Eigenvalues of $AB$ and $BA$ where $A$ and $B$ are square matrices

Show that if $$A,B \in M_{n \times n}(K)$$, where $$K=\mathbb{R}, \mathbb{C}$$, then the matrices $$AB$$ and $$BA$$ have same eigenvalues.

I do that like this:

let $$\lambda$$ be the eigenvalue of $$B$$ and $$v\neq 0$$

$$ABv=A\lambda v=\lambda Av=BAv$$

the third equation is valid, because $$Av$$ is the eigenvector of $$B$$. Am I doing it right?

• It looks correct and the right approach. Jun 5, 2014 at 16:43
• @DonAntonio The proof isn't correct.
– user63181
Jun 5, 2014 at 16:47
• I can't see any proof, leave alone it is correct or not, @user63181. Yet I can see now that I misread and the OP was attempting something that doesn't seem to help him to prove what he asked. Jun 5, 2014 at 16:49
• I think the OP was attempting the argument given here math.stackexchange.com/a/124903/49610 As for showing that $AB$ and $BA$ have the same characteristic polynomial, there's a beautiful, purely algebraic proof, in an answer I can't find for the life of me, using the fact that $$\begin{pmatrix}AB-t & 0 \\ 0 & -t\end{pmatrix}, \quad \begin{pmatrix}BA-t & 0 \\ 0 & -t \end{pmatrix}$$ can be cleverly factored as products of the same two matrices. Jun 5, 2014 at 17:17
• Also have a look at the following interesting blog post by Qiaochu Yuan:qchu.wordpress.com/2012/06/05/ab-ba-and-the-spectrum/… Jun 5, 2014 at 19:18

Here is a proof similar to what the OP has tried:

Let $\lambda$ be any eigenvalue of $AB$ with corresponding eigenvector $x$. Then

$$ABx = \lambda x \Rightarrow \\ BABx = B\lambda x \Rightarrow\\ BA(Bx) = \lambda (Bx)$$

which implies that $\lambda$ is an eigenvalue of $BA$ with a corresponding eigenvector $Bx$, provided $Bx$ is non-zero. If $Bx = 0$, then $ABx = 0$ implies that $\lambda = 0$.

Thus, $AB$ and $BA$ have the same non-zero eigenvalues.

• Can something similar be said about singular-values? Update: math.stackexchange.com/questions/2448088/… Sep 27, 2017 at 20:08
• Your proof has no interest because you don't show the equality of multiplicities of the eigenvalues of $AB,BA$. Moreover I don't see why the eigenvalue $0$ is a problem: if $\det(AB)=0$, then $\det(BA)=0$ too.
– user91684
Apr 13, 2018 at 9:40
• @loupblanc My primary aim was to give a proof similar to the one OP attempted. About the zero eigenvalue problem, yes, if $A$ and $B$ are square matrices, that's true. But the proof I've shown is more general and works even if $A$ and $B$ are rectangular such that $AB$ and therefore $BA$ are square. Apr 14, 2018 at 11:57
• Remark that since the OP gave the green chevron to a complete solution, it's because he (she) understood that he was on a wrong path. Note also that to show the complete version of the result is natural because, generically, $AB$ and $BA$ are similar. The purpose of my post was not to annoy you but rather to point out that in the problems where this result is used, one always needs to consider the full spectrum (with multiplicity). Then, I propose that those who upvoted your post should not be allowed to use the considered result in its complete form.
– user91684
Apr 14, 2018 at 13:01
• Let $A=\begin{bmatrix}0&1\\0&0\end{bmatrix}$, $B=\begin{bmatrix}0&0\\0&1\end{bmatrix}$. Then $AB = A \ne 0$, but $BA = 0$. So $AB$ and $BA$ are not always similar. Nevertheless, you are right that they will always have the same spectrum. I will attempt to update my answer with a proof of this (one that has not been given in any of the other answers). I think the best proof is Alternative proof #2 in math.stackexchange.com/a/822198/152030. Apr 15, 2018 at 4:07

It suffices to show that $AB$ and $BA$ have the same characteristic polynomial. First assume that $A$ is invertible then

$$\chi_{AB}(x)=\det(AB-xI)=\det A\det(B-xA^{-1})\\=\det(B-xA^{-1})\det A=\det(BA-xI)=\chi_{BA}(x)$$ Now since $\operatorname{GL}_n(K)$ is dense in $\operatorname{M}_n(K)$ then there's a sequence of invertible matrices $(A_n)$ convergent to $A$ and by the continuity of the $\det$ function we have $$\chi_{AB}(x)=\det(AB-xI)=\lim_{n\to\infty}\det(A_nB-xI)=\lim_{n\to\infty}\det(BA_n-xI)\\=\det(BA-xI)=\chi_{BA}(x).$$

• While this proof works, I think there is a little bit too much machinery at play here, as the original problem is probably intended for a student who has just started learning about eigenvalues. Jun 5, 2014 at 21:13

Alternative proof #1:

If $n\times n$ matrices $X$ and $Y$ are such that $\mathrm{tr}(X^k)=\mathrm{tr}(Y^k)$ for $k=1,\ldots,n$, then $X$ and $Y$ have the same eigenvalues.

See, e.g., this question.

