StatisticsProbability

Knowledge guide

Poisson Splitting: Why Random Category Counts Can Be Independent

Learn why independent labeling of a Poisson total produces independent category counts, and when conditioning on the total changes that result.

A random total and independent labels

Suppose N∼Poisson⁡(μ)N\sim\operatorname{Poisson}(\mu) counts events in a fixed interval. Give every event one of mm mutually exclusive labels, independently of other events and the total count. Label jj has probability rjr_j, with ∑jrj=1\sum_jr_j=1. If CjC_j counts label-jj events, then

Cj∼Poisson⁡(μrj),C_j\sim\operatorname{Poisson}(\mu r_j),

and the CjC_j are mutually independent. The mean in each category is the total mean multiplied by the category probability. Both the mean and variance of that category count equal μrj\mu r_j.

Why this does not contradict a fixed total

When the total is fixed at N=nN=n, the category counts have a multinomial distribution. Learning one category count constrains the remaining counts because they must add to nn.

The unconditional distribution includes the variation in the Poisson total. For nonnegative integers c1,…,cmc_1,\ldots,c_m and n=∑jcjn=\sum_jc_j,

P(C1=c1,…,Cm=cm)=e−μμnn!n!∏jcj!∏jrjcj=∏je−μrj(μrj)cjcj!.P(C_1=c_1,\ldots,C_m=c_m) =e^{-\mu}\frac{\mu^n}{n!}\frac{n!}{\prod_jc_j!}\prod_jr_j^{c_j} =\prod_j e^{-\mu r_j}\frac{(\mu r_j)^{c_j}}{c_j!}.

This factorization proves unconditional independence. Conditional dependence and unconditional independence describe two different distributions.

New example: automatically routed service requests

Consider a theoretical model in which a service center receives a Poisson mean of 12 requests per hour. Each request is independently labeled billing, technical, or general with probabilities 0.5, 0.3, and 0.2. The hourly counts are independent Poisson variables with means 6, 3.6, and 2.4. Their means sum to 12.

If we instead condition on exactly 12 requests arriving, their category counts must sum to 12 and are multinomial. These parameters describe an illustrative model, not observations from a real service center.

Thinning and complementary counts

Keeping each event independently with probability rr produces a Poisson kept count with mean μr\mu r and an independent discarded count with mean μ(1−r)\mu(1-r). A later independent filter with retention probability ss leaves mean μrs\mu rs.

Checks before applying the theorem

The total must follow a Poisson model, the labels must be independent across events, and the label probabilities must apply consistently. A fixed total, correlated labels, or varying unmodeled detection probabilities does not automatically satisfy this theorem. A Poisson mean is an expected count, so a fractional value is valid and does not claim a fractional observed event.

Related question

Apply this knowledge

Use the concept guide to understand the reasoning, then return to the complete question and worked answer.

Poisson Errors: Estimating Typos Missed by Two Proofreaders

Sources

These references support the core concepts and interpretation boundaries explained above.