odd integer

M +1
⎡ M +1 ⎤ ⎡ ⎛ 1 ⎞ ⎤ − j ω2M
⇒ H (e ) = j2 ∑ h ⎢
2
jω
− k ⎥ ⋅ sin ⎢ω ⎜ k − ⎟ ⎥ ⋅ e
k =1 ⎣ 2 ⎦ ⎣ ⎝ 2 ⎠⎦
Example
h [ n ] = δ [ n ] − δ [ n − 1]
⇒ H ( e jω ) = 1 + ( e jω ) = 1 − e− jω
ω
ω −j
= j 2sin ⋅e 2
2

– Location of zeros for FIR linear-phase systems

M
H ( z ) = ∑ h [ n] z − n
{h [ n]}
n =0
If are symmetric,
= z − M H ( z −1 )
M o
H ( z ) = ∑ h [ M − n] z − n = ∑ h[k ] z k −M
n =0 k =M
If z0 = re jθ is a zero of H ( z ) , then
H ( z 0 ) = z0
−M
H z0( )=0
−1
That is, z0−1 = r −1e − jθ is also a zero of H ( z ).

When h [ n ] is real and z0 is a zero of H ( z ) , z0∗ = re− jθ will also be a zero of H ( z ) ,
and so will (z )
∗ −1
0 .
Therefore, when h [ n ] is real, there will be four conjugate reciprocal zeros of the
form
(1 − re jθ
z −1 )(1 − re − jθ z −1 )(1 − r −1e jθ z −1 )(1 − r −1 z − jθ z −1 )
unless r = 1.

If {h [ n ]} are anti-symmetric, we have
H ( z ) = −∑ h [ M − n ] z − n = − ∑ h [ k ] z k − M = − z − M H ( z −1 )
M o
n =0 k =M
In particular, when z = 1, H (1) = − H (1) and if z = −1, H ( −1) = ( −1) H ( −1).

M +1
Thus, H ( z ) must have a zero at z = 1.
It also must have a zero at z = −1 ⇔ ω = π when M is even.
type - II type - III type - IV

• Relation of FIR linear-phase systems to minimum-phase systems
– Any FIR linear-phase system can be represented by

H ( z ) = H min ( z ) H max ( z ) H uc ( z )
where H max ( z ) = H min ( z −1 ) z − M i , M i = no. of zeros of H min ( z )
and H uc ( z ) has only M 0 zeros on the unit circle.
⇒ The order of M is 2 M i + M 0

Example: Decomposition of a linear-phase system

H min ( z ) = 1.252 (1 − 0.9e j 0.6π z −1 )(1 − 0.9e − j 0.6π z −1 )(1 − 0.8e j 0.8π z −1 )(1 − 0.8e − j 0.8π z −1 )
⇒ H max ( z ) = ( 0.9 ) (1 − 1.1111e j 0.6π z −1 )(1 − 1.1111e − j 0.6π z −1 )(1 − 1.25e j 0.8π z −1 )(1 − 1.25e − j 0.8π z −1 )
2
The overall system H ( z ) = H min ( z ) H max ( z ) has linear phase

– Causal and linear phase FIR filter

⎧ w[ M − n ] 0 ≤ n ≤ M M
w[ n ] = ⎨ ⇒ W ( e jω ) = We ( e jω ) e 2
− jω
⎩ 0
even function of ω
The impulse response of a desired filter is also symmetric; hd [ n ] = hd [ M − n ]
ωM
Hd (e ) = H (e )e
−j
jω jω 2
e
The overall response

H ( e jω ) =
1
2π
∫ H d ( e jθ )W e j (ω −θ ) dθ( )
θM (ω −θ ) M
1
H e ( e ) e We e ( )
−j j (ω −θ ) −j
= ∫
jθ 2
e 2 dθ
2π
ωM
1
H e ( e )We e ( )
j (ω −θ ) −j
= ∫
jθ
e 2 dθ
2π
ωM
= Ae ( e )e
−j
jω 2
Ae ( e jω ) = ∫ π H ( e )W ( e ) dθ
1 π θ (ω θ )
j j −
where
2π − e e
The resulting response has generalized linear phase and its magnitude is also an
even real function.

Example Linear-phase lowpass filter
⎧ − j ω2M
⎪e , ω < ωc ωM
H d ( e jω ) = ⎨
1
⇒ hd [ n ] =
ωc −j
∫ω e 2
e jω n dω
⎪⎩ 0 , ω c < ω ≤ π 2π − c
⎡ ⎛ M ⎞⎤
sin ⎢ω c ⎜ n − ⎟⎥
= ⎣ ⎝ 2 ⎠⎦
π ⎛⎜ n − ⎞⎟
M
⎝ 2 ⎠
h [ n ] = hd [ n ]ω [ n ] h [ n ] = hd [ n ] w[ n ]

– Filter design by Kaiser method

High-pass filter
⎧0 0 ≤ ω ≤ ωc
⎪
H hp ( e ) = ⎨ − j ω M
jω
⎪⎩e 2 ωc ≤ ω ≤ π
ωM
H hp ( e )=e − H ep ( e jω )
−j
jω 2
⎛ M⎞ ⎛ M⎞
sin π ⎜ n − ⎟ sin ω c ⎜ n − ⎟
⇒ hhp [ n ] = ⎝ 2 ⎠
− ⎝ 2 ⎠
π ⎛⎜ n − ⎞⎟ π ⎛⎜ n − ⎞⎟
M M
⎝ 2 ⎠ ⎝ 2 ⎠
• Design example:
If ω s = 0.35π , ω p = 0.5π and δ 1 = δ 2 = δ = 0.021

⇒ A = 33.56
⇒ β = 0.5842 ( A − 21) + 0.07886 ( A − 21)
0.4
= 2.5974
33.56 − 8
M= = 23.73 ⇒ M = 24
2.285 × 0.15π

High-pass filter example
A = 33.56
β = 2.5974
M = 24 ⇒ type I FIR
Actual error=0.0213>0.021

High-pass Kaiser high-pass filter example with M=25 (type-II)

• Use of a type-II Kaiser HPF
may not be appropriate

– FIR filter with generalized response
G1 ⎛ M⎞
N =4 sin ω ⎜ n − ⎟
⎝ 2 ⎠
k
⇒ hd [ n ] = ∑ ( Gk − Gk +1 )
N
G2
π ⎛⎜ n − ⎞⎟
k =1 M
G4
⎝ 2 ⎠
G3 where GN +1 = 0
0 ω1 ω2 ω3 π ⇒ h [ n ] = hd [ n ]ω [ n ]
– Differentiator
ωM
Hd (e ) = jω e
−j
jω 2
−π <ω <π
⎡ ⎛ M ⎞⎤ ⎡ ⎛ M ⎞⎤
cos ⎢π ⎜ n − ⎟⎥ sin ⎢π ⎜ n − ⎟ ⎥
⇒ hd [ n ] = ⎣ ⎝ 2 ⎠⎦
− ⎣ ⎝ 2 ⎠⎦
2
n−
M ⎛ M⎞
π ⎜n − ⎟
2 ⎝ 2 ⎠
In the case of using a symmetric window of size M + 1, h [ n ] = hd [ n ] ⋅ ω [ n ] results in

h [ M − n] = −h [ n]
Kaiser formulae were developed for frequency response with simple magnitude
discontinuities, but still works on differentiator

Kaiser window design of a differentiator

The amplitude error: E (ω ) = ω − Ao ( e ) ; 0 ≤ ω ≤ π
jω
where Ao ( e jω ) is the amplitude of the filter H ( z ) and has zeros at z = ±1
By not imposing constraint H(z) to have a zero at z=-1, it is possible to design a

filter having better approximation to the desired response, while using a less
number of filter taps.

Example: Impulse response of Kaiser-windowed differentiatore

• Optimum approximation of FIR filters
– Design a filter that best approximates the desired response for a given M
– Rectangular window provides the best-mean squared approximation:
⎧h [ n ] , 0 ≤ n ≤ M
h [ n] = ⎨ d
minimizes ⎩ 0 , otherwise
∫ π H (e ) − H (e )
1 π 2
ε =2 jω jω
dω
2π − d
∞
= ∑ e [ n]
n =−∞
2
⎧⎪hd [ n ] − h [ n ]; 0 ≤ n ≤ M
e [ n] ⎨
⎪⎩hd [ n ]; otherwisw
Do not permit individual control errors in

different bands

• Consider the design of a type-I FIR filter with zero phase, i.e., he [ n ] = he [ − n ]
– Corresponding frequency response
Ae ( e ) = ∑ h [ n]e
L L
jω
e
− jω n
= he [ 0] + ∑ 2he [ n ] cos (ω n )
n =− L n =0
is a real, even and periodic function of ω
A causal system can be obtained by delaying it by L samples

h [ n ] he [ n − L ] = he [ 2 L − n ]
– From a polynomial approximation
cos ω n = Tn ( cos ω ) ; Tn ( x ) = cos ( n cos −1 x )
A ( e jω ) can be represented as an L-th degree trigonometric polynomial
A( e ) = ∑a
L L
jω
k cos ω = P ( x ) x=cos ω ; P ( x ) = ∑ ak x k
k
k =0 k =0
– Define an error function by
E (ω ) = W (ω ) H d ( e jω ) − H ( e jω )
2
where W (ω ) is a weighting function

– Example
⎧1, 0 ≤ ω ≤ ω p
Hd (e ) = ⎨
jω
⎩0, ωs ≤ ω ≤ π
⎧1
⎪ ; 0 ≤ ω ≤ ωp δ1
W (ω ) = ⎨ K ; K=
⎪⎩1, δ2
ωs ≤ ω ≤ π
Maximum weighted absolute approximation error
is δ = δ2

– Minimax or Chevyshev criterion

min
{he [ n];0≤n≤ L} ω∈F ( max E(ω) )
F = {ω : 0 ≤ ω ≤ ω p , ωs ≤ ω ≤ π }
Find a set of impulse response that minimizes δ
– Alternative theorem:
Let denote a closed subset comprising the disjoint union of closed subsets of real
axis x. Then m
P ( x ) = ∑ ak x k
k =0
is an m-th order polynomial
For desired function DP ( x ) and positive function WP ( x) which are continuous

on FP , and weighted error defined by
EP ( x ) = WP ( x ) ⎡⎣ DP ( x ) − P ( x ) ⎤⎦
a necessary and sufficient condition that be a unique r-th order polynomial
minimizing E = max EP ( x ) is that EP ( x ) has at least r+2 alternations
x∈FP
At least r+2 values xi in FP such that x1 < x2 < < xr + 2 and

EP ( xi ) = − EP ( xi +1 ) = ± E , i = 1,2, ,r + 2
– Example: Consider 5-th order polynomials Pj ( x ) that

approximate unity for −1 ≤ x ≤ −0.1 and zero for
0.1 ≤ x ≤ 1
Assume that WP ( x) = 1 for these two regions
The optimum 5-th order polynomial has at least 7
alternations of the error in the region in FP
P1 ( x ) has 3 alternations in −1 ≤ x ≤ −0.1 and

2 alternations in 0.1 ≤ x ≤ 1
P2 ( x ) also has 5 alternations; 3 in −1 ≤ x ≤ −0.1

and 2 in 0.1 ≤ x ≤ 1
0.1 ≤ x ≤ 1 ⇒ optimum polynomial

• Optimum type-I lowpass filters
– Consider polynomial P ( x ) defined by

L=7
L
P ( cos ω ) = ∑ ak ( cos ω )
k
k =0
– Assume that
⎧1; cos ω p ≤ cos ω ≤ 1
DP ( cos ω ) = ⎨
⎩0; − 1 ≤ cos ω ≤ cos ωs
⎧1/ K ; cos ω p ≤ cos ω ≤ 1
WP ( cos ω ) = ⎨
⎩0; − 1 ≤ cos ω ≤ cos ωs
EP ( cos ω ) = WP ( cos ω ) ⎡⎣ DP ( cos ω ) − P ( cos ω ) ⎤⎦

FP = {ω : 0 ≤ ω ≤ ω p , ωs ≤ ω ≤ π }
– Equi-ripple approximation
A set of coefficients {ak } is designed to make EP ( cos ω ) have at least ( L + 2 )
alternations on FP

• Optimum type-I lowpass filters (cont.)
– Redraw A e
jω
( ) in terms of x = cos ω
L=7

– Max. possible number of alternations of the error

is ( L + 3) since an L-th degree polynomial can
have at most ( L − 1) points with zero slope

∂ L
P ( cos ω ) = − sin ω ∑ kak ( cos ω )
k −1
∂ω k =0
L −1
= − sin ω ∑ ( k + 1) ak +1 ( cos ω )
k
k =0
Has zeros at ω = 0, ω = π and roots of
( L − 1) -th order polynomial

– Alternations always occur at ω = ω p and ω = ωs
– The filter will be equi-ripple except possibly at
ω = 0 or ω = π

• Parks-McClellan algorithm
– The alternation theorem indicates that the optimum filter Ae e

jω
( ) will satisfies
W (ωi ) ⎡⎣ H d ( e jωi ) − Ae ( e jωi ) ⎤⎦ = ( −1) δ ; i = 1,2,

i +1
,L + 2
⇒ H d ( e jωi ) = Ae ( e jωi ) + ( −1) δ / W (ωi )

i +1
Ae ( e ) = ∑ h [ n]e
L L L L
jω
e
− jω n
= he [ 0] + ∑ 2he [ n ] cos (ω n ) = ∑ ak cos ω = ∑ ak x k
k
n =− L n =0 k =0 k =0
jω
– Hd e i ( ) can be rewritten as
⎡ 2 L 1 ⎤
⎢1 x x x
W (ω1 ) ⎥
1 1 1
⎡ H d ( e jω1 ) ⎤ ⎢ ⎥ ⎡a ⎤
⎢ ⎥ ⎢ −1 ⎥ 0
⎢ H d ( e ) ⎥ ⎢1 x2 x2
jω2 2 L ⎢ ⎥
x2 ⎥
W ( ω2 ) ⎢ ⎥ a
⎢ ⎥=⎢
1
⎥⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ jωL + 2 ⎥
⎢⎣ H d ( e )⎥⎦ ⎢⎢ 2 L ( −1) ⎥
L+2
⎥ ⎣δ ⎦
⎢ L+2 L+2
1 x x x L+2
W (ωL+ 2 ) ⎥⎦
⎣
– Solution for ak and δ can be found by means of polynomial interpolation

• Parks-McClellan algorithm (cont.)
– For a given set of extremal frequencies, the P-M algorithm determines

L+2
∑b H (e ω )
k d
j k
L+2
1
δ= bk = ∏
k =1
;
bk ( −1)
k +1
L+2
xk − xi
∑k =1 W ( ωk )
i =1
i≠k
– For given ak and δ , the error function has magnitude ±δ at L + 2 frequencies ωi
– Ae ( e jω ) has values 1 ± K δ for 0 ≤ ω ≤ ω p and ±δ for ωs ≤ ω ≤ π

– An L-th order trigonometric polynimial Ae ( e ) can be obtained by interpolating
jω
E (ωi )
L +1
d k Ck
∑ x − xk
Ae ( e ) = P ( cos ω ) = Lk =+11
jω
dk
∑k =1 x − xk
( −1) δ ,
k +1 L +1
x = cos ω , Ck = H d ( e ) W (ω ) 1
jωk
− dk = ∏ = bk ( xk − xL+ 2 )
k i =1 xk − xi
i≠k

ak and δ
Ae ( e jω ) 1 ± Kδ
Example: L = 7

• Characteristics of optimum FIR filters
– Size of the filter tap

−10log (δ1δ 2 ) − 13
M=
2.324 (ωs − ω p )

• Filter design examples by Parks-McClellan algorithm

δ1 = 102 ; δ 2 = 10−3
ω p = 0.4π ; ωs = 0.6π
M = 26 M = 27

• Filter design example (cont.)
– Compensation of zero-order holder
⎧ ω/2
⎪ ; 0 ≤ ω ≤ ωp
H d ( e jω ) = ⎨ sin (ω / 2 )
⎪ ωs ≤ ω ≤ π
⎩0;
δ1 = 102 ; δ 2 = 10 −3 ; ω p = 0.4π ; ωs = 0.6π
M = 28 ⇒ L = 14

• Optimum bandpass filter design
⎧0; 0 ≤ ω ≤ 0.3π
⎪
H d ( e jω ) = ⎨1; 0.35π ≤ ω ≤ 0.3π
⎪0; 0.7π ≤ ω ≤ π
⎩
⎧1; 0 ≤ ω ≤ 0.3π
⎪
W ( e jω ) = ⎨1; 0.35π ≤ ω ≤ 0.3π
⎪0.2; 0.7π ≤ ω ≤ π
⎩
M = 74 ⇒ L = 37
– Local extrema may occur in the transition

regions
– The approximation need not be equi-ripple

in the approximation regions

• Optimum approximation of FIR filters
– Design a filter that best approximates the desired response for a given M
– Rectangular window provides the best-mean squared approximation:
⎧h [ n ] , 0 ≤ n ≤ M
h [ n] = ⎨ d
minimizes ⎩ 0 , otherwise
∫ π H (e ) − H (e )
1 π 2
ε =2 jω jω
dω
2π − d
∞
= ∑ e [ n]
n =−∞
2
⎧⎪hd [ n ] − h [ n ]; 0 ≤ n ≤ M
e [ n] ⎨
⎪⎩hd [ n ]; otherwisw
Do not permit individual control errors in

different bands

• Consider the design of a type-I FIR filter with zero phase, i.e., he [ n ] = he [ − n ]
– Corresponding frequency response
Ae ( e ) = ∑ h [ n]e
L L
jω
e
− jω n
= he [ 0] + ∑ 2he [ n ] cos (ω n )
n =− L n =0
is a real, even and periodic function of ω
A causal system can be obtained by delaying it by L samples

h [ n ] he [ n − L ] = he [ 2 L − n ]
– From a polynomial approximation
cos ω n = Tn ( cos ω ) ; Tn ( x ) = cos ( n cos −1 x )
A ( e jω ) can be represented as an L-th degree trigonometric polynomial
A( e ) = ∑a
L L
jω
k cos ω = P ( x ) x=cos ω ; P ( x ) = ∑ ak x k
k
k =0 k =0
– Define an error function by
E (ω ) = W (ω ) H d ( e jω ) − H ( e jω )
2
where W (ω ) is a weighting function

– Example
⎧1, 0 ≤ ω ≤ ω p
Hd (e ) = ⎨
jω
⎩0, ωs ≤ ω ≤ π
⎧1
⎪ ; 0 ≤ ω ≤ ωp δ1
W (ω ) = ⎨ K ; K=
⎪⎩1, δ2
ωs ≤ ω ≤ π
Maximum weighted absolute approximation error
is δ = δ2

