a

r
X
i
v
:
m
a
t
h
/
0
4
0
5
0
5
6
v
1


[
m
a
t
h
.
N
T
]


4

M
a
y

2
0
0
4
Almost All Palindromes
Are Composite

William D. Banks

Dept. of Mathematics, University of Missouri

Columbia, MO 65211, USA
bbanks@math.missouri.edu
Derrick N. Hart
Dept. of Mathematics, University of Missouri
Columbia, MO 65211, USA
hart@math.missouri.edu
Mayumi Sakata
Dept. of Mathematics, University of Missouri
Columbia, MO 65211, USA
sakata@math.missouri.edu
February 1, 2008
Never odd or even...
Abstract
We study the distribution of palindromic numbers (with respect
to a xed base g 2) over certain congruence classes, and we derive
a nontrivial upper bound for the number of prime palindromes n x
as x . Our results show that almost all palindromes in a given
base are composite.

MSC Numbers: 11A63, 11L07, 11N69

Corresponding author
1
1 Introduction
Fix once and for all an integer g 2, and consider the base g representation
of an arbitrary natural number n N:
n =
L1

k=0
a
k
(n)g
k
.
Here a
k
(n) {0, 1, . . . , g1} for each k = 0, 1, . . . , L1, and we assume that
the leading digit a
L1
(n) is nonzero. The integer n is said to be a palindrome
if its digits satisfy the symmetry condition:
a
k
(n) = a
L1k
(n), k = 0, 1, . . . , L 1.
Let P N denote the set of palindromes (in base g), and for every positive
real number x, let
P(x) = {n x| n P}.
In this paper, we study the distribution of palindromes in congruence classes.
Using estimates for twisted Kloosterman sums to bound exponential sums
over the set P
L
of palindromes with precisely L digits, we show that the set
P(x) becomes uniformly distributed (as x ) over the congruence classes
modulo p, where p > g is any prime number for which the multiplicative
order ord
p
(g) of g in the group (Z/pZ)

is at least 3p
1/2
; see Corollary 4.4 to
Theorem 4.3 for a precise statement of this result. We remark that, thanks to
the work of Pappalardi [5], almost all primes p satisfy the stronger condition
ord
p
(g) p
1/2
exp((log p)
c
) where c is any constant less than (1 log 2)/2;
see also [2, 4].
Using a variation of these techniques, we also show that the set P(x) becomes
uniformly distributed (as x ) over the congruence classes modulo q,
where q 2 is any integer relatively prime to g(g
2
1); see Corollary 4.5.
This latter result, although weaker than that obtained for primes p satisfying
the condition ord
p
(g) 3p
1/2
, allows us to deduce the main result of this
paper: almost all palindromes in a given base are composite. More precisely,
in Theorem 5.1, we show that
#
_
n P(x) | n is prime
_
= O
_
#
P(x)
log log log x
log log x
_
, x ,
2
where the implied constant depends only on the base g. This result appears
to be the rst of its kind in the literature.
Acknowledgments. The authors would like to thank Florian Luca and
Igor Shparlinski, whose valuable observations on the original manuscript led
to signicant improvements in our estimates. During the preparation of this
paper, W. B. was supported in part by NSF grant DMS-0070628.
2 Preliminary Estimates
For any integer q 2, let e
q
(x) denote the exponential function exp(2ix/q),
which is dened for all x R. For any integer c that is relatively prime to
q, let c denote an arbitrary multiplicative inverse for c modulo q; that is,
c c 1 (mod q). Finally, let d(q) be the number of positive integral divisors
of q, and let ord
q
(g) be the smallest integer t 1 such that g
t
1 (mod q).
Lemma 2.1. For all q 2 with gcd(q, g) = 1 and all a, b Z, we have

ordq(g)

k=1
e
q
(ag
k
+ bg
k
)

d(q)q
1/2
gcd(a, b, q)
1/2
.
Proof. Consider the twisted Kloosterman sum
K

(a, b; q) =

1cq
gcd(c,q)=1
(c) e
q
(ac + bc),
where is a Dirichlet character modulo q. The Weil-Estermann bound

