betaFeedback

Reading v1View latest version

WorksPhysics and Astronomy

Grow, Compress, and Perturb: A Hybrid Quantum-Classical Pathway to Scalable Quantum Chemistry on Quantum Computers

Abstract

While Variational Quantum Eigensolver (VQE) variants are among the leading candidates for electronic structure calculations on quantum computers, gradient-based operator selection limits their scalability. To solve this, we introduce the Hamiltonian-gradient-free VULCAN (Variational Upbuilding, Layered Compression, Adaptive N-body perturbation) layered VQE framework. Our method dynamically generates ansatz operators using Variational Double Bracket Flow (VDBF) equations. Combined with a linearly scaling minimal complete pool, it creates a pool 4-12 times smaller than QEB (Qubit-Excitation-Based) pools. By appending entire layers simultaneously, Hamiltonian gradient measurements are completely eliminated. Circuit depth is managed via a grow-then-compress scheme using efficient state-overlap measurements, and qubit requirements are reduced by projecting the Hamiltonian onto a dominant-determinant subspace. Using high-performance emulation up to 32 logical qubits, we achieve chemical accuracy at just 5-20% of the measurement cost of batched ADAPT-VQE for hydrogen chains, BeH2 and H2O. For the 32-qubit C3H6, where batched ADAPT-VQE does not reach chemical accuracy, a single VULCAN layer combined with a GPU-accelerated stochastic Epstein-Nesbet (ENPT2) correction does. This stochastic ENPT2 post-processing step enables early termination at chemical accuracy.

Introduction

Accurate calculation of ground state energies for strongly correlated many-body systems represents a major computational bottleneck for classical computing, as the exponential growth of the Hilbert space renders exact classical diagonalization intractable for large-scale electronic structure problems. Quantum computing offers an alternative by mapping exponentially large electronic states onto a linear number of qubits. Fully quantum algorithms such as Quantum Phase Estimation (QPE) [1] achieve this in principle, but their reliance on deep circuits and large numbers of logical qubits currently places them beyond the capabilities of the Noisy Intermediate-Scale Quantum (NISQ) devices [2]. The Variational Quantum Eigensolver (VQE) [3] has consequently emerged as the leading hybrid quantum-classical algorithm for near-term quantum computation.

Despite its promise, the practical deployment of VQE is hindered by fundamental bottlenecks [4, 5], particularly in ansatz design. Standard hardware-efficient ansätze construct highly expressive but unstructured variational manifolds, which severely limit trainability. This lack of physical structure leads to the barren plateau phenomenon, where gradients vanish exponentially with system size, and to a critical sensitivity to the initial reference state.

The ADAPT-VQE [6] family of algorithms [7, 8, 9] addresses this by dynamically constructing problem-tailored circuits. Iteratively adding only the most relevant operators from a predefined pool restricts the variational manifold to physically relevant subspaces, yielding highly compact ansätze that remain resilient to barren plateaus. This adaptive construction introduces a severe trade-off: selecting each operator requires evaluating a polynomially scaling number of energy gradients. Recent measurement protocols have elegantly reduced this cost [10], but the overall measurement overhead remains high, and under shot noise may be noisy enough to produce sub-optimal operator choices. With this in mind, some of us introduced a gradient-approximated variant of ADAPT-VQE denoted GGA-VQE (Greedy Gradient-free Adaptive VQE)[11] to mitigate statistical noise and bypass the measurement-heavy classical optimization of standard ADAPT-VQE. GGA-VQE uses an analytical 1D energy landscape extrapolation (requiring only a few measurements) to select a pool operator and its optimal angle simultaneously, freezing previous parameters rather than re-optimizing the entire circuit. However, this greedy, local-only optimization strategy makes the algorithm prone to trapping in shallow local minima for strongly correlated systems, where it requires additional post-processing sweeps, such as sequential re-optimization, and it still suffers from an O(M)O(M) measurement overhead per step across candidate operators in pool MM unless special Hamiltonian symmetries can be exploited. In that context, we introduce VULCAN (Variational Upbuilding, Layered Compression, Adaptive N-body perturbation) – a hybrid quantum-classical framework based on a “grow, compress, and perturb” architecture that directly mitigates these bottlenecks. Its hardware efficiency comes from removing the gradient-screening step altogether: no energy gradients are measured to select operators, which reduces the number of Pauli-term evaluations needed on the quantum computer. To do so, we leverage some ingredients that have recently been proposed.

First, Minimal Complete Pools (MCPs) [7, 11, 12, 13] provide a robust mathematical framework for constructing generator sets. While scaling only linearly with the number of qubits NN, the use of MCPs has long been limited by the intractable search for the required generators. This has changed recently with the availability of a general mathematical framework [12] scaling polynomially with the number of qubits to enable the fast verification of the completeness of a user-defined pool of Pauli string operators with respect to a target Lie algebra. However, our previous work integrating MCPs into both adaptive [12] and fixed Lie-algebraic [12, 14] frameworks to reduce quantum resources, showed that their restricted size shifts the computational burden onto the choice of starter operators. Identifying the optimal starters remains an open challenge in the pursuit of rapid convergence and chemical accuracy for strongly correlated systems. Existing static pre-screening strategies, such as the Unitary Selective Coupled-Cluster (USCC) method [15], can isolate physically relevant starters from Hamiltonian matrix elements, but require preliminary VQE runs, whose optimized amplitudes must be computed and stored before screening can begin [12, 14]. This reliance on preliminary computational overhead simply replaces one bottleneck with another, rendering static pre-screening highly undesirable for scalable NISQ hardware deployment. Therefore, in order to eliminate this bottleneck, we propose as an alternative to generate this starter pool dynamically, without any preliminary classical optimization. VULCAN is inspired by MB-ADAPT-VQE [12], which grows the ansatz by adding, at each iteration, a batch of operators drawn from an MCP built around physically motivated starters. We keep this MCP-based construction, but replace both the pre-screening of the starters and the gradient-based selection, as detailed below, and compress each new layer using state-overlap measurements to limit the circuit depth.

Furthermore, in the spirit of non-iterative VQE approaches such as NI-DUCC (Non-Iterative Disentangled Unitary Coupled-Cluster) [14], we choose to retain a layered product-of-exponentials ansatz form, while departing from it in three respects: i) the number of layers is grown until convergence rather than fixed in advance; ii) each new layer is greedily optimized with preceding layers held frozen, bounding the classical optimization cost per iteration; and iii) the generators populating each layer are derived from the Hamiltonian itself rather than chosen beforehand. Specifically, we employ the Variational Double Bracket Flow (VDBF) equation [16, 17] to expose the variationally helpful excitations. Because entire layers are appended at once, the ADAPT selection phase is bypassed and no Hamiltonian gradient is ever measured. The resulting ansatz reads

∣ψ⟩=∏k=1n(∏j=1Mexp⁡(iθj,kPj))∣ψ0⟩,(1)\lvert\psi\rangle= \prod_{k=1}^{n}\left(\prod_{j=1}^{M}\exp(i\theta_{j,k}P_j)\right)\lvert\psi_0\rangle, \tag*{(1)}

where ∣ψ0⟩\lvert\psi_0\rangle is the Hartree–Fock state, the PjP_j are the MM pool operators expressed as Pauli strings, and the θj,k\theta_{j,k} are the variational parameters.

Appending complete layers grows circuit depth quickly, so the approach pairs growth with three mechanisms that trade one hardware resource against another. Depth is controlled by a grow-then-compress scheme inspired by Overlap-ADAPT-VQE [9], in which candidate operators are ranked by the gradient of the overlap with a target state rather than by energy gradients. Because our pool consists of single Pauli strings, the magnitude of each such gradient is obtained from a single overlap evaluation, without reference to the Hamiltonian. Qubit count may be reduced further by projecting the Hamiltonian onto the subspace spanned by the dominant determinants of a one-layer ansatz. Finally, a stochastic implementation of the second-order Epstein-Nesbet perturbative correction (ENPT2) [18] is introduced to shift the remaining error onto classical GPU-accelerated hardware. Leveraging the optimized VQE energy and wavefunction to preserve quantum resources, it permits an early termination of the quantum optimization. Together these give a tunable trade-off — measurement cost against circuit depth, qubit count, and classical post-processing, along which the user may move according to which resource is scarcest.

We term this framework VULCAN. In noiseless emulation, on linear hydrogen chains, BeH₂ and H₂O, VULCAN reaches chemical accuracy at 5–20% of the Pauli-term evaluation count of batched ADAPT-VQE with QEB pools, and overlap compression brings the circuit depth below that baseline for four of the seven systems where it was applied. For H₂O in the 6-31G basis (26 qubits) and C₃H₆ (32 qubits), stochastic ENPT2 reaches chemical accuracy after a single layer. Subspace projection reaches chemical accuracy in exact diagonalization with 23–50% fewer qubits. Section II presents these results, and Section IV details the pool construction, compression loop, projection scheme and perturbative correction.

Results

Algorithmic Workflow and Conceptual Schematic:

The VULCAN workflow is executed in three primary phases, illustrated in Figure 1, with an optional projection step inserted between the first two.

Algorithmic pipeline showing Phase 1 pool generation, Phase 2 grow-then-compress, and Phase 3 correction

Figure 1. Overview of the algorithmic pipeline. Phase 1 generates the operator pool classically: the VDBF flow generator is evaluated as a commutator with the target Hamiltonian and split into individual Pauli strings; operators are added to selected starters among them until a minimal complete pool is obtained, and the remaining VDBF operators are then appended. An optional projection step, between Phases 1 and 2, selects kk excitation patterns from the first-layer ansatz, takes HprojH_{\mathrm{proj}} as the projection of HH on the 2k2^k determinants they generate, and projects the pool onto the reduced register. Phase 2 grows the ansatz one complete layer at a time, optimizing only the parameters of the newly appended layer. The optimized state serves as the target of an overlap-guided compression step, which inserts generators by decreasing overlap gradient until the infidelity gain becomes negligible or the compressed layer reaches a set fraction of the grown layer; the compressed layer is then re-optimized for energy before the next growth iteration. Phase 3 applies a classical, GPU-accelerated ENPT2 correction, permitting early termination of Phase 2.

  • Phase 1: Operator Pool Initialization (classical). The algorithm employs the flow equation, which is evaluated as a commutator with the target Hamiltonian, yielding a linear combination of Pauli operators. These are split into individual Pauli strings, operators are added to selected starters among them until a minimal complete pool of 2N−42N - 4 generators is obtained, and the remaining VDBF operators are then appended. The pool is produced in seconds on a classical computer and requires no preliminary VQE run.

