3
$\begingroup$

The example below is taken from Björk (2009). Let Radon-Nikodym derivative be $$L=\frac{dP}{dQ} \;\; \text{on} \; \mathcal F$$ or written analogously $$P(A) = \int_AL(\omega)dQ(\omega) \;\; \text{for all} \; A\in \mathcal F.$$

For finite sample space $\Omega$ this simplifies to $$P(A) = \sum_{\omega\in A}L(\omega)Q(\omega).$$

Let $\Omega = \{ 1,2,3 \}$, $\mathcal F = 2^\Omega$ and $G=\left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$ and $$P(1)=1/4, \;\; P(2)=1/2, \;\; P(3)=1/4,$$ $$Q(1)=1/3, \;\; Q(2)=1/3, \;\; Q(3)=1/3.$$

We are computing $L$ on $G$. Using above formula for $A=\{1\}$ we have that $$1/4 = L(1)\cdot 1/3,$$ which leads to $L(1)=3/4.$

Proceeding $$P(\{2,3\}) = L(2)Q(2) + L(3)Q(3),$$ which results in $$L(2) + L(3) = 9/4.$$ Björk writes $L(2)=9/8$ and $L(3)=9/8$. Could you please explain why $L(\omega)$ is constant over the subsets $A\in G$?

$\endgroup$
2
$\begingroup$

Let ${\mathcal{F}} = 2^{\Omega}$ and let $\mathcal{G} =\left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$. Both are sigma-algebras of subsets of $\Omega$.

The book/question is confusing since the way you have written $P(A) = \sum_{\omega\in A}L(\omega)Q(\omega)$ makes it seem like $Q$ is a mapping from $\Omega$, but probability measures are mappings from sigma-algebras, not the sample space itself. So switching between continuous and discrete makes things more confusing in my opinion, since one is a Lebesgue Integral and the other is a normal summation.

So in the case of ${\mathcal{F}} = 2^{\Omega}$, since every $\omega \in \Omega$ is also a set $F \in \mathcal{F}$, then the expression you wrote works fine. So that is why we can write the $L^{\mathcal{F}}(\omega ) = \frac{P(\omega)}{Q(\omega)}$ expressions and evaluate them as usual. And we can see that since $\mathcal{F}$ is the power set of $\Omega$, we know exactly which $\omega \in \Omega$ occurs if we know which sets in $\mathcal{F}$ occur or do not occur. So $L^{\mathcal{F}}$ is measurable with respect to $\mathcal{F}$.

In the case of $L^{\mathcal{G}}$, the only sets in $\mathcal{G}$ are $\left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$, so given knowledge of whether those occur we need to be able to calculate $L^{\mathcal{G}}(\omega)$. So now assuming that the probability space is equipped with sigma-algebra $\mathcal{G}$, we can only calculate $P(A)$ & $Q(A)$ for $A \in \left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$.

So \begin{equation} P(\{2,3\}) = \frac{3}{4} = \int_{\{2,3\}}L^{\mathcal{G}}(\omega)dQ(\omega) = \int_{\Omega}1_{\{2,3\}}(\omega)L^{\mathcal{G}}(\omega)dQ(\omega) \end{equation} and imagine that $L^{\mathcal{G}}(\omega)$ is a simple function, we can write it as $L^{\mathcal{G}}(\omega) = L^{\mathcal{G}}(1)1_{\{1\}}(\omega) + L^{\mathcal{G}}(2)1_{\{2\}}(\omega) + L^{\mathcal{G}}(3)1_{\{3\}}(\omega)$ and so, defining $S_k = \{\omega |L^{\mathcal{G}}(\omega) = L^{\mathcal{G}}(k)\}$ \begin{equation} \int_{\Omega}1_{\{2,3\}}(\omega)L^{\mathcal{G}}(\omega)dQ(\omega) = \sum_{k = {1,2,3}}L^{\mathcal{G}}(k)Q(\{2,3\} \cap S_k)\\ = L^{\mathcal{G}}(1)Q(\emptyset) + L^{\mathcal{G}}(2)Q({2}) + L^{\mathcal{G}}(3)Q({3}) \end{equation} So now we see the issue of $Q$ being only defined on $\left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$. So the only way for the expression to make sense and thus equal $P(\{2,3\}) = \frac{3}{4}$, we need for $L^{\mathcal{G}}(\omega)$ as a simple function to be written: $L^{\mathcal{G}}(\omega) = L^{\mathcal{G}}(1)1_{\{1\}}(\omega) + L^{\mathcal{G}}(2)1_{\{2,3\}}(\omega) = L^{\mathcal{G}}(1)1_{\{1\}}(\omega) + L^{\mathcal{G}}(3)1_{\{2,3\}}(\omega)$, since then we'd have: \begin{equation} \int_{\Omega}1_{\{2,3\}}(\omega)L^{\mathcal{G}}(\omega)dQ(\omega) = \sum_{k = {1,2}}L^{\mathcal{G}}(k)Q(\{2,3\} \cap S_k)\\ = L^{\mathcal{G}}(1)Q(\emptyset) + L^{\mathcal{G}}(2)Q(\{2,3\}) \end{equation} which is a valid expression.

So finally $\frac{3}{4} = L^{\mathcal{G}}(2)Q(\{2,3\}) = L^{\mathcal{G}}(2) \frac{2}{3} $, so $L^{\mathcal{G}}(2) = L^{\mathcal{G}}(3) = \frac{9}{8}$

So the book and this answer is basically a long, convoluted way to say that since $L^{\mathcal{G}}(\omega)$ must be $\mathcal{G}$-measurable, then the only way to calculate $L^{\mathcal{G}}(2)$ and $L^{\mathcal{G}}(3)$ with only the knowledge of whether sets in $\left\{ \Omega, \emptyset, \{1\}, \{2,3\} \right\}$ occur, is if $L^{\mathcal{G}}(2) = L^{\mathcal{G}}(3)$. This is since when $\{2,3\}$ occurs, we do not know which of $2$ or $3$ occurs, just that one of them does. So the only way to know $L^{\mathcal{G}}(2)$ and $L^{\mathcal{G}}(3)$ is if they are equal. Then we can say that when $\{2,3\}$ occurs that $L^{\mathcal{G}}(\omega \in \{2,3\}) = L^{\mathcal{G}}(2) = L^{\mathcal{G}}(3)$

$\endgroup$
  • $\begingroup$ I think you are calculating $L$ on $\mathcal F$ while I on $G$. $\endgroup$ – tosik Apr 1 at 6:22
  • $\begingroup$ I am taking another look. Give me a bit $\endgroup$ – Slade Apr 1 at 17:08

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.