2
$\begingroup$

Let $(X,\mu)$ be a probability measure space and $A$ be a measurable subset of $X$ such that $0 \le \mu(A) < p < 1$.

Question

When is it true that there exists a measurable $B \subseteq X$ such that $A \subseteq B$ and $\mu(B)=p$ ?

$\endgroup$
  • 2
    $\begingroup$ It is true if there are no atoms (that means: any subset of positive measure has a subset of positive but smaller measure.) $\endgroup$ – Fedor Petrov Nov 29 at 14:20
  • $\begingroup$ Ah, great thanks. I've just come across essentially this same fact at the moment, the so-called Sierpiński's Theorem, a kind of "Intermediate-Value Theorem" for measures. $\endgroup$ – dohmatob Nov 29 at 14:26
  • $\begingroup$ (now I remember I wrote a short section in a wiki article on this: en.wikipedia.org/wiki/Atom_(measure_theory)#Non-atomic_measures) $\endgroup$ – Pietro Majer Nov 29 at 14:43
  • $\begingroup$ Yes, that's I was referring to in my comment. $\endgroup$ – dohmatob Nov 29 at 14:44
5
$\begingroup$

An equivalent condition is that the probability space $(X, \mathcal{M}, \mu)$ is non-atomic, meaning that for every $A\in \mathcal{M}$ with $\mu(A)>0$ there exists $B \in \mathcal{M}$ such that $B\subset A$ and $0<\mu(B)<\mu(A)$.

If $(X, \mathcal{M}, \mu)$ is non-atomic then $\mu: \mathcal{M}\to [0,1]$ has a monotone section: that is, a map $E:[0,1]\to \mathcal{M}$ such that for all $0\le t\le s\le1$ one has $\emptyset=E_0\subset E_t\subset E_s\subset E_1=X$ and $\mu(E_t)=t$.

In fact, any "partial monotone section of $\mu$" (meaning a family $E:S\to \mathcal{M}$ as above, but with possibly smaller domain $S\subset [0,1]$) can be extended to a monotone section of $\mu$ (which in particular solves your problem).

I think this result is essentially due to Sierpiński; the proof uses the Zorn lemma on the set of graphs of the partial monotone sections of $\mu$, ordered by "extension" i.e. inclusion of graphs.

By the way, the maximality argument via Zorn lemma generalizes to a proof of the Lyapunov convexity theorem: if a $\mathbb{R}^n$ valued bounded vector measure $\mu=(\mu_1,\dots,\mu_n)$ is non atomic (meaning, all $\mu_i$ are non-atomic bounded measures), then the range of $\mu$ is a compact convex set.

$\endgroup$
  • $\begingroup$ Thanks for the input. The question has been answered in the comments. Someone also downvoted it for the fun of it... $\endgroup$ – dohmatob Nov 29 at 14:41
  • $\begingroup$ Thanks for mentioning the Chebychev convexity theorem. New good stuff. $\endgroup$ – dohmatob Nov 29 at 14:55
  • $\begingroup$ Which minimal restrictions on my problem (i.e on a non-atomic measure) would remove the need for the (somewhat controversial) Zorn's lemma in the proof ? $\endgroup$ – dohmatob Nov 29 at 14:57
  • $\begingroup$ What you want is to prove that a non-atomic probability measure is divisible, meaning the range is [0,1]. I guess the axiom of Dependent Choice is sufficient $\endgroup$ – Pietro Majer Nov 29 at 16:10
  • 1
    $\begingroup$ What exactly was proved by Chebyshev? $\endgroup$ – Fedor Petrov Nov 29 at 16:39
2
$\begingroup$

An argument for Sierpiński's Intermediate Value Theorem for non-atomic measures without using Zorn's lemma (but possibly implicitly using countable axiom of choice, that I am not qualified to judge), it was told to me by Alexander Kuznetsov.

Denote for a measurable subset $C\subset X$ $$f(C,q):=\sup_{D\subset C,\mu(D)\leqslant q} \mu(D).$$ We want to prove that $f(C,q)=q$ if $\mu(C)\geqslant q$. Choose consecutively disjoint subsets $C_1,C_2,\ldots$ in $C$ so that $$ q-\sum_{j<i}\mu(C_j)\geqslant \mu(C_i)\geqslant f\left(C\setminus \cup_{j<i} C_j,q-\sum_{j<i}\mu(C_j)\right)-\frac 1i$$ for all $i=1,2,\ldots$. Then for $C_0=\cup C_i$ we have $\mu(C_0)=\sum \mu(C_i)\leqslant q$. Assume that $\mu(C_0)=q-a$ for $a>0$. The set $C\setminus C_0$ has positive measure, thus it contains a subset $E$ for which $0<\mu(E)<a$ (since our measure is non-atomic, the set $C\setminus C_0$ may be partitioned onto two subsets of positive measure, take the part with smaller measure and partition it again and so on). But it means that for large $i$ the choice of $C_i$ was not $1/i$-optimal in above sense: we could use $E$ instead. A contradiction.

$\endgroup$
  • $\begingroup$ Great constructive answer. Thanks! $\endgroup$ – dohmatob Nov 29 at 17:17

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.