J. Math. Anal. Appl.

363 (2010) 225229

Contents lists available at ScienceDirect
Journal of Mathematical Analysis and
Applications
www.elsevier.com/locate/jmaa
Variations on a theme of Boole and SteinWeiss
L. Colzani
a
, E. Laeng
b,
, L. Monzn
c
a
Universit di Milano Bicocca, Dipartimento di Matematica e Applicazioni, Via Cozzi 53, 20125 Milano, Italy
b
Politecnico di Milano, Dipartimento di Matematica F. Brioschi, Via Bonardi 9, 20133 Milano, Italy
c
Department of Applied Mathematics, 526 UCB, University of Colorado, Boulder, CO 80309-0526, USA
a r t i c l e i n f o a b s t r a c t
Article history:
Received 15 May 2009
Available online 19 August 2009
Submitted by L. Grafakos
Keywords:
Hilbert transform
Distribution function
L
p
norm
We give an alternative proof of a theorem of Stein and Weiss: The distribution function of
the Hilbert transform of a characteristic function of a set E only depends on the Lebesgue
measure |E| of such a set. We exploit a rational change of variable of the type used by
George Boole in his paper On the comparison of transcendents, with certain applications to the
theory of denite integrals together with the observation that if two functions f and g have
the same L
p
norm in a range of exponents p
1
< p < p
2
then their distribution functions
coincide.
2009 Elsevier Inc. All rights reserved.
1. Introduction
The Hilbert transform on the line is the operator H dened by the principal value integral
Hf (x) =
1

+
_

f (x y)
y
dy. (1.1)
The analogue on the circle is the conjugate operator C dened by
C f (x) =
1
2
2
_
0
f (y)
tan((x y)/2)
dy. (1.2)
E.M. Stein and G.L. Weiss proved in [6] that the distribution function of the Hilbert transform of a characteristic func-
tion
E
, where E is a measurable set of nite Lebesgue measure, satises

_
x R:

H
E
(x)

>t
_

=
2|E|
sinh(t)
. (1.3)
Similarly, if E is a measurable subset of the circle, they proved that the distribution function of the conjugate function
of its characteristic function satises

_
x [0, 2):

C
E
(x)

>t
_

= 4arctan
_
sin(|E|/2)
sinh(t)
_
. (1.4)
*
Corresponding author.
E-mail addresses: [email protected] (L. Colzani), [email protected] (E. Laeng), [email protected] (L. Monzn).
0022-247X/$ see front matter 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2009.08.017
226 L. Colzani et al. / J. Math. Anal. Appl. 363 (2010) 225229
The above formulas are completely explicit and surprisingly non-sensitive to the detailed structure of the set E. In [5]
L.H. Loomis had previously proven that H is an operator of weak type (1, 1) by analyzing the spacing of the roots of a
suitable polynomial, a technique somewhat similar to the one used in [6] to prove (1.3) and (1.4). In their book [7] Stein
and Weiss give also a different proof, based on complex analytic properties of C. A complex analytic proof of (1.3) can be
found in an article of A.P. Calderon [2, p. 434]. Here we give yet another proof via the following lemmas, that might have
some independent interest.
Lemma 1.1. Let a
1
< b
1
<a
2
< b
2
< <a
n
<b
n
be 2n distinct real numbers, then
+
_

F
_
n

k=1
x b
k
x a
k
_
dx =
_
n

k=1
(b
k
a
k
)
_
+
_

F (x)
(x 1)
2
dx,
where F is any function such that one of the two integrals in the above equality converges.
Lemma 1.2. Let a
1
< b
1
<a
2
< b
2
< <a
n
<b
n
be 2n distinct real numbers and 1 < p < . Then
+
_

