For which integers $d \geq 1$ can we find real matrices $R_1, \dotsc, R_d$ of size $d \times d$ such that for any unit vector $v \in \mathbb{R}^d$, $$R_1 v, \dotsc, R_d v$$ is an orthonormal basis? Note that the chosen set of $R_i$s has to work simultaneously for all $v$. (Does this phenomenon have a name?)

It may not be obvious at first, but this question seems to be related to associative division algebras (i.e., reals, complex numbers, and quaternions). Indeed, I know that such matrices $R_i$ can be found when $d=1,2,4$:

  • $d=1$ is trivial—simply take $$R_1 = \begin{pmatrix}1\end{pmatrix}$$
  • $d=2$ is based on complex numbers—we represent $1$ and $i$ by $2 \times 2$ real matrices as follows: $$ R_1 = \begin{pmatrix}1&0\\0&1\end{pmatrix}\qquad R_2 = \begin{pmatrix}0&-1\\1&0\end{pmatrix} $$
  • $d=4$ is based on quaternions—we represent $1$, $i$, $j$, $k$ by $4 \times 4$ matrices as follows: $$ R_1 = \begin{pmatrix}1&0&0&0\\0&1&0&0\\0&0&1&0\\0&0&0&1\end{pmatrix}\qquad R_2 = \begin{pmatrix}0&1&0&0\\-1&0&0&0\\0&0&0&-1\\0&0&1&0\end{pmatrix}\qquad\\ R_3 = \begin{pmatrix}0&0&1&0\\0&0&0&1\\-1&0&0&0\\0&-1&0&0\end{pmatrix}\qquad R_4 = \begin{pmatrix}0&0&0&1\\0&0&-1&0\\0&1&0&0\\-1&0&0&0\end{pmatrix}\qquad $$

Is this possible for any other $d$? I.e., can we find $R_i$s even when there is no associative division algebra of dimension $d$?

  • $\begingroup$ I think associativity is irrelevant. What you want is probably closer to en.wikipedia.org/wiki/Composition_algebra. $\endgroup$ – Qiaochu Yuan Aug 19 '15 at 9:30
  • $\begingroup$ I disagree. I can't use octonions for $d=8$ since they are not associative and thus can't be represented by matrices (because matrix multiplication is associative). $\endgroup$ – Māris Ozols Aug 19 '15 at 10:07
  • 1
    $\begingroup$ I have to take this back - octonions indeed work (see the answer by Adam Przeździecki). $\endgroup$ – Māris Ozols Aug 19 '15 at 15:07
  • $\begingroup$ This phenomenon has some interesting applications in quantum information which I was not aware of at the time I asked the question (see dx.doi.org/10.1103/PhysRevA.65.022316). $\endgroup$ – Māris Ozols Feb 2 '16 at 20:26
  1. Consider the matrix $\sum_i a_iR_i$. One can show that it sends every vector of length 1 to a vector of length $\sqrt{\sum_i a_i^2}$. It follows that if the norm of $a=(a_i)$ is 1, then the matrix $\sum_i a_iR_i$ is orthogonal. In particular, all the matrices $R_i$ are orthogonal.
  2. By multiplication from the left or from the right with an orthogonal matrix, we can assume without loss of generality that $R_1=Id$ (the collection $\{SR_iT\}$ where $S$ and $T$ are orthogonal matrices will still satisfy the condition.
  3. Consider now $R_2$. Then $R_2$ cannot have a real eigenvalue $\lambda$, because then $R_2-\lambda R_1$ will have nontrivial kernel, which contradicts 1. It follows that, since $R_2$ is orthogonal, it can be written as the direct sum of $2\times 2$ rotation matrices. If one of these rotation angles is $\theta$, then by considering 1. for the matrix $R_2+R_1$ we reach the conclusion that $\cos(\theta)=0$. So $\theta=90$ or $270$. In any case, it follows that $R_2^2 = -1$, and the same is true for $R_i$ $i=3,\ldots, d$ of course. We also get that $d$ is even.
  4. We can now multiply the collection $\{R_i\}$ from the left by $-R_2$. Again we will receive a collection of matrices which satisfies the condition, and contains the matrix $-R_2^2=Id$. But then it follows that all the non-unital matrices in this collection square to -1. In other words, we get that $(R_iR_j)^2=-1$ for every $i\neq j$, $i,j>1$.
  5. Consider now the algebra $\mathbb{R}\langle X_1,\ldots X_{d-1}\rangle/(X_i^2=-1, X_iX_j + X_jX_i=0 (i\neq j))$. This algebra is a twisted group algebra with the group $G=(\mathbb{Z}/2)^{d-1}$. This algebra is semisimple, and one can prove directly that its center is 2 dimensional and generated by $X:=X_1\cdots X_{d-1}$.
    If $r=2 \mod 4$, then $X^2=-1$. The center is then isomorphic with $\mathbb{C}$, so we get a central simple algebra of dimension $2^{d-2}$ over $\mathbb{C}$, which must be the matrix algebra $M_n(\mathbb{C})$, where $n= 2^{(d-2)/2}$. This algebra has pnly one irreducible representation, and it is of dimension $2^{(d-2)/2}$ over $\mathbb{C}$, and therefore of dimension $2^{d/2}$ over $\mathbb{R}$. This rules out the possibility that $r=6,10,14...$.
    If $r=0 \mod 4$, then $X^2 = 1$, and the center is isomorphic with $\mathbb{R}\oplus\mathbb{R}$, and the algebra splits as the direct sum of two central simple $\mathbb{R}$-algebras, each of dimension $2^{(d-2)}$. Such an algebra will either be isomorphic with $M_n(\mathbb{R})$ where $n = 2^{(d-2)/2}$, or with $M_n(D)$ where $D$ is the quaternion algebra and $n= 2^{(d-6)/2}$. In the first case the algebra has a unique simple representation of dimension $2^{(d-2)/2}$, and in the second case it has a unique irreducible representation of dimension $2^{(d+2)/2}$. For $d>8$ both dimensions are bigger than $d$ so there is no solution. For $d=8$ we can study carefully the algebra and find a solution, the one given by Adam. In order to analyze the algebra we can, for example, consider the subalgebra generated by $X_1,X_2$ (in the proper quotient) which is isomorphic with the quaternion algebra. Then the algebra will be a tensor product of the quaternion algebra and its centralizer. This reduces the question to studying an algebra of a smaller dimension, and we can find out that at the end we get a matrix algebra over the reals, and so we have a solution for $d=8$.
  • $\begingroup$ I'm not following the last point entirely. Is this the $\mathcal{C}\ell_{0,d-1}(\mathbb{R})$ Clifford algebra? Can you give a reference why it is a direct sum of two central simple algebras isomorphic to $M_n(D)$? How do you get $2^{d/2}$ in the end? $\endgroup$ – Māris Ozols Aug 19 '15 at 14:03
  • $\begingroup$ It seems that there must be a gap in your argument, since $d=8$ is actually possible (see the answer by Adam Przeździecki). Could you elaborate a bit more on step 5? $\endgroup$ – Māris Ozols Aug 19 '15 at 15:43
  • $\begingroup$ You right, I was not careful enough. I edited my answer. $\endgroup$ – Ehud Meir Aug 19 '15 at 16:17

One more example: for $d=8$ based on octonions. And this are all possible dimensions. $R_1$ takes vectors of length one to vectors of length one hence it is an isometry. By multiplying from the left by $R_1^{-1}$ we may assume that $R_1=\mathop{\rm Id}$. Then for every $v\in S^{d-1}$, the unit sphere in $\mathbb{R}^d$, the map $v\mapsto(R_2v,R_3v,\ldots,R_dv)$ gives a parallelization of the tangent bundle of $S^{d-1}$. Then by a celebrated Bott-Kervaire-Milnor theorem such a parallelization exists only when $d\in\{1,2,4,8\}$.

Edit: More explicitly, the solution for $d = 8$ is given by $$ \sum_{i=1}^8 a_i R_i = \begin{pmatrix} a_1 & a_2 & a_3 & a_4 & a_5 & a_6 & a_7 & a_8 \\ -a_2 & a_1 & -a_4 & a_3 & -a_6 & a_5 & a_8 & -a_7 \\ -a_3 & a_4 & a_1 & -a_2 & -a_7 & -a_8 & a_5 & a_6 \\ -a_4 & -a_3 & a_2 & a_1 & -a_8 & a_7 & -a_6 & a_5 \\ -a_5 & a_6 & a_7 & a_8 & a_1 & -a_2 & -a_3 & -a_4 \\ -a_6 & -a_5 & a_8 & -a_7 & a_2 & a_1 & a_4 & -a_3 \\ -a_7 & -a_8 & -a_5 & a_6 & a_3 & -a_4 & a_1 & a_2 \\ -a_8 & a_7 & -a_6 & -a_5 & a_4 & a_3 & -a_2 & a_1 \\ \end{pmatrix} $$

  • $\begingroup$ Could you elaborate more on how this works (e.g., what exactly are the matrices $R_i$)? Since octonions are not associative, I don't think you can represent them by matrices as matrix multiplication is associative. $\endgroup$ – Māris Ozols Aug 19 '15 at 13:34
  • 1
    $\begingroup$ You may view the octonions $\bf{O}$ as a vector space $\mathbb{R}^8$ plus a bilinear map $\bf{O}\times\bf{O}\to\bf{O}$. You also have the standard basis $\{e_i\mid i=0,1,\ldots,7\}$ of $\bf{O}$. The bilinear map is not associative, but all you need is that each $e_i$ acts linearly from the left on $\bf{O}$ and that $\forall_{v\in\bf{O}}\{e_0v,e_1v,\ldots,e_7v\}$ is an orthonormal basis. Thus you may put $R_i=e_{i-1}$. $\endgroup$ – Adam Przeździecki Aug 19 '15 at 14:02
  • $\begingroup$ I checked that this indeed works, so I added an explicit expression to your answer. Do you have any good reference for why there is no solution other than $d \in \{1,2,4,8\}$? $\endgroup$ – Māris Ozols Aug 19 '15 at 14:46
  • $\begingroup$ @AdamPrzeździecki Cześć Adam ! wikipedia article en.wikipedia.org/wiki/Parallelizable_manifold says that it was proved in 1958 by Kervaire, Bott and Milnor that the only paralelizable spheres are $S^1$, $S^3$ and $S^7$. $\endgroup$ – Marek Mitros May 27 '16 at 15:23

This is an alternative way to see this question. By argument given by the current answer given by Ehud Meir we may assume that $R_1$ is the identity matrix. Then the hypotheses imply that for $i=2,\cdots,d$ the maps $v\rightarrow R_i v$ are $d-1$ independent vector fields on the unit sphere $S^{d-1}$. But it is known that the maximal number of independent vector fields on the unit sphere $S^{n-1}$ is $n-1$ if and only if $n=1, 2, 4 $ or $8$. This follows from a much stronger result due to J. F. Adams (Annals of Mathematics, 1962)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.