19

Let $\displaystyle f(x)=\frac{1+\cos(2\pi x)}2$ for $x\in\mathbb R$, and $f^n=\underbrace{ f \circ \cdots \circ f}_{n}$. Is it true that for Lebesgue almost every $x$, $\displaystyle\lim_{n \to \infty} f^n(x)=1$?

I'm more inclined to believe that the answer is "yes".

This is the Problem $5$ of $2021$ Miklós Schweitzer. Recently, a related question reminds me of this problem. After spending some time on it, I found that it is a hard problem, as always as Miklós Schweitzer does. Almost every problem from that competition is very difficult for me.

First of all, for a fixed $x_0\in\mathbb R$, if $f^n(x_0)$ is convergent, then its limit $\ell$ must be a fixed point of $f$. Since $f(x)=\cos^2(\pi x)\in[0,1]$, we must have $\ell\in[0,1]$. Let's find the fixed points of $f$. Let $g(x)=f(x)-x$ for $x\in[0,1]$, then we need to find the zeroes of $g$. Since $g'(x)=-\pi\sin(2\pi x)-1$, $g'$ has two zeroes $\eta_1,\eta_2\in[0,1]$ with $1/2<\eta_1<3/4$, $3/4<\eta_2<1$, and $\sin(2\pi\eta_1)=\sin(2\pi\eta_2)=-1/\pi$. Hence, $g$ is decreasing in $[0, \eta_1)$, then increasing in $(\eta_1, \eta_2)$, and then decreasing in $(\eta_2,1]$. Note that $g(1/2)=-1/2<0, g(1)=0$, we know that $g(\eta_1)<0$ and $g(\eta_2)>0$. Therefore, we can find three zeroes of $g$, named by $\ell_1$, $\ell_2$ and $\ell$ with $\ell_1\in(0,1/2)$, $\ell_2\in(\eta_1, \eta_2)$ and $\ell=1$.

The graphs

We can find the locations the fixed points $\ell_1, \ell_2$ more accurately. Indeed, since $$g\left(\frac14\right)=\cos^2\left(\frac\pi4\right)-\frac14=\frac12-\frac14>0,\qquad g\left(\frac13\right)=\cos^2\left(\frac\pi3\right)-\frac13=\frac14-\frac13<0,$$ we have $\ell_1\in(1/4,1/3)$, hence $$f'(\ell_1)=-\pi\sin(2\pi\ell_1)<-\pi\sin\left(\frac{2\pi}3\right)=-\frac{\sqrt 3}2\pi<-1.$$ Also, $$g\left(\frac56\right)=\cos^2\left(\frac56\pi\right)-\frac56=\frac34-\frac56<0,\qquad g\left(\frac{11}{12}\right)=\frac{1+\cos\left(\frac{11}6\pi\right)}2-\frac{11}{12}=\frac{\sqrt3-2}4>0,$$ we have $\ell_2\in(5/6,11/12)$, hence $$f'(\ell_2)=-\pi\sin(2\pi\ell_2)>-\pi\sin\left(\frac{11\pi}6\right)=\frac{1}2\pi>1.$$

The following are not rigorous.

Therefore, locally, near $\ell_1$, $f$ behaves like $-A(x-\ell_1)$ with $A>1$. Consider the map $f_1: x\mapsto -Ax$, then $f_1^n(x)$ converges if and only if $x=0$. This indicates that, for fixed $x_0$, if the sequence $\{f^n(x_0)\}$ doesn't reach $\ell_1$, it will not converge to $\ell_1$ ; a similar analysis on $\ell_2$ indicates that $\{f^n(x_0)\}$ will not converge to $\ell_2$ if it doesn't touch $\ell_2$; hence, if $\{f^n(x_0)\}$ converges without touching $\ell_1, \ell_2$, then the limit should must be $\ell=1$. I think the ideas in this paragraph can be written down rigorously, although I don't know how to write a clean one.

Another question I've not had any ideas: What if $\{f^n(x_0)\}$ diverges? To finish the problem, even if we write down a proof about the above paragraph, we also need to prove that for a.e. $x$, the sequence $\{f^n(x)\}$ is convergent.

Any help would be appreciated!

Feng
  • 13,705
  • Dear downvoter, would you kindly point out why my question is being downvoted so that I can improve the quality of my question? – Feng Oct 25 '22 at 02:01
  • 4
    Hi Feng, sorry you have to experience that. I personally don't see a problem with the quality of your problem, in fact is one of the most interesting in a long time. Unfortunately, these unpleasantries do occur in MSE. Don't get discouraged. – Mittens Oct 25 '22 at 02:46
  • @OliverDíaz Thanks! You're so nice! – Feng Oct 25 '22 at 02:50
  • 1
    With the substitution $u=1-2x$ the iteration function becomes $g:[-1, 1] \to [-1, 1]$, $g(u) = \cos(\pi u)$. But that does not make it simpler :) – Martin R Oct 25 '22 at 09:01
  • Can something be said about $\int_{0}^{1} f^{n}(x) dx$? If this quantity is close to $1$ exponentially fast as $n$ increases, then your claim would be true. – acreativename Oct 25 '22 at 12:19
  • @acreativename Thanks for yoour comment. Indeed, I tried this at the first glance of the problem. Even $\int_0^1 f(f(x)),dx$ is not easy to compute. – Feng Oct 25 '22 at 12:24
  • @MartinR Thanks for your comment. It looks simpler if we use $g$ to compute the pre-images. At least, we don't need to use the square root symbols very often in $g-$formulation. Maybe it will simplify the problem to some extent, although I'm not sure about that. – Feng Oct 25 '22 at 12:29

2 Answers2

2

It suffices for iteration purposes to restrict $f$ to $[0,1]$ as $f(.) \in [0,1]$.

Set

$$A:= \{x: x\in[0,1], \exists n \in \mathbb{N};\text{ } f^{n}(x) \in (l_{2},1]\} $$

where $l_{2}$ is the second largest fixed point of $f$ (which is roughly $0.89$). If $A$ has measure $1$ then almost all points converge to $1$, this fact has been discussed by people before me. $\textbf{Assume}$ for the sake of contradiction that $m(A) < 1$.

$\textbf{Definition 1:}$ When $V \subset [0,1]$ and $n \in \mathbb{N}$ then

$$f^{n}(V) := \{y: \exists x \in V\text{ so that } y = f^{n}(x)\}$$

$$f^{-n}(V) := \{z: f^{n}(z) \in V\}$$

Note that $[0,1] = A \cup B$ where $$B=\bigcap_{j=1}^{\infty}f^{-j}([0,l_{2}]).$$ Observe that $f^{-n-1}([0,l_{2}]) \subseteq f^{-n}([0,l_{2}])$ for all $n \geq 1$, indeed if $x \in f^{-n-1}([0,l_{2}])\cap (f^{-n}([0,l_{2}]))^c$ then $f^{n}(x) > l_{2}$ and therefore $f^{n+1}(x) > l_{2} \notin [0,l_{2}]$ which is impossible.

Getting our hands dirty we can verify that $[0,l_{2}] \subseteq [0,0.9]$, $f^{-1}([0,l_{2}]) \subseteq [0.1, 0.9]$ and $f^{-2}([0,l_{2}]) \subseteq f^{-1}([0.1,0.9]) \subseteq [0.1,0.4] \cup [0.6,0.9].$

Thus $B \subseteq [0.1,0.4] \cup [0.6,0.9]$, furthermore $$B = \bigcap_{j=1}^{\infty}f^{-j}([0.1,0.4] \cup [0.6,0.9]).$$ Set

$$C = [0.1,0.4] \cup [0.6,0.9].$$

On $C$ we have $1.84 \leq |f'(.)| \leq \pi$. Furthermore like before we have $f^{-n-1}(C) \subseteq f^{-n}(C)$, indeed if $x \in f^{-n-1}(C) \cap (f^{-n}(C))^{c} $ then $f^n(x) \not \in C$ and $f^{n+1}(x) \not \in C$ which is impossible.

Since we know that $m(A) < 1$ and $m(A)+m(B) = 1$, it follows that $m(B)>0$.

$\textbf{Claim 1:}$ $$m(f^{-n}(C)) \rightarrow m(B).$$

$\textbf{Proof:}$ Note that the characteristic functions $1_{f^{-n}(C)}(.)$ decrease downward to $1_{B}(.)$ (as $n\rightarrow \infty$). Thus by the monotone convergence theorem we have that $$\int_{0}^{1}1_{f^{-n}(C)}(x)dx \rightarrow \int_{0}^{1}1_{B}(x)dx = m(B).$$

$$\Rightarrow m(f^{-n}(C)) \rightarrow m(B).$$

$\textbf{Definition 2:}$ For an interval $I \subseteq [0,1]$ with non empty interior, define

$$P_I := \frac{m(A\cap I)}{|I|} = \frac{m(A\cap [a,b])}{b-a} = \frac{\int_{I}1_{A}(x) dx}{|I|}$$

where $I = [a,b]$.

$\textbf{Definition 3:}$ For an interval $[a,b] \subseteq [0,1]$ we set $T((\gamma_{j})_{j=1}^{n}, [a,b])$, where $\gamma_{j} \in \{0,1\}$ as follows: if $n = 1$ then $T((\gamma_{j})_{j=1}^{1}, [a,b]) = \begin{cases} f^{-1}([a,b])\cap[0,\frac{1}{2}] & \text{ if }\gamma_{1} = 0 \\ f^{-1}([a,b]) \cap[\frac{1}{2},1] & \text{ if }\gamma_{1} = 1 \end{cases}$

and for $n > 1$ we define $$T((\gamma_{j})_{j=1}^{n}, [a,b])= T((\gamma_{j})_{j=2}^{n},T((\gamma_{k})_{k=1}^{1}, [a,b]))$$

$\textbf{Example 1:}$ $T((0,1),[0,1]) = T((1), T(0,[0,1]))= T((1),[0,\frac{1}{2}])= [\frac{1}{2},\frac{3}{4}]$

$\textbf{Observation 1:}$ $f^{-n}([a,b])$ is a union of $2^n$ disjoint closed intervals, each of the form $T((\gamma_{j})_{j=1}^{n}, [a,b])$ for some sequence $(\gamma_{j})_{j=1}^{n} \in \{0,1\}^{n}$

$\textbf{Lemma 1:}$ $$\lim_{n \rightarrow \infty}\min( \min_{V \in \{0,1\}^{n}}P_{T(V,[0.1,0.4])}, \min_{V \in \{0,1\}^{n}}P_{T(V,[0.6,0.9])}) = 0$$

$\textbf{Proof:}$ Suppose that there exists a sequence of naturals $n_{1}< n_{2},...$ and an $\alpha > 0$ so that

$$\min( \min_{V \in \{0,1\}^{n_{j}}}P_{T(V,[0.1,0.4])}, \min_{V \in \{0,1\}^{n_{j}}}P_{T(V,[0.6,0.9])}) \geq \alpha,$$

then for such $n_{j}$ note that

$$f^{-n_{j}}(C) = \bigcup_{V \in \{0,1\}^{n_{j}}}T(V,[0.1,0.4]) \cup \bigcup_{W \in \{0,1\}^{n_{j}}}T(W,[0.6,0.9])$$

a union of $2^{n_{j}+1}$ intervals, if each such interval $U$ had $P_{U} \geq \alpha$ then we have the inequality; $$m(B) \leq m(f^{-n_{j}}(C))(1-\alpha)$$

Taking limits as in claim $1)$ we have

$$m(B) \leq m(B)(1-\alpha),$$

which is only possible when $\alpha = 0$ or $m(B) = 0$, as $m(B) \neq 0$ we have a contradiction.

$\textbf{Lemma 2:}$ If $I$ is a subinterval of $C$ (which is $[0.1,0.4] \cup [0.6,0.9]$) then when $\gamma \in \{0,1\}$, $T((\gamma), I)$ has length between $\frac{|I|}{\pi}$ and $\frac{|I|}{1.84}$.

$\textbf{Proof:}$ $T((0),I)$ and $T((1),I)$ are two distinct intervals with the same size (as $T((1),I) = 1 - T((0),I)$). If $f([u,v]) = I$ where $u,v \geq \frac{1}{2}$ then $f([u,v])= [\cos^2(\pi u), \cos^2(\pi v)]$ (here $[u,v]$ corresponds to $T((1),I)$). Thus $|f([u,v])| = |f'(\eta)||u-v|$ for some $\eta \in [u,v]$. On $C$ we have $1.84 \leq |f'(\eta)| \leq \pi$. Thus $$1.84 |T((\gamma),I)| \leq |I| \leq \pi |T((\gamma),I)|.$$

$\textbf{Definition 4:}$ Given an interval $I \subseteq [0,1]$ we set $M_{I} := \sup_{x \in I}|f'(x)|$ and $m_{I} := \inf_{x \in I}|f'(x)|$.

$\textbf{Observation 2:}$ We have $M_{I}-m_{I} \leq 2\pi^2 |I|$ by the mean value theorem applied to $f'(.)$.

