1
$\begingroup$

Consider $P(X,Y)$ discrete and $Z = f(Y)$ with $f$ deterministic. The function $f$ identifies a partition of the elements of the alphabet $\mathcal{Y}$ of $Y$. Each outcome $z \in \mathcal{Z}$ is a subset $z \subseteq \mathcal{Y}$. Let $Z' = f'(Y)$ be identical to $Z$ except for two elements $z_1', z_2' \in \mathcal{Z}'$ that are derived from $z_1, z_2 \in Z$ by "moving" one element $\bar{y} \in z_1$ from $z_1$ to $z_2$. This means that $z_1' = z_1 \setminus \{\bar{y}\}$ and $f'(\bar{y}) = z_2' = z_2 \cup \{\bar{y}\}$. Does the following relationship hold?

$$ P(X=x, Y=y \mid Z=z_1) = P(X=x, Y=y \mid Z'=z_1') \cdot \frac{P(Z=z_1)}{P(Z'=z_1')} \quad \text{for} \quad y \ne \bar{y} $$

If correct, does it also imply $P(X=x, Y=y, Z=z_1) = P(X=x, Y=y, Z=z_1')$ for $y \ne \bar{y}$? If not, please, provide a counterexample.

$\endgroup$
1
  • $\begingroup$ How is the function $f'$ formally defined (without using informal terms such as "moving")? In particular, what is $f'(\bar y)$? $\endgroup$ Commented Dec 12, 2021 at 4:17

1 Answer 1

1
+50
$\begingroup$

$\newcommand{\Y}{\mathcal{Y}}\newcommand{\ZZ}{\mathcal{Z}}$It appears that for all $u\in\Y$ and all $z\in\ZZ$ we have \begin{equation*} f(u)=z\iff u\in z \tag{1} \end{equation*} and \begin{equation*} f'(u)=\begin{cases} z_1'&\text{ if }u\in z_1'=z_1\setminus\{\bar y\}, \\ z_2'&\text{ if }u\in z_2'\cup\{\bar y\}, \\ f(u)&\text{ if }u\notin z_1'\cup z_2'=z_1\cup z_2. \end{cases} \tag{2} \end{equation*}

Then the identity \begin{equation*} P(X=x,Y=y|Z=z_1)=P(X=x,Y=y|Z'=z_1')\frac{P(Z=z_1)}{P(Z'=z_1')} \tag{3} \end{equation*} does not hold in general, even if $y\ne\bar y$. Indeed (assuming $P(Z=z_1)\ne0$ and $P(Z'=z_1')\ne0$), we can rewrite (3) as \begin{equation*} P(X=x,Y=y,Z=z_1)=P(X=x,Y=y,Z'=z_1')\frac{P(Z=z_1)^2}{P(Z'=z_1')^2}. \tag{5} \end{equation*} However, \begin{equation*} P(Z=z_1)=P(f(Y)=z_1)=P(Y\in z_1)=P(Y\in z_1'\cup\{\bar y\}) =P(Y\in z_1')+P(Y=\bar y) \end{equation*} and \begin{equation*} P(Z'=z_1')=P(f'(Y)=z_1')=P(Y\in z_1'), \end{equation*} so that \begin{equation*} \text{$P(Z'=z_1')=P(Z=z_1)$ iff $P(Y=\bar y)=0$. }\tag{6} \end{equation*} Also, \begin{multline*} P(X=x,Y=y,Z=z_1)=P(X=x,Y=y,f(y)=z_1) \\ =P(X=x,Y=y)1(f(y)=z_1)=P(X=x,Y=y)\,1(y\in z_1) \end{multline*} and \begin{multline*} P(X=x,Y=y,Z'=z_1')=P(X=x,Y=y,f'(y)=z_1') \\ =P(X=x,Y=y)\,1(f'(y)=z_1')=P(X=x,Y=y)\,1(y\in z_1'), \end{multline*} so that \begin{equation*} P(X=x,Y=y,Z=z_1)=P(X=x,Y=y,Z'=z_1') \tag{7} \end{equation*} if $y\ne\bar y$. Thus, (5) does not hold for all $y\ne\bar y$ unless $P(Y=\bar y)=0$, and hence (3) does not hold for all $y\ne\bar y$ unless $P(Y=\bar y)=0$.

On the other hand, the identity \begin{equation*} P(X=x,Y=y|Z=z_1)=P(X=x,Y=y|Z'=z_1')\frac{P(Z'=z_1')}{P(Z=z_1)}, \end{equation*} which is equivalent to (7) (which also was in question), will always hold for all $y\ne\bar y$.

$\endgroup$
2
  • $\begingroup$ Thanks! I have definitely messed up the apexes in the question. Your last relationship is what I wanted to prove. $\endgroup$
    – Cesare
    Commented Dec 15, 2021 at 9:24
  • $\begingroup$ All right. Are you fully satisfied with this answer? $\endgroup$ Commented Dec 15, 2021 at 18:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .