There is a subtlety with the definition of ccsd_double_amps generated by run_pyscf (and the _pyscf_molecular_data behind it). For a user to correctly instantiate the packed ccsd amplitudes in openfermion, he would actually need to do
uccsd_singlet_get_packed_amplitudes(ccsd_single_amps, (1/2)*ccsd_double_amps, num_qubits, num_electrons)
on data generated from run_pyscf. Notice the factor 1/2; In the literature, I don't think there is a clear winner on the prefactor in front of T_2, the CC doubles operator. I have seen 1, or 1/2, or 1/4.
However, as a simple test, I have compared several different methods, and for H2 and LiH, and found good agreement between 'classical' CCSD and quantum circuit UCCSD with the packed amplitudes generated by uccsd_singlet_get_packed_amplitudes(ccsd_single_amps, ccsd_double_amps, num_qubits, num_electrons), for input data from OpenfermionPsi4 and other test hdf5 files around, where there is no need for a pre-factor, and the assumptions/conventions therefore seem aligned.
Now, only for data generated by openfermionpyscf there is a need for this factor 0.5 (sanity check: omitting the factor gives results which are worse even than HF). Hope I explained that clearly, and perhaps someone can independently check this.
A possible solution is to adjust the property ccsd_double_amps of PyscfMolecularData and insert the factor there, in order to be consistent with what uccsd_singlet_get_packed_amplitudes assumes.
babbush
commented
5 years ago
There is a subtlety with the definition of
ccsd_double_ampsgenerated byrun_pyscf(and the_pyscf_molecular_databehind it). For a user to correctly instantiate the packed ccsd amplitudes in openfermion, he would actually need to douccsd_singlet_get_packed_amplitudes(ccsd_single_amps, (1/2)*ccsd_double_amps, num_qubits, num_electrons)on data generated fromrun_pyscf. Notice the factor 1/2; In the literature, I don't think there is a clear winner on the prefactor in front of T_2, the CC doubles operator. I have seen 1, or 1/2, or 1/4.However, as a simple test, I have compared several different methods, and for H2 and LiH, and found good agreement between 'classical' CCSD and quantum circuit UCCSD with the packed amplitudes generated by
uccsd_singlet_get_packed_amplitudes(ccsd_single_amps, ccsd_double_amps, num_qubits, num_electrons), for input data from OpenfermionPsi4 and other test hdf5 files around, where there is no need for a pre-factor, and the assumptions/conventions therefore seem aligned.Now, only for data generated by
openfermionpyscfthere is a need for this factor 0.5 (sanity check: omitting the factor gives results which are worse even than HF). Hope I explained that clearly, and perhaps someone can independently check this. A possible solution is to adjust the propertyccsd_double_ampsofPyscfMolecularDataand insert the factor there, in order to be consistent with whatuccsd_singlet_get_packed_amplitudesassumes.