13
$\begingroup$

In ${\bf M}_n(\mathbb R)$, let us consider the usual operator norm $$\|A\|=\sup\frac{\|Ax\|}{\|x\|},$$ where $\|x\|$ is the Euclidian norm.

The closed unit ball $B$ is the set of contractions (in the terminology used by operator theorists). It is a convex compact subset of ${\bf M}_n(\mathbb R)$. By Krein-Milman (finite dimensional case), it is the convex hull of its subset ${\rm ext}(B)$ of extremal points. It turns out that ${\rm ext}(B)$ is the orthogonal group ${\bf O}_n(\mathbb R)$.

Now, remember that ${\bf O}_n(\mathbb R)$ has two connected components, a positive one ${\bf SO}_n(\mathbb R)$ and a negative one ${\bf O}_n^-(\mathbb R)$.

What is the convex hull of ${\bf SO}_n(\mathbb R)$ ?

Clearly, it is a compact convex subset, included in $B$. It is a strict subset of $B$, because it does not meet ${\bf O}_n^-(\mathbb R)$. The title refers to the "positive part" of $B$, but this could be inappropriate, in the sense that it could meet the convex hull of ${\bf O}_n^-(\mathbb R)$ non-trivially.

Remark also that this convex hull is invariant under multiplication at right or left by an element of ${\bf SO}_n(\mathbb R)$. Therefore it would be enough to decide which diagonal matrices ${\rm diag}(a_1,\ldots,a_n)$ with $|a_1|\le a_2\le\cdots\le a_n$ it contains.

When $n=2$, ${\bf SO}_n(\mathbb R)$ is a circle and its convex hull is a disk, obviously a much smaller set (even from the dimensional point of view) than $B$.

$\endgroup$
6
  • $\begingroup$ If $B$ is open, then $B$ has no extremal points, and so is not the convex hull of the set of its extremal points. $\hspace{.2 in}$ If $B$ is closed, then $B$ has members that are not contractions. $\;\;$ $\endgroup$
    – user5810
    Feb 29, 2012 at 8:44
  • 5
    $\begingroup$ Ricky, it is customary in many places/books/texts/whatever in analysis to use the word "contraction" to mean "distance non-increasing". This is how almost all practising operator theorists and most functional analysts I've met use the word, for instance $\endgroup$
    – Yemon Choi
    Feb 29, 2012 at 8:44
  • 4
    $\begingroup$ The answer for $n=3$ is given in $\S 4.1$ of arxiv.org/abs/0911.5436. In $\S 4.4$ of ibid, there's a discussion of some properties of the convex hull of $SO(n)$ for larger $n$. $\endgroup$ Feb 29, 2012 at 15:31
  • $\begingroup$ Merci, Laurent ! $\endgroup$ Feb 29, 2012 at 15:47
  • 1
    $\begingroup$ A few basic remarks: (a) the set of extremal points of the convex hull of $SO_n$ is exactly $SO_n$ (idem for $O^-_n$); (b) for $n$ even these sets are symmetric; (c) for all $n\ge 2$ the convex hull of $SO_n$ and $O^-_n$ have 0 in their intersection, as we see looking just at the convex hull of diagonal $\pm 1$-matrices (d) for $n\ge 3$ the convex hull of $SO_n$ and $O^-_n$ have nonempty interior (hence their intersection is a neighborhood of 0). (This is because it contains 0 and using [I skip details] that the representation in $\mathbf{R}^n$ is absolutely irreducible for $n\ge 3$.) $\endgroup$
    – YCor
    Nov 10, 2016 at 5:06

1 Answer 1

11
$\begingroup$

I'm a bit late in answering this. But in case there is still interest, please have a look at:

Saunderson, Parrilo, Willsky. Semidefinite descriptions of the convex hull of rotation matrices, SIAM J. Optimization, 25(3), 1314-1343, 2015.

This paper provides explicit spectrahedral representations for conv $SO(n)$ (Theorem 1.3).

$\endgroup$
1
  • $\begingroup$ Thanks a lot ! Apparently, the authors ignored my MO question. $\endgroup$ Nov 9, 2016 at 17:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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