Let $f: \cal D_0 \to [0, \infty]$ be a convex function on a compact set $\cal{D}_0$ and let $\cal D \subseteq \cal D_0$. I think the following holds: $$ \inf_{x \in \cal{D}}\ f(x) = \inf_{x \in \overline{\cal{D}}}\ f(x) \quad \textrm{if} \quad \inf_{x \in \cal{D}}\ f(x) < \infty.$$ Is this correct and can anybody give me a hint on how to show this?
-
2In what sort of space is $\mathcal{D}$ residing? – ncmathsadist Jan 03 '14 at 19:00
-
1If $f(x) = e^x$ and ${\cal D} = \mathbb{R}$, then a minimum does not exist. – copper.hat Jan 03 '14 at 19:05
-
@ncmathsadist: Let us just say that there is some topology that allows to define the closure operator, do I need more? – Marco Tomamichel Jan 03 '14 at 19:15
-
@ncmathsadist: I now define $\cal D$ as a subset of a compact set - this is what I had in mind but did not state. – Marco Tomamichel Jan 03 '14 at 19:21
-
@copper.hat: I updated the question to exclude this case. $\cal D$ should be a subset of a compact set. – Marco Tomamichel Jan 03 '14 at 19:24
-
@MarcoTomamichel: The statement is true more generally with some assumptions on the topological structure of ${\cal D}$. – copper.hat Jan 03 '14 at 19:35
1 Answers
We must have $\inf_{x \in \overline{\cal D}} f(x) \le \inf_{x \in {\cal D}} f(x)$.
I am assuming that the topological structure is such that if $\bar{x} \in \overline{\cal D}$ and $x \in {\cal D}$, then $tx+(1-t) \bar{x} \in {\cal D}$ for all $t \in [0,1)$.
Let $\bar{x} \in \overline{\cal D}$ be such that $f(\bar{x}) < \infty$. Since $\inf_{x \in {\cal D}} f(x) < \infty$, there is some point $x_0 \in {\cal D}$ such that $f(x_0) < \infty$. Consider $\phi:[0,1] \to [0,\infty)$ defined by $\phi(t) = f(t x_0 + (1-t) \bar{x})$ (note that the range of $\phi$ is $[0,\infty)$, not $[0, \infty]$ since $f$ is convex). Since $\phi$ is convex, it is upper semi-continuous. Since $\phi(t) \ge \inf_{x \in {\cal D}} f(x)$ for all $t \in [0,1)$, we have $\phi(1) \ge \inf_{x \in {\cal D}} f(x)$.
Hence we have $f(\bar{x}) \ge \inf_{x \in {\cal D}} f(x)$ for all $\bar{x} \in \overline{\cal D}$.
- 172,524
-
Is this really an assumption on the topological structure or does it follow from the definition of closure? – Marco Tomamichel Jan 03 '14 at 19:36
-
I think $\phi$ is not continuos (on a closed domain) but rather upper semi-continuous, which should suffice here. – Marco Tomamichel Jan 03 '14 at 19:38
-
1@MarcoTomamichel: Without any further assumptions, the closure can be pretty 'wild'. It does depend on the topology, first countable would be sufficient and holds in many useful cases. – copper.hat Jan 03 '14 at 19:43
-
@MarcoTomamichel: My apologies, I need more caffeine. I will remove the above remark momentarily. – copper.hat Jan 03 '14 at 19:45
-
@MarcoTomamichel: I have updated my answer to state usc. rather than continuity. – copper.hat Jan 03 '14 at 19:47
-
My problem actually requires a weak-* topology, which is apparently not first countable. – Marco Tomamichel Jan 03 '14 at 19:51
-
1@MarcoTomamichel: However, the condition I needed above is true in this case. – copper.hat Jan 03 '14 at 20:01