  • Phase 2: Grow-then-compress loop (hybrid). The algorithm enters an iterative cycle of ansatz construction and compression. Rather than selecting individual operators via Hamiltonian gradients, each iteration appends one complete pool layer to the ansatz and minimizes the energy over the new parameters, with preceding layers frozen. The resulting state serves as the target for a compression sub-loop, in which generators are inserted one at a time in order of decreasing overlap gradient and fitted by infidelity minimization, until the infidelity stops improving or the compressed layer reaches a set fraction of the grown layer. The compressed layer is then re-optimized for energy. The loop terminates when the @referencesWinvalid

energy change between iterations falls below ϵ\epsilon or a maximum iteration count is reached.

  • Optional: Subspace projection. To maximize viability on NISQ hardware, for systems whose qubit count exceeds the available register, a single ansatz layer is optimized greedily on a classical emulator to identify its dominant excitation patterns. The Hamiltonian is projected onto the subspace spanned by all combinations of these patterns applied to the Hartree–Fock state. The original pool is projected onto the subspace: generators whose excitation lies in the span of the selected patterns are kept and rewritten on the reduced register, and the workflow resumes at Phase 2 (growth and compression) on that register.

  • Phase 3: Classical Post-Processing (classical, GPU). To achieve a better accuracy without adding quantum gate depth, the algorithm applies a second-order perturbation theory. The optimized energy seeds a second-order Epstein-Nesbet correction, evaluated exactly for small systems and by importance sampling for large ones. Because this step is classical, it adds no quantum resource cost and allows Phase 2 to be terminated early.

Numerical setup:

All simulations were performed with Hyperion, a GPU-accelerated quantum emulator optimized for low-level statevector operations [19]. Hamiltonian integrals were computed with OpenFermion-PySCF, and the Jordan–Wigner transformation maps the fermionic operators to qubit space. All results are obtained by noiseless state-vector emulation. Unless otherwise stated, the STO-3G basis set is used. We benchmarked VULCAN on linear hydrogen chains (H6_6, H8_8, H10_{10}) at d=0.85d = 0.85 and 3.03.0 Å, BeH2_2 at 3.03.0 Å, H2_2O in the 6-31G basis (26 qubits), and C3_3H6_6 in an active space of 18 electrons in 16 spatial orbitals (32 qubits). The subspace projection results for H4_4, BeH2_2 and H2_2O (STO-3G), and H2_2O (6-31G, 26 qubits) in Section II E demonstrate the reduction in qubit count achieved by projecting the Hamiltonian onto the 2k2^k determinants generated by kk excitation patterns of a single-layer ansatz. Table I compares the VULCAN pool sizes with those of the QEB pool, which are 4 to 25 times larger. The baseline throughout is batched ADAPT-VQE [20] using the QEB pool, in which the bb highest-gradient operators are selected at each iteration. The batch sizes bb listed in Table I are chosen so that further increases yield no significant improvement in the ADAPT-VQE convergence rate for the systems studied; the baseline therefore represents the best case with respect to batch size.

Performance is measured against exact full configuration interaction (FCI) energies using two hardware-resource metrics. The first is the CNOT gate count, which governs circuit depth and hence sensitivity to decoherence. The second metric is a hardware-agnostic proxy, the cumulative number of Pauli-term evaluations, in the figures and tables, cumulative hardware circuit executions denotes this Pauli-term evaluation count.

One energy evaluation ⟨ψ∣H∣ψ⟩\langle\psi|H|\psi\rangle requires NHN_H quantum measurements, where NHN_H is the number of measurable Pauli strings in the Hamiltonian, evaluating one ADAPT gradient also costs nn measurements per candidate operator, whereas evaluating one overlap gradient in our framework costs a single measurement (Section IV C). Energy evaluations performed by the classical optimizer are included for both methods. Finally, the ENPT2 correction is computed classically on GPU with Hyperion and incurs no additional quantum resource cost.

Measurement cost: gradient-free layered growth

At each iteration, the complete VULCAN pool (generated as described in Section IV B) is appended to the ansatz and the new parameters are optimized, while the parameters of previously added layers are kept fixed. Because the pool is physics-informed and complete, the ansatz reaches chemical accuracy without the multi-iteration descent required by batched ADAPT-VQE (Figure 2–Figure 5, right panels, orange squares versus blue curves). For stretched hydrogen chains (H6_6, H8_8, H10_{10} at d=3.0d=3.0 Å), chemical accuracy is achieved at the first iteration, except for equilibrium geometries, it is reached within the first few iterations.

Energy error with respect to the FCI energy for the BeH₂ molecule at an internuclear distance of $d = 3.0$ Å

Figure 2. Energy error with respect to the FCI energy as a function of the number of CNOTs (left panel) and of the cumulative hardware circuit executions (right panel) for the BeH₂ molecule at an internuclear distance d=3.0d = 3.0 Å. Blue circles (“Standard QEB”): batched ADAPT-VQE with the QEB pool; orange squares (“Standard Vulcan”): VULCAN without compression; green circles: VULCAN growth iterations; large pink ‘X’: compressed iterations; thin ‘×’: PT2-corrected energies. Numbers label the iterations. The grey band indicates chemical accuracy.

Plots of energy error versus number of CNOT gates and cumulative hardware circuit executions for H₂O

Figure 5. Same as Fig. 2 for the H₂O molecule in the 6-31G basis (26 qubits).

The low measurement cost follows from the elimination of Hamiltonian gradient evaluations. In VULCAN, the only quantum measurements are energy evaluations, each costing NHN_H Pauli-string measurements. By contrast, each batched ADAPT-VQE iteration evaluates gradients for every operator in a pool 4 to 25 times larger, at a cost of NHN_H measurements per operator.

As shown in Table II (Standard column), the measurement cost at chemical accuracy is a small fraction of the ADAPT-VQE baseline for all systems studied of Table II. For the hydrogen chains at 0.850.85 Å, the ratio ranges from 8% (H6_6) to 18% (H8_8), with H10_{10} at 10%, and it is 5% for H2_2O. For stretched geometries (d=3.0d = 3.0 Å), where convergence is typically harder, the advantage is equally pronounced: 10% for BeH2_2, 12% for H6_6 and H10_{10}, and 19% for H8_8.

These reductions are a direct consequence of bypassing the gradient selection phase entirely, and are obtained without compression, circuit depth is addressed in the following section.

Circuit depth: overlap-guided compression

While VULCAN reduces measurement cost substantially (Section II C), appending a complete pool at each iteration produces circuits that are deeper than those of batched ADAPT-VQE. At the first chemically accurate iteration, the CNOT count is 1.3–2.9 ×\times that of the baseline for equilibrium geometries and 1.0–1.9×\times for stretched geometries (Table II, Standard column). We now show how overlap-guided compression reduces this depth penalty, and removes it at stretched geometries.

Compression operates in two phases, as illustrated in Figure 2–Figure 4: a growth phase (green circles), where the complete pool is appended and optimized, and a compression phase (large filled pink ‘X’ symbols), where operators are re-selected by overlap gradient and the ansatz is pruned to produce a shallower circuit. Numerical markers in the figures link each compression target to its corresponding compressed ansatz. Compression inserts generators one at a time until the infidelity gain becomes negligible or the compressed layer reaches a set fraction of the grown layer (typically 50%). The initial growth iteration (labeled 1) corresponds to the first uncompressed VULCAN iteration. Details of the overlap compression algorithm are given in Section IV C.

Six-panel convergence plots for H₆, H₈, and H₁₀ hydrogen chains

Figure 4. Same as Figure 2 for the hydrogen chains H₆, H₈ and H₁₀ (top to bottom) at an internuclear distance d=3.0d = 3.0 Å.

For the equilibrium hydrogen chains, compression does not bring the CNOT count below the baseline (H6_6: 1.29, H8_8: 1.94, H10_{10}: 1.45; Table II, Compressed column). Compression is more effective for the stretched geometries: the ratio drops to 0.90 for BeH2_2, 0.49 for H6_6, 0.29 for H8_8 and 0.14 for H10_{10}. At stretched geometries, the chemically accurate states obtained can be spin-contaminated, and compression reproduces any contamination of its target.

It is worth noting that compression is not free in measurement cost. The overlap-guided compression adds circuit executions, raising the cumulative measurement cost from the uncompressed values reported in Section II C: for example, H8_8 at 0.85 Å goes from 0.18 to 0.59 of the baseline, and H10_{10} from 0.10 to 0.35. Crucially, the compressed measurement cost remains below the batched ADAPT-VQE baseline in every case tested. The user therefore faces a tunable trade-off: stop after growth for minimal measurement cost, or compress for a shallower circuit at a higher, but still sub-baseline, measurement budget. Overlap compression reduces gate count at fixed qubit number. For systems whose register size itself exceeds the available hardware, a complementary reduction is needed: projecting the Hamiltonian onto a smaller subspace before the grow-then-compress loop begins.

Qubit count: subspace projection

Given the operator pool from Phase 1, we select kk excitation patterns from the most important generators of the first-layer ansatz. The K=2kK = 2^k determinants obtained by applying all their combinations to the Hartree–Fock state span the subspace, and HprojH_{\mathrm{proj}} is the corresponding K×KK \times K submatrix of HH, a Hamiltonian on kk qubits.

The subspace is grown one qubit at a time. At each kk, a second-order Epstein–Nesbet estimate of the energy contribution from the determinants excluded by the projection is computed. Two variants of this estimate are available: one that takes the exact ground state of the current HprojH_{\mathrm{proj}} as its reference, and a cheaper one that uses the Hartree–Fock state directly. When the estimate falls below a tolerance τ\tau, the growth stops (Section IV D).

Figure 7–10 show the results for H₄ (STO-3G, 8 qubits), BeH₂ (STO-3G, 14 qubits), H₂O (STO-3G, 14 qubits), and H₂O (6-31G, 26 qubits). Each figure contains three panels.

Exact diagonalization vs $k$

(a)

VULCAN vs CNOTs

(b)

Exact diagonalization vs Pauli terms

(c)

Figure 7. Adaptive-kk projection for the H4_4 molecule at 0.85 Å (STO-3G, 8 qubits). (a) As a function of the number of qubits kk kept by the projection: energy error of the ground state of the projected Hamiltonian HprojH_{\mathrm{proj}} (∣E0−EFCI∣|E_0-E_{\mathrm{FCI}}|), PT2 estimate of the energy lost by the projection with the exact subspace ground state (“External PT2”) or the Hartree–Fock state (“External PT2 (HF)”) as reference, and error of one VULCAN iteration on HprojH_{\mathrm{proj}}. (b) Error of one VULCAN iteration on HprojH_{\mathrm{proj}} as a function of the number of CNOTs; the star marks one VULCAN iteration on the full Hamiltonian, and the dashed line its CNOT count. (c) Exact-diagonalization error as a function of the number of Pauli terms of HprojH_{\mathrm{proj}}; the dashed line marks the number of terms of the full Hamiltonian. The grey band indicates chemical accuracy.

