Suppose $y>1$ is some approximation to $x=\sqrt{2}+1$. Give a brief reason (not a proof) why one should expect $(1/y)+2$ to be a closer approximation to $x$ than $y$ is.

After testing this out for a bit, it looks like we can let $y_{n+1}=\frac{1}{y_n}+2$ and $\lim_{n\to\infty}y_n=\sqrt{2}+1$, but this does not give me any intuitive idea as to why $y_{n+1}$ should be a better approximation to $x$ than $y_n$ is.

Can anyone give a brief reason for this improvement in aproximation, especially a more "intuitive" one than simple numerical data?

  • 1
    $\begingroup$ A side note: If $x=\sqrt{2}+1$, then $x=(1/x)+2$ (rationalizing the denominator for $1/(\sqrt{2}+1)$ makes this more apparent). $\endgroup$ – interrogative Feb 15 '16 at 17:15
  • $\begingroup$ A good intuition is to think that if your initial $y_n$ is in $(1-\sqrt(2),1+\sqrt(2))$, then the $y_n's$ increase towards $x$, and therefore $|y_{n+1}-x|<|y_n-x|$. You can see this by setting up $y_{n+1}>y_{n}=1/y_n+2>y_n$ which is quadratic and easily solvable. The same argument applies for $y_{n+1}<y_n$ Hope this helps! $\endgroup$ – Fede Poncio Feb 15 '16 at 17:35

Let$ y<1+\sqrt{2}$,

So $1+\sqrt{2} - y=D$ (say).

$1/y > \sqrt{2}-1$

And $1/y+2>1+\sqrt{2}$

So let $1/y+2-(1+\sqrt{2}) =d.$

So we have to prove d < D.

Let D-d>0.

Simplifying you get : 2$\sqrt{2} >y+1/y $

Which is true as $1-\sqrt{2}<y<1+\sqrt{2}$.

You can do similarly for $y>1+ \sqrt{2}$

  • $\begingroup$ I think this is the most natural/easiest way to approach this problem. One thing though: I think you left out one important piece at the very end..."which is true as $y<1+\sqrt{2}$." It really should be "which is true as $1<y<\sqrt{2}+1$" because $2\sqrt{2}>y+1/y$ holds for all $y\in(\sqrt{2}-1,\sqrt{2}+1)$. $\endgroup$ – interrogative Feb 15 '16 at 17:59
  • $\begingroup$ @interrogative Yes you are right. $\endgroup$ – Sudhanshu Feb 15 '16 at 18:22

Let $y=\sqrt2+1+\epsilon=x+\epsilon$ with $\epsilon\to 0$ so $$\frac1y=\frac1{\sqrt 2+1+\epsilon}=\frac{\sqrt{2}-(1+\epsilon)}{2-(1+\epsilon)^2}=\frac{\sqrt2-1-\epsilon}{1-2\epsilon-\epsilon^2}$$ so that $$z=\frac1y+2= \frac{\sqrt2+1-5\epsilon-2\epsilon^2}{1-2\epsilon-\epsilon^2}\sim(\sqrt2+1-5\epsilon-2\epsilon^2)(1+2\epsilon)\sim x+(2\sqrt 2-3)\epsilon$$ and then $$|z-x|\sim(3-2\sqrt2)|\epsilon|<|z-y|\sim(4-2\sqrt2)|\epsilon|$$


Let $y_n=\frac{\sqrt{2}+1}{1+\delta}$, with $\delta$ small (not necessarily positive or nonzero, but surely not $-1$), then \begin{align}y_{n+1}&=\frac1{y_n}+2\\ &=\frac{1+\delta}{\sqrt{2}+1}+2\\ &=(1+\delta)\frac{\sqrt{2}-1}{(\sqrt{2}-1)(\sqrt{2}+1)}+2\\ &=(1+\delta)(\sqrt{2}-1)+2\\ &=\sqrt{2}+1+\delta(\sqrt{2}-1) \end{align} We get $|y_n-(\sqrt{2}+1)|=|\frac{\delta}{\delta+1}|(\sqrt{2}+1)$ and $|y_{n+1}-(\sqrt{2}+1)|=|\delta|(\sqrt{2}-1)$ so what we want is $$\left|\frac{\delta}{\delta+1}\right|(\sqrt{2}+1)>|\delta|(\sqrt{2}-1)$$ and this is true since $$\left|\frac{1}{\delta+1}\right|(\sqrt{2}+1)^2>1$$ for small enough $\delta$ (we're good if $|\delta|<2+2\sqrt{2}$). And since this is true (if you're starting with a good enough approximation, that is. It'll still work for larger starting values, but you'll have to show that $\delta$ gets small enough), we'll get better and better approximations, since the errors get smaller and smaller.

Note that this was secretly induction. I was unable to do a base case, since you didn't provide $y_0$ or $y_1$.


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.