Configuration Interaction and Cluster Models

• Configuration Interaction and Cluster Models are complementary quantum many-body methods that use linear and exponential wavefunction expansions to capture electron correlation.
• They blend CI’s exact treatment of valence correlations with cluster models’ efficient all-order summation of core and core–valence effects for improved accuracy.
• Hybrid CI+CC frameworks enable systematic improvements in treating static and dynamic correlations, offering sub-percent accuracy for atomic and molecular properties.

Configuration interaction (CI) and cluster models form two foundational and complementary strategies for the non-perturbative treatment of electronic correlation in atoms, molecules, and extended systems. Contemporary electronic structure theory has witnessed a proliferation of hybrid and cluster-based models, especially for systems with strong static and dynamic correlation, multireference character, or relativistic effects. This article catalogs the principles, mathematical formulations, algorithmic approaches, and roles of configuration interaction and cluster models—both individually and in synergy—across key classes of quantum many-body methods.

1. Theoretical Foundations: CI, Cluster Expansion, and their Mapping

Configuration interaction methods employ a linear expansion of the correlated many-fermion wave function in a basis of Slater determinants, while coupled-cluster (CC) and more general cluster models employ an exponential (or cluster-operator) ansatz to systematically resum classes of correlation diagrams.

• CI expansion:

$$|\Psi_{\text{CI}}\rangle = \sum_I c_I |\Phi_I\rangle$$

where $|\Phi_I\rangle$ runs over determinants (or configuration state functions), and $c_I$ are variational coefficients.

• Cluster (CC) expansion:

$$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$

where $T_n$ are n-body excitation operators acting on the reference $|\Phi_0\rangle$.

These ansätze are formally equivalent in the FCI (full configuration interaction, or full-rank) limit, related via the cluster decomposition. The explicit mapping between CI and CC amplitudes involves recursively subtracting all disconnected contributions from the CI coefficients, a process that rapidly becomes combinatorially expensive at high excitation rank (Lehtola et al., 2017). The hierarchy of connected cluster amplitudes, $T_k$, encodes irreducible $k$-body correlations; the rapidity of their decay with $k$ directly indicates the feasibility of CC truncation for a given system.

2. Hybrid CI-Cluster (CI+CC) Models: Effective Hamiltonians and All-Order Techniques

Hybrid models combine the respective strengths of CI and CC: the former's exact treatment of valence–valence correlation, and the latter's all-order resummation of core and core–valence diagrams. This structure is explicit in the effective Hamiltonian approaches for atomic (and molecular) systems with a small number of valence electrons (Porsev et al., 2012, Dzuba, 2014, Liu et al., 3 Feb 2025).

• The method begins with a Dirac–Fock or Hartree–Fock reference and partitions the Hamiltonian into a frozen core and active valence space.
• Core and core–valence correlation effects are summed to all orders via linearized CCSD (or higher) cluster amplitudes, yielding an energy-dependent effective Hamiltonian $H_{\text{eff}}(E) = H_{\text{core}} + \Sigma(E)$.
• Valence–valence correlations are solved exactly by configuration interaction—in practice, a large-scale CI diagonalization in the effective Hamiltonian.
• This approach yields sub-percent accuracy for level energies, response properties such as polarizabilities, and transition matrix elements in divalent and trivalent atoms/ions, e.g., Mg, Si$|\Phi_I\rangle$0, Ba, Lu, and highly charged ions such as Ir$|\Phi_I\rangle$1 (Porsev et al., 2012, Dzuba, 2014, Liu et al., 3 Feb 2025).

CI+CC/CI+all-order frameworks are systematically improvable by enlarging the CI space or including higher excitations in CC; error budgets are predominantly determined by the omission of higher-order (e.g., triple) cluster corrections (Porsev et al., 2012, Dzuba, 2014).

3. Cluster Analysis, Cluster-Decomposed CI, and Tailoring Cluster Models

Cluster analysis—extracting cluster amplitudes from CI wave functions or selected CI spaces—forms the basis for several modern externally corrected or tailored coupled cluster (TCC) methods. These methods recognize that the dominant missing pieces in truncated CC expansions for strongly correlated systems often reside in higher excitation ranks, especially in multi-reference regimes (Aroeira et al., 2020, Gururangan et al., 2021, Erhart et al., 20 Jun 2025).