Panel (a) shows the energy error of the exact diagonalization of HprojH_{\mathrm{proj}} as a function of subspace size kk, together with the two PT2 estimates and the error of a single VULCAN iteration on HprojH_{\mathrm{proj}}. In all three STO-3G systems, the exact-diagonalization error drops rapidly with kk and enters the chemical accuracy band well before the full register is reached: at k=4k = 4 out of 8 for H₄, k=7k = 7 out of 14 for BeH₂, and k=7k = 7 out of 14 for H₂O. The PT2 correction estimates the energy lost by the projection. With the exact subspace ground state as reference, this estimate agrees closely with the actual energy loss at every kk for H₄ and H₂O, and for BeH₂ once the dominant determinants are included (k≥7k \ge7); below that it underestimates the loss by about a factor of two. The Hartree–Fock-based estimate is cheaper, since it does not require diagonalizing HprojH_{\mathrm{proj}}, and follows the same trend, although it can lie above or below the actual loss. Both fall below the tolerance at the same kk, providing an automatic stopping criterion.

Panel (b) shows the error of a single VULCAN iteration on HprojH_{\mathrm{proj}} as a function of CNOT count. The star marks the same single-iteration calculation on the full Hamiltonian. For H₄, the projected calculation reaches chemical accuracy at a fraction of the full-register gate cost. For stretched BeH₂, a single VULCAN iteration does not reach chemical accuracy on the projected register, and increasing kk beyond 7 does not help: the exact error at k=7k = 7 is already 3.4×10−43.4 \times10^{-4} Ha, so the remaining error comes from the single iteration. A second iteration on the 7-qubit register reaches chemical accuracy (3.8×10−43.8 \times10^{-4} Ha) with 680 CNOTs, against 1084 CNOTs for two iterations on the full 14-qubit register.

These circuits can be shortened further without loss of accuracy. Choosing the basis of the subspace that minimizes the CNOT count of the projected pool, and compressing the projected layer with the overlap-guided procedure (see Supplementary Material), reduces the cost by 40–74% at the same energy (Table III). The resulting circuits require 1.4 to 6 times fewer CNOTs than ADAPT-VQE with the QEB pool on the full register, and 2.6 to 10 times fewer than VULCAN on the full register, on a smaller register (half the qubits for the STO-3G systems, 20 of 26 for H₂O in 6-31G).

Panel (c) shows the exact-diagonalization error as a function of the number of Pauli terms in HprojH_{\mathrm{proj}}. As kk decreases, the projected Hamiltonian becomes substantially simpler than the full Hamiltonian (dashed line). For H₂O, the full Hamiltonian contains 2109 Pauli terms, while HprojH_{\mathrm{proj}} at k=7k = 7 contains only 256. This reduces the per-evaluation measurement cost, since each energy measurement scales with the number of Hamiltonian terms.

Figure 10 extends the projection to H₂O in the 6-31G basis, a 26-qubit system. Panel (a) shows the exact-diagonalization error reaching chemical accuracy at k=20k = 20 out of 26 qubits, confirming that the projected subspace captures the essential correlation energy even for this larger system. Panel (c) confirms the Hamiltonian simplification: the full system contains 25035 Pauli terms, reduced to 8911 at k=20k = 20.

Exact diagonalisation vs $k$

(a) Exact diagonalisation vs kk

VULCAN vs CNOTs

(b) VULCAN vs CNOTs

Exact diagonalisation vs Pauli terms

(c) Exact diagonalisation vs Pauli terms

Figure 10. Same as Figure 7 for the H2_2O molecule (6-31G, 26 qubits).

Panel (b) shows the circuit cost. A single VULCAN iteration on the projected register at k=20k = 20 requires 7658 CNOTs, about the same as the 7728 CNOTs of the same iteration on the full 26-qubit register, and the cost grows quickly with kk: at k=23k = 23, where the projection keeps the whole space and recovers the full-register energy, the same iteration requires 27456 CNOTs. Fewer qubits therefore do not give a shallower circuit here. This is not due to the pool, since each projected generator performs the same rotation as on the full register, but to the cost of each generator: on the full register a generator acts on at most four spin orbitals, whereas on the projected register it acts on a growing fraction of the qubits. This cost can be reduced by choosing a different basis for the same subspace and compressing the layer: at k=20k = 20, the circuit then requires 2012 CNOTs, against 4043 for ADAPT-VQE with the QEB pool at the same accuracy (Table III).

Projection consistently reduces the qubit register and the Hamiltonian complexity. For the STO-3G systems, it also gives shallower circuits; for H₂O in 6-31G, this requires the optimized basis. Once the basis is optimized and the circuit compressed, the projected circuit is shorter than ADAPT-VQE with the QEB pool at the same accuracy in all four systems (Table III), and the PT2 estimate provides a stopping criterion for the growth of the subspace in all of them.

Early termination: Stochastic PT2

As the number of ansatz layers increases, the marginal energy improvement from each additional layer diminishes. Rather than deepening the circuit further, we apply a second-order Epstein–Nesbet perturbation theory (ENPT2) correction classically, using only the energy and wavefunction already obtained from the preceding quantum

MoleculeCNOTsCumulative hardware executions
StandardCompressedPT2StandardCompressedPT2
H₆ - 0.85 Å1.301.291.290.080.240.24
H₈ - 0.85 Å2.901.940.480.180.590.02
H₁₀ - 0.85 Å1.721.450.140.100.350.002
BeH₂ - 3.0 Å1.920.900.900.100.350.35
H₆ - 3.0 Å0.990.49-0.120.19-
H₈ - 3.0 Å1.390.29-0.190.22-
H₁₀ - 3.0 Å1.390.14-0.120.13-
H₂O (6-31G)1.59-0.530.05-0.0005

Table 2. Fraction of CNOT counts and cumulative hardware executions at the first iteration where chemical accuracy is reached for each method with respect to the same metric with batched ADAPT-VQE with the QEB pool. Standard means the standard VULCAN, and Compressed means the compressed VULCAN. The PT2 results are calculated using the compressed VULCAN, except for H₂O, for which the standard VULCAN was used. Bold numbers mean that our method needed fewer CNOTs or measurements than batched ADAPT-VQE with the QEB pool to achieve chemical accuracy. –: not computed (no compression for H₂O) or not needed (PT2 at stretched geometries, where chemical accuracy is reached without correction).

SystemQubits full → kkVULCAN (full register)QEB (full register)Original basis (no compression)Optimized basis (+ compression)
H₄8 → 4142 (2.1×10−52.1 \times10^{-5})130 (1.3×10−41.3 \times10^{-4})62 (1.3×10−41.3 \times10^{-4})22 (1.3×10−41.3 \times10^{-4})
BeH₂ (3.0 Å)14 → 71084 (8.4×10−58.4 \times10^{-5})593 (8.5×10−58.5 \times10^{-5})680 (3.8×10−43.8 \times10^{-4})410 (3.8×10−43.8 \times10^{-4})
H₂O (STO-3G)14 → 71598 (4.7×10−54.7 \times10^{-5})338 (2.1×10−42.1 \times10^{-4})302 (2.1×10−42.1 \times10^{-4})152 (2.2×10−42.2 \times10^{-4})
H₂O (6-31G)26 → 207728 (2.5×10−32.5 \times10^{-3})4043 (3.5×10−33.5 \times10^{-3})7658 (3.5×10−33.5 \times10^{-3})2012 (3.5×10−33.5 \times10^{-3})

Table 3. CNOT counts, with the energy error with respect to FCI (Ha) in parentheses (STO-3G basis unless stated). Qubits: size of the full register and of the projected register kk. VULCAN (full register): uncompressed VULCAN on the full register, with the same number of iterations as the original-basis column (for one iteration, the star in panel (b) of Figs. 7–10), given with its own error; unlike the other columns, it is not taken at the accuracy of the original-basis column. QEB (full register): ADAPT-VQE with the QEB pool, one operator added per iteration, stopped at the first iteration whose error is at most that of the original-basis column. Original basis (no compression): one VULCAN iteration on the kk-qubit register with the projected pool (two iterations for BeH₂, where a single iteration stops at 1.7×10−21.7 \times10^{-2} Ha); its error sets the accuracy at which the QEB and optimized-basis columns are compared. Optimized basis (+ compression): the same iterations in the optimized basis of the subspace, rebuilt by overlap-guided compression on the kk qubits and stopped at the first circuit reaching the error of the original-basis column (Supplementary Material). Bold: fewest CNOTs at equal accuracy.

Six-panel plots showing results for hydrogen chains H₆, H₈ and H₁₀

Figure 3. Same as Figure 2 for the hydrogen chains H₆, H₈ and H₁₀ (top to bottom) at an internuclear distance d=0.85d = 0.85 Å.

optimization. The ENPT2 correction adds small contributions of determinants outside the variational state, as in selected configuration interaction (sCI) methods. Because this step adds no quantum circuit executions, it effectively shifts the remaining error onto classical GPU hardware. PT2-corrected energies are indicated in Figure 2–Figure 5 by thin pink ‘×’ markers, distinct from the larger filled ‘X’ of the compressed iterations, with numerical labels identifying the compressed VULCAN iteration used as the seed. For equilibrium geometries, the correction consistently brings the energy below chemical accuracy using circuits from the first or second compressed iteration. The corresponding measurement and CNOT costs are reported in Table II (PT2 columns) and reflect only the quantum resources of the seed iteration, not the classical PT2 cost. PT2 enables early termination of the quantum loop, reaching chemical accuracy from a shallower seed circuit with far less accumulated measurement cost. For H₈ at 0.85 Å, chemical accuracy via compression alone requires iterating until 0.59 of the baseline measurement budget has been spent, applying PT2 to the first compressed iteration instead reaches the same accuracy at a cumulative cost of only 0.02. For H₁₀ at 0.85 Å, the corresponding reduction is from 0.35 to 0.002. For stretched geometries, where chemical accuracy is already reached without correction, PT2 is omitted.

For larger systems, exact evaluation of the ENPT2 sum is prohibitive due to exponential scaling. We employ a stochastic variant based on importance sampling (Section IV E), which evaluates only a small fraction of the perturbative terms.

We test the stochastic variant on two larger systems: H₂O in the 6-31G basis, a 26-qubit problem (Figure 5), and C₃H₆ (Figure 6).

Plots of energy error versus number of CNOT gates and cumulative hardware circuit executions for C₃H₆

Figure 6. Same as Fig. 2 for the C₃H₆ molecule in the STO-3G basis. The calculations use an active space of 18 electrons in 16 spatial orbitals (32 qubits). The FCI reference energy is computed within the same active space.

For H₂O, the ENPT2 sum contains 2262^{26} terms, one for each basis state of the 26-qubit register. The stochastic method samples a small subset of these terms, with probabilities proportional to the squared couplings ∣H0j∣2\lvert H_{0j}\rvert^2, and estimates the total correction from that subset alone. As shown in Table IV, evaluating fewer than 0.01% of the terms (about 1700 to 6700 denominators) is sufficient: for the first VULCAN iteration, the estimated correction stabilizes at approximately −2.5×10−3-2.5 \times10^{-3} Ha with a standard error of a few 10−510^{-5} Ha, and varies by about 10−410^{-4} Ha across subset sizes, well below chemical accuracy. Chemical accuracy is reached already at the first VULCAN iteration combined with this correction, at a measurement cost of 0.0005 of the batched ADAPT-VQE baseline. The second-iteration corrections show comparable convergence. This is the same 26-qubit system for which subspace projection (Figure 10) reduces the register to k=20k = 20 qubits while preserving chemical accuracy in exact diagonalization. On that register, the circuit in the optimized basis needs fewer CNOTs than one VULCAN iteration on the full register (2012 against 7728, Table III), but stops at 3.5×10−33.5 \times10^{-3} Ha.

1st Iteration
ff (×10−4\times10^{-4})ΔE\Delta E (×10−3\times10^{-3})σ\sigma (×10−5\times10^{-5})
0.25-2.4006.5
0.5-2.4684.6
0.75-2.4963.4
1.0-2.5063.0
2nd Iteration
ff (×10−4\times10^{-4})ΔE\Delta E (×10−3\times10^{-3})σ\sigma (×10−5\times10^{-5})
0.25-1.6636.5
0.5-1.7164.8
0.75-1.7963.8
1.0-1.7294.9

Table IV. Stochastic ENPT2 correction for H2_2O (6-31G, 26 qubits), seeded by the first and second VULCAN iterations. ff is the fraction of the 2262^{26} terms of Eq. (15) whose denominators are evaluated, ΔE\Delta E is the estimated correction in Ha, and σ\sigma is the standard error of the estimator (Algorithm 2).

For C₃H₆ (STO-3G, active space of 18 electrons in 16 orbitals, 32 qubits), evaluating only 1484 denominators, the stochastic correction reduces the energy error to 1.0×10−31.0 \times10^{-3} Ha at the first VULCAN iteration, after 9.6×1059.6 \times10^{5} circuit executions (Figure 6). Batched ADAPT-VQE remains 6.4×10−36.4 \times10^{-3} Ha above the FCI energy after 359 iterations and 3.3×10113.3 \times10^{11} circuit executions. The computational cost of this step is reported in Section IV E.

The combination of a single VULCAN layer with stochastic ENPT2 therefore provides a practical route to chemical accuracy, within the emulation regime, for systems where the full grow-then-compress loop would be expensive: the quantum circuit remains shallow, and importance sampling reduces the number of classical denominator evaluations, although memory still scales with the full state vector (Section IV E).

One can see that the formula for the PT2 (Eq. (15)) diverges if the denominator becomes too small. Although we did not observe this divergence in any of the systems tested, we also implemented the Diagonal Unitary Correction (DUC) formula (Eq. 25 of [18]) to prevent any potential instabilities. The DUC yields identical results to Eq. (15), even when combined with the stochastic method.

Discussion

In this work, VULCAN offers an alternative approach to the gradient-based operator selection of ADAPT-VQE with a physics-derived, complete pool whose generators are produced in seconds on a classical computer. The cost of this method is a set of trade-offs: between measurement budget, circuit depth, qubit count, and classical post-processing, that the user navigates according to whichever hardware resource is the scarcest. When the dominant constraint is measurement budget, the uncompressed layered ansatz is the appropriate operating point: it reaches chemical accuracy at 5–20% of the baseline measurement cost, but with deeper circuits. When the circuit depth is the binding constraint, overlap compression recovers depths comparable or shorter than batched ADAPT-VQE for four of the seven systems, at a measurement cost that remains below the baseline in every case tested. When the required number of qubits exceeds the available register, subspace projection reduces the register before any quantum circuit is executed, with the PT2-guided stopping criterion on the dimension of the projected subspace. Finally, when GPU resources are available, stochastic ENPT2 permits early termination of the quantum loop entirely: for H2_2O in the 6-31G basis, a single layer suffices once the perturbative correction is applied classically.

Projection, compression, and perturbative correction act on different resources and are complementary. For H2_2O in 6-31G, projection reduces the 26-qubit register to 20 qubits (Figure 10), and stochastic ENPT2 separately reaches chemical accuracy after one layer on the full register (Figure 5). On the projected registers, choosing the basis of the subspace and compressing the projected layer gives circuits shorter than both VULCAN and ADAPT-VQE with the QEB pool on the full register, for all four systems tested (H4_4, BeH2_2, H2_2O in STO-3G and 6-31G, Table III).

We benchmarked VULCAN on H6_6, H8_8, H10_{10}, BeH2_2, H2_2O and C3_3H6_6, comparing against batched ADAPT-VQE with QEB pools, an operator pool 4 to 25 times larger than ours with gradient-based selection optimized to its best-case batch size. Despite this, VULCAN reached chemical accuracy at 5–20% of the baseline measurement cost for all systems of Table II. After overlap compression, the CNOT count fell below the baseline for four of the seven systems, all at stretched geometries, and for stretched hydrogen chains the improvement grew with chain length, reaching 0.14× for H10_{10}. Stochastic ENPT2 extended this to 26 qubits (H2_2O, 6-31G), where chemical accuracy was achieved after a single layer at 0.05% of the baseline measurement cost, and to 32 qubits (C3_3H6_6), where batched ADAPT-VQE does not reach chemical accuracy.

Three-panel comparison for the BeH$_2$ molecule at 3.0 Å

Figure 8. Same as Figure 7 for the BeH2_2 molecule at 3.0 Å (STO-3G, 14 qubits).

Three-panel comparison for the H$_2$O molecule

Figure 9. Same as Figure 7 for the H2_2O molecule (STO-3G, 14 qubits).

Outlook. Although VULCAN is ultimately aimed at hardware implementations, all presented results are based on noiseless state-vector emulations, with costs reported as Pauli-term evaluations rather than shot counts. The behavior of VULCAN under shot noise and hardware error, particularly the optimization of layers containing up to ∼103\sim10^3 parameters, remains to be established. Because reproducing all proposed VULCAN features demands significant quantum resources and careful, hardware-dependent measurement management, this will be the focus of a separate hardware implementation study. Even without these incoming enhancements and as a preliminary proof of concept, we evaluated the total energy of dihydrogen (H2_2) on the Quantinuum System Model H2 trapped-ion (see Supplementary Material) and found a good match (−1.12666-1.12666 Ha) with the emulated theoretical values (−1.12836-1.12836 Ha), 1.7×10−31.7 \times10^{-3} Ha above the FCI energy, close to chemical accuracy (1.6×10−31.6 \times10^{-3} Ha).

Several technical aspects will also require dedicated studies beyond the scope of the present work. First, the ordering of Pauli-string exponentials within a layer modifies the ansatz and may impact convergence; exploring this dependence will be a valuable direction. Second, the single-Pauli VULCAN pool does not explicitly conserve particle number, SzS_z, or total spin. While this has no visible effect at equilibrium geometries, in the strongly correlated regime of stretched chains, states achieving chemical accuracy in energy can exhibit spin contamination and partially deviate from the (Nα,Nβ)(N_\alpha, N_\beta) sector of the reference state. Such spin contamination is a well-known phenomenon not specific to VULCAN, appearing similarly in other symmetry-unadapted methods like unrestricted Hartree–Fock or nonspin-adapted DMRG. A rigorous analysis of these effects in challenging, highly correlated benchmark systems will be highly useful. Third, all energies above are converged within a finite basis set. Reaching chemical accuracy with respect to the complete basis set additionally requires a basis-set correction. In the Supplementary Material (Section 2), we combine VULCAN with the density-based basis-set correction (DBBSC) [21], which estimates the missing energy from the density of the quantum state at no additional quantum cost. Results demonstrate the full applicability of the DBBSC approach to VULCAN as it was demonstrated for ADAPT-VQE[21].

Finally, although an extensive classical ENPT2 correction is a highly desirable strategy for early termination of the quantum sequence, developing a fully quantum ENPT2 evaluation method will become increasingly relevant as quantum processors advance toward fault-tolerant quantum computing (FTQC).

Concluding Remarks. Overall, future work will fofocus on three primary directions: (1) deploying VULCAN on physical NISQ hardware to directly test noise resilience; (2) integrating projection with the full grow-then-compress loop for systems beyond minimal basis sets; and (3) restoring the symmetries of the state. This last objective does not require altering the pool or the circuit architecture; instead, penalty terms can be incorporated into the cost function to be minimized, analogous to constrained VQE. The following table, Table V summarizes: i) the usual VQE bottlenecks addressed in this work when dealing with quantum chemistry; ii) the VULCAN component that solves each of them, and iii) the remaining open questions.

VQE BottleneckTechnical solution (VULCAN)Results in this work
Energy-gradient measurements to select operators (ADAPT-VQE)Whole pool appended as one layer, no operator selectionChemical accuracy at 5–20% of the measurement cost of batched ADAPT-VQE (hydrogen chains, BeH2_2, H2_2O, up to 26 qubit simulations)
Choice of starters for MCPsPhysically motivated starters generated from the Hamiltonian by the VDBF flow equation, completed into an MCPPool produced with no preliminary VQE runs, 4–25 times smaller than the QEB pool
Circuit depth of layered ansätzeState-overlap compressionCNOT count below batched ADAPT-VQE for four of the seven tested systems
Number of qubits of the registerProjection of the Hamiltonian onto a dominant-determinant subspaceChemical accuracy in exact diagonalization with 23–50% fewer qubits and 60–88% fewer Pauli terms (H4_4, BeH2_2, H2_2O; up to 26 qubits)
Circuit cost on the projected registerOptimized subspace basis and cost-aware compression40–74% fewer CNOTs than in the original basis; 1.4–6 times fewer than ADAPT-VQE (QEB) on the full register, at the same accuracy.
Number of iterations and measurements to reach chemical accuracyENPT2 correctionChemical accuracy after a single VULCAN layer: H2_2O (6-31G, 26 qubits) at 0.05% of the cost of batched ADAPT-VQE, C3_3H6_6 (32 qubits), where batched ADAPT-VQE does not reach it
High number of terms in the ENPT2 correction for larger systemsStochastic ENPT2Fewer than 0.01% of the terms evaluated for H2_2O (6-31G); 1484 denominators for C3_3H6_6

Table V. Summary of the VQE bottlenecks in quantum chemistry addressed by VULCAN and corresponding results obtained in this work.

