Augmenting AVL trees
How we’ve thought about trees so far
Good for determining ancestry
Can be good for quickly finding an element
Other kinds of uses? • Any thoughts?
• Finding a minimum/maximum…
– (heaps are probably just as good or better)
• Finding an average?
• More complicated things?!!!11one
Enter: idea of augmenting a tree
Augmenting
• Can quickly compute many global properties that seem to need knowledge of the whole tree!
• Examples:
– size of any sub-tree
– height of any sub-tree
– averages of keys/values in a sub-tree
– min+max of keys/values in any sub-tree, …
• Can quickly compute any function f(u) so long as you only need to know f(u.left) and f(u.right)!
Augmenting an AVL tree
• Can augment any kind of tree
• Only balanced trees are guaranteed to be fast
• AfteraugmentinganAVLtreetocompute f(u), we can still do all operations in O(lg n)!
Simple first example
• We are going to do one simple example
A regular AVL tree
• Then, you will help with a harder one! already does this
• Problem: augment an AVL tree so we can do:
– Insert(key): add key in O(lg n)
– Delete(key): remove key in O(lg n)
– Height(node): get height of sub-tree rooted at node in O(1)
How do we do this?
Store some extra data at each node… but what?
Can we compute this function quickly? • Function we want to compute: Height(u) = H(u)
• If someone gives us H(uL) and H(uR), can we compute H(u)?
u
• What formula should we use?
• Ifuisaleafthen
uL
uR
– H(u) = 0 • Else
H(uL)
H(u)=?
– H(u) = max{H(uL), H(uR)}+1
H(uR)
Augmenting AVL tree to compute H(u)
• Each node u contains
– key: the key
– left, right: child pointers
– h: height of sub-tree rooted at u
• How? Insert(d)
d, 21301
d, 23
Insert(a) Insert(e) Insert(b) Insert(c)
2 a a a , , , 012
e , 0
b b , , 1 1
e , 0
b, 01
a, 20
c, 0
c, 0
The usual stuff…
Secret sauce!
Algorithm idea:
• From the last slide, we develop an algorithm
Insert(key):
1 2 3
BST search for where to put key
Insert key into place like in a regular AVL tree
Fix balance factors and rotate as you would in AVL insert, but fix heights at the same time.
(Remember to fix heights all the way to the root. Don’t stop before reaching the root!)
• (When you rotate, remember to fix heights of all nodes involved, just like you fix balance factors!)
Harder problem: scheduling conflicts
• Your calendar contains a bunch of time intervals [lo,hi] where you are busy
• Wewanttobeableto quickly tell whether a new booking conflicts with an earlier booking.
Breaking the problem down
• YoumustdesignadatastructureDto efficiently do:
– Insert(D; x): Insert interval x into D.
– Delete(D; x): Delete interval x from D.
– Search(D; x): If D contains an interval that overlaps with x, return any such interval.
Otherwise, return null.
The hard part
• AllfunctionsmustruninO(lgn)
Figuring out the data structure – 1
• Iterativeprocess;HARDtogetrightthefirsttime! • Need a way to insert intervals into the tree
– Use low end-point of interval as the key • Example tree:
8, 16
29, 36
48, 52
26, 36
60, 80
30, 34
Figuring out the data structure – 2
• Whatfunctiondowewanttocompute?
– Does an interval x intersect any interval in the tree?
• Whatinfoshouldwestoreateachnodeu?
– Mhi(u) = Maximum high endpoint of any node in the
subtree.
• Howcanweusetheinfostoredateachnode to compute the desired function (by looking at a small number of nodes)?
• Start by computing whether an interval x intersects any interval in a subtree.
Algorithm for Search within a subtree Returns an interval in the subtree rooted at u
Search(lo, hi, u): that intersects [lo, hi] if u is null then return null
Algorithm for Search within a subtree Returns an interval in the subtree rooted at u
Search(lo, hi, u): that intersects [lo, hi]
if u is null then return null
if [lo, hi] intersects [lo(u), hi(u)] then return [lo(u), hi(u)]
Algorithm for Search within a subtree Returns an interval in the subtree rooted at u
Search(lo, hi, u): that intersects [lo, hi]
if u is null then return null
if [lo, hi] intersects [lo(u), hi(u)] then return [lo(u), hi(u)] else (no intersection)
if lo < lo(u) return Search(lo, hi, left(u))
lo hi
Every node v on this hi
lo(u) hi(u)
u
side has lo(v) > hi
Algorithm for Search within a subtree Returns an interval in the subtree rooted at u
Search(lo, hi, u): that intersects [lo, hi]
if u is null then return null
if [lo, hi] intersects [lo(u), hi(u)] then return [lo(u), hi(u)] else (no intersection)
if lo < lo(u) return Search(lo, hi, left(u)) else (lo ≥ lo(u))
Every node v on this side has hi(v) < lo
if lo > Mhi(left(u)) then return Search(lo, hi, right(u))
u
lo(u) hi(u)
lo k lo hi
Algorithm for Search within a subtree Returns an interval in the subtree rooted at u
Search(lo, hi, u): that intersects [lo, hi]
if u is null then return null
if [lo, hi] intersects [lo(u), hi(u)] then return [lo(u), hi(u)]
else (no intersection)
if lo < lo(u) return Search(lo, hi, left(u))
lo(u) hi(u)
else (lo ≥ lo(u))
if lo > Mhi(left(u)) then return Search(lo, hi, right(u))
else (lo ≤ Mhi(left(u))
return Search(lo, hi, left(u))
v
lo hi lo(v) hi(v) = Mhi(left(u))
u
Search(lo, hi, u):
Final algorithm for Search
if u is null then return null
if [lo, hi] intersects [lo(u), hi(u)] then return [lo(u), hi(u)] else (no intersection)
if lo < lo(u) return Search(lo, hi, left(u)) else (lo ≥ lo(u))
if lo > Mhi(left(u)) then return Search(lo, hi, right(u)) else (lo ≤ Mhi(left(u))
Search(D, x=[lo, hi]):
return Search(lo, hi, left(u))
return Search(lo, hi, root(D))
Algorithms for Insert and Delete
• Insert(D, x=[lo, hi]):
– Do regular AVL insertion of key lo, also storing hi.
– Set Mhi of the new node to hi.
– Fix balance factors and perform rotations as usual, but also update Mhi(u) whenever you update the balance factor of a node u.
– Update Mhi(u) for all ancestors, and for every node involved in a rotation, using formula: Mhi(u) = max{hi(u), Mhi(left(u)), Mhi(right(u))}.
• Delete(D, x=[lo, hi]): similar to Insert
Why O(lg n) time?
• Insert/Delete: normal AVL operation = O(lg n)
– PLUS: update Mhi(u) for each u on path to the root • Length of this path ≤ tree height, so O(lg n) in an AVL tree
– PLUS: update Mhi(u) for each node involved in a rotation
• At most O(lg n) rotations (one per node on the path from the root ≤ tree height)
• Each rotation involves a constant number of nodes • Therefore, constant times O(lg n), which is O(lg n).
• Search
– Constant work + recursive call on a child
– Single recursive call means O(tree height) = O(lg n)