This is Problem 4.26 on p.58 of The Theory of Algebraic Numbers by Harry Pollard and Harold G. Diamond (Dover edition).

Prove that $\left|\sqrt2-\dfrac ab\right|\geq\dfrac1{3b^2}$ for all positive integers $a,b$.

Here's what I've done so far. Split into cases:

  • (i) $\dfrac ab<\sqrt2$
  • (ii) $\dfrac ab>\sqrt2+\dfrac13$
  • (iii) $\sqrt2<\dfrac ab<\sqrt2+\dfrac13$.

(i) I followed the method used by the authors in proving a similar result in that chapter. Let $f(x)=x^2-2$; this is the minimum polynomial of $\sqrt2$ in $\mathbb Q$ so $f\left(\sqrt2\right)=0$. By the mean-value theorem: $$\frac{f\left(\sqrt2\right)-f\left(\dfrac ab\right)}{\sqrt2-\dfrac ab}\ =\ f'(\xi)$$ for some $\dfrac ab<\xi<\sqrt2$. We have $f'(\xi)=2\xi<2\sqrt2<3$, and so $$\left|f\left(\frac ab\right)\right|\ <\ 3\left|\sqrt2-\frac ab\right|$$ But $$\left|f\left(\frac ab\right)\right|=\left|\frac{a^2}{b^2}-2\right|=\frac{|a^2-2b^2|}{b^2}\geq\frac1{b^2}$$ Hence $$\left|\sqrt2-\frac ab\right|\ \geq\ \frac1{3b^2}$$ as required.

(ii) Done immediately as $$\left|\sqrt2-\frac ab\right|=\frac ab-\sqrt2\,> \frac13\geq\frac1{3b^2}$$

(iii) This is where I'm stuck.

So I've basically done much of the hard work above and the final piece is all I need to complete the jigsaw. I would be grateful for some help. Thanks.


Do the case distinction with $1/12$ instead of $1/3$. This does not affect (i) and (iii) works in the same way as (i), since you still have $2 \xi < 3$ as $2(\sqrt{2} +1/12)<3$.

Of course, you need to reconsider (ii) then. But it is hardly more difficult, as you have $\frac{a}{b} - \sqrt{2} \ge \frac{1}{12}$ and $\frac{1}{12} = \frac{1}{3 \cdot 4^2} \ge \frac{1}{3b^2}$, except if $b=1$. Yet then if $b=1$, you have that $a/b$ is an integer so that in this case $a/b \ge 2$ and the difference is at least $1/3$ in this case too.

There may be a shorter method, but this was what came to mind.


Prove that $$\left|\sqrt2-\dfrac ab\right|>\dfrac1{3b^2}$$ for integers $a$ and $b$ with $b\neq 0$.

Without loss of generality, assume that $b>0$. Note that $$\left|\sqrt{2}-\frac{a}{b}\right|=\frac{\left|a^2-2b^2\right|}{b(a+\sqrt{2}b)}\,.\tag{*}$$ If $a<0$, then $$\left|\sqrt{2}-\frac{a}{b}\right|>\sqrt{2}>\frac{1}{3b^2}\,.$$ If $0\leq a\leq b$, then, using (*), $$\left|\sqrt{2}-\frac{a}{b}\right|\geq\frac{b^2}{b(b+\sqrt{2}b)}\geq \frac{1}{1+\sqrt{2}}>\frac{1}{3}\geq\frac{1}{3b^2}\,.$$ If $a\geq \frac{3}{2}b$ and $b>1$, then $$\left|\sqrt{2}-\frac{a}{b}\right|\geq\frac{3}{2}-\sqrt{2}>\frac{1}{12}\geq\frac{1}{{3b^2}}\,.$$ If $a>1$ and $b=1$, then $$\left|\sqrt{2}-\frac{a}{b}\right|\geq 2-\sqrt{2}>\frac{1}{3}\geq\frac{1}{3b^2}\,.$$ If $b<a<\frac{3}{2}b$, then, using (*), $$\left|\sqrt{2}-\frac{a}{b}\right|\geq \frac{1}{b\left(a+\sqrt{2}b\right)}>\frac{1}{b\left(\frac{3}{2}b+\sqrt{2}b\right)}=\frac{1}{\left(\frac{3}{2}+\sqrt{2}\right)b^2}>\frac{1}{3b^2}\,.$$

Indeed, the smallest constant $\gamma>0$ such that $$\left|\sqrt{2}-\frac{a}{b}\right|\geq \frac{1}{\gamma\,b^2}$$ for all integers $a$ and $b$ with $b\neq 0$ is $\gamma=\frac{3}{2}+\sqrt{2}$. The equality holds if and only if $a=3$ and $b=2$.

