6
$\begingroup$

In this question, the matrices are square and real and the polynomials have real coefficients, but feel free to mention other fields if that is interesting.

Let $\chi(M)$ denote the characteristic polynomial of a matrix $M$.

For two pairs of matrices, let us write $(A_1,B_1)\stackrel\chi\sim (A_2,B_2)$ if $$ \chi(p(A_1,B_1)) = \chi(p(A_2,B_2)) $$ for every polynomial $p$ in two non-commuting variables.

A sufficient but not necessary condition for $(A_1,B_1)\stackrel\chi\sim (A_2,B_2)$ is that $A_2=X^{-1}A_1X$ and $B_2=X^{-1}B_1X$ for some non-singular $X$.

Q0. I'm sure this is not my invention. Where is it studied?

Q1. Can this relation be characterised in a finite way (without calling on every polynomial)?

Q2. How can this relation be tested for? Assume exact real arithmetric or, if you prefer, assume the matrices are rational.

$\endgroup$

2 Answers 2

4
$\begingroup$

The problem of simultaneous conjugation has been studied for more than 50 years. First off, it has been proved that the ring of $GL(n)$-invariants on $M(n)\times M(n)$ or, more generally, on $M(n)^m$ is generated by the coefficients of the characteristic polynomials $\chi(p(A_1,\ldots,A_m))$ where $p$ runs through all non-commutative monomials. In characteristic zero, this is due to to K. S. Sibirskiı̆ (Algebraic invariants of a system of matrices, Sibirsk. Mat. Ž. 9 (1968), 152–164) and, independently, by C. Procesi (The invariant theory of n × n matrices, Advances in Math. 19 (1976), 306–381). In positive characteristic and even over $\mathbb Z$ this is due to S. Donkin (Invariants of several matrices, Invent. Math. 110 (1992), 389–401).

Thus the equivalence relation $\overset\chi\sim$ is simply that given by the GIT-quotient $M(n)^m\to M(n)^m/\!/GL(n)$. In particular, the usual theorems of GIT apply: Two $m$-tuples of matrices are equivalent if and only if the closures of their simultaneous conjugacy classes intersect. In particular, if these conjugacy classes are closed then $\overset\chi\sim$ coincides with simultaneous conjugacy.

The question of when a simultanoues conjugacy class is closed has been studied by R. W. Richardson (Conjugacy classes of n-tuples in Lie algebras and algebraic groups. Duke Math. J.57(1988), 1–35). He shows this to be the case if and only if the subalgebra of $M(n)$ generated by $A_1,\ldots,A_m$ is semisimple. This holds for example if $M(n)$ is generated by $A_1,\ldots,A_m$ which, for $m\ge2$, is the generic case.

As for the computational aspects of your questions it has been shown by Procesi (loc. cit.) in characteristic zero that it suffices to check the traces of $p(A_1,\ldots,A_m)$ where $p$ is a monomial of degree at most $2^n-1$. As far as I remember this is far from optimal and the true bound should be polynomial in $n$.

Finally, for testing equivalence of two $m$-tuples it suffices to test a so-called separating set of invariants, i.e., invariants which separate everything which can be separated by invariants. For this, a paper by H. Derksen and V. Makam (Algorithms for orbit closure separation for invariants and semi-invariants of matrices Algebra Number Theory 14 (2020), 2791–2813) might be useful. In Thm.~1.14 they derive the bound $O(n\log n)$ on the degree of sparating invariants.

$\endgroup$
1
  • $\begingroup$ A splendidly thorough answer, thank you! $\endgroup$ Commented Aug 3, 2023 at 5:43
6
$\begingroup$

In most (but not all) cases, this relation is just joint similarity. This fact relies on the following proposition.

Proposition: Let $K$ be a field. Then the only automorphisms of the $K$-algebra $M_d(K)$ are inner automorphisms. In other words, if $\phi:M_d(K)\rightarrow M_d(K)$ is a $K$-algebra automorphism, then there is some invertible $X$ with $\phi(Y)=XYX^{-1}$ for all $Y\in M_d(K)$.

Here is a proof of the above proposition.

Suppose that $(A_1,\dots,A_r),(B_1,\dots,B_r)$ are $d\times d$-matrices over a field $K$ where $(A_1,\dots,A_r),(B_1,\dots,B_r)$ both generate the algebra $M_d(K)$. Then I claim that $(A_1,\dots,A_r)\stackrel\chi\sim(B_1,\dots,B_r)$ if and only if there is some invertible $X$ where $B_j=XA_jX^{-1}$ for $1\leq j\leq r$ (if $B_j=XA_jX^{-1}$ for $1\leq j\leq r$, then we say that $(A_1,\dots,A_r),(B_1,\dots,B_r)$ are jointly similar). The direction $\leftarrow$ is clear, so we just need to assume $(A_1,\dots,A_r)\stackrel\chi\sim(B_1,\dots,B_r)$ in order to establish the direction $\rightarrow$.

