J. Math. Anal. Appl.

285 (2003) 349–355

www.elsevier.com/locate/jmaa

A formula for the Fourier transform of certain odd

differentiable functions ✩
Ryan Berndt
University of Minnesota, 127 Vincent Hall, 206 Church Street, SE, Minneapolis, MN 55455, USA
Received 22 October 2002
Submitted by L. Grafakos

In this paper we show a Fourier transform formula for a certain class of differentiable
functions K(x) = B(x)/x defined on the real line R. It relates the global behavior of the
Fourier transform of K to the function B. The precise formulation is contained in (2) and
(3) below. In particular, the theorem recovers the essential behavior of the Fourier transform
of the kernel of the Hilbert transform, and may be thought of as a generalization of this.
Before we state the theorem we set notation.
We write the Fourier transform of a func-
tion f as fˆ and it is given by the formula fˆ(z) = f (x)e−izx dx. The independent variable
z is an element of R. For A and B real we write A ≈ B to mean there exist positive con-
stants C1 and C2 such that C1 B
A
C2 B.

Theorem 1. Let K be positive for x > 0 and odd. Suppose there exist constants c1 and c2 ,
0 < c1
c2 < 2, such that
K(x) K(x)
c1
−K  (x)
c2 for x > 0. (1)
x x
Then, K̂ exists as a tempered distribution and can be identified with a function (which we
also call K̂) such that for z > 0,

i K̂(z) ≈ B(π/z). (2)

Since K̂ is odd, this describes K̂ for all z ∈ R so that we may in fact write

K̂(z) = −i sgn(z)B(π/z)h(z), (3)

where h is a real function that is bounded above and below by positive constants.

✩
This work was done while the author was writing his doctoral thesis under the direction of Max Jodeit Jr. at
the University of Minnesota.
E-mail address: [email protected].

0022-247X/$ – see front matter  2003 Elsevier Inc. All rights reserved.
doi:10.1016/S0022-247X(03)00243-9
350 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355

Before we prove the theorem, we consider some examples of functions K that sat-
isfy (1). An important one is given by B(x) = 1/π so that K is the kernel of the Hilbert
transform. It is known that K̂(z) = −i sgn(z). This agrees with our theorem and (3). An-
other example is given by an odd version of the Riesz potentials. That is, for 0 < α < 1
we may take K(x) = x −1+α for x > 0 and extend K to be odd for x < 0. In this case
c1 = 1 − α and c2 = 1 + α suffice to satisfy (1).
Other functions satisfying (1) include

1 + log2 |x|
K(x) = ,
x
 
3 + sin 1 + log2 |x| 1 1
K(x) = , and K(x) = √ + 3/2 .
x 2 x 2x
In all three of these examples we may take c1 = 1/2 and c2 = 3/2.
Although condition (1) is a technically useful formulation of our hypothesis, we may
formulate (1) in another way that is somehow more pleasing. If K(x) = B(x)/x and if
there exists a constant c0 , 0 < c0 < 1, such that
  
B (x)
c0 B(x) for x
= 0, (4)
|x|
then K satisfies (1) with c1 = 1 − c0 and c2 = 1 + c0 . Conversely, (1) implies (4) with
c0 = max{1 − c1 , c2 − 1}.
We are interested in the global behavior of K̂. However, there are results in the literature
that describe the asymptotic behavior of K̂ near 0 or ∞. We write f (x) ∼ g(x) as x → ∞
to mean f (x)/g(x) → 1 as x → ∞. We say (x) is a slowly varying function if for any
λ > 0, (λx)/(x) → 1 as x → ∞. Then,

Theorem 2 ([3, Theorem 1] and [5]). If K(x)  0 for x  0 and K is odd, bounded,
and decreases to 0 as x → ∞ then K(x) ∼ (x)/x α as x → ∞ for some slowly varying
function  if and only if K̂(z) ∼ zα−1 (1/z)Γ (1 − α) cos(πα/2) as z → 0.

This theorem can also be found in [1, p. 240]. The hypotheses on K in Theorem 2
imply [1, p. 27] that K ∈ L1 (R). For a theorem that is similar but that relates the asymp-
totic behavior of K̂ at infinity to the asymptotic behavior of K at zero see the paper of
Cline [2]. Theorem 1 is in the spirit of these asymptotic results, but Theorem 1 applies to
nonintegrable functions and gives global results.
Finally, we should note that functions B satisfying (4) are not necessarily slowly vary-
 
ing. In fact B(x) = 3 + sin 1 + log2 |x| satisfies (4) but limx→∞ B(λx)/B(x) does not
exist. This example is essentially from Seneta [4, p. 49].

