Real Analysis Exchange

Vol. 29(2), 2003/2004, pp. 557–587

F. Bayart, UFR de Mathématiques, UMR AGAT CNRS, Université des

Sciences et Technologies de Lille, F-59665, Villeneuve d’Ascq cedex, France.
email: [email protected]
S. V. Konyagin, Department of Mechanics and Mathematics, Moscow State
University, Moscow 119992, Russia. email: [email protected]
H. Queffélec, UFR de Mathématiques, UMR AGAT CNRS, Université des
Sciences et Technologies de Lille, F-59665, Villeneuve d’Ascq cedex, France.
email: [email protected]

CONVERGENCE ALMOST EVERYWHERE

AND DIVERGENCE EVERYWHERE OF
TAYLOR AND DIRICHLET SERIES

Abstract
Recent results concerning the convergence almost everywhere or di-
an nit appeared in the liter-
P
vergence everywhere of Dirichlet series
ature, Prevealing significant differences with the case of trigonometric
series an eint . In this work, we prove in several cases the optimality
of these results. We P also discuss the statistical effect of a change of
signs, by considering ±an nit . According to the way (probabilistic or
topological) this change of signs is made, the properties of the resulting
series are quite different, and can also be applied to the theory of power
series.

1 Introduction
Dvoretzky and Erdös ([DE]) proved the following.
P+∞
(1) If 0 |an |2 = ∞, and if |an | is non-increasing,
P∞ then there exists a
choice of signs εn = ±1 such that the series 0 εn an eint diverges for
each t ∈ R. We may even have unbounded divergence everywhere.

In a second paper, they gave an idea of the size of “admissible signs” in (1).
Key Words: Dirichlet series, quasi-sure property, everywhere divergence
Mathematical Reviews subject classification: 42A32, 42A20, 42A61
Received by the editors March 4, 2003
Communicated by: Alexander Olevskii

557
558 F. Bayart, S. V. Konyagin and H. Queffélec

PN
(2) If lim log1 N 1 |an |2 > 0, then for almost all choices of signs εn = ±1,
P∞
the series 0 εn an eint diverges everywhere, and the condition on (an )
is optimal.
P∞ 2
On the other hand, Carleson ([Ca]) proved that, if 0 |an | < ∞, then
P ∞ int
0 an e converges for almost all t (with respect to Lebesgue measure).
Later, Hedenmalm and Saksman ([HS]) extended Carleson’s result to Dirichlet
series as follows.
P∞ P∞
(3) If 0 |an |2 < ∞, then 1 an n−1/2+it converges for almost all t.

A simplified proof of (3) was given in [KQ], as well as an extension of (1) to

Dirichlet series, as follows.
P∞
(4) If 1 |an |2 = ∞, and if n1/2 |an | is non-increasing,
P∞ then there exists a
choice of signs εn = ±1 such that the series 1 εn an n−1/2+it diverges
unboundedly everywhere.

The aims of this paper are the following: First, we extend (2) to Dirichlet
series, with a change of scale, and show the optimality of the result. Second, we
show that (3) P is optimal in two respects:
P∞ denoting by H the space of Dirichlet
∞
series f (s) = 1 an n−s P∞such that 1 |an |2
= kf k2 < ∞, we show that, for
−1/2+it
some f ∈ H, the series 1 an n may diverge for t ∈ E, a prescribed set
of measure zero on the line. Moreover, even if f belongs to the smaller space
H∞ (to be defined later), the translation 1/2 P∞ cannot be completely dispensed
with. Third, we extend (1) to show that 0 εn an eint diverges everywhere
for quasi-all choices of signs, and give a similar extension for Dirichlet series.
This topological point of view turns out to be better adapted if we want to
exhibit Taylor series with a strange behavior (see Theorem 4.3). Finally, part
5 is devoted to some concluding remarks and questions.
A word of explanation is necessary for the terms “almost all” and “quasi-
all” which we used. Let Ω = {−1, 1}N be the set of all choices of signs
ω = (εn (ω))n≥0 , equipped with its natural topology (product of the discrete
topology on {−1, 1}) and with its natural probability (product of the probabil-
ities 21 (δ−1 + δ1 ) on each factor), for which the variables ω 7→ εn (ω) appear as
independent random variables such that P (εn = 1) = P (εn = −1) = 1/2. Ω is
then a Baire space (since it is compact) and a probability space. A property
(Q) will be said to hold:

a) For almost all choices of signs if it holds for ω ∈ A, with P (A) = 1.

b) For quasi-all choices of signs if it holds for ω ∈ A, a dense Gδ subset of Ω.

Convergence and Divergence of Taylor and Dirichlet Series 559

2 Almost Sure Divergence Everywhere of Dirichlet Se-

ries
In this section, we shall prove the following result.

Theorem 2.1. Let (an )n≥1 be a sequence of complex numbers.

N
1 X
a) If limN →∞ |an |2 = γ > 0 (with γ = ∞ allowed), then for
log log N 1
X∞
almost all choices of signs εn = ±1, the series εn an nit diverges for
1
each t ∈ R.
>
b) The result of a) is optimal in the following sense : if δN − → 0, there exists
N
1 X
a sequence (an ) such that limN →∞ |an |2 > 0, but for each
δN log log N 1
∞
X
ω ∈ Ω, the series εn (ω)an nit converges for at least one t ∈ R.
1

Preliminary comment: In this statement, the translation 1/2 of (3) and (4)
has disappeared, which is predictable for the following reason: Suppose that
∞
X √
the hypotheses are fulfilled, but that n−1 |an |2 < ∞ (e.g., an = j if n =
1
∞
j
X
22 , an = 0 otherwise). Then, for each t ∈ R, the series εn (ω)an n−1/2+it
1
converges for almost all ω ∈ Ω by the three series Theorem [Ka]; now, by
Fubini’s Theorem, this implies that, almost surely with respect to ω, the series
X∞
εn (ω)an n−1/2+it converges for almost all t. Divergence for every t cannot
1
∞
X
take place for the series εn (ω)an n−1/2+it .
1
Proof. (of Theorem 2.1) We begin with b); M. Weiss [W] proved that, if
∞
X n
bn → 0, then bn ei2 t converges for some real t. We will use this result as
0
j
follows. Set Nj = 22 , δj∗ = supn≥Nj δn , bj =
p ∗
δj , and take

an = bj if n = Nj ; an = 0 if n ∈
/ {N0 , N1 , . . . }.
560 F. Bayart, S. V. Konyagin and H. Queffélec

Observe that the hypothesis of b) holds for an ; in fact, if Nj ≤ N < Nj+1 ,

one has
N j
1 X C X ∗ C
|an |2 ≥ ∗ δk ≥ ∗ jδj∗ = C,
δN log log N 1 δj × j jδj
k=1
where C > 0 is a numerical constant. And if εn = ±1, setting θj = εNj , we
have
∞ ∞
X X j
εn an nit = θj bj ei2 (t log 2) ,
n=1 j=0

and since θj bj → 0, this series converges for some t, by the result of M. Weiss.
The proof of a) is more difficult, although we follow more or less closely
the trigonometric case; we first assume that γ < ∞. Let l > 0 to be adjusted
later; define an increasing sequence (pk ) of integers by p0 = 0, and by the
condition that pk+1 is the first integer > pk such that
X
|an |2 > l. (2.1)
pk <n≤pk+1

∞ pk
X 1X
This is possible since |an |2 = ∞. Il follows that |an |2 → l (without
1
k 1
loss of generality, we can assume that an → 0). Let now (Ns ) be an increasing
Ns
1 X
sequence of integers such that |an |2 → γ as s → ∞, and Ks be
log log Ns 1
defined by pKs ≤ Ns < p1+Ks . Observe that
 
pKs Ns
1 X 1 X X
|an |2 ≥  |an |2 − |an |2 
log log pKs 1 log log Ns 1 pKs <n<p1+Ks
N
!
s
1 X
2
≥ |an | − l ,
log log Ns 1

so that
pK Ns
!
1 Xs log log pKs 1 X
|an |2 ≥ 2
|an | − l .
Ks 1 Ks log log Ns 1
log log pKs
Letting s tend to infinity now gives l ≥ γlim . One thus can find an
Ks
infinite set J of positive integers such that
 
log log pj 2l 2jl
j ∈ J =⇒ ≤ =⇒ log pj ≤ exp . (2.2)
j γ γ
Convergence and Divergence of Taylor and Dirichlet Series 561

∞
X
Now, we fix T > 0, and will prove that, almost surely, εn (ω)an nit diverges
1
for each t ∈ [−T, T ]; this will clearly imply a), and the proof will be based on
the following crucial Lemma of minimax.
 1/2
X
Lemma 2.2. Set qj =  |an |2  and
pj <n≤pj+1

X
Bj (t) = Bj (t, ω) = εn (ω)an nit .
pj <n≤pj+1

Suppose that µ + 1 ∈ J. Then, one has

! !
|Bj (t)|
P inf sup ≤ 1 ≤ λT Clµ (1 − B)µ , (2.3)
|t|≤T µ
2 <j≤µ
q j /4

where 0 < B < 1 is a numerical constant, Cl > 1 is a constant that tends to

>
one as l −
→ 0, and λT > 0 only depends on T .
Proof of the Lemma. We shall discretize the infimum over t, and shall
need for that an estimate for the derivative of Bj , under the form of the fol-
lowing sublemma.
Pν it
Sublemma. Let Q(t) = n=1 εn bn n , ν ≥ 2, be a random Dirichlet poly-
2 1/2
Pν

nomial, and let T ≥ 1; set kf kT for sup|t|≤T |f (t)|, and kbk2 = 1 |bn |
for b = (b1 , . . . , bν ). Then (E denoting expectation) one has
p
E (kQkT ) ≤ CT0 kbk2 log ν, (2.4)

where the constant CT0 only depends on T .

Proof of the Sublemma. Let N ≥ 2 be an integer to be chosen later, and

tk = kT T
N , −N ≤ k ≤ N ; (tk ) is a N −net of [−T, T ]. According to a well-known
estimate of Salem and Zygmund [SaZ], we have that
!
p
E sup |Q(tk )| << kbk2 log N ,
|k|≤N

where AN << BN means that there exists a constant κ, which does not depend
T
on N , such that AN ≤ κBN . Moreover, if t ∈ [−T, T ] and |t − tk | ≤ N , one
562 F. Bayart, S. V. Konyagin and H. Queffélec

Pν
has |Q(t) − Q(tk )| ≤ |t − tk |kQ0 k∞ ≤ N
T
1 |bn | log n ≤
T 1/2
Nν log νkbk2 , by
the Cauchy-Schwarz inequality. Therefore,

T ν 1/2 log ν
kQkT ≤ sup |Q(tk )| + kbk2 ,
|k|≤N N

and
T ν 1/2 log ν
 
p
E (kQkT ) << kbk2 log N + .
N
 
The choice N = T ν 1/2 log ν + 1 (where [.] stands for the integer part) gives
the result, observing that log(T ν) ≤ log(ν T ) ≤ T log ν.

Let us return to the proof of the Lemma. We! abbreviate inf |t|≤T as inf t ,
|Bj (t)|
and denote by Eµ the event inf sup ≤ 1 appearing in the left
t µ
2 <j≤µ qj /4
kT
hand-side of (2.3). Let N ≥ 2 to be chosen later, tk = N , −N ≤ k ≤ N and
(µ being kept fixed) FN be the event
!
|Bj (tk )|
inf sup ≤1
|k|≤N µ
2 <j≤µ
qj /2

and GN be the event

kBj0 kT N
sup > .
µ
2 <j≤µ
qj /4 T

First, we observe that

Eµ ⊂ FN ∪ GN . (2.5)
T
In fact, if ω ∈ FNc ∩ GcN , and if −T ≤ t ≤ T , let k be such that |t − tk | ≤ N,
q
and j ∈] µ2 , µ] be such that |Bj (tk )| > 2j . Then

|Bj (t)| ≥ |Bj (tk )| − |Bj (t) − Bj (tk )|

≥ |Bj (tk )| − |t − tk |kBj0 kT
qj T qj N qj
> − = ,
2 N 4 T 4
so that ω ∈ Eµc . Second, we will show that

CT00
 
4l
P (GN ) ≤ µClµ , where Cl = exp . (2.6)
N γ
Convergence and Divergence of Taylor and Dirichlet Series 563

X
In fact, setting Ij =]pj , pj+1 ], we have Bj0 (t) = (i log n)an εn nit , and the
n∈Ij
sublemma gives
 1/2
X 3/2
E kBj0 kT ≤ CT0 log pj+1  log2 n|an |2  ≤ CT0 (log pj+1 ) qj ,

p

n∈Ij

so that by Markov’s inequality

 
X qj N
P (GN ) ≤ P kBj0 kT >
µ 4 T
2 <j≤µ
X 4T 0 3/2
≤ C qj (log pj+1 )
µ N qj T
2 <j≤µ

4T 0 3/2
≤ C µ (log pµ+1 )
N T
4T CT0
 
3(µ + 1)l
≤ µ exp ,
N γ

where we used (2.2) since µ + 1 ∈ J. This gives (2.6), with CT00 = 4T CT0 .
Finally, we will show that

P (FN ) ≤ 3N (1 − δ)µ , where 0 < δ < 1 is a numerical constant. (2.7)

 
[ \
In fact, FN =  Fj,k , where Fj,k is the event
µ
|k|≤N 2 <j≤µ

n qj o
Fj,k = ω; |Bj (tk )| ≤ .
2
For fixed k, the Fj,k are independent, since the blocks Bj have pairwise disjoint
supports Ij . Therefore we have
X Y
P (FN ) ≤ P (Fj,k ).
µ
|k|≤N 2 <j≤µ

c
We shall majorize P (Fj,k ) by minorizing P (Fj,k ) with the help of the Paley-
Zygmund inequality ([Ka])