– Minimax or Chevyshev criterion

min
{he [ n];0≤n≤ L} ω∈F ( max E(ω) )
F = {ω : 0 ≤ ω ≤ ω p , ωs ≤ ω ≤ π }
Find a set of impulse response that minimizes δ
– Alternative theorem:
Let denote a closed subset comprising the disjoint union of closed subsets of real
axis x. Then m
P ( x ) = ∑ ak x k
k =0
is an m-th order polynomial
For desired function DP ( x ) and positive function WP ( x) which are continuous

on FP , and weighted error defined by
EP ( x ) = WP ( x ) ⎡⎣ DP ( x ) − P ( x ) ⎤⎦
a necessary and sufficient condition that be a unique r-th order polynomial
minimizing E = max EP ( x ) is that EP ( x ) has at least r+2 alternations
x∈FP
At least r+2 values xi in FP such that x1 < x2 < " < xr + 2 and
EP ( xi ) = − EP ( xi +1 ) = ± E , i = 1,2,", r + 2
– Example: Consider 5-th order polynomials Pj ( x ) that

approximate unity for −1 ≤ x ≤ −0.1and zero for
0.1 ≤ x ≤ 1
Assume that WP ( x) = 1 for these two regions
The optimum 5-th order polynomial has at least 7
alternations of the error in the region in FP

2 alternations in 0.1 ≤ x ≤ 1
P2 ( x ) also has 5 alternations; 3 in −1 ≤ x ≤ −0.1

and 2 in 0.1 ≤ x ≤ 1
0.1 ≤ x ≤ 1 ⇒ optimum polynomial

• Optimum type-I lowpass filters
– Consider polynomial P ( x ) defined by

L=7
L
P ( cos ω ) = ∑ ak ( cos ω )
k
k =0
– Assume that
⎧1; cos ω p ≤ cos ω ≤ 1
DP ( cos ω ) = ⎨
⎩0; − 1 ≤ cos ω ≤ cos ωs
⎧1/ K ; cos ω p ≤ cos ω ≤ 1
WP ( cos ω ) = ⎨
⎩0; − 1 ≤ cos ω ≤ cos ωs
EP ( cos ω ) = WP ( cos ω ) ⎡⎣ DP ( cos ω ) − P ( cos ω ) ⎤⎦

FP = {ω : 0 ≤ ω ≤ ω p , ωs ≤ ω ≤ π }
– Equi-ripple approximation
A set of coefficients {ak } is designed to make EP ( cos ω ) have at least ( L + 2 )
alternations on FP

– Redraw A e
jω
( ) in terms of x = cos ω
L=7

– Max. possible number of alternations of the error

is ( L + 3) since an L-th degree polynomial can
have at most ( L − 1) points with zero slope

∂ L
P ( cos ω ) = − sin ω ∑ kak ( cos ω )
k −1
∂ω k =0
L −1
= − sin ω ∑ ( k + 1) ak +1 ( cos ω )
k
k =0
Has zeros at ω = 0, ω = π and roots of
( L − 1) -th order polynomial

– Alternations always occur at ω = ω p and ω = ωs
– The filter will be equi-ripple except possibly at
ω = 0 or ω = π

• Parks-McClellan algorithm
– The alternation theorem indicates that the optimum filter Ae e

jω
( ) will satisfies
W (ωi ) ⎡⎣ H d ( e jωi ) − Ae ( e jωi ) ⎤⎦ = ( −1) δ ; i = 1,2,", L + 2

i +1
⇒ H d ( e jωi ) = Ae ( e jωi ) + ( −1) δ / W (ωi )

i +1
Ae ( e ) = ∑ h [ n]e
L L L L
jω
e
− jω n
= he [ 0] + ∑ 2he [ n ] cos (ω n ) = ∑ ak cos ω = ∑ ak x k
k
n =− L n =0 k =0 k =0
jω
– Hd e i ( ) can be rewritten as
⎡ 1 ⎤
⎢1 x x 2
" x L
W (ω1 ) ⎥
1 1 1
⎡ H d ( e jω1 ) ⎤ ⎢ ⎥ ⎡a ⎤
⎢ ⎥ ⎢ −1 ⎥ 0
⎢ H d ( e ) ⎥ ⎢1 x2 x2 " x2
jω2 2 L ⎢ ⎥
⎥
W ( ω2 ) ⎢ ⎥ a
⎢ ⎥=⎢
1
⎥⎢ ⎥
⎢ # ⎥ ⎢# # # " # # ⎥⎢ # ⎥
⎢ jωL + 2 ⎥
⎢⎣ H d ( e )⎥⎦ ⎢⎢ ( −1) ⎥
L+2
⎥ ⎣δ ⎦
⎢ L+2 L+2
2
" L
W (ωL+ 2 ) ⎥⎦
1 x x x L+2
⎣
– Solution for ak and δ can be found by means of polynomial interpolation

– For a given set of extremal frequencies, the P-M algorithm determines

L+2
∑b H (e ω )
k d
j k
L+2
1
δ= bk = ∏
k =1
;
bk ( −1)
k +1
L+2
xk − xi
∑k =1 W ( ωk )
i =1
i≠k
– For given ak and δ , the error function has magnitude ±δ at L + 2 frequencies ωi
– Ae ( e jω ) has values 1 ± K δ for 0 ≤ ω ≤ ω p and ±δ for ωs ≤ ω ≤ π

– An L-th order trigonometric polynimial Ae ( e ) can be obtained by interpolating
jω
E (ωi )
L +1
d k Ck
∑ x − xk
Ae ( e ) = P ( cos ω ) = Lk =+11
jω
dk
∑k =1 x − xk
( −1) δ ,
k +1 L +1
x = cos ω , Ck = H d ( e ) W (ω ) 1
jωk
− dk = ∏ = bk ( xk − xL+ 2 )
k i =1 xk − xi
i≠k

ak and δ
Ae ( e jω ) 1 ± Kδ
Example: L = 7

• Characteristics of optimum FIR filters
– Size of the filter tap

−10log (δ1δ 2 ) − 13
M=
2.324 (ωs − ω p )

• Filter design examples by Parks-McClellan algorithm

δ1 = 102 ; δ 2 = 10−3
ω p = 0.4π ; ωs = 0.6π
M = 26 M = 27

• Filter design example (cont.)
– Compensation of zero-order holder
⎧ ω/2
⎪ ; 0 ≤ ω ≤ ωp
H d ( e jω ) = ⎨ sin (ω / 2 )
⎪ ωs ≤ ω ≤ π
⎩0;
δ1 = 102 ; δ 2 = 10 −3 ; ω p = 0.4π ; ωs = 0.6π
M = 28 ⇒ L = 14

• Optimum bandpass filter design
⎧0; 0 ≤ ω ≤ 0.3π
⎪
H d ( e jω ) = ⎨1; 0.35π ≤ ω ≤ 0.3π
⎪0; 0.7π ≤ ω ≤ π
⎩
⎧1; 0 ≤ ω ≤ 0.3π
⎪
W ( e jω ) = ⎨1; 0.35π ≤ ω ≤ 0.3π
⎪0.2; 0.7π ≤ ω ≤ π
⎩
M = 74 ⇒ L = 37
– Local extrema may occur in the transition

regions, which may not be acceptable
– The approximation need not be equi-ripple

in the approximation regions

Seoul National University 6. Discrete Fourier transform
Discrete Fourier transform

• Discrete Fourier series representation of periodic sequences
– Let x [ n ] be a periodic sequence with period N such that x [ n ] = x [ n + mN ] for any

integer m
Example A periodic complex exponential

2π kn
ek [ n ] = e = ek [ n + mN ]; k is an interger
j
N
Note that ek [ n ] = ek +AN [ n ] for any integer A

– A set of N periodic complex exponentials {e [ n]} ,
k k = 0,1,2,...N − 1, defines all the
frequency components
2π kn
1 N −1 i
⇒ x [ n ] = ∑ X [ k ]e N
j
N k =0
N −1 2π mn
−j
X [ k ] = ∑ x [ n ] e N
n =0
– Fourier series representation

2π kn
1 N −1 i
x [ n ] = ∑ X [ k ]e N
j
N k =0

N −1 2π rn N −1 2π kn 2π rn
−j 1 N −1 −j
∑ x [ n]e = ∑ ∑ X [ k ]e
j
N N
e N
n =0 n =0 N k =0
N −1 2π ( k − r ) n
1 N −1 j
= ∑ X [k ] ∑ e
N
k =0 N k =0
2π ( k − r ) n
1 N −1 j ⎧1, k − r = AN
Since ∑e N
=⎨
N n =0 ⎩0, otherwise
N −1 2π rn
∑ x [ n] e = X [ r ]
j
N
n =0
The Fourier series coefficient can be obtained by

N −1 2π kn
−j
X [ k ] = ∑ x [ n ] e N
n =0
N −1 2π ( k + N ) n N −1 2π kn
−j −j
X [ k + N ] = ∑ x [ n ] e N
= ∑ x [ n ] e N − j 2π n
e = X [ k ]
n −0 n −0
The Fourier series coefficients is a periodic sequence

– Discrete Fourier series (DFS)
Define by x [ n ] ←⎯⎯
DFS
→ X [ k ]
N −1 2π
−j
Analysis equation : X [ k ] = ∑ x [ n ] wNkn ; wN = e N
n =0
1 N −1
Synthesis equation : x [ n ] = ∑ X [ k ]wN− kn
N k =0
Example Periodic pulse train

∞
⎧1, n = mM
x [ n ] = ∑ δ [ n − mN ] = ⎨
r =−∞ ⎩0, otherwise
Since x [ n ] = δ [ n ] , 0 ≤ n ≤ N − 1
N −1
X [ k ] = ∑ δ [ n ] wN− kn = 1
n =0
2π
1 1 N −1 − kn 1 N −1 j N kn
⇒ x [ n ] = ∑ X [ k ] wN = ∑ wN = ∑ e
− kn
N N k =0 N k =0
Example Duality in DFS
Let the DFS coefficient be a periodic impulse train
∞
Y [ k ] = N ∑ δ [ k − rN ]
r =−∞
N −1
⇒ y [ n ] = ∑ N δ [ k ] wN− kn =wN−0 = 1, ∀n
k =0
Y [ k ] = Nx [ n ] ↔ y [ n ] = X [ n ]

Example Periodic rectangular pulse train
πk
4 4 −j
2π
1−W 5k sin 4π k
X [ k ] = ∑W nk
= ∑e 10
kn
= =
10 2 e − j 10
n −0
10
n =0 1−W k
πk
sin
10
10
Any periodic sequence can be represented as a sum of complex exponential

sequences
• Properties of DFS
– Linearity
ax1 [ n ] + bx2 [ n ] ←⎯⎯
DFS
→ aX [ k ] + bX [ k ]
– Shift of a sequence
N −1 2π N −1 2π 2π 2π
−j −j ( n−m ) − j −j
∑ x [ n − m] e = ∑ x [ n − m ] e = X [ k ] e = X [ k ]WNkm
kn km
N N N N
e
n =0 n=0
⇒ x [ n − m ] ←⎯⎯DFS
→ X [ k ]WNkm
– Duality
x [ n ]W − nA ←⎯⎯
N
DFS
→ X [ k − A ]
1 N −1 N −1
x [ n ] = ∑ X [ k ]wN ⇒ Nx [ − n ] = ∑ X [ k ]WNkn
− kn
N k =0 k =0
N −1
⇒ Nx [ − k ] = ∑ x [ n ]WNkn
k =0
If x [ n ] ←⎯⎯
DFS
→ X [ k ] , X [ n ] ←⎯⎯
DFS
→ Nx [ − k ]
– Symmetry properties
x * [ n ] ←⎯⎯
DFS
→ X * [ − k ]
x * [ −n ] ←⎯⎯
DFS
→ X * [ k ]

⇒ If x [ n ] is real, X [ k ] = X * [ − k ]
X [ k ] = X [ − k ]
– Periodic convolution
N −1 N −1
x3 [ n ] = ∑ x1 [ m ] x2 [ n − m ] = ∑ x2 [ m ] x1 [ n − m ]
m =0 m=0
⎛
N −1 N −1
⎞
X 3 [ k ] = ∑ ⎜ ∑ x1 [ m ] x2 [ n − m ] ⎟WNkn

n =0 ⎝ m =0 ⎠
N −1 N −1
= ∑ x1 [ m ] ∑ x2 [ n − m ]WNkn
m =0 n =0
N −1
= X 2 [ k ] ∑ x1 [ m ]WNkm = X 1 [ k ] X 2 [ k ]
m =0
N −1
⇒ ∑ x1 [ m ] x2 [ n − m ] ←⎯⎯
DFS
→ X 1 [ k ] X 2 [ k ]
m =0
Duality
1 N −1
x1 [ n ] x2 [ n ] ←⎯⎯
→ ∑ X 1 [ m ] X 2 [ k − m ]
DFS
N m =0
Difference between aperiodic convolution and periodic convolution
• The sum is over an interval 0 ≤ m ≤ N − 1
• x 2 [ n − m ] = x 2 ⎡⎣ − ( m − n ) ⎤⎦

Example Periodic convolution

– Properties of DFS

– Properties of DFS (cont.)

• Fourier transform (F.T.) of periodic signals
– A sequence should be absolutely summable to guarantee the uniform convergence of

its Fourier transform
– Periodic signals are not absolutely summable

– The F.T. of a periodic signal can be interpreted as an impulse train in the frequency
domain with impulse values proportional to the DFS coefficients for the sequence
– If x [ n ] is periodic with period N, the Fourier transform of x n [ ] is defined by
2π ∞
⎛ 2π k ⎞
x [ n ] ⇔ X ( e jω ) = ∑
X [ k ] δ ⎜ ω − ⎟
N k =−∞ ⎝ N ⎠
Although the F.T. of a periodic sequence does not converge in the normal sense,
the introduction of impulses enables to use the F. T.
Note: X e
jω
( ) has a necessary periodicity with period 2π
2π ⎛ 2π k ⎞ jωn
X ( e jω )e jωn dω =
1 2π −ε 1 2π −ε
∫ ∫ ∑ X [ k ]δ ⎜ ω − ⎟ e dω ; 0 < ε < 2π / N
2π 0−ε 2π 0−ε N k ⎝ N ⎠
1 ∞ 2π −ε ⎛ 2π k ⎞ jωn
= ∑ X [k ] ∫ δ ⎜ω − ⎟e dω
N k =−∞ 0 −ε
⎝ N ⎠
2π
1 ∞
= ∑ X [ k ] e N = x [ n ]
j kn
N k =−∞
Example: Fourier transform of a periodic impulse train

∞
p [ n ] = ∑ δ [ n − mN ] ⇒ P [ k ] = 1, ∀ k
m = −∞
2π ∞
⎡ 2π k ⎤
⇒ P e jω =( ) ∑ δ
N k = −∞ ⎢⎣
ω −
N ⎥⎦
– For a finite-length sequence x [ n ] , ( i.e ., x [ n ] = 0 for n < 0 and n ≥ N )
x [ n ] = x [ n ] ∗ p [ n ]
∞ ∞ ∞
= ∑ ∑ δ [A − mN ]x [ n − A] = ∑ x [ n − mN ]
m =−∞ A =−∞ m =−∞
X ( e jω ) = X ( e jω ) P ( e jω )
2π ∞
⎛ 2π k ⎞
= X (e ) jω
∑ δ ⎜⎝ ω − ⎟
N k =−∞ N ⎠
2π ∞
⎛ j 2Nπ k ⎞ ⎛ 2π k ⎞
= ∑ ⎜ ⎟ ⎜⎝
X e δ ω −
N ⎠
⎟
N k =−∞ ⎝ ⎠
⎛ j 2Nπ k ⎞ N −1 2π
⇒ X [ k ] = X ( e jω )
−j
⎟ ∑ [ ]
k
2π = X ⎜e =
x n e N
ω=
N
k
⎝ ⎠ n =0
Thus, X [ k ] is obtained by sampling the Fourier transform of a finite sequence x [ n ]
Example: Relationship between the DFS coefficients and the FT of one period
sin ( 5ω / 2 ) − j 2ω
( )
4
⎧1; 0 ≤ n ≤ 4
x [n ] = ⎨ ⇒ X e
jω
= ∑ e − jω n
=
sin (ω / 2 )
e
⎩ 0; otherwise n=0

• Sampling the Fourier transform
– Since X ( e jω ) is periodic with period 2π , X [ k ] is also periodic in k with period N.

⎛ j 2Nπ k ⎞ N −1 2π
X [ k ] = X ( e jω )
−j k
– Since
2π = X ⎜ e ⎟ = ∑ x [ n ]e N ,
ω= k
N ⎝ ⎠ n =0
2π 2π
1 N −1 ⎡ ∞ − j mk ⎤ j kn
x [ n ] = ∑ ⎢ ∑ x [ m ] e N ⎥e N
N k =0 ⎣ m=−∞ ⎦
∞ 2π
1 N −1 j N ( k −m )n
= ∑ x [ m] ∑ e
m =−∞ N k =0
∞ ∞
= ∑ x [ m ] ∑ δ [ n − m − AN ]
m =−∞ A =−∞
∞
= x [ n ] ∗ ∑ δ [ n − AN ]
A =−∞
∞
= ∑ x [ n − AN ]
A =−∞
The samples of the F.T. of x [ n ] can be thought of as DFS coefficients of a periodic

sequence x [ n ] obtained through periodic replication of x [ n ]

– A periodic sequence X [ k ] can be interpreted as equally spaced samples of the
Fourier transform of one period of x [ n ]
⎧ x [ n ] , 0 ≤ n ≤ N − 1
x [ n] = ⎨
⎩ 0 , otherwise
N −1 N −1
X (e jω
) = ∑ x [ n] e − jω n
= ∑ x [ n ] e− jω n
n =0 n =0
2π
⇒ X [ k ] = X ( e jω ) 2π k ⇐ Sampling of the Fourier transform at a rate of
ω=
N ω
Example ⎧1, 0 ≤ n ≤ 4
x [ n] = ⎨
⎩0, otherwise
5ω
2 e− j 2ω ⇒ X [ k ] = sin ( kπ / 2 ) e− j 4 kπ /10
sin
X ( e jω ) = ∑ e− jω n
4
=
n =0 sin
ω sin ( kπ /10 )
2

• Fourier representation of finite-duration sequences: Discrete Fourier transform (DFT)

⎧ x [ n ] , 0 ≤ n ≤ N − 1 ∞
x [ n] = ⎨ ⇔ x [ n ] = ∑ x [ n − AN ] = x ⎡⎣( ( n ) ) N ⎤⎦
⎩ 0 , otherwise A =−∞
where x ⎡⎣( ( n ) ) N ⎤⎦ x ⎡⎣( n modulo N ) ⎤⎦
– x [ n ] can be visualized as warping the finite sequence around the cylinder

– Since the Fourier series coefficients X [ k ] of x [ n ] is also a periodic sequence with
a period of N, the DFT of x [ n ] can be obtained by
⎪⎧ X [ k ] , 0 ≤ k ≤ N − 1
X [k ] = ⎨
⎪⎩0 , otherwise
∞
X [ k ] = X ⎡⎣( ( k ) ) N ⎤⎦ ⇔ X [ k ] = ∑ X [ k − mN ]
m =−∞
N −1
1 N −1
– Since X [ k ] = ∑ x [ n ]WN and x [ n ] = ∑ X [ k ]WN− nk ,
nk

n =0 N k =0
⎧ N −1 ⎧ 1 N −1
⎪∑ [ ] ⎪ ∑ X [ k ]WN , 0 ≤ n ≤ N − 1
− nk
x n W nk
, 0 ≤ n ≤ N − 1
X [ k ] = ⎨ k =0 and x [ n ] = ⎨ N k =0
N
⎪ ⎪
⎩ 0 , otherwise ⎩ 0 , otherwise
x [ n ] ←⎯⎯
DFT
→ X [k ]

