Median of medians explained

Median of Medians
Alt:	Illustration showing vertical rectangles with 5 circles in each. The middle circle is black and the others are white/empty. The very middle rectangle has an arrow pointing to its black circle and an 'x' at the end of the arrow.
Caption:	Median of medians
Class:	Selection algorithm
Data:	Array
Best-Time:	O(n)
Time:	O(n)
Space:	O( log n) auxiliary
Optimal:	Yes

In computer science, the median of medians is an approximate median selection algorithm, frequently used to supply a good pivot for an exact selection algorithm, most commonly quickselect, that selects the kth smallest element of an initially unsorted array. Median of medians finds an approximate median in linear time. Using this approximate median as an improved pivot, the worst-case complexity of quickselect reduces from quadratic to linear, which is also the asymptotically optimal worst-case complexity of any selection algorithm. In other words, the median of medians is an approximate median-selection algorithm that helps building an asymptotically optimal, exact general selection algorithm (especially in the sense of worst-case complexity), by producing good pivot elements.

Median of medians can also be used as a pivot strategy in quicksort, yielding an optimal algorithm, with worst-case complexity

O(n log n) . Although this approach optimizes the asymptotic worst-case complexity quite well, it is typically outperformed in practice by instead choosing random pivots for its average O(n) complexity for selection and average O(n log n) complexity for sorting, without any overhead of computing the pivot.

Similarly, Median of medians is used in the hybrid introselect algorithm as a fallback for pivot selection at each iteration until kth smallest is found. This again ensures a worst-case linear performance, in addition to average-case linear performance: introselect starts with quickselect (with random pivot, default), to obtain good average performance, and then falls back to modified quickselect with pivot obtained from median of medians if the progress is too slow. Even though asymptotically similar, such a hybrid algorithm will have a lower complexity than a straightforward introselect up to a constant factor (both in average-case and worst-case), at any finite length.

The algorithm was published in, and thus is sometimes called BFPRT after the last names of the authors. In the original paper the algorithm was referred to as PICK, referring to quickselect as "FIND".

Motivation

Quickselect is linear-time on average, but it can require quadratic time with poor pivot choices. This is because quickselect is a divide and conquer algorithm, with each step taking

O(n) time in the size of the remaining search set. If the search set decreases exponentially quickly in size (by a fixed proportion), this yields a geometric series times the O(n) factor of a single step, and thus linear overall time. However, if the search set decreases slowly in size, such as linearly (by a fixed number of elements, in the worst case only reducing by one element each time), then a linear sum of linear steps yields quadratic overall time (formally, triangular numbers grow quadratically). For example, the worst-case occurs when pivoting on the smallest element at each step, such as applying quickselect for the maximum element to already sorted data and taking the first element as pivot each time.

If one instead consistently chooses "good" pivots, this is avoided and one always gets linear performance even in the worst case. A "good" pivot is one for which we can establish that a constant proportion of elements fall both below and above it, as then the search set decreases at least by a constant proportion at each step, hence exponentially quickly, and the overall time remains linear. The median is a good pivot – the best for sorting, and the best overall choice for selection – decreasing the search set by half at each step. Thus if one can compute the median in linear time, this only adds linear time to each step, and thus the overall complexity of the algorithm remains linear.

The median-of-medians algorithm computes an approximate median, namely a point that is guaranteed to be between the 30th and 70th percentiles (in the middle 4 deciles). Thus the search set decreases by at least 30%. The problem is reduced to 70% of the original size, which is a fixed proportion smaller. Applying the same algorithm on the now smaller set recursively until only one or two elements remain results in a cost of

(n)/(1-0 . 7) ≈ 3 . 33 {n}

Algorithm

As stated before, median-of-medians is used as a pivot selection strategy in the quickselect algorithm, which in pseudocode looks as follows. function nthSmallest(list, n) index := select(list, 1, length of list, n) // Use select(list, 0, length of list - 1, n - 1) if list is zero-based numbering return list[index]

Be careful to handle left, right and n when implementing. The following pseudocode assumes that left, right, and the list use one-based numbering and that select is initially called with 1 as the argument to left and the length of the list as the argument to right. Note that this returns the index of the n'th smallest number after rearranging the list, rather than the actual value of the n'th smallest number.

function select(list, left, right, n) loop if left = right then return left pivotIndex := pivot(list, left, right) pivotIndex := partition(list, left, right, pivotIndex, n) if n = pivotIndex then return n else if n < pivotIndex then right := pivotIndex - 1 else left := pivotIndex + 1

Subroutine is the actual median-of-medians algorithm. It divides its input (a list of length) into groups of at most five elements, computes the median of each of those groups using some subroutine, then recursively computes the true median of the

(n)/(5) medians found in the previous step:. Note that calls ; this is an instance of mutual recursion.

function pivot(list, left, right) // for 5 or less elements just get median if right − left < 5 then return partition5(list, left, right) // otherwise move the medians of five-element subgroups to the first n/5 positions for i from left to right in steps of 5 // get the median position of the i'th five-element subgroup subRight := i + 4 if subRight > right then subRight := right median5 := partition5(list, i, subRight) swap list[median5] and list[left + floor((i − left) / 5)] // compute the median of the n/5 medians-of-five mid := floor((right − left) / 10) + left + 1 return select(list, left, left + floor((right − left) / 5), mid)

Partition helper functions

See also: Dutch national flag problem.

