TD$Δ$SCF: Time-Dependent Density Functional Theory with a Non-Aufbau Reference for near-degenerate states

Abstract: Near-degenerate electronic structures remain a major challenge for conventional single-reference density functional theory (DFT). To address this problem, we propose time-dependent $Δ$SCF (TD$Δ$SCF), a novel linear-response scheme in which a non-Aufbau $Δ$SCF determinant serves as the reference for a subsequent TDDFT calculation. In contrast to collinear spin-flip (SF)-TDDFT, this formulation preserves the usual Coulomb and exchange-correlation response contributions while describing the target states from an electronically excited reference. We examine the performance of TD$Δ$SCF for several prototypical problems involving near-degeneracy, including the torsional potential of ethylene, singlet--triplet gaps of representative diradicals, geometry optimizations of benzyne isomers, and bond-dissociation curves of hydrogen fluoride and F$_2$. Across these tests, TD$Δ$SCF shows markedly weaker functional dependence than SF-TDDFT and often yields a more balanced description of challenging singlet states. In particular, it provides smooth torsional potentials, improved singlet--triplet gaps, a consistent monocyclic structure for singlet $m$-benzyne, and a more satisfactory description of bond dissociation without the spurious low-lying states found in SF-TDDFT. At the same time, the method exhibits a systematic tendency to overestimate singlet energies and can lose accuracy when the underlying $Δ$SCF reference is not well suited to the final state. We also identify a numerical instability that can arise in non-Aufbau calculations and trace its origin to the exchange-correlation potential near uncompensated nodal regions. These results highlight both the promise and the practical limitations of TD$Δ$SCF as a low-cost method for singlet states with near-degenerate electronic structures.

• The paper demonstrates that TDΔSCF mitigates functional sensitivity in TDDFT by using a non-Aufbau reference to accurately model near-degenerate states.
• It employs a linear-response framework retaining the full DFT kernel, yielding smoother torsional curves and reliable singlet–triplet energy gaps in diradicals.
• Numerical tests on ethylene, benzyne, and bond dissociation benchmark cases validate the method’s performance despite challenges like orbital instability and overestimated singlet energies.

TDΔSCF: Linear-Response TDDFT with Non-Aufbau References for Near-Degenerate States

Introduction

The paper "TD$Δ$SCF: Time-Dependent Density Functional Theory with a Non-Aufbau Reference for near-degenerate states" (2603.29948) presents a rigorous investigation into a novel linear-response protocol—TD$Δ$SCF: Time-Dependent Density Functional Theory with a non-Aufbau self-consistent field reference—targeting electronic structure problems where near-degeneracy or static correlation severely impedes the reliability of conventional single-reference DFT/TDDFT methodologies. The formalism is compared extensively with standard collinear spin-flip TDDFT (SF-TDDFT), with a focus on fundamental differences in kernel structure, functional dependence, the quality of predicted potential energy surfaces (PES), evaluation of singlet–triplet (S–T) gaps, molecular structure optimization, bond dissociation, and the exposure of an intrinsic numerical instability when using non-Aufbau determinants.

Formulation of TD$Δ$SCF and Relationship to SF-TDDFT

TD$Δ$SCF is constructed by taking as its reference determinant a variationally optimized non-Aufbau (open-shell) solution obtained via constrained SCF (e.g., MOM, STEP), and performing spin-conserving (single excitation) linear-response TDDFT in the Tamm–Dancoff approximation (TDA). In contrast, collinear SF-TDDFT employs a high-spin reference (typically a triplet $M_S=1$), with the target (low-spin) states accessed via spin-flip excitation operators; here, the Coulomb and standard XC kernel contributions vanish in the effective coupling—a result which leaves the SF protocol highly sensitive to the admixture of Hartree–Fock exchange and yields pronounced functional dependence and stability problems, especially for pure and semi-local functionals.