– Example
N −1 − j 2π
1 − e − j 2π k ⎧ N , k = 0, ± N , ±2 N , ⋅⋅ ⋅
X [ k ] = ∑ e N
= 2π k
=⎨
n =0
1− e
−j
N ⎩0, otherwise
N=5; x [ n]
x [ n ]
X [ k ] X [k ]

– Example: N=10
x ⎡⎣ n ⎤⎦
x ⎡⎣ n ⎤⎦
X [k ]

• Properties of the DFS
– Linearity
x [ n ] ←⎯⎯
DFT
→ X [ k ] ⇒ α x1 [ n ] + β x2 [ n ] ←⎯⎯
DFT
→α X 1 [ k ] + β X 2 [ k ]
N1 N2 N ≥ max ⎡⎣ N1, N 2 ⎤⎦ with zero padding
– Circular shift
x1 [ n ] = x [ n − m ] ←⎯⎯
DFT
→ X ( e jω ) e − jω m
2π
x1 [ n ] ←⎯⎯
→ X1 [k ] = X [k ]e
−j km
Let DFT N
Since x1 [ n ] = 0 outside 0 ≤ n ≤ N -1, x1 [ n ] cannot be obtained by simply time shifting

x [ n]
⇒ x1 [ n ] = x1 ⎡⎣( ( n ) ) N ⎤⎦ ←⎯⎯

DFT
→ X 1 [ k ] = X 1 ⎡⎣( ( k ) ) N ⎤⎦
2π m( ( k ) ) 2π km 2π km
⇒ X 1 [ k ] = X ⎡⎣( ( k ) ) N ⎤⎦ e = X ⎡⎣( ( k ) ) N ⎤⎦ e
−j −j −j
= X [ k ] e
N
N N N
⎧⎪ x1 [ n ] = x ⎡( ( n − m ) ) ⎤ , 0 ≤ n ≤ N − 1
⇒ x1 [ n ] = ⎨ ⎣ N⎦
2π km
⇒ x ⎡⎣( ( n − m ) ) N ⎤⎦ ←⎯
−j
⎯
DFT
→ X [k ]e N
, 0 ≤ n ≤ N − 1: circular shift

– Example: circular shift

– Duality
x [ n ] = X ⎣⎡( ( n ) ) N ⎦⎤ ←⎯ ⎯
DFT
→ X [ k ] = X ⎣⎡( ( k ) ) N ⎦⎤
⇒ X [ k ] ←⎯⎯ DFT
→ Nx [ − k ]
Let x1 [ n ] = X [ n ] ⇒ x1 [ n ] ←⎯⎯

DFT
→ X 1 [ k ] = Nx [ − k ]
⎧ Nx [ − k ] , 0 ≤ k ≤ N − 1
X1 [k ] = ⎨
⎩0, otherwise
⎧⎪ N x ⎡( ( − k ) ) ⎤ , 0 ≤ k ≤ N − 1
=⎨ ⎣ N⎦
⎪⎩0, otherwise
⇒ X [ n ] ←⎯⎯ DFT
→ N x ⎡⎣( ( − k ) ) N ⎤⎦ , 0 ≤ k ≤ N − 1
– Symmetry
DFT of x [ n ] = DFS coefficients of x [ n ] = x ⎡⎣( ( n ) ) N ⎤⎦
⇒ x∗ [ n ] ←⎯⎯
DFT
→ X ∗ ⎡⎣( ( − k ) ) N ⎤⎦ , 0 ≤ n ≤ N − 1
x∗ ⎡⎣( ( − n ) ) N ⎤⎦ ←⎯⎯
DFT
→ X ∗ [k ]
{ x [ n ] + x ∗ [ − n ]}; x [ n ] = x ⎡⎣( ( n ) ) N ⎤⎦
1
xe [ n ] =
2
x0 [ n ] = { x [ n ] − x ∗ [ − n ]}
1
2
xe [ n ] = xe [ n ] , 0 ≤ n ≤ N − 1
=
1
2
{
x ⎡⎣( ( n ) ) N ⎤⎦ + x∗ ⎡⎣( ( n ) ) N ⎤⎦}
= { x [ n ] + x∗ [ N − n ]} ⇐ ( ( − n ) ) N = ( N − n ) ; ( ( n ) ) N = n; 0 ≤ n ≤ N − 1
1
2
→ X e [ k ] = { X [ k ] + X ∗ [ k ]} = Re{ X [ k ]}
1
xe [ n ] ←⎯⎯
DFT
2
Similarly,
xo [ n ] = xo [ n ] , 0 ≤ n ≤ N − 1
⇒ x0 [ n ] ←⎯⎯
DFT
→ j Im { X [ k ]}
– N-point Circular convolution

N −1
x3 [ n ] = ∑ x1 [ m ] x2 [ n − m ] , 0 ≤ n ≤ N −1
m =0
N −1
= ∑ x1 ⎡⎣( ( m ) ) N ⎤⎦ x2 ⎡⎣( ( n − m ) ) N ⎤⎦
m=0
N −1
= ∑ x1 [ m ] x2 ⎡⎣( ( n − m ) ) N ⎤⎦
m=0
Δ x1 [ n ] N x2 [ n ]
N −1
= ∑ x1 ⎡⎣( ( n − m ) ) N ⎤⎦ x2 [ m ]
m=0
= x2 [ n ] N x1 [ n ]
Example
x1 [ n ] = δ [ n − 1]

Example: N-point circular convolution

⎧1 0 ≤ n ≤ L − 1
x1 [ n ] = x2 [ n ] = ⎨
⎩0 otherwise
N=L:
N −1
⎧N , k = 0
X 1 [ k ] = X 2 [ k ] = ∑1 ⋅ WNkn = ⎨
n =0 ⎩0, otherwise
⎧N 2, k = 0
⇒ X 3 [k ] = X1 [k ] X 2 [k ] = ⎨
⎩0, otherwise
⇒ x3 [ n ] = N , 0 ≤ n ≤ N − 1
N =2 L: Linear convolution ?
x1 [ n ] N x2 [ n ] ←⎯⎯
DFS
→ X1 [k ] X 2 [k ]
1 N −1
⇔ x1 [ n ] x2 [ n ] ←⎯⎯
DFS
→ X 3 [k ] = ∑ X 1 [A ] X 2 ⎣⎡( ( k − A ) ) N ⎦⎤
N A =0
1
X1 [k ] N X 2 [k ]
N

Example: N(=2L)-point circular convolution

⎧1 0 ≤ n ≤ L − 1
x1 [ n ] = x2 [ n ] = ⎨
⎩0 otherwise
x1 [ n ] N x2 [ n ] ←⎯⎯ DFT
→ X1 [k ] X 2 [k ]
1 N −1
x1 [ n ] x2 [ n ] ←⎯⎯
DFT
→ X 3 [ k ] = ∑ X 1 [ A ] X 2 ⎣⎡( ( k − A ) ) N ⎦⎤
N A =0
1
= X1 [k ] N X 2 [k ]
N
• Linear convolution using the DFT
x1 [ n ]
∞
x3 [ n ] = ∑ x [ m] x [ n − m]
1 2
m =−∞
0 L
x2 [ n ]
0 L + p −1
0 p
– N Δ ( L + p − 1)-point circular convolution with zero padding
⎧ x [ n ] , 0 ≤ n ≤ N − 1
x p [ n] = ⎨ ; x [ n ] = ∑ x [ n − mN ]
⎩ 0, otherwise m
X 3 ( e jω ) = X 1 ( e jω ) X 2 ( e jω ) ⇒ X 3 [ k ] = X 3 ( e jω ) 2π k , 0 ≤ k ≤ N −1
ω=
N
= X 1 ( e jω ) X 2 ( e jω ) 2π k
ω=
N
= X1 [k ] X 2 [k ]
⇒ x3 p [ n ] = x1 [ n ] N x2 [ n ]

– The circular convolution of two finite-length sequences can be equivalent to linear

convolution of the two sequences.
⎧⎪∑ x3 [ n − mN ] , 0 ≤ n ≤ N − 1
x3 p [ n ] = ⎨ m
⎪⎩0, otherwise
– Calculation of the output of an LTI using the DFT

hn
zero padding
0 p
y [ n] = h [ n] ∗ x [ n]
xn
L −1
0
L L
– How to resolve the issue when x [ n ] is infinite?
Assume x [ n ] = 0 for n < 0

The sequence x [ n ] can be decomposed in to a sum of shifted finite-length
∞ ⎧ x [ n + mL ] , 0 ≤ n ≤ N − 1
segments of length L x [ n ] = ∑ xm [ n − mL ]; xm [ n ] = ⎨
m =0
⎩ 0 , otherwise
– Illustration of warping around due to improper circular convolution
N=L N = L + p −1

• Linear convolution using the discrete Fourier transform
– How to resolve the issue when x [ n ] is infinite?

