Professional Documents
Culture Documents
As we have seen, any (sufficiently smooth) function f (t) that is periodic can be built out of sins and
coss. We have also seen that complex exponentials may be used in place of sins and coss. We shall now
use complex exponentials because they lead to less writing and simpler computations, but yet can easily be
converted into sins and coss. If f (t) has period 2, its (complex) Fourier series expansion is
Z
ik t
1
ck e
f (t) =
f (t)eik t dt
with
ck = 2
(1)
k=
ik t
ck eik t ei2k = ck eik t . We now develop an expansion for non-periodic functions, by allowing complex
exponentials (or equivalently sins and coss) of all possible periods, not just 2, for some fixed . So, from
now on, do not assume that f (t) is periodic.
For simplicity well only develop the expansions for functions that are zero for all sufficiently large |t|.
With a little more work, one can show that our conclusions apply to a much broader class of functions. Let
L > 0 be sufficiently large that f (t) = 0 for all |t| L. We can get a Fourier series expansion for the part
of f (t) with L < t < L by using the periodic extension trick. Define FL (t) to be the unique function
determined by the requirements that
i) FL (t) = f (t) for L < t L
ck (L)eik L t
where
ck (L) =
1
2L
k=
f (t)eik L t dt
(2)
If we can somehow take the limit L , we will get a representation of f that is is valid for all ts , not
just those in some finite interval L < t < L. This is exactly what we shall do, by the simple expedient
of interpreting the sum in (2) as a Riemann sum approximation to a certain integral. For each integer k,
and denote by = L
the spacing, k+1 k , between any two
define the k th frequency to be k = k L
R
successive frequencies. Also define f() = f (t)eit dt. Since f (t) = 0 for all |t| L
Z L
Z
Z
1
1
1
1
1
ck (L) = 2L
f (k ) = 2
f (k )
f (t)eik L t dt = 2L
f (t)ei(k L )t dt = 2L
f (t)eik t dt = 2L
L
In this notation,
f (t) = FL (t) =
1
2
f(k )eik t
(3)
k=
for any L < t < L. As we let L , the restriction L < t < L disappears, and the right hand
R
1
it
y=
0 1 2 3
February 25, 2007
1
it
2 f ()e
small slices of width and approximate, as in the above figure, the area under the part of the graph of
1
1
it
with between k and k + by the area of a rectangle of base and height 2
f (k )eik t .
2 f ()e
R
P
1
1
it
ik t
f ()e d
where
f () =
f (t)eit dt
(4)
f (t) = 2
The function f is called the Fourier transform of f . It is to be thought of as the frequency profile of the
signal f (t).
Example 1 Suppose that a signal gets turned on at t = 0 and then decays exponentially, so that
f (t) =
eat
0
if t 0
if t < 0
f (t)e
it
dt =
at it
dt =
t(a+i)
et(a+i)
1
dt =
=
(a + i) 0
a + i
Since a + i has modulus a2 + 2 and phase tan1 a , we have that f() = A()ei() with A() =
1
1
and () = tan1 a . The amplitude A() and phase () are plotted below.
|a+i| =
a2 + 2
A()
()
Example 2 Suppose that a signal consists of a single rectangular pulse of width 1 and height 1. Lets
say that it gets turned on at t = 21 and turned off at t = 12 . The standard name for this normalized
rectangular pulse is
1
1
1
rect(t) = 1 if 2 < t < 2
0 otherwise
t
1
21
2
It is also called a normalized boxcar function. The Fourier transform of this signal is
d
=
rect()
rect(t) e
it
dt =
1/2
e
1/2
it
1/2
ei/2 ei/2
eit
=
=
dt =
i 1/2
i
sin 2
2
d
when 6= 0. When = 0, rect(0)
=
R 1/2
dt = 1. By lH
opitals rule
1/2
d
lim rect()
= lim 2
1
cos 2
sin 2
d
= lim 2 2
= 1 = rect(0)
0
d
so rect()
is continuous at = 0. There is a standard function called sinc that is defined(1) by sinc =
sin
d
d
. In this notation rect() = sinc 2 . Here is a graph of rect().
