Let $A\in M_{n}(\mathbb{Q})$, meaning that $A$ is $n\times n$ matrix with entries in the rational numbers $\mathbb{Q}$. Suppose that $A$ satisfies two conditions:

  • $\det(A)\neq 0$
  • For every integer $k$, there exists a $B\in M_{n}(\mathbb{Q})$ such that $B^k = A$.

Does it follow that $A = I_n$? Here, $I_n$ is the $n\times n$ identity matrix.

Motivation. The analogous problem where $\mathbb{Q}$ everywhere is replaced by $\mathbb{Z}$ has an affirmative answer. More precisely, let's prove the following statement:

Claim. Let $A\in M_{n}(\mathbb{Z})$, meaning that $A$ is $n\times n$ matrix with integer entries. Suppose that $\det(A)\neq 0$ and that for every integer $k$, there exists $B\in M_{n}(\mathbb{Z})$ such that $B^k = A$. Then $A = I_n$.

Proof. Since $\det(A)\neq 0$, there are only finitely many primes dividing $\det(A)$. Let $p$ be any prime which does not divide $\det(A)$. After reducing mod $p$, we still get a non-singular matrix $\overline{A}$, so $\overline{A}\in\operatorname{GL}_{n}(\mathbb{Z}/p\mathbb{Z})$. Let $k=\# \operatorname{GL}_{n}(\mathbb{Z}/p\mathbb{Z})$ be the cardinality of the group of invertible matrices with entries in $\mathbb{Z}/p\mathbb{Z}$. By hypothesis, there exists $B\in M_{n}(\mathbb{Z})$ such that $B^k=A$. After reducing mod $p$, we get $(\overline{B})^k = \overline{A}$. By Lagrange's theorem, $(\overline{B})^k = \overline{I_{n}}$. Therefore, $\overline{A} = \overline{I_{n}}$, so that each entry of $A-I_{n}$ is divisible by $p$. Since we have infinitely many choices for the prime $p$, it follows that each entry of $A-I_{n}$ must in fact be zero, yielding $A=I_n$.

Remark 1. The hypothesis $\det(A)\neq 0$ is necessary. Otherwise, we can just take any diagonal matrix $A= \operatorname{diag}(1, 1, .. 1, 0, 0 .., 0)$ with any number of $1$s and $0$s. Certainly $A^k = A$ for every $k$ but $A\neq I_{n}$ if there is at least one zero on the diagonal.

Remark 2. It is tempting to modify the above argument for the case of $\mathbb{Q}$, but it doesn't seem to work immediately. Again using the fact that $\det(A)\neq 0$, there are only finitely many primes appearing in the numerator and denominator of $\det(A)$, so we can pick a prime $p$ which doesn't appear there. Reducing the original matrix $A$ mod $p$ still works fine. But the problem arises at the step when we use hypothesis to get a matrix $B$ such that $B^k = A$ (where $k$ clearly depends on p -- this is important). That matrix $B$ might have the prime $p$ show up in the denominator of some of its entries, so reduction mod $p$ doesn't make sense.

Reference. I learnt the proof above from "Mathematical Bridges" by Andreescu, Mortici and Tetiva. The problem appears as Exercise 18 in Chapter 6, and the solution is presented few pages later. So I am also adding the (contest-math) tag, since that is the theme of the aforementioned book.

