Heapsort

The following is a simple way to implement the algorithm in pseudocode. Arrays are zero-based and swap is used to exchange two elements of the array. Movement 'down' means from the root towards the leaves, or from lower indices to higher. Note that during the sort, the largest element is at the root of the heap at a[0], while at the end of the sort, the largest element is in a[end].

procedure heapsort(a, count) is
    input: an unordered array a of length count
 
    (Build the heap in array a so that largest value is at the root)
    heapify(a, count)

    (The following loop maintains the invariants that a[0:end] is a heap and every element
     beyond end is greater than everything before it (so a[end:count] is in sorted order))
    end ← count - 1
    while end > 0 do
        (a[0] is the root and largest value. The swap moves it in front of the sorted elements.)
        swap(a[end], a[0])
        (the heap size is reduced by one)
        end ← end - 1
        (the swap ruined the heap property, so restore it)
        siftDown(a, 0, end)

The sorting routine uses two subroutines, heapify and siftDown. The former is the common in-place heap construction routine, while the latter is a common subroutine for implementing heapify.

(Put elements of 'a' in heap order, in-place)
procedure heapify(a, count) is
    (start is assigned the index in 'a' of the last parent node)
    (the last element in a 0-based array is at index count-1; find the parent of that element)
    start ← iParent(count-1)
    
    while start ≥ 0 do
        (sift down the node at index 'start' to the proper place such that all nodes below
         the start index are in heap order)
        siftDown(a, start, count - 1)
        (go to the next parent node)
        start ← start - 1
    (after sifting down the root all nodes/elements are in heap order)

(Repair the heap whose root element is at index 'start', assuming the heaps rooted at its children are valid)
procedure siftDown(a, start, end) is
    root ← start

    while iLeftChild(root) ≤ end do    (While the root has at least one child)
        child ← iLeftChild(root)   (Left child of root)
        swap ← root                (Keeps track of child to swap with)

        if a[swap] < a[child] then
            swap ← child
        (If there is a right child and that child is greater)
        if child+1 ≤ end and a[swap] < a[child+1] then
            swap ← child + 1
        if swap = root then
            (The root holds the largest element. Since we assume the heaps rooted at the
             children are valid, this means that we are done.)
            return
        else
            Swap(a[root], a[swap])
            root ← swap          (repeat to continue sifting down the child now)

The heapify procedure can be thought of as building a heap from the bottom up by successively sifting downward to establish the heap property. An alternative version (shown below) that builds the heap top-down and sifts upward may be simpler to understand. This siftUp version can be visualized as starting with an empty heap and successively inserting elements, whereas the siftDown version given above treats the entire input array as a full but "broken" heap and "repairs" it starting from the last non-trivial sub-heap (that is, the last parent node).

Difference in time complexity between the "siftDown" version and the "siftUp" version.

Also, the siftDown version of heapify has O(n) time complexity, while the siftUp version given below has O(n log n) time complexity due to its equivalence with inserting each element, one at a time, into an empty heap.[4] This may seem counter-intuitive since, at a glance, it is apparent that the former only makes half as many calls to its logarithmic-time sifting function as the latter; i.e., they seem to differ only by a constant factor, which never affects asymptotic analysis.

To grasp the intuition behind this difference in complexity, note that the number of swaps that may occur during any one siftUp call increases with the depth of the node on which the call is made. The crux is that there are many (exponentially many) more "deep" nodes than there are "shallow" nodes in a heap, so that siftUp may have its full logarithmic running-time on the approximately linear number of calls made on the nodes at or near the "bottom" of the heap. On the other hand, the number of swaps that may occur during any one siftDown call decreases as the depth of the node on which the call is made increases. Thus, when the siftDown heapify begins and is calling siftDown on the bottom and most numerous node-layers, each sifting call will incur, at most, a number of swaps equal to the "height" (from the bottom of the heap) of the node on which the sifting call is made. In other words, about half the calls to siftDown will have at most only one swap, then about a quarter of the calls will have at most two swaps, etc.

