Journal of Mathematical Sciences, Vol. 187, No.

1, November, 2012

Fourier transform versus Hilbert transform

Elijah Liflyand

Presented by M. M. Malamud

Dedicated to the 75-th birthday of R. M. Trigub

Abstract. Certain relations between the Fourier transform of a function and the Hilbert transform of
its derivative are revealed. They concern the integrability/non-integrability of both transforms. Certain
applications are discussed.

Keywords. Fourier transform, integrability, Hilbert transform, Hardy space.

1. Introduction

In this note, we are going to compare the Fourier transform of a function and the Hilbert transform
of a related function. For this, let us start with some known results. The first one is given in [11, Th. 2]
(see also [5]). We shall use the following T -transform of a function g(u) defined on (0, ∞):

t/2 
3t/2
g(t + s) − g(t − s) g(s)
T g(t) = ds = ds,
s s−t
0 t/2

where the integral is understood in the improper (principal value) sense, that is, as limδ→0+ δ for the
first integral on the right and with obvious modification for the second one.

Theorem 1.1. Let f : R+ → C be locally absolutely continuous, of bounded variation, and

limt→∞ f (t) = 0. Let also T f  ∈ L1 (R+ ). Then the cosine Fourier transform of f
∞
fc (x) = f (t) cos xt dt (1.1)
0

is Lebesgue integrable on R+ , with

fc L1 (R+ )  f  L1 (R+ ) + T f  L1 (R+ ) . (1.2)

Here and in what follows, we use the notation “  ” and “  ” as abbreviations for “ ≤ C ” and
“ ≥ C ”, with C being an absolute positive constant.
Theorem 1.1, as well as the whole paper [11], was one of the first steps in the project on transferring
the results on the integrability of trigonometric series to Fourier transforms suggested to the author
by R. M. Trigub. For the better understanding of the motivation, we refer the reader to [20].
Translated from Ukrains’kiı̆ Matematychnyı̆ Visnyk, Vol. 9, No. 2, pp. 209–218, April–May, 2012.
Original article submitted September 26, 2011

1072 – 3374/12/1871–0049 
c 2012 Springer Science+Business Media New York 49
 Let us now turn to the Hilbert transform of an integrable function g

1 g(t)
Hg(x) = dt,
π t−x
R

where the integral is also understood in the improper (principal value) sense, now as limδ→0+ |t−x|>δ .
It is not necessarily integrable, and when it is, we say that g is in the (real) Hardy space H 1 (R).
Now, we recall the necessary conditions for the integrability of the Fourier transform and the Hilbert
transform. For the integrability of the Fourier transform (see, e.g., [12]; for its initial form for Fourier
series, see, e.g., [8]), in addition to the obvious uniform continuity of the Fourier transform and its
vanishing at infinity, it is necessary that T f (t) exist and be finite for all t. For a related result, see
also [16, Ch. I, §4.1].
As for the Hilbert transform, if g ∈ H 1 (R), then it satisfies the cancellation property


g(t) dt = 0. (1.3)
R

This cancellation property was apparently first mentioned in [9].

An odd function always satisfies (1.3). However, not every odd integrable function belongs to
H 1 (R) (for a counterexample, see, e.g., [13]). Paley–Wiener’s theorem (see [15]; for the alternative
proof and discussion, see Zygmund’s paper [22]) asserts the following.

Theorem 1.2. If g ∈ L1 (R) is a function that is odd and monotonically decreasing on R+ , then
Hg ∈ L1 , i.e., g is in H 1 (R).

Recently, in [14, Thm. 6.1], this theorem has been extended to a class of functions more general
than monotone ones; for further generalizations, see [13].
The Hilbert transform of odd functions is frequently studied (see, e.g., [10, 3.5] and references
therein). The main goal of this note is to reveal new relations between Theorems 1.1 and 1.2. We also
establish that the belonging of the odd extension of f  to a wider subspace of L1 than H 1 not only
ensures the integrability of the cosine Fourier transform of f but provides a necessary and sufficient
condition for this; see Theorem 3.1.
We restrict ourselves to the cosine Fourier transforms. Similar analysis of the sine Fourier transform
needs a certain additional treatment and will be published elsewhere.

2. Hilbert transform and Fourier transform

Let us extend f  in Theorem 1.1 from R+ to the other half-axis as the odd function Φ, that is,


f  (t), t > 0,
Φ(t) = 
−f (−t), t < 0.

Theorem 2.1. Let a function f (t) defined on [0, ∞) be vanishing at infinity, locally absolutely con-
tinuous, and of bounded variation. Let the cosine Fourier transform of f be non-integrable on [0, ∞).
Then Φ ∈ H 1 (R).

