Lecture VIII: Fourier series

Maxim Raginsky
BME 171: Signals and Systems
Duke University
September 19, 2008
Maxim Raginsky Lecture VIII: Fourier series
This lecture
Plan for the lecture:
1
Review of vectors and vector spaces
2
Vector space of continuous-time signals
3
Vector space of T-periodic signals
4
Complete orthonormal systems of functions
5
Trigonometric Fourier series
6
Complex exponential Fourier series
Maxim Raginsky Lecture VIII: Fourier series
Review: vectors
Recall vectors in n-space R
n
. Each such vector u can be uniquely
represented as a linear combination of n unit vectors e
1
, . . . , e
n
:
u =
1
e
1
+
2
e
2
+ . . . +
n
e
n
,
where
1
, . . . ,
n
are real numbers (coecients). These can be
computed using the scalar (or dot) product as follows:

k
= u e
k
, k = 1, . . . , n
Note that the vectors e
1
, . . . , e
n
are:
normalized e
k
e
k
= 1 for all k
orthogonal e
k
e
m
= 0 for k = m
e
1
e
2
u
v
x
y
6.5 3
3
5
0
Example: u, v in R
2
u = 3e
1
+ 5e
2
v = 6.5e
1
+ 3e
2
Maxim Raginsky Lecture VIII: Fourier series
Vector spaces
By denition, vectors are objects that can be added together and
multiplied by scalars:
if u =

n
k=1

k
e
k
and v =

n
k=1

k
e
k
, then we can form their sum
u +v =
n

k=1
(
k
+
k
)e
k
if u =

n
k=1

k
e
k
and is a scalar, then we can form the vector
u =
n

k=1

k
e
k
More generally, a collection of objects that can be added and/or
multiplied by scalars is called a vector space.
Maxim Raginsky Lecture VIII: Fourier series
Signal spaces
We have already seen one example of a vector space the space of
continuous-time signals. We can easily verify:
we can form the sum of any two signals x
1
(t) and x
2
(t) to get
another signal
x(t) = x
1
(t) + x
2
(t)
we can multiply any signal x(t) by a constant c to get another signal
z(t) = cx(t)
Unlike the n-space R
n
, the vector space of all continuous-time signals is
huge. In fact, it is innite-dimensional.
Maxim Raginsky Lecture VIII: Fourier series
The space of periodic signals
Let us consider all T-periodic signals. Any such signal x(t) satises
x(t + T) = x(t) for all t
for some given T > 0.
It is easy to see that T-periodic signals form a vector space:
if x
1
(t) and x
2
(t) are T-periodic, then
x
1
(t + T) + x
2
(t + T) = x
1
(t) + x
2
(t),
so their sum is T-periodic
if x(t) is T-periodic, then
cx(t + T) = cx(t),
so any scaled version of x(t) is also T-periodic
If we impose even more conditions on our T-periodic signals (the
so-called Dirichlet conditions, which hold for all signals encountered in
practice), then we can represent signals as innite linear combinations of
orthogonal and normalized (orthonormal) vectors.
Maxim Raginsky Lecture VIII: Fourier series
Complete orthonormal sets of functions
First of all, we can dene a scalar (or dot) product of two T-periodic
signals x
1
(t) and x
2
(t) as
x
1
, x
2
=
_
T
0
x
1
(t)x
2
(t)dt
(note that we can integrate over any whole period, not necessarily from
t = 0 to t = T).
Then we can take any sequence of T-periodic functions (signals)