(a, b; q)

d(q)q
1/2
gcd(a, b, q)
1/2
holds for such sums (the original proofs by Weil [7] and Estermann [3] carry
over for twisted sums with only slight modications). Averaging over all
Dirichlet characters modulo q for which (g) = 1, it follows that
ord
q
(g)
(q)

(a, b; q) =

1cq
cg
k
(mod q), k
e
q
(ac + bc) =
ordq(g)

k=1
e
q
(ag
k
+ bg
k
).
The result follows.
3
Lemma 2.2. The following bound holds for all q 2, k 2 and h Z
provided that q | h:

k1

a=0
e
q
(ha)

k exp
_

4 gcd(h, q)
2
q
2
_
.
Proof. Let us write
s(q, k, h) =

k1

a=0
e
q
(ha)

.
If d = gcd(h, q), then s(q, k, h) = s(q/d, k, h/d), hence it suces to prove the
assertion for the special case where gcd(h, q) = 1, which we now assume.
Without loss of generality, we may also suppose that k q. Indeed, if
k q + 1, then we can express k = mq + r with 0 r q 1 and simply
observe that
s(q, k, h) = s(q, r, h) r q 1 (q + 1) exp(4/q
2
) k exp(4/q
2
).
If gcd(h, q) = 1 and 2 k q, we have
s(q, k, h)
2
=
k1

a,b=0
e
q
(h(a b)) = k +
k1

a,b=0
a=b
cos
_
2h(a b)
q
_
k + k(k 1) cos(2/q);
therefore,
s(q, k, h)
2
k
2

1
k
+
_
1
1
k
_
cos(2/q)
1
2
(1 + cos(2/q)) .
Using the fact that 1 + cos x 2 exp(x
2
/4) for 0 x , we obtain the
desired result.
3 Exponential Sums over Palindromes
For every L 1, let P
L
denote the set of palindromes (in base g) with
precisely L digits; that is,
P
L
= {n P | g
L1
n < g
L
}.
4
Lemma 3.1. Let q 2 be an integer such that p > g for every prime divisor
p of q. Then for every c Z such that
ord
q
(g) > d(q)q
1/2
gcd(c, q)
1/2
,
the exponential sum
S
L
(c) =

nP
L
e
q
(cn)
satises the bound

S
L
(c)

#
P
L

(L2ordq(g)1)/4
c
,
where

c
=
1
g
+
(g 1) d(q)q
1/2
gcd(c, q)
1/2
g ord
q
(g)
.
Proof. Since
S
2L
(c) =
g1

a
0
=1
g1

a
1
=0
. . .
g1

a
L1
=0
e
q
_
L1

k=0
ca
k
_
g
k
+ g
2L1k
_
_
=
g1

a
0
=1
e
q
_
ca
0
_
1 +g
2L1
__
L1

k=1
g1

a
k
=0
e
q
_
ca
k
_
g
k
+ g
2L1k
__
and
S
2L+1
(c) =
g1

a
0
=1
g1

a
1
=0
. . .
g1

a
L
=0
e
q
_
ca
L
g
L
+
L1

k=0
ca
k
_
g
k
+ g
2Lk
_
_
=
g1

a
0
=1
e
q
_
ca
0
(1 +g
2L
)
_
g1

a
L
=0
e
q
(ca
L
g
L
)
L1

k=1
g1

a
k
=0
e
q
_
ca
k
_
g
k
+ g
2Lk
__
,
it follows that

S
2L+
(c)

(g 1)g

L1

k=1

g1

a=0
e
q
_
ca
_
g
k
+ g
2L+1k
__

for all L 1 and = 0 or 1.

5
Put N = ord
q
(g), and write L 1 = Nm + , where m = (L 1)/N and
0 < N. Then, using the arithmetic-geometric mean inequality, we derive
that

S
2L+
(c)

2
(g 1)
2
g
2+2
Nm

k=1

g1

a=0
e
q
_
ca
_
g
k
+ g
2L+1k
__

2
(g 1)
2
g
2+2
_
_
1
Nm
Nm

k=1

g1

a=0
e
q
_
ca
_
g
k
+ g
2L+1k
__

2
_
_
Nm
= (g 1)
2
g
2+2
_
T
Nm
_
Nm
,
where
T =
Nm

k=1
g1

a,b=0
e
q
_
c(a b)
_
g
k
+ g
2L+1k
__
= gNm+ m
g1

a,b=0
a=b
N

k=1
e
q
_
c(a b)
_
g
k
+ g
2L+1k
__
.
Since p > g for every prime p | q, it follows that gcd(a b, q) = 1 whenever
a = b. Using Lemma 2.1, we therefore obtain that

gNm+ (g 1)gmd(q)q
1/2
gcd(c, q)
1/2
.
Consequently,

S
2L+
(c)

