I am reading the book Markov Chains and Stochastic Stability from Meyn and Tweedie. They define Markov chains on a measurable state space $(E,\Sigma)$ (Chapter 3.4) and they define it on the space $\Omega = \prod_{i \in \mathbb{N}}E, $ with an $\sigma$-algebra $\mathcal{A}$ which is the smallest $\sigma$-algebra that contains all cylinder sets with only finitly many sets different from $E$ $$A_1 \times A_2 \times \dots A_n \times E \times E \times \dots$$

Then they define the Markov chain as a family of random variables $(X_n)_{n \in \mathbb{N}}$ where for $\omega=(x_n)_{n \in \mathbb{N}}\in \Omega$ they set $$X_n(\omega)=x_n .$$

Thus, all Markov chains are defined on the same set $\Omega$ and the random variables $(X_n)$ are also always the same. Now if they talk about a certain initial distribution $\mu$ and transition kernel $p(x,A)$; then they assoicate a Markov chain to it by constructing a specific measure $\mathbb{P}_\mu$. Thus, by this definition, two Markov chains only differ on the probability measure of the probability space.

My problem is that in the book they define the term $$ \mathcal{F}_n = \sigma(X_0,\dots,X_n) \subseteq \mathcal{B}(X^{n+1})$$ and they say

which is the smallest $\sigma$-field for which the random variable $\{X_0,\dots,X_n\}$ is measurable. In many cases $\mathcal{F}_n$ will coincide with $\mathcal{B}(X^{n+1})$, although this depend in particular on the initial measure $\mu$ choosen for a particular chain.

How can $\mathcal{F}_n$ depend on the initial measure? The random variable is already defined as $X_n(\omega)=x_n$, and thus the measurability of $\{X_0,\dots,X_n\}$ depends only on $\Sigma$ and $\mathcal{A}$ where does the intial measure $\mu$ comes into play?

Update: After seeing the answers, I think it is a good thing to provide my question with an example. Lets consider the case where $E=\{1,2\}$ and $\Omega = E \times E$, then the random variables $X_0$ and $X_1$ are already defined as above, in particular $X_0$ is defined as $$ X_0 ((1,1))=X_0((1,2))=1$$ and $$X_0((2,1))=X_0((2,2))=2.$$ Now if $\mathbb{P}_\mu$ is the probability that $X_0 = 1 $, then we must have $$ \mathbb{P}_\mu[\{(1,1),(1,2)\}]=1.$$ But this is completely independent from defining $\mathcal{F}_0$ (or $\mathcal{F}_n$). In this case we always have $$\mathcal{F}_0 = \{\{(1,1),(1,2)\},\{(2,1),(2,2)\},E,\emptyset \} $$ which does not depend on $\mu$. It seems to me that in the answers one believes that $\mathbb{P}_\mu[\{(2,1),(2,2)\}]=0$ implies somehow that this set should not belong to $\mathcal{F}_0$, but I think this is not correct.

  • 1
    $\begingroup$ Working with the canonical process, the sigma field does not depend on the measure indeed : you are right. $\endgroup$ – Olivier Aug 14 at 9:21
  • $\begingroup$ This being said, one could also consider the sigma-algebra generated by the measurable sets with non-null measure if we want to force the measure into, and perhaps that's what the authors have in mind... $\endgroup$ – Olivier Aug 14 at 9:26

If we take $\mu$ to be the distribution of $X_0$, then $\sigma(X_0)$ depends on $\mu$ already: $\sigma(X_0)$ is the smallest sigma-algebra for which $X_0$ is measurable. Assume $X_0\equiv 1$ ($\mu$ is a point measure at 1), without loss of generality. Then $\sigma(X_0)=\{\Omega,\emptyset\}$. Assume $X_0\in\{1,2\}$, and $A=X_0^{-1}(1)$, then $\sigma(X_0)$ must contain at least the elements $\{\Omega,A,A^c,\emptyset\}$.