  • 6
    $\begingroup$ Try $A=\pmatrix{1&1\\0&1}$. $\endgroup$ – Lord Shark the Unknown Jul 7 '18 at 17:23
  • 1
    $\begingroup$ @LordSharktheUnknown Very nice! Indeed, if we let $B = \pmatrix{1 & \frac{1}{k} \\ 0 & 1}$, then $B^k = A$. You should post this as an answer when you get a chance. Also, I wonder if these are the only examples. In other words, do you think that with the conditions above, we can at least conclude that $A$ is unipotent? (You proved that we cannot conclude that $A$ is identity) $\endgroup$ – Prism Jul 7 '18 at 17:34
  • $\begingroup$ Let $S\in M_2(\mathbb Q)$ by any invertible matrix.The first family is $S\begin{pmatrix} 1 & a\\ 0 & 1\end{pmatrix}S^{-1}$. The second family is $S\begin{pmatrix} 2 & 1\\ -1 & 0\end{pmatrix} S^{-1}$. The first construction can be seen to extend to any $n\geq 2$. Might be the same for the second family. $\endgroup$ – Yong Hao Ng Jul 18 '18 at 11:29
  • $\begingroup$ The solutions $A = (B_k)^k$ are $$ B_k = \begin{pmatrix} 1 & a/k \\ 0 & 1\end{pmatrix} \implies (B_k)^k = \begin{pmatrix} 1 & a\\ 0 & 1\end{pmatrix} $$ and $$ B_k = \frac{1}{k}\begin{pmatrix} k+1 & 1 \\ -1 & k-1\end{pmatrix} \implies (B_k)^k = \begin{pmatrix} 2 & 1 \\ -1 & 0 \end{pmatrix} $$ respectively. This extends to solving $SAS^{-1} = (SB_kS^{-1})^k$, explaining the existence of the $S$ matrices. $\endgroup$ – Yong Hao Ng Jul 18 '18 at 11:31

The general result is as follows:

Let $A\in M_n(\mathbb{Q})$; then

For every integer $k$, there exists $B\in M_n(\mathbb{Q})$ such that $B^k=A$ IFF $A(A-I_n)^n=0$.

You can see several proofs

in english in https://artofproblemsolving.com/community/c7h42444 , grobber's post.

in french in "La revue de mathématiques spéciales" n° 117-2 , R535, where, with a colleague, we gave a solution using Mahler's measure.


For convenience, we call a matrix $A$ with solutions $A=B^k$ for all $k\geq 1$ exceptional. Notice that $$ \det(A) = \det(B)^k \implies \sqrt[k]{\det(A)} = \det(B) \in \mathbb Q $$ This is only possible if $\det(A)=1$, so we assume that.


Let $T_n(b)$ denote an $n$ by $n$ matrix with all entries zero except for top right entry $b$. Then for each $n\geq 2$ there is an infinite family of exceptional matrices in $M_n(\mathbb Q)$, of the form $$ A = S(I_n+T_n(b))S^{-1} $$ where $S$ can be any invertible matrix in $M_n(\mathbb Q)$. The solutions $B_k$ are of the form $$ B_k = S(I_n + T_n(b/k))S^{-1} $$ When $b\neq 0$, we have $A\neq I_n$, hence an infinite family of counterexamples and answering the original question. We remark that this family is unipotent.

For dimension $n=2$, we give another infinite family of exceptional matrices $$ A = S\begin{pmatrix}2 & 1\\ -1 & 0\end{pmatrix} S^{-1} $$ where $S$ can any invertible matrix in $M_2(\mathbb Q)$. It has the solutions $$ B_k = S \begin{pmatrix} \frac{k+1}{k} & \frac{1}{k}\\ \frac{-1}{k} & \frac{k-1}{k} \end{pmatrix}S^{-1} $$ We remark that this family is also unipotent.

In addition, for $n=2$ we prove that the above two are the only types possible. By applying suitable conjugation matrices, we get a reduced form where we prove that it is either of the given forms above or $A$ will fail to remain in $M_2(\mathbb Q)$ upon repeated squareroots (equivalently $A=B^{2^k}$ has no solutions for some large enough $k$), or fail $A=B^3$ for one particular case.

A. Reduction of $M_2(\mathbb Q)$ via conjugation

Let $S\in M_n(\mathbb Q)$ with $\det(S)\neq 0$, so that $S$ is invertible and $S^{-1}\in M_n(\mathbb Q)$. Then for all $k\geq 1$, $$ SAS^{-1} = S(B_k)^kS^{-1} = (SB_kS^{-1})^k $$ Therefore $A' = SAS^{-1}$ is also exceptional with solutions $SB_kS^{-1}$ for each $k\geq 1$.