In TD$Δ$SCF, the response manifold is built on the promoted, often open-shell, SCF determinant (Figure 1), so that for near-degenerate configurations (e.g., biradicals, bond-breaking, twisted π systems), both the static correlation and the dynamical response from conventional hybrid or (meta-)GGA functionals remain present in the excitation kernel.

Figure 1: Schematic comparison of SF-TDDFT and TD$Δ$SCF reference choices and excitation spaces.

Performance for Near-Degenerate and Strong Correlation Problems

Ethylene Torsion Benchmark

Torsion about the C=C bond in ethylene is a canonical example of a system with changing multiconfigurational character. Standard DFT/TDDFT curves show unphysical cusps near 90°, indicative of the breakdown of the single-reference ansatz. TD$Δ$SCF yields a smooth, physically plausible torsional PES for all three functionals (BLYP, B3LYP, BHHLYP), closely reproducing the MRCISD high-level reference near the degeneracy point without the abrupt non-differentiability seen in SF-TDDFT (Figure 2), especially for pure functionals.

Figure 2: Ethylene torsional barriers calculated using TD$Δ$SCF and SF-TDDFT. TD$Δ$SCF gives smooth barriers, free of cusps present in SF-BLYP.

While SF-TDDFT dramatically overestimates the torsional barrier with pure functionals, TD$Δ$0SCF exhibits much weaker functional sensitivity, underestimating barriers primarily due to deterioration of the $Δ$1 end-point—an area still adequately described by ground-state DFT. Re-evaluating barrier heights with KS-DFT as reference for the planar geometry restores quantitative agreement across functionals, revealing that TD$Δ$2SCF's deficiencies are localized to the non-degenerate regime, consistent with its design focus.

Singlet–Triplet Gaps in Diradicals and Benzyne Isomers

S–T gaps for a set of atomic and molecular diradicals are systematically evaluated. Conventional DFT struggles due to single-determinant bias, with systematic errors growing with HF exchange fraction. SF-TDDFT, especially with BLYP, fails catastrophically (MAE > 200 kcal/mol); the error decreases with hybrid functionals but remains functional-dependent and biased toward underestimating gaps, especially for atoms and one-center diradicals. In contrast, TD$Δ$3SCF consistently produces accurate gaps with a weak dependence on the underlying functional (MAEs: 4.3–4.7 kcal/mol for BLYP/B3LYP/BHHLYP; Figure 3), always avoiding the catastrophic failures of SF-TDDFT.

Figure 3: Comparison of computed and experimental singlet–triplet energy gaps ($Δ$4) for a benchmark set, highlighting the strong functional dependence and severe errors of SF-TDDFT (BLYP) and the modest, weakly functional-dependent errors of TD$Δ$5SCF.

Importantly, TD$Δ$6SCF does systematically overestimate S–T gaps compared to experiment, reflecting a bias inherited from the non-Aufbau nature of the reference and lack of optimization on the singlet surface in the response step.

Optimization of Benzyne Isomer Singlet Geometries

Geometry optimizations on o-, m-, and p-benzyne singlet surfaces further reveal functional and methodological differences. For o- and p-benzyne, both protocols yield comparable, reliable geometries (deviation <0.04 Å, <6° from SF-CCSDT reference). For m-benzyne, SF-TDDFT with B3LYP predicts an unphysical bicyclic structure, consistent with known functional pathologies; TD$Δ$7SCF, by contrast, retains the monocyclic minimum preferred by high-level benchmark and quantum Monte Carlo calculations, with systematic agreement among various functionals (Figure 4).

Figure 4: Singlet-state optimized geometrical parameters of benzyne isomers. TD$Δ$8SCF predicts a monocyclic minimum for singlet m-benzyne for all functionals, unlike SF-B3LYP, which stabilizes an erroneous bicyclic structure.

