10

I've been studying some math by myself this summer, and have recently been doing some reading about the groups $\text{GL}(2,\mathbb{Z})$, $\text{SL}(2,\mathbb{Z})$, etc. I've been trying to get a better grasp of conjugation in these groups, but have been unable to figure out something relatively simple:

Are the following two matrices conjugate in $\text{SL}(2,\mathbb{Z})$? If not maybe in $\text{GL}(2,\mathbb{Z})$?

$$\begin{bmatrix} a & b\\ c & d \end{bmatrix}, \begin{bmatrix} a & c\\ b & d \end{bmatrix}\in\text{SL}(\mathbb{Z}[t,t^{-1}]) $$

If so is there a nice conjugating element? Thanks in advance! I don't have the greatest linear algebra background, but am loving group theory, which might be why I am not getting this.

Thomas Preu
  • 2,002
  • 1
    Each matrix is just the transpose of the other. – Geoffrey Trang Aug 06 '22 at 20:09
  • 2
    @GeoffreyTrang I'm sorry if this is obvious (which it probably is)--but does their being each others transpose mean they are conjugate in $\text{SL}(2,\mathbb{Z})$? I know that it means they are similar, but I am not sure inside which group. – user672836 Aug 06 '22 at 20:26
  • 4
    A priori it means they're conjugate inside $GL_2(\mathbb{C})$ and you'd have to do more work to conclude anything stronger than this. – Qiaochu Yuan Aug 06 '22 at 20:48
  • 1
    FWIW, one can check that $$\begin{bmatrix} be & bf\ bf & ce+df-af \end{bmatrix}^{-1} \begin{bmatrix} a & b\ c & d \end{bmatrix}\begin{bmatrix} be & bf\ bf & ce+df-af \end{bmatrix} =
    \begin{bmatrix} a & c\ b & d \end{bmatrix}$$ assuming $e,f$ are chosen so that the new matrix is nonsingular.
    – Greg Martin Aug 06 '22 at 21:06
  • 3
    @GregMartin The determinant of your transformation matrix is divisible by $b$. So independently of the choices of $e,f$, if $b\not\in{\pm 1}$ the conjugation will not take place over $\mathbb{Z}$. If $b=0$ you cannot even form the inverse. I guess you are aware of all that, I just wanted to point this out explicitly. – Thomas Preu Aug 07 '22 at 04:37

2 Answers2

3

Two such matrices are not always conjugate.

If the matrices $$ A= \begin{pmatrix} a & b \\ c & d \end{pmatrix}, \quad A^T= \begin{pmatrix} a & c \\ b & d \end{pmatrix} $$ are conjugate in $GL_2(\mathbb{C})$, then $XA=A^TX$ for some matrix $$ X= \begin{pmatrix} x & y \\ u & v \end{pmatrix} \in GL_2(\mathbb{C}). $$ Let $b\neq0$ or $c\neq0$. We have $u=y$ and $$ xb+y(d-a)-vc=0. $$

Let $$ A= \begin{pmatrix} t^2+t+1 & t \\ 1+t & 1 \end{pmatrix}. $$ The equality $xt-y(t^2+t)-v(1+t)=0$ must be fulfilled for some $x,y,v\in\mathbb{C}$ simultaneously not equal to $0$. If $t$ is an independent variable, this is not possible.

Correction. I replaced the matrix $A$.

kabenyuk
  • 10,712
  • 1
    Does this not contradict the fact that a matrix and its transpose are conjugate over $\mathbb{C}$? – user672836 Aug 07 '22 at 14:08
  • No, it doesn't. If $t\in\mathbb{C}$ and $t\neq0$, then $v=(1+2/t)x$ and the equation $vx^2-y^2=1$ has solutions (so that $\operatorname{det}(X)=1$). – kabenyuk Aug 07 '22 at 15:09
  • 1
    Ah ok -- so for any particular $t$, the matrices are conjugate, but there is no general conjugating element is what you're saying? – user672836 Aug 07 '22 at 15:16
  • Yes, we can also say that. – kabenyuk Aug 07 '22 at 15:21
  • Ah. I was more asking if there is any $A\in\text{SL}(2, \mathbb{Z})$ for which $A$ and $A^T$ are not conjugate, either in $\text{GL}(2, \mathbb{Z})$ or in $\text{SL}(2, \mathbb{Z})$. Can we apply this argument there? Would it come down to solving the equation you gave as a Diophantine equation? – user672836 Aug 07 '22 at 15:25
  • 2
    This is a very interesting question. I think that if $A\in SL_2(\mathbb{Z})$, then $A$ and $A^T$ are conjugate in $SL_2(\mathbb{Z})$. The solution to this question seems to come down to calculations in real quadratic fields. But maybe there is a simpler way. – kabenyuk Aug 07 '22 at 16:13
  • If possible, could you post this calculation? Is it basically using the existing solution to the conjugacy problem in $\text{SL}(2,\mathbb{Z}$ with the periodic part of the continued fraction of the slope of the expanding eigenvector? – user672836 Aug 09 '22 at 04:56
  • @Thomas, I replaced the matrix $A$. – kabenyuk Aug 09 '22 at 16:26
2

I am not sure what adjoining $[t,t^{-1}]$ accomplishes, so I will ignore that.

Set $G=\text{GL}(2,\mathbb{Z})$ and $H=\text{SL}(2,\mathbb{Z})$.
Call $M\in \text{Mat}(2,\mathbb{C})$ "transpose-conjugate by $S\subset\text{Mat}(2,\mathbb{C})$" ("tc by $S$" for short) if and only if $\exists T\in S: T^{-1}MT=M^T$.
Call $M\in \text{Mat}(2,\mathbb{C})$ "transpose-conjugate" ("tc" for short), if and only if $M$ is transpose-conjugate by $G$.

The answer has three parts:

  1. $M_1:=\begin{pmatrix}1&s\\0&1\end{pmatrix}\in H$ for any $s\in\mathbb{Z}\setminus\{0\}$ is not tc by $H$, but is tc by $G$.
  2. $M_2:=\begin{pmatrix}2&5\\-9&-22\end{pmatrix}\in H$ is not tc.
  3. Further remarks on the property tc.

Part 1. Let $M:=M_1:=\begin{pmatrix}1&s\\0&1\end{pmatrix}\in H$ for any $s\in\mathbb{Z}\setminus\{0\}$. Take a general $T:=\begin{pmatrix}a&b\\c&d\end{pmatrix}\in H$. The determinant is $D:=\text{det}(T)=ad-bc=1$. As kabenyuk said, we need to test: $$ N:=MT-TM^T\stackrel{\text{?}}{=}0. $$ We get $N:=\begin{pmatrix}-bs+cs&sd\\-sd&0\end{pmatrix}$. Since $s\neq 0$ we have $d=0$. Furthermore $-b+c=0$. Since $d=0$ we have $1\stackrel{\text{!}}{=}\det(T)=a\cdot 0-b\cdot c=-bc$. The system $[-b+c=0,1=-bc]$ has no solutions in $\mathbb{Z}$ (but $(b,c)\in\{(i,i),(-i,-i)\}$ over $\mathbb{C}$ are solutions). Thus $M\in H$ cannot be conjugated to $M^T$ via matrices in $H$, i.e. $M$ is not tc by $H$.
However $M$ and $M^T$ are conjugate via matrices in $G\setminus H$, e.g. $T=\begin{pmatrix}0&1\\1&0\end{pmatrix}$ does the job for all $s$.

Part 2. Now look at $M:=M_2:=\begin{pmatrix}2&5\\-9&-22\end{pmatrix}\in H$. Take a general $T:=\begin{pmatrix}a&b\\c&d\end{pmatrix}\in G$. The determinant is $D:=\text{det}(T)=ad-bc\in\{\pm 1\}$. Test again: $$ N:=MT-TM^T\stackrel{\text{?}}{=}0. $$ We get $N:=\begin{pmatrix}-5b+5c&9a+24b+5d\\-9a-24c-5d&-9b+9c\end{pmatrix}$. The diagonal entries boil down to $b=c$. Substituting this in the off-diagonal terms shows, that both yield essentially the same equations $9a+24c+5d=0$. We distinguish two cases according to $D\in\{\pm 1\}$ and use Maple to solve the resulting systems:
case A. $D=1$: Maple produces two (families of) solutions, for both $d$ is a free parameter. Set $\sigma=\sqrt{11d^2-9}$, then $(a,b,c,d)=$ $$\left( 3d\pm\frac{8}{9}\sigma,-\frac{4}{3}d\mp\frac{1}{3}\sigma,-\frac{4}{3}d\mp\frac{1}{3}\sigma,d \right). $$ As kabenyuk pointed out, we need to do some computations in quadratic extensions of $\mathbb{Q}$: We need to have $q(d):=\sigma^2=11d^2-9$ a square (in $\mathbb{Z}$). If $q(d)$ is a square it is so $\bmod 11$, thus we need $-9\equiv 2\bmod 11$ to be a square, but the squares $\bmod 11$ are $\{0,1,3,4,5,9\}$. Thus $q(d)$ is never a square and we get no integral (not even rational) transformation matrices. This settles A.
case B. $D=-1$: Except for changing $\sigma=\sqrt{11d^2+9}$ the solutions provided by Maple have the same form as before: $(a,b,c,d)=$ $$\left( 3d\pm\frac{8}{9}\sigma,-\frac{4}{3}d\mp\frac{1}{3}\sigma,-\frac{4}{3}d\mp\frac{1}{3}\sigma,d \right). $$ We need $q(d):=\sigma^2=11d^2+9$ to be a square (in $\mathbb{Z}$). It will turn out that for all $d$ with $q(d)\in\mathbb{Z}$ all possible $a$ are proper fractions with denominator $3$. Here is a list indexed by $j$ (the meaning of $j$ becomes clearer after discussing Pell's equation) with the smallest (in magnitude) $d$ with $q(d)\in\mathbb{Z}$, associated pairs of $a$ and the remainder of the numerator $\tilde{a}\bmod 3$ of these $a$. $$\begin{bmatrix} j & d & a & \tilde{a}\bmod 3 \\ -3 & -3591 & \left\{-\frac{559}{3},-\frac{64079}{3}\right\}& \{1,2\}\\ -2 & -180 & \left\{-\frac{28}{3},-\frac{3212}{3}\right\} & \{1,2\}\\ -1 & -9 & \left\{-\frac{1}{3},-\frac{161}{3}\right\} & \{1,2\}\\ 0 & 0 & \left\{\pm\frac{8}{3}\right\} & \{1,2\}\\ 1 & 9 & \left\{\frac{1}{3},\frac{161}{3}\right\} & \{1,2\}\\ 2 & 180 & \left\{\frac{28}{3},\frac{3212}{3}\right\} & \{1,2\}\\ 3 & 3591 & \left\{\frac{559}{3},\frac{64079}{3}\right\} & \{1,2\} \end{bmatrix}$$ A generalized Pell equation is of the form $X^2-MY^2=N$. See the wikipedia page or lecture notes by Keith Conrad. Here and here and there are some resources to compute solutions for (non-generalized) Pell equations. If we set $M=11, N=9$ any solution $(X,Y)$ will produce a suitable $d=Y$ and any $d$ can be realized this way.
From the general theory we know that there is a bounded set $F=\{(X_{F,i},Y_{F,i})\}_{i=1}^m$ of fundamental solutions to the generalized equation $X^2-MY^2=N$ and a generating solution $(X',Y')$ to the normalized Pell equation $X'^2-MY'^2=1$ giving rise to $u=X'+Y'\sqrt{11}$ such that every solution to the generalized Pell equation $X^2-MY^2=N$ satisfies $X+Y\sqrt{11}=(X_{F,i}+Y_{F,i}\sqrt{11})(X'+Y'\sqrt{11})^{j}$ for some $1\leq i\leq m$ and $j\in\mathbb{Z}$. Doing some computer calculations we get in our case: $$ F:=\{(\pm3,0)\}, u=10+3\sqrt{11} \quad\Rightarrow\quad X+Y\sqrt{11}=\pm 3(10+3\sqrt{11})^{j} $$ Up to sign we can index the solutions by the exponent $j\in\mathbb{Z}$.
The $Y$-part of these solutions gives $d$. Because $a=3d\pm\frac{8}{9}\sigma=3d\pm\frac{8}{9}\sqrt{11d^2+9}$ we need to show, that $11d^2+9=11Y^2+9\not\equiv 0\bmod 27$ in order to prove that $3a\tilde{a}\not\equiv 0\mod 3$, which implies $a\in\mathbb{Q}\setminus\mathbb{Z}$. We show by induction the following

Lemma: If $(X,Y)$ is a solution to $X^2-11Y^2=9$ then $X\equiv \pm 3\bmod 27$ and $Y\in\{0,\pm 9\}\bmod 27$.
proof of Lemma: for $j=0$ we have the solutions $(\pm 3,0)$ which obviously satisfy the lemma.
For $j+1>0$ we assume that a solution $(X_j,Y_j)$ with index $j$ satisfies the lemma. The solution $X_{j+1},Y_{j+1}$ satisfies $X_{j+1}+Y_{j+1}\sqrt{11}=\pm(X_j+Y_j\sqrt{11})(10+3\sqrt{11})$, i.e. $X_{j+1}=\pm(10X_j+33Y_j),Y_{j+1}=\pm(3X_j+10Y_j)$. Computing $\bmod 27$ together with the induction hypothesis on $(X_j,Y_j)$ shows that the property of the lemma also holds for $j+1$.
For $j-1<0$ we have $X_{j-1}+Y_{j-1}\sqrt{11}=\pm(X_j+Y_j\sqrt{11})(10+3\sqrt{11})^{-1}=\pm(X_j+Y_j\sqrt{11})(10-3\sqrt{11})$, i.e. $X_{j+1}=\pm(10X_j-33Y_j),Y_{j+1}=\pm(-3X_j+10Y_j)$. Computing $\bmod 27$ again induction goes through.

By the lemma we have $Y\in\{0,\pm 9\}\bmod 27$ which implies $11d^2+9=11Y^2+9\equiv 9\not\equiv 0\bmod 27$. This finishes case B. $\text{det}(T)=-1$.

In summary this proves, that $M_2=\begin{pmatrix}2&5\\-9&-22\end{pmatrix}\in H$ is not tc by $G$.

Part 3.

property 1: Symmetric matrices are trivially tc, they can be ignored when studying tc.

property 2: TFAE:

  1. $M=\begin{pmatrix}e&f\\g&h\end{pmatrix}$ is tc,
  2. $M^T$ is tc,
  3. $-M$ is tc,
  4. any conjugate $M'=U^{-1}MU$ is tc, in particular $\begin{pmatrix}h&g\\f&e\end{pmatrix}=\begin{pmatrix}0&1\\1&0\end{pmatrix}^{-1}M\begin{pmatrix}0&1\\1&0\end{pmatrix}$ is tc,
  5. $M':=\begin{pmatrix}e&-f\\-g&h\end{pmatrix}$ is tc.

The equivalences 1. $\Leftrightarrow$ 2. $\Leftrightarrow$ 3. are easy. 1. $\Leftrightarrow$ 4. follows from $$ T^{-1}MT=M^T \Leftrightarrow T'^{-1}M'T'=M'^T, \text{ where } T'=U^{-1}TU^{-T} $$ and 1. $\Leftrightarrow$ 5. from $$ \begin{pmatrix}a&b\\c&d\end{pmatrix}^{-1}\begin{pmatrix}e&f\\g&h\end{pmatrix}\begin{pmatrix}a&b\\c&d\end{pmatrix}=\begin{pmatrix}E&F\\G&H\end{pmatrix}\Rightarrow $$ $$ \begin{pmatrix}a&-b\\-c&d\end{pmatrix}^{-1}\begin{pmatrix}e&-f\\-g&h\end{pmatrix}\begin{pmatrix}a&-b\\-c&d\end{pmatrix}=\begin{pmatrix}E&-F\\-G&H\end{pmatrix} $$

property 3: If any entry of $M=\begin{pmatrix}e&f\\g&h\end{pmatrix}\in H$ vanishes then $M$ is tc. This can be seen from the following list of $M\in H=\text{SL}(2,\mathbb{Z})$ and $T\in G=\text{GL}(2,\mathbb{Z})$ which satisfy $T^{-1}MT=M^T$ (Note that part 1. is part of this list): $$ M:=\begin{pmatrix}0&\pm 1\\\mp 1&h\end{pmatrix}, T:=\begin{pmatrix}1&0\\0&-1\end{pmatrix},\quad M:=\begin{pmatrix}\pm 1&0\\g&\pm 1\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&0\end{pmatrix}, $$ $$ M:=\begin{pmatrix}\pm 1&f\\0&\pm 1\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&0\end{pmatrix},\quad M:=\begin{pmatrix}e&\pm 1\\\mp 1&0\end{pmatrix}, T:=\begin{pmatrix}1&0\\0&-1\end{pmatrix}. $$ This remains true even for $M\in G\setminus H$ as implied by these matrices where $2\mid g_e\in\mathbb{Z}$ and $2\nmid g_o\in\mathbb{Z}$ together with some easy to check further cases: $$ M:=\begin{pmatrix}1&0\\g_e&-1\end{pmatrix}, T:=\begin{pmatrix}1&\frac{g_e}{2}\\\frac{g_e}{2}&\left(\frac{g_e}{2}\right)^2-1\end{pmatrix},\quad M:=\begin{pmatrix}1&0\\g_o&-1\end{pmatrix}, T:=\begin{pmatrix}2&g_o\\g_o&\frac{g_o^2-1}{2}\end{pmatrix}. $$

These properties imply: If we want to study tc for $M=\begin{pmatrix}e&f\\g&h\end{pmatrix}\in H$ we may assume wlog: $$0<e\neq h,|e|\leq|h|, 0<f\neq g, |f|\leq |g|.$$ ($e\neq h$ is implied by one of the example families below.)

One can use these properties to search for $M\in H$ with "small entries" that might not be tc. For $\max\{|e|,|f|,|g|,|h|\}\leq 25$ and $|eh|\leq 50$ a heuristic search only left three candidates for not-tc up to the symmetries of property 2: $$ M_2=\begin{pmatrix}2&5\\-9&-22\end{pmatrix},\quad \begin{pmatrix}2&5\\7&18\end{pmatrix},\quad \begin{pmatrix}2&5\\9&23\end{pmatrix}. $$ For $\max\{|e|,|f|,|g|,|h|\}\leq 15$ without restrictions on the diagonal the heuristic search leaves us with two candidates for not-tc: $$ \begin{pmatrix}7&9\\10&13\end{pmatrix},\quad \begin{pmatrix}9&7\\14&11\end{pmatrix}. $$

Magma verifies quickly that all of these candidates are also not tc:

G := GeneralLinearGroup(2,Integers());
ML := [elt<G|2,5,-9,-22>,elt<G|2,5,7,18>,elt<G|2,5,9,23>,elt<G|7,9,10,13>,elt<G|9,7,14,11>,elt<G|4,3,13,10>];
M := ML[2];
MT := Transpose(M);
b,T := AreGLConjugate(M, MT);
M;
MT;
b;

Finally here is a list of tc $M\in G$ together with some $T\in G$ as witnesses (here $x,y,z\in\mathbb{Z}$ are any integers, such that $\det(M)\in\{-1,1\}$): $$ M:=\begin{pmatrix}1&1\\x&x+1\end{pmatrix}, T:=\begin{pmatrix}-1&1\\1&0\end{pmatrix},\quad M:=\begin{pmatrix}1&1\\x&x-1\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&x-2\end{pmatrix}, $$ $$ M:=\begin{pmatrix}1&x\\x^2&x^3+1\end{pmatrix}, T:=\begin{pmatrix}x&-1\\-1&0\end{pmatrix},\quad M:=\begin{pmatrix}x&y\\z&x\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&0\end{pmatrix}, $$ $$ M:=\begin{pmatrix}3&2\\4&3\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&0\end{pmatrix},\quad M:=\begin{pmatrix}3&1\\8&3\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&0\end{pmatrix}, $$ $$ M:=\begin{pmatrix}2&3\\9&14\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&4\end{pmatrix},\quad M:=\begin{pmatrix}4&3\\9&7\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&1\end{pmatrix}, $$ $$ M:=\begin{pmatrix}7&3\\9&4\end{pmatrix}, T:=\begin{pmatrix}0&1\\1&-1\end{pmatrix},\quad M:=\begin{pmatrix}14&3\\9&2\end{pmatrix}, T:=\begin{pmatrix}3&2\\2&1\end{pmatrix}, $$ $$ M:=\begin{pmatrix}28&3\\9&1\end{pmatrix}, T:=\begin{pmatrix}3&1\\1&0\end{pmatrix}, $$

Remark: This answer went through a lot of editing to make it clearer and to reflect simplifying insights.

Thomas Preu
  • 2,002
  • What I meant in my comment was the following observation. (I am using the notations from my answer.)

    Since $xv-y^2=\pm1$ and $bx+(d-a)y-cv=0$, then $$ bx^2+(d-a)xy-cy^2=\pm c $$ or $$ cv^2-(d-a)vy-by^2=\pm b. $$ The discriminant of this quadratic form is $\operatorname{tr}^2(A)-4$. Then we can do computations as it is done in the book Number theory Borevich Z.I., Shafarevich I.R. (Chap. 2, Sect 7). Frankly speaking, I did these calculations only for a few simple cases.

    – kabenyuk Aug 09 '22 at 16:23
  • I think the article is useful in this matter – kabenyuk Aug 09 '22 at 17:02
  • I find the massive wall of text in your answer hard to read. Can you please directly and concisely say at the very start of the post if you have actually proved an example works or if you have not (and if you have, exactly for which matrix or matrices)? At the end of your post you say there is a proof, but earlier in the post you say you just have plausibility arguments, and you use the same letter $M$ to denote different matrices. I can prove the very first matrix you call $M$ is an example by converting the problem into a task in algebraic number theory. – KCd Aug 10 '22 at 02:19
  • For your first matrix $M$, my calculations with ideals in a certain quadratic ring led me to the Pell equation $X^2 - 11Y^2 = 1$, and the assumption that $M$ and $M^\top$ are conjugate in ${\rm GL}_2(\mathbf Z)$ would imply $X \equiv 0 \bmod 3$, but because $X + Y\sqrt{11} = \pm(10 + 3\sqrt{11})^j$ for some $j \in \mathbf Z$ it turns out by induction on $j$ that $X \equiv \pm 1 \bmod 3$, so that's a contradiction. – KCd Aug 10 '22 at 11:14
  • The end of your updated post lists "three candidates" (called $M_2$) that might be transpose-conjugate, but you don't state clearly what the status is of the 2nd and 3rd examples in that list of three matrices. – KCd Aug 11 '22 at 21:23
  • Your property 3 says the $M$ with a vanishing entry are “not tc” when you meant to say that they are tc. By the way, do you know you can ping someone in a comment using @ in front of their username? I have been checking this page each day to see if you edited your post because I never got a comment that you responded to anything I wrote here. – KCd Aug 13 '22 at 12:07
  • @KCd Thank you for your remarks, I try to incorporate them. As I am only participating regularly at MSE since about a week I am not sure how to use different parts of the site in and out. From here and there I got the impression I should keep comments to a minimum. Apologies if that was inapropriate. – Thomas Preu Aug 13 '22 at 12:30