Methods

The layered ansatz

The VULCAN ansatz is defined in Eq. (1). Each product ∏j=1Mexp⁡(iθj,kPj)\prod_{j=1}^{M}\exp(i\theta_{j,k}P_j) constitutes one layer. Layers are appended one at a time. When a new layer kk is added, only its parameters {θj,k}j=1M\{\theta_{j,k}\}_{j=1}^{M} are optimized, all preceding layers are held frozen. This greedy scheme bounds the number of free parameters per optimization to the pool size MM, independent of the total circuit depth.

In the compressed variant, each grown layer is replaced by a shorter layer built by overlap-guided insertion (Section IV C). The compressed ansatz retains the product-of-exponentials form, with a sequence (jk,1,…,jk,Lk)(j_{k,1},\ldots,j_{k,L_k}) of LkL_k pool indices, in which an index may appear more than once:

∣ψ⟩=∏k=1n(∏l=1Lkexp⁡(iθl,kPjk,l))∣ψ0⟩,(2)|\psi\rangle= \prod_{k=1}^{n}\left(\prod_{l=1}^{L_k}\exp(i\theta_{l,k}P_{j_{k,l}})\right)|\psi_0\rangle, \tag*{(2)}

where the most recently inserted generator acts last, and Lk≤max⁡(1,⌊fM⌋)L_k \le\max(1,\lfloor fM\rfloor), where ff is the compression fraction (typically f=0.5f=0.5).

Operator pool generation

The necessity of constructing operator pools based on the MCP framework, defined as the minimal generating set for a target Lie algebra, is driven by the severe measurement overhead inherent in gradient-based adaptive techniques. In standard protocols, iterative operator selection requires evaluating an energy gradient for every pool element, a process that scales naively O(N4)O(N^4) with the number of qubits and rapidly becomes computationally intractable. Because MCPs scale linearly with the number of qubits NN, they have the potential to mitigate this bottleneck.

Existing pools and the starter challenge

The expressivity and convergence rate of adaptive variational algorithms depend on the choice of the operator pool. Traditionally, these pools are constructed using heuristic or physically motivated criteria to align with the target problem’s structure. Prominent examples in the literature include:

  • Fermionic pools [6]: Comprise physically motivated operators, typically the set of all unique, spin-complemented one- and two-body excitations.

  • Qubit pool [7]: Derived by decomposing the Fermionic-ADAPT operators via the Jordan-Wigner mapping, treating the individual Pauli strings as the fundamental operators of the pool.

  • Qubit-Excitation-Based (QEB) pools [8]: Employ exponential single- and double-qubit excitation evolutions, expressed directly through qubit creation and annihilation operators mapping.

  • Coupled Exchange Operator (CEO) pool [22]: Constructed from linear combinations of qubit excitations that act on the exact same set of spin-orbitals, designed to strictly preserve specific symmetries.

A limitation of these pools is that their size scales polynomially or worse with respect to the number of qubits. To overcome this, recent works [7, 12, 13, 14, 20] have introduced MCPs. These pools scale linearly with the number of qubits yet retain sufficient expressivity to reach any state of the Hilbert space. However, selecting the optimal elements to seed an MCP remains a challenge. Existing strategies such as the Unitary Selective Coupled-Cluster (USCC) approach [15] isolate physically relevant starters via amplitude thresholds and symmetry filtering from a preliminary VQE run. While effective, this requires computing and storing optimized amplitudes before screening can begin [12, 14], replacing one bottleneck with another.

Dynamic generation via flow equations

To overcome the limitations of static pre-screening, we dynamically synthesize the most relevant operators using flow generators. By leveraging algebraic commutators derived from the target Hamiltonian itself, we generate a compact, highly expressive operator pool without the need for prior classical optimizations.

The foundation of our pool generation relies on the Variational Double Bracket Flow (VDBF) framework. Originally introduced by Wegner [16] to diagonalize Hamiltonians, double bracket flows have recently been adapted by Shrikhande et al. [17] as a means to efficiently estimate ground state energies via classical Pauli propagation. The flow is defined by the following differential equation

ddsH(s)=[G(s),H(s)]=[[HD(s),H(s)],H(s)](3)\frac{d}{ds}H(s)=[G(s),H(s)]=[[H_D(s),H(s)],H(s)] \tag*{(3)}

where G(s)G(s) is an anti-Hermitian flow generator, describing the evolution of the Hamiltonian along a flow parameter ss. To drive the system towards a specific target state rather than achieving full diagonalization, one defines HD=ρH_D=\rho, where ρ\rho is the projector onto a reference computational basis state. Taking the zero state ρ=∣0N⟩⟨0N∣\rho=|0^N\rangle\langle0^N| and expanding the projector in the Pauli basis gives

∣0⟩⟨0∣=I+Z2,(4)|0\rangle\langle0|=\frac{I+Z}{2}, \tag*{(4)}

so that for an NN-qubit system

∣0N⟩⟨0N∣∝I+∑iZi+∑i<jZiZj+⋯ .(5)\lvert0^N\rangle\langle0^N\rvert\propto I + \sum_i Z_i + \sum_{i<j} Z_iZ_j + \cdots. \tag*{(5)}

To circumvent the exponential cost of evaluating the full projector, we truncate it to retain only the one-body ZiZ_i terms, yielding the state-specific flow generator

G(s)=∑i[Zi,H(s)](6)G(s) = \sum_i [Z_i,H(s)] \tag*{(6)}

In practice we use this generator only at s=0s=0 and with the projector truncated to its one-body terms; no flow is integrated. At this order the Pauli support of G(0)G(0) consists of the strings PZiPZ_i with P∈HP \in H and {P,Zi}=0\{P,Z_i\}=0, i.e. the Hamiltonian terms that couple the reference to other determinants. This screening is related to the direct interaction set used to select generators, as used in [17], but differs in that no energy-gradient ranking is performed and the complete set is appended as a layer. Higher orders (s>0s>0, or two-body projector terms) would expose further excitations at a higher classical cost and with a larger pool. Since the purpose of this step is to supply physically relevant starters for the MCP, the lowest order is sufficient.

Classical VDBF algorithms such as the one proposed by [17] remain in the Heisenberg picture, using G(s)G(s) to sequentially rotate the Hamiltonian via Pauli propagation. Although relying strictly on this one-body approximated flow for exact classical Hamiltonian diagonalization can lead to slow convergence, our method avoids this pitfall. We do not use G(s)G(s) as a rigid evolution operator, but rather as a mathematical tool to identify a targeted Pauli operator pool. The subsequent variational relaxation of the parameters θj,k\theta_{j,k} within our layered quantum ansatz recovers the higher-order correlations that the truncated classical flow would otherwise capture too slowly. We therefore remain in the Schrödinger picture, repurposing G(s)G(s) not as an evolution operator but as a generator to construct a highly expressive variational manifold.

Evaluating the commutator in the equation above yields a linear combination of Pauli operators. These operators are isolated as individual Pauli strings and added to our candidate pool. In classical VDBF implementations, the coefficients of the Pauli rotations are fixed by the numerical integration of the flow equation. In contrast, by injecting these operators into our layered VQE framework, their respective amplitudes θj,k\theta_{j,k} become free variational parameters.

The resulting candidate operators are then symmetry-filtered with respect to the Hartree–Fock reference: a candidate is retained only if, applied to ∣ψ0⟩\lvert\psi_0\rangle, it yields a determinant with the same numbers of α\alpha and β\beta electrons. Some of them serve as starters for a minimal complete pool built with the construction algorithm of [12], which guarantees that the final pool generates the full target Lie algebra, the remaining candidates are added to it, as detailed below. This filter guarantees the correct sector for the first excitation out of the reference only. An individual Pauli string does not commute with the particle-number or spin operators, and acting on other determinants it can change NαN_\alpha and NβN_\beta. The pool therefore does not preserve the total particle number, SzS_z or total spin. The state may thus acquire weight outside the (Nα,Nβ)(N_\alpha,N_\beta) sector of the reference, as well as spin contamination.

Pool structure and size. The pool is the union of two sets. The first is a minimal complete pool (MCP) of 2N−42N-4 generators, constructed with the selected starters following Ref. [12]; it scales linearly with NN and guarantees that the pool generates the target Lie algebra. The second consists of the remaining VDBF excitations added to the MCP, which make the ansatz converge faster; their number is bounded by the number of Hamiltonian Pauli strings and grows as O(N4)O(N^4). Table 1 reports both. The linear scaling therefore refers to the MCP; the layer size is dominated by the added excitations and remains 4 to 25 times smaller than the QEB pool.

MoleculeNNMCP (2N−42N - 4)Added excitationsVULCAN totalQEB (bb)
BeH2_2142473971092 (10)
H6_61220129149561 (10)
H8_816284745021940 (30)
H10_{10}2036125112875035 (50)
H2_2O26481278132615275 (50)
C3_3H6_632601420148036456 (5)

Table 1. Number of qubits NN and pool sizes. The VULCAN pool is the union of a minimal complete pool (MCP) of 2N−42N - 4 generators, which scales linearly with NN and guarantees that the pool generates the target Lie algebra, and the excitations added from the VDBF commutator of Eq. (6), whose number follows the number of Hamiltonian Pauli strings. The number in parentheses after the QEB pool size is the batch size bb used for batched ADAPT-VQE. Pool sizes are independent of the molecular geometries considered in this work.

Grow-then-compress loop

The ansatz is constructed iteratively in a layered fashion, drawing operators from the complete pool generated in Section IV B. Each iteration proceeds through three steps: target acquisition, overlap-guided growth, and greedy re-optimization.

a. Target acquisition. At iteration nn, the algorithm appends the full pool as a new layer to the current ansatz and optimizes the new parameters greedily, with all preceding layers frozen. The resulting state defines the uncompressed target ∣Φtarget⟩|\Phi_{\mathrm{target}}\rangle for the subsequent compression.

b. Overlap-guided compression. The compressed layer is built on top of the frozen base ∣ψbase⟩|\psi_{\mathrm{base}}\rangle, which is the last accepted compressed ansatz, or ∣ψ0⟩|\psi_0\rangle at the first iteration. All states are real, and the generators Ak=iPkA_k=iP_k are real antisymmetric matrices (PkP_k contains an odd number of YY operators), so that exp⁡(iθPk)=exp⁡(θAk)\exp(i\theta P_k)=\exp(\theta A_k) maps real states to real states. We therefore minimize the infidelity

C(θ)=1−⟨Φtarget∣ψ(θ)⟩2,(7)C(\theta)=1-\langle\Phi_{\mathrm{target}}|\psi(\theta)\rangle^2, \tag*{(7)}

where ∣ψ(θ)⟩|\psi(\theta)\rangle is the base followed by the generators inserted so far. New generators are always inserted at the outermost position (acting last) with θ=0\theta=0. For such an insertion,

∂C∂θk∣θk=0=−⟨Φtarget∣ψ⟩gk,gk=2⟨Φtarget∣Ak∣ψ⟩,(8)\left.\frac{\partial C}{\partial\theta_k}\right|_{\theta_k=0}=-\langle\Phi_{\mathrm{target}}|\psi\rangle g_k,\qquad g_k=2\langle\Phi_{\mathrm{target}}|A_k|\psi\rangle, \tag*{(8)}

and since ⟨Φtarget∣ψ⟩\langle\Phi_{\mathrm{target}}|\psi\rangle is common to all candidates, ranking by the overlap gradient ∣gk∣|g_k| is equivalent to ranking by the infidelity gradient.

