Vibrational Configuration Interaction (VCI)

• Vibrational configuration interaction (VCI) is a variational method that expands the vibrational wavefunction in product bases to recover intermode correlations and anharmonic effects.
• VCI workflow involves constructing an anharmonic potential energy surface, solving a VSCF problem, and diagonalizing the Hamiltonian in a truncated excitation space.
• Variants such as L-VCI, DVCI, and VHCI optimize convergence, reduce computational cost, and enhance applications in spectroscopic and tunneling studies.

Searching arXiv for recent and foundational papers on vibrational configuration interaction to ground the article. VCI, in the molecular literature cited here, denotes vibrational configuration interaction: a variational method in which the vibrational wavefunction is expanded in a basis of product configurations and the vibrational Hamiltonian is diagonalized to obtain anharmonic vibrational energies and wavefunctions. In practice, VCI is typically the post-VSCF correlation step, but the same variational structure also appears in localized-mode formulations, selected-CI variants, hybrid tunneling schemes, cavity vibrational calculations, and DFT-based anharmonic solvers. A distinct nuclear-structure usage of the same acronym denotes variational configuration interaction, built from projected variational pair states rather than vibrational modals (Panek et al., 2016, Moitra et al., 2020, Eraković et al., 2022, Yu, 2022, Chasman et al., 2010).

1. Variational definition and relation to VSCF

In standard molecular usage, VCI is the vibrational analogue of electronic configuration interaction. The reference state is usually a vibrational self-consistent field (VSCF) product state, and VCI recovers the intermode correlation that remains beyond that mean-field description. The basic expansion is written as

$$|\Psi_{\lambda}\rangle = \sum_I C_I^{(\lambda)} |\Phi_I^{\mathrm{VSCF}}\rangle,$$

or, in localized-mode notation,

$$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$

with coefficients obtained variationally from Hamiltonian diagonalization (Moitra et al., 2020, Panek et al., 2016).

The practical workflow is correspondingly sequential. One constructs an anharmonic potential energy surface (PES), solves the VSCF or L-VSCF problem, builds the VCI Hamiltonian in the resulting modal basis, and diagonalizes that Hamiltonian in a chosen excitation space to obtain anharmonic vibrational energies. In benchmark-quality calculations, the final VCI eigenvectors are then used for state assignment, identification of overtone and combination-band character, and diagnosis of Fermi resonances through the dominant CI coefficients (Nandi et al., 31 Jan 2025).

Within this framework, full VCI means variational diagonalization in a large product basis

$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$

where the configuration space $\mathcal V$ is specified by an excitation or pruning rule. This variational foundation remains intact in selected or adaptive variants; what changes is the organization of the basis and the Hamiltonian action, not the underlying physical theory (Tran et al., 2023, Garnier, 2018).

2. Hamiltonians, coordinates, and modal bases

The vibrational Hamiltonian used in many VCI calculations is the Watson Hamiltonian. In simplified L-VSCF/L-VCI work it is written as

$$\hat H=-\frac12\sum_i^M \frac{\partial^2}{\partial q_i^2}+V(\boldsymbol q),$$

whereas full-dimensional cavity calculations use the exact Watson Hamiltonian in mass-scaled rectilinear normal modes and then enlarge the problem by adding cavity-photon coordinates through the Pauli–Fierz Hamiltonian (Panek et al., 2016, Yu, 2022).

The PES entering VCI is commonly represented as an $n$-mode expansion,

$$V(\boldsymbol{q})= \sum_i^M V_i^{(1)}(q_i) +\sum_{i<j}^M V_{ij}^{(2)}(q_i,q_j) +\sum_{i<j<k}^M V_{ijk}^{(3)}(q_i,q_j,q_k) +\cdots,$$

or, equivalently in Franck–Condon work,

$$V(\mathbf{q}) = \bar V^0 + \sum_{i=1}^M \bar V^{m_i} + \sum_{j>i=1}^M \bar V^{m_i m_j} + \cdots .$$

A central technical point is that a truncated $n$-mode PES is not invariant under coordinate transformation, so separate truncated PESs in normal-mode and localized-mode coordinates are not equivalent (Panek et al., 2016, Moitra et al., 2020).

The choice of coordinates strongly affects convergence. Localized modes are obtained from normal modes by a unitary transformation,

$$\tilde{\boldsymbol Q}=\boldsymbol Q\,\boldsymbol U,$$

chosen to maximize a localization measure. In this basis, the Hessian is no longer diagonal, but the coordinates become chemically interpretable and the mutual coupling between modes is often reduced (Panek et al., 2016).

