2
$\begingroup$

Suppose that $(X,d)$ is a Polish metric space and $A$ is a set of continuous bounded functions $f:X\to \mathbb{R}$. Suppose that $\mu:X\to[0,1]$ is a Borel probability measure. Define

$$\sup A:X\to (-\infty,+\infty],\quad \sup A(x):=\sup\{f(x)\colon f\in A\}, $$ On the other hand, we have the essential suppremum, that is, a Borel measurable function $$\text{ess.sup}_\mu A:X\to (-\infty,+\infty], $$ which is unique up to $\mu$-null sets an such that $$ f\le \text{ess.sup}_\mu A\quad\mu\text{-a.s. for all }f\in A $$ and if another Borel measurable function $Y$ satisfies the condition above, then $\text{ess.sup}_\mu A\le Y$ $\mu$-a.s.

My question is: Is it true that $$\sup A=\text{ess.sup}_\mu A$$ $\mu$-almost surely?

$\endgroup$
  • $\begingroup$ Maybe I'm being naive, but is there any specific reason to suspect that $\sup A$ is even (essentially) measurable? $\endgroup$ – Jochen Glueck Oct 15 '19 at 20:31
  • $\begingroup$ $\sup A$ is even lower semicontinuous as a supremum of continuous functions, in particular Borel-measurable. $\endgroup$ – Dieter Kadelka Oct 15 '19 at 21:10
  • $\begingroup$ @DieterKadelka: Good point, thank you. $\endgroup$ – Jochen Glueck Oct 15 '19 at 21:55
1
$\begingroup$

The claim now is correct.

For the following proof we only need that $X$ is separable metric and $A$ a family of lower semicontinuous (l.s.c.) functions on $X$. (Of course these assumptions may be further weakened.) To simplify notation we may assume that $0 \leq f \leq 1$ for each $f \in A$. Let $g$ be an arbitrary representant of ess sup$_{f \in A} f$ and $h := \sup_{f \in A} f$, which is l.s.c. again, hence Borel-measurable.

By V.I.Bogachev, Measure Theory I (2007), 4.7.1 there is an at most countable subset $\{f_n\} \subset A$ such that $\sup_n f_n = g$ $\mu$-a.e., hence $g \leq h$ $\mu$-a.e. If not $h \leq g$ $\mu$-a.e., then there is $B \in \cal{B}(X)$ with $\mu(B) > 0$ and $g(x) < h(x)$ for $x \in B$. But then there are rational $0 \leq p < q \leq 1$ with $\mu(g \leq p, h \geq q) > 0$. Let $C := \{x \in X \colon g(x) \leq p, h(x) \geq q\}$. Since $X$ is separable metric there is $x \in C$ with $\mu(U \cap C) > 0$ for any open neighbourhood of $x$. Let $f \in A$ with $f(x) > \frac{p+q}{2}$, then $U := \{y \in X \colon f(y) > \frac{p+q}{2}\}$ is an open neighbourhood of $x$, hence $\mu(U \cap C) > 0$. But this contradicts the assumption that $g(y) \leq p$ for $\mu$-a.e. $y \in C$.

$\endgroup$
  • $\begingroup$ Very nice proof $\endgroup$ – Littlefield Oct 18 '19 at 19:58
1
$\begingroup$

In general no. Assume that $X$ is a set with atomless nontrivial probability measure on $\cal{P}(X)$, endowed with the discrete metric, and $A := \{1_{\{x\}} \colon x \in X\}$. Then ess sup $_\mu A \equiv 0$ and $\sup A \equiv 1$. The answer may be true with additional assumptions. (For the existence of such $X$ see f.i. Jech (1978), Set Theory.)

$\endgroup$
  • $\begingroup$ But, in general, $1_{\{x\}}$ is not continuous. Notice that I assume $A$ a set of continuous bounded functions. It is right that with the trivial metric $1_{\{x\}}$ is continuous. I wanted to include that the metric space is Polish. I have edit the question with a Polish metric space. $\endgroup$ – Littlefield Oct 15 '19 at 16:13
  • $\begingroup$ At least you need another assumption: $Y := supp~ \mu = X$. Otherwise let $f$ be any continuous function on $X$ with $f|Y \equiv 0$, $f \geq 0$ and not $f \equiv 0$. Take $A = \{f\}$. Then again ess sup$_\mu A = 0$ but $\sup A \not\equiv 0$. $\endgroup$ – Dieter Kadelka Oct 15 '19 at 16:43
  • 1
    $\begingroup$ I think that this is not a problem. Note that I want $\mu$-a.s. equality. So in your counterexample you have a.s. equality. $\endgroup$ – Littlefield Oct 15 '19 at 19:28
  • $\begingroup$ That's right. Since $X$ is polish $\mu$ is regular and you can assume that $X$ is compact metric and supp $\mu = X$. Maybe this simplifies the proof. Further you may assume that $0 \leq f \leq 1$ for all $f \in A$. $\endgroup$ – Dieter Kadelka Oct 15 '19 at 21:16

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.