Each insertion step proceeds as follows. (i) The overlap gradient gkg_k is evaluated for every generator of the full pool; the pool is restored at every step, so a generator may be inserted more than once. (ii) Candidates are tried in order of decreasing ∣gk∣|g_k|: each is placed at the outermost position with θ=0\theta=0, and all parameters of the compressed layer are re-optimized with the L-BFGS-B algorithm to minimize CC. (iii) A candidate is rejected, and the next one tried, if its optimized angle satisfies ∣θ∣<θrej|\theta|<\theta_{\mathrm{rej}}. The pass ends when the infidelity gain of an insertion falls below ΔCmin⁡\Delta C_{\min}; when max⁡k∣gk∣<gnull\max_k|g_k|<g_{\mathrm{null}}, i.e. no generator improves the fidelity to first order; after nrejn_{\mathrm{rej}} consecutive rejections; or when the compressed layer reaches its maximum length max⁡(1,⌊fM⌋)\max(1,\lfloor fM\rfloor), where ff is the compression fraction (typically f=0.5f=0.5).

c. Cost of the overlap gradient. In standard ADAPT-VQE, the energy gradient of a single candidate costs NHN_H measurements. The overlap gradient bypasses the Hamiltonian entirely. Because AkA_k and the states are real, ⟨Φtarget∣Ak∣ψ⟩\langle\Phi_{\mathrm{target}}|A_k|\psi\rangle is real, and since eπ2Ak=Ake^{\frac{\pi}{2}A_k}=A_k for a Pauli generator,

∣gk∣2=4∣⟨Φtarget∣eπ2Ak∣ψ⟩∣2.(9)|g_k|^2=4\left|\langle\Phi_{\mathrm{target}}|e^{\frac{\pi}{2}A_k}|\psi\rangle\right|^2. \tag*{(9)}

The magnitude of each overlap gradient is therefore obtained from a single overlap evaluation, reducing the cost from NHN_H measurements to 1. The sign of gkg_k, which is not needed for ranking, is not recovered. In emulation, all gkg_k are computed exactly.

d. Greedy re-optimization. The surviving compact layer undergoes a final energy optimization. To maintain computational efficiency, only the parameters of the current layer are optimized; all preceding layers remain frozen. The outer loop then returns to target acquisition, using the compressed ansatz as the new baseline, and terminates when the energy change between iterations falls below a convergence threshold ε\varepsilon or a maximum iteration count is reached.

Subspace projection

For systems whose qubit count exceeds the available register, we project the Hamiltonian onto a smaller subspace before the grow-then-compress loop begins.

Subspace construction

Starting from the Hartree–Fock determinant ∣ψ0⟩|\psi_0\rangle, we define the set of all determinants reachable by repeated application of the excitation patterns present in the Hamiltonian’s Pauli strings.

We select kk linearly independent excitation patterns B0,…,Bk−1B_0,\ldots,B_{k-1} and define the projected subspace as the K=2kK=2^k determinants obtained by applying all possible combinations of these patterns to ∣ψ0⟩|\psi_0\rangle:

Sk=span⁡{∣D1⟩,∣D2⟩,…,∣DK⟩},(10)S_k=\operatorname{span}\{|D_1\rangle,|D_2\rangle,\ldots,|D_K\rangle\}, \tag*{(10)}

where each ∣Di⟩|D_i\rangle is a computational-basis determinant obtained by applying a combination of kk linearly independent excitation patterns B0,…,Bk−1B_0,\ldots,B_{k-1} to ∣ψ0⟩|\psi_0\rangle. For example, with k=2k=2 excitation patterns B0B_0 and B1B_1, the subspace contains four determinants: ∣ψ0⟩|\psi_0\rangle itself, B0∣ψ0⟩B_0|\psi_0\rangle, B1∣ψ0⟩B_1|\psi_0\rangle, and B0B1∣ψ0⟩B_0B_1|\psi_0\rangle. Although each pattern is chosen in the sector of the reference, combinations of patterns can produce determinants with different numbers of α\alpha and β\beta electrons. For H4_4 at k=4k=4, for instance, B0B_0 (1α1β→2α2β1\alpha1\beta\to2\alpha2\beta) and B3B_3 (0α0β→2α2β0\alpha0\beta\to2\alpha2\beta) each conserve Nα=Nβ=2N_\alpha=N_\beta=2, but applying both to ∣ψ0⟩|\psi_0\rangle flips orbital 2α2β2\alpha2\beta twice and empties 0α0β1α1β0\alpha0\beta1\alpha1\beta, giving a determinant with no electrons; 4 of the 16 determinants of S4S_4 lie outside the (Nα,Nβ)(N_\alpha,N_\beta) sector. SkS_k therefore contains such determinants, and HprojH_{\mathrm{proj}} is defined on all of them.

The projected Hamiltonian, HprojH_{\mathrm{proj}}, is obtained by computing the matrix elements of HH in the basis {∣D1⟩,…,∣DK⟩}\{|D_1\rangle,\ldots,|D_K\rangle\} and re-expressing the resulting K×KK\times K operator as a Pauli sum on kk qubits. The projected HprojH_{\mathrm{proj}} typically contains far fewer Pauli terms than the original HH. The reduced problem can then be passed to the grow-then-compress loop (Section IV C) or solved by exact diagonalization for validation.

Selection of excitation patterns

The quality of the projected subspace depends on which kk excitation patterns are chosen. We rank candidates using a two-stage procedure: a greedy sweep that determines optimal angles, followed by a beam search that ranks the patterns by importance.

a. Step 1: greedy sweep. A single application of HH to ∣ψ0⟩|\psi_0\rangle yields the coupling ∣⟨ψ0∣HAj∣ψ0⟩∣|\langle\psi_0|HA_j|\psi_0\rangle| for every pool generator Aj=iPjA_j=iP_j. The generators are sorted by decreasing coupling magnitude. This ordering defines the sequence in which the generators appear in the product of exponentials.

Starting from ∣ψ⟩=∣ψ0⟩|\psi\rangle=|\psi_0\rangle, we sweep through the sorted generators one by one, performing a greedy energy minimization inspired by greedy VQE. For each generator AjA_j:

  1. Compute ∣ϕ⟩=Aj∣ψ⟩|\phi\rangle=A_j|\psi\rangle and orthogonalize against the current state: ∣ϕ⊥⟩=(∣ϕ⟩−⟨ψ∣ϕ⟩∣ψ⟩)/∥⋅∥|\phi_\perp\rangle=(|\phi\rangle-\langle\psi|\phi\rangle|\psi\rangle)/\lVert\cdot\rVert. If the perpendicular norm is below 10−810^{-8}, the generator is assigned θj=0\theta_j=0.

  1. Minimize the energy exactly in the two-dimensional space {∣ψ⟩,∣ϕ⊥⟩}\{|\psi\rangle,|\phi_\perp\rangle\}: 2θj=atan2⁡(C,−B)2\theta_j=\operatorname{atan2}(C,-B), where B=(H00−H11)/2B=(H_{00}-H_{11})/2 and C=H01C=H_{01}.

  1. Update ∣ψ⟩←cos⁡θj∣ψ⟩−sin⁡θj∣ϕ⊥⟩|\psi\rangle\leftarrow\cos\theta_j|\psi\rangle-\sin\theta_j|\phi_\perp\rangle so that subsequent generators are evaluated against the improved state.

Each generator requires one application of HH, so the full sweep costs M+1M+1 HH-applications. The coupling determines the order; the greedy sweep determines the angles. No angle is ever revisited.

b. Step 2: beam search and importance ranking. The state has the form of a product ∏j(cos⁡θjI−sin⁡θjAj)∣ψ0⟩\prod_j(\cos\theta_j I-\sin\theta_j A_j)|\psi_0\rangle.

To score each generator’s contribution, we expand this product approximately using a beam search of width 2k2^k (Figure 11). At each generator, every surviving branch splits into two: one where the generator is applied (weight ×∣sin⁡θj∣\times|\sin\theta_j|) and one where it is not (weight ×∣cos⁡θj∣\times|\cos\theta_j|). Only the 2k2^k heaviest branches are kept. Generators with θj=0\theta_j=0 are skipped. Each generator’s importance I(Pj)I(P_j) is the total weight of the surviving branches in which PjP_j was applied. The generators are sorted by decreasing importance, and the first kk whose excitation patterns are linearly independent are selected. Only the excitation pattern of each selected generator is used, the subspace SkS_k is the full span of all 2k2^k combinations of these kk patterns applied to ∣ψ0⟩|\psi_0\rangle.

Beam search for scoring generator importance

Figure 11. Beam search for scoring generator importance, illustrated with M=3M = 3 generators and beam width 2k=42^k = 4 (k=2k = 2). At each generator PjP_j, every surviving branch splits into two: one where PjP_j is applied (orange, weight ×∣sin⁡θj∣\times\lvert\sin\theta_j\rvert) and one where it is not (blue, weight ×∣cos⁡θj∣\times\lvert\cos\theta_j\rvert). Only the 2k2^k heaviest branches survive (solid); the rest are pruned (dashed). Each generator’s importance is the total weight of the surviving branches that include it. The generators are ranked by this score, and the first kk in-sector, linearly independent excitation patterns define the basis of the projected subspace.

The sweep runs once regardless of kk. When the subspace is grown adaptively, only the beam search is repeated at each new kk, since the beam width 2k2^k must match the current subspace size. In Figure 11, we give an example of this branching process for a three-generator example with k=2k=2.

Adaptive-kk growth with PT2 stopping

Rather than fixing kk in advance, we grow the subspace one excitation pattern at a time. At each kk, we evaluate how much energy is missing from the subspace using a second-order Epstein–Nesbet estimate:

ΔEk(2)=∑D∉Sk∣⟨D∣H∣ψref⟩∣2Eref−⟨D∣H∣D⟩.(11)\Delta E_k^{(2)}=\sum_{D\notin S_k}\frac{|\langle D|H|\psi_{\mathrm{ref}}\rangle|^2}{E_{\mathrm{ref}}-\langle D|H|D\rangle}. \tag*{(11)}

The numerators are obtained from a single sparse application of HH to ∣ψref⟩|\psi_{\mathrm{ref}}\rangle within the reachable determinant set. The diagonal energies ⟨D∣H∣D⟩\langle D|H|D\rangle are read from the diagonal terms of HH.

