9
$\begingroup$

Let $p$ be a prime number. For every integer $m$, there are integers $u_1$, $u_2$, such that $\lvert u_1\rvert, \lvert u_2\rvert < \sqrt{p}$ and $$m \equiv u_1u_2^{-1} \pmod{p}.$$

Proof of this statement is at the end. I'm interested in generalizing this statement. For example,

Let $p$ be a prime number, $f = \sqrt[3]{p}$. For every integer $m$, there are integers $u_1$, $u_2$, $u_3$, such that $\lvert u_1\rvert, \lvert u_2\rvert, \lvert u_3\rvert < f$ and $$m \equiv u_1^{\pm 1}u_2^{\pm 1}u_3^{\pm 1} \pmod{p}.$$

But this is false. The question is, how large should $f$ be to make the statement true? Can we choose $f = c\sqrt[3]{p}$ for some constant $c$? Probably, no. I think, something like $f = \log(p)\sqrt[3]{p}$ should work. The same question for the products of the form $u_1^{\pm 1}u_2^{\pm 1}\dotsm u_n^{\pm 1}$.

Proof.

Consider the pairs $(a, b)$ such that $0 \le a, b < \sqrt{p}$. There are $(\lfloor\sqrt{p}\rfloor + 1)^2 > p$ such pairs. Hence, we can choose two distinct pairs $(a_1, b_1)$, $(a_2, b_2)$ such that $$a_1m + b_1 \equiv a_2m + b_2 \pmod{p}.$$ Let $u_2 = a_1 - a_2$, $u_1 = b_2 - b_1$. Then $$m \equiv u_1u_2^{-1} \pmod{p}.$$ I tried to generalize this proof, but did not succeed.

Update

This article deals with a similar problem. Theorem 1 implies that $f = p^{\beta + o(1)}$ works if $n \ge 14$, where $\beta = \frac{1}{4}e^{-1/2} \approx 0.1516$. This result is the best possible (in term of $\beta$) until at least the Burgess bound on the smallest quadratic nonresidue is improved.

$\endgroup$
6
  • 5
    $\begingroup$ Interestingly, no matter what $n$ is, proving something like ``$f=p^{0.1}$ suffices'' is hopeless: this is because, with our present knowledge, nothing prevents all of $1,2,3,\dotsc, p^{0,1}$ to be quadratic residues. $\endgroup$
    – Seva
    Commented Jul 25, 2023 at 20:15
  • 1
    $\begingroup$ Thanks everyone for the replies! I wanted to generalize the proof of the case $n = 2$. But it seems that for $n \ge 3$ the problem becomes very hard. $\endgroup$ Commented Jul 26, 2023 at 11:11
  • 1
    $\begingroup$ The bottom line is that the answer depends on how large short character sums can be. See Aurel's excellent answer and my supplement to it. $\endgroup$
    – GH from MO
    Commented Jul 26, 2023 at 23:49
  • $\begingroup$ Re, @‍Aurel's excellent answer and @GHfromMO's supplement to it. $\endgroup$
    – LSpice
    Commented Jul 27, 2023 at 1:01
  • 2
    $\begingroup$ Is there a theorem to the effect that, let $n$ be any positive integer, then for every integer $m$, there are integers $u_1,u_2$, such that $|u_1|,|u_2|<\sqrt n$ and $m\equiv u_1u_2\bmod n$? $\endgroup$ Commented Jul 28, 2023 at 3:53

2 Answers 2

7
$\begingroup$

$\newcommand{\Z}{\mathbb{Z}}$Let $U$ be a subset of the integers $u$ that are nonzero mod $p$ and satisfy $|u|<f$. Consider $V$ the space of functions on $(\Z/p\Z)^\times$, and the following operator on $V$ $$ T \colon F \mapsto \frac{1}{|U|}\sum_{u\in U,\, e=\pm 1}F(u^ex), $$ which is self-adjoint for the natural inner product on $V$.

The eigenvalues of $T$ are the $$ \lambda_\chi = \frac{2}{|U|}\sum_{u\in U}\Re(\chi(u)), $$ where $\chi$ ranges over complex characters of $(\Z/p\Z)^\times$.

Let $\Lambda$ be an upper bound on $|\lambda_\chi|$ for nontrivial $\chi$.

By considering the action on $T$ on the indicator function of two singletons, and decomposing orthogonally to the constant function in $V$, you see that if $\Lambda^n < \frac{1}{p-1}$, then every $m$ is a product of $n$ values $u^{\pm 1}$ for $u\in U$. In other words, your problem is directly related to short sums of Dirichlet characters.

For instance, if you take $U$ the set of primes up to $f$, and assume the Generalised Riemann Hypothesis, then you get $\Lambda = O((\log p)/\sqrt{f})$ (maybe up to some log factors, or after replacing $T$ by a sum with smoothing factors) and therefore you can take $f = O(p^{2/n}(\log p)^2)$.

You should ask someone more knowledgeable about character sums for more bounds.

$\endgroup$
0
4
$\begingroup$

$\newcommand{\Z}{\mathbb{Z}}$Let me add more detail and precision to Aurel's nice answer. First, the characters of $(\Z/p\Z)^\times$ form an eigenbasis of $V$ with respect to $T$. Indeed, $T\chi=\lambda_\chi \chi$. Now if $I_m$ denotes the indicator function of an arbitrary element $m\in(\Z/p\Z)^\times$, then $(T^n I_m)(1)$ equals $|U|^{-n}$ times the number of representations $m=u_1^\pm\dotsb u_n^\pm$ with $u_1,\dotsc,u_n\in U$. On the other hand, the decomposition $$I_m=\frac{1}{p-1}\sum_\chi\overline{\chi(m)}\chi$$ shows that $$(T^n I_m)(1)=\frac{1}{p-1}\sum_\chi\overline{\chi(m)}\lambda_\chi^n.$$ As $\lambda_{\chi_0}=2$, there exists a representation $m=u_1^\pm\dotsb u_n^\pm$ with $u_1,\dotsc,u_n\in U$ as long as $$(p-2)\Lambda^n<2^n.$$ In other words, $U$ is a good set as long as $$\max_{\chi\neq\chi_0}\left|\sum_{u\in U}\Re(\chi(u))\right|<\frac{|U|}{(p-2)^{1/n}}.$$

$\endgroup$
4
  • 1
    $\begingroup$ I suppose, $F_m$ was meant to be $I_m$? Also, one gets a slightly better bound factoring out $\Lambda^{n-2}$ and applying Parseval to the remaining sum of the squares. $\endgroup$
    – Seva
    Commented Jul 27, 2023 at 14:08
  • $\begingroup$ @Seva: Thank you, I changed notation while writing :-) Please feel free to insert an "Added" section to my post, or add it as a separate answer, as a slightly better bound is always welcome. $\endgroup$
    – GH from MO
    Commented Jul 27, 2023 at 14:36
  • 1
    $\begingroup$ It looks like I should have divided by operator $T$ by $2$! $\endgroup$
    – Aurel
    Commented Jul 27, 2023 at 22:07
  • $\begingroup$ @Aurel: Indeed, but it is only a matter of convenience. $\endgroup$
    – GH from MO
    Commented Jul 27, 2023 at 22:29

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .