Torino
Vol. 1, 60 (2002)
P. Codeca  N. Taddia
L
2
EXPANSIONS IN SERIES OF FRACTIONAL PARTS
Abstract. In this paper we consider the problemof L
2
expansions in series
of fractional parts for functions f ∈ L
2
([0, 1], R) which are odd with
respect to the point
1
2
. Sufﬁcient conditions are given, but we also prove
by the BanachSteinhaus theorem that this is not always possible.
1. Introduction
Let :
(1) P(x) = −
1
π
¸
m≥1
sin(2πmx)
m
=
¸
x −[x] −
1
2
if x / ∈ Z
0 if x ∈ Z
In a paper of 1936 [5] H. Davenport considered the formal identity
(2)
¸
n≥1
α(n)P(nx) = −
1
π
¸
n≥1
¸
d/n
dα(d)
n
sin(2πnx)
which is obtained by substituting into (2) the expansion of P given in (1) and collecting
together the terms for which the product mn has the same value. Davenport considered
the cases in which α(n) has the values
µ(n)
n
,
λ(n)
n
and
(n)
n
, where, as usual, µ,λ and
denote, respectively, the M¨ obius, Liouville and Von Mangoldt funtions.
In [5] Davenport proved the following results:
i) The functions
¸
N
n=1
µ(n)
n
P(nx) and
¸
N
n=1
λ(n)
n
P(nx) are uniformly bounded in
N and x.
ii) The identities
¸
n≥1
µ(n)
n
P(nx) = −
1
π
sin(2πx)
¸
n≥1
λ(n)
n
P(nx) = −
1
π
¸
n≥1
sin(2πn
2
x)
n
2
hold almost everywhere.
17
18 P. Codeca  N. Taddia
In a subsequent paper [6], Davenport, using Vinogradov’s method and the Siegel
Walﬁsz theorem, showed that the identities ii) hold everywhere and that the two series
converge uniformly in R. The proof of the weaker properties ii) (the almost everywhere
convergence) depends on L
2
convergence, namely on the estimate [5]:
R
N
2
=
1
0

N
¸
n=1
µ(n)
n
P(nx) +
1
π
sin(2πx)
2
dx = O(
log(N)
N
)
The aim of the present paper is just to consider the formal identity (2) for general α(n)
and from the point of view of L
2
convergence. In order to state in a more precise way
the problems studied and the results obtained we give the following deﬁnitions.
Le us deﬁne:
L
2
0
= { f ∈ L
2
([0, 1], R), f (x) = −f (1 − x) a.e. on [0, 1]}
that is L
2
0
is the subspace of the functions f ∈ L
2
([0, 1], R) which are odd with respect
to the point
1
2
. It is easy to see that L
2
0
is a closed subspace of L
2
([0, 1], R) and so it is
a Hilbert space.
Moreover, if n is a nonnegative integer, we denote with M
+
(n) the greatest prime
which divides n and, as usual, we set (x, y) =
¸
{n≤x,M
+
(n)≤y}
1.
We shall consider the following problems:
α) Is it possible to apply the GramSchimdt procedure to the functions P(nx) and,
if (ϕ
n
(x))
n
is the resulting orthonormal system, is it true that
f (x) =
¸
n≥1
b(n)ϕ
n
(x)
in the L
2
sense for every f ∈ L
2
0
?
β) Is it possible to give sufﬁcient conditions for expanding (in the L
2
sense) in series
of fractional parts the functions belonging to certain subsets of L
2
0
?
γ ) Is it possible to expand (in the L
2
sense) every f ∈ L
2
0
in series of fractional
parts ?
Theorem 1 below answers positively to question α), while in Theorem 2 we show that
it is possible to give a positive answer to question β).
On the contrary, the answer to question γ ) is negative. This is proved in Theorem 3,
whose proof depends on the BanachSteinhaus theorem and on an asymptotic formula
with an estimate of the error term (see Lemma 1) for the difference (x, y) −(
x
2
, y)
with y ﬁxed.
L
2
expansions in series 19
REMARK 1. The identity
(3)
¸
n≥1
µ(n)
n
P(nx) = −
1
π
sin(2πx) x ∈ R
has interesting applications. For example, it was used in [4] to prove a result of linear
independence for the fractional parts, namely the fact that the determinant:
(4) det(P(
md
q
)) = 0 0 ≤ d ≤ m ≤
q −1
2
Actually, for the proof of (4), it is enough to know that (3) holds for rational x and this,
in turn, is equivalent to the result L(1, χ) = 0 (here χ is a caracter mod q and L(s, χ)
is the corresponding L function). Moreover it can be shown that also the weaker result
that (3) holds in the L
2
sense can be used to prove results like the following one.
Let h : [0, 1] → R be continuous and of bounded variation. Let us also suppose that
h(x) = h(1−x), ∀x ∈ [0, 1] and that
1
0
f (x)dx = 0. Then the following implication
holds:
d
¸
a=1
h(
a
d
) = 0 ∀ d ≥ 1 ⇒h(x) = 0, ∀ x ∈ [0, 1]
It is well known (see [3] and [2]) that this property is false if we drop the bounded
variation condition.
2. Notation and results
If a and b are positive integers (a, b) will denote the greatest common divisor of a and
b, and dn means that d divides n. With τ(n) we will indicate the number of divisors
of n, as usual.
We have obtained the following results.
THEOREM 1. The following properties hold:
i) The functions f
n
(x) = P(nx), n = 1, 2, ... ,are linearly independent.
ii) The set of functions f
n
(x) = P(nx), n = 1, 2, ... is complete in L
2
0
.
iii) Let (ϕ
n
(x)) be the orthonormal system obtained from f
n
(x) = P(nx) by the
GramSchmidt procedure. Then we have
(5)
N
¸
n=1
b(n)ϕ
n
(x) − f (x)
L
2 →0 when N → +∞
for every f ∈ L
2
0
, where the (b(n)) are the Fourier coefﬁcients of f (x) with
respect to the system (ϕ
n
(x)).
20 P. Codeca  N. Taddia
Theorem 1 solves the problem of expanding every function f ∈ L
2
0
in series of
linear combination of fractional parts, since we have
ϕ
n
(x) =
n
¸
h=1
γ
h,n
P(hx)
where the γ
h,n
are suitable coefﬁcients ([8], p.305).
As for the problem of the expansion of functions f ∈ L
2
0
in series of fractional parts
we have the following theorems.
THEOREM 2. Let f (x) ∼
¸
k≥1
a(k) sin(2πkx) be the Fourier series of f ∈ L
2
0
.
Set
R
N
(x) =
N
¸
n=1
α(n)P(nx) − f (x)
with α(n) ∈ R. Then we have
(6) R
N
2
=
1
2
¸
m≥1

1
π
¸
nm
n≤N
α(n)
n
m
+a(m)
2
If R
N
→0 when N →∞we must have
(7) α(n) = −π
¸
dn
µ(d)
d
a(
n
d
).
Finally, let
a
∗
(n) =
¸
h≥1
h≡0(n)
a(h)
2
.
