Density-fitted ADC(2)

ADC2, IPADC2 and EAADC2 implement strict second-order ADC for neutral excitation, ionization and electron attachment from real RHF, ROHF and UHF references, including noncanonical occupied and virtual orbitals. It uses the same MethodBase and InitialGuessMixin contracts as EOMCC. Its CIS/CISd initial guesses operate on the bare DF Hamiltonian; no coupled-cluster amplitudes or HBar are computed.

from ccpy import Molecule, RHF, ADC2

mol = Molecule("O 0 0 0; H 0 0 .96; H .93 0 -.24", "6-31g",
               auxiliary_basis="def2-universal-jkfit")
adc = ADC2(RHF(mol), nfrozen=1)
adc.initial_guess(method="cisd", multiplicity=1, nacto=2, nactu=2,
                  roots_per_irrep={"A": 3}, roots=[0, 2])
totals = adc.energy()         # SCF + MP2 correlation + ADC excitation energy
excitation_energies = adc.omega
virtual_density = adc.virtual_density(root=0)  # {"a": Dvv_alpha, "b": Dvv_beta}

roots_per_irrep, roots, use_symmetry, nacto and nactu have the same meaning as in EOMCC. method="cis" is also supported. Closed-shell restricted references support singlet/triplet guesses; open-shell references can use multiplicity=-1 for unadapted guesses. Spin-adapted UHF guesses do not make a spin-contaminated UHF ADC Hamiltonian spin-pure.

Charged states use the existing bare-Hamiltonian guess API:

from ccpy import IPADC2, EAADC2

ip = IPADC2(mf, nfrozen=1, formulation="full")
ip.initial_guess(method="ipcisd", nacto=2, nactu=2,
                 roots_per_irrep={"A": 2})
ionized_totals = ip.energy()
ea = EAADC2(mf, nfrozen=1, formulation="full")
ea.initial_guess(method="eacisd", nacto=2, nactu=2,
                 roots_per_irrep={"A": 2})
attached_totals = ea.energy()

ipcis and eacis are also supported. Charged operators use the alpha removal/attachment branch of CCpy’s existing IP/EA API. Their default guess multiplicity is -1. The signed energy is omega = E(target)-E(parent): IP roots are ionization potentials, while positive electron affinities equal -ea.omega. Root indices and irrep selection follow the EOMCC interface.

Formulations and root targeting

All three methods accept formulation="folded" (default) or formulation="full". Folded mode targets states dominated by singles (EE), one hole (IP), or one particle (EA). Its default guesses are respectively cis, ipcis and eacis. Full mode defaults to cisd, ipcisd and eacisd and retains explicit doubles/satellite vectors for targeting states with strong higher-excitation character. Explicit guess selection remains available in either mode; a folded guess with no primary component is rejected.

adc = ADC2(mf)                          # folded, CIS guesses
satellites = IPADC2(mf, formulation="full")
satellites.initial_guess(method="ipcisd", nacto=3, nactu=3,
                         roots_per_irrep={"A": 5}, roots=[4])

Writing the full matrix as [[A,B],[B.T,D]], folded mode evaluates

\[R_2=(\omega I-D)^{-1}B^\dagger R_1,\qquad [A+B(\omega I-D)^{-1}B^\dagger]R_1=\omega R_1.\]

D is diagonal in the internal semicanonical basis. Each outer energy iteration solves the primary-space Hermitian problem with CCpy’s eomcc_davidson and uses a Newton energy update including the resolvent derivative. Only primary vectors enter the Davidson subspace; satellite work arrays are reconstructed during matrix actions. Final normalized full vectors are retained for densities and FNO root following. This is not an amplitude-direct ground-state implementation: MP2 doubles remain stored. Satellite poles raise an explicit error recommending full mode.

Full mode uses the expanded linear problem with sequential orthogonal root locking. Folded roots use overlap following without Euclidean primary-space locking, because different roots are orthogonal in the reconstructed full-space metric. Duplicate converged roots raise an error. Every final eigenpair is checked against the original expanded matrix action; unconverged states cannot supply FNO densities or corrections. davidson_solver="standard" is the supported mode.

Equations and storage

The implementation uses antisymmetrized spin-orbital integrals g_ijab = <ij||ab> and first-order MP2 amplitudes

\[t_{ij}^{ab} = -g_{ijab}/\Delta_{ij}^{ab}, \qquad \Delta_{ij}^{ab} = \epsilon_a+\epsilon_b-\epsilon_i-\epsilon_j.\]

The ADC secular matrix has the strict ADC(2) orders

\[\begin{split}M = \begin{pmatrix} M_{11}^{(0)}+M_{11}^{(1)}+M_{11}^{(2)} & M_{12}^{(1)}\\ M_{21}^{(1)} & M_{22}^{(0)} \end{pmatrix}.\end{split}\]

The singles action is evaluated through the CIS action, occupied/virtual self-energy intermediates, and symmetrized products of the MP2-amplitude and g_imae singles-space actions. The satellite block is diagonal in the internal semicanonical orbitals. Singles–doubles couplings are evaluated as staged contractions with the DF oo, ov and vv factors. The native implementation builds neither the secular matrix nor ovvv/vvvv integral tensors.

