1 Introduction
It is known that if for every n, the sequence $\binom {2n}{n}$ can be computed in $O(\log ^k{n})$ arithmetic operations for a fixed constant k, then integers can be factored efficiently [Reference Lipton3, Reference Shamir5]. We ask if there exist linearly recurrent sequences which contain many small factors like $\binom {2n}{n}$ . If such sequences exist, they can be used instead of $\binom {2n}{n}$ to factor integers. This is because the nth term of any linearly recurrent sequence can be computed in $O(\log {n})$ arithmetic operations using repeated squaring of the companion matrix [Reference Bostan and Mori1]. We first set up some notation to formally state our question.
Let $P(n)$ be the largest prime factor of n and $s_y(n)$ be the largest divisor d of n with $P(d)\le y$ . Thus, $s_y(n)$ is the y-smooth part of n. Given a sequence ${\textbf {u}}=(u_n)_{n\ge 0}$ of positive integers we ask whether we can find $c>1$ and K such that
contains many elements. For example, if $u_n=\binom {2n}{n}$ is the sequence of middle binomial coefficients, then ${\mathcal A}_{2,2,{\textbf {u}}}$ contains all the positive integers. The main question we tackle in this paper can be formally stated as follows.
Question 1.1 Does there exist a linearly recurrent sequence u such that ${\mathcal A}_{K,c,{\textbf {u}}}$ is infinite?
Here, we address the problem in the simplest case namely $u_n=a^n-1$ for some positive integer a. Our results are easily extendable to all Lucas sequences, in particular, the sequence of Fibonacci numbers.
To start we recall the famous $ABC$ -conjecture. Put
for the algebraic radical of n.
Conjecture For all $\varepsilon>0$ there exists a constant $K_{\varepsilon }$ such that whenever $A,B,C$ are coprime nonzero integers with $A+B=C$ , then
Throughout this paper, $a>1$ is an integer and $u_n=a^n-1$ .
We have the following result.
Theorem 1.1 Assume the $ABC$ conjecture. Then for any $K>0$ , $c>1$ , the set ${\mathcal A}_{K,c,{\textbf {u}}}$ is finite.
One can ask what can one prove unconditionally. Maybe we cannot prove that ${\mathcal A}_{K,c,{\textbf {u}}}$ is finite but maybe we can prove that it is thin, that is that it does not contain too many integers. This is the content of the next theorem.
Theorem 1.2 We have
In particular, if one wants to find for all large N an interval starting at N of length k, that is $[N+1,\ldots ,N+k]$ which has nonempty intersection with ${\mathcal A}_{K,c,{\textbf {u}}}$ then infinitely often one should take $k>\exp (\log N/(157\log \log N))$ . But if the $ABC$ conjecture is true, one will no longer find elements of ${\mathcal A}_{K,c,{\textbf {u}}}$ in the above interval for large N no matter how large k is. Regarding Theorem 1.2, see [Reference Shamir6] for a more general result which applies to any linearly recurrent sequence but which gives a slightly weaker bound when specialised to our sequence u.
2 Proofs
2.1 The proof of Theorem 1.1
We apply the $ABC$ conjecture to the equation
for $n\in {\mathcal A}_{K,c,{\textbf {u}}}$ with the obvious choices. Note that
We then have
We may of course assume that $1<c<a$ . Then
We choose $\varepsilon>0$ small enough so that $\log a-(1+\varepsilon )\log (a/c)>0$ . Then, we get
The next lemma shows that the left–hand side above is $\le n^{2/3+o(1)}$ as $n\to \infty $ . This is unconditional and finishes the proof of Theorem 1.1.
Lemma 2.1 We have
as $n\to \infty $ .
Proof Let $\ell _p$ be the order of a modulo p; that is the smallest positive integer k such that $a^k\equiv 1\pmod p$ . Since primes p participating in $S_{K,a}(n)$ have $p\mid a^n-1$ , it follows that $\ell _p\mid n$ . Since also such primes are $O(n)$ , it follows that
where $P_{K,a}(n):=\{p\le Kn: \ell _p\mid n\}$ . To estimate $P_{K,a}(n)$ we fix a divisor d of n and look at primes $p\le Kn$ such that $\ell _p=d$ . Such primes p have the property that $p\equiv 1\pmod d$ by Fermat’s Little Theorem. In particular, the number of such (without using results on primes in progressions) is at most
However, since these primes divide $a^d-1$ , the number of them is $O(d)$ . Thus, for a fixed d the number of such primes is
Summing this up over all divisors d of n we get that
as $n\to \infty $ , where we used $d(n)$ for the number of divisors of n and the well-known estimate $d(n)=n^{o(1)}$ as $n\to \infty $ (see Theorem 315 in [Reference Hardy, Wright, Heath-Brown and Silverman2]). Hence,
as $n\to \infty $ , which is what we wanted.□
Remark 2.2 The current Lemma 2.1 was supplied by the referee. Our initial statement was weaker. The combination between Lemma 2.1 and estimate (2.1) shows that we can even take K growing with n such as $K=n^{1-\varepsilon }$ in the hypothesis of Theorem 1.1 and retain its conclusion. This has been also noticed in [Reference Murty and Wong4].
2.2 The proof of Theorem 1.2
It is enough to prove an upper bound comparable to the upper bound from the right–hand side of (1.1) for $\#({\mathcal A}_{K,c,{\textbf {u}}}\cap (N/2,N])$ as then we can replace N by $N/2$ , then $N/4$ , etc. and sum up the resulting inequalities. So, assume that $n\in (N/2,N]$ . We estimate
On the one hand, since $s_{KN}(u_n)\ge s_{Kn}(u_n)\ge c^{n}\ge c^{N/2}$ for all $n\in {\mathcal A}_{K,c,{\textbf {u}}}$ , we get that
Next, writing $\nu _p(m)$ for the exponent of p in the factorisation of m, we have
Let $o_p:=\nu _p(u_{\ell _p})$ . It is well-known that if p is odd then
(see, for example, (66) in [Reference Shparlinski7]). In particular, if $p\mid u_n$ , then $p^{o_p}\mid u_n$ . Furthermore, for each $k\ge 0$ , the exact power of p in $u_n$ is $o_p+k$ if and only if $\ell _p p^k$ divides n and $\ell _p p^{k+1}$ does not divide n. When $p=2$ , we may assume that a is odd (otherwise $\nu _2(u_n)=0$ for all $n\ge 1$ ), and the right–hand side of the above formula needs to be ammended to
Thus, for odd p,
A similar formula holds for $p=2$ . In particular, for $p=2$ , we have
Thus, the prime $p=2$ contributes a summand of size $O(N)$ to the right-hand side of (2.2). From now on, we assume that p is odd. The first cardinality in the right-hand side of formula (2.3) above is
The remaining cardinalities on the right-above can be bounded as
Thus,
We thus get
It remains to bound S. Since $p^{o_p}\mid a^{\ell _p}-1$ , we get that $p^{o_p}<a^{\ell _p}$ so $o_p\log p\ll \ell _p$ . Hence,
We get the first nontrivial upper bound on $\#({\mathcal A}_{K,c,{\textbf {u}}}\cap (N/2,N]$ , namely
so
To do better, we need to look more closely at $o_p\log p/\ell _p$ for primes $p\le KN$ . We split the sum S over primes $p\le KN$ in two subsums. The first is over the primes in the set $Q_1$ consisting of p such that $o_p\log p/\ell _p<1/y_N$ , where $y_N$ is some function of N which we will determine later. We let $Q_2$ be the complement of $Q_1$ in the set of primes $p\le Kn$ . The sum over primes $p\in Q_1$ is
For $Q_2$ , we use the trivial estimate
and it remains to estimate the cardinality of $Q_2$ . Note that $Q_2$ consists of primes p such that $o_p>\ell _p/(y_N\log p)\gg \ell _p/(y_N\log N)$ . We put $\ell _p$ in dyadic intervals. That is $\ell _p\in (2^i,2^{i+1}]$ for some $i\ge 0$ . Then primes $p\le KN$ in $Q_2$ with such $\ell _p$ have the property that $o_p\gg 2^i/(y_N\log N)$ . Hence,
which gives
Summing up over all the i, we get
where I is maximal such that $(2^I,2^{I+1}]$ contains an element p of $Q_2$ . By a result of Stewart (see Lemma 4.3 in [Reference Shparlinski7]),
Thus,
Choosing $y_N:={\displaystyle {\exp \left (c\frac {\log N}{\log \log N}\right )}}$ with a positive constant c to be determined later, we get
Choosing $c:=1/156$ , we get
which is what we wanted.
Acknowledgement
We thank the referee for suggesting the current Lemma 2.1 with its proof and for pointing out reference [Reference Murty and Wong4] and Professor Igor E. Shparlinski for pointing out reference [Reference Shparlinski6]. F.L. worked on this paper while visiting the Max Planck Institute for Software Systems in Saarbrücken, Germany in Fall of 2020. F.L. thanks the Institute for hospitality and support.