Quantum HF/DFT-embedding algorithms

for electronic structure calculations: Scaling

up to complex molecular systems
Cite as: J. Chem. Phys. 154, 114105 (2021); https://doi.org/10.1063/5.0029536
Submitted: 14 September 2020 . Accepted: 11 February 2021 . Published Online: 15 March 2021

Max Rossmannek, Panagiotis Kl. Barkoutsos, Pauline J. Ollitrault, and Ivano Tavernelli

ARTICLES YOU MAY BE INTERESTED IN

Modeling nonadiabatic dynamics with degenerate electronic states, intersystem crossing,

and spin separation: A key goal for chemical physics
The Journal of Chemical Physics 154, 110901 (2021); https://doi.org/10.1063/5.0039371

Limitations of Hartree–Fock with quantum resources

The Journal of Chemical Physics 154, 044112 (2021); https://doi.org/10.1063/5.0018415

Quantum orbital-optimized unitary coupled cluster methods in the strongly correlated

regime: Can quantum algorithms outperform their classical equivalents?
The Journal of Chemical Physics 152, 124107 (2020); https://doi.org/10.1063/1.5141835

J. Chem. Phys. 154, 114105 (2021); https://doi.org/10.1063/5.0029536 154, 114105

© 2021 Author(s).
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

Quantum HF/DFT-embedding algorithms

for electronic structure calculations:
Scaling up to complex molecular systems
Cite as: J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536
Submitted: 14 September 2020 • Accepted: 11 February 2021 •
Published Online: 15 March 2021

Max Rossmannek,1,2 Panagiotis Kl. Barkoutsos,1 Pauline J. Ollitrault,1,3 and Ivano Tavernelli1,a)

AFFILIATIONS
1
IBM Quantum, IBM Research – Zurich, 8803 Rüschlikon, Switzerland
2
Department of Chemistry, University of Zurich, Winterthurerstrasse 190, 8057 Zürich, Switzerland
3
Laboratory of Physical Chemistry, ETH Zurich, 8093 Zürich, Switzerland

a)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
In the near future, material and drug design may be aided by quantum computer assisted simulations. These have the potential to target
chemical systems intractable by the most powerful classical computers. However, the resources offered by contemporary quantum com-
puters are still limited, restricting the simulations to very simple molecules. In order to rapidly scale up to more interesting molecular
systems, we propose the embedding of the quantum electronic structure calculation into a classically computed environment obtained at
the Hartree–Fock (HF) or density functional theory (DFT) level of theory. This result is achieved by constructing an effective Hamilto-
nian that incorporates a mean field potential describing the action of the inactive electrons on a selected Active Space (AS). The ground
state of the AS Hamiltonian is then determined by means of the variational quantum eigensolver algorithm. We show that with the pro-
posed HF and DFT embedding schemes, we can obtain significant energy corrections to the reference HF and DFT calculations for a
number of simple molecules in their strongly correlated limit (the dissociation regime) as well as for systems of the size of the oxirane
molecule.
Published under license by AIP Publishing. https://doi.org/10.1063/5.0029536., s

I. INTRODUCTION of reach due to the inherent exponential scaling of the Hilbert space
associated with the electronic structure calculations. While several
Quantum chemistry simulations allow for the prediction of approximate methods have been developed in the past to circum-
important chemical processes throughout, for instance, the elu- vent this issue, these often break down when considering strongly
cidation of reaction mechanisms by means of the calculation of correlated systems such as transition metal complexes4 and compli-
ground or excited state electronic structure properties.1 A variety cated catalytic processes.5 In the past decades, quantum comput-
of research and industrial applications such as chemical catalysis, ing has emerged as a new potential computational paradigm for
material design, drug discovery, and photo-chemical processes for the solution of many problems in chemistry and physics for which
solar energy conversion, just to name a few,2,3 could take advantage classical algorithms have an unfavorable scaling. In particular, quan-
of these methods. Since the development of the first computers, the tum computing has been shown to be a useful resource in a vari-
research on quantum chemistry has blossomed and a large variety ety of research areas such as chemistry,4,6 drug discovery,7 strongly
of algorithms have been developed aspiring to achieve more accu- correlated systems,8,9 field theory,10,11 material science,12 and many
rate solutions of the Schrödinger equation. However, despite many others.
theoretical and algorithmic advances, the solutions of many inter- Despite these recent advances and the possibility to execute
esting and relevant problems in chemistry and physics remain out calculations on quantum devices (e.g., Ref. 13), the application of

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-1

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

quantum algorithms is still in its infancy. In fact, most of the research the total set of orbitals (N tot ) is considered for the wavefunction
in chemistry relies on hybrid quantum–classical algorithms,14 which Ansatz.
use highly optimized classical (number crunching) functionalities In this work, we propose HF and DFT-based quantum embed-
together with quantum algorithms for the representation and opti- ding schemes based on the well known notion of an Active Space
mization of the system wavefunction. The most well known quan- (AS),45,46 which defines the set of active orbitals described by the
tum chemistry algorithm that provides the means to leverage state- quantum algorithm. To this end, we construct an effective Hamil-
of-the-art quantum hardware is the Variational Quantum Eigen- tonian, which incorporates a mean field potential of the inactive
solver (VQE).15 electrons and, thus, fully replaces the explicit mapping of the corre-
For the representation of the many-body wavefunction in sponding orbitals in the quantum register. The quantum algorithm
quantum circuits, some of the approaches derived in quantum is therefore restricted to a subset of active orbitals, which, however,
chemistry can be mapped directly to quantum algorithms. In par- feel the presence of the environment through the action of the mean
ticular, the Hartree–Fock (HF) method has proven to pose a use- field potential generated by the inactive electrons of the environ-
ful starting point for the mapping of electronic structure problems ment. Similar approaches of the HF embedding have been proposed
in the qubit space using the so-called second quantization formal- in the literature47,48 mainly based on Dynamical Mean Field The-
ism. Among the most commonly used post-HF expansions of the ory (DMFT)49 and Density Matrix Embedding Theory (DMET)50
many-electron wavefunction in quantum computing is the Coupled for the high-level description of the subsystem. The latter aims at
Cluster (CC) Ansatz,16–18 which allows for a systematic and con- a similar HF embedding scheme. However, while the focus of its
trolled inclusion of higher order configurations starting from the authors was on the development of a self-consistent HF embed-
uncorrelated HF Slater determinant. Several quantum implemen- ding approach,51 in this work, we only consider iterative embedding
tations of CC have already been introduced in the literature,15,19–25 within the framework of DFT. Concerning the DMFT approach,
including schemes for the optimization of the one-electron molecu- this is based on Green’s function techniques and therefore it is not
lar basis functions.9,26 It should also be noted that, in the quantum particularly suited for the kind of molecular application of inter-
implementation of CC, the Trotterization leads to a dependence of est to this work. Additionally, during the preparation of this paper,
the results on the ordering of the cluster operators.25,27 In addition another related approach appeared in the literature.52 In this case,
to the classically inspired expansions, pure native quantum repre- the authors propose a DFT embedding scheme similar to ours, which
sentations of the many-electron wavefunctions that can be better however uses a different Ansatz to resolve the double counting prob-
optimized for the available quantum hardware platforms have been lem of the correlation terms. Furthermore, they do not update the
proposed.21,28,29 However, some of these wavefunction Ansätze can embedding potential in a self-consistent manner as we do in this
exhibit vanishing small energy gradients, leading to difficulties in work.
the optimization process (i.e., Barren plateaus). To overcome this This paper is organized as follows: In Sec. II, we outline the the-
problem, other CC techniques have been developed, which con- ory and the implementation of the proposed AS schemes for quan-
struct the operators directly in the qubit representation.30,31 These tum electronic structure calculations embedded in HF and DFT.
techniques generally result in significantly shorter circuit depths. We split the derivation into two parts: one for the HF embedding
The combination of the VQE algorithm with the different wave- scheme and the other for the DFT embedding scheme. Section III
function Ansätze showed already interesting results in the calcula- lists the technical details of our numerical methods. In Sec. IV, we
tion of ground19,28,32–34 and excited state properties15,35–44 of simple present and discuss results on the dissociation of simple molecular
molecules (up to a few atoms). However, this protocol does not allow systems, such as water, molecular nitrogen, molecular oxygen, and
us to scale to larger systems using the currently available classical oxirane obtained with both types of embedding schemes. Section V
simulators of quantum circuits (limited to a maximum of about 50 summarizes and concludes.
qubits) or the available quantum computers (also limited to a few
tens of qubits). Therefore, in order to leverage the potential advan-
tage of the available quantum algorithm, we explore the possibility
of an embedding scheme in which only a portion of the full system is II. THEORY
represented by the high-level quantum computing approach, while In this work, we propose two embedding schemes for quantum
the rest is treated with an efficient but (necessarily) approximated electronic structure algorithms based on HF and Kohn–Sham (KS)
classical representation of the electronic structure, such as HF or DFT Molecular Orbitals (MOs). The subsystem solved by means of
Density Functional Theory (DFT). This embedding approach is of the quantum approach [such as quantum Unitary Coupled Cluster
particular relevance when the complex, highly correlated, subsystem Singles and Doubles (q-UCCSD)1,15,21 ] is embedded in the poten-
can be localized in a well defined subspace of the complete set of tial generated by the environment (i.e., the remaining electrons),
one-electron orbitals used to represent the many-electron wavefunc- which is computed within the HF or DFT framework. Our solu-
tion. In this case, an accurate description of the electronic structure tions are based on the Range-Separation (RS) technique for the
is obtained at lower cost in terms of the number of orbitals, namely, two-electron integrals,53 which allows for a rigorous partitioning of
4
O(Nqc ) for the quantum subsystem (where N qc is the number of the problem into a subsystem (i.e., the AS) and its environment.
2 3
orbitals encoded in the quantum circuit) and O(Nenv ) to O(Nenv ) If this partitioning is done wisely, we can achieve a good level of
for the environment (where N env is the number of orbitals assigned accuracy for many properties of interest while significantly reduc-
to the classical processor and N tot = N qc + N env ); this is to be com- ing the costs of the calculation. Furthermore, in the case of the DFT
4
pared with the O(Ntot ) scaling when no embedding is used and embedding scheme (which is the main target of this work), we will

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-2

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

extend the algorithm to include the self-consistent optimization of

the embedding potential, leading to more accurate energies and den-
sities. In the following, we will call active electrons the electrons that
are part of the AS, while the remaining ones will be referred to as
inactive.