There is a subroutine called that can, in linear time, group a list (ranging from indices left to right) into three parts, those less than a certain element, those equal to it, and those greater than the element (a three-way partition). The grouping into three parts ensures that the median-of-medians maintains linear execution time in a case of many or all coincident elements. Here is pseudocode that performs a partition about the element list[pivotIndex]:

function partition(list, left, right, pivotIndex, n) pivotValue := list[pivotIndex] swap list[pivotIndex] and list[right] // Move pivot to end storeIndex := left // Move all elements smaller than the pivot to the left of the pivot for i from left to right − 1 do if list[i] < pivotValue then swap list[storeIndex] and list[i] increment storeIndex // Move all elements equal to the pivot right after // the smaller elements storeIndexEq := storeIndex for i from storeIndex to right − 1 do if list[i] = pivotValue then swap list[storeIndexEq] and list[i] increment storeIndexEq swap list[right] and list[storeIndexEq] // Move pivot to its final place // Return location of pivot considering the desired location n if n < storeIndex then return storeIndex // n is in the group of smaller elements if n ≤ storeIndexEq then return n // n is in the group equal to pivot return storeIndexEq // n is in the group of larger elements

The subroutine selects the median of a group of at most five elements; an easy way to implement this is insertion sort, as shown below. It can also be implemented as a decision tree.

function partition5(list, left, right) i := left + 1 while i ≤ right do j := i while j > left and list[j−1] > list[j] do swap list[j−1] and list[j] j := j − 1 i := i + 1 return left + (right - left) / 2

Properties of pivot

Of the

(n)/(5) groups, half the number of groups \left( (1)/(x 2) (n)/(= 5) (n)/(10) \right) have their median less than the pivot (Median of Medians). Also, another half the number of groups \left( again, (1)/(x 2) (n)/(= 5) (n)/(10) \right) have their median greater than the pivot. In each of the (n)/(10) groups with median less than the pivot, there are two elements that are smaller than their respective medians, which are smaller than the pivot. Thus, each of the (n)/(10) groups have at least 3 elements that are smaller than the pivot. Similarly, in each of the (n)/(10) groups with median greater than the pivot, there are two elements that are greater than their respective medians, which are greater than the pivot. Thus, each of the (n)/(10) groups have at least 3 elements that are greater than the pivot. Hence, the pivot is less than 3 x (n)/(10) elements and greater than another 3 x (n)/(10) elements. Thus the chosen median splits the ordered elements somewhere between 30%/70% and 70%/30%, which assures worst-case linear behavior of the algorithm. To visualize:

bgcolor=gray

12		bgcolor=gray	15		bgcolor=gray	11		bgcolor=gray	2		bgcolor=gray	9		bgcolor=gray	5		bgcolor=gray	0		bgcolor=gray	7		bgcolor=gray	3		bgcolor=gray	21		bgcolor=gray	44		bgcolor=gray	40		bgcolor=gray	1		bgcolor=gray	18		bgcolor=gray	20		bgcolor=gray	32		bgcolor=gray	19		bgcolor=gray	35		bgcolor=gray	37		bgcolor=gray	39
	bgcolor=gray	13		bgcolor=gray	16		bgcolor=gray	14		bgcolor=gray	8		bgcolor=gray	10		bgcolor=gray	26		bgcolor=gray	6		bgcolor=gray	33		bgcolor=gray	4		bgcolor=gray	27		bgcolor=white	49		bgcolor=gray	46		bgcolor=white	52		bgcolor=gray	25		bgcolor=white	51		bgcolor=gray	34		bgcolor=gray	43		bgcolor=white	56		bgcolor=white	72		bgcolor=white	79
Medians	bgcolor=gray	17		bgcolor=gray	23		bgcolor=gray	24		bgcolor=gray	28		bgcolor=gray	29		bgcolor=gray	30		bgcolor=gray	31		bgcolor=gray	36		bgcolor=gray	42		bgcolor=red	47		bgcolor=white	50		bgcolor=white	55		bgcolor=white	58		bgcolor=white	60		bgcolor=white	63		bgcolor=white	65		bgcolor=white	66		bgcolor=white	67		bgcolor=white	81		bgcolor=white	83
	bgcolor=gray	22		bgcolor=gray	45		bgcolor=gray	38		bgcolor=white	53		bgcolor=white	61		bgcolor=gray	41		bgcolor=white	62		bgcolor=white	82		bgcolor=white	54		bgcolor=white	48		bgcolor=white	59		bgcolor=white	57		bgcolor=white	71		bgcolor=white	78		bgcolor=white	64		bgcolor=white	80		bgcolor=white	70		bgcolor=white	76		bgcolor=white	85		bgcolor=white	87
	bgcolor=white	96		bgcolor=white	95		bgcolor=white	94		bgcolor=white	86		bgcolor=white	89		bgcolor=white	69		bgcolor=white	68		bgcolor=white	97		bgcolor=white	73		bgcolor=white	92		bgcolor=white	74		bgcolor=white	88		bgcolor=white	99		bgcolor=white	84		bgcolor=white	75		bgcolor=white	90		bgcolor=white	77		bgcolor=white	93		bgcolor=white	98		bgcolor=white	91

