I am working on a problem from the book "probability and random process" by Geoffrey Grimmett.

suppose X, Y are independent random variables take values of non-negative integers. and they have the following property. $$P\left(X=k|X+Y=n\right)=\left(\begin{array}{c} n\\ k \end{array}\right)p^{k}\left(1-p\right)^{n-k} $$ which is a binomial distribution, prove X, and Y are poisson random variables.

I know the result that "conditioning on X+Y, X or Y obeys binomial distribution" from the property of poisson process. However, I don't know how to prove from formula to poisson.

Any help will be appreciated, thanks.

  • $\begingroup$ Poisson variables take the value $0$ -- I presume you meant non-negative integers? $\endgroup$ – joriki Apr 4 '12 at 15:58
  • $\begingroup$ For any fixed value of $n$, the conditional probability mass function of $X$ given $X+Y=n$ is not a valid mass function (the total probability $\sum_{k>0} P\{X=k|X+Y=n\}$ is less than $1$). So some different conditions must be imposed, e.g. nonnegative random variables rather than positive random variables. $\endgroup$ – Dilip Sarwate Apr 4 '12 at 16:23
  • $\begingroup$ thanks, @joriki and Dilip. it should be non-negative integer, and no other conditions are available in the question. $\endgroup$ – fast tooth Apr 4 '12 at 16:45

The given conditional distributions fix the ratios of probabilities $P(X=k,Y=l)$ with common sum $k+l=n$. Since $X$ and $Y$ are independent, these are given by $P(X=k)P(Y=l)$. Thus we have


Taking logarithms, we find that $\log P$ is determined by a linear recurrence:

$$\log P(Y=l+2)=2\log P(Y=l+1)-\log P(Y=l)-\log\beta_l\;.$$

Since you know that conditionalizing independent Poisson distributions on the sum yields a binomial distribution, you already know that one solution of this inhomogenous recurrence relation is (the logarithm of) a Poisson distribution. The homogeneous recurrence relation is solved by $\log P (Y=l)=al+b$. Adding this to the known solution for the inhomogeneous recurrence relation and using normalization shows that $Y$ must be Poisson-distributed, and analogously for $X$.

  • $\begingroup$ wow. brilliant answer. thanks a lot! Can you tell me the motivation of this solution? $\endgroup$ – fast tooth Apr 4 '12 at 23:06
  • $\begingroup$ @johnniac: I'll try. Generally speaking, $n^2$ values have to be represented by $2n$ values, so it's plausible that these $2n$ values are completely determined; we just have to find a way to disentangle them. If you write down the array of probabilities $P(X=k,Y=l)=P(X=k)P(Y=l)$, it's clear that forming ratios allows you to get rid of one of the variables. Once you have ratios that only contain one of the variables, it's a matter of showing that those ratios fix all the values, and of remembering that multiplication and division can be turned into addition and subtraction by taking logarithms. $\endgroup$ – joriki Apr 4 '12 at 23:51
  • $\begingroup$ hi, @joriki, I see what you are talking about, I try to go through the proof by myself, I found the distribution for Y. but on the final step Nomalization, it seems I have a lot variables to be determined, a,b, and poisson($\lambda$), how can I show such a distribution must be a poisson? $\endgroup$ – fast tooth Apr 5 '12 at 15:18
  • $\begingroup$ @johnniac: There's no need to explicitly determine any variables. You have a Poisson distribution and you multiply it by a geometric progression; the result is, simply by its form, a (possibly denormalized) Poisson distribution. $\endgroup$ – joriki Apr 5 '12 at 16:59
  • $\begingroup$ yes, you are right, thanks for you help, you are really good at it! $\endgroup$ – fast tooth Apr 5 '12 at 17:51

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.