2
(g 1)
2
g
2+2
_
gNm+ (g 1)gmd(q)q
1/2
gcd(c, q)
1/2
Nm
_
Nm
= (g 1)
2
g
2Nm+2+2
_
N + (g 1) d(q)q
1/2
gcd(c, q)
1/2
gN
_
Nm
= (g 1)
2
g
2L+22

L1
c
.
Since
#
P
2L+
= (g 1)g
L+1
, the result follows.
6
Lemma 3.2. Let q 2 be an integer such that gcd
_
q, g(g
2
1)
_
= 1. Then
for every c Z such that q | c, the exponential sum
S
L
(c) =

nP
L
e
q
(cn)
satises the bound

S
L
(c)

#
P
L
exp
_

(L 5) gcd(c, q)
2
q
2
_
.
Proof. As in the proof of Lemma 3.1, we have

S
2L+
(c)

(g 1)g

L1

k=1

g1

a=0
e
q
_
ca
_
g
k
+ g
2L+1k
__

for all L 1 and = 0 or 1. Let

B = {1 k L 1 | q divides c(g
k
+ g
2L+1k
)},
G = {1 k L 1 | q does not divide c(g
k
+ g
2L+1k
)}.
Using Lemma 2.2 to estimate individual terms in the preceding product when
k G, and using the trivial estimate when k B, we obtain that

S
2L+
(c)

(g 1)g
(+
#
G+
#
B)
exp
_

4 gcd(c, q)
2 #
G
q
2
_
=
#
P
2L+
exp
_

4 gcd(c, q)
2 #
G
q
2
_
.
Now let f = q/ gcd(c, q). Since q does not divide c, we have f 2, and the
stated condition on q implies that ord
f
(g
2
) 2. Thus, if k and both lie in
B, then
(g
2
)
k
g
2L+1
(g
2
)

(mod f), k (mod ord

f
(g
2
)).
We therefore see that
#
B 1 +(L 2)/2 = L/2 ,
#
G L 1 L/2 L/2 1 (2L + 5)/4,
and the result follows.
7
4 Distribution of Palindromes
Proposition 4.1. Let p > g be a prime number such that ord
p
(g) 3p
1/2
.
Then for every L 10p 5, the following estimate holds for all a Z:

#
_
n P
L
| n a (mod p)
_

#
P
L
p

<
#
P
L
p
(0.99)
L
.
Proof. Using the relation
1
p
p1

c=0
e
p
(cm) =
_
1 if m 0 (mod p),
0 otherwise,
it follows that
#
_
n P
L
| n a (mod p)
_
=

nP
L
1
p
p1

c=0
e
p
(c(n a))
=
1
p
p1

c=0
e
p
(ca)

nP
L
e
p
(cn)
=
#
P
L
p
+
1
p
p1

c=1
e
p
(ca)S
L
(c),
where S
L
(c) is the exponential sum considered in Lemma 3.1. Therefore

#
_
n P
L
| n a (mod p)
_

#
P
L
p

1
p
p1

c=1

S
L
(c)

#
P
L
p
p1

c=1

(L2 ordp(g)1)/4
c
,
where for each c = 1, . . . , p 1, we have

c
=
1
g
+
2(g 1)p
1/2
g ord
p
(g)

1
g
+
2(g 1)
3g
=
2g + 1
3g

5
6
since g 2. Also,
(L 2 ord
p
(g) 1)/4 (L 2p + 1)/4 L/5, L 10p 5.
8
Consequently,

#
_
n P
L
| n a (mod p)
_

#
P
L
p

#
P
L
p
(p 1)
_
5
6
_
L/5
.
Finally, remarking that the condition ord
p
(g) 3p
1/2
implies that p 11,
we have
(p 1)
_
5
6
_
L/5
< (0.99)
L
, L 10p 5.
This completes the proof.
Proposition 4.2. Let q 2 be an integer such that gcd
_
q, g(g
2
1)
_
= 1.
Then for every L 10 +2q
2
log q, the following estimate holds for all a Z:

#
_
n P
L
| n a (mod q)
_

#
P
L
q

<
#
P
L
q
exp
_

L
2q
2
_
.
Proof. Using the relation
1
q
q1

c=0
e
q
(cm) =
_
1 if m 0 (mod q),
0 otherwise,
it follows that
#
_
n P
L
| n a (mod q)
_
=

nP
L
1
q
q1

c=0
e
q
(c(n a))
=
1
q
q1

c=0
e
q
(ca)

nP
L
e
q
(cn)
=
#
P
L
q
+
1
q
q1

c=1
e
q
(ca)S
L
(c),
where S
L
(c) is the exponential sum considered in Lemma 3.2. If 1 c q1,
then q | c, hence by Lemma 3.2 we derive the estimate:

S
L
(c)

#
P
L
q
exp
_
log q
(L 5) gcd(c, q)
2
q
2
_

#
P
L
q
exp
_
log q
L 5
q
2
_

#
P
L
q
exp
_

L
2q
2
_
,
9
the last inequality following from the stated condition on L. The result
follows immediately.
Theorem 4.3. Let q 2 be a xed integer, and suppose that there exist
constants A 1 and
_
2/3 < 1, depending only on q, such that

#
_
n P
L
| n a (mod q)
_

#
P
L
q

#
P
L
A
L
for all L 1 and a Z. Then for some constant B 1 that depends only
on g, the following estimate holds for all x 1 and a Z:

#
_
n P(x) | n a (mod q)
_

#
P(x)
q

#
P(x) AB
(log x)/(2 log g)
.
Proof. We remark that the condition
_
2/3 guarantees that g
2
is
bounded below by an absolute constant greater than 1; since g 2, we
have
g 1
g
2
1

g 1
2
3
g 1
3.
For all L 1, x y > 0, and a Z, let us denote
P
a
= {n P | n a (mod q)},
P
a,L
= {n P
a
| g
L1
n < g
L
},
P
a
(x) = {n P
a
| n x},
P
a
(y; x) = {n P
a
| y < n x}.
We also denote
P(y; x) = {n P | y < n x}.
In what follows, the implied constants in the symbol O may depend on
g but are absolute otherwise. We recall that the notation U = O(V ) for
positive functions U and V is equivalent to U cV for some constant c.
Let a Z be xed in what follows, and suppose that g
2M+1
x < g
2M+
,
where M is an integer and = 0 or 1. We observe that
#
P(x) =
#
P(g
2M+1
) +
#
P(g
2M+1
; x), (1)
10
and that
#
P
a
(x) =
#
P
a
(g
2M+1
) +
#
P
a
(g
2M+1
; x).
Our goal is to estimate

#
P
a
(x)
#
P(x)
q

(2)

#
P
a
(g
2M+1
)
#
P(g
2M+1
)
q

#
P
a
(g
2M+1
; x)
#
P(g
2M+1
; x)
q

.
Since the integer g
2M+1
is not a palindrome (a fact that is only used to
simplify our notation), we have by a straightforward calculation:
#
P(g
2M+1
) = g
M
+ g
M+1
2. (3)
On the other hand,
#
P
a
(g
2M+1
) =
2M+1

L=1
#
P
a,L
=
M1

=0
#
P
a,2+1
+
M+1

=1
#
P
a,2
=
M1

=0
_
#
P
a,2+1

#
P
2+1
q
+
#
P
2+1
q
_
+
M+1

=1
_
#
P
a,2

#
P
2
q
+
#
P
2
q
_
=
#
P(g
2M+1
)
q
+
M1

=0
_
#
P
a,2+1

#
P
2+1
q
_
+
M+1

=1
_
#
P
a,2

#
P
2
q
_
.
Using the hypothesis of the theorem, it therefore follows that

#
P
a
(g
2M+1
)
#
P(g
2M+1
)
q

M1

=0
#
P
2+1
A
2+1
+
M+1

=1
#
P
2
A
2
.
Since
M1

=0
#
P
2+1

2+1
+
M+1

=1
#
P
2

2
=
M1

=0
(g 1)g

2+1
+
M+1

=1
(g 1)g
1

2
<
g 1
g
2
1
_
g
M

2M+1
+ g
M+1

2M+2
_
= O
_
g
M

2M
_
,
11
we see that

#
P
a
(g
2M+1
)
#
P(g
2M+1
)
q

= O
_
Ag
M

2M
_
. (4)
We now turn to the more delicate estimation of
#
P
a
(g
2M+1
; x). To this
end, put M = K + L, where K and L are positive integers to be selected
later. Examining the base g representation of an arbitrary palindrome n in
P
2M+
, we see that n may be expressed either in the form
n = n
1
+ g
K+
n
2
+ g
K+2L+
n
3
,
or the form
n = n
1
+ g
K+2L+
n
3
,
where
1 n
1
< g
K
, g
K1
n
3
< g
K
, n
1
+ g
K
n
3
P
2K
, (5)
and, in the former case, n
2
P
2L+2
for some 0 L + 1. The in-
tegers n
1
, n
2
, n
3
, are uniquely determined by n. We call n
3
the K-signature
of n and write s
K
(n) = n
3
. The integer n
1
is uniquely determined by n
3
to-
gether with the rst and third conditions of (5); we call n
1
the K-complement
of n
3
and write c
K
(n
3
) = n
1
.
Note that the number of palindromes n P
2M+
with a xed K-signature
s
K
(n) = n
3
is precisely
1 +
L+1

=0
#
P
2L+2
= 1 +
L+1

=0
(g 1)g
L+1
= g
L+
. (6)
Now, given x in the range g
2M+1
x < g
2M+
, let y be the palindrome in
P
2M+
dened by
y = y
1
+ g
K
(g
2L+
1) +g
K+2L+
y
3
,
where
y
3
=
_
_
_
_
x/g
K+2L+
_
+ 1 if g
2M+1
x < g
2M+1/2
,
_
x/g
K+2L+
_
1 if g
2M+1/2
x < g
2M+
,
12
and y
1
= c
K
(y
3
). If x lies in the smaller range, then x < y, while y < x if x
lies in the larger range. In either case, we have

#
P(g
2M+1
; x)
#
P(g
2M+1
; y)

= O(g
L
) (7)
and

#
P
a
(g
2M+1
; x)
#
P
a
(g
2M+1
; y)

= O(g
L
),
since there are at most O(1) distinct K-signatures for palindromes between
x and y. Consequently,

#
P
a
(g
2M+1
; x)
#
P(g
2M+1
; x)
q

(8)
=

#
P
a
(g
2M+1
; y)
#
P(g
2M+1
; y)
q

+ O(g
L
).
Now, if n P(g
2M+1
; y), then its K-signature lies in the range
g
K1
s
K
(n) y
3
.
Thus,
#
P(g
2M+1
; y) = (y
3
g
K1
+ 1)g
L+
. (9)
On the other hand, if n P
a
(g
2M+1
; y) with s
K
(n) = n
3
, then either
n = n
1
+ g
K+
n
2
+ g
K+2L+
n
3
a (mod q)
or
n = n
1
+ g
K+2L+
n
3
a (mod q),
depending on the form of n. In the latter case, there is at most one such
palindrome n (for each xed K-signature n
3
), while in the former case, since
n
2
g
K
_
a c
K
(n
3
) g
K+2L+
n
3
_
(mod q),
the number of such palindromes n is
#
P
b,2L+2
for each 0 L+ 1,
where
b = b(n
3
, ) = g
K
(a c
K
(n
3
) g
K+2L+
n
3
).
13
Hence, using (6), we derive that
#
P
a
(g
2M+1
; y) =
y
3

