Gong, Chen, and Wu (2016) introduced Cholesky decomposition (CD) of the two-electron repulsion integrals into the valence bond self-consistent field (VBSCF) method, removing the integral-transformation bottleneck at a decomposition threshold of 10−610^{-6} with total-energy errors below 10−610^{-6} Hartree, and left a linked three-part agenda open: extend CD to the post-VBSCF dynamic-correlation methods VBCI and VBPT2 with accuracy and efficiency validation; accelerate the shifted computational hotspots, namely Hamiltonian and overlap matrix construction over the nonorthogonal valence bond structure basis and the structure-coefficient updates, for large active spaces; and characterize the tradeoff among decomposition threshold, number of Cholesky vectors, and wall-clock time across broader chemical spaces, deriving threshold-selection strategies. A 2025 follow-up from the same group integrated resolution-of-the-identity (a factorization of the two-electron integrals over an auxiliary atom-centered basis) and chain-of-spheres-exchange approximations into VB orbital optimization, extending VB applicability beyond 100 atoms, and a companion 2025 work introduced block-orthogonalized orbitals for post-VBSCF calculations along an alternative route. However, no accessible evidence shows a validated CD or other low-rank integral implementation inside VBCI or VBPT2, nor a systematic threshold-strategy study across diverse chemical spaces. The post-VBSCF low-rank extension, the large-active-space matrix-construction bottleneck, and the threshold-selection question therefore remain open.

Background

In valence bond (VB) theory the electronic wavefunction is written as a linear combination of chemically interpretable VB structures, each a spin-adapted product of localized, generally overlapping (nonorthogonal) one-electron orbitals, in the spirit of Lewis structures. The valence bond self-consistent field (VBSCF) method variationally optimizes both the coefficients of the VB structures and the orbitals themselves; it captures static (nondynamic) electron correlation but recovers little dynamic correlation. Dynamic correlation is added by post-VBSCF methods: VBCI (valence bond configuration interaction) adds excited VB structures built from the optimized VB orbitals and redetermines the structure coefficients, while VBPT2 (valence bond second-order perturbation theory) evaluates a perturbative second-order correction on the frozen VBSCF wavefunction. These methods are implemented in the XMVB program family of the Xiamen VB group. A central computational obstacle is the two-electron repulsion integral (ERI) tensor (pq∣rs)(pq|rs) over basis functions: at every optimization iteration it must be transformed into the nonorthogonal VB orbital basis, for every pair of VB structures, and because VB orbitals are nonorthogonal none of the simplifications available in molecular-orbital methods apply. The Cholesky decomposition (CD) technique approximates the ERI tensor as a sum of MM products of Cholesky vectors, where the decomposition threshold δ\delta controls the rank MM and hence the accuracy: smaller δ\delta means more vectors and higher accuracy at greater cost. CD had long been used in conventional molecular-orbital quantum chemistry; Gong, Chen, and Wu (J. Comput. Chem. 37, 2157, 2016) applied it to VBSCF and showed on ethylene and polyene test sets (C2nH2n+2C_{2n}H_{2n+2} and C2nH4n−2C_{2n}H_{4n-2}, n=1n=1-77) that a threshold of 10−610^{-6} keeps the total-energy error below 10−610^{-6} Hartree and that the integral transformation ceases to be the bottleneck. Once the integral transformation ceases to be the bottleneck, the computational hotspot for large active spaces (many VB structures and orbitals) shifts to constructing the Hamiltonian and overlap matrices over the nonorthogonal VB structure basis and to updating the structure coefficients, and how CD-type integral treatment should be carried into VBCI and VBPT2 remains open: excited VB structures multiply the number of required integral transformations, and perturbative denominators add further transformed intermediates. A broader open question of practical importance is the choice of the threshold δ\delta: the tradeoff among δ\delta, the resulting number of Cholesky vectors MM, wall-clock time, memory, and achieved accuracy was characterized only on a narrow set of test molecules, so no threshold-selection strategy existed for diverse chemical spaces and VB wavefunction classes.

