Suppose that the number of events $N$ is a Geometric random variable with mean $\frac{1-p}{p}$. Further suppose that each event can be classified into one of $m$ types with probabilities $p_{1},p_{2},\ldots,p_{m}$ independent of all other events. Then we consider the number of events $N_{1},N_{2},\ldots,N_{m}$ corresponding to event types $1,2,\ldots,m$, respectively,

What can we say about $N_{i}$?

What is the distribution of $N_{i}$?

Are $N_{1},N_{2},\ldots,N_{m}$ mutually independent random variables?

What is the sign of $cov(N_{1},N_{2})$?

My attempt: For fixed $N=n$, the conditional joint distribution of $(N_{1},\ldots,N_{m})$ is multinomial with parameters $(n,p_{1},\ldots,p_{m})$. Also, for fixed $N=n$, the conditional marginal distribution of $N_{j}$ is binomial with parameters $(n,p_{j})$.

I want to show that the joint pf is the product of the marginal pfs, establishing mutual independence

The joint pf of $(N_{1},\ldots,N_{m})$ is given by

\begin{align} \mathbb{P}(N_{1}=n_{1},\ldots,N_{m}=n_{m}) &= \mathbb{P}(N_{1}=n,\ldots,N_{m}=n_{m}|N=n)\mathbb{P}(N=n)\\ &= \left( \frac{n!}{n_{1}!\cdots n_{m}!}p_{1}^{n_{1}}\cdots p_{m}^{n_{m}}\right)(1-p)^{n}p \\ &= \mbox{I do not know how to continue} \end{align} where $n=n_{1}+n_{2}+\cdots +n_{m}$. Similarly, as I fail with the above, try to calculate the marginal pf of $N_{j}$

\begin{align} \mathbb{P}(N_{j}=n_{j}) &= \sum_{n=n_{j}}^{\infty} \mathbb{P}(N_{j}=n_{j}|N=n)\mathbb{P}(N=n)\\ &= \sum_{n=n_{j}}^{\infty} \binom{n}{n_{j}}p_{j}^{n_{j}}(1-p_{j})^{n-n_{j}}p(1-p)^{n} \\ &= \mbox{I do not know how to continue} \end{align}

  • $\begingroup$ The $N_j$'s cannot be independent because, else, if you take the case $p_1=\cdots=p_m$, the sum of Geometric iid variates is a Negative Binomial $(k,p_1)$ variate. $\endgroup$
    – Xi'an
    Feb 22, 2019 at 10:10

2 Answers 2


You have made a good start. To finish this working you just pull out all terms that don't depend on the summation index $n$ and then apply a well-known summation identity for the binomial coefficients. For all $n_k = 0,1,2,3,...$ you have:

$$\begin{align} \mathbb{P}(N_k=n_k) &= \sum_{n=0}^{\infty} \mathbb{P}(N_{k}=n_k|N=n)\mathbb{P}(N=n) \\[6pt] &= \sum_{n=0}^{\infty} \text{Bin}(n_k | n, p_j) \cdot \text{Geom}(n|p) \\[6pt] &= \sum_{n=n_k}^{\infty} \binom{n}{n_k} p_k^{n_k}(1-p_k)^{n-n_k} \cdot p(1-p)^{n} \\[6pt] &= p \Big( \frac{p_k}{1-p_k} \Big)^{n_k} \sum_{n=n_k}^{\infty} \binom{n}{n_k} \Big[ (1-p_k) (1-p) \Big]^{n} \\[6pt] &= p \Big( \frac{p_k}{1-p_k} \Big)^{n_k} \frac{[ (1-p_k) (1-p) ]^{n_k}}{[ 1 - (1-p_k) (1-p) ]^{n_k+1}} \\[6pt] &= p \cdot \frac{(p_k - p \cdot p_k)^{n_k}}{(p+p_k - p \cdot p_k)^{n_k+1}} \\[6pt] &= \frac{p}{p+p_k - p \cdot p_k} \cdot \Big( \frac{p_k - p \cdot p_k}{p+p_k - p \cdot p_k} \Big)^{n_k} \\[6pt] &= \frac{p}{p+p_k - p \cdot p_k} \cdot \Big( 1 - \frac{p}{p+p_k - p \cdot p_k} \Big)^{n_k} \\[6pt] &= \text{Geom} \Big( n_k \Big| \frac{p}{p+p_k - p \cdot p_k} \Big). \\[6pt] \end{align}$$

So we see that:

$$N_k \sim \text{Geom} \Big( \frac{p}{p+p_k - p \cdot p_k} \Big).$$

You can see here that the parameter in the resulting geometric distribution for $N_k$ is no larger than the parameter in the original geometric distribution for $N$. They are equal in the special case where $p_k=1$. This is unsurprising, given that the latter probability represents the probability that a given item in the original set of $N$ objects is included in the set of $N_k$ objects.

  • $\begingroup$ You very much appreciate your answer, but I have a question, according to what you calculated What kind of distribution has $N_{k}$? $\endgroup$ Feb 22, 2019 at 6:34
  • $\begingroup$ I have added further information to show the distributional family. $\endgroup$
    – Ben
    Feb 22, 2019 at 7:10

This is only a long remark that is not directly related with the original question, fully addressed by the earlier answer of Ben:

The Geometric distribution is infinitely divisible. This means that, for $X\sim\mathcal{G}(P)$, $X$ can always be written as $$X=X_1+\dots+X_n\qquad X_i\stackrel{\text{iid}}{\sim}P_n$$for any $n$. The simplest way to prove this is to look at the moment generating function $$\varphi(z)=\mathbb{E}[e^{zX}]=\dfrac{p}{1-(1-p)e^z}$$ Indeed, $\varphi^t$ $(t>0)$ appears as a Negative Binomial $\mathcal{NB}(t,p)$ moment generating function (in the general definition of a Negative Binomial distribution with real parameter $t$).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.