Data Structures Dana Shapira Hash Tables 1 Element

Data Structures Dana Shapira Hash Tables 1

Element Uniqueness Problem n Let n Determine whethere exist i j such that xi=xj n Sort Algorithm n Bucket Sort for (i=0; i<m; i++) T[i]=NULL; for (i=0; i<n; i++){ if (T[xi]= = NULL) T[xi]= i else{ output (i, T[xi]) return; } } What happens when m is large or when we are dealing with real numbers? ? 2

Hash Tables h n Notations: n U universe of keys of size |U|, n K an actual set of keys of size n, n T a hash table of size m U n Use a hash-function h: U {0, …, m-1}, h(x)=i that computes the slot i in array T where element x is to be stored, for all x in U. Set of array indices n h(k) is computed in O(|k|) = O(1). h(x 1) h(x 2) h(x 3) h(x 4) x 1 x 2 x 3 x 4 3

Example 0 n n h: U {0, …, m-1} h(x)=x mod 10 (what is m? ) input: 17, 62, 19, 81, 53 Collision: x ≠ y but h(x) = h(y). m « |U|. Solutions: 1. Chaining 2. Open addressing 1 81 2 62 3 4 53 5 6 7 8 17 9 19 4

Collision-Resolution by Chaining 1 81 62 12 53 17 37 57 19 n n n Insert(T, x): Delete(T, x): Search(T, x): Insert new element x at the head of list T[h(x. key)]. Delete element x from list T[h(x. key)]. Search list T[h(x. key)]. 5

Analysis of Chaining Simple Uniform Hashing n Any given element is equally likely to hash to any slot in the hash table. n The slot an element hashes to is independent of where other elements hash. Load factor: α = n/m (elements stored in the hash table / number of slots in the hash table) 6

Analysis of Chaining n Theorem: In a hash table with chaining, under the assumption of simple uniform n n hashing, both successful and unsuccessful searches take expected time Θ(1+α) on the average, where α is the hash table load factor. Proof: Unsuccessful Search: Under the assumption of simple uniform hashing, any key k is equally likely to hash to any slot in the hash table. The expected time to search unsuccessfully for a key k is the expected time to search to the end of list T[h(k)] which has expected length α. expected time - (1 + α) including time for computing h(k). Successful search: The number of elements examined during a successful search is 1 more than the number of elements that appear before k in T[h(k)]. expected time - (1 + α) 7 n Corollary: If m = (n), then Insert, Delete, and Search take expected constant time.

Designing Good Hash Functions n. Example: n. Input = reals drawn uniformly at random from [0, 1) n. Hash function: h(x) = mx n. Often, the input distribution is unknown. n. Then we can use heuristics or universal hashing. 8

The Division Method Hash function: h(x) = x mod m m = 2 k h(x) = the lowest k bits of x n. Heuristic: m = prime number not too close to a power of 2 n 9

The Multiplication Method Hash function: h(x) = m (cx mod 1) , for some 0 < c < 1 n Optimal choice of c depends on input distribution. n. Heuristic: Knuth suggests the inverse of the golden ratio as a value that works well: Example: x=123, 456, m=10, 000 h(x) = 10, 000 ·(123, 456· 0. 61803… mod 1) = = 10, 000 ·(76, 300. 0041151… mod 1) = = 10, 000 · 0. 0041151… = 41. 151. . . = 41 10

Efficient Implementation of the Multiplication Method h(x) = m (cx mod 1) w bits n. Let w be the size of a machine word n. Assume that key x fits into a machine word n. Assume that m = 2 p * n. Restrict ourselves to values of c of the form c = s / 2 w n. Then cx = sx / 2 w nsx is a number that fits into two machine words nh(x) = p most significant bits of the lower word w bits Fractional part Integer part after multiplying by m = 2 p w bits p bits 11

Example h(x) = m (cx mod 1) nx = 123456, p = 14, m = 214 = 16384, w = 32, n. Then sx = (76300 ⋅ 232) + 17612864 n. The 14 most significant bits of 17612864 are 67; that is, h(x) = 67 x = 0000 0001 1110 0010 0100 0000 s = 1001 1110 0011 0111 1001 1011 1001 sx = 0000 0001 0010 1010 0000 1100 0001 0000 1100 0000 0100 0000 h(x) = 00 0000 0100 0011 = 67 12

Open Addressing n. All elements are stored directly in the hash table. ➥Load factor α cannot exceed 1. n. If slot T[h(x)] is already occupied for a key x, we probe alternative locations until we find an empty slot. n. Searching probes slots starting at T[h(x)] until x is found or we are sure that x is not in T. n. Instead of computing h(x), we compute h(x, i) n i -the probe number. 13 h(x)

