Journal of Number Theory 130 (2010) 24042427

Contents lists available at ScienceDirect

Journal of Number Theory
www.elsevier.com/locate/jnt
Weyl sums over integers with ane digit restrictions
Michael Drmota
a,,1
, Christian Mauduit
b
a
Vienna University of Technology, Institute of Discrete Mathematics and Geometry, Wiedner Hauptstrasse 8-10, A-1040 Vienna, Austria
b
Universit de la Mditerrane, Institut de Mathmatiques de Luminy, CNRS UMR 6206, Campus de Luminy, Case 907,
13288 Marseille Cedex 9, France
a r t i c l e i n f o a b s t r a c t
Article history:
Received 16 December 2009
Revised 26 April 2010
Available online 2 July 2010
Communicated by David Goss
MSC:
11A63
11K06
05A15
05A16
Keywords:
Digital expansions
Uniform distribution modulo 1
Generating functions
Saddle point method
For any given integer q 2, we consider sets N of non-negative
integers that are dened by ane relations between their q-adic
digits (for example, the set of non-negative integers such that
the number of 1s equals twice the number of 0s in the binary
representation). The main goal is to prove that the sequence
(n)
nN
is uniformly distributed modulo 1 for all irrational
numbers . The proof is based on a saddle point analysis of certain
generating functions that allows us to bound the corresponding
Weyl sums.
2010 Elsevier Inc. All rights reserved.
1. Introduction
Let q 2 be a given integer and let
n =
L

j=0

j
(n)q
j
*
Corresponding author.
E-mail addresses: michael.drmota@tuwien.ac.at (M. Drmota), mauduit@iml.univ-mrs.fr (C. Mauduit).
URL: http://www.dmg.tuwien.ac.at/drmota/ (M. Drmota).
1
Supported by the Austrian Science Foundation FWF, project S9604, part of the Austrian National Research Network Analytic
Combinatorics and Probabilistic Number Theory.
0022-314X/$ see front matter 2010 Elsevier Inc. All rights reserved.
doi:10.1016/j.jnt.2010.04.006
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2405
be the q-ary digital expansion of n with digits
j
(n) {0, 1, . . . , q 1},
L
(n) = 0, and L = L(n) =
[log
q
n] denotes the length of the expansion of n. Further, for {0, 1, . . . , q 1} let
|n|

:= card
_
j N: 0 j L,
j
(n) =
_
denote the number of digits of n that equal .
For example, the q-ary sum-of-digits function is given by
s
q
(n) =
L

j=0

j
(n) =
q1

=0
|n|

.
Several works concern the study of statistical properties of sequences of integers dened by digital
properties: distribution in residue classes [12,15,16,2022] uniform distribution modulo 1 [2,57,18,
13,19] and study of the associated exponential sums [1,3,4,14]; see also [23] for a description of the
links to spectral analysis and properties of symbolic dynamical systems.
The purpose of this paper is to study, for any xed irrational number , the distribution modulo 1
of the sequence (n)
nN
, where N is a set of integers dened by ane properties of their digits.
Denition 1. We say that the system L = (L
k
)
k=1,...,K
of linear forms on R
q
dened for every
(x
0
, . . . , x
q1
) R
q
by
L
k
(x
0
, x
1
, . . . , x
q1
) =
q1

=0
a
k,
x

, k = 1, . . . , K
(with a
k,
Z for (k, ) {1, . . . , K} {0, . . . , q 1}) is complete if
(i) the family of vectors formed by (a
1,0
, . . . , a
1,q1
), . . . , (a
K,0
, . . . , a
K,q1
) and (1, . . . , 1) is linearly
independent over Q.
(ii) span
Z
{L
k
(n
0
, . . . , n
q1
): k = 1, . . . , K, (n
0
, . . . , n
q1
) Z
q
} =Z
K
(where span
Z
(A) denotes the set
of all nite linear combinations of elements of A with integer coecients).
Let L be a complete system of linear forms over R
q
and = (
1
, . . . ,
K
) be a K-tuple of non-
negative real numbers.
Denition 2. We say that is L-admissible if the system of equations
L
k
(x
0
, x
1
, . . . , x
q1
) =
k
, k = 1, . . . , K,
x
0
+ + x
q1
= 1
has a positive solution x
0
> 0, x
1
> 0, . . . , x
q1
> 0.
Example 1. If K = 1, L= (L
1
) with L
1
(x
0
, . . . , x
q1
) =

q1
=0
x

, then = (
q1
2
) is L-admissible.
Example 2. If K 1, L = (L
1
, . . . , L
k
) with L
k
(x
0
, . . . , x
q1
) = x
0
x
k
for k = 1, . . . , K, then =
(0, . . . , 0) is L-admissible.
Example 3. If K = 1, L= (L
1
) with L
1
(x
0
, . . . , x
q1
) = x
0
2x
1
, then = (0, . . . , 0) is L-admissible.
2406 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
For any complete system L of linear forms over R
q
, for any L-admissible K-tuple (R
+
)
K
and
for any K-tuple = (
1
, . . . ,
K
) Z
K
we dene the set of integers
N =N(L, , )
=
_
n N: L
k
_
|n|
0
, |n|
1
, . . . , |n|
q1
_
= [
k
log
q
n] +
k
, k = 1, . . . , K
_
. (1)
In what follows we will always assume that L is complete and that is L-admissible.
In Section 3 we will give the following estimate for card{n N: n < N}:
Theorem 1. There exist positive constants C
1
, C
2
, and with 0 < 1 depending only on L, , and such
that for any integer N 2 we have
C
1
N

(log
q
N)
K/2
card{n N: n < N} C
2
N

(log
q
N)
K/2
.
In Section 4 we prove our main result:
Theorem 2. For any irrational number the sequence (n)
nN
is uniformly distributed modulo 1.
Such a kind of theorem has been proved in [13] in the particular case of sequences of integers
with an average sum of digits. More precisely, for any b : N R such that
q1
2
+b() N for any
1 and such that the sequence (
b()

