0
$\begingroup$

Given a random variable $X$ with $\mathsf{supp}\, X \subseteq [0,1]$ and $n$ positive numbers $h_1,\cdots,h_n$ with $\sum_{i=1}^n h_i=1$, I want to know some sufficient conditions for decomposing $X$ into $X=\sum_{i=1}^nh_iW_i$, where components $W_i$ are mutually independent and satisfy $\mathsf{supp} \, W_i \subseteq [0,1]$.

A particular appearling case is decomposition of a uniform distribution on $[0,1]$.


Some personal remarks may be meaningful.

  1. For a Bernoulli distribution $B(p)$ with any $p\in (0,1)$, there does not exist such a decomposition.
  2. If $n=2$, for any pair $(h_1,h_2)$, we can decompose a random variable $X$ uniformly distributed on $[0,1]$ as follows:

we first calculate the binary respresentation $\{b_i\}_{i=1}^{+\infty}$ of $h_1$, i.e., $h_1=\sum_{i=1}^{+\infty} b_i\cdot 2^{-i}$. Then let $$ W_1=\sum_{i=1}^{+\infty}\frac{b_i}{h_1} \cdot 2^{-i} \cdot B_i, $$ and $$ W_2=\sum_{i=1}^{+\infty}\frac{(1-b_i)}{h_2} \cdot 2^{-i} \cdot B_i, $$ where $B_i$'s are independent copies of a Bernoulli random variable $B(\frac{1}{2})$. It is clear that $W_1$ is independent of $W_2$ and $X=h_1W_1+h_2W_2$ in distribution.

$\endgroup$
2
  • $\begingroup$ In your second example, in general $X$ will not equal $h_1W_1+h_2W_2$ in distribution. Indeed, for almost all $h_1$, the distribution will have an infinitely smooth density. $\endgroup$ Oct 21 '21 at 12:31
  • $\begingroup$ @losif Pinelis I have corrected my previous post. $\endgroup$
    – Ryan Chen
    Oct 21 '21 at 13:05
0
$\begingroup$

On your general question, there are several books which address it:

Yu. V. Linnik and I. V. Ostrovskii, Decompositions of random variables and vectors,

Lukacz, Characteristic functions,

Ramachandran, Advanced theory of characteristic functions.

$\endgroup$

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.