Topic 3

The -function & convolution.

Impulse response & Transfer function
In this lecture we will described the mathematic operation of the convolution of
two continuous functions. As the name suggests, two functions are blended or
folded together.
We will then discuss the impulse response of a system, and show how it is related
to the transfer function of the system.
First though we will dene a special function called the -function or unit impulse.
It is, like the Heaviside step function u(t), a generalized function or distribution
and is best dened by considering another function in conjunction with it.
3.1 The -function
Consider a function
g(t) =
_
1/w 0 < t < w
0 otherwise
One thing of note about g(t) is that
_
w
0

g(t)dt = 1.
The lower limit 0

is a innitesimally small amount less than zero. Now, suppose

that the width w gets very small, indeed as small at 0
+
, an number an innitesimal
amount bigger than zero. At that point, g(t) has become like the function, a
very thin, very high spike at zero, such that
_

(t)dt =
_
0
+
0

(t)dt = 1
1
3/2
w 1/
w 0
g(t)
0
(t)
(a) (b)
Figure 3.1: As w becomes very small the function g(t) turns into a -function (t) indicated by
the arrowed spike.
In some sense it is akin to the derivative of the Heaviside unit step
_
t

(t)dt = u(t) .
More formally the delta function is dened in association with any arbitrary function
f (t), as
The delta function ...
_

f (t)(t)dt = f (0) .
Picking out values of a function in this way is called sifting of f (t) by (t). We
can also see that
_

f (t)(t )dt = f () ,
a result that we will return to.
0
(t)
f(t)
0
f(t)
(t )

t
(t )
0

t
A
(a) (b) (c)
Figure 3.2: (a,b) Sifting. (c) With an amplitude A.
Although the -function is innitely high, very often you will see a described as
the unit -function, or see a -function spike with an amplitude A by it. This is to
denote a delta-function where
_

(t)dt = 1 or
_

A(t)dt = A.
3/3
3.2 Properties of the -function
Fourier transform of the delta function:
FT [(t)] = 1
Proof: Use the denition of the -function and sift the function f (t) = e
it
:
_

(t)e
it
dt = e
i0
= 1 .
Symmetry: The -function has even symmetry.
(t) = (t)
Parameter Scaling:
(at) =
1
|a|
(t)
Proof: To prove this return to the fundamental denition,
_

f (t)(t)dt = f (0)
If a 0, substitute (at) for t (no swap in limits)
_

f (at)(at)d(at) = a
_

f (at)(at)dt = f (0)
But
_

f (at)(t)dt = f (0) a (at) = (t).

Now if a < 0, substitute (at) for t (but need to swap limits as a negative)
_

f (at)(at)d(at) = a
_

f (at)(at)dt = f (0)
But
_

f (at)(t)dt = f (0) a (at) = (t).

So |a| (at) = (t) covers both cases, and the stated denition follows immedi-
ately.
3/4
3.3 Fourier Transforms that involve the -function
Fourier Transform of e
i
0
t
FT
_
e
i
0
t

=
_

e
i(
0
)t
dt = 2(
0
) .
Fourier Transform of 1
FT [1] = 2() .
You could obtain this either by putting
0
= 0 just above, or by using the dual
property, FT [1] = 2(), then the even symmetry property () = ().
Fourier Transform of cos
0
t
FT [cos
0
t] = FT
_
1
2
_
e
i
0
t
+ e
i
0
t
_
_
= ((
0
) + ( +
0
)) .
Fourier Transform of sin
0
t
FT [cos
0
t] = FT
_
1
2i
_
e
i
0
t
e
i
0
t
_
_
= i ((
0
) ( +
0
)) .
Fourier Transform of Complex Fourier Series yes, this can be useful!
FT
_
n=

n=
C
n
e
in
0
t
_
= 2
n=

n=
C
n
( n
0
) .
3.4 Convolution
We turn now to a very important technique is signal analysis and processing. The
convolution of two functions f (t) and g(t) is denoted by f g. The convolution
is dened by an integral over the dummy variable .
The convolution integral. The value of f g at t is
(f g)(t) =
_

f ()g(t )d
3/5
The process is commutative, which means that
(f g)(t) (g f )(t) or
_

f ()g(t )d
_

f (t )g()d
3.4.1 Example 1
It is easier to see what is going on when convolving a signal f with a function g
of even or odd symmetry. However, to get into a strict routine, it is best to start
with an example with no symmetry.
[Q] Find and sketch the convolution of f (t) = u(t)e
at
with g(t) = u(t)e
bt
,
where both a and b are positive.
[A] Using the rst form of the convolution integral, the short answer must be
the unintelligible
f g =
_

u()e
a
u(t )e
b(t)
d .
First, make sketches of the functions f () and g(t ) as varies. Function f ()
looks just like f (t) of course. But g(t ) is a reected (time reversed) and
shifted version of g(t). (The reection is easy enough. To check that the shift
is correct, ask yourself where does the function g(p) drop? The answer is at
p = 0. So g(t ) must drop when t = 0, that is when = t.)
0

f( )
)

g(
0

)
g(
0

g(t )
t
(a) (b) (c) (d)
Figure 3.3:
We now multiply the two functions, BUT we must worry about the fact that t is
a variable. In this case there are two dierent regimes, one when t < 0 and the
other when t 0. Figure 3.4 shows the result.
So now to the integration. For t < 0, the function on the bottom left of Figure
3.4 is everywhere zero, and the result is zero. For t 0
_

u()e
a
u(t)e
b(t)
d = e
bt
_
t
0
e
(ba)
d =
e
bt
b a
_
e
(ba)t
1
_
3/6
f( ) g(t )
0

g(t ) f( )

t
t 0
t
<0
f( ) g(t )
0

t
g(t ) f( )

t >0
t 0
Figure 3.4:
So
f g(t) =
_ _
e
at
e
bt
_
/(b a) for t 0
0 for t < 0
It is important to realize that the function at the bottom right of Figure 3.4 is
NOT the convolution. That is the function you are about to integrate over for a
particular value of t. Figure 3.5 shows the t > 0 part of the convolution for b = 2
and a = 1.
0
0.05
0.1
0.15
0.2
0.25
0 1 2 3 4 5 6
(exp(-x)-exp(-2*x))
Figure 3.5: f g(t) plotted for t > 0 when b = 2 and a = 1.
3/7
3.4.2 Example 2
[Q] Derive an expression for the convolution of an arbitrary signal f (t) with the
function g(t) shown in the gure. Determine the convolution when f (t) = A, a
constant, and when f (t) = A + (B A)u(t).
[A] Follow the routine. Function f () looks exactly like f (t), but g(t ) is
reected and shifted. Multiply and integrate over from to . Because g
only has nite range, we can pinch in the limits of integration, and the convolution
becomes
f g =
_
t
ta
f ()d +
_
t+a
t
f ()d
f(t)
t
g(t)
t a
a
1
1
g(t )
f(t)
t
1
1
t+a
ta

t
Figure 3.6:
When f (t) = A, a constant, it is obvious by inspection that the convolution is zero
for all t. When f (t) = A + (B A)u(t), we have to be more careful because
there is a discontinuity in the function.
From Figure 3.7(a):
The convolution is zero for all t < a and all t > a (Diagram positions
1,2,5).
The maximum value is when t = 0 (Position 4). By inspection, or using the
integrals above, (f g)(t = 0) = a(B A).
For a < t < 0, (Position 3)
f g =
_
t
ta
(A)d+
_
0
t
Ad+
_
t+a
0
Bd = aA+tA+(t+a)B = (a+t)(BA)
showing that the increase in correlation value is linear. Symmetry tells us that
the decrease for 0 < t < a will also be linear.
One can see from Figure 3.7(b) that this convolution provides a rudimentary de-
tector of steps in the signal. The maximum in the correlation is proportion to the
step size, and would go negative if the step was downwards.
3/8
g(t )
4: t=0
f( ) f( ) g(t )
A
B
a a a 2a 0 0 0
1: 2: t=a a<t<0 3: 5: t>a

t<a
f
*
g
t
a a
a(BA)
(a) (b)
Figure 3.7:
3.5 The Impulse Response Function
You are very used to describing systems using the transfer function in the frequency
domain. So used to it in fact, that you may have forgotten to wonder why cant
we describe a systems response in the time domain? The answer is You can, but
you didnt know about convolution until now.
In the time domain, a system is described by its Impulse Response Function
h(t). This function literally describes the response of system at time t to an unit
impulse or -function input administered at time t = 0.
Suppose that now is time t, and you administered an impulse to the system at
time in the past. The response now is y(t) = h(t ).
t
0
Now
t
Whack
t
x(t)
Succession of whacks

Figure 3.8:
Suppose you administered a succession of impulses of dierent strengths x. The
portion dy(t) of the response due to impulse a time earlier is
dy(t) = x()h(t )d ,
so that the total response now is
y(t) =
_
t

x()h(t )d ,
where all t.
3/9
Now suppose that h is causal. If > t then h(t ) = 0, and therefore, provided
h(t) is causal,
the time response y(t) to an input x(t) is
y(t) =
_

x()h(t )d = x h .
The output is the input CONVOLVED with the Impulse Response
Function.
3.6 How do we connect this up ...
Can we reconcile the following things you now know about systems and signals?
1. The temporal output is the temporal input CONVOLVED with the Impulse
Response Function.
2. The frequency domain output is the frequency domain input MULTIPLIED
by the Transfer Function.
3. The frequency domain signal is the Fourier Transform of the temporal signal
Mathematically, it must be that the FT of a convolution is a product.
y(t) = x(t) h(t)

FT FT
? ?
Y () = X()H()
3.7 The Fourier Transform of a Convolution
To prove this we need to develop the integration for the Fourier Transform of a
convolution. Now x h is a perfectly respectable function of t, so
FT [x h] =
_

t=
[(x h)(t)] e
it
dt
=
_

t=
_

=
x()h(t )de
it
dt
3/10
For absolute clarity, lets switch the order of integration, then write t = +p. In
the inner integral is a constant, so that dt = dp.
FT [x h] =
_

=
_

t=
x()h(t )e
it
dtd
=
_

=
_

p=
x()h(p)e
i
e
ip
dpd
=
_

=
x()e
i
d
_

p=
h(p)e
ip
dp
= FT [x] FT [h]
= X() H()
So, the rst important thing we discover is
The time-convolution/frequency-modulation property of Fourier
transform
f (t) g(t) F()G()
or FT [f (t) g(t)] = FT [f (t)] FT [g(t)] = F()G()
3.8 Impulse response and Transfer Function
The second important connection is that
The Fourier Transform of the Impulse Response Function is the Trans-
fer Function
h(t) H()
or FT [h(t)] = H()
3/11
*h(t) x(t) y(t)
X( Y( ) ) ) H(
Figure 3.9:
3.9 The time-modulation/frequency-convolution property
There is a further result involving convolution that we state now.
The modulation/convolution property of Fourier transform
f (t)g(t)
1
2
F() G()
3.10 Example1
[Q1] Show that (t) f (t) = f (t).
[A1] One could do this the long way by writing
_

f ()(t )d, etc, but a

quicker way is to argue that
FT [(t) f (t)] = FT [(t)] FT [f (t)] = 1 . F()
FT
1
[FT [(t) f (t)]] = FT
1
[F()]
(t) f (t) = f (t)
We could also show that f (t) (t ) = f (t ) .
Both of these are manifestations of the
Sifting property of -functions
f (t) (t ) = f (t )
3/12
Previously this was expressed as an integral but we now recognize that integral
as the convolution integral!
3/13
3.11 More examples
[Q2] Convolve the two top hat functions f (t) and g(t) shown in below, and plot
the resulting convolved signal r (t). Find its Fourier Transform R().
1
a a
f(t)
2
a/2 a/2
g(t)
Figure 3.10:
[A2] Plotted below as functions of are f (), the reversed and shifted g(t )
for t = 3a/2, 3a/2 < t < a/2 and t = a/2, and the resulting f ()g(t ).
g(t )
1
a
f( )

t
a
g(t )
1
a a
f( )

t
g(t )
1
a a
f( )

t
a t+a/2
0 a
fg
fg
fg
Figure 3.11:
From the plots it is obvious that
for all t < 3a/2 f g is zero, so the convolution integral is zero.
For 3a/2 < t < a/2 the integral is
_
t+a/2
a
2d = 2
_
t +
3a
2
_
.
3/14
At t = a/2 there is complete overlap, at which point the integral is 2a.
It will continue constant at 2a until t = 0.
The convolution has even symmetry.
The result is given in Figure 3.12.
2a
0 3a/2 3a/2 a/2 a/2
g(t) f(t)
*
Figure 3.12:
To nd the Fourier transform of this convolution, we can use f g F()G().
Function f (t) is a top hat of height 1, but its width is 2a. We can use the
Parameter Scaling Property from Lecture 2 that says if
If f (t) F() then FT [f (t)] =
1
||
F
_

_
We have to be VERY careful. It is tempting to say that because our function is 2
times wider than the standard top hat, the scaling must be 2. WRONG!
The scaling is = 1/2. For the unit top-hat of half-width a/2
() =
2

sin
_
a
2
_
= a sinc
_
a
2
_
.
So for our parameter scaled version
F() =
1
|(1/2)|

_

(1/2)
_
= 2
2
2
sin
_
2a
2
_
=
2

sin (a) = 2a sinc

_
a

_
.
Function g(t) is a top hat of height 2 and half-width a/2. From Lecture 2, Section
2.4,
G() = 2() =
4

sin
_
a
2
_
= 2a sinc
_
a
2
_
.
So the Fourier Transform of the convolution is
FT [f g] =
8

2
sin (a) sin
_
a
2
_
= 4a
2
sinc
_
a
2
_
sinc
_
a

_
.
3/15
[Q3] A signal f (t) is amplitude modulated by cos
0
into a signal y(t). Find the
resulting Fourier Transform Y () in terms of F().
[A3] We could use the method given in Lecture 2 Amplitude modulation by a
cosine. Instead, lets use the property at Section 3.6 of this Lecture.
f (t)g(t)
1
2
F() G()
where our g(t) = cos
0
t. By writing
cos
0
t =
1
2
_
e
i
0
t
+ e
i
0
t
_
one can determine that
FT [cos
0
t] = (( +
0
) + (
0
)) .
(
0
) (+
0
)

0 0

Figure 3.13: Frequency spectrum of cos
0
t Hence
FT [f (t) cos
0
t] =
1
2
F() (( +
0
) + (
0
))
=
1
2
(F( +
0
) + F(
0
))
So suppose f (t) had a amplitude spectrum as sketched below. After amplitude
modulation, the amplitude spectrum would split into two parts.
Cosine
Modulation
F( )
0
F( )
0
0

0
Figure 3.14: An amplitude spectrum split into two by modulation with a cosine wave.
3/16
3.12 Summary
We started by developing the denition of the unit delta function (t). Although
this is an innitely high spike, it is also innitely narrow, so that the integral over
it is unity.
The delta function was more formally dened via its sifting properties.
We looked at properties of the delta function and at Fourier transforms that involve
the delta function
We then dened the convolution integral, and looked at how to lay out the
integral graphically.
We dened the Impulse Response function and showed that the temporal output
of a system is the temporal input convolved with the input.
Finally, by showing that the FT of a convolution of two temporal function is the
product of their individual FTs, we found that our old friend the Transfer Function
is the Fourier Transform of the Impulse Response.