4
$\begingroup$

Each of the scalar random variables, $ Y $, $ X $, $ U $, and $ V $, is continuous and possibly has $ \mathbb{R} $ as its support. The random variable, $Z$, could be vector valued, but continuous.

I have the following \begin{align}\nonumber Y = \beta X + U\\\nonumber X = \pi Z + V, \end{align} where $ Z \perp (U,V) $ ( $ \perp $ denotes independence). But $ U\not\perp X$. In the second equation, given $ Z $, there is one-to-one mapping between $ X $ and $ V $.

Given the set-up, I need to know whether the following conditional independence holds: $ U\perp X\mid V $.

My question is: since, given $ Z $, there is one-to-one mapping between $ X $ and $ V $, is it true that $ \sigma(X, Z) = \sigma(V, Z) $, where $ \sigma(X, Z) $ is the sigma algebra generated by $ (X,Z) $ and $ \sigma(V, Z) $ that generated by $ (V,Z) $?

If $ \sigma(X, Z) = \sigma(V, Z) $, then I could write $ U\mid X,Z \sim U\mid V,Z $. But since $ Z \perp (U,V) $, I could write $ U\mid V,Z \sim U\mid V $. From this could we deduce that $ U\perp X\mid V $.

$\endgroup$
0
$\begingroup$

The answer is yes. Indeed, $Z\perp(U,V)$ implies $\pi Z\perp(U,V)$. So, without loss of generality $\pi Z=Z$ and $X=Z+V$ (the condition $Y=\beta X+U$ is irrelevant and not needed here). So, the desired conditional independence $(U\perp X)|V$ can be rewritten as $(U\perp Z+V)|V$, which means that \begin{equation} E\Big(\big(f(U)g(Z+V)\big)|V\Big)\overset{\text{(?)}}=E(f(U)|V)\times E\big(g(Z+V)|V\big) \tag{-1} \end{equation} almost surely (a.s.) for all nonnegative Borel-measurable functions $f,g$, that is, \begin{equation*} E[f(U)g(Z+V)h(V)]\overset{\text{(?)}}=E[E\big(f(U)|V\big)\,E\big(g(Z+V)|V\big)\,h(V)]; \tag{0} \end{equation*} here and in what follows, $f,g,h$ are any nonnegative Borel-measurable functions.

By the condition $Z\perp(U,V)$, \begin{equation} \begin{aligned} E[f(U)g(Z+V)h(V)]&=\iint P(U\in du,V\in dv)f(u)h(v)\tilde g(v) \\ &=E[f(U)h(V)\tilde g(V)], \end{aligned} \tag{1} \end{equation} where \begin{equation*} \tilde g(v):=\int P(Z\in dz)Eg(z+v)=Eg(Z+v). \end{equation*} In particular, (1) with $f=1$ implies that \begin{equation*} E\big(g(Z+V)|V\big)=E\tilde g(V) \end{equation*} a.s. So, by the definition of the conditional expectation, the right-hand side of (0) equals the last expression in (1). Thus, (0) is proved.


If the random variables (r.v.'s) $U,V,Z$ are discrete, then the proof can be written in a simpler and more transparent way, as follows. Let us use notations such as $p_V$ to denote the probability mass function (pmf) of the r.v. $V$; $p_{U,V}$ to denote the joint pmf of the r.v.'s $U,V$; $p_{U,X|V}$ to denote the joint conditional pmf of the r.v.'s $U,X$ given $V$; etc. Then the desired conditional independence $(U\perp X)|V$ can be rewritten as $p_{U,X|V}=p_{U|V}p_{X|V}$ or, equivalently, as \begin{equation*} \frac{p_{U,X,V}}{p_V}\overset{\text{(?)}}=\frac{p_{U,V}}{p_V}\frac{p_{X,V}}{p_V}. \tag{0'} \end{equation*} Since $X=Z+V$ and $Z\perp(U,V)$, we have $p_{U,X,V}(u,x,v)=p_{U,V}(u,v)p_Z(x-v)$ and $p_{X,V}(x,v)=p_V(v)p_Z(x-v)$ for all $u,x,v$, whence (0') follows.

$\endgroup$
  • $\begingroup$ Many Thanks, Losif, for the reply. First, as I can understand, to show the desired conditional independence, one has to show the following: \begin{equation*}\nonumber E(f(U)g(Z+V)|V)\overset{\text{(?)}}=E(f(U)|V)\,E(g(Z+V)|V) \end{equation*} almost surely (a.s.) for all nonnegative Borel-measurable functions $f,g$. Immediately after the above you write, ``that is, \begin{equation*} Ef(U)g(Z+V)h(V)\overset{\text{(?)}}=EE(f(U)|V)E(g(Z+V)|V)h(V); \tag{0} \end{equation*} here and in what follows, $f,g,h$ are any nonnegative Borel-measurable functions." $\endgroup$ – user_akt Feb 24 at 23:37
  • $\begingroup$ I do not understand why is $ h(V) $ is required here. Secondly, it would easier for me to understand if there are parenthesis along with the expectation operator in (0). Also, which equation is equation (1), in which I have to assume $ f =1 $. I shall be grateful, if you could reply. $\endgroup$ – user_akt Feb 24 at 23:37
  • $\begingroup$ The equivalence of formulas (-1) and (0) follows right by the definition of the conditional expectation; cf. e.g. the 1st display in section "Conditional expectation with respect to a random variable" at en.wikipedia.org/wiki/…; a condition guaranteeing that both integrals in that display exist is missing there (to that end, it was required in my answer that $f,g,h$ be nonnegative). $\endgroup$ – Iosif Pinelis Feb 25 at 1:05
  • $\begingroup$ Previous comment continued: The roles of $X$, $Y$, $f$, and $g(Y)$ on that Wikipedia page are played, respectively, by $f(U)g(Z+V)$, $V$, $h$, and the right-hand side of formula (-1) in the transition from (-1) to (0) in my answer. I have also added some parentheses and/or brackets after the expectations sgins. Please let me know if something is still unclear. As for the missing tag (1), I guess it was due to my mistaken use of MathJax. Now this tag is restored. $\endgroup$ – Iosif Pinelis Feb 25 at 1:08
  • $\begingroup$ I have also added some parentheses and brackets after the expectation signs, as you requested (they are not needed, though, according to a standard convention). $\endgroup$ – Iosif Pinelis Feb 25 at 1:14

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.