Math 55b: Honors Advanced Calculus and Linear Algebra

Parsevals identity for Fourier transforms

Parsevals identity for Fourier transforms can be stated as follows:
Theorem. There exists a linear operator f

f on L
2
([0, 1]) that coincides
with the Fourier transform on bounded L
1
functions and satises the identity
(

f, g) = 2(f, g)
for all f, g L
2
([0, 1]).
This identity, or its special case |

f|
2
= 2|f|
2
, is known as Parsevals identity.
The two forms are equivalent, since applying the special case to f g and f ig
recovers the general case. (We have seen this polarization trick last term;
remember quarter squares?) Yet another equivalent form is: the operator
f (2)
1/2
f is unitary. This is actually not quite equivalent: a unitary
operator must not only preserve the norm but be bijective. On an innite-
dimensional Hilbert space there are operators that are norm-preserving but
not surjective, such as (a
1
, a
2
, a
3
, . . .) (0, a
1
, a
2
, . . .) on l
2
. But the map
f (2)
1/2
f is known to be bijective, because its fourth iterate is the identity
by the inversion formula.
Since both sides of Parsevals identity are continuous in f, it is enough to prove
the identity for f in a dense subset of L
2
([0, 1]). We know that the continuous
functions supported on nite intervals constitute such a subset. Let f, then, be
a continuous function of compact support. Write (f, f) as a convolution:
(f, f) =

f(t)f(t) dt =

f(t)f
1
(t) dt = (f f
1
)(0),
where f
1
(t) := f(t) is also a continuous function of compact support. Thus the
same is true of g := f f
1
. Therefore g =

f

f
1
. We readily nd that

f
1
() =

f(),
and conclude that g() = |

f()|
2
for all R. Since g is continuous we may
apply Fejers theorem for Fourier transforms to nd
g(0) =
1
2
lim
R

R
R

1
||
R

f()|
2
d.
But |

f()|
2
0 for all , so the limit exists if and only if

f()|
2
d converges
in which case it equals that integral. Thus
(f, f) = g(0) =
1
2

f()|
2
d =
1
2
(

f,

f)
as desired, and we are done.
1
(The same approach can also prove (

f, g) = 2(f, g) directly once it is known
that

f, g are in L
2
so that

f g is L
1
by Cauchy-Schwarz.)
As an application we prove:
Theorem. For any L
1
function f : R C, its Fourier transform

f() ap-
proaches 0 as || .
Recall that we have already shown that

f is bounded by

|f(t)| dt. Thus

again it is enough to prove our claim for f in a dense subset S of L
1
(R). Indeed
for each we may then approximate f by some f
1
S to within /2 in the L
1
norm; then |

f()

f
1
()| /2 for all , since

f

f
1
is the Fourier transform
of f f
1
. If it is known that

f
1
() 0 as || , then there exists R such
that |

f
1
()| < /2 for || > R; thus |

f()| < for || > R. Since is arbitrary,

it will follow that f() 0 as || as desired.
We may use the L
2
functions in L
1
(R) as the dense subset S, since the bounded,
compactly supported functions are dense in L
1
(R) and a compactly supported
integrable function is automatically square integrable. By Parseval, if f
1
is in
this S then so is

f
1
. But we know already (Lemma 46.3) that

f
1
is uniformly
continuous. Well, a uniformly continuous L
2
function on R must tend to 0 as
|| . Indeed if it didnt we would have an > 0 and innitely many
n
such
that |

f
1
(
n
)| and |
m

n
| > 1 for m = n. Since

f
1
is uniformly continuous,
there exists > 0 such that |

f
1
()

f
1
(

)| < /2 if |

| < . We may assume

that < 1/2. Then the -balls about the
n
are disjoint and we have
>

f
1
()|
2
d

n=1

n+
n
|

f
1
()|
2
d

n=1

n+
n
(/2)
2
d =

n=1

2
2
,
which is impossible. Our theorem is thus proved.
We could have used other subsets S, such as the functions of bounded variation,
for which we know

f
1
() 1/||. We use the L
2
functions because this approach
generalizes most readily to the Fourier transform on an arbitrary locally compact
abelian group.
2