I'm trying to solve the following problem, which appears in a course about HMM's (problem 6.2).

Let $P,Q$ be probability measures on $(\Omega,\mathcal{G})$ such that $P$ is absolutely continuous with respect to $Q$ and let $\mathcal{G}'\subset\mathcal{G}$ a sub-$\sigma$-field. Show that $$\frac{dP_{\mid\mathcal{G}'}}{dQ_{\mid\mathcal{G}'}}=E_Q\left[\frac{dP}{dQ}\mid\mathcal{G}'\right]$$

I'm not even sure I understand the question.

1) Is $P_{\mid\mathcal{G}'}$ defined by $P_{\mid\mathcal{G}'}(A)=P[A\mid\mathcal{G}']=E_P[1_A\mid\mathcal{G}']$ (and idem for $Q$)?

2) Can we say that $$E_Q\left[\frac{dP}{dQ}\mid\mathcal{G}'\right]=\int_{\Omega}\frac{dP}{dQ}dQ_{\mid\mathcal{G}'}$$ (using the fact that conditional expectation is the expectation under the conditional ditribution)?

3) Using the definition of conditional expectation, I should prove that for any $A\in\mathcal{G}'$, we have $$E_Q\left[\frac{dP}{dQ}1_A\right]=E\left[\frac{dP_{\mid\mathcal{G}'}}{dQ_{\mid\mathcal{G}'}}1_A\right]$$ But under what measure is the expectation on the RHS? And how do I know that $\frac{dP_{\mid\mathcal{G}'}}{dQ_{\mid\mathcal{G}'}}$ even exists? I should show that first right?

Anyway, it seems that I'm a bit lost on this very simple measure theory problem, so I'd appreciate any help.

  • $\begingroup$ I suppose that you refer to $\frac{d P_{|\mathcal{G}'}}{d Q_{|\mathcal{G}'}}$ instead of $\frac{d P_{|\mathcal{G}'}}{d P_{|\mathcal{G}'}}$. $\endgroup$ – user178826 Oct 18 '16 at 17:26
  • $\begingroup$ Yes of course. I corrected, thanks. $\endgroup$ – Augustin Oct 18 '16 at 17:49
  • $\begingroup$ $P_{|\mathcal G'}$ is the restriction of $P$ to $\mathcal G'$. $\endgroup$ – John Dawkins Oct 18 '16 at 19:11
  • $\begingroup$ Isn't it the conditional probability? $\endgroup$ – Augustin Oct 18 '16 at 19:34
  • $\begingroup$ Because this property is used page 84 for the EM algorithm, and it seems to me that in this context, it is a conditional probability, given the observations. $\endgroup$ – Augustin Oct 18 '16 at 19:43

You are given two probability measures $P,Q$ on the measurable space $\left(\Omega,\mathcal{G}\right)$ s.t. $P\ll Q$ and a $\sigma$-algebra $\mathcal{G'}$ s.t. $\mathcal{G'}\subset\mathcal{G}$.

$P_{|\mathcal{G'}}$ is simply the restriction of $P$ to $\mathcal{G'}$, that is $$P_{|\mathcal{G'}}:\mathcal{G'}\rightarrow [0,1]$$ $$A\mapsto P(A)$$

According to the Radon-Nikodým-theorem, the density of $P$ with respect to $Q$ exists, is unique and denoted by $\frac{dP}{dQ}$. Now, how can one compute the density $\frac{dP_{|\mathcal{G'}}}{dQ_{|\mathcal{G'}}}=\frac{dP}{dQ}_{|\mathcal{G'}}$?

We have $\frac{dP}{dQ}_{|\mathcal{G'}}=E_Q[\varphi|\mathcal{G'}]$, where $\varphi$ is the density of $P$ with respect to $Q$: Let $A\in\mathcal{G'}$. Then $Q(A)=\int_A \varphi dQ=\int_A E[\varphi|\mathcal{G'}] dQ$.

It follows that $E_Q[\frac{dP}{dQ}_{|\mathcal{G'}} 1_A]=E_Q[E_Q[\varphi|\mathcal{G'}]1_A]=E_Q[\varphi 1_A]$


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.