Using $\mathrm{tr}(UV)=\mathrm{tr}(VU)$, it is easy to see that $$\mathrm{tr}[(AB)^k]=\mathrm{tr}(\underbrace{ABAB\cdots AB}_{\text{k-times}}) =\mathrm{tr}(\underbrace{BABA\cdots BA}_{\text{k-times}})=\mathrm{tr}[(BA)^k].$$ Now use the above with $X=AB$ and $Y=BA$.

Alternative proof #2:

$$\begin{bmatrix} I & A \\ 0 & I \end{bmatrix}^{-1} \color{red}{\begin{bmatrix} AB & 0 \\ B & 0 \end{bmatrix}} \begin{bmatrix} I & A \\ 0 & I \end{bmatrix} = \color{blue}{\begin{bmatrix} 0 & 0 \\ B & BA \end{bmatrix}}.$$ Since the $\color{blue}{\text{red matrix}}$ and the $\color{red}{\text{blue matrix}}$ are similar, they have the same eigenvalues. Since both are block triangular, their eigenvalues are the eigenvalues of the diagonal blocks.

• Your second proof is very interesting, and it shows that for every commutative ring with identity $|Ex-AB|=|Ex-BA|$. But this proof is tricable and it is not clear how you guess for it. So I will be grateful if you give suggestive reasoning. May 19, 2017 at 18:10
• @MikhailGoltvanitsa I think it is based on Schur Decomposition. A similar technique is used in this post. Sep 22, 2018 at 14:16
• @IshanSingh, Thank you very much! Sep 22, 2018 at 18:17

If $A$ is invertible, user63181 already showed that

$$\det(AB-xI)=\det(BA-xI)$$

We now prove this equality in general.

Fix $x$

Let $$P_x(y):=\det[(A-yI)B-xI]-\det[B(A-yI)-xI] \,.$$

Then $P_x(y)$ is a polynomial of degree at most $n$ in $y$. Whenever $y$ is not an eigenvalue of $A$ the matrix $A-yI$ is invertible, thus by the first pat $P_x(y)=0$. Hence $P_x$ has infinitely many roots, and hence $P_x \equiv 0$.

This proves that $P_x(0)=0$ which is exactly what you need to prove.

• Very nice and algebraic solution! May 19, 2017 at 18:01

Here is a more "algebraic" approach from the other answers by user63181 and N. S. which, as far as I can see generalizes to other fields (where the continuity argument might fail) although I thought the argument by continuity was cool!

First we use the following characterization of non-zero eigenvalues: $\lambda$ is an eigenvalue for $AB$ iff $I - \lambda AB$ is not invertible.

I claim the following: $I - \lambda AB$ is (not) invertible iff $I - \lambda BA$ is (not) invertible.

Proof: Suppose $I - \lambda AB$ is invertible, then let $U := 1 + \lambda B (I - \lambda AB)^{-1}A$. Now show that $U$ is an inverse to $I - \lambda BA$ (by multiplying out and using distributivity). The converse direction follows by letting $V := 1 + \lambda A(1-\lambda BA)^{-1}B$.

From this it follows that $AB$ and $BA$ have the same non-zero eigenvalues.

• Your proof has no interest because you don't show the equality of multiplicities of the eigenvalues of $AB,BA$. Moreover I don't see why the eigenvalue $0$ is a problem: if $\det(AB)=0$, then $\det(BA)=0$ too.
– user91684
Apr 13, 2018 at 9:41

Observe that there is some $$v$$ such that $$ABv=\lambda v$$ iff there are $$v,w$$ that solve the linear system: $$\begin{cases}Aw=\lambda v, \\Bv=w.\end{cases}\tag1$$ On the other hand, from (1), if $$\lambda\neq0$$, one also immediately gets $$BAw=\lambda w$$.

It follows that $$ABv=\lambda v$$ for some $$\lambda\neq0$$ iff $$BAw=\lambda w$$, where $$w=Bv$$ (and thus also $$v=\frac1\lambda Aw$$).

Example

For example, consider $$A\equiv \begin{pmatrix}1&0\\0&0\end{pmatrix}, \qquad B\equiv \frac12\begin{pmatrix}1&1\\1&1\end{pmatrix}.$$ Then $$AB = \frac12\begin{pmatrix}1&1\\0&0\end{pmatrix}, \qquad BA = \frac12\begin{pmatrix}1&0\\1&0\end{pmatrix}.$$ Thus $$AB e_1=\frac12 e_1$$ and $$AB(e_1-e_2)=0$$, while $$BA(e_1+e_2)=\frac12(e_1+e_2)$$ and $$BA e_2=0$$. And we can notice how the nonzero eigenvectors of $$AB$$ are related as spelled out before: $$Be_1= \frac12(e_1+e_2)$$, and $$A(e_1+e_2)=e_1$$.

On the other hand, this same example shows how the result fails for zero eigenvalues. It remains true that both matrices have kers of the same dimensions, but the correponding eigenvectors are not related as in the nonzero eigenvalues case. Here the ker of $$AB$$ is spanned by $$e_1-e_2$$, but $$B(e_1-e_2)=0$$, which thus clearly does not give the ker of $$BA$$, which is spanned by $$e_2$$.

I'd also observe here that the reason we don't get the stated relation between eigenvectors of $$AB$$ and $$BA$$ is that we can have situations where $$ABv=0$$ because $$Bv=0$$ (which is also what happens in the example). If however we have $$ABv=0$$ but $$Bv\neq0$$, then we can see that we recover (at least partially) the result, because $$(BA)(Bv)=0$$.