Semidefinite programming applications to Hartree-Fock and linear scaling electronic structure theories.

摘要

Semidefinite programming (SDP) is a relatively modern subfield of convex optimization which has been applied to many problems in the reduced density matrix (RDM) formulation of electronic structure. SDPs deal with minimization (or maximization) of linear objective functions of matrices, subject to linear equality and inequality constraints and positivity constraints on the eigenvalues of the matrices. Energies of chemical systems can be expressed as linear functions of RDMs, whose eigenvalues are electron occupation numbers or their products which are expected to be non-negative. Therefore, it is perhaps not surprising that SDPs fit rather naturally in the RDM framework in electronic structure. This dissertation presents SDP applications to two electronic structure theories.;The first part of this dissertation (chaps. 1-3) reformulates Hartree-Fock theory in terms of SDPs in order to obtain upper and lower bounds to global Hartree-Fock energies. The upper and lower bounds on the energies are frequently equal thereby providing a first-ever certificate of global optimality for many Hartree-Fock solutions. The SDP approach provides an alternative to the conventional self-consistent field method of obtaining Hartree-Fock energies and densities with the added benefit of global optimality or a rigorous lower bound. Applications are made to the potential energy curves of (H 4)2, N2, C2, CN, Cr2 and NO2. Energies of the first-row transition elements are also calculated. In chapter 4, the effect of using the Hartree-Fock solutions that we calculate as references for coupled cluster singles doubles calculations is presented for some of the above molecules.;The second part of this dissertation (chap. 5) presents a SDP approach to electronic structure methods which scale linearly with system size. Linear scaling electronic structure methods are essential in order to make calculations on large systems feasible. Among these methods the so-called density matrix based ones seek to minimize the energy as a function of the one-electron reduced density matrix for a given effective one-electron Hamiltonian, subject to trace and idempotency constraints which ensure the correct number of electrons and a Slater determinant pre-image, respectively. We reformulate this minimization as a SDP, generalize it to non-orthogonal bases and exploit sparsity to ensure linear scaling. Since this approach relaxes the idempotency constraint it eliminates the errors and problems associated with popular linear scaling approaches which choose to enforce idempotency. Computation times with the SDP approach are presented for one-dimensional hydrogen chains.