Weiss Ch. 5
Hash tables are a very clever storage and lookup mechanism with potential for O(1) lookup and insertion. Algorithms that guarantee the lower bound are complex, and new hashing schemes with very good performance, like Cuckoo and Hopscotch (I'm not making any of this up, you know) are too complex for analysis and rely on experiments to justify their governing heuristics.
Cryptographic hashing is a very heavy version of the idea that concentrates on good hash functions, which are rather a secondary consideration (at best) in hash tables. Crypto hashes are found in public-key crypto signatures; Bob sends Alice his encrypted signature and an encrypted hash value along with his (differently) encrypted document. Alice decrypts the document using her private key, and the signature using Bob's public key (!). She applies the known hash function to the decrypted document and should derive the (now-decrypted) hash value in Bob's signature. Else it's a wrong or corrupted doc: e.g. signature and hash have been pasted on to a false doc.
Our PPT goes over the practical, simple, common-or-garden academic hashing ideas. The text also presents modern methods and the usual sort of analysis we see for hashing schemes. This lecture is aimed at outlines and the analysis.
Load factor λ of a hash-table =
Thus in separate chaining with lists:
No external storage, just probe forward to next table location.
Big hazard here is primary clustering, which is an obvious phenomenon and weakness. If a "cluster" starts, (say 2 adjacent elements fill up), it's more likely an new entry will land on the cluster, after which it's guaranteed to wind up at the end of the cluster, enlarging it...oops.
With linear probing, insertions and unsuccessful searches must take
the same amount of time, which turns out to be
.5(1 +1/(1-λ)2). Successful searches should perform fewer probes on average than unsuccessful, and in fact they take
.5(1 +1/(1-λ)). Calculations are "somewhat involved".
FOCS has zip on hashing analysis; Knuth vol 3 chapter 6.4 has about 46 pages on hashing, with diagrams, code, and relatively a lot of analysis.
Imagine a purely random collision resolution strategy. Impractical but analyzable. Fraction of empty cells is (1 - λ), so number we expect to probe for an unsuccessful search is 1/ (1 - λ). Now insertions happen after unsuccessful searches, so we can use this formula to get average cost of successful search.
λ starts at 0 with an empty table and then increases; the average performance thus starts well but degrades. Also notice that items (quickly) inserted during a low value of λ are going to be easier to find in a successful search: small number of probes both times.
Taking account of the varying number of probes we'll need takes an
integral over the values of λ at work during the insertions,
so the mean insertion time is:
I(λ) = ∫λ0 (1/(1-x)) dx = 1/λ ln(1/(1-λ)).
Both these ideal probing formulae are better than the clustering
quoted before; we can see how clustering hurts us. Typical
Number of probes vs. load factor for linear (dashed) and random probing: S(uccessful), U(nsuccessful) = I(nsert).
On collision, step quadratically, for instance by 1,4,9,16,...,i2
No guarantee of finding empty cell if λ > 1/2, or worse for non-prime table sizes. Here's related result:
With prime table size and quadratic probing, a new element can be inserted if λ ≤ 1/2.
tablesize is odd prime > 3, h(x) is hash function, or
h0(x) first try, and h(x) + i2 for subsequent
Pick two locations h(x) + i2 and h(x) + j2,
0 ≤ i,j ≤ ⌊ tablesize/2 ⌋.
Now we're claiming
these locations are distinct, so let's assume they're not and look for
contradictions. That is, assume the locations are the same but i
≠ j. Then
h(x) + i2 = h(x) + j2 (mod tablesize)
i2 = j2 (mod tablesize)
i2 - j2 = 0 (mod tablesize)
(i-j) (i+j) = 0 (mod tablesize)
Since tablesize is prime, (i-j) or (i+j) = 0 (mod tablesize)
But they're distinct so first can't be true, and they're both less than half the tablesize so the second can't be true. Thus the first tablesize/2 locations are distinct under quadratic probing.
Two hash functions, say h and g: if h(x) gives collision, then use g(x), often by stepping by g(x), so look next at i*g(x).
Need to be a bit careful, (as always, actually) about hash functions:
g(x) shouldn't ever = 0, for instance. Non-prime table sizes can yield
non-probable elements by running out of 'new' locations early (e.g. in
W. p. 187). OK would be
g(x) = R - (x mod R), R a prime smaller than table-size. Theoretically double hashing is interesting since experimentally it works really well, just a bit more complex with those two hash functions.
Various signals (load factor, failed insertion) trigger rehashing, which is simply creating a new, bigger table, hashing all elements of old table into the new one, and recycling the old one. O(N), so a pain but amortized over the O(N) inserts we've likely had already, its cost ls low.
You get hashSet and hashMap abstractions, but need to provide "equals" and "hashCode".
CB has found hash tables to be attractive substitutes for all sorts of other data structures, and rather addictive: esp. in Perl.
We want guaranteed, worst case O(1) access (routers, hardware,...) Just for example, consider separate chaining with load factor of 1, to see our problem. f we distribute N balls into N bins randomly, what is the expected number of balls in the fullest bin (i.e. the length of the list in separate chaining)? The bad news is it's Θ(logN/loglogN), or nearly logN.
Balls and Bins Stolen with thanks from Avrim at CMU.
Perfect hashing is a classical solution to this problem.
Assume we know the N table items in advance. We'll take the time to construct a hash table with worst-case O(1) access performance. If our separate chains, or lists, have at most a constant number of items we're done. Also with more lists, they'd be shorter. So how to make sure we don't need too many lists, and how to make sure we don't just get unlucky anyway?
For second problem, assume table size of M, so M lists. We choose M large enough (see below) so that there will be no collisions with probability at least 1/2.
We can use the principles of universal hashing (Weiss 5.8) to generate good hash functions. In any case, we simply hash our N items into the table -- if we get a collision we give up, empty the table out, and start all over with a new hash function. If M, the table length, is big enough, we can reduce the expected number of attempts to 2, and all this work is charged off to the insertion algorithm.
So how large must M be?
Weiss's theorem 5.2:
If N balls are placed into M = N2 bins, probability that no bin has more than one ball is < 1/2. (easy, cute counting proof p. 193).
Gulp. must our table size grow as N2 !??. Well... suppose we just use N bins, but each bin points to a hash table! Bwah-hah-ha! These only have to be as big as the square of the number of collisions, and each is constructed with a different hash function until IT is collision free. You can prove (again, simple counting) Weiss Theorem 5.3, that the total size of the secondary hash tables will be at most 2N.
Last four sections of Chapter 5 are of varying difficulty. Extendible hashing is very easy. Universal hashing is mostly modular mathematics, not too hard. Cuckoo and Hopscotch algorithms are complex algorithms, long sections enlivened by figures and code: good investment for pledge party ice-breaking conversation.