The heapsort algorithm itself has O(n log n) time complexity using either version of heapify.

 procedure heapify(a,count) is
     (end is assigned the index of the first (left) child of the root)
     end := 1
     
     while end < count
         (sift up the node at index end to the proper place such that all nodes above
          the end index are in heap order)
         siftUp(a, 0, end)
         end := end + 1
     (after sifting up the last node all nodes are in heap order)
 
 procedure siftUp(a, start, end) is
     input:  start represents the limit of how far up the heap to sift.
                   end is the node to sift up.
     child := end 
     while child > start
         parent := iParent(child)
         if a[parent] < a[child] then (out of max-heap order)
             swap(a[parent], a[child])
             child := parent (repeat to continue sifting up the parent now)
         else
             return

Note that unlike siftDown approach where, after each swap, you need to call only the siftDown subroutine to fix the broken heap; the siftUp subroutine alone cannot fix the broken heap. The heap needs to be built every time after a swap by calling the heapify procedure since "siftUp" assumes that the element getting swapped ends up in its final place, as opposed to "siftDown" allows for continuous adjustments of items lower in the heap until the invariant is satisfied. The adjusted pseudocode for using siftUp approach is given below.

 procedure heapsort(a, count) is
    input: an unordered array a of length count
 
    (Build the heap in array a so that largest value is at the root)
    heapify(a, count)

    (The following loop maintains the invariants that a[0:end] is a heap and every element
     beyond end is greater than everything before it (so a[end:count] is in sorted order))
    end ← count - 1
    while end > 0 do
        (a[0] is the root and largest value. The swap moves it in front of the sorted elements.)
        swap(a[end], a[0])
        (rebuild the heap using siftUp after the swap ruins the heap property)
        heapify(a, end)
        (reduce the heap size by one)
        end ← end - 1

The most important variation to the basic algorithm, which is included in all practical implementations, is a heap-construction algorithm by Floyd which runs in O(n) time and uses siftdown rather than siftup, avoiding the need to implement siftup at all.

Rather than starting with a trivial heap and repeatedly adding leaves, Floyd's algorithm starts with the leaves, observing that they are trivial but valid heaps by themselves, and then adds parents. Starting with element n/2 and working backwards, each internal node is made the root of a valid heap by sifting down. The last step is sifting down the first element, after which the entire array obeys the heap property.

The worst-case number of comparisons during the Floyd's heap-construction phase of heapsort is known to be equal to 2n − 2s₂(n) − e₂(n), where s₂(n) is the number of 1 bits in the binary representation of n and e₂(n) is number of trailing 0 bits.[5]

The standard implementation of Floyd's heap-construction algorithm causes a large number of cache misses once the size of the data exceeds that of the CPU cache.[6]^: 87 Much better performance on large data sets can be obtained by merging in depth-first order, combining subheaps as soon as possible, rather than combining all subheaps on one level before proceeding to the one above.[7][8]

Bottom-up heapsort is a variant which reduces the number of comparisons required by a significant factor. While ordinary heapsort requires 2n log₂ n + O(n) comparisons worst-case and on average,[9] the bottom-up variant requires n log₂n + O(1) comparisons on average,[9] and 1.5n log₂n + O(n) in the worst case.[10]

If comparisons are cheap (e.g. integer keys) then the difference is unimportant,[11] as top-down heapsort compares values that have already been loaded from memory. If, however, comparisons require a function call or other complex logic, then bottom-up heapsort is advantageous.

This is accomplished by improving the siftDown procedure. The change improves the linear-time heap-building phase somewhat,[12] but is more significant in the second phase. Like ordinary heapsort, each iteration of the second phase extracts the top of the heap, a[0], and fills the gap it leaves with a[end], then sifts this latter element down the heap. But this element comes from the lowest level of the heap, meaning it is one of the smallest elements in the heap, so the sift-down will likely take many steps to move it back down.[13] In ordinary heapsort, each step of the sift-down requires two comparisons, to find the minimum of three elements: the new node and its two children.

Bottom-up heapsort instead finds the path of largest children to the leaf level of the tree (as if it were inserting −∞) using only one comparison per level. Put another way, it finds a leaf which has the property that it and all of its ancestors are greater than or equal to their siblings. (In the absence of equal keys, this leaf is unique.) Then, from this leaf, it searches upward (using one comparison per level) for the correct position in that path to insert a[end]. This is the same location as ordinary heapsort finds, and requires the same number of exchanges to perform the insert, but fewer comparisons are required to find that location.[10]

Because it goes all the way to the bottom and then comes back up, it is called heapsort with bounce by some authors.[14]

function leafSearch(a, i, end) is
    j ← i
    while iRightChild(j) ≤ end do
        (Determine which of j's two children is the greater)
        if a[iRightChild(j)] > a[iLeftChild(j)] then
            j ← iRightChild(j)
        else
            j ← iLeftChild(j)
    (At the last level, there might be only one child)
    if iLeftChild(j) ≤ end then
        j ← iLeftChild(j)
    return j

