Given a closed, convex, non-empty set $K\subseteq\mathbb{R}^n$ the support function $h_K:\mathbb{R}^n\to (-\infty,\infty]$ is defined as $$h_K(y) := \sup_{x\in K} \langle x,y\rangle$$

It is easy to see that $h_K$ is convex, $\mathbb{R}_{\geq 0}$-homogenuous and that all sublevel sets $\{y \mid h_K(y)\leq C\}$ are closed (in other words: $h_K$ is lower semicontinuous). I have several questions which I haven't been able to answer satisfactorily through google. Note that I'm interested in the case of general $K$ not just compact ones.

  1. Do these properties characterize support functions? There is an obvious way to define a convex set given such a function $h$ by setting $K:=\{x \mid \forall y: \langle x,y\rangle \leq h(y)\}$ and obviously $h_K\leq h$, but I cannot prove equality.
  2. Is there an description of $h_{K_1\cap K_2}$ in terms of $h_{K_1}$ and $h_{K_2}$ similarly to the descriptions $h_{K_1+K_2} = h_{K_1}+h_{K_2}$, $h_{conv(K_1\cup K_2)} = \max\{h_{K_1},h_{K_2}\}$ ? I think that something like $h_{K_1\cap K_2} = \sup\{h \leq \min\{h_{K_1},h_{K_2}\} \mid h \,\text{convex}\}$ might be true but I cannot prove it.

1. Yes, this is true. This can be derived from the involutive property of the Legendre-Fenchel transform: if $f:\mathbb R^n\to (-\infty,+\infty]$ is a closed convex function, then $f^{**}=f$ (Theorem 12.2 in Rockafellar's Convex Analysis). Recall that $$f^*(x) = \sup\{\langle x,y\rangle - f(y) : y\in\mathbb R^n\}$$ and a convex function is closed if its epigraph is.

With every convex closed set $K$ we can associate a closed convex function $\delta_K$ by letting $\delta_K(x)=0$ when $x\in K$ and $\delta_K(x)=+\infty$ otherwise. Observe that $\delta_K^*$ is exactly $h_K$.

Given $h$ as in your question, define $K= \{x : \langle x,y\rangle \le h(y) \ \forall y\}$. Observe that $h^*(x)=0$ when $x\in K$, because the supremum is attained by $y=0$. If $x\notin K$, then there is $y$ such that $\langle x,y\rangle > h(y)$. Considering large multiples of such $y$, we conclude that $h^*(x)=\infty$.

Thus, $h^* = \delta_K$. By the involutive property, $h=h^{**} = \delta_K^* = h_K$.

By the way, this result is Theorem 13.2 in Rockafellar's book.

2. Yes, this is correct. Another way to state this fact: the epigraph of $h_{K_1\cap K_2}$ is the convex hull of $\operatorname{epi} h_{K_1} \cup \operatorname{epi} h_{K_2}$. To see this, observe that the epigraph of $h_K$ is determined by its intersection with the horizontal plane $z=1$. This intersection is nothing but $K^\circ$, the polar of $K$. It remains to use the fact that $(K_1\cap K_2)^\circ = \operatorname{conv}(K_1^\circ \cup K_2^\circ)$. See also Corollary 16.5.1 in Rockafellar's book.

  • $\begingroup$ There is something off in the second case of your proof of statement 1: You implicitly claim that $\{y \mid h(y)<\infty\}$ is closed when you assert that $x$ has a neighbourhood contained in $\{h=\infty\}$. That is not true: Consider the parabola $K:=\{(x_1,x_2)\in\mathbb{R}^2 \mid x_2\geq x_1^2\}$. Then $h_K$ is finite exactly on the set $\{(y_1,y_2) \mid y_2<0\}\cup\{0\}$ which is not closed. $\endgroup$ – Johannes Hahn May 14 '14 at 0:06
  • $\begingroup$ Theorem 2.29 in Hörmander's book: "Notions of convexity" treats the intersection, the support function is the lower semi-continuous regularization of $$x \mapsto \inf_{x_1 + x_2 = x \in \mathbb{R}^n} \{h_1(x_1) +h_2(x_2)\}$$. $\endgroup$ – Jorge E. Cardona Jan 30 '17 at 15:09

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.