1
$\begingroup$

In my studies of convex cones I have recently met the following problem:

A non-empty set $ K\subseteq \mathbb{R}^d $ is called a convex cone if for all members $ x,y\in K $ and all non-negative real scalars $ 0 \leq\alpha,\beta \in \mathbb{R} $ we have $ \alpha x + \beta y \in K $. A convex cone is called pointed if we have $ K \cap -K = \{0\} $ and we denote by "ri" the relative interior. I am asked to prove that if $ K $ is a closed pointed cone, then there exists a pointed cone $ K' $ such that $ K \backslash \{0\} \subseteq ri(K') $ and of course $ K-K $ is the Minkowski difference meaning $ K-K = \{ k_1 -k_2 | k_1,k_2 \in K \} $. Now I have no idea how to proceed on this and have no idea how to formulate a solution for this, I can see intuitively why this mght hold maybe using hyperplane separation but not formally, could someone please assist on this? I thank all helpers.

$\endgroup$
3
+50
$\begingroup$

I assume it suffices to work with full dimensional $K\subseteq\mathbb{R}^d$, so that the dual $K^*$ is pointed. Since $K$ is pointed, $K^*$ is full dimensional, so pick any set of $d$ linearly independent vectors $v_1,v_2,\dots,v_d$ in $K^*$ that lie in the relative interior of $K^*$ to get a subcone $K'=\text{cone}\{v_1,v_2,\dots,v_d\}$ that is still full dimensional.

Then $K'\subsetneq K^*\implies (K')^*\supsetneq K^{**}=K$ since dualizing is inclusion reversing and $K'$ is pointed since $K^*$ was full dimensional. (I guess the full dimensional assumption is not strictly necessary, but then we need to be careful about the vectors we choose.)

$\endgroup$
  • 1
    $\begingroup$ Thanks for this nice answer, could you please just tell me why pointed implies full dimensionality of dual and vice versa? Is it a theorem of sorts? $\endgroup$ – kroner Oct 15 '16 at 19:19
  • 1
    $\begingroup$ and could you please tell me why the inclusion you wrote implies inclusion in relative interior? $\endgroup$ – kroner Oct 15 '16 at 19:28
  • 1
    $\begingroup$ Take a line $L$ contained in a cone $C$. Dualizing reverses inclusion, so $ L\subset C \iff C^*\subset L^*$ and $L^*$ is a hyperplane given by $\{x\in(\mathbb{R}^*)^d\mid \langle x,l\rangle=0\}$ where $l\in L$ is any nonzero point. $\endgroup$ – Tachyon Oct 15 '16 at 19:48
  • $\begingroup$ thanks could you please tell me how do we obtain from your answer specifically that $ K \subseteq ri((K^*)') $? Because I need a pointed cone whose relative interior contains ours $\endgroup$ – kroner Oct 15 '16 at 19:50
  • 1
    $\begingroup$ We know $K\subset (K')^*$. Thus the relative interior of $K$ is contained in the relative interior of $(K')^*$. So if there is a point $p$ in $K\backslash ri((K')^*)$, then it must be a boundary point of both, so in particular there is some $x\in\mathbb{R}^d$ such that $\langle x,p\rangle =0$. But such a point is not in the interior of $K^*$, contrary to our choice of generators for $K'$. $\endgroup$ – Tachyon Oct 15 '16 at 19:57

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.