1
d
rect()
2
h()
=
h(t)e
it
dt =
= f() +
g ()
f (t) + g(t) eit dt =
f (t)e
it
dt +
g(t)eit dt
(L)
Time Shifting
Suppose that we build a new function h(t) = f (t t0 ) by time shifting a function f (t) by t0 . The easy
way to check the direction of the shift is to note that if the original signal f (t) has a jump when its argument
t = a, then the new signal h(t) = f (t t0 ) has a jump when t t0 = a, which is at t = a + t0 , t0 units to
the right of the original jump.
t0
f (t)
h(t) = f (t t0 )
a + t0
h()
=
=e
it
h(t)e
it0
dt =
f (t t0 )e
it
f()
dt =
f (t )e
it
dt = e
it0
(T)
(1)
There are actually two different commonly used definitions. The first, which we shall use, is sinc =
sinc = sin
. It is sometimes called the normalized sinc function.
sin
.
The second is
Scaling
If we build a new function h(t) = f
of h is
h()
=
h(t)e
it
dt =
= f()
it
dt =
f (t )eit dt where t = t , dt = dt
(S)
Example 3 Now consider a signal that consists of a single rectangular pulse of height H, width W and
centre C.
y
H
y = rHWC (t)
W
2
C+
W
2
The function rHWC (t) = H rect tC
(with rect the normalized rectangular pulse of Example 2) has height
W
1
=
,
i.e.
t
=
C 12 W and so is the specified signal. By combining properties (L),
H and jumps when tC
W
2
(T) and (S), we can determine the Fourier transform of rHWC (t) = H rect tC
for any H, C and W . We
W
build it up in three steps.
First we consider the special case in which H = 1 and C = 0. Then we have R1 (t) = rect Wt . So,
according to the scaling property (S) with
f (t) = rect(t),
t
W
=f
t
W
=f
with = W
we have
1 () = h()
d
R
= f() = W rect(W
) = W W2 sin W2 =
tC
W
sin W2
1 () = W eiC rect
2 () = h()
d W
R
= eit0 f() = eiC R
tC
W
= HR2 (t) is
W
2 () = HW eiC rect(W
d
rHWC () = H R
) = HW eiC W2 sin W2 = eiC 2H
sin 2
Example 4
Now suppose that we have a signal that consists of a number of rectangular pulses of various heights
and widths. Here is an example
February 25, 2007
y
2
y = s(t)
t
2
1
0
1
2
3
4
We can write this signal as a sum of rectangular pulses. As we saw in the last example, rHWC (t) = H rect
is a single rectangular signal with height H, centre C and width W . So
s(t) = s1 (t) + s2 (t) + s3 (t) where
tC
W
H = 2, W = 1, C = 1.5 for n = 1
H = 1, W = 2, C = 1
for n = 2
H = 0.5, W = 2, C = 3
for n = 3
W
So, using (L) and rHWC () = eiC 2H
sin 2 , which we derived in Example 3,
s() = s1 () + s2 () + s3 () =
4 i 32
e
Differentiation
If we build a new function h(t) = f (t) by differentiating an old function f (t), then the Fourier transform
of h is
Z
Z
h() =
h(t)eit dt =
f (t)eit dt
h()
=
it
Z
u dv = uv
v du =
(D)
Example 5 The differentiation property is going to be useful when we use the Fourier transform to solve
differential equations. As an example, lets take another look at the RLC circuit
R
+
x(t)
y(t)
thinking of the voltage x(t) as an input signal and the voltage y(t) as an output signal. If we feed any input
signal x(t) into an RLC circuit, we get an output y(t) which obeys the differential equation
LCy (t) + RCy (t) + y(t) = x(t)
February 25, 2007
(5)
5
You should be able to quickly derive this equation, which is also (16) in the notes Fourier Series, on your
own. Take the Fourier transform of this whole equation and use that
the Fourier transform of y (t) is i y() and
the Fourier transform of y (t) is i times the Fourier transform of y (t) and so is 2 y()
So the Fourier transform of (5) is
LC 2 y() + iRC y() + y() = x
()
This is trivially solved for
y() =
x
()
LC 2 +iRC+1
(6)
which exhibits classic resonant behaviour. The circuit acts independently on each frequency component
of the signal. The amplitude |
y ()| of the frequency part of the output signal is the amplitude |
x()| of
1
1
the frequency part of the input signal multiplied by A() = |LC2 +iRC+1| =
.