If the condition
(8)
¸
n≥1
τ(n)
a
∗
(n)
1
n
< +∞
holds then
(9) R
N
= f −
N
¸
n=1
α(n)P(nx) →0
when N →+∞, with α(n) as in (7).
THEOREM 3. There exist functions f ∈ L
2
0
such that
(10) sup
N
R
N
= +∞
where R
N
(x) =
¸
N
n=1
α(n)P(nx) − f (x) and the α(n) are given by (7).
L
2
expansions in series 21
REMARK 2. Condition (8) of Theorem2 is not a very restrictive one. It is certainly
satisﬁed, for instance, if there exists (b(n)) ⊂ ℓ
2
such that b(n +1) ≤ b(n), a(n) ≤
b(n), ∀n ≥ 1, since in this case we have
a
∗
(n) =
¸
h≥1
a(hn)
2
≤
¸
h≥1
1
n
n
¸
r=1
b
2
((h −1)n +r)
=
b
2
n
This means that (8) holds in particular for all bounded variation functions, since in this
case we have a(n) = O(
1
n
) ([1], vol.I, p. 72).
Another example is given by the class
β
of functions which satisfy a Lipschitz con
dition of order β > 0: in this case we have ([1], vol.I, p.215)
(
¸
k≥n
a
k

2
)
1
2
= O(
1
n
β
)
which gives a
∗
(n) = O(
1
n
2β
) and (8) holds again.
It should be noted that under condition (8) we can assert that the formal identity
(2) holds almost everywhere for a subsequence. This is an obvious corollary of (9), (7)
and M¨ obius inversion formula.
REMARK 3. The proof of Theorem 3 is based on BanachSteinhaus theorem and
so we do not give an explicit example.
The contrast between (5) of Theorem 1 and (10) of Theorem 3 can be explained as
follows.
If f ∈ L
2
0
, the minimum value of the difference
f −
N
¸
n=1
c
n
P(nx)
can be obtained directly by solving the linear system ([10], p. 83)
(11)
N
¸
n=1
a
m,n
c
n
= b
m
, m = 1, ..., N
where
a
m,n
=
1
0
P(nx)P(mx)dx =
1
12
(n, m)
2
nm
and
b
m
=
1
0
P(mx) f (x)dx
or equivalently by considering the sum
σ
N
(x) =
N
¸
n=1
b(n)ϕ
n
(x) =
N
¸
n=1
b(n)(
n
¸
h=1
γ
h,n
P(hx)) =
N
¸
n=1
c
n
P(nx)
22 P. Codeca  N. Taddia
which appears in (5) of Theorem 1. Generally the α(n) given by (7) of Theorem 2
are not the solution of system (11): for example it is easy to check directly that when
f (x) = −
1
π
sin(2πx) we have
α(n) =
µ(n)
n
, b
m
=
¸
1
2π
2
if m = 1
0 otherwise
and the numbers µ(n)/n do not satisfy condition (11), and consequently the difference
f −
¸
N
n=1
α(n)P(nx) is not the minimal one.
Usually, if the function is not a very irregular one (see condition (8) of Theorem 2)
the sum S
N
(x) =
¸
N
n=1
α(n)P(nx) will be a good approximation to σ
N
(x) and it will
happen that S
N
−σ
N
→0 when N →+∞. But there exist very irregular functions
f ∈ L
2
0
for which we have
sup
N
R
N
= sup
N
S
N
−σ
N
= +∞
although we think that is not easy to give an explicit example.
Let us now prove the results stated.
Proof of Theorem 1. Proof of property i).
We want to prove that the relation
(12)
n
¸
k=1
c
k
P(kx) = 0 a.e. on [0, 1] c
k
∈ R
implies c
k
= 0 ∀k = 1, ..., n. If n = 1 the implication is obviously true. Suppose now
that the equality (12) holds with n ≥ 2.
If 0 ≤ x <
1
n
we have [kx] = 0 for every k ≤ n and from (12) follows
n
¸
k=1
c
k
(kx −
1
2
) = 0 a.e. on [0,
1
n
)
from which we obtain
(13)
n
¸
k=1
c
k
= 0
If
1
n
≤ x <
1
n−1
we have
[kx] =
¸
1 if k = n
0 if k ≤ n −1
and so from (12) follows
n−1
¸
k=1
c
k
(kx −
1
2
) +c
n
(nx −
3
2
) = 0 a.e. on [
1
n
,
1
n −1
)
L
2
expansions in series 23
which in turn implies
(14)
n−1
¸
k=1
c
k
+3c
n
= 0
Relations (13) and (14) give c
n
= 0 and the desidered implication follows by induction.
This gives property i).
Proof of property ii).
Take f ∈ L
2
0
and set δ(n) =
1
0
f (x)P(nx)dx. Since
P(x) = −
1
π
¸
k≥1
sin(2πkx)
k
we have
(15) δ(n) = −
1
π
¸
k≥1
1
k
1
0
sin(2πknx) f (x)dx = −
1
2π
¸
k≥1
c(kn)
k
where the c(n) are the Fourier coefﬁcients of f . The relation (15) can be inverted and
we have
(16) c(n) = −2π
¸
k≥1
µ(k)
δ(nk)
k
provided the sufﬁcient condition
(17)
¸
k≥1
τ(k)c(nk)
k
< +∞ ∀n ≥ 1
is satisﬁed ([7], Theorem 270). But (17) is certainly true by Schwarz inequality, and so
(16) holds. If we suppose δ(n) = 0 ∀n ≥ 1 from (16) follows c(n) = 0 ∀n ≥ 1, but
this means that f ∈ L
2
0
has all the Fourier coefﬁcients equal to zero. The completeness
of the trigonometrical system implies f (x) = 0 a.e. and property ii) follows.
Proof of property iii).
If (ϕ
n
(x)) is the orthonormal system obtained from P(nx) by the GramSchmidt pro
cedure we have ([8], p. 305)
ϕ
n
(x) =
n
¸
h=1
γ
h,n
P(hx)
and viceversa
P(nx) =
n
¸
k=1
γ
′
h,n
ϕ
k
(x)
24 P. Codeca  N. Taddia
where the γ
h,n
and γ
′
h,n
are suitable coefﬁcients. This implies, by property ii), that
the orthonormal system (ϕ
n
(x)) is complete in L
2
0
and so Parseval identity holds. This
proves (5) and concludes the proof of Theorem 1.
Let us now prove Theorem 2.
Proof of Theorem 2. Formula (6) is simply Parseval identity for R
N
(x) =
¸
N
n=1
α(n)P(nx) −f (x) in fact we have
(18)
c
N
(m) =
√
2
1
0
R
N
(x) sin(2πmx)dx =
= −
1
√
2
1
π
(
¸
nm
n≤N
α(n)
n
m
) +a(m)
To justify (18) we note that
√
2
0
1
P(nx) sin(2πmx)dx = −
√
2
π
¸
k≥1
1
k
1
0
sin(2πknx) sin(2πmx) =
=
¸
−n
π
√
2m
if nm
0 otherwise
Since we have R
N
2
=
¸
m≥1
c
N
(m)
2
from (18) we obtain (6). Let us prove (7).