Using the conjugation argument, we will show that any exceptional matrix $A= \begin{pmatrix} a & b \\ c & d \end{pmatrix}\in M_2(\mathbb Q)$ can be transformed to one of the $2$ types of (reduced) exceptional matrix: $$ \begin{pmatrix} a & b \\ 0 & 1/a \end{pmatrix}, \begin{pmatrix} a & 1 \\ -1 & 0 \end{pmatrix} $$ for $c=0$ and $c\neq 0$ respectively.

For the former case, if $a=1$ then we have the construction given at the beginning of this answer. For the latter case, if $a=2$ then we have the other construction given at the start.

For all the rest of the cases, we show that we cannot take squareroots infinitely, hence contradicting that $A$ is exceptional. (For one particular case it fails due to no solutions to $A=B^3$.)

Therefore any exceptional matrix is (conjugate)-reduced to either $A_1=\begin{pmatrix} 1 & b\\ 0 & 1\end{pmatrix}$ or $A_2=\begin{pmatrix} 2 & 1\\ -1 & 0\end{pmatrix}$. This in turn means that all exceptional matrices are generated via $SA_1S^{-1}$ or $SA_2S^{-1}$.

B. Solving repeated squareroots

We want to analyze repeated squareroots starting from $A$, which turns out to be enough to classify the matrices for $n=2$. Consider a solution to $$ R^{2^k} = A $$ for some $k\geq 1$ (we will later let $k\rightarrow \infty$). Define $R_j=R^{2^{k-j}}$, so that $$ R_{j}= (R_{j+1})^2 $$ Then we have a series of equations $$ \begin{align*} A &= R_{0}=(R_1)^2\\ R_{1} &= (R_{2})^2\\ R_{j} &= (R_{j+1})^2 \end{align*} $$ Since $\det(A)=1$, each successive squareroot must have $\det(R_j) = \pm 1$. However since each of them (except the last, $R_k=R$) is a square, it must be $\det(R_j)=1$.

Let $tr(\cdot)$ be the matrix-trace function. The key result we want to derive is that

Proposition. Let the system of equations $$ \begin{align*} A &= R_{0}=(R_1)^2\\ R_{1} &= (R_{2})^2\\ R_{j} &= (R_{j+1})^2 \end{align*} $$ have a solution for $\det(R_j)=1$ and $A,R_j\in M_2(\mathbb Q)$, as $0\leq j\rightarrow \infty$. Then $tr(R_j)$ must be integral for each $0\leq j < k$. Moreover, the $tr(R_j)$'s satisfy the recurrence $$ tr(R_{j+1})^2 = tr(R_j) + 2 $$

Proof. By the Cayley-Hamilton theorem in dimension $n=2$, for any $R_i$ we have $$ (R_i)^2 - tr(R_i)R_i + \det(R_i)I = 0 $$ Therefore a necessary condition for $R_j = (R_{j+1})^2$, together with $\det(R_{j+1})=1$, is $$ \begin{align*} R_j - tr(R_{j+1})R_{j+1} + I &= 0\\ tr(R_{j+1})R_{j+1} &= R_j + I \end{align*} $$ Let $R_j = \begin{pmatrix} a & b \\ c & d\end{pmatrix}$ and taking determinant, $$ \begin{align*} tr(R_{j+1})^2\det(R_{j+1}) &= \det\begin{pmatrix} a+1 & b \\ c & d+1\end{pmatrix} \\ tr(R_{j+1})^2 &= (a+1)(d+1) - bc \\ &= a+d+1 + (ad-bc)\\ &= a+d+2 \\ &= tr(R_j)+2 \end{align*} $$ Therefore we have a recurrence relation $$ tr(R_{j+1})^2 = tr(R_j) + 2 $$ Note that $R_j\in M_2(\mathbb Q)$ so that $tr(R_j)\in \mathbb Q$.

