My main question is

Q1. Let $B$ be a Borel-measurable subset of $\mathbb R$. Is there a sequence of open sets $U_n$ independent of any measure such that for all Borel probability measures $\mu$ on $\mathbb R$, $\mu(B \Delta U_n) \to 0$?

Any open set is trivially a measure-independent limit of open sets.

Any closed set is an intersection of a nested sequence of open sets and hence is a measure-independent limit.

I wonder if Q1 can be answered negatively by showing that $\mathbb Q$ is a counter example.

Q1a. Is $\mathbb Q$ a counter example?

If $B$ is a measure-independent limit of $U_n$ then $1_B$ is a pointwise-limit of the sequence $1_{U_n}$ because of Dirac delta measures. Maybe this can be used to show that $\mathbb Q$ is a counter example, but I don't know how.

Motivation: last paragraph of the question in https://mathoverflow.net/questions/167823/the-borel-sigma-algebra-of-the-set-of-probability-measures?rq=1

  • 1
    $\begingroup$ I added the tag "descriptive-set-theory" to your question, because (see my answer) your question is linked to the Borel hierarchy. Hopefully, someone more qualified than I can produce a concrete example of a Borel set which is not a $G_{\delta, \sigma}$, which then produces a concrete counterexample. $\endgroup$ – PhoemueX Dec 20 '14 at 19:25

I can sharpen PhoemueX's very nice answer a little bit. In particular, yes, $\mathbb{Q}$ is a counterexample.

You are quite right that, if a set $B$ is to be the measure-independent limit of a sequence of sets $A_n$, we must have $1_{A_n} \to 1_B$ pointwise. In particular, as PhoemueX says, we must have $B = \bigcup_{k \ge 1} \bigcap_{n \ge k} A_n$. So if $A_n = U_n$ are open then $B$ must be $G_{\delta,\sigma}$.

But we can say more: it must also be that $B = \bigcap_{k \ge 1} \bigcup_{n \ge 1} A_n$. In fact, it is a nice little exercise to show that for any sequence of sets $A_n$, we have $$\limsup_{n \to \infty} 1_{A_n} = 1_{\limsup A_n}, \qquad \liminf_{n \to \infty} 1_{A_n} = 1_{\liminf A_n}$$ where $$\limsup A_n := \bigcap_{k \ge 1} \bigcup_{n \ge k} A_n, \qquad \liminf A_n := \bigcup_{k \ge 1} \bigcap_{n \ge k} A_n.$$ So in particular if $1_B = \lim_{n \to \infty} 1_{A_n}$ pointwise, then we have $1_B = \limsup 1_{A_n} = \liminf 1_{A_n}$ and hence $B = \limsup A_n = \liminf A_n$.

And when $A_n = U_n$ are open, then $\limsup U_n$ is actually $G_\delta$. So in order for a set to be a measure-independent limit of open sets, it must be both $G_\delta$ and $G_{\delta,\sigma}$.

However, $\mathbb{Q}$ is not $G_\delta$. Since $\mathbb{Q}$ is dense, if it were $G_\delta$ it would be a dense $G_\delta$, or in other words comeager. But since $\mathbb{Q}$ is countable it is also meager. This contradicts the Baire category theorem: $\mathbb{R}$ is not meager, so no subset of $\mathbb{R}$ can be simultaneously meager and comeager.

  • $\begingroup$ Nice, I should have thought of that myself! +1 $\endgroup$ – PhoemueX Dec 20 '14 at 21:00

This is not a complete answer, but maybe helpful nevertheless:

As you noted, it is a necessary condition that $\chi_{U_n}$ (the characteristic function/indicator function of $U_n$) fulfils $\chi_{U_n} \to \chi_B$ pointwise.

But conversely, if this is the case, then

$$ \mu(B \Delta U_n) = \int |\chi_B - \chi_{U_n} | \, d\mu \to 0 $$

by dominated convergence (the integrand is dominated by $2$, which is integrable, because $\mu$ is a finite measure).

Hence, what you are effectively asking is if for every Borel set $B$, there is some sequence $(U_n)_n$ of open sets with $\chi_{U_n} \to \chi_B$ pointwise.

Now note that every open set $U$ fulfils $\chi_U = \lim_n f_n$ pointwise for a suitable sequence $(f_n)_n$ of continuous functions (basically, take $U = \bigcup_n K_n$ with $K_n$ compact and $K_n \subset K_{n+1}$ and use (e.g.) Urysohns Lemma to construct $f_n \in C_c(U)$ with $f_n \equiv 1 $ on $K_n$).

Hence, every functio $\chi_U$ is of Baire class 1 (see http://en.wikipedia.org/wiki/Baire_function).

So if what you are asking is true, then every indicator function $\chi_B$ with $B$ Borel would be of Baire class (at most) two.

I highly doubt that this is true, but have not found an explicit counterexample/source where this is stated.

EDIT: Also, the pointwise limit $\chi_B = \chi_{U_n}$ implies that for $x \in B$, we have $\chi_{U_n}(x) \to 1$. Because of $\chi_{U_n}(x) \in \{0,1\}$, this yields some $n_x \in \Bbb{N}$ with $\chi_{U_n}(x) = 1$ for all $n \geq n_x$ and hence $x \in \bigcap_{n\geq n_x} U_n$. Hence,

$$ x \in \bigcup_{k \geq 1}\bigcap_{n \geq k} U_n. $$

A similar argument shows that if $x \notin \bigcup_{k} \bigcap_{n \geq k} U_n$, then $x \notin B$. Hence,

$$ B = \bigcup_{k \geq 1} \bigcap_{n \geq k}U_n, $$

which implies that $B$ is a $G_{\delta, \sigma}$ set, or (other notation, same statement) a $\Sigma_3^0$ set (see http://en.wikipedia.org/wiki/Borel_hierarchy and http://en.wikipedia.org/wiki/G%CE%B4_set for the notation used here).

The article http://en.wikipedia.org/wiki/Borel_hierarchy#Boldface_Borel_hierarchy claims (the last bullet point in that paragraph):

If $X$ is an uncountable Polish space, it can be shown that $\mathbf{\Sigma}^0_\alpha$ is not contained in $\mathbf{\Pi}^0_\alpha$ for any $\alpha < \omega_1$, and thus the hierarchy does not collapse.

This implies that not every Borel set is a $G_{\delta, \sigma}$ set. Hence, your claim is false.


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.