See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/2434730

Squarefree Values Of Polynomials All Of Whose Coefficients Are 0 And 1

Article  in  Acta Arithmetica · September 1995

DOI: 10.4064/aa-74-3-191-205 · Source: CiteSeer

CITATIONS READS
15 29

2 authors, including:

Sergei Konyagin
Russian Academy of Sciences
245 PUBLICATIONS   3,560 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Sergei Konyagin on 20 December 2012.

The user has requested enhancement of the downloaded file.

 SQUAREFREE VALUES OF POLYNOMIALS ALL
OF WHOSE COEFFICIENTS ARE 0 AND 1

Michael Filaseta and Sergei Konyagin

1. Introduction

Let n be a non-negative integer and consider the set of polynomials

n
X
Sn = ff (x) = "j xj : "j 2 f0; 1g for each j and "0 = 1g:
j =0
The condition "0 = 1 ensures that the elements of Sn are not divisible by x. Let
1
[
S= Sn :
n=0
There are interesting open problems concerning the polynomials in S . Using the main
result in [1] (with base 2) or using the well-known explicit formula for the number of
irreducible polynomials of degree  n modulo 2, one can easily show that there are at least
on the order of 2n =n irreducible polynomials in Sn . Odlyzko (private communication) has
asked whether almost all polynomials in S are irreducible? In other words, does
lim jff (x) 2 Sn : f (x) is irreduciblegj = 1?
n!1 2n
It is not even known how to establish that the limit (or the limit supremum) is positive.
Another open problem, posed by Odlyzko and Poonen [2], is to determine whether it is
true that if  is a root with multiplicity > 1 of some polynomial f (x) in S , then  is a
root of unity.
The purpose of this paper is to establish two results concerning the polynomials in S .
First, we shall show
Typeset by AMS-TEX
1
Theorem 1. Let b = 3, 4, or 5. Then there are innitely many polynomials f (x) 2 S
for which f (b) is squarefree. Moreover, for such b, the density of polynomials f (x) 2 S for
which f (b) is squarefree is
Y ?1
(1) lim jff (x) 2 S n : f (b) is squarefreegj 6
= 2 1
1 ? p2 :
n!1 2n pjb

There are other trivial values of b for which one can obtain similar results (when jbj  2),
but we do not know how to establish the analogous results for b  6. As an immediate
consequence of Theorem 1, we deduce the

Corollary. Let b = 3, 4, or 5. There are innitely many squarefree numbers in base b

consisting only of the digits 0 and 1.

The arguments can be modied slightly to allow for the possibility that "0 = 0 in the
denition of Sn. Thus, for b = 3, 4, or 5, we can obtain the density of squarefree numbers
in base b among the positive integers consisting only of the digits 0 and 1 in base b. For
b = 4, the density is 1=2 times the expression on the right-hand side of (1); for b = 3 and
5, the density is 3=4 times the expression on the right-hand side of (1).
It is of some interest to know a corresponding result for base 10. By applying an
argument similar to what we will use for b = 4 in Theorem 1, it can be shown that there
are innitely many squarefree numbers which consist only of the digits 0, 1, and 2. In
fact, if d1 , d2 , and d3 are any three distinct digits not equal to 0, 4, and 8 in some order,
then there are innitely many squarefree numbers m in base 10 with each digit of m being
either d1 , d2 , or d3 . We will not address this issue further here.
Our second theorem concerns squarefree polynomials in S (polynomials without any
roots having multiplicity > 1). We shall see how to obtain the next result as a fairly direct
consequence of our approach to establishing Theorem 1.
2
Theorem 2. Almost all polynomials in S are squarefree. In other words,
lim jff (x) 2 Sn : f (x) is squarefreegj = 1:
n!1 2n
In the next section, we give a proof of Theorem 1 for the case that b = 3. In the process,
we will establish some preliminaries for the cases b = 4 and 5. The remainder of the proof
of Theorem 1 is given in Section 3. In Section 4, we will establish Theorem 2 using a
lemma (Lemma 9) which aided in the proof of Theorem 1.