Assume x [ n ] = 0 for n < 0
The sequence x [ n ] can be decomposed in to a sum of shifted finite-length
segments of length L
∞
x [ n ] = ∑ xm [ n − mL ];
m =0
⎧ x [ n + mL ] , 0 ≤ n ≤ N − 1
where xm [ n ] = ⎨
⎩ 0 , otherwise

– Overlap-add method y [ n] = h [ n] ∗ x [ n]
xm [ n ] = x [ n + mL ] , 0 ≤ n ≤ L − 1
∞
y [ n ] = ∑ ym [ n − mL ]
N -point circular convolution
m =0
ym [ n ] = h [ n ] ∗ xm [ n ]

– Overlap-save method y [ n ] = h [ n ] ∗ x [ n ]
xm [ n ] = x ⎡⎣ n + m ( L − p + 1) − p + 1⎤⎦
0 ≤ n ≤ L −1
N -point circular convolution
∞
y [ n ] = ∑ ym ⎡⎣ n − m ( L − p + 1) + p − 1⎤⎦
m =0
ym [ n ] = h [ n ] ∗ xm [ n ]

Homework #7 :
• The discrete Cosine transform (DCT)
8 - 29, - 31, - 34, - 36,
– General class of finite-length transform - 39, - 49, & - 64
N −1
Α [ k ] = ∑ x [ n ] φk∗ [ n ] Due by June 9
n =0
1 N −1
x [ n] =
∑ Α [ k ] φk [ n ]
N k =0 ⎧1, m = k
where {φk [ n ]} are referred to the basis sequences s.t.
1 N −1
∑ φk [ n ]φm∗ [ n ] = ⎨
2π kn N n =0 ⎩0, m ≠ k
e.g. DFT: φk [ n ] = e
j
N
– DCT definition
π kn
The basis sequences {φk [ n ]} are cosine functions cos
N
Assume that x [ n ] are periodic and even symmetric
– Type-I periodic sequence: even periodic symmetry about n=0, N-1, 2(N-1),
x1 [ n ] = xα ⎡⎣( ( n ) )2 N − 2 ⎤⎦ + xα ⎡⎣( ( − n ) )2 N − 2 ⎤⎦
xα [ n ] = α [ n ] x [ n ]
⎧1
⎪ n = 0 and N -1
α [ n] = ⎨ 2
⎩⎪ 1 1 ≤ n ≤ N − 2

DCT-1 is defined by
N −1
π kn
X 1c [ k ] = 2∑α [ n ] x1 [ n ] cos , 0 ≤ k ≤ N −1
n =0 N −1
1 N −1 π kn
x1 [ n ] = ∑ α [ k ] X 1c [ k ] cos , 0 ≤ n ≤ N −1
N − 1 k =0 N −1
– Type-II periodic sequence: periodic with period 2N

x2 [ n ] = x ⎡⎣( ( n ) )2 N ⎤⎦ + x ⎡⎣( ( − n − 1) )2 N ⎤⎦
DCT-2 is defined by
N −1 π k ( 2n + 1)
X 2c [ k ] = 2∑ x2 [ n ] cos , 0 ≤ k ≤ N −1
n =0 2N
1 N −1 π k ( 2n + 1)
x2 [ n ] = ∑ β [ k ]X 2c [ k ] cos , 0 ≤ n ≤ N −1
N k =0 2N
⎧1
⎪ , k =0
β [k ] = ⎨ 2
⎩⎪ 1 , 1 ≤ k ≤ N − 1

– Relationship between DFT and DCT-1

x1 [ n ] = xα ⎡⎣( ( n ) )2 N − 2 ⎤⎦ + xα ⎡⎣( ( − n ) )2 N − 2 ⎤⎦ = x1 [ n ] , 0 ≤ n ≤ 2 N − 3
2(N-1)-point DFT of x1 [ n ] :
X 1 [ k ] = X α [ k ] + X α∗ [ k ] = 2 Re{ X α [ k ]} , 0 ≤ k ≤ 2 N − 3
N −1
2π kn
X α [ k ] = 2∑α [ n ] x [ n ] cos
n =0 2 ( N − 1)
= X 1c [ k ]
The DCT-1 of N-point sequence is identical to 2(N-1)-point DFT of x1 [ n ].
2π kn
2 N −3
1
x1 [ n ] = ∑ X [k ]e
j
2( N −1)
2 ( N − 1) k =0 1
π kn
⎡ N −1 c π kn
( N −1) ⎤
π kn 2 N −3
[ 2 N − 2 − k ]e
2 N −3 j
1
⎢ ∑ X [ k ]e + ∑ X 1 [ 2( N − 1) − k ]e ∑ ( N −1)
j j
( N −1) X 1c
= c
⎥
2 ( N − 1) ⎣ k =0 1 k =N
⎦
k=N
N −2 ( 2 N − 2 − m )π n
= ∑ X 1c [ m ]e
j
1 ⎡ N −1 c π kn N −2
N −1 ⎤
π kn N −1
∑ [ ] ∑ [ ]
j j
= X k e N −1
+ X c
k e
2 ( N − 1) ⎢⎣ k =0 ⎥
1 1 m =1
k =1 ⎦ N −2 − nmπ
⎤ =∑ [ m]e
j
1 ⎡ N − 2 c ⎛ j πNkn−1 − j πNkn−1 ⎞ π ( N −1) n X 1c N −1
∑ [ ]⎜ 1[ ] 1[ 1] e
j
= + ⎟ + c
+ c
− N −1 m =1
2 ( N − 1) ⎢⎣ k =1 1 ⎥
X k e e X 0 X N
⎝ ⎠ ⎦
⎧1
1 N −1 π nk ⎪ , k = 0 and N − 1
= ∑ α [ k ] X 1c [ k ] cos , 0 ≤ n ≤ N − 1; α [k ] = ⎨ 2
N − 1 k =0 N −1
⎩⎪1 , 1 ≤ k ≤ N − 2
Note : DCT-1 involves only real-valued coefficients

⇒ reduce the implementation complexity.
– Relationship between DFT and DCT-2

x2 [ n ] = x ⎣⎡( ( n ) )2 N ⎦⎤ + x ⎣⎡( ( − n − 1) )2 N ⎦⎤ = x2 [ n ] , 0 ≤ n ≤ 2 N − 1
2N-point DFT of 2N-point sequence x2 [ n ]

X 2 [ k ] = X [ k ] + X * [ k ] e j 2π k / 2 N
= e jπ k / 2 N ( X [ k ] e − jπ k / 2 N + X * [ k ] e jπ k / 2 N )
= 2e jπ k / 2 N Re{ X [ k ] e − jπ k / 2 N } , 0 ≤ k ≤ 2 N − 1
N −1
π k (2n + 1)
Re{ X [ k ] e − jπ k / 2 N } = ∑ x [ n ] cos
k =0 2N
⇒ X 2 [ k ] = 2Re { X [ k ] e
c − jπ k / 2 N
}; 0 ≤ k ≤ N − 1
⇒ X 2c [ k ] = X 2 [ k ] e − jπ k / 2 N ; 0 ≤ k ≤ N − 1
N −1
π (2 N − k )(2n + 1)
X 2c [ 2 N − k ] = 2∑ x [ n ] cos = − X 2c [ k ]; 0 ≤ k ≤ 2 N − 1
k =0 2N
⎧ X 2c [ 0] , k =0
⎪ c 1 2 N −1
⎪ X 2 [k ]e k = 1, ⋅ ⋅⋅, N − 1 x2 [ n ] = ∑ X 2 [ k ] e j 2π k / 2 N ,
jπ k / 2 N
, ⇔
⇒ X 2 [k ] = ⎨ 2 N k =0
⎪0, k=N
n = 0,1, ⋅ ⋅ ⋅, 2 N − 1
⎪− X c [ 2 N − k ] e jπ k / 2 N , k = N + 1, ⋅ ⋅⋅,2 N − 1
⎩ 2
– Energy compaction property of DCT-2

Since DCT-2 coefficients of a finite-length sequence are concentrated in the low
indices than DFT, DCT-2 is often used for data compression applications in
preference to DFT
N −1
1 N −1
∑ x [ n] ∑ β [ k ] X 2c [ k ]
2 2
=
n=0 N k =0
32 - point DFT
Example x [ n ] = a cos (ω 0 n + φ )
n
⇐ 32 − point DCT

Energy concentration can be examined by measuring the truncation error of the

transformations.
1 N −1
xm [ n ] = ∑ Tm [ k ] X [ k ] e j 2π k / 2 N , n = 0,1, ⋅ ⋅ ⋅,2 N − 1; m = 1,3,5, ⋅ ⋅ ⋅, N − 1
N k =0
⎧1, 0 ≤ k ≤ ( N −1− m)/ 2

⎪
Tm [ k ] = ⎨0, ( N +1− m)/ 2 ≤ k ≤ ( N −1+ m)/ 2
⎪
⎩1, ( N +1+ m)/ 2 ≤ k ≤ N −1
1 N −1− m π k (2n + 1)
xmc [ n ] = ∑ β [ k ] X 2c [ k ] cos ; 0 ≤ n ≤ N −1
N k =0 2N
1 N −1
MS truncation error : EDFT [ m ] = ∑ x [ n ] − xm [ n ]
2
N n =0
1 N −1
EDCT [ m ] = ∑ x [ n ] − xmc [ n ]
2
N n =0

Discrete Fourier transform

• Discrete Fourier series representation of periodic sequences
– Let x [ n ] be a periodic sequence with period N such that x [ n ] = x [ n + mN ] for any

integer m
Example A periodic complex exponential

2π kn
ek [ n ] = e = ek [ n + mN ]; k is an interger
j
N
Note that ek [ n ] = ek +AN [ n ] for any integer A

– A set of N periodic complex exponentials {e [ n]} ,
k k = 0,1,2,...N − 1, defines all the
frequency components
2π
1 N −1 i
⇒ x [ n ] = ∑ X [ k ]e N
j kn
N k =0
N −1 2π mn
−j
X [ k ] = ∑ x [ n ] e N
n =0
– Fourier series representation