1/4
)
1
is bounded, then Theorem 1.2 from [13] says that for any
irrational number the sequence (n)
nE
b
is uniformly distributed modulo 1, where
E
b
=
_
n N: s
q
(n) =
q 1
2
[log
q
n] +b
_
[log
q
n]
_
_
.
It is easy to verify that in the particular case where q = 2 or 3, L and dened as in Example 1
and = (0), our theorem is a consequence of Theorem 1.2 from [13] but that these results are
formally disjoint when q 4. Nevertheless the study of [13] concerns the case of integers whose sum
of digits is close to the expected value and our work generalizes this study to the case of integers
whose sum of digits (or any other linear combination of digits) is far from the expected value.
In the particular case where L and are dened as in Example 2 and for = (0, . . . , 0), our
theorem corresponds to Theorem 4.2 (for the set E
q1
) from [19]. The main theorem from [19] can be
understood as a uniform distribution result in the case where the set N dened by (1) is generated
by a deterministic q-innite automaton corresponding to a random walk of zero average on a d-
dimensional lattice (see [19] for denitions of these notions).
For example when q = 3, L and as in Example 2 and = (0, . . . , 0), the set N = {n N: |n|
0
=
|n|
1
= |n|
2
} is generated by the deterministic 3-innite automaton (that is depicted in Fig. 1) with 0
as initial state and 0 as unique nal state.
The theorem we prove here is a generalization of this result to the case of any random walk on a
q-dimensional lattice (the more general case of d-dimensional lattices, with d q, corresponds to the
generalization suggested in Section 5 of our paper).
Indeed for L and as in Example 3 and = (0), the set N = {n N: |n|
0
= 2|n|
1
} is generated
by the deterministic q-innite automaton that is depicted in Fig. 2 with 0 as initial state and 0 as
unique nal state. It is also linked to the random walk on the lattice Z with probability transitions
(
1
3
,
2
3
). It follows in particular from our main theorem that for any irrational number the sequence
(n)
|n|
0
=2|n|
1
is uniformly distributed modulo 1 (such a kind of result was out of reach of the methods
developed in [13] and [19]).
It follows from Weyls criterion that in order to prove Theorem 2, it is enough to show that for
every irrational number we have
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2407
Fig. 1. Innite 3-automaton generating n N with |n|
0
= |n|
1
= |n|
2
.
Fig. 2. Innite q-automaton generating n N with |n|
0
= 2|n|
1
.

nN, nN
e(n) =o
_
card{n N: n N}
_
as N , where we denote e(x) = e
2ix
(for general references to the theory of uniformly dis-
tributed sequences we refer to [17] and [11]). For this purpose we use a specic saddle point method
applied to properly chosen generating functions. In Section 2 we set up the generating functions re-
lated to our problem. Then in Section 3 we collect some technical properties that are necessary to
apply a saddle point analysis which leads to a proof of Theorem 1. A variation of this method leads
then in Section 4 to a proof of Theorem 2. Finally we comment on some generalizations of Theorem 2
concerning missing digits (Section 5.1) and non-integer coecients (Section 5.2).
2. Generating functions
We rst present explicit formulas for
S
N
(x
0
, x
1
, . . . , x
q1
, y) =

0<n<N
x
|n|
0
0
x
|n|
1
1
x
|n|
q1
q1
y
n
.
Lemma 1. Dene T
,N
(x
0
, x
1
, . . . , x
q1
, y) for N q

recursively by
T
,q
(x
0
, x
1
, . . . , x
q1
, y) =

r<
_
x
0
+ x
1
y
q
r
+ + x
q1
y
(q1)q
r
_
,
T
,q
j (x
0
, x
1
, . . . , x
q1
, y) = x
j1
0
_
x
0
+ x
1
y
q
j
+ + x
1
y
(1)q
j
_
T
j,q
j (x
0
, x
1
, . . . , x
q1
, y) for 1 <q and j < , and by
T
,q
j
+N
(x
0
, x
1
, . . . , x
q1
, y) = T
,q
j (x
0
, x
1
, . . . , x
q1
, y) + x
j1
0
x

y
q
j
T
j,N
(x
0
, x
1
, . . . , x
q1
, y) for 1 <q and N

<q
j
.
2408 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
Then we have
S
q
(x
0
, x
1
, . . . , x
q1
, y) =

j<
_
x
1
y
q
j
+ + x
q1
y
(q1)q
j
_
T
j,q
j (x
0
, x
1
, . . . , x
q1
, y),
S
q
(x
0
, x
1
, . . . , x
q1
, y) = S
q
(x
0
, x
1
, . . . , x
q1
, y) +
_
x
1
y
q

+ + x
1
y
(1)q

_
T
,q
(x
0
, x
1
, . . . , x
q1
, y) for 2 <q, and
S
q

+N
(x
0
, x
1
, . . . , x
q1
, y) = S
q
(x
0
, x
1
, . . . , x
q1
, y)
+ x

y
q
j
T
,N
(x
0
, x
1
, . . . , x
q1
, y) for N

<q

.
Proof. First we give an alternative denition for T
,N
(x
0
, x
1
, . . . , x
q1
, y). Suppose that we consider
all number n <q

in the form n =
0
(n) +
1
(n)q + +
1
(n)q
1
. Similarly to the above we set
|n|
,
:= card
_
j N: 0 j <,
j
(n) =
_
.
Of course, if n <q

and = 0 then |n|

,
= |n|

. However, for = 0 we usually have |n|

,0
= |n|
0
since
|n|
,0
takes all zero digits up to q 1 into account. Now set (for N q

)
T
,N
(x
0
, x
1
, . . . , x
q1
, y) =

n<N
x
|n|
,0
0
x
|n|
,1
1
x
|n|
,q1
q1
y
n
.
With help of this denition the proof of Lemma 1 is immediate. 2
Corollary 1. Suppose that x
0
, x
1
, . . . , x
q1
are complex numbers that are suciently close to the positive real
axis and x
0
+ + x
q1
= 1. Then we have

0<n<N
x
|n|
0
0
x
|n|
1
1
x
|n|
q1
q1
= G(x
0
, . . . , x
q1
, log
q
N) (x
0
+ + x
q1
)
log
q
N

x
1
+ + x
q1
x
0
+ x
1
+ + x
q1
1
,
where G(x
0
, x
1
, . . . , x
q1
, t) is a function that is analytic in x
0
, x
1
, . . . , x
q1
and continuous and periodic in t
(with period 1).
Furthermore, if I is any closed interval of positive real numbers with min I > 1/q, then, for every > 0
such that there is at least one j with |arg(x
j
)| , there exist > 0 and C > 0 such that

0<n<N
x
|n|
0
0
x
|n|
1
1
x
|n|
q1
q1

C
_
|x
0
| +|x
1
| + +|x
q1
|
_
(1) log
q
N
(2)
uniformly for all x
j
with |x
j
| I .
Proof. We rst provide a corresponding representation for T
,N
. Suppose that the q-adic expansion
of N

is given by
N

=
1
q
k
1
+
2
q
k
2
+ +
L
q
k
L
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2409
with digits
j
such that 0 <
j
<q and exponents k
1
>k
2
> >k
L
0, then it directly follows that
T
k
1
+1,N
(x
0
, x
2
, . . . , x
q1
, 1) = (x
0
+ + x

1
1
)X
k
1
+ x
k
1
k
2
1
0
x

1
(x
0
+ + x

2
1
)X
k
2
+ x
k
1
k
3
2
0
x

1
x

2
(x
0
+ + x

3
1
)X
k
3
+
+ x
k
1
k
L
L+1
0
x

1
x

L1
(x
0
+ + x

L
)X
k
L
,
where X abbreviates X = x
0
+ x
1
+ + x
q1
. Further we have
S
q
(x
0
, x
2
, . . . , x
q1
, 1) = (X x
0
)
X

1
X 1
and (for 2)
S
q
(x
0
, x
2
, . . . , x
q1
, 1) = (X x
0
)
X

1
X 1
+(x
1
+ + x
1
)X

= (X x
0
)
X

1
X 1
x
0
X

+(x
0
+ + x
1
)X

.
Consequently, if N is given by
N =
0
q
k
0
+
1
q
k
1
+ +
L
q
k
L
then we have
S
N
(x
0
, x
2
, . . . , x
q1
, 1) = (X x
0
)
X
k
0
1
X 1
x
0
X
k
0
+(x
0
+ + x

0
1
)X
k
0
+ x
k
0
k
1
1
0
x

0
(x
0
+ + x

1
1
)X
k
1
+ x
k
0
k
2
2
0
x

0
x

1
(x
0
+ + x

2
1
)X
k
2
+
+ x
k
0
k
L
L
0
x

0
x

L1
(x
0
+ + x

L
)X
k
L
. (3)
For 0 t < 1 let the q-adic expansion of q
t
be given by
q
t
=
0
+

j1

j
q
k
j
with digits 0 <
j
<q and exponents 0 <k
1
<k
2
< and set
2410 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
G(x
0
, x
2
, . . . , x
q1
, t) = X
t
_
X(1 x
0
)
X 1
+(x
0
+ + x

0
1
)
+
x

0
x
0
(x
0
+ + x

1
1
)
_
X
x
0
_
k
1
+
x

0
x
0
x

1
x
0
(x
0
+ + x

2
1
)
_
X
x
0
_
k
2
+
_
.
It is an easy exercise to show that G is continuous in t and can be periodically extended to a (con-
tinuous) function with period 1 provided x
0
, . . . , x
q1
are suciently close to the positive real line. In
fact G is Hlder continuous with a positive exponent depending on x
0
, . . . , x
q1
(compare with [10]).
Furthermore, by denition it follows that
S
N
(x
0
, . . . , x
q1
, 1) = G(x
0
, . . . , x
q1
, log
q
N) X
log
q
N

X x
0
X 1
.
Finally, if we assume that |x
j
| I and |arg(x
j
)| for some j and for some closed interval I of
positive real numbers then the representation (3) implies (2) almost immediately. Note that min I >
1/q implies that |x
0
| + +|x
q1
| > 1. 2
Corollary 2. Set
P
N
(z
1
, . . . , z
K
, y) =

n<N
K

k=1
z
L
k
(|n|
0
,...,|n|
q1
)
k
y
n
.
Then we have
P
N
(z
1
, . . . , z
K
, y) = S
N
_
K

k=1
z
a
k,0
k
, . . . ,
K

k=1
z
a
k,K
k
, y
_
.
Consequently, there exists a function H(z
1
, . . . , z
k
, t) that is analytic in z
1
, . . . , z
k
(if they are suciently close
to the positive real axis) and continuous and periodic in t (with period 1) such that
P
N
(z
1
, . . . , z
K
, 1) = H(z
1
, . . . , z
k
, log
q
N) F (z
1
, . . . , z
k
)
log
q
N

F (z
1
, . . . , z
k
)

K
k=1
z
a
k,0
k
F (z
1
, . . . , z
k
) 1
, (4)
where we assume that
F (z
1
, . . . , z
k
) =
q1

=0
K

k=1
z
a
k,
k
= 1.
Furthermore, if J is any closed interval of positive real numbers with the property that F (|z
1
|, . . . , |z
K
|) > 1
for all z
k
with |z
k
| J (1 k K), then, for every > 0 such that there is at least one k with |arg(z
k
)| ,
there exist > 0 and C > 0 such that

P
N
(z
1
, . . . , z
K
, 1)

C F
_
|z
1
|, . . . , |z
k
|
_
(1) log
q
N
(5)
uniformly for all z
k
with |z
k
| J .
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2411
Proof. We just have to note that if we set x

K
k=1
z
a
k,
k
then we obtain
q1

=0
x
|n|

=
K

k=1
z

q1
=0
a
k,
|n|

k
and can apply Corollary 1. In particular note that Denition 1(ii) implies that (2) translates to (5). 2
In what follows we will make the assumption that
a
k,0
= 0 (1 k K). (6)
This implies that x
0
in S
N
(x
0
, . . . , x
q1
) is substituted by

K
k=1
z
a
k,0
k
= 1. Hence, F (z
1
, . . . , z
K
) is of the
form
F (z
1
, . . . , z
K
) = 1 +
q1

=1
K

k=1
z
a
k,
k
.
In particular, we always have
F (z
1
, . . . , z
K
) > 1
for all positive real numbers z
1
, . . . , z
K
.
The assumption (6) is no real restriction. If we start with the general linear forms
L
k
(x
0
, x
1
, . . . , x
q1
) =
q1

=0
a
k
x

,
then the slightly modied linear forms
L
k
(x
0
, x
1
, . . . , x
q1
) =
q1

=0
(a
k,
a
k,0
)x

=
q1

=1
(a
k,
a
k,0
)x

have the property that the corresponding coecients a

k,
=a
k,
a
k,0
satisfy a
k,0
= 0 and the condi-
tion (1) translates to
L
k
_
|n|
0
, |n|
1
, . . . , |n|
q1
_
= L
k
_
|n|
0
, |n|
1
, . . . , |n|
q1
_
a
k,0
[log
q
n]
= [
k
log
q
n] +
k
a
k,0
[log
q
n]
=
_
(
k
a
k,0
) log
q
n
_
+
k
+ O(1), (7)
where the O(1)-term depends on n and k. This means that if we replace the linear forms L
k
by L
k
then (1) is replaces by (7) that is almost of the same form. In fact, the following calculations could be
worked out, too, by using (7) instead of (1). However, in this case it would be necessary to keep track
of k and n which would make notations even more involved. Therefore we have decided to work with
(1) and, of course, with (6).
2412 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
3. Estimate of card{n N: n < N}: Saddle point approximations
Our rst goal is to give a precise estimate for the number
card{n N: n < N},
that is, to prove Theorem 1. For this purpose, for every integral multi-index m = (m
1
, . . . , m
K
) we
consider the sets
V
m
(N) =
_
n < N: L
k
_
|n|
0
, . . . , |n|
q1
_
=m
k
, 1 k K
_
and their cardinalities card V
m
(N). With help of the generating function P(z
1
, . . . , z
K
, 1) we can ob-
tain these numbers by the use of K-fold Cauchy integration:
card V
m
(N) = card
_
n < N: L
k
_
|n|
0
, . . . , |n|
q1
_
=m
k
, 1 k K
_
=
1
(2i)
K
_