If R
N
→0 from (6) it follows immediatly that
1
π
(
¸
nm
α(n)
n
m
) = −a(m) ∀m ≥ 1
from which we obtain (7), since the Dirichlet inverse of
1
n
is
µ(n)
n
.
Let us now suppose that condition (7) holds: in this case (6) becomes
(19) R
N
2
=
1
2
∞
¸
m=N+1

1
π
¸
nm
n≤N
α(n)
n
m
+a(m)
2
when N →+∞. From (19) it follows
(20) R
N
→0 ⇐⇒
∞
¸
m=N+1

¸
nm
n≤N
α(n)
n
m

2
→0
since (a(m)) ∈ ℓ
2
. But we also have
(21) −
1
π
¸
nm
n≤N
α(n)
n
m
=
¸
dm
d≤N
a(d)
d
m
(
¸
k(m/d)
k≤(N/d)
µ(k))
L
2
expansions in series 25
From (20) and (21) follows
(22) R
N
→0 ⇐⇒
∞
¸
m=N+1
a
′
N
(m)
2
→0
when N →∞, where
a
′
N
(m) =
1
m
¸
dm
d≤N
da(d)(
¸
k(m/d)
k≤(N/d)
µ(k)).
Let us now prove that (8) implies (9). First we note that
(23) a
′
N
(n) ≤
¸
hn
a(h)τ(
n
h
)
h
n
≡ γ (n)
uniformly in N. From the deﬁnition of γ (n) in (23) follows easily
¸
n≥1
γ
2
(n) =
¸
n≥1
¸
hn
kn
a(h)a(k)τ(
n
h
)τ(
n
k
)
hk
n
2
≤
≤ (
¸
m≥1
τ
2
(m)
m
2
)
¸
¸
h,k≥1
a(h)a(k)τ(
k
(h, k)
)τ(
h
(h, k)
)
(h, k)
2
hk
= (24)
= (
¸
m≥1
τ
2
(m)
m
2
)
¸
say, if we remember that τ(nm) ≤ τ(n)τ(m). If (h, k) = δ so that h = rδ, k = sδ,
(r, s) = 1 we also have
¸
≤
¸
r,s≥1
τ(r)τ(s)
rs
¸
δ≥1
a(rδ)a(sδ) ≤
≤
¸
r,s≥1
τ(r)τ(s)
rs
(
¸
δ≥1
a(rδ)
2
)
1
2
(
¸
δ≥1
a(sδ))
1
2
= (25)
=
¸
¸
r≥1
τ(r)
√
a∗(r)
r
2
< +∞
if condition (9) holds. From (24) and (25) follows that the series
¸
n≥1
γ
2
(n) is con
vergent, but this implies that condition (22) is satisﬁed, since from (23) it follows ob
viously that
∞
¸
n=N+1
a
′
N
(n)
2
≤
∞
¸
n=N+1
γ
2
(n) →0
26 P. Codeca  N. Taddia
when N →+∞. This proves (10) and concludes the proof of Theorem 2.
Proof of Theorem 3. We need the following two lemmas.
LEMMA 1. Let p
1
< p
2
< ... < p
n
be the prime numbers up to p
n
, with p
1
=
2. If M
+
(k) denotes the greatest prime divisor of the integer k, with the convention
M
+
(1) = 1, set
(x, p
n
) =
¸
k≤x
M
+
(k)≤p
n
1
Then we have
(26) (x, p
n
) −(
x
2
, p
n
) = c(n) ln
n−1
(x) + O(ln
n−2
(x)) when x →+∞
where
c(n) =
¸
1 if n = 1
1
(n−1)!
¸
n
k=2
(ln( p
k
))
−1
if n ≥ 2
Proof. Formula (26) is obviously true if n = 1, because in this case we have
(x, p
1
) −(
x
2
, p
1
) = 1 ∀x ≥ 2
Now we will prove that if (26) holds for the integer n then it holds also for n + 1 and
the lemma will follow by induction. We have
(27) (y, p
n+1
) =
¸
p
k
1
1
... p
kn
n
p
k
n+1
n+1
≤y
1 =
[ln y/ ln p
n+1]
¸
k
n+1
=0
y
p
k
n+1
n+1
, p
n
with k
j
≥ 0 for j = 1, ..., n +1. From (27) with y = x and y =
x
2
we obtain
(x, p
n+1
) −(
x
2
, p
n+1
) =
[ln
x
2
/ ln p
n+1
]
¸
k
n+1
=0
(
x
p
k
n+1
n+1
, p
n
) −(
x
2p
k
n+1
n+1
, p
n
)
+
+
¸
[ln
x
2
/ ln p
n+1
]<k
n+1
≤[ln x/ ln p
n+1
]
(
x
p
k
n+1
n+1
, p
n
) =
¸
1
(x) +
¸
2
(x), (28)
where
¸
2
(x) is zero if the sum is empty. Consider now
¸
1
(x): by induction hypoth
esis we have
¸
1
(x) = c(n)
[ln
x
2
/ ln p
n+1
]
¸
k
n+1
=0
ln
n−1
x
p
k
n+1
n+1
+ O
ln
n−1
x
=
= c(n)
n−1
¸
h=0
(−1)
h
(
n −1
h
) ln
n−1−h
x · ln
h
p
n+1
¸
[ln
x
2
/ ln p
n+1
]
¸
k
n+1
=0
k
h
n+1
+ (29)
+ O
ln
n−1
x
L
2
expansions in series 27
where, if n = 1 the above expression must obviously be interpreted as
¸
1
(x) =
c(1)
ln p
2
ln x + O(1)
We now recall that
1
h
+2
h
+... +(m −1)
h
=
=
1
h +1
m
h+1
+(
h +1
1
)B
1
m
h
+(
h +1
2
)B
2
m
h−1
+...
(30)
where the B
j
are the Bernoulli numbers ([9], p. 65). If we use the expression (30) in
(29) we obtain
[ln(
x
2
)/ ln p
n+1
]
¸
k
n+1
=1
k
h
n+1
=
1
h +1
ln x
ln p
n+1
h+1
+ O(ln
h
x) when x →+∞
and if we substitute this in (29) we have
¸
1
(x) =
c(n)
ln p
n+1
n−1
¸
h=0
(−1)
h
(
n −1
h
)
1
h +1
ln
n
x + O(ln
n−1
x) =
=
c(n)
n ln p
n+1
ln
n
x + O(ln
n−1
x) (31)
in view of the identity
n−1
¸
h=0
(−1)
h
(
n −1
h
)
1
h +1
=
1
n
n−1
¸
h=0
(−1)
h
(
n
h +1
) =
1
n
which holds since
n−1
¸
h=0
(−1)
h
(
n
h
) = (1 −1)
n
= 0
Let us now consider the sum
¸
2
(x) which appears in (28).
