Cello

Cello is a library of succinct data structures, oriented in particular for string searching and other string operations.

Usually, searching for patterns in a string takes O(n) time, where n is the length of the string. Indices can speedup the search, but take additional space, which can be costly for very large strings. A data structure is called succinct when it takes n + o(n) space, where n is the space needed to store the data anyway. Hence succinct data structures can provide additional operations with limited space overhead.

It turns out that strings admit succinct indices, which do not take much more space than the string itself, but allow for O(k) substring search, where k is the length of the substring. Usually, this is much shorter, and this considerably improves search times. Cello provide such indices and many other related string operations.

An example of usage would be:

let
  x = someLongString
  pattern = someShortString
  index = searchIndex(x)
  positions = index.search(pattern)

echo positions

Many intermediate data structures are constructed to provide such indices, though, and as they may be of independent interest, we describe them in the following.

Notice that a string here just stands for a (usually very long) sequence of symbols taken from a (usually small) alphabet. Prototypical examples include

• genomic data, where the alphabet is A, C, G, T or
• time series, where each value is represented by a symbol, such as HIGH, MEDIUM, LOW, or UP, DOWN
• where only two values are available, it is often convenient to store the data as bit sequences to save space.

At the moment all operations are implemented on

type AnyString = string or seq[char] or Spill[char]

where spills are just memory-mapped sequences. The library may become generic in the future, although this is not a priority.

Notice that Cello is not Unicode-aware: think more of searching large genomic strings or symbolized time series, rather then using it for internationalized text, although I may consider Unicode operations in the future.

Versions

Cello recent version (>= 0.2) requires Nim >= 0.20. For usage with Nim up to 0.19.4, use Cello 0.1.6.

Basic operations

The most common operations that we implement on various kind of sequence data are rank and select. We first describe them for sequences of bits, which are the foundation we use to store more complex kind of data.

For bit sequences, rank(i) counts the number of 1 bits in the first i places. The number of 0 bits can easily be obtained as i - rank(i). Viceversa, select(i) finds the position of the i-th 1 bit in the sequence. In this case, there is not an obvious relation to the position of the i-th 0 bit, so we provide a similar operation select0(i).

To ensure that rank(select(i)) == i, we define select(i) to be 1-based, that is, we count bits starting from 1.

As a reference, we implement rank and select on Nim built-in sets, so that for instance the following is valid:

let x = { 13..27, 35..80 }

echo x.rank(16)  # 3
echo x.select(3) # 16

More generally, one can define 'rank' and select for sequence of symbols taken from a finite alphabet, relative to a certain symbol. Here, rank(c, i) is the number of symbols equal to c among the first i symbols, and select(c, i) is the position of the i-th symbol c in the sequence.

Again, we give a reference implementation for strings, so that the following is valid:

let x = "ABRACADABRA"

echo x.rank('A', 8)   # 4
echo x.select('A', 4) # 8

Notice that in both cases, the implementation of rank and select is a naive implementation which takes O(i) operations. More sophisticated data structures allow to perform similar operations in constant (for rank) or logarithmic (for select) time, by using indices. Succinct data structures allow to do this using indices that take at most o(n) space in addition to the sequence data itself, where n is the sequence length.

Data structures

We now describe the succinct data structures that will generalize the bitset and the string examples above. In doing so, we also need a few intermediate data structures that may be of independent interest.

Bit arrays

Bit arrays are a generalization of Nim default set collections. They can be seen as an ordered sequence of bool, which are actually backed by a seq[int]. We implement random access - both read and write - as well as naive rank and select. An example follows:

var x = bits(13..27, 35..80)

echo x[12]   # false
echo x[13]   # true
x[12] = true # or incl(x, 12)
echo x[12]   # true
x[12] = false

echo x.rank(16)    # 3
echo x.select(3)   # 16
echo x.select0(30) # 90

Int arrays

Int arrays are just integer sequences of fixed length. What distinguishes them by the various types seq[uint64], seq[uint32], seq[uint16], seq[uint8] is that the integers can have any length, such as 23.

