Abstract
Hybrid quantumclassical embedding methods for correlated materials simulations provide a path towards potential quantum advantage. However, the required quantum resources arising from the multiband nature of d and f electron materials remain largely unexplored. Here we compare the performance of different variational quantum eigensolvers in ground state preparation for interacting multiorbital embedding impurity models, which is the computationally most demanding step in quantum embedding theories. Focusing on adaptive algorithms and models with 8 spinorbitals, we show that state preparation with fidelities better than 99.9% can be achieved using about 2^{14} shots per measurement circuit. When including gate noise, we observe that parameter optimizations can still be performed if the twoqubit gate error lies below 10^{−3}, which is slightly smaller than current hardware levels. Finally, we measure the ground state energy on IBM and Quantinuum hardware using a converged adaptive ansatz and obtain a relative error of 0.7%.
Introduction
Eigenstate preparation for Hamiltonian systems is one promising application of noisy intermediatescale quantum (NISQ) computers to achieve practical quantum advantage^{1,2,3,4,5,6,7}. One of the representative hybrid quantumclassical algorithms to achieve this task is the variational quantum eigensolver (VQE). It attempts to find the ground state of a given Hamiltonian H within a variational manifold of states that are generated by parametrized quantum circuits U(θ) acting on a reference state \(\left\vert {{{\Psi }}}_{0}\right\rangle\). The parameters θ are obtained by classically minimizing the energy cost function \(E({{{{{{{\boldsymbol{\theta }}}}}}}})=\left\langle {{{\Psi }}}_{0} {U}^{{{{\dagger}}} }({{{{{{{\boldsymbol{\theta }}}}}}}})HU({{{{{{{\boldsymbol{\theta }}}}}}}}) {{{\Psi }}}_{0}\right\rangle\) that is measured on quantum hardware^{2,3,4,8}. The quality of a VQE calculation is tied to the ability of the variational ansatz to represent the ground state with high fidelity. In quantum computational chemistry, the unitary coupled cluster ansatz truncated at single and double excitations (UCCSD) has been extensively studied, owing to the success of the classical coupled cluster algorithm^{9,10,11}. It was found that the application of UCCSD ansatz is limited by the rapid circuit growth with system size and the deteriorating accuracy in the presence of static electron correlations^{8,12,13}. Therefore, alternative variants have been developed, including hardwareefficient ansätze, that improve the trainability and expressivity of the wave function ansatz^{3,13,14,15,16,17,18,19,20}.
Indeed, it was found that compact and numerically exact variational ground state ansätze can be adaptively constructed for specific problems using approaches like the adaptive derivativeassembled pseudotrotter (ADAPT) ansatz^{13,16}. The adaptive ansatz is typically obtained by successively appending parametrized unitaries to a variational circuit with generators chosen from a predefined operator pool. In practice, the ADAPTVQE algorithm works well with an operator pool composed of fermionic excitation operators in the UCCSD ansatz. The extended qubitADAPT VQE approach^{16} utilizes an operator pool composed of Pauli strings in the qubit representation of fermionic excitation operators in the UCCSD ansatz, which is shown to be capable of generating significantly more compact ansätze than the original ADAPTVQE method at the price of introducing more variational parameters. As the circuit complexity (i.e., the number of twoqubit operations in the circuit) is a determining factor for practical calculations on NISQ devices, qubitADAPT is preferable and chosen for the comparative study in this work. Regarding the scalability of the qubitADAPT method towards larger system sizes, we note that reference^{18} reports a favorable linear systemsize scaling for the adaptive ansatz complexity of nonintegrable mixedfield Ising model using the adaptive variational quantum imaginary time evolution method (AVQITE). AVQITE is known to generate variational circuits of comparable complexity as qubitADAPT VQE. As a first step to investigate the scalability in fermionic models, we here study qubitADAPT VQE for fermionic models with two and three spinful orbitals.
An alternative approach to constructing efficient wavefunction ansätze for problems in condensed matter physics is to exploit the sparsity of the Hamiltonian. Interacting electron systems are often simulated with reduced degrees of freedom, represented, for example, by a singleband Hubbard model. This simplified model features a sparse Hamiltonian including nearestneighbor hopping and onsite Coulomb interactions only. Motivated by the simplicity of the Trotterized circuits for dynamics simulations due to Hamiltonian sparsity, the Hamiltonian variational ansatz (HVA) has been proposed by promoting the time in Trotter circuits to independent variational parameters^{21}. The HVA ansatz has attracted much attention and turns out to be very successful in reaching a compact state representation for sparse Hamiltonian systems including local spin models^{21,22,23}. Here, we propose to combine the flexibility of an adaptive approach with the efficiency of the HVA by designing a “Hamiltonian commutator” (HC) operator pool that contains pairwise commutators of operators that appear in the Hamiltonian.
To obtain a realistic description of correlated quantum materials, which typically contain partially filled dorbitals such as transition metal compounds, or forbitals such as rareearth and actinide systems, it is important to go beyond the singleorbital description of a simple Hubbard model^{24}. Intriguing physics arises from the local Hund’s coupling of electrons in different atomic orbitals. Examples are bad metallic behavior with suppressed quasiparticle coherence and orbitalselective Mott transitions or superconducting pairing, which naturally require a multiorbital description^{25,26,27,28}. A multiorbital model including additional interorbital hoppings and Hund’s couplings will necessarily make the Hamiltonian less sparse and consequently the HVA ansatz more complicated. Nevertheless, the complexity of material simulations can be greatly reduced by quantum embedding methods which map the infinite system to coupled subsystems, typically a noninteracting effective medium and some manybody interacting impurity models^{24,29,30,31,32,33,34,35,36,37}. These quantum embedding approaches have proven to be very effective to simulate correlated electron systems, including energies, electronic structure, magnetism, superconductivity, and spectral properties of multiple competing phases. The computational load in these approaches is shifted from the solution of a full lattice system to that of an interacting multiorbital impurity model. Classical algorithms for solving the impurity problem, however, are not scalable, which can be more tractable with quantum computers^{35,38}.
In this paper, we compare the VQE circuit complexity for ground state preparation of multiorbital manybody impurity models with a fixed HVA versus a qubitADAPT ansatz with different operator pools. An HC operator pool compatible with HVA is proposed to allow a fair comparison between qubitADAPT and fixed ansatz HVA calculations. For comparison, we also include results from UCCSD and qubitADAPT calculations with a simplified UCCSD pool. To connect with quantum embedding methods for realistic materials simulations, we use the Gutzwiller embedding approach^{33,39,40,41,42,43,44} to generate the impurity models that we employ for our benchmark^{35,45}. The quantum calculation we perform is general and could also be applied to other embedding methods. Numerical results from a noiseless state vector simulator and quantum assembly language (QASM)based simulator with quantum sampling noise are presented. Important techniques for efficient circuit simulations of qubitADAPT VQE are discussed, including ways to simplify generators and reduce the operator pool size. We further investigate the impact of realistic gate noise by performing qubitADAPT VQE simulations with a realistic noise model including amplitude and dephasing channels. Finally, we measure the energy cost function of the converged VQE ansatz for the e_{g} model composed of eight spinorbitals on the IBM quantum processing unit (QPU) ibmq_casablanca and on Quantinuum hardware.
Results and discussion
Quantum embedding model
Here we focus on a specific quantum embedding method: the wellestablished Gutzwiller variational embedding approach for correlated material simulations^{33,39,40,41,42,43,44}, which is known to be equivalent to rotationally invariant slaveboson theory at the saddle point approximation^{46,47}. Recently, our group has developed a hybrid Gutzwiller quantumclassical embedding approach (GQCE)^{35}. GQCE maps the ground state solution of a correlated electron lattice system to a coupled eigenvalue problem of a noninteracting quasiparticle Hamiltonian and one or multiple finitesize interacting embedding Hamiltonians^{44}. Within GQCE one employs a quantum computer to find the ground state energy and the singleparticle density matrix of the interacting embedding Hamiltonian, for example, using VQE.
The embedding Hamiltonian describes an impurity model consisting of a physical manybody \({N}_{{{{{{{{\mathcal{S}}}}}}}}}\)orbital subsystem (\({\hat{{{{{{{{\mathcal{H}}}}}}}}}}_{{{{{{{{\mathcal{S}}}}}}}}}\)) coupled with a \({N}_{{{{{{{{\mathcal{B}}}}}}}}}\)orbital quadratic bath (\({\hat{{{{{{{{\mathcal{H}}}}}}}}}}_{{{{{{{{\mathcal{B}}}}}}}}}\)):
with
Here α, β, γ, δ are composite indices for sites and spatial orbitals in the physical subsystem. Likewise, the bath sites and orbitals are labeled by a, b, and σ is the spin index. The fermionic ladder operators \({\hat{c}}^{{{{\dagger}}} }\) and \({\hat{f}}^{{{{\dagger}}} }\) are used to distinguish the physical and bath orbital sites.
The onebody component and twobody Coulomb interaction in the physical subsystem are specified by matrix ϵ and tensor V. The quadratic bath and its coupling to the subsystem are defined by matrix λ and \({{{{{{{\mathcal{D}}}}}}}}\), respectively. Compared with typical quantum chemistry calculations, the embedding Hamiltonian is much sparser since the twobody interaction only exists between electrons in the physical subsystem.
For clarification, we name the abovedefined embedding Hamiltonian system as (\({N}_{{{{{{{{\mathcal{S}}}}}}}}},{N}_{{{{{{{{\mathcal{B}}}}}}}}}\)) impurity model, where (\({N}_{{{{{{{{\mathcal{S}}}}}}}}},{N}_{{{{{{{{\mathcal{B}}}}}}}}}\)) is the number of spatial orbitals in the system and bath models. Within GQCE, the ground state solution of the embedding Hamiltonian at halfelectron filling is needed, which is achieved by a chemical potential absorbed in the onebody Hamiltonian coefficient matrices ϵ and λ in Eq. (1).
