ACTA ARITHMETICA

LXIV.3 (1993)
Cyclotomic polynomials with large coecients
by
Helmut Maier (Athens, Ga.)
Dedicated to Paul Erdos
on the occasion of his eightieth birthday
1. Introduction. Let
n
(z) =

(n)
m=0
a(m, n)z
m
be the nth cyclotomic
polynomial. Let
A(n) = max
0m(n)
[a(m, n)[ and S(n) =

0m(n)
[a(m, n)[ .
The coecients a(m, n) and especially A(n) and S(n) have been the subject
of numerous investigations (see [1] and the references given there). Until
recently all these investigations concerned very thin sets of integers n. In [3]
the author could establish a property valid for a set of integers of asymptotic
density 1. Let (n) be any function dened for all positive integers such that
lim
n
(n) = 0. Then S(n) n
1+(n)
for a set of integers of asymptotic
density 1. Here we deal with properties valid for sequences of positive lower
density.
Theorem. For any N > 0, there are c(N) > 0 and x
0
(N) 1 such that
cardn x : A(n) n
N
c(N)x,
for all x x
0
(N).
2. A certain set of candidates. Let N > 0 be given. In this section
we identify a certain set of integers in which a large subset will later be
shown to have A(n) n
N
. To describe the set, we x a positive odd integer
(2.1) K = K(N)
(to be determined later) and set
L = 20K , =
1
100L
, =

L
2
.
The author was supported in part by an NSF grant.
228 H. Mai er
The letter p always denotes prime numbers and (m) denotes the number
of distinct prime factors of m. Basic for our construction is the set S.
Let
(2.2) S = S(N, x) = n = mp
1
. . . p
L
x : x
1
L

< p
i
x
1
L
+
,
(m) = (n) = 1 , (m) (1 + )(log log x) .
Lemma. For each N, there are eectively computable constants c
0
(N)
> 0 and x
0
(N) > 0 such that for all x x
0
(N),
cardS c
0
(N)x.
Pr o o f. Let T = T (x) be the set of n x which have all of the properties
of elements of S, but the condition (n) = 1 fails. This implies that at least
two of the primes p
1
, . . . , p
L
are the same. We have
(2.3) cardS
1
L!

p
1
. . .

p
L

mx/p
1
...p
L
(m)=1,(m)(1+) log log x
1 cardT
where

denotes a sum over p

i
with x
1
L

< p
i
x
1
L
+
.
Clearly
cardT x

p
1
p
2
= o(x) for x .
From [2] and [4] we know that the inner sum in (2.3) is at least x/10p
1
. . . p
L
for all x x
1
(N), where x
1
(N) is a constant depending only on the choice
of N. Thus
(2.4) cardS
x
10L!
_

1
p
_
L
o(x) for x .
Now

1
p
= log
_
1
L
+
_
log
_
1
L

_
+ O
_
1
log x
_
,
so there is some number c
1
(N) such that

1
p
c
1
(N) > 0
for all x x
2
(N), where x
2
(N) is a constant depending only on N. The
lemma now follows from (2.4).
3. Investigation of log [
n
(z)[ on the unit circle. We start with the
well-known identity
(3.1)
n
(z) =

d | n
(1 z
d
)
(n/d)
Cyclotomic polynomials with large coecients 229
for all complex z for which both sides are dened. We write e
2i
= e()
and obtain
(3.2) log [
n
(e())[ =

d | n

_
n
d
_
log [1 e(d)[ .
To show that A(n) is large it would suce to show there is some with
log [
n
(e())[ large. The terms on the right of (3.2) will be large for (n/d)
= 1 and |d| small. (Here || denotes the distance to the nearest integer.)
Simple choices for the pair (, d) however do not work because of a certain
cancellation eect which has already been described in [3]. We repeat its
description for the convenience of the reader.
Assume r [ n/d, (n/d) = 1, (n) ,= 0, (r) 2 and d = k +
with k an integer, [[ 1/2. Thus [[ = |d|. Also assume that [r[
1/2. For t [ r we have e(dt) = e(t) = 1 + 2it + O((t)
2
). This implies
log [1 e(dt)[ = log(t) + O(1). Thus

