betaFeedback

WorksPhysics and Astronomy

GENIE: learned quantum circuits as embedding impurity solvers, and why compactness, not correlation strength, decides when they work

Abstract

Sample-based quantum diagonalization has been used as an impurity solver in density matrix embedding theory (DMET), but always with fixed-form circuits and without comparison to classical selection at the same subspace size. We introduce GENIE, a DMET impurity solver in which a generative transformer learns the sampling circuit, trained on the quantum-selected configuration interaction energy. On N2, hydrogen chains, water and a hydrogen ring, we compare GENIE at equal determinant budget with classical selected configuration interaction (CIPSI, ASCI, heat-bath CI) and with DMET combined with sample-based quantum diagonalization (SQD). GENIE selects: for N2 at 200 determinants it is within a factor of 1.7 of CIPSI and 25 to 1000 times more accurate than the determinants a fixed circuit actually samples, while standard SQD reaches good energies mainly by filling the subspace. GENIE does not beat the best classical selection, and its circuits require fault-tolerant hardware. Its failures are governed by the compactness of the wavefunction relative to the budget, not by correlation strength. A cheap estimate from coupled-cluster amplitudes predicts this compactness. A router built on it never recommended GENIE where it then failed, apart from seed-to-seed variance, and it transferred unchanged to a blind test on new systems. We also identify two benchmarking pitfalls that can create apparent quantum advantages of one to two orders of magnitude.

Introduction

Sample-based quantum diagonalization uses a quantum computer only to propose determinants: a circuit is prepared and measured, the bitstrings span a subspace, and the Hamiltonian is diagonalized classically in that subspace. In its original form, quantum-selected configuration interaction (QSCI) [1], and in its large-scale form with configuration recovery, sample-based quantum diagonalization (SQD) [2], the energy is variational and no expectation value has to be estimated on the device. SQD experiments on up to 77 qubits [2] and its Krylov variants [3, 4] have made the approach the main route to chemistry on present hardware.

To treat molecules larger than a device, SQD has been used as the impurity or fragment solver of quantum embedding: density matrix embedding theory (DMET) [5, 6, 7] for hydrogen rings and cyclohexane conformers [8] and for ligand-like molecules [9], the localized active space (LAS) method [10, 11] for multi-metal complexes [12], and DFT embedding [13]. All of these sample fixed-form circuits, in practice the local unitary cluster Jastrow (LUCJ) ansatz [14], and report the error against the exact (FCI) energy of the embedded problem. The only learned component so far is a classical restricted Boltzmann machine that post-processes the samples [15].

An energy close to the embedded FCI does not reveal why it is good. It can arise from selection, the circuit sampling the important determinants, or from filling, the subspace becoming a large fraction of the space. Classical selected configuration interaction (SCI), such as CIPSI [16], adaptive sampling CI (ASCI) [17] and heat-bath CI (HCI) [18], selects deterministically and is the natural yardstick. Critical assessments found QSCI expansions less compact than classical SCI and sampling inefficient for peaked states [19], and a recent review places most claimed QSCI advantages within the classically simulable regime [20]; strong classical SCI [21] raises the bar further, while advantages for sparse ground states have been shown for constructed Hamiltonians [22].

A different line of work makes the circuit itself learnable. The generative quantum eigensolver (GQE) [23] trains a transformer to emit sequences of operators from a pool. Trained on the QSCI energy, it produces subspaces up to 50 % smaller than HCI for N2_2 in active spaces of up to 32 qubits with 83–98 % fewer two-qubit gates than LUCJ [24], and has been scaled to 40 qubits [25]. These studies work in a complete active space, without embedding, and compare against HCI only.

Here we combine the two lines and ask a sharper question than “is the energy good”: does the circuit select, how does it compare with the best classical heuristics at the same number of determinants, and can one tell beforehand whether it will work? Our contributions are: (i) GENIE, a DMET impurity solver whose sampling circuit is learned by GQE, with the embedding Hamiltonian expressed in the embedded mean-field basis so that operator pool, reference determinant and Hamiltonian are consistent (§ 1); (ii) an equal-budget protocol that compares every GENIE run at its final subspace size NN with an oracle, CIPSI, ASCI and HCI in two orbital bases, and with DMET+SQD restricted to the determinants actually sampled (Secs. III A–III C); (iii) the finding that failures are governed by the compactness of the impurity wavefunction relative to the budget rather than by correlation strength (Sec. III E); and (iv) an amplitude-based estimate and router that predict these failures in advance, validated in a blind test on new systems (Secs. III F and III G). Section III H discusses the scope of an impurity-level study and the LAS alternative, and Sec. III I resources and hardware feasibility; the appendices contain every data table behind the figures.