0
(t),
1
(t),
2
(t), . . . that are
1
normalized:
k
,
k
=
_
T
0

2
k
(t)dt = 1
2
orthogonal:
k
,
m
=
_
T
0

k
(t)
m
(t)dt = 0 if k = m
3
complete: if a signal x(t) is such that
x,
k
=
_
T
0
x(t)
k
(t)dt = 0 for all k, then x(t) 0
Maxim Raginsky Lecture VIII: Fourier series
Fourier series
Let {
k
(t)}

k=1
be a complete, orthonormal set of functions. Then any
well-behaved T-periodic signal x(t) can be uniquely represented as an
innite series
x(t) =

k=0

k
(t).
This is called the Fourier series representation of x(t). The scalars
(numbers)
k
are called the Fourier coecients of x(t) (with respect to
{
k
(t)}) and are computed as follows:

k
= x,
k
=
_
T
0
x(t)
k
(t)dt
In analogy to vectors in n-space, you can think of
k
as the projection
(or component) of x(t) in the direction of
k
(t).
Maxim Raginsky Lecture VIII: Fourier series
Fourier coecients
To derive the formula for
k
, write
x(t)
k
(t) =

m=0

m
(t)
k
(t),
and integrate over one period:
_
T
0
x(t)
k
(t)dt
. .
x,
k

=
_
T
0
_

m=0

m
(t)
k
(t)
_
dt
=

m=0

m
_
T
0

m
(t)
k
(t)dt
. .

m
,
k

=
k
,
where in the last line we use the fact that {
k
} form an orthonormal
system of functions.
Maxim Raginsky Lecture VIII: Fourier series
Convergence of Fourier series
It can be proved that, because the functions {
k
} form a complete
orthonormal system, the partial sums of the Fourier series
x(t) =

k=0

k
(t)
converge to x(t) in the following sense:
lim
N
_
T
0
_
x(t)
N

k=1

k
(t)
_
2
dt = 0
So, we can use the partial sums
x
N
(t) =
N

k=1

k
(t)
to approximate x(t).
Maxim Raginsky Lecture VIII: Fourier series
Trigonometric Fourier series
Fact: the sequence of T-periodic functions {
k
(t)}

k=0
dened by

0
(t) =
1

T
and
k
(t) =
_
_
_
_
2
T
sin(k
0
t), if k 1 is odd
_
2
T
cos(k
0
t), if k > 1 is even
is complete and orthonormal. Here,

0
=
2
T
is called the fundamental frequency. Orthonormality is quite easy to
show, completeness not so much.
Thus, any well-behaved T-periodic signal x(t) can be represented as an
innite sum of sinusoids (plus a constant term
0

0
):
x(t) =

k=0

k
(t)
Maxim Raginsky Lecture VIII: Fourier series
Trigonometric Fourier series
A more common way of writing down the trigonometric Fourier series of
x(t) is this:
x(t) = a
0
+

k=1
a
k
cos(k
0
t) +

k=1
b
k
sin(k
0
t)
Then the Fourier coecients can be computed as follows:
a
0
=
1
T
_
T
0
x(t)dt
a
k
=
2
T
_
T
0
x(t) cos(k
0
t)dt
b
k
=
2
T
_
T
0
x(t) sin(k
0
t)dt
Recall that
0
= 2/T.
Maxim Raginsky Lecture VIII: Fourier series
Trigonometric Fourier series
To relate this to the orthonormal representation in terms of the
k
, we
note that we can write
a
0
=
1

