# Find $f(f(\cdots f(x)))=p(x)$

$\newcommand{\nest}{\operatorname{nest}}$Let's define a function $\nest(f, x, k)$, which takes a function $f$, an input $x$, and a non-negative integer $k$, and calls $f$ on $x$ repeatedly ($k$ times). For example, $$\nest(f, x, 0) = x\\ \nest(f, x, 1)=f(x)\\ \nest(f, x, 2)=f(f(x))\\ \nest(f, x, 3)=f(f(f(x)))$$ Formally, this function can be written as $$\nest(f, x, k)= \begin{cases} x & \text{if } k=0\\ \nest(f, f(x), k-1) & \text{otherwise} \end{cases}$$

For a given $k$ and a polynomial $p$, how can I find a function $f: \mathbb C \to \mathbb C$ such that $\nest(f, x, k)=p(x)$?

If it's not possible to do so in the general case, is it possible with $p(x)=c+x^2$?

• I don't know about $c+x^2$ but I found the function $f(x)=\sqrt{x^2+\frac{c}{k}}$ for $p(x)=\sqrt{c+x^2}$ Jul 19, 2018 at 20:11
• It's not to hard to find a function $f$ of that sort. For $k = 2$, for example, start with two arbitrary numbers $a$ and $b$, and define $f$ to act according to the chain $a \mapsto b \mapsto a^2 +c \mapsto b^2 + c \mapsto (a^2 + c)^2 + c \mapsto \cdots$. Then take two other starting numbers that don't appear in that chain and repeat. (If separate chains eventually overlap, then join them into a tree structure.) Whether any function $f$ with a remotely interesting functional form exists, of course, is a different question. Jul 19, 2018 at 20:17
• Doubt you can find one for $p(x)=c+x^2$. At least if you want $f$ to be holomorphic. Such functions fold the complex plane onto itself. Compositions will fold the plane a power of that original number. For example, if $f$ doesn't fold the plane, neither will any of its further nestings. If $f$ folds the complex plane with two layers, the 2nd nesting will fold the plane with four layers. $3$ layers will result in $9$ layers on the next nesting. $2$ is not a tenable power for foldings of a nesting. Probably can be made rigorous with the Argument principle.
– user123641
Jul 19, 2018 at 20:29
• I tried to find a series expansion for $f(x)$ with $k=2$ and $p(x)=1+x^2$ over $[0,1]$ using $11$ equally spaced points and minimizing the SSR. I get $f(x)\approx0.643+0.033x+0.940x^2-0.201x^3$. It fits so well that I'd assume $f(x)$ exists in this case. But whether it has a closed form or exists over all of $\mathbb{C}$ is another question. Regardless, you might be able to find approximations to your $f(x)$'s Graph: desmos.com/calculator/vf5fdpjbwi
– Jam
Jul 19, 2018 at 20:41
• What you're looking for are called fractional iterates of the (polynomial) function; see e.g. this WIkipedia bit or this one or this previous math.se question Jul 19, 2018 at 21:06

You are really asking about functional iterates, normally denoted by $$f^ n(x)\equiv \operatorname{nest}(f, x, n)$$. I gather you are interested in problem solving methods over rigor.

They are normally obtained by functional iteration through Schröder's equation, but even your simple quadratic $$p(x)=x^2+c$$ does not have closed solutions, except, e.g., for $$c=-2$$, a "chaotic" logistic map , as you may see from the examples in the WP article linked above.

In that celebrated special case, a closed form (p302) was found by Ernst Schroeder himself (1870).

Namely, for
$$p(x)= x^2-2,$$ it follows directly that for $$y=\frac{x\pm \sqrt{x^2-4}}{2}$$ that is $$x=y+y^{-1},$$ one has $$p(x)=y^2+y^{-2}\equiv p^1 (x).$$ Whence,
$$p^n (x)= y^{2^n}+ y^{-2^n}.$$

More formally, in E.S.'s language of conjugacy, $$\psi(p(x))=g(\psi(x)),\\ \psi(x)=\frac{x\pm \sqrt{x^2-4}}{2}\\ g(y)=y^2 \qquad \Longrightarrow \\ g^n(y)=y^{2^n};$$ so that $$p(x)= \psi^{-1} \circ g \circ \psi (x)$$, and $$p^n= \psi^{-1} \circ g^n \circ \psi ~.$$

I am restricting this to real variables and domains where the objects treated make sense. Your particular question $$f^k (x)=x^2-2=p(x)$$ then can produce $$p^{1/k}(x)$$ for suitable domains for you to explore. There are, of course, a plethora of solution-seeking texts on the subject, like C Efthimiou's Introduction to Functional Equations, AMS 2011, ISBN: 978-0-8218-5314-6 , online. A conjugacy iteration approximation method is available in our 2011 paper: Approximate solutions of Functional equations.

• Not sure that you present Schroeder's example at the end in an optimal way, anecdotally because you seem to use the notation $f$ for another object than the OP, and mainly because your function $\psi$ is not defined unambiguously. Remember that the OP asks for some function $f:\mathbb C\to\mathbb C$ such that $f(f(z))=p(z)$ for every $z$ in $\mathbb C$, where $p:\mathbb C\to\mathbb C$ denotes the polynomial $p(z)=z^2-2$. Schroeder's approach (and everybody's approach after him) starts with the remark that, for every $z\ne0$, $$p(z+z^{-1})=z^2+z^{-2}$$ Thus, if one can find some ...
– Did
Jul 25, 2018 at 18:41
• ... function $\psi:\mathbb C\to\mathbb C^*$ such that, for every $z$ in $\mathbb C$, $$z=\psi(z)+\psi(z)^{-1}\tag{1}$$ then the function $f$ defined formally by $$f(z)=\psi(z)^{\sqrt2}+\psi(z)^{-\sqrt2}$$ is a solution. Now, this approach obviously meets some serious problems that must be solved before one can consider that it addresses the question, such as the fact that, for every complex number $z$ not in the subset $[2,+\infty)$ of the real line, no solution $\psi(z)$ of the identity $(1)$ is real and nonnegative hence, one must explain how one should define $\psi(z)^\sqrt2$, ...
– Did
Jul 25, 2018 at 18:41
• I mean, in routine physics applications, one restricts to real fields and avoids juicy definitional ambiguities of plain powers, the case for $\psi^{1/k}$... Jul 25, 2018 at 19:00
• FYI, even restricting things to $\mathbb R$, there is actually no solution to the equation $f(f(x))=x^2-2$ for every $x$ in $\mathbb R$. See there for some explanations.
– Did
Jul 25, 2018 at 19:24
• How are Julia sets related to the (in)existence of a functional square root of $x\mapsto x^2-2$?
– Did
Jul 25, 2018 at 19:48

I will discuss quadratic $$p(x)$$, although the general method applies to $$p(x)$$ with an attractive fixed point. First consider $$p(x)=x(a-x)$$, with $$|a|<1$$, then there is an attractive fixed point at $$0$$. We look for a function $$H(x)$$ such that equation 1 is obeyed:

$$H(p(x))=aH(x)$$. ( 1)

We can then put Equation 2:

$$p^n(x)=H^{-1} (a^nH(x))$$. ( 2)

It is straightforward to expand $$H(x)$$ as a power series $$H(x)= x+x^2/(a(a-1)+...$$, but the higher order terms are relatively complex rational functions of a.

One can compute $$H(x)$$ by computing $$y=p^n(x)$$ with $$n$$ sufficiently large that the power series is accurate for $$H(y)$$ then put $$H(x)=a^{-n}H(y)$$.

I did numerical experiments for $$a=1/2$$, and the taylor series for $$H(x)$$ out to order 10. Equation 1 was obeyed to high accuracy, $$H(x)$$ appears smooth on its domain, which is the basin of the 0 attractor. $$H(x)$$ is singular on the boundary of its domain, the Julia set, and I don't think it can be continued beyond there.

$$H(x)=0$$ for all 0's of $$p^n(x)=0$$ and any integer $$n$$. It appears to be analytic on its domain. Equation 2 allows us to differentiate with respect to n and create a continuous flow map in the complex plane. I have not had time to study the inverse map. Clearly it must be multiple valued, but I believe it has regions where single valued definitions exist. In particular, for $$p^n(x)=H^{-1} (a^nH(x))$$ and $$n$$ small, we can use $$x$$ as the starting point of an iterative method to find $$H^{-1} (a^nH(x))$$.

The function $$x^2+c$$ has a quadratic fixed point at $$\infty$$. In this case we transform the mapping to $$q(x)=x^2/(1+cx^2)$$ and develop a power series about $$0$$. Because the fixed point is quadratic the functional equation is different. I have not studied this case much yet.

Similar to tippy2tina's answer, one can easily numerically generate such functions on $$\mathbb R$$ when they converge monotonically to a fixed-point if a fixed-point exists. Let $$f^k(x)=\operatorname{nest}(f,x,k)$$. In this case consider for $$c\le-1/4$$:

$$p(x)=x^2+c,\\q(x)=p^{-1}(x)=\sqrt{x-c}$$

where the inverse, $$q$$, has $$q^n(x)\to\beta=\frac{1+\sqrt{1-4c}}2$$ monotonically and

$$q'(\beta)=\frac1{2\beta}$$

describing its rate of convergence

$$q^{n+1}(x)-\beta\sim\frac{q^n(x)-\beta}{2\beta}$$

which implies we can have

$$q^{n+\alpha}(x)-\beta\sim\frac{q^n(x)-\beta}{(2\beta)^\alpha}$$

and thus we may define

$$q^\alpha(x)=\lim_{n\to\infty}p^n\left(\beta+\frac{q^n(x)-\beta}{(2\beta)^\alpha}\right)$$

In your case, you desire to find

$$p^{1/k}(x)=q^{-1/k}(x)=\lim_{n\to\infty}p^n\left(\beta+\frac{q^n(x)-\beta}{(2\beta)^{-1/k}}\right)$$

which will converge to a real solution over $$[a,\infty)$$ for some $$a\le0$$. Note that $$p$$ has two inverses, so only a part of the solution is accounted for, while the other part cannot be real due to issues described in the linked posts.

Graphically for solving in the case of $$c=-3.6$$ and $$k=3:$$