Distortion

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Stanford University

Winter 2009-2010

Signal Processing and Linear Systems I

Lecture 11: Frequency Response of LTI Systems

February 7, 2011

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 1

Linear Time-Invariant Systems, Revisited

• A linear time-invariant system is completely characterized by its impulse


response h(t).

• For a linear system with an input signal x(t), the output is given by the
convolution
� ∞
y(t) = (x ∗ h)(t) = x(τ )h(t − τ ) dτ
−∞

x(t) y(t)
∗h(t)

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 2


• The Fourier transform of the convolution is

Y (jω) = H(jω)X(jω)

where X(jω) is the input spectrum, Y (jω) is the output spectrum, and
H(jω) is the Fourier transform of the impulse response h(t).

• H(jω) is called the frequency response or transfer function of the system.


Each frequency in the input spectrum X(jω) is
– Scaled by the system amplitude response |H(jω)|,

|Y (jω)| = |H(jω)||X(jω)|

– Phase shifted by the system phase response� H(jω),

� Y (jω) = � H(jω) + � X(jω)

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 3

to produce the output spectrum Y (jω).

• If the input is to a system is a complex exponential ejω0t, the input


spectrum is
� �
X(jω) = F ejω0t
= 2πδ(ω − ω0).

The output spectrum is

Y (jω) = H(jω)(2πδ(ω − ω0))


= H(jω0)(2πδ(ω − ω0)).

The ouput signal is

y(t) = F −1 [Y (jω)]
= F −1 [H(jω0)(2πδ(ω − ω0))]

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 4


= H(jω0)ejω0t
� H(jω0 ))
= |H(jω0)|ej(ω0t+

A sinusoidal input ejω0t to an LTI system produces a sinusoidal output


at the
– Same frequency,
– Scaled in amplitude, and
– Phase shifted.
This corresponds to multiplication by a complex number H(jω0).

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 5

Frequency Response Example

An input signal

x(t) = 2 cos(t) + 3 cos (3t/2) + cos(2t)

is applied to a system with an impulse response h(t)

2
h(t) = sinc2(t/π)
π

Find the output signal (x ∗ h)(t).

First, the frequency response or transfer function of the system is


� � � �
2 2
F sinc (t/π) =
2
π∆(πω/2π) = 2∆(ω/2)
π π

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 6


so
H(jω) = 2∆(ω/2)
The input spectrum is

X(jω) = 2π [δ(ω − 1) + δ(ω + 1)] + 3π [δ(ω − 3/2) + δ(ω + 3/2)]


+π [δ(ω − 2) + δ(ω + 2)]

The output spectrum is the product of the input spectrum, and the transfer
function, as shown on the next page:

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 7

X( j!)
4!
3! 3!
2! 2! 2!
! !
−2 −1 0 1 2 !
×
H( j!) = 2"(!/2)
2

−2 −1 0 1 2 !

=
Y ( j!) = H( j!)X( j!)
4!

3!/2 2! 2! 2!
3!/2
−2 −1 0 1 2 !

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 8


The output signal spectrum is then


Y (jω) = 2π [δ(ω − 1) + δ(ω + 1)] + [δ(ω − 3/2) + δ(ω + 3/2)]
2

and the output signal is

3
y(t) = 2 cos(t) + cos(3t/2).
2

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 9

Another Example

A signal
x(t) = e−tu(t)
is applied to a zero-state system with an impulse response

h(t) = 2e−2tu(t)

What is the spectrum of the output Y (jω) = H(jω)X(jω), and the output
signal y(t)?

