PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

Parallel quantum chemistry on noisy intermediate-scale quantum computers

Robert Schade ,1 Carsten Bauer ,1 Konstantin Tamoev,2 Lukas Mazur ,1 Christian Plessl ,1,3 and Thomas D. Kühne2
1
Paderborn Center for Parallel Computing, Paderborn University, 33098 Paderborn, Germany
2
Dynamics of Condensed Matter and Institute for Sustainable Systems Design, Chair of Theoretical Chemistry,
Paderborn University, 33098 Paderborn, Germany
3
Department for Computer Science, Paderborn University, 33098 Paderborn, Germany

(Received 9 February 2022; accepted 22 July 2022; published 29 August 2022)

A parallel hybrid quantum-classical algorithm for the solution of the quantum-chemical ground-state energy
problem on gate-based quantum computers is presented. This approach is based on the reduced density-matrix
functional theory (RDMFT) formulation of the electronic structure problem. For that purpose, the density-matrix
functional of the full system is decomposed into an indirectly coupled sum of density-matrix functionals for
all its subsystems using the adaptive cluster approximation to RDMFT. The approximations involved in the
decomposition and the adaptive cluster approximation itself can be systematically converged to the exact result.
The solutions for the density-matrix functionals of the effective subsystems involves a constrained minimization
over many-particle states that are approximated by parametrized trial states on the quantum computer similarly
to the variational quantum eigensolver. The independence of the density-matrix functionals of the effective
subsystems introduces a new level of parallelization and allows for the computational treatment of much larger
molecules on a quantum computer with a given qubit count. In addition, for the proposed algorithm techniques
are presented to reduce the qubit count, the number of quantum programs, as well as its depth. The evaluation
of a density-matrix functional as the essential part of our approach is demonstrated for Hubbard-like systems on
IBM quantum computers based on superconducting transmon qubits.

DOI: 10.1103/PhysRevResearch.4.033160

I. INTRODUCTION schemes to effectively realize quantum many-particle states

on quantum hardware have been suggested [7,9,10].
Quantum computers have recently emerged as a power-
In this paper, we assess the potential of QC for
ful resource for solving computational problems that have
ab initio quantum chemistry, in which the exponential com-
eluded an efficient treatment on classical hardware due to
plexity lies in the quantum mechanical treatment of electrons
their computational complexity. A particularly promising ap-
in the calculation of the total energies and nuclear forces
plication for quantum computing (QC) is quantum chemistry,
of many-fermion systems. Within the framework of the re-
whose focus is on solving the electronic structure problem
duced density-matrix functional theory (RDMFT) [11–14] we
represented by different interacting many-fermion Hamilto-
propose a hybrid quantum-classical algorithm for comput-
nians. Here, the utilization of traditional numerical methods
ing the reduced density-matrix functional (RDMF) in which
is significantly hampered by its computational cost that scales
the quantum mechanical expectation values of the reduced
exponentially or, for quantum Monte Carlo methods, by the
density matrix are evaluated on a quantum computer. This
infamous fermion sign problem [1,2]. Consequently, quantum
makes the approach very similar to the VQE in the sense
chemistry simulations regularly reach and push the limits of
that a parametrized trial state for the many-particle state is
the most powerful high-performance computers.
prepared on the quantum computer and the parameters of this
Recently, significant research efforts have been aimed at
state are modified with a minimization algorithm running on
developing novel quantum algorithms that allows one to
a classical computer till the measured total energy is mini-
solve electronic structure problems more efficiently. Most
mal. While the approach proposed here introduces additional
notably, quantum phase-estimation methods [3,4] and vari-
equality constraints on the minimum compared to the VQE,
ational quantum eigensolvers (VQE) [5–8] for computing
the formulation allows for approximations like the adaptive
the (ground-state) energy of atoms and molecules have been
cluster approximation (ACA) [15].
put forward. Complementing these algorithmic advances,
The proposed combination of an RDMFT-based algorithm
and the ACA in RDMFT approximately decomposes the
RDMF of the full system into a sum of RDMFs of smaller
effective systems. This on the one hand drastically reduces the
Published by the American Physical Society under the terms of the required qubit count, but also makes the problem inherently
Creative Commons Attribution 4.0 International license. Further parallelizable. Thus, much larger molecules can be treated on
distribution of this work must maintain attribution to the author(s) a quantum computer with a given qubit count compared to
and the published article’s title, journal citation, and DOI. a traditional VQE. However, due to its VQE-like nature, the

2643-1564/2022/4(3)/033160(15) 033160-1 Published by the American Physical Society

ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

proposed algorithm inherits the noise tolerance and suitability but not on the one-particle Hamiltonian ĥ that includes the
for near-term quantum computers. external potential.
The structure of the paper is as follows. First, we introduce Several approximations of the RDMF have been developed
the employed RDMFT-based approach in Sec. II. Afterwards, and even simple variants suffice to describe strong elec-
in Sec. III, we introduce our hybrid quantum-classical algo- tronic correlations beyond the abilities of local or semilocal
rithm for computing the RDMF on quantum hardware. The exchange-correlation functionals [16,17]. However, promis-
subsequent sections are then concerned with the efficient im- ing functionals have also shown pathological behaviours [18].
plementation of this algorithm on noisy intermediate-scale It is therefore worthwhile to take a step back to the definition
quantum (NISQ) devices. Concretely, we discuss various pos- of the RDMF [13,19],
sibilities for reducing the number of qubit operations (Sec. IV) 
and the number of quantum programs (Sec. V). Lastly, we F Ŵ [ρ (1) ] = min  Pi i |Ŵ |i  (4)
{Pi ,|i }→ρ (1) , Pi =1
showcase exemplary results for the evaluation of the RDMF i
i
of the half-filled Hubbard chain as obtained in noise-free
as the constrained minimum over an ensemble of fermionic
quantum simulation,s as well as genuine simulations on NISQ
many-particle wave functions |i  with ensemble probabili-
hardware by IBM in Sec. VI.
ties Pi . Beyond normalization, the major requirement is that
the ensemble {Pi , |i } must correspond to the given one-
II. REDUCED DENSITY-MATRIX FUNCTIONAL THEORY particle reduced density matrix ρ (1) via
A. Theoretical framework (1)

ρα,β = Pi i |ĉβ† ĉα |i . (5)
RDMFT for fermionic systems uses the one-particle re- i
duced density matrix, i.e.,
Focusing on the case of a nondegenerate ground state and zero
(1)
ρα,β = ĉβ† ĉα , (1) electron temperature, the minimization can be simplified to
only a single many-particle wave function
with the fermionic annihilation operator ĉα for electrons in
the one-particle basis state with index α and creation operator F Ŵ [ρ (1) ] = min |Ŵ |, (6)
ĉβ† for electrons in the one-particle basis state with index β, as |→ρ (1)
the basic quantity. Assuming a finite-dimensional one-particle
with the constraints
basis of Nχ states, i.e., α, β ∈ {1, ..., Nχ }, the one-particle
reduced density matrix ρ (1) ∈ C Nχ ×Nχ is hermitian. As such, (1)
ρα,β = |ĉβ† ĉα |. (7)
RDMFT can be seen as an extension to density functional the-
ory (DFT) in the sense that beyond the electron density, which Let us note in passing that the generalization of RDMFT to
(1)
corresponds to the diagonal elements of ρα,β , all elements finite temperatures is relatively straightforward [11,20].
of the one-particle reduced density matrix are considered. For an efficient minimization of the total energy in Eq. (3),
It is suitable for the description of strong local electronic the derivatives of the RDMF F Ŵ [ρ (1) ] with respect to the
(1)
correlations in solids and molecules because the one-particle elements of the one-particle reduced density matrix ρα,β , are
reduced density matrix explicitly contains information about required. Within our hybrid classical-quantum algorithm—
the orbital occupancies that are essential for the description of to be introduced in Sec. III—these derivatives are readily
electronic correlations. Similar to DFT, RDMFT is an exact available as Lagrange multipliers during the constrained min-
theory if the underlying functionals are not approximated. imization procedure.
Given a many-particle Hamiltonian of an N-particle system Thus far, no approximations have been introduced and the
at zero temperature, constrained minimization in Eq. (4) still entails the exponen-
tial complexity of the many-particle problem. However, as we
Ĥ = ĥ + Ŵ , (2) will show in the following section, the minimization problem
which is composed of a noninteracting (single-particle) required for the RDMF is well suited for quantum computers
 and can be solved via a hybrid quantum-classical algorithm
operator ĥ = α,β hα,β ĉα† ĉβ and a two-particle interaction
 on NISQ quantum hardware.
operator Ŵ = 21 α,β,γ ,δ Uα,β,δ,γ ĉα† ĉβ† ĉγ ĉδ , the ground-state
energy EN (ĥ + Ŵ ) of the system can be expressed as [11,13]
B. RDMFT-based ab inito molecular dynamics simulations
EN (ĥ + Ŵ ) = min Tr(ρ h) + F [ρ ].
(1) Ŵ (1)
The overall goal of this method is to perform ab initio
ρ (1) ,0ρ (1) 1,Tr(ρ (1) )=N
molecular dynamics simulations. Hence, analytical nuclear
(3)
forces are required that can be evaluated within RDMFT as
Here, the condition 0  ρ (1)  1 is shorthand for the
ensemble-representability of the one-particle reduced density dEN
matrix that is, its eigenvalues fi (occupations) must fulfill Fi = − (8a)
dRi
0  fi  1. This way, the exponential many-particle com-
d   (1)   (1) 
plexity of the fermionic problem is absorbed into the RDMF =− Tr ρmin h + F Ŵ ρmin , (8b)
F Ŵ [ρ (1) ], which is the analogon of the exchange-correlation dRi
functional in DFT. Since F Ŵ [ρ (1) ] is universal, it only depends with the position Ri of atom i and the one-particle reduced
(1)
on the interaction and the one-particle reduced density matrix, density matrix ρmin in the minimum of Eq. (3). The derivatives

033160-2
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

ρ (1) of the full system. The nonlocal interactions Ŵnon−local ,

and the corresponding RDMF can either be similarly de-
composed further into semilocal RDMFs or be evaluated
approximately from approximate parametrized RDMFs [16],
or even approximate parametrized density functionals.
The following section shows how the ACA [15] can be
ˆ
used to evaluate a local or semilocal RDMF F W̃ [ρ (1) ] by
creating a smaller effective system for which the RDMF has
FIG. 1. Schematic representation of the decomposition of the
interaction Hamiltonian Ŵ into local terms Ŵlocal,i and a nonlocal
to be evaluated.
term Ŵnon−local for a situation with six one-particle states and two
states per local interaction term. Black dots represent the one-particle D. Adaptive cluster approximation
states. The starting point of the ACA [15] is a (semi)local RDMF
F Ŵlocal [ρ (1) ] with an interaction
of the RDMF can be expressed as
1 
 (1)  (1) Ŵlocal = Uα,β,δ,γ ĉα† ĉβ† ĉγ ĉδ , (13)
d Ŵ  (1)   ∂F Ŵ ρmin ∂ρmin,β,α 2 α,β,γ ,δ∈C
F ρmin =
dRi (1)
α,β ∂ρmin,β,α
∂Ri
which only includes a limited number of orbitals Nint = |C|
1  ∂Uα,β,δ,γ (2) that is much lower than the total number of one-particle basis
+ ρmin,γ δαβ , (9) states Nχ of the one-particle reduced density matrix ρ (1) .
2 α,β,γ ,δ ∂Ri
The goal of the ACA is to approximate the RDMF by
(2)
where ρmin is the two-particle reduced density matrix in the modifying the one-particle reduced density matrix ρ (1) into
(1)
minimum. The required derivatives of the RDMF with respect a much smaller one-particle reduced density matrix ρACA so
to the one-particle reduced density matrix are given by the that
Lagrange multipliers of the density-matrix constraints, i.e.,  (1) 
[ρ ] = F Ŵlocal ρACA
Ŵlocal (1)
FACA ≈ F Ŵlocal [ρ (1) ] (14)
 (1) 
∂F Ŵ ρmin and
(1)
= −λα,β . (10)
∂ρmin,β,α
∂FACA
Ŵlocal (1)
[ρ ] ∂F Ŵlocal [ρ (1) ]
(1)
≈ (1)
. (15)
Thus, an algorithm for the solution of the constrained mini- ∂ρβ,α ∂ρβ,α
mization problem given in Eq. (6) is required that produces
the Lagrange multipliers. In Sec. III we propose to use the This is achieved by systematically constructing an environ-
augmented Lagrangian approach [21,22] for this purpose. ment by setting up a unitary transformation of the one-particle
states, which do not interact in Ŵlocal so that the transformed
C. Local approximation of the RDMF
density matrix has a banded shape. This increases the near-
sightedness [28] of the one-particle reduced density matrix
While for the VQE one of the only known methods to and drastically mediates the impact of a truncation of most of
reduce the number of required qubits is to employ symme- the noninteracting one-particle states. If only the interacting
tries [23], our RDMFT-based formulation allows additional states are kept, it is denoted as zeroth-order ACA (ACA0 ),
avenues to drastically reduce the qubit count. Starting from whose the zeroth-order ACA density matrix has Nint one-
the general case of a RDMF F Ŵ [ρ (1) ], with a given interaction particle states. The first-order ACA is defined by keeping all
Hamiltonian Ŵ , it is important to recognize that many of the one-particle states that have nonzero elements in the trans-
important electronic correlation effects stem from strong local formed density matrix with interacting one-particle states.
electronic interactions [24–26]. We follow here the local- Hence, the first-order ACA density matrix has at most 2Nint
approximation approach of Blöchl and coworkers [19,27]. one-particle states. Thus, the solution of the RDMF in the
The electron-electron interaction can then be decomposed n-order ACA would require at most (n + 1)Nint one-particle
into local terms Ŵlocal,i and nonlocal terms Ŵnon−local , i.e., states and, hence, greatly reduce the number of required
 qubits.
Ŵ = Ŵlocal,i + Ŵnon−local (11) Figure 2 compares the convergence of the RDMF in the
i (1)
nth order ACA, i.e., F Ŵlocal,1 [ρACA(n) ] with the naive cluster
as schematically shown in Fig. 1. With the separation given approximation, where only the first n sites of the chain are
in Eq. (11) the RDMF can be approximated as a sum of local (1)
considered for the RDMF F Ŵlocal,1 [ρnaive,n ]. Although both
RDMFs and a remainder containing the nonlocal interactions RDMFs have the same number of one-particle states nNint and
as hence the same computational complexity, the RDMF in the

F Ŵ [ρ (1) ] ≈ F Ŵlocal,i [ρ (1) ] + F Ŵnon−local [ρ (1) ]. (12) ACA converges much more smoothly and quickly with the
i
number of one-particle states. The numerical results suggest
that the additive error
The RDMFs F Ŵlocal,i [ρ (1) ] have only a local interaction but are   (1)  
still functionals of the full one-particle reduced density matrix (n) = F Ŵlocal,1 ρACA(n) − F Ŵlocal,1 [ρ (1) ] (16)

033160-3
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

(1)
FIG. 2. Convergence of the RDMF F Ŵlocal,1 [ρACA(n) ] of the
ground-state density matrix for a half-filled 16-site Hubbard chain
(U/t = 1) for the local interaction Ŵlocal,1 on the first site of chain: the
(1)
nth order ACA is compared to the value of the RDMF F Ŵlocal,1 [ρnaive,n ]
with a naive truncation, where the truncated density matrix ρnaive,n (1)

only considers the first n sites of the chain.

converges like O(e−n ). In other words, the required ACA-

level n and, hence, the qubit count is in O( log( −1 )).
The corresponding implementation for the local approx-
imation and the ACA for the Hubbard model is publicly
available [29].

E. Comparison of an RDMFT-based approach to

wave-function-based approaches
The proposed RDMFT-based approach requires a two-level
minimization procedure expressed by Eq. (3) and Eq. (21),
respectively. This is in contrast to most wave function based
approaches either on classical or quantum computers via
the VQE. The two-level nature together with the fact that FIG. 3. Schematic representation of our hybrid quantum-
constrained minimization problem are commonly more com- classical algorithm for the total energy minimization with the RDMF.
plicated to solve than unconstrained minimization procedures The RDMF is approximated with the local approximation and the
shows that the proposed approach is more involved than for ACA so that the number of required qubits is drastically reduced in
example a conventional VQE procedure. comparison to existing VQE-like approaches.
The impact of the inner level, i.e., the constrained mini-
mization of the RDMF can be mediated in practice by reusing III. HYBRID QUANTUM-CLASSICAL
the optimal parameters from the previous total-energy mini- ALGORITHM FOR THE RDMF
mization step. Apart from the straightforward availability of
analytical nuclear forces, the main advantage of the proposed We propose to evaluate the density-matrix functionals
W̃ˆ
approach lies in its parallel nature and the ability to treat much F [ρ (1) ] in Eq. (12) on quantum computers with a VQE-like
larger systems than the underlying wave function represen- approach by solving the constrained minimization problem
tation (either classical or quantum-computing based) would given in Eq. (6). For this purpose, a parametrized ansatz
allow. |(u) is chosen for the fermionic many-particle wave state
As already noted in Sec. II C, our formulation within the |. It is parameterized by a vector of real parameters u.
framework of RDMFT also permits to utilize parametrized Several different efficient parametrizations are possible and
one-particle reduced density matrix functionals [16,30] for are discussed in Sec. IV B. A schematic of the computational
some of the local or nonlocal RDMFs in Eq. (12), which have approach is shown in Fig. 3.
already been shown to describe static and dynamic correlation The minimum of the parametrized constrained minimiza-
well. Thus, precious quantum computing resources can be tion
reserved for orbitals or spatial regions with otherwise hard to ˆ
F W̃ [ρ (1) ] =
u min (u)|W̃ˆ |(u) (17)
describe strong electronic correlations. u:|(u)→ρ (1)

