a set of pairwise commuting polynomials

Question

The set $\{1,x,x^2,x^3,\dots\}$ has the following properties:

• any two of these polynomials commute

• for any $n\ge 0$ there is exactly one polynomial of degree $n$ inside

Is this the only example? If not, is there a general form of such sets? I was thinking of this problem during a long flight, it seems very natural and might have been investigated, but I didn't found this result so far.

---

The above question has a negative answer - as flawr mentioned (see below), the Chebyshev polynomials of the first kind are another example.

Also, Chris Culter pointed out, that we can replace any polynomial $f$ in our "good" set with a conjugate of the form $\sigma f \sigma^{-1}$, where $\sigma(x)=ax+b$ with $a\neq 0$.

---

So now, a conjecture:

$\{1,x,x^2,x^3,\dots\}$ and $\{T_0,T_1,T_2,\dots\}$ and the conjugates among them are the only examples.

---

---

Is the above conjecture true? Is such a thing already known?

---

---

Edit after a clarification request:

The field is $\mathbb{R}$.

I consider polynomials of real coefficients and one real variable.

Polynomials $P$ and $Q$ commute, if $P(Q(x))=Q(P(x))$ for every $x\in\mathbb{R}$.

Answer 1

Yes, every such polynomial chain is similar to (exactly one of) the power polynomials $\{x^1,x^2,\ldots\}$ or the Chebyshev polynomials $\{T_1, T_2,\ldots\}$.

---

Doing some internet digging, I found that the result was discovered by Block and Thielman; there's a nice overview of commuting polynomials, leading up to the result, in this article: https://digitalworks.union.edu/cgi/viewcontent.cgi?article=1784&context=theses

The theorem itself appears on p33, as the Block-Thielman theorem. I present my own streamlined version of the results below.

In detail, you can establish that:

1. For any quadratic polynomial $q(x)$, and for every degree $k$, there is at most one degree-$k$ polynomial that commutes with $q$.

   It follows that a chain of polynomials (which mutually commute and have exactly one term of each degree) is uniquely determined by its quadratic term (!).

2. Every quadratic polynomial is similar to a unique monic polynomial of the form $x^2+c$.

3. In a chain, the value of $c$ must be equal to 0 or -2. In the first case, the chain is similar to the power chain. In the second case, it's similar to the Chebyshev chain.

Here's a sketch of how to prove #3 once you know #1 and #2. Take a chain $\{p_1, p_2, p_3, \ldots\}$, which by #1 we know is uniquely determined by its quadratic $p_2$. By #2, $p_2$ is similar to a monic quadratic of the form $x^2+c$. Apply that similarity transform to every term in the chain, obtaining a new chain:

$$\{q_1, x^2+c, q_3, q_4, \ldots\}$$

and consider the degree-three term $q_3$. Because $q_3$ commutes with $x^2 + c$, we get a functional equation stating that the following two polynomials are equal everywhere:

$$q_3\circ q_2 = q_2\circ q_3$$

When two polynomials are equal, their coefficients must be equal. Hence we have a system of six equations relating $c$ to the coefficients of $q_3(x)$. ($q_m\circ q_n$ has degree $m+n$.)

The equations simplify. In fact, we can use the fact that $q_3$ commutes with $x^2+c$ to solve for all of $q_3$'s coefficients (see below). The rest of the system tells us that $c$ was in fact highly constrained:

$$c(c+2) = 0$$

That is, $c=0$ or $c=-2$. At a higher level, this shows that a chain of polynomials can only exist when its quadratic term is similar to $x^2$ or to $x^2-2$— nothing else.

In the first case, the quadratic is similar to $x^2$, so the chain is similar to the power chain (by uniqueness #1). In the second case, the quadratic is similar to $x^2-2$. This is in turn similar to the second Chevyshev polynomial $T_2(x)=2x^2-1$, through the similarity $\alpha(x)=2x$, so the chain is similar to the Chebyshev chain.

---

We can derive the coefficients of $q_3$ from the fact that it commutes with $x^2+c$. Take note of the following facts:

4. For even polynomials $f(-x)=f(x)$, all the odd coefficients vanish. For odd polynomials $f(-x)=-f(x)$, all the even coefficients vanish. (N.B. Not all polynomials are odd or even functions.)

