CIPSI-driven CC(P;Q) Method

• The paper introduces a hybrid quantum chemistry approach that integrates CIPSI selection with coupled-cluster moment expansion to capture essential higher excitation effects for near-CCSDT accuracy.
• The methodology partitions the Hilbert space into a compact P-space for nonperturbative treatment and a Q-space for moment-based corrections, enhancing precision.
• The approach ensures dramatic cost reductions (up to 80× speedup) and robust error control, making it effective for strongly correlated and multistate quantum chemical problems.

The CIPSI-driven CC($P$;$Q$) methodology is a hybrid quantum chemistry approach that merges the coupled-cluster ($\mathrm{CC}$) moment expansion formalism with the deterministic selected configuration interaction method known as CIPSI (Configuration Interaction by Perturbatively Selecting Iteratively). Its primary objective is to achieve CCSDT- (and higher-) quality electronic energies with significantly reduced computational cost, by efficiently capturing the most important higher-than-doubly excited determinants required for an accurate description of correlation effects, especially in strongly correlated or multireference regimes (Gururangan et al., 2021, Gururangan et al., 2023, Priyadarsini et al., 2024, Priyadarsini et al., 17 Jan 2026).

1. Theoretical Basis: Partitioned Coupled-Cluster Formalism

The CC($P$;$Q$) scheme partitions the $N$-electron Hilbert space into two complementary subspaces relative to a reference Slater determinant $|\Phi_0\rangle$ (typically Hartree–Fock):

• $\mathcal{H}^{(P)}$: The "model" or $P$-subspace, which includes $|\Phi_0\rangle$ and a selected set of excited determinants $\{|\Phi_\mu\rangle\}$ crucial for nonperturbative treatment.
• $\mathcal{H}^{(Q)}$: The $Q$-subspace, defined as the remaining portion of Hilbert space, $\mathcal{H} \setminus \mathcal{H}^{(P)}$.

Within this partitioning, the cluster operator is restricted to $P$,

$$T^{(P)} = \sum_{|\Phi_\mu\rangle\in\mathcal{H}^{(P)}} t_\mu E_\mu$$

and the CC equations are solved in $P$ using biorthogonal projections to yield amplitudes $t_\mu$ and energy $E^{(P)}$.

Residual correlation from the $Q$-space is recovered through a noniterative moment-based correction,

$$\Delta E^{(P;Q)} = \sum_{\nu\in Q} \ell_\nu(P)\, \mathcal{M}_\nu(P),$$

where $$\mathcal{M}_\nu(P) = \langle\Phi_\nu| \overline{H}^{(P)} |\Phi_0\rangle$$ are the $Q$-space moments and $\overline{H}^{(P)} = e^{-T^{(P)}} H e^{T^{(P)}}$. The left weights $\ell_\nu(P)$ involve the biorthogonal de-excitation operator $\Lambda^{(P)}$. The final energy is $E^{(P+Q)} = E^{(P)} + \Delta E^{(P;Q)}$ (Gururangan et al., 2021, Gururangan et al., 2023, Priyadarsini et al., 2024).

This framework is state-specific and rigorously size-extensive for arbitrary selection of $P$.

2. CIPSI Determinant Selection Algorithm

CIPSI is used to construct a compact set of determinants whose inclusion in $P$ captures the leading non-dynamical correlation effects:

1. Iterative Expansion: Starting from a minimal reference space ($V_{\text{int}}^{(0)}$), a CI wave function $$|\Psi_k\rangle = \sum_{I\in V_{\text{int}}^{(k)}} c_I |\Phi_I\rangle$$ is optimized by diagonalizing $H$ in $V_{\text{int}}^{(k)}$.
2. External Screening: All external single and double excitations $|\Phi_\alpha\rangle$ are generated, and their importance is ranked via the Epstein–Nesbet second-order energy contribution,

$$e_{\alpha,k}^{(2)} = \frac{|\langle\Phi_\alpha| H | \Psi_k\rangle|^2}{E_{\text{var},k} - \langle\Phi_\alpha| H |\Phi_\alpha\rangle}.$$

1. Determinant Addition: Determinants with the largest $|e_{\alpha,k}^{(2)}|$ are added to the internal space, and the process repeats until a stopping criterion is met (e.g., maximum $N_{\text{det}}$ or perturbative correction $<\eta$).
2. Final $P$-Space: $P$ is defined as all singles and doubles, union with the selected triples (and quadruples for CCSDTQ) present in the final CIPSI list. Remaining triples and higher go into $Q$ (Gururangan et al., 2021, Gururangan et al., 2023, Priyadarsini et al., 2024).

This process systematically focuses the iterative treatment on those high-level excitations most essential for electronic structure, as determined by their perturbative contributions.

3. Integration of CIPSI and CC($P$;$Q$) and Algorithmic Workflow

The methodology proceeds as follows:

1. Run CIPSI to produce a compact $V_\text{int}^*$, identifying key triple (and quadruple) excitations.
2. Form $P$-space as all singles/doubles plus the selected higher excitations from $V_\text{int}^*$.
3. Solve CC($P$) for amplitudes $t_\mu$ and left operator $\Lambda^{(P)}$.
4. Build $Q$-space as the complementary high-rank excitations.
5. Evaluate moments and corrections: Compute $\{\mathcal{M}_\nu(P), \ell_\nu(P)\}$ over $Q$ to obtain the moment expansion correction $\Delta E^{(P;Q)}$.
6. Iterate if unconverged: Increase $N_{\text{det}}$ and repeat until $|E^{(P+Q)}_\text{new} - E^{(P+Q)}_\text{old}| <$ threshold (e.g., $10^{-6} E_h$) or the fraction of leading triple amplitudes in $P$ exceeds a chosen fraction (e.g., 90%) (Gururangan et al., 2021, Priyadarsini et al., 2024).

