Coincidences in random shuffling revisited

This is just a follow-up from my last post, collecting and cleaning up my thoughts on the behavior of “coincidences” in the random shuffled order of songs in a playlist (i.e., consecutive songs by the same artist).  But first, it is useful to begin with the slightly different and more common problem mentioned by Chris in the comments.  Consider the following game:

You and a friend each have a brand new deck of cards.  You shuffle your deck, and your friend shuffles his.  Then each of you deals the top card of your respective decks face up next to each other on the table.  If the two cards match, then you win one dollar.  Continue dealing together, one card at a time through the entire deck, winning an additional dollar whenever two cards match.  How much should you be willing to pay to play this game?  What is the probability that you lose your initial wager (i.e., no two cards match)?

This is the “hat check problem” in not-so-clever disguise– n people check their hat at a restaurant, and the hats are returned randomly; what is the expected number of people who get their own hat, and what is the probability that no one gets their own hat?  Or, more abstractly, given a random permutation \pi \in S_n, what is the expected number of fixed points of \pi (i.e., the number of elements j such that \pi(j)=j), and what is the probability that \pi has no fixed points?

Both of these questions have “nice” answers: the expected number of fixed points is 1, independent of n, and the probability of no fixed points approaches 1/e very quickly as n\to\infty.  (In other words, you should be willing to play $1 to play the card game above, with less than a 37% chance of losing money.)  The former result is a nice application of linearity of expectation, the latter of inclusion-exclusion.  But I think it is useful to find ways to present interesting mathematics like this to students in ways that require less machinery, such as the notions of “random variables,” “expected value,” or the fact that it is “linear.”

For example: imagine creating a large table, with n! rows, one for each possible permutation (or arrangement of cards in a shuffled deck, or distribution of hats to customers, etc.).  Each row of the table contains a single number, the number of fixed points of the corresponding permutation.  We want to compute the average of these n! numbers (i.e., compute their sum and divide by n!).  At first glance, this seems hard to do; the numbers in the table range from 0 to n, with varying numbers of permutations having each possible number of fixed points.

But now let’s expand the table, making it wider so that instead of just a single column of numbers, we have n columns, with a 1 in column j if j is a fixed point of the corresponding permutation, and 0 otherwise.  If we sum the entries in each row, we get the original single-column table.  The useful trick is to instead sum the entries in each column: for each column j, there are (n-1)! permutations that fix j.  Thus, the sum of entries in the table (i.e., the total number of 1s) is n(n-1)! = n!; dividing by the number of rows n!, the average number of fixed points is 1.

Using this same table of 0s and 1s, let’s re-label the columns by defining the subsets of permutations F_1, F_2, ..., F_n, where F_j = \{\pi : \pi(j)=j\}.  Then the entry in “row \pi, column j” is 1 if \pi\in F_j, and 0 otherwise.  The cardinality of each F_j is (n-1)!… but more generally, the cardinality of the intersection of any k of these subsets is (n-k)! (why?).  This is what makes the inclusion-exclusion formula for the probability of no fixed points so elegant:

\frac{1}{n!} \sum_{k=0}^n (-1)^k {n \choose k}(n-k)!

= 1 - \frac{1}{1!} + \frac{1}{2!} - \frac{1}{3!} + ... + (-1)^n \frac{1}{n!}

Why is this useful?  So far, none of this is new.  But coming back now to the original problem of shuffling songs in a playlist: instead of fixed points of the form \pi(j)=j, we are interested in “coincidences,” or “adjacencies” of the form \pi(j)+1 = \pi(j+1), where \pi(j) is the position in the shuffled playing order of Song j.  We define the corresponding “adjacency-preserving” subsets A_1, A_2, ..., A_n, where A_j = \{\pi : \pi(j)+1 = \pi(j+1)\}, with the last subset A_n = \{\pi: \pi(n)+1 = \pi(1)\} corresponding to viewing the natural order of songs as cyclic.

The key observation is that these subsets have structure very similar to the F_j; indeed, the cardinality of each A_j is also (n-1)!, so that the “sum over columns” approach above yields an expected number of coincidences of 1, independent of n, if we view the natural song order cyclically (i.e., if we “allow” the coincidences indicated by A_n), or (n-1)/n if we don’t.

More generally, the cardinality of the intersection of any k of the A_j is also (n-k)!except when k=n.  In that case, instead of exactly one permutation with all n possible fixed points (i.e., the identity), there is no shuffled playlist with all n possible coincidences (why?).

Fortunately, this has very little effect on the essential result: the inclusion-exclusion formula for the probability of no coincidences (allowing the cyclic one) looks just like that for no fixed points, except that it is missing the final (-1)^n \frac{1}{n!} term.

 

This entry was posted in Uncategorized. Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s