(red = "(one of the two possible) median of medians", gray = "number < red", white = "number > red")

5-tuples are shown here sorted by median, for clarity. Sorting the tuples is not necessary because we only need the median for use as pivot element.

Note that all elements above/left of the red (30% of the 100 elements) are less, and all elements below/right of the red (another 30% of the 100 elements) are greater.

Proof of linear running time

The median-calculating recursive call does not exceed worst-case linear behavior because the list of medians has size

(n)/(5) , while the other recursive call recurses on at most 70% of the list. Let T(n) be the time it takes to run a median-of-medians Quickselect algorithm on an array of size n . Then we know this time is:T(n) \leq T(n/ 5 ) + T(n ⋅ 7 / 10 ) + c ⋅ n,

where

• the

T\left( (n)/(5) \right) part is for finding the true median of the (n)/(5) medians, by running an (independent) Quickselect on them (since finding the median is just a special case of selecting a k-smallest element)

• the

O(n) term c ⋅ n is for the partitioning work to create the two sides, one of which our Quickselect will recurse (we visited each element a constant number of times in order to form them into n/5 groups and take each median in O( 1 ) time).

• the

T\left( (7 n)/(10) \right) part is for the actual Quickselect recursion (for the worst case, in which the k-th element is in the bigger partition that can be of size (7 n)/(10) maximally)

By induction:

T(n) \leq 10 ⋅ c ⋅ n \in O(n).

Analysis

The key step is reducing the problem to selecting in two lists whose total length is shorter than the original list, plus a linear factor for the reduction step. This allows a simple induction to show that the overall running time is linear.

The specific choice of groups of five elements is explained as follows. Firstly, computing median of an odd list is faster and simpler; while one could use an even list, this requires taking the average of the two middle elements, which is slower than simply selecting the single exact middle element. Secondly, five is the smallest odd number such that median of medians works. With groups of only three elements, the resulting list of medians to search in is length

(n)/(3) , and reduces the list to recurse into length (2)/(3) n , since it is greater than 1/2 × 2/3 = 1/3 of the elements and less than 1/2 × 2/3 = 1/3 of the elements. Thus this still leaves n elements to search in, not reducing the problem sufficiently. The individual lists are shorter, however, and one can bound the resulting complexity to O(n log n) by the Akra–Bazzi method, but it does not prove linearity.

Conversely, one may instead group by

g = seven, nine, or more elements, and this does work. This reduces the size of the list of medians to (n)/(g) , and the size of the list to recurse into asymptotes at 3n/4 (75%), as the quadrants in the above table approximate 25%, as the size of the overlapping lines decreases proportionally. This reduces the scaling factor from 10 asymptotically to 4, but accordingly raises the c term for the partitioning work. Finding the median of a larger group takes longer, but is a constant factor (depending only on g ), and thus does not change the overall performance as n grows. In fact, considering the number of comparisons in the worst case, the constant factor is (2 g(g -1 ))/(g -3) .

If one instead groups the other way, say dividing the

n element list into 5 lists, computing the median of each, and then computing the median of these – i.e., grouping by a constant fraction, not a constant number – one does not as clearly reduce the problem, since it requires computing 5 medians, each in a list of (n)/(5) elements, and then recursing on a list of length at most (7)/(10) n . As with grouping by 3, the individual lists are shorter, but the overall length is no shorter – in fact longer – and thus one can only prove superlinear bounds. Grouping into a square of \sqrt{n} lists of length \sqrt{n} is similarly complicated.

References

• Blum . M. . Manuel Blum. Floyd . R. W. . Robert W. Floyd. Pratt . V. R. . Vaughan Pratt. Rivest . R. L. . Ron Rivest. Tarjan . R. E. . Robert Tarjan . Time bounds for selection . 10.1016/S0022-0000(73)80033-9 . Journal of Computer and System Sciences . 7 . 4 . 448–461 . August 1973 . free .

External links

• "Lecture notes for January 30, 1996: Deterministic selection", ICS 161: Design and Analysis of Algorithms, David Eppstein