Let $X$ be a random variable on $\mathbb{R}$ and $Y$ another random variable, independent from $X$ and also defined on $\mathbb{R}$. Let $F_X$ be the CDF of $X$. I'm interested in calculating the distribution of $F_X(Y) = \mathbb{P}(X \leq Y \mid Y)$. Is there a general way of calculating this?

We know that if $X$ and $Y$ are continuous and identically distributed, $F_X(Y) \sim Uniform[0, 1]$, but what about in general? Or maybe, if we look at a simplified example:

Let $X \sim N(0,1)$, $Y \sim N(\mu, 1)$, $X$ independent of $Y$, what is the distribution of $F_X(Y)$?

On this particular example, I did some simulations in R programming language and I have a sense that $F_X(Y)$ is Beta distributed, but couldn't find a way to prove it. I also obtained similar results when $X$,$Y$ were Chi-squared, Exponential, or Weibull distributed.

Any help or further literature is greatly appreciated!

  • $\begingroup$ For starters, $F_X(Y)$ takes on values in $(0,1)$ or $(0,1]$ or $[0,1)$ or $[0,1]$ and so it is not too surprising that Beta distributions spring to mind when doing simulations. $\endgroup$ – Dilip Sarwate Aug 24 '17 at 18:07
  • $\begingroup$ (1) By varying $X$ and $Y$ you can produce literally any distribution on $[0,1]$. (2) It is not the case that $F_X(Y)$ must be uniform when identically distributed: you must assume both $X$ and $Y$ have continuous distributions. (3) The ways of computing $F_X(Y)$ are no different than the methods available to compute any distribution function. Which ones will work depends on how you have specified $F_X$ and $F_Y$. $\endgroup$ – whuber Aug 24 '17 at 20:40
  • $\begingroup$ Thanks for your feedback! (1) - I agree; (2) True, I edited my question; (3) Yes, and Jarle's answer gave me a push further in deriving the result. $\endgroup$ – Damjan Aug 24 '17 at 21:14

The cdf of $Z=F_X(Y)$ is \begin{align} F_Z(z) &=P(Z\le z) \\&=P(F_X(Y) \le z) \\&=P(Y \le F_X^{-1}(z)) \\&=F_Y(F_X^{-1}(z)). \end{align} If $F_X$ is not continuous and strictly increasing, for example if $X$ is discrete, then $F_X^{-1}$ is the generalised inverse of $F_X$, that is, $F_X^{-1}(z)=\mathrm{inf}(y \in \mathbb R:F_X(y)\ge z)$, also known as the quantile function of $X$.

In the case that $X$ and $Y$ are both continuous and the support of $Y$ is a subset of the support of $X$, the pdf of $Z$ becomes $$ f_Z(z) = \frac d{dz}F_Z(z)=f_Y(F_X^{-1}(z))\frac d{dz}F_X^{-1}(z)=\frac{f_Y(F_X^{-1}(z))}{f_X(F_X^{-1}(z))}. \tag{1} $$ If both $X$ and $Y$ are continuous but $Y$ has larger support than $X$, for example if $X\sim\mathrm{exp}(1)$ and $Y\sim N(1,1)$, then $Z$ has a pdf given by (1) but in addition a point mass of size $P(Y\le 0)=\phi(-1)=0.1586$ in $z=0$.

R code illustrating this:

n <- 1e+4
X <- rexp(n,shape=1,scale=1)
Y <- rnorm(n,mean=1)
Z <- pexp(Y)

enter image description here

  • $\begingroup$ +1 -- but it might be worth pointing out that this derivation implicitly assumes the distributions are continuous (and not just when you differentiate: that assumption sneaks in at the point you refer to $F_X^{-1}$). $\endgroup$ – whuber Aug 24 '17 at 20:58
  • $\begingroup$ Great, thanks! So, for my example, we get $F_X^{-1} (z) = \sqrt{2} Erf^{-1} (2z-1)$, which is defined for $z \in (0,1)$ (reference: en.wikipedia.org/wiki/Normal_distribution#Quantile_function). Consequently, $F_Z(z) = \frac{1}{2}[1 + Erf[\frac{1}{\sqrt{2}}(\sqrt{2}Erf^{-1}(2z-1)-\mu)]]$, which doesn't seem like a distribution of a Beta distributed variable. Similarly, we can also calculate $f_Z(z)$. $\endgroup$ – Damjan Aug 24 '17 at 21:28

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.