Professional Documents
Culture Documents
f (n) =
_
1
0
f (x)e
2inx
dx, (n ).
We think of
f as dening a function
f : . The Fourier series of f is the formal
series
f (n)e
2inx
.
We rewrite this series down without any assertion that it actually converges: the question
of convergence is something we will postpone for now. The only observation we make
now is this: if it is known a priori that f can be expressed as
f (x) =
a
n
e
2inx
for some constants a
n
, and if in addition we know that this series converges uniformly
(for example because
|a
n
| < ), then the Fourier coecients of the function f must
be the a
n
, because
f (m) =
_
1
0
f (x)e
2imx
dx
=
_
1
0
_
a
n
e
2inx
_
e
2imx
dx
=
n
a
n
_
1
0
e
2i (nm)x
dx
= a
m
.
We used the uniform convergence to interchange the integration with the summation
at the third line, and then we used the fact that
_
e
2i (nm)x dx
= 0 for n = m.
Before trying to say more about Fourier series and their convergence, we record some
basic facts about the coecients.
Proposition 2.1. The Fourier coecients satisfy:
(a) |
f (n)| f
L
1
()
, for all n;
(b)
f (n) 0 as |n| .
3
Proof. The rst property is straightforward. The second property is the iemann-
Lebesgue lemma. We previously stated the iemann-Lebesgue lemma in the context
of the Fourier transform, and the present statement follows immediately. Indeed, apply-
ing the previous version to the function f (x)
[0,1]
(x) teaches us that
_
1
0
f (x)e
2i x
dx
goes to 0 as || , and the Fourier coecients in the present context are just the
values of this expression at integer values of .
Proposition 2.2. Suppose that F is the indenite integral of a periodic integrable function f : so
F(x) =
_
x
0
f for all x . To ensure that F is also periodic, assume that
_
1
0
f = 0. Then the
Fourier coecients of F and its L
1
-derivative f are related by
f (n) = 2in
F(n).
Proof. We can prove this using integration by parts (as justied in Problem Set 6):
F(n) =
_
1
0
F(x)e
2inx
dx
=
_
1
0
f (x)
_
e
2inx
/(2in)
_
dx +
_
F(x)
_
e
2inx
/(2in)
__
1
0
=
1
2in
f (n) +0.
Corollary 2.3. If f is k times continuously dierentiable, then n
k
f (n) 0 as |n| .
3. Examples
The square wave. Let f be the periodic function that is 1 on [1/4, 1/4] and 1 on
(1/4, 3/4). The Fourier coecients are
f (n) =
_
1
0
e
2inx
dx +2
_
1/4
1/4
e
2inx
dx
=
_
0, n even,
2(1)
(n1)/2
/(n), n odd.
4 3. Examples
Written in terms of cosines, the Fourier series looks like
4
cos(2x)
4
3
cos(6x) +
4
5
cos(10x)
2 1 1 2
1.0
0.5
0.5
1.0
Figure 1: The partial sum of the Fourier series for the square wave, summed to |n| = 11.
The triangle wave. Let g be the indenite integral of the previous square wave f , with
g(0) = 0. The Fourier coecients of g are
g(n) =
_
0, n even,
i (1)
(n+1)/2
/(n)
2
, n odd.
2 1 1 2
0.2
0.1
0.1
0.2
Figure 2: The partial sum of the Fourier series for the triangle wave, summed to
|n| = 11.
Written in terms of sines, the Fourier series of the triangle wave looks like
2
2
sin(2x)
2
9
2
sin(6x) +
2
25
2
sin(10x)
5
Unlike the case of the square wave, this sum is uniformly absolutely convergent.
4. Partial sums and Cesro means
We write s
n
f for the nth partial sum of the Fourier series of f . The partial sum is
interpreted in the symmetric sense, going from n to n:
(s
n
f )(x) =
n
m=n
f (m)e
2imx
.
The nth Cesro mean of the Fourier series is the average of the rst n partial sums:
(
n
f )(x) =
1
n
_
(s
0
f )(x) + +(s
n1
f )(x)
_
.
We can give alternative expressions for the partial sums and Cesro means. We have
(s
n
f )(x) =
n
m=n
_
1
0
f ( y)e
2imy
e
2imx
dy
=
_
1
0
_
n
m=n
e
2im(xy)
_
f ( y) dy
=
_
1
0
D
n
(x y) f ( y) dy.
Here D
n
is the Dirichlet kernel, dened by the formula that arises above:
D
n
(x) =
n
n
e
2imx
.
Summing this geometric series, we nd
D
n
(x) =
sin((2n +1)x)
sin(x)
.
The expression for s
n
f that appears above is a convolution. For periodic functions f and
g in this context, the convolution f g is dened by
( f g)(x) =
_
1
0
f (x y)g( y) dy.
6 5. Pointwise convergence of Cesro means
So we can write
s
n
f = D
n
f . (1)
The Cesro sums can be written similarly. We have
n
f = (1/n)(D
0
+. . . D
n
) f
= F
n
f
where F
n
is the Fejr kernel,
F
n
= (1/n)(D
0
+ + D
n1
).
Summing the series again, we nd
F
n
=
1
n
_
sin(nx)
sin(x)
_
2
.
We summarize this discussion with a proposition:
Proposition 4.1. For f L
1
(), the partial sums s
n
f and the Cesro means
n
f of the Fourier
series of f are given by
s
n
f = D
n
f
and
n
f = F
n
f
respectively, where denotes convolution on and the functions D
n
and F
n
are the Dirichlet kernel
and the Fejr kernel
D
n
(x) =
sin((2n +1)x)
sin(x)
F
n
=
1
n
_
sin(nx)
sin(x)
_
2
.
5. Pointwise convergence of Cesro means
The material on Cesro sums that we developed for the Fourier transform adapts in a
straightforward way to the case of Fourier series.
Theorem 5.1. If f is a periodic integrable function and x
0
is a Lebesgue point of f , then the
Cesaro means (
n
f )(x
0
) converge to f (x
0
) as n .
7
Proof. We draw on what we learned when dealing the Fourier transform. The question
is to show that
(F
n
f )(x
0
) f (x
0
)
as n . The integral
_
1
0
F
n
is 1 (an easy exercise if you use the denitions of D
n
and
F
n
as sums, rather than the formulae in terms of sin). So an equivalent question is to
show
_
1/2
1/2
( f (x
0
y) f (x
0
))F
n
( y) dy 0.
The function F
n
is non-negative, so we are okay if we can show
_
1/2
1/2
| f (x
0
y) f (x
0
)|F
n
( y) dy 0.
To make use of our earlier material directly, we can quietly forget that our functions
extend periodically past the interval [1/2, 1/2] and simply think of the functions on
given by
g( y) =
_
| f (x
0
y) f (x
0
)|, |y| 1/2
0, |y| 1/2
and
G
n
( y) =
_
F
n
( y), |y| 1/2
0, |y| 1/2
Then g is integrable on and has 0 as a Lebesgue point. And G
n
satises the conditions
for an approximation to the identity": that is, with = 1/n, we have
G
n
( y) Const. min(1/, |y|
2
).
It follows from our previous results about convolutions on that
_
g( y)G
n
( y) 0
as n , which is what we need.
We have the following corollaries.
Corollary 5.2. If f is a continuous periodic function, then the Cesro means of the Fourier series,
n
f , converge to f everywhere.
Corollary 5.3. If f L
1
() and
f (n) = 0 for all n, then f = 0 a.e.
8 6. Pointwise convergence of the partial sums
Corollary 5.4. If the partial sums, s
n
f (x
0
), converge as n and x
0
is a Lebesgue point of
f , then the limit is the correct one:
lim
n
(s
n
) f (x
0
) = f (x
0
).
Proof. If the partial sums of any series converge, then the Cesro means also converge,
to the same value. So the corollary follows from the theorem.
As a simple criterion which is a corollary of the previous one:
Corollary 5.5. If f is continuous and the Fourier coecients are absolutely summable (that is
|
f (n)| is nite), then (s
n
f )(x) converges to f (x) as n , for all x.
6. Pointwise convergence of the partial sums
For an integrable function f L
1
(), it may not be the case that the partial sums of the
Fourier series converge almost everywhere. Furthermore, there are continuous periodic
functions f for which (s
n
f )(x) fails to converge for some x. (We sketched an example
of this phenomenon in class.)
If we wish to see (s
n
f )(x) converge to f (x) as n , we should impose some con-
straints on the behavior of f near x. There are various criteria which work. Here is
one. This is Dinis criterion.
Theorem 6.1. If f is a periodic integrable function, then the partial sums of the Fourier series,
(s
n
f )(x
0
), converge to f (x
0
) as n provided that the function
f (x) f (x
0
)
x x
0
(2)
is an integrable function of x on the interval [x
0
1/2, x
0
+1/2].
Before giving the proof, we make some comments. The denominator in (2) goes to zero
only as x x
0
; so, since the numerator is integrable, the integrability of the quotient
depends only on the behavior near x
0
. In other words, the quotient is integrable on
[x
0
1/2, x
0
+ 1/2] provided that is integrable on [x
0
, x
0
+ ] for any positive .
Second the criterion applies in many basic cases: it applies
if f is dierentiable at x
0
;
9
if f is continuous at x
0
and the left- and right-derivatives exist;
if f is Hlder continuous at x
0
with any exponent (0, 1). By denition, this
means that there is a constant C with
| f (x) f (x
0
)| C|x x
0
|
.
The other remark we make is that, on the interval [1/2, 1/2], the function sin(y) is
comparable to the function y:
|2y| | sin(y)| |y|.
So, writing x as x
0
y, Dinis criterion is equivalent to the integrability of
| f (x
0
y) f (x
0
)|
sin(y)
(3)
as a function of y on [1/2, 1/2].
Proof of the Theorem (Dinis Criterion). The rst simple steps of the argument are the same
as our analysis of the Cesro means, but with the Dirichlet kernel D
n
replacing F
n
. These
rst steps tell us that the convergence of the partial sums is equivalent to having
_
1/2
1/2
( f (x
0
y) f (x
0
))D
n
( y) dy 0
as n . We write D
n
( y) as a quotient of sines; the above integral is
_
1/2
1/2
_
f (x
0
y) f (x
0
sin(y)
_
sin((2n +1)y) dy.
The expression that appears in the large parentheses is the same expression that appears
in (3) and that is integrable by hypothesis. So the above integral has the general shape
_
h( y) sin((2n +1)y) dy
for some integrable function h. This integral converges to 0 as n by the iemann-
Lebesgue lemma. This concludes the proof.
10 7. The Poisson summation formula
7. The Poisson summation formula
We return to integrable functions on , rather than periodic functions. The Poisson
summation formula states that, if f : is integrable and
f denotes its Fourier
transform, then under mild conditions on f , we have
n=
f (n) =
n=
f (n). (4)
We will give the proof, and at the same time elucidate what sort of mild conditions
we must impose.
The rst step is to construct from the integrable function f a periodic function F L
1
()
by the formula
F(x) =
n=
f (x +n). (5)
Does this sum make sense? It does, but not necessarily for all x. If write f
n
for the
function on [0, 1] dened by f (x +n), then
f
n
L
1
[0,1]
=
_
| f | < ,
which says that the sum
f
n
is absolutely summable in L
1
[0, 1]. Absolute summability
implies summability, so there does exist an F L
1
[0, 1] such that
n=
f
n
= F
in the sense that the partial sums converge in L
1
norm. It also follows from this that, for
almost all x, the series
f (x +n) is absolutely convergent and converges to F(x).
Having constructed this periodic function F, the key to the Poisson summation formula
is the following simple lemma:
Lemma 7.1. The nth Fourier coecient,
F(n), of the periodic function F is equal to
f (n), where
F(n) = lim
N
_
1
0
_
N
m=N
f
m
(x)
_
e
2inx
dx
= lim
N
_
N+1
N
f (x)e
2inx
dx
=
_
f (x)e
2inx
], dx
=
f (n).
We can now understand why the Poisson summation formula (4) might hold. The sum
that appears on the left of (4) is supposed to be F(0): that is, if the sum (5) which
denes F converges at x = 0 to the correct value, then
f (n) = F(0).
On the other hand, because
f (n) =
F(n), the sum on the right of (4) is the Fourier series
of the periodic function F at x = 0. So, if the Fourier series of F converges at x = 0 to the
correct value, then
f (n) =
F(n)
= F(0).
To summarize this, we have (for example), the following sucient conditions (in two
variants) for all this to work. First variant:
Proposition 7.2. The Poisson summation formula (4) holds provided that (a) the sum dening
F in (5) converges at x = 0 to F(0), and (b) the function F(x) satises Dinis criterion at
x = 0.
The second variant uses Corollary 5.5.
Proposition 7.3. The Poisson summation formula (4) holds provided that (a) the sum dening
F in (5) converges to everywhere to a continuous function F, and (b) the sum
f (n) converges
absolutely.
The conditions of the second version of the Proposition hold, for example, if both f
and
f have decay like (1 +|x|)
1
as |x| .
12 8. Two examples of the Poisson summation formula
8. Two examples of the Poisson summation formula
As a rst example of the Poisson summation formula (an example where f is discontin-
uous) take
f (x) =
_
1, |x| 1/4
0, |x| < 1/4,
i.e. the characteristic function of [1/4, 1/4]. (The sum that denes F converges rather
trivially to the periodic function that is 1 on [1/4, 1/4] and 0 elsewhere in its periodic
range.) We certainly have
f (n) = 1.
The Fourier transform of f is the function
f ( ) =
sin(/2)
.
By Dinis criterion, the Fourier series of F converges at x = 0, so we have
f (n) =
f (n),
or in other words
1 =
n=
sin(n/2)
n
(where the term on the right for n = 0 is interpreted using LHpitals rule). Apart from
n = 0, the non-zero contributions come from odd n, and the sum is
1 =
1
2
+2
k=0
(1)
k
(2k +1)
,
i.e.,
4
= 1
1
3
+
1
5
1
7
+ .
This is a series for that can also be deduced from the Taylor series for arctan.
As a second example, we take the Gaussian function, with the variable scaled by R > 0:
f (x) = e
(Rx)
2
.
The Fourier transform is
f ( ) =
1
R
e
(/R)
2
.
13
So the Poisson summation formula gives
e
n
2
R
2
=
1
R
e
n
2
/R
2
.
This formula is more often expressed in terms of t = R
2
:
e
n
2
t
=
1
e
n
2
/t
.
If we dene
(t ) =
e
n
2
t
then the formula reads
(t ) =
1
t
(1/t ).
This is Jacobis identity for the theta function.
We can adapt this example a little. If we apply the summation formula to the function
f (x) = e
(Rx)
2
e
2i ax
,
we obtain
e
n
2
R
2
e
2ina
=
1
R
e
(na)
2
/R
2
.
Writing t = R
2
again and replacing a by x, we have
e
n
2
t
e
2inx
=
1
e
(nx)
2
/t
.
The function H(t, x) that appears on the left of this equality satises the dierential
(4)
H
t
=
2
H
x
2
,
as is easy to see by dierentiating term by term. This equation is the heat equation (though
the factor 4 is not standard). The function H(t, x) is the fundamental solution of the heat
equation in the periodic case. That is, it represents the temperature on a circular wire at
time t, when the heat is initially all concentrated at a single point at x = 0. The function
(t ) represents the temperature at x = 0 as a function of time.