EE518 Digital Signal Processing University of Washington

Autumn 2001 Dept. of Electrical Engineering

Lecture 10 B: Freq. Response of LTI systems, Phase & Group Delay, Inverse Systems
Oct 31, 2001

Prof: J. Bilmes <[email protected]> TA: Mingzhou Song <[email protected]>

10B.1 Frequency Response of LTI Systems

An LTI system is characterized by

y[n] = x[n] ∗ h[n]

Y (e jω ) = X(e jω )H(e jω )

Magnitude and phase are changed separately.

|Y (e jω )| = |X(e jω )||H(e jω )|
]Y (e jω ) = ]X(e jω ) + ]H(e jω )

Note

H(e jω ) =|H(e jω )|e jθ(ω)

]H(e jω ) =θ(ω) = arg[H(e jω )]

The magnitude response |H(e jω )| and phase response ]H(e jω ) are all real functions of ω, unlike H(e jω ) which is a
complex function. Affects of θ(ω) only to ??(p1) a multiple of 2π, since e j· is periodic in 2π.
Sometime the magnitude response is measured in dB, i.e.

20 log10 |H(e jω )|dB

Ideal Filters
|ω| < ωc

1
Hl p (e jω ) =
0 else
Correspondingly,
sin ωc n
hl p [n] = −∞ < n < ∞
πn
Hhp (e jω ) = 1 − Hl p (e jω )
Since
F
δ[n] ←→ 1
sin ωc n
hhp [n] = 1 − −∞ < n < ∞
πn

How to design these on computer? Mostly chapter 7 stuff. Basic idea is we need to determine from ideal magnitude
and phase response to produce realizable filters, i.e., need to allow for types of distortion.
Note above filters are all non-causal, how can we account for this in some cases?

10B-1
 10B-2

10B.2 Phase Distortion and Delay

hid [n] = δ[n] = ideal delay system

Hid (e jω ) = e− jωnd
|Hid (e jω )| = 1
]Hid (e jω ) = −ωnd |ω| < π (periodic with period 2π)

This is phase distortion. But is it bad? This is nice distortion. Delay other parts of system. Ok if don’t need real-time
performance.
Ex: ideal LPF with delay distortion.

e− jωnd |ω| < ωc


Hl p (e jω ) =
0 else

Correspondingly,
sin ωc (n − nd )
hl p [n] = −∞ < n < ∞
π(n − nd )
Definition 10.1 (Linear Phase). Phase is linear in frequency ω.

]Hid (e jω ) = −ωnd (10B.1)

When
]Hid (e jω ) = −φ0 − ωnd
It is also considered “linear” (it is affine actually, but still easy to deal with.)
So linear phase is “good”, how can we tell?
Definition 10.2 (Phase Delay).
θ(ω) arg[H(e jω )]
τ p (ω) = − =− (10B.2)
ω ω

Time delay is caused by phase in certain cases.

Definition 10.3 (Group Delay).
d d
τg (ω) = grd[H(e jω )] = − {arg[H(e jω )]} = − θ(ω) (10B.3)
dω dω
Group delay indicates the degree of nonlinearity of the phase response (also the delay of a narrowband group of
frequencies.)

Ex: The input of an LTI system is

1 1
x[n] = A cos(ω0 n + φ) = Ae j(ω0 n+φ) + e− j(ω0 n+φ)
2 2
Then the output is (by eigenfunction property)
1 1
y[n] = AH(e jω0 )e j(ω0 n+φ) + AH(e jω0 )e− j(ω0 n+φ) = A|H(e jω0 )| cos(ω0 n + θ(ω0 ) + φ)
2 2
The magnitude of the output is scaled by |H(e jω0 )| and the phase is changed by θ(ω0 ).
Note
θ(ω0 )
  
lag = −θ(ω0 ), y[n] = A|H(e jω0 )| cos(ω0 n − − + φ)
ω0
 10B-3

So
θ(ω0 )
τ p (ω0 ) = −
ω0
is phase delay at ω = ω0 .
Note: each frequency of sinusoid could undergo a different delay, so output signal might not look very much like
input.
Ex: consider frequency response of some signal. We can think of this as being a sum of modulated chunks. To better
understand group delay, let’s take one of these chunks and analyze them. Consider sinusoidal envelope and sinusoid:
x[n] = A cos(ω0 n) cos(ωc n) ωc >> ω0
x[n] is known as envelope signal and cos(ωc n) is known as carrier signal. Using cosine identities
A A
x[n] = cos(ωl n) + cos(ωu n)
2 2
where
ωl = ωc − ω0 ωu = ωc + ω0

