Fast spectral projection algorithms for density-matrix computations (Q1306101): Difference between revisions

The goal of this paper is to construct fast algorithms for computing spectral projectors. These projectors are to be used to compute density matrices as in the Kohn-Sham scheme where the ground state density of an arbitrary interacting system of \(n\) electrons is sought. The ordinary sign function is defined on the reals, it equals \(1\) on \(]0,\infty[\), \(0\) at \(0\) and \(-1\) on \(]-\infty,0[\). For a matrix or operator of the form \[ T(x,y)= \sum_j \lambda_j\psi_j(x)\overline\psi_j(y) \] with \(\lambda_j\) real and \(\{\psi_j\}\) orthogonal, one defines \[ \text{sign}(T)(x, y)= \sum_j \text{sign}(\lambda_j) \psi_j(x)\overline\psi_j(y). \] The algorithm to compute \(\text{sign}(T)\) is a polynomial recursion with \(T_0= {T\over\|T\|_2}\) and \(T_{k+ 1}= (3T_k- T^3_k)/2\) for \(k= 0,1,\dots\)\ . One of the key ideas in the paper is to preserve sparsity in this iteration by using wavelets or other representations such as local cosine bases. A one-dimensional numerical example is given to support the claims made in the paper; the extension to multiple dimensions is also discussed.