The Art Gallery Guardian

Find the period of a nice eventually periodic sequence

A sequence is periodic if s_i = s_{i+p} for all i, where p>0 is called a period. A sequence is eventually periodic if there exists a n and p, such that s_i = s_{i+p} for all i>n. The sequence with index above n is the periodic part.

A sequence is called u-normal, if there exists s_i=s_{i+p} for some p>0 for all i in a interval of length u, then the sequence starting at s_i is part of the periodic part.

When does u-normal sequence comes up? Consider we have a recurrence relation that produces a sequence. Say it is of the form a_n = f(a_{n-1},a_{n-2},\ldots,a_{n-u}). The sequence a_1,\ldots is u-normal.


Given a oracle that can take input i and return the ith element in a u-normal eventually periodic sequence a. Find the smallest lexicographic pair (n,p) where p>0, such that a_i=a_{i+p} for all i>n.

One can solve this problem in O(u \log \frac{n}{u}) time. First, consider the subsequence a_u,a_{2u},\ldots. We guess an upper bound on n through exponential search in the subsequence. There is a O(u) time algorithm to decide if n'\geq n. For example using the KMP algorithm. We can quickly locate a n' such that n\in [n'-u,n'].

Let's take the sequence a_{n'-u},\ldots,a_{n'+3u}. We just need to solve the following problem. Given a O(u) length string, find the longest suffix that appears at least twice in the sequence. Let such suffix be s. We know |s|\geq 3u, which s has to overlap with any other occurrence of the sequence. The claim is that the partial match table in the KMP algorithm would give us such information. Hence we can obtain n. p can also be obtained in the same time. Note that KMP algorithm only uses the fact one can check equality of two elements. So the sequence can contain elements from very general space.

The total running time is O(\log \frac{n}{u}) calls to O(u) time string matching. The total running time is therefore O(u\log \frac{n}{u}).

In many applications, we do not get oracle access to ith index of the sequence. But we can read the sequence from the beginning as a list. In that case, we don't do binary search, but linear search. Advance the index by u and obtain n', and test if n is no larger than the current point. If so, again we have n\in [n'-u,n'] and reduce to the previous problem. If not, advance the index by u again and repeat. This gives us a O(n+u) time algorithm.

Posted by Chao Xu on 2019-02-05.
Tags: algorithms, infinite sequence.