In the numerical simulations presented here, we choose a Gutzwiller embedding Hamiltonian for the degenerate \({{{{{{{\mathcal{M}}}}}}}}\)band Hubbard model. The noninteracting density of states of the lattice model adopts a semicircular form \(\rho (\omega )=\frac{2{{{{{{{\mathcal{M}}}}}}}}}{\pi D}\sqrt{1{(\omega /D)}^{2}}\) as shown in Fig. 1a, which corresponds to the Bethe lattice in infinite dimensions. In the following, we set the half band width D = 1 as the energy unit. In physical systems, D is of the order of a few eV. The Coulomb matrix V takes the Kanamori form specified by Hubbard U and Hund’s J parameters: V_{αααα} = U, V_{ααββ} = U−2J, and V_{αβαβ} = V_{αββα} = J for α ≠ β. Here we have assumed spin and orbital rotational invariance (within the e_{g} or t_{2g} manifold) for simplicity and to limit the interaction parameter space.
The embedding Hamiltonian, as illustrated in Fig. 1b, is represented with \(2{{{{{{{\mathcal{M}}}}}}}}\) spatial orbitals: \({{{{{{{\mathcal{M}}}}}}}}\) degenerate physical orbital plus \({{{{{{{\mathcal{M}}}}}}}}\) degenerate bath orbitals. The symmetry of the model reduces matrices ϵ, λ and \({{{{{{{\mathcal{D}}}}}}}}\) to single parameters proportional to identity.
In the following, we set the electron filling for the lattice model to \({{{{{{{\mathcal{M}}}}}}}}+1\), which is one unit larger than halffilling, and fix the ratio of the Hund’s to Hubbard interaction to J/U = 0.3 and U = 7. These parameters put the model deep in the correlationinduced bad metallic state, with physical properties distinct from doped Mott insulators^{25}. It represents a wide class of strongly correlated materials, such as iron pnictides and chalcogenides, where Hund’s coupling significantly reduces the lowenergy quasiparticle coherence scale^{26,48,49}. Hund’s metal physics is far beyond a static meanfield description and requires treating the localized and itinerant characters of electrons on equal footing, which can be realized in the quantum embedding approach adopted here.
In the calculations below, we consider \({{{{{{{\mathcal{M}}}}}}}}=2\) and \({{{{{{{\mathcal{M}}}}}}}}=3\), which correspond to e_{g} and t_{2g} orbitals in cubic crystal symmetry, respectively. The associated \(({N}_{{{{{{{{\mathcal{S}}}}}}}}},{N}_{{{{{{{{\mathcal{B}}}}}}}}})=(2,2)\) and (3, 3) impurity models have in total 8 and 12 spinorbitals. The two models host nontrivial manybody ground states and represent important checkpoints along the path to achieve a practical quantum advantage in correlated materials simulations through a hybrid quantumclassical embedding framework. In quantum simulations reported below, parity encoding which exploits the symmetry in a total number of electrons and spin zcomponent is used to transform the fermionic Hamiltonian to qubit representation.
Variational quantum eigensolvers
GQCE leverages quantum computing technologies to solve for the ground state of the embedding Hamiltonian, specifically the energy and oneparticle density matrix. Note that the ground state is always prepared at halffilling for the embedding system, which is determined by the Gutzwiller embedding algorithm and is independent of the actual electron filling of the physical lattice model^{33,44}. For this purpose, we benchmark multiple versions of VQE with fixed or adaptively generated ansatz to prepare the ground state of the above embedding Hamiltonian. We consider VQE calculations with fixed UCCSD ansatz and the associated qubitADAPT VQE using a simplified UCCSD operator pool. The calculations are naturally performed in the molecular orbital (MO) basis representation, where the reference Hartree–Fock (HF) state becomes a simple tensor product state and fermionic excitation operators can be naturally defined. However, using a MO representation comes at the cost of reducing the sparsity of the embedding Hamiltonian compared to the atomic orbital (AO) basis representation. To take advantage of the Hamiltonian sparsity in AO representation, we consider a generalized form of the HVA and the associated qubitADAPT VQE with a modified HC operator pool.
VQE algorithm
For an N_{q}qubit system with Hamiltonian \(\hat{{{{{{{{\mathcal{H}}}}}}}}}\), VQE amounts to minimizing the cost function \(E({{{{{{{\boldsymbol{\theta }}}}}}}})=\langle {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}] \,\hat{{{{{{{{\mathcal{H}}}}}}}}}\, {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\rangle\) with respect to the variational parameters θ, as schematically illustrated in Fig. 2. Here, \(\left\vert {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\right\rangle =\hat{U}({{{{{{{\boldsymbol{\theta }}}}}}}})\left\vert {{{\Psi }}}_{0}\right\rangle\) is obtained by application of a parametrized quantum circuit \(\hat{U}({{{{{{{\boldsymbol{\theta }}}}}}}})\) onto a reference state \(\left\vert {{{\Psi }}}_{0}\right\rangle\). The cost function is evaluated on a quantum computer and the optimization is performed classically using E(θ) as input. The accuracy of VQE is therefore tied to the variational ansatz \(\left\vert {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\right\rangle\) and to the performance of the classical optimization, e.g., how often the cost function is called during the optimization and how well the approach converges to the global (as opposed to a local) minimum of E(θ).
UCCSD ansatz
The UCCSD ansatz takes the following form:
The operator \(\hat{T}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\) consists of single and double excitation operators with respect to the HF reference state \(\left\vert {{{\Psi }}}_{0}\right\rangle\):
Here p, q and \(\bar{p},\bar{q}\) refer to the occupied and unoccupied MOs, respectively, with spin included implicitly. \({f}_{j}(\{\hat{\sigma }\})={\sum }_{k}{w}_{jk}{\hat{P}}_{k}\) is a weighted sum of Pauli strings (\({\hat{P}}_{k}\in {\{I,X,Y,Z\}}^{\otimes {N}_{q}}\)) for the qubit representation of the fermionic excitation operator associated with parameter θ_{j}. Here θ_{j} runs over the set of parameters \({\theta }_{p}^{\bar{p}}\) and \({\theta }_{pq}^{\bar{p}\bar{q}}\). For the impurity model without spin–orbit interaction, only excitation operators which conserve a respective number of electrons in the spinup and spindown sectors need to be considered. In practical implementation, a singlestep Trotter approximation is often adopted to construct the UCCSD circuit:
Furthermore, the final circuit state generally depends on the order of the unitary gates. In the calculations reported here, we apply gates with singleexcitation operators first following the implementation in Qiskit^{50}.
QubitADAPT VQE with simplified UCCSD pool
VQEUCCSD is a useful reference point for quantum chemistry calculations. However, the fixed UCCSD ansatz has limited accuracy and often involves deep quantum circuits for implementations. Various approaches have been proposed to construct a more compact variational ansatz with systematically improvable accuracy. In this work, we will focus on the qubitADAPT VQE method^{16}, where the ansatz takes a similar pseudoTrotter form:
With qubitADAPT, the ansatz is recursively expanded by adding one unitary at a time, followed by reoptimization of parameters. The additional unitary is constructed with a generator selected from a predefined Pauli string pool which gives maximal energy gradient amplitude \( g{ }_{\max }\) at the preceding ansatz state. The ansatz expansion process iterates until convergence, which is set by \( g{ }_{\max } < 1{0}^{4}\) here. Note that we have set the half bandwidth of the original noninteracting lattice model to D = 1, such that \( g{ }_{\max } \sim 0.1\) meV in physical systems with D ~ 1 eV.
The computational complexity of qubitADAPT VQE calculations is tied to the size of the operator pool, which consists of a set of Pauli strings. Naturally, one can construct an operator pool using all the Pauli strings in the qubit representation of fermionic single and double excitation operators. However, the dimension of this UCCSDcompatible pool is usually quite big and scales as \({{{{{{{\mathcal{O}}}}}}}}({N}_{q}^{4})\). Here we propose a muchsimplified operator pool, which consists of Pauli strings from singleexcitation and paired doubleexcitation operators only. The pair excitation involves a pair of electrons with opposite spins, which are initially occupying the same spatial MO, hopping together to another initially unoccupied spatial MO. To further reduce the circuit depth, only one Pauli string is chosen from each qubit representation of the fermionic excitation operator. The qubit representation is a weighted sum of equallength Pauli strings, and a specific choice of which one of them does not seem to be important in practical calculations reported here. This simplified pool containing operators arising from the UCC ansatz restricted to single and paired double excitation operators (sUCCSpD)^{51,52} greatly reduces the number of Pauli strings compared to the UCCSD pool. The dimension of this sUCCSpD pool scales as \({{{{{{{\mathcal{O}}}}}}}}({N}_{q}^{2})\). For the (2, 2) e_{g} impurity model, the pool size reduces from 152 for UCCSD to 56 for sUCCSpD, and for the (3, 3) t_{2g} impurity model it reduces from 828 to 192. The code to perform the above qubitADAPT VQE calculations at the state vector level with examples are available in figshare^{53}.