Since 0 < ln 2/ ln p
n+1
< 1, ∀n ≥ 1 the sum, if it is not empty, contains at most
one term, namely k
n+1
= [ln x/ ln p
n+1
] : for this value of k
n+1
we obviously have
(x/p
k
n+1
n+1
) ≤ p
n+1
which implies that
(32)
¸
2
(x) =
x
p
k
n+1
n+1
, p
n
≤ ( p
n+1
, p
n
) = O(1)
uniformly in x. From (28), (31) and (32) it follows that
(x, p
n+1
) −(
x
2
, p
n+1
) =
c(n)
n ln p
n+1
ln
n
x + O(ln
n−1
x)
which proves the lemma.
28 P. Codeca  N. Taddia
LEMMA 2. Let
S(x) =
¸
x
2
<h,k<x
(h,k)=1
1
hk
.
Then we have
(33) S(x) =
6
π
2
(ln 2)
2
+o(1)
when x →+∞.
Proof. If we set
σ(x) =
¸
x
2
<h,k<x
1
hk
we can write
σ(x) =
¸
n≤x
1
n
2
S(
x
n
)
and, by M¨ obius inversion formula, we have
S(x) =
¸
n≤x
µ(n)
n
2
σ(
x
n
)
Since it is
σ(x) =
¸
¸
x
2
<h<x
1
h
2
= (ln 2)
2
+ O(
1
x
)
we have
(34) S(x) =
¸
n≤
√
x
µ(n)
n
2
(ln 2)
2
+o(1) =
¸
¸
n≥1
µ(n)
n
2
(ln 2)
2
+o(1)
when x →+∞. From 34 follows (33) if we remember that
¸
n≥1
µ(n)
n
2
=
6
π
2
We return to the proof of Theorem 3. It is based on the BanachSteinhaus theorem:
for N ≥ 1 we consider the bounded linear transformation
N
¸
ℓ
2
→ℓ
2
a →
N
(a) = a
′
L
2
expansions in series 29
deﬁned by
(35) a
′
(n) =
¸
¸
0 if 1 ≤ n ≤ N
1
n
¸
dn
d≤N
da(d)
¸
k(n/d)
k≤(N/d)
µ(k)
if n > N
where we have set a = (a(n)) ∈ ℓ
2
and a
′
= (a
′
(n)). The linearity of
N
is obvious:
let us prove that is also bounded. We have
a
′
(n) ≤
N
n
¸
dn
d≤N
a(d) ≤
N
n
¸
¸
d≤N
a(d)
2
1
2
· N
1
2
≤ N
3
2
a
n
from which the conclusion is immediatly obtained by squaring and summing over
n. We will now prove that the family of linear transformation
N
is not uniformly
bounded. Let p
1
< p
2
< ... < p
n
< ... be the sequence of prime numbers and let
M
+
(n) the greatest prime divisor of the integer n. We deﬁne
(36) a
N
(n) =
¸
1 if
N
2
< n ≤ N and M
+
(n) ≤ p
r
0 otherwise
for p
r
and N ﬁxed.
From formula (35) follows that in this case a
′
N
=
N
(a
N
) is given by
a
′
N
(n) =
¸
¸
0 if 1 ≤ n ≤ N
1
n
¸
dn
N
2
<d≤N
d a
N
(d) if n > N
since the inner sum
¸
k(n/d)
k≤(N/d)
µ(k) reduces to µ(1) = 1.
Consider now
N
(a
N
)
2
=
∞
¸
n=N+1
a
′
N
(n)
2
we have
N
(a
N
)
2
=
¸
N
2
<d
1
,d
2
≤N
a
N
(d
1
)a
N
(d
2
)
¸
{n≥N+1, n≡0 (d
1
), n≡0 (d
2
)}
1
(n/d
1
)
1
(n/d
2
)
=
=
¸
N
2
<d
1
,d
2
≤N
a
N
(d
1
)a
N
(d
2
)
d
1
d
2
(d
1
, d
2
)
2
¸
¸
¸
¸
m>
N(d
1
d
2
)
d
1
d
2
m
−2
If
N
2
< d
1
≤ N and
N
2
< d
2
≤ N we have
N
d
1
d
2
(d
1
, d
2
) < 4 and it follows
(37)
N
(a
N
)
2
≥ c
1
¸
N
2
<d
1
,d
2
≤N
a
N
(d
1
)a
N
(d
2
)
d
1
d
2
(d
1
, d
2
)
2
30 P. Codeca  N. Taddia
where c
1
=
¸
m≥4
m
−2
. From formula (33) of Lemma 2 we have
(38) S(
N
d
) =
¸
N
2d
<h,k≤
N
d
(h,k)=1
1
hk
≥ c
2
> 0
where c
2
is an absolute constant, if
N
d
is sufﬁciently large, say
N
d
> m. Let
f
r
(n) =
¸
1 if M
+
(n) ≤ p
r
0 otherwise
so that
(x, p
r
) =
¸
n≤x
M
+
(n)≤p
r
1 =
¸
n≤x
f
r
(n)
Consider now formula (37): collecting together the terms with the same greatest com
mon divisor we obtain, taking p
r
> m,
N
(a
N
)
2
≥ c
1
N
¸
d=1
¸
¸
¸
¸
¸
N
2d
<h,k≤
N
d
(h,k)=1
a
N
(hd)a
N
(kd)
hk
≥
≥ c
1
¸
N
pr
≤d≤
N
m
f
r
(d)S(
N
d
) ≥ c
1
c
2
¸
N
pr
≤d≤
N
m
f
r
(d) (39)
if we remember deﬁnition (36) and formula (38). From formula (26) of Lemma 1
follows
¸
N
pr
≤d≤
N
m
f
r
(d) ≥
[ln p
r
/ ln2]−1
¸
k=[ln m/ ln2]+1
¸
¸
¸
N
2
k+1
<d≤
N
2
k
f
r
(d)
=
=
[ln p
r
/ ln2]−1
¸
k=[ln m/ ln2]+1
(
N
2
k
, p
r
) −(
N
2
k+1
, p
r
)
=
=
[ln p
r
/ ln2]−1
¸
k=[ln m/ ln2]+1
c(r) ln
r−1
(
N
2
k
) + O(ln
r−2
(
N
2
k
))
= (40)
= c(r) ln
r−1
N
[
ln p
r
ln 2
] −[
ln m
ln 2
] −1
+ O(ln
r−2
N) (41)
for N →+∞, where c(r) is the constant speciﬁed in (26) of Lemma 1. From (39) and
(40) follows
(42)
N
(a
N
)
2
(a
N
)
2
≥ c
1
c
2
c(r) ln
r−1
N
[
ln p
r
ln 2
] −[
lnm
ln2
] −1
+ O(ln
r−2
N)
c(r) ln
r−1
N + O(ln
r−2
N)
L
2
expansions in series 31
since, by the deﬁnition (36) of a
N
(n) and by lemma 1 we have
(a
N
)
2
= c(r) ln
r−1
N + O(ln
r−2
N)
From (42) we obtain
(43) liminf
N→+∞
N
(a
N
)
2
(a
N
)
2
≥ c
1
c
2
[
ln p
r
ln 2
] −[
ln m
ln 2
] −1
= L(r)
where c
1
,c
2
and m are absolute constants. Since sup
r
L(r) = +∞formula (43) proves
that the linear transformation
N
are not uniformly bounded and this implies, by the
BanachSteinhaus theorem, the existence of a sequence a = (a(n)) ∈ ℓ
2
such that
(44) sup
N
N
(a) = +∞
If f (x) ∼
¸
n≥1
a(n) sin(2πnx) and R
N
(x) =
¸
N
n=1
α(n)P(nx) − f (x), where the
coefﬁcients α(n) are given by (7) of Theorem 2, it is easy to see that
(45)
N
(a)
2
≤ 4 R
N
2
+2
¸
∞
¸
n=N+1
a(n)
2
since
R
N
2
=
1
2
∞
¸
n=N+1
 −a
′
(n) +a(n)
2
and
N
(a)
2
=
∞
¸
n=N+1
a
′
(n)
2
if we remember (20), (22) and (35). From (44) and (45) follows (10). This concludes
the proof of Theorem 3.