2. Some Preliminaries and the Case b = 3

Let n be a positive integer. For integers b and m with m  2, we dene t(n) = t(n; m; b)

as the number of f (x) 2 Sn for which m divides f (b). We begin with an estimate for t(n).
Suppose rst that m and b are integers which are not relatively prime. Then there is a
prime p which divides both m and b. Observe that for every f (x) 2 Sn , we have f (b)  1
(mod p). Hence, for every f (x) 2 Sn , m does not divide f (b), and we deduce that t(n) = 0.
The next lemma deals with the remaining situation where m and b are relatively prime
integers.
Lemma 1. Let m and b be relatively prime integers with m  2. Then
n
t(n) = 2m (1 + o(1))

as n approaches innity.
Proof. Since
?1
mX 
m if mja
e
2 iaj=m =
j =0 0 otherwise,
we obtain
?1
mX ?1
mX
t(n) = m1 e2if (b)j=m = m1
X X
e2if (b)j=m :
f (x)2Sn j =0 j =0 f (x)2Sn
3
On the other hand, from the denition of Sn , we have
X n  
e if (b)j=m = e2ij=m Y 1 + e2ibk j=m :
2

f (x)2Sn k=1

Observe that when j = 0, the right-hand side is 2n . Hence,

n
t(n) = 2m + E;

where
?1
mX n 
E = m1
Y 
e2ij=m 1 + e2ib j=m :
k

j =1 k=1
It remains to show that E = o(2n ).
For each j 2 f1; 2; : : : ; m ? 1g, we rewrite the absolute value of the product above as
    
Yn   Yn  n  

 1+ e2ib k
j=m  =  eib k
j=m   Y
eib j=m + e?ib
k k
j=m 
   
k=1 k=1 k=1
n
Y 
= 2n cos(bk j=m) :
k=1

Since m and b are relatively prime and 1  j  m ? 1, the expression bk j=m is a rational
number which diers from an integer by at least 1=m. Therefore,
 
cos( bk j=m)  jcos(=m)j :

Since m  2, this last expression is < 1. We obtain

 
1 m
X?1 Yn  
2ibk j=m 
jE j  m 
 1+e 
j =1 k=1 

2n mX ?1 Yn  
=m cos(bk j=m)  2n jcos(=m)jn ;
j =1 k=1
and the lemma easily follows. 

4
Lemma 2. Let b be a positive integer, and let B be a real number > 0. Denote by S (B; n)
the number of f (x) 2 Sn such that f (b) is not divisible by p for every prime p  B . Then
2

 
1 ? p12 + o (2n ) :
Y
S (B; n) = 2n
pB; p b -

Lemma 2 follows from Lemma 1 by an easy sieve argument and we omit the details.
Observe that
Y

1

Y 1
 
6 Y 1 ?1  
1 ? p2 = 1 ? p2 (1 + O(1=B )) = 2 1 ? p2 (1 + O(1=B )) :
pB; p b
- pb - pjb
Fix " > 0. By choosing B suciently large and then choosing n suciently large, we
deduce from Lemma 2 that S (B; n) diers from
6  2n Y 1 ? 1 ?1
 

2 pjb p2
by  "2n . Thus, to prove Theorem 1, it suces to show that the number of f (x) 2 Sn
such that f (b) is divisible by p2 for some prime p > B is  "2n . For such an estimate we
may suppose that B is arbitrarily large; more specically, we can take B  B0 where B0
is an arbitrary constant depending only on ". The proof of Theorem 1 for the case b = 3
therefore follows from the following lemma.
Lemma 3. Let " > 0, and let B be suciently large. Then there are  "2n polynomials
f (x) 2 Sn for which there exists an integer d > B such that d jf (3). 2

Proof. Let d be an integer > B . Let r be the positive integer satisfying

3r=2 < d  3(r+1)=2 :

We x "r ; "r+1 ; : : : ; "n 2 f0; 1g arbitrarily and consider f (x) = nj=0 "j xj 2 Sn . Observe
P

that for any choice of "0 ; "1 ; : : : ; "r?1 2 f0; 1g, we have
r?1
X
0 "j 3j < d2 :
j =0
5
Also, for distinct choices of the r?tuple ("0 ; "1 ; : : : ; "r?1 ) with each "j 2 f0; 1g, the num-
Pr?1 j
bers j =0 "j 3 are distinct; hence, they are distinct modulo d2 . We deduce that with
"r ; "r+1 ; : : : ; "n 2 f0; 1g xed, there is at most one choice of ("0 ; "1 ; : : : ; "r?1 ) such that
f (3) is divisible by d2 . It follows that there are at most 2n?r+1 choices for f (x) 2 Sn such
that f (3) is divisible by d2 . The inequality 3(r+1)=2  d > B implies that r is large. Hence,

2n?r+1 = 2n+1 2?r = 2n+1 (3r=2 )?2 log 2= log 3  2n+1 (3(r+1)=2 )?5=4  2n+1 d?5=4 :

We deduce that the number of f (x) 2 Sn such that f (3) is divisible by d2 for some integer
d > B is
X
 2n+1 d?5=4 :
d>B
Since B is suciently large and 1
P ?5=4 converges, we deduce that this last expression
d=1 d
is  "2n , completing the proof of the lemma. 

