Lecture 8:Hashing

1
 The Search Problem
• Find items with keys matching a given search
key
– Given an array A, containing n keys, and a
search key x, find the index i such as x=A[i]
– As in the case of sorting, a key could be part
of a large record.

2
 Applications
• Keeping track of customer account information
at a bank
– Search through records to check balances and
perform transactions
• Keep track of reservations on flights
– Search to find empty seats, cancel/modify
reservations
• Search engine
– Looks for all documents containing a given word

3
 Special Case: Dictionaries

• Dictionary = data structure that supports mainly

two basic operations: insert a new item and
return an item with a given key
• Queries: return information about the set S:
– Search (S, k)
– Minimum (S), Maximum (S)
– Successor (S, x), Predecessor (S, x)
• Modifying operations: change the set
– Insert (S, k)
– Delete (S, k) – not very often
4
 Direct Addressing
• Assumptions:
– Key values are distinct
– Each key is drawn from a universe U = {0, 1, . . . , m - 1}
• Idea:
– Store the items in an array, indexed by keys

• Direct-address table representation:

– An array T[0 . . . m - 1]
– Each slot, or position, in T corresponds to a key in U
– For an element x with key k, a pointer to x (or x itself) will be placed
in location T[k]
– If there are no elements with key k in the set, T[k] is empty,
represented by NIL

5
 Operations
Alg.: DIRECT-ADDRESS-SEARCH(T, k)
return T[k]

Alg.: DIRECT-ADDRESS-INSERT(T, x)
T[key[x]] ← x

Alg.: DIRECT-ADDRESS-DELETE(T, x)
T[key[x]] ← NIL
• Running time for these operations: O(1)

6
 Comparing Different Implementations
• Implementing dictionaries using:
– Direct addressing
– Ordered/unordered arrays
– Ordered/unordered linked lists

Insert Search
direct addressing O(1) O(1)
ordered array O(N) O(lgN)
ordered list O(N) O(N)
unordered array O(1) O(N)
unordered list O(1) O(N)
7
 Hash Tables
• When K is much smaller than U, a hash table
requires much less space than a direct-
address table
– Can reduce storage requirements to |K|
– Can still get O(1) search time, but on the average
case, not the worst case

8
 Hash Tables
Idea:
– Use a function h to compute the slot for each key
– Store the element in slot h(k)

• A hash function h transforms a key into an index in a hash

table T[0…m-1]:
h : U → {0, 1, . . . , m - 1}
• We say that k hashes to slot h(k)
• Advantages:
– Reduce the range of array indices handled: m instead of |U|
– Storage is also reduced

9
 Example: HASH TABLES
0

U
(universe of keys) h(k1)
h(k4)

K k1
h(k2) = h(k5)
(actual k4 k2
keys)
k5 k3 h(k3)

m-1

10
 Do you see any problems
with this approach?
0

U
(universe of keys) h(k1)
h(k4)

K k1
h(k2) = h(k5)
(actual k4 k2
keys) Collisions!
k5 k3 h(k3)

m-1

11
 Collisions
• Two or more keys hash to the same slot!!
• For a given set K of keys
– If |K| ≤ m, collisions may or may not happen,
depending on the hash function
– If |K| > m, collisions will definitely happen (i.e., there
must be at least two keys that have the same hash
value)
• Avoiding collisions completely is hard, even with
a good hash function

12
 Handling Collisions
• We will review the following methods:
– Chaining
– Open addressing
• Linear probing
• Quadratic probing
• Double hashing
• We will discuss chaining first, and ways to
build “good” functions.

13
Handling Collisions Using Chaining
• Idea:
– Put all elements that hash to the same slot into a
linked list

– Slot j contains a pointer to the head of the list of all

elements that hash to j 14
Collision with Chaining - Discussion
• Choosing the size of the table
– Small enough not to waste space
– Large enough such that lists remain short
– Typically 1/5 or 1/10 of the total number of elements
• How should we keep the lists: ordered or not?
– Not ordered!
• Insert is fast
• Can easily remove the most recently inserted elements

15
 Insertion in Hash Tables
Alg.: CHAINED-HASH-INSERT(T, x)
insert x at the head of list T[h(key[x])]

• Worst-case running time is O(1)

• Assumes that the element being inserted isn’t already
in the list
• It would take an additional search to check if it was
already inserted

16
 Deletion in Hash Tables
Alg.: CHAINED-HASH-DELETE(T, x)
delete x from the list T[h(key[x])]

• Need to find the element to be deleted.