Hamiltonian variational ansatz
The Hamiltonian sparsity in the AO basis naturally motivates the application of the Hamiltonian variational ansatz^{21}, which generally takes the form of multilayer Trotterized annealinglike circuits. While different ways of designing specific HVA forms have been developed, we propose the following ansatz with L layers for the impurity model:
Here \(\hat{{{{{{{{\mathcal{H}}}}}}}}}=\mathop{\sum }\nolimits_{j = 1}^{{N}_{{{{{{{{\rm{G}}}}}}}}}}{\hat{h}}_{j}\), with \({\hat{h}}_{j}\) being a subgroup of Hamiltonian terms which share the same coefficient and mutually commute. Such ansatz construction aims to differentiate the physical and bath orbitals while retaining the degeneracy information among the orbitals in a systematic way. For each layer of unitaries, we first apply the multiqubit rotations that are generated by the interacting part of the Hamiltonian, since these act as entangling gates. For the (\({{{{{{{\mathcal{M}}}}}}}},{{{{{{{\mathcal{M}}}}}}}}\)) impurity model, two reference states have been tried: \(\left\vert {{{\Psi }}}_{0}^{{{{{{{{\rm{(I)}}}}}}}}}\right\rangle\) is a simple tensor product state with \({{{{{{{\mathcal{M}}}}}}}}\) physical orbitals fully occupied and the bath orbitals empty; \(\left\vert {{{\Psi }}}_{0}^{{{{{{{{\rm{(II)}}}}}}}}}\right\rangle\) is the ground state of the noninteracting part of \(\hat{{{{{{{{\mathcal{H}}}}}}}}}\), which is equivalent to the oneelectron core Hamiltonian in quantum chemistry. We did not find any significant difference between the two choices of reference state in practical simulations of the impurity models. Therefore, only HVA calculations with the reference state \(\left\vert {{{\Psi }}}_{0}^{{{{{{{{\rm{(I)}}}}}}}}}\right\rangle\) are reported here. We adopt the gradientbased Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm as the classical optimizer. Proper parameter initialization for HVA optimization is crucial, as barren plateaus and local energy minima are generally present in the variational energy landscape. In practice, we find that a uniform initialization of the parameters, such as setting all to π/7, overall works well for simulations reported here.
Inspired by the idea of adaptive ansatz generation^{13}, we also tried constructing and optimizing an Llayer HVA ansatz by adaptively adding layers from 1 to L. Specifically, the calculation starts with optimizing a singlelayer ansatz, followed by appending another layer to the ansatz while keeping the first layer at previously obtained optimal angles. The twolayer ansatz is then optimized with the parameters for the new layer initialized randomly or uniformly. The procedure continues with the optimization of llayer ansatz leveraging the (l−1)layer solution until the ansatz reaches L layers.
Let the number of cost function evaluations for optimizing an llayer ansatz be \({N}_{l}^{(2)}\). The total number of function evaluations amounts to \({N}^{(2)}=\mathop{\sum }\nolimits_{l = 1}^{L}{N}_{l}^{(2)}\). In practice, we find that the direct optimization of the Llayer ansatz using a uniform initialization takes N^{(1)} function evaluations with \({N}^{(1)} \sim {N}_{L}^{(2)} < {N}^{(2)}\), and reaches the same accuracy. Starting with L layers is therefore more efficient than growing the ansatz layer by layer.
Intuitively, this can be related to the fact that successive HVA optimization introduces discontinuities in the variational path toward the ground state whenever a new layer of unitaries is added. Since the energy gradient associated with new variational parameters that are initialized to zero (for continuity) vanishes (see the “Methods” section), they have to be initialized away from zero. In other words, the (l−1)layer HVA solution is not a good starting point for the optimization of the llayer ansatz. The open source code to perform the above HVA calculations at the state vector level with examples are available in figshare^{54}.
Hamiltonian commutator pool
It has been demonstrated that the qubitADAPT VQE in the MO basis outperforms VQEUCCSD calculations regarding circuit complexity and numerical accuracy^{13,16}. Motivation by this observation, we compare the corresponding qubitADAPT VQE with a Hamiltoniancompatible pool in AO basis and HVA calculations. Following HVA, we choose the simple tensor product state \(\left\vert {{{\Psi }}}_{0}^{{{{{{{{\rm{(I)}}}}}}}}}\right\rangle\) as the reference state. In the qubitADAPT step, the energy gradient criterion \({g}_{\theta }=2{{{{{\mathrm{Im}}}}}} [\langle {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}] \,\hat{P}\hat{{{{{{{{\mathcal{H}}}}}}}}}\, {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\rangle ]\) to append a new unitary generated by \(\hat{P}\) vanishes due to symmetry with Ψ[θ], if the number of PauliY operators in the Pauli string \(\hat{P}\) is even^{13,55}. This can be simply shown by the following argument. Because the impurity model in this study respects timereversal symmetry and spinflip (Z_{2}) symmetry, both Hamiltonian \(\hat{{{{{{{{\mathcal{H}}}}}}}}}\) and wavefunction are real (\(\hat{{{{{{{{\mathcal{H}}}}}}}}}={\hat{{{{{{{{\mathcal{H}}}}}}}}}}^{* },{{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]={{\Psi }}{[{{{{{{{\boldsymbol{\theta }}}}}}}}]}^{* }\)). The Pauli string \(\hat{P}\) is also real (\(\hat{P}={\hat{P}}^{* }\)) if it has an even number of PauliY operators. Consequently, the expectation value of \(\langle {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}] \,\hat{P}\hat{{{{{{{{\mathcal{H}}}}}}}}}\, {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\rangle\) is real and g_{θ} vanishes if the associated generator \(\hat{P}\) has an even number of PauliY operators.
By construction, the sUCCSpD pool consists of Pauli strings of an odd number of Y’s. However, the Hamiltonian of the impurity models studied here is all real. Consequently, all the Pauli strings in the qubit representation of the Hamiltonian contain an even number of Y’s, which excludes the option of directly constructing the operator pool from the Hamiltonian operators. Nevertheless, the practical usefulness of HVA implies that the Hamiltonianlike pool can be constructed by commuting the Hamiltonian terms, which we call the Hamiltonian commutator (HC) pool \({{{{{{{{\mathscr{P}}}}}}}}}_{{{{{{{{\rm{HC}}}}}}}}}\). Mathematically \({{{{{{{{\mathscr{P}}}}}}}}}_{{{{{{{{\rm{HC}}}}}}}}}\) is constructed in the following manner:
Here \({{{{{{{{\mathscr{P}}}}}}}}}_{{{{{{{{\rm{H}}}}}}}}}\) is the set of Pauli strings \(\{\hat{{P}_{h}}\}\) present in the qubit representation of Hamiltonian \(\hat{{{{{{{{\mathcal{H}}}}}}}}}={\sum }_{h}{w}_{h}{\hat{P}}_{h}\). \({N}_{Y}(\hat{P})\) counts the number of Y operators in the Pauli string \(\hat{P}\). Therefore, the size of \({{{{{{{{\mathscr{P}}}}}}}}}_{{{{{{{{\rm{HC}}}}}}}}}\) can scale as \({N}_{{{{{{{{\rm{H}}}}}}}}}^{2}\), where N_{H} is the total number of Hamiltonian terms. Clearly, the pool \({{{{{{{{\mathscr{P}}}}}}}}}_{{{{{{{{\rm{HC}}}}}}}}}\) should only be applied to sparse Hamiltonian systems. The dimension of the HC pool is 56 for the e_{g} impurity model, and 192 for the t_{2g} model.
Quantum circuit implementation
Performing a calculation on a quantum computer always needs to deal with the presence of noise. Even for ideal faulttolerant quantum computers, quantum sampling (or shot) noise is present due to a finite number of measurements that are used to estimate expectation values. The current noisy quantum devices exhibit additional noise originating from qubit relaxation and dephasing as well as hardware imperfections when implementing unitary gate operations. In this subsection, we describe several techniques adopted in our simulations to most efficiently use the available quantum resources and stabilize the calculations against sampling noise. We discuss how to mitigate gate noise in the final subsection.
Measurement circuit reduction
The quantum circuit implementation for VQE and its adaptive version amounts to the direct measurement of the Hamiltonian as a weighted sum of Pauli string expectation values, \(\langle \hat{{{{{{{{\mathcal{H}}}}}}}}}\rangle ={\sum }_{h}{w}_{h}\langle {\hat{P}}_{h}\rangle\), with respect to parametrized circuits U[θ]. Here, \(\hat{{{{{{{{\mathcal{H}}}}}}}}}={\sum }_{h}{w}_{h}{\hat{P}}_{h}\) is the Hamiltonian in qubit representation. Because the number of shots (or repeated measurements) scales with the desired precision ϵ as \({N}_{{{{{{{{\rm{sh}}}}}}}}}\propto \frac{1}{{\epsilon }^{2}}\) due to the central limit theorem, N_{sh} is often huge in practical calculations. Therefore, it is desirable to group the Pauli strings into mutually commuting sets such that the number of distinct measurement circuits is reduced to a minimum. Indeed, many techniques to achieve such measurement reduction have been developed^{56,57,58,59,60,61}. In this work, we adopt the measurement reduction strategy based on the Hamiltonian integral factorization^{61}, which shows a favorable linear systemsize scaling of the number of distinct measurement circuits and embraces a diagonal representation for the operators to be measured.
Specifically, we transform the physical subsystem Hamiltonian as follows:
with \({\widetilde{\epsilon }}_{\alpha \beta }={\epsilon }_{\alpha \beta }\frac{1}{2}{\sum }_{\gamma }{V}_{\alpha \gamma \gamma \beta }\). A typical way to simplify the measurement of the twobody terms \({\hat{{{{{{{{\mathcal{H}}}}}}}}}}_{{{{{{{{\mathcal{S}}}}}}}}}^{(2)}\) in Eq. (11) is to perform nested matrix factorization for the Coulomb V tensor. Namely, we first rewrite \({\hat{{{{{{{{\mathcal{H}}}}}}}}}}_{{{{{{{{\mathcal{S}}}}}}}}}^{(2)}\) in the following factorized form by diagonalizing the real symmetric positive semidefinite supermatrix V_{(αβ),(γδ)}:
Here l runs through the L positive eigenvalues of the supermatrix V, and the lth component of the auxiliary tensor \({{{{{{{\mathcal{L}}}}}}}}\) is obtained by multiplying the lth eigenvector with the square root of lth positive eigenvalue. Each tensor component, \({{{{{{{{\mathcal{L}}}}}}}}}^{(l)}\), which is a real symmetric matrix, is subsequently diagonalized to reach the following decomposition:
Here, we have defined \({\hat{n}}_{m\sigma }^{(l)}\equiv {\sum}_{\alpha \beta }{U}_{\alpha m}^{(l)}{U}_{\beta m}^{(l)}{\hat{c}}_{\alpha \sigma }^{{{{\dagger}}} }{\hat{c}}_{\beta \sigma }\). The index m goes through the M_{l} nonzero eigenvalues \({\lambda }_{m}^{(l)}\) and associated eigenvectors \({U}_{m}^{(l)}\), which determines the singleparticle basis transformation for the lth component. The whole embedding Hamiltonian of Eq. (1) can then be cast into the following doublyfactorized form with a unitary transformation similar to Eq. (13) for the onebody part:
which is composed of L + 1 groups characterized by unique singleparticle basis transformations {U^{(l)}}, including one from the singleelectron component. This form allows efficient measurement of the Hamiltonian expectation value using \(L+1\propto {{{{{{{\mathcal{O}}}}}}}}(N)\) distinct circuits for a generic quantum chemistry problem with a singleparticle basis dimension given by N.
