A DIOPHANTINE PROBLEM WITH PRIME VARIABLES

ALESSANDRO LANGUASCO & ALESSANDRO ZACCAGNINI

arXiv:1206.0252v1 [math.NT] 1 Jun 2012

Abstract. We study the distribution of the values of the form 1 p1 + 2 p2 + 3 pk3 ,

where 1 , 2 and 3 are non-zero real number not all of the same sign, with 1 /2
irrational, and p1 , p2 and p3 are prime numbers. We prove that, when 1 < k < 4/3,
these value approximate rather closely any prescribed real number.

Dedicated to Prof. R. Balasubramanian on the occasion of his 60th birthday

1. Introduction
The problem that we want to study in this paper can be stated in general as follows: given r non-zero real numbers 1 , . . . , r , and positive real numbers k1 , . . . , kr ,
approximate a given real number by means of values of the form
1 pk11 + + r pkr r ,

(1)

where p1 , . . . , pr denote primes. If = 1/k1 + 1/k2 + + 1/kr is small the goal is to

show that

 k
1 p11 + 2 pk22 + + r pkr r  <
(2)
has infinitely many solutions for every fixed > 0. If is large one expects to be able
to prove the stronger result that, in fact, some 0 is admissible in (2): more precisely,
it should be possible to take as a small negative power of maxj pj . The number of
variables r also plays a role, of course. Some hypothesis on the irrationality of at least
one ratio i /j is necessary, and also on signs, if one wants to approximate to all real
numbers and not only some proper subset. We will make everything precise in due course.
Many such results are known, with various types of assumptions and conclusions, and
we now give a brief description of a few among them. Vaughan [15] has r = 3 and kj = 1
for all j, the non-zero coefficients j not all of the same sign with 1 /2 irrational. In
this case is essentially (maxj pj )1/10 . The paper [16] contains more elaborate results of
the same kind, with the same integral exponent k 1 for all primes. Baker and Harman
[1] and Harman [7] have a result similar to Vaughans [15] with = (maxj pj )1/6 and
= (maxj pj )1/5 respectively. This has been recently improved to = (maxj pj )2/9 by
Matomaki [12].
The papers by Br
udern, Cook and Perelli [2], Cook [3], Cook and Fox [4], Harman [8],
Cook and Harman [5] all deal with the number of exceptional real numbers that
can not be well approximated by values of type (1), but in this case there are many
differences with the results quoted above. First, does not depend on the primes pj
but rather on (it is a small negative power of , in fact), but, in their setting, this is
essentially equivalent to the alternative statement as we shall see presently. It is more
important, of course, to define carefully what exceptional means. Actually, the results
apply to suitable sequences of positive real numbers n with limit +, and it is shown
that the number of exceptional elements in the sequence, that is, elements that can not
be approximated within the prescribed precision, is small in a strong quantitative sense.
2010 Mathematics Subject Classification. Primary: 11P55.
Key words and phrases. Diophantine problems with prime variables.
1

The assumption is that the coefficients j are all positive, which is not a restriction in
this case, that the ratio 1 /2 is irrational and algebraic, and that kj is the same positive
integer k for all j. The assumption on 1 /2 is needed to deal with some exponential
sums on the so-called minor arc.
Tolev [14] has r = 3, the coefficients j all equal to 1 and all the exponents kj equal
to a constant k (1, 15/14). The conclusion is that all sufficiently large real numbers
can be approximated, with a negative power of .
Parsell [13] considers two primes and a large number of powers of 2, so that in a
sense = 2 + , but r is large and is arbitrary but fixed. This has been improved
in [?LanguascoZaccagnini2010c] by the present authors, who showed that a smaller
number of powers of 2 is needed. In a similar vein, Languasco and Settimi [9] have the
corresponding result with one prime, two squares of primes and a large number of powers
of 2. Finally, the present authors [10] have a result with one prime and three squares of
primes, so that = 5/2, r = 4 and = (maxj pj )1/18+ , while in [11] they deal with
one prime, the square of a prime and the k-th power of a prime with k (1, 33/29) and
= (maxj pj )(3329k)/(72k)+ . In all of these papers, it is assumed that one, carefully
chosen, among the ratios i /j is irrational.
Our main result is the following Theorem.
Theorem 1. Let 1 < k < 4/3 be a real number and assume that 1 , 2 , and 3 are
non-zero real numbers, not all of the same sign and that 1 /2 is irrational. Let be
any real number. For any > 0 the inequality




1 p1 + 2 p2 + 3 pk3  max pj 3/102/(5k)+
(3)
j

has infinitely many solutions in prime variables p1 , p2 and p3 .