n
3
=g
K1
L+1

=0
#
P
b,2L+2
+ O(g
K
)
=
y
3

n
3
=g
K1
_
1
q
+
L+1

=0
#
P
2L+2
q
_
+
y
3

n
3
=g
K1
L+1

=0
_
#
P
b,2L+2

#
P
2L+2
q
_
+ O(g
K
)
=
#
P(g
2M+1
; y)
q
+
y
3

n
3
=g
K1
L+1

=0
_
#
P
b,2L+2

#
P
2L+2
q
_
+ O(g
K
).
Using the hypothesis of the theorem, it therefore follows that

#
P
a
(g
2M+1
; y)
#
P(g
2M+1
; y)
q

y
3

n
3
=g
K1
L+1

=0
#
P
2L+2
A
2L+2
+ O(g
K
)
=
y
3

n
3
=g
K1
L+1

=0
(g 1)g
L+1
A
2L+2
+ O(g
K
)
< A(y
3
g
K1
+ 1)
_
g 1
g
2
1
g
L+

2L++2
_
+ O(g
K
),
and consequently,

#
P
a
(g
2M+1
; y)
#
P(g
2M+1
; y)
q

= O
_
Ag
M

2L
_
+ O(g
K
).
Using this estimate together with (2), (4) and (8), it follows that

#
P
a
(x)
#
P(x)
q

= O
_
Ag
M

2L
+ g
L
+ g
K
_
.
We now choose integers K = M/2 + O(1) and L = M/2 + O(1) such that
K + L = M. Since g
2
> 1 and A 1, we have
max{g
K
, g
L
} = O(g
M/2
) = O
_
Ag
M/2
(g
2
)
M/2
_
= O
_
Ag
M

M
_
,
14
therefore

#
P
a
(x)
#
P(x)
q

= O
_
Ag
M

M
_
.
To complete the proof, we need only observe that

M
= O
_

(log x)/(2 log g)

_
for x in the range g
2M+1
x < g
2M+
, and using (1), (3), (7) and (9)
together with our choice of y
3
, it follows that
#
P(x) = g
M
+
x
g
M
+ O(g
M/2
);
thus g
M
= O
_
#
P(x)
_
.
Using Theorem 4.3, we can now derive two immediate corollaries.
Corollary 4.4. Let p > g be a prime number such that ord
p
(g) 3p
1/2
.
Then for some constant C > 0, depending only on g, the following estimate
holds for all x 1 and a Z:

#
_
n P(x) | n a (mod p)
_

#
P(x)
p

#
P(x) C (0.99)
log x
2 log g
10p
.
Proof. Using the trivial estimate

#
_
n P
L
| n a (mod p)
_

#
P
L
p

#
P
L
for 1 L 10p 6, it follows from Proposition 4.1 that the estimate

#
_
n P
L
| n a (mod p)
_

#
P
L
p

#
P
L
(0.99)
L10p+6
holds for all L 1 and a Z. The result now follows immediately from
Theorem 4.3.
Corollary 4.5. Let q 2 be an integer such that gcd
_
q, g(g
2
1)
_
= 1.
Then for some constant C > 0, depending only on g, the following estimate
holds for all x 1 and a Z:

#
_
n P(x) | n a (mod q)
_

#
P(x)
q

#
P(x) C q exp
_

log x
4q
2
log g
_
.
15
Proof. Using the trivial estimate

#
_
n P
L
| n a (mod q)
_

#
P
L
q

#
P
L
for 1 L < 10 + 2q
2
log q, it follows from Proposition 4.2 that the estimate

#
_
n P
L
| n a (mod q)
_

#
P
L
q

#
P
L
exp
_

(L 10 2q
2
log q)
2q
2
_
holds for all L 1 and a Z. The result now follows immediately from
Theorem 4.3.
5 Prime Palindromes
We now come to the main result of this paper.
Theorem 5.1. As x , we have
#
_
n P(x) | n is prime
_
= O
_
#
P(x)
log log log x
log log x
_
,
where the implied constant depends only on g.
Proof. As in the proof of Theorem 4.3, all implied constants in the symbol
O may depend on g but are absolute otherwise.
Assuming that x is suciently large, let
h = e log log log x , y = e
1
(log x)
1/4h
= exp
_
log log x
4e log log log x
_
1+o(1)
.
Let
Q = Q(y) =

g
3
<py
p,
where the product runs over prime numbers. Note that gcd
_
Q, g(g
2
1)
_
= 1.
By Mertens formula (see Theorem 11 in I.1.6 of [6]), we have the estimate
(Q)
Q
=

