Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Choose a random number uniformly over the unit interval $x_1 \in [0,1]$. Now chose a second number that is dependent on the first, such that it is chosen uniformly over the interval $x_2 \in [0, x_1]$. What is the probability distribution $x_2$? Is it a distribution that is already well-known?

Here's what I got so far - I examined small discrete cases, say with the integers $\{1,2,3\}$. In this case the probabilities work out to:

\begin{array}{rcl} P(1,3) &=& (1/3)(1/1 + 1/2 + 1/3) \cr P(2,3) &=& (1/3)(1/2 + 1/3) \cr P(3,3) &=& (1/3)(1/3) \end{array}

This generalizes to something like $$ P(n,k) = \frac{1}{k} \sum_{q=n}^k \frac{1}{q} $$ which can be reduced to the Digamma function $$ = \frac{\Psi(k+1) - \Psi(n)}{k} $$

because of this, I would not be surprised if the answer was some combination of gamma functions. I'm stuck here, since I'm not quite sure how if taking the limit of large $k$ will (upon normalization) give the answer I'm looking for.

share|cite|improve this question
up vote 4 down vote accepted

$x_2$ is equal in distribution to $U_1 U_2$ where $U_1$ and $U_2$ are independent uniform$(0,1)$ random variables. It is very easy to find the distribution of $U_1 U_2$...

EDIT: Continuing this iteratively, so that $x_{n+1}$ is uniform on $(0,x_n)$, the infinite sum $x:=x_1+x_2+x_3+\cdots$ (which is finite with probability $1$, by the monotone convergence theorem) is equal in distribution to $U_1 + U_1 U_2 + U_1 U_2 U_3 + \cdots$, where the $U_i$ are independent uniform$(0,1)$ rv's. The distribution of $x$ is the Dickamn distribution.

EDIT: Here are three ways to compute the distribution function of $x_2$ (the first one is the direct approach). It is given, for $0 < x \leq 1$, by $F(x)=x - x \log x$; hence $x_2$ has probability density function $f(x)=-\log(x)$, $0 < x < 1$ (as leonbloy already found).

Approach 1): Since conditioned on $x_1 = s$, $0 < s < 1$, $x_2$ is uniformly distributed on $(0,s)$, the law of total probability gives, for $0 < x \leq 1$, $$ {\rm P}(x_2 \le x) = \int_0^1 {{\rm P}(x_2 \le x|x_1 = s)ds} = \int_0^x {{\rm P}(x_2 \le x|x_1 = s)ds} + \int_x^1 {{\rm P}(x_2 \le x|x_1 = s)ds} $$ $$ = \int_0^x {1ds} + \int_x^1 {\frac{x}{s}ds} = x - x\log x. $$

Approach 2): $x_2$ is distributed as $U_1 U_2$, where $U_1$ and $U_2$ are independent uniform$(0,1)$ rv's. Hence, $$ {\rm P}(x_2 \le x) = {\rm P}(U_1 U_2 \le x) = \int_0^1 {{\rm P}(U_1 U_2 \le x|U_1 = s)ds} = \int_0^1 {P\bigg(U_2 \le \frac{x}{s}\bigg)ds} $$ $$ = \int_0^x {P\bigg(U_2 \le \frac{x}{s}\bigg)ds} + \int_x^1 {P\bigg(U_2 \le \frac{x}{s}\bigg)ds} = \int_0^x {1ds} + \int_x^1 {\frac{x}{s}ds} = x - x \log x. $$ (So, approaches 1) and 2) are quite similar: both rely on the law of total probability.)

Approach 3): Let $U_i$, $i=1,2$, be as above, and note that $-\log U_i$ is exponential$(1)$. Then, for any $0 < x \leq 1$, $$ {\rm P}(x_2 \le x) = {\rm P}(U_1 U_2 \le x) = {\rm P}(\log U_1 + \log U_2 \le \log x) = {\rm P}( - \log x \le \xi_1 + \xi_2 ), $$ where $\xi_1$ and $\xi_2$ are independent exponential$(1)$ rv's. The random variable $\xi_1+\xi_2$ has gamma density function $ye^{-y}$, $y > 0$. Hence, $$ {\rm P}(x_2 \le x) = \int_{ - \log x}^\infty {ye^{ - y} dy} = -e^{-y}(y+1) \big|_{ - \log x}^\infty = x - x \log x. $$ This approach can be useful for determining the distribution of $U_1 \cdots U_n$, where $U_i$ are independent uniform$(0,1)$ rv's.

share|cite|improve this answer

It can be interesting to attack from problem from a discrete analog; but it's not a very simple way here. What your problem statement gives you directly is the conditional probability $P(x_2 | x_1)$ and the marginal $P(x_1)$. Then...

$$P(x_2) = \int P(x_2,x_1) dx_1 = \int P(x_2 | x_1) P(x_1) dx_1$$

You have $P(x_1)=1$ in $[0,1]$ and $P(x_2 | x_1) = 1/x_1 $ with $0 \le x_2 \le x_1 \le 1$, so you get $P(x_2) = - \log(x_2)$

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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