# Algorithm for length of longest common subsequence

The case of multiple strings. A slight modification of the dynamic programming algorithm for two strings is used as a subroutine. Here is the pseudo code:

# The modified dynamic programming algorithm for longest common subsequence.
# Inputs: x, y - the strings
xL, yL - their respective length vectors
lcs(x, xL, y, yL)
u = |x|
v = |y|
T = 2-dimensional table with default value 0 for all elements.
for i in 0..u-1
for j in 0..v-1
if x[i] == y[j]
T[i, j] = T[i - 1, j - 1] + 1
else
T[i, j] = max(T[i - 1, j], T[i, j - 1])
# limit the value of the table element by the length vectors:
T[i, j] = min(T[i, j], xL[i], yL[j])
# Take the element wise minimum of the vectors for each of the strings and the bottom row and right-most column respectively.
return x, element-wise-minimum-of(xL, T[_, v - 1]), y, element-wise-minimum-of(yL, T[u - 1, _])

# Algorithm to compute the length of the longest common subsequence of multiple strings.
# Inputs: strings - a list of strings
mlcs(strings):
if for any string x: |x| == 0
return 0
R = a dictionary indexed by a string x, where the value is initialized to be the length vector of x (initially [1,2,..,|x|]).
N = initially a copy of R.
while true
for i in 0..|strings|
for j in i + 1..|strings|
x, y = strings[i], strings[j]
_, xL, _, yL = lcs(x, R[x], y, R[y])
R[x], R[y] = xL, yL
if for every string x: N[x] is element wise equal to R[x]
break
else
N[x] = R[x]
return the last element of any of the vectors in R


What follows is an implementation in the python language.

import collections

def lcs(x, xL, y, yL):
u = len(x)
v = len(y)
T = collections.defaultdict(int)
for i in range(u):
for j in range(v):
if x[i] == y[j]:
T[i, j] = T[i - 1, j - 1] + 1
else:
T[i, j] = max(T[i - 1, j], T[i, j - 1])
T[i, j] = min(T[i, j], xL[i], yL[j])
return (x, [min(xL[i], T[i, v - 1]) for i in range(u)]), (y, [min(yL[j], T[u - 1, j]) for j in range(v)])

def mlcs(strings):
if any(len(x) == 0 for x in strings):
return 0
R = {x: list(range(1, len(x) + 1)) for x in strings}
N = {x: R[x] for x in strings}
while True:
for i in range(len(strings)):
for j in range(i + 1, len(strings)):
x, y = strings[i], strings[j]
(_, xL), (_, yL) = lcs(x, R[x], y, R[y])
R[x] = xL
R[y] = yL
if all(all(N[x][i] == R[x][i] for i in range(len(x))) for x in strings):
break
else:
for x in strings:
N[x] = R[x][:]
return R[strings[-1]][-1]


The idea is to initialize for each input string a vector which holds information for each position of that string about what is the maximum length of a LCS up to that position. These are updated each time a pair-wise LCS computation is done. This in turn is done for all pairs until all the vectors are stable (ie. don't change) between iterations.

So far testing gives me mixed results. The correct length at first seems to be produced pretty reliably, but when trying to use the function mlcs as an oracle in an algorithm for retrieving an actual longest sequence, sometimes the result is not what is expected. This may or may not indicate that there are cases for which an incorrect length is produced.

So on to the question: Is the idea of the length computing algorithm sound?

• en.wikipedia.org/wiki/… – adrianN Jul 14 '16 at 7:41
• Posting a big blob of code is a poor fit for this site, I think. Could you reformulate your problem using concise pseudo-code? – adrianN Jul 14 '16 at 7:43
• Also, please add what you have tried towards proving correctness of your algorithm, and where you get stuck. – Raphael Jul 14 '16 at 9:10
• By sound I meant that it would always produce the correct result, which is as we have seen not true, so I am going to add an answer to this question. – user72935 Aug 16 '16 at 6:22

The algorithm is incorrect. That is it may return a length that is greater than the actual length of a longest common subsequence. Here is a test case which fails to produce the correct length: ['baaa', 'aaba', 'baba'].

• This answer can be improved by explaining why the algorithm fails conceptually. – Raphael Aug 16 '16 at 9:36
• I doubt that algorithm has any concept, so it is hard to fail conceptually. And "is incorrect" with a counterexample is absolutely good enough. – gnasher729 Aug 17 '16 at 14:03

I didn't try to read your pseudocode, but based on your description, if I understand your idea correctly, it has a (probably fatal) problem: Even if all pairs $(i, j)$ of strings have a common subsequence with length $\ge k$, it may be that the LCS of the set of all strings has length $< k$.

Here's an example:

AB
BC
CA


Every pair of strings has an LCS of length 1, but no character appears in all 3 strings, so the LCS of the set of all 3 strings is 0.

• So, initially we have the strings AB, BC, CA with the associated vectors [1 2], [1 2], [1 2]. Now, AB×BC gives AB=[0 1], BC=[1 1]. Then AB×CA gives AB=[0 0], CA =[0 0] and finally BC×CA gives BC=[0 0] and CA=[0 0], and we've reached a non-changing state of the vectors and we can conclude that the answer is 0. – user72935 Jul 14 '16 at 11:28
• Then I didn't fully understand your idea. Since you know that the function returns incorrect values sometimes, I suggest using standard debugging techniques to isolate a small counterexample, which will allow to see whether it is a fundamental problem with the algorithm or just a bug in your implementation of it. (To automate this I would generate all possible inputs in increasing size order, looking for the first one for which your algorithm and a known-good algorithm return different results.) – j_random_hacker Jul 14 '16 at 11:39
• I haven't been able to pin-point the problem. I have been concentrating on exploring with the implementation which can be tracked at github.com/batprey/lcs and now it seems the code is at a point where a sequence the length of an embedded common sequence in the test strings can be found (maybe I should say most of the time to be cautious). But embedding a sequence in the test strings does not guarantee that there is not a sequence even longer than that common to the strings which should be found instead were the algorithm working correctly. – user72935 Jul 19 '16 at 20:31
• Also, as to your good suggestions @j_random_hacker indeed I must test more systematically. – user72935 Jul 19 '16 at 20:32
• Your counterexample ['baaa', 'aaba', 'baba'] is a good one -- that should be small enough that you can figure out by hand what your algorithm should be doing with it. – j_random_hacker Jul 22 '16 at 13:31