# $k$ socks $n$ drawers

Consider the following problem:

There are $k$ socks and $n$ drawers which can only contain 1 sock (very small drawer indeed). The drawers are aligned. You put socks in the drawers with uniform probability. You are wondering the law of probability of the number of drawers containing 1 sock and the drawer to its right have also a sock in it.

More formally, write $[1:n]=[1;n]\cap\mathbb{N}$, and let $\eta$ be a configuration of the sock & drawer, meaning $\eta = (\eta_{i})_{i\in [1:n]}\in \Omega=\{0;1\}^{[1:n]}$ where $\eta_{i}=1$ means that there is a sock in the drawer $i$. Furthermore, denote $\mathbb{P}$ the law of probabilty over $\Omega$ such that $\forall \eta,\sigma \in \Omega:\mathbb{P}(\eta)=\mathbb{P}(\sigma)$

We are looking at the law of probability of the following random variable : $$\sum_{i=1}^{n-1} \eta_i \eta_{i+1}$$ under the condition $\sum \eta_i=k$

Using $w$ for empty drawers and $z$ for non-empty ones and $u$ for non-empty drawers with a non-empty drawer to the right we obtain the generating function

$$(1+w+w^2+\cdots) \\ \times \sum_{q\ge 0} (z+z^2u+z^3u^2+\cdots)^q (w+w^2+w^3+\cdots)^q \\ \times (1+z+z^2u+z^3u^2+\cdots).$$

This is

$$\frac{1}{1-w} \sum_{q\ge 0} z^q \frac{1}{(1-uz)^q} w^q \frac{1}{(1-w)^q} \\ \times \left(1+ z\frac{1}{1-uz}\right) \\ = \frac{1}{1-w} \frac{1}{1-wz/(1-w)/(1-uz)} \frac{1+z-uz}{1-uz} \\ = \frac{1+z-uz}{(1-w)(1-uz)-wz}.$$

Observe that when we put $u=1$ and $w=z$ we obtain

$$\frac{1}{(1-z)^2-z^2} = \frac{1}{1-2z}$$

which is a useful sanity check and shows that we have accounted for all strings. Differentiate with respect to $u$ and set $u=1$ to count non-empty neighbors:

$$\left.\left(\frac{1+z-uz}{1-w-uz+uwz-wz}\right)'\right|_{u=1} \\ = \left.-\frac{z}{1-w-uz+uwz-wz} - \frac{1+z-uz}{(1-w-uz+uwz-wz)^2} (wz-z)\right|_{u=1} \\ = -\frac{z}{1-w-z} - \frac{1}{(1-w-z)^2}(wz-z) = \frac{z^2}{(1-w-z)^2}.$$

We thus require

$$[z^k] [w^{n-k}] \frac{z^2}{(1-w-z)^2}$$

where we may assume that $k\ge 2$ since with $k=1$ or $k=0$ the expectation is zero. We get

$$[z^{k-2}] [w^{n-k}] \frac{1}{(1-w-z)^2} \\ = [z^{k-2}] \frac{1}{(1-z)^2} [w^{n-k}] \frac{1}{(1-w/(1-z))^2} \\ = [z^{k-2}] \frac{1}{(1-z)^2} (n-k+1) \frac{1}{(1-z)^{n-k}} \\ = (n-k+1) [z^{k-2}] \frac{1}{(1-z)^{n-k+2}} \\ = (n-k+1) {k-2+n-k+1\choose n-k+1} \\ = (n-k+1) {n-1\choose n-k+1} = (n-1) {n-2\choose n-k}.$$

We get for the expectation

$$(n-1) {n\choose k}^{-1} {n-2\choose n-k} = (n-1) \frac{k! (n-k)!}{n!} \frac{(n-2)!}{(k-2)! (n-k)!}$$

which is

$$\bbox[5px,border:2px solid #00A000]{ \frac{1}{n} k(k-1).}$$

An elementary immediate proof is sure to appear given the simplicity of this result now that it has been computed.

The interpretation of the question that was used here is documented in the following Maple code.

with(combinat);

EN :=
proc(n, k)
local choice, res, nb, pos;

res := 0;

for choice in choose(n, k) do
nb := 0;

for pos to k-1 do
if choice[pos] + 1 = choice[pos+1] then
nb := nb+1;
fi;
od;

res := res + nb;
od;

res;
end;

F :=
(n,k) ->
coeftayl(coeftayl(z^2/(1-w-z)^2, w=0, n-k), z=0, k);

ENX := (n,k) -> EN(n,k)/binomial(n,k);

FX := (n,k) -> 1/n*k*(k-1);


$$[z^k] [w^{n-k}] \left.\frac{1+z-uz}{(1-w)(1-uz)-wz}\right|_{u=1} \\ = [z^k] [w^{n-k}] \frac{1}{(1-w)(1-z)-wz} = [z^k] [w^{n-k}] \frac{1}{1-w-z} \\ = [z^k] \frac{1}{1-z} [w^{n-k}] \frac{1}{1-w/(1-z)} \\ = [z^k] \frac{1}{1-z} \frac{1}{(1-z)^{n-k}} = [z^k] \frac{1}{(1-z)^{n-k+1}} \\ = {k+n-k\choose n-k} = {n\choose n-k} = {n\choose k}$$

which is the correct result.

$$[u^m] [z^k] [w^{n-k}] \frac{1+z-uz}{(1-w)(1-uz)-wz}.$$

This is

$$[u^m] [z^k] [w^{n-k}] \frac{1+z-uz}{1-uz-w(1+z-uz)} \\ = [u^m] [z^k] \frac{1}{1-uz} [w^{n-k}] \frac{1+z-uz}{1-w(1+z-uz)/(1-uz)} \\ = [u^m] [z^k] \frac{(1+z-uz)^{n-k+1}}{(1-uz)^{n-k+1}} \\ = [u^m] [z^k] \left(1+\frac{z}{1-uz}\right)^{n-k+1}.$$

Extracting coefficients we find

$$\sum_{p=0}^{n-k+1} {n-k+1\choose p} [u^m] [z^k] \frac{z^p}{(1-uz)^p} \\ = \sum_{p=0}^{k} {n-k+1\choose p} [u^m] [z^{k-p}] \frac{1}{(1-uz)^p}.$$

We must have $m=k-p$ or $p=k-m$ and get

$${n-k+1\choose k-m} {m+k-m-1\choose k-m-1} \\ = {n-k+1\choose k-m} {k-1\choose m}.$$

We can compute higher moments with these data. We already have $\mathrm{E}[X]$ and obtain for $\mathrm{E}[X(X-1)]$

$$\sum_{m=2}^{k-1} m(m-1) {n-k+1\choose k-m} {k-1\choose m}.$$

This is zero when $k=1$ or $k=2$ so we may assume that $k\ge 3$ in calculating the sum.

$$\sum_{m=2}^{k-1} m(m-1) {n-k+1\choose k-m} \frac{(k-1)(k-2)}{m(m-1)} {k-3\choose m-2} \\ = (k-1)(k-2) \sum_{m=2}^{k-1} {n-k+1\choose k-m} {k-3\choose m-2} \\ = (k-1)(k-2) \sum_{m=1}^{k-2} {n-k+1\choose m} {k-3\choose k-m-2}.$$

We evaluate this using the Egorychev method and put

$${k-3\choose k-m-2} = \frac{1}{2\pi i} \int_{|z|=\epsilon} \frac{1}{z^{k-m-1}} (1+z)^{k-3} \; dz.$$

Observe that this vanishes when $m\ge k-1$ and when $m=0$ where we get $[z^{k-2}] (1+z)^{k-3}=0$ and hence we may set the range of $m$ from zero to infinity, getting for the sum

$$(k-1)(k-2) \frac{1}{2\pi i} \int_{|z|=\epsilon} \frac{1}{z^{k-1}} (1+z)^{k-3} \sum_{m\ge 0} {n-k+1\choose m} z^m \; dz \\ = (k-1)(k-2) \frac{1}{2\pi i} \int_{|z|=\epsilon} \frac{1}{z^{k-1}} (1+z)^{n-2} \; dz \\ = (k-1)(k-2) \times {n-2\choose k-2}.$$

We get for the second factorial moment

$$(k-1)(k-2) {n\choose k}^{-1} {n-2\choose k-2} \\ = (k-1)(k-2) \frac{k! (n-k)!}{n!} \frac{(n-2)!}{(n-k)!(k-2)!}.$$

This is

$$\frac{1}{n}\frac{1}{n-1} k(k-1)^2(k-2).$$

We have for the variance

$$\mathrm{Var}[X] = \mathrm{E}[X^2]-\mathrm{E}[X]^2 = \mathrm{E}[X(X-1)] + \mathrm{E}[X] -\mathrm{E}[X]^2$$

so that in the present case we obtain

$$\bbox[5px,border:2px solid #00A000]{ \frac{1}{n}\frac{1}{n-1} k(k-1)^2(k-2) + \frac{1}{n} k(k-1) - \frac{1}{n^2} k^2(k-1)^2.}$$

This simplifies to

$$\frac{k(k-1)(n-k)(n+1-k)}{n^2(n-1)}.$$

• Thank you very much for your answer. The way of computing the generating function is instructive and natural. – Salsifis Nov 10 '16 at 8:04
• @MarkoRiedel: Nice presentation! (+1) – Markus Scheuer Nov 14 '16 at 14:16

Let $B(n,k)$ be the set of $n$-bit strings with exactly $k$ ones. For $b=b_1b_2\ldots b_n\in B(n,k)$ let

$$f(b)=|\{i\in[n-1]:b_i=b_{i+1}=1\}|\;;$$

if all we want is the expected value of $f(b)$ when $b$ is chosen from the uniform distribution on $B(n,k)$, we want

$$\binom{n}k^{-1}\sum_{b\in B(n,k)}f(b)\;,$$

which isn’t hard to compute.

For any $i\in[n-1]$ there are $\binom{n-2}{k-2}$ strings $b=b_1b_2\ldots b_n\in B(n,k)$ such that $b_i=b_{i+1}=1$, so

\begin{align*} \sum_{b\in B(n,k)}f(b)&=\sum_{i\in[n-1]}|\{b\in B(n,k):b_i=b_{i+1}=1\}|\\ &=\sum_{i\in[n-1]}\binom{n-2}{k-2}\\ &=(n-1)\binom{n-2}{k-2}\;, \end{align*}

and hence

$$\binom{n}k^{-1}\sum_{b\in B(n,k)}f(b)=\frac{k!(n-k)!}{n!}\cdot\frac{(n-1)!}{(k-2)!(n-k)!}=\frac{k(k-1)}n\;.$$

If we want the actual distribution of $f(b)$, i.e.,

$$|\{b\in B(n,k):f(b)=m\}|$$

for each $m\in\{0,1,\ldots,k-1\}$, we have to work harder.

Suppose that $b\in B(n,k)$ has $\ell$ maximal blocks of ones. Each $1$ in a block except the first is the second $1$ of an adjacent pair, so $f(b)=k-\ell$, and $|\{b\in B(n,k):f(b)=m\}|$ is the number of $b\in B(n,k)$ having $k-m$ blocks of ones.

There are $\binom{k-1}{\ell-1}$ ways to divide $k$ ones into $\ell$ non-empty blocks. We now put a $0$ in each of the $\ell-1$ spaces between adjacent blocks of ones; that leaves $n-k-\ell+1$ zeroes to be distributed at will amongst the $\ell+1$ slots before the first block of ones, between adjacent blocks of ones, and after the last block of ones. This can be done in

$$\binom{(n-k-\ell+1)+(\ell+1)-1}{(n-k-\ell+1)-1}=\binom{n-k+1}{n-k-\ell+1}=\binom{n-k+1}\ell$$

ways, so there are

$$\binom{k-1}{\ell-1}\binom{n-k+1}\ell$$

strings in $B(n,k)$ with $\ell$ blocks of ones, and

\begin{align*} |\{b\in B(n,k):f(b)=m\}|&=\binom{k-1}{k-m-1}\binom{n-k+1}{k-m}\\ &=\binom{k-1}m\binom{n-k+1}{k-m}\;. \end{align*}

As a quick sanity check note that

$$\sum_m\binom{k-1}m\binom{n-k+1}{k-m}=\binom{n}k\;,$$

as it should.

The probability of choosing a word $b$ such that $f(b)=m$ is therefore

$$\binom{n}k^{-1}\binom{k-1}m\binom{n-k+1}{k-m}\;.$$

• Thank you for suggesting the complete coefficient extraction. I was able to confirm it using the generating function, consult my modified answer. (+1). – Marko Riedel Nov 10 '16 at 0:11
• @Marko: It’s always reassuring when we get the same answer! (I’d already upvoted yours, but it’s even better now.) – Brian M. Scott Nov 10 '16 at 0:24
• I have added some additional material concerning the variance where I hope I have not introduced any mistakes. – Marko Riedel Nov 10 '16 at 2:12
• Thank you very much for both of your answers. They are very clear and very instructive. – Salsifis Nov 10 '16 at 8:03
• @Salsifis: You're very welcome. – Brian M. Scott Nov 10 '16 at 16:18