3. The Cases b = 4 and b = 5

In this section, we complete the proof of Theorem 1. We will improve on the argument
given for Lemma 3 to obtain the desired result. We note that the work in this section
allows us also to handle the case b = 3 here, but we have chosen to indicate the proof of
the case b = 3 separately in the previous section partially because of its simplicity and
partially because the case b = 3 of Theorem 1 by itself can be used to obtain Theorem 2
(see Section 4).
As in the previous section, we x " > 0 and consider B to be suciently large. Analogous
to Lemma 3, we want to show for b = 4 and b = 5 that the number of f (x) 2 Sn such that
f (b) is divisible by d2 for some d > B is  "2n .
6
 For b  3, we dene
8 9
<X1 =
S (b) = : "j bj : "j 2 f0; 1g; all but nitely many "j are 0;
j =0
and

S 0 = S 0 (b) = fm1 ? m2 : m1 ; m2 2 S (b); m1 > m2 g

8 9
<X1 =
=: "j bj 2 Z
+
: "j 2 f?1; 0; 1g; all but nitely many "j are 0; :
j =0
For r and t positive integers, we consider the set

X (r; t) = X (r; t; b) = fu 2 \ [br?1 ; br ) : gcd(b; u) = 1 and tu2 2 S 0 g:

Z

The next several lemmas serve to estimate the size of X (r; t). In the end, we will need a
more intricate estimate for the case b = 5 than for the case b = 4; in particular, for the
case b = 5, we will need to strengthen our next lemma which is a preliminary bound on
jX (r; t)j.
Lemma 4. Let b  3, r  2, and t  1 be integers. Then
jX (r; t)j  3r+1 b2:

Proof. For any positive integers m and s, m is in S 0 if and only if bs m is in S 0 . Thus, we

may suppose that b t, and we do so. We may also suppose that jX (r; t)j 6= 0. Let u be
-

in X (r; t). Then tu2 is in S 0 . By the denition of S 0 , an element of S 0 is either relatively

prime to b or it is divisible by b. Thus, the conditions gcd(b; u) = 1 and tu2 2 S 0 imply
gcd(b; t) = 1.
We write
1
X
tu =2
k bk ;
k=0
7
where each k = k (u) is in f?1; 0; 1g. There are 3r+1 b2 dierent values for the (r +
2)?tuple (u0 ; 00 ; 01 ; : : : ; 0r ) where u0 is a non-negative integer < b2 and 0k 2 f?1; 0; 1g
for k 2 f0; 1; : : : ; rg. Consider a xed such (r + 2)?tuple. The lemma will follow if we can
show that there is at most one u 2 X (r; t) for which u  u0 (mod b2 ) and k (u) = 0k for
every k 2 f0; 1; : : : ; rg.
Let u and v be in X (r; t) with u  v (mod b2 ) and k (u) = k (v) for every k 2
f0; 1; : : : ; rg. We want to show that u = v. Let p be a prime divisor of b. Then gcd(b; u) = 1
implies p u. Since gcd(u ? v; u + v) = gcd(u ? v; 2u), we deduce that if p divides both
-

u ? v and u + v, then p = 2. Also, u  v (mod b2 ) implies p2 j(u ? v) so that in the case

p = 2, we have 4 (u + v). Since gcd(b; t) = 1, it follows that gcd(br+1 ; t(u + v)) is either 1
-

or 2 and, hence, divides b. The condition k (u) = k (v) for every k 2 f0; 1; : : : ; rg implies
?

br+1 j tu2 ? tv2 . We deduce br j(u ? v). The conclusion u = v now follows since u and v
are positive integers < br . 
Lemma 5. Let j and s be positive integers. Let K be a set of s?tuples ( ; : : : ; s ) 1

satisfying the two conditions:

(i) For each i 2 f1; 2; : : : ; sg, i 2 f1; 2; 3g.
(ii) For each i 2 fj + 1; j + 2; : : : ; sg, if i?j 2 f2; 3g, then i 2 f1; 2g.
Then

