Poisson-multinomial conditioning (source code)

= Poisson-multinomial conditioning
{c}
{wiki}

Let $X_j$ be <independent> <Poisson random variables> of <means> $\lambda_j\geq0$, with $\Lambda=\sum_j\lambda_j>0$ for finitely many categories. Conditional on $S=\sum_jX_j=n$, the count <vector> has the <multinomial distribution> with <probabilities> $p_j=\lambda_j/\Lambda$:
$$
\mathbb P(X_1=x_1,\ldots,X_k=x_k\mid S=n)
=\frac{n!}{\prod_jx_j!}\prod_j\left(\frac{\lambda_j}{\Lambda}\right)^{x_j},\qquad \sum_jx_j=n.
$$
Divide the <independent> Poisson joint mass by the <Poisson distribution> mass $e^{-\Lambda}\Lambda^n/n!$ of $S$. The exponential factors cancel, leaving the formula. Before conditioning the counts are <independent>; after a positive fixed total, distinct positive-probability categories have <covariance> $-np_ip_j$.

For <means> $2,3$, conditioning on total $4$ gives the first count a <binomial distribution> with parameters $4,2/5$. A zero-mean category is always zero. If $\Lambda=0$, every count is zero <almost surely>; conditioning on total zero is degenerate, while conditioning on a positive total has zero <probability> and does not define the displayed law.