7
$\begingroup$

I took this result from Minkowski's book on Geometry of numbers:

Two arbitrary real quantitites $a$ and $b$ may be made to approach as near as we wish in value the two fractions $\frac{x}{z}$ and $\frac{y}{z}$, that have the same denominator and at the same time in such a way that: $$ \left| \frac{x}{z} - a \right| < \sqrt{\frac{8}{19}} \frac{1}{z^{3/2}} \hspace{0.5in}\left| \frac{y}{z} - b \right| < \sqrt{\frac{8}{19}} \frac{1}{z^{3/2}}$$

I interpret this as we are trying to approximate both $a, b \in \mathbb{R}$ with two fractions $\frac{x}{z},\frac{y}{z} \in \mathbb{Q}$ with the same denominator to $O(z^{-3/2})$.

How to prove this result? Is $\sqrt{\frac{8}{19}}$ optimal?

I will settle for worse constant just to see how geometry of numbers works here.

$\endgroup$

2 Answers 2

8
$\begingroup$

With the constant $1$, this is Minkowski's higher dimensional extension of Dirichlet's approximation theorem:

If $\alpha_1, \ldots,\alpha_n$ are real numbers, then there are rationals $p_i/q$ with $|\alpha_i - p_i/q| < q^{-(1+1/n)}$. If one at least among the $\alpha_i$ is irrational, there are infinitely many such $n$-tuples $(p_1/q,\ldots,p_n/q)$.

The proof is a simple pigeonholing. With an integer parameter $Q$, consider the $Q^n+1$ points $(1,\ldots,1)$ and $(\{\alpha_1 x\}, \ldots,\{\alpha_n x\})$, $0 \leq x < Q^n$ integer, of the unit cube $[0,1]^n$. Some two such points must be contained by a common cube of side $1/Q$. Taking their difference produces a positive integer $q < Q^n$ and integers $p_1, \ldots, p_n \in \mathbb{Z}$ (arising as the appropriate integer parts) with $|q\alpha_i - p_i| \leq 1/Q$ for all $i = 1,\ldots,n$. We have $1/Q < q^{-1/n}$, giving the inequality. Moreover, we may by clearing the common factor assume $(p,q_1,\ldots,q_n) = 1$ for the tuple thus produced, and if say $\alpha_1$ is irrational it is clear that for a fixed tuple $|q\alpha_1 - p_1| \leq 1/Q$ can hold only for finitely many $Q$. Hence, letting $Q \to \infty$, infinitely many tuples $(q;p_1,\ldots,p_n)$ are thus obtained when some $\alpha_i \notin \mathbb{Q}$.


Better constants can be obtained. For $n = 1$ the optimal constant is $1/\sqrt{5}$ as you know (and then you have the Lagrange-Markov spectrum). For $n = 2$ the optimal constant is unknown as far as I am concerned; this is stated on page 41 of Wolfgang Schmidt's book "Diophantine Approximation" (LNM 785), to which I can refer you for an improvement of the constant $1$ in the general case. There, in particular, the value $2/3$ is shown to work for $n = 2$ (which however is slightly bigger than the constant you cite from Minkowski book), and a reference is given to the literature proving that the optimal constant lies between the values $\sqrt{2/7} \approx 0.53$ and $0.615 < 0.649 \approx \sqrt{8/19}$.

So no, the constant $\sqrt{8/19}$ is not optimal in this result.

$\endgroup$
7
$\begingroup$

This is a supplement to Vesselin Dimitrov's answer. For $n=2$ the infimum of admissible constants is between $(2/7)^{1/2}\approx 0.5345$ and $8/13\approx 0.6154$, and it has been conjectured that $(2/7)^{1/2}$ is the truth. The lower bound is due to Cassels (J. London Math. Soc. 30, (1955), 119-121), the upper bound is due to Nowak (Manuscripta Math. 36 (1981/82), 33-46).

P.S. Schmidt's book "Diophantine approximation" on page 41 refers to Mack's earlier result, which was slightly improved by Nowak. Schmidt is somewhat misleading to give $0.615$ as an upper bound, because Mack's upper bound was $\approx 0.6155$, and Nowak's upper bound is $\approx 0.6154$, both bigger than $0.615$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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