MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Some background: my coauthors and I are working on a problem which deals with the exponential growth rates of certain infinite products of matrices. One of the sub-problems which arises in this project is to prove that a particular function from the unit interval to the reals, which describes the growth rates of a family of related infinite matrix products, does not have any line segments in its graph. Now, the existence of a line segment would imply the existence of a pair of matrices some of whose characteristics coincide in a precise way. We have found an analytic proof that the graph cannot contain line segments, but it is long and displeasingly ad-hoc, and we would rather like to replace it with a shorter argument! Hence, we have found ourselves wondering whether there is an algebraic obstacle which would prevent these coincidences from occurring.

The question which arises from this research, then, is:

Do there exist matrices $A,B \in SL(2,\mathbb{Z})$ which satisfy the following constraints: the matrices $A$ and $B$ have only positive entries, do not commute, and satisfy $$\rho(AB)=\rho(A)\rho(B),$$ $$\mathbb{Q}\left(\rho(A)\right)=\mathbb{Q}\left(\rho(B)\right) \neq \mathbb{Q},$$ where $\rho$ denotes spectral radius?

The non-existence of such a pair of matrices would imply the non-existence of the line segments mentioned above.

One of my coauthors suspects that the spectral radius constraints on $A$ and $B$ cannot be met unless $A=C^k$, $B=C^\ell$ for some $C \in SL(2,\mathbb{Z})$ and $k,\ell \in \mathbb{N}$, which would contradict the hypothesis that $A$ and $B$ do not commute. Any solutions which end up being used in our paper will of course be gratefully acknowledged!

share|cite|improve this question
Some thoughts. I believe the second condition requires that tr A = tr C^k and tr B = tr C^l for some C in SL(2, Z). If we then write A = C^k + X, B = C^l + Y where X, Y are traceless, the first condition is equivalent to the statement that X C^l + C^k Y + XY has trace zero. I think I can show this is impossible if X = 0. – Qiaochu Yuan Jun 8 '10 at 16:00
Ian Agol's answer has been incorporated into the article "On a Devil’s staircase associated to the joint spectral radii of a family of pairs of matrices" which will be published in Journal of the European Mathematical Society in late 2012. – Ian Morris Dec 14 '11 at 11:06
up vote 10 down vote accepted

This is true: if one has two matrices $A, B \in SL_2 C$, and one knows the three traces $tr(A),tr(B),tr(AB)$, then this uniquely determines the matrices $A,B$ up to conjugacy if $A$ and $B$ generate a non-elementary discrete group (except in a few degenerate cases which won't occur for positive matrices). See for example Maclachlan-Reid. Since $tr(X) = \rho(X)+\rho(X)^{-1}$ for a positive matrix $X$, this implies that the three matrices have the same traces as two commuting matrices with the same traces, and therefore $A,B$ must be conjugate to a pair of commuting matrices. In fact, as your coauthor suspects, then they must be powers of the same matrix.

Addendum: I'll give some details of the argument, and I forgot a condition ($A$ and $B$ might generate a solvable group otherwise, although this is not possible for discrete groups).

Assume that $A, B \in SL_2 Z$. We want to show that $tr(A),tr(B),tr(AB)$ uniquely determine $A$ and $B$ up to conjugacy in $SL_2 C$.

Since $A$ is a positive matrix in $SL_2 Z$, we may conjugate it in $SL_2 R$ to be of the form $$ A=\begin{pmatrix} \lambda & 0 \\\ 0 & \lambda^{-1} \end{pmatrix} $$ where $\lambda >1$ (because $tr(A)>2$). We conjugate $B$ by the same matrix, and relabel so that $$ B=\begin{pmatrix} a & b \\\ c & d \end{pmatrix}. $$

If $A$ and $B$ commute, then we check that $b=c=0$, so in particular $ad=1$. Conversely, if $ad=1$, then we conclude that $bc=0$, and therefore $b=c=0$ since $A,B$ cannot generate a noncommutative solvable group since $SL_2 Z$ is discrete (see below).
From now on, let's assume that $A$ and $B$ don't commute.

The centralizer of $A$ in $SL_2 C$ consist of matrices of the form $$ \begin{pmatrix} z & 0 \\\ 0 & z^{-1} \end{pmatrix}. $$

We may conjugate $B$ by the centralizer of $A$ to be of the form $$ B=\begin{pmatrix} a & 1 \\\ ad-1 & d \end{pmatrix}. $$ Here I'm making use of the fact that $A,B$ are conjugate into $SL_2 Z$ to conclude that they don't generate a solvable group. This follows from the Margulis lemma (or an easy exercise). If $B$ had off-diagonal entries which were zero, then $\langle A, B\rangle$ would be solvable. The point is that then we may conjugate by elements of the centralizer of $A$ in $SL_2 C$ to arrange the upper right entry to be 1.

Thus, we have $$tr(A)=\lambda+\lambda^{-1}, tr(B)=a+d, tr(AB)=\lambda a +\lambda^{-1} d.$$

Now we compute $$a= \frac{\lambda tr(AB)-tr(B)}{\lambda^2-1} , d= \frac{\lambda^{-1} tr(AB)-tr(B)}{\lambda^{-2}-1}.$$

Thus, $tr(A), tr(B), tr(AB)$ canonically determine $A,B$ up to conjugacy (of course $\lambda$ is uniquely determined by $tr(A)$ given $\lambda >1$).

Now, we need to check that $\rho(AB) \neq \rho(A)\rho(B)$, where $\rho(A)=\lambda, \rho(B)=\mu$. If not, $tr(AB)=\lambda\mu +(\lambda\mu)^{-1}$ (again since $AB$ is positive). Plugging into the above equations, we conclude that $a=\mu, d=\mu^{-1}$, a contradiction.

I'm sure this argument can be simplified, but I wanted to carry through the sketch of the argument I gave the first go around.

Also, one may see that any subgroup of commuting matrices in $SL_2 Z$ is either powers of a parabolic element, a hyperbolic element, or a finite-order element.

share|cite|improve this answer
Thanks! Could you provide a reference for the general result which you use here (or a name or keyword, so I can look it up myself)? – Ian Morris Jun 8 '10 at 17:00
I think the statement is false in SL_2(C); in the setup in my comment above one can take k = l = 1, C = [[2 3][3 5]], X = 0, and Y = [[b+c b][c -b-c]] where b^2+3bc+c^2 = 0. – Qiaochu Yuan Jun 8 '10 at 17:42
@Qiaochu: Did you check that your pairs of matrices are nonconjugate? Remember, there is a 1-parameter family of matrices commuting with C. I think you've found a parametrization of the matrices obtained by conjugating B by this family (although I didn't check your formulae). – Ian Agol Jun 9 '10 at 23:25
Qiaochu's examples don't commute. Also, he might as well take the square root of C, [[1,1][1,2]]. – Ben Wieland Jun 10 '10 at 0:33
Agol: thank you for this excellent and very useful extended answer! – Ian Morris Jun 11 '10 at 12:33

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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