Methods

Density matrix embedding

The whole molecule is treated at the unrestricted Hartree–Fock (UHF) level. The impurity orbitals are the active orbitals chosen with atomic valence active space (AVAS) projection [26] onto atom-resolved labels (e.g. the 1s functions of the central atoms of a hydrogen chain). A Schmidt decomposition of the reference one-particle density matrix gives a bath of at most the same size; impurity plus bath span the embedding space of nn orbitals, in which the interacting Hamiltonian is

Hemb=Ecore+∑pqσhpqapσ†aqσ+12∑pqrsστ(pq∣rs)apσ†arτ†asτaqσ,(1)H_{\mathrm{emb}} = E_{\mathrm{core}} + \sum_{pq\sigma} h_{pq} a^\dagger_{p\sigma}a_{q\sigma} + \frac{1}{2}\sum_{pqrs\sigma\tau} (pq|rs) a^\dagger_{p\sigma}a^\dagger_{r\tau}a_{s\tau}a_{q\sigma}, \tag*{(1)}

with the frozen environment folded into EcoreE_{\mathrm{core}} and hpqh_{pq} [5, 6, 7]. Unlike the textbook mean-field bath, the Schmidt decomposition acts on a correlated reference density, a CASCI in the AVAS space, and the bath size is chosen adaptively from the singular-value spectrum; the scheme is one-shot, without a correlation potential (App. A). An embedded RHF is solved in the embedding space with EcoreE_{\mathrm{core}} as constant.

Orbital basis

Equation (1) is naturally expressed in the raw (Schmidt) basis. We rotate it into the embedded RHF molecular-orbital (MO) basis, h→CThCh \to C^{\mathsf{T}}hC and likewise for (pq∣rs)(pq|rs). Energies are invariant, but this step is essential for a learned circuit: the circuit starts from “the lowest orbitals occupied” and its excitation operators are indexed by occupied and virtual MOs. In the raw basis the starting determinant lies 20 000–40 000 kcal mol−1^{-1} above RHF and GENIE errors were 40–95 kcal mol−1^{-1}; after rotation they fall by up to five orders of magnitude (App. B, Table V). A complete-active-space GQE pipeline is unaffected, because its basis already is the MO basis, which is why the problem is easy to miss when such a pipeline is moved into an embedding.

RR (Å)raw basis (3 seeds)MO basis (2 seeds)
1.146±0.846 \pm0.80.04–0.05
1.882±0.382 \pm0.30.005–0.009
2.195±0.0495 \pm0.040.0013–0.0015
2.578±0.678 \pm0.60.0003

Table V. N₂/STO-3G, GENIE error (kcal mol−1^{-1}) before and after the basis rotation (CCSD angles, 80 operators, no cap).

Quantum-selected configuration interaction

A state ∣ψ⟩|\psi\rangle prepared on 2n2n qubits (Jordan–Wigner) is measured NsN_s times. After discarding bitstrings with the wrong electron numbers and completing the set under α↔β\alpha\leftrightarrow\beta exchange (spin completion), the distinct determinants form S\mathcal{S}, ∣S∣≤N|\mathcal{S}| \leq N, where NN is the budget (cap). The QSCI energy is the lowest eigenvalue of the projected Hamiltonian,

ES=min⁡eig⁡(⟨DI|Hemb|DJ⟩I,J∈S),(2)E_{\mathcal{S}} = \min\operatorname{eig}\left(\left\langle D_I\middle|H_{\mathrm{emb}}\middle|D_J\right\rangle_{I,J\in\mathcal{S}}\right), \tag*{(2)}

which is variational, ES≥EFCIembE_{\mathcal{S}} \geq E_{\mathrm{FCI}}^{\mathrm{emb}}. SQD [2] instead diagonalizes in the Cartesian product of all sampled α\alpha strings with all sampled β\beta strings (“product” subspace) and recovers configurations broken by noise; we also evaluate the “list” subspace of the sampled (α,β)(\alpha,\beta) pairs only, which isolates what the circuit selects.

GENIE: generative circuits for the impurity

Circuit generation. The operator pool consists of fermionic single and double excitation generators κj=τj−τj†\kappa_j = \tau_j - \tau_j^\dagger of the embedded RHF reference, each with a fixed angle θj\theta_j, taken either from the CCSD amplitudes of the embedded RHF (“CCSD angles”) or from the set {±1,±0.4,±0.1}\{\pm1,\pm0.4,\pm0.1\} (“free angles”). A decoder-only transformer generates a sequence j=(j1,…,jng)\mathbf{j}=(j_1,\ldots,j_{n_g}) autoregressively, and the circuit is