T
_
T
0
x(t)
0
(t)dt =
1

0
a
k
=
_
2
T
_
T
0
x(t)
2k
(t)dt =
_
2
T

2k
b
k
=
_
2
T
_
T
0
x(t)
2k1
(t)dt =
_
2
T

2k1
Thus, we have
x(t) = a
0
+

k=1
a
k
cos(k
0
t) +

k=1
b
k
sin(k
0
t)
=
1

T
0
(t) +
_
2
T
_

k=1

2k

_
T
2

2k
(t) +

k=1

2k1

_
T
2

2k1
(t)
_
=

k=1

k
(t)
Maxim Raginsky Lecture VIII: Fourier series
Symmetry properties
Things to watch out for when computing the Fourier coecients:
if x(t) is an even function, i.e., x(t) = x(t) for all t, then all its
sine Fourier coecients are zero:
b
k
=
2
T
_
T/2
2/T
x(t) sin(k
0
t)dt = 0
if x(t) is an odd function, i.e., x(t) = x(t), then all its cosine
Fourier coecients are zero:
a
k
=
2
T
_
T/2
2/T
x(t) cos(k
0
t)dt = 0,
and
a
0
=
1
T
_
T/2
2/T
x(t)dt = 0
Maxim Raginsky Lecture VIII: Fourier series
Trigonometric Fourier series: example
Consider a 2-periodic signal x(t) given by repeating the square wave:
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
a
0
= 0 and a
k
= 0 for all k
(the signal has odd
symmetry)
b
k
=
_
0
1
sin(kt)dt
_
1
0
sin(kt)dt
=
1
k
[cos(kt)]
0
1
+
1
k
[cos(kt)]
1
0
=
1
k
_
cos(k) 1 + cos(k) 1
_
=
1
k
(2 cos(k) 2)
=
4 sin
2
(k/2)
k
x(t) =

k=1
4 sin
2
(k/2)
k
sin(kt) =

k odd
4
k
sin(kt)
Maxim Raginsky Lecture VIII: Fourier series
Trigonometric Fourier series: example
Approximating x(t) by partial sums of its Fourier series:
1 0.5 0 0.5 1
1.5
1
0.5
0
0.5
1
1.5
N = 10
1 0.5 0 0.5 1
1.5
1
0.5
0
0.5
1
1.5
N = 15
1 0.5 0 0.5 1
1.5
1
0.5
0
0.5
1
1.5
N = 50
1 0.5 0 0.5 1
1.5
1
0.5
0
0.5
1
1.5
N = 150
Note the Gibbs phenomenon: the Fourier series (over/under)shoots the
actual value of x(t) at points of discontinuity. In signal processing, this
eect is also called ringing.
Maxim Raginsky Lecture VIII: Fourier series
Complex exponential Fourier series
Another useful complete orthonormal set is furnished by the complex
exponentials:

k
(t) =
1

T
e
jk
0
t
, k = . . . , 2, 1, 0, 1, 2, . . .
where
0
= 2/T, as before.
Note that these functions are complex-valued, and we need to redene
the dot product as
x
1
, x
2
=
_
T
0
x
1
(t)x

2
(t)dt,
where x

1
(t) denotes the complex conjugate of x
2
(t). Then it is
straightforward to show that

k
,
m
=
1
T
_
T
0
e
jk
0
t
e
jm
0
t
dt =
_
1, k = m
0, k = m
Maxim Raginsky Lecture VIII: Fourier series
Complex exponential Fourier series
Thus, we can expand any T-periodic x(t) as
x(t) =

k=
c
k
e
jk
0
t
The Fourier coecients are given by
c
k
=
1
T
_
T
0
x(t)e
jk
0
t
dt
To derive this, multiply the series representation of x(t) on the right by
e
jk
0
t
and integrate from 0 to T.
Maxim Raginsky Lecture VIII: Fourier series
Symmetry properties for real signals
Consider a real-valued T-periodic signal x(t). Then
c
k
=
1
T
_
T
0
x(t)e
jk
0
t
dt and c
k
=
1
T
_
T
0
x(t)e
jk
0
t
dt = c

k
Write c
k
in polar form:
c
k
= A
k
e
j
k
, A
k
= |c
k
|,
k
= c
k
Then
c
k
= c

k
A
k
= A
k
and c
k
= c
k
Thus, for real signals the amplitude spectrum A
k
has even symmetry,
while the phase spectrum
k
has odd symmetry.
Maxim Raginsky Lecture VIII: Fourier series
Amplitude and phase spectra
We can use the amplitudes A
k
and the phases
k
to represent the
spectrum (or frequency content) of x(t) graphically as follows:

0
2
0
3
0

0
2
0
3
0
3
0
2
0

0
3
0
2
0

0
0
0
A()
()

Observe that the amplitude spectrum A() has even symmetry, while
the phase spectrum () has odd symmetry.
Maxim Raginsky Lecture VIII: Fourier series