References
[1] BARY N.K., A treatise on trigonometric series, vol. I, Pergamon Press, London
1964.
[2] BATEMAN P.T. AND CHOWLA S., Some special trigonometric series related to
the distribution of prime numbers, J.London Math Soc. 38 (1963), 372–374.
[3] BESICOVITCH A.S., Problemon continuity, J.London Math Soc. 36 (1961), 388–
392.
[4] CODECA’ P., DVORNICICH R. AND ZANNIER U., Two problems related to the
nonvanishing of L(1, χ), J. de th´ eorie des nombres de Bordeaux 1 (1998), 49–
64.
32 P. Codeca  N. Taddia
[5] DAVENPORT H., On some inﬁnite series involving arithmetical functions,
Q.J.Math. 8 (1937), 8–13.
[6] DAVENPORT H., On some inﬁnite series involving arithmetical functions,
Q.J.Math. 8 (1937), 313–320.
[7] HARDY G.H. AND WRIGHT E.M., An introduction to the theory of numbers,
Oxford University Press, Oxford 1971.
[8] NAGY BELA SZ., Introduction to real functions and orthogonal expansions, Ox
ford University Press, New York 1965.
[9] RADEMACHER H., Topics in analytic number theory, SpringerVerlag, New York
1973.
[10] RUDIN W., Real and complex analysis, McGrawHill, New York 1966.
AMS Subject Classiﬁcation: 42C15, 41A58.
Paolo CODECA and Nicola TADDIA
Dipartimento di Matematica
via Machiavelli, 35
44100 Ferrara, ITALY
email: cod@dns.unife.it
email: tad@dns.unife.it
Lavoro pervenuto in redazione il 10.07.1999 e in forma deﬁnitiva il 30.03.2000.
1]. f (x) = − f (1 − x) a. 1]} 0 that is L 2 is the subspace of the functions f ∈ L 2 ([0. y) with y ﬁxed. the answer to question γ ) is negative. showed that the identities ii) hold everywhere and that the two series converge uniformly in R. We shall consider the following problems: α) Is it possible to apply the GramSchimdt procedure to the functions P(nx) and.N. Moreover. . The proof of the weaker properties ii) (the almost everywhere convergence) depends on L 2 convergence. Le us deﬁne: L 2 = { f ∈ L 2 ([0. This is proved in Theorem 3. we denote with M + (n) the greatest prime which divides n and. 1].M + (n)≤y} 1. y) = {n≤x. On the contrary. using Vinogradov’s method and the SiegelWalﬁsz theorem.18 P. as usual. Davenport. It is easy to see that L 2 is a closed subspace of L 2 ([0. In order to state in a more precise way the problems studied and the results obtained we give the following deﬁnitions. namely on the estimate [5]: 2 1 N RN = 0  n=1 1 log(N) µ(n) P(nx) + sin(2π x)2d x = O( ) n π N The aim of the present paper is just to consider the formal identity (2) for general α(n) and from the point of view of L 2 convergence. if (ϕn (x))n is the resulting orthonormal system. if n is a nonnegative integer.e. while in Theorem 2 we show that it is possible to give a positive answer to question β). Codeca . we set (x. 1]. R) and so it is 0 a Hilbert space. Taddia In a subsequent paper [6]. y) − ( 2 . R) which are odd with respect 0 1 to the point 2 . on [0. is it true that f (x) = in the L 2 sense for every f ∈ L 2 ? 0 β) Is it possible to give sufﬁcient conditions for expanding (in the L 2 sense) in series of fractional parts the functions belonging to certain subsets of L 2 ? 0 γ ) Is it possible to expand (in the L 2 sense) every f ∈ L 2 in series of fractional 0 parts ? b(n)ϕn (x) n≥1 Theorem 1 below answers positively to question α). whose proof depends on the BanachSteinhaus theorem and on an asymptotic formula x with an estimate of the error term (see Lemma 1) for the difference (x. R).
Let us also suppose that 1 h(x) = h(1 − x).L 2 expansions in series 19 R EMARK 1. . as usual. 1] → R be continuous and of bounded variation. χ) is the corresponding L function). is complete in L 2 . Notation and results If a and b are positive integers (a. For example. With τ (n) we will indicate the number of divisors of n. The identity (3) n≥1 µ(n) 1 P(nx) = − sin(2π x) x ∈ R n π has interesting applications. for the proof of (4). in turn. 1] d a=1 It is well known (see [3] and [2]) that this property is false if we drop the bounded variation condition. 2. ii) The set of functions f n (x) = P(nx). it was used in [4] to prove a result of linear independence for the fractional parts.. Moreover it can be shown that also the weaker result that (3) holds in the L 2 sense can be used to prove results like the following one. . We have obtained the following results. is equivalent to the result L(1. Then the following implication holds: d a h( ) = 0 ∀ d ≥ 1 ⇒ h(x) = 0. n = 1. and dn means that d divides n. n = 1.. . Then we have N (5) n=1 b(n)ϕn (x) − f (x) L2 → 0 when N → +∞ for every f ∈ L 2 .. 2. Let h : [0. ∀x ∈ [0. 2.are linearly independent. it is enough to know that (3) holds for rational x and this. where the (b(n)) are the Fourier coefﬁcients of f (x) with 0 respect to the system (ϕn (x)).. ∀ x ∈ [0. . T HEOREM 1. 1] and that 0 f (x)d x = 0. The following properties hold: i) The functions fn (x) = P(nx). b) will denote the greatest common divisor of a and b. 0 iii) Let (ϕn (x)) be the orthonormal system obtained from f n (x) = P(nx) by the GramSchmidt procedure. χ) = 0 (here χ is a caracter mod q and L(s. namely the fact that the determinant: (4) det(P( md )) = 0 q 0≤d ≤m≤ q −1 2 Actually.
