Rephrasing the strong Markov property with regular conditional distributions

56 Views Asked by At

This question is about definition 17.12 (of the strong markov property) and the following remark 17.13 in A. Klenke's Probability Theory.

For completeness, the definition: Let $E$ a polish space, and let $I\subset[0, \infty)$ be closed under addition. A Markov process $\left(X_t\right)_{t \in I}$ with values in $E$ and distributions $\left(ℙ_x, x \in E\right.$ ) has the strong Markov property if, for every a.s. finite stopping time $\tau$, every bounded $\mathcal{B}(E)^{\otimes I}$-$\mathcal{B}(\mathbb{R})$ measurable function $f: E^I \rightarrow \mathbb{R}$ and every $x \in E$, we have \begin{equation*} _x\left[f\left(\left(X_{\tau+t}\right)_{t \in I}\right) \mid \mathcal{F}_\tau\right]=_{X_\tau}[f(X)]:=\int_{E^I} \kappa\left(X_\tau, d y\right) f(y) . \tag{$*$} \end{equation*} Here, $κ: E \times \mathcal B(E)^{⊗ I} → [0,1]$ is defined by $κ(x,A) = ℙ_x[X ∈ A]$, which is a stochastic kernel by definition of a Markov process.

The remark following this definition states that if $I$ is countable, then the strong Markov property holds if and only if, for every almost surely finite stopping time $\tau$, we have \begin{equation*} \mathcal{L}_x\left[\left(X_{\tau+t}\right)_{t \in I} \mid \mathcal{F}_\tau\right]=\mathcal{L}_{X_\tau}\left[\left(X_t\right)_{t \in I}\right]:=\kappa\left(X_\tau, \cdot\right) . \tag{$\diamond$} \end{equation*} Here, $\mathcal L_x[(X_{\tau+t})_{t \in I} | \mathcal F_τ]$ denotes a regular conditional distribution (r.c.d.) of $\left(X_{\tau+t}\right)_{t \in I}$ given the sub-σ-algebra $ \mathcal F_τ$ (this notation is defined at the end of the definition of a Markov process).


The question is: why does the remark emphasise that the index set $I$ be countable? I understand that this would guarantee existence of the r.c.d. (since then $E^I$ is polish, so $(X_{\tau+t})_{t \in I}$ takes values in a polish space and thus has a r.c.d. given $\mathcal F_τ$), but I don't see why we need this guarantee: It seems to me that if the strong Markov property holds, then $(ω,A) ↦ κ(X_τ(ω),A)$ is a stochastic kernel which by $(*)$ is a r.c.d. of $(X_{\tau+t})_{t \in I}$ given $\mathcal F_τ$. On the other hand, if $(\diamond)$ holds for every a.s. finite stopping time $τ$, then $(*)$ follows from properties of regular conditional distributions (specifically Theorem 8.38 in the same book).