# Confusion about centraliser of $D_5$ in $GL_4(\mathbb{Z})$

I am trying to follow a derivation on a very old paper. My knowledge of group theory is limited, I have the basis but not much experience with advanced concepts.

We are working in 4 dimensions, so the paper quotes the 4D representations of the generators of $D_5$, $r$ (a rotation by $\pi/5$) and $p$ (the reflection $x \rightarrow x, y \rightarrow y$) as: $$\mathcal{R}(r) = \left (\begin{array}{ccc} 0 & 0 & 0 & -1 \\ 1 & 0 & 0 & -1 \\ 0 & 1 & 0 & -1 \\ 0 & 0 & 1 & -1 \\ \end{array} \right ),$$

$$\mathcal{R}(p) = \left (\begin{array}{ccc} 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ \end{array} \right ).$$

Question 1: I thought all dihedral groups had irreducible representations of at most 2 dimensions, hence all >2 dimensional representations should be constructed from direct sums of these (so block diagonals)? Where do they get these expressions from?

Then we calculate the centraliser of $D_5$ in $GL_4(\mathbb{Z})$, i.e. the largest subgroup of $GL_4(\mathbb{Z})$ which commutes with $D_5$.
Unfortunately they skip directly to the answer because apparently "computing these groups in straightforward".
The generators of the centraliser $C$ are given by:

$$\mathcal{R}(\delta) = \left (\begin{array}{ccc} -1 & 1 & 0 & -1 \\ 0 & 0 & 1 & -1 \\ -1 & 1 & 0 & 0 \\ -1 & 0 & 1 & -1 \\ \end{array} \right ),$$

$$\mathcal{R}(\tau) = \left (\begin{array}{ccc} -1 & 0 & 0 & 0 \\ 0 & -1 & 0 & 0 \\ 0 & 0 & -1 & 0 \\ 0 & 0 & 0 & -1 \\ \end{array} \right ).$$

Question 2: where do these come from? Is it there something very trivial that I'm not getting? Where would I even start?

My attempts:
1)I started this by building a 4D representation of $D_5$ just by direct summing its 2D irreps. I am using $r_{2D}$ as the 2D rotation matrix and $p_{2D} = \left (\begin{array}{cc} 1 & 0 \\ 0 & -1 \\ \end{array} \right ),$, to then make $\mathcal{R}(r_{4D}) = \left (\begin{array}{cc} r_{2D} & 0 \\ 0 & r_{2D} \\ \end{array} \right ),$ $\mathcal{R}(p_{4D}) = \left (\begin{array}{ccc} 1 & 0 & 0 & 0 \\ 0 & -1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & -1 \\ \end{array} \right ),$ etc.

2) I used the generic element of $GL_4(\mathbb{Z})$ as $\beta = \left (\begin{array}{ccc} a & b & c & d \\ e & f & g & h \\ i & l & m & n \\ q & s & t & u \\ \end{array} \right ).$
Then I brute-forced computed the relationships between the entries by requiring $h$ to commute with every element $d$ of $D_5$: $h^{-1}dh = d$.
I found that the matrix above is reduced to $\beta = \left (\begin{array}{ccc} a & 0 & c & 0 \\ 0 & a & 0 & c \\ s & 0 & u & 0 \\ 0 & s & 0 & u \\ \end{array} \right ).$ However this is where I stop, for I don't know how to compute the generator of the group of these matrices.

• This group has a permutation representation on 5 points, use that to get 5x5 permutation matrices. Then note that there's a 1 dimensional subspace fixed by these matrices given by the vectors with all coordinates the same. The other 4 dimensions give you the 4x4 matrices. Another way to see that is to look at how these matrices act on (1,0,0,0), (0,1,0,0), (0,0,1,0), (0,0,0,1), and (-1,-1,-1,-1) – Josh B. Oct 28 '17 at 21:54
• Where do you get the 1 dimensional subspace from? And why also the vector (-1, -1, -1, -1=? – SuperCiocia Oct 28 '17 at 22:44
• The 1 dimensional subspace is all the vectors with the same coordinates. – Josh B. Oct 28 '17 at 23:04
• I won't lie, I don't really see how everything unfolds from your answer. But if that's all you've got time for, then I thank you anyway. – SuperCiocia Oct 28 '17 at 23:34
• This was only a comment to try to motivate where the initial matrices come from. I don't know how to answer your question 2. Hopefully someone else more knowledgeable will come along. – Josh B. Oct 28 '17 at 23:40

$\def\glz{\operatorname{GL}_4(\mathbb Z)}$ $\def\inv{^{-1}}$

The question is about filling in some details about assertions made in this paper. The answer got quite long as I added more details. Guide: the part of the answer through the first edit shows that the commutant of the copy of $D_5$ in $\operatorname{GL}_4(\mathbb C)$ is two dimensional, abelian, and generated as a unital algebra by a certain matrix $M = R + R\inv$. However, the copy of $D_5$ and the matrix $M$ both lie in $\glz$, and the subsequent part of the argument in the 4th postlude shows that $M$ and $- 1$ (minus the identity matrix) generate the centralizer of $D_5$ in $\glz$ as a group. In between, I added more detail about the first part, and also explained the construction of the 4 dimensional representation of $D_5$ given in the paper.

I'm going to write $R$ and $J$ for your two matrix generators. $R$ is diagonalizable with eigenvectors the 4 primitive 5th roots of 1, call them $\omega, \omega^{-1}, \omega^2, \omega^{-2}$. In the basis of eigenvectors, matrices commuting with your $D_5$ in particular commute with $R$, so are diagonal. Also in the basis of eigenvectors, $J$, up to some scaling which is probably irrelevant, permutes the eigenvectors in pairs (belonging to inverse pairs of eigenvalues). Thus your diagonal commuting matrices have only two distinct diagonal entries. Thus the commutant is commutative, generated by two commuting idempotents of rank 2.

Now, still in the basis of eigenvectors, $M = R + R^{-1}$ is a diagonal matrix with just two eigenvalues $\omega + \omega^{-1}$ and $\omega^2 + \omega^{-2}$ corresponding to the range of the two desired idempotents. So in fact, the whole commutant is generated by polynomials in $M$.

Note that this is where $\mathcal R(\delta)$ came from, it is $R + R^{-1}$, in the original basis, not the basis of eigenvectors.

I've considered the problem from the point of view of the commuting algebra, but the centralizer subgroup just consists of the invertible elements in the commuting algebra. I've also worked over the complex numbers. Anyway, I think this is close to what you need.

Edit: Looking at this from more of a group representations point of view, we can get to the same place more conceptually, perhaps. Clearly $R + R^{-1}$ is in the commutant for any representation. So $R + R^{-1}$ acts as a scalar in each irreducible representation. In a 2 dimensional irreducible representation, the value of that scalar is the character of $R$, which is $2 \cos(2 \pi/k)$, where $k$ can be taken as a label for the representation. In either of the two one dimensional representations, $R + R^{-1}$ acts as the scalar $2$.

Now in our particular representation, by examining the spectrum of $R + R^{-1}$ or of $R$, you can see that both 2 dimensional irreducible representations occur with multiplicity 1, so it follows that $R + R^{-1}$ actually generates the commutant (as an algebra).

2nd edit: You asked how to use polynomials in $R + R^{-1}$ to generate the whole commutant. In fact linear polynomials will do, because of the special structure here, as I will next explain.

Mini-mini-course in representation theory: Suppose I have a group representation $\rho : G \to \operatorname{GL}(V)$, and it decomposes as a direct sum of mutually inequivalent irreducible subrepresentations, $V = W_1 \oplus W_2 \oplus \cdots \oplus W_n$. Let $P_j$ be the projection operator which is $1$ on $W_j$ and zero on $W_i$ for $i \ne j$. Thus $P_j^2 = P_j$, $P_i P_j = 0$ for $i \ne j$ and $\sum_j P_j = 1$, where I write $1$ for the identity operator. Then the entire commutant $\mathcal C$ of $\rho(G)$ consists of linear combinations of $P_1,\dots, P_n$.

Mini-mini-course in linear algebra Now suppose I have a collection of projection operators as above $P_1,\dots, P_n$ with $P_j^2 = P_j$, $P_i P_j = 0$ for $i \ne j$ and $\sum_j P_j = 1$. No matter if they came from a group representation or from somewhere else. Let $W_j = P_j V$, so $V$ is the direct sum of the subspaces $W_j$. Let $\mathcal C$ be the algebra generated by the projection operators $P_j$; in fact, it consists of linear combinations of the operators $P_j$. Suppose someone hands me a particular linear combination $T= \sum_i \alpha_i P_i$ with all the $\alpha_i$ distinct. Then I claim that I can recover the $P_j$ and hence all of $\mathcal C$ from polynomials in $T$, and if $n = 2$ I can do it with linear polynomials.

