Let $X$ be a random variable on a probability space $(\Omega,\Sigma,\mathbb{P})$.

Let $F_X$ denote the probability distribution function of $X$ given by:

$$F_X(y) = P(X\leq y) \text{ for } y \in \mathbb{R}$$

Then, I came across the following:

Almost everywhere differentiability of $F_X$ does not imply that the probability density function of $X$, denoted by $f_X$, exists. For example, if $X$ is a random variable such that $P(X = \frac{1}{2}) = 1$, then $F_X$ is differentiable almost everywhere but $f_X$ does not exist.

What is the reason that $f_X$ does not exist for $X$ such that $P(X = \frac{1}{2}) = 1$?


If the density of $X$ is $f$, then $$ \int_{(-\infty,1/2)\cup(1/2,+\infty)} f(x)\,dx = 0 $$ and $$ \int_{\{1/2\}} f(x)\,dx = 1. \tag 1 $$ The problem is the that integral of a function over and interval of length $0$ cannot be anything but $0$, but $(1)$ says it must be $1$.

Even if $F_X$ is continuous everywhere and differentiable almost everywhere, there may be no density. The Cantor distribution is a standard example of that. One way to describe the Cantor distribution is that it's the distribution whose base-$3$ expansion is $0.d_1 d_2 d_3 \ldots$ where $d_i$, $d_i=1,2,3,\ldots$ is $0$ or $2$, each with probability $1/2$, and these digits are independent.

  • $\begingroup$ @hermes He uses this to derive a contradiction. $\endgroup$ – Eric Auld Aug 15 '15 at 6:01
  • $\begingroup$ Actually, probability is separate on discrete and continuous case. On continuous case probability is always $0$ on single point. $\endgroup$ – Math Wizard Aug 15 '15 at 6:13

It would have to be a function such that $\int_0^a f_X = 0$ for all $a<1$ and $\int_0^a f_X = 1$ for all $a\geq 1$. No such function exists (only the Dirac delta "function")


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.