Hence, $\mathcal{F}_n$ depends on $\mu$.

Update: I believe I now better understand the question: If the RVs are already defined, they must already be measurable functions between two measure spaces. But using the generator operator $\sigma$ delivers the smallest sigma-algebra w.r.t. which the function/RV is measurable. So, if we define

$$ X_0{:}\ (\Omega,\mathcal{B}(\Omega)) \to (\mathbb{R},\mathcal{B}(\mathbb{R})) $$

but then say (as you do) that only one state of $X_0$ occurs, a smaller sigma-algebra might suffice.

Let us take your example with $E=\{1,2\}$, $\Omega=E\times E$, but now define $X_0((a,b))=a+b$. In this most general case, $\mathcal{F}_0$ is generated by $\{\emptyset,\Omega,\{(1,1)\},\{(2,1),(1,2)\},\{(2,2)\}\}$. But if you assume that $X_0\equiv 2$ (by properly choosing $\mu$), then $\mathcal{F}_0$ is generated by the smaller familiy of sets $\{\emptyset,\Omega,\{(1,1)\},\{(2,1),(1,2),(2,2)\}\}$. This was not apparent from your example, because you treated only two possible states for $X_0$.

  • 1
    $\begingroup$ +1. Why is this downvoted? Being short and simple are plusses for an answer when it is to the point, no? $\endgroup$ – Did Jun 2 '15 at 9:16
  • $\begingroup$ I have updated my question with an example which fits to your example and shows why $\sigma(X_0)$ should not depend on $\mu$. @Did I dont know - I have neither downvoted nor upvoted anything so far. $\endgroup$ – Adam Jun 2 '15 at 9:33
  • $\begingroup$ Here, $X_n(\omega)=x_n$, this is the key assumption you miss in the second example. $\endgroup$ – Olivier Aug 14 at 9:23

$X_0$ is the random variable variable with measure $\mu$ so $\mathcal{F_n}$ does depend on its definition.

If $\mu$ was a discrete random variable it could give very different possible paths based on the number of outcomes. If $\mu$ allowed for one only one outcome $0$ then the minimum sigma field wrt which it is measurable is $\{\phi, \{0\}, Z \ \{0\}, Z\}$ where $Z$ is the state space. In comparison, a continuous random variable might allow for many, many more outcomes still, i.e. a much large sigma algebra.

To be concrete, compare Brownian Motion with a fixed initial condition, versus normally distributed initial conditions.

  • $\begingroup$ The measure $\mu$ only gives a weight to a specific path. I don't see how this changes $\mathcal{F}_n$. For instance, $\mathcal{F}_0$ is the smallest sigma algebra, such that for any $A\in \Sigma$ we have $X_0^{-1}(A)\in \mathcal{F}_0$ this is obviously independet of $\mu$, because $X_0$ is already defined as a projection, right? $\endgroup$ – Adam Jun 2 '15 at 8:35
  • $\begingroup$ @Adam I updated to clarify the point. $\endgroup$ – muaddib Jun 2 '15 at 8:40
  • $\begingroup$ I still do not understand your point. I have updated my question with an example to clarify why I think $\mathcal{F}_n$ is independent of $\mu$. $\endgroup$ – Adam Jun 2 '15 at 9:26
  • $\begingroup$ @Adam, I'm not sure what to say, I don't really see why you don't think $X_0$ changes with the definition of $\mu$. $\endgroup$ – muaddib Jun 2 '15 at 9:34
  • $\begingroup$ In order to assure that $X_0$ is measurable, you also need that $X_0^{-1}(\{2\}) \in \mathcal{F}_0$, right? Also note that my defined $\mathcal{F}_0$ is not maximal, since $\{(1,1),(2,2)\} \notin \mathcal{F}_0$. $\endgroup$ – Adam Jun 2 '15 at 9:37

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.