Selecting orbitals in CASCI with spin!=0

Hello pyscf developers and users, I have a problem regarding CASCI/CASSCF calculation with spin!=0.

Description

The problem appears when trying to select an active space without the orbitals with occ=1. from an ROHF mean-field (output from print(myhf.mo_occ), should be orbitals=8,9 with base=1). The energy from CASCI is higher than the Hartree-Fock energy (-111.723322367153 > -112.275444059104 in the example below), which is not expected. This is not happening when switching to spin=0 or selecting an active space containing orbitals 8 and 9.

Minimal example

from pyscf import gto, scf, mcscf

xyz ="""
C 0. 0. -0.52900402
H 0. 0.93762420 -1.12377976
H 0. -0.93762420 -1.12377976
O 0. 0. 0.67769796
"""

mol = gto.M(
    atom = xyz,
    basis = "sto3g",
    spin=2
    )

myhf = scf.RHF(mol)
myhf.kernel()

active = [7, 10]
mycas = mcscf.CASCI(myhf, 2, 2)
mo = mycas.sort_mo(active)
mycas.kernel(mo)

converged SCF energy = -112.275444059104
CASCI E = -111.723322367153  E(CI) = -0.843243057452625  S^2 = 2.0000000

More information

Performing the same trick with CCSD is giving consistent results, at least versus the HF energy.

from pyscf import cc
ccsd = cc.CCSD(myhf, frozen=[i for i in range(12) if i+1 not in active])
ccsd.ccsd()

E(UCCSD) = -112.275526525089  E_corr = -8.246598538875619e-05

[xubwa]

For ccsd, you are still starting from the HF determinant, the hilbert space still contains the HF determinant with a few extras, that’s why you are still getting a similar to HF but lower energy. In CASCI, it’s like you remove the two active electrons from the hf det and repopulate them in your new active space. Hope this helps you understand where the different comes from.

[AlexandreF-1qbit]

Hi @hebrewsnabla, thank you for responding. How can I check that? I also suspect that the energy is another solution to ROHF and I agree that the [8,9] should be the lowest CAS(2,2) triplet root. I wonder if it is the expected behavior (I find this confusing when comparing with the behavior of CCSD with frozen orbitals, as seen in my example above). If it is indeed the expected behavior, we could close the issue.