t | r

_
n
dt
_
log [1 e(dt)[
=

t | r

_
n
dt
_
(log + log t) + O
_

t | r
1
_
= O
_

t | r
1
_
,
where we use

t | r
(t) = 0 and (for (r) 2)

t | r
(t) log t = 0 .
Thus the large contribution (n/d) log [1e(d)[ is cancelled by other terms.
A method to avoid this cancellation eect is to choose and d such that
for t < t
0
we have [t[ 1/2 but for t t
0
, [t[ > 1/2. This leads to
estimates of incomplete convolutions

t | r
t<t
0

_
n
dt
_
(log + log t)
which can be made large by an appropriate choice of d, r, t
0
and . For the
remaining sum

t | r
tt
0

_
n
dt
_
log [1 e(dt)[ ,
we have to show that the terms are small for appropriate choice of . This
will be done by showing that |dt| is not too small.
230 H. Mai er
Definitions. Let S(m
0
) = n S : n = m
0
p
1
. . . p
L
, y
0
= x

1
L
(K+1)
.
For each m
0
we dene an interval
I(m
0
) = [m
1
0
+ m
1
0
y
0
, m
1
0
+ 2m
1
0
y
0
] .
For n S with n = mp
1
. . . p
L
, set (n) = p
1
p
2
. . . p
L
. If n S(m
0
), we
write
(3.3) log [
n
(e())[ =

0
+

1
+

2
,
where

0
=

d=m
0
t,t | (n)
(t)K

_
n
d
_
log [1 e(d)[ ,

1
=

| m
0
m

=m
0

_
n
m

_
log [1 e(m

)[ ,

2
=

all other divisors

d | n

_
n
d
_
log [1 e(d)[ .
We shall investigate these three sums for I(m
0
).
4. The main part

0
. Let t [ (n) with (t) = Kl, 0 l K. Then
t is the product of K l distinct primes from [x
1
L

, x
1
L
+
]. Therefore
(4.1) t [x
Kl
L
(1)(Kl)
, x
Kl
L
(1)+(Kl)
] .
Moreover, if I(m
0
), then m
0
t [t + ty
0
, t + 2ty
0
], so that m
0
t
[ty
0
, 2ty
0
], where means fractional part. We write m
0
t = ty
0
with
1 2. We have e(m
0
t) = e(m
0
t) = 1 + 2ity
0
+ O((ty
0
)
2
) and
thus
(4.2) log [1 e(m
0
t)[ = log ty
0
+ O(1).
From (4.1) and the denition of y
0
we get
(4.3) ty
0
[x

l+1
L
(Kl1)

L
(Kl)
, x

l+1
L
(Kl1)

L
+(Kl)
].
For l = 0, that is, for (t) = K, we use the upper bound in (4.3) and
together with (4.2) we get
log [1 e(m
0
t)[
_

1
L
(K 1)

L
+ K
_
log x + O(1)
1
L
log x,
for x suciently large. There are
_
L
K
_
divisors t [ (n) with (t) = K and
Cyclotomic polynomials with large coecients 231
for each we have (n/(m
0
t)) = 1. Thus we get
(4.4)

t | (n)
(t)=K

_
n
m
0
t
_
log [1 e(m
0
t)[
_
L
K
_
L
1
log x.
For 1 l K, that is, (t) = K l, from (4.2) and (4.3) we get
[ log [1e(m
0
t)[ [
_
l + 1
L
+(Kl1)

L
+(Kl)
_
log x+O(1)
3l
L
log x,
for x suciently large. Since there are
_
L
Kl
_
divisors t [ (n) with (t) =
K l we get
(4.5)

t | (n)
(t)=Kl

_
n
m
0
t
_
log [1 e(m
0
t)[

3l
L
_
L
K l
_
log x.
We study the ratio of these upper bounds for consecutive l-values. For l 1,
3(l + 1)
L
_
L
K l 1
__
3l
L
_
L
K l
_
=
_
1 +
1
l
_
K l
L (K l) + 1
<
2K
L K
=
2
19
.
From this, (4.4) and (4.5) we obtain for x suciently large