The expectation value of \(\hat{{{{{{{{\mathcal{H}}}}}}}}}\) is obtained by measuring each group l independently in the variational state \(\left\vert {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\right\rangle\). The variational state is transformed to the same representation used in the lth group by applying a series of Givens rotations, \(\{{e}^{{\theta }_{\mu \nu }({\hat{c}}^{{{{\dagger}}} }_{\mu \sigma }{\hat{c}}_{\nu \sigma }h.c.)}\}\), with the set of {θ_{μν}} determined by the singleparticle transformation matrix U^{(l)}. Here μ and ν are generic indices for physical and bath orbital sites. Therefore, the number of distinct measurement circuits is N_{c} = L + 1. As an example, we have N_{c} = 4 for e_{g} model. We refer to the “Methods” section for further details.
In practice, it is advantageous to isolate the onebody and twobody terms that contain only density operators before the double factorization procedure, because they are already in a diagonal representation. For the e_{g} model we have carried out the doublefactorization with explicit calculations in the “Methods” section and we ultimately find N_{c} = 3 for the e_{g} model. This can be compared with the Hamiltonian measurement procedure using the mutual qubitwise commuting groups: operators that commute with respect to every qubit site are placed in the same group. This commuting Pauli approach generally needs \({N}_{c}\propto {{{{{{{\mathcal{O}}}}}}}}({N}^{4})\) distinct circuits for Hamiltonian measurement. And for the e_{g} model, it requires N_{c} = 5.
Noiseresilient optimization
Although classical optimization approaches such as BFGS, which rely on a computation of the energy gradient, are effective, they rely on very accurate cost function evaluations. Because of the inherent noise in quantum computing, optimization algorithms that are robust to cost function noise are highly desirable. In the noisy quantum simulations reported here, we adopt two optimization techniques that are more tolerant to noise than BFGS: sequential minimal optimization (SMO)^{62} and Adadelta^{63}. Because of their similar performance in the noisy simulations, we only discuss SMO in the main text and leave the discussions of Adadelta in the “Methods” section.
SMO is the first technique we use for our noisy quantum simulations. Tailored to the qubitADAPT ansatz of Eq. (8) where each variational parameter is associated with a single Pauli string generator, the optimization consists of N_{sw} sweeps of sequential single parameter minimization of the cost function. At a specific optimization step with varying parameter θ_{j}, while keeping others fixed, the cost function has a simple form of \(a\cos (2{\theta }_{j}b)+c\), with the optimal \({\theta }_{j}^{* }=b/2\) if a < 0 and (b + π)/2 otherwise. To determine the parameters a, b, and c, one requires knowledge of function values for at least three mesh points in the range of [−π/2, π/2). In practice, we use eight uniformly spaced mesh points to better mitigate the effect of noise in the cost function. Consequently, least square fitting is used to determine the values of a, b and c. In SMO calculations, we use the number of sweeps as the parameter to control the convergence, which we set to N_{sw} = 40. Alternative control parameters, such as energy and gradient, usually are required to be evaluated at higher precision, which can be challenging and introduce additional quantum computation overhead.
In this work, we perform noisy simulations with classical optimizations that include sampling noise due to a finite number of measurements or shots (N_{sh}) as well as both sampling and gate noise. The purpose is to investigate the performance of the qubitADAPT algorithm in the presence of sampling and gate noise and to separate the effects of sampling noise, which is controlled by a single parameter N_{sh} from the effect of gate noise. The code with the circuit implementation of qubitADAPT VQE with examples on QASM simulator and quantum hardware are available at figshare^{64}.
Statevector simulations
In this section, we present numerical simulation results using a statevector simulator, which is equivalent to a faulttolerant quantum computer with an infinite number of measurements (N_{sh} = ∞). Figure 3 shows the ground state energy calculations of the (2, 2) e_{g} and (3, 3) t_{2g} impurity models using VQEHVA as well as qubitADAPT VQE with sUCCSpD and HC pools. The reference UCCSD energy is 0.029 higher than the exact ground state energy E_{GS} for the e_{g} model and 0.128 higher for the t_{2g} model. This implies that both models are in the strong electron correlation region. For calculations of the e_{g} model, the energy converges below 10^{−5} with N_{θ} = 20 variational parameters for VQEHVA, N_{θ} = 59 for ADAPTsUCCSpD, and N_{θ} = 31 for ADAPTHC. Although the qubitADAPT VQE calculation on a statevector simulator is in principle deterministic, the operator selection from a predefined operator pool can introduce some randomness due to the numerical accuracy and near degeneracy of scores (i.e., the associated gradient components) for some operators. As a result, the converged N_{θ} can slightly change by about one between runs.
As a simple estimation of the circuit complexity for NISQ devices, we provide the number of CNOT gates N_{cx} assuming full qubit connectivity, which can be realized in trapped ion systems. The converged circuit has N_{cx} = 288 for VQEHVA, N_{cx} = 292 for ADAPTsUCCSpD, and N_{cx} = 150 for ADAPTHC. As a reference, the UCCSD ansatz has N_{θ} = 26 and N_{cx} = 1096. The HVA calculation converges with the smallest number of variational parameters, but the number of CNOT gates (N_{cx}) is in between that of ADAPTHC and ADAPTsUCCSpD because each variational parameter in HVA is associated with a generator composed of a weighted sum of Pauli strings. The ADAPTHC calculation starts from a reference state \(\left\vert {{{\Psi }}}_{0}^{{{{{{{{\rm{(I)}}}}}}}}}\right\rangle\), a simple tensor product state on an AO basis, with energy higher than the HF reference state used by ADAPTsUCCSpD, yet ADAPTHC converges faster to the ground state. In fact, the initial state fidelity, defined as f ≡ ∣〈Ψ_{0}∣ Ψ_{GS}〉∣^{2}, is 0.19 for ADAPTHC, compared with 0.76 for ADAPTsUCCSpD. Therefore, the final ansatz complexity does not show a simple positive correlation with the initial state fidelity, which implies that both the Hamiltonian structure and operator pool are determining factors.
Compared with ADAPTsUCCSpD, the advantage of ADAPTHC becomes more prominent when applied to the t_{2g} model. To reach energy convergence below 10^{−5}, ADAPTHC needs N_{θ} = 270 parameters and N_{cx} = 2052 CNOTs, while ADAPTsUCCSpD requires as many as N_{θ} = 1020 parameters and N_{cx} = 8066 CNOTs. For reference, the UCCSD ansatz has N_{θ} = 117 parameters and N_{cx} = 9200 CNOTs. The HVA calculation is carried out with up to L = 10 layers, which amounts to N_{θ} = 70 and N_{cx} = 2420, and the energy converges close to 10^{−6}.
We emphasize that strong electron correlation effects are present in our chosen model that lies deep in the bad metallic state^{48,49}. This state cannot be accurately captured within a meanfield description and hence requires the application of an appreciable number of unitary gates to the reference state. Generally, the circuit depth of a variational ansatz is tied to both the complexity of the problem (i.e. the complexity of the ground state wavefunction) and the desired state fidelity. As shown in Fig. 4, when we require a state fidelity close to 99.9% or an energy error close to 0.001, which is typically necessary for practical calculations, one observes a sharp rise of N_{θ} when the system is tuned from the weak correlation (U < 1) to the strong correlation (U > 2) regime by increasing Hubbard U.
Simulations with shot noise
The ADAPT VQE calculations are often reported at the statevector level, and a systematic study including the effect of noise is not yet available^{13,16,65,66,67}. Here we present qubitADAPT VQE calculations of the (2, 2)e_{g} model including shot noise.
Figure 5 shows the representative convergence behavior of the qubitADAPT energy with an increasing number of variational parameters N_{θ} calculated using different numbers of shots per observable measurement: Fig. 5a is for N_{sh} = 2^{12}, and Fig. 5b is for N_{sh} = 2^{16}. We use SMO for the classical optimization. The adaptive ansatz energy E overall decreases as the circuit grow and more variational parameters are used. The energy uncertainty is tied to the number of shots N_{sh}. The energy spread roughly reduces by a factor of 4 when N_{sh} increases from 2^{12} to 2^{16}, consistent with the 16fold increase in N_{sh} due to the central limit theorem.
The energy points shown include not only the final SMO optimized energies of the qubitADAPT ansatz with N_{θ} parameters but also the intermediate energies after each of the N_{sw} = 40 sweeps during SMO optimizations to provide more detailed convergence information. The abovereported N_{sh} is referred to as measurements for SMO optimizations. At the operator screening step of the qubitADAPT calculation to expand the ansatz by appending an additional optimal unitary, we fix N_{sh} = 2^{16} shots for energy evaluations in all cases and determine the energy gradient by the parametershift rule^{68}.
To further assess the quality of the qubitADAPT ansatz obtained in these QASM simulations, we plot in Fig. 5c the ansatz energies evaluated using a statevector simulator at the end of each noisy SMO optimization. The four solid curves are calculated using the variational parameters that are obtained by QASM optimizations with different numbers of shots N_{sh} as indicated and noiseless optimization results are shown for comparison as the dashed line. While there is no clear order of the energies during the early stages of the simulation, the final convergence is consistently improved with more shots. Specifically, the error converges close to and below 10^{−3} for N_{sh} = 2^{14} and 2^{16} and the fidelity f improves beyond 99.9%. The associated singleparticle density matrix elements also converge to an accuracy better than 10^{−2}.
Similar QASM simulations of qubitADAPT VQE have been performed using the Adadelta optimizer, as specified in the “Methods” section. Generally, we find the numerical results and the dependence on the number of shots to be comparable to SMO. Compared with SMO, Adadelta can potentially take advantage of multiple QPUs by evaluating the gradient vector in parallel.
Discussion of optimal pool size
One important factor determining the computational load of qubitADAPT VQE calculations is the size of the operator pool N_{p}. One simple strategy to reduce N_{p} is to strip off Pauli Z’s in the pool of operators because they contribute negligibly to the ground state energy as pointed out in refs. ^{16,66}. This reduces N_{p} of the Hamiltonian commutator (HC) pool from 56 to 16 for the e_{g} model, and from 192 to 60 for the t_{2g} model, due to a large degeneracy. Furthermore, some qualitative guidance has been laid out in the literature to construct a minimal complete pool (MCP) of size 2(N_{q}−1)^{16,69}, where N_{q} is the number of qubits. Indeed, we find that an MCP can be constructed using a subset of operators in the HC pool.
We discover a dichotomy that the reduction of the pool size can potentially make the optimization of the qubitADAPT ansatz more challenging, especially in the presence of noise. Figure 6 compares qubitADAPT calculations using three different pool sizes of dimensions 56, 16, and 10, which were introduced above. Figure 6a shows the qubitADAPT energies with increasing N_{θ} from statevector simulations of the e_{g} model using the three pools. All the simulations converge with 31 parameters and final CNOT gate numbers N_{cx} = 150, 98, and 62 which decrease for the smaller pools. The details of the convergence rate of the three runs differ significantly. When the pool dimension decreases, the region of N_{θ} with minimal energy change expands, as seen by the almost flat segments of the curves of Fig. 6a. The minimal energy gain implies that small noise in the cost function evaluation could deteriorate the parameter optimization.
Indeed as shown in Fig. 6b, the qubitADAPT energy from noisy simulation converges slower as the pool size decreases. The flat segments in the energy curves become more evident owing to the stochastic energy errors. We further analyze the quality of the qubitADAPT ansatz by evaluating the energy at optimal angles obtained in noisy simulations, as plotted in Fig. 6c. The energy difference is 0.001, 0.027, 0.135 at N_{θ} = 31 where the statevector simulation converges, and 0.0006, 0.001, 0.005 at the end of N_{θ} = 40 for calculations with pools of size 56, 16, and 10, respectively.
Our analysis clearly shows the strikingly distinct convergence behaviors of qubitADAPT calculations using different complete operator pools in the presence of sampling noise. This indicates that the optimal pool in practical calculations can be a tradeoff between choosing a small pool size and guaranteeing sufficient connectivity of the operators in the pool.
Simulations with noise models
Besides the inherent sampling noise in quantum computing, NISQ hardware is subject to various other error effects. These include coherent errors due to imperfect gate operations as well as stochastic errors due to qubit decoherence, dephasing, and relaxation. Here, we perform a preliminary investigation of the impact of hardware imperfections on qubitADAPT VQE calculations by adopting a realistic decoherence noise model proposed by Kandala et al. in ref. ^{3}. The model includes an amplitudedamping channel (\(\rho \to \mathop{\sum }\nolimits_{i = 1}^{2}{E}_{i}^{a}\rho {E}_{i}^{a{{{\dagger}}} }\)) and a dephasing channel (\(\rho \to \mathop{\sum }\nolimits_{i = 1}^{2}{E}_{i}^{d}\rho {E}_{i}^{d{{{\dagger}}} }\)). These act on the qubit density matrix following each singlequbit or twoqubit gate operation. The Kraus operators are given as:
The error rates \({p}^{{\rm {a}}}=1{e}^{\tau /{T}_{1}}\) and \({p}^{{\rm {d}}}=1{e}^{2\tau /{T}_{\phi }}\) are determined by the gate time τ, the qubit relaxation time T_{1} and the dephasing time T_{ϕ} = 2T_{1}T_{2}/(2T_{1}−T_{2}), where T_{2} is the qubit coherence time. For the sake of simplicity of the analysis, we choose a uniform singlequbit gate error rate \({p}_{1}^{{\rm {a}}}={p}_{1}^{{\rm {d}}}\equiv {p}_{1}=1{0}^{4}\), which is close to the value found in current hardware. We also assume a uniform twoqubit error rate \({p}_{2}^{{\rm {a}}}={p}_{2}^{{\rm {d}}}={p}_{2}\) that we vary between 10^{−4} and 10^{−2}, in order to study the impact of twoqubit noise on the VQE optimization.
Figure 7a shows a typical qubitADAPT energy curve E−E_{GS} during optimization as a function of the number of variational parameters N_{θ} obtained in noisy simulations with p_{2} = 10^{−2}, 10^{−3}, and 10^{−4}. Here, E_{GS} is the exact ground state energy. The results with only singlequbit noise are also shown for reference. Figure 7b contains the associated exact energies for the ansatz states, which we obtain by evaluating the VQE ansatz on a statevector simulator.
For p_{2} = 10^{−2}, which represents the current hardware noise level, the noisy energy increases with N_{θ}, indicating that the error rate is too large to get reliable energy estimation. Nevertheless, as shown in the corresponding statevector analysis in Fig. 7b, one still observes a sizable energy reduction in the early stage of the optimization. The evaluated ansatz state fidelity is found to improve from 0.19 in the initial state to about 0.70 with 4 < N_{θ} < 9. When further increasing N_{θ}, however, the statevector ansatz energy shows an upward trend due to noise accumulation, signifying a failure of the noisy optimization. For a smaller error rate, p_{2} = 10^{−3}, which was demonstrated recently with the IBM Falcon device^{70}, the noisy energy initially decreases and reaches a minimum near N_{θ} = 7. This is again followed by an upturn as the number of variational parameters N_{θ} grows. On the other hand, the corresponding statevector analysis shows a clear continuous energy improvement up to N_{θ} = 25, followed by saturation with small fluctuations. We find the ansatz state fidelity saturates near 0.97. Similar observations apply to the noisy simulations with other twoqubit error rates. The statevector analysis shows that the energy converges at an error ≈ 3 × 10^{−3} with a fidelity ≈ 0.997 for p_{2} = 10^{−4}. When including only singlequbit errors, we find an error ≈ 1 × 10^{−3} with a fidelity ≈0.9992.
The observed improvement of the ansatz (revealed using statevector analysis), even though the noisy energy expectation value increases, is intriguing. This effect is most clearly seen in results for p = 10^{−3} between 7 ≤ N_{θ} < 25. It demonstrates the robustness of VQE to certain types of noise effects and can be rationalized as follows. Assuming for simplicity a global depolarizing error channel, we can relate the expectation value of an observable \(\bar{\langle O\rangle }\) with respect to a noisy density matrix to the noiseless result 〈O〉 as \(\bar{\langle O\rangle }=(1p)\langle O\rangle +\frac{p}{{2}^{n}}{\rm {Tr}}[O]\)^{71,72}. Since any observable can be shifted to be traceless (Tr[O] = 0), 〈O〉 is equivalent to \(\bar{\langle O\rangle }\) up to a constant scaling factor. The noise thus only rescales the energy landscape of the variational ansatz and maintains the optimal parameters. The fact that we find the ansatz energy to saturate in the statevector analysis with finite p_{2} is caused by our choice of noise model, which includes noise effects beyond a global depolarizing channel. This observation of state improvement during optimization masked by noisy energy expectation values suggests that with reasonably small error rates, expensive error mitigation techniques may be restricted to the final converged state at the end of VQE calculations to ensure accurate observable measurements.
Estimating ground state energy on NISQ devices
As a further step to benchmark the realistic noise effect on qubitADAPT VQE calculations of the multiorbital quantum impurity models, we measure the Hamiltonian expectation value of the e_{g} model with a converged qubitADAPT ansatz on the IBM quantum device ibmq_casablanca. The ansatz with optimal parameters is obtained with the HC pool using statevector simulations. The converged qubitADAPT ansatz used for the ground state energy estimate has 32 parameters, and the associated 32 generators for multiqubit unitary gates are listed in the “Methods” section.
To reduce the noise in the cost function measurement, it is essential to utilize a range of error mitigation techniques. We employ the standard readout error mitigation using the full confusion matrix approach, as implemented in Qiskit^{50}. The adopted measurement circuits based on Hamiltonian integral factorization also allow convenient symmetry detection and filtering with respect to how well the ansatz preserves the total electron number N_{e} = 4 and total spin zprojection S_{z} = 0. The gate error is mitigated using zero noise extrapolation (ZNE) with Richardson secondorder polynomial inference^{73,74}. The noise scale factor increases from 1 to 2 and 3 for each measurement circuit by local random unitary folding following the implementation in Mitiq^{75,76}. Because of the random gate folding and the stochastic SWAP mapping during transpilation to native gates^{50}, we perform ten runs for each measurement circuit at each noise level to smooth out the nondeterministic effects with averaging. For each run, we apply N_{sh} = 2^{14} shots for the measurements.