U(j)=∏k=ng1exp⁡(θjkκjk),∣ψ(j)⟩=U(j)∣ΦHF⟩.(3)U(\mathbf{j}) = \prod_{k=n_g}^{1}\exp(\theta_{j_k}\kappa_{j_k}), \qquad|\psi(\mathbf{j})\rangle= U(\mathbf{j})|\Phi_{\mathrm{HF}}\rangle. \tag*{(3)}

Reward and policy update. Each generated circuit is sampled, the QSCI energy ES(j)E_{\mathcal{S}}(\mathbf{j}) of Eq. (2) is computed classically, and the transformer is updated with a group-relative policy-gradient (GRPO) objective, as in the GQE-for-QSCI trainer [24], rather than the logit-matching loss of the original GQE [23]. Token probabilities are πθ(j)=softmax⁡(−βzj)\pi_\theta(\mathbf{j})=\operatorname{softmax}(-\beta\mathbf{z}_{j}) for transformer outputs zj\mathbf{z}_j and an inverse temperature β\beta. For a batch of BB sequences with energies EbE_b, the advantages are Ab=(Eˉ−Eb)/σEA_b=(\bar{E}-E_b)/\sigma_E (batch mean and standard deviation), and with probability ratios rbk=πθ(jbk)/πθold(jbk)r_{bk}=\pi_\theta(j_{bk})/\pi_{\theta_{\mathrm{old}}}(j_{bk}) the loss is

L=−1ng∑klog⁡πθ(jb∗k)−min⁡(⟨rbkAb⟩,⟨clip⁡(rbk,1−ϵ1,1+ϵh)Ab⟩),(4)\mathcal{L} = -\frac{1}{n_g}\sum_k \log\pi_\theta(j_{b^*k}) -\min\left(\left\langle r_{bk}A_b\right\rangle,\left\langle\operatorname{clip}(r_{bk},1-\epsilon_1,1+\epsilon_h)A_b\right\rangle\right), \tag*{(4)}

where b∗b^* is the lowest-energy sequence of the batch, ⟨⋅⟩\langle\cdot\rangle averages over sequences and positions, and ϵ1=0.2\epsilon_1=0.2,

Figure 1. The GENIE pipeline. (a) Classical embedding: an AVAS impurity and its Schmidt bath define the embedding Hamiltonian HembH_{\mathrm{emb}} of nn orbitals (2n2n qubits); the embedded RHF provides CCSD amplitudes (operator angles and compactness estimate) and the orbitals into which HembH_{\mathrm{emb}} is rotated, so that operator pool, reference determinant and Hamiltonian share one basis. (b) Learned circuit: a transformer emits a sequence of ngn_g excitation operators; the circuit is sampled, the determinants are filtered and spin-completed, and the QSCI energy in at most NN determinants is the signal that updates the transformer. After training, the best determinant set is refined, and its energy is compared with the exact solution of the embedding. (c) Before training, the router compares the estimated compactness nestn_{\mathrm{est}} with the budget NN; after training, every result is compared with classical selection and with DMET+SQD at the same NN on the same Hamiltonian.

ϵh=0.28\epsilon_h = 0.28. The first term reinforces the best circuit found so far in the batch; the second is a clipped (PPO-style) surrogate. β\beta starts at 0.5 and is raised in steps of 0.02 by a scheduler that targets an energy variance of 10−510^{-5} within a batch, so that sampling sharpens as the batch converges. Optimizer: AdamW (learning rate 5×10−65 \times10^{-6}, weight decay 0.01). A replay buffer keeps the best sequences.

Refinement. As in [24], the QSCI solutions of the circuits of one epoch are merged into one sparse wavefunction (local refinement), which is then merged with the result of all previous epochs (global refinement). Each merge solves the generalized eigenvalue problem in the non-orthogonal basis of the input wavefunctions and keeps the NN determinants with the largest weights. Refinement is purely classical and generates no new determinants: the final set S∗S^* consists of determinants sampled by some circuit during training. It is therefore the product of all circuits executed in training, not of a single circuit, for H14H_{14} at 2.0 Å the best single circuit samples 1400–3400 distinct determinants, the refined set holds 10410^4, and the quantum cost of a GENIE point is that of the whole training run (Sec. III).

Unless stated, ng=80n_g = 80, 120 epochs, 100 circuits per epoch (12 000 circuits per run, NS=105N_S = 10^5 shots each, i.e. 1.2×1091.2 \times10^9 shots), a QSCI cap of N=104N = 10^4 (or budgets of 100 and 200), and three seeds (App. J, Table XVI); runs that stop before 120 epochs are rejected by the pipeline.

settingvalue
operator poolfermionic excitations (DMET pool), JW
anglesCCSD amplitudes, or {±1,±0.4,±0.1}\{\pm1,\pm0.4,\pm0.1\}
operators per circuit ngn_g80 (160, 320 depth test)
epochs120
samples / batch / buffer / warm-up100 / 100 / 100 / 100
lossGRPO, clip [0.8, 1.28], + NLL of best sequence
inverse temperature β\beta0.5, +0.02 per step, target variance 10−510^{-5}
optimizerAdamW, lr 5×10−65 \times10^{-6}, weight decay 0.01
QSCI cap NN10410^4 (100, 200 budget runs)
orbital basisembedded MO
seeds4, 5, 6

Table XVI. GENIE settings.