• Worst-case running time:
– Deletion depends on searching the corresponding
list

17
 Searching in Hash Tables

Alg.: CHAINED-HASH-SEARCH(T, k)

search for an element with key k in list

T[h(k)]

• Running time is proportional to the length of

the list of elements in slot h(k)

18
 Analysis of Hashing with Chaining:
Worst Case
• How long does it take to search T
for an element with a given key? 0

• Worst case:
– All n keys hash to the same slot
– Worst-case time to search is (n),
plus time to compute the hash
chain
function

m-1

19
 Analysis of Hashing with Chaining:
• Average case
Average Case
– depends on how well the hash function
distributes the n keys among the m slots T
n0 = 0
• Simple uniform hashing assumption:
– Any given element is equally likely to n2
hash into any of the m slots (i.e., probability n3
of collision Pr(h(x)=h(y)), is 1/m)
nj
• Length of a list:
T[j] = nj, j = 0, 1, . . . , m – 1 nk
• Number of keys in the table:
n = n0 + n1 +· · · + nm-1 nm – 1 = 0

• Average value of nj:

E[nj] = α = n/m 20
 Load Factor of a Hash Table
• Load factor of a hash table T:
T
 = n/m 0

– n = # of elements stored in the table

chain
– m = # of slots in the table = # of linked lists chain

•  encodes the average number of chain

elements stored in a chain
chain
•  can be <, =, > 1
m-1

21
 Case 1: Unsuccessful Search
(i.e., item not stored in the table)
Theorem
An unsuccessful search in a hash table takes expected time (1   )
under the assumption of simple uniform hashing
(i.e., probability of collision Pr(h(x)=h(y)), is 1/m)
Proof
• Searching unsuccessfully for any key k
– need to search to the end of the list T[h(k)]
• Expected length of the list:
– E[nh(k)] = α = n/m
• Expected number of elements examined in an unsuccessful search is α
• Total time required is:
– O(1) (for computing the hash function) + α  (1   )
22
Case 2: Successful Search

23
 Analysis of Search in Hash Tables
• If m (# of slots) is proportional to n (# of
elements in the table):

• n = O(m)

• α = n/m = O(m)/m = O(1)

 Searching takes constant time on average

24
 Hash Functions
• A hash function transforms a key into a table
address
• What makes a good hash function?
(1) Easy to compute
(2) Approximates a random function: for every input,
every output is equally likely (simple uniform hashing)
• In practice, it is very hard to satisfy the simple
uniform hashing property
– i.e., we don’t know in advance the probability
distribution that keys are drawn from

25
 Good Approaches for Hash Functions

• Minimize the chance that closely related keys

hash to the same slot
– Strings such as pt and pts should hash to
different slots
• Derive a hash value that is independent from
any patterns that may exist in the
distribution of the keys

26
 The Division Method
• Idea:
– Map a key k into one of the m slots by taking the
remainder of k divided by m
h(k) = k mod m
• Advantage:
– fast, requires only one operation
• Disadvantage:
– Certain values of m are bad, e.g.,
• power of 2
• non-prime numbers
27
 Example - The Division Methodm97 m

• If m = 2p, then h(k) is just the least 100

significant p bits of k
– p = 1 m = 2
 h(k) = {0, 1} , least significant 1 bit of k
– p = 2m = 4
 h(k) ={0, 1, 2, 3} , least significant 2
bits of k
 Choose m to be a prime, not close to a
power of 2 k mod 97
 Column 2: k mod 100
 Column 3:
28
 The Multiplication Method
Idea:
• Multiply key k by a constant A, where 0 < A < 1
• Extract the fractional part of kA
• Multiply the fractional part by m
• Take the floor of the result
h(k) = = m (k A mod 1)

fractional part of kA = kA - kA

• Disadvantage: Slower than division method

• Advantage: Value of m is not critical, e.g., typically 2p
29
Example – Multiplication Method

30
 Universal Hashing
• In practice, keys are not randomly distributed
• Any fixed hash function might yield Θ(n) time
• Goal: hash functions that produce random
table indices irrespective of the keys
• Idea:
– Select a hash function at random, from a designed
class of functions at the beginning of the execution

31
Universal Hashing

(at the beginning

of the execution)

32
Definition of Universal Hash Functions

H={h(k): U(0,1,..,m-1)}

33
Universal Hashing – Main Result

With universal hashing the chance of collision

between distinct keys k and l is no more than
the 1/m chance of collision if locations h(k)
and h(l) were randomly and independently
chosen from the set {0, 1, …, m – 1}