Figure 8a shows the Richardson extrapolation for the ground state energy with measured points at noise scale factors λ = 1, 2, 3, taking all 10 runs for each λ into account. The estimated energy has an absolute error Δ(E) = 0.6 ± 1.4 compared with the exact result indicated by the horizontal dashed line. This corresponds to a relative error of 3%. The standard deviation is obtained by fitting the sample points with a secondorder polynomial using the SciPy function curve_fit which takes both the mean values and standard deviations into account^{77}. In the postprocessing for the mean value of the energy cost function from statistical samplings, we first apply readout calibration, followed by symmetry filtering which discards the configurations with total electron number N_{e} ≠ 4 or total spin S_{z} ≠ 0. We observe that the ten runs can be divided into two groups based on the average N_{e} and S_{z} evaluated before symmetry filtering, as shown in Fig. 8c and d. A subgroup of five runs denoted by square symbols has much less bias away from the correct conserved quantum numbers N_{e} = 4 and S_{z} = 0 than the other five runs shown as circles. A more accurate ground state energy can be obtained when restricting to this optimal subgroup, as shown in Fig. 8b. The estimated energy error reduces significantly to Δ(E) = 0.1 ± 0.2, with a relative error of 0.7%.
In the above calculations on QPU, the circuits are transpiled into the basis gates of ibmq_casablanca device using the qubit layout and coupling map illustrated in the inset of Fig. 8a. Due to the limited qubit connectivity between nearest neighbors, each of the three transpiled measurement circuits for the e_{g} model contains about 350 CNOT gates, which amounts to over twofold increase compared to about 150 CNOTs without qubit swapping. Therefore, we also benchmark the calculations on other types of QPUs with full qubit connectivity such as trappedion devices. As an initial reference, we perform an energy estimation with the same ansatz on Quantinuum’s trappedion Honeywell System Model H12. The transpiled circuits have about 150 twoqubit ZZMax gates as expected. Due to limited access to the device, we apply only N_{sh} = 450 shots per circuit for the measurements without utilizing any error mitigation. The energy thus obtained is −17.6 ± 2, which should be compared with data points in Fig. 8a at a scale factor 1, and is found to be located near the lower end of that range. Here the error bar is estimated using multiple runs of simulations with the associated system Model H12 emulator (H12E) including a realistic noise model.
Conclusions
In an effort towards performing hybrid quantumclassical simulations of realistic correlated materials using a quantum embedding approach^{29,30,31,32,33,34,35,36,37}, we assess the gate depth and accuracy of variational ground state preparation with fixed and adaptive ansätze for two representative interacting multiorbital, e_{g} and t_{2g}, impurity models. To take advantage of the sparsity of the Hamiltonian in the atomic orbital representation in real space, we consider the HVA ansatz and an adaptive variant in the qubitencoded atomic orbital basis. An HC pool composed of pairwise commutators of the Hamiltonian terms is developed to allow fair comparison between the qubitADAPT and HVA ansatz. For reference, the standard UCCSD and related qubitADAPT calculations using UCCSDcompatible pools are also presented. The qubitADAPT calculation with an HC pool generally produces the most compact circuit representation with a minimal number of CNOTs in the final converged circuit. The fixed HVA ansatz follows very closely and has the additional advantage of requiring the least variational parameters N_{θ}.
To address the effect of quantum shot noise, we report QASM simulations of qubitADAPT VQE in the presence of shot noise for different numbers of shots (N_{sh}) that allow controlling the stochastic error. For our benchmark, we adopt stateoftheart techniques such as lowrank tensor factorization to reduce the number of distinct measurement circuits and a noise resilient optimization including sequential minimal optimization and Adadelta. We find a modest number of shots N_{sh} = 2^{14} per measurement circuit can lead to a variational representation of the ground state with fidelity f > 99.9%.
We further discuss ways to simplify the pool operators and reduce the pool size using e_{g} model as an example. It is pointed out that a minimal complete pool, as defined in refs. ^{16,69}, can be constructed using a subset of the HC pool. While a simplified pool can reduce the quantum computation resource in the adaptive operator screening procedure, it can make classical optimization more complicated, especially in the presence of noise. This suggests both the dimension and connectivity of operators are joint determining factors to design a practically optimal pool.
To assess the effects of realistic noise on VQE calculations of multiorbital impurity models, we perform qubitADAPT VQE calculations with a realistic decoherence noise model that includes amplitude and dephasing error channels. We find the impact of twoqubit errors to dominate over those of singlequbit errors, also since they are larger in NISQ hardware. We report that practically useful results can be obtained for p_{2} = 10^{−3}, which is close to current hardware levels. Importantly, we observe that the classical optimization continues to improve the ansatz even in a regime, where the noisy energy expectation value starts to rise. We reveal this behavior by executing the ansatz state on statevector simulators. Such persisting ansatz state improvement masked by noise shows that VQE is robust to certain noise effects and implies that costly error mitigation methods can potentially be reserved for the evaluation of expectation values in the final converged state.
Finally, we measure the energy for a converged qubitADAPT ansatz of the e_{g} model on the ibmq_casablanca QPU and Quantinuum’s H12 device. Using the results from IBM hardware, we obtain an error of 0.1 (0.7%) for the total energy by adopting error mitigation techniques such as zeronoise extrapolation, combined with a careful postselection based on symmetry and the conservation of quantum numbers.
Moving forward, the full qubitADAPT VQE calculations of quantum impurity models will be extended from noisy QASM simulations to simulations that include devicespecific noise effects beyond our decoherence model and finally to experiments on real hardware. Our study shows that an array of error mitigation techniques, including readout calibration, zeronoise extrapolation^{73,74}, and potentially probabilistic error cancellation^{74,78,79}, Clifford data regression^{80,81}, and probabilistic machinelearningbased techniques^{82}, need to be adopted to reach sufficiently accurate results. This is especially important when using VQE as an impurity solver in a quantum embedding approach as sufficiently accurate impurity model results are needed in order to enable the convergence of the classical selfconsistency loop. Our results constitute an important step forward in demonstrating highfidelity ground state preparation of impurity models on quantum devices. This is essential for realizing correlated material simulations through hybrid quantumclassical embedding approaches, where the ground state preparation of a generic felectron impurity model consisting of 28 spinorbital is on the verge of achieving practical quantum advantage^{35}.
Methods
Energy gradient of HVA
Here we show that the outermost lth layer gradient component vanishes (\({\left.\frac{\partial {{{{{{{\mathcal{E}}}}}}}}({{{{{{{\boldsymbol{\theta }}}}}}}})}{\partial {\theta }_{lj}}\right\vert }_{{{{{{{{{\boldsymbol{\theta }}}}}}}}}_{l} = 0}=0\)) for an llayer HVA ansatz \(\left\vert {{{\Psi }}}_{l}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\right\rangle ={{{\Pi }}}_{j = 1}^{{N}_{{{{{{{{\rm{G}}}}}}}}}}{e}^{i{\theta }_{lj}{\hat{h}}_{j}}\left\vert {{{\Psi }}}_{l1}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\right\rangle\):
Because the system Hamiltonian \(\hat{{{{{{{{\mathcal{H}}}}}}}}}\) under study is real due to timereversal symmetry, HVA is also real by construction. Therefore, \(\langle {{{\Psi }}}_{l1}[{{{{{{{\boldsymbol{\theta }}}}}}}}] \,\hat{{{{{{{{\mathcal{H}}}}}}}}}{\hat{h}}_{j}\, {{{\Psi }}}_{l1}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\rangle\) is real, and \({\left.\frac{\partial {{{{{{{\mathcal{E}}}}}}}}({{{{{{{\boldsymbol{\theta }}}}}}}})}{\partial {\theta }_{lj}}\right\vert }_{{{{{{{{{\boldsymbol{\theta }}}}}}}}}_{l} = 0}\) vanishes. Note that the exactly same reason motivates the development of the HC pool for qubitADAPT calculations.
Hamiltonian factorization of the impurity model
Here we explain explicitly how the Hamiltonian factorization is obtained using the e_{g} model as an example, whose Hamiltonian takes the following specific form:
Here \({\hat{n}}_{i\sigma }={\hat{c}}_{i\sigma }^{{{{\dagger}}} }{\hat{c}}_{i\sigma }\) and \({\hat{n}}_{i\sigma }^{f}={\hat{f}}_{i\sigma }^{{{{\dagger}}} }{\hat{\,f}}_{i\sigma }\) are the electron occupation number operators for the physical and bath orbitals, respectively. The factorization procedure is only needed for the singleparticle hybridization term (17) and the pair hopping and spinflip terms (18), as the rest are already in the diagonal representation.
The hybridization term (17) can be written in a diagonal form through singleparticle rotations on the physical and bath orbitals as follows:
where \({\hat{n}}_{m\sigma }^{(0)}={\hat{c}}_{m\sigma }^{{{{\dagger}}} (0)}{\hat{c}}_{m\sigma }^{(0)}\) and the rotated fermionic operators \({\hat{c}}_{m\sigma }^{(0)}\) are given by,
This can be derived conveniently in the matrix formulation:
The pair hopping and spinflip terms of the second line of Eq. (18) can be rewritten as:
with
The above expression is obtained by diagonalizing the Coulomb supermatrix of V_{(αβ),(γδ)} with densitydensity elements set to zero, V_{(αα),(γγ)} ≡ 0, which gives a single eigenvector associated with nonzero eigenvalue. Following the similar derivation in Eq. (23), the pair hooping and spinflip terms have the following diagonal representation:
with \({\hat{n}}_{m\sigma }^{(1)}={\hat{c}}_{m\sigma }^{{{{\dagger}}} (1)}{\hat{c}}_{m\sigma }^{(1)}\) and
Finally, we can represent the embedding Hamiltonian for e_{g} model in the following doublyfactorized form:
With the Hamiltonian integral factorization we find that three distinct measurement circuits are needed for the Hamiltonian expectation value: (i) the diagonal terms in the original atomic orbital basis, (ii) the hybridization terms in the basis of \({c}_{m\sigma }^{(0)}\)(22), (iii) the pair hopping and spinflip terms in the basis of \({c}_{m\sigma }^{(1)}\)(27).
Quantum simulation with Adadelta optimizer
In the main text, we reported the qubitADAPT VQE calculation with shots using the SMO optimizer. Here we additionally perform the calculations using the Adadelta optimization method, which is potentially tolerant to cost function errors^{63}. Below we describe the implementation of the algorithm followed by the results.
The algorithm minimizes the cost function along the steepest decent direction in parameter space, with a parameter update at step t as θ_{t} = θ_{t−1}−w_{t} ⊙ g_{t}. The gradient vector is determined from the derivative of the energy function along every parameter direction g_{t} = ∇_{θ}E(θ_{t}), where \(E({{{{{{{\boldsymbol{\theta }}}}}}}})=\langle {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}] \,\hat{{{{{{{{\mathcal{H}}}}}}}}}\, {{\Psi }}[{{{{{{{\boldsymbol{\theta }}}}}}}}]\rangle\) is the estimated energy. The set of parameterdependent adaptive learning rates is determined as \({{{{{{{{\bf{w}}}}}}}}}_{t}=\frac{\sqrt{{{\Delta }}{{{{{{{{\boldsymbol{\theta }}}}}}}}}_{t1}+\epsilon }}{\sqrt{{{{{{{{{\bf{s}}}}}}}}}_{t}+\epsilon }}\), where the leaked average of the square of rescaled gradients at the previous step is obtained as \({{\Delta }}{{{{{{{{\boldsymbol{\theta }}}}}}}}}_{t1}=\beta {{\Delta }}{{{{{{{{\boldsymbol{\theta }}}}}}}}}_{t2}+(1\beta ){({{{{{{{{\bf{w}}}}}}}}}_{t1}\odot {{{{{{{{\bf{g}}}}}}}}}_{t1})}^{2}\), and that of gradients is evaluated as \({{{{{{{{\bf{s}}}}}}}}}_{t}=\beta {{{{{{{{\bf{s}}}}}}}}}_{t1}+(1\beta ){{{{{{{{\bf{g}}}}}}}}}_{t}^{2}\). The operator ⊙ denotes elementwise product. The Adadelta algorithm involves a hyperparameter ϵ to regularize the ratio in determining w_{t}, which is set to 10^{−8}, and a mixing parameter set to β = 0.9. The leaked averages are all initialized to zero. We fix the number of steps in Adadelta optimization to N_{s} = 250 in our simulations. Considering that the evaluation of one gradient component associated with a variational parameter involves cost function measurements at two distinct parameter points following the parametershift rule, the quantum computational resource for Adadelta optimization is comparable to SMO with N_{sw} = 60.
Figure 9 shows the representative convergence behavior of qubitADAPT energy with an increasing number of variational parameters N_{θ} calculated using a number of shots N_{sh} = 2^{16} per observable. The adaptive ansatz energy E decreases as the circuit depth increases with more variational parameters. The energy points shown include not only the final Adadelta optimized energies of the qubitADAPT ansatz with N_{θ} parameters but also intermediate energies for the 250 Adadelta steps to provide a detailed view of the convergence. For the operator screening step of the qubitADAPT calculation we fix N_{sh} = 2^{16} for energy evaluations in all cases, and determine the energy gradient by the parametershift rule^{68}. The final energy error from the calculations with Adadelta is E−E_{GS} = 4.4 × 10^{−3}. This is comparable with the result from the SMO optimizer.
The ground state ansatz of (2, 2) e _{g} model used on ibmq_casablanca
The qubitADAPT ansatz takes the pseudoTrotter form. The converged ansatz for the e_{g} model which we used for the calculations on IBM quantum hardware ibmq_casablanca is composed of 32 generators for the multiqubit unitary gates, which are listed here with parity encoding (in the order that they appear in the ansatz):

IIIZXY, IYXIII, XYZIII, IIZYXZ, IXYIII,

ZXYIIZ, XYIIZZ, XYIIIZ, IIIIYX, IZXYXX,

IIXZYI, IIXIIY, IIXIZY, IIZYXZ, IIIZYX,

YXIIII, IZXIZY, IIXIZY, IIYIIX, IIZXYI,

IZXYXX, IZYIZX, ZYXIII, ZYIIZX, IIYIIX,

IIIIXY, IIXIIY, IIXIYZ, IIXZYI, YXXIZX,

IIXIYZ, YXXIIX.
Data availability
All the data to generate the figures are available at figshare^{83}. Data supporting the calculations are available together with the codes at figshare^{53,54,64}. All other data are available from the corresponding authors on reasonable request.
Code availability
All the computer codes developed and used in this work are available opensource at figshare^{45,53,54,64}.
References
AspuruGuzik, A., Dutoi, A. D., Love, P. J. & HeadGordon, M. Simulated quantum computation of molecular energies. Science 309, 1704–1707 (2005).
Peruzzo, A. et al. A variational eigenvalue solver on a photonic quantum processor. Nat. Commun. 5, 1–7 (2014).
Kandala, A. et al. Hardwareefficient variational quantum eigensolver for small molecules and quantum magnets. Nature 549, 242–246 (2017).
O’Malley, P. J. et al. Scalable quantum simulation of molecular energies. Phys. Rev. X 6, 031007 (2016).
Preskill, J. Quantum computing in the NISQ era and beyond. Quantum 2, 79 (2018).
McArdle, S., Endo, S., AspuruGuzik, A., Benjamin, S. C. & Yuan, X. Quantum computational chemistry. Rev. Mod. Phys. 92, 015003 (2020).
Cerezo, M. et al. Variational quantum algorithms. Nat. Rev. Phys 3, 625–644 (2021).
McClean, J. R., Romero, J., Babbush, R. & AspuruGuzik, A. The theory of variational hybrid quantumclassical algorithms. New J. Phys. 18, 023023 (2016).
Hoffmann, M. R. & Simons, J. A unitary multiconfigurational coupledcluster method: theory and applications. J. Chem. Phys. 88, 993–1002 (1988).
Bartlett, R. J., Kucharski, S. A. & Noga, J. Alternative coupledcluster ansätze ii. the unitary coupledcluster method. Chem. Phys.Lett. 155, 133–140 (1989).
Bartlett, R. J. & Musiał, M. Coupledcluster theory in quantum chemistry. Rev. Mod. Phys. 79, 291 (2007).
Romero, J. et al. Strategies for quantum computing molecular energies using the unitary coupled cluster ansatz. Quantum Sci. Technol. 4, 014008 (2018).
Grimsley, H. R., Economou, S. E., Barnes, E. & Mayhall, N. J. An adaptive variational algorithm for exact molecular simulations on a quantum computer. Nat. Commun. 10, 3007 (2019).
Ryabinkin, I. G., Yen, T.C., Genin, S. N. & Izmaylov, A. F. Qubit coupled cluster method: a systematic approach to quantum chemistry on a quantum computer. J. Chem. Theory Comput. 14, 6317–6326 (2018).
Lee, J., Huggins, W. J., HeadGordon, M. & Whaley, K. B. Generalized unitary coupled cluster wave functions for quantum computation. J. Chem. Theory Comput. 15, 311–324 (2018).
Tang, H. L. et al. Qubitadaptvqe: an adaptive algorithm for constructing hardwareefficient ansätze on a quantum processor. PRX Quantum 2, 020310 (2021).
Zhang, F. et al. Shallowcircuit variational quantum eigensolver based on symmetryinspired hilbert space partitioning for quantum chemical calculations. Phys. Rev. Res. 3, 013039 (2021).
Gomes, N. et al. Adaptive variational quantum imaginary time evolution approach for ground state preparation. Adv. Quantum Technol. 4, 2100114 (2021).
Fedorov, D. A., Peng, B., Govind, N. & Alexeev, Y. VQE method: a short survey and recent developments. Mater. Theory 6, 2 (2022).
Tilly, J. et al. The variational quantum eigensolver: a review of methods and best practices. Phys. Rep. 986, 1–128 (2022).
Wecker, D., Hastings, M. B. & Troyer, M. Progress towards practical quantum variational algorithms. Phys. Rev. A 92, 042303 (2015).
Ho, W. W. & Hsieh, T. H. Efficient variational simulation of nontrivial quantum states. SciPost Phys. 6, 29 (2019).
Wiersema, R. et al. Exploring entanglement and optimization within the hamiltonian variational ansatz. PRX Quantum 1, 020319 (2020).
Kent, P. R. & Kotliar, G. Toward a predictive theory of correlated materials. Science 361, 348–354 (2018).
Yin, Z., Haule, K. & Kotliar, G. Kinetic frustration and the nature of the magnetic and paramagnetic states in iron pnictides and iron chalcogenides. Nat. Mater. 10 12, 932–5 (2011).
Georges, A., de’Medici, L. & Mravlje, J. Strong correlations from Hund’s coupling. Annu. Rev. Condensed Matter Phys. 4, 137–178 (2013).
de’ Medici, L., Giovannetti, G. & Capone, M. Selective Mott physics as a key to iron superconductors. Phys. Rev. Lett. 112 17, 177001 (2014).
Sprau, P. O. et al. Discovery of orbitalselective cooper pairing in FESE. Science 357, 75 – 80 (2017).
Georges, A., Kotliar, G., Krauth, W. & Rozenberg, M. J. Dynamical meanfield theory of strongly correlated fermion systems and the limit of infinite dimensions. Rev. Mod. Phys. 68, 13 (1996).
Kotliar, G. et al. Electronic structure calculations with dynamical meanfield theory. Rev. Mod. Phys. 78, 865 (2006).
Sun, Q. & Chan, G. K.L. Quantum embedding theories. Acc. Chem. Res. 49, 2705–2712 (2016).
Knizia, G. & Chan, G. K.L. Density matrix embedding: a simple alternative to dynamical meanfield theory. Phys. Rev. Lett. 109, 186404 (2012).
Lanatà, N., Yao, Y.X., Deng, X., Dobrosavljević, V. & Kotliar, G. Slave boson theory of orbital differentiation with crystal field effects: application to uo_{2}. Phys. Rev. Lett. 118, 126401 (2017).
Lee, T.H., Ayral, T., Yao, Y.X., Lanata, N. & Kotliar, G. Rotationally invariant slaveboson and density matrix embedding theory: unified framework and comparative study on the onedimensional and twodimensional Hubbard model. Phys. Rev. B 99, 115129 (2019).
