AA tree · Lablypedia

AA trees are a variation of the red–black tree, a form of binary search tree which supports efficient addition and deletion of entries. Unlike red–black trees, red nodes on an AA tree can only be added as a right subchild. In other words, no red node can be a left sub-child. This results in the simulation of a 2–3 tree instead of a 2–3–4 tree, which greatly simplifies the maintenance operations. The maintenance algorithms for a red–black tree need to consider seven different shapes to properly balance the tree:

An AA tree on the other hand only needs to consider two shapes due to the strict requirement that only right links can be red:

Balancing rotations

Whereas red–black trees require one bit of balancing metadata per node (the color), AA trees require O(log(log(N))) bits of metadata per node, in the form of an integer "level". The following invariants hold for AA trees:

The level of every leaf node is one.
The level of every left child is exactly one less than that of its parent.
The level of every right child is equal to or one less than that of its parent.
The level of every right grandchild is strictly less than that of its grandparent.
Every node of level greater than one has two children.

A link where the child's level is equal to that of its parent is called a horizontal link, and is analogous to a red link in the red–black tree. Individual right horizontal links are allowed, but consecutive ones are forbidden; all left horizontal links are forbidden. These are more restrictive constraints than the analogous ones on red–black trees, with the result that re-balancing an AA tree is procedurally much simpler than re-balancing a red–black tree.

Insertions and deletions may transiently cause an AA tree to become unbalanced (that is, to violate the AA tree invariants). Only two distinct operations are needed for restoring balance: "skew" and "split". Skew is a right rotation to replace a subtree containing a left horizontal link with one containing a right horizontal link instead. Split is a left rotation and level increase to replace a subtree containing two or more consecutive right horizontal links with one containing two fewer consecutive right horizontal links. Implementation of balance-preserving insertion and deletion is simplified by relying on the skew and split operations to modify the tree only if needed, instead of making their callers decide whether to skew or split.

<b id="mwKA">function</b> skew <b id="mwKQ">is</b> <b id="mwKg">input:</b> T, a node representing an AA tree that needs to be rebalanced. <b id="mwKw">output:</b> Another node representing the rebalanced AA tree. <b id="mwLA">if</b> nil(T) <b id="mwLQ">then</b> <b id="mwLg">return</b> Nil <b id="mwLw">else if</b> nil(left(T)) <b id="mwMA">then</b> <b id="mwMQ">return</b> T <b id="mwMg">else if</b> level(left(T)) == level(T) <b id="mwMw">then</b> <i id="mwNA">Swap the pointers of horizontal left links.</i> L = left(T) left(T) := right(L) right(L) := T <b id="mwNQ">return</b> L <b id="mwNg">else</b> <b id="mwNw">return</b> T <b id="mwOA">end if</b> <b id="mwOQ">end function</b>

Skew:

<b id="mwPw">function</b> split <b id="mwQA">is</b> <b id="mwQQ">input:</b> T, a node representing an AA tree that needs to be rebalanced. <b id="mwQg">output:</b> Another node representing the rebalanced AA tree. <b id="mwQw">if</b> nil(T) <b id="mwRA">then</b> <b id="mwRQ">return</b> Nil <b id="mwRg">else if</b> nil(right(T)) <b id="mwRw">or </b> nil(right(right(T))) <b id="mwSA">then</b> <b id="mwSQ">return</b> T <b id="mwSg">else if</b> level(T) == level(right(right(T))) <b id="mwSw">then</b> <i id="mwTA">We have two horizontal right links. Take the middle node, elevate it, and return it.</i> R = right(T) right(T) := left(R) left(R) := T level(R) := level(R) + 1 <b id="mwTQ">return</b> R <b id="mwTg">else</b> <b id="mwTw">return</b> T <b id="mwUA">end if</b> <b id="mwUQ">end function</b>

Split:

3 more chapters are free with a Lably account.

Insertion

Insertion begins with the normal binary tree search and insertion procedure. Then, as the call stack unwinds (assuming a recursive implementation of the search), it's easy to check the validity of the tree and perform any rotations as necessary. If a horizontal left link arises, a skew will be performed, and if two horizontal right links arise, a split will be performed, possibly incrementing the level of the new root node of the current subtree. Note, in the code as given above, the increment of level(T). This makes it necessary to continue checking the validity of the tree as the modifications bubble up from the leaves.

