Advanced Data Structures and Implementation

Top-Down Splay Trees Red-Black Trees Top-Down Red Black Trees Top-Down Deletion Deterministic Skip Lists AA-Trees Treaps k-d Trees Pairing Heaps

Top-Down Splay Tree

Direct strategy requires traversal from the root

down the tree, and then bottom-up traversal to implement the splaying tree. Can implement by storing parent links, or by storing the access path on a stack. Both methods require large amount of overhead and must handle many special cases. Initial rotations on the initial access path uses only O(1) extra space, but retains the O(log N) amortized time bound.

Case 1: Zig

Y X

Y
XR
YL Yr

YL

Yr

XR

If Y should become root, then X and its right sub tree are made left children of the smallest value in R, and Y is made root of center tree. Y does not have to be a leaf for the Zig case to apply.

Case 2: Zig-Zig

L Y
Z

X
XR

Y
ZL Zr

X
YR XR

YR

ZL

Zr

The value to be splayed is in the tree rooted at Z. Rotate Y about X and attach as left child of smallest value in R

Case 3: Zig-Zag (Simplified) L Y

YL

X
XR

L Y
YL

X
Z
XR

ZL

Zr

ZL

Zr

The value to be splayed is in the tree rooted at Z. To make code simpler, the Zig-Zag rotation is reduced to a single Zig. This results in more iterations in the splay process.

Splay Trees Implementation

See page 457 for generic SplayTree class
implementation. Page 458 shows Top-Down splaying method. Refer to page 459 for insertion method into the Top-Down Splay Tree. Page 460 shows deletion method from the tree.

Reassembling the Splay Tree L

X R L
XL XR

R
XL XR

When the value to be splayed to the root is at the root of the center tree, we have reached the point where we are ready to reassemble the tree. This is accomplished by a) making XL the right child of the maximum element in L, b) making XR the left child of the minimum element in R, and then making L and R the left and right children of X

Operation 1: Zig-Zig
L A B Ar Br Dl Cr E Fl F Fl Gl G H Er Gl Hl Xl X Xl Xr X Xr F G H R L C B D Cr A R

C
D Dl

E
Er Br Ar

Hl

Rotate B around A and make L child of minimum element in R (which is now empty)

L is still empty, and R is now the tree rooted at B. Note that R contains nodes > X but not in the right subtree of X.

Operation 2: Zig-Zag
L D Dl F E Er G H Br C Cr B A Ar Dl Fl Gl R L D F G H E Er C Cr X Xl X Xl Xr Xr Br B A Ar R

Fl Gl

Hl

Hl

Just perform Zig (simplified Zig-Zag)

L was previously empty and it now consists of node D and Ds left subtree

After X reaches root:

L D Dl F Fl Hl Gl G H Cr E Br Ar Xl X Xr C B A R

Er

This configuration was achieved by doing Zig Zig (of F, G) followed by a Zig (node H)

Reassemble XL becomes right sub tree of H, XR becomes left sub tree of E, and then L, R reattached to X
B

D Dl F Fl Hl Gl C G H Xl Cr E Er Br Ar A

Note that this is not the same tree as was obtained by doing BU splaying.

Red-Black Tree


Popular alternative to the AVL tree. Operations take O(log N) time in worst case. Height is at most 2log(N+1). A red-black tree is a binary search tree with one extra attribute for each node: the color, which is either red or black. The root is black. If node is red, its children must be black. Every path from a node to a null reference must contain the same number of black nodes. Basic operations to conform with rules are color changes and tree rotations.

Theorem 1 In a red-black tree, at least half the nodes on any path from the root to a leaf must be black. Proof If there is a red node on the path, there must be a corresponding black node.

Theorem 2 In a red-black tree, no path from any node, N, to a leaf is more than twice as long as any other path from N to any other leaf.
Proof: By definition, every path from a node to any leaf contains the same number of black nodes. By Theorem1, a least the nodes on any such path are black. Therefore, there can no more than twice as many nodes on any path from N to a leaf as on any other path. Therefore the length of every path is no more than twice as long as any other path.

Theorem 3 A red-black tree with n internal nodes has height h <= 2 lg(n + 1). Proof: Let h be the height of the red-black tree with root x. By Theorem 1, bh(x) >= h/2 From Theorem 1, n >= 2bh(x) - 1 Therefore n >= 2 h/2 1 n + 1 >= 2h/2 lg(n + 1) >= h/2 2lg(n + 1) >= h

Bottom-Up Insertion
Cases: 0: X is the root color it black 1: Both parent and uncle are red color parent and

uncle black, color grandparent red, point X to grandparent, check new situation 2 (zig-zag): Parent is red, but uncle is black. X and its parent are opposite type children color grandparent red, color X black, rotate left on parent, rotate right on grandparent 3 (zig-zig): Parent is red, but uncle is black. X and its parent are both left or both right children color parent black, color grandparent red, rotate right on grandparent

Top-Down Red-Black Trees

In T-Down insertion, the corrections are done

while traversing down the tree to the insertion point. When the actual insertion is done, no further corrections are needed, so no need to traverse back up the tree. So, T-Down insertion can be done iteratively which is generally faster. Insertion is always done as a leaf (as in ordinary BST insertion).

Process
On the way down, when we see a node X that

has two red children, we make X red and its two children black. If Xs parent is red, we can apply either the single or double rotation to keep us from having two consecutive red nodes. Xs parent and the parents sibling cannot both be red, since their colors would already have been flipped in that case.

Example: Insert 45
30
15 10 Two red children 40 20 50 55 70

60
65 80

85
90

Example (Cont.)
30 15 10 20 50 40 55 60 65 80 70 85 90

flip colors two red nodes

Example (Cont.): Do a single rotation

30 15 10 20 50 60 70 65 80 85 90

40
5

55

Example (Cont.): Now Insert 45

30 15 10 20 50 60 70 65 80 45 85 90

40
5

55

Note
Since the parent of the newly inserted node was

black, we are done. Had the parent of the inserted node been red, one more rotation would have had to be performed. Although red-black trees have slightly weaker balancing properties, their performance in experimentally almost identical to that of AVL trees.

Implementation of Top-Down RedBlack Trees

See pages 464-467

Top-Down Deletions
Recall that in deleting from a binary search tree, the only

nodes which are actually removed are leaves or nodes with exactly one child. Nodes with two children are never removed. Their contents are just replaced. If the node to be deleted is red, there is no problem just delete the node. If the node to be deleted is black, its removal will violate property. The solution is to ensure that any node to be deleted is red.

Deterministic Skip Lists

A probabilistically balanced
linked list. Invented in 1986 by William Pugh. Definition: Two elements are linked if there exists at least one link going from one to another. Definition: The gap size between two elements linked at height h is equal to the number of elements of height h-1 between them.

Skip List
d) xtra pointers every eighth item - full structure NIL

9
12

21
17

19

25

26

e) skip list - same link distribution, random choice 6 3 7 NIL 26

12

17

19

25
21

Search time
In the deterministic version (a-d):
in in in in a, we need to check at most n nodes b, at most n/2+1 nodes c, at most n/4+2 nodes general, at most log N nodes

Efficient search, but impractical insertion

and deletion.

Levels
A node with k forward level
pointers is called a level k node. If every (2i)th node has a pointer 2i nodes ahead, they have the following distribution: 1 2 3

percent
50 25 12.5

Central idea in skip lists

Choose levels of nodes randomly, but in
the same proportions (as in e). A nodes i th forward pointer, points to the
next node of level i or higher. Insertions and deletions require only local modifications. A nodes level never changes after first being chosen.

Insertion
To perform insertion, we must make sure
that when a new node of height h is added, it doesnt create a gap of four heights of h node (in 1-2-3 deterministic skip list). See page 269 fig. 12.19 For implementation of Skip List see pages 472-474.

AA-Trees
Also known as binary B-tree (BB-tree). BB-tree is a red-black tree with one extra condition: any

node may have at most one red child. Some conditions to make it simpler (p.475): - only right child can be red - code functions recursively - instead of color store information in small integer: - one if the node is a leaf - the level of its parent, if the node is red - one less then the level of its parent, if the node is black

Advantages
AA-trees simplify algorithms by:
- eliminating half of the restructuring cases - simplifying deletion by removing an annoying case if an internal node has only one child, that child must be a red right child We can always replace a node with the smallest child in the right sub tree (it will either be a leaf or have a red child)

Links in AA-tree
A horizontal link is a connection between a node

and a child with equal levels. Horizontal links are right references. There cannot be two consecutive horizontal links. Nodes at level 2 or higher must have two children. If a node has no right horizontal link, its two children are at the same level.

Example
30 70

15

50

60

85

10

20

35

40

55

65

80

90

Insertion in AA-tree
A new item is always inserted at the bottom

level. In the previous example, inserting 2 will create a horizontal left link. In the previous example, inserting 45 generates consecutive right links. After inserting at the bottom level, we may need to perform rotations to restore the horizontal link properties.

skew remove left horizontal links

split remove consecutive

horizontal links
X P

P
G X G

More on skew & split

skew removes a left horizontal link. skew might also create consecutive right

horizontal links. First we must apply skew and then use split, if necessary. After a split, the middle node increases a level, which may create a problem for the original parent.

Implementation of AA-trees
Refer to pages 476 480 for detailed
implementation techniques. See page 477 & 479 for more examples of left and right rotations of AA-trees.

Treaps
Binary search tree. Like skip list, it uses random numbers and
gives O (log N) expected time for any input. Slower than balanced search tree. Although deletion is much slower, it is still O (log N) expected time.

Definition of a treap
Each node stores an item, left and right link, and
a priority that is randomly assigned when the node is created. Treap is a binary search tree with the property that the node priorities satisfy heap order: any nodes priority must be at least as large as its parents. See pages 481-483 for implementation details.

k d Trees
Multidimensional b-tree. Branching of odd levels is
done with respect to the first key, and branching on even levels is done with respect to the second key. Root is arbitrary chosen to be an odd level. Can be visually represented:

Some facts about k-d trees

Can have any number of dimensions. In practice searches tend to be very

efficient. For a randomly constructed tree, the average running time of a partial match query is O (M+kN^(1-1/k)). See pages 484-485 for implementation details.

Pairing Heaps
A min (max) pairing heap is a min (max)
tree in which operations are done in a specified manner. 8
4 5 3 6

Insert
Create 1-element max tree with new item
and meld with existing max pairing heap.
9

+ insert(2) =

Insert (Cont.)
Create 1-element max tree with new item
and meld with existing max pairing heap.
14 9

+ insert(14) =
7 6 7 7

3 6 3

IncreaseKey (Node, theAmount)

Since nodes do not have parent fields, we
cannot easily check whether the key in the Node becomes larger than that in its parent. So, detach the Node from sibling doublylinked list and meld.

IncreaseKey (Node, theAmount)

9

theNode
6 4 2 6

If theNode is not the root,

1

remove sub tree rooted at theNode from its sibling list.

IncreaseKey (Node, theAmount)

9 18 2 6

6 3 1

IncreaseKey (Node, theAmount)

18 9 2 6 2 6 3 4 5 1 1 3

Pairing heaps
See pages 488 491 for implementation
details.