All simulations are noiseless state-vector simulations with CUDA-Q [27]. Because QSCI energies are variational, the best of kk seeds is a valid result at kk-fold cost.

Equal-budget protocol and baselines

For every finished GENIE run with final subspace size NN we evaluate, on the same HembH_{\mathrm{emb}} and at the same NN, the variational energies of (a) an oracle, the NN determinants with largest ∣cI∣|c_I| in the exact embedded ground state, re-diagonalized (a reference, not a method); (b) CIPSI [16], grown by Epstein–Nesbet second-order contributions until exactly NN determinants; (c) ASCI [17], fixed-size and seeded with CIPSI, which can drop determinants; and (d) a fixed-size HCI [18] (App. C). Selection is basis dependent, so each selector is run in both the raw and the MO basis and the better value is the reference. Errors are quoted against the embedding FCI.

For DMET+SQD we use the same embedding and an LUCJ circuit [14] with all-pairs interactions and two repetitions, CCSD-initialized and optimized with the linear method of ffsim [28], 10410^{4} or 10510^{5} shots, and six configuration-recovery iterations [2] (Qiskit addon [29]). Noise is modeled on the FakeSherbrooke backend (App. E).

Compactness and router

We characterize the exact impurity state ∑IcI∣DI⟩\sum_I c_I \lvert D_I \rangle by its leading weight w1=max⁡I∣cI∣2w_1 = \max_I |c_I|^2 and by its compactness n99n_{99}, the smallest number of determinants, ordered by ∣cI∣|c_I|, whose weights sum to 0.99. Since n99n_{99} requires the exact state, we estimate it from CCSD amplitudes (MP2 amplitudes if CCSD does not converge). With determinant-level single and double coefficients cSc_S and cDc_D (from t1t_1 and τ2=t2+t1t1\tau_2 = t_2 + t_1t_1; App. § 1.3.6) and disconnected quadruples approximated by products,

nest=1+#{∣cS∣>τ}+#{∣cD∣>τ}+#{∣cDcD′∣>τ},n_{\mathrm{est}} = 1 + \#\{|c_S| > \tau\} + \#\{|c_D| > \tau\} + \#\{|c_Dc_{D'}| > \tau\},

with τ=0.01\tau= 0.01. The router recommends GENIE if 3nest≤N3n_{\mathrm{est}} \le N; GENIE with a deeper circuit or a larger budget if nest≤N<3nestn_{\mathrm{est}} \le N < 3n_{\mathrm{est}}; and a fragment method (LAS) otherwise. The factor three is the observed accuracy of nestn_{\mathrm{est}} (Sec. III F).

Numerical results

We study N2_2/STO-3G at R=1.1R = 1.1–2.52.5 Å, where the AVAS N 2s2p2s2p impurity plus bath spans the full valence space (9–10 orbitals, 18–20 qubits), so that the embedding error vanishes and the solver is isolated; N2_2/cc-pVDZ (13 of 28 orbitals, 26 qubits, 2.9×1062.9 \times10^6 determinants); linear H10_{10} and H14_{14}/STO-3G with central-atom impurities at R=1.0R = 1.0, 1.5 and 2.0 Å (up to 20 qubits,

63 504 determinants); and [Fe(H2_2O)4_4]2_2bpm4+^{4+} for the LAS route [11, 12]. All systems and embeddings are listed in Table IV. Software: PySCF [30], CUDA-Q [27], Qiskit [29], ffsim [28] and our package QuEnAIS [31].