Taddia Theorem 1 solves the problem of expanding every function f ∈ L 2 in series of 0 linear combination of fractional parts.n P(hx) h=1 where the γh. T HEOREM 2. . There exist functions f ∈ L 2 such that 0 (10) where R N (x) = sup N N n=1 α(n)P(nx) R N = +∞ − f (x) and the α(n) are given by (7). T HEOREM 3. let a ∗ (n) = If the condition (8) n≥1 h≥1 a(h)2.N.20 P. Codeca .n are suitable coefﬁcients ([8].305). Then we have (6) RN 2 n=1 α(n)P(nx) − f (x) = 1 2 m≥1  1 π α(n) nm n + a(m)2 m n≤N If (7) R N → 0 when N → ∞ we must have α(n) = −π dn µ(d) n a( ). As for the problem of the expansion of functions f ∈ L 2 in series of fractional parts 0 we have the following theorems. d d Finally. with α(n) as in (7). h≡0(n) τ (n) a ∗ (n) 1 < +∞ n holds then N (9) RN = f − n=1 α(n)P(nx) → 0 when N → +∞. Let f (x) ∼ k≥1 a(k) sin(2πkx) N Set be the Fourier series of f ∈ L 2 . p. since we have n ϕn (x) = γh. 0 R N (x) = with α(n) ∈ R.
n = and 0 P(nx)P(mx)d x = 1 1 (n. If f ∈ L 2 . m = 1. (7) and M¨ bius inversion formula. m)2 12 nm bm = N P(mx) f (x)d x 0 or equivalently by considering the sum N n N σ N (x) = n=1 b(n)ϕn (x) = b(n)( n=1 h=1 γh. the minimum value of the difference 0 N f − cn P(nx) n=1 can be obtained directly by solving the linear system ([10]. ∀n ≥ 1. N 1 where am. vol.I.. a(n) ≤ b(n). This is an obvious corollary of (9). p. 1 It should be noted that under condition (8) we can assert that the formal identity (2) holds almost everywhere for a subsequence. The contrast between (5) of Theorem 1 and (10) of Theorem 3 can be explained as follows.I.n P(hx)) = cn P(nx) n=1 . p. 83) N (11) n=1 am. if there exists (b(n)) ⊂ ℓ2 such that b(n + 1) ≤ b(n).215) ( k≥n ak 2 ) 2 = O( 1 1 ) nβ which gives a ∗ (n) = O( n2β ) and (8) holds again. . vol. since in this case we have a ∗ (n) = h≥1 a(hn)2 ≤ h≥1 1 n n r=1 b2 ((h − 1)n + r ) = b n 2 This means that (8) holds in particular for all bounded variation functions.. The proof of Theorem 3 is based on BanachSteinhaus theorem and so we do not give an explicit example. 72). It is certainly satisﬁed.L 2 expansions in series 21 R EMARK 2. since in this 1 case we have a(n) = O( n ) ([1].. for instance.n cn = bm . Condition (8) of Theorem 2 is not a very restrictive one. Another example is given by the class β of functions which satisfy a Lipschitz condition of order β > 0: in this case we have ([1]. p. o R EMARK 3.
) n from which we obtain n (13) k=1 ck = 0 If 1 n ≤x< 1 n−1 we have [kx] = 1 0 if k = n if k ≤ n − 1 and so from (12) follows n−1 k=1 1 3 ck (kx − ) + cn (nx − ) = 0 2 2 1 1 a.. Let us now prove the results stated. If n = 1 the implication is obviously true.e. on [0.. and consequently the difference N f − n=1 α(n)P(nx) is not the minimal one. on [ . Suppose now that the equality (12) holds with n ≥ 2. n. Usually.e.e. . ) n n−1 . n bm = 1 2π 2 0 if m = 1 otherwise and the numbers µ(n)/n do not satisfy condition (11). Proof of Theorem 1. on [0. Taddia which appears in (5) of Theorem 1.N. 1] ck ∈ R implies ck = 0 ∀k = 1. if the function is not a very irregular one (see condition (8) of Theorem 2) N the sum SN (x) = n=1 α(n)P(nx) will be a good approximation to σ N (x) and it will happen that SN − σ N → 0 when N → +∞. We want to prove that the relation n (12) k=1 ck P(kx) = 0 a. Generally the α(n) given by (7) of Theorem 2 are not the solution of system (11): for example it is easy to check directly that when 1 f (x) = − π sin(2π x) we have α(n) = µ(n) . Codeca . Proof of property i).. But there exist very irregular functions f ∈ L 2 for which we have 0 sup N R N = sup N SN − σ N = +∞ although we think that is not easy to give an explicit example.22 P. 1 If 0 ≤ x < n we have [kx] = 0 for every k ≤ n and from (12) follows n k=1 1 ck (kx − ) = 0 2 1 a.
Proof of property ii). If (ϕn (x)) is the orthonormal system obtained from P(nx) by the GramSchmidt procedure we have ([8].n P(hx) h=1 n P(nx) = γh. Since 0 P(x) = − we have (15) δ(n) = − 1 π 1 k 1 0 1 π k≥1 sin(2πkx) k k≥1 sin(2πknx) f (x)d x = − 1 2π k≥1 c(kn) k where the c(n) are the Fourier coefﬁcients of f . p. 305) n ϕn (x) = and viceversa γh.n ϕk (x) k=1 ′ . This gives property i). and property ii) follows. Proof of property iii). but this means that f ∈ L 2 has all the Fourier coefﬁcients equal to zero. 1 Take f ∈ L 2 and set δ(n) = 0 f (x)P(nx)d x. The completeness 0 of the trigonometrical system implies f (x) = 0 a.e. and so (16) holds.L 2 expansions in series 23 which in turn implies n−1 (14) k=1 ck + 3cn = 0 Relations (13) and (14) give cn = 0 and the desidered implication follows by induction. If we suppose δ(n) = 0 ∀n ≥ 1 from (16) follows c(n) = 0 ∀n ≥ 1. But (17) is certainly true by Schwarz inequality. The relation (15) can be inverted and we have (16) c(n) = −2π µ(k) k≥1 δ(nk) k provided the sufﬁcient condition (17) k≥1 τ (k)c(nk) < +∞ k ∀n ≥ 1 is satisﬁed ([7]. Theorem 270).
Taddia where the γh.n are suitable coefﬁcients. that the orthonormal system (ϕn (x)) is complete in L 2 and so Parseval identity holds. by property ii).N. Codeca . This 0 proves (5) and concludes the proof of Theorem 1. But we also have (21) − 1 π α(n) nm n = m a(d) dm d ( m µ(k)) k(m/d) n≤N d≤N k≤(N/d) . Let us now prove Theorem 2. Formula (6) is simply Parseval identity for R N (x) = α(n)P(nx) − f (x) in fact we have √ 1 2 0 R N (x) sin(2πmx)d x = c N (m) = (18) 1 n 1 = − √ π ( nm α(n) m ) + a(m) 2 n≤N N n=1 ′ To justify (18) we note that √ 2 0 1 √ 2 P(nx) sin(2πmx)d x = − π = −n √ π 2m k≥1 1 k 1 0 sin(2πknx) sin(2πmx) = 0 2 if nm otherwise from (18) we obtain (6).24 P. Proof of Theorem 2. From (19) it follows (20) R N → 0 ⇐⇒ ∞ m=N+1  α(n) nm n 2  →0 m n≤N since (a(m)) ∈ ℓ2 . This implies. n Let us now suppose that condition (7) holds: in this case (6) becomes 2 (19) RN = 1 2 ∞ m=N+1  1 π α(n) nm n + a(m)2 m n≤N when N → +∞. Let us prove (7).n and γh. since the Dirichlet inverse of n is µ(n) . Since we have If RN 2= m≥1 c N (m) R N → 0 from (6) it follows immediatly that 1 ( π α(n) nm n ) = −a(m) m ∀m ≥ 1 1 from which we obtain (7).
k) hk kn (24) ≤ = ( m≥1 ( m≥1 τ 2 (m) m2 τ 2 (m) ) m2 ) h.s≥1 τ (r )τ (s) rs δ≥1 a(r δ)a(sδ) ≤ a(r δ)2) 2 ( 1 r. (r. if we remember that τ (nm) ≤ τ (n)τ (m).L 2 expansions in series 25 From (20) and (21) follows (22) when N → ∞. k) = δ so that h = r δ. k = sδ. k) (h. From the deﬁnition of γ (n) in (23) follows easily γ 2 (n) n≥1 = n≥1 hn n n hk a(h)a(k)τ ( )τ ( ) 2 ≤ h k n k h (h. s) = 1 we also have ≤ (25) ≤ r. k)2 a(h)a(k)τ ( )τ ( ) = (h. From (24) and (25) follows that the series n≥1 γ 2 (n) is convergent. First we note that ′ (23) a N (n) ≤ hn n h a(h)τ ( ) ≡ γ (n) h n uniformly in N. k≤(N/d) d≤N Let us now prove that (8) implies (9). but this implies that condition (22) is satisﬁed. since from (23) it follows obviously that ∞ n=N+1 2 √ τ (r ) a∗(r ) < +∞ r a N (n)2 ≤ ′ ∞ n=N+1 γ 2 (n) → 0 . If (h.k≥1 say.s≥1 τ (r )τ (s) ( rs δ≥1 δ≥1 a(sδ)) 2 = 1 = r≥1 if condition (9) holds. where a N (m) = ′ R N → 0 ⇐⇒ ∞ m=N+1 a N (m)2 → 0 ′ 1 m da(d)( dm k(m/d) µ(k)).
pn ) − . We have [ln y/ ln pn+1 ] (27) (y..... L EMMA 1. pn ) + (x). set (x.N. p1 ) = 1 2 Now we will prove that if (26) holds for the integer n then it holds also for n + 1 and the lemma will follow by induction. From (27) with y = x and y = (x. This proves (10) and concludes the proof of Theorem 2. p1) − ( . If M + (k) denotes the greatest prime divisor of the integer k.. pnn kn+1 pn+1 ≤y 1= y kn+1 pn+1 x 2 .26 P. pn ) = ( n+1 2 pn+1 k ( [ln x / ln pn+1 ]<kn+1 ≤[ln x/ ln pn+1 ] 2 x kn+1 pn+1 1 (x) + where 2 (x) is zero if the sum is empty. pn ) = 1 M + (k)≤ pn k≤x Then we have (26) where c(n) = 1 1 (n−1)! n −1 k=2 (ln( pk )) (x. pn+1 ) = 2 [ln x / ln pn+1 ] 2 kn+1 =0 we obtain x . with p1 = 2. < pn be the prime numbers up to pn . Proof of Theorem 3. with the convention M + (1) = 1. . n + 1. Taddia when N → +∞. Let p1 < p2 < . pn ) − x ( . We need the following two lemmas. pn kn+1 =0 with k j ≥ 0 for j = 1. Consider now esis we have x [ln 2 / ln pn+1 ] 1 (x): by induction hypoth 1 (x) = = + c(n) kn+1 =0 n−1 lnn−1 x kn+1 pn+1 + O lnn−1 x = (29) c(n) h=0 O ln n−1 x [ln 2 / ln pn+1 ] n−1 h (−1)h ( ) lnn−1−h x · lnh pn+1 kn+1 + h kn+1 =0 x .. because in this case we have x ∀x ≥ 2 (x. pn+1) − (28) + x ( . pn ) = c(n) lnn−1 (x) + O(lnn−2 (x)) 2 if n = 1 if n ≥ 2 when x → +∞ Proof. 2 ( x kn+1 pn+1 . pn+1 ) = k k p11 .. Codeca . Formula (26) is obviously true if n = 1.
(31) and (32) it follows that (x. pn ≤ ( pn+1. pn ) = O(1) uniformly in x. Since 0 < ln 2/ ln pn+1 < 1. if it is not empty. = 1 2 h+1 (30) where the B j are the Bernoulli numbers ([9]. From (28). if n = 1 the above expression must obviously be interpreted as 1 (x) = c(1) ln x + O(1) ln p2 We now recall that 1h + 2h + . pn+1 ) − which proves the lemma. pn+1 ) = lnn x + O(lnn−1 x) 2 n ln pn+1 . contains at most one term. 65).L 2 expansions in series 27 where. x c(n) ( ... + (m − 1)h = 1 h+1 h+1 m h+1 + ( )B1 m h + ( )B2 m h−1 + . ∀n ≥ 1 the sum... If we use the expression (30) in (29) we obtain [ln( x )/ ln pn+1 ] 2 kn+1 =1 h kn+1 = 1 h+1 n−1 ln x ln pn+1 h+1 + O(lnh x) when x → +∞ and if we substitute this in (29) we have (x) 1 = = c(n) ln pn+1 (−1)h ( h=0 1 n−1 ) lnn x + O(lnn−1 x) = h h+1 (31) c(n) lnn x + O(lnn−1 x) n ln pn+1 n−1 in view of the identity n−1 h=0 1 1 n−1 (−1) ( ) = h h+1 n h n−1 (−1)h ( h=0 1 n )= h+1 n which holds since (−1)h ( h=0 n ) = (1 − 1)n = 0 h Let us now consider the sum 2 (x) which appears in (28). namely kn+1 = [ln x/ ln pn+1 ] : for this value of kn+1 we obviously have kn+1 (x/ pn+1 ) ≤ pn+1 which implies that (32) 2 (x) = x kn+1 pn+1 . p.
hk (h.k)=1 S(x) = 6 (ln 2)2 + o(1) π2 x 2 <h.28 P. by M¨ bius inversion formula. It is based on the BanachSteinhaus theorem: for N ≥ 1 we consider the bounded linear transformation N ℓ2 → ℓ2 ′ a → N (a) = a . If we set σ (x) = we can write σ (x) = n≤x x <h. Let S(x) = Then we have (33) when x → +∞.N.k<x 1 hk 1 x S( ) n2 n and. we have o S(x) = Since it is σ (x) = µ(n) x σ( ) n2 n n≤x x 2 <h<x we have (34) S(x) = 2 1 1 = (ln 2)2 + O( ) h x µ(n) (ln 2)2 + o(1) n2 √ n≤ x µ(n) (ln 2)2 + o(1) = n2 n≥1 when x → +∞. Taddia L EMMA 2. Proof. From 34 follows (33) if we remember that n≥1 6 µ(n) = 2 n2 π We return to the proof of Theorem 3. Codeca .k<x 2 1 .
d2 )2 d1 d2 ≥ c1 m> N(d1 d2 ) d1 d2 N d1 d2 (d1 . d2 ) < 4 and it follows m −2 (37) N (a N ) N 2 <d1 . We will now prove that the family of linear transformation N is not uniformly bounded.L 2 expansions in series 29 deﬁned by (35) a (n) = ′ 0 1 n if 1 ≤ n ≤ N dn da(d) ′ k(n/d) µ(k) if n > N N d≤N k≤(N/d) ′ where we have set a = (a(n)) ∈ ℓ2 and a = (a (n)). n≡0 (d2 )} 1 1 = (n/d1 ) (n/d2 ) = If N 2 N 2 <d1 . We have N a (n) ≤ n ′ is obvious: dn d≤N N a(d) ≤ n d≤N a(d) 2 1 2 ·N2 ≤ N2 1 3 a n from which the conclusion is immediatly obtained by squaring and summing over n... We deﬁne (36) a N (n) = 1 0 if N < n ≤ N and M + (n) ≤ pr 2 otherwise for pr and N ﬁxed. ′ From formula (35) follows that in this case a N = N (a N ) is given by 0 if 1 ≤ n ≤ N ′ 1 a N (n) = d a N (d) if n > N dn n N 2 <d≤N since the inner sum Consider now k(n/d) k≤(N/d) µ(k) reduces to µ(1) = 1. < pn < . n≡0 (d1 ). N (a N ) 2 = ∞ n=N+1 a N (n)2 ′ we have N (a N ) 2 = a N (d1 )a N (d2 ) N 2 <d1 .d2 ≤N a N (d1 )a N (d2 ) (d1 . Let p1 < p2 < .d2 ≤N N 2 < d1 ≤ N and < d2 ≤ N we have 2 a N (d1 )a N (d2 ) (d1 . The linearity of let us prove that is also bounded.d2 ≤N {n≥N+1. d2 )2 d1 d2 ... be the sequence of prime numbers and let M + (n) the greatest prime divisor of the integer n.
N. if fr (n) = so that N d is sufﬁciently large. N N (a N ) 2 ≥ c1 d=1 N <h. From formula (33) of Lemma 2 we have S( N )= d 1 ≥ c2 > 0 hk N d N <h. Let (x. pr ) = M + (n)≤ pr n≤x Consider now formula (37): collecting together the terms with the same greatest common divisor we obtain.k)=1 where c2 is an absolute constant. From formula (26) of Lemma 1 follows [ln pr / ln 2]−1 fr (d) N pr N ≤d≤ m ≥ k=[ln m/ ln 2]+1 [ln pr / ln 2]−1 N 2k+1 <d≤ N k 2 = (40) (41) = ( k=[ln m/ ln 2]+1 [ln pr / ln 2]−1 N .k≤ N d 2d (h. taking pr > m.k)=1 a N (hd)a N (kd) ≥ hk N ) ≥ c1 c 2 d N pr (39) ≥ c1 N pr N ≤d≤ m fr (d)S( fr (d) N ≤d≤ m if we remember deﬁnition (36) and formula (38). Taddia where c1 = (38) m≥4 m −2 . Codeca . pr ) = 2k+1 c(r ) lnr−1 ( k=[ln m/ ln 2]+1 N N ) + O(lnr−2 ( k )) = k 2 2 = c(r ) lnr−1 N [ ln pr ln m ]−[ ] − 1 + O(lnr−2 N) ln 2 ln 2 for N → +∞. pr ) − 2k fr (d) = ( N . where c(r ) is the constant speciﬁed in (26) of Lemma 1. say 1 0 if M + (n) ≤ pr otherwise 1= fr (n) n≤x > m.k≤ N d 2d (h.30 P. From (39) and (40) follows (42) N (a N ) 2 2 (a N ) ≥ c1 c2 c(r ) lnr−1 N [ ln pr ] − [ ln m ] − 1 + O(lnr−2 N) ln 2 ln 2 c(r ) lnr−1 N + O(lnr−2 N) .
388– 392.. [2] BATEMAN P. 49– e 64. [3] B ESICOVITCH A..c2 and m are absolute constants..K. (22) and (35). the existence of a sequence a = (a(n)) ∈ ℓ2 such that (44) sup N N (a) = +∞ (45) N If f (x) ∼ n≥1 a(n) sin(2πnx) and R N (x) = n=1 α(n)P(nx) − f (x). 36 (1961). 38 (1963). Since supr L(r ) = +∞ formula (43) proves that the linear transformation N are not uniformly bounded and this implies. AND C HOWLA S.L 2 expansions in series 31 since. χ). it is easy to see that N (a) 2 ≤4 RN 2 since RN and N (a) 2 +2 ′ ∞ n=N+1 a(n)2 = 1 2 ∞ n=N+1  − a (n) + a(n)2 ∞ 2 = n=N+1 a (n)2 ′ if we remember (20). by the BanachSteinhaus theorem.London Math Soc.. Two problems related to the nonvanishing of L(1. Some special trigonometric series related to the distribution of prime numbers. I. Problem on continuity. vol. de th´ orie des nombres de Bordeaux 1 (1998). J. 372–374. London 1964. where the coefﬁcients α(n) are given by (7) of Theorem 2. This concludes the proof of Theorem 3. .S. [4] C ODECA’ P.T. A treatise on trigonometric series. J. From (44) and (45) follows (10). References [1] BARY N. by the deﬁnition (36) of a N (n) and by lemma 1 we have (a N ) From (42) we obtain (43) lim inf N (a N ) 2 2 2 = c(r ) lnr−1 N + O(lnr−2 N) ln m ln pr ]−[ ] − 1 = L(r ) ln 2 ln 2 N→+∞ (a N ) ≥ c 1 c2 [ where c1 .. DVORNICICH R. AND Z ANNIER U.London Math Soc. J. Pergamon Press.
Oxford University Press.1999 e in forma deﬁnitiva il 30. New York 1966. 8 (1937). Q.Math. Oxford University Press. [6] DAVENPORT H. McGrawHill.J. [9] R ADEMACHER H.. ITALY email: cod@dns.M. [7] H ARDY G. 8 (1937).N. New York 1973.2000.unife. Codeca . [8] NAGY B ELA S Z . An introduction to the theory of numbers.it Lavoro pervenuto in redazione il 10. AND W RIGHT E. 41A58.. 313–320.03. Oxford 1971.. Q. SpringerVerlag. 8–13. New York 1965. Introduction to real functions and orthogonal expansions.Math. AMS Subject Classiﬁcation: 42C15.. 35 44100 Ferrara. .07.32 P. On some inﬁnite series involving arithmetical functions. [10] RUDIN W..H.J.unife.. Paolo CODECA and Nicola TADDIA Dipartimento di Matematica via Machiavelli.it email: tad@dns. Real and complex analysis. Topics in analytic number theory. Taddia [5] DAVENPORT H. On some inﬁnite series involving arithmetical functions.