Is it possible to find the probability distribution of the random variable $X$ that solves the following equation? $$ X = Bin(X, p) + Bin(X, 1-p), $$ where $Bin(X,p)$ is a random variable distributed according to a binomial distribution with parameters $X$ and $p$, $p \in (0, 1)$, the two random variables on the right side are independent.


The solutions do not depend on $p$ in $(0,1)$. These are the Poisson distributions and the Dirac mass at $0$.

To show this, first recall that, for every random variable $Y$ binomial $(n,p)$ and every $|s|\leqslant1$, the generating function of $Y$ is $$E(s^Y)=(ps+q)^n.$$ Considering $g(s)=E(s^X)$, the identity in distribution in the question translates as $$g(s)=g(ps+q)g(qs+p). $$ Considering $h(t)=-\log g(1-t)$, one gets $$h(t)=h(pt)+h(qt). $$ Iterating this, one gets, for every $n\geqslant1$, $$ h(t)=\sum_{k=0}^n{n\choose k}h(p^kq^{n-k}t). $$ Assume that $X$ is integrable with $\lambda=E(X)$, then $g(1-t)=1-\lambda t+o(t)$ when $t\to0$. If $\lambda=0$, $X=0$ with full probability. Otherwise, $h(t)\sim\lambda t$ when $t\to0$.

Fix some $\lambda'\lt\lambda$, then $h(t)\geqslant\lambda' t$ for every $t$ small enough. When $n\to\infty$, $p^kq^{n-k}\to0$ uniformly in $k$, hence, for every $t$ in $(0,1)$, for every $n$ large enough, $$ h(t)\geqslant\sum_{k=0}^n{n\choose k}\lambda'p^kq^{n-k}t=\lambda' t. $$ The same iteration shows that, for every $\lambda''\gt\lambda$, $h(t)\leqslant\lambda''t$, thus, $h(t)=\lambda t$ for every $t$ in $(0,1)$, that is, $g(s)=\mathrm e^{-\lambda(1-s)}$ for every $s$ in $(0,1)$, which means that $X$ is Poisson with parameter $\lambda$.

If $X$ is a non integrable solution, then, for every $\lambda$, $h(t)\geqslant\lambda t$ for every $t$ small enough, and yet again the same reasoning yields $h(t)\geqslant\lambda t$ for every $t$, hence $h(t)$ is infinite, that is, $g(s)=0$, which is absurd.

| cite | improve this answer | |
  • $\begingroup$ It seems to me that the solution (i.e. the parameter $\lambda$ of the poisson distribution) does not depend on $p$, right? $\endgroup$ – QuantumLogarithm Mar 11 '14 at 8:31
  • $\begingroup$ Correct, for every value of $p$, the set of solutions is the same. $\endgroup$ – Did Mar 11 '14 at 8:33
  • $\begingroup$ In case the equation is $X_1=Bin(X_2,p) + Bin(X_3,1−p)$, where $X_1$, $X_2$ and $X_3$ are identically distributed, but $X_2$ and $X_3$ are not independent, nothing can be done to determine the distribution of those variables, right? $\endgroup$ – QuantumLogarithm Mar 11 '14 at 10:29
  • $\begingroup$ Unless one specifies the dependence of $(X_2,X_3)$ (and even then this is not assured). $\endgroup$ – Did Mar 11 '14 at 10:32
  • $\begingroup$ If $X$ is integrable then did you use a Taylor expansion for $g(1-t)$? Don't you have then $o(t^2)$? $\endgroup$ – QuantumLogarithm Mar 14 '14 at 12:45

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.