systembasisRR (Å)impurity / fragmentsembedding (orbitals; Nα,NβN_{\alpha},N_{\beta})qubits; determinants
N₂STO-3G1.1–2.5AVAS N 2s2p (10e,8o)full valence space, 9–10 orbitals18–20; ≤7 056\leq7\,056
N₂cc-pVDZ1.1, 2.1AVAS N 2s2p8 impurity + 5 bath = 13 of 28; (7,7)26; 2 944 656
H₁₀STO-3G1.0, 1.5, 2.01s of atoms 4, 5≤6\leq6 of 10≤12\leq12; small
H₁₄STO-3G1.01s of atoms 5–89 of 14; (4,4)18; 15 876
H₁₄STO-3G1.5, 2.01s of atoms 5–810 of 14; (5,5)20; 63 504
H₂O (blind)6-31G0.96, 1.44, 1.92AVAS O 2s2p + H 1s9 orbitals; (5,5)18; 15 876
H₁₂ ring (blind)STO-3G1.0, 1.5, 2.01s of six adjacent atoms9–10 orbitals18–20; ≤\leq63 504
H₁₀, H₁₄ (LAS)STO-3G1.0–2.0H₂ units (2o, 1α1β) or two halves (5o/7o)two Fe 3d: (5o; 4α2β), (5o; 2α4β)4 or 10/14 per fragment
[Fe(H₂O)₄]₂bpym⁴⁺6-31Gequilibriumtwo Fe 3d: (5o; 4α2β), (5o; 2α4β)AVAS (12e,10o)10 per fragment

Table IV. Systems, embeddings and problem sizes.

Selection at equal budget

Figure 2 shows the central comparison. At 200 determinants (about 3% of the space) GENIE is within a factor of 1.7 of CIPSI at every geometry [Fig. 2(b)], including 2.5 Å where the leading determinant carries only 10% of the weight. ASCI is 1–5 times better than both and coincides with the oracle: GENIE reaches CIPSI but not ASCI quality. One 200-determinant seed in three fails (17 kcal mol−1^{-1} at 2.5 Å). At 100 determinants GENIE fails at 1.8–2.5 Å [Fig. 2(a)], with errors of 17–50 kcal mol−1^{-1} against 0.5–2 kcal mol−1^{-1} for CIPSI; only one seed at 2.1 Å succeeds. Without a cap, GENIE keeps 600–1100 determinants and reaches 10−210^{-2}–10−410^{-4} kcal mol−1^{-1}; with free angles it reaches the embedding FCI using 25–51% of the space.

Figure 2. N2_2/STO-3G, error against the embedding FCI at fixed budgets of (a) 100 and (b) 200 determinants: GENIE (CCSD angles, 80 operators, three seeds), CIPSI and ASCI (MO basis), and the oracle (dashed). Dotted line: 1 kcal mol−1^{-1}. Full data in Table VI.

Two benchmarking pitfalls, and seed variance

Three effects can make a quantum selection scheme look much better or worse than it is (Figure 3). First, the basis of the Hamiltonian: in the raw Schmidt basis the learned circuit starts from a determinant tens of thousands of kcal mol−1^{-1} above RHF and fails by 46–95 kcal mol−1^{-1}; in the MO basis the same pipeline reaches 10−210^{-2}–10−410^{-4} kcal mol−1^{-1} (Figure 3(a)). Second, the basis of the baseline: CIPSI selected in the raw basis is 2–36 times worse than in the MO basis, HCI and ASCI behave alike (Figure 3(b)). Against raw-basis CIPSI, GENIE appeared 6–38 times better, an advantage that disappears under a fair baseline. Our fixed-size HCI is moreover 3–65 times worse than CIPSI on the same problems (Table VI). This fixed-size variant is weaker than production semistochastic HCI, which we did not run; the point is that HCI’s quality depends on its implementation and size control, so advantages measured against HCI alone [24, 32] should be read with care.

Figure 3. (a) GENIE error for N2_2/STO-3G with the embedding Hamiltonian in the raw Schmidt basis and in the embedded MO basis (no cap, CCSD angles, 80 operators; Table V). (b) Classical selectors at N=200N = 200, selected in the raw basis (open) and in the MO basis (filled); same Hamiltonian, energies are basis invariant, only the selection changes. (c) Five GENIE seeds in a plain CAS(10e,8o) N2_2 problem at 2.1 Å, N=200N = 200, against CIPSI and the oracle.

RRNNGENIE (seeds)HCI rawHCI MOCIPSI rawCIPSI MOASCI rawASCI MOoracle MO
1.810022.76 / 21.51 / 21.4917.898.3810.372.017.811.761.92
2.110018.36 / 0.76 / 16.7413.796.107.460.666.140.530.57
2.510021.14 / 49.93 / 24.989.891.561.040.541.010.100.11
1.82000.334 / 0.431 / 0.26512.951.562.500.2752.290.2450.253
2.12000.099 / 0.092 / 0.07810.955.242.950.0812.050.0570.058
2.52000.065 / 17.31 / 0.0359.641.320.630.0390.360.0120.015
1.11 109 / 1 0650.041 / 0.0460.0110.00020.0020.00010.0020.00010.0001
1.8964 / 1 0350.009 / 0.0050.0510.00020.0030.00010.0020.00010.0001
2.1660 / 6720.0013 / 0.00150.0800.00000.0030.00000.0030.00000.0000
2.5586 / 6010.0003 / 0.00030.175−0.0001ᵃ0.002−0.0001ᵃ0.001−0.0001ᵃ−0.0001ᵃ

Table VI. N2_2/STO-3G, all selectors at equal size NN (kcal mol−1^{-1} vs. embedding FCI). “raw”/“MO”: selection basis. GENIE: CCSD angles, 80 operators, 120 epochs, seeds 4/5/6 (4/5 for the uncapped runs). Bold: best classical.