(E(X))2
P (X ≥ λE(X)) ≥ (1 − λ)2 ,
E(X 2 )
564 F. Bayart, S. V. Konyagin and H. Queffélec

for 0 < λ < 1 and X a positive random variable. This gives

   
c 1 1
P (Fj,k ) =P |Bj (tk )| > qj = P |Bj (tk )|2 > E(|Bj (tk )|2 )
2 4
2 2


9 E(|Bj (tk )| ) 1
≥ 4
≥ = δ0 ,
16 E(|Bj (tk )| ) 16
by Khintchine’s inequality ([Dis]). Therefore
X
P (FN ) ≤ (1 − δ 0 )µ/2 ≤ 3N (1 − δ)µ ,
|k|≤N
√
where δ is defined by 1 − δ 0 = 1 − δ. It now follows from (2.5), (2.6) and
(2.7) that
CT00
P (Eµ ) ≤ P (FN ) + P (GN ) ≤ 3N (1 − δ)µ + µClµ .
N
We optimize this inequality with the choice
h
1/2 i
N = CT00 µClµ (1 − δ)−µ + 1,

and this gives (2.3) of Lemma 2.2 with an appropriate numerical constant
B ∈]0, 1[, and with an appropriate λT .
Lemma 2.2 easily gives the conclusion in a) of Theorem 2.1. Let us first fix
>
l > 0 so small that Cl (1 − B) = q < 1; this is possible since Cl → 1 as l −
→ 0.
µ
Then, (2.3) gives PP(Eµ ) ≤ λT q (once and for all, we restrict

ourselves to
µ + 1 ∈ J), so that µ P (Eµ ) < ∞, implying P limµ Eµc = 1, by the Borel-
Cantelli Lemma. Equivalently, there exists Ω0 ⊂ Ω, with P (Ω0 ) = 1, such
that for ω ∈ Ω0 and for µ large enough (depending on ω), one has
!
|Bj (t, ω)|
inf sup > 1.
|t|≤T µ
2 <j≤µ
qj /4

Fix t ∈ [−T, T ], and ω ∈ Ω0 . We can find arbitrarily

√
large µ, and arbitrarily
qj l
large j ∈]µ/2, µ], such that
P |Bj (t, ω)|it > 4 > 4 (in view of (2.1)). Since Bj
is a block of the series εn (ω)an n , this proves that the Cauchy criterion
fails for this series, which is therefore divergent.
N
1 X
If we now suppose that limN →∞ |an |2 = +∞, then for almost
log log N 1
X∞
all choices of signs εn = ±1, the series εn an nit is unboundedly divergent
1
Convergence and Divergence of Taylor and Dirichlet Series 565

for each t ∈ R. Indeed, in this case, the previous proof gives, for any 0 <
γ < ∞, a subset Ωγ of Ω with P (Ωγ ) = 1 such that, for each t ∈ [−T,√T ]
and each ω ∈ Ωγ , there exists arbitrarily large j such that Bj (t, ω) > 4l ,
with l = κγ, where  κ is a numerical constant (l just depends on γ by the
4l
condition exp γ (1 − B) < 1, where B is a numerical constant). Setting
T
γn = n, and Ω∞ = n Ωγn , for each t in [−T, T ] and each ω in Ω∞ , the series
P ∞ it
1 ε n (ω)an n is unboundedly divergent.

3 Optimality of the Hedenmalm-Saksman Theorem; the

Case of H∞ .
In this section, we are going to prove the optimality of (3) in the introduction
and, answering a question of Hedenmalm ([He]), we will show that, even for the
smaller space H∞ , the translation factor n−1/2 , or at least some translation
factor n−ε , cannot be dispensed with.

Theorem 3.1. Let E ⊂ PR be a set of Lebesgue-measure zero. Then, there

∞
exists a Dirichlet series 1 an n−s such that:
P∞ 2
a) 1 |an | < ∞.
P∞ −1/2+it
b) 1 an n diverges for t ∈ E.

Proof. In [KQ] (see c) of Theorem 2.1), to prove a Dvoretzky-Erdös type

result, we had successively built:

i) a sequence of blocks Bk = pk <n≤pk+1 |an |n−1/2 of the divergent series

P

|an |n−1/2 , such that 1 < Bk ≤ 2, and that the sequence of integers
P
∞
X 1
(pk ) is not too lacunary: pk+1 = ∞.
1
log pk

1
ii) a sequence (lk ) of “lengths” adapted to the the pk ’s: lk = , so
4 log pk+1
pk
P∞
that 1 lk = ∞.

iii) a sequence (∆k ) of intervals of length 2lk , covering R infinitely many

times.
X
iv) a sequence of blocks Dk (t) = εn an n−1/2+it , with Ik =]pk , pk+1 ], εn =
n∈Ik
±1, |Dk | big at the center of ∆k and oscillating little on ∆k .
566 F. Bayart, S. V. Konyagin and H. Queffélec

Here, we use exactly the same ingredients, but in a different order. Since E
is negligible, we can find a sequence of closed intervals ∆k = [tk − lk ; tk + lk ]
such that 0 < lk < 1/4 and that:
∞
X
lk < ∞ (3.1)
1
Each point of E belongs to infinitely many ∆k ’s. (3.2)

(This should be thought of as a converse of the Borel-Cantelli Lemma). Now,

motivated by ii), we define inductively an increasing sequence of positive in-
tegers by h i
p1 = 4; pk+1 = pk e1/4lk , k ≥ 1. (3.3)

We denote by Ik the block of integers ]pk , pk+1 ] and define a sequence (bn ) of
positive numbers (the moduli of the an ’s) by

bn = 0 for n = 1, . . . , 4; bn = lk n−1/2 for n ∈ Ik , k ≥ 1.

We claim that:
∞
X
b2n < ∞ (3.4)
1
X 1
n−1/2 bn ≥ for k large enough (k ≥ k0 ). (3.5)
16
n∈Ik

In fact,
X X 1 pk+1 lk
b2n = lk2 ≤ lk2 log ≤
n pk 4
n∈Ik pk <n≤pk+1
P∞
in view of (3.3), and (3.4) follows since 1 lk < ∞. Moreover
X X 1 lk pk+1 lk 1 1
n−1/2 bn = lk ≥ log ≥ =
n 2 pk 2 8lk 16
n∈Ik n∈Ik

for k ≥ k0 . We can nowXadjust blockwise the real signs εn = ±1 in such a way

that, setting Dk (t) = εn bn n−1/2+it , we have
n∈Ik

1 X 1
|Dk (tk )| ≥ bn n−1/2 ≥ , (3.6)
2 32
n∈Ik
Convergence and Divergence of Taylor and Dirichlet Series 567

for k ≥ k0 . Due to the fact that |Dk | oscillates little on the small interval ∆k
around tk , it follows that
1
|Dk (t)| ≥ for t ∈ ∆k and k ≥ k0 . (3.7)
64
Indeed, setting
 it
X n
Ck (t) = Dk (t)p−it
k = ε n bn n −1/2
,
pk
n∈Ik
X
and Sk = bn n−1/2 , we have for t ∈ ∆k and k ≥ k0
n∈Ik
   
|Dk (t)| − |Dk (tk )| = |Ck (t)| − |Ck (tk )|
≤ |Ck (t) − Ck (tk )|
X n pk+1 Sk
≤ bn n−1/2 |t − tk | log ≤ lk log Sk ≤ ,
pk pk 4
n∈Ik

so that
Sk Sk Sk Sk 1
|Dk (t)| ≥ |Dk (tk )| −≥ − = ≥ ,
4 2 4 4 64
in view of (3.5) and (3.6). Finally, set an = εn bn ; (3.2), (3.4) and (3.7) show
that the sequence (an ) has the required properties.
Remark. Let now E be a null-set of the circle. It is known ([KK]) that there
is a continuous function whose Fourier series diverges on E. The method of
proof of the previous Theorem can be used to obtain in a simple way such
a function f , maybe not continuous, but in H 2 . Just define arcs ∆k on the
circle satisfying
h (3.1)
i and (3.2). Inductively define integers pk by p1 = 1 and
1
pk+1 = pk + 4lk . Set Ik =]pk , pk+1 ], bn = 0 for n = 0, . . . , 4, bn = lk for
 
X 
intk  1
n ∈ Ik . Adjust real signs εn = ±1 blockwise so that  ε n bn e  ≥ for

  8
n∈Ik
k large enough. This implies, as for Dirichlet series, that
 
X 
int  1
ε n bn e  ≥ for t ∈ ∆k ,

 16


n∈Ik

∞
X
so that f (t) = εn bn eint has the required properties.
0
568 F. Bayart, S. V. Konyagin and H. Queffélec

P∞Now −s let us turn to the case of H∞ , the space of Dirichlet series f (s) =
1 an n , with convergence and boundedness of f in the half-plane C0 =
{s ∈ C; <(s) > 0}. We have the inclusion H∞ ⊂ H, with contraction of
norms: f ∈ H∞ =⇒ kf kH ≤ kf k∞ = sup<(s)>0 |f (s)|. More precisely
(see [HLS]), H∞ is the set of multipliers of H. So, the space H∞ should be
considered as a much smaller space than H. In spite of this, we have the
following result, which answers in the negative a question of H.Hedenmalm
([He]).
P∞
Theorem
P∞ 3.2. There exists a Dirichlet series f (s) = 1 an n−s ∈ H∞ , such
that 1 an nit diverges for each t ∈ R. We even have that f is continuous on
the closed half-plane C0 .