Suppose that $tr(R_j)\not\in\mathbb Z$ for some $j$. We write $tr(R_j) = u_j/v_j$ for some $u_j,v_j\in \mathbb Z$ and $\gcd(u_j,v_j)=1$. Then $v_j\geq 2$. Now $$ tr(R_{j+1})^2 = tr(R_j)+2 = \frac{u_j+2v_j}{v_j} $$ Since $\gcd(u_j+2v_j,v_j) = \gcd(u_j,v_j) = 1$, no further cancellation is possible and we must have $$ tr(R_{j+1}) = \pm\frac{\sqrt{u_j+2v_j}}{\sqrt{v_j}} = \frac{u_{j+1}}{v_{j+1}} $$ If $\sqrt{v_j} \not \in \mathbb Z$, then $tr(R_{j+1})\not\in\mathbb Q$, contradicting $R_{j+1}\in M_2(\mathbb Q)$. So we assume $v_{j+1} = \sqrt{v_j}\in\mathbb Z$. Likewise, we must also have $u_{j+1} = \pm \sqrt{u_j+2v_j} \in\mathbb Z$.

Since we simply took a squareroot resulting in a rational, it must be the case that $\gcd(u_{j+1},v_{j+1}) = 1$, so we can reapply the same arguments for $tr(R_{j+2}) = u_{j+2}/v_{j+2}$. i.e. $$u_{j+2} = \pm \sqrt{u_{j+1} + 2v_{j+1}} \in \mathbb Z, v_{j+2} = \sqrt{v_{j+1}} \in \mathbb Z$$

Hence by induction we can see that the denominator is squareroot-ed every iteration. If $R_j\not \in \mathbb Z$ for some $j$, then this iterative squareroot process must result in an irrational denominator eventually, i.e. $tr(R_m) \not\in \mathbb Q$ for some large enough $m>j$. This contradicts $R_m\in M_2(\mathbb Q)$, therefore it must have been the case that $R_j\in \mathbb Z$ for all $j\geq 0$. $$\tag*{$\square$}$$

Let $A=\begin{pmatrix} a & b \\ c & d\end{pmatrix}$ be exceptional. We want to classify $A$ using the conjugation and repeated squareroots arguments established earlier.

C1. Case $n=2$ and $c=0$

First, we look at the case $c=0$. Since $\det(A)=1$, we get $d=1/a$, giving us
$$ A= \begin{pmatrix} a & b \\ 0 & 1/a \end{pmatrix} $$ Since $A$ is exceptional, we can take squareroots infinitely. By the proposition in B., we must have $tr(A) \in \mathbb Z$, say $tr(A) = t\in\mathbb Z$.

Write $a = u/v$ with $u,v\in\mathbb Z$ and $\gcd(u,v)=1$. Then $$ \begin{align*} \frac{u^2+v^2}{uv} &= a + \frac{1}{a} = t\in \mathbb Z\\ u^2+v^2 &= tuv \end{align*} $$ This shows that $u$ divides $v$ and vice-versa, which is only possible if $u,v = \pm 1$. Therefore $a=1$ or $a= -1$. If $a=-1$, then $tr(A) = -2$. Taking the first squareroot as $R_1$, the formula gives $$ tr(R_1)^2 = tr(A) +2 = 0 \implies tr(R_1) = 0 $$ But now for the second squareroot $R_2$ we have $$ tr(R_2)^2 = tr(R_1)+2 = 2 \implies tr(R_2) = \pm \sqrt{2}, $$ contradicting $tr(R_2)\in\mathbb Q$. Therefore the only possibility is $a=1$. We have seen that this results in the family of solution given at the start of this answer.

C2. Case $n=2$ and $c\neq 0$

Choosing the conjugation matrices $$ S= \begin{pmatrix} 0 & 1 \\ c & -a \end{pmatrix}, S^{-1} = \begin{pmatrix} a/c & 1/c \\ 1 & 0 \end{pmatrix}, $$ we get a new exceptional matrix $A'$ $$ A'= SAS^{-1} = \begin{pmatrix} a+d & 1 \\ bc-ad & 0 \end{pmatrix} = \begin{pmatrix} a+d & 1 \\ -1 & 0 \end{pmatrix} $$ where the last equality is due to the condition $\det(A)=1$.