Third, seed variance is large [Fig. 3(c)]: five seeds span a factor of 300; the best run beats CIPSI by 10×, the median is 3× worse and one in five fails (App. C). A single favorable run is therefore not evidence of an advantage; we report all seeds, and since QSCI energies are variational, the best of kk seeds is a legitimate result at kk-fold cost.

Selection versus filling

Restricted to the determinants it samples, DMET+SQD is 1.4–100 times less accurate than CIPSI at equal size (Table VIII), and more shots do not help at 2.5 Å: 37.4 and 36.4 kcal mol−1^{-1} for 139 and 309 determinants. The optimized fixed-form circuit does not reach the important determinants of a stretched bond. With fewer determinants GENIE is 25–50 times (1.8 Å) and 500–1000 times (2.5 Å) more accurate (Figure 4), i.e. it behaves like CIPSI rather than like the sampler.

Error versus subspace size for SQD, CIPSI, and GENIE

Figure 4. N2_2/STO-3G at (a) 1.8 and (b) 2.5 Å: error versus subspace size for the determinants actually sampled by DMET+SQD (“list” subspace, optimized LUCJ, 10410^4 and 10510^5 shots), GENIE (budgets 100 and 200) and CIPSI. Full data in Table VIII.

RRshotsproduct SQDlist SQDlist dim (sampled)CIPSI @ dimlist/CIPSI⟨S2⟩\langle S^2\rangle
1.110410^41.2–3.46.3 (5.6–6.6)81 (62)4.51.40.01
1.110510^50.27–0.531.6 (1.5–1.8)248 (175)0.483.40.001
1.810410^414.7–27.513.6 (12.7–14.1)197 (154)2.651.2
1.810510^52.5–9.910.9 (9.9–11.6)361 (294)0.56190.5–1.1
2.510410^40.5–15537.4 (37.1–37.6)139 (107)0.80472.4
2.510510^50.22–0.2936.4 (35.9–37.0)309 (221)0.37992.4–2.6

Table VIII. N2_2/STO-3G, DMET+SQD with optimized LUCJ (10410^4/10510^5 shots, seeds 1–3; kcal mol−1^{-1} vs. embedding FCI). “CIPSI @ dim”: CIPSI (raw basis) at the size of the sampled list; ⟨S2⟩\langle S^2\rangle of the list state.

The good energies usually reported for SQD come from the product subspace. In small embeddings the α×β\alpha\times\beta product of the sampled strings covers most of the space; adding random exploration shots made every tested system exact, including N2_2/cc-pVDZ, by covering 99.9–100% of spaces of up to 2.9×1062.9 \times10^6 determinants (App. D). This is exact diagonalization, not compression. Per seed, the product energy scatters over up to two orders of magnitude while the list energy is stable and far from CIPSI (Figure 5(a)). Noise makes this explicit (App. E, Table X). Under a FakeSherbrooke noise model (N2_2, 1.1 Å; routed LUCJ with 3 499 CX gates, estimated circuit fidelity 0.05 at one tenth of present-day error rates and 10−1310^{-13} at present-day rates), the product subspace becomes the full space and its energy exact. The sampled list improves with noise: from 5.5–6.7 kcal mol−1^{-1} with 87–95 determinants without noise to 0.05–0.20 kcal mol−1^{-1} with 5 600 determinants (39% of the space) at one tenth of present-day rates, and 0.2–2.0 kcal mol−1^{-1} with 7 600 (53%) at present-day rates. Better energies from a noisier device are the signature of filling: noise spreads the samples over the space, and the classical diagonalization does the rest.

Figure 5. (a) N2_2/STO-3G, DMET+SQD with optimized LUCJ: range over three seeds of the product-subspace energy (bars), the sampled-list energy (diamonds, mean and range) and CIPSI at the size of the list (squares). (b) N2_2/cc-pVDZ (26 qubits): error versus subspace fraction for DMET+SQD (unoptimized LUCJ, 10410^{4}/ 10510^{5} shots), CIPSI (raw basis) at the same sizes, GENIE (budget 200 and uncapped), and the best of CIPSI and ASCI in the MO basis at GENIE’s sizes; filled 1.1 Å, open 2.1 Å.

ssfidelityvalid shotsproductlist
01100 %676–729 (5 %): 1.7–2.987–95 (0.6 %): 5.5–6.7
0.10.05213 %14 400 (100 %): 0.0005 570–5 639 (39 %): 0.05–0.20
11.4×10−131.4 \times10^{-13}1.5 %14 400 (100 %): 0.0007 550–7 630 (53 %): 0.21–2.0

Table X. N2_2/STO-3G, 1.1 Å, DMET+SQD under noise (all-pairs LUCJ, 2 repetitions, 10410^4 shots, seeds 1–3; subspace size, fraction of the 14 400-determinant space, error in kcal mol−1^{-1}). Valid shots: fraction of bitstrings with the correct electron numbers.