033160-4
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

Algorithm 1. RDMF with augmented Lagrangian. the augmented Lagrangian can be written as
 1
set initial parameters u, initial penalties μ, initial Lagrange L(u, λ, μ) = W (u) + λi ci (u) + μi (ci (u))2 , (22)
multipliers λ 2 i
i
while not converged do
solve unconstrained problem minu L(u, λ, μ) with the objective function W (u), the Lagrange multipliers
λ ← λ + μc(u)  multiplier update λi , and penalty parameters μi . After setting initial penalty
μ ← f (μ, c(u))  penalty update parameters μ0 = μinitial and Lagrange multipliers λ0 = λinitial
end while the main loop of the augmented Lagrangian begins and re-
quires in every loop iteration k the solution of the auxiliary
unconstrained problem
is an upper bound to the exact constrained minimum of Eq. (6)
if the parametrization is sufficiently flexible to allow the ful- uk = argminu L(u, λk , μk ). (23)
fillment of the density-matrix constraints, i.e., there exists a u
so that The constraint violations ci (uk ) in this iteration are then
used to update the Lagrange multipliers as λi,k+1 = λi,k +
(1)
ρβ,α = (u)|ĉα† ĉα |(u) (18) μi,k ci (uk ) (first-order multiplier update). Additionally, the
penalty parameters μk can be modified for example with a
for the given ρ (1) . constant factor β > 1, i.e., μk+1 = βμk before the algorithm
Similar to the VQE this can be used to construct a hybrid proceeds to the next iteration k → k + 1. Thus, for the so-
quantum-classical algorithm, where the parameters u are op- lution of the constrained minimization only unconstrained
timized on a classical computer, but all expectation values problems in Eq. (23) have to be solved. Existing techniques
of quantum mechanical observables in the state |(u) are from noisy unconstrained minimization such as the simulta-
evaluated on a quantum computer. Relevant observables to be neous perturbation stochastic approximation (SPSA) [31,32]
evaluated on the quantum computer are the elements of the or related noise-resistant algorithms can be used.
one-particle reduced density matrix The augmented Lagrangian approach as a method for con-
(1)
ρβ,α (u) = (u)|ĉα† ĉα |(u) (19) strained minimization has the advantage that no derivatives
of the objective function or constraints are required and that
and elements of the two-particle reduced density matrix ρ (2) , exact constraint satisfaction is not required in every iteration.
i.e., A major advantage compared to the use of penalty methods
(2)
is that due to the presence of the Lagrange multipliers the
ρα,β,γ ,δ (u) = (u)|ĉγ ĉδ ĉα ĉβ |(u).
† †
(20) penalty parameters do not have to be increased to infinity,
Only those elements are required for which the corresponding which drastically reduces the problem of ill conditioning. The
matrix elements of the interaction Hamiltonian are nonzero. cost to be paid for this advantage is the necessity for multiple
With these quantities, the expectation value of the interac- solutions of the unconstrained subproblems in Eq. (23) that,
however, can be mediated by a warm-start strategy. The aug-
tion Hamiltonian W̃ (u) = (u)|W̃ˆ |(u) can be estimated. mented Lagrangian approach for the RDMF usually converges
These expectation values carry the exponential complexity of in between five and 10 outer iterations of the augmented
the fermionic many-particle problem and can be evaluated Lagrangian [33].
efficiently with gate-based quantum computers. A further important property of the augmented Lagrangian
The practical evaluation of expectation values requires two approach is the fact that the Lagrange multipliers λi corre-
steps. In the first step, the state |(u) is set up on the quantum spond to the derivatives of the RDMF with respect to the
computer. The second step is to measure the fermionic oper- elements of the one-particle reduced density matrix, which is
ators in the prepared state. To measure fermionic observables required for an efficient minimization of the total energy in
on gate-based quantum computers they have to be transformed Eq. (3).
to bosonic qubit operators, which in turn have to be trans-
formed with additional gates to Pauli-z operators that can
finally be measured. Details about fermionic transformations IV. REDUCTION OF THE NUMBER
are discussed in Sec. IV A. OF QUBIT OPERATIONS
With the measurements of the expectation values, the con- A. Fermionic transformations
strained minimization for the RDMF in Eq. (17) can be written
as A well-known way to map fermionic operators to bosonic
operators is the Jordan-Wigner transformation [34]. It maps
W̃ˆ
F [ρ (1) ] = min W (u). (21) each one-particle state to a qubit, but the number of required
u:ρ (1) (u)=ρ (1)
additional qubit operations scales linearly with the number
To solve this constrained minimization problem we propose to of qubits. The parity transformation [35] is equivalent to the
use the augmented Lagrangian approach [21,22], as described Jordan-Wigner transformation in this regard. A promising al-
in Algorithm 1. To formulate the problem in terms of the ternative is the Bravyi-Kitaev transformation [35,36], which
standard constrained optimization problem we map the com- in contrast to the Jordan-Wigner transformation requires only
plex density-matrix constraints ρ (1) (u) − ρ (1) = 0 to equality a logarithmically scaling number of additional qubit opera-
constraints ci for real and imaginary parts individually. Then tions and becomes advantageous for large problems.

033160-5
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

of a controlled rotation into two controlled Pauli gates and

additional single-qubit gates.

C. Increasing the flexibility of the trial states

We are not aware of any method that can answer the ques-
tion if a chosen hardware-efficient trial state can represent the
ground state or any other desired state of a system in poly-
nomial time by only using classical computations. Thus, any
FIG. 4. Quantum circuit representation of the hardware-efficient additional degrees of freedom in the hardware-efficient trial
trial state composed of single-qubit Euler rotations U q,i (u) and
state that do not affect the computation on a quantum com-
parametrized quantum entanglers Uent (u).
puter and only consume polynomial time are useful. Within
the ACA, there are Nint interacting one-particle states and
nNint noninteracting one-particle states. We therefore propose
B. Hardware-efficient trial states here to introduce an additional degree of freedom by means
As a variational ansatz for the hybrid-classical algorithm, of an unitary transformation of the noninteracting states that
described in Sec. III, we have implemented hardware-efficient is optimized in every step of the solution of the unconstrained
trial states [7]. For the latter, the state is parametrized by subproblems to minimize the augmented Lagrangian. This
quantum gates that are natively supported by the specific does not increase the number of observables to be measured,
quantum device at hand. Only the fermionic measurement because the interactions remain local on the Nint sites. The
operators are transformed for example with transformations computational effort for this additional step is polynomial
such as the Jordan-Wigner transformation to bosonic qubit in (n + 1)Nint . The additional freedom through the rotation
operators. This leads to short quantum programs for the state of part of the one-particle basis makes a given parametrized
preparation. By contrast, some other popular trial states, such hardware-efficient trial state more flexible in the sense that
as the unitary coupled-cluster ansatz [9,10], where the trial it can represent fermionic states that were not representable
state is formulated with fermionic operators that are then without the additional unitary transform.
transformed to qubit operators, tend to produce longer quan-
tum programs for the state preparation. Our RDMFT-based V. REDUCTION OF THE NUMBER
approach proposed in Sec. II is independent of the ansatz for OF QUANTUM PROGRAMS
the trial state.
Specifically, we parametrize the N-qubit wave function as At variance to the VQE, where only the expectation value
of the Hamiltonian Ĥ is required, the proposed evaluation
of the RDMF on quantum computers requires the individual
d N−1 N−1
estimation of all elements of the one-particle reduced density
|(u) = U q,d−i+1 (u) Uent U q,0 (u)|0 (1)
matrix, i.e., ρα,β and the expectation value of the interaction
i=1 q=0 q=0
(24) Hamiltonian Ŵ .
in which the index i enumerates d repetitive blocks of single-
qubit Euler rotations U q,i (u) followed by entanglers Uent , A. Measurements for the interaction Hamiltonian
which are composed of a sequence of two-qubit gates to create Measurements for the expectation value of the interaction
an entanglement between the qubits (Fig. 4). Crucially, we Ŵ  or the local interaction Ŵlocal,i  require in the worst case
choose the U q,i (u) as a sequence of single-qubit gates in O(Nχ4 ) quantum programs and O(Nint,i 4
) quantum programs,
accordance with the features of the quantum computer that respectively. The measurement of the expectation value of the
is available to us. (local) interaction Hamiltonian is very similar to the measure-
In the form of Eq. (24), the blocked structure of the conven- ment of the total energy in the VQE because the interaction
tional hardware-efficient trial states leads to an extendability is contained in the total energy. Thus, the same techniques
issue: parameters that were optimized for a trial state with a developed for the VQE can be applied here for the interaction
given number of blocks can not be used as a starting point energy, e.g., the approach proposed by Izmaylov et al. [37]
for a trial state with more blocks due to the presence of the brings the scaling down to O(Nχ3 ) and O(Nint,i 3
), respectively.
entanglers. However, one can work around this caveat by However, in this paper we will instead focus on the measure-
replacing the conventional entanglers Uent by parametrized ment of the individual elements of the one-particle reduced
entanglers Uent (u) that for a certain choice of the parameters density matrix, which is qualitatively different from the mea-
are identical to the identity gate. In this case, a shallow ini- surement of the total energy or interaction energy.
tial trial state can be readily extended with these “disabled”
entanglers, which will then be tuned away from identity as
part of the constrained minimization. A simple choice for B. Measurements for the one-particle reduced density matrix
the parametrized entanglers is the controlled rotations, i.e., The number of measurements to evaluate an individual ele-
CRX or CRZ, instead of the basic controlled Pauli gates ment of the one-particle reduced density matrix is independent
CNOT or CZ. Note that one drawback of parametrizing the of the system size for the common fermionic transformations.
entanglers is the doubling of required basic two-qubit gates Thus, the number of measurements for each evaluation of the
for the implementation of Uent (u) due to the decomposition one-particle reduced density matrix scales quadratically with