A. Hartree–Fock embedding
In this section, we derive the so-called inactive Fock operator.
The goal of this operator is to embed the quantum computation into
a classically computed environment treated at the HF level of theory,
through the notion of an AS. While this method is not new and dif-
ferent variants of it have been implemented before in other software
packages,52,54,55 in the following, we summarize the key concepts that
are needed for its implementation within the framework of quantum
computing in QISKIT.56 This section also lays down the fundamentals
for the implementation of the DFT embedding scheme presented in
Sec. II B.
The benefit of this embedding scheme lies in outsourcing
the calculation of the inactive electrons to the classical HF driver,
while the quantum computation is restricted to the AS. In this
way, less qubit resources are necessary to investigate the electronic
energy of a molecular system, making the entire calculation much FIG. 1. Separation of the MOs into active and inactive components. The active
more efficient while keeping a good level of accuracy. Figure 1 orbitals (blue box) are mapped onto the qubit space and treated with the q-UCCSD
depicts the separation of the orbitals into the active and inactive approach, while the inactive ones (orange box) are part of the HF/DFT embedding
and are evaluated classically. Effective Core Potentials (ECPs) can be used in
spaces. replacement of all inactive core electrons (white box) with the aim of reducing the
The total electronic energy, E, is defined by the expectation computational cost.
value of the system Hamiltonian, Ĥ,

1
E = ⟨Ψ∣Ĥ∣Ψ⟩ = ∑ hpq Dpq + ∑ gpqrs dpqrs , (1)
pq 2 pqrs integrals, hpq ; the active one- and two-electron density matrices,
DA and dA , replace D and d; and the constant energy offset, EI , is
where Ψ is the wavefunction, hpq and g pqrs are the one- and two- added.
electron integrals, respectively, and D and d are the one- and two- Therefore, the Hamiltonian that we evaluate on the quantum
particle density matrices. computer (qc) takes the form
To achieve the implementation of the HF embedding, we split
† † †
the one-electron density into an active and inactive part, D = DA I
Ĥqc = ∑ Fuv âu âv + ∑ guvxy âu âv âx ây , (5)
+ DI . In the MO basis, the latter simplifies to DIiq = 2δiq , where we uv uvxy

use Helgaker’s index notation46 in which i, j, k, l denote inactive, u,

v, x, y denote active, and p, q, r, s denote general MOs. As shown in where â†u and âu are the creation and annihilation Fermionic oper-
Appendix A 1, inserting this into Eq. (1) leads to ators (later mapped to the qubit space using the parity transforma-
tion57 ). Note that all indices are restricted to the AS, significantly
1 reducing the required quantum resources.
E = EI + ∑ Fuv
I
DAuv + A
∑ guvxy duvxy , (2) The extension for the unrestricted formalism is obtained in a
uv 2 uvxy
similar manner and is outlined in Appendix A 2.
where we define the inactive Fock operator, B. Density functional theory embedding
I
Fpq = hpq + ∑(2giipq − giqpi ), (3) In order to extend the embedding to work with DFT, we need
i to introduce a RS of the two-electron integrals, g pqrs .53 To this end,
we split the two-electron operator, ĝpq , into a Long-Range (LR) and
and the inactive energy, a Short-Range (SR) part,

1 1 μ,LR μ,SR
EI = ∑ hjj + FjjI = I I
∑(hij + Fij )Dij . (4) ĝpq = = ĝpq + ĝpq , (6)
j 2 ij ∣r̂p − r̂q ∣

Comparing Eqs. (1) and (2), we observe the following differences. where μ is the RS parameter of unit a.u.−1 . This is necessary in
In Eq. (2), the inactive Fock operator, F I , replaces the one-electron order to avoid a double counting of the correlation terms, which are

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-3

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

present in both DFT and Wave Function Theory (WFT). Since DFT as
is known to be accurate for SR interactions,53 we can use it to treat SR
δEcoul+xc
the SR part, while the LR interactions are calculated with WFT. ∫ [ρ]Δρ(⃗r)d⃗r = ∑(jSR SR
pq + νxc,pq )ΔDpq . (13)
δρ(⃗r) pq
Our derivation of the following equations follows that of
Hedegård et al.58 closely. Additionally, we provide our detailed Because of the non-linearity of Eq. (12b), the density needs to be
derivations in Appendix B. updated in an iterative, self-consistent manner. Therefore, we define
With the RS of the two-electron integrals in place, we can split the density and the density matrix at the iteration step i as
the total electronic energy into two terms,
ρ(i+1) = ρ(i) + Δρ(i) , (14a)
μ,LR μ,SR
E= EWFT + Ecoul+xc,DFT . (7) (i+1) (i) (i)
Dpq = Dpq + ΔDpq . (14b)

Note that Eq. (7) provides an adiabatic connection between the pure This leads to the final form of the total electronic energy,
DFT and the pure WFT solutions through the coupling parameter, μ.
1
However, in order to simplify the notation, we drop the superscript E = ∑(hij + FijI,LR )DIij
μ since it is anyways implied by the separation into LR and SR. 2 ij
Analogous to Sec. II A, we can introduce an AS in the WFT SR (i) 1 I,SR I
+ Exc [ρ ] + ∑ j Dij
part, 2 ij ij
1 A,(i),SR SR A,(i)
I,LR
E = EWFT A,LR
+ EWFT SR
+ Ecoul+xc,DFT . (8) − ∑[( juv + νxc,uv [ρ(i) ])Duv
uv 2
I,LR A,(i),SR A,(i+1)
Note that the difference between Eqs. (2) and (8) is that WFT − (Fuv + jI,SR
uv + juv
SR
+ νxc,uv [ρ(i) ])Duv ]
only treats the LR part. Thus, the inactive Fock operator, defined in 1 LR A,(i+1)
Eq. (3), becomes + ∑ guvxy duvxy , (15)
2 uvxy

I,LR LR LR
Fpq = hpq + ∑(2giipq − giqpi ). (9)
i

In order to properly combine the SR-DFT and LR-WFT calcu-

SR SR
lations, we need to handle the non-linearity of Ecoul+xc,DFT = Ecoul+xc
with respect to the electronic density, ρ,

SR SR SR
Ecoul+xc [ρ + Δρ] ≠ Ecoul+xc [ρ] + Ecoul+xc [Δρ], (10)

where Δρ is the correction to the density obtained from the WFT

calculation. However, a linear approximation can be obtained with
the following replacement:

SR
SR SR δEcoul+xc
Ecoul+xc [ρ + Δρ] − Ecoul+xc [ρ] ≈ ∫ [ρ]Δρ(⃗r)d⃗r. (11)
δρ(⃗r)

The right-hand side of Eq. (11) can then be expressed in terms of the
Coulomb integrals,

SR FIG. 2. Illustration of the DFT embedding scheme. During initialization (0), a DFT
δEcoul calculation of the full system is performed using a classical code, providing the
jSR
pq = ⟨ϕp ∣
SR
[ρ]∣ϕq ⟩ = ∑ gpqrs Drs , (12a)
δρ(⃗r) rs initial density, ρ(0) , and the one- and two-electron integrals, hpq and gpqrs . The den-
sity is then split into inactive, I, and active parts, A; and the two-electron integrals
are separated into long-range, LR, and short-range, SR, components. In step (1),
and the exchange integrals, the inactive SR energy contribution is calculated at the DFT level of theory. The
resulting “active” density component (a) is used in step (2) to initialize the VQE
optimization. This returns the active LR energy contribution and the updated elec-
SR
SR δExc SR tronic density, which is used as a new input for the DFT calculation, (b). Steps (1)
νxc,pq = ⟨ϕp ∣ [ρ]∣ϕq ⟩ = νxc,pq [ρ], (12b) and (2) are repeated until convergence.
δρ(⃗r)

