An application benchmark for fermionic quantum simulations

I. INTRODUCTION

• The benchmark addresses the lack of application-specific measures of device usefulness for variational fermionic simulation, using a model solvable by the Bethe ansatz.The one-dimensional Fermi-Hubbard model is described as relatively simple while capturing essential complexity in preparing strongly correlated fermionic ground states.Source: §I · p. 1
• A chain-length turning point is expected when bounded circuits and hardware limits prevent the VQE energy from continuing to approach the infinite-chain limit.Before the turning point, the energy should converge toward that limit; the resulting length marks an effective fermionic length for the device.Source: §I · p. 1
• The proposed Sycamore variational layer combines a variable X rotation, a tunable two-qubit gate, and final tunable Z phases.Parallel two-qubit elements are parameterized so each layer completes in fixed time, allowing multilayer composition.Source: §I · pp. 1–2

II. THE BENCHMARK

• The benchmark estimates effective fermionic length from VQE energy densities for increasingly large one-dimensional Fermi-Hubbard chains, against the infinite-chain reference.The length at which the estimate ceases to improve characterizes the largest chain for which the chosen device and ansatz provide the best estimate of the infinite-chain energy density.Source: §II · p. 2

A. Representation of the one-dimensional FHM

• The one-dimensional model has L sites and 2L spin orbitals, which the canonical Jordan-Wigner encoding maps to qubits.This mapping expresses the Hamiltonian using operators acting on at most two qubits.Source: A. Representation of the… · pp. 2–3, Figure 2 · p. 3

• The variational energy is evaluated from expectation values of the Fermi-Hubbard Hamiltonian for a parameterized ansatz state.Source: B. The variational energy minimization… · p. 3
• Minimizing E(theta) over circuit parameters yields E(theta*), whose deviation from the analytic infinite-chain energy density estimates finite-chain error.Source: B. The variational energy minimization… · p. 4
• O(1/(L epsilon^2)) measurements estimate chain energy density to error epsilon, so fixed-error estimates require fewer measurements for longer chains.The energy can be evaluated using only three measurement bases.Source: B. The variational energy minimization… · p. 4

• Accurate simulation of the one-dimensional Fermi-Hubbard model can provide estimates of the energy per site in the thermodynamic limit, without specifying a complete accuracy dependence.Source: C. Definition of the benchmark · p. 4
• Finite-size error scales as O(1/L), while decoherence error increases with chain length, producing an optimal chain length L*.Source: C. Definition of the benchmark · p. 4
• The authors hypothesize that preparing the exact ground state requires at least a linear-depth circuit, based on the need for non-Gaussian fermionic states.They note that numerical evidence supports this hypothesis.Source: C. Definition of the benchmark · p. 4
• L* is the effective fermionic length: the chain size at which a hardware-efficient ansatz gives its best estimate of infinite-chain energy density.The metric reflects both device capability and VQE optimization quality.Source: C. Definition of the benchmark · p. 4
• The proposed benchmark is illustrated with a hardware-efficient ansatz and VQE strategy tailored to Google’s Sycamore architecture.Source: C. Definition of the benchmark · p. 4

• Hardware-efficient ansatzes use a processor’s native parameterizable operations to build shallow variational circuits.Source: §III · p. 4
• The section formalizes the ansatz class and gives a specific example intended for fermionic-system simulation.Source: §III · p. 4
• The Sycamore example describes how the device can be operated to implement the proposed ansatz.Source: §III · p. 4

A. VQE ansatz and optimization strategy

• Hardware-efficient ansatzes use native parameterizable operations for shallow circuits, with qubit connectivity, gate elements, layered edges, and a noise-limited depth.The framework also encompasses devices with parallel parameterized multiqubit operations, such as shuttling-ion systems.Source: A. VQE ansatz and optimization strategy · pp. 4–5
• Global-cost gradients can vanish exponentially for ansatzes approximating 2-designs, while shallow circuits with local costs can retain polynomially vanishing gradients.The 1D Fermi-Hubbard Hamiltonian is local, motivating gradient-based optimization with a few layers.Source: A. VQE ansatz and optimization strategy · p. 5
• Layer-by-layer optimization retains prior-layer parameters and adds randomly initialized layers until energy convergence or the evaluation limit is reached.The sequential procedure is intended to maintain significant overlap with the ground state.Source: A. VQE ansatz and optimization strategy · p. 5
• The proposed training method offloads fermionic Gaussian operations to a classical processor, reserving the device’s variational resource for non-Gaussian operations.Source: A. VQE ansatz and optimization strategy · p. 5

B. Example of Sycamore

• Sycamore’s tunable-coupler physics motivates an ansatz whose controls are optimized directly, potentially mitigating cross-talk and coherent errors.The processor has 54 transmon qubits, with gates implemented using tunable couplers.Source: B. Example of Sycamore · pp. 5–6
• The ideal two-qubit element combines exchange and dispersive interactions, flanked by single-qubit X rotations at the start and Z rotations at the end.The sequence of variational elements can generate a universal gate set.Source: B. Example of Sycamore · pp. 5–6, Figure 3 · p. 6
• The coupler amplitude is bounded by ±Φmax and layer runtime by fixed tend, while tinit controls the experimental implementation of θ and φ.The final Z rotations can compensate for frequency shifts induced by flux controls.Source: Figure 4 · p. 6, B. Example of Sycamore · p. 6
• The numerical ansatz repeats the ABCD edge-pattern sequence; its maximum layer count is constrained by coherent depth, approximately T2 divided by gate time.The four edge sets correspond to the layer patterns used in the benchmark.Source: B. Example of Sycamore · pp. 6–7, Figure 5 · p. 7