033160-6
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

the number of one-particle basis states. Consequently, without

the ACA of the RDMF, the number of measurements would
scale quadratically with the overall system size in terms of
one-particle basis states.
With the ACA, the number only scales as


Nlocal

O (n + 1)2 Nint,i
2
, (25)
i=1

where Nlocal is the number of local RDMFs, n the chosen

order of the ACA, Nint,i the number of interacting one-particle
states in the ith local interaction Ŵlocal,i . The number of local
RDMFs is linear in terms of the overall system size. Provided
some order n of the ACA is sufficient for a proper description
of the system, then the overall number of measurements for
the one-particle reduced density matrices scale linearly with
the system size because the number of measurements for each FIG. 5. The number of quantum programs for the measurement
local RDMF is independent of the overall system size. Thus, of the one-particle reduced density matrix for a given number of
this linear-scaling behavior of the ACA drastically reduces the one-particle states at different levels of commutativity of Pauli
qubit requirements, as well as the overall measurement count. strings: without commutativity, Pauli strings acting on disjoint qubits
(DISJOINT), qubit-wise commutativity (QWC) and general commu-
1. Combination of measurements tativity (GC), respectively. The Jordan-Wigner transformation has
been used and the implied minimum clique cover problem has been
The number of quantum programs can be further reduced
solved with polynomial-time algorithms from NetworkX.
by combining commuting observables [6,7,38] instead of
measuring only a single observable per quantum program
execution. We will show here that all elements of the one- polynomial time. We use here the algorithms implemented in
particle reduced density matrix can be measured with only the NetworkX-library [40]. For the remainder of the section,
O(Nχ ) quantum programs instead of O(Nχ2 ), where Nχ is the we use the Jordan-Wigner transformation and assume an even
number of one-particle states. The fermionic transformation number of one-particle states. Other fermionic transforma-
turns the elements of the one-particle reduced density matrix tions give similar results.
ρ̂β,α = ĉα† ĉβ into weighted sums of Pauli strings, e.g., ĉ1† ĉ1 → Figure 5 shows the dependence of the number of measure-
1/21 ⊗ ... + 1/2Ẑ ⊗ 1 ⊗ ... in the Jordan-Wigner transfor- ment programs on the number of one-particle states Nχ for the
mation. Pairwise commuting Pauli strings can be measured measurement of the one-particle reduced density matrix for
simultaneously, i.e., in a single quantum program. We define the three commutativity levels. In total 2Nχ2 − Nχ Pauli strings
three different levels of commutation: have to be measured for the elements of the one-particle re-
(i) DISJOINT: Disjoint-qubit commutativity means that duced density matrix. When using qubit-wise commutativity,
the Pauli operators in the two Pauli strings act on different the number is reduced to roughly Nχ2 , but still scales quadrat-
qubits, e.g., Ẑ ⊗ 1 and 1 ⊗ Ẑ commute, but not Ẑ ⊗ 1 and ically. Using the general commutativity of Pauli strings, the
Ẑ ⊗ Ẑ. number of programs is reduced to approximately 2Nχ because
(ii) QWC: Qubit-wise commutativity is satisfied if each with each quantum program about Nχ Pauli strings can be
Pauli operator in the first Pauli string commutes with the Pauli measured.
operator on the same qubit of the second Pauli string, e.g., Ẑ ⊗ a. Construction of measurement programs. When measur-
1 and Ẑ ⊗ Ẑ commute, but not Ẑ ⊗ 1 and X̂ ⊗ Ẑ, respectively. ing just a single Pauli string, the measurement circuits are
(iii) GC: general commutativity between two Pauli strings, straightforward to set up by first transforming the x- and
e.g., Ẑ ⊗ X̂ ⊗ X̂ and 1 ⊗ Ŷ ⊗ Ẑ commute. y-Pauli operators to Pauli-z operators with single-qubit gates
The employed commutation level can be used to define a and then reducing the multiqubit z measurements to single-
graph where the Pauli strings that have to be measured are qubit z measurements with two-qubit gates. The freedom in
the nodes and the two nodes are connected by an edge if they the placement of the two-qubit gates can be used to optimize
commute at the selected commutation level. The problem of for the coupling topology of the given quantum computer.
finding the lowest possible number of disjoint groups of pair- When measuring multiple Pauli strings per quantum pro-
wise commuting Pauli strings, i.e., sets of Pauli strings that gram, we follow the general idea of the construction given
can be measured simultaneously is equivalent to the minimum by Gokhale et al. [38] based on the stabilizer formalism
clique cover problem. The minimum clique cover problem can [41,42]. However, we have improved on the formulation and
be solved with heuristic graph coloring algorithms. implementation of the construction in several aspects: we
All Pauli strings in a clique are then measured with one have implemented it purely over GF(2), we have added the
quantum program and the total number of quantum pro- missing phase row, our construction also works if the set of
grams is given by the number of cliques. Even though the Pauli strings is not complete or linearly independent, and we
minimum clique cover problem is NP-hard [39], greedy algo- have identified freedom in some aspects of the construction
rithms have been shown to give good approximate results in to optimize for a given quantum computer. Because of our

033160-7
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

modifications in the underlying algorithm, we present the Algorithm 2. Construction of measurement programs.
construction here as a whole.
The construction utilizes the stabilizer matrix formalism, Input: N commuting Pauli strings P1 , ..., PN
which writes a set of N commuting Pauli strings that act on Output: measurement circuit consisting of the gates that have
Nq qubits as a stabilizer matrix S ∈ GF(2)2Nq +1×N , i.e., been applied to S
S ← stabilizer matrix of P1 ,
..., PN with ri = 0  prep.
⎡ ⎤
Z1,1 Z1,2 ... Z1,N RZX ← rk
SZ
⎢ .. .. ⎥ SX
⎢ . . ... ⎥ while rk(SX ) < RZX do  rank max.
⎢ ⎥
⎢Z ... ZNq ,N ⎥ choose some qubit q to apply Hadamard to
⎢ Nq ,1 ZNq ,2 ⎥
⎢ ⎥ S ← H (q)S
S=⎢
⎢ X1,1 X1,2 ... X1,N ⎥ ⎥. (26) end while
⎢ . .. ⎥ P, L, U ← PLU decomposition of SZX  PLU-decomp.
⎢ .. . ... ⎥
⎢ ⎥ T ← transpositions in P
⎢ ⎥
⎣XNq ,1 XNq ,2 ... XNq ,N ⎦ for each transposition i → j in T do  permutation
r1 r2 ... rN S ← SWAP(i, j)S
end for
The elements are defined as follows: L −1 ← inv(L)
(1) Zi, j = 1 iff the j-th Pauli-string has the Pauli-z or for each (i, j), i = j with Li,−1j = 1 do  row reduc.
Pauli-y acting on the ith qubit. S ← CNOT(i, j)S
end for
(2) Xi, j = 1 iff the jth Pauli-string has the Pauli-x or Pauli-
if SX is not diagonal then  diag. reduc.
y acting on the ith qubit.
reduce SX to diagonal form with CNOTs
(3) r j defines that the jth single-qubit measurement needs
end if
to be multiplied with (−1)r j . reduce SZ to zero with phase gates and CNOTs
(4) Otherwise all elements are zero. for each qubit q do  X-Z-flip
For example the Pauli strings X̂ ⊗ Ẑ ⊗ Ŷ ⊗ 1, Ŷ ⊗ Ẑ ⊗ S ← H(q)S
X̂ ⊗ 1, 1 ⊗ X̂ ⊗ Ẑ ⊗ Ŷ , and 1 ⊗ Ŷ ⊗ Ẑ ⊗ X̂ correspond to end for
the stabilizer matrix for each qubit q do  sign-step
⎡ ⎤ if rq = 1 then
0 1 0 0
S ← Y(q)S
⎢1 1 0 1⎥
⎢ ⎥ end if
⎢ ⎥
⎢1 0 1 1⎥ end for
⎢ ⎥
⎢0 0 1 0⎥
⎢ ⎥
⎢ ⎥
S=⎢ ⎢ 1 1 0 0 ⎥,
⎥ (27)
⎢ ⎥ (2) transforming the SX to upper triangular form (PLU-
⎢0 0 1 1 ⎥
⎢ ⎥ decomp. and row reduc.),
⎢1 1 0 0 ⎥
⎢ ⎥ (3) reducing SX to diagonal form (diag. reduc.),
⎢ ⎥
⎣0 0 1 1 ⎦ (4) transforming the SZ to a zero matrix (Z reduction),
(5) exchanging the elements of the SZ and Sz matrix so that
0 0 0 0
SX = 0 and SZ diagonal (X-Z flip)
(6) and using the information in the phase row to get the
where ri = 0 has been chosen as a convention. The action of
correct signs in the measurements (sign step).
quantum gates now corresponds to changes in the stabilizer
The result is a stabilizer matrix, where SZ is diagonal, SX =
matrix [42]:
0 and ri = 0, which means that the Pauli strings Pi can be
(i) CNOT(c,t): ∀i ∈ 1, ..., N do ri → ri + Zt,i Xc,i (1 +
measured as single-qubit measurements at the qubits q, where
Xt,i + Zc,i ), Xt,i → Xt,i + Xc,i , Zc,i → Zc,i + Zt,i .
SZ,q,i = 1.
(ii) Hadamard gate H(q): ∀i ∈ 1, ..., N do ri → ri +
The initial step is the construction of the stabilizer matrix
Xq,i Zq,i and swap Xq,i and Zq,i .
S from the set {Pi } of N Pauli strings. It should be noted that
(iii) Phase gate S(q): ∀i ∈ 1, ..., N do ri → ri + Xq,i Zq,i ,
all algebra has to be performed over GF(2) instead of R. The
Zq,i → Zq,i + Xq,i . T
All operations are performed in GF(2). The inclusion of the initial rank of the matrix [SZ SX ] represents the maximal
phase row r1 , ..., rN guarantees that the sign of the measure- rank of the SX matrix that can be reached by exchanging rows
ments is correct. of the SX and SZ matrix with Hadamard gates. Next, these
The algorithm for the construction of the measurement Hadamard gates are applied to qubits such that the rank of SX
T
circuits is described in Algorithm 2. To simplify the notation becomes equal to the initial rank of [SZ SX ] . The qubits
we define the first Nq rows of the stabilizer matrix as matrix where the Hadamard gates have to be applied can be found in
SZ , the following Nq rows as the matrix SX , and the last row polynomial time with the simple Algorithm 3.
as the phase row. For the example given in Eq. (27) the ranks are
T
Specifically, the algorithm performs the following steps: rk([SZ SX ] ) = 4 and rk(SX ) = 2. Applying Hadamard
(1) maximizing the rank of the SX matrix (rank max.), gates at the first and second qubit results in the stabilizer

033160-8
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

Algorithm 3. Maximizing the rank of SX with Hadamards. q0 : H × • S H Y

q1 : H × • S H
Input: stabilizer matrix S
Output: set H of qubits, where Hadamards need to be applied q2 : × S H Y
so that SX has maximal q3 : × • • S H
rank
SZ c : /4
RZX ← rk 3 2 1 0
SX
H ←∅
FIG. 6. Measurement circuit constructed with Algorithm 2 for
while rk(SX ) < RZX do
the group of Pauli strings X̂ ⊗ Ẑ ⊗ Ŷ ⊗ 1, Ŷ ⊗ Ẑ ⊗ X̂ ⊗ 1, 1 ⊗ X̂ ⊗
for each qubit q not yet in H do
Ẑ ⊗ Ŷ , and 1 ⊗ Ŷ ⊗ Ẑ ⊗ X̂ , respectively.
if rk((H (q)S)X ) > rk(SX ) then
S ← H (q)S
H ←H ∪q The row reduction with L −1 , i.e., CNOT(0,3) then reduces the
end if stabilizer matrix to
end for ⎡ ⎤
end while 1 0 0 0
⎢1 1 0 0 ⎥
⎢ ⎥
⎢0 0 1 0⎥
⎢ ⎥
⎢1 0 1 1⎥
⎢ ⎥
Srow red. = ⎢1 1 0 1⎥ (31)
⎢0 1 0 0⎥
matrix ⎢ ⎥
⎢0 0 1 1⎥
⎢ ⎥
⎡ ⎤ ⎣0 0 0 1⎦
1 1 0 0 0 1 0 1
⎢0 0 1 1⎥
⎢ ⎥
⎢1 0 1 1⎥ and with the diagonal reduction, i.e., CNOT(3,2), CNOT(3,1),
⎢ ⎥
⎢0 0 1 0⎥ and CNOT(1,0) finally to
⎢ ⎥
Srank max. = ⎢0 1 0 0⎥ (28) ⎡ ⎤
⎢1 1⎥ 1 0 0 0
⎢ 1 0 ⎥ ⎢0 1 0 0⎥
⎢1 0⎥ ⎢ ⎥
⎢ 1 0 ⎥ ⎢0 0 1 0⎥
⎣0 0 1 1⎦ ⎢ ⎥
⎢0 0 0 1⎥
0 1 0 1 ⎢ ⎥
Sdiag. red. = ⎢1 0 0 0⎥, (32)
⎢0 1 0 0⎥
⎢ ⎥
⎢0 0 1 0⎥
with rk(SX ) = 4. The PLU decomposition of Srank max.,X in the ⎢ ⎥
⎣0 0 0 1⎦
exemplary case is
0 1 0 1
⎛ ⎞⎛ ⎞ where SX = 1. With phase gates at all qubits, the SZ matrix
0 1 0 0 1 0 0 0 can be reduced to a zero matrix and with Hadamard gates at
⎜1 0 0 0⎟⎜0 1 0 0⎟
=⎝
1⎠⎝0 0⎠
Srank max.,X all qubits the SZ and SX matrix are flipped to obtain
0 0 0 0 1 ⎡ ⎤
0 0 1 0 1 0 0 1 1 0 0 0
⎛ ⎞ ⎢0 1 0 0⎥
1 1 0 1 ⎢ ⎥
⎢0 0 1 0⎥
⎜0 1 0 0⎟ ⎢ ⎥
×⎝ ⎢0 0 0 1⎥
1⎠
(29)
0 0 1 ⎢ ⎥
SX−Z−flip = ⎢0 0 0 0⎥. (33)
0 0 0 1 ⎢0 0 0 0 ⎥
⎢ ⎥
⎢0 0 0 0 ⎥
⎢ ⎥
⎣0 0 0 0 ⎦
and, hence, the stabilizer matrix after permutation with
0 1 0 1
SWAP(0,1) and SWAP(2,3) is
The final step is to apply Y gates at the second and fourth
⎡ ⎤ qubits to obtain the proper signs of the measurements. The
0 0 1 1 resulting measurement circuit for this example is shown in
⎢1 1 0 0⎥
⎢ ⎥ Fig. 6.
⎢0 0 1 0⎥
⎢ ⎥ b. Optimization of measurement programs. The naive ap-
⎢1 0 1 1⎥ plication of the construction of measurement programs can,
⎢ ⎥
Sperm. = ⎢1 1 0 1⎥. (30) however, leads to an O(N 2 ) gate count in the worst case [38].
⎢0 0⎥
⎢ 1 0 ⎥ Therefore, we propose here to use the degrees of freedom in
⎢0 1⎥
⎢ 0 1 ⎥ the construction to optimize the measurement circuits for a
⎣1 0 0 0⎦ given quantum computer. One important degree of freedom is
0 1 0 1 the ordering of the measurement qubits, i.e., what single-qubit

033160-9
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

FIG. 8. Hardware efficient trial state used in this work with 16

real parameters u.