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-4

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

where we have ordered the terms such that the top line contains all In all applications, we selected the STO-3G basis set, which
contributions, which remain constant for the duration of the whole enabled the calculation of the exact reference curves with full
iterative procedure, the second and third lines correspond to the SR- CI (FCI). Furthermore, the optimal value of the range-separation
DFT, and the remaining lines correspond to the LR-WFT energy parameter, μ, was selected case-by-case by scanning the interval
terms, respectively. 0.01–500 and finally picking the value that gives the lowest energy
Figure 2 summarizes the implementation of this DFT embed- (see the supplementary material for more details).
ding scheme. The initialization step includes all the pre-calculations
and the computation of the constant inactive LR energy contribution
[first line of Eq. (15)]. The resulting energy terms of steps (1) and IV. RESULTS AND DISCUSSION
(2) in Fig. 2 correspond to lines two and three, and four and five of In this section, we present results on the dissociation of simple
Eq. (15), respectively. These two calculations iterate, upon exchang- molecular systems of increasing size, namely, molecular nitrogen,
ing the active electronic density, ρA , until the total electronic energy molecular oxygen, water, and oxirane using the proposed HF and
reaches convergence. DFT embedding schemes (see Secs. II A and II B, respectively).
To enable the comparison with exact results obtained with FCI,
we limited our calculations to the relatively small basis set STO-
III. NUMERICAL METHODS 3G. The results obtained with both embedding schemes are com-
The HF and DFT embedding schemes have been implemented pared with standard HF and DFT calculations as well as with other
in the development version 0.8 of QISKIT Aqua Chemistry. The source commonly used, correlated, post-HF solutions such as Coupled
code will be made available in the Github repository.59 For the clas- Cluster Singles and Doubles (CCSD) and Complete Active Space
sical computing backend, we choose PySCF60 since it allows quick Self-Consistent Field (CASSCF). In particular, we demonstrate that
prototyping within Python, the same programming language used even within the embedding approximation, the quantum CC algo-
for QISKIT. rithm can solve some pathologies of classical CCSD in the strongly
All the results presented hereafter are obtained by means of correlated regime, such as the dissociation of N2 and O2 . Finally,
diagonalizing the Hamiltonian with the NumPyEigensolver algo- with the study of oxirane, we show how the proposed embedding
rithm, as implemented in QISKIT. In the case of the non-iterative schemes can extend the use of quantum algorithms to molecular
HF embedding scheme, we also run VQE simulations with the sizes that would exceed the capabilities of state-of-the-art quantum
statevector backend.59 This backend implements an exact, i.e., computers.
noiseless, simulation of the quantum circuit and, thus, is expected
to converge to the same result as the NumPyEigensolver approach A. A benchmark system: Water
when a suitable wavefunction Ansatz is chosen.
We start with the investigation of the AS of water, analyzing
A. Hartree–Fock embedding the dependence of the energy on the number of molecular orbitals
and electrons included in the quantum description. Standard post-
In all simulations using the HF embedding, we use the parity
HF methods such as Complete Active Space Configuration Interac-
fermion-to-qubit mapping57 and the q-UCCSD Ansatz21 for the rep-
tion (CASCI) and CASSCF as implemented in PySCF60 are used to
resentation of the electronic wavefunction. In QISKIT, the q-UCCSD
benchmark the HF embedding scheme. A study on the effect of the
Ansatz is implemented in a well-defined manner whose details are
size of the basis set from STO-3G68 to 6-31G∗ 69 and cc-pVTZ70 is
outlined in Ref. 21 and Appendix C. Furthermore, qubits are tapered
given in the supplementary material (Figs. S1 and S2). In the fol-
off61 in order to maximally reduce the computational costs. The clas-
lowing, we restrict our calculations to the STO-3G basis. Finally,
sical optimizer L-BFGS-B62 is used for the optimization of the VQE
we apply both HF and DFT embedding approaches to the study of
parameters.
the symmetric double-proton dissociation, focusing our attention
on the dissociation limit, where strong correlation effects become
B. Density functional theory embedding important.
In all DFT embedding applications, we use the RS-XCF (Range-
Separated Exchange-Correlation Functional) ldaerf scheme,63,64 as
implemented in the xcfun library65 for the separation of the LDA 1. General properties of active-space embedding
(Local Density Approximation) functional into its short and long techniques
range components (see Sec. II B). This approach achieves the split- Figure 3 summarizes the main results obtained for water in the
ting of the two-electron integrals by means of the error function, STO-3G basis. In general, the ground state energies obtained with
which is a common approach in RS-DFT.53,66,67 The use of the LDA HF embedding (blue triangles in Fig. 3) are in qualitatively good
functional is solely motivated by the current technical limitations agreement with the classical CASCI results60 (orange crosses). The
of the PySCF code. Future extensions to allow the use of arbitrary CASSCF60 energy values (green circles) are consistently lower (or
DFT functionals are under investigation. Nonetheless, the proposed equal) than the CASCI ones since the former also includes the opti-
scheme is fully independent from the nature of the selected func- mization of the orbital coefficients, which are kept fixed in CASCI.
tional and all applications presented in the following should be con- More interestingly, all statevector-based VQE calculations using the
sidered as proof-of-principle demonstrations extendable to any type q-UCCSD Ansatz (brown pentagons in Fig. 3) also converge to the
of DFT functional. exact solutions.

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-5

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 4. Energy diagram of the KS-MOs of H2 O in the STO-3G basis. The solid
lines correspond to occupied MOs, while the dotted lines represent virtual ones.

Figure 4 depicts these orbitals in an energy diagram. Note that we do

not display properly the HF-MOs, but their difference to the shown
KS-MOs is minor.
Figure 5 depicts the dissociation curves from d = 0.5 Å up to
d = 2.0 Å, as computed by the proposed DFT- and HF embedding
schemes for a selection of ASs. As a comparison, we also show the
FIG. 3. Electronic structure energies (in Hartree) of a water molecule obtained with corresponding profiles computed with CCSD and FCI (computed
the HF embedding scheme for different choices of the AS, which are reported in with PySCF). In general, increasingly larger ASs provide results
the upper panel. We report absolute (right axis) and relative energies with respect
in closer agreement with the exact solution, as given by FCI even
to HF (left axis). The coloring follows the same scheme used in Figs. 1 and 2:
orange for the inactive HF orbitals that define the embedding, blue for the orbitals though, as mentioned in Subsection IV A 1, the optimal choice of
belonging to the AS, and black for the remaining virtual ones. The classical HF AS cannot be determined a priori.71,72 This becomes more evident
reference (black, solid) is obtained with RHF in the STO-3G basis. As a reference, looking at the lower panels of Fig. 5 where we plot the energy dif-
we also show the results obtained with FCI, CASCI, and CASSCF. ferences with respect to the exact value. In addition, note that the
DFT embedding scheme is not variational in contrast to HF-based
embedding methods.

It is worth mentioning that the choice of the AS has a strong,

and not always predictable, effect on the outcome. In particular, it
is hard to predict a priori the optimal AS that maximizes the per-
formance of the embedding calculation.71,72 In the absence of a sys-
tematic procedure, we can explore all possible AS configurations,
as shown in Fig. 3, for the case of water in the STO-3G basis set.
However, this procedure is only possible for relatively small sys-
tems, while it becomes impractical in the most general cases (see,
for instance, the additional results in the supplementary material,
Figs. S3–S5), leading to a large degree of uncertainty in the appli-
cation of the embedding approach. On the other hand, recently, we
witnessed the development of new numerical techniques, which aim
at automatizing the selection of the optimal AS in molecular sys-
tems.33,72–75 The combination of these approaches with our embed-
ding scheme goes however beyond the scope of this work and will
become the subject of future investigations.

2. The double-dissociation of water

FIG. 5. (Upper panels) Symmetric dissociation profiles of a water molecule at a
In this section, we present the symmetric dissociation profile of fixed H–O–H angle (α = 105.4○ ) in the STO-3G basis set. The left panel reports the
a water molecule at fixed H–O–H angle of α = 105.4○ in the STO- energies obtained with the DFT embedding scheme at μ = 7.0, while the right panel
3G basis set using different approaches. For the DFT embedding shows the same curves obtained with the HF embedding scheme, respectively.
scheme outlined in Sec. III B, we swept over the range-separation The size of the ASs is reported in the legend with the format: (number of electrons,
parameter, μ, and found the optimal value for this system to be μopt number of orbitals). The HF (gray, dotted), DFT (gray, dashed-dotted) (LDA/VWN
XCF), CCSD (blue, dashed), and FCI (red, solid) references were computed with
= 7.0. These results are reported in Fig. S6 of the supplementary
PySCF. (Lower panels) Energy differences with respect to the exact FCI reference.
material together with a sketch of the KS-MOs (see Fig. S10).