They are backed by a bit array, and can be used to store many integer numbers of which an upper bound is known without wasting space. For instance, a sequence of positive numbers less that 512 can be backed by an int array where each number has size 9. Using a seq[uint16] would almost double the space consumption.

Most sequence operations are available, but they cannot go after the initial capacity. Here is an example:

var x = ints(200, 13) # 200 ints at most 2^13 - 1

x.add(123)
x.add(218)
x.add(651)
echo x[2]   # 651
x[12] = 1234
echo x[12]   # 1234

echo x.len       # 13
echo x.capacity  # 200

RRR

The RRR bit vector is the first of our collections that is actually succinct. It consists of a bit arrays, plus two int arrays that stores rank(i) values for various i, at different scales.

It can be created after a bit array, and allows constant time rank and logarithmic time select and select0.

let b: BitArray = ...
let r = rrr(b)

echo r.rank(123456)
echo r.select(123456)
echo r.select0(123456)

To convince oneself that the structure really is succinct, stats(rrr) returns a data structures that shows the space taken (in bits) by the bit array, as well as the two auxiliary indices.

Reference

Wavelet tree

The wavelet tree is a tree constructed in the following way. An input string over a finite alphabet is given. The alphabet is split in two parts - the left and the right one, call them L and R.

For each character of the string, we use a 1 bit to denote that the character belongs to R and a 0 bit to denote that it belongs to L. In this way, we obtain a bit sequence. The node stores the bit sequence as an RRR structures, and has two children: the one to the left is the wavelet tree associated to the substring composed by the characters in L, taken in order, and similarly for the right child.

This structure allows to compute rank(c, i), where c is a character in the alphabet, in time O(log(l)), and select(c, i) in time O(log(l)log(n)) where l is the size of the alphabet and n is the size of the string. It also allows O(log(l)) random access to read elements of the string.

It can be used as follows:

let
  x = "ACGGTACTACGAGAGTAGCAGTTTAGCGTAGCATGCTAGCG"
  w = waveletTree(x)

echo x.rank('A', 20)   # 7
echo x.select('A', 7)  # 20
echo x[12]             # 'G'

Reference

Rotated strings

The next ingredient that we need it the Burrows-Wheeler transform of a string. It can be implemented using string rotations, so that's what we implement first. It turns out that this implementation is too slow for our purposes, but rotated strings may be useful anyway, so we left them in.

A rotated strings is just a view over a string, rotated by a certain amount and wrapping around the end of the string. If the underlying string is a var, our implementation reuses that memory (which is then shared) to avoid the copy of the string. We just implement random access and printing:

var
  s = "The quick brown fox jumps around the lazy dog"
  t = s.rotate(20)

echo t[10] # n
echo t[20] # u

t[18] = e

echo s # The quick brown fox jumps around the lezy dog
echo t # jumps around the lezy dogThe quick brown fox

Suffix array

The suffix array of a string is a permutation of the numbers from 0 up to the string length excluded. The permutation is obtained by considering, for each i, the suffix starting at i, and sorting these strings in lexicographical order. The resulting order is the suffix array.

Here the suffix array is represented as an IntArray. It can be obtained as follows:

let
  x = "this is a test."
  y = suffixArray(x)

echo y # @[7, 4, 9, 14, 8, 11, 1, 5, 2, 6, 3, 12, 13, 10, 0]

Sorting the indices may be a costly operation. One can use the fact that the suffixes of a string are a quite special collection to produce more efficient algorithms. Other than the sort-based one, we offer the DC3 algorithm.

Notice that at the moment DC3 is not really optimized and may be neither space nor time efficient.

