In single-reference coupled-cluster (CC) methods, one has to solve a set of non-linear polynomial equations in order to determine the so-called amplitudes which are then used to compute the energy and other properties. Although it is of common practice to converge to the (lowest-energy) ground-state solution, it is also possible, thanks to tailored algorithms, to access higher-energy roots of these equations which may or may not correspond to genuine excited states. Here, we explore the structure of the energy landscape of variational CC (VCC) and we compare it with its (projected) traditional version (TCC) in the case where the excitation operator is restricted to paired double excitations (pCCD). By investigating two model systems (the symmetric stretching of the linear H 4 molecule and the continuous deformation of the square H 4 molecule into a rectangular arrangement) in the presence of weak and strong correlations, the performance of VpCCD and TpCCD are gauged against their configuration interaction (CI) equivalent, known as doubly-occupied CI (DOCI), for reference Slater determinants made of ground-or excited-state Hartree-Fock orbitals or state-specific orbitals optimized directly at the VpCCD level. The influence of spatial symmetry breaking is also investigated.
I. COUPLED CLUSTER AND STRONG CORRELATIONSingle-reference (SR) coupled-cluster (CC) methods offers a reliable description of weakly correlated systems through a well-defined hierarchy of systematically improvable models. [1][2][3][4][5] On top of this hierarchy stands full CC (FCC), which is equivalent to full configuration interaction (FCI), and consequently provides, at a very expensive computational cost, the exact wave function and energy of the system in a given basis set. Fortunately, more affordable methods have been designed and the popular CCSD(T) method, which includes singles, doubles and non-iterative triples, is nowadays considered as the gold standard of quantum chemistry for ground-state energies and properties. 6,7 Despite its success for weakly correlated systems, it is now widely known that CCSD(T) flagrantly breaks down in the presence of strong correlation as one cannot efficiently describe such systems with a single (reference) Slater determinant. This has motivated quantum chemists to design multi-reference CC (MRCC) methods. [8][9][10][11][12] However, it is fair to say that these methods are computationally demanding and still far from being black-box.Because SRCC works so well for weak correlation, it would be convenient to be able to treat strong correlation within the very same framework. This is further motivated by the fact that one can compensate the poor quality of the reference wave function by simply increasing the maximum excitation degree of the CC expansion. However, this is inevitably associated with a rapid growth of the computational cost, and hence one cannot always afford this brute-force strategy. The development of SR-based methods for strong correlation is ongoing and some of them (usually based on the...