50
Proof. First, let us give a relation for the Hilbert transform of an odd integrable function useful in
applications. It is apparently known but has nowhere been published in this precise form.
Lemma 2.1. Let g be an odd integrable function. Then, for any 0 < a < 1 < b and x > 0,
 bx
1 g(t) 1 g(t)
Hg(x) = dt = dt + Γ(x), (2.1)
π t−x π t−x
R ax

where Γ is such that

∞ ∞
2
|Γ(x)| dx ≤ |g(t)| dt. (2.2)
π min(b − 1, 1 − a)
0 0

Proof. By the oddness, we have, for x > 0,

∞  
−bx ∞  
 g(t) 
 + dt dx
 t−x 
0 −∞ bx
∞  ∞ 
 ∞ t/b
 tg(t)  dx
=2  dt dx ≤ 2 t|g(t)| dt
 t −x
2 2  t − x2
2
0 bx 0 0
∞ t/b ∞
2b2 dx 2
≤ 2 t|g(t)| 2
dt ≤ |g(t)| dt. (2.3)
b −1 t b−1
0 0 0

Similarly,
∞  0 ax 
 ∞
 g(t)  2
 + dt dx ≤ |g(t)| dt. (2.4)
 t−x  1−a
0 −ax 0 0

Finally,
∞  −ax
  ∞ bx ∞ ∞
 g(t)  1 1 2
 dt dx ≤ |g(t)| dt dx ≤ |g(t)| dt ≤ |g(t)| dt. (2.5)
 t−x  x ln(b/a) min(b − 1, 1 − a)
0 −bx 0 ax 0 0

The lemma is proved.

Now, taking a = 1/2 and b = 3/2, we see that the ”leading term” on the right-hand side of (2.1)
1
is π T g(x). By this,

1
Hg(x) = T g(x) + Γ(x).
π
We thus can rewrite the right-hand side of (1.2) as ΦH 1 (R) . This has been observed in [11] and later
on in [5].
Since f (t) is of bounded variation, f  (t) exists almost everywhere and is integrable on [0, ∞). It
now follows from (1.2) that (T f  )(x) is non-integrable. Finally, the above reasoning proves that the
Hilbert transform of Φ is non-integrable, which completes the proof.

51
3. Wider integrability space

The space of integrable functions g with integrable T g is one of the widest spaces the belonging of
the derivative f  to which ensures the integrability of the cosine Fourier transform of f. However, the
possibility of existence (or non-existence) of a wider space of such type is of considerable interest. Let us
show that such a space does exist, moreover, it is the widest possible, that is, provides a necessary and
sufficient condition for the integrability of the cosine Fourier transform. In fact, it has been introduced,
in essence, (for different purposes) in [7] as

1 |
g (x)|
Q = {g : g ∈ L (R), dx < ∞}. (3.1)
|x|
R

With the obvious norm


|
g (x)|
gL1 (R) + dx,
|x|
R

it is a Banach space and ideal in L1 (R). What we will actually use is the space Q0 of the odd functions
from Q; such functions naturally satisfy (1.3).

Theorem 3.1. Let f : R+ → C be locally absolutely continuous, of bounded variation, and

limt→∞ f (t) = 0. Then the cosine Fourier transform of f given by (1.1) is Lebesgue integrable on
R+ if and only if f  ∈ Q0 .

Proof. The assumptions of the theorem give a possibility to integrate by parts. This yields

∞
1 1
fc (x) = − f  (t) sin xt dt = − fs (x).
x x
0

Integrating both sides over R+ completes the proof.

The discussion and comments are in order. At first sight, Theorem 3.1 does not seem to be a result
at all, a technical reformulation of (1.1) at most. This is not the case after the appearance of the
analysis of Q in [7]. Indeed, the well-known extension of Hardy’s inequality (see, e.g., [6, (7.24)])

|
g (x)|
dx  gH 1 (R) (3.2)
|x|
R

implies that H 1 (R) ⊆ Q ⊆ L10 (R), where the latter is the subspace of functions g in L1 (R) which
satisfy the cancellation property (1.3). It is shown in [7], by using the duality arguments, that both
inclusions are proper.
It is doubtful that Q (or Q0 ) may be defined in terms of f itself rather than its Fourier transform.
Therefore, it is of interest to find certain proper subspaces of Q0 , wider than H 1 , the belonging to
which being easily verifiable.

Remark 3.1. It is worth noting that the mentioned embedding immediately proves Theorem 1.1.
The initial proof in [11] is quite complicated.

52
4. Applications

We present possible applications of both Theorems 2.1 and 3.1.

4.1. Application of Theorem 2.1

An example, quite complicated, of a monotone f with non-integrable cosine Fourier transform is
given in [17, Ch. VI, §6.11, Th. 125]. That construction was quite interesting but apparently appeared
since the necessary condition for the integrability of the Fourier transform given in the introduction
had not been known.
Using that necessary condition, we can construct simpler examples. For s > 0, let the even function
f (t) = | ln−s (t − 1)| on (1, 3/2), f (t) = 0 when 0 ≤ t ≤ 1, and, say, infinitely smooth on (1, ∞). The
only problematic point for the necessary condition is x = 1. At this point, that condition reduces to
the convergence/divergence of the integral

1/2
1
dt.
t lns (t)
0

This integral diverges when s ≤ 1. Therefore, the odd integrable function Φ(t), the derivative of f,
which equals 0 when 0 ≤ t ≤ 1, is (very) good on (1, ∞), is (t − 1)−1 | ln−1−s (t − 1)| on (1, 3/2), and
has non-integrable Hilbert transform if 0 < s ≤ 1.

4.2. Application of Theorem 3.1

As in [11], such theorems can be most naturally applied to the integrability of trigonometric series.
Given a trigonometric series
∞

a0 /2 + an cos nx, (4.1)
n=1

find assumptions on the sequences of coefficients {an } under which the series is the Fourier series of
an integrable function.
We can relate this problem to a similar one for Fourier integrals as follows. First, given series (4.1)
with the null sequence of coefficients being in an appropriate sequence space, set for x ∈ [n, n + 1]

A(x) = an + (n − x)Δan , a0 = 0,

where Δan = an − an+1 . So, we construct a corresponding function by means of a linear interpolation
of the sequence of coefficients. Of course, one may interpolate not only linearly, but there are no
problems so far where this might be of importance.
Second, for functions of bounded variation, in order to pass from series to integrals and vice versa,
we will make use of the following result due to Trigub [18, Th. 4] (for its extension, see [19]; an earlier
version, for functions with compact support, is due to Belinsky [3]):
 
 +∞ +∞

 −ixt −ikx 
sup  ϕ(t)e dt − ϕ(k)e   ϕBV . (4.2)
0<|x|≤π  −∞

−∞

53
Relation (4.2) allows us to pass from estimating the trigonometric series (4.1) to estimating the Fourier
transform of A(t). Let us see how an analog of Theorem 3.1 for trigonometric series follows in this
manner.
To this extent, we introduce an analog of the space Q0 for sequences. We denote
 π  ∞ 

  dx
q0 = d = {dn } ∈ 1 :  dn sin xn  <∞ , (4.3)
 x
0 n=1

with
π  
 ∞  dx
dq0 = d1 +  dn sin xn  .
 x
0 n=1

Theorem 4.1. Let {an } be a null sequence of bounded variation. Then (4.1) is the Fourier series of
an integrable function if and only if {Δan } ∈ q0 .
Proof. The proof goes along the above-given lines. With {an } in hand, where a0 = 0 is taken with no
loss of generality, we build A(t). Then, by (4.2),
π   ∞ π  ∞ 

 ∞  1  
 a cos nx  dx
| 
A (x)| dx =  A
(t) sin xt dt  dx.
 n  c
x 
0 n=1 0 0 0

The inner integral on the right is equal to

∞

n+1 ∞
2 sin(x/2)
Δan sin xt dt = Δan sin(n + 1/2)x.
x
n=1 n n=1

Since {an } is a sequence of bounded variation, that is, {Δan } ∈ 1 , we can replace sin(n + 1/2)x on
the right by sin xn, which completes the proof.

In fact, the obtained result has already been pointed out in [1]. However, it was used there as
a technical detail or, more precisely, as a starting point for deriving certain sufficient conditions in
terms of amalgamated space. As above, the present consideration became possible after the mentioned
results in [7]. Of course, as Theorem 1.1 can be readily derived from Theorem 3.1, many known results
on the integrability of trigonometric series immediately follow from Theorem 4.1. We omit the details.

4.3. On the Zygmund–Bochkarev results

Integrability of a Fourier transform is closely related to the absolute convergence of Fourier series.
There are two main types of conditions for the latter (see, e.g., [2, Ch. IX]). One is due to S. N. Bernstein
and states that each periodic function from the Lip α class with α > 12 is expanded in the absolutely
convergent Fourier series. More generally, the next sufficient condition holds true:
1
ω(f, t)
dt < ∞,
t3/2
0

where ω(f, t) is the module of continuity of the function f with step t.

54
 Later on, A. Zygmund discovered a bit different condition: If, in addition, f is of bounded varia-
tion, any positive Lipschitz smoothness ensures the absolute convergence; more precisely, a sufficient
condition is
1 
ω(f, t)
dt < ∞. (4.4)
t
0

Both conditions are sharp on the corresponding whole class (see, e.g., [8]). However, if, for Bernstein-
type results, the sharpness has been proved in various ways (for instance, an example of a function
from Lip 12 whose Fourier series is not absolutely convergent was built by Bernstein by using the Leg-
endre symbols), the only counterexample in the case where (4.4) is not valid for a function of bounded
variation and its Fourier series does not converge absolutely has been given by Bochkarev in a very
sophisticated way [4]. Of course, Bochkarev’s construction allows one to build a counterexample of a
function with the given module of continuity. This is of importance, when the class defined by (4.4) is
studied. However, the whole class of even functions of bounded variation with integrable Fourier trans-
form (or, more or less equivalently, for functions with compact support, with absolutely convergent
Fourier series) is wider — just Q0 , for which our ansatz gives a possibility to construct counterexamples
easier. We observe firstly that Theorem 2.1 provides not only the examples of functions of bounded
variation with f  ∈ H 1 but also with the same f  ∈ Q0 . Since, by Theorem 3.1, the latter is necessary
and sufficient, for such examples (4.4) is definitely not valid. Finally, if one wishes to have a counterex-
ample for the Fourier series rather than for the Fourier transform, it remains to take a corresponding
f to be supported on [0, a], a < π. In virtue of the known Wiener’s result [21, Ch. II, §11], for such
functions, the absolute convergence of their Fourier series (after π-periodization) is equivalent to the
integrability of the Fourier transform.

REFERENCES
1. B. Aubertin and J. J. F. Fournier, “Integrability theorems for trigonometric series,” Stud. Math., 107, 33–59
(1993).
2. N. K. Bary, A Treatise on Trigonometric Series, Macmillan, New York, 1964.
3. E. S. Belinsky, “On asymptotic behavior of integral norms of trigonometric polynomials,” in Metric Ques-
tions of the Theory of Functions and Mappings [in Russian], Naukova Dumka, Kiev, 1975, pp. 15–24.
4. S. V. Bochkarev, “On a problem of Zygmund,” Math. USSR-Izv., 7, 629-–637 (1973).
5. S. Fridli, “Hardy Spaces Generated by an Integrability Condition,” J. Approx. Theory, 113, 91–109 (2001).
6. J. Garcia-Cuerva and J. L. Rubio de Francia, Weighted Norm Inequalities and Related Topics, North-
Holland, Amsterdam, 1985.
7. R. L. Johnson and C. R. Warner, “The convolution algebra H 1 (R),” J. Funct. Spaces Appl., 8, 167–179
(2010).
8. J.-P. Kahane, Séries de Fourier Absolument Convergentes, Springer, Berlin, 1970.
9. H. Kober, “A note on Hilbert’s operator,” Bull. Amer. Math. Soc., 48:1, 421–426 (1942).
10. V. Kokilashvili and M. Krbec, Weighted Inequalities in Lorentz and Orlicz Spaces, World Scientific, Singa-
pore, 1991.
11. E. Liflyand, “Fourier transforms of functions from certain classes,” Anal. Math., 19, 151–168 (1993).
12. E. Liflyand and S. Tikhonov, “The Fourier transforms of general monotone functions,” in Analysis and
Mathematical Physics, Birkhäuser, Basel, 2009, pp. 373–391.

55
13. E. Liflyand and S. Tikhonov, “Weighted Paley–Wiener theorem on the Hilbert transform,” C. R. Acad. Sci.
Paris, Ser. I, 348, 1253–1258 (2010).
14. E. Liflyand and S. Tikhonov, “A concept of general monotonicity and applications,” Math. Nachr., 284,
1083–1098 (2011).
15. R. E. A. C. Paley and N. Wiener, “Notes on the theory and application of Fourier transforms,” Trans.
Amer. Math. Soc., 35, 348–355 (1933).
16. E. M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton Univ. Press,
Princeton, 1971.
17. E. C. Titchmarsh, Introduction to the Theory of Fourier Integrals, Clarendon Press, Oxford, 1937.
18. R. M. Trigub, “Multipliers of Fourier series and approximation of functions by polynomials in spaces C and
L,” Soviet Math. Dokl., 39, 494–498 (1989).
19. R. M. Trigub, “A Generalization of the Euler-Maclaurin formula,” Math. Notes, 61, 253–257 (1997).
20. R. M. Trigub and E. S. Belinsky, Fourier Analysis and Appoximation of Functions, Kluwer, Dordrecht,
2004.
21. N. Wiener, The Fourier Integral and Certain of Its Applications, Dover, New York, 1959.
22. A. Zygmund, “Some Points in the theory of trigonometric and power series,” Trans. Amer. Math. Soc., 36,
586–617 (1934).

Elijah Liflyand
Department of Mathematics, Bar-Ilan University,
52900 Ramat-Gan, Israel
E-Mail: liflyand@math.biu.ac.il

56