• Cluster decomposition:

The CI coefficients $|\Phi_I\rangle$2, $|\Phi_I\rangle$3, ... are recursively mapped to cluster amplitudes $|\Phi_I\rangle$4, $|\Phi_I\rangle$5, ... iteratively subtracting lower-rank disconnected contributions (Lehtola et al., 2017, Aroeira et al., 2020).

• Externally corrected CC (EC-CC, TCC):

Higher-order cluster amplitudes (e.g., triples $|\Phi_I\rangle$6 and quadruples $|\Phi_I\rangle$7) are extracted from selected CI, ACI, or QSCI wavefunctions and injected as frozen amplitudes into the CC equations for the lower ranks, e.g.,

$|\Phi_I\rangle$8

$|\Phi_I\rangle$9 fixed during the iterations (Aroeira et al., 2020, Erhart et al., 20 Jun 2025).

• Quantum-classical tailored CC (QSCI-TCC):

Determinant amplitudes from quantum-selected CI (QSCI), efficiently sampled on quantum hardware, are mapped onto the cluster operator in the active space; the remaining amplitudes are optimized classically (Erhart et al., 20 Jun 2025, Vaquero-Sabater et al., 22 May 2026). This construction maintains polynomial cost for dynamical correlation while embedding strong static correlation from the quantum device.

These strategies support balanced descriptions of strong and weak correlation, are robust across the dissociation of multibonded systems (H$c_I$0O, N$c_I$1), and remain accurate when conventional CCSD or CCSD(T) fails (Erhart et al., 20 Jun 2025, Gururangan et al., 2021).

4. Selected and Clustered CI Methods: Truncation, Screening, and Locality

Selected configuration interaction (SCI) and cluster-based CI models tackle the combinatorial explosion of determinant spaces by exploiting physical structure (locality, entropy, sparsity) and adaptive selection (Abraham et al., 2020, Zgid et al., 2012, Damour et al., 2024).

• Selected CI (SCI, e.g. CIPSI, ASCI, HCI):

Iteratively builds the determinant space by adding those configurations with the largest estimated perturbative energy contributions (Epstein-Nesbet, PT2), often using thresholds on the marginal gain per iteration (Gururangan et al., 2021, Damour et al., 2024). This adaptive process converges rapidly to FCI-quality solutions for moderate systems by focusing effort where it matters most.

• Cluster-based SCI (TPSCI Editor's term):

Orbitals are grouped into clusters, and the CI expansion is built in the tensor-product basis of many-body cluster states. A compressed, high-rank local state basis (e.g., from Tucker/SVD decomposition) allows TPSCI to treat large and strongly correlated systems—such as bond breaking or conjugated $c_I$2 systems—with far fewer parameters than Slater-determinant SCI (Abraham et al., 2020).

• SCI for open systems and resonances:

SCI has been extended to treat electronic resonances by formulating the selection and extrapolation in the presence of a non-Hermitian (complex absorbing potential, CAP) Hamiltonian. The CAP-SCI protocol converges both the real (position) and imaginary (width) parts of resonance energies, exposing the importance of high-order correlation even for unbound states (Damour et al., 2024).

In the context of dynamical mean-field theory (DMFT) and quantum impurity solvers, truncated CI serves as a controlled, polynomial-cost alternative to ED, enabling systematic and efficient solvers for impurity and bath models with large active spaces (Zgid et al., 2012).

5. Pair and Cluster-Based Models: pCCD, DOCI, and Their Interplay

Pair Coupled Cluster Doubles (pCCD) and Doubly Occupied Configuration Interaction (DOCI) exploit the "seniority-zero" subspace, wherein all electrons are paired and static correlation is well captured by geminal-based wave functions (Shepherd et al., 2016, Nowak et al., 2022).

• DOCI:

$c_I$3

where $c_I$4 are all determinants with completely paired occupations.

• pCCD:

$c_I$5

with $c_I$6 denoting pair creation/annihilation operators.

