Let $(X,\mathscr X,\mathbb P)$ be a probability space, $(Y,\mathscr Y)$ a measurable space, and $h:X\times Y\to\mathbb R$ a real-valued function measurable with respect to the product $\sigma$-algebra $\mathscr X\otimes\mathscr Y$ (where $\mathbb R$ is endowed with the Borel $\sigma$-algebra).

Moreover, let $\mathscr F$ and $\mathscr G$ be two $\sigma$-subalgebras of $\mathscr X$ such that if $E$ is in one of $\mathscr F$ or $\mathscr G$, but not in the other, then $\mathbb P(E)$ is either $0$ or $1$.

Conjecture: For every function $f:X\to Y$ that is $\mathscr F/\mathscr Y$-measurable, there exists a function $g:X\to Y$ that is $\mathscr G/\mathscr Y$-measurable such that $$h(x,f(x))=h(x,g(x))\quad\text{$\mathbb P$-a.s.}$$

The idea is that the $\sigma$-algebras $\mathscr F$ and $\mathscr G$ “almost coincide,” so that a function measurable with respect to one of them can be “slightly” modified into a function measurable with respect to the other.

Remark: A more natural version of the conjecture would require that $f(x)=g(x)$ $\mathbb P$-a.s., but without further assumptions on $(Y,\mathscr Y)$, it may happen that the set $\{x\in X\,|\,f(x)=g(x)\}$ is not even $\mathscr X$-measurable. I believe this modified conjecture holds if, for instance, $Y$ is a separable and metrizable topological space and $\mathscr Y$ is the Borel $\sigma$-algebra on it. That said, I am curious about whether the more general conjecture holds without assuming any kind of topological structure on $(Y,\mathscr Y)$. This explains why I use the function $h$ to make sense of the idea that $f$ and $g$ “almost coincide.”

This conjecture strikes me as something that “should” be true, but I am concerned that a counterexample involving weird $\sigma$-algebras on very large sets may ruin it. I would be grateful for any thoughts, remarks, or references.


I believe the answer is yes. First of all, either $\mathscr F=\mathscr G$ or both $\sigma$-algebras consist of sets of measure $0$ or $1$. Indeed, suppose that a set $E$ is in $\mathscr F$ but not in $\mathscr G$ and $\mathbb P(E)=1$ (otherwise, take $E^c$). For every set $A\in \mathscr F$, either $A\cap E$ or $A^c\cap E$ is not in $\mathscr G$, so $\mathbb P(A)$ is either $0$ or $1$. If $\mathscr G\not\subseteq \mathscr F$ then, by a similar argument, $\mathbb P(A)$ is either $0$ or $1$ for all $A\in \mathscr G$.

Let $\Pr:=\mathbb P\circ (x\to (x,f(x)))^{-1}$. Since $$\Pr(x\in A, y\in B)=\mathbb P(x\in A, f(x)\in B)=\mathbb P(x\in A)\mathbb P(f(x)\in B),$$ because $f^{-1}(B)\in \mathscr F$ has measure $0$ or $1$, $\Pr=\mathbb P\times \mu$ is a product measure on $\mathscr X \times \mathscr Y$ and, moreover, $\mu(B)=0$ or $1$ for all $B\in\mathscr Y$.

Given a set $A\in \mathscr X\times \mathscr Y$, let $A_x=\{y : (x,y)\in A\}$. Since $\mu(A_x)\in \{0,1\}$, the rectangle $r(A):=\{x: \mu(A_x)=1\}\times Y$ differs from $A$ by a set of measure zero. Therefore, any simple function on the product space is $\Pr$-almost surely equal to a function of $x$ and, as a result, the same is true for all measurable functions, i.e. $h(x,y)=\bar h(x)$ a.s.. This implies that, for some $y_0$, $h(x,y) = h(x,y_0)$ a.s. and we can take $g(x)\equiv y_0$.

| cite | improve this answer | |
  • $\begingroup$ Thank you, @D_809, the first part of your answer is crystal clear! I am not sure, though, that the existence of such a set $B\in\mathscr Y$ can be guaranteed if $\mathscr Y$ is not countably generated. Let $(X,\mathscr X,\mathbb P)$ be $[0,1]$ with the Borel $\sigma$-algebra and the Lebesgue measure, $(Y,\mathscr Y)$ be $[0,1]$ with the $\sigma$-algebra of countable and co-countable sets, and $\mathscr F=\mathscr Y$. If $f$ is the identity, then your condition states that there exists a minimal co-countable subset (and so a maximal countable subset) of $[0,1]$, which is clearly not the case. $\endgroup$ – triple_sec Dec 21 '19 at 18:44
  • $\begingroup$ Yes, you are right. Let me think some more. $\endgroup$ – D_809 Dec 21 '19 at 22:59
  • $\begingroup$ I made a second attempt. Let me know if this makes sense. $\endgroup$ – D_809 Dec 22 '19 at 0:23
  • $\begingroup$ This is awesome, much appreciated! $\endgroup$ – triple_sec Dec 22 '19 at 1:35

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.