# $\det(AB-BA)=0$?

Let $A,B\in M_{n}(\mathbb{Q})$. If $(A-B)^2=AB$, prove that $\det(AB-BA)=0$.

I considered the function $f:\mathbb{Q}\rightarrow \mathbb{Q}$, $f(x)=\det(A^2+B^2-BA-xAB)$ and I obtained that: $$f(0)=\det(A^2+B^2-BA)=\det(2AB)=2^n\det(AB)$$ $$f(1)=\det(A^2+B^2-BA-AB)=\det((A-B)^2)=\det(AB)$$ $$f(2)=\det(A^2+B^2-BA-2AB)=\det((A-B)^2-AB)=\det(AB-AB)=0$$

I don't have any other idea.

• Yes so how does that equal $AB$ though? – K Split X Apr 18 '17 at 23:39
• @KSplitX We're assuming it is. In this case If $P$ then $Q$ where $P: (A-B)^2=AB$ and $Q: det(AB-BA)=0$. – Sentinel135 Apr 18 '17 at 23:45
• You can easily find such matrices over $\mathbb{R}$, so it is plausible such rational matrices exist as well. For example $A = \dfrac{3+\sqrt{5}}{2}I$ and $B=I.$ – Arin Chaudhuri Apr 18 '17 at 23:45
• I wonder if there is an example of such $A$ and $B$, even with complex entries, such that $AB-BA\neq 0$. – Jonas Meyer Apr 19 '17 at 3:09
• @JonasMeyer I found one, see my edited answer. The entries are real numbers though. – i707107 Apr 21 '17 at 1:07

## 3 Answers

I remember that I have seen this, but cannot remember where it was. It might be on this site before.

The idea is to use $(-3\pm\sqrt 5)/2$. We begin with

\begin{align} (A+xB)(A+x'B) &= A^2 + x BA + x' AB + xx' B^2 \\ &= A^2 + (x' + x) AB+x(BA-AB)+xx'B^2.\end{align}

Now we find numbers $x$, $x'$ such that $x'+x = -3$, $xx'=1$. These numbers are the roots of the quadratic equation $$\lambda^2 +3\lambda +1 = 0.$$ Thus, we have $x = \frac{-3+\sqrt 5}2$ and $x'= \frac{-3-\sqrt 5}2$.

With these, we have \begin{align} (A+xB)(A+x'B) &= A^2 -3AB + B^2 + x(BA-AB) \\ &= BA-AB + x(BA-AB) = (BA-AB)(1+x).\end{align} We now have that $$\det(A+xB)(A+x'B) =q \in \mathbb{Q},$$ and $$q=(1+x)^n \det(BA-AB)$$ Then it follows by $A, B\in M_n(\mathbb{Q})$ and $(1+x)^n\in\mathbb{R} \backslash \mathbb{Q}$ that $\det(BA-AB)=0$.

An example with $AB\neq BA$

For the example that @Jonas Meyer requested, here it is: $$A=\begin{pmatrix} 1 & 1 \\ 0 & \frac{-3+\sqrt 5}2\end{pmatrix}, \ \ B=\begin{pmatrix} \frac{3+\sqrt 5}2 & 1 \\ 0 & -1\end{pmatrix}.$$ Then $$A-B= \begin{pmatrix} \frac{-1-\sqrt 5}2 & 0 \\ 0 & \frac{-1+\sqrt 5}2\end{pmatrix}$$, $$(A-B)^2 = \begin{pmatrix} \frac{3+\sqrt 5}2 & 0 \\ 0 & \frac{3-\sqrt 5}2\end{pmatrix} = AB.$$ But, $$BA =\begin{pmatrix} \frac{3+\sqrt 5}2 & \sqrt 5 \\ 0 & \frac{3-\sqrt 5}2\end{pmatrix} \neq AB.$$

• Could you clarify how you conclude that $q\in\mathbb Q$? – stewbasic Apr 19 '17 at 2:23
• Its conjugate inside $\mathbb{Q}(\sqrt 5)$ is the same with itself. – i707107 Apr 19 '17 at 2:25
• Ah right, thanks. So is there a counterexample over $\mathbb C$ then? – stewbasic Apr 19 '17 at 2:26
• @SangchulLee Arin Chaudhuri's comment is not a counterexample to this because it gives $AB=BA$. – i707107 Apr 19 '17 at 2:29
• Not sure right now, but some modification of Arin's example would give a counterexample. – i707107 Apr 19 '17 at 2:30