3
j 
p s
jK j  p 1+ 2 :
1+ 2
Proof. For each t 2 f1; 2; : : : ; j g, consider the elements (1 ; : : : ; s ) of K and dene Kt as
the set of [(s ? t + j )=j ]?tuples (t ; j +t ; : : : ; [(s?t)=j ]j +t ). Thus, jK j  jt=1 jKt j. Also,
Q

observe that the number of components in each element of K is the sum over t of the
number of components in each element of Kt . In other words,
j 
(2) s=
X s ? t+ j:
t=1 j
8
Fixing t 2 f1; 2; : : : ; j g, we consider the elements ( 1 ; 2 ; : : : ; [(s?t+j )=j ]) of Kt . For each
i 2 f1; 2; : : : ; [(s?t+j )=j ]g, we dene Ni as the number of dierent choices for 1 ; 2 ; : : : ; i
which arise. In other words, Ni is the number of i?tuples ( 1 ; 2 ; : : : ; i ) obtained from
the rst i components of the elements of Kt . Thus, jKt j = N[(s?t+j )=j ]. By condition (i),
N1  3. By conditions (i) and (ii), N2  7 (there are  3 choices for ( 1 ; 2 ) with 1 = 1
and  4 choices for ( 1 ; 2 ) with 1 2 f2; 3g). Fix i 2 f3; 4; : : : ; [(s ? t + j )=j ]g. Let M be
the number of (i ? 1)?tuples ( 1 ; 2 ; : : : ; i?1 ) with i?1 = 1. Observe that M  Ni?2 .
By condition (i), there are  3M possible i?tuples ( 1 ; 2 ; : : : ; i ) with i?1 = 1. On the
other hand, by condition (ii), there are  2(Ni?1 ? M ) possible i?tuples ( 1 ; 2 ; : : : ; i )
with i?1 2 f2; 3g. Therefore,

Ni  3M + 2(Ni?1 ? M ) = 2Ni?1 + M  2Ni?1 + Ni?2 :

p
Recall that N1  3 and N2  7. An easy induction argument now gives Ni  3(1+ 2)i?1 .
Thus,
p 
3p

p [(s?t+j )=j ]
jKt j = N[(s?t+j)=j]  3(1 + 2) s?t)=j] =
[(
1+ 2 :
1+ 2
The lemma now follows from jK j  jt=1 jKt j and (2). 
Q

Lemma 6. Let b be an odd integer  5, and let r and j be positive integers with j  r.
Let a and t be positive integers and suppose that bj jja. Then the number of positive
integers u < br with gcd(b; u) = 1 and such that both tu2 and t(u + a)2 are in S 0 is
p
 (b ? 1)3j (1 + 2)r?j .
Proof. As in the proof of Lemma 4, we may suppose that gcd(b; t) = 1 and do so. Let u
be as in the statement of the lemma. Let

D(u) = t(u + a)2 ? tu2 = ta(2u + a):

9
Since tu2 and t(u + a)2 are in S 0 , we have
1
X 1
X
(3) tu =
2
k bk and t(u + a) =
2
k bk
k=0 k=0
for some integers k and k in f?1; 0; 1g. We write
r?1
X 1
X
(4) u= uk bk and D(u) = dk bk
k=0 k=0
where, for each non-negative integer k, uk 2 [0; b ? 1] and

(5) dk = k ? k 2 [?2; 2]:

Note that since b  5, D(u) has a unique representation as in (4) with dk 2 [?2; 2].
Suppose now that v is a positive integer < br with v 6= u and gcd(b; v) = 1 and such that
both tv2 and t(v + a)2 are in S 0 . Let ` be the non-negative integer satisfying b` jj(v ? u).
Then D(v) ? D(u) = 2ta(v ? u) so that

b`+j jj (D(v) ? D(u)) :

Viewing the numbers u0 ; u1 ; : : : ; u`?1 in (4) as xed, we deduce that the numbers d0 ; d1 ; : : : ;
d`+j ?1 are uniquely determined. Furthermore, the number u` uniquely determines the
value of d`+j and dierent values of u` lead to dierent values of d`+j . In particular, there
is at most one choice of u` which leads to d`+j = 0. We refer to such a choice of u` as
\nice."
We keep the notation above and still view u0 ; u1 ; : : : ; u`?1 as xed. Suppose that `  1.
Since b is an odd integer relatively prime to tu, we obtain that gcd(b; t(u + v)) = 1 so
?

that b` jj tv2 ? tu2 . Hence, the numbers 0 ; 1 ; : : : ; `?1 in (3) are uniquely determined.
Dierent values of u` lead to dierent values of ` . We are interested only in u for which
tu2 2 S 0 so that ` 2 f?1; 0; 1g. Therefore, there are at most 3 dierent values of u` such
that tu2 2 S 0 .
10
 Since ` and ` are in f?1; 0; 1g, for each d` 2 f?2; ?1; 0; 1; 2g, there are at most 3 ?jd` j
values of ` such that (5) holds. In particular, we deduce that if `  j and u`?j is not nice
(so that d` 6= 0), then there are at most two values of ` , and hence at most two values of
u` , for which tu2 and t(u + a)2 are both in S 0 .
Since b u, there are at most b ? 1 choices for u0 in (4). Fix u0 and consider the choices
-

for u1 ; : : : ; ur?1 as in (4) with u as in the lemma. For ` 2 f1; 2; : : : ; r ? 1g and for any
given u1 ; : : : ; u`?1 , there are at most 3 dierent values of u` , say i = i (u0 ; u1 ; : : : ; u`?1 )
where i is a positive integer  3. At most one such u` is nice, and if such a choice
of u` exists we can suppose that it is 1 and do so. We dene ` (u` ) = i where i 2
f1; 2; 3g with u` = i . Observe that u in (4) is uniquely determined by the value of
(1 (u1 ); 2 (u2 ); : : : ; r?1 (ur?1 )) (where we are still viewing u0 as xed). Also, if ` 2
fj + 1; j + 2; : : : ; r ? 1g and `?j (u`?j ) 2 f2; 3g (so that u`?j is not nice), then ` (u`)  2.
Thus, the set of (r ? 1)?tuples (1 (u1 ); : : : ; r?1 (ur?1 )) satises the conditions of the set
K in Lemma 5 with s = r ? 1. Recalling that there are  b ? 1 choices for the value of
u0 , we deduce that the number of u < br with gcd(b; u) = 1 and such that both tu2 and
t(u + a)2 are in S 0 is

3
j 
p r?1 p
 (b ? 1) p 1+ 2 < (b ? 1)3j (1 + 2)r?j ;
1+ 2
establishing the lemma. 
Lemma 7. Let b be a positive integer  3. Let r and ` be positive integers with 1  `  r.
Let t be a positive integer. Then there exist 3r?`+2 intervals each of length < 2b` with the
union of these intervals containing all numbers u for which br?1  u < br and tu2 2 S 0 .
Proof. Let s be the positive integer satisfying
bs?1 < t  bs :
b?1 b?1
11
For u < br and tu2 2 S 0 , we obtain
r s?
2 X
+ 1
tu =
2
k bk for some k 2 f?1; 0; 1g:
k=0
Fix k for r + s + ` ? 2  k  2r + s ? 1. Let
r s?
2 X
+ 1 s+`?3
r+X r s?
2 X
+ 1 s+`?3
r+X
= k bk ? bk and = k bk + bk :
k=r+s+`?2 k=0 k=r+s+`?2 k=0
For br?1  u < br and tu2 2 S 0 , we deduce that tu2 is in some such [;  ] so that u 2 [ ;  ]
where
hp p i  
[ ;  ] = =t; =t \ br?1 ; br :

Observe that
s+`?3
r+X r+s+`?2
? = 2 bk < 2b b ? 1 :
k=0
Therefore,
p
 ?  =t ? =t =
p  ? p p
t( =t + =t)
?    ?  < 2br+s+`?2=(b ? 1) = 2b`:
<  t tbr?1 br+s?2=(b ? 1)
Hence, the 3r?`+2 choices for r+s+`?2 ; : : : ; 2r+s?1 , each in f?1; 0; 1g, lead to 3r?`+2
intervals [ ;  ] of length < 2b` satisfying the conditions of the lemma. 

Since b  3, it is not dicult to check that the intervals in the proof of Lemma 7 above
are disjoint. On the other hand, it is already clear in the statement of Lemma 7 that we
may consider these intervals to be disjoint.