Traditional VCI uses harmonic-oscillator eigenfunctions as one-mode basis functions. A recent DFT-oriented formulation keeps the same variational idea but replaces the usual harmonic basis with Gaussian-type polynomial functions,

$$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$0

which span the same space as the wave functions of the quantum harmonic oscillator. Because this basis is non-orthogonal, the resulting problem is a generalized eigenvalue problem,

$$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$1

with analytic overlap and kinetic-energy matrix elements and anharmonic potential terms evaluated by Gauss–Hermite quadrature (Withanage et al., 3 Sep 2025).

3. Excitation spaces, convergence, and scaling

VCI accuracy is controlled by two coupled truncations: the representation of the PES and the size of the configuration space. Common excitation-space notations include VCI[$$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$2], where $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$3 is the maximum number of simultaneously excited modes, and VCI-S, VCI-SD, VCI-SDT, and higher ranks, which restrict the number of simultaneously excited modes. Equivalent cutoffs can also be formulated through $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$4 and $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$5, limiting per-mode and total quanta (Moitra et al., 2020, Panek et al., 2016).

These truncations determine both the attainable accuracy and the computational burden. The basis dimension grows rapidly with the number of modes and basis functions per mode, and full VCI is therefore expensive even for relatively small molecules. The cited literature emphasizes the same bottlenecks repeatedly: the construction of the PES, the growth of the excitation space, and the resulting cost of building and diagonalizing the VCI Hamiltonian (Panek et al., 2016, Tran et al., 2023).

The dependence on truncation is application-specific. In vibrationally resolved X-ray absorption, VCI[1] is exact within a 1M PES but cannot describe combination bands or multi-mode correlation, whereas VCI[4] introduces overtone and combination-band structure such as $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$6, $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$7, and $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$8 in formaldehyde and acetaldehyde (Moitra et al., 2020). In single-well tunneling calculations, the practical local-state calculation can be kept manageable by a VCISD-type truncation with the highest excitation in each mode limited to 6, precisely because the basis is centered on a single minimum rather than spanning multiple wells simultaneously (Eraković et al., 2022).

Localized-mode coordinates modify this convergence pattern in a systematic way. For ethene and furan, localized modes reduce average intermode coupling, accelerate convergence of the $$\Psi_i^{\rm L\text{-}VCI}(\tilde{\boldsymbol q}) =\sum_{I}^{N_{\text{states}}} c_I^{(i)} \psi_{\boldsymbol n_I}^0(\tilde{\boldsymbol q}),$$9-mode PES, and make the convergence of the VCI excitation expansion smoother, with significantly smaller low-order errors. In the cited ethene example, localized-mode PESs with up to three-mode terms achieve roughly the same quality as normal-mode PESs with up to four-mode terms (Panek et al., 2016).

4. Algorithmic reformulations and extensions

Several major developments keep the VCI variational structure but reorganize basis construction, Hamiltonian application, or coupling evaluation.

Variant	Defining feature	Role in the calculation
L-VCI	localized-mode coordinates	smoother PES and excitation-space convergence
DVCI	dual/local-force-field factorization	targeted eigenpairs without storing the full $$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$0 block
VHCI	heat-bath selection with semistochastic PT2	pruned approximation to full VCI
VSCF/VCI + instanton	single-well VCI local states plus semiclassical inter-well couplings	tunneling spectra in multiple wells

Localized-mode VCI retains the standard VSCF/VCI sequence but changes the coordinate system before PES construction and modal optimization. The resulting reduction in mode coupling is the reason localized coordinates can support low-cost hybrid PES models and better low-order VCI approximations (Panek et al., 2016).

Dual vibration configuration interaction (DVCI) is an iterative, residual-driven, factorized reformulation of VCI. Its central idea is a dual Hamiltonian representation,

$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$1

which allows the residual

$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$2

to be evaluated without constructing the large secondary Hamiltonian block explicitly. The paper reports memory reduction by more than a factor of 15 while retaining sub-$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$3 accuracy (Garnier, 2018).

Vibrational heat-bath configuration interaction (VHCI) places VCI in the selected-CI family. Configurations are added if

$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$4

and the semistochastic PT2 correction removes the memory bottleneck of the perturbative space. This method is explicitly presented as a selected, high-accuracy approximate VCI. The same paper also formulates VHCI with VSCF modals rather than harmonic-oscillator modals, but finds that the minor improvements to accuracy are outweighed by the much higher computational cost associated with the loss of sparsity (Tran et al., 2023).

A different extension appears in multiwell tunneling theory. There, VSCF/VCI is not used as a stand-alone global solver; instead it supplies the single-well vibrational energies and wavefunctions that define the diagonal blocks of a larger tunneling matrix, while the off-diagonal couplings are supplied semiclassically through instanton/JFI wavefunctions and a generalized Herring formula (Eraković et al., 2022).

5. Spectroscopic, tunneling, and cavity applications

VCI is widely used as the vibrational engine that turns anharmonic PESs into observable spectra. In vibrationally resolved K-edge X-ray absorption, ground- and core-excited-state PESs are built from CVS-CC2, CVS-CCSD, CVS-CCSDR(3), or CVS-CC3 calculations selected by ADGA, VCI is solved on each surface, and Franck–Condon factors are computed from overlaps of the resulting anharmonic vibrational wavefunctions,

$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$5

This makes VCI not merely a postprocessing step but a direct probe of PES quality, because the vibrational progression depends sensitively on the shape of the excited-state surface (Moitra et al., 2020).

In multiwell tunneling problems, the same variational machinery supplies accurate local vibrational states that can then be coupled across barriers. The combined VSCF/VCI–instanton method was benchmarked on a two-dimensional double well, malonaldehyde, and asymmetrically deuterated malonaldehyde. Its key result is that harmonic local energies may suffice in symmetric cases because local errors cancel, but once the wells are asymmetric or near resonance, accurate VCI local energies become essential for correct splittings, avoided crossings, and state mixing (Eraković et al., 2022).

VCI is also used as a state-analysis tool in full-dimensional molecular spectroscopy. For ethylene glycol, VSCF/VCI on a full-dimensional machine-learned PES shows that the OH-stretch region is dominated by nearly pure anharmonic OH fundamentals, whereas the CH-stretch region is a strongly resonant manifold with substantial Fermi resonance and combination-band character. In the global minimum conformer $$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$6, the OH-stretch states appear at 3629 and 3681 cm$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$7, compared with harmonic values 3831 and 3871 cm$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$8, while the lower CH band at 2797–2798 cm$$|\Psi\rangle = \sum_{I\in \mathcal V} c_I |I\rangle,$$9 is identified as a strongly mixed state dominated by $\mathcal V$0 rather than a pure CH stretch (Nandi et al., 31 Jan 2025).

For floppy complexes, VCI is used to extract spectra from multidimensional high-level PESs in situations where harmonic analysis is inadequate. In CH$\mathcal V$1–Rg complexes, VCI based on CCSD(T)-F12a PESs shows especially large anharmonicity and zero-point vibrational energy effects for CH$\mathcal V$2–He; the paper states that the anharmonic correction for the intermolecular bending mode can be nearly 50% for He and less than 3% for Kr. The same work introduces configuration-averaged VSCF (CAVSCF) to avoid symmetry-broken modals for degenerate states before the final VCI diagonalization (Meisner et al., 2020).

Under vibrational strong and ultrastrong coupling, cavity VSCF/VCI extends the modal product basis to a joint molecule-plus-photon space. For water, using a 4-mode representation for the effective potential and a 3-mode representation for the dipole moment, the field-free cVSCF/VCI spectrum reproduces bending near 1594 cm$\mathcal V$3, symmetric stretch near 3656 cm$\mathcal V$4, and asymmetric stretch near 3756 cm$\mathcal V$5. Inside the cavity, the VCI coefficients resolve Rabi splitting, polarization-selective coupling, and cavity-induced mixing between molecular modes (Yu, 2022).

6. Limits, accuracy, and terminological scope

VCI is variational and systematically improvable, but its reliability is bounded by the quality of the PES and by the truncations imposed on both the PES representation and the excitation space. This is explicit in vibrationally resolved XAS, where CC2, CCSD, CCSDR(3), and CC3 can generate noticeably different core-excited PESs, and in some cases CC2 and CC3 even predict dissociative core-excited surfaces that make Franck–Condon analysis unreliable (Moitra et al., 2020).

A related practical limit is that a larger one-mode basis is not always better. In the combined VSCF/VCI–instanton framework, the authors note that for a fitted $\mathcal V$6-mode potential too many harmonic-oscillator functions can probe unphysical regions and produce intruder states. This underscores that VCI convergence is governed not only by basis size but also by the fidelity of the underlying local PES representation (Eraković et al., 2022).

Selected and factorized variants address feasibility rather than changing the underlying variational idea. DVCI, VHCI, localized-mode L-VCI, and the DFT-based Gaussian-polynomial formulation all preserve the core VCI logic: define a configuration basis, evaluate Hamiltonian matrix elements, and diagonalize. This suggests a stable conceptual core across quite different implementations, with the main differences lying in coordinate choice, basis representation, selection rules, and treatment of couplings (Garnier, 2018, Tran et al., 2023, Panek et al., 2016, Withanage et al., 3 Sep 2025).

Finally, the acronym itself is not unique across subfields. In the molecular literature assembled here, VCI means vibrational configuration interaction. In the nuclear-structure paper “Pair-vibrational states in the presence of neutron-proton pairing,” however, VCI means variational configuration interaction, an approximate many-body diagonalization method built from projected intrinsic pair-condensate states,

$\mathcal V$7

followed by diagonalization in a nonorthogonal basis of optimized configurations. The shared acronym therefore denotes analogous variational-CI logic, but not the same degrees of freedom, Hamiltonians, or physical observables (Chasman et al., 2010).