18.03 Class 26, Apr 12

Fourier Series


A function  f(t)  is "periodic of period 2L"  if  f(t+2L) = f(t).
Waves, vibrations, ... many things phenomena are periodic.

A basic example of a periodic function is given by  cos(t),  which
is periodic of minimal period  2pi.  The is such a basic example
that for today we'll consider only period  2pi:  L = pi .

Other examples are  cos(nt),  n = 2, 3, ...;  sin(nt), n = 1, 2, ...,
and the constant function  1 = cos(0t).

We know a lot about these functions. We know for example how to solve
LTI equations having these as signals. So by superposition we know
how to deal with linear combinations of them as well. The most general
linear combination is

(a0/2) + a1 cos(t) + a2 cos(2) + ... + b1 sin(t) + b2 sin(2t) + ...

(We'll see in a minute why it was smart to write the constant term this way.)
This is a "Fourier series."

Theorem: Essentially any periodic function with period  2pi  can be written
as a Fourier series:  f(t) = this sum.

This will allow us to use very general periodic functions as signals.
To be effective though we have to figure out how to compute  an  and  bn
from  f(t).

a0  is a good place to start:  All the terms beyond the first one have
zero average (over any interval of length  2pi).  Thus

a0/2 = average of f(t) = (1/2pi) integral_{-pi}^pi f(t) dt    so

a0 = (1/pi) integral_{-pi}^pi f(t) dt  .

Recalling that  cos(0t) = 1,  it might be sensible to try to compute  an
by looking at

integral_{-pi}^pi f(t) cos(nt) dt

Putting the various terms of the proposed Fourier series for  f(t) into
this integral one at a time, we see that we need the following trigonometric
integrals:  for  n, k > 0:

integral_{-pi}^pi cos(kt) cos(nt) dt = pi  if  n = k
                                            otherwise

integral_{-pi}^pi sin(kt) cos(nt) dt = 0

integral_{-pi}^pi sin(kt) sin(nt) dt = pi  if  n = k
                                        0   otherwise

We find that only one term is nonzero in the sum:

integral_{-pi}^pi f(t) cos(nt) = an pi .

(Notice that this is correct for  n = 0  as well as  n > 0.  This is
the reason for writing the constant term as  a0/2.)  Similarly,

integral_{-pi}^pi f(t) sin(nt) = bn pi .

That's it: each coefficient turns out to be, necessarily, the integral
of a certain trig function times  f(t).


Example:  sq(t) = 1  for    0 < t < pi,
                  -1  for  -pi < t < 0 .

The average is  0 :  a0 = 0.

sq(t)  is an odd function:
A function  f(t)  is "odd"  if  f(-t) = -f(t),
                     "even"  if  f(-t) =  f(t).

Any odd function has average (over  -pi < t < pi)  zero.

Even times odd  is odd
Even times even is even
Odd  times odd  is even

and

integral_{-pi}^pi odd dt  = 0

integral_{-pi}^pi even dt = 2 integral_0^pi even dt

If  f(t)  is any odd function then  f(t) cos(nt)  is again odd and has
zero integral:  an = 0.   (Similarly, if  f(t)  is an even function then
bn = 0.)

Thus  an = 0,  and

bn = (2/pi) integral_0^pi sin(nt) dt

    = - (2/(pi n)) [cos(nt)]^pi_0.

Since  cos(n pi) = +1  if  n  is even
                    -1  if  n  is odd

bn = - (2/(pi n)) [1 - 1] = 0         if  n  is even
    = - (2/(pi n)) [-1 -1] = 4/(n pi)  if  n  is odd.

We have computed the Fourier series for sq(t):

sq(t) = (4/pi)( sin(t) + (1/3) sin(3t) + (1/5) sin(5t) + ... )

I showed some slides of how this builds up.

Once you have computed one Fourier series you can find lots of others.
For example suppose  f(t)  is periodic of period  2pi  and  f(t) = 0  for
0 < t < pi,  f(t) = -1  for  pi < t < 2pi.

Then also  f(t) = -1  for  -pi < t < 0.  We could compute the Fourier
coefficients just as we just did; or we can realize that

f(t) = -(1/2) + (1/2) sq(t)   and so

f(t) = -(1/2) + (2/pi)( sin(t) + (1/3) sin(3t) + ...) .


Fourier series constitute a new kind of "transform." Notice the analogy
with Laplace transform:

Laplace: (functions of t > 0)  ---->  (Functions of  s)

Fourier: (functions periodic of period  2pi) ---> (a0 , a1, ...,  b1, b2, ...)

Each coefficient  an , bn , is like a value of the Laplace transform;
it is given by integrating  f(t)  multiplied, not by  e^{-st}  but rather
by a sinusiodal function.  We are in a better situation here, in that
we can be explicit about the inverse transform; this is given exactly by
the Fourier series.