34
 Designing a Universal Class
of Hash Functions
• Choose a prime number p large enough so that every possible
key k is in the range [0 ... p – 1]
Zp = {0, 1, …, p - 1} and Zp* = {1, …, p - 1}
• Define the following hash function
ha,b(k) = ((ak + b) mod p) mod m,
 a  Zp* and b  Zp The class Hp,m of hash
functions is universal
• The family of all such hash functions is
Hp,m = {ha,b: a  Zp* and b  Zp}
• a , b: chosen randomly at the beginning of execution
35
 Example: Universal Hash Functions
E.g.: p = 17, m = 6

ha,b(k) = ((ak + b) mod p) mod m

h3,4(8) = ((38 + 4) mod 17) mod 6

= (28 mod 17) mod 6

= 11 mod 6

=5
36
 Advantages of Universal Hashing

• Universal hashing provides good results on

average, independently of the keys to be stored
• Guarantees that no input will always elicit the
worst-case behavior
• Poor performance occurs only when the
random choice returns an inefficient hash
function – this has small probability
37
 Open Addressing
• If we have enough contiguous memory to store all the keys (m >
N)  store the keys in the table itself
• No need to use linked lists anymore
• Basic idea:
– Insertion: if a slot is full, try another one,
until you find an empty one
– Search: follow the same sequence of probes
– Deletion: more difficult ... (we’ll see why)

• Search time depends on the length of the

probe sequence!
38
Generalize hash function notation:
• A hash function contains two arguments now:
(i) Key value, and (ii) Probe number

h(k,p), p=0,1,...,m-1

• Probe sequences
<h(k,0), h(k,1), ..., h(k,m-1)>
– Must be a permutation of <0,1,...,m-1>
– There are m! possible permutations
– Good hash functions should be able to
produce all m! probe sequences

39
 Common Open Addressing Methods

• Linear probing
• Quadratic probing
• Double hashing

• Note: None of these methods can generate

more than m2 different probing sequences!

40
 Linear probing: Inserting a key
• Idea: when there is a collision, check the next available
position in the table (i.e., probing)

h(k,i) = (h1(k) + i) mod m

i=0,1,2,...
• First slot probed: h1(k)
• Second slot probed: h1(k) + 1
• Third slot probed: h1(k)+2, and so on
probe sequence: < h1(k), h1(k)+1 , h1(k)+2 , ....>

• Can generate m probe sequences maximum, why?

wrap around
41
 Linear probing: Searching for a key
• Three cases:
(1) Position in table is occupied with an
element of equal key 0

(2) Position in table is empty

h(k1)
(3) Position in table occupied with a h(k4)
different element
• Case 2: probe the next higher index h(k2) = h(k5)
until the element is found or an
h(k3)
empty position is found
• The process wraps around to the m-1
beginning of the table

42
 Linear probing: Deleting a key
• Problems
– Cannot mark the slot as empty 0
– Impossible to retrieve keys inserted
after that slot was occupied
• Solution
– Mark the slot with a sentinel value
DELETED
• The deleted slot can later be used for
insertion m-1
• Searching will be able to find all the
keys
43
 Primary Clustering Problem
• Some slots become more likely than others
• Long chunks of occupied slots are created
 search time increases!!

initially, all slots have probability 1/m

Slot b:
2/m

Slot d:
4/m

Slot e:
5/m

44
Quadratic probing

i=0,1,2,...

45
 Double Hashing
(1) Use one hash function to determine the first slot
(2) Use a second hash function to determine the
increment for the probe sequence
h(k,i) = (h1(k) + i h2(k) ) mod m, i=0,1,...
• Initial probe: h1(k)
• Second probe is offset by h2(k) mod m, so on ...
• Advantage: avoids clustering
• Disadvantage: harder to delete an element
• Can generate m2 probe sequences maximum
46
 Double Hashing: Example
h1(k) = k mod 13 0
1 79
h2(k) = 1+ (k mod 11) 2
h(k,i) = (h1(k) + i h2(k) ) mod 13 3
4 69
• Insert key 14: 5 98

h1(14,0) = 14 mod 13 = 1 6
7 72
h(14,1) = (h1(14) + h2(14)) mod 13 8
9
= (1 + 4) mod 13 = 5 10
14

h(14,2) = (h1(14) + 2 h2(14)) mod 13 11 50

12
= (1 + 8) mod 13 = 9
47