Prerequisites : Cartesian Tree

Cartesian Sort is an Adaptive Sorting as it sorts the data faster if data is partially sorted. In fact, there are very few sorting algorithms that make use of this fact.

For example consider the array {5, 10, 40, 30, 28}. The input data is partially sorted too as only one swap between “40” and “28” results in a completely sorted order. See how Cartesian Tree Sort will take advantage of this fact below.

Below are steps used for sorting.

**Step 1 : ** Build a (min-heap) Cartesian Tree from the given input sequence.

**Step 2 : ** Starting from the root of the built Cartesian Tree, we push the nodes in a priority queue.

Then we pop the node at the top of the priority queue and push the children of the popped node in the priority queue in a pre-order manner.

- Pop the node at the top of the priority queue and add it to a list.
- Push left child of the popped node first (if present).
- Push right child of the popped node next (if present).

**How to build (min-heap) Cartesian Tree?**

Building min-heap is similar to building a (max-heap) Cartesian Tree (discussed in previous post), except the fact that now we scan upward from the node’s parent up to the root of the tree until a node is found whose value is smaller (and not larger as in the case of a max-heap Cartesian Tree) than the current one and then accordingly reconfigure links to build the min-heap Cartesian tree.

**Why not to use only priority queue?**

One might wonder that using priority queue would anyway result in a sorted data if we simply insert the numbers of the input array one by one in the priority queue (i.e- without constructing the Cartesian tree).

But the time taken differs a lot.

Suppose we take the input array – {5, 10, 40, 30, 28}

If we simply insert the input array numbers one by one (without using a Cartesian tree), then we may have to waste a lot of operations in adjusting the queue order everytime we insert the numbers (just like a typical heap performs those operations when a new number is inserted, as priority queue is nothing but a heap).

Whereas, here we can see that using a Cartesian tree took only 5 operations (see the above two figures in which we are continuously pushing and popping the nodes of Cartesian tree), which is linear as there are 5 numbers in the input array also. So we see that the best case of Cartesian Tree sort is O(n), a thing where heap-sort will take much more number of operations, because it doesn’t make advantage of the fact that the input data is partially sorted.

**Why pre-order traversal?**

The answer to this is that since Cartesian Tree is basically a heap- data structure and hence follows all the properties of a heap. Thus the root node is always smaller than both of its children. Hence, we use a pre-order fashion popping-and-pushing as in this, the root node is always pushed earlier than its children inside the priority queue and since the root node is always less than both its child, so we don’t have to do extra operations inside the priority queue.

Refer to the below figure for better understanding-

// A C++ program to implement Cartesian Tree sort // Note that in this program we will build a min-heap // Cartesian Tree and not max-heap. #include<bits/stdc++.h> using namespace std; /* A binary tree node has data, pointer to left child and a pointer to right child */ struct Node { int data; Node *left, *right; }; // Creating a shortcut for int, Node* pair type typedef pair<int, Node*> iNPair; // This function sorts by pushing and popping the // Cartesian Tree nodes in a pre-order like fashion void pQBasedTraversal(Node* root) { // We will use a priority queue to sort the // partially-sorted data efficiently. // Unlike Heap, Cartesian tree makes use of // the fact that the data is partially sorted priority_queue <iNPair, vector<iNPair>, greater<iNPair>> pQueue; pQueue.push (make_pair (root->data,root)); // Resembles a pre-order traverse as first data // is printed then the left and then right child. while (! pQueue.empty()) { iNPair popped_pair = pQueue.top(); printf("%d ",popped_pair.first); pQueue.pop(); if (popped_pair.second->left != NULL) pQueue.push (make_pair(popped_pair.second->left->data, popped_pair.second->left)); if (popped_pair.second->right != NULL) pQueue.push (make_pair(popped_pair.second->right->data, popped_pair.second->right)); } return; } Node *buildCartesianTreeUtil(int root, int arr[], int parent[], int leftchild[], int rightchild[]) { if (root == -1) return NULL; Node *temp = new Node; temp->data = arr[root]; temp->left = buildCartesianTreeUtil(leftchild[root], arr, parent, leftchild, rightchild); temp->right = buildCartesianTreeUtil(rightchild[root], arr, parent, leftchild, rightchild); return temp ; } // A function to create the Cartesian Tree in O(N) time Node *buildCartesianTree(int arr[], int n) { // Arrays to hold the index of parent, left-child, // right-child of each number in the input array int parent[n],leftchild[n],rightchild[n]; // Initialize all array values as -1 memset(parent, -1, sizeof(parent)); memset(leftchild, -1, sizeof(leftchild)); memset(rightchild, -1, sizeof(rightchild)); // 'root' and 'last' stores the index of the root and the // last processed of the Cartesian Tree. // Initially we take root of the Cartesian Tree as the // first element of the input array. This can change // according to the algorithm int root = 0, last; // Starting from the second element of the input array // to the last on scan across the elements, adding them // one at a time. for (int i=1; i<=n-1; i++) { last = i-1; rightchild[i] = -1; // Scan upward from the node's parent up to // the root of the tree until a node is found // whose value is smaller than the current one // This is the same as Step 2 mentioned in the // algorithm while (arr[last] >= arr[i] && last != root) last = parent[last]; // arr[i] is the smallest element yet; make it // new root if (arr[last] >= arr[i]) { parent[root] = i; leftchild[i] = root; root = i; } // Just insert it else if (rightchild[last] == -1) { rightchild[last] = i; parent[i] = last; leftchild[i] = -1; } // Reconfigure links else { parent[rightchild[last]] = i; leftchild[i] = rightchild[last]; rightchild[last]= i; parent[i] = last; } } // Since the root of the Cartesian Tree has no // parent, so we assign it -1 parent[root] = -1; return (buildCartesianTreeUtil (root, arr, parent, leftchild, rightchild)); } // Sorts an input array int printSortedArr(int arr[], int n) { // Build a cartesian tree Node *root = buildCartesianTree(arr, n); printf("The sorted array is-\n"); // Do pr-order traversal and insert // in priority queue pQBasedTraversal(root); } /* Driver program to test above functions */ int main() { /* Given input array- {5,10,40,30,28}, it's corresponding unique Cartesian Tree is- 5 \ 10 \ 28 / 30 / 40 */ int arr[] = {5, 10, 40, 30, 28}; int n = sizeof(arr)/sizeof(arr[0]); printSortedArr(arr, n); return(0); }

Output :

The sorted array is- 5 10 28 30 40

**Time Complexity : ** **O(n)** best-case behaviour (when the input data is partially sorted), **O(n log n)** worst-case behavior (when the input data is not partially sorted)

**Auxiliary Space : **We use a priority queue and a Cartesian tree data structure. Now, at any moment of time the size of the priority queue doesn’t exceeds the size of the input array, as we are constantly pushing and popping the nodes. Hence we are using O(n) auxiliary space.

**References :**

https://en.wikipedia.org/wiki/Adaptive_sort

http://11011110.livejournal.com/283412.htmlhttp://gradbot.blogspot.in/2010/06/cartesian-tree-sort.htmlhttp://www.keithschwarz.com/interesting/code/?dir=cartesian-tree-sorthttps://en.wikipedia.org/wiki/Cartesian_tree#Application_in_sorting

This article is contributed by **Rachit Belwariar**. If you like GeeksforGeeks and would like to contribute, you can also write an article and mail your article to contribute@geeksforgeeks.org. See your article appearing on the GeeksforGeeks main page and help other Geeks.

Please write comments if you find anything incorrect, or you want to share more information about the topic discussed above