Linear Probing Hash function: h(k, i) = (h'(k) + i) mod m, where h' is an original hash function. n Benefits: Easy to implement n Problem: Primary Clustering - Long runs of occupied slots build up as table becomes fuller. 14

Quadratic Probing Hash function: h(k, i) = (h'(k) + c 1 i + c 2 i 2) mod m, where h' is an original hash function. n Benefits: No more primary clustering n Problem: Secondary Clustering - Two elements x and y with h'(x) = h'(y) have same probe sequence. 15

Double Hashing Hash function: h(k, i) = (h 1(k) + ih 2(k)) mod m, where h 1 and h 2 are two original hash functions. n h 2(k) has to be prime w. r. t. m; that is, gcd(h 2(k), m) = 1. n Two methods: Choose m to be a power of 2 and guarantee that h 2(k) is always odd. Choose m to be a prime number and guarantee that h 2(k) < m. n Benefits: No more clustering n Drawback: More complicated than linear and quadratic probing 16

Analysis of Open Addressing n Uniform hashing: The probe sequence h(k, 0), …, h(k, m – 1) is equally likely to be any permutation of 0, …, m – 1. n Theorem: In an open-address hash table with load factor α < 1, the expected number of probes in an unsuccessful search is at most 1 / (1 – α), assuming uniform hashing. Proof: Let X be the number of probes in an unsuccessful search. Ai = “there is an i-th probe, and it accesses a non-empty slot” 17

Analysis of Open Addressing – Cont. 18

Analysis of Open Addressing – Cont. Corollary: The expected number of probes performed during an insertion into an open-address hash table with uniform hashing is 1 / (1 – α). 19

Analysis of Open Addressing – Cont. Theorem: Given an open-address hash table with load factor α < 1, the expected number of probes in a successful search is (1/α) ln (1 / (1 – α)), assuming uniform hashing and assuming that each key in the table is equally likely to be searched for. n A successful search for an element x follows the same probe sequence as the insertion of element x. n Consider the (i + 1)-st element x that was inserted. n The expected number of probes performed when inserting x is at most n Averaging over all n elements, the expected number of probes in a successful search is 20

Analysis of Open Addressing – Cont. 21

Universal Hashing n. A family of hash functions is universal if for each pair k, l of keys, there at most | | / m functions in such that h(k) = h(l). n. This means: For any two keys k and l and any function h chosen uniformly at random, the probability that h(k) = h(l) is at most 1/m. n P= (| | / m )/ | | =1/m n. This is the same as if we chose h(k) and h(l) uniformly at random from [0, m – 1]. 22

Analysis of Universal Hashing Theorem: For a hash function h chosen uniformly at random from a universal family , the expected length of the list T[h(x)] is α if x is not in the hash table and 1 + α if x is in the hash table. Proof: Indicator variables: Yx = the number of keys ≠ x that hash to the same slot as x 23

Analysis of Universal Hashing Cont. If x is not in T, then |{y T : x ≠ y}| = n. Hence, E[Yx] = n / m = α. If x is in T, then |{y T : x ≠ y} = n – 1. Hence, E[Yx] = (n – 1) / m < α. The length of list T[h(x)] is one more, that is, 1 + α. 24

Universal Family of Hash Functions n Choose a prime p so that m = p. n Let such n For each n define the hash function as follows: = | | = Example: m=p=253 a=[248, 223, 101] x=1025=[0, 2, 1] . 25

Universal Family of Hash Functions Theorem: The class is universal. Proof: Let such that x ≠ y, w. l. o. g n For all a 1, …, ar there exists a single a 0 such that n For all n there exists a single w such that z·w=1(mod m) for mr values 26 n The number of hash functions h in , for which h(k 1) = h(k 2) is at most mr/ mr+1 =1/m

Universal Family of Hash Functions Choose a prime p so that m < p. For any 1 ≤ a < p and 0 ≤ b < p, we define a function ha, b(x) = ((ax + b) mod p) mod m. Let Hp, m be the family Hp, m = {ha, b : 1 ≤ a < p and 0 ≤ b < p}. Theorem: The class Hp, m is universal. 27

Summary n Hash tables are the most efficient dictionaries if only operations Insert, Delete, and n n n Search have to be supported. If uniform hashing is used, the expected time of each of these operations is constant. Universal hashing is somewhat complicated, but performs well even for adversarial input distributions. If the input distribution is known, heuristics perform well and are much simpler than universal hashing. For collision-resolution, chaining is the simplest method, but it requires more space than open addressing. Open addressing is either more complicated or suffers from clustering effects. 28