Proof. We modify a construction due to Lick ([Li]). The following well-

known inequality ([B, vol.1, p.90]) is crucial:
 
N
X sin nt 
 ≤ A, ∀t ∈ R, ∀N ≥ 1, (3.8)

n 


n=1

A being a constant. We will use three sequences (hk ), (Hk ), and (nk ) having
the following properties:

1. (hk ) is an increasing sequence of integers such that k≥1 log1hk < ∞ (one
P
2
can take hk = 2k ).

2. (Hk ) is a sequence of positive real numbers such that

 
1
a) exp is rational.
Hk
b) Hk ≥ 4khk .

Such a sequence clearly exists.

3. (nk ) is a sequence of integers such that:

 
l
a) nk exp is an integer for 1 ≤ l ≤ 2hk .
Hk
 
2hk
b) nk+1 > nk exp .
Hk
X ∞
c) n−σ
k < ∞ for each σ > 0.
k=1
Convergence and Divergence of Taylor and Dirichlet Series 569

 
1 pk
If exp = , one can for example take nk = ck (qk )2hk , the integers ck
Hk qk
being chosen so large that b) and c) hold. Those three sequences being fixed,
let us set, for k = 1, 2, . . . :
1 e−s e−(hk −1)s e−(hk +1)s e−2hk s
Pk (s) = + + ··· + − − ··· −
hk hk − 1 1 1 hk
hXk −1 h −1
e−js X e−(2hk −j)s
k

= − (3.9)
j=0
hk − j j=0
hk − j
hk
X e−(hk −n)s − e−(hk +n)s
= .
n=1
n

For s = it purely imaginary, one clearly has

hk
X sin nt
Pk (s) = 2ie−ihk t ,
n=1
n

so that (3.8) and the maximum modulus principle give

|Pk (s)| ≤ 2A, for each s such that <(s) > 0. (3.10)

We now modify the definition of Pk to obtain a Dirichlet polynomial, by setting

 
1 s
Qk (s) = Pk n−s
k .
log(hk ) Hk
Equivalently,

k −1
hX    −s
1 1 j
Qk (s) =  exp nk
log(hk ) j=0 hk − j Hk

hXk −1    −s
1 2hk − j
− exp nk .
j=0
h k − j Hk

Condition 3.a) guarantees that the Qk ’s are Dirichlet polynomials, while con-
dition 3.b) implies that their spectra are disjoint. We can therefore consider
the Dirichlet series
∞
X X∞
f (s) = Qk (s) = an n−s ,
k=1 1

which will be the required counterexample. In fact:

570 F. Bayart, S. V. Konyagin and H. Queffélec

P∞
• If s = σ + it, with σ > 0, n=1 an n−s is absolutely convergent.
To prove this, denote by Q∗k (s) the sum of the moduli of the terms
appearing in Qk (s). We have


k −1 
hX
    −s 
1 1 j
Q∗k (s) = exp nk
  
log(hk ) j=0  hk − j Hk
 


k −1 
hX
    −s 
 1 2hk − j
+ exp nk

 hk − j Hk
 

j=0
hk
2 X 1
≤ n−σ
k ≤ 4n−σ
k ,
log hk j=1
j

and condition 3.c) ensures that

∞
X ∞
X ∞
X
|an n −s
|= Q∗k (s) ≤4 n−σ
k < ∞.
n=1 k=1 k=1

• f ∈ H∞ and f is continuous on C0 .
To prove this, simply note that, in view of (3.10):

2A
<(s) ≥ 0 =⇒ |Qk (s)| ≤ ,
log hk

so that the series defining f converges normally on C0 , and that, for

s ∈ C0 , one has
∞
X 2A
|f (s)| ≤ = M < ∞.
log hk
k=1

∞
X
• an nit diverges for each t ∈ R.
1
hk |t|
To prove this, take t ∈ R. For k large enough, one has ≤ 1, in view
Hk
Convergence and Divergence of Taylor and Dirichlet Series 571

of 2.b). So that, considering the “first half” of Qk , one has

k −1
hX    it 
 1 1 j
exp nk
 
log hk j=0 hk − j Hk
 

k −1
hX  
 1 1 ijt 
= exp

log hk j=0 hk − j Hk


k −1
hX  
1 1 jt
≥ cos
log hk j=0 hk − j Hk
k −1
hX
1 1
≥ cos 1 ≥ δ > 0,
log hk j=0 hk − j

where δ is an absolute constant (we used the fact that, for 0 ≤ j ≤ hk −1,
 ≤ hk |t| ≤ 1). This shows that P∞ an nit fails to verify
 jt 
one has  1
Hk  Hk
the Cauchy criterion and therefore diverges.
Remarks.
an nit is boundedly divergent at each real
P
1. In our example, the series
number t. In fact, for each integer N , there exists an integer r such that
N
X
an nit = Q1 (−it) + · · · + Qr (−it) + αr ,
1

where αr is a partial sum of Qr+1 (−it), so that

N 
X 
an nit  ≤|Q1 (−it)| + · · · + |Qr (−it)| + Q∗r+1 (−it)
 

 
1
∞
X 1
≤2A + 4.
log hk
k=1

This is in contrast with what happens for FourierPseries. According to

∞ int
a Theorem of Marcinkiewicz ([Kö]), if an → 0, 0 an e cannot be
boundedly divergent everywhere. We shall return to this in the last
section.
P∞
2. If one considers general Dirichlet series f (s) = n=1 an e−λn s , conver-
gent and with bounded sum f in the half-plane C0 , and if one looks at
572 F. Bayart, S. V. Konyagin and H. Queffélec

the convergence on the boundary iR of C0 , three fairly distinct behaviors

appear:
a) If λn = n, we have convergence almost everywhere on the boundary
by Carleson’s Theorem.
b) If λn = log n, we may have everywhere divergence on iR, as indicated
by Theorem 3.2.
c) If λn = log pn , where pn is the nth prime number, Kronecker’s The-
orem implies that we have absolute convergence everywhere on the
boundary. (More generally, Bohr’s Theorem implies that, in all the
X∞
examples of Theorem 3.2, we necessarily have |apn | < ∞.
n=1

4 Quasi-Sure Divergence Everywhere of Taylor and

Dirichlet Series; an Application
In this section, we will consider the set of “admissible” signs ω = (εn (ω)) in
(1), . . . , (4) from the topological, i.e. quasi-sure, point of view, which turns
out to be better adapted to some counterexamples. Let X be a σ−compact
topological space; i.e., X = ∪∞ k=1 Xk , where (Xk ) is an increasing sequence of
compact subsets of X. We first prove a general theorem (Recall that a series
is said to be unboundedly divergent if its partial sums are unbounded.) for a
series (fn )n≥1 of continuous functions :X → C. It is convenient to set

Sn = f1 + · · · + fn , δN (t) = sup |Sn (t) − SN (t)| and δ(t) = lim inf δN (t).
n>N N →∞

Theorem 4.1. Let (fn )n≥1 be a sequence of continuous functions :X → C.

Then we have the following:
a) If δ(t) = ∞ for all t, then for quasi-all choices of signs ω, the series
X∞
εn (ω)fn (t) is unboundedly divergent everywhere on X.
1

b) If there is a continuous function ϕ such that δ(t) > ϕ(t) > 0 for all t,
∞
X
then for quasi-all choices of signs ω, the series εn (ω)fn (t) is divergent
1
everywhere on X.
Convergence and Divergence of Taylor and Dirichlet Series 573

Proof.
a) We shall construct an increasing sequence (Nk )k≥0 of integers such that
 
∀t ∈ Xk , sup Sn (t) − SN (t) > k.
k−1
(4.1)
Nk−1 <n≤Nk

Take N0 = 0. We assume that N0 , . . . , Nk−1 have been defined, and de-

scribe how to choose Nk . Let
∗
 
δN (t) = sup Sn (t) − SN (t) .
k−1
Nk−1 <n≤N

∗ −1
By our assumption, the continuous functions (δN (t) + 1) decrease to 0
for each t ∈ Xk , and Dini’s Theorem for the compact set Xk implies the
∗ −1
existence of N > Nk−1 such that (δN (t) + 1) < (k+1)−1 for each t ∈ Xk .
Just take Nk = N to get (4.1). Now, set

Ok = {ω ∈ Ω; εn (ω) = 1 if Nk−1 < n ≤ Nk } ,

and [
Ωk = Ol .
l≥k

Ok is an open set, and Ωk is a dense open set. In fact, if α ∈ Ω and if N

is a positive integer, take l ≥ k such that Nl−1 ≥ N , and define ω ∈ Ω by
εn (ω) = εn (α) if n ≤ N , εn (ω) = 1 if n > N , then ω ∈ Ωk , which shows
that α ∈ Ωk . The set \
A= Ωk = lim Ok
k≥1

is therefore quasi-sure and for ω ∈ A, and t ∈ R, one has

n
 X   
sup

εn (ω)fn (t) =

sup Sn (t) − SN (t) > k
 k−1
Nk−1 <n≤Nk Nk−1 <n≤Nk
Nk−1 +1

for arbitrarily large values of k such that t ∈ Xk , which proves the un-
X∞
bounded divergence of εn (ω)fn (t).
1

b) We will now construct the sequence (Nk ) so that

  ϕ(t)
∀t ∈ Xk , sup Sn (t) − SN
k−1
(t) ≥ . (4.2)
Nk−1 <n≤Nk 2
574 F. Bayart, S. V. Konyagin and H. Queffélec

Take N0 = 0 and having chosen N0 , . . . , Nk−1 , construct Nk by a variant

of Dini’s Theorem. For each t, there exists N (t) > Nk−1 such that

∗ δ(t) ϕ(t)
δN (t) (t) > > ,
2 2
and by continuity

∗ 0 ϕ(t0 )
δN (t) (t ) > for t0 ∈ U (t), an open neighbourhood of t.
2
Take a finite cover U (t1 ), . . . , U (tr ) of Xk and define Nk = maxj≤r N (tj )
to get (4.2); the rest of the proof is unchanged.
P Conclusion b) no longer
holds under the sole assumption that fn (t) diverges everywhere. See the
recent work of Keleti and Matrai ([KM]) for a counterexample.

If we specialize Theorem 4.1 to Taylor and Dirichlet series and to X = R,

we get the following.
∞
X
Theorem 4.2. Let (an )n≥0 be a sequence of complex numbers with |an |2 =
0
∞.
a) If (|an |) is non-increasing, then for quasi-all choices of signs ω, the series
X∞
εn (ω)an eint is unboundedly divergent everywhere.
0

b) If (n1/2 |an |)n≥1 is non-increasing, then for quasi-all choices of signs ω, the
X∞
series εn (ω)an n−1/2+it is unboundedly divergent everywhere.
1

Proof. Let (fn (t)) be either an eint or an n−1/2 nit . In both cases, it fol-
lows respectively from [DE] or [KQ] P that there exists a choice of real signs
θn = ±1 such that the series θn fn (t) is unboundedly divergent every-
where.PNow, Theorem 4.1 (applied to θn fn ) shows that, for quasi-all ω, the
series εn (ω)θn fn (t) is unboundedly Pdivergent everywhere (in t); and we
have the same conclusion for the series εn (ω)fn (t), since the “translation”
(εn (ω)) 7→ (εn (ω)θn ) is a homeomorphism of Ω onto itself.
Theorem 4.2 will be applied to give examples of Taylor series with patho-
logical (although generic!) properties, which would probably be very difficult
to obtain explicitly.
Convergence and Divergence of Taylor and Dirichlet Series 575

P∞
Theorem 4.3. There exists a Taylor series 0 bn z n = f (z), with radius of
convergence 1, having the following three properties:

a) The series converges at z = 1 and diverges unboundedly for |z| = 1 and

z 6= 1.

b) f is unbounded on any circle γρ (θ) = 1 − ρ + ρe2iπθ , 0 <

Pρ∞< 1, 0 ≤ θ ≤ 1,
internally tangent to D at 1. In particular, the series 0 bn z n converges
uniformly on no circle γρ .

c) The unit circle |z| = 1 is a natural boundary for f .

Proof. Let (ρj )j≥1 be a sequence of values of ρ which decreases to zero,

and let Z 1
In,j = |γρj (θ)|n |1 − γρj (θ)|dθ.
0

We claim that
∞
X
In,j = ∞ for every j. (4.3)
n=0

In fact, setting e(θ) = e2iπθ , one has

∞ 1 1
|1 − γρj (θ)| |1 − γρj (θ)|
X Z Z
In,j = dθ ≥ 2
dθ
n=0 0 1 − |γρj (θ)| 0 1 − |γρj (θ)|
1 Z 1/2
ρj |1 − e(θ)| |θ|
Z
≥ dθ ≥ δj dθ = ∞
0 2ρj (1 − ρj )(1 − cos 2πθ) 0 θ2

(δj depending only on ρj ).

Lemma 4.4. There exists a sequence (wn )n≥0 , non-increasing and tending to
zero, and such that
∞
X ∞
X
wn2 = ∞ and wn In,k = ∞ for k ≥ 1. (4.4)
n=0 n=0

Proof of Lemma. In view of (4.3), we can find a sequence 0 = n1 < n2 < . . .

of integers such that
X
In,k ≥ j for k = 1, . . . , j. (4.5)
nj ≤n<nj+1
576 F. Bayart, S. V. Konyagin and H. Queffélec

Define (wn ) by wn = j −1/2 for nj ≤ n < nj+1 . We have

∞ ∞ ∞
X X nj+1 − nj X 1
wn2 = ≥ = ∞.
0 1
j 1
j

And, for fixed k, as soon as j ≥ k, (4.5) implies that

X X
wn In,k ≥ j −1/2 and In,k ≥ j 1/2 ,
nj ≤n<nj+1 nj ≤n<nj+1
P∞
so that n=0 wn In,k = ∞.
Now, for ω ∈ Ω, we define
∞
X ∞
X
n
fω (z) = (1 − z) εn (ω)wn z = bn z n . (4.6)
0 0

We claim that, quasi-surely, fω has the three properties a), b), c) of Theorem
4.3. First of all (omitting ω), one has

b0 = ε0 w0 , bn = εn wn − εn−1 wn−1 for n ≥ 1,

so that b0 + · · · + bn = εn wn → 0, and that the series converges at z = 1.