Lemma 8. Let b be an odd integer  5. Let r and t be positive integers. Then

p !
jX (r; t)j  exp log 3(log b + log(1 + 2)) r ;
log(3b)
12
where the implied constant depends on b but not on r or t.
Proof. Consider an arbitrary positive integer `  r. By Lemma 7, X (r; t) is contained in
the union of 3r?`+2 disjoint intervals [ i ; i ], with 1  i  3r?`+2 , where each interval is
of length < 2b` . For each i 2 f1; 2; : : : ; 3r?`+2 g and k 2 f1; 2; : : : ; b ? 1g, we set

Xi;k (r; t) = fu 2 X (r; t) : u 2 [ i ; i ] and u  k (mod b)g:

Let ni;k = jXi;k (r; t)j. Then
? +2 b?1 r?`+2
b?1
ni;k (ni;k ? 1) = 3 X X
r `
3X X
2 jf(u; v) : u 2 Xi;k (r; t); v 2 Xi;k (r; t); and u < vgj
i=1 k=1 i=1 k=1
? +2 b?1
3rX`
X X
= jf(u; v) : u 2 Xi;k (r; t); v 2 Xi;k (r; t); and v ? u = agj
1a< b i=1 k=1
2 `
bja
X
 jf(u; v) : u 2 X (r; t); v 2 X (r; t); and v ? u = agj :
a< b
1 2 `
bja
From Lemma 6, we now deduce that
? +2 b?1
3rX
ni;k (ni;k ? 1)  X̀ X (b ? 1)3j (1 + p2)r?j
`
X

i=1 k=1 2 j =1 1a<2b`

bj jja
X̀ p p
 2b`?j (b ? 1)3j (1 + 2)r?j  b` (1 + 2)r :
j =1
Therefore,
? +2 b?1 ? +2 b?1  
3rX
?
`
3rX`

jX (r; t)j =
X
ni;k 
X
1+ n i;k (n i;k 1)
i=1 k=1 i=1 k=1 2
? +2 b?1 ? +2 b?1
=
3r `
X X
1+
3r `
X X ni;k (ni;k ? 1)  3r?` + b` (1 + p2)r :
i=1 k=1 i=1 k=1 2
We choose " p #
` = (log 3 ?log(3
log(1 + 2)) r + 1
b)
to obtain the lemma. 

13
Lemma 9. Let b = 4 or 5. Let " > 0, and let B = B(") be suciently large. Then the
number of f (x) 2 Sn such that f (b) is divisible by d for some integer d > B is  "2n .
2

Proof. Since B is suciently large, the number of f (x) 2 Sn as in the lemma is 0 unless n is
also large. We therefore consider n large. Let r be a positive integer for which br > B . We
consider the integers d such that br?1  d < br . For f (x) 2 Sn , we have 0 < f (b)  bn+1
so that if f (b) is divisible by d2 (which is  b2r?2 ), then r  (n + 3)=2. We therefore
suppose, as we may, that r  (n + 3)=2.
Recall that each f (x) 2 Sn has constant term 1 so that if f (b) is divisible by d2 , then
gcd(b; d) = 1. If f (b) = td2 , then we also have that 1  t = f (b)=d2  bn?2r+3 so that
d 2 X (r; t) for some positive integer t  bn?2r+3 . We use Lemmas 4 and 8 to obtain that
the number of f (x) 2 Sn for which there exists a d 2 [br?1 ; br ) such that d2 jf (b) is
8
n?2r 3r
?2r+3
bnX <4
>
p ! for b = 4
 jX (r; t)j  > 5n?2r exp log 3(log 5 + log(1 + 2)) r for b = 5:
t=1 : log 15
In either case, if r > n=(2:4), the above expression on the right is easily  2n =(nB ). We
restrict our attention now to r  n=(2:4). We note that our method for obtaining this
bound on r is not the best possible, and it would be easy to replace 2:4 with a larger
number; however, 2:4 will be sucient for what follows.
Pn
Let s denote a positive integer  n ? 2r. We consider f (x) = j =0 "j x 2 Sn with
j

