This is a seemingly basic question that I realise I don't know the answer to:

Let $\mathcal{B}$ be the Borel $\sigma$-algebra of $\mathbb{R}$.

Is it the case that for every sub-$\sigma$-algebra $\mathcal{G}$ of $\mathcal{B}$ there exists a countably generated sub-$\sigma$-algebra $\mathcal{G}'$ of $\mathcal{G}$ such that for every probability measure $\mathbb{P}$ on $\mathcal{B}$, $\mathcal{G} \subset \sigma(\mathcal{G}' \cup \mathcal{N}_{\mathbb{P}})$?

(Here, $\mathcal{N}_\mathbb{P}$ denotes the set of $\mathbb{P}$-null sets.)

Although less important for me, it would also be interesting if possible to know the following:

  • If the answer to the above question is no (in ZFC): Is the answer to the above question being yes consistent with ZF+DC?
  • But if the answer is yes: Does it still remain yes if we replace "every probability measure $\mathbb{P}$ on $\mathcal{B}\,$" with "every probability measure $\mathbb{P}$ on $\mathcal{G}'\,$"?

I've now managed to work out that the answer is no:

Proposition. Let $\rho$ be a probability measure on $(\mathbb{R},\mathcal{B})$, and let $\mathcal{G}$ be a sub-$\sigma$-algebra of $\mathcal{B}$ such that

  • $\rho(A) \in \{0,1\}$ for all $A \in \mathcal{G}$;
  • for $\rho$-almost every $x \in \mathbb{R}$ there exists $A \in \mathcal{G}$ such that $x \in A$ and $\rho(A)=0$.

Then for every countably generated sub-$\sigma$-algebra $\mathcal{G}'$ of $\mathcal{G}$ there exists a probability measure $\mathbb{P}$ on $(\mathbb{R},\mathcal{B})$ such that $\mathcal{G} \not\subset \sigma(\mathcal{G}' \cup \mathcal{N}_\mathbb{P})$.

So, for instance, we could take $\mathcal{G}$ to be the countable-cocountable algebra (with $\rho$ being any atomless probability measure).

Proof. We can represent $\mathcal{G}'$ as $\sigma(N_i : i \in \mathbb{N})$ for some sequence of $\rho$-null sets $N_i \in \mathcal{G}$. Let $N:=\bigcup_{i=1}^\infty N_i$. It is clear that $$ \mathcal{D} \ := \ \bigcup_{B \subset N} \{B , \mathbb{R} \!\setminus\! B\} $$ is a $\sigma$-algebra, and therefore $\mathcal{G}' \subset \mathcal{D}$. Fix $x \in \mathbb{R} \setminus N$ and $A \in \mathcal{G}$ such that $x \in A$ and $\rho(A)=0$. Let $\mathbb{P}=\frac{1}{2}\delta_x + \frac{1}{2}\rho$. Then $\mathbb{P}(A)=\frac{1}{2}$. However, for any $B \in \mathcal{D}$, if $B \subset N$ then $\mathbb{P}(B)=0$ but if $\mathbb{R} \!\setminus\! B \subset N$ then $\mathbb{P}(B)=1$. So in particular, every member of $\mathcal{G}'$ has $\mathbb{P}$-trivial measure, and so $A \not\in \sigma(\mathcal{G}' \cup \mathcal{N}_\mathbb{P})$. QED

[As for the additional question: the above proof is valid in ZF+DC.]

  • $\begingroup$ You say that the proof is valid in ZF+DC, but I don't see what is a non-binary measure on the countable-cocountable algebra. Sure, I will believe you if you tell me that you can find one with Zorn's lemma (or some weaker choice principle of the flavor of BPI), but not with only DC. $\endgroup$ – Asaf Karagila Mar 5 '17 at 21:58
  • 1
    $\begingroup$ What about a non-trivial convex combination of two distinct Dirac masses? Or, as in my proof, the equal-weight convex combination of a Dirac mass and the standard binary measure? Is there a misunderstanding? $\endgroup$ – Julian Newman Mar 5 '17 at 23:51

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.