In this paper, we consider the open problem of the compelxity of the LLL algorithm in the case when the approximation parameter of the algorithm has its extreme value 1. This case is of interest because the output is then the strongest lovasz-reduced basis. Experiments reported by Lagarias and Odlyzko (13) seem to show that the algorithm remain polynomial in average. However no bound better than a naive exponential order one is estalished for the worst-case complexity of the optimal LLL algorithm, even for fixed small dimension (higher than 2). Here we prove that, for any fixed dimension n, the number of iterations of the LLL algorithm is linear with respect to the size of the input. It is easy to deduce from (17) that the lienar order is optimal. Moreover in 3 dimensions, we give a tight bound for the maximum number of iterations and we characterize precisely the output basis. Our bound also improves the known one for the usual (non-optimal) LLL algorithm.
展开▼
