Let $\mu$ be a non-atomic probability measure defined on $\mathbb R$. Let $I$ simply be an interval of length $\beta$, i.e., $ I = [i \beta,(i+1)\beta]$ for some $i \in \mathbb{Z}$, and $B$ any Borel set. I have the following fact given to me, which I was not able to refute yet and might well be correct:

$\mu(I \cap B) \rightarrow \mu(I) \dfrac{\mathrm{Len}(I \cap B)}{\mathrm{Len}(I)}$ as $\beta\downarrow 0$.

Here $\mathrm{Len}$ denotes the 1-D volume or the exterior measure.

This is a bit hard to imagine. The intuition (which may fail most of the times at measure theory) says that this is roughly dividing the probability assigned to interval $I$ by its length so that we have a 'unit/uniform' measure now. Then we assign this unit measure to the length of $I \cap B$ simply by multiplication. How can we prove this result, if this is true? If not, what are the minimum number of assumptions we need such that this is true?

  • 2
    $\begingroup$ How are you defining $Len(I \cap B)$? $\endgroup$ May 19 '21 at 14:40
  • $\begingroup$ @user6247850 I did not do that yet, I was hoping for the 1-D volume (exterior measure). $\endgroup$ May 19 '21 at 15:03
  • 3
    $\begingroup$ If $\mu$ is non-atomic then as $\beta \to 0$, $\mu(I \cap B) \to 0$. So in some sense you want to claim that this is the "leading order term" in the asymptotic expansion as $\beta \to 0$. (Or refute that, too.) $\endgroup$
    – Ian
    May 19 '21 at 15:31
  • $\begingroup$ So it seems like perhaps you want $\sim$ in place of $\to$ in the statement. I am also not quite sure of the role of $i$; is it fixed once and for all, or what? $\endgroup$ May 24 '21 at 4:17
  • $\begingroup$ @NateEldredge thanks for your comment. $i$ is defined such that $I$ is an interval of length $\beta$. I can change my wording there to be more specific. $\endgroup$ May 24 '21 at 15:31

I interpret your question as follows.

Let $\mu$ be a non-atomic probability measure on $\mathbb{R}$. Let $B$ be a Borel subset of $\mathbb{R}$. Then, $$\lim_{\beta \downarrow 0} \frac{\mu([0,\beta]\cap B)/\mu([0,\beta])}{|[0,\beta]\cap B|/|[0,\beta]|} = 1,$$ where $|\cdot|$ denotes Lebesgue measure.

This claim is definitely false. One can take $\mu = \sum_{k=1}^\infty 2^{-k}\text{Unif}([2^{-2^k-1},2^{-2^k}])$, where $\text{Unif}(I)$ denotes the uniform probability measure over $I$, and $B = \cup_{k=1}^\infty [2^{-2^k-2},2^{-2^k}]$.


I think it may not be correct for all cases. Consider a positive decreasing sequence $\beta_n \to 0$ for $n \to \infty$. We have that $\textrm{Len}(I_n)=\beta_n$. Let's consider $I_n=[0,\beta_n]$ (i.e. $i=0$). We have that $I_1 \supset I_2 \supset I_3 \supset ...$ and $I_\infty=\cap_{n \in \mathbb{N}}I_n=\{0\}$. So $\lim_{n \to \infty}\mu(I_n)=\mu(\{0\})$ and $\lim_{n \to \infty}\mu(I_n \cap B)=\mu(B \cap \{0\})$. Let $B$ be $(-\varepsilon,0]$ for some $\varepsilon >0$. We have that $\textrm{Len}(I_n \cap B)=\textrm{Len}(\{0\})=0,\forall n \in \mathbb{N}$. So $$\lim_{n \to \infty}\frac{\textrm{Len}(I_n \cap B)}{\textrm{Len}(I_n)}\mu(I_n)=\lim_{n \to \infty}0 = 0$$ which is not necessarily equal to $\mu(B \cap \{0\})=\mu(\{0\})$.

  • $\begingroup$ Thanks for your answer! However, isn't probability measure on a singleton always $0$? $\endgroup$ May 19 '21 at 15:02
  • $\begingroup$ A probability measure is a measure on a set $X$ such that $\mu(X)=1$, nothing else $\endgroup$
    – Snoop
    May 19 '21 at 15:06
  • $\begingroup$ That is a good point, thank you. Which additional assumptions are needed so that the statement above is correct? $\endgroup$ May 19 '21 at 15:12
  • $\begingroup$ I think the measure should be non-atomic, am I right? $\endgroup$ May 19 '21 at 15:22
  • 1
    $\begingroup$ Added that, thanks so much! $\endgroup$ May 19 '21 at 15:27

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.