Constrained Optimization Algorithms for Orbital Optimization in Quantum Chemistry

Abstract

We present a modular constrained-orbital-optimization framework for quantum chemistry. The formulation separates the correlated electronic-structure solver from the orbital optimizer: the solver supplies one- and two-particle reduced density matrices, while the molecular orbitals are updated on the orthonormality-constrained Stiefel manifold with an implicit steepest-descent algorithm. Because the orbital optimizer only requires reduced density matrices, MP2, CASCI, and DMRG can be treated within the same interface. For CASCI solvers, the approach is closely related to optimal-orbital full configuration interaction and CASSCF\cite{helgaker_MulticonfigurationalSelfConsistentField_2000a}, but uses a solver-independent constrained-optimization update rather than CAS-specific orbital-rotation equations. When conventional CASSCF orbital-rotation iterations converge to higher-energy local solutions, CO-CAS can recover lower-energy stationary solutions. We also introduce a modified direct inversion in the iterative subspace procedure to accelerate macro-iteration convergence and a dynamical-weighting scheme to improve state-averaged excited-state calculations. Applications to LiF, H$_2$O, and pyrazine show that orbital optimization lowers energies relative to fixed-orbital MP2, CASCI, and DMRG references while improving convergence and potential-energy-curve smoothness.