1
L
_
L
K
_
log x
3
L
_
L
K 1
_
log x

i=0
_
2
19
_
i
(4.6)
=
1
L
_
L
K
_
log x
57
17

1
L
_
L
K 1
_
log x
=
_
1
57
17
K
L K + 1
_
1
L
_
L
K
_
log x >
14
17L
_
L
K
_
log x.
5. The divisors of m
0
. Our aim now is to show that

1
is small for
n S, I(m
0
). By denition we have

1
=

| m
0
m

=m
0

_
n
m

_
log [1 e(m

)[ .
Note that for I(m
0
) and m

[ m
0
, m

< m
0
we have 0 < m

< 1.
Thus
e(m

) = 1 + 2im

+ O((m

)
2
) .
From this we get
[1 e(m

)[ = 2m

(1 + O(m

))
232 H. Mai er
and so
log [1 e(m

)[ = log m

+ log + O(1) .
Thus for all n S,

1
=
_
n
m
0
_

m

| m
0
m

=m
0

_
m
0
m

_
(log m

+ log ) + O((log x)
(1+) log 2
) ,
since (m
0
) (1 + ) log log x.
We have (since (m
0
) = 1 implies m
0
is not a prime or prime power)

| m
0
m

=m
0

_
m
0
m

_
log m

= log m
0
and

| m
0
m

=m
0

_
m
0
m

_
log = log .
Since log(m
0
) 1, this yields
(5.1)

1
(log x)
(1+) log 2
for n S , I(m
0
) .
6. The divisors d = m

t with (t) K. The remaining divisors

in

2
are of two kinds. The rst kind are of the form m

t with m

[ m
0
,
m

< m
0
, t [ (n) and 1 (t) K. We treat the contribution of these
divisors in this section, leaving the treatment for the remaining divisors,
which are of the form m

t with (t) > K, for the nal section.

Let C > 0 be a constant that we will soon choose as a large absolute
constant. If we have |m
1
0
m

t| > 2(log x)
C
for (t) K, then we also
have
(6.1) |m

t| (log x)
C
for all I(m
0
). Indeed,
[m

t m
1
0
m

t[ 2m
1
0
y
0
m

t y
0
t = o((log x)
C
)
for any C.
We study the exceptional set
(6.2) S
E
(m
0
) = n S(m
0
) : |m
1
0
m

t| 2(log x)
C
for some m

[ m
0
, m

< m
0
, t [ (n) , 1 (t) K .
We shall replace the inequality with a congruence. Let n S
E
and suppose
|m
1
0
m

t| 2(log x)
C
. Let m
1
0
m

t = k + where k is an integer and

Cyclotomic polynomials with large coecients 233
[[ 2(log x)
C
. Then
t =
m
0
m

k +
m
0
m

.
Note that r := (m
0
/m

) is an integer. Thus
(6.3) t r mod
m
0
m

, [r[ 2
m
0
m

(log x)
C
.
We estimate the cardinality of S
E
(m
0
) by writing S
E
(m
0
) as a union
of subsets. For a given m

[ m
0
with m

< m
0
and a given integer g with
1 g K, let
S
E
(m
0
, m

, g) = n S(m
0
) : |m
1
0
m

t| 2(log x)
C
for some t [ (n) with (t) = g .
Note that if n = m
0
tu S
E
(m
0
, m

, g) then u x/(m
0
t) and
(6.4) x
g(
1
L
)
< t x
g(
1
L
+)
and (6.3) holds for some integer r. Thus
[S
E
(m
0
, m

, g)[

ux/(m
0
t)
1
x
m
0

t
1
t
where

denotes a sum over t satisfying (6.4) and satisfying (6.3) for some
integer r. Since t [ (n) and (t) 1 one has t > m
0
/m

, so that possible
solutions of (6.3) with t m
0
/m

do not occur in the sum

. Thus for a
xed r, we have

(r)
t
1
t

g log x
m
0
/m

uniformly in r. Since each prime divisor of t exceeds m

0
/m

, we see that
r = 0 is not a possibility in (6.3) and so the set of possible values of r is
empty when 2(m
0
/m

)(log x)
C
< 1. Therefore
[S
E
(m
0
, m

, g)[ g
m

m
2
0
xlog x

|r|2|m
0
/m

|(log x)
C
1 .
Thus

m
0
[S
E
(m
0
)[

m
0

| m
0
K

g=1
[S
E
(m
0
, m

, g)[

m
0

| m
0
K
2

m
0
x(log x)
1C
<

m
0
(m
0
)
m
0
x(log x)
1C
x(log x)
3C
.
234 H. Mai er
7. The larger divisors. Here we study the divisors d = m

t with
m

[ m
0
, t [ (n), (t) > K. In contrast to the last section we here have to
remove exceptional -values from I(m
0
).
Given a xed triplet (m
0
, m

, t) with m

[ m
0
, t [ (n), (t) > K, we
want an estimate for E(m
0
, m

, t), where E(m

0
, m

, t) = I(m
0
) :
|m

t| (log x)
C
, and denotes the Lebesgue measure. We have
E(m
0
, m

, t) =
_
I(m
0
) :

_
k
m

t

(log x)
C
m

t
,
k
m

t
+
(log x)
C
m

t
_
for some k Z
_
.
We determine the number of integers k for which
(7.1)
_
k
m

t

(log x)
C
m

t
,
k
m

t
+
(log x)
C
m

t
_
I(m
0
) ,= .
From the denition of I(m
0
), such k-values satisfy
m

t
m
0
+
y
0
m

t
m
0
(log x)
C
k
m

t
m
0
+
2y
0
m

t
m
0
+ (log x)
C
.
Since (t) > K, we have y
0
m

t/m
0
as x . Thus for large x,
the number of integers k satisfying (7.1) is at most 2y
0
m

t/m
0
. Thus for
large x, we have
(7.2) E(m
0
, m

, t)
2y
0
m
0
(log x)
C
.
For n S, let
J(n) = I(m
0
)
_
m

| m
0
t | (n),(t)>K
E(m
0
, m

, t) .
Thus from (7.2) we have
J(n)
y
0
m
0

| m
0
t | (n),(t)>K
2y
0
m
0
(log x)
C

y
0
m
0
_
1
2(m
0
)2
L
(log x)
C
_

y
0
m
0
(1 2
L+1
(log x)
1C
)
from the denition of S. Thus for x large, we have J(n) ,= .
We now use the results of this section and the previous section to es-
timate

2
. Let n S S
E
and let J(n). Then from (6.1) and the
denition of J(n), we have for each divisor d of n in the sum

2
that
|d| (log x)
C
.
Cyclotomic polynomials with large coecients 235
Thus for these values of d and we have
[log [1 e(d)[ [ log log x.
Recalling the denition of

2
in (3.3) we conclude that

2
(m
0
)2
L
log log x (log x)
(1+2) log 2
.
Combining this estimate with our estimates (4.6) and (5.1) for

0
and

1
,
we have for x suciently large, n S S
E
, and J(n),
log(nA(n)) log S(n) log [
n
(e())[ >
1
2
_
L
K
_
L
1
log x.
From Sections 2 and 6 we have [S S
E
[
K
x for any xed K. Thus by
choosing K suciently large, we have the Theorem.
Acknowledgments. The author wants to thank Professor Carl Pomer-
ance for valuable advice concerning the presentation of this paper.
References
[1] P. T. Bat eman, C. Pomer ance and R. C. Vaughan, On the size of the coecients
of the cyclotomic polynomial , in: Topics in Classical Number Theory, Vol. I, G. Halasz
(ed.), Colloq. Math. Soc. Janos Bolyai 34, North-Holland, 1984, 171202.
[2] G. H. Har dy and S. Ramanuj an, The normal number of prime factors of a num-
ber n, Quart. J. Math. (Oxford) 48 (1917), 7692.
[3] H. Mai er, The coecients of cyclotomic polynomials, in: Analytic Number Theory,
Proc. Conf. in Honor of Paul T. Bateman, Progr. Math. 85, Birkhauser, 1990, 349
366.
[4] K. Pr achar, Primzahlverteilung, Springer, Berlin 1957.
DEPARTMENT OF MATHEMATICS
UNIVERSITY OF GEORGIA
ATHENS, GEORGIA 30602
U.S.A.
Received on 26.6.1992
and in revised form on 1.2.1993 (2273)