Coupling with conditional distributions

187 Views Asked by At

I have a question regarding coupling a random variable with the same random variable conditioned on a particular event. The example I will be discussing is taken from "Coupling and Poisson Approximation" by Janson.

Let $r$ balls be thrown independently of each other into $n$ boxes labelled $1,...,n$. Let $p_i$ denote the probability of hitting box $i$. Defined the indicator $I_i = I(\text{box $i$ is empty})$, and let $W=\sum_{i=1}^n I_i$. For $j\neq i$, we may create a coupling of $I_i$ and $I_i | I_j=1$ by removing all balls from box $j$ and redistributing them according to the original distribution, repeating until box $j$ is empty.

While seems obvious that this should be a coupling of $I_i$ and $I_i | I_j=1$, it is not clear to me how I would write down the joint distribution for the coupling if I wanted to rigorously verify that, especially since in the conditional space, our probabilities are normalised by $\mathbb{P}(I_j=1)$. There are several examples of this type in the paper, so I'm hoping if I can understand this simpler one, I can understand the others. Thanks.

1

There are 1 best solutions below

2
On BEST ANSWER

It's equivalent but easier to see if you consider "fresh balls" from a countable sequence instead of reusing the old ones that were thrown in box $j$. Here is a coupling. Consider $k$ boxes and a given box $j\le k$ of interest. Consider a countable ordered sequence of iid balls $B_1,B_2, B_3,...$ thrown into the boxes.

  • For the distribution $n$ balls conditionally on box $j$ being empty, consider the first $n$ balls of the sequence not falling into box $j$. In other words, if a ball falls in box $j$, forget it and continue throwing balls using the next "fresh" one from the sequence.
  • For the unconditional distribution, consider only the first $n$ balls $B_1, B_2,...,B_n$ no matter where they are thrown.