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-6

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

Table I reports dissociation energies and non-parallelity errors HF embedding in general produces better estimates than the DFT
(NPEs) for the different methods. The NPE is defined as the dif- embedding due to the fact that we use HF orbitals for the refer-
ference between the maximum and minimum error over the entire ence FCI calculation. In addition, in general, the HF embedding
energy profile evaluated with respect to the exact calculation.76 For improves upon the initial HF-estimate more than the DFT embed-
brevity, we only report a selection of ASs, while the complete set of ding does. Finally, the quality of the CASSCF results suggests that
calculations can be found in Table S1 of the supplementary mate- future extensions to incorporate an iterative optimization of the
rial. While all HF embedding calculations lead to better dissocia- orbitals could yield further improvements, especially in the case of
tion energies than standard HF (ΔEHF = 559.7 mHa), DFT embed- small ASs.
ding with CAS(2, 3) and CAS(4, 4) performs worse than standard
DFT (ΔEDFT = 361.7 mHa). With the exclusion of these last two B. The dissociation of molecular nitrogen
cases, we observe a systematic and significant improvement of all and oxygen
results as the size of the AS increases. In particular, the HF embed-
As a second example, we investigate the dissociation of molec-
ding can provide nearly exact results (ΔEFCI = 247.1 mHa) with
ular nitrogen and oxygen, which show strong correlation character
CAS(8, 6) and CAS(10, 7), which is not a surprise as the latter cor-
at large distances. In fact, it is well known in the literature that for
responds to the FCI calculation. On the other hand, we observe
both systems CCSD fails to produce the correct dissociation profile
that the DFT embedding cannot reach the same accuracy even with
(cf. Ref. 9 and the references therein).
the larger AS. This is due to the nature of the LDA-DFT-orbitals,
For the determination of the range-separation parameter μ,
which, by definition, cannot reproduce the FCI result based on
we follow the procedure outlined in Sec. III B. The optimal val-
HF orbitals. However, in the case of large molecular systems with
ues were found to be μopt [N2 ] = 7.0 and μopt [O2 ] = 8.0, respec-
hundreds of electrons, the DFT embedding will become the only
tively (see Figs. S7 and S8 of the supplementary material). A sketch
viable option for a correct description of a molecular embedding
of the KS molecular orbitals is also given in the supplementary
potential.
material (Figs. S11 and S12). Figures 6 and 7 depict energy dia-
Additional insights on the accuracy of the different embedding
grams for the KS-MOs of nitrogen and oxygen, respectively, and
schemes can also be gained from the analysis of the NPE results.
provide a useful indication of the relative energies of the orbitals
The values that show an improvement compared to the standard
involved in our choices of AS. In Figs. 8 and 9, we report the dis-
HF and DFT calculations are highlighted in bold. In agreement
sociation profiles between d = 1.0 Å and d = 3.0 Å computed with
with the observations based on the dissociation energies and on the
the different AS sizes together with the standard HF, DFT, and
profiles shown in Fig. 5, we observe a systematic improvement of
CCSD results and the reference FCI calculation. In the case of N2 ,
the quality of the results by increasing the AS [from CAS(6, 5) to
the curve produced with CCSD is qualitatively wrong, showing an
CAS(10, 7)]. Even though the above results were obtained with μ
unphysical drop in energy as the distance increases beyond about
values selected through energy minimization, other techniques exist
d = 2 Å. This result is in agreement with the results reported in
for this purpose,67 which could lead to a further improvement of our
Ref. 9. The CCSD situation of O2 is slightly better, even though the
results.
deviation from the exact FCI curve is sizable along the entire disso-
It is worth mentioning that this behavior cannot be general-
ciation profile. Concerning the standard HF and DFT calculations,
ized to all situations. In fact, as it will also become evident for the
we also observe large discrepancies compared to the exact curve in
other systems, increasing the number of electrons while keeping the
both systems, with the first one showing large overestimation of the
number of orbitals constant implies a reduction of the number of
dissociation energies due to the complete neglect of the correlation
virtual orbitals available for the excitations. In addition, the order-
energy.
ing of the molecular orbitals may change upon stretching, shifting
in and out new orbitals from the selected AS, and leading to sud-
den changes in the energy profiles. Once more, we observe that the

TABLE I. Dissociation energies, ΔE, and NPE values for the symmetric double-
stretch of water. ΔE are computed as the difference between the energies at equilib-
rium, d = 0.948 Å, and at d = 2.0 Å. The NPE values are given within parentheses.
The results for HF, DFT, FCI, and CCSD are ΔE HF = 559.7 (347.3) mHa, ΔE DFT
= 361.7 (122.8) mHa, ΔE FCI = 247.1 (0.0) mHa, and ΔE CCSD = 236.0 (12.0) mHa,
respectively. Situations that show an improvement of the NPE compared to standard
HF and DFT are highlighted in bold.

CAS μ = 7.0 (NPE) HF-Emb. (NPE) CASSCF (NPE)

(2, 3) 557.0 (347.2) 452.5 (239.7) 368.6 (165.0)

(4, 4) 393.4 (181.1) 210.5 (118.3) 244.0 (117.4)
(6, 5) 239.6 (20.3) 235.5 (11.6) 245.0 (2.1)
(8, 6) 253.7 (11.7) 247.1 (0.1) 247.1 (0.1) FIG. 6. Energy diagram of the KS-MOs of N2 in the STO-3G basis. The solid lines
(10, 7) 254.1 (12.2) 247.1 (0.0) 247.1 (0.0) correspond to occupied MOs, while the dotted lines represent virtual ones. In red
are the orbitals involved in at least one of the ASs considered.

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-7

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 7. Energy diagram of the KS-α-SOs of O2 in the STO-3G basis. The solid
lines correspond to energy levels for which both α and β SO are occupied, the
dashed lines are singly occupied, and the dotted lines represent virtual orbitals. In
red are the orbitals involved in at least one of the ASs considered.

FIG. 9. (Upper panels) Dissociation profiles of an oxygen molecule in the STO-3G

basis set. The left panel reports the energies obtained with the DFT embedding
As for the case of water, increasing the size of the AS in the HF scheme at μ = 8.0, while the right panel shows the same curves obtained with the
embedding scheme, we observe a gradual increase in the accuracy of HF embedding scheme, respectively. The size of the ASs is reported in the legend
the calculations, reaching for the larger AS [CAS(6, 6) and CAS(8, with the format: (number of electrons, number of orbitals). The HF (gray, dotted),
DFT (gray, dashed-dotted) (LDA/VWN XCF), CCSD (blue, dashed), and FCI (red,
7) in the case of N2 and CAS(8, 6) and CAS(10, 7) in the case of O2 ] solid) references were computed with PySCF. (Lower panels) Energy differences
results within a few mHa from the reference curve. We also observe with respect to the exact FCI reference.
for O2 some discontinuities in the dissociation profiles at about d = 2
Å in both CAS(2, 3) and CAS(4, 4) calculations. This effect is due to
a swap in the ordering of the KS orbitals induced by the stretching of
the bond, which causes an abrupt change in the composition of the
AS. This has a particular impact on the energy in the case of small In Tables II and III, we summarize the results of the N2 and
AS composed of only a few KS orbitals around the Fermi level. O2 calculations reporting the values of the dissociation energies
and NPEs for the different AS sizes, together with the CASSCF
results. These results confirm our previous observations, which
point toward a systematic increase in the quality of the results with
the increase in the AS size. However, we need to stress once more
that due to the small size of the basis set, the different number of vir-
tual orbitals in CAS(n, m) when the number of electrons n increases
at fixed value of m, and the issue with the orbital crossing men-
tioned above, we cannot always expect a monotonic decrease in the
error as a function of parameters n and m. As already mentioned
above, in many cases, a careful analysis of the orbitals included in the
AS and their occupancy is required in order to achieve the desired
accuracy.

TABLE II. Dissociation energies, ΔE, and NPE values for the dissociation of the
nitrogen molecule. ΔE are computed as the difference between the energies at equi-
librium, d = 1.078 Å, and at d = 3.0 Å. The NPE values are given within parentheses.
The results for HF, DFT, FCI, and CCSD are ΔE HF = 1009.0 (828.9) mHa, ΔE DFT =
542.5 (342.3) mHa, ΔE FCI = 201.4 (0.0) mHa, and ΔE CCSD = −10.2 (221.8) mHa,
respectively. Situations that show an improvement of ΔE and the NPE compared to
standard HF and DFT are highlighted in bold.
FIG. 8. (Upper panels) Dissociation profiles of a nitrogen molecule in the STO-3G
basis set. The left panel reports the energies obtained with the DFT embedding CAS μ = 7.0 (NPE) HF-Emb. (NPE) CASSCF (NPE)
scheme at μ = 7.0, while the right panel shows the same curves obtained with the
HF embedding scheme, respectively. The size of the ASs is reported in the legend (2, 4) 597.8 (417.8) 597.4 (416.5) 597.4 (416.9)
with the format: (number of electrons, number of orbitals). The HF (gray, dotted), (4, 5) 431.8 (243.4) 432.8 (249.7) 441.9 (250.9)
DFT (gray, dashed-dotted) (LDA/VWN XCF), CCSD (blue, dashed), and FCI (red, (6, 6) 180.8 (66.6) 172.5 (29.0) 186.7 (15.4)
solid) references were computed with PySCF. (Lower panels) Energy differences
with respect to the exact FCI reference.
(8, 7) 189.7 (16.4) 180.4 (21.1) 194.6 (6.8)

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-8

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

TABLE III. Dissociation energies, ΔE, and NPE values for the dissociation of the
oxygen molecule. ΔE are computed as the difference between the energies at equi-
librium, d = 1.218 Å, and at d = 3.0 Å. The NPE values are given within parentheses.
The results for HF, DFT, FCI, and CCSD are ΔE HF = 687.9 (598.5) mHa, ΔE DFT =
340.6 (202.8) mHa, ΔE FCI = 138.3 (0.0) mHa, and ΔE CCSD = 384.8 (248.2) mHa,
respectively. Situations that show an improvement of ΔE and the NPE compared to
standard HF and DFT are highlighted in bold.

CAS μ = 8.0 (NPE) HF-Emb. (NPE) CASSCF (NPE)

(2, 3) 367.3 (338.0) 687.9 (598.5) 687.9 (598.5)

(4, 4) 93.1 (147.7) 365.5 (260.7) 54.8 (274.0)
(6, 5) 135.2 (50.4) 138.9 (37.1) 143.5 (41.0)
(8, 6) 119.7 (18.6) 120.5 (17.9) 133.6 (4.8)
(10, 7) 124.8 (13.6) 125.3 (14.2) 136.8 (1.6)

C. The dissociation of oxirane

FIG. 11. (Upper panels) Dissociation profiles of an oxirane molecule in the STO-
As a final example, we consider the oxirane molecule, C2 H4 O, a 3G basis set. The left panel reports the energies obtained with the DFT embedding
system that is not possible to simulate with a brute force all-electron scheme at μ = 6.0, while the right panel shows the same curves obtained with the
calculation with state-of-the-art quantum computers. In fact, the HF embedding scheme, respectively. The size of the ASs is reported in the legend
with the format: (number of electrons, number of orbitals). The HF (gray, dotted),
number of qubits and the required circuit depth for the representa- DFT (gray, dashed-dotted) (LDA/VWN XCF), CCSD (blue, dashed), and FCI (red,
tion of the system wavefunction would by far exceed the possibility solid) references were computed with PySCF. The DFT reference line has been
of current hardware. On the other hand, we can show that by using shifted by −0.3Ha to improve the scaling of the figure. (Lower panels) Energy
DFT embedding with a suitable choice of the active space, we can differences with respect to the exact FCI reference.
achieve a sizable improvement of the energetics associated with the
cleavage of the C–C bond in oxirane compared to standard DFT.
Once more, due to the current limitation of our embedding schemes
and the requirements for the calculation of the FCI reference, we larger systems. However, we expect that in order to achieve chem-
limit the DFT calculation to the LDA functional and the basis set ically accurate results, the use of better DFT functionals (beyond
to STO-3G. The KS-MOs obtained under these conditions are pre- the simple LDA used here) for the description of the embedding
sented in Fig. S13 of the supplementary material, and as previously potential will become of increasing importance. In addition, in this
encountered, Fig. 10 illustrates an energy diagram of this system. case, the accuracy of the dissociation profile increases smoothly with
Figure 11 depicts the dissociation profiles of oxirane between the size of the AS. Similarly, as for O2 , we also notice discontinu-
d = 1.0 Å and d = 3.0 Å for a selection of ASs. For the DFT embed- ities between d = 2.75 Å and d = 3.0 Å for the DFT embedding
ding scheme, the analysis performed in the supplementary material CAS(2, 6) and the HF embedding CAS(6, 6) calculations due to the
(see Fig. S9) suggests an optimal range-separation parameter of μopt crossing of molecular orbital levels and the consequent changes in
= 6.0. The overall behavior is similar to the one observed for the pre- the composition of the AS.
vious systems, suggesting that the scheme can be safely extended to Table IV summarizes the dissociation energies and NPEs for
the profiles reported in Fig. 11. The values highlighted in bold cor-
responds to the best results achievable with the current implementa-
tion of the embedding schemes. As expected, due to the dimension

TABLE IV. Dissociation energies, ΔE, and NPE values for the dissociation of the
oxirane molecule. ΔE are computed as the difference between the energies at equi-
librium, d = 1.482 Å, and at d = 3.0 Å. The NPE values are given within parentheses.
The results for HF, DFT, FCI, and CCSD are ΔE HF = 562.0 (308.8) mHa, ΔE DFT =
343.9 (97.9) mHa, ΔE FCI = 268.9 (0.0) mHa, and ΔE CCSD = 267.4 (5.4) mHa, respec-
tively. Situations that show an improvement of ΔE and the NPE compared to standard
HF and DFT are highlighted in bold.

CAS μ = 6.0 (NPE) HF-Emb. (NPE) CASSCF (NPE)

(2, 6) 343.1 (256.8) 352.9 (99.8) 345.9 (92.8)

(4, 6) 351.3 (97.8) 318.3 (82.8) 238.7 (112.5)
FIG. 10. Energy diagram of the KS-MOs of oxirane in the STO-3G basis. The solid (6, 6) 328.9 (75.5) 229.1 (93.0) 233.8 (35.1)
lines correspond to occupied MOs, while the dotted lines represent virtual ones. In (8, 6) 316.2 (61.8) 330.1 (64.6) 263.2 (83.3)
red are the orbitals involved in at least one of the ASs considered.

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-9

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

of the system (with 18 valence electrons) and the size of the larger SUPPLEMENTARY MATERIAL
active space [CAS(8, 6)], we cannot expect to achieve results com-
The supplementary material reports on additional HF-
parable with FCI for both embedding schemes. However, we notice
embedding calculations obtained for larger basis sets as well as
a sizable improvement of the HF embedding dissociation energies
additional results obtained with different choices of the AS. Also
[from the HF value of 562.0 mHa to 229.1 mHa in CAS(6, 6)] and a
provided are detailed plots for the selection of the range-separation
smaller, but still significant, correction of about 60% in the DFT case.
parameters μ employed in our calculations and sketches of the MOs
Further studies with more accurate DFT functionals will be needed
for all investigated systems, as well as, the atomic coordinates used
to assess the potential of the DFT embedding in more complex
in the calculations.
systems.

ACKNOWLEDGMENTS
V. CONCLUSIONS The authors thank Valery Weber and Jürg Hutter for useful
In this work, we introduced an embedding scheme that enables discussions as well as Manfred Sigrist who advised M.R. during a
the partitioning of electronic structure calculations into an Active significant part of this work.
Space (AS) subsystem treated with a high level quantum algorithm I.T., P.J.O., and M.R. acknowledge financial support from
and an environment described at the HF or DFT level of theory. the Swiss National Science Foundation (SNF) through Grant No.
In this way, we can restrict the quantum calculations to a critical 200021-179312.
subset of molecular orbitals that can fit on state-of-the-art quan-
tum computers, while the remaining electrons provide the embed- APPENDIX A: HARTREE–FOCK EMBEDDING
ding potential computed using a classical algorithm. Since for most
In this section, we provide more detailed derivations of the HF
chemical processes the quality of the electronic structure predictions
embedding. First, we derive the restricted spin case and generalize
depends on a small set of frontier orbitals, this scheme will allow
the equations for unrestricted spins in the second part of this section.
the solution of interesting quantum chemistry problems where the
AS can be described with a quantum algorithm presenting a favor-
able scaling in the number of active electrons. We demonstrate the 1. Restricted spins
performance of the embedding schemes in the case of the dissoci- We introduce the splitting of the one-electron density matrix,
ation of a few test molecular systems, namely, molecular nitrogen, D = DA + DI , into Eq. (1), one term at a time. The simplest case is
molecular oxygen, water, and oxirane, highlighting the benefits of the one of the one-electron contribution, which becomes
the recursive update of the embedding potential to enhance the
A I
convergence of the ground state calculations. The improvements in ∑ hpq Dpq = ∑ hvq Dvq + ∑ hjq Djq
accuracy obtained in the strongly correlated bond-breaking regime pq vq jq
are significant in all tested systems and for both embedding schemes, = A
∑ huv Duv + 2 ∑ hjj , (A1)
demonstrating the potential of this approach. It is important to uv j
mention that the use of the iterative range-separated DFT embed-
ding requires the tuning of an extra parameter, which cannot be set where we use the fact that a density matrix element vanishes when
a priori. Further investigation is needed to automatize this technique any of its indices correspond to a virtual orbital. We can proceed
for general use in larger molecular systems. Of particular relevance analogously with the two-electron terms as in
are the results obtained for oxirane, which demonstrate the appli-
cability of the proposed quantum embedding scheme to complex 1 1 1 1
∑ gpqrs dpqrs = ∑ gpqjs dpqjs + ∑ gjqus djqus + ∑ gvqus dvqus .
organic molecules. 2 pqrs 2 pqjs 2 jqus 2 vqus
Improvements of the proposed embedding scheme can be
obtained through the combination of the iterative update of the (A2)
embedding potential with the simultaneous optimization of the
We can now express the two-electron density matrices in terms of
active orbitals, as done, for instance, in the multiconfigurational
one-electron ones,
self-consistent field (MCSCF) approach. We believe that the pro-
posed HF and DFT embedding schemes will provide a fundamen-
dpqjs = Dpq Djs − δqj Dps
tal framework for the scaling up of quantum electronic structure
calculations to large molecular systems with an arbitrary number = 2δjs Dpq − δqj δsq Dpq
of electrons (i.e., as many as HF or DFT calculations can deal = (2δjs − δqj δsq )Dpq (A3a)
with).
The possibility of partitioning the solution of the electronic and
structure problem into an active component (defined by the AS)
treated by means of a quantum computing algorithm and an inert djqus = dusjq = Dus Djq − δsj Duq
environment component solved at the HF or DFT level of theory (as = 2δjq Dus − δsj Duq
presented here and in other previous studies) will make it possible
to use quantum computers in the solution of important problems in = 2δqj δsq Duq − δsj Duq
physics, chemistry, biology, and medicine. = (2δqj δsq − δsj )Duq , (A3b)

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-10

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

where we omit the superscripts I and A for brevity. These expres- 1 nσ nσ σσ 1 nσ nσ nσ nσ