In the notation above, we have r = 3, = 2+1/k and = (maxj pj )3/102/(5k)+ . We use

the variant of the circle method introduced by Davenport and Heilbronn [6] to deal with
these problems, where the variables are not necessarily integral. The following lemmas
are the two key ingredients of the proof. They relate a suitable L2 -average of the error
on the major arc to a generalized version of the so-called Selberg integral, which is a
well-known and widely used tool in this context: see [2], [?LanguascoZaccagnini2010c],
[9], [10]. The same argument, with comparatively minor changes, can be used with
1 p1 + 2 p2 + 3 p23 (with the same hypothesis on the ratio 1 /2 and on signs as above),
and = (maxj pj )1/18+ .
Before the statement, we need to define the relevant quantities, beginning with the
exponential sums. As usual, we write e() = e2i . For any real k 1 we let
X
X
Sk () =
log p e(pk )
and
Uk () =
e(nk )
(4)
Xpk X

Xnk X

where is a small, fixed positive constant, which may depend on the coefficients j . Then
we set
Z 2X 
2
1/k
1/k
1/k
1/k
Jk (X, h) =
((x + h) ) (x ) ((x + h) x ) dx.
(5)
X

This is the generalized version of the Selberg integral referred to above: the classical
function is J1 (X, h).
Lemma 1. Let k 1 be a real number. For 0 < Y 1/2 we have
Z Y

X 2/k2 log2 X
1 
|Sk () Uk ()|2 d k
,
+ Y 2 X + Y 2 Jk X,
Y
2Y
Y
2

where Jk (X, h) is defined in (5).

This is Theorem 1 of [11].
Lemma 2. Let k 1 be a real number and be an arbitrarily small positive constant.
There exists a positive constant c1 = c1 (), which does not depend on k, such that

 log X 1/3 
Jk (X, h) k h2 X 2/k1 exp c1
log log X
uniformly for X 15/(6k)+ h X.
This is the special case C = 12/5 of Theorem 2 of [11].
2. Proof of Theorem 1
In order to prove that (3) has infinitely many solutions, it is sufficient to show the
existence of an increasing sequence Xn with limit + such that (3) has at least a solution
with maxj pj [Xn , Xn ]. This sequence actually depends on rational approximations
for 1 /2 : more precisely, we recall that there are infinitely many pairs of integers a and
q such that (a, q) = 1, q > 0 and


1
 1 a
  2.
2 q
q

We take the sequence X = q 5k/(k+2) (dropping the useless suffix n) and then, as customary,
define all of the circle-method parameters in terms of X. We may obviously assume that
q is sufficiently large. The choice of the exponent 5k/(k + 2) is justified in the discussion
following the proof of Lemma 4. As usual, we approximate to Sk using the function
Z X 1/k
Tk () =
e(tk ) dt
(X)1/k

and notice the simple inequality

Tk () k, X 1/k1 min X, ||1 .

(6)

Since the variables are not integers, we cannot count exact hits as in the standard
applications of the circle method, only near misses, so that we need some measure of
proximity. For > 0, we detect solutions of (3) by means of the function
b () = max(0, ||),
K

which, as the notation suggests, is the Fourier transform of

 sin() 2
K () =

for 6= 0, and, by continuity, K (0) = 2 . This relation transforms the problem of

counting solutions of the inequality (3) into estimating suitable integrals. We recall the
trivial, but crucial, property


2
2
K () min , ||
.
(7)
When X is an interval, a half line, or the union of two such sets we let
Z
I(, , X) =
S1 (1 )S1 (2 )Sk (3 )K ()e() d.
X

The starting point of the method is the observation that

Z
X


I(, , R) =
log p1 log p2 log p3 K ()e (1 p1 + 2 p2 + 3 pk3 ) d
R

Xp1 ,p2 X
Xpk3 X

Xp1 ,p2 X
Xpk3 X

log p1 log p2 log p3 max(0, |1 p1 + 2 p2 + 3 pk3 |)

(log X)3 N (X),

where N (X) denotes the number of solutions of the inequality (3) with p1 , p2 [X, X]
and pk3 [X, X]. In other words, I(, , R) provides a lower bound for the quantity
that we are interested in.
We now give the definitions that we need to set up the method. More definitions
will be given at appropriate places later. We let P = P (X) = X 5/(6k) , = (X) =
X 3/102/(5k)+ , and R = R(X) = 2 (log X)3/2 . The choice for P is justified at the end
of 6, the one for at the end of 7 and the one for R at the end of 8. See also 9 for a
fuller discussion. We now decompose R as M m t where
h P Pi

P  P 
M= ,
,
m = R,

,R ,
t = R \ (M m),
X X
X
X
so that
I(, , R) = I(, , M) + I(, , m) + I(, , t).
The sets M, m and t are called the major arc, the intermediate (or minor) arc and the
trivial arc respectively. In 3 we prove that the major arc yields the main term for
I(, , R). We show in 7 that the contribution of the intermediate arc does not cancel
the main term, exploiting the hypothesis that 1 /2 is irrational to prove that |S1 (1 )|
and |S1 (2 )| can not both be large for m: see Lemma 4 for the details. The trivial
arc, treated in 8, only gives a rather small contribution.
From now on, implicit constants may depend on the coefficients j , on , k and .
3. The major arc
We write
I(, , M) =

S1 (1 )S1 (2 )Sk (3 )K ()e() d

T1 (1 )T1 (2 )Tk (3 )K ()e() d

Z


+
S1 (1 ) T1 (1 ) T1 (2 )Tk (3 )K ()e() d
ZM


+
S1 (1 ) S1 (2 ) T1 (2 ) Tk (3 )K ()e() d
ZM


+
S1 (1 )S1 (2 ) Sk (3 ) Tk (3 ) K ()e() d

= J1 + J2 + J3 + J4 ,

say. We will give a lower bound for J1 and upper bounds for J2 , J3 and J4 . For brevity,
since the computations for J3 are similar to, but simpler than, the corresponding ones
for J2 and J4 , we will skip them.
4

4. Lower bound for J1

The lower bound J1 2 X 1+1/k is proved in a classical way. We have
Z
J1 =
T1 (1 )T1 (2 )Tk (3 )K ()e() d
M
Z
=
T1 (1 )T1 (2 )Tk (3 )K ()e() d
R
Z +

+O
|T1 (1 )T1 (2 )Tk (3 )|K () d .
P/X

Using inequalities (6) and (7), we see that the error term is
Z +


d
2 1/k1
2 1+1/k 2
2 1+1/k
X

X
P
=
o

X
.
3
P/X
For brevity, we set D = [X, X]2 [(X)1/k , X 1/k ] and rewrite the main term in the form
Z
Z Z

e (1 t1 + 2 t2 + 3 tk3 ) K () d dt1 dt2 dt3

Z DZ R
= max(0, |1 t1 + 2 t2 + 3 tk3 |) dt1 dt2 dt3 .
D

We now proceed to show that the last integral is 2 X 1+1/k . Apart from trivial permutations or changes of sign, there are essentially two cases:
(1) 1 > 0, 2 < 0, 3 < 0;
(2) 1 > 0, 2 > 0, 3 < 0.
We briefly deal with the second case, the other one being similar. A suitable change
of variables shows that
Z
Z
du1 du2 du3
J1
max(0, |1 u1 + 2 u2 + 3 u3 |)
11/k
u3
D
Z
Z
max(0, |1 u1 + 2 u2 + 3 u3 |) du1 du2 du3 ,
X 1/k1
D

where D = [X, (1 )X]3 , for large X. For j = 1, 2, let aj = 2|3|/|j |, bj = 3aj /2

and Ij = [aj X, bj X]. Notice that if uj Ij for j = 1, 2, then


1 u1 + 2 u2 4|3 |X, 6|3|X

so that, for every such choice of (u1 , u2 ), the interval [a, b] with endpoints /|3 | +
(1 u1 + 2 u2 )/|3 | is contained in [X, (1 )X]. In other words, for u3 [a, b] the values
of 1 u1 + 2 u2 + 3 u3 cover the whole interval [, ]. Hence, for any (u1 , u2 ) I1 I2
we have
Z (1)X
Z
1
max(0, |1 u1 + 2 u2 + 3 u3 |) du3 = |3 |
max(0, |u|) du 2 .
X

Finally,
2

J1 X

1/k1

ZZ

I1 I2

which is the required lower bound.

5

du1 du2 2 X 1+1/k ,

5. Bound for J2
We recall definition (4) and notice that the Euler summation formula implies that
Tk () Uk () 1 + ||X
Using (7) we see that
J2

for any k 1.

(8)



S1 (1 ) T1 (1 ) |T1 (2 )| |Tk (3 )| d

ZM


S1 (1 ) U1 (1 ) |T1 (2 )| |Tk (3 )| d
2
M
Z


2
U1 (1 ) T1 (1 ) |T1 (2 )| |Tk (3 )| d
+
M

= (A2 + B2 ),

say. In order to estimate A2 we use Lemmas 1 and 2. By the Cauchy inequality and (6)
above, for any fixed A > 0 we have
Z
Z P/X 
1/2
2 1/2  P/X
2
2


A2
S1 (1 ) U1 (1 ) d
|T1 (2 )| |Tk (3 )| d
P/X

(log X)A

1/2 Z

1/X

2+2/k

d +

P/X

P/X

1/X

X 1+1/k
X 2/k2 1/2
d

A
4
(log X)A/2

by Lemma 2, which we can use
provided that X/P X 1/6+ , that is, P X 5/6 . This
proves that 2 A2 = o 2 X 1+1/k . Furthermore, using inequalities (6) and (8) we see that
Z 1/X
Z P/X
B2
|T1 (2 )| |Tk (3 )| d + X
|T1 (2 )| |Tk (3 )| d
0

1 1+1/k
X
+ X 1/k
X


so that 2 B2 = o 2 X 1+1/k .

1/X

P/X

1/X

d
X 1/k log P,

6. Bound for J4
Inequality (7) implies that
Z




2
S1 (1 ) S1 (2 ) Sk (3 ) Tk (3 ) d
J4
ZM




S1 (1 ) S1 (2 ) Sk (3 ) Uk (3 ) d
2
M
Z




2
S1 (1 ) S1 (2 ) Uk (3 ) Tk (3 ) d
+
M

= (A4 + B4 ),

say. The Parseval inequality and trivial bounds yield, for any fixed A > 0,
Z
1/2
Z 

2 1/2  
Sk (3 ) Uk (3 )2 d


S1 (1 ) d
A4 X
M

P  X 1/2
X(X log X)1/2 Jk X,
A X 1+1/k (log X)1/2A/2
X
P
6

by Lemmas 1 and 2 which we can use provided
that X/P X 15/(6k)+ , that is, P
X 5/(6k) . This proves that 2 A4 = o 2 X 1+1/k . Furthermore, using (8), the Cauchy
inequality and trivial bounds we see that
Z 1/X
Z P/X










B4
S1 (1 ) S1 (2 ) d + X
S1 (1 ) S1 (2 ) d
0

X +P

Z

P/X 

1/X


S1 (1 )2 d

1/X

1/2

S1 (2 )2 d
P X log X.

P/X 

1/X

Hence B4 P X log X, so that taking P = o X 1/k (log X)1 we get 2 B4 = o 2 X 1+1/k .
We may therefore choose
P = X 5/(6k) .
(9)
7. The intermediate arc
We need to show that |S1 (1 )| and |S1 (2 )| can not both be large for m,
exploiting the fact that 1 /2 is irrational. We achieve this using a famous result by
Vaughan about S1 ().
Lemma 3 (Vaughan [17], Theorem 3.1). Let be a real number and a, q be positive
integers satisfying (a, q) = 1 and | a/q| < q 2 . Then

X
p
S1 () + Xq + X 4/5 log4 X.
q
Lemma 4. Let 1 k < 4/3. Assume that 1 /2 is irrational and let X = q 5k/(k+2) ,
where q is the denominator of
a convergent of the continued fraction for 1 /2 . Let
V () = min |S1 (1 )|, |S1(2 )| . Then we have
sup V () X 4/5+1/(10k) log4 X.
m

Proof. Let m and Q = X 2/51/(5k) P . By Dirichlets Theorem, there exist integers

ai , qi with 1 qi X/Q and (ai , qi ) = 1, such that |i qi ai | Q/X, for i = 1, 2. We
remark that a1 a2 6= 0 otherwise we would have M. Now suppose that qi Q for
i = 1, 2. In this case we get
1
a2
a1
a2 q1 a1 q2 = (1 q1 a1 )
(2 q2 a2 )
2
2
2
and hence
  2



 



1
a2 q1 a1 q2  2 1 +  1  Q < 1
(10)
 2  X


2
2q
for sufficiently large X. Then, from the law of best approximation and the definition of
m, we obtain
X (k+2)/(5k) = q |a2 q1 | q1 q2 R Q2 R X (k+2)/(5k) ,

(11)

which is absurd. Hence either q1 > Q or q2 > Q. Assume that q1 > Q. Using Lemma 3
on S1 (1 ), we have


p
X
4/5
V () |S1 (1 )| sup
log4 X
+ Xq1 + X
q1
Q<q1 X/Q
X 4/5+1/(10k) (log X)4 .

The other case is totally similar and hence Lemma 4 follows.

7

Lemma 5. For j = 1 and 2 we have

Z
|S1 (j )|2 K () d X log X and

|Sk (3 )|2K () d X 1/k (log X)3 .

Proof. We have to split the range [P/X, R] into two intervals in order to use (7) efficiently.
In the first case we have
Z 1/
Z
Z R
d
2
2
2
|S1 (j )| K () d
|S1 (j )| d +
|S1 (j )|2 2

m
P/X
1/
by (7), for j = 1, 2. By periodicity
Z 1/
Z
2
2

|S1 (j )| d
P/X

|S1 ()|2 d X log X,

by the Prime Number Theorem (PNT). We also have

Z n
Z +
Z R
X
1
2 d
2 d
|S1 ()|2 d X log X,
|S1 ()| 2
|S1 (j )| 2
2

(n 1) n1
|j |/
1/
n|j |/

again by the PNT. This proves the first part of the statement. For the second part, we
argue in a similar way, replacing the PNT by an appeal to (iii) of Lemma 7 in Tolev [14].

Now let
X1 = { [P/X, R] : |S1 (1 )| |S1 (2 )|}

X2 = { [P/X, R] : |S1 (1 )| |S1 (2 )|}

so that [P/X, R] = X1 X2 and

Z 


Z




S1 (1 )S1 (2 )Sk (3 )K () d.
+
I(, , m)
X2

X1

Cauchys inequality gives

Z
Z
max |S1 (1 )|
X1

X1

4/5+1/(10k)

X1

|S1 (2 )| K () d
4

(log X) (X log X)

1/2

(X

1/2 Z

1/k

X1

|Sk (3 )|2K () d

1/2

(log X)3 )1/2

X 13/10+3/(5k) (log X)6

by Lemmas 4 and 5. The computation on X2 is similar and gives the same final result.
Summing up,




I(, , m) X 13/10+3/(5k) (log X)6 ,


and this is o 2 X 1+1/k provided that
= (X 3/102/(5k) (log X)6 ).

(12)

8. The trivial arc

Using the Cauchy inequality and a trivial bound for Sk (3 ) we see that
Z +




|S1 (1 )| |S1(2 )| |Sk (3 )| K () d
I(, , t) 2
R
Z +
1/2 Z +
1/2
1/k
2
X
|S1 (1 )| K () d
|S1 (2 )|2 K () d
R

1/2

1/2

X 1/k C1 C2 ,

say, where in the last but one line we used the inequality (7), and, for j = 1, 2, we set
Z +
|S1 ()|2
Cj =
d.
2
|j |R
We argue as in the proof of Lemma 5. Using the PNT we have
Z n
X
X log X
1
.
|S1 ()|2 d
Cj
2
(n 1) n1
|j |R
n|j |R

Collecting these estimates, we conclude that



X 1+1/k log X


.
I(, , t)
R
Hence, the choice
R = 2 (log X)3/2
is admissible.

(13)

9. Remark on the choice of the parameters

The choice X = q 5k/(k+2) with 1 k < 4/3 arises from the bounds (10) and (11).
Their combination prevents us from choosing the optimal value X = q 2 . This is justified
as follows: neglecting log-powers, let X = q a(k) , Q = X b(k) , = X c(k) , and recall the
choices P = X 5/(6k) in (9) and R = 2 (log X)3/2 in (13) which are due, respectively,
to the bound for B4 and for the trivial arc. Then, essentially, we have to maximize k
subject to the constraints

a(k) 1

0 b(k) 5/(6k)

c(k) 0

2b(k) 1 1/a(k)
by (10),

2b(k) + 2c(k) 1/a(k) by (11),

c(k) 1 1 1 b(k) by (12),

2
2k
2

which is a linear optimization problem in the variables 1/a(k), b(k), c(k) and 1/k. The
solution for this problem is 1/a(k) = (k + 2)/(5k), b(k) = (2k 1)/(5k), c(k) = (4
3k)/(10k), for 1/k 3/4, and this is equivalent to the statement of the main Theorem.
References

[1] R. C. Baker and G. Harman, Diophantine approximation by prime numbers, J. London Math. Soc.
25 (1982), 201215.
[2] J. Br
udern, R. J. Cook, and A. Perelli, The values of binary linear forms at prime arguments, Proc.
of Sieve Methods, Exponential sums and their Application in Number Theory, 1997, pp. 87100.
[3] R. J. Cook, The value of additive forms at prime arguments, J. Theor. Nombres Bordeaux 13 (2001),
7791.
[4] R. J. Cook and A. Fox, The values of ternary quadratic forms at prime arguments, Mathematika
48 (2001), 137149.
[5] R. J. Cook and G. Harman, The values of additive forms at prime arguments, Rocky Mountain J.
Math. 36 (2006), 11531164.
[6] H. Davenport and H. Heilbronn, On indefinite quadratic forms in five variables 21 (1946), 185193.
[7] G. Harman, Diophantine approximation by prime numbers, J. London Math. Soc. 44 (1991), 218
226.
9

[8] G. Harman, The values of ternary quadratic forms at prime arguments, Mathematika 51 (2004),
8396.
[9] A. Languasco and V. Settimi, On a Diophantine problem with one prime, two squares of primes and s
powers of two, preprint, to appear in Acta Arithmetica (2012). http://arxiv.org/abs/1103.1985.
[10] A. Languasco and A. Zaccagnini, A Diophantine problem with a prime and three squares of primes,
2012. submitted.
[11] A. Languasco and A. Zaccagnini, L2 -norms of exponential sums over prime powers, 2012. submitted.
[12] K. Matom
aki, Diophantine approximation by primes, Glasgow Math. J. 52 (2010), 87106.
[13] S. T. Parsell, Diophantine approximation with primes and powers of two, New York J. Math. 9
(2003), 363371.
[14] D. Tolev, On a Diophantine inequality involving prime numbers, Acta Arith. 61 (1992), 289306.
[15] R. C. Vaughan, Diophantine approximation by prime numbers. I, Proc. London Math. Soc. 28
(1974), 373384.
[16] R. C. Vaughan, Diophantine approximation by prime numbers. II, Proc. London Math. Soc. 28
(1974), 385401.
[17] R. C. Vaughan, The Hardy-Littlewood method, second, Cambridge U. P., 1997.

Alessandro LANGUASCO
Universit`a di Padova
Dipartimento di Matematica
Via Trieste 63
35121 Padova, Italy
E-mail: languasco@math.unipd.it
Alessandro ZACCAGNINI
Universit`a di Parma
Dipartimento di Matematica
Parco Area delle Scienze, 53/a
Campus Universitario
43124 Parma, Italy
E-mail: alessandro.zaccagnini@unipr.it

10