On N2_2/cc-pVDZ (26 qubits, 2.9×1062.9 \times10^6 and 8.5×1058.5 \times10^5 determinants at 1.1 and 2.1 Å), GENIE is again far ahead of DMET+SQD at similar size, 6.2 against 29 kcal mol−1^{-1} and 22.5 against 94 kcal mol−1^{-1} at about 200 determinants, and without cap reaches 0.48 and 0.62 kcal mol−1^{-1} with 0.3 and 1.2% of the space Figure 5(b), Table IX.

RRmethoderrorsubspace
1.1SQD 10410^4 / 10510^5 shots28.9 / 19.70.005 / 0.01 %
1.1CIPSI (raw) at those sizes5.4 / 1.5
1.1GENIE, budget 200 / no cap6.25 / 0.480.007 / 0.34 %
2.1SQD 10410^4 / 10510^5 shots94.1 / 88.50.02 / 0.03 %
2.1CIPSI (raw) at those sizes11.2 / 7.9
2.1GENIE, budget 200 / no cap22.5 / 0.620.023 / 1.2 %
1.1CIPSI / ASCI (MO) at GENIE sizes6.16, 0.021 / 6.14, 0.021
2.1CIPSI / ASCI (MO) at GENIE sizes17.2, 0.053 / 11.7, 0.053

Table IX. N2_2/cc-pVDZ (26 qubits). DMET+SQD with an unoptimized LUCJ circuit (all pairs, 2 reps; optimization would need ∼74\sim74 GB), seed 1, and GENIE (seed 4). kcal mol−1^{-1} vs. embedding FCI; subspace as fraction of the space (2.94×1062.94 \times10^6 at 1.1 Å, 8.5×1058.5 \times10^5 at 2.1 Å, where the bath is smaller).

RR (Å)w1w_1n99n_{99}NNngn_gGENIECIPSIASCI
1.00.89112200804.604.504.50
1.50.464 76910410^4803.0–3.22.552.50
1.50.464 76910410^41603.032.552.50
2.00.1413 37610410^48068–707.005.95
2.00.1413 37610410^41607.927.005.95
2.00.1413 37610410^43208.097.005.95

Table 1. H14_{14} (DMET, 18–20 qubits): GENIE against classical selection at equal size NN (kcal mol−1^{-1} vs. embedding FCI, MO basis). w1w_1, n99n_{99}: leading weight and number of determinants holding 99 % of the exact embedded state (MO basis); ngn_g: operators per circuit. Full data in Table XIII.

RRNNngn_ganglesseedGENIECIPSIASCIoracle
1.06 53880CCSD40.0190.00040.00030.0003
1.06 27380CCSD50.0250.00070.00060.0006
1.07 95680free4, 50.0000.0000.0000.000
1.020080CCSD44.604.504.504.53
1.510 00080CCSD4 / 53.10 / 3.192.552.502.65
1.510 00080free4 / 53.04 / 3.092.552.502.65
1.510 000160CCSD43.032.552.502.65
1.520080CCSD460.552.251.860.7
2.010 00080CCSD4 / 570.1 / 67.97.005.958.76
2.010 00080free4 / 570.2 / 70.27.005.958.76
2.010 000160CCSD47.927.005.958.76
2.010 000320CCSD48.097.005.958.76
2.020080CCSD4136.044.544.4113.1

Table XIII. H14_{14}, all GENIE runs and classical selectors at equal size (kcal mol−1^{-1} vs. embedding FCI; MO basis; HCI skipped, App. C). All runs 120 epochs. Runs interrupted before 120 epochs are rejected by the pipeline (two such runs, 4 and 32 epochs, were rerun).

Against classical selection in the MO basis at the same sizes it is behind, however: tied with CIPSI and ASCI at 200 determinants and 1.1 Å (6.25 vs. 6.14), 1.9 times behind at 2.1 Å (22.5 vs. 11.7), and 23 and 12 times behind uncapped (0.48 vs. 0.021 and 0.62 vs. 0.053 kcal mol−1^{-1}). In the larger space, the 80-operator circuit reaches chemical accuracy but not the compact expansion that CIPSI finds.