A,σ
∑ ∑ gpqrs dpqrs = ∑ ∑(gkkjj − gkjjk ) + ∑ ∑(gjjuv − gjvuj )Duv
sions can then be inserted into Eq. (A2) to obtain 2 pq rs 2 k j j uv
1 1 A 1 nσ nσ
∑ gpqrs dpqrs = ∑ gpqjj Dpq − ∑ gpjjq Dpq + ∑ gjjuq Duq + A,σσ
∑ ∑ guvxy duvxy . (A9)
2 pqrs pqj 2 pqj jqu 2 uv xy
1 A 1 A
− ∑ gjquj Duq + ∑ guvxy duvxy . (A4) The second case of opposite spins, σ ≠ τ, behaves slightly differ-
2 jqu 2 uvxy
ent due to the differing two-electron excitation operator, Eq. (A7d).
Upon inspection, it becomes clear that the first two sums only yield Thus, the expression of the two-electron density matrices in terms
non-zero contributions when p = q = k or when p = u and q = v. In the of one-electron ones analogous to Eq. (A1) becomes
latter case, this causes the sums to coincide with the third and fourth στ
dpqjs = Dσpq Dτjs = δjs Dσpq (A10a)
terms. These observations allow us to simplify Eq. (A4) to become
and
1 1 1 στ
A
∑ gpqrs dpqrs = 2 ∑(gkkjj − gkjjk ) + 2 ∑(gjjuv − gjvuj )Duv djqus = Dσjq Dτus = δjq Dτus , (A10b)
2 pqrs kj 2 jvu 2
which leads to the first expression of the two-electron contribution
1 A of opposite spins,
+ ∑ guvxy duvxy . (A5)
2 uvxy 1 nσ nτ 1 nσ nτ 1 nσ nτ
στ σ A,τ
∑ ∑ gpqrs dpqrs = ∑ ∑ gpqjj Dpq + ∑ ∑ gjjus Dus
Finally, we can substitute Eqs. (A1) and (A5) into Eq. (1), yielding 2 pq rs 2 pq j 2 j us
⎡
⎢
⎤
⎥ A 1 nσ nτ A,στ
E = ∑[2hjj + ∑(2gkkjj − gkjjk )] + ∑⎢ huv + ∑ (2g jjuv − g jvuj )⎥Duv + ∑ ∑ gvqus dvqus . (A11)
j
⎢
uv ⎢ j
⎥
⎥ 2 vq us
k ⎣ ⎦
1 A
Using the same arguments as before, we can once again simplify this
+ ∑ guvxy duvxy . (A6) equation to become
2 uvxy
1 nσ nτ στ 1 nσ nτ 1 nσ nτ A,σ
This equation simplifies to Eq. (2) with the use of the inactive Fock ∑ ∑ gpqrs dpqrs = ∑ ∑ gkkjj + ∑ ∑ guvjj Duv
operator, Eq. (3), and the inactive energy, Eq. (4). 2 pq rs 2 k j 2 uv j
1 nσ nτ A,τ 1 nσ nτ A,στ
2. Unrestricted spins + ∑ ∑ gjjuv Duv + ∑ ∑ guvxy duvxy . (A12)
2 j uv 2 uv xy
For unrestricted spins, we have to remove the implicit sum-
mation over the spin state, σ, and calculate with the one- and Finally, we can obtain the two-electron contributions for unre-
two-electron density matrices for each spin separately, stricted spins by combining Eqs. (A9) and (A12) into
n
Dσpq = ⟨Ψ∣Êpq
σ
∀σ ∈ {α, β}, (A7a) 1 1 nα nα β
∣Ψ⟩ ∑ gpqrs dpqrs = ∑[∑(gkkjj − gkjjk ) + ∑ gkkjj ]
2 pqrs 2 j k k
στ
dpqrs = ⟨Ψ∣êστ
pqrs ∣Ψ⟩ ∀σ, τ ∈ {α, β}, (A7b) n n nα
1 β β
+ ∑[∑(gkkjj − gkjjk ) + ∑ gkkjj ]
2 j k
σ
Êpq = â†pσ âqσ ∀σ ∈ {α, β}, (A7c) k

nα ⎡
⎢ n α nβ ⎤
⎥ A,α
σ σ σ + ∑⎢ ⎢∑(guvjj − gujjv ) + ∑ guvjj ⎥Duv
Êpq Êrs − δqr Êps , σ=τ ⎥
êστ
pqrs = { σ τ ∀σ, τ ∈ {α, β}. uv ⎢
⎣ j j ⎥
⎦
Êpq Êrs , σ≠τ
nβ ⎡ nβ n α
⎤
⎢ ⎥ A,β
(A7d) + ∑⎢ ⎢∑(guvjj − gujjv ) + ∑ guvjj ⎥Duv
⎥
uv ⎢ j j ⎥
Thus, in contrast to the closed shell description with restricted spins ⎣ ⎦
in Appendix A 1, we now have to keep track of the spin state where 1 A
+ ∑ guvxy duvxy . (A13)
each index iterates over. This is indicated by the additional super- 2 uvxy
scripts, σ and τ, and summation labels, nσ and nτ . For brevity, we
In full analogy to the restricted spin case, this allows us to define
refrain from explicitly denoting that σ, τ ∈ {α, β} for the remainder
the inactive Fock operator and energy as
of this derivation.
nσ nτ
Analogous to Eq. (A1), the one-electron contribution can be I,σ
Fpq = hpq + ∑(giipq − giqpi ) + ∑ giipq (A14)
written per spin state as i i
nσ nσ nσ
σ A,σ and
∑ hpq Dpq = ∑ huv Duv + ∑ hjj . (A8)
pq uv j 1 β
EI = I,σ
∑ ∑ hjj + Fjj
For the two-electron contributions, we have to differentiate between 2 σ=α j
two cases. In the first case, the spins of both electrons are aligned,
σ = τ, and the resulting equation can be derived in full analogy to 1 β I,σ I,σ
= ∑ ∑(hij + Fij )Dij . (A15)
Eq. (A5), 2 σ=α ij

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-11

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

APPENDIX B: DENSITY FUNCTIONAL THEORY 1

E = EI,LR + Exc
SR (i)
[ρ ] + I,SR I
∑ j Dij
EMBEDDING 2 ij ij
In this appendix, we provide more detailed steps deriving the 1 A,(i),SR A,(i) SR A,(i)
− ∑ juv Duv − ∑ νxc,uv [ρ(i) ]Duv
embedding equations of the iterative DFT embedding scheme. In 2 uv uv
doing so, we follow the work of Hedegård et al.58 rather closely. We A,(i+1) A,(i),SR A,(i+1)
focus on the steps necessary to arrive at the final form of the total + EA,LR + ∑ jI,SR
uv Duv + ∑ juv Duv
uv uv
electronic energy, Eq. (15), after the introduction of the linear model,
SR A,(i+1)
Eq. (13). + ∑ νxc,uv [ρ(i) ]Duv , (B5)
(i) A,(i) uv
We start by noting that ΔDpq = ΔDuv since the inactive
I
part of the density matrix, Dij , is constant by definition. Thus, we where we have re-ordered the terms such that the upper two lines
can express the total electronic energy after exploiting inherent contain all the inactive terms and the lower lines contain all the
properties of the one-electron density matrices as active ones. After insertion of the expressions for the LR energy
contributions, EI and EA , according to Eqs. (4) and (2), respec-
E = EI,LR + EA,(i+1),LR + Ecoul+xc
SR
[ρ(i) ] tively, we arrive at the final expression of the total electronic energy,
(i),SR SR A,(i) Eq. (15).
+ ∑(juv + νxc,uv [ρ(i) ])ΔDuv . (B1) The extension of these equations to unrestricted spins is
uv
similarly straight forward as in the case of the HF embedding
To ease the implementation of Eq. (B1), we can rewrite the equation (cf. Appendix A 2).
and group its terms into active and inactive ones. To do so, we start
by rewriting the Coulomb part of the third term by expanding D(i) APPENDIX C: QISKIT’S EXCITATION OPERATOR
= DI + DA,(i) twice, ORDER IN THE q-UCCSD ANSATZ
1 In QISKIT, the order of the excitation operators of the q-UCCSD
J SR [ρ(i) ] =
(i) SR (i)
∑ D gpqrs Drs Ansatz is always constructed in a reproducible manner. For the pur-
2 pqrs pq
poses of this discussion, we assume that the number of electrons is
1 (i) ⎛ SR I SR A,(i) ⎞ stored as a pair in nelec = (nalpha, nbeta) and the number of
= ∑D ∑ gpqij Dij + ∑ gpquv Duv spin orbitals accordingly in the pair norbs. Furthermore, we assume
2 pq pq ⎝ ij uv ⎠
a block-ordering of the spin orbitals in the qubit register, meaning
1⎛ I SR I A,(i) SR I that the first half of the qubits maps to α-spin orbitals and the second
= ∑ D g Dij + ∑ Dxy gxyij Dij half maps to β-spin orbitals.
2 ⎝ klij kl klij xyij
With these assumptions in place, we can express QISKIT’s order
A,(i) A,(i) A,(i) ⎞ of excitation operators, as presented in Fig. 12. First, all α-spin
+ ∑ DIkl gkluv
SR SR
Duv + ∑ Dxy gxyuv Duv . (B2)
kluv xyuv ⎠

We can proceed analogously with the Coulomb part of the fourth

term,
(i),SR A,(i) SR (i) A,(i+1) A,(i)
∑ juv ΔDuv = ∑ Dpq gpquv (Duv − Duv )
uv pquv
A,(i+1) A,(i)
= ∑ DIij gijuv
SR
Duv − DIij gijuv
SR
Duv
ijuv
SR A,(i) A,(i+1) A,(i)
SR A,(i)
+ ∑ Dxy gxyuv Duv − Dxy gxyuv Duv .
xyuv
(B3)

By gathering and canceling matching terms of Eqs. (B2) and (B3)

through the use of the symmetry, g pqrs = g rspq , we arrive at the final
expression of the SR Coulomb contributions,

A,(i) 1 I,SR I 1 A,(i),SR A,(i)

J SR [ρ(i) ] + ∑ juv
(i),SR
ΔDuv = ∑ jij Dij − ∑ juv Duv
uv 2 ij 2 uv
A,(i+1) A,(i),SR A,(i+1)
+ ∑ jI,SR
uv Duv + ∑ juv Duv .
uv uv
(B4)
FIG. 12. A Python code outlining the order of the excitation operators of the
q-UCCSD Ansatz as constructed in QISKIT.
Inserting Eq. (B4) into Eq. (B1) finally leads to

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-12

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

14
D. Wecker, M. B. Hastings, and M. Troyer, Phys. Rev. A 92, 042303 (2015).
15
A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love,
A. Aspuru-Guzik, and J. L. O’Brien, Nat. Commun. 5, 4213 (2014).
16
J. Čížek, J. Chem. Phys. 45, 4256 (1966).
17
H. G. Kümmel, “A biography of the coupled cluster method,” in Recent Progress
in Many-Body Theories—Proceedings of the 11th International Conference, edited
by R. F. Bishop, T. Brandes, K. A. Gernoth, N. R. Walet, and Y. Xian (World
Scientific Publishing, Singapore, 2002), pp. 334–348.
18
R. J. Bartlett and M. Musiał, Rev. Mod. Phys. 79, 291 (2007).
19
P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero, J. McClean, R. Barends,
J. Kelly, P. Roushan, A. Tranter, N. Ding, B. Campbell, Y. Chen, Z. Chen,
B. Chiaro, A. Dunsworth, A. G. Fowler, E. Jeffrey, E. Lucero, A. Megrant, J.
Y. Mutus, M. Neeley, C. Neill, C. Quintana, D. Sank, A. Vainsencher, J. Wen-
ner, T. C. White, P. V. Coveney, P. J. Love, H. Neven, A. Aspuru-Guzik, and J.
M. Martinis, Phys. Rev. X 6, 031007 (2016).
20
J. Romero, R. Babbush, J. R. McClean, C. Hempel, P. J. Love, and A. Aspuru-
Guzik, Quantum Sci. Technol. 4, 014008 (2018).
21
FIG. 13. Illustration of the ordering of the q-UCCSD excitations in QISKIT. (Top) P. K. Barkoutsos, J. F. Gonthier, I. Sokolov, N. Moll, G. Salis, A. Fuhrer,
Model system composed of four electrons in six orbitals [nelec = (2, 2) and norbs M. Ganzhorn, D. J. Egger, M. Troyer, A. Mezzacapo, S. Filipp, and I. Tavernelli,
= (3, 3)]. (Bottom) Corresponding ordered list of excitations, as produced by the Phys. Rev. A 98, 022322 (2018).
22
code in Fig. 12. N. Moll, P. Barkoutsos, L. S. Bishop, J. M. Chow, A. Cross, D. J. Egger, S. Filipp,
A. Fuhrer, J. M. Gambetta, M. Ganzhorn, A. Kandala, A. Mezzacapo, P. Müller,
W. Riess, G. Salis, J. Smolin, I. Tavernelli, and K. Temme, Quantum Sci. Technol.
3, 030503 (2018).
23
single-excitations are applied, followed by all β-spin ones. Then, M. Kühn, S. Zanker, P. Deglmann, M. Marthaler, and H. Weiss, J. Chem. Theory
all double-excitations are applied where the nested for loops are Comput. 15, 4764 (2019).
24
ordered such that the β-spin orbitals are iterated on the inner levels. N. P. Bauman, E. J. Bylaska, S. Krishnamoorthy, G. H. Low, N. Wiebe, C. E.
To make this implementation more concrete, we present the out- Granade, M. Roetteler, M. Troyer, and K. Kowalski, J. Chem. Phys. 151, 014107
(2019).
put of the described code for the example of nelec = (2, 2) and 25
F. A. Evangelista, G. K.-L. Chan, and G. E. Scuseria, J. Chem. Phys. 151, 244112
norbs = (3, 3) in Fig. 13. (2019).
26
W. Mizukami, K. Mitarai, Y. O. Nakagawa, T. Yamamoto, T. Yan, and
Y-y. Ohnishi, “Orbital optimized unitary coupled cluster theory for quantum
DATA AVAILABILITY computer,” Phys. Rev. Research 2, 033421 (2019).
27
The data that support the findings of this study are available A. F. Izmaylov, M. Díaz-Tinoco, and R. A. Lang, Phys. Chem. Chem. Phys. 22,
from the corresponding author upon reasonable request. 12980 (2020).
28
A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and
J. M. Gambetta, Nature 549, 242 (2017).
REFERENCES 29
M. Ganzhorn, D. Egger, P. Barkoutsos, P. Ollitrault, G. Salis, N. Moll, M. Roth,
1 A. Fuhrer, P. Mueller, S. Woerner, I. Tavernelli, and S. Filipp, Phys. Rev. Appl. 11,
J. D. Whitfield, J. Biamonte, and A. Aspuru-Guzik, Mol. Phys. 109, 735
044092 (2019).
(2011). 30
2 I. G. Ryabinkin, T.-C. Yen, S. N. Genin, and A. F. Izmaylov, J. Chem. Theory
T. Helgaker, S. Coriani, P. Jørgensen, K. Kristensen, J. Olsen, and K. Ruud, Chem.
Comput. 14, 6317 (2018).
Rev. 112, 543 (2012). 31
3
K. D. Vogiatzis, M. V. Polynski, J. K. Kirkland, J. Townsend, A. Hashemi, C. Liu, I. G. Ryabinkin, R. A. Lang, S. N. Genin, and A. F. Izmaylov, J. Chem. Theory
and E. A. Pidko, Chem. Rev. 119, 2453 (2018). Comput. 16, 1055 (2020).
32
4
M. Reiher, N. Wiebe, K. M. Svore, D. Wecker, and M. Troyer, Proc. Natl. Acad. C. Hempel, C. Maier, J. Romero, J. McClean, T. Monz, H. Shen, P. Jurcevic, B. P.
Sci. U. S. A. 114, 7555 (2017). Lanyon, P. Love, R. Babbush, A. Aspuru-Guzik, R. Blatt, and C. F. Roos, Phys. Rev.
5 X 8, 031022 (2018).
M. Boero, M. Parrinello, and K. Terakura, J. Am. Chem. Soc. 120, 2746 33
(1998). R. Sagastizabal, X. Bonet-Monroig, M. Singh, M. A. Rol, C. C. Bultink, X. Fu,
6
A. Aspuru-Guzik, R. Lindh, and M. Reiher, ACS Cent. Sci. 4, 144 (2018). C. H. Price, V. P. Ostroukh, N. Muthusubramanian, A. Bruno, M. Beekman,
7 N. Haider, T. E. O’Brien, and L. DiCarlo, Phys. Rev. A 100, 010302 (2019).
Y. Cao, J. Romero, and A. Aspuru-Guzik, IBM J. Res. Dev. 62(6), 1 (2018). 34
8 F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, S. Boixo,
J.-M. Reiner, F. Wilhelm-Mauch, G. Schön, and M. Marthaler, Quantum Sci.
M. Broughton, B. B. Buckley, D. A. Buell, B. Burkett, N. Bushnell, Y. Chen,
Technol. 4, 035005 (2019).
9 Z. Chen, B. Chiaro, R. Collins, W. Courtney, S. Demura, A. Dunsworth,
I. O. Sokolov, P. K. Barkoutsos, P. J. Ollitrault, D. Greenberg, J. Rice, M. Pistoia,
D. Eppens, E. Farhi, A. Fowler, B. Foxen, C. Gidney, M. Giustina, R. Graff,
and I. Tavernelli, J. Chem. Phys. 152, 124107 (2020).
10 S. Habegger, M. P. Harrigan, A. Ho, S. Hong, T. Huang, W. J. Huggins, L. Ioffe,
C. Kokail, C. Maier, R. van Bijnen, T. Brydges, M. K. Joshi, P. Jurcevic, C. A. S. V. Isakov, E. Jeffrey, Z. Jiang, C. Jones, D. Kafri, K. Kechedzhi, J. Kelly, S. Kim,
Muschik, P. Silvi, R. Blatt, C. F. Roos, and P. Zoller, Nature 569, 355 (2019). P. V. Klimov, A. Korotkov, F. Kostritsa, D. Landhuis, P. Laptev, M. Lindmark,
11
S. V. Mathis, G. Mazzola, and I. Tavernelli, “Toward scalable simulations E. Lucero, O. Martin, J. M. Martinis, J. R. McClean, M. McEwen, A. Megrant,
of lattice gauge theories on quantum computers,” Phys. Rev. D 102, 094501 X. Mi, M. Mohseni, W. Mruczkiewicz, J. Mutus, O. Naaman, M. Neeley, C. Neill,
(2020). H. Neven, M. Y. Niu, T. E. O’Brien, E. Ostby, A. Petukhov, H. Putterman, C. Quin-
12
R. Babbush, N. Wiebe, J. McClean, J. McClain, H. Neven, and G. K.-L. Chan, tana, P. Roushan, N. C. Rubin, D. Sank, K. J. Satzinger, V. Smelyanskiy, D. Strain,
Phys. Rev. X 8, 011044 (2018). K. J. Sung, M. Szalay, T. Y. Takeshita, A. Vainsencher, T. White, N. Wiebe, Z. J.
13
See https://quantum-computing.ibm.com/ for IBM, IBM Quantum Experience; Yao, P. Yeh, and A. Zalcman, “Hartree-Fock on a superconducting qubit quantum
accessed 16 March 2020. computer,” Science 369, 1084 (2020).

