4
$\begingroup$

Can all matrices be written as the sum of commuting matrices?

All invertible matrices $A$ can be written as a sum of matrices $A=B+C$ such that $B$ and $C$ commute, $BC=CB$.

One example might be $$B=\frac{A+A^{-1}}{2}, \quad C=\frac{A-A^{-1}}{2},$$ but I'm not certain if this is the only solution.

My question is:

Can non-invertible matrices be written as the sum of commuting matrices?

Edit: In response to Lulu's comment, I need to add another condition that neither of $B$ or $C$ commute with all matrices.

$\endgroup$
4
  • 13
    $\begingroup$ Well, $A=A+0$ and $A,0$ commute. More broadly, if $C$ is any matrix that commutes with every matrix then $A=(A-C)+C$. Are there more conditions? $\endgroup$ Commented Apr 7, 2023 at 10:34
  • 1
    $\begingroup$ Related $\endgroup$ Commented Apr 7, 2023 at 10:56
  • 1
    $\begingroup$ @lulu I seem to have overlooked that option! Thank you for your answer. $\endgroup$ Commented Apr 7, 2023 at 11:20
  • 1
    $\begingroup$ I suppose I should add the requirement that one of the matrices does not commute with all matrices $\endgroup$ Commented Apr 7, 2023 at 11:26

1 Answer 1

2
$\begingroup$

The answer to your question is negative in the following two pathological cases:

  1. When the matrices are $1\times1$, the answer is negative because all matrices commute.
  2. When $A$ is a $2\times2$ matrix over $GF(2)$ and it is not a scalar matrix, the answer is also negative. In this case, $A$ is similar to either $\operatorname{diag}(1,0)$ or a companion matrix. Therefore $A$ only commutes with polynomials in itself. Now, in order that $B$ commutes with $C=A-B$, we must have $AB=BA$. Hence $B$ must be a polynomial in $A$. So, by Cayley-Hamilton theorem, $B=pA+qI_2$ and $C=A-B=(1-p)A-qI_2$ for some scalars $p,q\in GF(2)=\{0,1\}$. But then one of $B$ or $C$ will be a scalar matrix that commutes with all matrices.

In all other cases, the answer to your question is affirmative.

If $A$ is a scalar matrix whose size is at least $2\times2$, i.e., if $A=aI_n$ for some $n\ge2$, we may take $B=\operatorname{diag}(b,a,\ldots,a)$ and $C=\operatorname{diag}(a-b,0,\ldots,0)$ for any $b\ne a$.

If $A$ is not a scalar matrix, its size is at least $2\times2$ and the field is not $GF(2)$, we may simply take $B=pA$ and $C=(1-p)A$ for any $p\not\in\{0,1\}$.

If $A$ is not a scalar matrix, its size is at least $3\times3$ and the field is $GF(2)$, there are two possibilities:

  1. $A$ can be diagonalised as $aI_k\oplus(1-a)I_{n-k}$ where $2\le k<n$. Then we may take $B=(1-a)\oplus aI_{k-1}\oplus (1-a)I_{n-k}$ and $C=\operatorname{diag}(1,0,\ldots,0)$.
  2. $A$ is not diagonalisable. Its minimal polynomial is then at least quadratic. If it is at least cubic, then $A-A^2$ and $A^2$ are not scalar matrices and we may take $B=A-A^2$ and $C=A^2$. If it is quadratic, since $A$ is not diagonalisable, the rational canonical form of $A$ must contain a $2\times2$ companion matrix $M$. Hence up to similarity we may assume that $A$ is in the form of $\pmatrix{M\\ &N}$. Take $B=\pmatrix{M\\ &N-I_{n-2}}$ and $C=\pmatrix{0\\ &I_{n-2}}$ and we are done.
$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.