Denote $r_0 = tr(A) = a+d$ and similarly let $r_j = tr(R_j)$, where $R_j$'s are the matrices obtained by repeated squareroots as in B.. By proposition B. again, we must have $r_j\in\mathbb Z$ and satisfy the recurrence $$ r_{j+1}^2 = r_j + 2 \implies r_{j+1} = \pm \sqrt{r_j + 2} $$ Since $r_{j+1}\in\mathbb Z$ clearly $r_j\geq -2$ for each $j$.

For $r_0=2$, corresponding to $A=\begin{pmatrix} 2 & 1\\ -1 & 0\end{pmatrix}$, we have seen at the start of the answer that it is an exceptional matrix.

For $r_0 = -2,0,1$, we get $r_1 = 0, \pm \sqrt{2}, \pm \sqrt{3}$ respectively. $r_1=0$ in turn gives $r_2 = \pm \sqrt{2}$, so all of them contradicts $r_j\in\mathbb Z$.

For $r_0 > 2$, then $r_1^2 > 4$. Since $r_j > -2$ we must have $r_1>2$ again, so by induction $r_j > 2$ for all $j$. The sequence of $r_{j+1} = \sqrt{r_j+2}$ becomes smaller than $3$ for some sufficiently large $m$, then $2 < r_m < 3$ results in $$ r_{m+1}^2 = r_m + 2 \implies 4 < r_{m+1}^2 < 5 $$ which contradicts the fact that $r_{m+1}\in\mathbb Z$.

Since $r_0\in\mathbb Z$, this leaves the final case of $r_0 = -1$. For this case we use a numeric solver to show directly that there are no solutions to $$ B^3 = \begin{pmatrix} -1 & 1\\ -1 & 0 \end{pmatrix} $$ Edit 1: Algebraic proof
Let $B$ be a solution to $A=B^3$ and $t=tr(B)$. Then $$\det(B)^3 = \det(A) = 1$$ Since $\det(B)\in\mathbb Q$ we conclude $\det(B)=1$. By Cayley-Hamilton: $$ \begin{align*} B^2-tB+I &= 0\\ B^3-tB^2+B &= 0\\ A-t(tB-I)+B &= 0\\ (t^2-1)B &= A+tI = \begin{pmatrix} -1+t & 1\\ -1 & t\end{pmatrix}\\ (t^2-1)^2 \det(B) &= \det(A+tI) = (t-1)t + 1\\ t^4-2t^2+1 &= t^2-t+1\\ t(t^3-3t+ 1) &= 0 \end{align*} $$ Since $t^3-3t+1$ has no solutions $\pmod 2$ it is irreducible in $\mathbb Z$ and hence also irreducible in $\mathbb Q$. Since $t\in \mathbb Q$, the only possibility is $t=0$. Therefore $-B = A\implies B = -A$. However we can check manually that $B^3 = -A^3 = -I \neq A$, hence there cannot be a solution $B$.
End edit 1

Hence, going through all integer possibilities of $r_0=a+d$, we see that only $r_0=2$ results in an exceptional matrix.

This completes all the cases.

  • $\begingroup$ Wow, what an impressive work! Thank you so much. There are many cool ideas here (my favorite is the trick using Cayley-Hamilton in Edit 1). $\endgroup$ – Prism Jul 18 '18 at 21:55
  • $\begingroup$ @Prism Thanks for the kind words, I apologize for the wall of text since it feels like they could have been more concise. I just noticed that since conjugation $SAS^{-1}$ does not change the trace, in some sense this is proving that the requirement is exactly just $tr(A)=2$. The unipotent condition then comes from the characteristic polynomial: $$ A^2 -2A+I = 0 \implies (A-I)^2 = 0 $$ Perhaps solving for trace directly might a more useful idea for higher dimensions, since the current method is too computational. $\endgroup$ – Yong Hao Ng Jul 19 '18 at 3:09

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.