"2r+s?2 ; "2r+s?1 ; : : : ; "n xed elements from f0; 1g. Thus, we obtain 22r+s?3 dierent val-
ues of f (b). Let N (d) denote the number of dierent (2r + s ? 3)-tuples ("1 ; "2 ; : : : ; "2r+s?3 ),
with each "j 2 f0; 1g, such that d2 jf (b). Suppose N (d)  1. Consider the f (x) counted
by N (d), and let f1 (x) denote the f (x) which minimizes the value of f (b). Then there are
N (d) ? 1 other f (x) counted by N (d) each having the property that d2 jf (b). For each of
these N (d) ? 1 dierent f (x), we obtain
0 < f (b) ? f1 (b)  b2r+s?2  d2 bs :
14
Thus, there are at least N (d) ? 1 dierent f (x) 2 Sn (with "2r+s?2 ; "2r+s?1 ; : : : ; "n xed)
such that f (b) ? f1 (b) = td2 for some positive integer t  bs . Dierent choices for f (x)
give dierent values for t. We deduce that there are at least N (d) ? 1 dierent t  bs for
which d 2 X (r; t).
With "2r+s?2 ; "2r+s?1 ; : : : ; "n still xed, we bound the number of f (x) 2 Sn such that
there is a d 2 [br?1 ; br ) for which d2 jf (b). This number is
X X X
 N (d) = (N (d) ? 1) + 1:
br?1 d<br b ?1 d<br
r
b ?1 d<br
r
N (d)1 N (d)1
From our comments above and from Lemmas 4 and 8, we deduce that
X X X X X
(N (d) ? 1)  1= 1
b ?1 d<br
r b ?1 d<br 1tbs
r 1tb b ?1 d<br
s r
N (d)1 d2X (r;t) d2X (r;t)
8
X
>
< 4s 3r p ! for b = 4
= jX (r; t)j  > 5s exp log 3(log 5 + log(1 + 2)) r for b = 5:
tbs
1 : log 15
Also,
X
1  br :
br?1 d<br
N (d)1
Letting "2r+s?2 ; "2r+s?1 ; : : : ; "n now vary, we deduce that the number of f (x) 2 Sn such
that there exists a d 2 [br?1 ; br ) for which d2 jf (b) is

 2n?2r?s4s 3r + 2n?2r?s4r for b = 4

and
p !
 2n?2r?s5s exp log 3(log 5 +loglog(1 + 2)) r + 2n?2r?s 5r for b = 5:
15
In the case b = 4, we choose
 
s = r log(4=3) + 1;
log 4
15
and in the case b = 5, we choose
" p !#
s = logr 5 log 5 ? log 5 +log
log(1 + 2) (log 3)
15 + 1:

It is easily checked that since 1  r  n=(2:4), in either case the choice of s is a positive
integer  n ? 2r. We obtain that the number of f (x) 2 Sn such that f (b) is divisible by
some d2 with br?1  d < br is

2n exp(?0:14r) for b = 4
 2n?2r?sbr  2n exp(?0:034r) for b = 5:
In either case, b = 4 or b = 5, since e2 > b, the above bound is  2n e?2r=100  2n b?r=100 .
Letting r vary over the positive integers for which br > B , we easily obtain now that
the number of f (x) 2 Sn such that f (b) is divisible by d2 for some d > B is  2n B ?1=100 .
Since B is suciently large, the proof of the lemma is complete. 

4. The Proof of Theorem 2

Let R be a xed real number  1. We begin by estimating the number of f (x) 2 Sn

divisible by the square of a non-constant polynomial in [x] of degree  R. We will show
Z

that there are o(2n ) such f (x).

Odlyzko and Poonen [2] have obtained extensive results about the roots of polynomials
in Sn . For our purposes, it suces to know that these roots are bounded in absolute value
Pm
by 2 which is easily established as follows. Let f (x) 2 Sn, and write f (x) = j =0 "j x
j

where m  n, "j 2 f0; 1g for each j , and "0 = "m = 1. If  2 and jj  2, then C

 
 m  ?1
mX m
jjj = jjm ? jjj j ??11
X j 
jf ()j  
 "j    jjm ?
 j =0  j =0

= jj
m+1 ? 2jjm + 1
jj ? 1 = (jj ?j2)j ?
jjm + 1 > 0:
1
16
Thus, f () 6= 0, and we deduce that all roots of the polynomials in Sn necessarily have
absolute value < 2.
Let g(x) 2 [x] of degree r 2 [1; R], and suppose that g(x) is a factor of some polynomial
Z

in Sn . It follows that the roots of g(x) are < 2. Also, since polynomials in Sn are monic,
the leading coecient of g(x) must be 1. Since the degree of g(x) is  R, it follows that
each coecient of g(x) has absolute value less than or equal to the product of 2R (an upper
bound on the absolute value of the product of the roots of g(x)) and 2R (an upper bound
on the number of combinations of r  R roots taken k at a time where k 2 f0; 1; : : : ; rg).
Since the absolute value of the coecients of g(x) are bounded by 4R and since g(x) has
degree  R, there are
 2  4R + 1 R+1
?

dierent possible values of g(x). To establish what we rst set out to show, it suces then
to obtain that for each such g(x), there are o(2n ) dierent possible f (x) 2 Sn divisible by
g(x)2 .
Fix g(x) as above. Suppose that f (x) = nj=0 "j xj 2 Sn is divisible by g(x)2 . We
P

consider the set Tn (f (x)) consisting of the polynomials w(x) = nj=0 "0j xj 2 Sn where
P

there is exactly one k 2 f1; 2; : : : ; ng for which "0k 6= "k . In other words, w(x) =
Pn 0 j 0̀
j =0 "j x 2 Tn (f (x)) if and only if there is a k 2 f1; 2; : : : ; ng such that " = "` for
every ` 2 f0; 1; : : : ; ng with ` 6= k and "0k = 1 ? "k . Thus, jTn (f (x))j = n. Since f (x)
is divisible by g(x)2 and f (x) has constant term 1, it must be the case that g(x) is not
divisible by x. If w(x) = nj=0 "0j xj 2 Tn (f (x)) and k 2 f1; 2; : : : ; ng with "0k 6= "k , then
P

f (x) ? w(x) = xk is not divisible by g(x)2 . We deduce that the elements of Tn (f (x)) are
not divisible by g(x)2 .
Now, suppose that f1 (x) and f2 (x) are distinct polynomials in Sn with each divisible
by g(x)2 . We show that Tn (f1 (x)) and Tn (f2 (x)) are disjoint. If the sets were not disjoint,
17
then there would be some w(x) which diers from each of f1 (x) and f2 (x) by a power of x.
By considering f1 (x) ? f2 (x), it follows that for some k and ` in f1; 2; : : : ; ng with k > `,
xk  x` = x` (xk?`  1) is divisible by g(x)2 . Since the roots of xk?`  1 are distinct and
since g(x) is not divisible by x, we deduce that g(x)2 cannot divide x` (xk?`  1). Hence,
Tn (f1 (x)) and Tn (f2 (x)) are disjoint.
For each f (x) 2 Sn divisible by g(x)2 , there correspond n polynomials, namely the
elements of Tn (f (x)), which are not divisible by g(x)2 , and these n polynomials are dierent
for dierent f (x). Thus, there are  2n =(n+1) polynomials in Sn divisible by g(x)2 . Hence,
there are o(2n ) polynomials in Sn divisible by g(x)2 and thus o(2n ) polynomials f (x) 2 Sn
which are divisible by the square of a polynomial of degree  R.
Fix " > 0. It suces to show that if R is suciently large, then there are  "2n
polynomials f (x) 2 Sn which are divisible by the square of a polynomial in [x] of degree
Z

> R. We will use Theorem 1 with b = 4 and the fact already established that the roots
of the polynomials in Sn have absolute value < 2. We note, however, the case b = 3
of Theorem 1 could be used instead of the case b = 4 if we use that the roots of the
polynomials in Sn have real parts < 1:5 (cf. [1] or [2]).
Let f (x) 2 Sn with f (x) divisible by the square of a polynomial g(x) 2 [x] of degree
Z

r > R. We may suppose that g(x) is monic (otherwise, replace g(x) with ?g(x)). Then
the roots of f (x) and hence g(x) have absolute value < 2. If 1 ; : : : ; r denote the roots
of g(x), then g(x) = rj=1 (x ? j ) and
Q

r
Y
jg(4)j = j4 ? j j  2r > 2R :
j =1

Since f (x) is divisible by g(x)2 , we deduce that f (4) is divisible by d2 for some integer
d > 2R . On the other hand, from Lemma 9 with b = 4, we obtain that for R suciently
large, there are  "2n such polynomials f (x) 2 Sn . Hence, Theorem 2 follows.
18
 Acknowledgments: Research of the rst author was supported by NSA Grant MDA904-
92-H-3011 and NSF Grant DMS-9400937. Research of the second author was supported by
a grant from the Cultural Initiative Fund and the Russian Academy for Natural Sciences
and by Grant MC5000 from the International Science Foundation. The authors also grate-
fully acknowledge NSF EPSCoR Grant EHR 9108772 and ONR Contract N0014-91-51343
which provided support for the second author while he visited the University of South
Carolina where this research took place.

References
1. J. Brillhart, M. Filaseta, and A. Odlyzko, On an irreducibility theorem of A. Cohn,
Can. J. Math. 33 (1981), 1055{1059.
2. A. M. Odlyzko and B. Poonen, Zeros of polynomials with 0, 1 coecients, L'Enseigne-
ment Mathematique 39 (1993), 317{348.

Michael Filaseta Sergei Konyagin

Mathematics Department Department of Mechanics and Mathematics
University of South Carolina State University
Columbia, SC 29208 Moscow 119899
U.S.A. Russia
[email protected] [email protected]

19

View publication stats