For small N, one can do this in a straightforward way for by storing a table of . For larger N, once memory becomes insufficient, one instead computes and at each step. This requires 3N of the basic steps to reach n = N, rather than 2N.
If we have not proved that the sequence is periodic by the time we have reached , then we know that D = m+p > N. This observation accounts for only one of the lower bounds in Table 2, namely , for which the lower bound is exactly .
A more useful method for obtaining lower bounds on m+p is based on the following observation. Let denote the maximum of the absolute values of the coefficients of . During the computation, maintain a list of the record values of , i.e., those n for which for all k < n. Clearly, if is a record, then are all distinct, so we know that D > n. In those cases for which a preliminary computation up to , say, had indicated that did not have a small period, our program continued to compute up to without employing Floyd's algorithm. If the last record occurred for , then this was larger than could have been obtained from Floyd's algorithm and the cost was less than the cost of using that algorithm. On the other hand, in those rare cases where the last record occurred for , the computation was repeated using Floyd's algorithm and a relatively small was usually detected. For example, with , a computation up to indicated that the last record occurred at n = 1782995, suggesting a periodic sequence. A recomputation of in the range found the values .
As mentioned above, a list was maintained of values of for which the computation of required quadruple precision. Let us call these n markers. Suppose one has computed , where N > m+2p, and that one of these markers, , occurs within the periodic part of the sequence, that is, for . Then, by periodicity, must be a marker and thus must occur within the list of markers. Thus, by scanning this list of for repeats, one can obtain p directly, as well as an upper bound on m. A recomputation of in a short range enables one to compute m exactly.
For example, with , we computed up to and determined that the last record occurred for n = 45622056, suggesting that the sequence was possibly periodic. A check of the list of markers revealed that for n = 39667761 and n + q = 132886569, and that no marker k between these values had . Thus, the period was revealed to be p = q = 93218808. Since is a marker, but is not, the list also showed that , and a binary search in this range revealed that m = 39420662.
On the other hand, for , we computed up to . The last record of (which was 81363), occurred for n = 1199978517, giving the lower bound for D = m + p listed in Table 2. The computation of values of required hours of CPU time on the Amdahl 5860. Even if , it does not seem practical to compute p in this case by a direct approach unless p should happen to be quite small relative to m+p.
In addition to the markers, a table of for multiples
of was also maintained, so that the computation could be restarted from any such
value without having to recompute the entire sequence. Although the markers here arose
in a natural way from the algorithm employed, one could clearly also use some more
artificial criterion for inventing markers. For example, one could store all those
pairs with the first component of divisible by 1000. Then, in a
computation up to one would expect about values of to be stored,
and that one of these values would occur within the period unless p was unusually