5. If a polynomial commutes with a degree $n$ monic polynomial, then its leading coefficient must be a root of $x^{n-1}-1$.

Then here's what we know about $q_3$:

• It commutes with $(x^2+c)$, obeying the equation $$q_3\circ(x^2+c) = (x^2+c)\circ q_3$$

• By #5, its leading coefficient must be 1 (the only root of $x-1$).

• By expanding out the equation, we find that $q_3(x)^2 \equiv q_3(-x)^2$. Hence $q_3(x)$ must be an odd or even function $=\pm q_3(-x)$. But it's cubic, so its degree-3 term can't vanish (see #4) — we conclude it's odd.

• It's an odd function, so its only nonzero coefficients are its leading coefficient and its linear coefficient. (See #4)

• The linear coefficient $b_1$ is the only unknown; our system lets us solve for it, finding $3c = 2b_1$.

---

Answer 2

the examples you mentioned are not the only examples: Consider the Chebyshev Polynomials $\{T_n \mid n \in \mathbb N\}$. These have the property that

$$T_n(T_m(x)) = T_{nm}(x)$$

and therefore they commute, but they are not repeated compositions of themselves.

Answer 3

You could replace each polynomial $f(x)$ in your set with the conjugate $f'(x)=f(x-1)+1$. This will yield $\{2,x,x^2-2x+2,x^3-3x^2+3x,\ldots\}$.

A bit more generally, let $\sigma(x)=ax+b, a\neq0$ be an invertible polynomial. Then you can conjugate like so: $f'(x)=\sigma(f(\sigma^{-1}(x)))$.

Answer 4

The answer is YES to all your questions : yes it is already known, and I'm pretty certain that I've already seen here on MSE if not a proof, at least a mention of it. But I can't seem to find it back so I rewrite a proof here.

Let $(f_k)_{k\geq 0}$ be a sequence of polynomials with $\deg(f_k)=k$ and $f_i(f_j(x))=f_j(f_i(x))$ for any $i,j$. Using a suitable conjugation, we may assume

$$f_0(x)=1\tag{1}$$

Then, for any $i$ we have $f_i(1)=1$, or in other words

$$f_i(x)=1+(x-1)g_i(x)\tag{2}$$

for some $g_i$ of degree $i-1$. There are two constants $c$ and $d$ such that $g_2(x)=cx+d$ (note that $c\neq 0$). It follows that $f_2(x)=cx^2+(d-c)x+1-d$, and hence $f_1(f_2(x))=g_1f_2(x)+1-g_1=(cg_1)x^2+g_1(d-c)x+1-g_1d$ while $f_2(f_1(x))=(cg_1^2)x^2+g_1(d+c-2g_1)x+(g_1^2-g_1)c+1-g_1d$. Comparing the leading coefficients and using $cg_1\neq 0$, we deduce that $g_1=1$, so

$$ f_1(x)=x \tag{3} $$

There are three constants $k,l,m$ such that $g_3=kx^2+lx+m$. It follows that $f_3(x)=kx^3 + (l-k)x^2 + (m-l)x + 1-m$, and hence

$$ \begin{array}{lcrrrr} kf_2(x)^3 & = & (kc^3)x^6 + 3kc^2(d-c)x^5 & + 3kc(c^2+d^2-3cd+c)x^4 & + k(d-c)(c^2+d^2-8cd+6c)x^3 & + \ldots \\ (l-k)f_2(x)^2 & = & & + (l-k)c^2 x^4 & + 2c(l-k)(d-c)x^3 & + \ldots \\ f_3(f_2(x)) & = & (kc^3)x^6 + 3kc^2(d-c)x^5 & + c\big(lc+k(3c^2+3d^2-9cd+2c)\big) x^4 & + (d-c)\big(2cl+k(c^2+d^2-8cd+4c)\big)x^3 & + \ldots \\ \end{array} $$

Similarly,

$$ \begin{array}{lcrrr} cf_3(x)^2 & = & (k^2c)x^6 + 2kc(l-k)x^5 + c(k^2+l^2+2mk-4kl)x^4 & + 2c(-l^2+(k+m)l+k(1-2m))x^3 & + \ldots \\ (d-c)f_3(x) & = & & + k(d-c)x^3 & + \ldots \\ f_2(f_3(x)) & = & (k^2c)x^6 + 2kc(l-k)x^5 + c(k^2+l^2+2mk-4kl)x^4 & + \big(kd+2c(-l^2+(k+m)l+k(\frac{1}{2}-2m))\big)x^3 & + \ldots \\ \end{array} $$

We thus obtain the system

$$ \left\lbrace \begin{array}{lcl} kc^3 &=& k^2c \\ 3kc^2(d-c) &=& 2kc(l-k) \\ c\big(lc+k(3c^2+3d^2-9cd+2c)\big) &=& c(k^2+l^2+2mk-4kl) \\ (d-c)\big(2cl+k(c^2+d^2-8cd+4c)\big) &=& kd+2c(-l^2+(k+m)l+k(\frac{1}{2}-2m)) \\ \end{array}\right. $$

Remembering that $ck\neq 0$, the first line yields $k=c^2$, the second then yields $l=\frac{c}{2}(3d-c)$, and the third yields $m=\frac{-c^2+6(1-d)c+3d(d+2)}{8}$. The fourth equation then becomes equivalent to $c^5 + (3d - 6)c^4 + (3d^2 - 12d + 8)c^3 + (d^3 - 6d^2 + 8d)c^2=0$, which nicely factorizes as $c^2(d+c-4)(d+c-2)(d+c)=0$. So $d$ must be one of $4-c,2-c$ or $-c$.

When $d=4-c$, putting $\sigma(x)=\frac{(4-d)x-(2-d)}{2}$, we have $\sigma f_2\sigma^{-1}(x)=2x^2-1$ and $\sigma f_3\sigma^{-1}(x)=4x^3-3x$.

When $d=2-c$, putting $\sigma(x)=(2-d)x-(1-d)$, we have $\sigma f_2\sigma^{-1}(x)=x^2$ and $\sigma f_3\sigma^{-1}(x)=x^3$.

When $d=-c$, putting $\sigma(x)=d(1-x)$, we have $\sigma f_2\sigma^{-1}(x)=x^2$ and $\sigma f_3\sigma^{-1}(x)=x^3$.

So we have shown that (up to conjugation) we can assume either (a) $f_2(x)=x^2, f_3(x)=x^3$, (b) $f_2(x)=T_2(x)$, $f_3(x)=T_3(x)$.

Suppose we are in case (a). Let $n\geq 4$. Then $f_n(x^2)=f_n(x)^2$. Comparing the leading coefficients, we see that $f_n$ is monic. Let $z\in{\mathbb C}$ be a complex root of $f_n$. Since $f_n(x^2)=f_n(x)^2$, we see that $z^2$ is a root of $f_n$ also, and by induction every $z^{2^m}$ for $m\geq 1$ is a root of $f_n$. Since there are only finitely many roots, by the pigeonhole principle we can find $m\lt m'$ such that $z^{2^m}=z^{2^{m'}}$, and if $z\neq 0$ then $z$ must be a root of unity. But a root of unity is of the form $exp(2pi r)$ where $r \in {\mathbb Q}/{\mathbb Z}$. Denote by $R$ the (possibly empty) subset of ${\mathbb Q}/{\mathbb Z}$ of all arguments that can be obtained in this way. Because of $f_n(x^2)=f_n(x)^2$, $R$ is stable by division by $2$, so it is either empty or infinite (which is impossible). Thus $0$ is the only complex root of $f_n$, $f_n(x)=x^n$ and we are done in this case.

Suppose we are in case (b). Let $n\geq 4$. Then $f_n(T_2(x))=T_2(f_n(x))$. Comparing the leading coefficients, we see that the leading coefficient of $f_n$ is $2^{n-1}$. Write $f_n(x)=\sum_{k=0}^n a_kx^{n-k}$ (so $a_0=2^{n-1}$). Then, for $1\leq d \leq n$ in the expansion of $f_n(T_2(x))-T_2(f_n(x))$, the coefficient of degree $2n-d$ in $x$ is of the form $-4a_0a_d+h(a_0,a_1,\ldots,a_{d-1})$ where $h$ is a polynomial in $a_0,a_1,\ldots,a_{d-1}$. Thus the value of $a_d$ can be deduced from the values of $a_0,a_1,\ldots,a_{d-1}$. It follows that there is at most one polynomial of degree $n$ commuting with $T_2$. But we already know that $T_n$ has this property, so $f_n=T_n$ and we are done.