Numerical evidence demonstrates energy equivalence of pCCD and DOCI across large Hubbard lattices, attributable to the exponential compressibility of the seniority-zero subspace. FCIQMC or selected CI restricted to seniority zero stochastically generalizes DOCI to larger systems (Shepherd et al., 2016). pCCD can be augmented with post-pCCD CI or CC corrections (e.g., pCCD-CISD+RDC) to recover missing dynamical correlation, outperforming CCSD for ground-state potential curves and spectroscopic constants (Nowak et al., 2022).

6. Multireference and Stochastic CI–Cluster Approaches

Multireference coupled-cluster and hybrid stochastic methods extend CI–CC combinations to problems with pronounced multi-configurational character (Filip et al., 2023).

• Hybrid CI–CCMC:

The wave function is expressed as $c_I$7, where the CAS or strongly correlated reference is handled by CI, and dynamic correlation is captured by a stochastic (e.g., CCMC) sampling of external amplitudes.

• Projection equations:

Separate coupled conditions evolve the CI and CC amplitudes, typically via Monte Carlo propagation, maintaining formal size extensivity outside the (truncated) CAS.

• Scaling and accuracy:

Hybrid CI–CCMC methods dramatically reduce the stochastic walker cost and noise compared to full MRCCMC, provided the CAS is chosen to capture as much static correlation as feasible.

This approach allows for black-box, systematically improvable multireference correlation schemes that bridge the gap between adaptive CI and cluster Monte Carlo.

7. CI–Cluster Diagnostics, Analysis, and Physical Interpretation

The decomposition of CI wave functions into cluster amplitudes (Lehtola et al., 2017, Kristiansen et al., 2024), as well as the computation of configuration weights in CC models (Kristiansen et al., 2024), provides a direct route for analyzing correlation complexity and the limits of single-reference CC and CI truncations.

• Norm fall-off of connected amplitudes ($c_I$8):

Rapid decay of $c_I$9 with $$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$0 implies that truncation at double or triple excitations is sufficient—typical for weakly correlated or near-equilibrium molecules. Slow decay (high-rank amplitudes sizable) signals failure of single-reference CC or CI models, mandating multireference or adaptive FCI strategies (Lehtola et al., 2017).

• Configuration weights in CC:

Projection of the CC state onto determinants yields weights analogous to $$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$1 in CI. In the single-reference regime, these weights match FCI probabilities to within $$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$2. In strongly correlated or separated-fragment cases, CC weights can become pathological unless a fully exponential (extended CC) bra is used (Kristiansen et al., 2024).

• Interpretation for truncation and method selection:

Diagnostics based on the cluster decomposition guide practitioners in choosing the appropriate truncation level in CC or CI, and identify the necessity for multi-reference (or hybrid) treatments in transition metal, bond-dissociation, or other multireference cases.

8. Algorithmic and Computational Considerations

Modern CI and cluster approaches universally exploit adaptive truncation, tensor factorizations, locality, and stochastic algorithms for scalability. Representative scaling and accuracy comparisons:

Method	Cost scaling	Correlation regime	System sizes
CI (FCI)	exponential	static+dynamic	$$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$318e/18o
SCI	variable	static; dynamical (with PT2)	$$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$4
CCSD(T)	O($$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$5)	weak/medium dynamic	dozens–hundreds o
CI+CC (hybrid)	polynomial	static+dynamic	2–3 valence e or large active spaces
QSCI-TCC	O($$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$6)	static (QSCI), dynamic (CC)	large (quantum resource)
TPSCI	compressed, $$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$7	strong local correlation	42e/42o $$|\Psi_{\text{CC}}\rangle = e^{T}|\Phi_0\rangle, \quad T = T_1 + T_2 + T_3 + \cdots$$8-systems

CI solvers allow for systematic control via excitation level, iterative expansion, or adaptive space selection; clusters enable all-order summation of selected amplitudes; hybrid and QSCI techniques provide a route for quantum advantage and new approaches to strong correlation (Erhart et al., 20 Jun 2025, Vaquero-Sabater et al., 22 May 2026).

---

Configuration interaction and cluster models—including their hybridizations—now form a versatile, systematically improvable landscape for the quantitative and interpretive treatment of quantum many-body problems, with distinct advantages in flexibility, resource scaling, and physical insight over purely perturbative or diagonalization-based approaches. Their integration with quantum algorithms and adaptive selection marks a new era in electronic structure theory, capable of accurately and efficiently addressing both static and dynamical correlation in challenging and emergent systems.