We present an algorithm to reduce the computational complexity for plane-wa
ve codes used in electronic structure calculations. The proposed algorithm
avoids the diagonalization of large Hermitian matrices arising in such prob
lems. The computational time for the diagonalization procedure typically gr
ows as the cube of the number of atoms, or the number of eigenvalues requir
ed. To reduce this computational demand, we approximate directly the occupa
tion operator corresponding to the eigenvectors associated with the occupie
d states in a certain subspace without actually computing these eigenvector
s. A smoothed Chebyshev-Jackson expansion of the Heaviside function of the
Hamiltonian matrix is used to represent the occupation operator. This proce
dure requires only matrix-vector products and is intrinsically parallelizab
le. (C) 1999 Elsevier Science B.V.