2 2
2 2 2
(1LC ) +R C
We shall shortly learn how to convert (6) into an equation that expresses the time domain output signal y(t)
in terms of the time domain input signal x(t).
d
Example 6 The Fourier transform, rect(),
of the rectangular pulse function of Example 2 decays rather
1
slowly, like for large . We can try suppressing large frequencies by eliminating the discontinuities at
t = 21 in rect(t). For example
g(t) =
0
if t 58
5
5
3
4(t + 8 ) if 8 t 8
1
if 83 t
4( 58 t) if 38 t 85
0
if t 58
y
1
y = g(t)
3
8
85 83
3
8
5
8
It is not very difficult to evaluate the Fourier transform g() directly. But it easier to use properties (L)(D),
since
y
if t 85
3
5
if 58 t 38
4
8
8
3
3
= s4 (t) + s5 (t)
if 8 t 8
g (t) = 0
t
5
3
8 8
5
3
4 if 8 t 8
5
y = g (t)
0
if t 8
(
H = 4, W = 14 , C = 21 for n = 4
where sn (t) = rHWC (t) with
for n = 5
H = 4, W = 14 , C = 21
By Example 3, the Fourier transform of
s4 () + s5 () =
February 25, 2007
8 i/2
e
By Property (D), the Fourier transform of g (t) is i times g(). So the Fourier transform of g(t) is
the Fourier transform of g (t):
1
i
g() =
2i 8 sin 2 sin 8 =
16
2
1
i
times
sin 2 sin 8
This looks somewhat like the the Fourier transform in Example 2 but exhibits faster decay for large .
1
g()
Parsevals Relation
The energy carried by a signal f (t) is
Z
|f (t)|2 dt =
f (t)f (t) dt
Subbing in that
f (t) =
we have that
Z
This formula,
Duality
1
2
|f (t)|2 dt =
1
2
1
2
f()eit d =
1
2
f() eit d =
f (t)f()eit d dt =
f() f() d
|f (t)|2 dt =
1
2
1
2
1
2
1
2
f()eit d
Z
f()
f (t)eit dt d
|f()| d
2
The duality property says that if we build a new timedomain function g(t) = f(t) by exchanging the
roles of time and frequency, then the Fourier transform of g is
g() = 2f ()
(Du)
To verify this, just write down just the definition of g() and the Fourier inversion formula (4) for f (t) and,
in both integrals, make a change of variables so that the integration variable is s:
g() =
g(t)e
it
dt =
f(t)eit dt
f (t) =
1
2
t=s
f(s) eis ds
Z
f(s) eist ds
=s
f()eit d =
1
2
So g(), which is given by the last integral of the first line is exactly 2 times the last integral of the second
line with t replaced by , which is 2f ().
February 25, 2007
Example 7 In this example, we shall compute the Fourier transform of sinc(t) = sint t . Our starting point
is Example 2, where we saw that the Fourier transform of the rectangular pulse rect(t) of height one and
d
d
width one is rect()
= 2 sin 2 = sinc 2 . So, by duality (Du), with f (t) = rect(t) and g(t) = f(t) = rect(t),
t
the Fourier transform of g(t) = sinc 2 is g() = 2f () = 2 rect() = 2 rect(), since rect(t) is
even. So we now know that the Fourier transform of sinc 2t is 2 rect(t). To find the Fourier transform
of sinc(t), we just need to scale the 12 out of 2t . So we apply the scaling property (S) with f (t) = sinc 2t
and h(t) = sinc(t) = f (2t) = f t where = 21 . By (S), the Fourier transform of h(t) = sinc(t) is
f() = 2 rect() = rect 2 .
Multiplication and Convolution
A very common operation in signal processing is that of filtering. It is used to eliminate high frequency
noise and also to eliminate sixty cycle hum, arising from the ordinary household AC current. It is also
used to extract the signal from any one desired radio or TV station. In general, the filter is described
by a function H()
of frequency. For example you might have H()
= 1 for desired frequencies and
H()
= 0 for undesirable frequencies. When a signal f (t) is fed into the filter, an output signal g(t), whose
H()
= LC2 +iRC+1 .
The question what is g(t)? remains. Of course it is the inverse Fourier transform
g(t) =
1
2
it
f()H()e
d
of f()H(),
but we would like to express it more directly in terms of the original timedomain signal f (t).
R
So lets substitute f() = f ( )ei d (which expresses f() in terms f (t)) and see if we can simplify
the result.
Z Z
it
f ( )ei H()e
d d
g(t) = 1
2
=
=
Z
1
f ( ) 2
i(t )
H()e
d d
f ( )H(t ) d
f ( )H(t ) d
f (t )H( ) (d ) =
f (t )H( ) d
We conclude that
g() = f()H()
= g(t) = (f H)(t)
(C)
Example 8 The convolution integral gives a sort of moving weighted average, with the weighting determined
by H. Its value at time t involves the values of f for times near t. I say sort of because there is no
R
requirement that H(t) dt = 1 or even that H(t) 0.
