5

Let $A,B\in M_{3x3}(\mathbb Z/p\mathbb Z)$ ($p$ a prime number). Find the probability $P$ that $AB=BA$ that is $P(AB=BA)$

$$A=\begin{pmatrix} a_{11} & a_{12} & a_{13} \\ a_{21} & a_{22} & a_{23} \\ a_{31} & a_{32} & a_{33} \\ \end{pmatrix} $$

$$B=\begin{pmatrix} b_{11} & b_{12} & b_{13} \\ b_{21} & b_{22} & b_{23} \\ b_{31} & b_{32} & b_{33} \\ \end{pmatrix} $$

Please I would really appreciate if you can help me with this problem. Any ideas or suggestions would be highly appreciated :)

Ben Grossmann
  • 225,327
user128422
  • 3,057
  • This is a nice problem---what have you tried so far? – Travis Willse Oct 14 '14 at 03:47
  • The first thing that came to my mind was just to multiply the matrices then, I have the coefficients of $AB$ and $BA$ then I took the first of them : $a_{11}b_{11}+a_{12}b_{21}+a_{13}b_{31}=b_{11}a_{11}+b_{12}a_{21}+b_{13}a_{31}$ wich gives us $a_{12}b_{21}+a_{13}b_{31}=b_{12}a_{21}+b_{13}a_{31}$ – user128422 Oct 14 '14 at 03:55
  • but I don´t know exactly in how many ways $a_{12}b_{21}+a_{13}b_{31}=b_{12}a_{21}+b_{13}a_{31}$ – user128422 Oct 14 '14 at 04:00
  • You can certainly do a case analysis for this equation. If $a_{12} \neq 0$, then you can solve uniquely for $b_{21}$ for any values of the six remaining variables, and if $a_{12} = 0$, you reduce the problem to a simpler equation. But you'd need to do this for every entry, and the equations are interdependent, so a brute-force analysis of the component equations is probably not the best way to proceed. – Travis Willse Oct 14 '14 at 04:03
  • Exactly there are a lot of equations :/ – user128422 Oct 14 '14 at 04:09
  • 1
    This may well be hard. It might be somewhat easier if you restrict to invertible matrices, which makes this a problem in group theory. Anyway, this article looks potentially relevant: http://projecteuclid.org/euclid.dmj/1077468920 (gated). – Travis Willse Oct 14 '14 at 04:19
  • Glad to help. Please post a comment if you learn anything more, I'd be quite interested to learn more. It might be possible to make observations special to this case since the field is so small. In fact, there are only $\sim 4 \cdot 10^8$ pairs of matrices, so direct computation wouldn't be prohibitive, but it probably wouldn't be very satisfying either. – Travis Willse Oct 14 '14 at 04:33
  • Why $3\times3$ matrices? Is $2\times2$ too easy? – bof Oct 15 '14 at 02:20
  • @bof Brute-forcing he $2 \times 2$ case only involves $3^8 = 6561$ cases, and doing so shows that the probability is $35/243 = 0.14403\ldots$. – Travis Willse Oct 15 '14 at 13:07
  • @Travis Is $p=3$ in this question? I missed that part. Nobody cares about other values of $p$, or they've all been solved? – bof Oct 15 '14 at 13:13
  • One observation that might help with the general reduction is that the number of matrices in $M_{3 \times 3}(\mathbb{Z} / 3\mathbb{Z})$ with which a given matrix commutes is an invariant of the orbit of the $GL(3, 3)$ conjugation action. Certainly the Jordan form is an invariant of the orbit, too, but I'm not sure whether Jordan forms actually parameterize the orbits. Either way, this technique should be a major improvement over brute-forcing for general matrix size and finite field. – Travis Willse Oct 15 '14 at 13:15
  • @bof Possibly OP is asking because this is a natural pilot case: Characteristic $2$ often behaves qualitatively differently from higher prime characteristics, and $3 \times 3$ is the first case that brute-forcing takes at least a little time on a typical desktop. – Travis Willse Oct 15 '14 at 13:24
  • @bof Of course, one could ask this question over any ring, including any $\mathbb{Z} / n \mathbb{Z}$, but one might expect the cases when this is a field (i.e. $n$ prime) to be especially nice. – Travis Willse Oct 15 '14 at 13:26
  • @Travis Must be my weak eyes (glaucoma, cataracts, etc.) I've read the OP's question repeatedly, and I see where he asks about a field $\mathbb Z/\mathbb Z_p$ ($p$ a prime number), but I don't see where he specialized it to $p=3$. Was that in a comment, maybe deleted by now? – bof Oct 15 '14 at 13:33
  • Oh, you're right, I misread the question myself! In that case, I'm not sure why not $2 \times 2$ matrices... – Travis Willse Oct 15 '14 at 13:34
  • My teacher gave us this problem. I think he just chose a random number but he could have certainly chosen another number – user128422 Oct 15 '14 at 16:37

3 Answers3

2

Let $P_p$ be the required probability. The couples $\{(A,uI_3+vA+wA^2)|A\in M_3\}$ are some solutions ; moreover, "generically", they are THE solutions. Then $P_p\approx \dfrac{p^9p^3}{p^{18}}=\dfrac{1}{p^6}$ ; moreover $P_p\geq 1/p^6$. Numerical experiments (for $p=2,3,5,7$) seem to "show" that $1/p^6\leq P_p\leq 2/p^6$.

  • They are THE solutions only if the minimal polynomial of $A$ coincides with the characteristic polynomial. – Sungjin Kim Oct 15 '14 at 21:41
  • @ i707107 , that you write is not a scoop; indeed, this property of cyclic endomorphisms is well-known. I speak about generic matrices; when the underlying field $K$ is infinite, then almost all matrices are cyclic because almost all matrices have distinct eigenvalues in an algebraic extension of $K$ (choose randomly a complex or real matrix and calculate the discriminant of its characteristic polynomial). –  Oct 15 '14 at 23:56
  • When $K$ is a finite field, "generic" means that the probability that a matrix is cyclic is close to 1. For instance, if $p=101$, then $99/100$ of the matrices of $M_3(\mathbb{Z}/p\mathbb{Z})$ have distinct eigenvalues. –  Oct 15 '14 at 23:57
2

This is perhaps not very satisfying, but brute-forcing the computation for $p = 2$ in Maple with the below code gives that there are $7456$ commuting (ordered) pairs of matrices in $M_{3 \times 3}(\mathbb{Z} / 2\mathbb{Z})$ (out of $(2^9)^2$ total pairs), so the probability in that case is $$\frac{7456}{(2^9)^2} = \frac{233}{8192} = 0.028442\ldots,$$ which agrees with loup blanc's experiments.

p := 2;
M := [seq(seq(seq(seq(seq(seq(seq(seq(seq(Mod(p, Matrix(3, (i, j) -> m||i||j), integer), m33 = 0..(p-1)), m32 = 0..(p-1)), m31 = 0..(p-1)), m23 = 0..(p-1)), m22 = 0..(p-1)), m21 = 0..(p-1)), m13 = 0..(p-1)), m12 = 0..(p-1)), m11 = 0..(p-1))]:
k := 0;
for i from 1 to nops(M) do
    for j from 1 to nops(M) do
        if op(convert(Mod(p, M[i].M[j] - M[j].M[i], integer[]), set)) = 0 then k := k + 1 end if;
    end do;
end do;
k;
k/

(Incidentally, if someone knows a better way to generate a list of all matrices in $M_{3 \times 3}(\mathbb{Z} / p \mathbb{Z})$ I'd be grateful to learn how.)

Travis Willse
  • 99,363
0

Edit on 10/13/2016

The number of similarity classes $|I|$ of $\mathrm{GL}_3(\mathbb{F}_p)$ below easily follows from this answer. In fact, without having to consider $8$ different types, we have $$ |I|=\sum_{\lambda \in P_3} p^{\ell(\lambda)} $$ where $P_3$ is the set of partition of $3$, and $\ell(\lambda)$ is the number of parts of $\lambda$.

Then $$ |I|=p^3+p^2+p. $$

An Upper bound for $p\geq 5$

Let $A\in M_3(\mathbb{Z}/p\mathbb{Z})$. Denote by $\mathbb{F}_p=\mathbb{Z}/p\mathbb{Z} $ and $\mathbb{F}_p[x]$ the polynomial ring over $\mathbb{F}_p$. We write $M^A=\mathbb{F}_p^3$ the finitely generated module over $\mathbb{F}_p[x]$ where $x \cdot v = A \cdot v$ for any $v\in\mathbb{F}_p^3$. It is well-known that the dimension of the space of commuting matrices $C_A = \{ B \in M_3(\mathbb{Z}/p\mathbb{Z}) | AB=BA\}$ depends only on the similarity class of $A$. The similarity class of $A$ falls into one of the following eight types:

  1. $aaa$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-a)}$ where $a\in\mathbb{F}_p$.

  2. $aa^2$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-a)^2}$ where $a\in\mathbb{F}_p$.

  3. $a^3$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)^3}$ where $a\in\mathbb{F}_p$.

  4. $C$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(C(x))}$ where $C(x)\in\mathbb{F}_p[x]$ is a monic cubic irreducible polynomial over $\mathbb{F}_p$.

  5. $abb$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-b)}\oplus \frac{\mathbb{F}_p[x]}{(x-b)}$ where $a, b\in\mathbb{F}_p$ and $a\neq b$.

  6. $ab^2$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-b)^2}$ where $a, b\in\mathbb{F}_p$ and $a\neq b$.

  7. $aB$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(B(x))}$ where $a \in\mathbb{F}_p$ and $B(x)$ is a monic quadratic irreducible polynomial over $\mathbb{F}_p$.

  8. $abc$-type: $M^A \simeq \frac{\mathbb{F}_p[x]}{(x-a)}\oplus \frac{\mathbb{F}_p[x]}{(x-b)}\oplus \frac{\mathbb{F}_p[x]}{(x-c)}$ where $a,b,c\in\mathbb{F}_p$ are distinct.

Let $\{I_1,I_2,\ldots, I_8\}$ be the set of the above eight types similarity classes, and let $I=\cup_{j=1}^8 I_j$ be the set of all distinct similarity classes. Define $G=GL_3(\mathbb{F}_p)$, and $G_A=\{H\in G| AH=HA\}$. By the orbit-stabilizer formula, $$ |\{(A,B)\in M_3(\mathbb{F}_p)^2 | AB=BA\}|=\sum_{A\in I} |G/G_A| |C_A|\leq \sum_{A\in I} |G| \frac{|C_A|}{|C_A|-3p^{\dim C_A -1}}. $$ For the last inequality, let $A\in I_j$ and $\dim C_A = m$. Let $\{I=B_1, \ldots, B_m\}$ be a basis for $C_A$. Consider $$ \det (c_1 B_1 + \cdots + c_m B_m) = 0. \ \ \ (*) $$ For any given $c_2,\ldots, c_m$, the equation $(*)$ is a cubic equation in $c_1$. Therefore, the number of solution to $(*)$ is at most $3p^{m-1}$. We remark that with a lot of effort, we will be able to find explicit formula for $|G_A|$. This is certainly more difficult than finding $|C_A|$.

Note that $$ 1\leq\frac{|C_A|}{|G_A|}\leq\frac{|C_A|}{|C_A|-3p^{\dim C_A -1}}=\frac1{1-\frac3p}=\frac p{p-3}. $$ The number of elements in eacy $I_j$ can be easily found: $$|I_1|= |I_2|= |I_3|=p, \ |I_4|= \frac{p^3-p}3, \ |I_5| =|I_6|=p(p-1), \ |I_7|=\frac{p(p^2-p)}2,$$ $$|I_8|=\frac{p(p-1)(p-2)}6.$$

Therefore, $$ |\{(A,B)\in M_3(\mathbb{F}_p)^2 | AB=BA\}|\leq \frac p{p-3}|G|\left(3p+\frac{p^3-p}3 +2p(p-1)+\frac{p(p^2-p)}2+\frac{p(p-1)(p-2)}6\right) $$ $$ =\frac p{p-3} |G| (p^3+p^2+p). $$

Combining these, we obtain that $$ \frac{ |G|(p^3+p^2+p)}{(p^9)^2}\leq \frac{|\{(A,B)\in M_3(\mathbb{F}_p)^2 | AB=BA\}|}{(p^9)^2}\leq \frac p{p-3}\frac{|G|(p^3+p^3+p)}{(p^9)^2}. $$ Since $|G|=(p^3-1)(p^3-p)(p^3-p^2)$, we have as $p\rightarrow\infty$, $$ \frac{|\{(A,B)\in M_3(\mathbb{F}_p)^2 | AB=BA\}|}{(p^9)^2}\sim \frac1{p^6}. $$

Sungjin Kim
  • 20,102