$\textbf{Lemma 3:}$ $$P_{I} \leq \frac{M_{T((\gamma),I)}}{m_{T((\gamma),I)}}P_{T((\gamma),I)}$$

$\textbf{Proof:}$ As before suppose $f([u,v]) = I$ where $u,v \geq \frac{1}{2}$ (as by symmetry $m([u,v]\cap A) = m([1-u,1-v] \cap A)$). Note here that $T((1),I) = [u,v]$. Observe that

$$P_{I} = \frac{\int_{I}1_{A}(x)dx}{|I|}$$

$$=\frac{\int_{u}^{v}1_{A}(\cos^2(\pi x))\pi \sin(2\pi x) dx}{|I|} \text{(by integration by substitution)}$$

$$= \frac{\int_{u}^{v}1_{A}(x)\pi \sin(2\pi x) dx}{|I|} \text{ (note that }1_{A}(\cos^2(\pi x))=1_{A}(x))$$

$$\leq \frac{M_{[u,v]}\int_{u}^{v}1_{A}(x)\,dx}{|I|}$$

$$= \frac{M_{[u,v]}P_{[u,v]}|v-u|}{|I|}$$

$$\leq \frac{M_{[u,v]}P_{[u,v]}|v-u|}{m_{[u,v]}|v-u|} \text{ (by M.V.T)}$$

$$= \frac{M_{T((1),I)}P_{T((1),I)}}{m_{T((1),I)}}$$

the case for $\gamma = 0$ is symmetrical.

$\textbf{Observation 3:}$ If $I \subseteq C$ ($I$ is a closed connected interval), we have by lemma 2 that $$P_{I} \leq \frac{1.84 + 2\pi^2|T((\gamma),I)|}{1.84}P_{T((\gamma),I)} \leq (1 + \frac{2\pi^2 |I|}{1.84^2})P_{T((\gamma),I)}$$. By induction and lemma 2, we gather that

$$P_{I} \leq (\prod_{j=0}^{n-1}(1+\frac{2\pi^2|I|}{1.84^{j+2}}))P_{T((\gamma_{j})_{j=1}^{n},I)}$$

Note that $\sum_{j=0}^{\infty} \frac{2\pi^2|I|}{1.84^{j+2}} < \infty$ thus

$\prod_{j=0}^{\infty}(1+\frac{2\pi^2|I|}{1.84^{j+2}}) = K_{I} < \infty.$

Now set $C_{1} = [0.1, 0.4]$ and $C_{2} = [0.6,0.9]$, as per lemma 1) there exists $i \in {1,2}$ ,a natural number sequence $n_{1}<n_{2}<...$, and a sequence of sequences $(V_{j})_{j=1}^{\infty}$, with $V_{j} \in \{0,1\}^{n_{j}}$ so that $P_{T(V_{j}, C_{i})} \leq \frac{1}{j}$. Thus we must have $P_{C_{i}} \leq \frac{K_{C_{i}}}{j}$ for all $j \in \mathbb{N}$ or $P_{C_{i}} = 0$, but this is impossible as we have $m(C_{i}\cap A) > 0 $ for $i = 1,2$.

Feng
  • 13,705
  • I am very, very sorry for my late reply. I can't believe that I almost missed such a perfect solution. I don't know why, I didn't see this answer the first time you answered. – Feng Sep 05 '23 at 03:08
1

Let $\ell_1<\ell_2<1$ be the three fixed points of $f$.The idea is to follow where certain subintervals of $[0,1]$ get mapped and to show that in a finite number of iterations, such subintervals fall into $[\ell_2,1]$ where the dynamics of the system is obvious.

  1. Notice that $[\ell_2,1]\mapsto [\ell_2,1]$ and since $f(x)>x$ and $f$ increases in $[\ell_2,1]$, for $x\in (\ell_2,1]$, $f^n(x)\xrightarrow{n\rightarrow\infty}1$.

The key now is to show that with the exception of preimages (of preimages) of the points $\ell_1$ and $\ell_2$ and of the root $z$, all other points get mapped into $(\ell_2,1]$.

  1. Since $f$ decreases in $[0,\ell_1]$ and $f(x)>x$ on $[0,\ell_1)$, there is a point $0<\ell^{(1)}_2<\ell_1$ such that $f(\ell^{(1)}_2)=\ell_2$. Notice that $[0,\ell^{(1)}_2]\mapsto[\ell_2,1]$. Then by (1), iterations of $f$ at points in $[0,\ell_2')$ converge to $1$.

  2. $f$ has a (the only one) zero $\ell_1<z<\ell_2$. Notice that $[\ell_1,z]\mapsto[0,\ell_1]$ and $[z,\ell_2]\mapsto[\ell_1,\ell_2]$, Thus, the interesting dynamics happens in $[\ell^{(1)}_2,\ell_2]$.

  3. There is $\ell_1<\ell^{(2)}_2<z$ such that $f(\ell^{(2)}_2)=\ell^{(1)}_2$. Notice that $[\ell^{(2)}_2,z]\mapsto[0,\ell^{(1)}_2]$ and from there we get to $[\ell_2,1]$.

  4. There are points $z<\ell^{(3)}_2<\ell^{(1)}_1<\ell_2$ such that $f(\ell^{(3)}_2)=\ell^{(1)}_2$ and $f(\ell^{(1)})=\ell_1$. Notice that $[z,\ell^{(3)}_2]\mapsto[0,\ell^{(1)}_2]$ and from there we get to $[\ell_2,1]$.

We may proceed this way by partitioning the remaining subintervals in $[\ell^{(1)}_2,\ell^{(2)}_2]$ and $[\ell^{(3)}_2,\ell_2]$ using higher order preimages of $z$ and $\ell_1$ and $\ell_2$. To make this much clearer, we may have to use a much better labeling of preimages to denote the order of a preimage., and possibly use symbolic dynamics.

Mittens
  • 39,145
  • 1
    Showing that the union of the pre-images exhausts the entire interval $[0, 1]$ (which the exception of the pre-images of $\ell_1$ and $\ell_2$) is the crucial part of this problem IMO. I must admit that this is not yet obvious to me from your argumentation. – Martin R Oct 24 '22 at 16:40
  • @FShrike: $f$ ha fixed points $\ell_1<\ell_2<1$ in $[0,1]$. The dynamics in $[\ell_2,1]$ is simple since $f$ increases and is above the identity. The rest is to carefully see where subintervals of $[0,\ell_1]$ and $[\ell_1,\ell_2]$ wet mapped. Subintervals of interest are built by using preimages of such fixed points. Ultimately, one wishes to se that each subinterval, after a finite number of steps, falls into $[\ell_2,1]$. – Mittens Oct 24 '22 at 16:43
  • @MartinR: I'll correct (3) shortly. – Mittens Oct 24 '22 at 16:48
  • 1
    The symmetry of the function implies that $\ell^{(1)}_2 = 1-\ell_2$. With $d = 1-\ell_2$ we have $(1/2-d, 1/2+d) \mapsto [0, d) \mapsto (1-d, 1]$ (which may be what you wrote, in a different notation). The preimage of $(1/2-d, 1/2+d)$ is two intervals (symmetric to $x=1/2$), the pre-preimage is four intervals, etc. But again, it is not obvious that the union of all those preimages is “almost all” of $[0, 1]$. – Martin R Oct 24 '22 at 17:32
  • @MartinR: That is correct. I was writing this as I was waiting to take a flight. Indeed, notation can be improved so that preimages of $(f^n)^{-1}(\ell_2)$ can be easily identifiable. I am not claiming it is easy. My numerical experiments (only with double precision, I have no better resources right now) do suggest that intervals of the form $[z',\ell'_2]$ (or with reversed order) abound between to different preimages of $\ell_2$ ($z'$ being a primate of the zero of $f$). I have not done any symbolic dynamics in decades but this problems reminds me of a paper of Mazurkiewicz. – Mittens Oct 24 '22 at 17:50
  • @OliverDíaz Thanks for your partial answer! After reading your ideas, I think you want to prove that $$m\left(\left(\bigcup_{n=1}^\infty f^{-n}(\ell_2,1)\right)\bigcap(0,1)\right)=1.$$ This is in some sort of convergence in measure. Indeed, I tried this before, to prove directly convergence in measure by calculating the pre-images. Your argument seems simpler because here you only need to compute $f^{-n}(\ell_2,1)$ but I was computing $f^{-n}(0,1-\epsilon)$, which is much harder. I hope that this method will work, but I’m not so confident about it, since it is so complicated in calculation. – Feng Oct 25 '22 at 00:49
  • 1
    @Feng: that is correct. In fact, I suspect that if $x\notin\bigcup_n (f^n)^{-1}({\ell_1,\ell_2})$ the there is $m\in\mathbb{N}$ such that $f^m(x)\in (\ell_2,1]$. – Mittens Oct 25 '22 at 00:52
  • 1
    @OliverDíaz: One has also to exclude all preperiodic points of all orders i.e. all $x\in [0, 1]$ such that $f^m(x) = f^n(x)$ for some $m < n$. For example, the solution $a \approx 0.2020420696475173$ of $f(f(x)) = x$ is periodic of order $2$: $a \mapsto f(a) \mapsto a$. – The set of preperiodic points is still countable, so if that were the only exceptions then the conjecture is proven. – Martin R Oct 25 '22 at 06:54