Let $(x, A) \mapsto P(x, A)$ be a probability kernel whose "target" (wikipedia terminology) is a product space $Y \times Z$, and say both $Y$ and $Z$ are compact metric spaces. For every $x$ there is the disintegration $(\mu_{x,y})_{y \in Y}$ characterized by $$P(x,A_1 \times A_2)=\int_{A_1}\mu_{x,y}(A_2) P(x,\pi^{-1}(dy))$$ where $\pi \colon Y \times Z \to Y$ is the canonical map.

Well, but I need that $\mu_{x,y}$ measurably depends on $(x,y)$, or in other words I need a kernel $(x,y,A_2) \to \mu_{x,y}(A_2)$. How could I justify this measurability ?

  • $\begingroup$ Not sure whether this answers your question, but Lemma 2 of "On the existence of good stationary strategies" by Sudderth may be of use. $\endgroup$ – Ilya Dec 5 '14 at 16:19

In my own work I needed to disintegrate a probability kernel and I found a good reference. The following is a simplified version of Theorem 1.25 in Kallenberg's 2017 book, Random Measures, Theory and Applications.

Theorem: Consider a $\sigma$-finite kernel $\rho: S \to T \times U$, where $T$ and $U$ are Borel spaces. There exist $\sigma$-finite kernels $\nu: S \to T$ and $\mu: S \times T \to U$ such that $\rho = \nu \otimes \mu$. The assertion remains true for any fixed $\sigma$-finite kernel $\nu: S \to T$ such that $\nu_s \sim \rho_s(\cdot \times U)$ for all $s \in S$.

Here a Borel space is any measurable space isomorphic to a Borel subset of the real line, for example, a Polish space.


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.