Suppose that $(A_1,\dots,A_r),(B_1,\dots,B_r)$ both generate the algebra $M_d(K)$. Then define an equivalence relation $\approx$ on $K\langle x_1,\dots,x_r\rangle$ where we set $p_0(x_1,\dots,x_r)\approx p_1(x_1,\dots,x_r)$ precisely when $$\chi((p_0(A_1,\dots,A_r)-p_1(A_1,\dots,A_r))q(A_1,\dots,A_r))=x^d$$ for each $q\in K\langle x_1,\dots,x_r\rangle$. Then we observe that $p_0(x_1,\dots,x_r)\approx p_1(x_1,\dots,x_r)$ if and only if $p_0(A_1,\dots,A_r)=p_1(A_1,\dots,A_r).$ Therefore, the mapping $\phi_0:K\langle x_1,\dots,x_r\rangle/\approx\rightarrow M_d(K)$ defined by letting $\phi(p(x_1,\dots,x_r)/\approx)=p(A_1,\dots,A_r)$ is a $K$-algebra isomorphism.

On the other hand, we can define an isomorphism $\phi_1:K\langle x_1,\dots,x_r\rangle/\simeq\rightarrow M_d(K)$ by letting $\phi_1(p(x_1,\dots,x_r)/\approx)=p(B_1,\dots,B_r)$. We therefore obtain an automorphism $\phi_1\phi_0^{-1}:M_d(K)\rightarrow M_d(K)$ where $$\phi_1\phi_0^{-1}(p(A_1,\dots,A_r))=\phi_1(p(x_1,\dots,x_r)/\approx)=p(B_1,\dots,B_r).$$ Since $\phi_1\phi_0^{-1}$ is an automorphism of the algebra $M_d(K)$, there is some invertible $X$ where $\phi_1\phi_0^{-1}(Y)=XYX^{-1}$ for all $X$. Therefore, we have $$p(B_1,\dots,B_r)=\phi_1\phi_0^{-1}(p(A_1,\dots,A_r))=X\cdot p(A_1,\dots,A_r)\cdot X^{-1}$$ for all $p$, so $(A_1,\dots,A_r)$ and $(B_1,\dots,B_r)$ are jointly similar.

I suspect that we can also prove our main result without relying on the fact that the only algebra automorphisms of $M_d(K)$ are inner, and perhaps I will come back to this answer with that other proof.

The hypothesis that $(A_1,\dots,A_r)$ generates $M_d(K)$ is a weak hypothesis that is satisfied in almost all cases. For example, if $A=(a_{i,j})$ and $B=(b_{i,j})$ are the matrices where $a_{i,j}=1$ if $j=i+1\mod d$ and $a_{i,j}=0$ otherwise and where $b_{i,j}=1$ iff $i=j=1$ and $b_{i,j}=0$ otherwise, then $\\{A^rBA^s,0\leq r<d,0\leq s<d\\}$ is the canonical basis for the vector space $M_d(K)$. If $K$ is an infinite field, then the set of all pairs $(A,B)$ of $d\times d$-matrices where $\\{A^rBA^s,0\leq r<d,0\leq s<d\\}$ is a basis for the vector space $M_d(K)$ is a non-empty Zariski open set. If $V$ is an affine space over an infinite field $K$, and $O_i$ is Zariski open in $V$ for $i\in I$, and $|I|<|K|$, then $\bigcap_{i\in I}O_i\neq\emptyset$ (reference). This means that subsets of proper Zariski closed set should be considered to be 'small' sets. We conclude that if $r>1$ and the field $K$ is infinite, then $(A_1,\dots,A_r)$ typically generates the algebra $M_d(K)$.

I personally find it easier to work with non-commuting matrices $(A_1,\dots,A_r)$ when $(A_1,\dots,A_r)$ generates the algebra $M_d(K)$.

$\endgroup$
1
  • $\begingroup$ Very interesting, thanks. Can you elaborate on "most (but not all)"? I know that joint similarity is not sufficient even if all of $A_1,B_1,A_2,B_2$ are non-singular. For example $A_1=\pmatrix{1&1\\0&1}$, $A_2=B_1=B_2=I$. $\endgroup$ Commented Aug 2, 2023 at 3:06

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .