We compare various algorithms for large matrix diagnalization, namely
the Davidson and Olsen-Davidson algorithms, and a Lanczos-like one, as
well as a new method introduced here. The numerical tests are perform
ed with the help of a model involving a vibronic Hamiltonian and a bat
h. For various cases, it is shown that the convergence of the Davidson
and Olsen-Davidson methods is sensitive to the adequacy of the trial
vector. The new method presents convergence properties similar to the
Davidson and Olsen-Davidson methods in their most favorable case, but
thes convergence properties are not destroyed by a poor guess of the t
rial vector.