To use an alternative algorithm, just pass an additional parameter, of type

type SuffixArrayAlgorithm* {.pure.} = enum
  Sort, DC3

like this

let
  x = "this is a test."
  y = suffixArray(x, SuffixArrayAlgorithm.DC3)

echo y # @[7, 4, 9, 14, 8, 11, 1, 5, 2, 6, 3, 12, 13, 10, 0]

Reference

Burrows-Wheeler transform

The Burrows-Wheeler transform of a string is a string one character longer, together with a distinguished character. Once one has a suffix array sa for the string s & '\0', where \0 is our distinguished character, the Burrows-Wheeler transform is the string which at the index i has the last character of the rotation of s by sa[i]. The distinguished index if the permutation of \0.

We recall the following two facts:

• the Burrows-Wheeler transform can be inverted - the exact algorithm is outside the purposes of this documentation
• whenever a character is a good predictor for the next one (in the original string), the string in the Burrows-Wheeler transform tends to have many repeated characters, which allows to compress it by run-length encoding.

An example of usage is this:

let
  s = "The quick brown fox jumps around the lazy dog"
  t = burrowsWheeler(s)
  u = inverseBurrowsWheeler(t)

echo t # gskynxeed\0 l in hh otTu c uwudrrfm abp qjoooza
echo u # The quick brown fox jumps around the lazy dog

Notice that for this to work we assume that s does not contain \0 itself. We use the fact that Nim strings are not null terminated, hence \0 is a valid character. Notice that printing the transformed string may not work as intended, since the terminal may interpret the embedded \0 as a string terminator.

Reference

FM indices

An FM index for a string puts together essentially all the pieces that we have described so far. The index itself holds a walevet tree for the Burrows-Wheeler transform of the string, together with a small auxiliary table having the size of the string alphabet.

It can be used for various purposes, but the simplest one is backward search. Given a pattern p (a small string) and possibly long string s, there is a way to search all occurrences of p in time O(L), where L is the length of p - the time is independent of s - using an FM index for s.

Every occurrence of p appears as the prefix of some rotation of s - hence all such occurrences correspond to consecutive positions into the suffix array for s. The first and last such positions can be found as follows:

let
  x = "mississippi"
  pattern = "iss"
  fm = fmIndex(x)
  sa = suffixArray(x)
  positions = fm.search(pattern)

echo positions.first # 2
echo positions.last  # 3

for j in positions.first .. positions.last:
  let i = sa[j.int]
  echo x.rotate(i)

# issippimiss
# ississippim

For economy, the FM index itself does not include the suffix array, as some applications do not require the latter. Still, it is quite frequent to need both; since computing the FM index requires the suffix array in any case, and computing the suffix array is quite costly, there is a way to get both at the same time. In the above example, we could write as well

let
  index = searchIndex(x)
  fm = index.fmIndex
  sa = index.suffixArray

The above type can be used to streamline search:

let
  index = searchIndex(x)
  positions = index.search(pattern)

echo positions # @[1, 4]

Reference

Applications

Here we describe a few applications of the above data structures, together with some other string utilities included in Cello.

Boyer-Moore-Horspool search

To make a comparison with naive string searching (without using indices), an implementation of Boyer-Moore-Horspool string searching is provided.

The Boyer-Moore algorithm and variations (such as the one used here, due to Horspool) scan a string linearly to find a pattern, but use a precomputed table based on the pattern to skip more than one charachter at a time. The key observation is that after making a comparison for the pattern in a given position, one already knows that some subsequent positions will not match for sure, hence can be skipped. The resulting algorithm is still O(n) in the length of the string, but may perform less than n actual comparisons.

The API mimics strutils.find and it is meant to be used as follows:

let
  x = "mississippi"
  pattern = "iss"

echo boyerMooreHorspool(x, pattern) # 1 (ississippi)
echo boyerMooreHorspool(x, pattern, start = 2)  # 4 (issippi)

Reference

Levenshtein similarity

The Levenshtein distance (or edit distance) between two strings is the minimum number of insertions, deletions or substitutions required to change one string into the other.