The general-reference spin-Fock partition diagonalizes occupied and virtual spin blocks internally, within irreps, and transforms guesses and state densities between that basis and the supplied working orbitals. The supplied occupied orbitals are unchanged. The occupied–virtual Fock block belongs to the perturbation, with t1 = -Fov/(epsilon_v-epsilon_o). Both secular matrices and virtual densities retain its first-order singles contributions. Thus ROHF support includes nonzero Fov; it uses this specified spin-Fock perturbation convention rather than assuming Brillouin’s theorem. Ground correlation includes the corresponding Fov*t1 contribution.

Ground-state MP2 doubles and spin-orbital doubles work arrays require O(o^2 v^2) storage; this is not an amplitude-direct ADC implementation. Packing retains only independent same-spin occupied/virtual pairs and all opposite-spin doubles. R[k].flatten() has a Euclidean independent-determinant metric. R[k].as_cluster() converts to CCpy’s redundant spin-block amplitudes; do not compare their unweighted norms.

correlation_energy is the raw DF-MP2 ground-state correlation energy. omega remains the raw ADC excitation energy, and raw_energy() returns uncompleted truncated-space totals. FNO energy() and completed_omega apply the configured MethodBase completion policy.

State densities and FNO integration

The virtual state-density blocks use the second-order, unrelaxed intermediate state representation (ISR) formulas, including the quadratic MP2-amplitude corrections to the singles contribution. They are not transition densities, MP2 ground densities, or simply CIS particle densities. Full state RDMs, transition moments, oscillator strengths and analytic ADC gradients are not implemented by this interface.

The built-in adc2 FNO provider is installed automatically unless a custom provider has already been registered:

from ccpy import CCSD, EOMCCSD, FNOOptions, FNOTarget

targets = (FNOTarget("EE", root=0, irrep="A", multiplicity=1),)
fno = FNOOptions(source="adc2", targets=targets, occupation_cutoff=1.e-6,
                 energy_completion="state_specific")
eom = EOMCCSD(mf, nfrozen=1, fno=fno, fno_targets=targets)
eom.initial_guess(multiplicity=1, nacto=2, nactu=2,
                  roots_per_irrep={"A": 1})
total = eom.energy()

The full-space provider uses the selected formulation and its corresponding bare-Hamiltonian guesses, with up to two active occupied and virtual orbitals for CISd. Set FNOTarget(..., formulation="full") for doubles/satellite state selection; its default is "folded". Targets with different formulations can be combined in one FNO policy. The reduced-space calculation projects each full-space ISR vector through the actual orbital transformation and uses it as its Davidson seed. A normalized projected overlap below 0.5 is treated as a lost root and raises an error. Degenerate-state identification may require a larger retained space or a custom provider with subspace tracking. Full/reduced calculations use the same frozen core, DF factors and perturbation convention. Multiple spin groups reuse one set of ground-state ADC intermediates.

The native ADC-FNO provider supports EE, IP and EA for RHF/ROHF/UHF and noncanonical references. DEA, DIP and SF ADC models remain unsupported and raise explicit capability errors. Ordinary MP2-FNO preparation remains available for all existing CC/EOM families.

Validation and references

tests/methods/test_adc2.py compares random full matrix actions and virtual state densities with PySCF 2.14.0’s independent UADC spin-block implementation on RHF and UHF references, using exactly the same DF integrals. EE also checks the ROHF branch. The charged PySCF ROHF path omits Fov contributions, so it is not used as an oracle for those terms. tests/methods/test_adc2_general_reference.py independently constructs the second-order BCH effective Hamiltonian and virtual density in small full Fock spaces, checking the charged non-Brillouin terms. It also checks matrix-action, density and Davidson-energy invariance under occupied/virtual rotations on RHF, ROHF and UHF references. The tests also check matrix symmetry, diagonal, determinant metric, Davidson residuals, a dense small-molecule spectrum, spin/irrep/root targeting, frozen core and full-retention ADC-FNO completion. PySCF is optional and used only for numerical tests.

The equations are based on J. Schirmer’s ADC construction and the transparent spin-block reference implementations by the PySCF ADC authors:

Attribution and the Apache-2.0 license are retained in THIRD_PARTY_NOTICES.md and ccpy/adc/LICENSE.apache2.

tests/methods/test_adc2_formulations.py checks folded/full energies, vectors and densities, explicit satellite roots, pole/guess guards and FNO selection with each formulation.