S.

Boyd

EE102

Lecture 8
Transfer functions and convolution

convolution & transfer functions

properties
examples
interpretation of convolution
representation of linear time-invariant systems

81

Convolution systems
convolution system with input u (u(t) = 0, t < 0) and output y:
y(t) =

h( )u(t ) d =
0

h(t )u( ) d
0

abbreviated: y = h u
in the frequency domain: Y (s) = H(s)U (s)
H is called the transfer function (TF) of the system
h is called the impulse response of the system

PSfrag replacements

block diagram notation(s):

u

Transfer functions and convolution

y
h

y
H

82

Properties
1. convolution systems are linear: for all signals u1, u2 and all , R,
h (u1 + u2) = (h u1) + (h u2)
2. convolution systems are causal: the output y(t) at time t depends only
on past inputs u( ), 0 t
3. convolution systems are time-invariant: if we shift the input signal u
over T > 0, i.e., apply the input

0
t<T
u
e(t) =
u(t T ) t 0

to the system, the output is

ye(t) =

0
t<T
y(t T ) t 0

in other words: convolution systems commute with delay

Transfer functions and convolution

83

4. composition of convolution systems corresponds to

multiplication of transfer functions
convolution of impulse responses
composition
PSfrag replacements
u

BA

ramifications:
can manipulate block diagrams with transfer functions as if they were
simple gains
convolution systems commute with each other
Transfer functions and convolution

84

Example: feedback connection

PSfrag replacements
u

in time domain, we have complicated integral equation

y(t) =

g(t )(u( ) y( )) d
0

which is not easy to understand or solve . . .

in frequency domain, we have Y = G(U Y ); solve for Y to get
Y (s) = H(s)U (s),

H(s) =

G(s)
1 + G(s)

(as if G were a simple scaling system!)

Transfer functions and convolution

85

General examples
first order LCCODE: y 0 + y = u, y(0) = 0
take Laplace transform to get
1
U (s)
Y (s) =
s+1
transfer function is 1/(s + 1); impulse response is et
Z t
integrator: y(t) =
u( ) d
0

transfer function is 1/s; impulse response is 1

delay: with T 0,
y(t) =

0
t<T
u(t T ) t T

impulse response is (t T ); transfer function is esT

Transfer functions and convolution

86

Vehicle suspension system

(simple model of) vehicle suspension system:
y
PSfrag replacements

b
u

input u is road height (along vehicle path); output y is vehicle height

vehicle dynamics: my 00 + by 0 + ky = bu0 + ku
assuming y(0) = 0, y 0(0) = 0, (and u(0) = 0),
(ms2 + bs + k)Y = (bs + k)U
TF from road height to vehicle height is H(s) =
Transfer functions and convolution

bs + k
ms2 + bs + k
87

DC motor
PSfrag replacements

J00 + b0 = ki
(J is rotational inertia of shaft & load; b is mechanical resistance of shaft
& load; k is motor constant)
assuming (0) = 0(0) = 0,
Js2(s) + bs(s) = kI(s),

(s) =

k
I(s)
2
Js + bs

i.e., transfer function H from i to is

H(s) =
Transfer functions and convolution

k
Js2 + bs
88

Circuit examples
consider a circuit with linear elements, zero initial conditions for inductors
and capacitors,
one independent source with value u
y is a voltage or current somewhere in the circuit
then we have Y (s) = H(s)U (s)
example: RC circuit

PSfrag replacements

RCy 0(t) + y(t) = u(t),

impulse response is L1
Transfer functions and convolution

1
1 + sRC

Y (s) =
=

1
U (s)
1 + sRC

1 t/RC
e
RC
89

to find H: write circuit equations in frequency domain:

resistor: v(t) = Ri(t) becomes V (s) = RI(s)
capacitor: i(t) = Cv 0(t) becomes I(s) = sCV (s)
inductor: v(t) = Li0(t) becomes V (s) = sLI(s)
in frequency domain, circuit equations become algebraic equations

Transfer functions and convolution

810

1F

example:
1

PSfrag replacements

v
1F

vout
v+

vin
1

lets find TF from vin to vout (assuming zero initial voltages for capacitors)
1
s
= Vin
by voltage divider rule, V+ = Vin
1 + 1/s
s+1
current in lefthand resistor is (using V = V+):
I=
Transfer functions and convolution

Vin V
=
1

s
1
Vin =
Vin
s+1
s+1
811

 I flows through 1Fk1, yielding voltage

1 (1)(1/s)
1
= Vin
Vin
s + 1 1 + 1/s
(s + 1)2
finally we have Vout

1
s2 + s 1
= V Vin
= Vin
2
(s + 1)
(s + 1)2

so transfer function is
1
1
s2 + s 1

=
1

H(s) =
(s + 1)2
s + 1 (s + 1)2
impulse response is
h(t) = L1(H) = (t) et tet
we have
vout(t) = vin(t)
Transfer functions and convolution

(1 + )e vin(t ) d
0
812

Interpretation of convolution

y(t) =

h( )u(t ) d
0

y(t) is current output; u(t ) is what the input was seconds ago
h( ) shows how much current output depends on what input was
seconds ago
for example,
h(21) big means current output depends quite a bit on what input was,
21sec ago
if h( ) is small for > 3, then y(t) depends mostly on what the input
has been over the last 3 seconds
h( ) 0 as means y(t) depends less and less on remote past
input
Transfer functions and convolution

813

Graphical interpretation

y(t) =

h(t )u( ) d
0

to find y(t):
flip impulse response h( ) backwards in time (yields h( ))
drag to the right over t (yields h(t ))
multiply pointwise by u (yields u( )h(t ))
integrate over to get y(t)

Transfer functions and convolution

814

h( )

h( )

u( )

u( )

h(t1 )

h(t2 )

placements

t1

t2

h(t3 )
y =uh

t3

Transfer functions and convolution

815

Example
communication channel, e.g., twisted pair cable
PSfrag replacements

impulse response:
1.5

0.5

PSfrag replacements
0
0

10

a delay 1, plus smoothing

Transfer functions and convolution

816

simple signalling at 0.5 bit/sec; Boolean signal 0, 1, 0, 1, 1, . . .

1
0.5
0
0

10

10

PSfrag replacements

0.5
0
0

output is delayed, smoothed version of input

1s & 0s easily distinguished in y
Transfer functions and convolution

817

simple signalling at 4 bit/sec; same Boolean signal

1
0.5
0
0

10

10

PSfrag replacements

0.5
0
0

smoothing makes 1s & 0s very hard to distinguish in y

Transfer functions and convolution

818

Linear time-invariant systems

consider a system A which is
linear
time-invariant (commutes with delays)
causal (y(t) depends only on u( ) for 0 t)
called a linear time-invariant (LTI) causal system
we have seen that any convolution system is LTI and causal; the converse
is also true: any LTI causal system can be represented by a convolution
system
convolution/transfer function representation gives universal description for
LTI causal systems
(precise statement & proof is not simple . . . )
Transfer functions and convolution

819