The second step in initialization is to generate a random matrix where each . is a data structure that is stored in addition to the table, though if is generated pseudorandomly, then storing the seed of the pseudorandom process is sufficient for to be regenerated when needed. The -element table is then generated as (where ``'' denotes equivalence modulo ). That is, is the table as described in Section 2; intuitively, the element in the -th ``row'' and -th ``column'' for and is .
To complete initialization, is deleted, and (or the seed needed to regenerate it), the table , and prime are stored for the next key regeneration attempt. In addition, is stored for some (different) one-way and collision-resistant function , so that when is reconstructed, it can be recognized as correct.
After a sufficient number of successful key reconstructions (see Section 4.2), the table is ``hardened'' as described in Section 2: if over a number of successful key reconstructions, each induced feature descriptor is consistent on the -th feature (i.e., is usually the same, as specified more precisely in Section 5), then element is assigned to be a random element of . This should usually not affect reconstruction for the correct user, since that user typically selects .