24
$\begingroup$

I was playing around on desmos the other day, and noticed that $\sin\left(\sin\left(x\right)\right)$ is basically a version of sin with a lower amplitude (which makes intuitive sense). To me, it seems intuitve that this curve, when composed infinite times, becomes a straight line, as the values at $x=\frac{\pi}{2}\mathbb{Z}$ would move (slowly) towards the values at $x=\pi\mathbb{Z}$ by virtue of them moving away from the peaks, but is there a way to go about properly proving this?

$\endgroup$
16
$\begingroup$

Good question. The straight line you are talking about is actually the line $y=0$.

One way to see this is to start at any point $x_0\in\mathbb{R}$ and apply a fixed point iteration with the formula $$x_{n+1} = \sin(x_n).$$

It's almost as if the maths Gods knew you were going to ask this question, because Wikipedia even has a nice diagram of what's going on in your exact example!

enter image description here

The diagram shows what happens with $x_0 = 2$. You get a staircase diagram converging towards the only root of the equation $x = \sin(x),\ $ namely $x = 0.$

But this was with a starting point of $x_0 = 2.$ What about other starting points? Well, we know that, for all $x \in \mathbb{R},\ -1 \leq \sin(x) \leq 1.$ If $x_0$ is such that $0\leq x_1=\sin(x_0)\leq 1,\ $ then we get a convergence similar to that in the diagram above (staircase). And if $x_0$ is such that $-1\leq x_1=\sin(x_0)\leq 0,\ $ then we also get something similar, but happening below the $x-$axis, but still towards the root $x=0$.

It follows that the curve that $\sin\left(\sin\left(\sin\left(\ldots x \right)\right)\right)$ converges to is in fact the $x$-axis, ie. the line $y=0.$

Why/when does this occur / how can we prove that we get convergence via this "fixed-point iteration" process?

It can be shown that a fixed-point iteration $x_{n+1} = g(x_n)$:

  • converges if $\ \vert g'(x)\vert < 1\ $ near the root and $x_1$ is sufficiently close to the root
  • diverges if $\ \vert g'(x) > 1\ $ near the root.

There's lots of material out there if you want to look up proofs of these things, but the above is a summary of when convergence happens.

In the $x_{n+1} = \sin(x_n)\ $ example, we have $g(x)=\sin(x).\ $ We know that $g'(0) = 1$, and $g'(x)<1$ for $x$ close to but not equal to the root. We have also seen that $x_1\in [-1,1]\ $ and is sufficiently close to the root. Therefore we get convergence towards $0$ for any starting value $x_0.$

$\endgroup$
2
  • $\begingroup$ This is a clever way to formalize this. Thanks! $\endgroup$ – Tuatarian May 7 at 16:10
  • 1
    $\begingroup$ @Tuatarian It's pretty much the first example you will see in a dynamical systems class (cosine is a little more interesting). $\endgroup$ – Brady Gilg May 7 at 18:53
8
$\begingroup$

We know that $\sin(\mathbb{R})=[-1,1]$ Since the sine function is odd, it will be enough to consider the interval $[0,1]$. The sine function is increasing on this interval so $\sin([0,1])=[0,\sin(1)]$ Since $|sin(x)|<|x|$ this will be a shorter interval than $[0,1]$. We can apply the process repeated getting a succession of smaller intervals, which must either converge to the origin, or to an interval $[0,y]$ with $y=\sin y$. But the only such $y$ is $y=0$, and the intervals shrink to the origin.

Thus, for any $\varepsilon>0$ there is $N$ such for $n>N,\ |\sin^{(n)}x|<\varepsilon$ for every $x\in \mathbb{R}$. Thus, you are correct in saying that the graph will look like a straight line, in fact the $x$-axis.