<b id="mwWQ">function</b> insert <b id="mwWg">is</b> <b id="mwWw">input:</b> X, the value to be inserted, and T, the root of the tree to insert it into. <b id="mwXA">output:</b> A balanced version T including X. <i id="mwXQ">Do the normal binary tree insertion procedure. Set the result of the</i> <i id="mwXg">recursive call to the correct child in case a new node was created or the</i> <i id="mwXw">root of the subtree changes.</i> <b id="mwYA">if</b> nil(T) <b id="mwYQ">then</b> <i id="mwYg">Create a new leaf node with X.</i> <b id="mwYw">return</b> node(X, 1, Nil, Nil) <b id="mwZA">else if</b> X < value(T) <b id="mwZQ">then</b> left(T) := insert(X, left(T)) <b id="mwZg">else if</b> X > value(T) <b id="mwZw">then</b> right(T) := insert(X, right(T)) <b id="mwaA">end if</b> <i id="mwaQ">Note that the case of X == value(T) is unspecified. As given, an insert</i> <i id="mwag">will have no effect. The implementor may desire different behavior.</i> <i id="mwaw">Perform skew and then split. The conditionals that determine whether or</i> <i id="mwbA">not a rotation will occur or not are inside of the procedures, as given</i> <i id="mwbQ">above.</i> T := skew(T) T := split(T) <b id="mwbg">return T</b> <b id="mwbw">end function</b>

Deletion

As in most balanced binary trees, the deletion of an internal node can be turned into the deletion of a leaf node by swapping the internal node with either its closest predecessor or successor, depending on which are in the tree or on the implementor's whims. Retrieving a predecessor is simply a matter of following one left link and then all of the remaining right links. Similarly, the successor can be found by going right once and left until a null pointer is found. Because of the AA property of all nodes of level greater than one having two children, the successor or predecessor node will be in level 1, making their removal trivial.

To re-balance a tree, there are a few approaches. The one described by Andersson in his original paper is the simplest, and it is described here, although actual implementations may opt for a more optimized approach. After a removal, the first step to maintaining tree validity is to lower the level of any nodes whose children are two levels below them, or who are missing children. Then, the entire level must be skewed and split. This approach was favored, because when laid down conceptually, it has three easily understood separate steps:

Decrease the level, if appropriate.
Skew the level.
Split the level.

However, we have to skew and split the entire level this time instead of just a node, complicating our code.

<b id="mweg">function</b> delete <b id="mwew">is</b> <b id="mwfA">input:</b> X, the value to delete, and T, the root of the tree from which it should be deleted. <b id="mwfQ">output:</b> T, balanced, without the value X. <b id="mwfg">if</b> nil(T) <b id="mwfw">then</b> return T <b id="mwgA">else if</b> X > value(T) <b id="mwgQ">then</b> right(T) := delete(X, right(T)) <b id="mwgg">else if</b> X < value(T) <b id="mwgw">then</b> left(T) := delete(X, left(T)) <b id="mwhA">else</b> <i id="mwhQ">If we're a leaf, easy, otherwise reduce to leaf case.</i> <b id="mwhg">if</b> leaf(T) <b id="mwhw">then</b> return right(T) <b id="mwiA">else if</b> nil(left(T)) <b id="mwiQ">then</b> L := successor(T) right(T) := delete(value(L), right(T)) value(T) := value(L) <b id="mwig">else</b> L := predecessor(T) left(T) := delete(value(L), left(T)) value(T) := value(L) <b id="mwiw">end if</b> <b id="mwjA">end if</b> <i id="mwjQ">Rebalance the tree. Decrease the level of all nodes in this level if</i> <i id="mwjg">necessary, and then skew and split all nodes in the new level.</i> T := decrease_level(T) T := skew(T) right(T) := skew(right(T)) <b id="mwjw">if not</b> nil(right(T)) right(right(T)) := skew(right(right(T))) <b id="mwkA">end if</b> T := split(T) right(T) := split(right(T)) return T <b id="mwkQ">end function</b>

<b id="mwkw">function</b> decrease_level <b id="mwlA">is</b> <b id="mwlQ">input:</b> T, a tree for which we want to remove links that skip levels. <b id="mwlg">output:</b> T with its level decreased. should_be = min(level(left(T)), level(right(T))) + 1 <b id="mwlw">if</b> should_be < level(T) <b id="mwmA">then</b> level(T) := should_be <b id="mwmQ">if</b> should_be < level(right(T)) <b id="mwmg">then</b> level(right(T)) := should_be <b id="mwmw">end if</b> <b id="mwnA">end if</b> return T <b id="mwnQ">end function</b>

A good example of deletion by this algorithm is present in the Andersson paper.

Performance

The performance of an AA tree is equivalent to the performance of a red–black tree. While an AA tree makes more rotations than a red–black tree, the simpler algorithms tend to be faster, and all of this balances out to result in similar performance. A red–black tree is more consistent in its performance than an AA tree, but an AA tree tends to be flatter, which results in slightly faster search times.