8
$\begingroup$

Let $\mathfrak{I}:=\big\{ \, f:=\sum_{k=0}^\infty f_k z^k \in\mathbb{C}[[z]]\; : \text{s.t. }\; f_0=0 \;\text{ and }\; f_1=1\big\}$. A most basic result about linearization states that, for any $f\in\mathfrak{I}$ and for any $\lambda\in\mathbb{C}$ not a root of unity, there exists a unique $h\in\mathfrak{I}$ that linearizes $f(\lambda z)$: $$h(\lambda z)=f(\lambda h(z)).$$ Existence and uniqueness of this conjugation in the setting of formal power series is indeed easily established, as the coefficients of $h$ can be determined inductively from the data $(f_1,f_2,\dots) $ expanding the composition: one finds $$h_1=1$$ $$h_{n+1}={1\over \lambda^{n}-1 } \sum_{k\ge2,\,j} f_k \lambda^{k-1} h_{j_1}\dots h_{j_k},$$ the latter sum being extended over all integers $k\ge2$ and multi-indices $j\in\mathbb{Z}_+^k$ of length $|j|:=j_1+\dots+j_k=n+1$, thus with all components $j_i\le n$. The above recursion implies that the coefficients $h_n$ should be rational functions of $\lambda$, with denominators in the form $\prod_{k=1}^{n-1}(\lambda^k-1)^{m_{n,k}}$, where the exponents $m_{k,n}$ may be larger than $1$, at least at a first glance. However, after experiments with various $f$ and up to $n=50$ I observed that the form of $h_n$ always shows the simpler denominator, with all $m_{k,n}=1$: $$h_n={P_n(\lambda)\over (\lambda-1)(\lambda^2-1)\dots(\lambda^{n-1}-1)},$$ for some polynomials $P_n(\lambda)$ of degree at most ${n\choose 2}$.

Is this really a general phenomenon? If so, why? Is there a simpler formula (maybe still recursive) for $h_n$ than the above recursion?

$\endgroup$
3
  • 1
    $\begingroup$ Late comment: I've seen this as a "general phenomenon" because I got used to look at this functional selfcomposition as a diagonalization problem of Carlemanmatrices. Carlemanmatrices are a special form, associated with some function $f(x)=c_0+c_1 x + c_2 x^2 + ... $ and in a first step rewritten as dot-product $f(x) = V(x) \cdot C $ where $V(x)$ is the vector $[1,x,x^2,x^3,...]$ and $C$ is the columnvector $[c_0,c_1,c_2,...]$. The Carlemanmatrix for $f(x)$, say $F$ is now constructed so that $V(x) \cdot F = V(f(x))$ (...cntd...) $\endgroup$ Sep 30, 2022 at 6:17
  • 1
    $\begingroup$ (...) and obviously iterations of $f(x)$ are now expressible by powers of $F$. Fractional iterates of $f()$ by fractional powers of $F$, and this is, for $c_0=0$ and $c_1<>1$, easily computable by a recusive procedure. We get then $V(f°^h(x))= V(x) \cdot M \cdot D^h \cdot W$ where for easiness of notation I write $W=M^{-1}$. If I understand you correctly, $M$ is then the Carlemanmatrix for your function $h()$. Thus, this is a very general concept and the occurences of polynomials $P_n(\lambda)$ of degree at most $\binom{n}{2}$ is then the general phenomenon. (...cntd...) $\endgroup$ Sep 30, 2022 at 6:25
  • 1
    $\begingroup$ (...cntd...) For the functions $t^x-1$ (with $t \ne \exp(1)$) I have detailed this in a workout go.helms-net.de/math/tetdocs/APT.htm with explicite display of coefficients for the polynomials (of small order). (for $t=\exp(1)$ resp $u=\log(t)$ having $u=1$ this cannot be used for fractional powers, one needs then application of the matrix-logarithm, another subject, another polynomials) $\endgroup$ Sep 30, 2022 at 6:31

2 Answers 2

6
$\begingroup$

You may find useful information in a recent article by D. Sauzin and al. "Explicit linearization of one-dimensional germs through tree-expansions" here, where they use "mould calculus" (introduced by J. Écalle 40 years ago) to write down the coefficients $h_n$ and explore the combinaotrial structure of their family.

$\endgroup$
0
1
$\begingroup$

I think it is very unusual for the $h_n$s to have a simpler formulation. Even in the case of a quadratic polynomial, the conjugating function $h$ need not be a polynomial. I'm afraid I don't have a general explanation of that for you, other than noticing it by working through some examples.

There is however a simpler construction of $h$ as a limit, rather than as a power series that works when $|\lambda|\neq 1$, namely let $$h_n=\lambda^{-n}\circ f^n.$$ Then, $h_n\to h$. This formulation is fairly well known so, perhaps, you are already aware of it. It's quite easy to work with programmatically.

Note, however that the case $|\lambda|=1$ is more subtle than you've stated. The existence of the linearization is then equivalent to the the fixed point being in the Fatou set. Indeed, if $\lambda$ is a root of unity, the fixed point is in the Julia set but, even if $\lambda$ is not a root of unity the fixed point might still be in the Julia set.

$\endgroup$
3
  • $\begingroup$ Thank you, yes, I am aware of what you are saying. Here I only look at the formal series expansion, with no issue of convergence, so the only condition on $\lambda$ is not being a root of unity. (However one can assume $|\lambda|\neq 1$ if useful). $\endgroup$ May 12, 2017 at 4:59
  • 2
    $\begingroup$ How does this address the question? As I understand it was not conjectured that $h$ is a polynomial. The question was about denominators of the coefficients of $h$. $\endgroup$ May 12, 2017 at 5:59
  • $\begingroup$ @AlexandreEremenko You're right, of course. The reply started as a comment but ended up being longer than I anticipated. $\endgroup$ May 12, 2017 at 12:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.