Suppose that you are working with the axiom of choice. If you randomly choose a subset of the real line, what is the probability that it will be measurable?

  • 13
    $\begingroup$ What probability distribution are you using? $\endgroup$ – Chris Eagle May 9 '11 at 21:45
  • 5
    $\begingroup$ The cardinality of the collection of all measurable sets is $2^{\mathfrak{c}}$ (at least all subsets of the Cantor set); the cardinality of the collection of all nonmeasurable sets is also at least $2^{\mathfrak{c}}$ (you have at least that many distinct Vitali sets). Hence both are of cardinality $2^{\mathfrak{c}}$. So you can't even argue as one sometimes does that a "random real" is almost certain to be transcendental (based on cardinality). I don't think the question can be given a reasonable answer. $\endgroup$ – Arturo Magidin May 9 '11 at 21:58
  • 1
    $\begingroup$ The answer has to be zero. If not, I don't believe in choice. $\endgroup$ – Michael Lugo May 9 '11 at 22:47
  • $\begingroup$ @Micheal: Why? Usually when one encounters a counter example, they turn out to be numerous... $\endgroup$ – Thomas Rot May 9 '11 at 22:54
  • 1
    $\begingroup$ What about the probability distribution where $P(x \in A)=1/2$ for each $x\in\mathbb{R}$? $\endgroup$ – mjqxxxx May 10 '11 at 14:21

A reasonable way to select a random subset $A$ of the real line might be the following. For each $t\in\mathbb{R}$, decide (independently for each $t$) to either put $t$ in $A$ or not, each with probability 1/2. This is what @mjqxxxx suggested.

To make this precise, for each $t\in\mathbb{R}$, we have the basic Bernoulli 1/2 probability space: $\Omega_t=\{0,1\}$, $\mathcal{F}_t=2^\Omega$, and $P_t$ is defined by $P_t(\{0\})=P_t(\{1\})=1/2$. The overall probability space is the product of these: $\mathbb{X}=\prod_{t\in\mathbb{R}}\Omega_t$, $\mathbb{F}=\bigotimes_{t\in\mathbb{R}}\mathcal{F}_t$, and $\mathbb{P}$ is the product measure. Note that $\mathbb{X}$ is the collection of all subsets of $\mathbb{R}$, i.e. the power set of $\mathbb{R}$.

Let $\mathcal{M}\subset\mathbb{X}$ denote the collection of Lebesgue measurable sets. (The same thing can be done for Borel measurable sets as well.) We would like to compute $\mathbb{P}(\mathcal{M})$.

Unfortunately, as will be shown below, $\mathcal{M}\notin\mathbb{F}$, and so $\mathbb{P}(\mathcal{M})$ is undefined. In other words, the set of measurable sets is not measurable, so we cannot speak of the probability that a random set is measurable. (At least not in this model.)

To prove this, define $\pi_t:\mathbb{X}\to\{0,1\}$ by $\pi_t(A)=1_A(t)$. Then $\pi_t$ are the projections, and, by definition, $\mathbb{F}=\sigma(\pi_t:t\in\mathbb{R})$. By Proposition 4.6 in "Probability and Stochastics" by Erhan Çinlar, a function $F:\mathbb{X}\to\mathbb{R}$ is measurable if and only if there exists a sequence $(t_n)$ in $\mathbb{R}$ and a measurable function $f:\{0,1\}^\infty\to\mathbb{R}$ such that $$ F(A) = f(\pi_{t_1}(A), \pi_{t_2}(A), \ldots), $$ for all $A\in\mathbb{X}$.

Now suppose $\mathcal{M}\in\mathbb{F}$ and define $F=1_{\mathcal{M}}$. Choose a sequence $(t_n)$ and a measurable function $f$ such that $$ 1_{\mathcal{M}}(A) = f(\pi_{t_1}(A), \pi_{t_2}(A), \ldots), $$ for all $A\in\mathbb{X}$. Let $B=\{t_1,t_2,\ldots\}$. Since $B$ is countable, it is Lebesgue measurable, and so $$ 1 = 1_{\mathcal{M}}(B) = f(\pi_{t_1}(B), \pi_{t_2}(B), \ldots) = f(1,1,\ldots). $$ Let $C$ be any non-measurable set. Since $t_n\in C\cup B$ for all $n$, we have $$ 1_{\mathcal{M}}(C\cup B) = f(\pi_{t_1}(C\cup B), \pi_{t_2}(C\cup B), \ldots) = f(1,1,\ldots) = 1, $$ and so $C\cup B$ is measurable. But $C^c\cap B$ is countable, and therefore measurable, from which we deduce that $$ C = (C\cup B) \cap (C^c \cap B)^c $$ is a measurable set, which is a contradiction.


I wanted to add some additional information which was too much for a comment, and which incidentally also addresses the comment from @Hendrik.

One might wonder if we can find a measurable collection $\mathcal{A}$ such that $\mathcal{A}\subset\mathcal{M}$. If we can and if, for example, we have $\mathbb{P}(\mathcal{A})=1$, then we can say that the random set is almost surely measurable. But it should be straightforward to adapt the above proof to show that if $\mathcal{A}\in\mathbb{F}$ and $\mathcal{A}\subset\mathcal{M}$, then $\mathcal{A}=\emptyset$. Similarly, if $\mathcal{A}\in\mathbb{F}$ and $\mathcal{M}\subset\mathcal{A}$, then $\mathcal{A}=\mathbb{X}$.

It follows from this that $\mathcal{M}$ is not in the completion of this probability space. To see this, suppose that $\mathcal{M}=\mathcal{A}\cup\mathcal{N}$, where $\mathcal{A}\in\mathbb{F}$ and $\mathcal{N}\subset\mathcal{B}$ for some $\mathcal{B}\in\mathbb{F}$ satisfying $\mathbb{P}(\mathcal{B})=0$. Then $\mathcal{A}\subset\mathcal{A}\cup\mathcal{N}=\mathcal{M}$, which implies $\mathcal{A}=\emptyset$. Thus, $\mathcal{M}=\mathcal{N}\subset\mathcal{B}$, which implies $\mathcal{B}=\mathbb{X}$. But this contradicts $\mathbb{P}(\mathcal{B})=0$.

  • 1
    $\begingroup$ What happens if you take the completion of the measure space $(\mathbb{X},\mathbb{P})$? $\endgroup$ – Hendrik Vogt Jun 30 '11 at 8:14

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.