Let $x$ be an irrational number, and $\beta$ strictly larger than its irrationality index, which means that for some $C>0$, for all $n\in \mathbb{Z}^*$, $$d(nx,\mathbb{Z})>C n^{-\beta}.$$ It is known that for a.e. irrational number $x$, the irrationality index is $1$. It is even known that some numbers satisf the above for $\beta=1$ (for instance, $x=\sqrt{2}$).

By arguments from measure theory, I have been able to prove that if $a_n$ satisfies $$\sum_{n=1}^{\infty}na_n<\infty,$$ almost every couple $(x,y)$ of $\mathbb{R}^2$ satisfies for some $C>0$ $$d(nx+my,\mathbb{Z})>Ca_{|n|+|m|},n,m\in \mathbb{Z}.$$ Ideally, I would like, as for a single number $x$, find irrational numbers $(x,y)$ such that this holds in the limiting case $a_n=n^{-2}$.

Has anyone an idea? Or has anyone a useful reference for such things?

  • 2
    $\begingroup$ I suspect the answer is no. Almost all real numbers $x$ have irrationality index $2$, and it seems very likely that one can a sum of two such numbers which is a Liouville number, for example. $\endgroup$ – Greg Martin Nov 12 '19 at 21:20
  • 1
    $\begingroup$ E.g., if $\alpha$ is Liouville, then chances are that $\alpha+\sqrt2$ and $\alpha-\sqrt2$ are both index $2$ while their sum is $2\alpha$. $\endgroup$ – Gerry Myerson Nov 12 '19 at 22:41
  • $\begingroup$ Ok thanks. So it is likely that even if I carefully choose $x$ and $y$, there will be no better bound than $0$? $\endgroup$ – kaleidoscop Nov 13 '19 at 5:39
  • $\begingroup$ I completely modified the question as I have made some progress on my side. $\endgroup$ – kaleidoscop Nov 27 '19 at 19:28
  • $\begingroup$ We must assume $(m,n) \neq (0,0)$, else $d(nx+my, {\bf Z}) = 0$ . . . $\endgroup$ – Noam D. Elkies Nov 28 '19 at 17:44

Yes, such $(x,y)$ exist; for example, $x = \root 3 \of 2$ and $y = x^2$. For $l,m,n \in \bf Z$, define $$ N(l,m,n) := l^3 + 2m^3 + 4n^3 - 6lmn \in {\bf Z}; $$ this is the algebraic norm $$ (l + mx + nx^2) (l + m\rho x + n \bar\rho x^2) (l + m\bar\rho x + n \rho x^2) $$ of $l + mx + nx^2$, where $\rho$ is the cube root of unity $e^{2\pi i/3} = (-1 + \sqrt{-3})/2$. But if $(m,n) \neq (0,0)$ then $l + mx + nx^2 \neq 0$, so $\left|N(l,m,n)\right| \geq 1$ and $\left|l + mx + nx^2\right| \gg (|l|+|m|+|n|)^{-2}$. Taking for $l$ the integer nearest to $-(mx+nx^2)$ we deduce that $d(mx+nx^2,{\bf Z}) \gg (|m|+|n|)^{-2}$, as claimed.

The same argument (which generalizes the familiar one for $\sqrt 2$) shows that in general if $x$ is an algebraic number of degree $D$ then $$ d\Bigl(\sum_{j=1}^{D-1} n_j x^j, {\bf Z}\Bigr) \gg \left( \sum_{j=1}^{D-1} \left| n_j \right| \right)^{1-D} $$ for $n_1,\ldots,n_{D-1} \in \bf Z$ not all zero. This is best possible up to the value of the implicit constant, because Dirichlet's celebrated "pigeonhole" argument shows that conversely for any $N$ one can find integers $n_1,\ldots,n_{D-1} \in [-N,N]$, not all zero, such that $d(\sum_{j=1}^{D-1} n_j x^j, {\bf Z}) \ll N^{1-D}$.

| cite | improve this answer | |
  • $\begingroup$ Sorry I did not provide a reference; none of this is new, but it's easier to recall the proofs than to find them in the literature. $\endgroup$ – Noam D. Elkies Nov 28 '19 at 17:46
  • $\begingroup$ I see, pretty neat, thanks! "Not new" but for me it is dropped from the sky... Do you know f by any chance such arguments extend even further to infinite sums? i.e. I'm looking for $x_j$ such that for integers $n_j$, the distance from $\sum_j x_j n_j$ to $\mathbb{Z}$ is bounded from below by the inverse value of some norm of $\|(n_j)\|$ on the space of integer sequences... $\endgroup$ – kaleidoscop Nov 29 '19 at 21:04
  • 1
    $\begingroup$ Seems the best $C$ is $1/(6x) = 0.13228342\ldots$. For large $k$, if $l+mx+nx^2 = (x-1)^k$ then its product with $(|m|+|n|)^2$ depends on the phase of $(e^{2\pi i/3}x - 1)^k$, which can be made arbitrarily close to the optimum. As with $\sqrt2$ for $d(nx,{\bf Z})$, the choice $x = \root3\of 2$ here is the most familiar but not quite the best; taking for $x$ the real root of $x^3-x+1$, or the root $2\cos(2\pi/7)$ of $x^3+x^2-2x-1$, should yield somewhat larger $C$. To be sure the exact value of $C$ is partly an artifact of the choice of quadratic factor: [cont'd] $\endgroup$ – Noam D. Elkies Dec 9 '19 at 21:34
  • 1
    $\begingroup$ changing $(|m|+|n|)^2$ to say $m^2+n^2$ or $\max(|m|,|n|)^2$ would change both the liminf and (probably) the sequence of algebraic units that attain it. But the liminf would remain finite and positive. $\endgroup$ – Noam D. Elkies Dec 9 '19 at 21:36
  • 1
    $\begingroup$ Very cool. Thank you. For $x = \root 3 \of 2$ I could only get down to .132283780 before running out of patience. Two more digits and an ISC lookup and I may have got there. $\endgroup$ – O. S. Dawg Dec 10 '19 at 5:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.