This goes by the name interpolation formulas.

Fix $i$ and consider
$$\prod_{j \ne i} \frac{ T - \alpha_j 1}{\alpha_i - \alpha_j},$$ which is a polynomial of degree $n-1$ in $T$. You can check that applied to a vector in $W_k$ for $k \ne i$, it gives zero, but applied to a vector in $W_i$ it gives back the vector. Thus the polynomial in $T$ is actually equal to $P_i$.

3rd postlude I see from your discussion/chat that you are still puzzled about the derivation of the 4d representation matrices. As "explained" in the physics paper, the representation was derived from the 5d permutation representation exactly as Josh B. suggested. First of all, you have to know that given a representation $V$ and a subrepresentation $W$, you get a quotient representation on $V/W$. Start with the permutation representation of $D_5$ on $F^5$, $F$ your favorite field, given by $$R: e_0 \to e_1 \to e_2 \to e_3 \to e_4 \to e_0$$ where $e_i$ are the standard basis elements of $F^5$ (with shifted labels) and $$J: e_0 \to e_0,\quad e_1 \to e_4 \to e_1, \quad e_2 \to e_3 \to e_2.$$ Now $w = \sum_{j = 0}^4 e_j$ is fixed by both $R$ and $J$. In fact, it's fixed by all permutation matrices. Let $W = F w$ and $V = F^5/W$. Give $V$ the basis $\overline e_1 = e_1 + W, \dots, \overline e_4 = e_4 + W$. Note $e_0 + W = -(\overline e_1 + \overline e_2 + \overline e_3 + \overline e_4)$, since $e_0 = -\sum_{j = 1}^4 e_j + w$. It follows that the matrices of $R$ and $J$ in $V = F^5/W$ with respect to the basis $\{\overline e_i\}_{1 \le i \le 4}$ are exactly those that you quoted from the paper.

4th postlude: I finished it off.

Demonstration that the centralizer of $D_5$ in $\glz$ is generated (as a group) by $M = R + R\inv$ and $- 1$. Recall that $\glz$ is the set of integer matrices with determinant $\pm 1$.

We already know that any matrix that commutes with $D_5$ pointwise is a linear combination of $M$ and the identity matrix. We need to find out when $a M + b 1$ has integer entries and determinant $\pm 1$. By inspection, $$a M + b 1 =\left( \begin{array}{cccc} b-a & a & 0 & -a \\ 0 & b & a & -a \\ -a & a & b & 0 \\ -a & 0 & a & b-a \\ \end{array} \right)$$ is integer valued if and only if $a$ and $b$ are integers. By computation, $$\det(a M + b 1) = \left(a^2+a b-b^2\right)^2,$$ and in particular $\det M = 1$. Moreover, $\det(a M + b 1) \ge 0$, so $$a M + b 1 \in \glz \implies \det(a M + b 1) = 1.$$

The minimal polynomial for $M$ is $x^2 + x -1$, as follows for example from the computation of the eigenvalues of $M$. Thus we have $$M\inv = (M + 1) \quad \text{and} \quad M^2 = (1 - M).$$

Observation 1. Suppose $a M + b 1 \in \glz$ then $a$ and $b$ are relatively prime.

Proof. If $a$ and $b$ have a common prime factor $p$, then $\det(a M + b 1)$ is divisible by $p^4$. qed

Suppose that $T = a M + b 1 \in \glz$. We want to show that $T$ is in the subgroup of $\glz$ generated by $M$ and $- 1$. If one of the coefficients $a, b$ is zero, then $T = \pm M$ or $T = \pm 1$, and we are done, If both of the coefficients $a, b$ are $\pm 1$, then $T = \pm M\inv$ or $T = \pm M^2$, by the discussion of the minimal polynomial, so again we are done. Now we can proceed by induction on $|a| + |b|$.

Let us consider the case that $a b > 0$. By extracting a factor of $-1$ (i.e. the matrix $-1$), we can assume $a, b > 0$, and relatively prime. Moreover, we are only interested in the case that $a b > 1$. In particular neither of $a, b$ is divisible by the other.

Observation 2. $b > a > 0$.

Proof. If $a >b$ then $$a ^2 + ab - b^2 > a^2 - b^2 = (a -b)(a + b) > a + b > a > 1.$$ Hence, $\det(T) >1$, a contradiction. qed

Now we take our element $T = a M + b 1 \in \glz$ and factor out $M\inv$: $$T = M\inv (a M^2 + b M) = M\inv ( a(1 - M) + b M) = M\inv ( (b - a) M + a 1).$$ The factor $(b - a) M + a 1$ is necessarily in $\glz$, and has positive coefficients whose sum $b$ is less than the sum of the coefficients of $T$. So our conclusion follows from the induction hypothesis.

Next consider the case that $a b < 0$. By extracting a factor of $-1$ if necessary, we can assume $a > 0 > b$, and $|a b| > 1$.

Observation 3. $a > -b$.

Proof. if $-b > a$, then $$b^2 - ab - a^2 > b^2 - a^2 = (b-a)(b+a) > 1.$$ Hence $\det(T) > 1$, a contradiction. qed

We take our element $T = a M + b 1 \in \glz$ and factor out $M$: $$T = M( a 1 + b M\inv) = M ( a 1 + b (M + 1)) = M ( b M + (a + b) 1).$$ The factor $b M + (a + b) 1$ is in $\glz$ and has coefficients $a + b > 0$ and $b <0$. The sum of absolute values of these coefficients is $(a + b) - b = a$, which is less than the corresponding sum for the coefficients of $T$, namely $a - b$. Again our conclusion follows from the induction hypothesis.

• First of all thanks for the answer. I can see and follow your reasoning for $R$, but I don't quite follow how you get $J$ and deduce that the commuting matrices can only have two distinct entries (are you counting the different roots of unity as distinct)? What exactly do you mean by 'commutant'? The general matrix that commutes with everything in $D_5$? How do I explicitly work with polynomails in $M$ to generate the whole commutant? Is my attemp at getting the generic commuting matrix from $GL$ wrong? – SuperCiocia Oct 31 '17 at 0:26
• Also, if $R(\delta)$ comes from $R+R^{-1}$, where does $R(\tau)$ come from and how is it related to $R(\delta)$? – SuperCiocia Oct 31 '17 at 0:27
• Also, how can $R(\delta)$ that I wrote in my question be simply $R + R^{-1}$, if some of the $\omega$ are not integers? – SuperCiocia Oct 31 '17 at 0:39
• As you say, "commutant" means the algebra of all matrices commuting with everything in $D_5$. To see what $J$ must do, apply $RJ = J R^{-1}$ to an eigenvector of $R$; conclude if $v$ is an eigenvector with eigenvalue $\lambda$, then $Jv$ is an eigenvector with eigenvalue $\lambda^{-1}$. An element of the commutant acts diagonally on the basis of eigenvectors, but must also commute with $J$, and $J$ interchanges pairs of eigenvectors; therefore the element of the commutant must act by the same scalar on each such pair. – fredgoodman Oct 31 '17 at 0:42
• But I think you can just as well forget everything before the edit, as the second argument is more efficient and more illuminating. $R(\delta) = R + R^{-1}$ in the original basis (just compute it!). I checked this after realizing that $R + R^{-1}$ is an important matrix. (In the basis of eigenvectors, it is diagonal with the two different eigenvalues, as I wrote.) As for $R(\tau)$, it is just a scalar matrix, so it doesn't do anything for us. I haven't yet looked at the original paper, so I don't know why the authors are interested in it. – fredgoodman Oct 31 '17 at 0:52

$\def\glz{\operatorname{GL}_4(\mathbb Z)}$ $\def\inv{^{-1}}$

I'm posting some further observations, as if my previous gigantic post wasn't enough.

We have specific matrices $R = \mathcal R(r)$ and $J = \mathcal R(p)$ in $\glz$ providing a representation of the dihedral group $D_5$. It is required to find the centralizer of $D_5$ in $\glz$

We also have the matrix $M = R + R\inv$, denoted $\mathcal R(\delta)$ in the original post, which is an element of $\glz$ commuting with $D_5$.

The first step is to find the commutant of $D_5$, that is the set of all matrices commuting with all elements of $D_5$.

Your original brute force method for computing the commutant of $R$ works fine. I asked Mathematica to do it, and was told the solutions are of the form $$\left( \begin{array}{cccc} a & b & 0 & -b \\ 0 & a+b & b & -b \\ -b & b & a+b & 0 \\ -b & 0 & b & a \\ \end{array} \right)$$ which by examination is $b M + (a + b) 1$. So this gives back the solution for the commutant, that it is the set of linear combinations of $M$ and the identity, without any of the "fancy" arguments from my previous answer.

The determinant of $a M + b 1$ is $D(a, b) = (a^2 + ab - b^2)^2$. So we have a Diophantine problem to determine the integral solutions of $D(a, b) = 1$.

I'm not going to provide a simpler solution to this problem than already appeared in my other answer (4th postlude), but I want to point out that Fibonacci numbers pervade this problem.

In fact, we can first observe that the sequence of powers of $M$ is a matrix Fibonacci sequence in the following sense. Using the minimal polynomial of $M$, namely $x^2 + x - 1$ repeatedly, we find that, with $(f_k)_{k \ge 0}$ denoting the Fibonacci sequence, $$(-M)^k = -f_k M + f_{k-1} \quad \text{and} \quad M^{-k} = f_k M + f_{k+1} 1,$$ for $k \ge 1$. So what we have to show is that an integer pair $(a, b)$ satisfies $D(a, b) = 1$ if and only if $$(a, b) = \pm (-f_k, f_{k-1}) \quad \text{or} \quad (a, b) = \pm(f_k, f_{k+1}),$$ for some $k$. (I am extending the Fibonacci sequence by taking $f_0 = 0$.) This requires more or less the same sort of work as was contained in my previous 4th postlude, and I am going to skip the details. (This is known: Jones, James P. Diophantine representation of the Fibonacci numbers. Fibonacci Quart. 13 (1975), 84–88. )

In any case the OP is interested in generalizing these results to $D_n$, $n >5$, and all of these Fibonacci phenomena are very much tied to the case of $D_5$.

Edit: I am adding this in response to some questions posed to me by the OP.

$\def\glz{\operatorname{GL}_4(\mathbb Z)} \def\inv{^{-1}} \def\boldG{\mathbf G} \def\boldn{\mathbf n} \def\Z{\mathbb Z} \def\R{\mathbb R} \def\boldq{\mathbf q}$ Let me start over with a few things.

The starting point in the paper is the set of points $\boldG_j = (\cos(2 \pi j/ 5), \sin(2 \pi j/5))$, $0 \le j \le 4$ in the real vector space $\mathbb R^2$. (It is also convenient to take the indices as representing residue classes mod 5.) These $\boldG_j$ satisfy $\sum_{j = 0}^4 \boldG_j = 0$, but $\{\boldG_j : 1 \le j \le 4\}$ is linearly independent over $\mathbb Z$, while of course linearly dependent over $\mathbb R$. So the set $P$ of integer linear combinations of $\{\boldG_j : 1 \le j \le 4\}$ is a complicated configuration of points in the plane, but is also a free $\mathbb Z$--module of rank $4$. The dihedral group $D_5$, generated by the geometric rotation $\tau$ by angle $2 \pi/5$ in the plane and the reflection through the $x$-axis $\sigma$ leaves invariant the set of $\{\boldG_j : 0 \le j \le 4\}$. The group $D_5$ acts on $\mathbb R^2$ by real linear automorphisms, and hence acts on $P$ by $\mathbb Z$--linear automorphisms. The matrices of the generators $\tau$, $\sigma$ with respect to the $\Z$--basis $\{\boldG_j : 1 \le j \le 4\}$ of $P$ are the matrices $M(\tau)$ and $M(\sigma)$ given in the paper.

We have a $\Z$--module isomorphism $\boldq : \Z^4 \to P$, $\boldq(\boldn) = \sum_{j = 1}^4 n_j \boldG_j$. This gives us a representation of $D_5$ on $\Z^4$ with the same representation matrices. We can regard this representation as a representation on $\mathbb R^4$ or on $\mathbb C^4$, or whatever, by using the same representation matrices.

There is also an element $\delta = \tau + \tau\inv$. (It lives in the the group ring consisting of formal linear combinations of the group elements of $D^5$.). It acts on $P$ or on $\Z^4$ by the matrix $M(\delta) = M(\tau) + M(\tau)\inv$.

All of these transformations $\tau, \sigma, \delta$ will act via the same matrices, either on $P$, with respect to the basis $\{\boldG_j : 1 \le j \le 4\}$ or on $\Z^4$ with respect to the standard basis.