= Support-function subgradients as exposed faces
{title2=$\partial\sigma_C(x)=\{p\in C:\langle p,x\rangle=\sigma_C(x)\}$}
Let $C$ be a nonempty closed <convex set> in finite dimensions. At any $x$ where its <support function> is finite,
$$
\partial\sigma_C(x)=\{p\in C:\langle p,x\rangle=\sigma_C(x)\}.
$$
Indeed $\sigma_C^*=\delta_C$ for the <indicator functional of a constraint set>. The equality case of the <Fenchel–Young inequality> says $p\in\partial\sigma_C(x)$ exactly when $\sigma_C(x)+\delta_C(p)=\langle p,x\rangle$, which is the displayed maximizing condition.
If $C$ is compact, the maximum is attained and this set is nonempty. For an unbounded closed convex set, a finite supremum need not be attained. Take $C=\{(u,v):v\ge e^u\}$, the <epigraph> of an <exponential function>, and $x=(0,-1)$. Then $\sigma_C(x)=0$, approached as $u\to-\infty$, but no point of $C$ has $v=0$. The subdifferential at that direction is therefore empty. The support function must be considered at its finite-valued directions, rather than assumed finite everywhere for every closed convex set.
Back to article page