Could anyone help me understand the definition on page 7 definition 2.25 here? I do not understand the notation $(P(a))(A)$ - what does this mean?

Also, is $P(a, A)$ a probability measure from the set of all Borel sets of $X$ to $[0,1]$?

I believe that $P$ is a function from $X \times B(X)\to [0,1]$?

Is this true?


Going strictly by what Definition 2.25 in the link says, $P$ is a function $P : \mathcal{X} \to \mathcal{P}(\mathcal{X})$, where $\mathcal{P}(\mathcal{X})$ is the space of probability measures on $(\mathcal{X}, \mathscr{B}(\mathcal{X}))$ (presumably $\mathcal{X}$ is a topological space and $\mathscr{B}(\mathcal{X})$ is the $\sigma$-algebra of Borel subsets of $\mathcal{X}$, but I didn't read the notes to check whether that's true). That means that for each $a \in \mathcal{X}$, $P(a)$ is a probability measure on $(\mathcal{X}, \mathscr{B}(\mathcal{X}))$, so for each $A \in \mathscr{B}(\mathcal{X})$, $(P(a))(A)$ is a number between $0$ and $1$.

You can think of $P$ as a "random probability measure" if you equip $\mathcal{P}(\mathcal{X})$ with a $\sigma$-algebra that makes $P$ a measurable function. Usually this is done by endowing $\mathcal{P}(\mathcal{X})$ with the $\sigma$-algebra $\mathcal{H}$ which is the smallest $\sigma$-algebra such that for each $B \in \mathscr{B}(\mathcal{X})$, the map $Q \mapsto Q(B)$ from $\mathcal{P}(\mathcal{X})$ to $[0, 1]$ is measurable (i.e., $\mathcal{H}$ is generated by all such maps). This way, $P : \mathcal{X} \to \mathcal{P}(\mathcal{X})$ is a measurable function, hence a "random element" of $\mathcal{P}(\mathcal{X})$.

The random probability measure approach is one way of defining the transition probability $P$. Another approach is using so-called transition kernels (AKA Markov kernels, etc.).

If $(\mathcal{X}, \mathcal{F})$ and $(\mathcal{Y}, \mathcal{G})$ are two measure spaces, then a transition kernel from $(\mathcal{X}, \mathcal{F})$ to $(\mathcal{Y}, \mathcal{G})$ is a function $Q : \mathcal{X} \times \mathcal{G} \to [0, 1]$ such that

  1. for each $x \in \mathcal{X}$, the map $Q(x, \cdot) : \mathcal{G} \to [0, 1]$ is a probability measure on $(\mathcal{Y}, \mathcal{G})$, and
  2. for each $G \in \mathcal{G}$, the map $Q(\cdot, G) : \mathcal{X} \to [0, 1]$ is $\mathcal{F}$-measurable.

Given a transition kernel $Q$ from $(\mathcal{X}, \mathcal{F})$ to $(\mathcal{Y}, \mathcal{G})$, we can define a random measure $P : \mathcal{X} \to \mathcal{P}(\mathcal{Y})$, where $\mathcal{P}(\mathcal{Y})$ stands for the space of probability measures on $(\mathcal{Y}, \mathcal{G})$, by defining $$ (P(a)(A)) = Q(a, A) $$ for each $a \in \mathcal{X}$ and $A \in \mathcal{G}$. Conversely, given a random measure $P : \mathcal{X} \to \mathcal{P}(\mathcal{Y})$, we can define a transition kernel $Q$ from $(\mathcal{X}, \mathcal{F})$ to $(\mathcal{Y}, \mathcal{G})$ by the same relation. Thus, there is a one-to-one correspondence between random probability measures and transition kernels.

Because of this, it is common to use notation like $P(a, A)$ when $P$ is a random probability measure to stand for $(P(a))(A)$; the notation $P(a, A)$ is easier to read, and you're not really doing anything wrong by treating the random probability measure as a transition kernel.


In homogeneous markov chain the system dynamics do not change over time, I think it would be just the expression of the transition kernel which is fixed with time.

Maybe the typing in this textbook would be more expressive:
enter image description here

  • $\begingroup$ Thanks for your answer, actually I am interested in understanding the notations and definition from the Hairer notes only, for my context. $\endgroup$ – riemann May 24 '19 at 9:45

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.