1

_

K
P(z
1
, . . . , z
K
, 1)
dz
1
z
m
1
+1
1

dz
K
z
m
K
+1
K
. (8)
Since P(z
1
, . . . , z
K
, 1) can be well approximated by a power F (z
1
, . . . , z
K
)
log
q
N
it is natural to do this
with help of a multivariate saddle point method.
We start with a preliminary lemma.
Lemma 2. Suppose that the system of equations
q1

=0
a
k
x

=
k
(1 k K), (9)
q1

=0
x

= 1 (10)
has a positive solution x
0
> 0, x
1
> 0, . . . , x
q1
> 0. Then there exist unique positive numbers z
1
>
0, . . . , z
K
> 0 such that
q1

=0
a
k
K

r=1
z
a
r
r
=
k
q1

=0
K

r=1
z
a
r
r
(1 k K). (11)
Proof. Let Z denote the set of solution (x
0
, . . . , x
q1
) of (9) and (10) with positive coordinates. By
assumption Z is not empty, in particular, it either consists of exactly one point (if q = K + 1) or it is
the intersection of a (q K 1)-dimensional hyperplane with the half spaces x
j
> 0, and, thus, can
be considered as an open set in a (q K 1)-dimensional space. Next consider the function
f (x
0
, . . . , x
q1
) =
q1

=0
x
j
log x
j
, (x
0
, . . . , x
q1
) Z.
Observe that f is a strictly concave positive function with unbounded derivative if one of the x
j
goes
to 0. Hence f attains its (only) maximum at some point (x

0
, . . . , x

q1
) Z.
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2413
Alternatively, this maximum can be calculated with help of Lagrange multipliers. Set

f (x
0
, . . . , x
q1
,
0
, . . . ,
K
) =
q1

=0
x
j
log x
j
+
0
_
q1

=0
x

1
_
+
K

k=1

k
_
q1

=0
a
k,
x

k
_
.
Then by Lagranges theorem there exist

0
, . . . ,

K
such that (x

0
, . . . , x

q1
) satises the system of
equations

f
x
j
= log x

j
+1 +

0
+
K

k=1

k
a
k, j
= 0 (0 j <q).
Hence, if we set z
k
= e

k
, we have
x

j
= e
1+

0
K

k=1
z
a
k, j
j
,
and since (9) and (10) imply that
q1

=0
a
k
x

=
k
q1

=0
x

,
it directly follows that (11) is satised for z
k
= e

k
. This is also the unique solution since every solu-
tion of (9) can be reinterpreted as a maximum of f on Z. 2
In what follows, we will denote by the (open) set of (
1
, . . . ,
K
) for which (9) has a unique
solution z
k
(
1
, . . . ,
K
) (1 k K) in the above sense. In fact, this is also a multivariate saddle point
as the proof of the following theorem shows.
Recall that we always assume that a
k,0
= 0, which implies that F (z
1
, . . . , z
K
) > 1 for all positive
real numbers z
1
, . . . , z
K
.
Theorem 3. Suppose that E is a compact subset of . Then uniformly for all integer vectors m =
(m
0
, . . . , m
q1
) Z
q
with
_
m
0
log
q
N
, . . . ,
m
q1
log
q
N
_
E
and as N we have
card V
m
(N) =
H( z
1
, . . . , z
K
, log
q
N)
(2 log
q
N)
K/2

1/2
F ( z
1
, . . . , z
K
)
log
q
N
z
m
1
1
z
m
K
K

_
1 + O
_
1
log N
__
(12)
2414 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
where
z
k
= z
k
_
m
0
log
q
N
, . . . ,
m
q1
log
q
N
_
(1 k K)
and

= det
_

2
log F ( z
1
e
t
1
, . . . , z
K
e
t
K
)
t
i
t
j

t
1
=0,...,t
K
=0
_
1i, jK
.
Proof. Our starting point is the representation (8), where we will use the circles of integration

k
=
_
z
k
: |z
k
| = z
k
_
(1 k K).
Due to the upper bound (5) we thus get an upper bound for those parts of the integral where
|arg(z
k
)| (for some k) of the form
C F ( z
1
, . . . , z
K
)
(1) log
q
N
.
Hence, these parts of the integral can be neglected.
For the remaining parts we use standard saddle point approximation on powers of functions
(see [8]). Note that ( z
1
, . . . , z
K
) is the saddle point of the function
(z
1
, . . . , z
K
) F (z
1
, . . . , z
K
)
log
q
N
z
m
1
1
z
m
K
K
= exp
_
log
q
N log
_
F (z
1
, . . . , z
K
)
_

k=1
m
k
log z
k
_
.
Hence, we directly obtain (12). 2
Remark 1. Theorem 3 has a slight extension. We also have
card V
m
(N) =
H( z
1
, . . . , z
K
, log
q
N)
(2 log
q
N)
K/2

1/2
F ( z
1
, . . . , z
K
)
log
q
N
z
m
1
1
z
m
K
K

_
1 + O
_
1
log N
__
(13)
where
z
k
= z
k
(
1
, . . . ,
K
) (1 k K)
and m
k

k
log
q
N = O(1). This means that we can vary m
k
a little bit without changing the saddle
points z
k
, that only depends on
1
, . . . ,
K
. This property will be frequently used in the sequel.
Remark 2. If we do not use the saddle point ( z
1
, . . . , z
K
) but any point (
1
, . . . ,
K
) of positive real
numbers we get an upper bound of the form
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2415
card V
m
(N)
H(
1
, . . . ,
K
, log
q
N)
(2 log
q
N)
K/2

1/2
F (
1
, . . . ,
K
)
log
q
N

m
1
1

m
K
K

_
1 + O
_
1
log N
__
, (14)
with
= det
_

2
log F (
1
e
t
1
, . . . ,
K
e
t
K
)
t
i
t
j

t
1
=0,...,t
K
=0
_
1i, jK
.
This follows from the fact that the absolute value of F (
1
e
it
1
, . . . ,
K
e
it
K
) can be estimated by

F
_

1
e
it
1
, . . . ,
K
e
it
K
_

F (
1
, . . . ,
K
) exp
_

1
2
K

i, j=1

i j
t
j
t
j
+ O
_
K

i=1
|t
i
|
3
__
, (15)
where

i j
=

2
log F (
1
e
t
1
, . . . ,
K
e
t
K
)
t
i
t
j

t
1
=0,...,t
K
=0
.
Of course, the constant implied by the term O(1/log N) depends (continuously) on
1
, . . . ,
K
.
The case
1
= =
K
= 0 is now easy to deal with. The corresponding asymptotic formula for
the numbers card{n N: n < N} is an immediate corollary of the above remark.
Corollary 3. Suppose that
1
= =
K
= 0 and let
1
, . . . ,
K
be given (xed) integers. Then (0, . . . , 0)
and we have
card{n N: n < N} = card
_
n < N: L
k
_
|n|
0
, . . . , |n|
q1
_
=
k
, 1 k K
_
=
H( z
1
, . . . , z
K
, log
q
N)
(2 log
q
N)
K/2

1/2
F ( z
1
, . . . , z
K
)
log
q
N
z

1
1
z

k
K
_
1 + O
_
1
log N
__
,
where z
k
= z
k
(0, . . . , 0) > 0 satises
q1

=0
a
k
K

r=1
z
a
r
r
= 0 (1 k K).
The next step is a little bit more involved. Suppose that there exists k with
k
> 0 and consider
the set
S =
_
k:
k
=0
_
q
(m
k
)/
k
: m Z, (m
k
)/
k
> 0
_
that is the union of geometric sequences.
2416 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
Let s
0
< s
1
< be an ordered version of the elements of S = {s
0
, s
1
, . . .}. Observe that if n > 0 is
an integer with s
j
n < s
j+1
then for all k with
k
> 0 there exists m
j,k
Z with
q
m
j,k

k
n <q
m
j,k
+1
k

k
.
For those k with
k
= 0 we set m
j,k
=
k
. If fact, this means that for all k {1, 2, . . . , K} and n
{s
j
, s
j
+1, . . . , s
j+1
1} we have
[
k
log
q
n] +
k
=m
j,k
.
Let m
j
= (m
j,1
, . . . , m
j,K
) denote the multi-index that collects these m
j,k
. More precisely this shows
that
card{n N: s
j
n < s
j+1
} = card V
m
j
(s
j+1
) card V
m
j
(s
j
).
Thus, we have proved the following lemma.
Lemma 3. Assume that there exists k with
k
> 0 and suppose that N is a positive integer with N = [s
J
] for
some s
J
S. Then we have
card{n N: n < N} =

j< J
_
card V
m
j
(s
j+1
) card V
m
j
(s
j
)
_
.
3.1. Proof of Theorem 1
Lemma 3 can be used to determine the asymptotic order of magnitude of the numbers card{n N:
n < N}. We will actually prove that there are two positive constants C
1
, C
2
with
C
1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N
card{n N: n < N}
C
2
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N
,
where
z
k
= z
k
(
1
, . . . ,
K
) (1 k K).
Thus, from Theorem 1 is explicitly given by
= log
q
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
.
Proof of Theorem 1. If
1
= =
K
= 0 then this estimate follows from Corollary 3.
If there is k with
k
> 0 then by Lemma 3 and (13) we get the upper bound:
card{n N: n < N}

j: s
j1
<N
card
_
V
m
j
(s
j+1
)
_

k:
k
>0

m
k
log
q
N+
k
card
_
V
([(m
k
)

/
k
]+

)
1K
_
q
(m
k
)/
k
__
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2417

k:
k
>0

m
k
log
q
N+
k
1
m
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
(m
k
)/
k

1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N
.
On the other hand we have
card{n N: n < N} card
_
V
m
j
(s
j+1
)
_
card
_
V
m
j
(s
j
)
_
+card
_
V
m
j+1
(s
j+2
)
_
card
_
V
m
j+1
(s
j+1
)
_
for every j with s
j+2
< N. Let k
0
be chosen such that
k
0
is largest. There clearly exists a constant
C such that s
j
= q
(m
k
0
)/
k
0
, s
j+1
= q
(m+1
k
0
)/
k
0
, and s
j
N/C. Further we have m
j,k
= [(m

k
0
)
k
/
k
0
] +
k
. Hence we can use the saddle point z
k
= z
k
(
1
, . . . ,
K
) (1 k K) and obtain
by (13):
card
_
V
m
j
(s
j+1
)
_
card
_
V
m
j
(s
j
)
_
=
H( z
1
, . . . , z
K
, log
q
s
j+1
)
(2 log
q
s
j+1
)
K/2

1/2
F ( z
1
, . . . , z
K
)
log
q
s
j+1
z
m
1
1
z
m
K
K
_
1 + O
_
1
log N
__

H( z
1
, . . . , z
K
, log
q
s
j
)
(2 log
q
s
j
)
K/2

1/2
F ( z
1
, . . . , z
K
)
log
q
s
j
z
m
1
1
z
m
K
K
_
1 + O
_
1
log N
__

1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N

_
H( z
1
, . . . , z
K
, log
q
s
j+1
) F ( z
1
, . . . , z
K
)
log
q
s
j+1
log
q
s
j
H( z
1
, . . . , z
K
, log
q
s
j
) + O
_
1
log N
__
.
Similarly we get
card
_
V
m
j+1
(s
j+2
)
_
card
_
V
m
j+1
(s
j+1
)
_

1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N

_
H( z
1
, . . . , z
K
, log
q
s
j+2
) F ( z
1
, . . . , z
K
)
log
q
s
j+2
log
q
s
j+1
H( z
1
, . . . , z
K
, log
q
s
j+1
) + O
_
1
log N
__
.
Since card(V
m
j
(s
j+1
)) card(V
m
j
(s
j
)) 0 and card(V
m
j1
(s
j+2
)) card(V
m
j+1
(s
j+1
)) 0, it follows
that
H( z
1
, . . . , z
K
, log
q
s
j+2
) F ( z
1
, . . . , z
K
)
log
q
s
j+2
log
q
s
j+1
H( z
1
, . . . , z
K
, log
q
s
j
)
C

log N
2418 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
for some constant C

> 0. Hence we get

card{n N: n < N}
1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N

_
_
F ( z
1
, . . . , z
K
)
1/
k
0
1
_
+ O
_
1
log N
__

1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N
.
This completes the proof of Theorem 1. 2
4. Uniform distribution modulo 1 of the sequence (n)
nN
As we remark at the end of Section 1 proving Theorem 2 is equivalent to proving that for any
irrational we have
W =

nN, n<N
e(n) =o
_
card{n N: n < N}
_
.
For this purpose we introduce the function
U(x
0
, . . . , x
q1
; t
0
, . . . , t
q1
) = x
0
e(t
0
) + + x
q1
e(t
q1
).
We will also use the shorthand notation U(x, t).
Lemma 4. Suppose that x
0
, . . . , x
q1
are positive real numbers. Then there exists a constant c > 0 that depends
continuously on x = (x
0
, . . . , x
q1
) such that for all real vectors t = (t
0
, . . . , t
q1
) and t
0
= (t
0,0
, . . . , t
0,q1
)

U(x, t)U(x, t +t
0
)

U(x, 0)
2
exp
_
c

0i<j<q
t
0,i
t
0, j

2
_
,
where x = min
kZ
|x k| denotes the distance to the nearest integer.
Proof. We rst consider |U(x, t)|
2
. By using the inequality |sin(t)|
2

t we obtain

U(x, t)

