I was wondering if the possibility of using one single diagonalization
for each scf step may be applicable to the siesta method:
http://dx.doi.org/10.1103/PhysRevB.86.174308
"Here we show how to avoid a major computational bottleneck: the
self-consistent-field optimization prior to force calculations. The
optimization-free quantum mechanical molecular dynamics method gives
trajectories that are almost indistinguishable from an "exact"
microcanonical Born-Oppenheimer molecular dynamics simulation even when
low-prefactor linear scaling sparse matrix algebra is used."
