## Selection, Revisited

### November 30, 2012

We studied a selection algorithm in a previous exercise. The algorithm is the same as quicksort, except that recursion follows only the partition in which the target of the selection is located. We used a randomized algorithm to select the pivot, which gives an expected O(n) time complexity but a worst-case O(n^2) time complexity. In today’s exercise we examine an algorithm due to Blum, Floyd, Pratt, Rivest and Tarjan from their 1973 paper “Time Bounds for Selection” in *Journal of Computer Systems Science* that provides guaranteed O(n) time complexity (actually, the time complexity is sub-linear, but the only claim is that it is linear).

The algorithm is the same as the previous exercise except in the selection of the pivot. Instead of a random pivot, the algorithm partitions the input into blocks of five elements, finds the median of each block by comparisons, then chooses the median of the medians (which is computed recursively) as the pivot. Thus, the algorithm is known as “selection by the median of the medians of five.”

Your task is to write a function that selects the *k*th item from a list. When you are finished, you are welcome to read or run a suggested solution, or to post your own solution or discuss the exercise in the comments below.

Pages: 1 2

[…] Pages: 1 2 […]

Even if the algorithm is very clear and straight forward, a careless implementation might contain an annoying bug with duplicates. For example, if at some point in the recursion, all the elements are equal suddenly it is not that easy to split the list in two sub-lists because one of them will be empty. I’ve used a very simple antidote by mapping the element a[i] into (a[i], i) transforming the initial list into a list of distinct pairs. Below is my Python implementation:

Another Python version. A key can be given to determine sort order.

Sub-linear? What do you mean by that? You can’t possibly select the kth item without examining all n items. Even the trivial cases of k=1 and k=n are linear.

A question: when you’re trying to find something other than the median, is it best to pivot about an approximate median (as this algorithm does), or is it possible to save time by choosing a different pivot? I couldn’t figure that out.