2
= U(x, t)U(x, t)
=
q1

j=0
x
2
j
+2

0i<j<q
x
i
x
j
cos
_
2(t
i
t
j
)
_
= U(x, 0)
2
4

0i<j<q
x
i
x
j
sin
_
(t
i
t
j
)
_
2
U(x, 0)
2

16

0i<j<q
x
i
x
j
t
i
t
j

2
U(x, 0)
2
exp
_
c
1

0i<j<q
t
i
t
j

2
_
,
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2419
where
c
1
=
16

2
min
0i<j<q
x
i
x
j
U(x, 0)
2
is a positive constant depending continuously on x. Since t
2
+t +t

1
2
t

2
for any real num-
bers t and t

it immediately follows that

U(x, t)U(x, t +t
0
)

U(x, 0)
2
exp
_

c
1
2

0i<j<q
_
t
i
t
j

2
+t
i
t
j
+t
0,i
t
0, j

2
_
_
U(x, 0)
2
exp
_

c
1
4

0i<j<q
t
0,i
t
0, j

2
_
.
This proves the lemma for c = c
1
/4. 2
Next we set
U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; s)
:= U
_
K

k=0
z
a
k,0
k
, . . . ,
K

k=0
z
a
k,q1
k
;
K

k=0
a
k,0
s
k
,
K

k=0
a
k,1
s
k
+ s, . . . ,
K

k=0
a
k,q1
s
k
+(q 1)s
_
,
that is, we substitute x

K
k=0
z
a
k,
k
and t

K
k=0
a
k,
s
k
+s (1 K).
Note that
U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; 0) = F
_
z
1
e(s
1
), . . . , z
k
e(s
K
)
_
. (16)
Furthermore, we have the following upper bound.
Lemma 5. Suppose that x
0
, . . . , x
q1
are positive real numbers and t
0
, . . . , t
q1
and are real numbers. Then
there exists a constant C
3
> 0 (that depends continuously on x
0
, . . . , x
q1
) with

S
N
_
x
0
e
it
1
, . . . , x
q1
e
it
q1
, e
i
_

C
3

log
q
N

j<
U
_
x
0
, . . . , x
q1
; t
0
, t
1
+q
j
, . . . , t
q1
+(q 1)q
j
_

.
Consequently, we have for positive real numbers z
1
, . . . , z
K
and real numbers s
1
, . . . , s
K
and

P
N
_
z
1
e
is
1
, . . . , z
K
e
is
K
, e
i
_

C
3

log
q
N

j<
U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

.
Proof. The estimate for S
N
follows immediately from the representations given in Lemma 1. The
upper bound for P
N
is just a rewritten version of the upper bound for S
N
. 2
2420 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
This estimate shows that if we are interested in upper bounds for
W
m
(N) =

nV
m
(N)
e(n)
=
1
_
0

1
_
0
P
N
_
z
1
e
is
1
, . . . , z
K
e
is
K
, e
i
__
z
1
e
is
1
_
m
1

_
z
K
e
is
K
_
m
K
ds
1
ds
K
,
then it is sucient to get proper upper bounds for integrals of the form
1
_
0

1
_
0

j<

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

ds
1
ds
K
. (17)
Following this idea we prove upper bounds for (17) in Lemmas 6 and 7 which will lead to upper
bounds for W
m
(N) in Lemma 8.
We have the following estimates.
Lemma 6. Suppose that z
1
, . . . , z
K
are positive real numbers. Then there exists a constant C
4
> 0 (that de-
pends continuously on z
1
, . . . , z
K
) such that for all integers 1 and all real numbers
1
_
0

1
_
0

U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; )

ds
1
ds
K

C
4

K/2
F (z
1
, . . . , z
K
)

.
Proof. Observe that Lemma 4 also implies that

U(x, t)

U(x, 0) exp
_
c

0i<j<q
t
i
t
j

2
_
.
Hence, we also get

U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; )

U(z
1
, . . . , z
K
; 0, . . . , 0; 0) exp
_
c

0i<j<q
_
_
_
_
_
K

k=1
(a
k,i
a
k, j
)s
k
+(i j)
_
_
_
_
_
2
_
.
By the linear independence assumption on the forms L
k
and by Denition 1(ii) there exist j
0
<
j
1
< < j
K
such that the matrix C = (a
k, j

a
k, j
0
)
1k,K
is regular and, thus, has determinant
det C =d = 0. Further, there exist
k
with
K

k=1
(a
k, j

a
k, j
0
)
k
= ( j

j
0
) (1 K).
Hence, there exist integers d
j
with
d(s
j
+
j
) =
K

=1
d
j
_
K

k=1
(a
k, j

a
k, j
0
)(s
k
+
k
)
_
.
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2421
Hence, for all j we have
_
_
d(s
j
+
j
)
_
_
2

=1
_
_
_
_
_
K

k=1
(a
k, j

a
k, j
0
)(s
k
+
k
)
_
_
_
_
_
2

0i<j<q
_
_
_
_
_
K

k=1
(a
k,i
a
k, j
)s
k
+(i j)
_
_
_
_
_
2
.
Consequently, there exists a constant c

> 0 with

U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; )

U(z
1
, . . . , z
K
; 0, . . . , 0; 0) exp
_
c

k=1
_
_
d(s
k
+
k
)
_
_
2
_
F (z
1
, . . . , z
k
) exp
_
c

k=1
_
_
d(s
k
+
k
)
_
_
2
_
,
so that
1
_
0

1
_
0

U(z
1
, . . . , z
K
; s
1
, . . . , s
K
; )

ds
1
ds
K
F (z
1
, . . . , z
k
)

1
_
0

1
_
0
exp
_
c

k=1
_
_
d(s
k
+
k
)
_
_
2
_
ds
1
ds
K
F (z
1
, . . . , z
k
)

_
1
_
0
e
c

ds
2
ds
_
K

C
4

K/2
F (z
1
, . . . , z
k
)

.
This completes the proof of the lemma. 2
Remark 3. Alternatively we can prove Lemma 6 by using property (16) and previous estimates for F .
Namely, by using (15) for |s
j
| (where > 0 is chosen suciently small) and the property that

F
_
z
1
e(s
1
), . . . , z
k
e(s
K
)
_

F (z
1
, . . . , z
K
)
1
for some > 0 if there is some j with |s
j
| (compare with (5)) the upper bound follows.
The next lemma is crucial for proving upper bound on Weyl sums.
Lemma 7. Suppose that z
1
, . . . , z
K
are positive real numbers. Then there exists a constant C
5
> 0 (that depends
continuously on z
1
, . . . , z
K
) such that for all integers 1 and each real number
2422 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
1
_
0

1
_
0

j<

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

ds
1
ds
K

C
5

K/2
F (z
1
, . . . , z
K
)

exp
_

c
4

j<
_
_
(q 1)q
j
_
_
2
_
.
Proof. For simplicity we assume that is a multiple of 4. The other cases can be handled in the
same way.
We split the product of the integrand into two parts:

j<, j0mod4

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_
U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j+1
_

j<, j2,3mod4

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

.
By applying Lemma 4 with t
0,
= (q 1)q
j
we get

j<, j0mod4

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_
U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j+1
_

