6
$\begingroup$

Is there a continuous function $f:[0,+\infty) \to \operatorname{SO}(n)$ whose image is dense in $\operatorname{SO}(n)$ and that is well behaved in certain ways?

  • For each $\varepsilon>0$ it doesn't take longer than necessary, or not much, to come within distance $\varepsilon$ of every point.

  • It is not too hard to compute. One can write software for it without being a genius.

$\endgroup$
3
  • $\begingroup$ Does it need to be 1:1 or is it okay if it self-intersects? $\endgroup$ Commented Mar 26, 2021 at 18:59
  • 1
    $\begingroup$ Consider the Cayley transform, which relates the special orthogonal group to its Lie algebra. Easy to compute, but tough on distances. $\endgroup$
    – Ben McKay
    Commented Mar 26, 2021 at 19:30
  • $\begingroup$ @StevenStadnicki : I have no objection to self-intersection. (rhyme not intended) $\endgroup$ Commented Mar 27, 2021 at 4:52

2 Answers 2

6
$\begingroup$

[EDIT: Dan Asimov notified me that this construction is similar to a construction in his 1985 paper entitled "The Grand Tour: a Tool for Viewing Multidimensional Data". The construction in the 1985 paper is somewhat more elegant than this one, avoiding the use of the exponential map and the sine function.]

We'll describe such a function $f$ as the composition of three continuous maps:

  • $h : [0, +\infty) \rightarrow [-1,1]^{\binom{n}{2}}$;
  • $g : [-1,1]^{\binom{n}{2}} \rightarrow \mathcal{A}$;
  • $j : \mathcal{A} \rightarrow SO(n)$;

where $\mathcal{A}$ is the space of antisymmetric matrices with entries in $[-1, 1]$.

Each of these three maps, and thus their composition $f$, is not only continuous but is in fact Lipschitz-continuous (unlike a spacefilling curve).

In reverse order:

  • $j(A) := \exp((\pi \sqrt{n}) A)$, where $\exp$ is the matrix exponential;
  • $g^{-1}(A)$ is the vector $v$ obtained by 'flattening' the entries in the upper triangle of $A$ into a vector of $\binom{n}{2}$ elements, and $g$ is the inverse of the function $g^{-1}$ just described;
  • $h(t) := (\sin(c_1 t), \sin(c_2 t), \dots, \sin(c_{\binom{n}{2}} t))$, where $c_1, c_2, \dots, c_{\binom{n}{2}}$ are a set of $\binom{n}{2}$ irrationals that are linearly independent over $\mathbb{Q}$.

The image of $h$ is dense in the hypercube as mentioned here:

https://en.wikipedia.org/wiki/Linear_flow_on_the_torus

and, because $g$ is a homeomorphism, it follows that $g(h(t))$ is dense in $\mathcal{A}$. As such, it remains to show that $j$ is surjective onto $SO(n)$; that would establish that $f(t) = j(g(h(t)))$ is dense in $SO(n)$.


Claim: Every matrix $R \in SO(n)$ is expressible as the matrix exponential of an antisymmetric matrix $A$ with Hilbert-Schmidt norm $\operatorname{tr}(A^T A) \leq \pi^2 n$.

Proof: By an orthogonal change of basis, we can assume that the matrix $R$ is block-diagonal, consisting of $1 \times 1$ blocks of the form:

$$ \begin{pmatrix} 1 \end{pmatrix} $$

and $2 \times 2$ blocks of the form:

$$ \begin{pmatrix} \cos{\theta} & \sin{\theta} \\ -\sin{\theta} & \cos{\theta} \end{pmatrix} $$

where $\theta \in [-\pi, \pi]$. As such, $R$ is the matrix exponential of a block-diagonal antisymmetric matrix with blocks of the form:

$$ \begin{pmatrix} 0 \end{pmatrix} $$

and:

$$ \begin{pmatrix} 0 & \theta \\ -\theta & 0 \end{pmatrix} $$

The result follows.

Corollary: Every orthogonal matrix is the matrix exponential of an antisymmetric matrix with entries in $[-\pi \sqrt{n}, \pi \sqrt{n}]$. As such, $j$ is indeed surjective onto $SO(n)$.

$\endgroup$
7
  • $\begingroup$ And then there's the question of how $c_1,\ldots,c_{\binom{n-1}2}$ should be chosen so as to satisfy the desideratum about $\text{“each }\varepsilon>0\text{.”}$ Specifically, I would probably want to reduce as fast as possible with respect to $t$ the distance between $\{f(s) : 0\le s\le t\}$ and the point in $\operatorname{SO}(n)$ that is farthest from that set. $\qquad$ $\endgroup$ Commented Mar 27, 2021 at 22:16
  • 1
    $\begingroup$ @MichaelHardy $\binom{n}{2} = \frac{1}{2}n(n-1)$ is already the number of elements strictly above the (all-zero) diagonal. $\endgroup$ Commented Mar 28, 2021 at 0:35
  • $\begingroup$ ok, So this gives an easy-to-compute curve that is dense in $\operatorname{SO}(n).$ And I commented above about choosing the $c_i\text{s}.$ But here I wonder whether this approach is best from the point of view of the $\text{“each }\varepsilon>0\text{”}$ part. At this point I'm contemplating posting a refinement of this question that is principally devoted to that. $\endgroup$ Commented Mar 28, 2021 at 1:26
  • $\begingroup$ Am I right in thinking that the reason you chose $[-1,1]$ rather than some larger interval is that that is just how big it had to be in order to give you all of $\operatorname{SO}(n) \text{?} \qquad$ $\endgroup$ Commented Mar 28, 2021 at 3:59
  • $\begingroup$ @MichaelHardy In terms of choosing $c_1, \dots, c_{\binom{n}{2}}$ to give the good $\varepsilon$-approximability, that's effectively a separate question unrelated to $SO(n)$. For two values, $1, \phi$ is known to be optimal, but I'm unsure how this generalises to higher dimensions. $\endgroup$ Commented Mar 28, 2021 at 12:09
4
$\begingroup$

Take a parameterization of $SO(n)$ that has domain a hypercube of dimension $\binom{n}{2}$, e.g. https://math.stackexchange.com/questions/965451/. Now take the image of a Brownian motion in the hypercube with reflecting boundary conditions. If you want determinism, pick your favorite family of closed curves that has a space-filling limit and just concatenate them.

$\endgroup$
6
  • $\begingroup$ I have qualms about space-filling curves in this problem. I said for any $\varepsilon>0,$ we should not take much longer than needed to come within $\varepsilon$ of every point. Can that be done with space-filling curves? $\endgroup$ Commented Mar 27, 2021 at 19:37
  • $\begingroup$ If randomness is allowed, one can pick an independent sequence of uniformly distributed orthogonal matrix and then just follow geodesics from each to the next. And maybe that's not bad for what I have in mind. I have some lesser qualms about that. $\endgroup$ Commented Mar 27, 2021 at 19:38
  • 1
    $\begingroup$ You take successive approximations to the actual space filling curve, one after the other. So you get decent convergence. FWIW my initial thoughts were more in line with Adam’s but I didn’t think of the slick idea of using sin on top of a 1D foliation of the torus, which fixes what would otherwise be a nasty continuity issue. $\endgroup$ Commented Mar 28, 2021 at 1:52
  • $\begingroup$ Which families of space-filling curves in $\operatorname{SO}(n)$ are understood? Some with nice properties? $\endgroup$ Commented Mar 28, 2021 at 3:56
  • $\begingroup$ You just need a family in the hypercube, and there are several of these, eg successive approximations to the Hilbert curve $\endgroup$ Commented Mar 28, 2021 at 12:17

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .