A normalization scheme for the non-symmetric s-step Lanczos algorithm.
The Lanczos algorithm is among the most frequently used techniques for computing a few dominant eigenvalues of a large sparse non-symmetric matrix. When variants of this algorithm are implemented on distributed-memory computers, the synchronization time spent in computing dot products is increasingly limiting the parallel scalability. The goal of s-step algorithms is to reduce the harmful influence of dot products on the parallel performance by grouping several of these operations for joint execution; thus, plummeting synchronization time when using a large number of processes. This paper extends the non-symmetric s-step Lanczos method introduced by Kim and Chronopoulos (J. Comput. Appl. Math., 42(3), 357-374, 1992) by a novel normalization scheme. Compared to the unnormalized algorithm, the normalized variant improves numerical stability and reduces the possibility of breakdowns.