The Fourier transform of the input is

� � 1
X(jω) = F e−tu(t) = .
1 + jω

The transfer function or frequency response is the Fourier transform of the

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 10


impulse response

� � 2
H(jω) = F 2e−2tu(t) = .
2 + jω

The Fourier transform of the output is then

2
Y (jω) = H(jω)X(jω) =
(1 + jω)(2 + jω)

This is illustrated on the next page:

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 11

1.5
1 !
0.5
|X( j!)| 0 0 ∠X( j!)
!0.5
!1 −!
!5 !4 !3 !2 !1 0 1 2 3 4 5
!
1.5
1
!
0.5
|H( j!)| 0 0 ∠H( j!)
!0.5
!1 −!
!5 !4 !3 !2 !1 0 1 2 3 4 5
!
1.5
1
!
0.5
|Y ( j!)| 0
0 ∠Y ( j!)
!0.5
!1
−!
!5 !4 !3 !2 !1 0 1 2 3 4 5
!

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 12


Note that the magnitude profiles are multiplied (Y (jω) is narrower than
either X(jω) or H(jω)) and that the phase profiles add.

We can find the time signal y(t) by noting that

2 2 2
Y (jω) = = −
(1 + jω)(2 + jω) 1 + jω 2 + jω

which you can check. Then

y(t) = 2(e−t − e−2t)u(t).

The last two steps we’ll cover in the section on Laplace transforms, later in
the quarter.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 13

Ideal Filters

There are several basic filter types we will encounter. Some of these are:
Ideal Lowpass:
H( j!)

−!c !c !

This suppresses all frequencies about a cutoff frequency ωc. It will be


important for reconstructing a continuous waveform from its samples.
Ideal Highpass:

H( j!)

−!c !c !

This suppresses all frequencies below ωc.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 14


Ideal Bandpass

2!c 2!c H( j!)

−!0 !0 !

This passes a band of frequencies. The bands are of width 2ωc, and are
centered at ±ω0. This is useful in communications, were we want to select
for a specific frequency range.

Each of these can be implemented as a convolution. The impulse response


of the filter is the inverse Fourier transform of the frequency response.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 15

Ideal Lowpass

The ideal lowpass


H( j!)

−!c !c !

can be written as
H(jω) = rect(ω/(2ωc))
The impulse response is then
�� � ��
� ��
2ωc 2π 2π � ω �
F −1
{rect(ω/(2ωc))} = F −1
rect
2π 2ωc 2ωc 2π
�ω � �ω �
c c
= sinc t
π π

This is plotted below:

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 16


�! � !c
!c c 1
sinc t "
" " ∼
t

2! ! ! 2! t
− −
"c "c "c "c

This has several practical problems. The impulse response is

• Non-causal: we’ll address this in a few slides.

• Infinite in duration: must be truncated somewhere.

• Decays very slowly, as 1/t.

Since we are going to have to truncate the impulse response, we’d like it to
decay as fast as possible, so that we minimize its length. We can do this by
making the response smoother.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 17

For example, this more practical filter


H( j!)

−2!c −!c !c 2!c !

has the same passband, but allows for a transition band of width ωc. This
can be written as a convolution of two rect’s,

H(jω) = rect(ω/3ωc) ∗ rect(ω/ωc)

(Convince yourself this is true!) The impulse response is then


� � �� � � ω ��
3ωc 3ωc ωc c
h(t) = sinc t sinc t
2π 2π 2π 2π

This will decay as 1/t2, which is much faster. Smoother transition bands
will result in faster decay, and even shorter impulse responses for a given
truncation error.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 18


Ideal Highpass Filter

The ideal highpass

H( j!)

−!c !c !

can be written as 1 minus the ideal lowpass

H(jω) = 1 − rect(ω/(2ωc))

The impulse response is then


�ω � �ω �
c c
h(t) = δ(t) − sinc t
π π

This is of course also not practical. It is plotted below:

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 19

!c
!(t) "c �" �
�! � " c
!c c !(t) − sinc t
− sinc t # #
" "

2! ! ! 2! t
− −
"c "c !c "c "c

"

The comments about the ideal lowpass apply.

In addition, the impulse must be approximated. If we do this by limiting


the range of frequencies, we get the next filter.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 20


Ideal Bandpass

The ideal bandpass

2!c 2!c H( j!)