It is computed by strutils.editDistance. Here we expose a similarity measure derived from it, defined as s = (L - e) / L, where L is the cumulative length of the two strings, and e is the edit distance. It is a number between 0 and 1, which is 1 only if the two strings are equal.

It is simply used as

let
  a = someString
  b = someOtherString
  s = levenhstein(a, b)

Ratcliff-Obershelp similarity

The Levenshtein similarity is a quite crude measure of whether two strings resemble each other. A better measure is given by the Ratcliff-Obershelp similarity which is defined as s = (2 * m) / L, where L is the cumulative length of the two strings, and m is the number of matching characters.

Matching characters are defined recursively: first we find the longest common substring lcs between the two and count the number of characters of lcs as matching. Then, recursively, we compare the number of matching characters in the chunks to the left of lcs and to the right of lcs.

For instance, when comparing ALEXANDRE and ALEKSANDER, we find the following sequence of longest common substrings:

• ALE
• AND
• R

giving a Ratcliff-Obershelp similarity of 2 * (3 + 3 + 1) / (9 + 10).

It is simply used as

let
  a = someString
  b = someOtherString
  s = ratcliffObershelp(a, b)

Reference

Jaro similarity

The Jaro similarity of two strings a and b is given by

0 if m == 0
((m / len(a)) + (m / len(b)) + ((m - t / 2) / m)) / 3 otherwise

where m is number of matching characters and t is the number of transpositions. Here two characters are considered matching if they are equal and their ì distance is less then max(len(a), len(b)) / 2. The substrings of a and b given by matching characters are permutations of each other. Characters that match but appear in different positions in these strings are considered transpositions.

For instance, when comparing ALEXANDRE and ALEKSANDER, we find the following matches inside a and b respectively: ALEANDRE, ALEANDER. Hence here m = 8, t = 2, so that the similarity is ((8 / 9) + (8 / 10) + (7 / 8)) / 3.

Reference

Jaro-Winkler similarity

The Jaro-Winkler similarity of two strings is a correction to the Jaro similarity that favours strings which have a long common prefix. If L is the length of the common prefix of two strings and J is the Jaro similarity, the Jaro-Winkler similarity is computed as

J + p * L * (1 - J)

where p is a constant factor, commonly set as p=0.1.

NB The Jaro Winkler similarity can be higher than 1, unlike the other metrics implemented in Cello.

Approximate search

We implement a naif form of approximate search for strings. The algorithm is as follows: when looking for a pattern we randomly select a substring of the pattern whose length is a given fraction (exactness) of the pattern itself. We then search for this substring exactly in the target string. If we find it, we focus on a window around this match having the same length as the pattern. We compare the similarity of the window with the pattern itself, using one of the similarity functions above. If this is above a given threshold (tolerance) we accept the match and return the position of the window; otherwise we try with another attempt. After a certain number of attempts fail, we return -1.

The algorithm is driven by the following type:

type
  Similarity {.pure.} = enum
    RatcliffObershelp, Levenshtein, LongestSubstring, Jaro, JaroWinkler
  SearchOptions = object
    exactness, tolerance: float
    attempts: int
    similarity: Similarity

and can be used like this:

let
  s = someLongString
  pattern = someShortString
  index = searchIndex(s)
  options = searchOptions(exactness = 0.2)
  position = index.searchApproximate(x, pattern, options)

echo position

The defaults are exactness = 0.1, tolerance = 0.7, attempts = 30 and similarity = Similarity.RatcliffObershelp

TODO

• Improve DC3 algorithm
• More applications of suffix arrays
• Construct wavelet trees in threads
• Make use of SIMD operations to improve performance
• Allow data structures to work on memory-mapped files
• Implement assembly on top of FM indices following this thesis

Thanks

The logo comes from cliparts.co