F (z
1
, . . . , z
K
)
/2
exp
_
c

j<, j0mod4

0<i<jq
_
_
(i j)(q 1)q
j

_
_
2
_
F (z
1
, . . . , z
K
)
/2
exp
_
c

j<, j0mod4
_
_
(q 1)q
j

_
_
2
_
.
Furthermore, by applying the inequality
|v
1
v
m
|
|v
1
|
m
+ +|v
1
|
m
m
,
we obtain (by applying Lemma 6):
1
_
0

1
_
0

j<, j2,3mod4

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

ds
1
ds
K

1
/2

j<, j2,3mod4
1
_
0

1
_
0

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

/2
ds
1
ds
K

K/2
F (z
1
, . . . , z
K
)
/2
.
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2423
Hence,
1
_
0

1
_
0

j<

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

ds
1
ds
K

C
5

K/2
F (z
1
, . . . , z
K
)

exp
_
c

j<, j0mod4
_
_
(q 1)q
j
_
_
2
_
.
Similarly we can deal with the other residue classes modulo 4. Combining these four estimates nally
proves the lemma. 2
We set
E(, ) =
1
4

j<
_
_
(q 1)q
j
_
_
2
.
For irrational it is clear that E(, ) as (compare with [13]).
Next we prove an upper bound for
W
m
(N) =

nV
m
(N)
e(n).
Lemma 8. Suppose that m
k
= [
k
log
q
N] +
k
. Then, there exist constants C
6
> 0 and C
7
> 0 (that depend
on
1
, . . . ,
K
and on
1
, . . . ,
K
) such that as N

W
m
(N)

C
6
card
_
V
m
(N)
_

_
e
c E(,[log
q
N]/2)
+e
C
7
log
q
N
_
.
Proof. We rst x positive numbers z
1
, . . . , z
K
. Recall that F (z
1
, . . . , z
K
) > 1. Hence, by Lemmas 5
and 7 we have

W
m
(N)

nV
m
(N)
e(n)

1
_
0

1
_
0
P
N
_
z
1
e
is
1
, . . . , z
K
e
is
K
, e
i
_ _
z
1
e
is
1
_
m
1

_
z
K
e
is
K
_
m
K
ds
1
ds
K

1
_
0

1
_
0

P
N
_
z
1
e
is
1
, . . . , z
K
e
is
K
, e
i
_

z
m
1
1
z
m
K
K
ds
1
ds
K

log
q
N
1
_
0

1
_
0

j<

U
_
z
1
, . . . , z
K
; s
1
, . . . , s
K
; q
j
_

ds
1
ds
K

log
q
N
1

K/2
F (z
1
, . . . , z
K
)

z
m
1
1
z
m
K
K
exp
_
cE
_
(q 1),
__

1
(log N)
K/2
F (z
1
, . . . , z
K
)
log
q
N
z
m
1
1
z
m
K
K
exp
_
cE
_
(q 1), [log
q
N]/2
__
2424 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
+
1
(log N)
K/2
F (z
1
, . . . , z
K
)
1
2
log
q
N
z
m
1
1
z
m
K
K

1
(log N)
K/2
F (z
1
, . . . , z
K
)
log
q
N
z
m
1
1
z
m
K
K
_
e
c E(,[log
q
N]/2)
+e
C
7
log
q
N
_
,
where C
7
> 0 depends on z
1
, . . . , z
K
. Now, if we choose z
k
= z
k
= z
k
(
1
, . . . ,
K
) we, thus, obtain the
proposed estimate. 2
4.1. Proof of Theorem 2
We are now ready to prove the nal step of Theorem 2, i.e., that for all irrational numbers we
have, as N ,

nN, nN
e(n) =o
_
card{n N: n N}
_
.
Proof of Theorem 2. First assume that
1
= =
K
= 0. Here we have with m= (
1
, . . . ,
K
)
W =

nN, n<N
e(n) = W
m
(N)
and we can directly apply Lemma 8.
Now suppose that there exists k with
k
> 0. With the same reasoning as in Lemma 3 we have (if
N = [s
J
] for some s
J
S)
W =

nN, n<N
e(n)
=

j J
_
W
m
j
(s
j+1
) W
m
j
(s
j
)
_
and consequently
|W|

j< J
_

W
m
j
(s
j+1
)

W
m
j
(s
j
)

_
.
Now, with help of Lemma 8 we get the upper bound

j< J

W
m
j
(s
j+1
)

k:
k
>0

m
k
log
q
N+
k

W
([(m
k
)

/
k
]+

)
1K
_
q
(m
k
)/
k
_

k:
k
>0

m
k
log
q
N+
k
1
m
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
(m
k
)/
k

_
e
cE(,(m
k
)/(2
k
))
+e
C
7
m/
k
_
= o
_
1
(log
q
N)
K/2
_
F ( z
1
, . . . , z
K
)
z

1
1
z

K
K
_
log
q
N
_
= o
_
card{n N: n N}
_
.
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2425
Similarly we can estimate the second sum

j< J
W
m
j
(s
j
). This proves the lemma if N = [s
J
] for some
s
J
S. If N is not of that form we just have to add
W
m
J
(N) W
m
J
(s
J
)
which can be handled with help of Lemma 8. 2
5. Generalizations
5.1. Missing digits
A rst generalization of Theorem 2 is to assume that some digits D {0, 1. . . , q 1} do not
appear, that is, we additionally assume that
|n|

= 0 for D (18)
(compare also with [19]). Formally this condition could be included into (1) without any change
of notation. However, then there is no positive solution of the corresponding system of equations
L
k
(x
0
, . . . , x
q1
) =
k
(1 k K) since (18) forces x

= 0 for all D. Nevertheless, we can work

with in the missing-digit-case almost in the same way as above.
First, it is clear that the generating function
S
D
N
_
(x
j
)
j / D
, y
_
=

n<N

j / D
x
|n|
j
j
y
n
is just obtained by using S
N
(x
1
, . . . , x
N
, y) and setting x

= 0 for D. In particular, we directly use

Lemma 1 and, hence, all subsequent considerations directly transfer.
After all we get precisely the same as Theorem 2. The only difference is that we have to consider
linear forms in the remaining variables x
j
, j / D. More precisely we have to assume that the system
L
D
= (L
D
k
)
k=1,...,K
L
D
k
=

/ D
a
k,
x

is complete (compare with Denition 1) and that the system

/ D
a
k,
x

=
k
, k = 1, . . . , K,

/ D
x

= 1
has a positive solution. Then the sequence (n)
nN
is uniformly distributed modulo 1, where N is
the set of positive integers with |n|
j
= 0 for j D and