If we send x[n] through an LTI system with |H(e jω )| = 1, then

A A
y[n] = cos(ωl n + θ(ωl )) + cos(ωu n + θ(ωu ))
2 2
θ(ωu ) + θ(ωl ) θ(ωu ) − θ(ωl )
=A cos(ωc n + ) cos(ω0 n + )
2 2
The output is the same form as the input, but each component has potentially different delays. To get an idea, if
ωc >> ω0 and ω0 is small, we can look at phase delay in terms of θc (ωc ).
Take Taylor expansion, we get
 
d
θc (ω) ≈ θc (ωc ) + θc (ω)|ω=ωc (ω − ωc ) + ε
dω
and let  
d
θ̂c (ω) = θc (ωc ) + θc (ω)|ω=ωc (ω − ωc )
dω
which is a first order approximation for θc (ω).
The phase delay at ω = ωc is
θ(ωu ) + θ(ωl ) θ(ωc )
− ≈−
2ωc ωc
Using θ̂c (ω), phase delay at ωc of carrier is phase delay as if that signal alone occurred.
θ(ωu ) − θ(ωl ) d
− ≈ − θc (ω)|ω=ωc = −θ0c (ωc ) = grd[H(e jω )]|ω=ωc
2ω0 dω
This is the group delay. So base signal is delayed in time by group delay.
Note: if θ(ω) is in radian and ω in rads/sec, −θ(ω)/ω is in units of seconds. And
dθ(ω)
τg (ω) = −
dω
is also in units of seconds. An example of the input and output is shown in the Fig 10B.1
In general, group delay gives delay in time (seconds, or can be converted to samples) of a group of neighboring close
frequencies.
Note: above example, the carrier signal (at ωc ) is delayed by τ p (ωc ), but note that x[n] doesn’t contain a frequency
component of ωc .
Ex: more group delay
See O&S Example 5.1 (p 245)
 10B-4

x[n]
1

0.5

−0.5

−1
0 20 40 60 80 100 120 140 160 180 200

y[n], τ =20 (samples), τ =10 (samples)

g p
1

0.5

−0.5

−1
0 20 40 60 80 100 120 140 160 180 200
n

Figure 10B.1: Group delay and phase delay of amplitude modulated signal.

10B.3 System Functions for Linear Constant Coefficient Difference Equa-

tions

Take z-transform of the linear constant coefficient difference equation

N M
∑ ak y[n − k] = ∑ bk x[n − k]
k=0 k=0

we get
N M
∑ akY (z)z−k = ∑ bk X(z)z−k
k=0 k=0
or
Y (z) ∑M bk z−k
H(z) = = k=0
X(z) ∑Nk=0 ak z−k

Note: poles come from recursively using y[n] in previous output, i.e., this is what gives the potential for instability.
H(z) can be factored as
b0 ∏M −1
k=0 (1 − ck z )
H(z) =
a0 ∏Nk=0 (1 − dk z−1 )

There exist zeros at z = ck and poles at z = dk . Each 1 − ck z−1 gives a zero at z = ck and a pole at z = 0. Each 1 − dk z−1
gives a pole at z = dk and a zero at z = 0.
How many poles and zeros?
b0 zN−M ∏M
k=0 (z − ck )
H(z) =
a0 ∏Nk=0 (z − dk )
So the number of zeros and poles are the same.
Stability and Causality
 10B-5

We need to specify this in addition to H(z) because there exist multiple possible impulse responses for a given H(z).
Causality requires ROC be outside outermost pole and where that pole lives determines the stability, since stability
needs
∞
∑ |h[n]| < ∞
n=−∞

Ex:
1 1
H(z) = 5 −1
= 1 −1 −1
1− 2z + z−2 (1 − 2 z )(1 − 2z )

There are three possible ROC’s. Stability requires an ROC of 1/2 < |z| < 2 which gives an a two sided sequence h[n].
Causality requires an ROC of |z| > 2 which gives a non-stable system.

10B.4 Inverse Systems

Given H(z), can we find Hi (z) such that

H(z)Hi (z) = 1, i.e.,Hi (z) = 1/H(z) and Hi (e jω ) = 1/H(e jω )

or
h[n] ∗ hi [n] = δ[n]
For
b0 ∏M −1
k=0 (1 − ck z )
H(z) =
a0 ∏k=0 (1 − dk z−1 )
N

the inverse is
a0 ∏Nk=0 (1 − dk z−1 )
Hi (z) =
b0 ∏M −1
k=0 (1 − ck z )

This means we can cascade the inverse system to the output of the original system and get an overall output which is
the same as input x[n], as shown below,

x[n] −→ [ H(z) ] −→ [ Hi (z) ] −→ x[n]

But can we do this always?

The answer is no. Because H(z) might have zeros on the unit circle and can’t recover what has been removed, e.g.,
the ideal lowpass filter. But for rational systems, many times we can recover.
What about ROC? if H(z) has an ROC, what is the ROC of Hi (z)?
The key idea is ROC’s must overlay. Any ROC for Hi (z) that overlaps H(z)’s ROC is ok.
Ex:
If
z−1 − 0.5
H(z) = |z| > 0.9
1 − 0.9z−1
then,
−2 + 1.8z−1
Hi (z) =
1 − 2z−1
There are two possible ROC’s, i.e., |z| > 2 and |z| < 2, and both overlap with the ROC of H(z), i.e., |z| > 0.9. But they
have different impulse responses.
For |z| < 2, the inverse system is stable but non-causal and has the impulse response

hi1 [n] = 2 · 2n u[−n − 1] − 1.8 · 2n−1 u[−n]

 10B-6

For |z| > 2, the inverse system is non-stable but causal and has the impulse response

hi2 [n] = −2 · 2n u[n] + 1.8 · 2n−1 u[n − 1]