Let $(X,d)$ be a metric space and $K:X\times \mathcal B(X)\to[0,1]$ is a stochastic kernel on $X$. We call this kernel absolute continuous with respect to a measure $\mu:\mathcal B(X)\to\mathbb R_{\geq0}$ if for any $B\in\mathcal B(X)$ such that $\mu(B) = 0$ holds $K(x,B) = 0$ for all $x\in X$.

What a necessary and sufficient conditions for such measure $\mu$ to exist given a kernel $K$?

  • 1
    $\begingroup$ I think that if $K(x,\cdot)$ is a measure for each $x$, then $\mu(\cdot)=\sum_{x\in X}K(x,\cdot)$ is a measure, and each $K(x,\cdot)$ is absolutely continuous with respect to $\mu$. $\endgroup$ – Jason Swanson Jun 30 '11 at 9:45

Suppose that $K(x,\cdot)$ is a measure for each $x$. Define $$ \mu(B) = \sum_{x\in X}K(x,B). $$ This possibly uncountable sum can be understood either as the supremum over all finite subsums, or as the integral with respect to counting measure on $X$. Those are equivalent definitions. The latter is useful, so that we may apply tools of integration theory, such as Fubini/Tonelli.

To check that $\mu$ is a measure, first note that $\mu(\emptyset)=0$. Let $\{B_n\}$ be pairwise disjoint. Then \begin{align*} \mu\bigg(\bigcup_{n=1}^\infty B_n\bigg) &= \sum_{x\in X}K\bigg(x,\bigcup_{n=1}^\infty B_n\bigg)\\ &= \sum_{x\in K}\sum_{n=1}^\infty K(x,B_n)\\ &= \sum_{n=1}^\infty\sum_{x\in X} K(x,B_n)\\ &= \sum_{n=1}^\infty\mu(B_n). \end{align*} Reversing the order of summation is justified by Tonelli's theorem.

Finally, if $\mu(B)=0$, then $K(x,B)=0$ for all $x$. Hence, $K$ is absolutely continuous with respect to $\mu$.


To address the comments, let us suppose that $x\mapsto K(x,U)$ is Borel measurable for each open $U$. We will show that this implies $x\mapsto K(x,B)$ is Borel measurable for every $B\in\mathcal{B}(X)$.

Let $f_B(x)=K(x,B)$. Let $$ \mathcal{D} = \{B\in\mathcal{B}(X): f_B\text{ is Borel measurable}\}. $$ Since $X$ is open, we have $X\in\mathcal{D}$. If $A,B\in\mathcal{D}$ and $B\subset A$, then $f_{A\setminus B}=f_A-f_B$ is measurable, which implies $A\setminus B\in\mathcal{D}$. And if $\{A_n\}_{n=1}^\infty$ is an increasing sequence in $\mathcal{D}$ with $A=\bigcup_{n=1}^\infty A_n$, then $f_A=\sup_n f_{A_n}$ is measurable, which shows that $A\in\mathcal{D}$. In all, we have shown that $\mathcal{D}$ is a Dynkin system that contains the open sets. So by the $\pi$-$\lambda$ theorem, $\mathcal{D}=\mathcal{B}(X)$.

If $\nu$ is a measure on $(X,\mathcal{B}(X))$, then we may now define $$ \mu(B) = \int_X K(x,B)\nu(dx). $$ The proof that $\mu$ is a measure is the same as what was done above for counting measure. However, now if $\mu(B)=0$, then we may only conclude that $K(x,B)=0$ for $\nu$-a.e. $x$. A word of warning: the exceptional null set may depend on $B$. More specifically, we can say that $$ \forall B,\;\mu(B)=0 \quad\Rightarrow\quad(K(x,B)=0\text{ for $\nu$-a.e. $x$}). $$ But we can *not* say that $$ (\forall B,\;\mu(B)=0 \quad\Rightarrow\quad K(x,B)=0)\text{ for $\nu$-a.e. $x$}. $$ In fact, there may not be a single $x\in X$ such that $K(x,\cdot)$ is absolutely continuous with respect $\mu$.

For example, let $X=\mathbb{R}$ and $K(x,B)=1_B(x)$. In other words, $K(x,\cdot)=\delta_x$, the point mass centered at $x$. Let $\nu=m$, where $m$ is Lebesgue measure. Then $$ \mu(B) = \int_{\mathbb{R}}1_B(x)m(dx) = m(B). $$ That is, $\mu=m$. It is true that for each fixed $B$ with Lebesgue measure zero, $\delta_x(B)=0$ for Lebesgue a.e. $x$. However, there is not a single $x\in\mathbb{R}$ such that $\delta_x(B)=0$ for every Lebesgue null $B$.

  • $\begingroup$ thank you, could we think about conditions on the measure $\nu$ to use it instead of the counting measure, or you prefer to see it as a new question? $\endgroup$ – Ilya Jun 30 '11 at 10:05
  • $\begingroup$ If $x\mapsto K(x,B)$ is measurable for each $B$, then you could define $\mu$ that way. But in the end, given $\mu(B)=0$, all you could conclude is that $K(x,B)=0$ for $\nu$-a.e. $x$. $\endgroup$ – Jason Swanson Jun 30 '11 at 10:11
  • $\begingroup$ what if $K(x,B)$ is continuous for all open $B$? $\endgroup$ – Ilya Jun 30 '11 at 10:25

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.