Fermionic character of the ansatz

• XX+YY rotations generate matchgates that construct fermionic Gaussian transformations, while non-nearest-neighbor rotations and ZZ interactions add non-Gaussian operations.The paper interprets parametrized ZZ interactions as variational electrostatic terms between fermions.Source: Fermionic character of the ansatz · p. 7
• The illustrated circuit orderings include interleaved horizontal sites and a vertical ordering with oblique sites.Source: Fermionic character of the ansatz · p. 7, Figure 6 · p. 7
• Orbital permutations that change the Jordan-Wigner ordering are fermionic Gaussian transformations that can be computed classically efficiently.Source: Fermionic character of the ansatz · p. 7

Hamiltonian and circuit orderings

• Circuit ordering maps simulated spin orbitals to physical qubits; the authors heuristically choose locality-preserving layouts, and EFL can depend on that choice.The paper notes that an optimal ordering can be found using a method in Appendix B.Source: Hamiltonian and circuit orderings · p. 8

IV. NUMERICAL EXPERIMENTS

• The numerical experiments test the proposed methods and report on reproducibility of the layer-by-layer training procedure.Source: §IV · p. 8

A. Algorithm description

• The numerical experiments use containerized, reproducible workflows, with noiseless circuits simulated without finite-sampling effects.The simulations use U=8, µ=4, and t=1.Source: A. Algorithm description · p. 8, Figure 7 · p. 8
• The workflow generates the Hamiltonian and particle-number constraint, maps them to qubits, builds the ansatz, and optimizes variational parameters with SLSQP.Source: Figure 7 · p. 8

B. Demonstration of the benchmark

• With 33 Sycamore ansatz layers, the effective fermionic length is L*=6; the interleaved layout performs best, while the 1×8 chain does not converge close to the ground state.Energy-density deviation decreases through length six for all tested orderings.Source: Figure 8 · p. 9, B. Demonstration of the benchmark · p. 9
• The bounded-depth model represents finite coherence time: as chain length grows, the ansatz performs more poorly because its depth cannot increase.Source: B. Demonstration of the benchmark · p. 9

Reproducibility of training results

• For the 1×8 chain, nine runs reached a best energy-density deviation of 0.15, while most ended near 0.18; 1×6 and smaller chains converged within 0.02.The authors attribute the 1×8 error to limited circuit depth, not random initialization.Source: Reproducibility of training results · p. 9

V. DISCUSSION

• The benchmark uses the 1D Fermi-Hubbard model to estimate effective fermionic length, and bounded-depth circuits recover the hypothesized benchmark behavior in numerical experiments.The benchmark is described as requiring modest measurements and the infinite-chain ground-state energy.Source: §V · pp. 9–10
• For the 1×2 chain, estimated sampling time is about 9 seconds at energy-density accuracy 10^-2 and around 15 minutes at 10^-3.The estimate assumes 5 kHz sampling, grouped co-measurable terms, and zero within-group covariance; the authors expect measurement counts to decrease for equal per-site accuracy on longer chains.Source: §V · p. 10
• The number of function evaluations required to train the ansatz may obstruct scalability, but addressing that burden is outside this paper’s scope.The authors identify opportunities for improvement without evaluating them here.Source: §V · p. 10
• Suggested training extensions include matrix-product-state or tensor-network initialization and adiabatic-assisted VQE for building longer chains.These are proposed as possible improvements to the training method.Source: §V · p. 10
• Simulations under more realistic noise conditions, ansatz error resilience, and interplay with error-mitigation techniques remain future work.The proposed use of auxiliary qubits as a bath to dissipate entropy from incoherent processes is a hypothesis for future study.Source: §V · p. 11
• A proposed alternative prepares an optimal Gaussian state classically, then refines it with parameterized ZZ and non-nearest-neighbor iSWAP gates.The authors say this approach cannot yield a higher-energy state than the efficiently computed mean-field state, but generally needs more measurements to converge than the benchmark.Source: §V · p. 11

Appendix A: Fermionic character of the ansatz

• The appendix introduces fermionic Gaussian transformations using Majorana operators and connects them to quadratic fermionic Hamiltonians and Bogoliubov transformations.For particle-number-conserving cases, the rotations are orbital rotations.Source: Appendix A · p. 13
• A Slater determinant is constructed from the vacuum in the diagonal basis of a quadratic fermionic Hamiltonian, and Gaussian-state density matrices admit a covariance-matrix representation.The appendix also describes the covariance matrix as a compact notation.Source: Appendix A · p. 13
• Nearest-neighbor matchgates are introduced through a two-qubit gate definition, with six generators specified for their nearest-neighbor action.The passage also states the matchgate determinant condition.Source: Appendix A · p. 13
• A full network of matchgates generates any SO(2L) rotation, parameterized by 2L² − L angles and implementable in linear depth on a linear qubit array.The passage identifies this as the general rotation case.Source: Appendix A · p. 13

Appendix B: Advanced training method

• The method offloads Gaussian transformations to a classical computer to focus the quantum ansatz on non-Gaussian resources and improve on a classically efficient mean-field state.The quantum state is parameterized by U(θ), while the Hamiltonian basis is optimized classically.Source: Appendix B · p. 14
• For ansatzes that do not conserve particle number, the cost function evaluates the parametrized quantum state against a classically basis-optimized Hamiltonian.The passage defines ρ(θ) as the state prepared on the quantum computer and H(φ) as the Hamiltonian represented in the optimized basis.Source: Appendix B · p. 14
• Training uses a modified Newton procedure with measured gradients and a Hessian; first and second derivatives with respect to θ can be measured using the parameter-shift rule.The Hessian is given as B = ∇∇⊺Ω(θ,φ).Source: Appendix B · p. 14