Related Observation

Conjecture. Let $\gamma_b:=\max\,\left\{\frac{1}{b^2\,\left|\sqrt{2}-\frac{a}{b}\right|}\,\Big|\,a\in\mathbb{Z}\right\}$ for each $b\in\mathbb{N}$. Define $$\Gamma:=\liminf_{b\to\infty}\,b\cdot\gamma_b\,.$$ Then, $$\Gamma=\inf\,\left\{b\cdot\gamma_b\,\big|\,b\in\mathbb{N}\right\}=2\,.$$

After checking all positive integers $b\leq 1500$, I find that $$\Gamma\leq 1+\frac{29}{41}\sqrt{2}\lesssim 2.0003\,.$$ Below is a plot illustrating the credibility of this conjecture. Th horizontal axis is $b$. The fuzzy blue line shows $\gamma_b$ versus $b$, and the nice orange line is $\dfrac{2}{b}$ versus $b$ for $b=1,2,\ldots,1500$.

Plot describing $\gamma_b$ (blue) and $\frac{2}{b}$ (orange) versus $b$ for $b=1,2,\ldots,1500$.

What is very strange is $\Gamma$ seems to be equal to $2$ as well if $\sqrt{2}$ is replaced by $\sqrt{3}$, $\sqrt{5}$, $\sqrt{6}$, $\sqrt{7}$, and $\sqrt{8}$. Could $\Gamma=2$ universally hold when $\sqrt{d}$ replaces $\sqrt{2}$ for any non-square positive integer $d$? I tried replacing $\sqrt{2}$ by $\frac{1+\sqrt{5}}{2}$, $\sqrt[3]{2}$, $\pi$, $\text{e}$, and $\ln(3)$ as well, and it looks like $\Gamma=2$ still holds.

On the other hand, $$\limsup\limits_{b\to\infty}\,\gamma_b=2\sqrt{2}\,.$$ If $\sqrt{2}$ is replaced by $\sqrt{d}$ for any non-square positive integer $d$, then $$\limsup\limits_{b\to\infty}\,\gamma_b=2\sqrt{d}\,.$$

  • 1
    $\begingroup$ The notation is a bit hard to untangle for me, but fwiw it is not hard to see that every number can be approximated to $\frac{1}{2b}$ by a rational of the form $\frac{a}{b}$, as $\frac{1}{b}\mathbb{Z} + [-1/2b,1/2b]$ covers the real line, which I believe gives the $\Gamma$. $\endgroup$
    – quid
    Aug 22 '16 at 16:29
  • $\begingroup$ @quid I think you are right. I didn't try hard enough to untangle my notations (as they came directly from the code I used to experiment). The answer seems to be as simple as you have noted. $\endgroup$ Aug 22 '16 at 16:33
  • 1
    $\begingroup$ Or, it's in the other direction and the worst case, in which case my argument above is only a plausibility argument (which one could likely turn into an actual one), in that for some $b$ the $\sqrt{2}$ will be very close to a number of the form $\frac{a+1/2}{b}$ and the approx is only marginally better than $1/2b$ for that $b$. I will stop guessing around now. I should have checked before commenting. :-) $\endgroup$
    – quid
    Aug 22 '16 at 16:39
  • $\begingroup$ The last statement seemingly can be proven from my answer since there are infinitely many $(a,b)$ : $|a^2-db^2|=1$ and thus $|\sqrt d - {a \over b}| < {1 \over kb^2}$ is equivalent to $k < \sqrt d + \sqrt{d + {1 \over b^2}}$ for arbitrarily large $b$. $\endgroup$ Aug 23 '16 at 15:59
  • 1
    $\begingroup$ Yes, and it has a more general form: if $r\in\mathbb{R}$ is algebraic over $\mathbb{Q}$ whose (monic) minimal polynomial $f(X)\in\mathbb{Q}[X]$ is of degree $n$, then there exists $K>0$ such that $$\left|r-\frac{a}{b}\right|\geq \frac{1}{K\,b^n}$$ holds for all $a,b\in\mathbb{Z}$ with $b>0$. $\endgroup$ Aug 23 '16 at 16:07

Here is the alternate way, without case analysis.

Suppose that for some $a,b$, $|\sqrt 2 - {a \over b}| < {1 \over kb^2}$. Rewriting the inequality, $$-{1 \over kb^2} < {a \over b} - \sqrt 2 < {1 \over kb^2}\\ 2-{2\sqrt 2 \over kb^2} + {1 \over k^2b^4} < {a^2 \over b^2} < 2 + {2\sqrt 2 \over kb^2} + {1 \over k^2b^4} \\ a^2 = 2b^2 + {1 \over k^2b^2} + {\delta \over k}$$ where $|\delta| < 2 \sqrt 2$. Also, from $\sqrt 2$ being irrational we have $$|a^2 - 2b^2| \ge 1 \\ |{1 \over k^2b^2} + {\delta \over k}| \ge 1 \\ |\delta| \ge k - {1 \over kb^2} \\ k - {1 \over kb^2} < 2 \sqrt2 \\ k < \sqrt2 + \sqrt{2 + {1 \over b^2}}$$ Finally for $k=3$ $b=1$ can be excluded separately and $\sqrt 2 + \sqrt{2 + {1 \over 4}} < 3.$


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.