Let $C\subseteq \mathbb R^n$ be non-empty, convex and compact. For $v\in S^{n-1}$, let $H_v$ be the supporting hyperplane in the direction of $v$ (i.e., $H_v$ is the boundary of the smallest closed half-space with outward normal $v$ that contains $C$). Let $U\subseteq S^{n-1}$ be the set of directions $v$ such that $H_v$ meets $C$ at exactly one point.

Main question: Does $S^{n-1}\backslash U$ have measure zero?

If not, then I have a second question: Is $U$ dense in $S^{n-1}$?

For $n=2$, it's easy to see that $S^{n-1}\backslash U$ is countable (otherwise there are uncountably many nondegenerate line segments in the boundary of $C$, and hence $C$ has infinite perimeter). But a cylinder shows that in general $S^{n-1}\backslash U$ need not be countable.

  • $\begingroup$ Does a general version of your argument show that the boundary of $C$ would have positive (Lebesgue) measure if $S^{n-1}\setminus U$ doesn't have measure zero? $\endgroup$ Oct 9, 2022 at 18:01
  • 2
    $\begingroup$ doesn't the result follow from the fact that the support function is lipschitz so it is differentiable almost everywhere? $\endgroup$
    – alesia
    Oct 9, 2022 at 21:26
  • 1
    $\begingroup$ @alesia: Yup, that looks like the answer to my question. Why don't you write it up as an official answer? $\endgroup$ Oct 10, 2022 at 0:17
  • 1
    $\begingroup$ @alesia I do not see how differentiability would imply the result. A convex function $f(x,y)=x^2$ is differentiable everywhere yet, at every point, the supporting hyperplane meets the graph along a line. Am I missing something? $\endgroup$ Oct 10, 2022 at 2:35
  • $\begingroup$ @PiotrHajlasz: alesia may be thinkng of something like Theorem 1.1 here: oyama.e.u-tokyo.ac.jp/notes/diffSuppFunc01.pdf The graph of your f(x,y) is not compact. If you restrict to a compact domain, then there are two ways of relevant ways of counting: by counting the normals and by counting the contact points. If you count contact points, you are right. But in my question, I am counting the normals. $\endgroup$ Oct 10, 2022 at 14:24

2 Answers 2


The support function of $C$, restricted to the unit sphere, is differentiable exactly at directions such that (the relevant) hyperplane normal to that direction has a single contact point with $C$.

Because $C$ is bounded, its support function is Lipschitz. Rademacher's theorem then says it is differentiable almost everywhere, giving a positive answer to the question.

  • $\begingroup$ The assumption that $C$ is bounded is not needed. Even if $C$ is unbounded, its boundary is locally Lipschitz and therefore by Rademacher's theorem, is locally differentiable almost everywhere. $\endgroup$
    – Deane Yang
    Oct 10, 2022 at 19:24
  • 1
    $\begingroup$ If $C$ is not bounded, $H_v$ might not be defined. $\endgroup$ Oct 10, 2022 at 19:46
  • $\begingroup$ Ah, yes. I see. $\endgroup$
    – Deane Yang
    Oct 10, 2022 at 23:09
  • $\begingroup$ @DeaneYang I posted an answer which gives a much stronger result. $\endgroup$ Oct 22, 2022 at 2:58

This is a consequence of the following result from

enter image description here

Ewald, G.; Larman, D. G.; Rogers, C. A., The directions to the line segments and of the r-dimensional balls on the boundary of a convex body in Euclidean space, Mathematika, Lond. 17, 1-20 (1970). ZBL0199.57002.

The above statement is copied from the monograph.

Schneider, Rolf, Convex bodies: the Brunn-Minkowski theory, Encyclopedia of Mathematics and its Applications 151. Cambridge: Cambridge University Press (ISBN 978-1-107-60101-7/hbk; 978-1-139-00385-8/ebook). xxii, 736 p. (2014). ZBL1287.52001.


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.