4
$\begingroup$

Let $K,T\subset\mathbb{R}^{n}$ be convex bodies (i.e. nonempty, compact and convex) containing the origin, and let $\lambda\in(0,1)$. Fix $\langle \cdot ,\cdot \rangle$ to be the standard dot product in $\mathbb{R}^n$ (though probably here any dot product would do). Let us denote by $K^\circ$ the dual to $K$ with respect to the fixed dot product, and recall that $K^\circ=\{y\in\mathbb{R}^n : \langle x, y\rangle \leq 1, \forall x\in K\}$.

Is it true that $$(\lambda K + (1-\lambda) T)^\circ\subset \lambda K^\circ + (1-\lambda)T^\circ?$$

So far I managed to show the equivalence of this question to the following: Denote by $h_K(x)=\sup_{y\in K} \langle x, y \rangle$ the support function of $K$, and by $g_K(x)=\inf \{ r>0: x\in rK\}$ the gauge function of $K$. By some very basic manipulations one can see that the original question is equivalent to asking whether or not it's true that $$h_{(\lambda K + (1-\lambda)T)^\circ}\leq \lambda h_{K^\circ} + (1-\lambda)h_{T^\circ}.$$ This is already quite believable. Using the fact that for convex bodies (as above) it holds that $h_{K^\circ}=g_K$, and that for such bodies $g_K$ is nearly a norm (i.e. it's sublinear, positively homogeneous and non-negative), then one obtains $$\Vert\cdot \Vert_{\lambda K +(1-\lambda)T} \leq \lambda\Vert\cdot\Vert_K + (1-\lambda)\Vert\cdot\Vert_T.$$ I don't know if the last inequality is true, but as far as I understand convexity, it has every right to be true. Is there a functional analysis approach I might be missing?

Lastly, if the above is not true in general I would be satisfied in answering this for $\lambda=\frac{1}{2}$.

$\endgroup$
3
$\begingroup$

Does this work?

First prove it when $n=1$. Now reduce the general case to this one as follows:

If $L$ is a line through the origin in $\mathbb R^n$ and $K$ is convex in $\mathbb R^n$ then the intersection of the dual of $K$ with $L$ is the dual (with respect to $L$) of the image of $K$ in $L$ under orthogonal projection.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I have to admit this answer quite surprised me, since it's so lean yet does not go through norms. Thanks! (P. S. hope I'm not breaking any rules in this none-helpful comment.) $\endgroup$ – Donjim Feb 28 '17 at 13:12
1
$\begingroup$

\begin{align*} h_{((1-\lambda)K + \lambda T)^\circ}(x) &= \sup_\xi \frac{\langle\xi,x\rangle}{h_{(1-\lambda)K + \lambda T}(\xi)}\\ &= \sup_\xi \frac{\langle\xi,x\rangle}{(1-\lambda)h_K(\xi) + \lambda h_T(\xi)}\\ &\le \sup_\xi \left[(1-\lambda)\frac{\langle\xi,x\rangle}{h_K(\xi)} + \lambda\frac{\xi,x\rangle}{h_T(\xi)}\right]\\ &\le (1-\lambda)h_{K^\circ}(x) + \lambda h_{T^\circ}(x)\\ &= h_{(1-\lambda)K^\circ+\lambda T^\circ}(x) \end{align*} Therefore, $$ ((1-\lambda)K + \lambda T)^{\circ} \subset (1-\lambda)K^\circ + \lambda T^\circ. $$ Equality holds if and only if $T = tK$ for some $t > 0$.

| cite | improve this answer | |
$\endgroup$

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.