Exponential characterization by a uniform random split (source code)

= Exponential characterization by a uniform random split

Let $X>0$ have <probability density function> $g$, and let $U$ be <independent> and uniform on $(0,1)$. Put $Y=XU$, $Z=X(1-U)$. Then $Y$ and $Z$ are <independent> if and only if they are identically distributed <exponential random variables> of some rate $c>0$, equivalently $X$ has <gamma density> $g(x)=c^2xe^{-cx}$.

To prove the nontrivial direction without substituting at a possibly singular boundary, use the <uniform split of a random total> formulas
$$
f_{Y,Z}(y,z)=\frac{g(y+z)}{y+z},\qquad
h(y)=f_Y(y)=f_Z(y)=\int_y^\infty\frac{g(t)}t\,dt.
$$
The <function> $h$ is continuous for positive $y$. <independence> forces $h(y)h(z)$ to depend only on $y+z$. Its positive-support endpoint is the same as that of $X$; if that endpoint were finite and nonzero, <independence> would make the endpoint of $Y+Z$ twice as large, a contradiction. Thus $h$ is positive at every positive argument.

Comparing equal sums gives $h(y+t)h(z)=h(y)h(z+t)$, initially <almost everywhere> and then everywhere by continuity. Hence $h(y+t)/h(y)=R(t)$ is <independent> of $y$, with $R(s+t)=R(s)R(t)$. The continuous positive solutions are exponentials. Since $h$ is a <probability density function>, it follows that $h(y)=ce^{-cy}$ with $c>0$; the joint <probability density function>[density] then gives $g(x)=c^2xe^{-cx}$. Conversely <beta-gamma independence> proves <independence> for this gamma total. No assumption that $h(0)$ is finite was used.