3
$\begingroup$

For any matrix $A$ in either $SO(2)$ or $SO(3)$, the expression $\frac{1}{\sqrt{2}}||\log A||_F$ gives the angle of the rotation represented by $A$, where $\log A$ is a matrix logarithm of $A$ and the norm is the Frobenius norm. From this result it follows that, for any given (skew-symmetric) matrix $X$ satisfying $\exp X = A$, every matrix $X'$ in the set $$ L_X= \left\{tX \;\middle|\; t=\frac{||X||_F+k2\pi\sqrt{2}}{||X||_F} \;,\; k\in\mathbb{Z}\right\} $$ also satisfies $\exp X' = A$ .

I know from numerical tests that this is not necessarily true for other $SO(n)$ groups. So my question is: given a logarithm $X$ of a matrix $A\in SO(n)$, for arbitrary $n>3$, are there any known parameterisations of a set of logarithms of $A$ that includes (but is not limited to) $X$?

I have two objectives here:

  1. I would ideally like to know how to parameterise all logarithms of $A$, or at least mathematically interesting subsets of them, out of a desire to improve my understanding of special orthogonal groups and algebras.
  2. I am writing tests for functions that compute matrix exponentials and principal matrix logarithms. My test data consists of orthogonal matrices and their principal matrix logarithms. For each $(A,X)$ test pair, I am generating lots of additional test pairs of the form $(A^k,kX)$. For large $k$, $kX$ is not a principal logarithm of $A^k$, so I cannot compare $kX$ to the computed logarithm of $A^k$. I could compare $\log A^k$ to $\log \exp kX$, but the equality of these expressions doesn't imply that $\log A^k$ was computed correctly (e.g. if the $\log$ function erroneously mapped everything to a constant matrix $C$, these two expressions would be equal). So I would like to know how to map $kX$ to the principal logarithm without using $\log$ or $\exp$ .
$\endgroup$
4
$\begingroup$

After a bit more reading, I've found an answer to my own question:

The principal logarithm of a matrix $A\in SO(n)$ is the logarithm whose eigenvalues have imaginary components in the interval $(-i\pi,i\pi)$. As the Lie algebra $so(n)$ is the set of real skew-symmetric (and hence normal) $n\times n$ matrices, any logarithm* $X\in so(n)$ of $A$ has a complex eigendecomposition: $$ \begin{align} X&=V\Lambda V^* \enspace,\\ \Lambda&=\mathrm{diag}(\lambda_1,-\lambda_1, \lambda_2, -\lambda_2, \ldots)\enspace, \end{align} $$ with an extra zero eigenvalue if $n$ is odd, where the non-zero eigenvalues are purely imaginary. Since $V$ is unitary, we have $$ A=e^X = e^{V\Lambda V^*} = Ve^\Lambda V^* \enspace, $$ where $$ e^\Lambda=\mathrm{diag}\left(e^{\lambda_1},e^{-\lambda_1},e^{\lambda_2},\ldots\right)\enspace. $$ So if $\lambda_j$ and $-\lambda_j$ are eigenvalues of $X$, you can construct a set $L_j$ of logarithms of $A$ as follows: $$ \begin{align} L_j &= \left\{V\Lambda^{(j)}_kV^* \;\middle|\; k\in\mathbb{Z} \right\} \\ \Lambda^{(j)}_k &= \mathrm{diag}\left(\ldots,-\lambda_{j-1},\;\lambda_j+2ik\pi, \;-(\lambda_j+2ik\pi), \;\lambda_{j+1},\ldots\right) \end{align} $$ By adjusting all pairs of positive/negative imaginary eigenvalues like this, you can compute the principal logarithm that corresponds to $X$.

* Note: A matrix $A\in SO(n)$ has real logarithms that are not in $so(n)$ (e.g. see Is a real logarithm of a special orthogonal matrix necessarily skew-symmetric?). A real logarithm $X$ is in $so(n)$ if and only if $e^{tX}\in SO(n)$ for all real numbers $t$.

$\endgroup$

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.