28
$\begingroup$

My question is on the existence (or not) of a function $f:\mathbb{R}\to\mathbb{R}$ which satisfy the equation: $$f(f(x))=x^2-x+1 \text{ for every }x\in\mathbb{R}$$

Supposing that such a map do exist I was able to prove until now that:

$f(x)=1$ iff $x=0$ or $x=1$

$f(x)\ne 0$ for every $x$

$f(x^2-x+1)=f^2(x)-f(x)+1$ for every $x$

$f(x)\ge\frac{3}{4}$ if $x\ge\frac{3}{4}$

For each $x$ it holds that $f(x)=f(1-x)$ or $f(x)+f(1-x)=1$

$f(x)<f(x^2-x+1)$ for all $x\in\mathbb{R}\setminus\{0,1\}$

$\endgroup$
  • 2
    $\begingroup$ Nice observations. $\endgroup$ – Samrat Mukhopadhyay Feb 21 '15 at 9:13
  • $\begingroup$ Why $f(x^2-x+1)=f^2(x)-f(x)+1$ for every $x$? In generally $f(x)$ is not linear function. $\endgroup$ – Leox Feb 21 '15 at 9:16
  • 1
    $\begingroup$ @Leox, that is because $f(f(f(x)))=f^{2}(x)-f(x)+1$ $\endgroup$ – Samrat Mukhopadhyay Feb 21 '15 at 9:18
  • 1
    $\begingroup$ For large $x$, $f(x)\approx x^{\sqrt2}$. $\endgroup$ – Yves Daoust Feb 21 '15 at 14:04
  • 1
    $\begingroup$ @abel Please explain why $f(\frac{1}{2})=\frac{3}{4}$. Also, I am not very sure about the range observation (it is the reverse inclusion). $\endgroup$ – aly Feb 21 '15 at 14:17
11
$\begingroup$

There are infinitely many solutions to $f(f(x))=x^2-x+1$ for $f\colon\mathbb{R}\to\mathbb{R}$ and, in fact, infinitely many differentiable solutions. However, there is only one continuously differentiable solution.

f(f(x))=x^2-x+1

First, let's start with the case where $f$ is assumed to be continuous. Then, we can say the following. I will write $g(x)=x^2-x+1$.

  1. $f(1)=1$
  2. $f$ is strictly increasing over $x\ge1/2$.
  3. $x\le f(x)\le g(x)$ for all $x$.
  4. $f(x)=f(1-x)$ for all $x$.
  5. $f^\prime(1)=1$.

The first property is noted in the question, and does not require continuity. If $f(1)=a$ then we have $f(a)=f(f(1))=g(1)=1$. So, $g(a)=f(f(a))=a$ and, the only fixed point of $g$ is $a=1$.

Next, if $a\not=b\in[1/2,\infty)$ and $f(a)=f(b)$ then, $g(a)=f(f(a))=f(f(b))=g(b)$, so $a=b$. So, $f$ is one-to-one and continuous on this range, implying that it is either strictly increasing or strictly decreasing. We can rule out the decreasing case - as noted in the question, $f(x)\ge3/4$ for $x\ge3/4$. If $f$ was decreasing over $x\ge1/2$ then, as it is bounded below, it would be bounded. So, $g(x)=f(f(x))$ would be bounded, which it isn't. This proves the second statement.

Over $x > 1$ we have $f(x)\not=x$ (otherwise, $x$ would have to be a fixed point of $g$). Then, we either have $f(x) > x$ or $f(x) < x$ for all such $x$. We can rule out the second case, as it would imply that $g(x)=f(f(x))<x$ which is false. By a similar argument, $f(x) > x$ for all $x < 1$. Also, if $f(x) > g(x)$ for any $x$, then we would have $g(x)=f(f(x))\ge f(x) > g(x)$, which is a contradiction. This proves statement 3.

As noted in the question, $f(x)=f(1-x)$ or $f(x)=1-f(1-x)$ for all $x$. In particular, for $x < 1/2$, we can use the inequality $f(1-x) > f(1/2)\ge1/2$ to see that $f(1-x) > 1 - f(1-x)$. So, by continuity we either have $f(x)=f(1-x)$ for all $x\le 1/2$ or $f(x)=1-f(1-x)$ for all such $x$. We can rule out the second case, as it implies that $f(x) < 1/2$ for all $x < 1/2$, which would give the contradiction $g(x)=f(f(x))<1/2$. We have proved statement 4 for all $x\le 1/2$. For $x\ge1/2$ we can replace $x$ by $1-x$ to prove the statement.

Finally, for statement (4) we can apply (3) to get, $$ \frac{(1+h)-1}{h}\le\frac{f(1+h)-f(1)}{h}\le \frac{g(1+h)-1}{h} $$ for all $h>0$. Both sides of this inequality tend to $1$ as $h\to0$. The same argument applies for $h < 0$, except that the inequalities are the other way round. So, $f^\prime(1)=1$.


We can now construct continuous solutions as follows. Pick any $a>1$ and set $b=g(a) > a$. Then, choose any $c\in(a,b)$. On the range $[a,c]$ we take $f$ to be any strictly increasing function with $f(a)=c$, $f(c)=b$. Extend to $(c,b]$ by taking $f(f(x))=g(x)$ for $x\in(a,c]$. The intervals $g^r([a,b])$ cover $(1,\infty)$ as $r$ runs through $\mathbb{Z}$, and we define $f(g^r(x))=g^r(f(x))$ to extend $f$ to all of $(1,\infty)$.

For differentiable $f$, we need to ensure that we choose differentiable $f\colon[a,c]\to[c,b]$ and line up the derivatives at the end points. Using $g^\prime(x)=f^\prime(f(x))f^\prime(x)$ at $x=a$, we require $g^\prime(a)=f^\prime(c)f^\prime(a)$.

We can construct $f$ on $[1/2,1)$ similarly. Set $a=1/2$ and $b=g(a) > a$. Choose any $c\in(a,b)$ and choose $f$ on $[a,c]$ to be strictly increasing with $f(a)=c$ and $f(c)=b$. We extend to $(c,b]$ as above, then use $f(g^r(x))=g^r(f(x))$ to extend to all of $[1/2,1)$.

Finally, we use (4) to extend to $x < 1/2$ and set $f(1)=1$ to complete the construction of $f$ on $\mathbb{R}$. The construction above can be done choosing $f$ to be differentiable in each of $[0,1/2)$ and $(1/2,\infty)$. As long as we choose the right-derivative of $f$ at $1/2$ to be $0$, then the extension to $x\le1/2$ satisfies $f^\prime(1/2)=0$. Finally, (5) implies that $f$ is differentiable at 1. However, in general $f$ will not be continuously differentiable at $1$.


Let me now show that there is a unique continuously differentiable solution. I'll concentrate on $x\ge1$, on which $g$ is strictly increasing. A similar argument applies to $x\in[1/2,1]$, and extends to all of $\mathbb{R}$ by (4).

First, we have $g^\prime(x)=f^\prime(f(x))f^\prime(x)$. Replacing $x$ by $f^{-1}(x)$ and $g^{-1}(x)$ gives \begin{align} f^\prime(x)&=\frac{g^\prime(f^{-1}(x))}{f^\prime(f^{-1}(x))},\\ f^\prime(f^{-1}(x))&=\frac{g^\prime(f^{-2}(x))}{f^\prime(f^{-2}(x))}. \end{align} Plugging the second equality into the first and using $g^{-1}=f^{-2}$ gives, $$ f^\prime(x)=\frac{g^\prime(f^{-1}(x))}{g^\prime(f^{-2}(x))}f^\prime(f^{-2}(x))=\frac{g^\prime(g^{-1}(f(x)))}{g^\prime(g^{-1}(x))}f^\prime(g^{-1}(x)). $$ Iterating this identity gives, $$ f^\prime(x)=f^\prime(g^{-r}(x))\prod_{k=1}^r\frac{g^\prime(g^{-k}(f(x))}{g^\prime(g^{-k}(x))} $$ for all $r\ge1$. If we now let $r$ go to infinity, then $g^{-r}(x)\to1$. If $f^\prime$ is continuous at $1$, then we can impose $f^\prime(g^{-r}(x))\to f^\prime(1)=1$ as $r\to\infty$. So, we get a differentiable equation for $f$, \begin{align} f^\prime(x)&=F(x,f(x)),\\ F(x,y)&=\prod_{k=1}^\infty\frac{g^\prime(g^{-k}(y))}{g^\prime(g^{-k}(x))}. \end{align} It can be seen that $F(x,y)$ is increasing in $y$. So, if $f$, $\tilde f$ were two continuously differentiable solutions and $\tilde f(x) > f(x)$ for some $x > 1/2$, then we see that $\tilde f^\prime(x) > f^\prime(x)$. It follows that $\tilde f(y) > f(y)$ for all $y > x$ and, therefore, $$ g(x)=\tilde f(\tilde f(x)) > \tilde f(f(x)) > f(f(x))=g(x) $$ which is a contradiction, so there is at most 1 continuously differentiable solution.

$\endgroup$
7
$\begingroup$

This paper shows that in general your functional equation has no solution for all $x\in\mathbb{C}$.
See also this, and this. They explain more about functional equations like your one.

$\endgroup$
  • $\begingroup$ If I am computing correctly $\Delta(g)=0<1$, where $g(x)=x^2-x+1$. As a conclusion there are continuous solutions on $\mathbb{R}$. Am I right? $\endgroup$ – aly Feb 25 '15 at 19:05
  • $\begingroup$ There are continuous solutions on $\mathbb{R}$. There are infinitely many of them. There are also infinitely many differentiable solution on $\mathbb{R}$. However, there is, I think, exactly one twice differentiable solution (I think - I'm just working out the details, and I'll post the answer). $\endgroup$ – George Lowther Feb 25 '15 at 22:17
  • $\begingroup$ @GeorgeLowther Great. Thank you very much for the information. $\endgroup$ – aly Feb 25 '15 at 23:10
  • $\begingroup$ Is this supposed to be an actual answer to the question ? $\endgroup$ – Gabriel Romon Mar 7 '15 at 17:43
1
$\begingroup$

Methods due to Jean Ecalle in the 1970's give the following: there is a real analytic solution for $x > 1.$ What makes it so hard is that the derivative at the fixpoint is exactly $1.$

There is also a real analytic solution for $0< x < 1.$

Where the two meet, the resulting joined function is continuous, differentiable, $C^1$ and actually $C^\infty,$ of Gevrey class.

The solution for $x > 1$ can be reflected across for $x < 0.$ Thus the curve is not $C^\omega$ at $0,$ either.

Carrying out the method is a big job. I did it for $f(f(x)) = \sin x.$ See my answer to my own question at https://mathoverflow.net/questions/45608/formal-power-series-convergence

$\endgroup$

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.