Yao, Y.X., Zhang, F., Wang, C.Z., Ho, K.M. & Orth, P. P. Gutzwiller hybrid quantumclassical computing approach for correlated materials. Phys. Rev. Res. 3, 013184 (2021).
Sakurai, R., Mizukami, W. & Shinaoka, H. Hybrid quantumclassical algorithm for computing imaginarytime correlation functions. Phys. Rev. Res. 4, 023219 (2022).
Vorwerk, C., Sheng, N., Govoni, M., Huang, B. & Galli, G. Quantum embedding theories to simulate condensed systems on quantum computers. Nat. Comput. Sci. 2, 424–432 (2022).
Bauer, B., Wecker, D., Millis, A. J., Hastings, M. B. & Troyer, M. Hybrid quantumclassical approach to correlated materials. Phys. Rev. X 6, 031045 (2016).
Bünemann, J., Weber, W. & Gebhard, F. Multiband Gutzwiller wave functions for general onsite interactions. Phys. Rev. B 57, 6896 (1998).
Fabrizio, M. Gutzwiller description of nonmagnetic Mott insulators: dimer lattice model. Phys. Rev. B 76, 165110 (2007).
Deng, X., Dai, X. & Fang, Z. LDA + Gutzwiller method for correlated electron systems. Europhys. Lett. 83, 37008 (2008).
Lanata, N. et al. γα isostructural transition in cerium. Phys. Rev. Lett. 111, 196801 (2013).
Lu, F., Zhao, J., Weng, H., Fang, Z. & Dai, X. Correlated topological insulators with mixed valence. Phys. Rev. Lett. 110, 096401 (2013).
Lanatà, N., Yao, Y.X., Wang, C.Z., Ho, K.M. & Kotliar, G. Phase diagram and electronic structure of praseodymium and plutonium. Phys. Rev. X 5, 011008 (2015).
Yao, Y.X. Python driver of Gutzwiller quantumclassical embedding simulation framework (PyGQCE). https://doi.org/10.6084/m9.figshare.11987616 (2020).
Kotliar, G. & Ruckenstein, A. E. New functional integral approach to strongly correlated fermi systems: the Gutzwiller approximation as a saddle point. Phys. Rev. Lett. 57, 1362 (1986).
Bünemann, J. & Gebhard, F. Equivalence of Gutzwiller and slaveboson meanfield theories for multiband Hubbard models. Phys. Rev. B 76, 193104 (2007).
de’ Medici, L., Mravlje, J. & Georges, A. Janusfaced influence of Hund’s rule coupling in strongly correlated materials. Phys. Rev. Lett. 107 25, 256401 (2011).
Lanatà, N. et al. Orbital selectivity in Hund’s metals: the iron chalcogenides. Phys. Rev. B 87, 045122 (2013).
Abraham, H. et al. Qiskit: an opensource framework for quantum computing. Zenodo https://github.com/Qiskit/qiskit (2019).
Stein, T., Henderson, T. M. & Scuseria, G. E. Seniority zero pair coupled cluster doubles theory. J. Chem. Phys. 140, 214113 (2014).
Henderson, T. M., Bulik, I. W. & Scuseria, G. E. Pair extended coupled cluster doubles. J. Chem. Phys. 142, 214116 (2015).
Yao, Y.X. Python code to perform QubitADAPT VQE calculations with statevector simulator. https://doi.org/10.6084/m9.figshare.19350509 (2022).
Yao, Y.X. & Getelina, J. C. Python code for VQE calculations with generic Hamiltonian Variational Ansatz. https://doi.org/10.6084/m9.figshare.19349846 (2022).
McArdle, S. et al. Variational ansatzbased quantum simulation of imaginary time evolution. npj Quantum Inf. 5, 75 (2019).
Izmaylov, A. F., Yen, T.C., Lang, R. A. & Verteletskyi, V. Unitary partitioning approach to the measurement problem in the variational quantum eigensolver method. J. Chem. Theory Comput. 16, 190–195 (2019).
Gokhale, P. et al. Minimizing state preparations in variational quantum eigensolver by partitioning into commuting families. arXiv:1907.13623 (2019).
Zhao, A. et al. Measurement reduction in variational quantum algorithms. Phys. Rev. A 101, 062322 (2020).
Crawford, O. et al. Efficient quantum measurement of Pauli operators in the presence of finite sampling error. Quantum 5, 385 (2021).
Yen, T.C. & Izmaylov, A. F. Cartan subalgebra approach to efficient measurements of quantum observables. PRX Quantum 2, 040320 (2021).
Huggins, W. J. et al. Efficient and noise resilient measurements for quantum chemistry on nearterm quantum computers. npj Quantum Inf. 7, 1–9 (2021).
Nakanishi, K. M., Fujii, K. & Todo, S. Sequential minimal optimization for quantumclassical hybrid algorithms. Phys. Rev. Res. 2, 043158 (2020).
Zeiler, M. D. Adadelta: an adaptive learning rate method. arXiv:1212.5701 (2012).
Mukherjee, A. & Yao, Y.X. Circuit Implementation of qubitADAPT VQE. https://doi.org/10.6084/m9.figshare.19351952 (2022).
Claudino, D., Wright, J., McCaskey, A. J. & Humble, T. S. Benchmarking adaptive variational quantum eigensolvers. Front. Chem. 8, 1152 (2020).
Yordanov, Y. S., Armaos, V., Barnes, C. H. & ArvidssonShukur, D. R. Qubitexcitationbased adaptive variational quantum eigensolver. Commun. Phys. 4, 1–11 (2021).
BonetMonroig, X. et al. Performance comparison of optimization methods on variational quantum algorithms. arXiv:2111.13454 (2021).
Mari, A., Bromley, T. R. & Killoran, N. Estimating the gradient and higherorder derivatives on quantum hardware. Phys. Rev. A 103, 012405 (2021).
Shkolnikov, V., Mayhall, N. J., Economou, S. E. & Barnes, E. Avoiding symmetry roadblocks and minimizing the measurement overhead of adaptive variational quantum eigensolvers. arXiv:2109.05340 (2021).
Finke, D. IBM Demonstrates 99.9% CNOT Gate Fidelity On A New Superconducting Test Device (Quantum Computing Report, 2021).
Urbánek, M. et al. Mitigating depolarizing noise on quantum computers with noiseestimation circuits. Phys. Rev. Lett. 127, 270502 (2021).
Vovrosh, J. et al. Simple mitigation of global depolarizing errors in quantum simulations. Phys. Rev. E 104 32, 035309 (2021).
Li, Y. & Benjamin, S. C. Efficient variational quantum simulator incorporating active error minimization. Phys. Rev. X 7, 021050 (2017).
Temme, K., Bravyi, S. & Gambetta, J. M. Error mitigation for shortdepth quantum circuits. Phys. Rev. Lett. 119, 180509 (2017).
LaRose, R., Mari, A., Karalekas, P. J., Shammah, N. & Zeng, W. J. Mitiq: a software package for error mitigation on noisy quantum computers. Quantum 6, 774 (2022).
GiurgicaTiron, T., Hindy, Y., LaRose, R., Mari, A. & Zeng, W. J. Digital zero noise extrapolation for quantum error mitigation. In 2020 IEEE International Conference on Quantum Computing and Engineering (QCE), 306–316 (IEEE, 2020).
Virtanen, P. et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nat. Methods 17, 261–272 (2020).
Endo, S., Benjamin, S. C. & Li, Y. Practical quantum error mitigation for nearfuture applications. Phys. Rev. X 8, 031027 (2018).
van den Berg, E., Minev, Z. K., Kandala, A. & Temme, K. Probabilistic error cancellation with sparse Pauli–Lindblad models on noisy quantum processors. arXiv:2201.09866 (2022).
Czarnik, P., Arrasmith, A., Coles, P. J. & Cincio, L. Error mitigation with Clifford quantumcircuit data. Quantum 5, 592 (2021).
Lowe, A. et al. Unified approach to datadriven quantum error mitigation. Phys. Rev. Res. 3, 033098 (2021).
Rogers, J. et al. Error mitigation in variational quantum eigensolvers using probabilistic machine learning. arXiv:2111.08814 (2021).
Mukherjee, A., Berthusen, N., Getelina, J. C., Orth, P. P. & Yao, Y.X. Data for Comparative Study of Adaptive Variational Quantum Eigensolvers for Multiorbital Impurity Models https://doi.org/10.6084/m9.figshare.19352222 (2022).
Acknowledgements
The authors acknowledge valuable discussions with Thomas Iadecola, Niladri Gomes, CaiZhuang Wang, and Nicola Lanatà. This work was supported by the US Department of Energy (DOE), Office of Science, Basic Energy Sciences, Materials Science and Engineering Division, including the grant of computer time at the National Energy Research Scientific Computing Center (NERSC) in Berkeley, CA, USA. The research was performed at the Ames National Laboratory, which is operated for the US DOE by Iowa State University under Contract No. DEAC0207CH11358. We acknowledge the use of the IBM Quantum Experience, through the IBM Quantum Researchers Program. The views expressed are those of the authors and do not reflect the official policy or position of IBM or the IBM Quantum team. This research also used resources from the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DEAC0500OR22725.
Author information
Authors and Affiliations
Contributions
A.M. and Y.Y. developed the codes and performed the simulations. N.F.B. and J.C.G. contributed to the HVA calculations. A.M., P.P.O., and Y.X.Y. analyzed the results. Y.X.Y., A.M., and P.P.O. wrote the paper with input from all authors. Y.X.Y. and P.P.O. conceived and supervised the project.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Communications Physics thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Mukherjee, A., Berthusen, N.F., Getelina, J.C. et al. Comparative study of adaptive variational quantum eigensolvers for multiorbital impurity models. Commun Phys 6, 4 (2023). https://doi.org/10.1038/s42005022010896
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s42005022010896
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.