/ D
a
k,
|n|

= [
k
log
q
n] +
k
, k = 1, . . . , K.
2426 M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427
5.2. Non-integer coecients
The restriction that the coecients a
k,
of the linear forms L
k
are integers was natural in the
context of Theorem 2. Nevertheless we can also consider general linear forms
L
k
(x
0
, . . . , x
q1
) =
q1

=0
a
k,
x

(1 k K)
and x intervals I
1
= [a
1
, b
1
], . . . , I
K
= [a
K
, b
K
] contained in the positive real line. A corresponding
set N can be then dened by the set of non-negative integers n with
L
k
_
|n|
0
, . . . , |n|
q1
_

k
log
q
n I
k
(1 k K),
where
1
, . . . ,
K
are given real numbers.
Instead of Cauchys formula we can then use the inverse Laplace transform. For example, if we set
(similarly to the above)
V
m
(N) =
_
n < N: L
k
_
|n|
0
, . . . , |n|
q1
_
m
k
I
k
_
,
where m = (m
1
, . . . , m
K
) is any vector of real numbers, then we have for all real numbers
s
0,1
, . . . , s
0,K

nV
m
(N)
y
n
=
1
(2i)
K
lim
T
1

s
0,1
i T
1
_
s
0,1
i T
1
lim
T
K

s
0,K
i T
K
_
s
0,K
i T
K
P
N
_
e
s
1
, . . . , e
s
K
, y
_
e
m
1
s
1
m
K
s
K
K

k=1
e
a
k
s
k
e
b
k
s
k
s
k
ds
1
ds
K
.
In particular, we can use s
0,k
= log z
k
, where z
k
= z
k
(m
1
/log
q
N, . . . , m
K
/log
q
N) are the saddle points
from above. Then these integrals can be asymptotically evaluated by a usual saddle point approxima-
tion, in particular if y = 1 and also if y = e().
Of course, there are some technical diculties that might occur. First note that the above integrals
are not absolutely convergent. This is due to the factor 1/s
k
of the Laplace transform
b
k
_
a
k
e
s
k
x
dx =
e
a
k
s
k
e
b
k
s
k
s
k
.
As usual, this can be handled by smoothing the characteristic functions of the intervals I
k
= [a
k
, b
k
].
Second, if there are rational relations between the coecients of the linear forms L
k
then we have
to deal with innitely many saddle points on the lines (s
k
) = s
0,k
. For example, if the coecients
a
k,
and m
k
are integers then
P
N
_
e
s
1
+2r
1
i
, . . . , e
s
K
+2r
K
i
, y
_
e
m
1
(s
1
+2r
1
i)m
K
(s
K
+2r
K
i)
= P
N
_
e
s
1
, . . . , e
s
K
, y
_
e
m
1
s
1
m
K
s
K
for all integers r
1
, . . . , r
K
. However, it is possible to deal with all these kinds of problems and prove
that the sequence (n)
nN
is uniformly distributed modulo 1.
M. Drmota, C. Mauduit / Journal of Number Theory 130 (2010) 24042427 2427
5.3. An open problem
In [9], the authors give an asymptotic equivalent for the number of prime numbers p, p < N, such
that the sum of digits of p is close to the expected value. Their method does not apply to the case
where the sum of digits (of any other linear combination of digits) of the prime numbers p is far
from the expected value, so the following problem is open:
Problem. Give estimates for card{p N: p < N, p prime}.
References
[1] J.-P. Allouche, P. Liardet, Generalized RudinShapiro sequences, Acta Arith. 60 (1) (1991) 127.
[2] J.-P. Allouche, M. Mends France, Suite de RudinShapiro et modle dIsing, Bull. Soc. Math. France 113 (3) (1985) 273283.
[3] J.-P. Allouche, M. Mends France, On an extremal property of the RudinShapiro sequence, Mathematika 32 (1) (1985)
3338.
[4] W.D. Banks, A. Conitti, I.E. Shparlinski, Character sums over integers with restricted g-ary digits, Illinois J. Math. 46 (3)
(2002) 819836.
[5] J. Coquet, On the uniform distribution modulo one of some subsequences of polynomial sequences, J. Number Theory 10 (3)
(1978) 291296.
[6] J. Coquet, On the uniform distribution modulo one of subsequences of polynomial sequences. II, J. Number Theory 12 (2)
(1980) 244250.
[7] J. Coquet, Graphes connexes, reprsentation des entiers et quirpartition, J. Number Theory 16 (3) (1983) 363375.
[8] M. Drmota, A bivariate asymptotic expansion of coecients of powers of generating functions, European J. Combin. 15
(1994) 139152.
[9] M. Drmota, C. Mauduit, J. Rivat, Primes with an average sum of digits, Compos. Math. 145 (2009) 271292.
[10] M. Drmota, J. Rivat, The sum-of-digits function of squares, J. Lond. Math. Soc. (2) 72 (2005) 273292.
[11] M. Drmota, R.F. Tichy, Sequences, Discrepancies, and Applications, Lecture Notes in Math., vol. 1651, Springer, Berlin, 1997.
[12] P. Erd os, C. Mauduit, A. Srkzy, On arithmetic properties of integers with missing digits. I. Distribution in residue classes,
J. Number Theory 70 (2) (1998) 99120.
[13] . Fouvry, C. Mauduit, Sur les entiers dont la somme des chiffres est moyenne, J. Number Theory 114 (2005) 135152.
[14] J.B. Friedlander, I.E. Shparlinski, On the distribution of DieHellman triples with sparse exponents, SIAM J. Discrete
Math. 14 (2) (2001) 162169.
[15] A.O. Gelfond, Sur les nombres qui ont des proprits additives et multiplicatives donnes, Acta Arith. 13 (1967/1968) 259
265.
[16] S. Konyagin, Arithmetic properties of integers with missing digits: distribution in residue classes, Period. Math. Hun-
gar. 42 (12) (2001) 145162.
[17] L. Kuipers, H. Niederreiter, Uniform Distribution of Sequences, John Wiley & Sons, New York, 1974.
[18] C. Mauduit, Automates nis et ensembles normaux, Ann. Inst. Fourier (Grenoble) 36 (2) (1986) 125.
[19] C. Mauduit, Proprits arithmtiques des substitutions et automates innis, Ann. Inst. Fourier 56 (2006) 25252549.
[20] C. Mauduit, C. Pomerance, A. Srkzy, On the distribution in residue classes of integers with a xed sum of digits, Ra-
manujan J. 9 (12) (2005) 4562.
[21] C. Mauduit, A. Srkzy, On the arithmetic structure of sets characterized by sum of digits properties, J. Number The-
ory 61 (1) (1996) 2538.
[22] C. Mauduit, A. Srkzy, On the arithmetic structure of the integers whose sum of digits is xed, Acta Arith. 81 (2) (1997)
145173.
[23] M. Quefflec, Substitution Dynamical SystemsSpectral Analysis, Lecture Notes in Math., vol. 1294, Springer-Verlag, Berlin,
1987.