$\endgroup$
3
  • $\begingroup$ +1. I like how concise and clear this answer is. $\endgroup$ – Taladris May 7 at 5:36
  • $\begingroup$ This is not a really conclusive argument. $y = \sin y$ is besides the point, because $\sin 0$ never actually turns up in the iteration process. The important thing is that there is a Lipschitz cone and the function stays strictly below it in an open region around 0. $\endgroup$ – leftaroundabout May 7 at 7:53
  • 2
    $\begingroup$ @leftaroundabout As I said, it's enough to consider the interval $[0,1]$, and $0$ certainly is one of the endpoints. $\endgroup$ – saulspatz May 7 at 11:17
6
$\begingroup$

Evaluation of the Limit

Let $a_{n+1}=\sin(a_n)$ and suppose $a_0\gt0$.

As shown in this answer, for $x\gt0$, $0\lt\sin(x)\lt x$. Thus, $a_n$ is a decreasing sequence bounded below by zero. Therefore, $\lim\limits_{n\to\infty}a_n$ exists. Since $\sin(x)$ is continuous, $$ \begin{align} \lim_{n\to\infty}a_n &=\lim_{n\to\infty}\sin(a_n)\tag{1a}\\ &=\sin\left(\lim_{n\to\infty}a_n\right)\tag{1b} \end{align} $$ Since $\sin(x)=x$ only at $x=0$, $(1)$ says that $$ \bbox[5px,border:2px solid #C0A000]{\lim_{n\to\infty}a_n=0}\tag2 $$ Since $\sin(x)$ is an odd function, $(2)$ holds for $a_0\lt0$.

Of course, if $a_0=0$, then $a_n=0$ for all $n\ge0$, and $(2)$ holds for $a_0=0$.


Asymptotics

As shown in this answer $$ \lim\limits_{x\to0}\frac{x-\sin(x)}{x^3}=\frac16\tag3 $$ If $a_0\ne0$, since $a_{n+1}=\sin(a_n)$, $(2)$ and $(3)$ say $$ \lim_{n\to\infty}\frac{a_n-a_{n+1}}{a_n^3}=\frac16\tag4 $$ An immediate consequence of $(2)$ and $(4)$ is $$ \begin{align} \lim_{n\to\infty}\frac{a_{n+1}}{a_n} &=\lim_{n\to\infty}\left(1-a_n^2\frac{a_n-a_{n+1}}{a_n^3}\right)\tag{5a}\\ &=1-0^2\cdot\frac16\tag{5b}\\[6pt] &=1\tag{5c} \end{align} $$ Thus, $(4)$ and $(5)$ give $$ \begin{align} \lim_{n\to\infty}\left(\frac1{a_{n+1}^2}-\frac1{a_n^2}\right) &=\lim_{n\to\infty}\frac{a_n^2-a_{n+1}^2}{a_n^2a_{n+1}^2}\tag{6a}\\ &=\lim_{n\to\infty}\frac{a_n-a_{n+1}}{a_n^3}\lim_{n\to\infty}\frac{a_n+a_{n+1}}{a_{n+1}}\lim_{n\to\infty}\frac{a_n}{a_{n+1}}\tag{6b}\\ &=\frac16\cdot2\cdot1\tag{6c}\\[3pt] &=\frac13\tag{6d} \end{align} $$ Finally, Stolz-Cesàro and $(6)$ say that $\lim\limits_{n\to\infty}\frac1{na_n^2}=\frac13$; that is, $$ \bbox[5px,border:2px solid #C0A000]{\lim_{n\to\infty}na_n^2=3}\tag7 $$ In fact, starting with $a_0=1$, we get $$ \begin{array}{r|l} n&na_n^2\\\hline 1&0.7080734183\\ 10&2.1433001582\\ 100&2.8511162950\\ 1000&2.9803925383\\ 10000&2.9976147637\\ 100000&2.9997198782\\ 1000000&2.9999678410 \end{array} $$


Bounds

Substituting $x\mapsto x/\pi$ in $(21)$ from this answer, we get $$ \cot(x)=\sum_{k\in\mathbb{Z}}\frac1{k\pi+x}\tag8 $$ Subtracting $(8)$ from $\frac1x$ and taking the derivative, we get $$ \frac1{\sin^2(x)}-\frac1{x^2}=\sum_{\substack{k\in\mathbb{Z}\\k\ne0}}\frac1{(k\pi+x)^2}\tag9 $$ Evaluating $(9)$ at $x=0$ gives $\frac2{\pi^2}\zeta(2)=\frac13$, which agrees with $(6)$.

Taking two derivatives of $(9)$ gives $$ \begin{align} \frac{\mathrm{d}^2}{\mathrm{d}x^2}\left(\frac1{\sin^2(x)}-\frac1{x^2}\right) &=\sum_{\substack{k\in\mathbb{Z}\\k\ne0}}\frac6{(k\pi+x)^4}\\ &\ge0\tag{10} \end{align} $$ Thus, $\frac1{\sin^2(x)}-\frac1{x^2}$ is an even convex function with a minimum of $\frac13$ at $x=0$.

Since $a_k^2$ is decreasing, $$ \frac13\le\frac1{a_{k+1}^2}-\frac1{a_k^2}\le\frac1{\sin^2(a_0)}-\frac1{a_0^2}\tag{11} $$ Summing $(11)$ gives $$ \frac n3\le\frac1{a_n^2}-\frac1{a_0^2}\le n\left(\frac1{\sin^2(a_0)}-\frac1{a_0^2}\right)\tag{12} $$ Since $\sin^2(x)\le1$, assume that $a_0^2\le1$, whence $\frac1{\sin^2(a_0)}-\frac1{a_0^2}\le\cot^2(1)$.

Solving $(12)$ for $a_n^2$ yields $$ \frac{\tau a_0^2}{\tau+na_0^2}\le a_n^2\le\frac{3a_0^2}{3+na_0^2}\tag{13} $$ where $\tau=\tan^2(1)\approx2.42551882$.


Graphical Verification

Limit $(7)$ holds for any $a_0\ne0$. Since $\sin(x)$ is an odd function, changing the sign of $a_0$, changes the sign of all $a_n$. Thus, we have that $$\newcommand{\sgn}{\operatorname{sgn}} \lim_{n\to\infty}\sqrt{n}\,a_n=\sqrt3\sgn(a_0)\tag{14} $$ Therefore, if we define $\sin^{\ast n}(x)$ to be the $n$-fold composition of $\sin(x)$, we get $$ \bbox[5px,border:2px solid #C0A000]{\lim_{n\to\infty}\sqrt{n}\sin^{\ast n}(x)=\sqrt3\sgn(\sin(x))}\tag{15} $$ Here is graphical verification of $(15)$:

enter image description here

Asymptotically, $(15)$ says that $$ \bbox[5px,border:2px solid #C0A000]{\sin^{\ast n}(x)\sim\sqrt{\frac3n}\sgn(\sin(x))}\tag{16} $$ which would be a more precise way of saying what you were trying to say.

$\endgroup$
3
  • $\begingroup$ This is also quite clever, +1 $\endgroup$ – Tuatarian May 8 at 19:59
  • $\begingroup$ Thanks. I have expanded the answer to include the functions you mentioned, although scaled so that we can better see what they look like. $\endgroup$ – robjohn May 8 at 22:35
  • $\begingroup$ Related: jonathan.bergknoff.com/journal/iterated-sine $\endgroup$ – Gary May 9 at 11:13
5
$\begingroup$

This is straightforward to prove, but needs attention to some detail.

Suppose we pick some $x_0$ and let $x_{n+1} = \sin x_n$. The idea is to show that $x_n \to 0$.

If $x_0 = 0$ then $x_n = 0 $ for all $n$, so suppose $x_0 \neq 0$. Note that $|x_n| \le 1$ for $n \ge 1$, so we might as well suppose $|x_0| \le 1$.

The mean value theorem shows that for for $x \neq 0$ and $|x| \le 1$ that $\sin x-0 = (\cos \xi) x$ for some $\xi \in (0, x)$ (or $(x,0)$ if $x <0$), and since $\cos \xi \in (\cos 1,1)$ (using $x\neq0$ and $|x| \le 1$ here) we see that $|\sin x| = \sin |x| < |x|$.

In particular, $|x_n| $ is non increasing and so $|x_n| \to t^*$ for some $t^*$. By continuity we have $\sin t^* = t^*$ and from the previous paragraph we see that we must have $t^* = 0$. Hence $x_n \to 0$ always.

$\endgroup$

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.