The statement is true without the assumption that $A,B$ have rational entries. As in i707107's answer, let $$x=\frac{-3-\sqrt{5}}2,\,x'=\frac{-3+\sqrt{5}}2$$ so that $x+x'=-3$, $xx'=1$. Let $X=A+xB$, $Y=A+x'B$. Then $$\begin{eqnarray*} (1+x')XY-(1+x)YX &=&(x'-x)(A^2+xx'B^2)+(x'(1+x')-x(1+x))AB\\ &&{}+(x(1+x')-x'(1+x))BA\\ &=&(x'-x)\left[A^2+B^2+(1+x'+x)AB-BA\right]\\ &=&(x'-x)\left[(A-B)^2-AB\right]\\ &=&0. \end{eqnarray*}$$ Thus $XY=kYX$ where $$k=\frac{1+x}{1+x'}$$ Note that $|k|>1$. Pick an eigenvector $v$ for $X$ whose eigenvalue $\lambda$ has maximal magnitude. Then $$X(Yv)=kYXv=(k\lambda)(Yv).$$ If $\lambda=0$ then $YXv=XYv=0$. If $\lambda\neq0$ then $|k\lambda|>|\lambda|$, so by assumption $k\lambda$ can't be an eigenvalue of $X$. This implies $Yv=0$, so again $XYv=0$ and $YXv=\lambda Yv=0$. In either case we have $(XY-YX)v=0$, so $XY-YX$ is singular. Finally since $$XY-YX=(x'-x)(AB-BA),$$ we conclude $AB-BA$ is singular.

To give a bit more insight, suppose we started with an arbitrary homogeneous degree 2 constraint on $A,B$: $$c_2A^2+c_1AB+c_1'BA+c_0B^2=0.$$ If we replace $A,B$ by commuting variables $a,b$, the corresponding polynomial would factor over $\mathbb C$: $$c_2a^2+(c_1+c_1')ab+c_0b^2=(\alpha a+\beta b)(\gamma a+\delta b).$$ Let $X=\alpha A+\beta B$ and $Y=\gamma A+\delta B$. If $A$ and $B$ commuted we'd have $XY=0$, but instead we get $$XY=(\alpha\delta-c_1)[A,B]$$ where $[A,B]=AB-BA$ is the Lie bracket. Note that $[X,Y]=(\alpha\delta-\beta\gamma)[A,B]$, so $$(\alpha\delta-\beta\gamma)XY=(\alpha\delta-c_1)[X,Y].$$ Unless a coefficient happens to vanish, this gives $XY=kYX$ for some $k$. When $k$ is not a root of unity this is quite a restrictive constraint (eg $XY$ must be singular).

• My first thought was that the result extends to (but may not beyond) $\Bbb{R}$, so this is interesting to see the answer carries over to all of $\Bbb{C}$. (+1) – Sangchul Lee Apr 19 '17 at 3:16
• @SangchulLee I think you could actually deduce the result over $\mathbb C$ from the result over $\mathbb Q$; extend to any finite extension $F$ of $\mathbb Q$ using the embedding $\rm{End}_{F}(F^n)\to\rm{End}_{\mathbb Q}(F^n)$, and then to the algebraic closure of $\mathbb Q$. Finally frame the problem as asking whether an algebraic variety is empty and use weak Hilbert's Nullstellensatz. – stewbasic Apr 19 '17 at 6:18
• May I ask how do you come up with this magnificent proof? It looks like a miracle. :) BTW, it is i707107 I think. – awllower Apr 19 '17 at 6:21
• @awllower I'll add a section at the end. – stewbasic Apr 19 '17 at 6:27
• Thanks for the clarifications. :D – awllower Apr 19 '17 at 6:52

In fact, there is a much stronger result.

Proposition. Let $a,b,c,d\in\mathbb{C}$. If $A,B\in M_n(\mathbb{C})$ satisfy $aA^2+bB^2+cAB+dBA=0_n$, then, generically (for example, randomly choose $a,b,c,d$), $A,B$ are simultaneously triangularizable (denoted by ST).

Proof. We follow the last part of the stewbasic's good post. We put $X=\alpha A+\beta B,Y=\gamma A+\delta B$; since generically $\alpha\delta\not= \beta\gamma$, it suffices to show that $X,Y$ are ST.

We obtain (generically) $XY=kYX$ for some complex number $k$. Generically, $k\not=1$ and $k$ is not a primitive root of unity; then, by a result from Drazin, $X,Y$ are ST.

Remark 1. Of course, there are $a,b,c,d$ s.t. $A,B$ are not necessarily ST. For instance, consider $AB+BA=0_2$ with $A=\begin{pmatrix}1&0\\0&-1\end{pmatrix},B=\begin{pmatrix}0&1\\1&0\end{pmatrix}$.

Remark 2. That is linked to the concept of quasi-commutative matrices. There are two non-equivalent definitions:

i) $A,B$ are quasi-commutative iff $AB-BA$ commute with $A,B$. By a result from McCoy, $A,B$ are always ST.

ii) $A,B$ are quasi-commutative iff $AB=kBA$ where $k$ is a complex number. This is the definition we are interested in.