Proof of Theorem 1. The function we will associate with i K̂(z) is

R
2 lim K(x) sin(xz) dx.
R→∞
0
 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355 351

We will justify this association at the end of the paper. Fix z > 0. For R > π/z we consider
the parts of the integral defined on [0, π/z] and [π/z, R]. We first show that
  π/z
2 c1 π
+ 2 B(π/z)
K(x) sin(xz) dx
B(π/z). (5)
π π 2 − c2
0
To prove (5) we use hypothesis (1). We then prove, using only that K is positive and
decreasing for x > 0, that
R
2
− B(π/z)
K(x) sin(xz) dx
0. (6)
π
π/z
We conclude that for R large enough
R
K(x) sin(xz) dx ≈ B(π/z). (7)
0
First, we prove the right half of (5). Since K(x) sin(xz) is positive for x ∈ (0, π/z),
π/z π/z
K(x) sin(xz) dx
z xK(x) dx.
0 0
Let ε > 0. Integration by parts yields
π/z   π/z
1 π2 π 1 2 1
xK(x) dx = K − ε K(ε) − x 2 K  (x) dx
2 z2 z 2 2
ε ε
  π/z
1π π c2

B + xK(x) dx.
2z z 2
ε
Now we bring the integral on the right back to the left so that
π/z  
π π
z xK(x) dx
B . (8)
2 − c2 z
ε

π/z
By Fatou’s lemma, z 0 xK(x) dx
cB(π/z), where c = π/(2 − c2 ). Incidentally, this
also proves that xK(x) = B(x) is integrable near 0—a fact we will use frequently below.
To establish the other half of (5), we write
π/z 
π/2z π/z
K(x) sin(xz) dx = K(x) sin(xz) dx + K(x) sin(xz) dx
0 0 π/2z
  π/2z
 π/z
π
K sin(xz) dx + K(x) sin(xz) dx,
z
0 π/2z
352 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355

applying the fact that K is decreasing. Next, since the integral of sin(xz) on [0, π/2z] is
1/z and K(x) = B(x)/x the first summand is equal to K(π/z)/z = (1/π)B(π/z). Hence,
π/z   π/z
1 π
K(x) sin(xz) dx  B + K(x) sin(xz) dx. (9)
π z
0 π/2z

Now, we want to show that

π/z  
1
K(x) sin(xz) dx  c + B(π/z)
π
π/2z

for some positive constant c. To this end we note that (1/π)B(π/z) = (1/z)K(π/z) =

π/z
π/2z K(π/z) sin(xz) dx. Hence, we want to show that

π/z

K(x) − K(π/z) sin(xz) dx

π/2z

is at least a positive constant times B(π/z). By the mean value theorem K(x) − K(π/z) =
K  (s)(x − π/z) for some point s ∈ (x, π/z). Then,
K(x) − K(π/z) = K  (s)(x − π/z) = −K  (s)(π/z − x)
   
K(s) π K(π/z) π
 c1 − x  c1 −x .
s z π/z z
Since K(π/z)/(π/z) = (z2 /π 2 )B(π/z) we now have
π/z π/z

z2
K(x) − K(π/z) sin(xz) dx  c1 2 B(π/z) (π/z − x) sin(xz) dx
π
π/2z π/2z
π
c1 c1
= B(π/z) (π − x) sin x dx = B(π/z).
π2 π2
π/2

Thus,
π/z  
1 c1
K(x) sin(xz) dx  + 2 B(π/z).
π π
π/2z

Therefore by (9),
π/z  
2 c1
K(x) sin(xz) dx  + 2 B(π/z),
π π
0
and we have established (5).
 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355 353

We now prove (6). Let kj = j π/z and bj = (2/j π)B(j π/z) for j = 1, 2, . . . . Let ν be
the integer such that νπ < Rz
(ν + 1)π . We write
R ν−1 j+1

k R
I := K(x) sin(xz) dx = K(x) sin(xz) dx + K(x) sin(xz) dx
π/z j =1 k νπ/z
j

:= Σν−1 + ρR .

kj+1
Since K(kj +1 )
K(x)
K(kj ) for x ∈ [kj , kj +1 ] and kj sin(xz) dx = (−1)j 2/z, we
have
kj+1
2K(kj +1 ) 2K(kj )
bj +1 =
K(x) sin(xz) dx
= bj for j even,
z z
kj
kj+1
−2K(kj ) −2K(kj +1 )
−bj =
K(x) sin(xz) dx
= −bj +1 for j odd.
z z
kj

To prove that I
0, we first note that if ν is odd then ρR
0, hence I
Σν−1 . If ν is

(ν+1)π/z
even then ρR
νπ/z K(x) sin(xz) dx, hence I
Σν . In either case I
Σm where
m is even. But, Σm
−b2 + b2 − b4 + b4 + · · · − bm + bm = 0. Consequently, I
0. To
prove the other inequality we note that if ν is even then ρR  0, hence I  Σν−1 . If ν is

(ν+1)π/z
odd then ρR  νπ/z K(x) sin(xz) dx, hence I  Σν . In either case I  Σm where m
is odd. But, Σm  −b1 + b3 − b3 + · · · + bm − bm . Therefore, I  −b1 . This proves (6)
and thereby (7).
We now show that the Fourier transform of K exists as a tempered distribution and

R
that we are justified in writing i K̂(z) = 2 limR→∞ 0 K(x) sin(xz) dx. Recall that imme-
diately after (8) we showed that xK(x) = B(x) is integrable on [0, 1]. We will use this fact
below. Let K(φ) be the linear functional on the Schwartz space S and be defined by the
formula

K(φ) = lim K(x)φ(x) dx.
ε→0
R→∞ ε
|x|
R

We first show that this integral is defined for φ ∈ S. Since K is odd,

 1

lim K(x)φ(x) dx = lim K(x) φ(x) − φ(−x) dx

ε→0 ε→0
R→∞ ε
|x|
R ε

R

+ lim K(x) φ(x) − φ(−x) dx.

R→∞
1

Since φ is smooth, |K(x)[φ(x) − φ(−x)]|
2φ  ∞ xK(x) which is integrable on [0, 1].
So, the first limit exists and is equal to the integral on [0, 1] by the dominated conver-
354 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355

gence theorem. Also, |K(x)[φ(x) − φ(−x)]|
K(1){|φ(x)| + |φ(−x)|} for x  1 and φ

is integrable. Hence, K is a linear functional defined on S and
R

K(φ) = lim K(x) φ(x) − φ(−x) dx.

R→∞
0

By definition, the Fourier transform of K is given by K(φ̂),

R

K(φ̂) = lim K(x) φ̂(x) − φ̂(−x) dx

R→∞
0
R 
= −2i lim K(x) φ(z) sin(xz) dz dx.
R→∞
0
Applying Fubini’s theorem along with the fact that xK(x) is integrable on [0, R] we get
 R
K(φ̂) = −2i lim φ(z) K(x) sin(xz) dx dz. (10)
R→∞
0
Finally, to pass the limit inside the first integral it suffices to find an integrable func-

R
tion independent of R that dominates |φ(z) 0 K(x) sin(xz) dx|. As we have seen,

R
| 0 K(x) sin(xz) dx|
c B(π/z) for some constant c. We show φ(z)B(π/z) is integrable
on [0, ∞]. The same methods can be applied to obtain integrability on [−∞, 0]. We have
∞ π ∞
     
φ(z)B(π/z) dz = φ(z)B(π/z) dz + φ(z)B(π/z) dz.
0 0 π

To estimate the part of the integral defined on [π, ∞] we let

 2  
φ2,0 = sup 1 + |z| φ(z)
z∈R

which is finite since φ ∈ S. Then, |φ(z)|z2
φ2,0 and so

∞ ∞
 
φ(z)B(π/z) dz
φ2,0 B(π/z) dz.
z2
π π
Now we change variables and use the same estimate for the integral of B as was given
in (8), except on the interval [0,1],
∞ 1
  π
φ(z)B(π/z) dz
πφ2,0 B(z) dz
φ2,0 B(1).
2 − c2
π 0
To estimate the part of the integral defined on [0, π] we apply the hypothesis (1),
 R. Berndt / J. Math. Anal. Appl. 285 (2003) 349–355 355

π π ∞
  B(z)
φ(z)B(π/z) dz
φ∞ B(π/z) dz = πφ∞ dz
z2
0 0 1
∞ ∞
K(z)
= πφ∞ dz
πc1−1 φ∞ −K  (z) dz
z
1 1
= πc1−1 φ∞ K(1).
Therefore, we may pass the limit inside the first integral in (10) and hence
  R 
i K̂(φ) = φ(z) 2 lim K(z) sin(xz) dx dz.
R→∞
0

The functional K̂ is a tempered distribution since there exists a constant C such that

R
|K̂(φ)|
C(φ∞ + φ2,0 ). So, we may identify i K̂(z) with 2 limR→∞ 0 K(z) ×
sin(xz) dx.

References

[1] N.H. Bingham, C.M. Goldie, J.L. Teugels, Regular Variation, Cambridge Univ. Press, Cambridge, 1987.
[2] D.B.H. Cline, Abelian Tauberian theorems relating the local behavior of an integrable function to the tail
behavior of its Fourier transform, J. Math. Anal. Appl. 154 (1991) 55–76.
[3] E.J.G. Pitman, On the behavior of the characteristic function of a probability distribution in the neighbourhood
of the origin, J. Austral. Math. Soc. 29 (1968) 423–443.
[4] E. Seneta, Regular Varying Functions, in: Lecture Notes in Mathematics, Vol. 508, Springer-Verlag, Berlin,
1976.
[5] K. Soni, R.P. Soni, Slowly varying functions and asymptotic behavior of a class of integral transforms II,
J. Math. Anal. Appl. 49 (1975) 477–495.