5
$\begingroup$

The Giry monad $G : \textbf{Meas} \to \textbf{Meas}$ maps a measurable space $(X, \mathcal{F})$ to its set of probability measures. The $\sigma$-algebra of $G(X, \mathcal{F})$ is the smallest algebra such that the evaluation map $p \mapsto p(E)$ is measurable for all $E \in \mathcal{F}$. I am aware that this condition is needed to define the integral in $\mu_X : \mathcal{G}^2X \to \mathcal{G}X$, but what is the intuitive meaning of the $\sigma$-algebra? Is there an intuitive example of a subset $A \subseteq \mathcal{G}(X, \mathcal{F})$ that is not in the $\sigma$-algebra?

$\endgroup$

1 Answer 1

3
$\begingroup$

(Disclaimer: I’m not at all seriously experienced with the Giry monad; I’ve not read much beyond the original Giry paper and what I’ve picked up on the street.)

Here are two kinds of sets which typically aren’t in the Giry $\sigma$-algebra $G(X)$:

  • $\{ p \mid p(E) \in A \}$, for some measurable $E \subseteq X$ and non-measurable $A \subseteq [0,1]$. (For specific spaces $X$, it’s not hard to check this is non-measurable; I guess that should be true in fairly wide generality, but I don’t quickly see how to prove that.)

  • $\{ p \mid p(E_i) \in A_i,\, \text{for all}\ i \in I \}$, for some arbitrary uncountable family of measurable sets $E_i \subseteq X$, $A_i \subseteq [0,1]$. For instance, take $X$ to be an uncountable set with the $\sigma$-algebra of just its countable and co-countable subsets. Then it’s not hard to check that for every measurable set $S$ of $G(X)$, the set $\{ p(\{x\}) \mid p \in S \}$ is an interval $[0,a]$ for all but countably many $x \in X$.

My overall intuition is that very crudely, the Giry $\sigma$-algebra is a bit like the (infinite-indexed) product topology: on the one hand there’s a “componentwise” constraint, as exemplified by the former counterexample above; on the other hand theres an “across-components” constraint, as seen with the latter.

$\endgroup$
1
  • 3
    $\begingroup$ Note that a natural example falling into the second category are singletons $\{p\}$ which aren't in the $\sigma$-algebra if $\mathcal{F}$ isn't countably generated. $\endgroup$ Jun 10, 2023 at 13:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.