This workflow allows rapid and black-box convergence to CCSDT/CCSDTQ energetics with only 1–3% of the high-rank excitation manifold treated nonperturbatively.

Step	Operation	Typical Scaling/Cost
CIPSI selection	Diagonalizations in $10^4$–$10^6$ det.	$O(N_s^3)$
CC(P) amplitude solve	Cluster amplitudes in compact $P$	$O(N^6) + O(n_t N^4)$
Q-space correction	Evaluation over $Q$	$O(n_Q N^6)$ ($n_Q$ small)

4. Extensions to Excited States: EOM-CC($P$;$Q$)

The CC($P$;$Q$) method extends naturally to excited state energetics via the equation-of-motion (EOM) CC formalism. For excited state $\mu$,

• EOM-CC($P$): The ground-state cluster operator $T^{(P)}$ is used, and the right ($R_\mu$) and left ($L_\mu$) excitation operators are also restricted to $P$.
• The noniterative Q-space correction addresses the missing triple (and quadruple) excitation relaxation for excited states analogously to the ground state.

The excited-state correction takes the form

$$\delta_\mu(P;Q) = \sum_{L\in Q} \ell_{\mu,L}(P) M_{\mu,L}(P),$$

where $M_{\mu,L}(P)$ and $\ell_{\mu,L}(P)$ are the appropriate EOM-CC moments and left weights (Priyadarsini et al., 17 Jan 2026).

Benchmark calculations on systems with multireference character (e.g., CH$^+$, CH, H$_2$O, cyclobutadiene) demonstrate that EOM-CC($P$;$Q$) can recover the CCSDT/EOMCCSDT energetics to within 0.1–1 m$E_h$ for both ground and excited states by noniteratively including a carefully selected fraction (often $\sim$1–4%) of the triple excitation manifold.

5. Computational Performance, Accuracy, and Error Control

The methodology yields dramatic reductions in cost relative to full CCSDT:

• On cyclobutadiene, CC($P$;$Q$) with $1\%$ of triples provides $80\times$ speedup over CCSDT, maintaining $<0.1$ m$E_h$ errors (Gururangan et al., 2023, Priyadarsini et al., 2024).
• For F$_2$ dissociation and automerization barriers, recovery of CCSDT energetics is achieved with only $0.5$–$2\%$ of triples in $P$, and the method remains effective as nondynamical correlation grows.
• In excited-state applications, inclusion of selected triples via CIPSI rapidly reduces principal errors, especially for states dominated by high-rank correlation.

Convergence and residual error are monitored through:

• Stability of $E^{(P+Q)}$ (e.g., changes $<10^{-6} E_h$).
• Monitoring the fraction of leading triple amplitudes in $P$ and magnitude of Q-space residuals $\max|\mathcal{M}_\nu(P)|$ or $\sum|\mathcal{M}_\nu|^2$.
• Comparison to extrapolated variational plus perturbatively corrected CIPSI energies (Gururangan et al., 2021, Priyadarsini et al., 2024).

6. Strengths, Applications, and Limitations

Key strengths:

• Deterministic, black-box, and self-improving approach—no user assignment of active orbitals required.
• Rapid convergence even in strongly correlated or multireference situations where perturbative triples corrections (CCSD(T), CR-CC(2,3)) fail.
• Size-extensive, error-controlled, and highly parallelizable workflow.
• Orders-of-magnitude CPU and memory savings relative to full CCSDT, both for ground and excited states.

Applications:

• Accurate potential energy surfaces and barriers (e.g., cyclobutadiene automerization, F$_2$ dissociation).
• Precise singlet–triplet gap calculations in biradicals, with sub-kcal/mol accuracy for closely lying states (Priyadarsini et al., 2024).
• Excited electronic states, including challenging multireference cases such as stretched bonds and near-degeneracies (Priyadarsini et al., 17 Jan 2026).

Limitations and ongoing work:

• Current implementations focus on triples (targeting CCSDT); extension to quadruples (CCSDTQ) and systematic validation on larger molecular sets is in progress.
• Determinant selection algorithms and thresholds ($N_{\text{det}}$, $f$, $\eta$) can influence efficiency and automation; further optimizations are underway.
• Moment corrections employ two-body Hamiltonian approximations in practice; potential improvements include higher-order treatments (Gururangan et al., 2023).

7. Comparison to Related Methodologies

• Compared to Quantum Monte Carlo–driven CC($P$;$Q$), the CIPSI-driven version is fully deterministic and exhibits more rapid, stable convergence when a compact CI wave function is possible (Gururangan et al., 2021).
• Unlike CR-CC(2,3), which uses all triples perturbatively, CIPSI-driven CC($P$;$Q$) iteratively incorporates key triple amplitudes, improving relaxation of $T_1$ and $T_2$ clusters and reducing size-consistency and nonparallelity errors.
• Adaptive, moment-driven selection (as in (Gururangan et al., 2023)) offers additional automation and error control by ranking determinants by their individual energy contributions $\delta_K = \ell_K(P) M_K(P)$, providing a flexible path toward chemical accuracy with minimal resources.

---

The CIPSI-driven CC($P$;$Q$) methodology thus provides a powerful, scalable, and systematically improvable route to high-level coupled-cluster energetics, extending its reliability to strongly correlated and multistate quantum chemical problems at a fraction of traditional computational expense (Gururangan et al., 2021, Gururangan et al., 2023, Priyadarsini et al., 2024, Priyadarsini et al., 17 Jan 2026).