Two choices of reference are available:

  • Exact reference: ∣ψref⟩|\psi_{\mathrm{ref}}\rangle is the ground state of HprojH_{\mathrm{proj}} at the current kk, obtained by dense diagonalization of the 2k×2k2^k\times2^k projected Hamiltonian. This captures the full energy missing from the subspace but requires diagonalization and one HH-application at every kk.

  • Hartree–Fock reference: ∣ψref⟩=∣ψ0⟩|\psi_{\mathrm{ref}}\rangle=|\psi_0\rangle. The couplings and diagonal energies are computed once and reused at every kk, requiring no diagonalization. This variant only sees determinants directly coupled to HF and gives a smaller estimate, but is substantially cheaper.

The growth stops when ∣ΔEk(2)∣|\Delta E_k^{(2)}| falls below a tolerance τ\tau or a maximum kmax⁡k_{\max} is reached.

The exact-reference estimate requires diagonalizing HprojH_{\mathrm{proj}} at every kk; it is used here to validate the stopping criterion. When it is used, the projected problem has in effect been solved classically, and the subsequent VULCAN run on HprojH_{\mathrm{proj}} serves to assess circuit cost. When HprojH_{\mathrm{proj}} is to be solved on a quantum device, the Hartree–Fock-reference estimate is the applicable criterion: it is cheaper, but it can lie above or below the missing energy (Figure 7–Figure 10), so a tighter tolerance τ\tau is advisable.

Solution on the projected register

On the projected register we do not build a new operator pool: we reuse the pool of Section IV B. Each generator of that pool is an excitation that turns one determinant into another. We keep only the generators whose excitation can be written as a combination of the kk selected patterns B0,…,Bk−1B_0,\ldots,B_{k-1}. Such a generator never takes the state outside the subspace SkS_k, so it can be rewritten as a Pauli string on the kk qubits, in the same way as HprojH_{\mathrm{proj}} itself. The other generators would leave SkS_k and are discarded. For H4_4 at k=4k=4, 14 of the 27 generators are kept.

Within the subspace, each kept generator performs exactly the same rotation as on the full register. The ansatz on the kk qubits is therefore the original ansatz restricted to SkS_k. Like the original pool (Section IV B), it does not conserve particle number (or spin): the subspace contains determinants outside of the (Nα,Nβ)(N_\alpha,N_\beta) sector of the reference, and the projected generators can couple them. This projected pool is then used as usual: we either diagonalize HprojH_{\mathrm{proj}} exactly, which gives the best energy available in the subspace, or run VULCAN iterations (Section IV C) on the kk qubits. Both are reported in Figure 7–Figure 10.

Fewer qubits do not automatically mean cheaper gates. For H4_4 at k=4k=4, a double excitation acting on 4 of the 8 spin orbitals becomes a string acting on 3 of the 4 projected qubits; for H2_2O (6-31G) at k=20k = 20, a projected generator costs about 19 CNOTs on average, against about 6 on the full register.

This cost depends on the choice of patterns, not on the subspace itself and therefore can be reduced: any other basis of the same span describes the same determinants and leaves every energy unchanged, but changes how many qubits each generator acts on. We therefore choose the basis that minimizes the CNOT count of the projected pool, and compress the resulting layer with the overlap-guided procedure of Section IV C, ranking candidates by gradient per added CNOT. Details of this method are given in the Supplementary Material.

Classical Post-Processing: ENPT2 and stochastic evaluation

We apply a second-order Epstein–Nesbet perturbation theory (ENPT2) correction classically to the VQE energy, following the formulation of Ryabinkin et al. [18]. This type of correction has been widely used in the closely related family of selected Configuration Interaction (sCI) methods [23, 24], where it refines the energy of a truncated wavefunction and vanishes identically when the full configuration space is included.

Formulation

Let H\mathcal{H} denote the Hilbert space of NN qubits and HH the Hamiltonian acting on it. Let U=∏jeiθjPjU = \prod_j e^{i\theta_jP_j} be the optimized ansatz unitary of Eq. (1), so that ∣Ψ⟩=U∣ψ0⟩\lvert\Psi\rangle= U\lvert\psi_0\rangle is the VQE state, and define the dressed Hamiltonian H~=U†HU\widetilde{H} = U^\dagger H U. The reference energy is the VQE energy,

E0=⟨ψ0∣H~∣ψ0⟩=⟨Ψ∣H∣Ψ⟩.(12)E_0 = \langle\psi_0\lvert\widetilde{H}\rvert\psi_0\rangle= \langle\Psi\lvert H\rvert\Psi\rangle. \tag*{(12)}

Let PX={I,X}⊗N\mathcal{P}_X = \{I, X\}^{\otimes N} be the set of all 2N2^N Pauli strings built from the identity and σx\sigma_x. Each Xj∈PXX_j \in\mathcal{P}_X maps ∣ψ0⟩\lvert\psi_0\rangle onto a determinant ∣Dj⟩=Xj∣ψ0⟩\lvert D_j\rangle= X_j\lvert\psi_0\rangle, whose diagonal energy and coupling are

Ej=⟨ψ0∣XjH~Xj∣ψ0⟩,(13)E_j = \langle\psi_0\lvert X_j\widetilde{H}X_j\rvert\psi_0\rangle, \tag*{(13)}
H0j=⟨ψ0∣H~Xj∣ψ0⟩.(14)H_{0j} = \langle\psi_0\lvert\widetilde{H}X_j\rvert\psi_0\rangle. \tag*{(14)}

The ENPT2 correction is then

ΔEENPT(2)=−∑Xj∈PX∖{I⊗N}∣H0j∣2Ej−E0.(15)\Delta E_{\mathrm{ENPT}}^{(2)} = - \sum_{X_j \in\mathcal{P}_X \setminus\{I^{\otimes N}\}} \frac{\lvert H_{0j}\rvert^2}{E_j - E_0}. \tag*{(15)}

For all systems considered in this work, we verified that the Diagonal Unitary Correction (DUC) formula (Eq. 25 of [18]) gives equivalent results.

Efficient evaluation via the auxiliary vector

The key computational insight is that all 2N2^N coupling coefficients H0jH_{0j} can be extracted simultaneously. Define the auxiliary state vector

∣w⟩=U†H∣Ψ⟩=H~∣ψ0⟩.(16)|w\rangle= U^\dagger H|\Psi\rangle= \widetilde{H}|\psi_0\rangle. \tag*{(16)}

Since ∣Dj⟩|D_j\rangle is a computational basis state, the coupling reduces to

H0j=⟨Dj∣w⟩.(17)H_{0j} = \langle D_j|w\rangle. \tag*{(17)}

A single evaluation of ∣w⟩|w\rangle therefore yields every numerator in Eq. (15). The denominators Ej−E0E_j-E_0 need only be computed for indices jj where ∣H0j∣|H_{0j}| exceeds a threshold ϵ\epsilon, each requiring one application of H~\widetilde{H}. The cost is dominated by the number of such active terms NactiveN_{\mathrm{active}}, which motivates the stochastic approach below when NactiveN_{\mathrm{active}} is large.

Algorithm 1 summarizes the deterministic evaluation.

Algorithm 1Deterministic ENPT2
Require:Hamiltonian HH, ansatz UU, Hartree–Fock state ∣ψ0⟩|\psi_0\rangle, threshold η\eta
Ensure:ENPT2 correction ΔE\Delta E
1:∣Ψ⟩←U∣ψ0⟩|\Psi\rangle\leftarrow U|\psi_0\rangle, E0←⟨Ψ∣H∣Ψ⟩E_0 \leftarrow\langle\Psi|H|\Psi\rangle, E0←⟨ψ0∣H∣ψ0⟩E_0 \leftarrow\langle\psi_0|H|\psi_0\rangle
2:∣w⟩←U†H∣Ψ⟩|w\rangle\leftarrow U^\dagger H|\Psi\rangle ▷ all couplings at once
3:ΔE←0\Delta E \leftarrow0
4:for j=1j=1 to 2N−12^N-1 do
5:if ∣⟨Dj∣w⟩∣2>η|\langle D_j|w\rangle|^2>\eta then
6:Ej←⟨ψ0∣XjH~Xj∣ψ0⟩E_j \leftarrow\langle\psi_0|X_j\widetilde{H}X_j|\psi_0\rangle
7:ΔE←ΔE−∣⟨Dj∣w⟩∣2/(Ej−E0)\Delta E \leftarrow\Delta E-|\langle D_j|w\rangle|^2/(E_j-E_0)
8:end if
9:end for
10:return ΔE\Delta E

Table 1 (PDF p. 18).

Stochastic evaluation

For large systems, the number of active terms grows prohibitively even after filtering out zero couplings. We estimate the sum in Eq. (15) stochastically using importance sampling with the distribution

qj=∣H0j∣2∑i∣H0i∣2(18)q_j=\frac{|H_{0j}|^2}{\sum_i|H_{0i}|^2} \tag*{(18)}

where only nonzero couplings are included. This distribution is near-optimal, since the variance of the estimator is minimized when qjq_j is proportional to the absolute value of the summand (see Supplementary Material for the derivation). Defining

fj=−∣H0j∣2Ej−E0,(19)f_j=-\frac{|H_{0j}|^2}{E_j-E_0}, \tag*{(19)}

the ENPT2 correction is ΔEENPT(2)=∑jfj\Delta E_{\mathrm{ENPT}}^{(2)}=\sum_j f_j, estimated by drawing NsN_s samples from qq and computing

