I'm trying to understand the solution to the following question:

Let X and Y be independent random variables, uniformly distributed on the interval $[0,1]$. Since $X \neq 0 $ almost surely, the random variable $Z=\frac{Y}{X}$ is well defined.

Compute $P(X < x | \sigma(Y)) $ and $ P(X < x | \sigma(Z)) $.

How do you calculate a conditional probability in the case where you are conditioning on a sigma algebra? How is the answer below obtained?

$$P(X < x | \sigma(Y)) = \min\{x,1\} $$ $$P(X < x | \sigma(Z)) = \min\{x^2,1\} I_{\{ Z \leq 1 \}} + \min\{xZ^2,1\}I_{\{ Z \geq 1\}} $$

  • $\begingroup$ Ah yes - I've corrected it now. Thanks. $\endgroup$ – kw3rti Jan 19 '17 at 19:40

Probability of an event B given a sigma algebra $\mathcal{F}$ is a random variable defined as $$P(B|\mathcal{F}) = E(I_B|\mathcal{F})$$ where $I_B$ is the indicator function.

In your case, we would have $$P(X<x|\sigma(Y))=E(I\{X<x\}|Y).$$

To get the conditional expectation, we use the following method (described in more detail here (A.2)):

We see how the expectation of $I\{X<x\}$ depends on $Y$ by calculating $$E(I\{X<x\}|Y=y)=P(X<x|Y=y)=P(X<x)=min \{x,1\}$$ for $x\geq0$. In the last step we omitted the condition, as $X$ and $Y$ are independent.

Now we should just replace $y$ with $Y$ in the last expression to get $E(I\{X<x\}|Y)$. However, there is no $y$ in the expression, because of independence, so we just have $E(I\{X<x\}|Y)=min \{x,1\}$

The second problem is quite harder. Similarly, you need to calculate $P(X<x|Z=z)$. To do that, I think you could use the Bayes' theorem for continuous random variables, as stated here, to get the conditional density $f_{X|Z=z}(x)$. Then, $P(X<x|Z=z)=\int_0^xf_{X|Z=z}(x)dx$ which should evaluate to $min\{x^2,1\}I_{\{z\leq1\}}+min\{xz^2,1\}I_{\{z>1\}}$. Finally, replacing $z$ with $Z$ you get the wanted solution.

  • $\begingroup$ Could you go straight to $P(X \leq x | \sigma (Y)) = P(X \leq x ) $ if X and Y are independent? Also in the case of $Z$, how do you deal with the fact that it is expressed in terms X and Y? Could you give me another hint? $\endgroup$ – kw3rti Jan 19 '17 at 19:51
  • $\begingroup$ You can go straight to $P(X<x)$ and that is justified by the calculation. You can prove the general case (for $X,Y$ independent $E(X|Y)=E(X)$) in the same manner as I did here. I'll get back to you regarding $Z$ a bit later when I have the time. $\endgroup$ – Blaza Jan 19 '17 at 20:08
  • $\begingroup$ I'm terribly sorry, I now see that my solution was incorrect. The law of total expectation would give the expectation of the random variable we want (the conditional probability), not the variable itself. I now corrected my answer and gave method to get the variable. $\endgroup$ – Blaza Jan 19 '17 at 21:01

The first identity is direct since $(X,Y)$ is independent hence $P(X<x\mid Y)=P(X<x)$ almost surely, end of story.

To show the second identity, since every distribution involved, conditional or not, has a PDF, a rather straightforward method is to compute the conditional PDF $f_{X\mid Z}$. This requires to know the joint PDF $f_{X,Z}$ and the marginal PDF $f_Z$, then $$f_{X\mid Z}(x\mid z)=\frac{f_{X,Z}(x,z)}{f_Z(z)}$$ and, by definition, $P(X<x\mid Z)=g_x(Z)$ where, for every $z$, $$g_x(z)=\int_0^xf_{X\mid Z}(\xi\mid z)d\xi$$ Sooo... to compute $f_{X,Z}$, we apply the classical Jacobian approach to the change of variable $(x,y)\to(x,z)=(x,y/x)$, which is such that $dxdy=xdxdz$ on the $(x,y)$-domain $0\leqslant x,y\leqslant1$, which is the $(x,z)$-domain $0\leqslant x\leqslant1$, $0\leqslant z\leqslant1/x$, hence $$f_{X,Z}(x,z)=x\mathbf 1_{0\leqslant x\leqslant1}\mathbf 1_{0\leqslant z\leqslant1/x}=x\mathbf 1_{z\geqslant0}\mathbf 1_{0\leqslant x\leqslant\min\{1,1/z\}}$$ Thus, $$f_Z(z)=\int_\mathbb Rf_{X,Z}(x,z)=\mathbf 1_{z\geqslant0}\int_0^{\min\{1,1/z\}}xdx=\tfrac12\min\{1,1/z\}^2\mathbf 1_{z\geqslant0}$$ and, for $x\geqslant0$, $$\int_0^xf_{X,Z}(\xi,z)d\xi=\mathbf 1_{z\geqslant0}\int_0^{\min\{1,1/z,x\}}\xi d\xi=\tfrac12\min\{1,x,1/z\}^2\mathbf 1_{z\geqslant0}$$ Dividing those two yields, for $z\geqslant0$, $$g_x(z)=\frac{\min\{1,x,1/z\}^2}{\min\{1,1/z\}^2}$$ hence, for every $x\geqslant0$, $$P(X<x\mid Z)=\frac{\min\{1,x,1/Z\}^2}{\min\{1,1/Z\}^2}=\frac{\min\{1,xZ,Z\}^2}{\min\{1,Z\}^2}$$ which is equivalent to the identity in your question.

  • 1
    $\begingroup$ (+1) Nice to see how such calculations are done. Why the downvote? $\endgroup$ – user144921 Jan 20 '17 at 22:19

Conditioning on the sigma algebra $\mathcal{G}$, a conditional probability is defined to be a $\mathcal{G}$ measurable function satisfying \begin{align} E(P(A \mid \mathcal{G}) \mathbb{1}_{G}) = P(A \cap G) \end{align} for any $G \in \mathcal{G}$.

Probabilities conditioned on sigma algebras are a bit tricky in the sense that you don't directly calculate them. Rather, you guess a $\mathcal{G}$ measurable function and verify that it satisfies the above condition. For example,

Let $G \in \sigma(Y)$. Then \begin{align*} E(\min(x,1) \mathbb{1}_G) &= E(\min(x,1)) \cdot E(\mathbb{1}_G) \\ &= (\mathbb{1}_{\{x \leq 1\}}x + \mathbb{1}_{\{x > 1\}}) \cdot P(G) \\ &= P(X < x) \cdot P(G) \\ &= P(\{X < x\} \cap G) \end{align*} where the first and last equalities follow by independence. Thus $P(X < x \mid \sigma(Y)) = \min(x,1)$.

Similarly for the second conditional probability, you need only to show that \begin{align*} E\Big((\min(x^2,1)\mathbb{1}_{\{Z\leq 1\}}+\min(x\cdot Z^2,1)\mathbb{1}_{\{Z\geq 1\}}) \mathbb{1}_G\Big) = P(\{X < x\} \cap G) \end{align*} for any $G \in \sigma(Z)$.

  • $\begingroup$ Correct me if I'm wrong, but the definition I found for the conditional probability is equivalent to this one, by the definition of conditional expectation? However, the conditional expectation definition I used does give a way of directly calculating this conditional probability, without knowing the solution in advance. $\endgroup$ – Blaza Jan 19 '17 at 18:51

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.