# Balls and bins conditioned on the number of non-empty bins

The expected number of occupied bins in the standard balls and bins problem (with $m$ balls into $N$ bins) is $N\left( {1 - {{\left( {1 - \frac{1}{N}} \right)}^m}} \right)$. How do I compute this expectation given that at least $k$ bins are non-empty after the $m$ balls were thrown?

In particular, how do I formulate the probability that some $i^{th}$ bin is empty given that at least $k$ bins are non-empty? in addition, it seems obivious, but how do I prove that the conditional expectation is a monotonic increasing function of $k$?

The standard argument to get the expected number of bins occupied is greatly simplified by the linearity of expectation. We only need to compute the probability that a single bin is filled. This is simplified by computing the probability that that bin is empty: $\left(1-\frac1n\right)^m$. Thus, the probability that that bin is filled is $1-\left(1-\frac1n\right)^m$. Linearity of expectation says that the expected number of bins filled would be $$n\left(1-\left(1-\frac1n\right)^m\right)\tag{1}$$ as you have stated.

However, when assuming that at least $k$ bins have been filled, we can not use all of the preceding simplifications.

Inclusion-Exclusion

One method of attack is using the Inclusion-Exclusion Principle. Let $S(i)$ be all the possible arrangements where bin $i$ is empty. Then we can compute $N(k)$, the size of the intersection of $k$ of the $S(i)$: there are $\binom{n}{k}$ ways to choose the empty bins and $(n-k)^m$ ways to put the $m$ marbles in to the remaining bins. Therefore, $$N(k)=\binom{n}{k}(n-k)^m\tag{2}$$ Now, to compute the number of arrangements in which exactly $k$ bins are filled, we compute the number of elements in exactly $n-k$ of the $S(i)$: \begin{align} &\sum_{j}(-1)^{j-n+k}\binom{j}{n-k}N(j)\\ &=\sum_{j}(-1)^{j-n+k}\binom{j}{n-k}\binom{n}{j}(n-j)^m\\ &=\sum_{j}(-1)^{j-n+k}\binom{k}{j+k-n}\binom{n}{k}(n-j)^m\\ &=\binom{n}{k}\sum_{j}(-1)^{k-j}\binom{k}{j}j^m\tag{3} \end{align} The number of ways to get at least $k$ bins filled is \begin{align} &\sum_{i\ge k}\sum_j(-1)^{i-j}\binom{n}{i}\binom{i}{j}j^m\\ &=\sum_i\sum_j(-1)^{k-j}\binom{-1}{i-k}\binom{n}{j}\binom{n-j}{n-i}j^m\\ &=\sum_j(-1)^{k-j}\binom{n}{j}\binom{n-j-1}{n-k}j^m\tag{4} \end{align} The number of bins times the number of ways to get at least $k$ bins filled is \begin{align} &\sum_{i\ge k}\sum_j(-1)^{i-j}i\binom{n}{i}\binom{i}{j}j^m\\ &=\sum_i\sum_j(-1)^{k-j}i\binom{-1}{i-k}\binom{n}{j}\binom{n-j}{n-i}j^m\\ &=\sum_i\sum_j(-1)^{k-j}[k+(i-k)]\binom{-1}{i-k}\binom{n}{j}\binom{n-j}{n-i}j^m\\ &=\sum_i\sum_j(-1)^{k-j}\binom{n}{j}\left[k\binom{-1}{i-k}-\binom{-2}{i-k-1}\right]\binom{n-j}{n-i}j^m\\ &=\sum_j(-1)^{k-j}\binom{n}{j}\left[k\binom{n-j-1}{n-k}-\binom{n-j-2}{n-k-1}\right]j^m\tag{5} \end{align} Dividing $(5)$ by $(4)$ yields the expected value $$k-\frac{\sum\limits_{j=0}^n(-1)^{k-j}\binom{n}{j}\binom{n-j-2}{n-k-1}j^m}{\sum\limits_{j=0}^n(-1)^{k-j}\binom{n}{j}\binom{n-j-1}{n-k}j^m}\tag{6}$$ where $\binom{-1}{k}=(-1)^k$ and $\binom{-2}{k}=(-1)^k(k+1)$ and $0^0=1$.

Example

For $n=6$, $m=4$, $k=2$: \begin{align} &2-\frac{\small\binom{6}{0}\binom{4}{3}0^4{-}\binom{6}{1}\binom{3}{3}1^4{+}\binom{6}{2}\binom{2}{3}2^4{-}\binom{6}{3}\binom{1}{3}3^4{+}\binom{6}{4}\binom{0}{3}4^4{-}\binom{6}{5}\binom{-1}{3}5^4{+}\binom{6}{6}\binom{-2}{3}6^4} {\small\binom{6}{0}\binom{5}{4}0^4{-}\binom{6}{1}\binom{4}{4}1^4{+}\binom{6}{2}\binom{3}{4}2^4{-}\binom{6}{3}\binom{2}{4}3^4{+}\binom{6}{4}\binom{1}{4}4^4{-}\binom{6}{5}\binom{0}{4}5^4{+}\binom{6}{6}\binom{-1}{4}6^4}\\ &=2-\frac{0-6+0-0+0-(-3750)+(-5184)}{0-6+0-0+0-0+1296}\\ &=\frac{670}{215}\doteq3.11627906976744 \end{align} whereas, without the knowledge that at least two bins were not empty, we would get $$6\left(1-\left(1-\frac16\right)^4\right)=\frac{671}{216}\doteq3.10648148148148$$ Not a terribly large difference, but with $4$ balls into $6$ bins, you wouldn't expect them to all land in one bin, so the knowledge that at least two bins were not empty is not very significant.

• Great answer! I guessed that the inclusion-exclusion principle should be involved, but I was not sure how to use it. Thanks for your elaborate explanation!
– r1c
Sep 12, 2013 at 8:30
• Can it be extended to the case where it is known that the balls are thrown $l$ by $l$ sets, where l|m and the balls in each set are known to fall in $l$ distinct bins? it is implied now that at least $l$ bins are occupied.
– r1c
Sep 12, 2013 at 9:27
• @Rob: I am not quite sure what you mean by "thrown $l$ by $l$ sets". Would you please explain what that means? Thanks.
– robjohn
Sep 12, 2013 at 9:44
• Sure. I have $m$ balls. I divide them into $m/l$ sets of balls (l|m), each set is of size $l$. Say that numbers are given to each set: $1,2,...,m/l$. I take the $l$ balls of set no. $1$ and throw these $l$ balls into $l$ bins (distinct, randomly chosen). Now I take the $l$ balls of set no. $2$ and throw them into $l$ bins (distinct, randomly chosen, independent of the bins that were chosen previously). I continue in this fashion until all the balls are thrown. For example, in the standard balls and bins, $l=1$.
– r1c
Sep 12, 2013 at 9:49
• I see. I will have to see if I can compute $N(k)$ for that.
– robjohn
Sep 12, 2013 at 10:11