Lecture 4 van Emde Boas 6.

046J Spring 2015

Lecture 4: Divide and Conquer:

van Emde Boas Trees

• Series of Improved Data Structures

• Insert, Successor

• Delete

• Space

This lecture is based on personal communication with Michael Bender, 2001.

Goal
We want to maintain n elements in the range {0, 1, 2, . . . , u − 1} and perform Insert,
Delete and Successor operations in O(log log u) time.
O(1)
• If n = nO(1) or n(log n) , then we have O(log log n) time operations

– Exponentially faster than Balanced Binary Search Trees

– Cooler queries than hashing

• Application: Network Routing Tables

– u = Range of IP Addresses → port to send (u = 232 in IPv4)

Where might the O(log log u) bound arise ?

• Binary search over O(log u) elements

• Recurrences
log u
– T (log u) = T 2 + O(1)
√
– T (u) = T ( u) + O(1)

Improvements
We will develop the van Emde Boas data structure by a series of improvements on
a very simple data structure.

Lecture 4 van Emde Boas 6.046J Spring 2015

Bit Vector
We maintain a vector V of size u such that V [ x ] = 1 if and only if x is in the set.
Now, inserts and deletes can be performed by just flipping the corresponding bit
in the vector. However, successor/predecessor requires us to traverse through the
vector to find the next 1-bit.

• Insert/Delete: O(1)

• Successor/Predecessor: O(u)

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 1 0 0 0 0 0 0 0 1 1 0 0 0 0 1

Figure 1: Bit vector for u = 16. THe current set is {1, 9, 10, 15}.

Split Universe into Clusters

√
We can improve performance by splitting up the range {0, 1, 2, . . . , u − 1} into u
√ √
clusters of size u. If x = i u + j, then V [ x ] = V.Cluster [i ][ j].
√
low( x ) = x mod u = j
x
high( x ) = √ =i
u
√
index (i, j) = i u + j

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 1 0 0 0 0 0 0 0 1 1 0 0 0 0 1

V.Cluster [0] V.Cluster [1] V.Cluster [2] V.Cluster [3]

√
Figure 2: Bit vector (u = 16) split into 16 = 4 clusters of size 4.

• Insert:

– Set V.cluster [ high( x )][low( x )] = 1 O(1)

Lecture 4 van Emde Boas 6.046J Spring 2015

– Mark cluster high( x ) as non-empty O(1)

• Successor:
√
– Look within cluster high( x ) O( u)
√
– Else, find next non-empty cluster i O( u)
√
– Find minimum entry j in that cluster O( u)
√
– Return index (i, j) Total = O( u)

Recurse
√
The three operations in Successor are also Successor calls to vectors of size u. We
can use recursion to speed things up.
√ √
• V.cluster [i ] is a size- u van Emde Boas structure (∀ 0 ≤ i < u)

• V.summary is a size- u van Emde Boas structure

• V.summary[i ] indicates whether V.cluster [i ] is nonempty

I NSERT (V, x )
1 Insert(V.cluster [ high( x )], low[ x ])
2 Insert(V.summary, high[ x ])

So, we get the recurrence:

√
T (u) = 2T ( u) + O(1)
log u
T ' (log u) = 2T ' + O(1)
2
=⇒ T (u) = T ' (log u) = O(log u)

S UCCESSOR (V, x )
1 i = high( x )
2 j = Successor (V.cluster [i ], j)
3 if j == ∞
4 i = Successor (V.summary, i )
5 j = Successor (V.cluster [i ], −∞)
6 return index (i, j)

Lecture 4 van Emde Boas 6.046J Spring 2015

√
T (u) = 3T ( u) + O(1)

 
' ' log u

T (log u) = 3T + O(1)
2
=⇒ T (u) = T ' (log u) = O((log u)log 3 ) ≈ O((log u)1.585 )

To obtain the O(log log u) running time, we need to reduce the number of re­
cursions to one.

Maintain Min and Max

We store the minimum and maximum entry in each structure. This gives an O(1)
time overhead for each Insert operation.

S UCCESSOR (V, x )

1 i = high( x )

2 if low( x ) < V.cluster [i ].max

3 j = Successor (V.cluster [i ], low( x ))

4 else i = Successor (V.summary, high( x ))

5 j = V.cluster [i ].min

6 return index (i, j)

√
T (u) = T ( u) + O(1)
=⇒ T (u) = O(log log u)

Don’t store Min recursively

The Successor call now needs to check for the min separately.

if x < V.min : return V.min (1)

Lecture 4 van Emde Boas 6.046J Spring 2015

I NSERT (V, x )
1 if V.min == None
2 V.min = V.max = x I O(1) time
3 return
4 if x < V.min
5 swap( x ↔ V.min)
6 if x > V.max
7 V.max = x)
8 if V.cluster [ high( x ) == None
9 Insert(V.summary, high( x )) I First Call
10 Insert(V.cluster [ high( x )], low( x )) I Second Call

If the first call is executed, the second call only takes O(1) time. So
√
T (u) = T ( u) + O(1)
=⇒ T (u) = O(log log u)

D ELETE (V, x )
1 if x == V.min I Find new min
2 i = V.summary.min
3 if i = None
4 V.min = V.max = None I O(1) time
5 return
6 V.min = index (i, V.cluster [i ].min) I Unstore new min
7 Delete(V.cluster [ high( x )], low( x )) I First Call
8 if V.cluster [ high( x )].min == None
9 Delete(V.summary, high( x )) I Second Call
10 I Now we update V.max
11 if x == V.max
12 if V.summary.max = None
13 else
14 i = V.summary.max
15 V.max = index (i, V.cluster [i ].max )

If the second call is executed, the first call only takes O(1) time. So
√
T (u) = T ( u) + O(1)
=⇒ T (u) = O(log log u)

Lecture 4 van Emde Boas 6.046J Spring 2015

Lower Bound [Patrascu & Thorup 2007]

Even for static queries (no Insert/Delete)
O(1)
• Ω(log log u) time per query for u = n(log n)

• O(n · poly(log n)) space

Space Improvements
We can improve from Θ(u) to O(n log log u).

• Only create nonempty clusters

– If V.min becomes None, deallocate V

• Store V.cluster as a hashtable of nonempty clusters

• Each insert may create a new structure Θ(log log u) times (each empty insert)
ˇ at]
– Can actually happen [Vladimir Cun´

• Charge pointer to structure (and associated hash table entry) to the structure

This gives us O(n log log u) space (but randomized).

Indirection
We can further reduce to O(n) space.

• Store vEB structure with n = O(log log u) using BST or even an array

=⇒ O(log log n) time once in base case

• We use O(n/ log log u) such structures (disjoint)

=⇒ O( log log
n

u · log log u ) = O( n ) space for small

• Larger structures “store” pointers to them

=⇒ O( log log
n
u · log log u ) = O( n ) space for large

• Details: Split/Merge small structures

6
MIT OpenCourseWare
http://ocw.mit.edu

6.046J / 18.410J Design and Analysis of Algorithms

Spring 2015

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.