Let $X$ be a real normed space equipped with $\lVert\cdot\rVert$ and $M$ be a convex set in $X$, such that $B(0,r) \subset M \subset B(0,R)$ for $r>0$. Here $B(a, t)$ stands for closed ball with center $a$ and radius $t$.

Further, let $\pi\colon S \rightarrow \partial M$ denote the central projection of the unit sphere onto boundary of $M$. We can show that the best Lipschitz constant $L(\pi)$ of the projection is at most $\frac{R(R+r)}{r}$. If $\mu$ is Minkowski functional of $M$, then $\pi(x) = \frac{x}{\mu(x)}$. Hence, for all $x,y \in S$ \begin{align*} \frac{\lVert \pi(x)-\pi(y) \rVert}{\lVert x-y \rVert} &= \frac{\lVert \mu(y)\,x-\mu(x)\,y \rVert}{\lVert x-y \rVert \,\mu(x)\mu(y)} \leqslant \frac{\lvert \mu(y)-\mu(x) \rvert \,\lVert x \rVert + \mu(x)\,\lVert x-y \rVert}{\lVert x-y \rVert \,\mu(x)\mu(y)} = \\ &= \frac{1}{\mu(x)\mu(y)} \frac{\lvert \mu(y)-\mu(x) \rvert}{\lVert x-y \rVert} + \frac{1}{\mu(y)} \leqslant \bigg( \frac{R}{r} \bigg)^2 \frac{1}{r} + R = \frac{R(R+r)}{r}, \end{align*} as for every $z \in X$ $$ \frac{\lVert z \rVert}{R} \leqslant \mu(z) \leqslant \frac{\lVert z \rVert}{r},$$ in particular, $\mu$ is $1/r$-Lipschitz.

So, we're getting closer to the point. Is this an exact bound of $L(\pi)$? No, at least in the case of X being inner product space. The exact bound then is $\frac{R^2}{r}$, it's proved in the paper "A note on starshaped sets" by Siniša Vrećica.

In fact, when dealing with $\frac{\lVert \pi(x)-\pi(y) \rVert}{\lVert x-y \rVert}$ we work inside $Y=\mathrm{span} \{x,y\}$ as $\pi(x),\pi(y) \in Y$. That means we can consider only two-dimensional normed spaces $X$. The proof of euclidean (planar) case of the estimate pretty uses trigonometry. (Siniša's proof looked too technical to me, so I proved it myself using just sines' theorem and elementary facts, like monotonicity of sine.)

It's also sufficient to consider convex bodies of the form $M = \mathrm{conv} (B(0,r) \cup \{z\})$, as that sets are the "worst" for $L(\pi)$. Also, WLOG, we can take $r=1$.

For example, due to my humble computations, for $X = \ell_\infty^2$ and $$M_1 = \mathrm{conv} (B(0,1) \cup \{(0,R)\}), \quad M_2 = \mathrm{conv} (B(0,1) \cup \{(R,R)\}), \quad R\geqslant 2$$ we have $L(\pi_1) = R(R-1)$ and $L(\pi_2) = R(R+1)/2$.

Now, I have no idea how to deal with non-euclidean case in general. It seems to me that the Lipschitz constant must be less than for euclidean case. Anyway, how can one obtain some estimations better than $\frac{R(R+r)}{r}$ (say, norm given by radial function of unit ball)? Or there is example when we can't do better?

  • 1
    $\begingroup$ You get actually estimate $R(R+r)/r$. (And it has to be homogeneous a priori.) $\endgroup$ – Fedor Petrov Feb 28 '16 at 9:05
  • $\begingroup$ @FedorPetrov, oh, right, my mistake. $\endgroup$ – quartermind Feb 28 '16 at 10:30
  • $\begingroup$ In general this is sharp estimate for all $r,R$. I have to draw a picture, it requires some time for me. $\endgroup$ – Fedor Petrov Feb 28 '16 at 17:50
  • $\begingroup$ "Monotonicity of sine"? All this time I assumed it was periodic... $\endgroup$ – Nik Weaver Feb 28 '16 at 18:23
  • $\begingroup$ On the Lipschitz constant of $\mu$, see jstor.org/stable/2316022 $\endgroup$ – Yoav Kallus Feb 28 '16 at 19:03

$r=1$, $R=OC/OB$. Unit ball is $DEBD'$ and symmetric part, $M$ is $DCD'$ and symmetric part, $x=B$, $y=G$, $\pi(x)=C$, $\pi(y)=H$, ray $OGH$ is very close to $OBC$.

We have $$\|C-H\|:\|B-G\|=\frac{CH}{OI}:\frac{BG}{OD}=\frac{CH\cdot BE}{BG\cdot CE}\cdot \frac{CE}{OI}\cdot \frac{OD}{BE}=\frac{OH}{OG}\cdot \frac{CE}{OI}\cdot\frac{OC}{CB}$$ (Menelaus theorem for triangle $EHG$ and line $CBO$ is used.) $OH/OG$ tends to $OC/OB$ when the ray $OGH$ tends to $OCB$. Next, $CE/OI=(CE/ED)\cdot (ED/OI)=(CB/BO)\cdot (1+OB/OC)$. Totally we get in the limit $(OC/OB)^2+(OC/OB)=R^2+R$.

We have used relation $ED/OI=1+OB/OC$, which follows, for example, from parallelogram $EBD'M$: $ED/OI=DM/OD'=(2DO-EB)/DO=2-EB/DO=2-CB/OC=1+OB/OC$.

enter image description here

  • $\begingroup$ This works for me, thanks! Your example of norm is very specific, I mean, I was likely to consider situations when $R \gg r$. We may ask, what if we fix a norm first (and $r=1$), and then release $R$ to infinity? But that's may be not so interesting, because our initial estimate can be at least asymptotically exact in non-euclidean case. $\endgroup$ – quartermind Feb 28 '16 at 19:27
  • $\begingroup$ It works for all $R>1$, I think. $\endgroup$ – Fedor Petrov Feb 28 '16 at 19:28

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.