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-13

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

35
P. J. Ollitrault, A. Kandala, C.-F. Chen, P. K. Barkoutsos, A. Mezzacapo, M. Pis- P. Eendebak, M. Everitt, I. F. Sertage, A. Frisch, A. Fuhrer, J. Gacon, J. Gam-
toia, S. Sheldon, S. Woerner, J. M. Gambetta, and I. Tavernelli, Phys. Rev. Res. 2, betta, B. G. Gago, J. Gomez-Mosquera, D. Greenberg, I. Hamamura, V. Havlicek,
043140 (2020). J. Hellmers, L. Herok, H. Horii, S. Hu, T. Imamichi, T. Itoko, A. Javadi-Abhari,
36 N. Kanazawa, A. Karazeev, K. Krsulich, P. Liu, Y. Luh, Y. Maeng, M. Mar-
O. Higgott, D. Wang, and S. Brierley, Quantum 3, 156 (2019).
37
R. M. Parrish, E. G. Hohenstein, P. L. McMahon, and T. J. Martínez, Phys. Rev. ques, F. J. Martin-Fernandez, D. T. McClure, D. McKay, S. Meesala, A. Mezza-
Lett. 122, 230401 (2019). capo, N. Moll, D. M. Rodriguez, G. Nannicini, P. Nation, P. J. Ollitrault, L. J.
38
K. M. Nakanishi, K. Mitarai, and K. Fujii, Phys. Rev. Res. 1, 033062 (2019). O’Riordan, H. Paik, J. Perez, A. Phan, M. Pistoia, V. Prutyanov, M. Reuter, J. Rice,
39 A. R. Davila, M. Rossmannek, R. H. P. Rudy, M. Ryu, N. Sathaye, C. Schn-
I. G. Ryabinkin, S. N. Genin, and A. F. Izmaylov, J. Chem. Theory Comput. 15,
abel, E. Schoute, K. Setia, Y. Shi, A. Silva, Y. Siraichi, S. Sivarajah, J. A. Smolin,
249 (2018).
40 M. Soeken, I. Sokolov, H. Takahashi, I. Tavernelli, C. Taylor, P. Taylour, K. Tra-
J. R. McClean, M. E. Kimchi-Schwartz, J. Carter, and W. A. de Jong, Phys. Rev.
bing, M. Treinish, W. Turner, D. Vogt-Lee, C. Vuillot, J. A. Wildstrom, J. Wil-
A 95, 042308 (2017).
41
son, E. Winston, C. Wood, S. Wood, S. Wörner, I. Y. Akhalwaya, and C. Zoufal
N. H. Stair, R. Huang, and F. A. Evangelista, J. Chem. Theory Comput. 16, 2236 (2019). “Qiskit: An open-source framework for quantum computing,” Zenodo.
(2020). https://doi.org/10.5281/zenodo.2562111.
42
R. Santagati, J. Wang, A. A. Gentile, S. Paesani, N. Wiebe, J. R. McClean, 57
S. B. Bravyi and A. Y. Kitaev, Ann. Phys. 298, 210 (2002).
S. Morley-Short, P. J. Shadbolt, D. Bonneau, J. W. Silverstone et al., Sci. Adv. 4, 58
E. D. Hedegård, S. Knecht, J. S. Kielberg, H. J. A. Jensen, and M. Reiher, J. Chem.
eaap9646 (2018).
43
Phys. 142, 224108 (2015).
T. Jones, S. Endo, S. McArdle, X. Yuan, and S. C. Benjamin, Phys. Rev. A 99, 59
See https://github.com/Qiskit/qiskit-nature/ for Qiskit, Qiskit Nature; accessed
062304 (2019).
44
29 January 2021.
J. Tilly, G. Jones, H. Chen, L. Wossnig, and E. Grant, “Computation of molecular 60
Q. Sun, J. Yang, and G. K.-L. Chan, Chem. Phys. Lett. 683, 291 (2017).
excited states on IBMQ using a discriminative variational quantum eigensolver,” 61
S. Bravyi, J. M. Gambetta, A. Mezzacapo, and K. Temme, “Tapering off qubits
Phys. Rev. A 102, 062425 (2020).
45 to simulate fermionic Hamiltonians,” arXiv:1701.08213v1 (2017).
B. O. Roos, P. R. Taylor, and P. E. M. Sigbahn, Chem. Phys. 48, 157 (1980). 62
46 J. L. Morales and J. Nocedal, ACM Trans. Math. Software 38, 1 (2011).
T. Helgaker, P. Jèrgensen, and J. Olsen, Molecular Electronic-Structure Theory 63
J. Toulouse, A. Savin, and H.-J. Flad, Int. J. Quantum Chem. 100, 1047
(John Wiley & Sons, Ltd., 2000).
47 (2004).
B. Bauer, D. Wecker, A. J. Millis, M. B. Hastings, and M. Troyer, Phys. Rev. X 6, 64
031045 (2016). S. Paziani, S. Moroni, P. Gori-Giorgi, and G. B. Bachelet, Phys. Rev. B 73, 155111
48 (2006).
N. C. Rubin, “A hybrid classical/quantum approach for large-scale studies of 65
quantum systems with density matrix embedding theory,” arXiv:1610.06910v2 U. Ekström, L. Visscher, R. Bast, A. J. Thorvaldsen, and K. Ruud, J. Chem.
(2016). Theory Comput. 6, 1971 (2010).
66
49
A. Georges and G. Kotliar, Phys. Rev. B 45, 6479 (1992). J. Heyd, G. E. Scuseria, and M. Ernzerhof, J. Chem. Phys. 118, 8207 (2003).
67
50
G. Knizia and G. K.-L. Chan, Phys. Rev. Lett. 109, 186404 (2012). E. Fromager, J. Toulouse, and H. J. A. Jensen, J. Chem. Phys. 126, 074111
51 (2007).
X. Wu, M. Lindsey, T. Zhou, Y. Tong, and L. Lin, “Enhancing robustness and 68
efficiency of density matrix embedding theory via semidefinite programming and W. J. Hehre, R. F. Stewart, and J. A. Pople, J. Chem. Phys. 51, 2657 (1969).
69
local correlation potential fitting,” Phys. Rev. B 102, 085123 (2020). P. C. Hariharan and J. A. Pople, Theor. Chim. Acta 28, 213 (1973).
70
52
H. Ma, M. Govoni, and G. Galli, npj Comput. Mater. 6, 85 (2020). T. H. Dunning, J. Chem. Phys. 90, 1007 (1989).
71
53
A. Savin and H.-J. Flad, Int. J. Quantum Chem. 56, 327 (1995). V. Veryazov, P. Å. Malmqvist, and B. O. Roos, Int. J. Quantum Chem. 111, 3329
54
T. Takeshita, N. C. Rubin, Z. Jiang, E. Lee, R. Babbush, and J. R. McClean, Phys. (2011).
72
Rev. X 10, 011004 (2020). C. J. Stein and M. Reiher, J. Chem. Theory Comput. 12, 1760 (2016).
73
55
M. Urbanek, D. Camps, R. V. Beeumen, and W. A. de Jong, “Chemistry on C. J. Stein and M. Reiher, J. Comput. Chem. 40, 2216 (2019).
74
quantum computers with virtual quantum subspace expansion,” J. Chem. Theory F. M. Faulstich, M. Máté, A. Laestadius, M. A. Csirik, L. Veis, A. Antalik,
Comput. 16, 5425–5431 (2020). J. Brabec, R. Schneider, J. Pittner, S. Kvaal, and Ö. Legeza, J. Chem. Theory
56
G. Aleksandrowicz, T. Alexander, P. Barkoutsos, L. Bello, Y. Ben-Haim, Comput. 15, 2206 (2019).
75
D. Bucher, F. J. Cabrera-Hernandez, J. Carballo-Franquis, A. Chen, C.-F. Chen, E. R. Sayfutyarova, Q. Sun, G. K.-L. Chan, and G. Knizia, J. Chem. Theory
J. M. Chow, A. D. Corcoles-Gonzales, A. J. Cross, A. Cross, J. Cruz-Benito, Comput. 13, 4063 (2017).
76
C. Culver, S. D. L. P. Gonzalez, E. D. L. Torre, D. Ding, E. Dumitrescu, I. Duran, X. Li and J. Paldus, J. Chem. Phys. 103, 1024 (1995).

J. Chem. Phys. 154, 114105 (2021); doi: 10.1063/5.0029536 154, 114105-14

Published under license by AIP Publishing