5
$\begingroup$

Let $A$ be a UFD, that is also a $k$-algebra, where $k$ is a field of characteristic $\not=2$ (for instance polynomials over $k$).

Is every involution in $\mathrm{GL}_n(A)$ diagonalisable?

This is of course true over the field of fractions of $A$.

In this question it is explained that $$M=\left(\begin{array}{rr} 3&-1\\-1 & 3\end{array}\right)$$ is diagonalisable over $\mathbb{Q}$ but not over $\mathbb{Z}$. However, it is not an involution.

$\endgroup$
8
  • 3
    $\begingroup$ @Vladimir Dotsenko: But 2 is invertible in $A$ by hypothesis. $\endgroup$
    – abx
    Dec 3, 2021 at 7:40
  • 3
    $\begingroup$ Isn't $A^{n}$ (identified with $n$-long column vectors with entries in $A$) expressible as ${\rm Im}(\frac{1+t}{2}) \oplus {\rm Im}(\frac{1-t}{2})$ as an $A$-module for any such involution $t$ ? $\endgroup$ Dec 3, 2021 at 12:10
  • 2
    $\begingroup$ Could you maybe use different letters for the UFD and the matrix? $\endgroup$ Dec 3, 2021 at 12:12
  • 3
    $\begingroup$ Following @GeoffRobinson's comment, if $A$ is a UFD with $2\in A^\times$ admitting a non-free f.g. projective $A$-module $P$, and $Q$ is another projective module such that $P\oplus Q\cong A^n$, then $\mathrm{id}_P \oplus (-\mathrm{id}_Q)$ should be counterexample. Conversely, if every f.g. projective $A$-module is free (e.g. if $A=F[x_1,\dots,x_r]$ with $F$ a field), then every involution is diagonalizable. $\endgroup$ Dec 3, 2021 at 12:22
  • 2
    $\begingroup$ By the way, the question is equivalent to asking whether every idempotent matrix $E \in M_{n}(A)$ is diagonalizable via an element of ${\rm GL}(n,A)$. Note that if $E$ is idempotent, then $2E-I$ is an involution, while if $t$ is an involution, then $t = 2E-I$ for some idempotent $E$ as in my comment above ( here we need $2$ invertible in $A$). $\endgroup$ Dec 4, 2021 at 12:53

1 Answer 1

8
$\begingroup$

Here is an answer based on my comment, and Geoff Robinson's earlier comment.

Let $A$ be a domain with $2\in A^\times$, and let $M$ be an $n\times n$ matrix with $M^2=I$. It is convenient to consider $M$ as an $A$-endomorphism of $A^n$ and $I$ as the identity endomorphism of $A^n$. Put $S_+ =\{x\in A^n\,:\,Mx=x\}$ and $S_-=\{x\in A^n\,:\,Mx=-x\}$. As noted in Geoff Robinson's comment, since $2\in A^\times$, we have $A^n=S_+\oplus S_-$. (This follows readily from the fact that $x=\frac{1}{2}(x+Mx)+\frac{1}{2}(x-Mx)$ for all $x\in A^n$ and $M^2=I$.) In particular, both $S_+$ and $S_-$ are f.g. projective $A$-modules.

Suppose that every f.g. projective $A$-module is free. Then $S_+$ and $S_-$ are free. Let $E_+$ and $E_-$ be bases for these spaces. Then $E=E_+\cup E_-$ is a basis of $A^n$ consisting of eigenvalues of $M$, which means that $M$ is a diagonalizable.

On the other hand, if there exists a f.g. projective $A$-module $P$ which is not free, then we can construct a non-diagonalizable involution $M:A^n\to A^n$ as follows. There exists a f.g. $A$-module $Q$ such that $A^n\cong P\oplus Q$ for some $n\in\mathbb{N}$. Now take $M$ to be the matrix corresponding to $\mathrm{id}_P\oplus (-\mathrm{id}_Q)$. For this $M$, we have $S_+=P$ and $S_-=Q$. If $M$ were diagonalizable, then $A^n$ would have a basis $E$ consisting of vectors $x\in A^n$ with $Mx\in \{\pm x\}$ (here we need $A$ to be a domain). Consequently, $E = (E\cap S_+)\cup (E\cap S_-)$, which means that $E_+:=E\cap S_+$ is a basis of $S_+=P$. This contradicts our choice of $P$, so $M$ is not diagonalizable.

There are examples of UFDs admitting f.g. projective modules which are not free --- see this question, for instance. Quoting from this answer, one can take $A=\mathbb{C}[a,b,c,x,y,z]/(ax+by+cz-1)$, the module $P$ to be the kernel of $(\alpha,\beta,\gamma)\mapsto a\alpha+b\beta+c\gamma:A^3\to A$, and $Q=(x,y,z)A$. One has $P\oplus Q=A^3$, so it should be possible to write the resulting $M\in \mathrm{M}_3(A)$ explicitly.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks Uriya First for the nice and detailed answer. And thanks to Geoff Robinson too. $\endgroup$ Dec 6, 2021 at 7:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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