H
_
n

k=1

[a
k
,b
k
]
(x)
_

p
dx = (p)
n

k=1
(b
k
a
k
),
where
(p) =
1

p
+
_

| log|x||
p
(x 1)
2
dx = 2p

_
0
t
p1
sinh(t)
dt =
4(1 2
p
)

p
(p) (p +1).
The second expression of is the one used in the Stein and Weiss paper. The third expression, written in terms of the
function of Riemann and the function of Euler, shows that (p) is a meromorphic function with poles at p = 1 and
p = 1, 2, 3, . . . . The fact that is holomorphic in a vertical strip will also play a role later.
The analogue of the above lemma for the conjugate operator is the following.
Lemma 1.3. If E is a nite union of disjoint intervals on the circle and 1 < p < +, then
2
_
0

C
E
(x)

p
dx =
2
_
0

1
2
log
_
1 cos x
1 cos (x |E|)
_

p
dx
= 4p
+
_
0
t
p1
arctan
_
sin(|E|/2)
sinh(t)
_
dt.
Lemma 1.4. Let us denote by f
p
= (
_
X
| f (x)|
p
dx)
1/p
the L
p
normof f in a measure space {X, dx}, by
f
(t) = |{x X: | f (x)| >t}|
the distribution function of f and by f

(s) = inf{t > 0:
f
(t) s} its decreasing rearrangement. Then, the values of f
p
in an
interval of exponents p uniquely determine
f
and f

. In other words, if f
p
= g
p
for p
1
< p < p
2
then
f
(t) =
g
(t) for all
t 0, which is equivalent to f

(s) = g

(s) for all s > 0.

2. Proofs and remarks
Proof of Lemma 1.1. The line y = t intersects the graph of y =

n
k=1
xb
k
xa
k
at n points x
1
(t), x
2
(t), . . . , x
n
(t), which are real
and distinct because our rational function is monotonic in between consecutive pairs of its vertical asymptotes. We can label
these points assuming that x
k
(t) ranges from a
k
to a
k+1
as the corresponding t ranges from to +. The point x
n
(t)
is exceptional and falls either in (, a
1
) or in (a
n
, +) depending on t being above or below the horizontal asymptote
y = 1, but it provides a bijection anyway. Summing up the n changes of variables given by x
1
(t), x
2
(t), . . . , x
n
(t) we obtain
+
_

F
_
n

k=1
x b
k
x a
k
_
dx =
+
_

F (t)
_
x

1
(t) + x

2
(t) + + x

n
(t)
_
dt. (2.1)
Clearly x
1
(t), x
2
(t), . . . , x
n
(t) are the roots of

n
k=1
(x b
k
) = t

n
k=1
(x a
k
), an algebraic equation in x whose leading
terms are (1 t)x
n
+ (

n
k=1
b
k
+t

n
k=1
a
k
)x
n1
+ = 0.
L. Colzani et al. / J. Math. Anal. Appl. 363 (2010) 225229 227
This implies that
x
1
(t) + x
2
(t) + + x
n
(t) =

n
k=1
b
k
t

n
k=1
a
k
1 t
=
n

k=1
a
k
+
1
1 t
n

k=1
(b
k
a
k
).
Therefore x

1
(t) + x

2
(t) + + x

n
(t) =
1
(t1)
2

n
k=1
(b
k
a
k
). Plugging this into (2.1) we obtain the required formula. 2
Proof of Lemma 1.2. A direct computation shows that
H
[a,b]
(x) =
1

b
_
a
dt
x t
dt =
1

log

x a
x b

.
Choosing F (x) = | log|x||
p
in the previous lemma, and observing that F (x) = F (1/x), we obtain
+
_

H
_
n

k=1

[a
k
,b
k
]
_
(x)

p
dx =
+
_

k=1
log

x b
k
x a
k

p
dx
=
1

p
+
_

log

k=1
x b
k
x a
k

p
dx =
1

p
n

k=1
(b
k
a
k
)
+
_

| log|x||
p
(x 1)
2
dx.
The integral in the above formula is convergent for p > 1 and yields the rst expression for (p) in the statement of the
lemma. It can also be written as
+
_

| log|x||
p
(x 1)
2
dx =
+
_
0
| log x|
p
_
1
(x 1)
2
+
1
(x +1)
2
_
dx = 2
+
_
0
| log x|
p
x
2
+1
(x
2
1)
2
dx
= 2
+
_

|t|
p
e
2t
+1
(e
2t
1)
2
e
t
dt =
+
_

|t|
p
e
t
+e
t
2
(e
t
e
t
)
2
4
dt = 2
+
_
0
t
p
cosht
sinh
2
t
dt = 2p
+
_
0
t
p1
sinht
dt,
which gives our second expression for (p). Finally, the third expression is essentially equivalent to the integral represen-
tation of the zeta function in B. Riemann ber die Anzhal der Primzahlen unter einer gegebener Grsse Monatsber (1859):
2p
+
_
0
t
p1
sinht
dt = 4p
+
_
0
e
t
t
p1
1 e
2t
dt = 4p
+
_
0
t
p1
+

n=0
e
(2n+1)t
dt
= 4p
+

n=0
+
_
0
e

_

2n +1
_
p1
d
2n +1
= 4p (p)
+

n=0
1
(2n +1)
p
= 4
_
+

k=1
1
k
p

+

k=1
1
(2k)
p
_
(p +1) = 4(1 2
p
)(p) (p +1). 2
Proof of Lemma 1.3. The proof is similar to the previous one for H, but we will sketch the main steps because the details
are slightly more complicated.
If E =

n
k=1
[a
k
, b
k
] with 0 a
1
<b
1
< <a
n
< b
n
< 2, then
C
E
(x) =
1
2
log
_
n

k=1
1 cos(x a
k
)
1 cos(x b
k
)
_
=
1
2
log
_
n

k=1
sin
2
(
xa
k
2
)
sin
2
(
xb
k
2
)
_
.
This function increases from to + in the interval a
k
< x < b
k
and decreases from + to in b
k
< x < a
k+1
.
The equation t =

n
k=1
sin
2
(
xa
k
2
)
sin
2
(
xb
k
2
)
has 2n solutions {s
k
(t)}
n
k=1
and {d
k
(t)}
n
k=1
with a
1
< s
1
(t) < b
1
< d
1
(t) < a
2
< s
2
(t) < b
2
<
d
2
(t) < for every t 0. Following [6], setting t = e

and taking square roots in the above equation, we see that {s

k
()}
n
k=1
are the n solutions of e
/2
=

n
k=1
sin(
xa
k
2
)
sin(
xb
k
2
)
, while {d
k
()}
n
k=1
are the n solutions of e
/2
= +

n
k=1
sin(
xa
k
2
)
sin(
xb
k
2
)
. Using Eulers
formula, this pair of equations can be rewritten as a pair of polynomial equations of degree n, establishing a connection
228 L. Colzani et al. / J. Math. Anal. Appl. 363 (2010) 225229
between the enpoints of the original intervals [a
k
, b
k
] and their roots. A computation now shows that, if |E| =

n
k=1
(b
k
a
k
),
then
n

k=1
_
s
k
() d
k
()
_
= i log
_
sinh(/2) + i sin(|E|/2)
sinh(/2) i sin(|E|/2)
_
.
This is formula (4.12) in [6]. Differentiating it with respect to we obtain
n

k=1
_
s

k
() d

k
()
_
=
sin(|E|/2) cosh(/2)
sin
2
(|E|/2) +sinh
2
(/2)
.
We conclude the proof with a multiple change of variable which is the analogue on the circle of what we did in
Lemmas 1.1 and 1.2 on the line, namely
C
E

p
p
=
2
_
0

1
2
log
_
n

k=1
sin
2
(
xa
k
2
)
sin
2
(
xb
k
2
)
_

p
dx =
1
(2)
p
+
_

||
p
n

k=1
_
s

k
() d

k
()
_
d
=
+
_

p
sin(|E|/2) cosh(/2)
sin
2
(|E|/2) +sinh
2
(/2)
d = 4
+
_
0
t
p
sin(|E|/2) cosh(t)
sin
2
(|E|/2) +sinh
2
(t)
dt
= 4
+
_
0
t
p
sin(|E|/2) cosh(t)
sinh
2
(t)
1 +
_
sin(|E|/2)
sinh(t)
_
2
dt = 4p
+
_
0
t
p1
arctan
_
sin(|E|/2)
sinh(t)
_
dt.
This proves the nal formula in the statement of the lemma. The previous formula is obtained observing that, since
C
E

p
only depends on |E|, we can substitute the original set E with an interval of the same measure. 2
Proof of Lemma 1.4. The distribution function (t) =
f
(t) = |{x X: | f (x)| > t}| determines the L
p
norms of f via the
formula
f
p
=
__
X

f (x)

p
dx
_
1/p
=
_
p
+
_
0
t
p1
(t) dt
_
1/p
. (2.2)
Vice versa, knowing the L
p
norms of f for all exponents p in an interval (p
1
, p
2
) determines its distribution function .
The idea is that the L
p
norm is the Mellin transform of the distribution function and this transform can be inverted. Let
F (z) =
+
_
0
t
z1
(t) dt =
+
_

e
zs

_
e
s
_
ds.
It is easy to see that if the integral which denes F (p) is absolutely convergent in an interval (p
1
, p
2
), then the above
integral is also absolutely convergent and denes a holomorphic function F (p + iq) in the vertical strip p
1
< p < p
2
and
<q < +. Moreover, F (p +iq) is the Fourier transform evaluated at q of the integrable function e
ps
(e
s
). By inverting
this transform, we can recover (t) from F (p + iq).
Finally, if f (x) and g(x) have the same L
p
(X) norms in an interval of exponents, then the associated functions F (p +
iq) and G(p + iq) are equal on the segment p
1
< p < p
2
and q = 0. Since these functions F and G are holomorphic,
they coincide everywhere. Since F and G determine the corresponding distribution functions, also these distributions must
coincide. 2
Lemmas 1.2 and 1.3, stated for nite unions of disjoint intervals, can be extended to measurable sets of nite Lebesgue
measure with a standard approximation argument that can be found in the original paper of Stein and Weiss [6, p. 271]. In
particular, this implies that both the Hilbert and the conjugate operators are of restricted type (p, p) and, by interpolation,
it leads to a proof of the theorem of M. Riesz on the boundedness of these operators on L
p
, for 1 < p < . Our last lemma
allows to recapture the full strength of the SteinWeiss results (1.3) and (1.4).
It is interesting that the crucial idea in the proof of the rst lemma already appears in a paper by George Boole [1]
(1857). Various versions of this trick have been known even before: For example Cauchy knew that if f is integrable on the
real line then
_
R
f (x) dx =
_
R
f (x 1/x) dx. Integral identities that exploit some suitable n to 1 rational change of variables
appear to have been rediscovered several times by several authors and used for various purposes (see, e.g., M.L. Glasser [3]
and [4]).
L. Colzani et al. / J. Math. Anal. Appl. 363 (2010) 225229 229
References
[1] G. Boole, On the comparison of transcendents, with certain applications to the theory of denite integrals, Philos. Trans. R. Soc. Lond. Ser. 147 (1857)
745803.
[2] A.P. Calderon, Singular integrals, Bull. Amer. Math. Soc. 72 (1966) 427465.
[3] M.L. Glasser, A remarkable property of denite integrals, Math. Comp. 40 (162) (1983) 561563.
[4] M.L. Glasser, Some useful properties of the Hilbert transform, SIAM J. Math. Anal. 15 (6) (1984) 12281230.
[5] L.H. Loomis, A note on Hilberts transform, Bull. Amer. Math. Soc. 52 (1946) 10821086.
[6] E.M. Stein, G. Weiss, An extension of a theorem of Marcinkiewicz and some of its applications, J. Math. Mech. 8 (1959) 263284.
[7] E.M. Stein, G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton University Press, Princeton, NJ, 1971.