Problem Statement

Complete low-rank (Cholesky-type) two-electron-integral treatment across the ab initio valence bond method family, extending the 2016 CD-VBSCF work and narrowed by subsequent progress. Concretely: (1) implement and validate the Cholesky decomposition (or an equivalent low-rank ERI factorization) within the post-VBSCF dynamic-correlation methods VBCI and VBPT2 - assessing accuracy against conventional exact-ERI references, efficiency gains, and integration with existing post-VBSCF machinery; (2) accelerate the remaining computational hotspots for large active spaces, namely the construction of the Hamiltonian and overlap matrices over the nonorthogonal VB structure basis and the evaluation of structure-coefficient updates, beyond what the 2025 acceleration of orbital-optimization-level VB by the resolution-of-the-identity (RI) and chain-of-spheres-exchange (COSX, a seminumerical grid-based evaluation of the exchange contribution) approximations already covers; and (3) characterize the tradeoff among the decomposition threshold δ\delta, the resulting number of Cholesky vectors MM, and wall-clock time and memory across chemically diverse spaces, and derive threshold-selection strategies that balance desired energy accuracy (including accuracy of relative energies) against computational cost, including their impact on the convergence behavior of VB optimizations for diverse VB wavefunctions. Strand (1): the low-rank vectors must be contracted through the additional integral classes that VBCI and VBPT2 require, with the accuracy of the resulting treatment established against conventional exact-ERI references and the cost of the additional transformed intermediates determined. Strand (2): constructing the Hamiltonian and overlap matrices over the nonorthogonal VB structure basis and updating the structure coefficients must achieve a cost scaling in the number of VB structures and orbitals lower than that of the current approach while preserving the accuracy of the unaccelerated result. Strand (3): this characterization must cover at least three system classes drawn from distinct chemistry families (for example, saturated hydrocarbons, conjugated π\pi systems, and heteroatom-containing molecules) and at least two VB wavefunction classes, and must yield an explicit accuracy-versus-cost threshold-selection rule. These strands jointly serve one objective: making accurate, dynamically correlated ab initio VB calculations routine for larger and chemically more diverse systems. Each strand is a sub-goal in its own right; the objective as a whole requires all three.

References

  1. Xiping Gong, Zhenhua Chen, Wei Wu, 'The application of Cholesky decomposition in valence bond calculation', Journal of Computational Chemistry 37, 2157-2162 (2016). DOI: 10.1002/jcc.24442, https://doi.org/10.1002/jcc.24442
  2. Chenru Ji, Yueyang Zhang, Fuming Ying, Chen Zhou, Wei Wu, 'Low-Rank Algorithms for Ab Initio Valence Bond Approaches', Journal of Chemical Theory and Computation 21, 2462-2471 (2025). DOI: 10.1021/acs.jctc.4c01787, https://doi.org/10.1021/acs.jctc.4c01787
  3. Chan Cao, Chen Zhou, Wei Wu, 'Block-orthogonalized orbitals and their applications in post-valence bond self-consistent field calculations', Chinese Science Bulletin 70, 5862-5870 (2025). DOI: 10.1360/tb-2024-1289, https://doi.org/10.1360/tb-2024-1289
  4. Chen Zhou, Xun Wu, Fuming Ying, Wei Wu, 'Ab Initio Valence Bond Theory for Strongly Correlated Systems', Journal of Chemical Theory and Computation 21, 6243-6257 (2025). DOI: 10.1021/acs.jctc.5c00596, https://doi.org/10.1021/acs.jctc.5c00596
  5. Zhenhua Chen, Wei Wu, 'Ab initio valence bond theory: A brief history, recent developments, and near future', The Journal of Chemical Physics 153, 090902 (2020). DOI: 10.1063/5.0019480, https://doi.org/10.1063/5.0019480
  6. T. B. Pedersen, S. Lehtola, I. Fdez. Galvan, R. Lindh, 'The versatility of the Cholesky decomposition in electronic structure theory', WIREs Computational Molecular Science 14, e1692 (2023). DOI: 10.1002/wcms.1692, https://doi.org/10.1002/wcms.1692
  7. Chenru Ji, Fuming Ying, Peifeng Su, Chen Zhou, Wei Wu, 'Implementation of molecular symmetry in valence bond calculation', Journal of the Chinese Chemical Society 70, 341-348 (2023). DOI: 10.1002/jccs.202200476, https://doi.org/10.1002/jccs.202200476
  8. Huaiyu Zhang, Chen Zhou, Y. Mo, Wei Wu, 'Performance of the VBSCF method for pericyclic and pi bond shift reactions', Journal of Computational Chemistry 40, 1123-1129 (2019). DOI: 10.1002/jcc.25729, https://doi.org/10.1002/jcc.25729
  9. Xuhui Lin, Xin Liu, Fuming Ying, Zhenhua Chen, Wei Wu, 'Explicit construction of diabatic state and its application to the direct evaluation of electronic coupling', The Journal of Chemical Physics 149, 044112 (2018). DOI: 10.1063/1.5035114, https://doi.org/10.1063/1.5035114