−!0 !0 !

can be considered an ideal lowpass filter that has been modulated to ω0.
The frequency response can be written
� � � �
ω + ω0 ω − ω0
H(jω) = rect + rect
2ωc 2ωc

By the modulation theorem, the impulse response is


�ω � �ω �
c c
h(t) = (2 cos(ω0t)) sinc t
π π
EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 21

The impulse response of a bandpass filter is simply a lowpass filter that has
been multiplied by a cosine!

!c �! � �! �
c !c
2 sinc t (2 cos(!0t)) sinc
c
t
" " " "

2! ! ! 2! t
− −
"c "c "c "c
!c �! �
c
−2
− sinc t
" "

Again, the bandpass filter would be easier to implement if it were based on


a more practical lowpass filter.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 22


Distortionless LTI Systems

The frequency response of a system includes

• Amplitude scaling by |H(jω)|

• Phase shift by � H(jω)

Often we would like a system to pass a signal without distortion

y(t) = Kx(t − td)

where K is some constant gain, and td is a constant delay. We get the


same signal out, amplified and delayed.
The Fourier transform of such a system is

Y (jω) = Ke−jωtd X(jω)

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 23

so the frequency response is

H(jω) = Ke−jωtd

or

|H(jω)| = K
� H(jω) = −ωtd

Hence, a distortionless system has a negative, linear phase as a function of


frequency.
K
|H( j!)| ∠H( j!)

! !

−!td

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 24


Practical Implementation of Non-Causal Filters

The ideal filters are not causal, so they can’t be implemented. In practice
they must be truncated and delayed. An ideal lowpass, and a truncated and
delayed lowpass are plotted below:

Ideal, Non-Causal Lowpass

td

Practical, Causal Lowpass

The result is a filter with the ideal frequency response multiplied by a


negative linear phase. This approximates a distortionless system over the
passband of the filter.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 25

Amplitude and Phase Distortion

So far we have mostly been concerned with the system amplitude response,
as in the communication example (Lecture 6, page 11):
1.5

1
Input, x(t)

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s
1.5
Impulse Response, h(t)

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s
1.5

1
Output, y(t)

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 26


The impulse response of this system is a shifted real, symmetric function,
so the phase is linear. Only amplitude scaling of the input spectrum is
occurring here.

The frequency response of a system can affect both the amplitude and
phase of the output signal.

Both of these are important, depending on the application.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 27

Phase Distortion and Group Delay

A distortionless system has a negative linear phase

� H(jω) = −ωtd

If the phase is not linear, the time delay is no longer a constant, and is a
function of frequency.

The time delay is the negative slope of � H(jω),

d
td(ω) = − � H(jω).

This frequency dependent delay is the group delay.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 28


∠H( j!) ∠H( j!)

! !

d
td (!) = − ∠H(!)
d!

Here high frequencies have a smaller delay.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 29

The effect of frequency dependent delays are shown below. The same input
(top) is applied to a linear phase system (middle), and system with linear
phase plus a small third order term (lower).
1.5

1
Input

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s
1.5
Delayed Output

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s
Phase Distorted Output

1.5

0.5

!0.5
0 1 2 3 4 5 6 7 8 9 10
Time, s

High and low frequencies aren’t aligned, and the transition gets broader.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 30


Importance of amplitude and phase distortion depends on application.

For audio or speech:

• Amplitude distortion is very is important.

• Humans are relatively insensitive to phase distortion.

For images or video:

• Amplitude distortion is relatively unimportant, as long as it is slowly


varying.

• Phase distortion is very important. Small amounts of non-linear phase


result in very blurry looking images.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 31

Example from MRI


Small additional quadratic phase produces substantial blurring. These have
zero, 1/2, and 1 cycle of additional quadratic phase. This is less than 1%
of the linear phase.

EE102A:Signal Processing and Linear Systems I; Win 09-10, Pauly 32

You might also like