S^Ns=1Ns∑m=1NsfImqIm.(20)\widehat{S}_{N_s}=\frac{1}{N_s}\sum_{m=1}^{N_s}\frac{f_{I_m}}{q_{I_m}}. \tag*{(20)}
Algorithm 2Stochastic ENPT2
Require:Hamiltonian HH, ansatz UU, batch size ΔNs\Delta N_s, target tolerance ϵ\epsilon, ψ0\psi_0 (Hartree–Fock state)
Ensure:Correction ΔE\Delta E, standard error σ\sigma
1:∣Ψ⟩←U∣ψ0⟩|\Psi\rangle\leftarrow U|\psi_0\rangle, E0←⟨Ψ∣H∣Ψ⟩E_0 \leftarrow\langle\Psi|H|\Psi\rangle, E0←⟨ψ0∣H∣ψ0⟩E_0 \leftarrow\langle\psi_0|H|\psi_0\rangle
2:∣w⟩←U†H∣Ψ⟩|w\rangle\leftarrow U^\dagger H|\Psi\rangle ▷ all couplings at once
3:W←∑j>0∣⟨Dj∣w⟩∣2W \leftarrow\sum_{j>0}|\langle D_j|w\rangle|^2, qj←∣⟨Dj∣w⟩∣2/Wq_j \leftarrow|\langle D_j|w\rangle|^2/W
4:S←∅S\leftarrow\emptyset, ΔE←0\Delta E\leftarrow0
5:repeat
6:ΔEprev←ΔE\Delta E_{\mathrm{prev}}\leftarrow\Delta E
7:Draw ΔNs\Delta N_s samples from qq and append to SS
8:Ns←∣S∣N_s\leftarrow|S|
9:for each unique state jj in SS not yet evaluated do
10:Ej←⟨ψ0∣XjH~Xj∣ψ0⟩E_j\leftarrow\langle\psi_0|X_j\widetilde{H}X_j|\psi_0\rangle
11:end for
12:mj←m_j\leftarrow occurrences of jj in SS
13:vj←−W/(Ej−E0)v_j\leftarrow-W/(E_j-E_0) ▷ value per draw of state jj
14:ΔE←1Ns∑jmjvj\Delta E\leftarrow\frac{1}{N_s}\sum_jm_jv_j
15:σ←1Ns(Ns−1)∑jmj(vj−ΔE)2\sigma\leftarrow\sqrt{\frac{1}{N_s(N_s-1)}\sum_jm_j(v_j-\Delta E)^2}
16:until σ≤ϵ\sigma\le\epsilon and ∣ΔE−ΔEprev∣≤ϵ|\Delta E-\Delta E_{\mathrm{prev}}|\le\epsilon
17:return ΔE,σ\Delta E,\sigma

Table 2 (PDF p. 18).

The variance of this estimator provides the error estimate σ\sigma reported in Table IV.

All couplings H0jH_{0j} are obtained from the auxiliary vector ∣w⟩|w\rangle as in the deterministic case. Only the sampled denominators Ej−E0E_j-E_0 need to be evaluated individually, which is what makes the stochastic approach efficient: the expensive per-term computation is restricted to a small random subset.

Algorithm 2 summarizes the procedure.

The loop terminates when two conditions are simultaneously satisfied: the standard error σ\sigma of the estimator falls below the tolerance ϵ\epsilon, ensuring that the statistical uncertainty is controlled, and the change ∣ΔE−ΔEprev∣|\Delta E-\Delta E_{\mathrm{prev}}| between successive rounds also falls below ϵ\epsilon, ensuring that the estimate has stabilized. Neither condition alone is sufficient: σ\sigma can be small while a strongly coupled state has not yet been drawn, causing the estimate to jump when it finally appears, and the change between rounds can be small by accident when successive rounds share most of their draws.

Several practical refinements are used in our implementation. A maximum draw budget Nsmax⁡N_s^{\max} guarantees termination. When EjE_j is close to E0E_0, the ENPT2 term diverges; in such cases we substitute the finite Diagonal Unitary Correction (DUC) formula (Eq. 25 of [18]).

Cost and scaling. Both variants require the auxiliary vector ∣w⟩|w\rangle in full, since the couplings are read from all 2N2^N of its components. Stochastic sampling therefore reduces the number of denominators evaluated, not the memory, which remains that of a dense state vector. Each denominator costs roughly one energy evaluation of the dressed Hamiltonian. For C3H6\mathrm{C}_{3}\mathrm{H}_{6} (32 qubits) on a single NVIDIA H100 GPU, E0E_0 took 26 s, ∣w⟩|w\rangle 10 s, and each denominator 28.5 s; the 1484 unique denominators took 11.8 h. The run stopped at its draw budget (5000 unique states) before reaching the target standard error of 10−610^{-6} Ha; the achieved σ=4.5×10−5\sigma= 4.5 \times10^{-5} Ha is nonetheless ∼35\sim35 times below chemical accuracy. The smallest denominator was 1.30 Ha and none was negative, so the DUC substitution was not triggered. ENPT2 thus applies in the regime where the ansatz can be emulated, systems beyond single-GPU memory require a distributed state vector, which will be presented in a forthcoming work.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.

Code availability

The classical emulation and algorithmic logic code utilized for this study are available from the corresponding author on reasonable request.

Acknowledgments

This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement No 810367), project EMC2 (JPP) and from HQI (JPP). This work was also supported by the French State through the France 2030 program (HYPERION project, i-Demo n°2 managed by Bpifrance). Computations have been performed at IDRIS (Jean Zay) on GENCI Grants: no A0150712052 (J.-P. P.) and no A0200815824 (G.M.). The authors thank Baptiste Claudon for his help on the hardware implementation. Computations on the Quantinuum System Model H2 machines have been made possible through the Quantinuum Startup Partner Program.

Competing interests

Jean-Philip Piquemal is a shareholder and co-founder of Qubit Pharmaceuticals. All remaining authors declare no competing interests.

References

References

  1. [1]J. Lee, D. W. Berry, C. Gidney, W. J. Huggins, J. R. McClean, N. Wiebe, and R. Babbush, Even more efficient quantum computations of chemistry through tensor hyper-contraction, PRX Quantum 2, 030305 (2021).
  2. [2]J. Preskill, Quantum computing in the NISQ era and beyond, Quantum 2, 79 (2018).DOI
  3. [3]A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A variational eigenvalue solver on a photonic quantum processor, Nature Communications 5, 4213 (2014).DOI
  4. [4]M. Haidar, M. J. Rancic, T. Ayral, Y. Maday, and J.-P. Piquemal, Open source variational quantum eigensolver extension of the quantum learning machine for quantum chemistry, Wiley Interdisciplinary Reviews: Computational Molecular Science 13, e1664 (2023).DOI
  5. [5]T. Harville, R. Khurana, V. F. Grizzi, and C. Liu, Recent developments in VQE: Survey and benchmarking, arXiv preprint arXiv:2602.11384 (2026).
  6. [6]H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, An adaptive variational algorithm for exact molecular simulations on a quantum computer, Nature Communications 10, 10.1038/s41467-019-10988-2 (2019).DOI
  7. [7]H. L. Tang, V. Shkolnikov, G. S. Barron, H. R. Grimsley, N. J. Mayhall, E. Barnes, and S. E. Economou, qubit-ADAPT-VQE: An adaptive algorithm for constructing hardware-efficient ansätze on a quantum processor, PRX Quantum 2, 020310 (2021).
  8. [8]Y. S. Yordanov, V. Armaos, C. H. W. Barnes, and D. R. M. Arvidsson-Shukur, Qubit-excitation-based adaptive variational quantum eigensolver, Communications Physics 4, 10.1038/s42005-021-00730-0 (2021).
  9. [9]C. Feniou, M. Hassan, D. Traoré, E. Giner, Y. Maday, and J.-P. Piquemal, Overlap-ADAPT-VQE: practical quantum chemistry on quantum computers via overlap-guided com-DOI
  10. [10]P. G. Anastasiou, N. J. Mayhall, E. Barnes, and S. E. Economou, How to really measure operator gradients in ADAPT-VQE, arXiv preprint arXiv:2306.03227 (2023).arxiv.org/abs/2306.03227
  11. [11]C. Feniou, M. Hassan, B. Claudon, A. Courtat, O. Adjoua, Y. Maday, and J.-P. Piquemal, Greedy gradient-free adaptive variational quantum algorithms on a noisy intermediate scale quantum computer, Scientific Reports 15, 18689 (2025).DOI
  12. [12]Y. Viswanathan, O. Adjoua, C. Feniou, S. Badreddine, and J.-P. Piquemal, An optimized construction of lie algebra generator pools for variational quantum eigensolvers in chemistry, Communications Physics 10.1038/s42005-026-02879-y (2026).DOI
  13. [13]V. O. Shkolnikov, N. J. Mayhall, S. E. Economou, and E. Barnes, Avoiding symmetry roadblocks and minimizing the measurement overhead of adaptive variational quantum eigensolvers, Quantum 7, 1040 (2023).DOI
  14. [14]M. Haidar, O. Adjoua, S. Badreddine, A. Peruzzo, and J.-P. Piquemal, Non-iterative disentangled unitary coupled-cluster based on lie-algebraic structure, Quantum Science and Technology 10, 025031 (2025).DOI
  15. [15]D. A. Fedorov, Y. Alekseev, S. K. Gray, and M. Otten, Unitary selective coupled-cluster method, Quantum 6, 703 (2022).DOI
  16. [16]F. Wegner, Flow-equations for Hamiltonians, Annalen der physik 506, 77 (1994).
  17. [17]C. Shrikhande, A. Bachhar, A. R. Jimenez, and N. J. Mayhall, Rapid ground state energy estimation with a Sparse Pauli Dynamics-enabled Variational Double Bracket Flow, arXiv preprint arXiv:2511.21651 (2025).
  18. [18]I. G. Ryabinkin, A. F. Izmaylov, and S. N. Genin, A posteriori corrections to the iterative qubit coupled cluster method to minimize the use of quantum resources in large-scale calculations, Quantum Science and Technology 6, 024012 (2021).
  19. [19]O. Adjoua, S. Badreddine, C. Feniou, I. Chollet, D. Traore, G. Michel, and J.-P. Piquemal, High performance quantum emulation for chemistry applications with Hyperion, arXiv:2604.01176 (2026).arxiv.org/abs/2604.01176
  20. [20]M. D. Sapova and A. K. Fedorov, Variational quantum eigensolver techniques for simulating carbon monoxide oxidation, Communications Physics 5, 199 (2022).DOI
  21. [21]D. Traore, O. Adjoua, C. Feniou, I.-M. Lygatsika, Y. Maday, E. Posenitskiy, K. Hammernik, A. Peruzzo, J. Toulouse, E. Giner, et al., Shortcut to chemically accurate quantum computing via density-based basis-set correction, Communications Chemistry 7, 269 (2024).DOI
  22. [22]M. Ramão, P. G. Anastasiou, L. P. Santos, N. J. Mayhall, E. Barnes, and S. E. Economou, Reducing the resources required by ADAPT-VQE using coupled exchange operators and improved subroutines, npj Quantum Information 11, 86 (2025).
  23. [23]A. D. Chien, A. A. Holmes, M. Otten, C. J. Umrigar, S. Sharma, and P. M. Zimmerman, Excited states of methylene, polyenes, and ozone from heat-bath configuration interaction, The Journal of Physical Chemistry A 122, 2714 (2018).DOI
  24. [24]Y. Garniron, T. Applencourt, K. Gasperich, A. Benali, A. Ferté, J. Paquier, B. Pradines, R. Assaraf, P. Reinhardt, J. Toulouse, et al., Quantum package 2.0: An open-source determinant-driven suite of programs, Journal of chemical theory and computation 15, 3591 (2019).DOI

Paper details

vlnc:W_rvefefu9
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 4

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents