I'm interested in the following type of case: there are 'n' continuous random variables which must sum to 1. What then would be the PDF for any one individual such variable? So, if $n=3$, then I am interested in the distribution for $\frac{X_1}{X_1+X_2+X_3}$, where $X_1, X_2$, and$ X_3 $are all uniformly distributed. The mean of course, in this example, is $1/3$, as the mean is just $1/n$, and though it is easy to simulate distribution in R, I do not know what the actual equation for the PDF or CDF is.

This situation is related to the Irwin-Hall distribution (https://en.wikipedia.org/wiki/Irwin%E2%80%93Hall_distribution). Only Irwin-Hall is the distribution of the sum of n uniform random variables, whereas I would like the distribution for one of n uniform r.v's divided by the sum of all n variables. Thanks.

  • 1
    $\begingroup$ If the $n$ continuous uniform random variables sum to $1$, then with $n=3$, $X_1+X_2+X_3 = 1$ and so the distribution of $\frac{X_1}{X_1+X_2+X_3} = X_1$ is the same as the distribution of $X_1$, right? $\endgroup$ – Dilip Sarwate Oct 5 '15 at 4:14
  • 1
    $\begingroup$ I should correct myself: the N uniform distributions don't sum to 1. I am assuming they are each uniform between 0 and 1, and so their sum may be anything from 0 to N. I am thinking of taking each uniform variable and dividing it by the sum of all N uniform variables to get a set of N random variables which sum to 1 and have expected value 1/N. Note: I removed the word 'uniform' from my first sentence. The distribution I'm looking for isn't uniform, but is derived from dividing one of N uniform variables by the sum of all N uniform variables, somehow. I'm just not sure how. $\endgroup$ – user3593717 Oct 5 '15 at 4:35
  • $\begingroup$ Where the $X_i$ are exponentially distributed, the vector of normalised variables has a Dirichlet distribution. This may be of interest in itself, but looked into might also provide tactics for this type of situation. $\endgroup$ – conjectures Oct 14 '15 at 22:40

The breakpoints in the domain make it somewhat messy. A simple but tedious approach is to build up to the final result. For $n=3,$ let $Y=X_2 + X_3,$ $W = {{X_2 + X_3} \over X_1},$ and $T = 1 + W.$ Then $Z = {{1} \over {T}}={{X_1} \over {X_1 + X_2 + X_3}}.$

The breakpoints are at 1 for $Y,$ 1 and 2 for $W,$ 2 and 3 for $T,$ and $1/3$ and $1/2$ for $Z.$ I found the complete pdf to be

$$f(z) = \begin{cases} \ \ \ \ \ {{1} \over {(1-z)^2}} \ , & \text{if} \ {0} \leq z \leq {1/3} \\\\ {{3z^3-9z^2+6z-1} \over {3z^3(1-z)^2}} \ , & \text{if} \ {1/3} \leq z \leq {1/2} \\\\ \ \ \ \ \ \ \ {{1-z} \over {3z^3}} \ , & \text{if} \ {1/2} \leq z \leq {1} \end{cases}$$

The cdf can then be found as $$F(z) = \begin{cases} \ \ \ \ \ \ \ \ \ \ \ {{z} \over {(1-z)}} \ , & \text{if} \ {0} \leq z \leq {1/3} \\\\ {{1} \over {2}}+{{-18z^3+24z^2-9z+1} \over {6z^2(1-z)}} \ , & \text{if} \ {1/3} \leq z \leq {1/2} \\\\ \ \ \ \ \ \ \ \ {{5} \over {6}} + {{2z-1} \over {6z^2}} \ , & \text{if} \ {1/2} \leq z \leq {1} \end{cases}$$

  • $\begingroup$ +1 Nice. Also, your density agrees beautifully with simulation. $\endgroup$ – Glen_b Oct 14 '15 at 23:43

Let $Y=\sum_{i=2}^n X_i$. We can find the cdf of $X_1/\sum_{i=1}^n X_i$ by calculating \begin{align*} P(\frac{X_1}{\sum_{i=1}^n X_i} \leq t) &= P(X_1 \leq t\sum_{i=1}^n X_i) \\ &= P((1-t)X_1 \leq t\sum_{i=2}^n X_i) \\ &= P(X_1 \leq \frac t{1-t}Y)\\ &= \int_0^1 P(x_1 \leq \frac t{1-t}Y)\ dx_1\\ &= \int_0^1 (1-F_Y(\frac{1-t}{t}x_1))\ dx_1\\ &= 1-\int_0^1 F_Y(\frac{1-t}{t}x_1)\ dx_1\\ \end{align*} We then differentiate and substitute the Irwin-Hall pdf to obtain the desired pdf: \begin{align*} f(t) &= \int_0^1 f_Y(\frac{1-t}{t}x_1)\cdot \frac{x_1}{t^2}\ dx_1\\ &= \frac{1}{t^2}\int_0^{1\wedge \frac{(n-1)t}{1-t}} \sum_{k=0}^{\lfloor \frac{1-t}{t}x_1\rfloor}\frac1{(n-2)!}(-1)^k\binom{n-1}k(\frac{1-t}{t}x_1-k)^{n-1} x_1\ dx_1 \end{align*} From here it gets a little messy, but you should be able to interchange the integral and summation and then perform a substitution (e.g, $u=\frac{tx_1}{1-t}-k$) to evaluate the integral and hence obtain an explicit formula for the pdf.



"the N uniform distributions don't sum to 1."

This is how I started(it's incomplete):

Consider $Y = \sum_{i=1}^n X_i$ and let $X=X_i$ by a slight abuse of notation.

Consider, $U = \frac{X}{Y}$ and $V =Y$:

$$ X=UV\\ Y=V $$

Then following transformation of variables:

$$ J = \begin{bmatrix} V & U\\ 0 & 1 \end{bmatrix} $$

The joint probability function of $(U,V)$ is given by:

$f_{U,V}(u,v) = f_{X,Y}(uv,v)|J|$

Where $X \sim U(0,1)$ and $Y \sim IrwinHall$

$$ f_X(x) = \begin{cases} 1 & 0 \leq x\leq 1\\ 0 & otherwise \end{cases} $$

And, $$ f_Y(y) = \frac{1}{2(n-1)!}\sum_{k=0}^n(-1)^k {n\choose k}(x-k)^{n-1} sign(x-k) $$

Thus, $$ f_{U,V}(u,v) = \begin{cases} \frac{1}{2(n-1)!}\sum_{k=0}^n(-1)^k {n\choose k}(uv-k)^{n-1} sign(uv-k) & 0 \leq uv \leq 1\\ 0 & otherwise \end{cases} $$

and $f_U(u) = \int f_{U,V}(u,v) dv$


Suppose we already know sum of $U(0,1)$ has a Irwin-Hall distribution. Now your question changes to find the pdf (or CDF) of $\frac{X}{Y}$ when X had a $U(0,1)$ distribution and $Y$ has a Irwin-Hall distribution.

First we need to find he joint pdf of $X$ and $Y$.

Let $Y_1=X_1\\Y_2=X_1+X_2\\Y_3=X_1+X_2+X_3$




$J=\begin{vmatrix} \frac{\partial X_1}{\partial Y_1} & \frac{\partial X_1}{\partial Y_2} &\frac{\partial X_1}{\partial Y_3} \\ \frac{\partial X_2}{\partial Y_1} & \frac{\partial X_2}{\partial Y_2} &\frac{\partial X_2}{\partial Y_3} \\ \frac{\partial X_3}{\partial Y_1} & \frac{\partial X_3}{\partial Y_2} &\frac{\partial X_3}{\partial Y_3} \end{vmatrix}=-1$

Since $X_1, X_2, X_3$ are i.i.d with $U(0,1),$ therefore, $f(x_1,x_2,x_3)=f(x_1)f(x_2)f(x_3)=1$

The joint distribution with $y_1,y_2,y_3$ is


Next let us integrate out the $Y_2$ and we can get the joint distribution of $Y_1$ and $Y_3$ i.e the joint distribution of $X_1$ and $X_1+X_2+X_3$

As suggested by whuber now I changed the the limits

$$h(y_1,y_3)=\int_{y_1+1}^{y_3-1} g(y_1,y_2,y_3)dy_2=\int_{y_1+1}^{y_3-1} 1 dy_2=y_3-y_1-2 \tag{1}$$

Now, we know the joint pdf of $X,Y$ i.e joint pdf $X_1$ and $X_1+X_2+X_3$ is $y_3-y_1-2$.

Next let find the pdf of $\frac{X}{Y}$

We need another transformation:

Let $Y_1=X\\Y_2=\frac{X}{Y}$

Then $X=Y_1\\Y=\frac{Y_1}{Y_2}$


$J=\begin{vmatrix} \frac{\partial x}{\partial y_1} & \frac{\partial x}{\partial y_2}\\ \frac{\partial y}{\partial y_1} & \frac{\partial y}{\partial y_2} \end{vmatrix}= \begin{vmatrix} 1 & 0\\ \frac{1}{y_2} & -\frac{y_1}{y_2^2} \end{vmatrix}=-\frac{y_1}{y_2^2}$

we already the joint distribution of $X,Y$ from above steps ref (1).



Next, we integrate the $y_1$ out we get the pdf of $y_2$ then we get the pdf of $\frac{X}{Y}$


This is the pdf of $X/Y$ i.e $\frac{X_1}{X1+X_2+X_3}$

We are not finish yet, what is $y_3$ in (2) then?

We know that $Y_3=X_1+X_2+X_3$ from the first transformation.

So at least we know $Y_3$ has a Irwin-Hall distribution.

I wonder can we plug the Irwin-Hall for $n=3$ pdf to (2) to get a explicit formula? or can we do some simulations from here as Glen suggested?

  • 2
    $\begingroup$ Simulation doesn't seem to agree with that pdf. $\endgroup$ – Glen_b Oct 5 '15 at 7:17
  • $\begingroup$ The logic and steps seem correct, but I feel uncomfortable about this solution. $\endgroup$ – Deep North Oct 5 '15 at 12:24
  • 2
    $\begingroup$ Where you integrated out $y_2$, you needed to account for the conditions $y_1\le y_2\le y_3$ and $y_3-1 \le y_2\le y_1+1$. $\endgroup$ – whuber Oct 5 '15 at 14:05

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.