H()
=
sin 2
The graph of this function was given in Example 2. It is a low pass filter in the sense that it lets through
most small frequencies and suppresses high frequencies. It is not really a practical filter, but I have chosen
it anyway because it is relatively easy to see the effect of this filter in tspace. From Example 2 we know
that
1
1
H(t) = 1 if 2 < t < 2
0 otherwise
So when this filter is applied to a signal f (t), the output at time t is
(f H)(t) =
=
f (t )H( ) d =
t+ 21
1
2
12
f (t ) d =
t 21
f ( ) d
t+ 21
where = t , d = d
f ( ) d
t 21
which is the area under the part of the graph of f ( ) from = t 21 to = t + 12 . To be concrete, suppose
that the input signal is
y
n
f ( ) = 1 if 0
0 if < 0
Then the output signal at time t is the area under the part of this graph from = t 12 to = t + 21 .
if t + 21 < 0, that is t < 21 , then f ( ) = 0 for all t 21 < < t + 21 (see the figure below) so that
(f H)(t) = 0
y
t 21
if t +
1
2
1
2
t+ 21
0 but t
0 < < t +
y = f ( )
1
2
t 21
if t
R t+ 21
t 21
1
2
y = f ( )
t+ 21
1
2
< < t +
1
2
and (f H)(t) =
1 d = 1.
y
R t+ 21
t 21
f ( ) d =
y = f ( )
t 21
t+ 21
0
(f H)(t) = t +
if t <
1
2
12
if 21 t
if t >
y = (f H)(t)
1
2
1
2
1
2
21
Example 9
For a similar, but more complicated example, we can through the procedure of Example 8 with H(t)
still being rect(t) but with f (t) replaced by the more complicated signal s(t) of Example 4. This time the
output signal at time t is
(s H)(t) =
s(t )H( ) d =
1
2
21
s(t ) d =
t+ 21
s( ) d
t 21
where = t
which is the area under the part of the graph of s( ) from = t 12 to = t + 21 . We can read off this
area from the figures below. In each figure, the left hand vertical dotted line is = t 12 and the right hand
vertical dotted line is = t + 12 . The first figure has t so small that the entire interval t 12 < < t + 12 is to
the left of = 2, where the graph of s( ) starts. Then, in each successive figure, we increase t, moving
the interval to the right. In the final figure t is large enough that the entire interval t 12 < < t + 21 is to
the right of = 4, where the graph of s( ) ends.
y
2
y
2
y = s( )
t 21
t+ 21
2 1
0
1
2
3
4
Case: t + 21 < 2, that is t < 25
area = 0
y
2
2 1
0
1
2
3
4
Case: 2 t + 12 1, that is 25 t 23
area = 2((t + 12 ) (2)) = 2t + 5
y
2
y = s( )
2 1
0
1
2
3
4
1
3
Case: 1 t + 2 0, that is 2 t 21
area = 2((1) (t 12 )) = 1 2t
2 1
0
1
2
Case: 1 t + 21 2, that is
area = 1
3
4
1
3
2
2
1
2
y
2
y = s( )
y = s( )
2 1
0
1
2
3
4
1
1
Case: 0 t + 2 1, that is 2 t
area = t + 21
y
2
y = s( )
2 1
0
1
2
3
4
Case: 2 t + 21 3, that is 23 t 52
area = [2 (t 12 ) + 21 [(t + 12 ) 2] = 74 21 t
10
y
2
y
2
2 1
0
1
2
1
Case: 3 t + 2 4, that is
area = 21
3
4
5
7
2 t 2
2 1
0
1
2
3
4
1
7
Case: 4 t + 2 5, that is 2 t 92
area = 12 [4 (t 21 )] = 49 12 t
y
2
y = s( )
1
0
1
2
3
Case: t 12 > 4, that is t >
t 21
t+ 21
9
2
area = 0
Impulses
y
y = (t)
t
the usual sense of the word. But it is possible to generalize the concept of a function (to something called
a distribution or a generalized function) so as to accommodate delta functions. One generalization involves
February 25, 2007
11
1
2
1
2
if < t <
y = (t)
otherwise
A treatment of these generalization procedures is well beyond the scope of this course. Fortunately, in practice
R
it suffices to be able to compute the value of the integral (t)f (t) dt for any continuous function f (t)
and that is easy. Because (t) = 0 for all t 6= 0, (t)f (t) is the same as (t)f (0). (Both are zero for t 6= 0.)