g
3
<py
_
1
1
p
_
= O
_
(log y)
1
_
= O
_
log log log x
log log x
_
, (10)
16
where (n) is the Euler function.
Now, if n P(x) is prime, either gcd(n, Q) = 1 or n is a prime divisor of
Q. We apply Bruns combinatorial sieve in the form given by Corollary 1.1
in I.4.2 of [6]:
#
_
n P(x) | n is prime
_
y +

q | Q
(q)2h
(q)A
q
,
where (q) is the Mobius function, (q) is the number of distinct prime
divisors of q, and
A
q
=
#
{n P(x) | n 0 (mod q)}.
By Corollary 4.5, we see that
A
q
=
#
P(x)
q
+ O
_
#
P(x) q exp
_

log x
4q
2
log g
__
.
If q | Q and (q) 2h, then
q y
2h
=
(log x)
1/2
e
2h
,
and since the number of such divisors q is bounded by y
2h
, we have

q | Q
(q)2h
q exp
_

log x
4q
2
log g
_

log x
e
4h
exp
_

e
4h
4 log g
_
= exp
_
log log x 4h
e
4h
4 log g
_
= O
_
1
log x
_
,
since h = e log log log x. Therefore,
#
_
n P(x) | n is prime
_
y +
#
P(x)

q | Q
(q)
q
+ O
_
_
_
_
#
P(x)

q | Q
(q)>2h
1
q
+ O
_
#
P(x)
log x
_
_
_
_
_
.
17
Since y = x
o(1)
and x
1/2
= O
_
#
P(x)
_
, the rst term in this estimate is
negligible. Also, using (10), we have
#
P(x)

q | Q
(q)
q
=
#
P(x)

g
3
<py
_
1
1
p
_
= O
_
#
P(x)
log log log x
log log x
_
.
Finally, we have

q | Q
(q)>2h
1
q

q | Q
(q)>2h
e
(q)2h
q
e
2h

py
(1 +e/p) exp
_
2h + e

py
1/p
_
.
Observing that

py
1
p
= (log log y)(1 +o(1)) = (log log log x)(1 +o(1)),
by our choice of h it follows that
#
P(x)

q | Q
(q)>2h
1
q

#
P(x) exp ((log log log x)(e + o(1))) = O
_
#
P(x)
(log log x)
2
_
.
This completes the proof.
6 Remarks and Open Problems
Using estimates from [1], it is possible to establish a version of Lemma 3.1 in
the case where q = p is prime with ord
p
(g) log p; this yields analogues of
Proposition 4.1 and Corollary 4.4 under the weaker assumption on ord
p
(g),
however the uniform constant 0.99 in those results must be replaced by a
term like exp((log log p)
c
) for some constant c > 0.
It seems natural to conjecture that the set of palindromes should behave as
random integers, thus one might expect that the asymptotic relation
#
_
n P(x) | n is prime
_
C
#
P(x)
log x
18
holds for some constant C > 0. While this question seems out of reach at
the moment, it should be feasible to derive the upper bound
#
_
n P(x) | n is prime
_
= O
_
#
P(x)
log x
_
using more sophisticated sieving techniques coupled with better estimates
for the distribution of palindromes in congruence classes. It is still an open
problem to show the existence of innitely many prime palindromes for any
xed base g 2.
References
[1] T. Cochrane, C. Pinner and J. Rosenhouse, Bounds on exponential
sums and the polynomial Waring problem mod p, J. London Math.
Soc. (2) 67 (2003) no. 2, 319336.
[2] P. Erdos and R. Murty, On the order of a (mod p), Number theory
(Ottawa, ON, 1996), 8797, CRM Proc. Lecture Notes 19, Amer.
Math. Soc., Providence, RI, 1999.
[3] T. Estermann, On Kloostermans sum, Mathematika 8 (1961) 8386.
[4] K.-H. Indlekofer and N. Timofeev, Divisors of shifted primes, Publ.
Math. Debrecen 60 (2002) no. 3-4, 307345.
[5] F. Pappalardi, On the order of nitely generated subgroups of Q

(mod p) and divisors of p 1, J. Number Theory 57 (1996) no. 2,

207222.
[6] G. Tenenbaum, Introduction to analytic and probabilistic number the-
ory, University Press, Cambridge, UK, 1995.
[7] A. Weil, On some exponential sums, Proc. Nat. Acad. Sci. U.S.A.
(1948) 204207.
19