Bond Dissociation: HF and F₂

The ability to recover the correct dissociation limit is a stringent test of a linear-response formalism. In restricted DFT, the dissociation limit is never achieved due to failure to describe correct near-degeneracy and static correlation. SF-TDDFT with conventional functionals produces strong functional dependence, unphysical ionic solutions (e.g., in HF), and spurious low-lying states (e.g., in F₂). It achieves proper dissociation only with high HF-exchange content and only for some bonds. TD$Δ$9SCF, in stark contrast, approaches the correct singlet-triplet degeneracy limit smoothly and without generating spurious roots, for all tested functionals (Figures 5 and 6). This robustness is achieved because the response is computed on an appropriate broken-symmetry reference.

Figure 5: Bond dissociation PES for hydrogen fluoride. TD$Δ$0SCF (all functionals) properly approaches the unrestricted dissociation limit, free from the ionic root instability present in SF-TDDFT (BLYP/B3LYP).

Figure 6: Bond dissociation PES for F₂. Only TD$Δ$1SCF (all functionals) shows physically reasonable asymptotic behavior without contamination by spurious low-lying roots as seen in SF-TDDFT.

Identification of a Numerical Instability in Non-Aufbau SCF Excitations

A key finding is the first identification and mechanistic analysis of a numerical instability in non-Aufbau SCF and TD$Δ$2SCF states—arising from divergence in the GGA exchange-correlation potential at points where the reduced gradient $Δ$3 becomes large on nodal manifolds of singly-occupied antibonding orbitals. In H₂ with a pure $Δ$4 open-shell configuration, the orbital energies of the antibonding orbital display sharp, grid-dependent discontinuities linked to near-zero density and finite density gradient at nodal points (Figures 7–9). In practical systems, this scenario is rare, but the effect is intrinsic, not merely a numerical artifact, and could be exacerbated in large or strongly correlated systems with pronounced unpaired character at particular spatial regions.

Figure 7: Non-Aufbau SCF orbital energies for H₂, showing sharp peaks in antibonding orbital energy, a direct manifestation of the $Δ$5 grid instability.

Figure 8: Scatter plots of $Δ$6 showing divergence at specific grid points with vanishing density and large $Δ$7.

Figure 9: Real-space visualization of $Δ$8 for H₂; pathology (diverging gradient) is localized at the antibonding nodal plane.

Implications, Limitations, and Theoretical Perspective

TD$Δ$9SCF establishes itself as a robust, low-cost formalism for challenging single-reference-inadequate problems, offering clearly superior functional insensitivity compared to collinear SF-TDDFT, as well as improved qualitative description for torsion, S–T gaps (especially in "same-center" biradicals), and bond dissociation. Its limitations are noteworthy: the systematic overestimation of singlet energies (hence S–T gaps), performance susceptible to the quality of the underlying non-Aufbau SCF reference, and potential for physically relevant, though rarely encountered, numerical instability when the open-shell density is highly nodal and not compensated by other electrons.

By not requiring explicit construction of noncollinear kernels or the treatment of spin-flip operators, TD$Δ$0SCF remains straightforward to implement in existing DFT/TDDFT architectures provided robust $Δ$1SCF optimization protocols (e.g., MOM/STEP/SGM) are available. This also implies that future developments—including automatic diagnosis of the quality of non-Aufbau references, regularization of $Δ$2 near nodal manifolds, and extension to analytic properties and nonadiabatic couplings—are feasible with minimal infrastructural changes. The approach complements ongoing efforts to generalize TDDFT linear-response or EOM-CC paradigms for strongly correlated and open-shell systems.

Conclusion

TD$Δ$3SCF defines a new linear-response framework for static correlation and near-degeneracy regimes, eliminating the unphysical functional dependence of SF-TDDFT and delivering accurate PESs, spectroscopic properties, and molecular geometries for problems where single-reference DFT fails. The method exposes and clarifies an intrinsic instability in open-shell DFT for extreme cases, an insight with ramifications for broader developments in density-based excited-state methods. With improvements in reference-state algorithms and regularization strategies for the exchange-correlation potential, TD$Δ$4SCF is poised for widespread applicability in quantum chemistry and materials science for the description of bond-breaking, conical intersections, and electronically nontrivial molecules.