Because f (0) is a constant, the area under (t)f (0) is f (0) times the area under (t), which we already said
is 1. So
Z
(t)f (t) dt = f (0)
()
=
(t)eit dt = eit t=0 = 1
By the time shifting property (T), the Fourier transform of (t t0 ) (where t0 is a constant) is eit0 . We
may come to the same conclusion by first making the change of variables = t t0 to give
Z
Z
=tt0
it
( )ei( +t0 ) d
(t t0 )e
dt =
H()
in response to the impulse input signal (t) is indeed
Z
( )H(t ) d = H(t)
( H)(t) =
where we have applied
H()
=
1
LC 2 +iRC+1
To make the numbers work out cleanly, lets choose R = 1, L = 6 and C = 5. Then
H()
=
1
6 2 +i5+1
1
(3i+1)(2i+1)
eat
0
if t 0
if t < 0
f() =
1
a+i
(7)
So we can determine the impulse response function H(t) for the RLC filter just by using partial fractions to
1
s. Since
write H() as a linear combination of a+i
a(2i + 1) + b(3i + 1)
a
(2a + 3b)i + (a + b)
1
=
+ b =
=
(3i + 1)(2i + 1)
3i + 1 2i+1
(3i + 1)(2i + 1)
(3i + 1)(2i + 1)
2a + 3b = 0, a + b = 1 b = 1 a, 2a + 3(1 a) = 0 a = 3, b = 2
H()
=
12
H()
=
1
3
and a = 12 ,
3
2
=
3i + 1 2i + 1
1
3
+ i
1
2
1
+ i
H(t) =
et/3 et/2
0
if t 0
if t < 0
Example 10 (again) Here is a sneakier way to do the partial fraction expansion of Example 10. We know
that H()
has a partial fraction expansion of the form
H()
=
1
a
b
=
+
(3i + 1)(2i + 1)
3i + 1 2i + 1
The fast way to determine a is to multiply both sides of this equation by (3i + 1)
1
b(3i + 1)
=a+
2i + 1
2i + 1
and then evaluate both sides at 3i + 1 = 0. Then the b term becomes zero, because of the factor (3i + 1)
in the numerator and we get
1
1
a=
= 1 =3
2i + 1 i= 1
3
3
1
1
1
b=
=
= 1 = 2
3i + 1 2i+1=0
3i + 1 i= 1
2
2
13
Signal
x(t) =
1
2
y(t) =
1
2
Fourier Transform
R
x() = x(t)eit dt
R
y() = y(t)eit dt
x()eit d
y()eit d
Linearity
Ax(t) + By(t)
A
x() + B y()
Time shifting
x(t t0 )
()
eit0 x
Frequency shifting
Scaling
Time shift & scaling
ei0 t x(t)
x t
x
x( 0 )
||
x()
tt0
)
i0 t
||eit0 x()
0
x
||e
Conjugation
x(t)
x
()
Time reversal
x(t)
x()
tDifferentiation
x(t)
x (t)
i x
()
x(n) (t)
(i)n x()
Differentiation
tx(t)
Convolution
tn x(t)
R
x( )y(t ) d
i dd x()
n
d
i d
x
()
Multiplication
x(t)y(t)
Duality
x
(t)
x()
y ()
R
1
()
y ( ) d
2 x
2x()
R
R
1
|x(t)|2 dt = 2
|
x()|2 d
Parseval
n
0
if t < 0,
eat if t > 0
at
if t < 0,
at
e u(t) = e
0
if t > 0
eat u(t) =
ea|t|
(
Boxcar in time
rect(t) =
General boxcar
rHWC (t) =
Boxcar in frequency
1
2
Impulse in time
(t t0 )
Single frequency
sinc
(t t0 ) x(t)
ei0 t
if |t| <
1
2
1
2
0 if |t| >
(
H if |t C| <
t
if |t C| >
1
t
sin
t
2
1
(a constant, Re a > 0)
a + i
1
(a constant, Re a < 0)
a + i
2a
(a constant, Re a > 0)
2
a + 2
sinc
W
2
W
2
sin
HW eiC sinc
rect() =
1
0
W
2
W
= eiC 2H
sin 2
if || < 1/2
if || > 1/2
eit0
eit0 x(t0 )
2 ( 0 )
14