The return value of the leafSearch is used in the modified siftDown routine:[10]

procedure siftDown(a, i, end) is
    j ← leafSearch(a, i, end)
    while a[i] > a[j] do
        j ← iParent(j)
    x ← a[j]
    a[j] ← a[i]
    while j > i do
        swap x, a[iParent(j)]
        j ← iParent(j)

Bottom-up heapsort was announced as beating quicksort (with median-of-three pivot selection) on arrays of size ≥16000.[9]

A 2008 re-evaluation of this algorithm showed it to be no faster than ordinary heapsort for integer keys, presumably because modern branch prediction nullifies the cost of the predictable comparisons which bottom-up heapsort manages to avoid.[11]

A further refinement does a binary search in the path to the selected leaf, and sorts in a worst case of (n+1)(log₂(n+1) + log₂ log₂(n+1) + 1.82) + O(log₂n) comparisons, approaching the information-theoretic lower bound of n log₂n − 1.4427n comparisons.[15]

A variant which uses two extra bits per internal node (n−1 bits total for an n-element heap) to cache information about which child is greater (two bits are required to store three cases: left, right, and unknown)[12] uses less than n log₂n + 1.1n compares.[16]

• Ternary heapsort uses a ternary heap instead of a binary heap; that is, each element in the heap has three children. It is more complicated to program, but does a constant number of times fewer swap and comparison operations. This is because each sift-down step in a ternary heap requires three comparisons and one swap, whereas in a binary heap two comparisons and one swap are required. Two levels in a ternary heap cover 3² = 9 elements, doing more work with the same number of comparisons as three levels in the binary heap, which only cover 2³ = 8. This is primarily of academic interest, or as a student exercise,[17] as the additional complexity is not worth the minor savings, and bottom-up heapsort beats both.
• Memory-optimized heapsort[6]^: 87 improves heapsort's locality of reference by increasing the number of children even more. This increases the number of comparisons, but because all children are stored consecutively in memory, reduces the number of cache lines accessed during heap traversal, a net performance improvement.
• Out-of-place heapsort[18][19][13] improves on bottom-up heapsort by eliminating the worst case, guaranteeing n log₂n + O(n) comparisons. When the maximum is taken, rather than fill the vacated space with an unsorted data value, fill it with a −∞ sentinel value, which never "bounces" back up. It turns out that this can be used as a primitive in an in-place (and non-recursive) "QuickHeapsort" algorithm.[20] First, you perform a quicksort-like partitioning pass, but reversing the order of the partitioned data in the array. Suppose (without loss of generality) that the smaller partition is the one greater than the pivot, which should go at the end of the array, but our reversed partitioning step places it at the beginning. Form a heap out of the smaller partition and do out-of-place heapsort on it, exchanging the extracted maxima with values from the end of the array. These are less than the pivot, meaning less than any value in the heap, so serve as −∞ sentinel values. Once the heapsort is complete (and the pivot moved to just before the now-sorted end of the array), the order of the partitions has been reversed, and the larger partition at the beginning of the array may be sorted in the same way. (Because there is no non-tail recursion, this also eliminates quicksort's O(log n) stack usage.)
• The smoothsort algorithm[21] is a variation of heapsort developed by Edsger W. Dijkstra in 1981. Like heapsort, smoothsort's upper bound is O(n log n). The advantage of smoothsort is that it comes closer to O(n) time if the input is already sorted to some degree, whereas heapsort averages O(n log n) regardless of the initial sorted state. Due to its complexity, smoothsort is rarely used.
• Levcopoulos and Petersson[22] describe a variation of heapsort based on a heap of Cartesian trees. First, a Cartesian tree is built from the input in O(n) time, and its root is placed in a 1-element binary heap. Then we repeatedly extract the minimum from the binary heap, output the tree's root element, and add its left and right children (if any) which are themselves Cartesian trees, to the binary heap.[23] As they show, if the input is already nearly sorted, the Cartesian trees will be very unbalanced, with few nodes having left and right children, resulting in the binary heap remaining small, and allowing the algorithm to sort more quickly than O(n log n) for inputs that are already nearly sorted.
• Several variants such as weak heapsort require n log₂ n+O(1) comparisons in the worst case, close to the theoretical minimum, using one extra bit of state per node. While this extra bit makes the algorithms not truly in-place, if space for it can be found inside the element, these algorithms are simple and efficient,[7]^: 40 but still slower than binary heaps if key comparisons are cheap enough (e.g. integer keys) that a constant factor does not matter.[24]
• Katajainen's "ultimate heapsort" requires no extra storage, performs n log₂ n+O(1) comparisons, and a similar number of element moves.[25] It is, however, even more complex and not justified unless comparisons are very expensive.