covers the electron-electron interaction of two electrons on the

same site. Hence, the interaction operator is

L
Ŵ = U n̂i,↑ n̂i,↓ . (35)
i=1

The half-filled Hubbard chain can be viewed as a model for a

chain of Hydrogen atoms, where the interaction strength U/t
is analogous to the distance between the atoms.
FIG. 7. Maximal gate count of the constructed measurement cir- We focus on the evaluation of the RDMF F Ŵ [ρ (1) ] for a
cuits for all elements of the one-particle reduced density matrix for a given one-particle reduced density matrix ρ (1) . This result is
given number of qubits. then used in the minimization of the total energy in Eq. (3).
As a physical example case we use the one-particle reduced
density matrix ρ0(1) of the ground-state of the chosen model.
measurement is measured at which qubit. The depth and the The ground state of the L-site half-filled Hubbard chain can
gate count of the constructed measurement circuits is sensitive be obtained efficiently for arbitrary interaction strengths with
to the order. Because the number of possible permutations MPS-DMRG techniques. We use here the MPS implementa-
scales exponentially with the number of qubits, an efficient tion in ITensor [44] through DMRGPY [45]. This defines the
heuristic for the order is required. Fixing the order so that
RDMF F Ŵ [ρ0(1) ] to be calculated.
there are no permutations in the PLU-decomposition is an
example of a simple heuristic. Figure 7 shows the maximal
gate count of the constructed measurement circuits for all B. Local approximation and ACA
elements of the one-particle reduced density matrix when Now the pipeline proposed in Sec. II, i.e., the local approx-
using the mentioned simple heuristic and different fermionic imation and the ACA can be applied,
transformations. Interestingly, the maximal gate count in a
measurement circuit only grows roughly linear with the qubit    L
 
F Ŵ ρ0(1) ≈ F U n̂i,↑ n̂i,↓ ρ0(1) (36a)
count.
i=1
The implementation for the presented algorithm based on
Qiskit [43] is publicly available [29]. Optimizations that ex- 
L
U n̂i,↑ n̂i,↓  (1) 
plore different heuristics and also consider the set of supported ≈ FACA(n) ρ0 (36b)
gates as well as the coupling topology of the given quantum i=1
computer will be addressed in future work. 
L
 (1) 
= F U n̂i,↑ n̂i,↓ ρ0,ACA(n) , (36c)
VI. RESULTS FOR A MODEL SYSTEM i=1

A. Definition of the model where we have chosen a single-site local approximation, i.e.,
Nint = 2 and n determines the order of the ACA. The result-
The Hubbard model [24–26] has proven to be an ex- (1)
ing L functionals F U n̂i,↑ n̂i,↓ [ρ0,ACA(n) ] to be computed consider
tremely valuable model system for materials with strong only 2(n + 1) one-particle states instead of 2L of the full
electronic correlations. Conventional DFT fails to qualita- system. The results discussed in Sec. II D show that the first-
tively describe these strong electronic correlations and the order ACA, i.e., n = 1 already gives results close to the exact
resulting phenomena like metal-insulator transitions. Thus, value and quantum computers with at least 4 qubits are widely
alternative correlated approaches are required. available. Thus, this work will consider the evaluation of the
We use here a L = 8-site half-filled Hubbard chain with the (1)
RDMF F Ŵ1 [ρ0,ACA(1) ] with Ŵ1 = U n̂1,↑ n̂1,↓ as an exemplary
Hamiltonian case.
 
L−1 
L
Ĥ = −t †
ĉi,σ ĉi+1,σ + U n̂i,↑ n̂i,↓ , (34) C. Trial state
σ ∈{↑,↓} i=1 i=1
As the hardware-efficient trial state we propose the variant
where n̂i,σ = ĉi,σ
†
ĉi,σ is the particle number operator for the shown in Fig. 8 tailored to IBM quantum computers
√ with the
spin channel σ on site i, t is the hopping parameter and U  0 Falcon r4L processor type, which have Rz -, X - and CNOT-
the interaction operator. The first sum represents the kinetic gates as basic gates and a linear topology for basic two-qubit
energy of electrons hopping between sites and the second sum gates [46]. The Bravyi-Kitaev transformation is used for the

033160-10
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

FIG. 10. Comparison of the RDMF Fqc,no noise obtained with the
FIG. 9. Convergence of the augmented Lagrangian L, the  inter- augmented Lagrangian and the trial state in Fig. 8 without noise
action energy Ŵ1  = U n̂1,↑ n̂i,↓ and the constraint violation i ci2 compared to the exact value Fexact (see Appendix B) for the RDMF
(1)
at the end of each solution of an unconstrained subproblem in F U n̂1,↑ n̂1,↓ [ρ0,ACA(1) ] defined in Sec. VI B.
a noiseless simulation for the situation defined in Sec. VI B and
U/t = 1. Fexact is the value of the RDMF obtained with a constrained
minimization over Slater determinants (see Appendix B). Solid lines Lagrange multipliers are initialized to zero, i.e., λ0 = 0, then
show the convergence when the initial Lagrange multipliers are cho- the augmented Lagrangian approach converges from a large
sen as zero and dashed lines the corresponding results if the initial constraint violation to the final result. However, if an initial
values of the Lagrange multipliers are chosen as the derivatives of guess for the Lagrange multipliers is obtained from the Müller
the Müller functional. functional [17], then the constraint violation is already small
after the first iteration of the augmented Lagrangian and the
convergence of the interaction energy Ŵ  is much quicker.
fermionic transformation and the two interacting one-particle The Müller functional [17] is an approximate parametrized
states are placed on the two center qubits. RDMF and its derivatives with respect to the one-particle
The chosen hardware efficient trial state is minimal in reduced density matrix, which correspond to the Lagrange
terms of the number of two-qubit gates. We were not able multipliers, can be obtained with polynomial cost.
to find a parametrization of the trial state that represents The results for the RDMF Fqc,no noise obtained with the
(1)
the one-particle reduced density matrix ρ0,ACA(1) of the cho- augmented Lagrangian and the trial state in Fig. 8 without
sen example case without the additional freedom generated noise are shown for a wide range of interactions strengths
by the additional unitary transform of the noninteracting U/t in Fig. 10. The chosen trial state together with the uni-
states introduced in Sec. IV C. One exception is the case of tary transformation of the noninteracting states introduced in
infinite interaction strengths, i.e., U/t → ∞, where the ACA- Sec. IV C obtains the density-functional that is nearly indis-
(1)
transformed density matrix ρ0,ACA(1) is diagonal. Hence, this tinguishable from the exact result for arbitrary U/t, i.e., in the
additional freedom makes the constraints in the constrained weakly interacting regime U/t  1, as well as in the strongly
minimization problem satisfiable for arbitrary interactions interacting and correlated regime U/t  1.
strengths and allows us to use such a simple trial state. The
implementation of the augmented Lagrangian and input files
for the presented results are publicly available [29]. E. Results on NISQ device
After showing that the proposed schema with the chosen
trial state converges accurately to the RDMF when exact ex-
D. Simulation without noise pectation values are considered, we now include noise into the
To show the suitability of the chosen trial state the con- simulation.
strained minimization with the augmented Lagrangian was We focus on IBM quantum computers [46] with a linear
run without noise, i.e., exact expectation values. Details for coupling topology like the ibmq_bogota, ibmq_manila, or
the constrained minimization can be found in Appendix C. ibmq_santiago machines, respectively. These are all 5-qubit
Figure 9 shows the convergence of the value of the augmented quantum computers with the processor type Falcon r4L. The
Lagrangian L and the expectation value of the interaction noisy simulations have been performed with the density-
Ŵ1  = U n̂1,↑ n̂i,↓ , i.e., the value of the RDMF in comparison matrix-based simulator in Qiskit 0.29.0 using the noise model
to the value Fexact of the RDMF obtained from a constrained of the ibmq_manila quantum computer, the qubit-wise com-
minimization over Slater determinants (see Appendix B). Ad- mutation of measurements, and 8192 shots per measurement.
ditionally, the overall constraint violation i ci2 is shown. The The noise simulation included depolarizing gate errors, ther-
convergence of the augmented Lagrangian is rapid so that mal relaxation errors, and single-qubit readout errors. The first
typically no more than 10 outer iterations are required. If the four qubits of the 5-qubit quantum computer were used.

033160-11
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

FIG. 11. Convergence of the expectation value Ŵ1  = U n̂1,↑ n̂i,↓ FIG. 12. Convergence of the expectation value Ŵ1  = U n̂1,↑ n̂i,↓
of the interaction and and the overall constraint violation i ci2 dur- of the interaction and the overall constraint violation i ci2 during
ing the solution of the unconstrained subproblems of the augmented the solution of the unconstrained subproblems of the augmented La-
(1) (1)
Lagrangian for the RDMF F U n̂1,↑ n̂1,↓ [ρ0,ACA(1) ] defined in Sec. VI B grangian for the RDMF F U n̂1,↑ n̂1,↓ [ρ0,ACA(1) ] defined in Sec. VI B for
for U/t = 1. The expectation values of the trial states were obtained U/t = 1 when executed with the ibmq_manila quantum computer.
from a noisy simulation on a classical computer. The correspond- Vertical-dashed lines indicate different outer iterations, i.e., updates
ing results obtained on a quantum computer are shown in Fig. 12. of the penalty parameters and Lagrange multipliers.
Vertical-dashed lines indicate different outer iterations, i.e., updates
of the penalty parameters and Lagrange multipliers.
commenced from a random starting point. Thus, we conclude
that the convergence to a local minimum rather than a global
minimum is very unlikely and that the observed behavior
Figure 11 shows the behavior of the expectation  value
points to a representability issue of the many-particle state
of the interaction Ŵ1  and the constrained violation i ci2
on the noisy quantum computer. This aspect is the subject of
during the constrained minimization when the quantum pro-
future research.
grams are simulated with the noise model of the real quantum
We therefore conclude that the evaluation of the RDMF
computer ibmq_manila. The corresponding calibration results
is possible on a quantum computer. The results obtained this
are given in Table I. The convergence when using the real
way can then be used in an iterative total energy minimization.
quantum computer to evaluate the expectation values during
The formulation of the total energy minimization problem
the minimization is very similar to the noisy simulation.
with the RDMF allows us to employ novel approximations
During the first outer iteration of the augmented La-
grangian, the constraint violation converges in both the noisy
simulation and real execution to a finite value. For a solution
of a constrained minimization problem with the augmented
Lagrangian approach this behavior is not unexpected, as can
be seen also in Fig. 9. However, after the penalty update, i.e.,
increase of the penalties on the constraints, the convergence
to essentially the same constraint violation in the second and
subsequent outer iterations is curious and points to a deeper is-
sue. Two possible explanations exist: Either the minimization
in the first outer iteration has converged to a local minimum
that is not the global minimum and subsequent iterations are
not able to escape this minimum, or due to the inclusion
of noise, the trial state can not represent a state that fulfills
the constraints. The dominant part of the overall constraint
violation stems from the constraints on the off-diagonal el-
(1) (1)
ements ρACA(1),1↑,2↑ and ρACA(1),1↓,2↓ , respectively, i.e., the
covalencies.
Using the converged parameters unoiseless of the trial state
from the noiseless constrained minimization as a starting point
for the constrained minimization with noise, the convergence FIG. 13. Convergence for the same situation as shown in Fig. 11
as shown in Fig. 13 is obtained. The converged values are very but starting from the converged parameters unoiseless of the noiseless
similar to the results shown in Fig. 11, where the minimization simulation.

033160-12
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

like the ACA to drastically reduce the qubit requirements of APPENDIX A: AVAILABILITY OF SOURCE
the problem and to parallelize it. CODE AND DATA
The source code of the software implementation of the
presented approach, input files for the presented results as well
VII. CONCLUSIONS AND OUTLOOK as the raw data are openly available at [29].
We have introduced a hybrid quantum-classical approach
based on an RDMFT formulation of the quantum total energy APPENDIX B: PRACTICAL DETAILS
problem and the ACA for gate-based quantum computers. OF THE EXACT RDMF
Using the latter, a drastic reduction of the necessary qubit
count is demonstrated. The measurements of the one-particle The exact reference of the RDMF has been obtained by
reduced density matrix that are required for the density-matrix parametrizing
constraints are shown to be obtainable with only a linear
number of quantum programs when the general commutativ- Ŵ
Fexact [ρ (1) ] = min (x)|Ŵ |(x) (B1)
Nχ
ity of observables is exploited. A construction algorithm for x∈C 2
the measurement programs is given and available degrees of
freedom for their optimization are introduced. with the equality constraints
The essential part of the proposed approach, the evalua-
(1)
tion of a local RDMF, is demonstrated with a Hubbard-type (x)|ĉα† ĉβ |(x) = ρβ,α , (B2a)
model system using noise-free simulations, simulations in- (x)|(x) = 1, (B2b)
cluding noise models of realistic quantum computers, as well
as with executions on real IBM quantum computers. While 
where |(x) = i xi |ni  is a sum over all Slater determinants
model systems such as the Hubbard model are convenient to
|00 · · · , |10 · · · ,..., respectively. The complex parameters
study the features and convergence behavior of the proposed
xi were represented by their real and imaginary parts and
algorithm, the goal of this approach is the application within
the constrained minimization was performed using the trust-
ab initio molecular dynamics calculations. The main differ-
region constrained minimization algorithm from SciPy 1.7.1
ence between the treatment of model systems and realistic
[52,53] with a convergence tolerance of 10−9 .
systems is the number of terms in the interaction Hamilto-
nian. The variational formulation based on RDMFT makes
the proposed algorithm very suitable for ab initio molecular APPENDIX C: PRACTICAL DETAILS OF THE
dynamics calculations because the forces can be evaluated in a AUGMENTED LAGRANGIAN APPROACH
straightforward way from already available quantities such as
the one- and two-particle reduced density matrix [47]. Noise a. Parameters for the Augmented Lagrangian. The initial
in the nuclear forces stemming from the noise of the quan- values for the parameters u0 were uniformly chosen at random
tum computer can be compensated for in ab initio molecular from [−π , π ]. The initial values for the Lagrange multipliers
dynamics simulations in the spirit of approximate computing λ0 were either chosen as zero, or as the numerical derivatives
[48], where the desired thermodynamic expectation values of the Müller functional [17] for the given one-particle re-
can nevertheless be accurately obtained by devising a prop- duced density matrix. The initial value for the penalties μ0
erly modified Langevin equation [49,50]. The investigation was chosen as 10 and the penalties were updated after each
of the representability of fermionic quantum states on noisy solution of the unconstrained problem by multiplying them
gate-based quantum computers, the optimization of the mea- with 1.5.
surement programs as well as the integration with molecular
dynamics programs like CP2K [51] are subject for future
1. Solution of the unconstrained subproblems
research.
In the case without noise, the unconstrained subproblems
were solved with the L_BFGS_B and COBYLA algorithms
implemented in Qiskit 0.29.0 [43,52]. The tolerance was set
ACKNOWLEDGMENTS to 10−3 and at most 10 000 iterations were permitted. The
The authors acknowledge the use of IBM Quantum ser- optimization of the unitary transform of the noninteracting
vices for this paper. The views expressed are those of one-particle states outlined in Sec. IV C has been performed
the authors, and do not reflect the official policy or posi- with the BFGS algorithm (with a convergence tolerance of
tion of IBM or the IBM Quantum team. In this paper the 10−9 ) from SciPy 1.7.1 [52] and the parametrization of unitary
ibmq_manila, which is one of the IBM Quantum Falcon Pro- matrices
cessors has been used. The authors gratefully acknowledge
the funding of this project by computing time provided by the U = eiH , (C1)
Paderborn Center for Parallel Computing (PC). This work is
partially funded by Paderborn University’s research award for where H is a parametrized arbitrary hermitian matrix.
“GreenIT”, as well as the Federal Ministry of Education and In the case with noise, the COBYLA algorithm imple-
Research (BMBF) and the state of North Rhine-Westphalia as mented in Qiskit 0.29.0 [43,52] was used with a convergence
part of the NHR Program. tolerance of 0.01.

033160-13
ROBERT SCHADE et al. PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

TABLE I. Calibration used for the ibmq_manila quantum computer.

Qubit Q0 Q1 Q2 Q3 Q4
T1 in μs 128.51 142.79 184.95 214.35 127.79
T2 in μs 74.34 79.95 24.85 69.18 44.4
Frequency in GHz 4.963 4.838 5.037 4.951 5.066
Anharmonicity in GHz –0.34335 –0.34621 –0.34366 –0.34355 –0.34211
Readout error 0.0259 0.0324 0.0190 0.02210 0.01920
SPAM error |1 → |0 0.0438 0.0482 0.0286 0.0306 0.0302
SPAM error |0 → |1 0.0080 0.0166 0.0094 0.0136 0.0082
Readout√ length in ns 5351.111 5351.111 5351.111 5351.111 5351.111
Identity, x, x errors × 104 1.920 3.970 2.571 1.834 3.155
CNOT error × 103 0 to 1 6.684 1 to 2 9.947 2 to 3 7.344 3 to 4 4.746
CNOT Gate time in ns 277.33 312.889 504.889 391.111 298.667

APPENDIX D: CALIBRATION OF THE IBMQ_MANILA QUANTUM COMPUTER

The thermal relaxation time constant T1 , the dephasing time constant T2 , and error rates from calibration of the ibmq_manila
quantum computer are shown in Table I.

[1] M. Troyer and U.-J. Wiese, Computational Complexity for simulating molecular electronic structure, Phys. Rev. A 95,
and Fundamental Limitations to Fermionic Quantum 020501(R) (2017).
Monte Carlo Simulations, Phys. Rev. Lett. 94, 170201 [11] T. L. Gilbert, Hohenberg-Kohn theorem for nonlocal external
(2005). potentials, Phys. Rev. B 12, 2111 (1975).
[2] F. Calcavecchia, F. Pederiva, M. H. Kalos, and T. D. Kühne, [12] A. J. Coleman, Structure of fermion density matrices, Rev.
Sign problem of the fermionic shadow wave function, Phys. Mod. Phys. 35, 668 (1963).
Rev. E 90, 053304 (2014). [13] M. Levy, Universal variational functionals of electron densities,
[3] S. Lloyd, Universal quantum simulators, Science 273, 1073 first-order density matrices, and natural spin-orbitals and solu-
(1996). tion of the v-representability problem, Proc. Natl. Acad. Sci.
[4] A. Aspuru-Guzik, A. Dutoi, P. Love, and M. Head-Gordon, USA 76, 6062 (1979).
Simulated quantum computation of molecular energies, Science [14] E. H. Lieb and F. Y. Wu, Absence of Mott Transition in an
309, 1704 (2005). Exact Solution of the Short-Range, One-Band Model in One
[5] A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, Dimension, Phys. Rev. Lett. 20, 1445 (1968).
P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A varia- [15] R. Schade and P. E. Blöchl, Adaptive cluster approximation
tional eigenvalue solver on a photonic quantum processor, Nat. for reduced density-matrix functional theory, Phys. Rev. B 97,
Commun. 5, 4213 (2014). 245131 (2018).
[6] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru-Guzik, [16] K. Pernal and K. J. H. Giesbertz, Reduced density matrix func-
The theory of variational hybrid quantum-classical algorithms, tional theory (RDMFT) and linear response time-dependent
New J. Phys. 18, 023023 (2016). rdmft (TD-RDMFT), Top. Curr. Chem. 368, 125 (2016).
[7] A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, [17] A. Müller, Explicit approximate relation between reduced two-
J. M. Chow, and J. M. Gambetta, Hardware-efficient variational and one-particle density matrices, Phys. Lett. A 105, 446
quantum eigensolver for small molecules and quantum mag- (1984).
nets, Nature (London) 549, 242 (2017). [18] E. Kamil, R. Schade, T. Pruschke, and P. E. Blöchl, Reduced
[8] J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y. Li, E. Grant, density-matrix functionals applied to the Hubbard dimer, Phys.
L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, The Rev. B 93, 085141 (2016).
Variational Quantum Eigensolver: A review of methods and [19] P. E. Blöchl, C. F. J. Walther, and T. Pruschke, Method to
best practices, arXiv:2111.05176. include explicit correlations into density-functional calculations
[9] P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero, based on density-matrix functional theory, Phys. Rev. B 84,
J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tranter, N. 205101 (2011).
Ding, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, [20] T. Baldsiefen, A. Cangi, and E. K. U. Gross, Reduced-density-
A. G. Fowler, E. Jeffrey, E. Lucero, A. Megrant, J. Y. Mutus, et matrix-functional theory at finite temperature: Theoretical
al., Scalable Quantum Simulation of Molecular Energies, Phys. foundations, Phys. Rev. A 92, 052514 (2015).
Rev. X 6, 031007 (2016). [21] M. Powell, A method for nonlinear constraints in minimization
[10] Y. Shen, X. Zhang, S. Zhang, J.-N. Zhang, M.-H. Yung, and K. problems, in Optimization (Academic Press, New York, 1969),
Kim, Quantum implementation of the unitary coupled cluster pp. 283–298.

033160-14
PARALLEL QUANTUM CHEMISTRY ON NOISY … PHYSICAL REVIEW RESEARCH 4, 033160 (2022)

[22] M. R. Hestenes, Multiplier and gradient methods, J. Optim. [39] R. Karp, Reducibility among combinatorial problems, in Com-
Theory Appl. 4, 303 (1969). plexity of Computer Computations, edited by R. Miller and J.
[23] S. Bravyi, J. M. Gambetta, A. Mezzacapo, and K. Temme, Thatcher (Plenum Press, New York, 1972) pp. 85–103.
Tapering off qubits to simulate fermionic Hamiltonians, [40] A. A. Hagberg, D. A. Schult, and P. J. Swart, Exploring network
arXiv:1701.08213. structure, dynamics, and function using networkx, in Proceed-
[24] M. C. Gutzwiller, Effect of Correlation on the Ferromagnetism ings of the 7th Python in Science Conference, edited by G.
of Transition Metals, Phys. Rev. Lett. 10, 159 (1963). Varoquaux, T. Vaught, and J. Millman ( Pasadena, CA, 2008)
[25] J. Hubbard, Electron correlations in narrow energy bands, Proc. pp. 11–15.
R. Soc. London A 276, 238 (1963). [41] D. Gottesman, Stabilizer codes and quantum error correction,
[26] J. Kanamori, Electron correlation and ferromagnetism of tran- Ph.D. thesis, California Institute of Technology, 1997.
sition metals, Prog. Theor. Phys. 30, 275 (1963). [42] S. Aaronson and D. Gottesman, Improved simulation of stabi-
[27] R. Schade, E. Kamil, and P. Blöchl, Reduced density-matrix lizer circuits, Phys. Rev. A 70, 052328 (2004).
functionals from many-particle theory, Eur. Phys. J.: Spec. Top. [43] Md. S. Anis, Abby-Mitchell, H. Abraham, AduOffei,
226, 2677 (2017). R. Agarwal et al., Qiskit: An open-source framework for
[28] E. Prodan and W. Kohn, Nearsightedness of electronic matter, quantum computing (2021), https://github.com/Qiskit/qiskit/
Proc. Natl. Acad. Sci. USA 102, 11635 (2005). blob/master/Qiskit.bib.
[29] R. Schade, C. Bauer, K. Tamoev, L. Mazur, C. Plessl, and T. D. [44] M. Fishman, S. R. White, and E. M. Stoudenmire, The
Kühne, pc2/rdmft_on_qc: Implementation and results, Zenodo ITensor software library for tensor network calculations,
(2021). arXiv:2007.14822.
[30] M. Piris, Global Method for Electron Correlation, Phys. Rev. [45] J. Lado, DMRGPY, https://github.com/joselado/dmrgpy (2021).
Lett. 119, 063002 (2017). [46] IBM Quantum, https://quantum-computing.ibm.com/ (2021).
[31] J. C. Spall, Multivariate stochastic approximation using a si- [47] T. D. Kühne, M. Krack, F. R. Mohamed, and M. Parrinello,
multaneous perturbation gradient approximation, IEEE Trans. Efficient and Accurate Car-Parrinello-like Approach to Born-
Automat. Contr. 37, 332 (1992). Oppenheimer Molecular Dynamics, Phys. Rev. Lett. 98, 066401
[32] J. C. Spall, A stochastic approximation technique for generating (2007).
maximum likelihood parameter estimates, in 1987 American [48] V. Rengaraj, M. Lass, C. Plessl, and T. D. Kühne, Accu-
Control Conference (IEEE, Minneapolis, MN, USA, 1987) pp. rate sampling with noisy forces from approximate computing,
1161–1167. Computation 8, 39 (2020).
[33] R. Schade, New methods for the ab-initio simulation of [49] T. D. Kühne and E. Prodan, Disordered crystals from first prin-
correlated systems, Ph.D. thesis, Georg-August University, ciples I: Quantifying the configuration space, Ann. Phys. 391,
Göttingen, 2019, http://dx.doi.org/10.53846/goediss-7327. 120 (2018).
[34] P. Jordan and E. P. Wigner, Über das paulische äquivalenzver- [50] I. O. Sokolov, P. K. Barkoutsos, L. Moeller, P. Suchsland,
bot, in The Collected Works of Eugene Paul Wigner: Part A: The G. Mazzola, and I. Tavernelli, Microcanonical and finite-
Scientific Papers, edited by A. S. Wightman (Springer, Berlin, temperature ab initio molecular dynamics simulations on
1993) pp. 109–129. quantum computers, Phys. Rev. Research 3, 013125 (2021).
[35] J. T. Seeley, M. J. Richard, and P. J. Love, The Bravyi-Kitaev [51] T. D. Kühne, M. Iannuzzi, M. Del Ben, V. V. Rybkin, P.
transformation for quantum computation of electronic structure, Seewald, F. Stein, T. Laino, R. Z. Khaliullin, O. Schütt, F.
J. Chem. Phys. 137, 224109 (2012). Schiffmann et al., Cp2k: An electronic structure and molecular
[36] S. B. Bravyi and A. Y. Kitaev, Fermionic quantum computation, dynamics software package - quickstep: Efficient and accurate
Ann. Phys. 298, 210 (2002). electronic structure calculations, J. Chem. Phys. 152, 194103
[37] A. F. Izmaylov, T.-C. Yen, R. A. Lang, and V. Verteletskyi, (2020).
Unitary partitioning approach to the measurement problem in [52] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T.
the variational quantum eigensolver method, J. Chem. Theory Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser,
Comput. 16, 190 (2020). J. Bright et al., SciPy 1.0: Fundamental algorithms for scientific
[38] P. Gokhale, O. Angiuli, Y. Ding, K. Gui, T. Tomesh, M. computing in python, Nat. Methods 17, 261 (2020).
Suchara, M. Martonosi, and F. T. Chong, Minimizing state [53] A. R. Conn, N. I. M. Gould, and P. L. Toint, Trust Region
preparations in variational quantum eigensolver by partitioning Methods (SIAM, Philadelphia, PA, 2000) https://epubs.siam.
into commuting families, arXiv:1907.13623. org/doi/pdf/10.1137/1.9780898719857.

033160-15