2π
1 N −1
x [ n ] = ∑ i X [ k ]e N
j kn
N k =0

N −1 2π N −1 2π kn 2π
−j 1 N −1 − j rn
∑ x [ n]e = ∑ ∑ X [ k ]e
rn j
N N
e N
n =0 n =0 N k =0
N −1 N −1 2π
1 j ( k − r )n
= ∑ X [ k ] ∑ e N
k =0 N k =0
1 N −1 j 2Nπ ( k −r )n ⎧1, k − r = AN
Since ∑e =⎨
N n =0 ⎩0, otherwise
N −1 2π
∑ x [ n] e = X [ r ]
j rn
N
n =0
The Fourier series coefficient can be obtained by

N −1 2π
−j
X [ k ] = ∑ x [ n ] e
kn
N
n =0
N −1 2π N −1 2π
− j (k +N )n −j
X [ k + N ] = ∑ x [ n ] e = ∑ x [ n ] e = X [ k ]
kn
N N − j 2π n
e
n −0 n−0
The Fourier series coefficients is a periodic sequence

– Discrete Fourier series (DFS)
Define by x [ n ] ←⎯⎯
DFS
→ X [ k ]
N −1 2π
−j
Analysis equation : X [ k ] = ∑ x [ n ] wNkn ; wN = e N
n =0
1 N −1
Synthesis equation : x [ n ] = ∑ X [ k ]wN− kn
N k =0
Example Periodic pulse train

∞
⎧1, n = mM
x [ n ] = ∑ δ [ n − mN ] = ⎨
r =−∞ ⎩0, otherwise
Since x [ n ] = δ [ n ] , 0 ≤ n ≤ N − 1
N −1
X [ k ] = ∑ δ [ n ] wN− kn = 1
n =0
2π
1 1 N −1 − kn 1 N −1 j N kn
⇒ x [ n ] = ∑ X [ k ] wN = ∑ wN = ∑ e
− kn
N N k =0 N k =0
Example Duality in DFS
Let the DFS coefficient be a periodic impulse train
∞
Y [ k ] = N ∑ δ [ k − rN ]
r =−∞
N −1
⇒ y [ n ] = ∑ N δ [ k ] wN− kn =wN−0 = 1, ∀n
k =0
Y [ k ] = Nx [ n ] ↔ y [ n ] = X [ n ]

Example Periodic rectangular pulse train
πk
4 4 −j
2π
1−W 5k sin 4π k
X [ k ] = ∑W nk
= ∑e 10
kn
= =
10 2 e − j 10
n −0
10
n =0 1−W k
πk
sin
10
10
Any periodic sequence can be represented as a sum of complex exponential

sequences
• Properties of DFS
– Linearity
ax1 [ n ] + bx2 [ n ] ←⎯⎯
DFS
→ aX [ k ] + bX [ k ]
– Shift of a sequence
N −1 2π N −1 2π 2π 2π
−j −j ( n−m ) − j −j
∑ x [ n − m] e = ∑ x [ n − m ] e = X [ k ] e = X [ k ]WNkm
kn km
N N N N
e
n =0 n=0
⇒ x [ n − m ] ←⎯⎯DFS
→ X [ k ]WNkm
– Duality
x [ n ]W − nA ←⎯⎯
N
DFS
→ X [ k − A ]
1 N −1 N −1
x [ n ] = ∑ X [ k ]wN ⇒ Nx [ − n ] = ∑ X [ k ]WNkn
− kn
N k =0 k =0
N −1
⇒ Nx [ − k ] = ∑ x [ n ]WNkn
k =0
If x [ n ] ←⎯⎯
DFS
→ X [ k ] , X [ n ] ←⎯⎯
DFS
→ Nx [ − k ]
– Symmetry properties
x * [ n ] ←⎯⎯
DFS
→ X * [ − k ]
x * [ −n ] ←⎯⎯
DFS
→ X * [ k ]

⇒ If x [ n ] is real, X [ k ] = X * [ − k ]
X [ k ] = X [ − k ]
– Periodic convolution
N −1 N −1
x3 [ n ] = ∑ x1 [ m ] x2 [ n − m ] = ∑ x2 [ m ] x1 [ n − m ]
m =0 m=0
⎛
N −1 N −1
⎞
X 3 [ k ] = ∑ ⎜ ∑ x1 [ m ] x2 [ n − m ] ⎟WNkn

n =0 ⎝ m =0 ⎠
N −1 N −1
= ∑ x1 [ m ] ∑ x2 [ n − m ]WNkn
m =0 n =0
N −1
= X 2 [ k ] ∑ x1 [ m ]WNkm = X 1 [ k ] X 2 [ k ]
m =0
N −1
⇒ ∑ x1 [ m ] x2 [ n − m ] ←⎯⎯
DFS
→ X 1 [ k ] X 2 [ k ]
m =0
Duality
1 N −1
x1 [ n ] x2 [ n ] ←⎯⎯
→ ∑ X 1 [ m ] X 2 [ k − m ]
DFS
N m =0
Difference between aperiodic convolution and periodic convolution
• The sum is over an interval 0 ≤ m ≤ N − 1
• x 2 [ n − m ] = x 2 ⎡⎣ − ( m − n ) ⎤⎦

Example Periodic convolution

– Properties of DFS

– Properties of DFS (cont.)

• Fourier transform of periodic signals
– A sequence should be absolutely summable to guarantee the uniform convergence of

its Fourier transform
– Periodic signals are not absolutely summable

– The F.T. of a periodic signal can be interpreted to be an impulse train in the frequency
domain with the impulse values proportional to the DFS coefficients for the sequence
– If x [ n ] is periodic with period N, the Fourier transform of x n [ ] is defined as
2π ∞
⎛ 2π k ⎞
x [ n ] ⇔ X ( e jω ) = ∑
X [ k ] δ ⎜ ω − ⎟
N k =−∞ ⎝ N ⎠
Although the F.T. of a periodic sequence does not converge in the normal sense,
the introduction of impulses enables to use the F. T.
Note: X e
jω
( ) has the necessary periodicity with period 2π
2π ⎛ 2π k ⎞ jωn
X ( e jω )e jωn dω =
1 2π −ε 1 2π −ε
∫ ∫ ∑ X [ k ]δ ⎜ ω − ⎟ e dω ; 0 < ε < 2π / N
2π 0−ε 2π 0−ε N k ⎝ N ⎠
1 ∞ 2π −ε ⎛ 2π k ⎞ jωn
= ∑ X [k ] ∫ δ ⎜ω − ⎟e dω
N k =−∞ 0 −ε
⎝ N ⎠
2π
1 ∞
= ∑ X [ k ] e N = x [ n ]
j kn
N k =−∞
Example: Fourier transform of a periodic impulse train

∞
p [ n ] = ∑ δ [ n − mN ] ⇒ P [ k ] = 1, ∀ k
m = −∞
2π ∞
⎡ 2π k ⎤
⇒ P e jω =( ) ∑ δ
N k = −∞ ⎢⎣
ω −
N ⎥⎦
– For a finite-length sequence x [ n ] , ( i.e ., x [ n ] = 0 for n < 0 and n ≥ N )
x [ n ] = x [ n ] ∗ p [ n ]
∞ ∞ ∞
= ∑ ∑ δ [A − mN ]x [ n − A] = ∑ x [ n − mN ]
m =−∞ A =−∞ m =−∞
X ( e jω ) = X ( e jω ) P ( e jω )
2π ∞
⎛ 2π k ⎞
= X (e ) jω
∑ δ ⎜⎝ ω − ⎟
N k =−∞ N ⎠
2π ∞
⎛ j 2Nπ k ⎞ ⎛ 2π k ⎞
= ∑ ⎜ ⎟ ⎜⎝
X e δ ω −
N ⎠
⎟
N k =−∞ ⎝ ⎠
⎛ j 2Nπ k ⎞ N −1 2π
⇒ X [ k ] = X ( e jω )
−j
⎟ ∑ [ ]
k
2π = X ⎜e =
x n e N
ω=
N
k
⎝ ⎠ n =0
Thus, X [ k ] is obtained by sampling the Fourier transform of the finite sequence x [ n ]
Example: Relationship between the DFS coefficients and the FT of one period
sin ( 5ω / 2 ) − j 2ω
( )
4
⎧1; 0 ≤ n ≤ 4
x [n ] = ⎨ ⇒ X e
jω
= ∑ e − jω n
=
sin (ω / 2 )
e
⎩ 0; otherwise n=0

Seoul National University 7. Fourier analysis of signals using DFT
Fourier Analysis of Signals Using DFT

• Discrete Fourier transform (DFT) can analyze the frequency content of continuous-time
signal
1 ∞ ⎛ ω 2π k ⎞
X (e ) = ∑ X c ⎜ +
jω
⎟
T k =−∞ ⎝ T T ⎠
V ( e jω ) = X ( e jθ )W ( e j (ω −θ ) ) dθ
1 π
2π ∫π
−
v [ n] = x [ n] w[ n ]
N −1 2π kn
= V ( e jω )
−j
V [ k ] = ∑ v [ n ]e N
2π k ; ω = ΩT
ω=
n=0 N

– Example

– Length of the DFT

N −1 2π kn
= V ( e jω )
−j
V [ k ] = ∑ v [ n] e N
2π k ; ω = ΩT
ω=
n =0 N
2π 2π
ΔΩ ≤ ⇒N≥
NT ΔΩT
Example: T=1/5000 and 10Hz resolution

2π 5000
10 ⋅ 2π ≤ ⇒N≥ = 500 ⇒ N = 512 = 29
NT T =1/ 5000 10
– Relationship between DFT values
N = 512; T = 1/ 5000
If V [11] = 2000 (1 + j ) ;
V [ k ] = V * [ ((− k )) N ] ⇒ V [ N − k ] = V * [ k ]
⇒ V [512 − 11] = V [501] = V * [11] = 2000 (1 − j )
Ω11 = 2π ⋅ 11 ⋅ 5000 / 512 = 2π ⋅ 107.4
⇒ X c ( Ω11 ) ≈ T ⋅ V [11] = 0.4 (1 + j )

• DFT analysis of sinusoidal signals
– Windowing and spectral sampling have an important effect on the analysis of

sinusoidal signals using the DFT
– Effect of windowing
sc (t ) = A0 cos(Ω 0t + θ 0 ) + A1 cos(Ω1t + θ1 )
⇒ x [ n ] = A0 cos(ω0 n + θ 0 ) + A1 cos(ω1n + θ1 )
After the windowing
v [ n ] = x [ n] w[ n]
= A0 w[ n ] cos(ω0 n + θ 0 ) + A1w[ n ] cos(ω1n + θ1 )
=
A0
2
( )
w[ n ] e jω0n+θ0 + w[ n ] e − j (ω0n+θ0 ) + 1 ( w[ n ] e jω1n+θ1 + w[ n ] e − j (ω1n+θ1 )
A
2
V ( e jω ) =
A0 ⎡
2 ⎣
( ) ( ) A
( )
W e ( 0 ) e jθ0 + W e ( 0 ) e − jθ0 ⎤ + 1 ⎡W e ( 1 ) e jθ1 + W e ( 1 ) e − jθ1 ⎤
j ω −ω j ω +ω
⎦ 2⎣
j ω −ω j ω +ω
⎦( )
it can be suffered from reduced resolution and leakage of spectrum

Example Leakage due to the use of a rectangular window
2π
ω0 =
6
2π
ω1 =
3

The resolution is primarily affected by the width of the main lobe of W e( )

jω
The degree of leakage depends on the relative amplitude of the main and side
lobe of W e jω ( )
These are associated with the window length L and the shape of the window

Example: Kaiser window
⎧ ⎡ n −α ⎤ ⎤
2
⎪ I0 ⎢β 1 − ⎡
⎢ ⎥ ⎥
⎪ ⎢ ⎣ α ⎦ ⎥⎦ L −1
w[n] = ⎨ ⎣ , 0 ≤ n ≤ L − 1, α =
⎪ I0 ( β ) 2
⎪
⎩ 0 , otherwise
A sl
24π ( Asl + 12 )
L +1
155 Δ ml
– Effect of spectral sampling
V [ k ] = V ( e jω ) 2π k ; Ω k = 2π k / ( NT )
ω=
N
Example N = 64
2π n 4π n
v [ n ] = cos + 0.75 cos ;
14 15
0 ≤ n ≤ 63

– Spectral sampling with frequencies matching DFT frequencies

2π n 2π n
Example v [ n ] = cos + 0.75 cos ; 0 ≤ n ≤ 63
16 8
N = 64

– DFT analysis by Kaiser window

2π n 4π n
Example v [ n ] = wK [ n ] cos + 0.75 wK [ n ] cos ; 0 ≤ n ≤ 63
14 15
L = 64 L = 32

– DFT analysis with 32-point Kaiser window and zero padding

2π n 4π n
Example v [ n ] = cos + 0.75 cos ; β = 5.48, L = 32 Kaiser window
14 15
N = 32 N = 128
N = 64 N = 1024
– DFT analysis with Kaiser window and zero padding

2π n 4π n
Example v [ n ] = cos + 0.75 cos ; β = 5.48 Kaiser window
14 15
N = 1024, L = 32 N = 1024, L = 54
N = 1024, L = 42 N = 1024, L = 64

• Fourier analysis of stationary random signals: Periodogram
– The sample mean and sample variance respectively defined by

2
1 L −1 1 L −1
mˆ X = ∑ x [ n ]; σˆ X = ∑ ( x [ n ] − mˆ X )
2
L n =0 L n =0
are unbiased and asymptotically unbiased estimators, respectively
– Periodogram analysis: estimation of the power spectrum
⎛ω ⎞ L −1
V (e ) = ∑ x [ n]w[ n] e
1
PXX (ω ) = PSS ⎜ ⎟ jω − jω n
T ⎝T ⎠ n =0
Periodogram is the F. T. of aperiodic correlation of windowed data sequence
L −1
V ( e jω ) =
1 1
I (ω ) = ∑ cvv [ m ]e − jωm
2
LU LU m =− L +1
L −1
cvv [ m ] = ∑ x [ n ]w[ n ] x [ n + m ] w [ n + m ]
n =0
Modified periodogram: w [ n ] is not a rectangular window

– Periodogram can be calculated by means of the DFT

1
I ( ωk ) = V [k ]
2
LU
– Properties of the periodogram:
L −1
E { I (ω )} = ∑ E {c [ m]}e
1 − jω m
vv
LU m =− L +1
L −1
E {cvv [ m ]} = ∑ w[ n ] w[ n + m ] E { x [ n ] x [ n + m ]}
n =0
L −1
= ∑ w[ n ] w[ n + m ]φxx [ m ]
n =0
= cww [ m ]φxx [ m]
L −1
φxx [ m ] = E { x [ n ] x [ n + m ]} ; cww [ m ] = ∑ w[ n ] w[ n + m ]
n =0
E { I (ω )} = ( ) dθ ; CWW ( e jω ) = W ( e jω )
1 π
PXX (θ ) CWW e
j (ω −θ )
2
LU ∫π
−
To make the periodogram unbiased,

L −1
1 L −1 2
∫ π W (e )
1 π 1
∑ w [ n] = 1 ⇒ U = ∑ w [ n]
2
jω
dω = 2
2π LU − LU n =0 L n =0

– As the window size increases, the peridogram becomes a consistent estimate

Example: white noise generation with rectangular windowing
L −1 2
1 1
I [ k ] = I ( ωk ) = V [ k ] = ∑ w[ n] x [ n] e
2 − j 2π kn / N
L L m =− L +1
L = 16 LL==256
256
L = 64 varL{=I 1024
(ωk )} 2
PXX (ω )

– Edge effect due to covolution of finite sequences

L −1
cvv [ m ] = ∑ w[ n ] w[ n + m ] x [ n ] x [ n + m ]
n =0
As m becomes close to L, the

calculation becomes inconsistent
This problem can be alleviated by

averaging multiple independent
periodogram estimates
xr [ n ] = w[ n ] x [ rR + n ]; 0 ≤ n ≤ Q − 1
1 K −1 1 L −1
I (ω ) = ∑ ∑ X r ( e jω )
2
K r =0 LU n =0

– Example ( )
x [ n ] = 0.5cos ( 2π n / 21 + θ ) + e [ n ]; e [ n ] ∼ U − 3, 3 ⇒ σ e2 = 1
A2π
PXX (ω ) = ⎡⎣δ (ω − ω0 ) + δ (ω + ω0 ) ⎤⎦ + σ e2
2
Q = 1024; U = 1
K = 31, L = 64
K = 1, L = 1024 K = 127, L = 16

Discrete-Time Signals and Systems

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

Discrete-Time Signals and Systems

Uploaded by

Copyright:

Available Formats

Seoul National University 1.

Discrete time signals and systems

1. Discrete-time signals and systems

– Discrete-time signals are represented as sequences of numbers

discrete-time signal analog signal

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

– An unit sample sequence is defined by

which is also referred to as a discrete-time impulse or simply as an impulse

– An arbitrary sequence can be represented as a sum of weighted and delayed

Ex) x [ n ] = m−3δ [ n + 3] + m1δ [ n − 1] + m4δ [ m − 4] m1

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Let x [ n ] be an exponential sequence given by

– Complex exponential sequence ( α = 1)

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

– A sequence is called periodic with an integer period of N, if

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

– Ideal delay system

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Linear time-invariant systems

– Let nk [ n ] be the response of the system to the input δ [ n − k ] , i.e., hk [ n ] = T {δ [ n − k ]}

– Time invariance implies

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Properties of LTI system

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

xn one sample Forward yn Backward

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Linear constant coefficients difference equations

– An Nth order linear const. difference eq. is represented by

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Frequency-domain representation of DTS

eigen-value eigen-function of the system

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

Ex: Ideal delay y [ n ] = x [ n − n0 ]

Since h [ n ] = δ [ n − n0 ] for an ideal delay system,

⇒ y [ n ] = A H ( e jω ) cos (ω 0 n + φ + θ ) , where θ =arg ⎡⎣ H ( e jω ) ⎤⎦

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

– Discrete-time LTI → periodicity of 2π

Ex: Ideal frequency selective filters (LPF)

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Representation of sequences by Fourier transform

= X R ( e jω ) + jX 1 ( e jω ) ← rectangular cord. form

– Note that Eq (A) is the inverse of Eq. (B)

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

If x [ n ] is absolutely summable (i.e., stable sequence), X ( e ) exists.

– Furthermore, x [ n ] can be showen to converge uniformly to a continuous function of ω .

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

Ex: Ideal LPF

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

– Theory of generalized functions

– Conjugate symmetric sequence: xe [ n ] = xe [ −n ]

– Conjugate anti-symmetric sequence: xo [ n ] = − xo [ −n ]

Note that X e ( e ) = X ( e ) and X o ( e ) = − X ( e )

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Fourier transform theorems

– Linearity: ax1 [ n ] + bx2 [ n ] ←⎯→ aX 1 ( e ) + bX 2 ( e )

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

⇒ Multiplication of two sequences is equivalent to periodic convolution of

420.461: Digital Signal Processing copyright@Yong-Hwan Lee

• Discrete-time Random Signals

– Consider an LTI system with impulse response h [ n ]

Note that only every Lth sample of u [ n ] is nonzero

If H(z) is causal with zeros at ck , H i ( z ) will be causal iff the Roc of H i ( z ) is

If H i ( z ) is stable, the RoC of H i ( z ) must include the unit circle, max

Example: A first-order IIR filter