X∞
Second, since wn2 = ∞ and since (wn ) is non-increasing, Theorem 4.1
0
implies the following.
∞
X
Quasi-surely, bn z n diverges unboundedly for |z| = 1 and z 6= 1. (4.7)
0

In fact, we have
N
X N
X −1
n
bn z =(1 − z) εn wn z n + εN wN z N
0 0
N
X −1
=(1 − z) εn wn z n + o(1),
0
P∞
and 0 εn wn z n is quasi-surely unboundedly divergent. We also have
∞
X
Quasi-surely, the unit circle is a natural boundary for bn z n . (4.8)
0
Convergence and Divergence of Taylor and Dirichlet Series 577

P∞ n
In fact, the result is well-known ([Ka]) for 0 εn (ω)wn z , which has the
radius of convergence 1, and the multiplication by (1 − z) does not affect the
result. Let us now fix j ≥ 1. We have

Quasi-surely, fω is unbounded on γρj . (4.9)

In fact, if it were not the case, the topological zero-one law (see for example
[Le]) would imply the existence of a constant C such that
∞
X
|1 − z| wn |z|n ≤ C for z ∈ γρj ;
0

i.e.,
∞
X
|1 − γρj (θ)| wn |γρj (θ)|n ≤ C for 0 ≤ θ ≤ 1.
0

Integrating with respect to θ and permuting, we would get

∞
X
wn In,j ≤ C,
0

contradicting Lemma 4.4. By the stability of quasi-sure properties, (4.7), (4.8)

and (4.9) hold simultaneously for ω ∈ A, where A is quasi-sure. But then,
properties a), b), c) of Theorem 4.3 hold for fω , ω ∈ A. Indeed, fix 0 < ρ < 1
and ω ∈ A. If fω were bounded, say by M , on the circle γρ , by the maximum
modulus principle, it would be bounded by M on γρj for any j such that
ρj < ρ, since γρj is interior to γρ ; and this would contradict (4.9). This ends
the proof of Theorem 4.3.

Comment. Among other things, Theorem 4.3 shows the optimality of Abel’s
non-tangential Theorem (see also part 5). There are explicit examples showing
this optimality, but one needs some luck to find them. For example, let a > 0
and let
∞  
X
n −az
bn z = exp .
0
1−z
One then has
∞  
X −az
(b0 + · · · + bn )z n = (1 − z)−1 exp ,
0
1−z
578 F. Bayart, S. V. Konyagin and H. Queffélec

P∞
and the right hand-side is the generating series 0 Ln (a)z n of the Laguerre
polynomials at a ([L]), so that b0 +· · ·+bn = Ln (a). But it follows for example
from the integral representation ([L])

ea ∞ n √
Z
Ln (a) = t J0 (2 at)e−t dt
n! 0

(where J0 Pis the Bessel functionP of order zero) that Ln (a) → 0 as n → ∞.

∞ ∞
Therefore, 0 bn converges and 0 bn = 0. Now, the fractional linear trans-
az
formation z 7→ = w takes the circle γρ , tangent to D at 1, to the vertical
1−z
a(1 − 2ρ)
line <(w) = . Therefore
2ρ
∞ 
X  a(1−2ρ)
bn z  = e−<w = e− 2ρ
n


 
0
P∞
for z ∈ γρ andP∞z 6= n1, while 0 bn = 0, which prevents the uniform con-
vergence of 0 bn z on γρ . On the other hand, it is easy to see that the
previous series converges at z if |z| = 1 and z 6= 1 (it is the Fourier series of
a C∞  function
 in the neighborhood of z on the circle), and moreover the sum
−az
exp 1−z of the series can be analytically continued at all the points of the
circle different from one; so we recover only part of the properties in Theorem
4.3.

5 Concluding Remarks and Questions

Let D be the open unit disk, and let E ⊂ D be such that 1 ∈ E and E ⊂ D∪{1}.
We will say that E is tangent to D at 1 if we have

|1 − z|
lim sup = ∞. (5.1)
z∈E,z→1 1 − |z|
P
The classical non-tangential Theorem of Abel [Din] states that, if an con-
verges and E ⊂ D is not tangent to D at 1 (we assume E ⊂ D ∪ {1}), then
an z n is uniformly convergent for z ∈ E. In the special case when E is a
P
circle γρ , internally tangent to D at 1, we used
P inn Theorem 4.3 the test-space
of choices of signs to get a Taylor series an z , convergent at 1, and not
uniformly convergent on γρ . In the general case, we will use the “larger” test-
space c0 , and the Banach-Steinhaus Theorem (which is a version of Baire’s
Theorem and of quasi-sure properties), to revisit classical results of Hardy
Convergence and Divergence of Taylor and Dirichlet Series 579

and Littlewood, showing at the same time that Abel’s Theorem is definitely
optimal.
Theorem 5.1. Let PE ⊂ D be any set which is tangent to D at 1. Then, there
∞
is a Taylor series 0 bn z n = f (z) such that:
1. The series has radius of convergence 1, and converges at z = 1.
2. f is not bounded on E. In particular, the series does not converge uni-
formly on E.
Proof. Let c0 be the Banach space of sequences a = (an )n≥0 tending to
zero atPinfinity, with its natural norm kak = supn |an |. We will search our
∞
series 0 bn z n in the form
∞
X ∞
X ∞
X
n n
bn z = (1 − z) an z = a0 + (an − an−1 )z n ,
0 0 1

with a = (an ) ∈ c0 . Let Lz (z ∈ E) be the linear form on c0 defined by

P∞
Lz (a) = (1 − z) 0 an z n . Since the dual space of c0 is `1 , we have
∞
X |1 − z|
kLz k = |1 − z| |z|n = . (5.2)
0
1 − |z|

It follows from (5.1) and (5.2) that supz∈E kLz k = ∞. The Banach-Steinhaus
Theorem now implies the existence of a ∈ c0 such that supz∈E |Lz (a)| = ∞.
If b0 = a0 and bn = an − an−1 for n ≥ 1, we have b0 + · · · + bn = an → 0, and
the sequence (bn ) fulfills the requirements of Theorem 5.1.
Remarks.
1. Theorem 5.1 can be applied to produce Dirichlet series with similar
properties. Indeed, if F ⊂ C0 , we will say that F is tangent to C0
s
at 0 if F ⊂ C0 ∪ {1} and lim sups→0,s∈F <(s) = ∞. Set φ(s) = 2−s , and
E = φ(F ). E is tangent to D at 1. Let f (z) be the Taylor series given by
Theorem 5.1, and set g(s) = f (2−s ). Then g has abscissa of convergence
0, converges at s=0, and is not bounded on F .
2. Theorem 5.1 can be compared with the following Theorem of Littlewood
([Lit]). For any curve E ⊂ D, tangent to D at 1, there exists a bounded,
holomorphic function f on D, such that, for almost all θ, the limit of
f (zeiθ ) as z → 1, z ∈ E, does not exist (while the non-tangential limit
of f (w) as w → eiθ exists for almost all θ).
580 F. Bayart, S. V. Konyagin and H. Queffélec

P∞ P∞
3. Of course, if 0 |bn | < ∞, the series 0 bn z n has a better behavior; it
converges
P∞ uniformly on D. Hardy and LittlewoodP[HL] proved that, even
∞
if 0 bn converges, no weaker condition than 0 |bn | < ∞ can force
the tangential uniform convergence at 1. More precisely, they proved
the following.
Theorem 5.2. Let (wn )n≥0 be aPnon-increasing sequence of positive
∞
numbers such Pthat wn → 0 and 0 wn = ∞. Then, there exists a
∞ n
Taylor series 0 b n z = f (z), with radius of convergence 1, and a
curve E ⊂ D, tangent to D at 1, such that:
P∞
a) 0 bn converges,
b) bn = O(wn ),
P∞
c) f is not bounded on E and in particular 0 bn z n does not converge
uniformly on E.

Preliminary Comment: The initial proof of Hardy and Littlewood

1
was complicated, and explicitly written only for wn = .
n log n

Proof of Theorem 5.2. We once more appeal to the Banach-Steinhaus

Theorem. Let X be the space of sequences (an )n≥0 such that an =
O(wn ), equipped with the norm
|an |
kak = sup . (5.3)
n wn
l∞ . Let Lz (z ∈ D) be the linear
(X, k.k) is a Banach space, isometric toP
∞
form on X defined by Lz (a) = (1 − z) 0 an+1 z n . It is plain that
∞
X
kLz k = |1 − z| wn+1 |z|n . (5.4)
0
P∞
Let now 0 < εj < 1, εj → 0. Since 0 wn = ∞, we can choose zj on
P∞
the circular arc |z − 1| = εj , |z| < 1, such that 0 wn+1 |zj |n ≥ εjj , and
j
(5.4) shows that kLzj k ≥ εj = j. By the Banach-Steinhaus Theorem,
εj
one can find a ∈ X such that supj |Lzj (a)| = ∞. Join the points zj by
a curve E; this curve will be tangent to D at 1. In fact, we have
∞
X |1 − zj |
j ≤ kLzj k ≤ |1 − zj | w1 |zj |n = w1 .
0
1 − |zj |
Convergence and Divergence of Taylor and Dirichlet Series 581

And set b0 = a1 , bn = an+1 − an if n ≥ 1, so that

|bn | ≤ |an+1 | + |an | ≤ kak(wn+1 + wn ) ≤ 2kakwn .

P∞ n
Since we have 0 bn z = Lz (a), and since b0 + · · · + bn = an+1 =
O(wn+1 ), the sequence (bn ) fulfills all the requirements of Theorem 5.2.

We will end up with the following observation. Theorems 2.1 and 4.2 have
shown that a change of sign can worsen the behavior of the series, which may
become everywhere divergent, but a change of sign may sometimes improve
the behavior, as indicated by the following.

Theorem 5.3. Let (an )n≥1 be a sequence of complex numbers such that |an |
decreases to zero. Then:

P∞
1. There exists a sequence (θn )n≥1 of complex signs such that 1 θn an nit
converges for each t ∈ R.

∞
X |an |
2. If moreover < ∞, one can choose the signs θn real (θn = ±1)
1
n
in 1.

Proof.

1. We will use the following.

Lemma 5.4. For each t ∈ R, there exists a constant Ct > 0 such that
 
XN 
n it 
(−1) n  ≤ Ct , for N = 1, 2, . . . (5.5)


 
n=1

Proof of the Lemma. The result is clear for t = 0; for t 6= 0, it is

582 F. Bayart, S. V. Konyagin and H. Queffélec

∞
X
well-known ([Ha]) that nit−1 is boundedly divergent; therefore,
1

2N
X N
X
(−1)n nit = (2n)it − (2n − 1)it
 
1 1
N  it !
X
it 1
= (2n) 1− 1−
1
2n
N  
X it
= (2n)it + O(n−2 )
1
2n
N
X
=it2it−1 nit−1 + O(1) = O(1),
1

which proves (5.5). Now, if we adjust θn , |θn | = 1, so as to have

P θn an =
(−1)n |an |, (5.5) and an Abel’s summation by parts show that θn an nit
converges everywhere.

2. We will use the following Lemma ([DC]).

Lemma 5.5. There exists a numerical constant C > 0 with the following
property. If z1 , . . . , zN are complex numbers of modulus less than one,
there exist real signs ε1 , . . . , εN such that

|ε1 z1 + · · · + εn zn | ≤ C for 1 ≤ n ≤ N. (5.6)

(The result was generalized in [BG]). We can now end the proof of 2.
as follows. Let bk = |a2k |, k = 0, 1, . . . Since |an | decreases, and since
P∞ |an | P∞
1 n < ∞, we clearly have 0 bk < ∞. Now, Lemma 5.5 allows us
to build a sequence (θn ) of real signs by blocks such that
 X 
θn an  ≤ Cbk , for k ≥ 0 and 2k ≤ N < 2k+1 .
 

 (5.7)
2k ≤n≤N

It follows that
 X 
n θn an  ≤ Ct0 bk , for k ≥ 0 and 2k ≤ N < 2k+1 .
it
 

 (5.8)
2k ≤n≤N
Convergence and Divergence of Taylor and Dirichlet Series 583

X
In fact, fix k and set Sn = θj aj if n ≥ 2k , S2k −1 = 0. We have,
2k ≤j≤n
for 2k ≤ N < 2k+1
X X
nit θn an = (Sn − Sn−1 )nit
2k ≤n≤N 2k ≤n≤N
X
Sn nit − (n + 1)it + SN N it ,


=
2k ≤n≤N −1

so that
 X  X Cbk |t|
nit θn an  ≤ + Cbk ≤ (C|t| + C)bk := Ct0 bk .
 

 n
2k ≤n≤N 2k ≤n≤N −1

θn an nit clearly follows from

P P
Now, since bk < ∞, the convergence of
(5.8).
The inequality (5.5) is in marked contrast with the case of power series.
According to a result of Lesigne and Petersen ([LP]), if a (necessarily bounded)
sequence (an )n≥0 of complex numbers satisfies
N 
X 
an eint  ≤ Ct ∀t ∈ R, ∀N ∈ N, (5.9)
 

 
0

then one must have lim|an | = 0 and a little bit more, in terms of ergodic
theory (see (5.14) below). We will reformulate their result, and give a slightly
shorter proof. Recall that a subset E of the natural numbers N is said to be
of uniform density zero if we have
 
|E ∩ [t, t + N ]|
lim sup = 0.
N →∞ t≥0 N +1

Theorem 5.6. Assume that a complex sequence (an )n≥0 satisfies (5.9). Then
there exists a subset E of N, with uniform density zero, such that

lim an = 0.
n→∞
(5.10)
n∈E
/

Equivalently,
 
|aj | + · · · + |aj+N |
lim sup = 0. (5.11)
N →∞ j∈N N +1
584 F. Bayart, S. V. Konyagin and H. Queffélec

Proof. At the beginning, we follow [LP]. Let K be a compact set of

the plane where (an ) takes its values, τ be the unilateral shift (bn ) 7→ (bn+1 )
on the cartesian product K N , X be the closed orbit of a = (an ) under τ ,
and µ be a τ −invariant measure on X, so that the composition operator
T : L2 (µ) → L2 (µ) given by T (g) = g ◦ τ is an isometry ([Wa]). One easily
deduces from (5.9) that, for any x = (xn ) ∈ X
 
Xn 
−ikθ 
xk e  ≤ 2C−θ = Dθ , ∀θ, ∀n. (5.12)


 
0

Now, denote by f ∈ L2 (µ) the function defined by f (x) = x0 , and by σ the

spectral measure of f , which is the positive measure on the circle T defined
by (via Bochner’s Theorem, see [Kat]):
σ̂(k) =< T k f, f > if k ≥ 0, σ̂(k) =< f, T −k f > if k < 0,
 
Z X n 
ikt k −ikθ 
where σ̂(k) = e dσ(t). (5.12) may be rewritten as  T f (x)e  ≤ Dθ ,

T 
k=0

so that if we square and integrate with respect to dµ(x), we get
Z Xn
2

ikt −ikθ 
e e  dσ(t) ≤ Dθ2 . (5.13)


T k=0


In fact, we have
Z X n
2 Z
 X
k −ikθ 
T f (x)e  dµ(x) = ei(l−k)θ T k f (x)T l f (x) dµ(x)


X k=0 
0≤k,l≤n X
X Z
= ei(l−k)θ eikt e−ilt dσ(t)
0≤k,l≤n T

Z Xn
2

ikt −ikθ 
=  e e  dσ(t).

T k=0


We now derive from (5.13) the following property (which implies that 0 ∈ K).
µ = δω , where ω = (0, 0, . . . , 0, . . . ). (5.14)
Indeed, suppose that this is not the case, and pick a point u 6= ω in the closed
support of µ. We can find an index k, a neighborhood V of u, and a positive
number ε such that, if x = (xn ) ∈ V , we have |xk | ≥ ε. This implies that
Z
|xk |2 dµ(x) ≥ ε2 µ(V ) > 0.
Convergence and Divergence of Taylor and Dirichlet Series 585

|x0 |2 dµ(x) > 0. Therefore, σ is a

R
Since µ is τ −invariant, we have as well
non zero-measure, since
Z
kσk = σ̂(0) =< f, f >= |x0 |2 dµ(x) > 0.

Let θ0 be a density point of σ. Testing (5.13) at θ0 and using the properties

of the Dirichlet kernel, we get
Z Xn 2
Dθ20 ≥ eik(t−θ0 )  dσ(t)
 

π
|t−θ0 |≤ n k=0

2
 π π 1
≥αn σ θ0 − , θ0 + ≥ αβn2 × = αβn,
n n n
where α, β are positive constants; but this is clearly impossible for big n,
proving (5.14) by contradiction. Here, we slightly diverge from [LP] as follows.
Since δω is the unique invariant probability measure for τ , the well-known
Theorem of Oxtoby ([Wa]) implies that, for any continuous function g : X →
C, we have

g(x) + g(τ x) + · · · + g(τ n x)

Z
→ g(ω) = g dµ uniformly on X. (5.15)
n+1

Testing (5.15) on the function g(x) = |x0 | and on the point x = τ j a, we

get (5.11). And, since (an ) is bounded, it is well-known ([Wa]) that (5.11) is
equivalent to (5.10).
Remark. Conclusion (5.10) is not a weakening of the information (5.14),
since a routine approximation argument shows that, conversely, (5.10) implies
(5.15) and therefore (5.14). We find this formulation (5.10) more suggestive
(see Question 2. below).
QuestionP1. Is the conclusion 2. in Theorem 5.3 still valid without the as-
∞
sumption 1 n−1 |an | < ∞?

Question 2. In Theorem 5.6, can one have the stronger conclusion an → 0?

If this were not the case, we would have an example of a trigonometric series
which is boundedly divergent everywhere, making a nice complement to the
Theorem of Marcinkiewicz quoted in section 3.

Acknowledgement. The second author was supported by RFFI Grant N.

02-01-00248. We thank L. Rodriguez-Piazza for stimulating conversations.
586 F. Bayart, S. V. Konyagin and H. Queffélec

References
[B] N. Bary, A treatise on trigonometric series, Pergamon Press, (1964).
[BG] I. Barany - V. S. Grinberg, On some combinatorial questions in finite-
dimensional spaces, Linear Algebra Appl., 41 (1981), 1–9.
[Ca] L. Carleson, On the convergence and growth of partial sums of Fourier
series, Acta Math., 116 (1966), 135–167.
[DE] A. Dvoretzky - P. Erdös, On power series which diverge everywhere
on the unit circle of convergence, Michigan Math J., 3 (1955-1956),
31–35.
[Din] P. Dienes, The Taylor series, Oxford, (1931).
[Dis] J. Diestel, Sequences and series in Banach spaces, Springer-Verlag,
1984.
[DC] A. Dvoretzky - H. Chojnacki, Sur les changements de signe d’une série
à termes complexes, C.R.A.S., 222 (1947) 516–518.
[Ha] G. H. Hardy, Divergent series, Oxford University Press, 2nd edition,
(1956).
[HL] G. H. Hardy - J. E. Littlewood, Abel’s Theorem and its converse, Proc.
Lond. Math. Soc., 2 (1920), 205–235.
[He] H. Hedenmalm, Topics in the theory of Dirichlet series, Visn. Khark.
Univ., Ser. Mat. Prykl. Mat. Mekh., 475 (2000), 195–203.
[HLS] H. Hedenmalm, P. Lindqvist, K. Seip, A Hilbert Space of Dirichlet se-
ries and a system of dilated functions in L2 (0, 1), Duke Math. Journal,
86 (1997), 1–36.
[HS] H. Hedenmalm - E. Saksman, Carleson’s convergence Theorem for
Dirichlet series, Pacific J. of Math., 208 (2003), 85–109.
[Ka] J. P. Kahane, Some random series of functions, 2nd edition, Cam-
bridge University Press, 1985.
[KK] J. P. Kahane - Y. Katznelson, Sur les ensembles de divergence des
séries trigonométriques, Studia Math., 26 (1966), 305–306.
[Kat] Y. Katznelson, Introduction to Harmonic Analysis, John Wiley and
Sons, 1968.
Convergence and Divergence of Taylor and Dirichlet Series 587

[KM] T. Keleti, T. Matrai, A nowhere convergent series of functions which

is somewhere convergent after a typical change of signs, Real Analysis
Exchange, 29 (2003/4), this issue.
[KQ] S. V. Konyagin - H. Queffélec, The translation 1/2 in the theory of
Dirichlet Series, Real Analysis Exchange, 27 (2001/2002), 155–175.

[Kö] T. W. Körner, The behavior of power series on their circle of conver-

gence, Springer Lecture Notes, 995, 56–94.
[L] N. N. Lebedev, Special functions and their applications, Dover, (1972).
[Le] P. Lefevre, Topological dichotomy and unconditional convergence,
Serdica Math. J., 25 (1999), 297–310.
[LP] E. Lesigne - K. Petersen, Bounded exponential sums, Osaka J. Math.,
27 (1990), 277–280.
[Li] D. Lick, A note on sets of convergence of Dirichlet series, Duke Math.
J., 31 (1964), 179–181.
[Lit] J. E. Littlewood, On a Theorem of Fatou, Journal London Math. Soc.,
2 (1927), 172–176.
[SaZ] R. Salem - A. Zygmund, Some properties of trigonometric series whose
terms have random signs, Acta Math., 91 (1954), 245–301.
[Wa] P. Walters, An introduction to ergodic theory, Springer-Verlag, (1982).
[W] M. Weiss, Concerning a Theorem of Paley on lacunary power series,
Acta Math., 102 (1959), 225–238.