A genuine embedding: hydrogen chains and circuit depth

For H14_{14}, DMET truncates the system to 9–10 of 14 orbitals and the embedding is genuine (Table I, Figure 6). Near equilibrium GENIE ties with CIPSI at 200 determinants, and with N≈7 000N \approx7\,000 free-angle circuits reach the embedding FCI. At 1.5 Å it is 1.2 times behind CIPSI at 10410^4 determinants. At 2.0 Å and 80 operators it fails: 68–70 kcal mol−1^{-1} against 6–7 kcal mol−1^{-1} for CIPSI and ASCI at the same 10410^4 determinants, identically for both seeds and both angle schemes. At 1.5 and 2.0 Å ev

Figure 6. H14_{14}/STO-3G, DMET with a four-atom impurity. (a) Error at N≤104N \leq10^4 versus bond length: GENIE (80 operators, CCSD and free angles, two seeds each), CIPSI, ASCI and oracle at the same NN. (b) 2.0 Å: GENIE error versus circuit length ngn_g; all runs 120 epochs.

ery 80-operator run fills the cap of 10410^4 determinants (15.7 % of the space), so the limit is not the number of determinants but which ones the circuit reaches. Doubling the circuit to 160 operators recovers CIPSI quality [Figure 6(b)]: 7.92 kcal mol−1^{-1} against 7.00 for CIPSI; 320 operators add nothing (8.09 kcal mol−1^{-1}). At 1.5 Å, where the 80-operator circuit already reaches the important determinants, 160 operators change nothing (3.03 against 3.04–3.19 kcal mol−1^{-1}): depth helps only where reach is the limit.

The training curves (Figure 7) show how. With 80 operators the best energy stops improving after a few epochs. With 160 and 320 operators it drops in steps, as the transformer finds sequences that reach new groups of determinants, and settles after 50–70 epochs. Refinement matters: the raw circuit’s subspace gives 14.6 kcal mol−1^{-1} at 160 operators, the refined set 7.9 kcal mol−1^{-1}. For N₂ at 2.5 Å the same curves show the seed failures of Sec. III B: two of three seeds at N=200N = 200 jump to below 0.1 kcal mol−1^{-1} within 25–45 epochs, the third never does [Fig. 7(b)].

Figure 7. Training curves: best error so far against epoch. (a) H₁₄ at 2.0 Å, N=104N = 10^{4}, with 80 (two seeds), 160 and 320 operators; solid after refinement, dashed for the raw circuit’s subspace; horizontal lines: CIPSI and ASCI at the same NN. (b) N₂ at 2.5 Å, three seeds each at N=100N = 100 (light) and N=200N = 200 (dark), with CIPSI at both sizes.

H₁₀ embeddings are small (≤12\le12 qubits) and every method is exact (Table XI).

systemRRw1estw_1^{\mathrm{est}}w1w_1 (exact)CCSD conv.nestn_{\mathrm{est}} (MP2)nestn_{\mathrm{est}} (CCSD)nestn_{\mathrm{est}} (used)n99n_{99} (MO)n99n_{99} (Schmidt)IPR (Schmidt)
H10_{10}1.00.9560.956yes202424681.09
H10_{10}1.50.094b^{b}0.840yes26363614101.40
H10_{10}2.00.6310.628yes36363617132.27
H14_{14}1.00.8990.894yes80138138112871.25
H14_{14}1.50.4600.458yes2401 1951 1954 7692 9814.57
H14_{14}2.00.1890.144yes31914 80014 80013 3767 64837.4
N2_21.10.7830.917yes44595925331.19
N2_21.80.2460.426no632416349905.11
N2_22.10.3780.192no4622046429117.3
N2_22.50.2900.099no5421154436422.2

Table XI. Diagnostics for every embedding with a GENIE run. b^{b}CCSD amplitude blow-up; the weight estimate is unreliable here, one reason it is not used for routing.

DMET+SQD on the same H₁₄ embeddings (10410^{4} shots, three seeds; Table XIV) repeats the N₂ picture at larger scale. Its sampled lists of 240–4 100 determinants are 5.4–6.1, 78.4–78.6 and 122–128 kcal mol−1^{-1} off at 1.0, 1.5 and 2.0 Å; its product subspaces are 2.3–2.8, 75–94 and 82–109 kcal mol−1^{-1} off although they hold 18–25 %, 10–12 % and 66–73 % of the space. GENIE with 10410^{4} determinants (15.7 %) reaches 3.0–3.2 kcal mol−1^{-1} at 1.5 Å and, with 160 operators, 7.9 kcal mol−1^{-1} at 2.0 Å.

RRsubspacesizefractionerror
1.0list238–2671.5–1.7%5.4–6.1
1.0product2 809–3 96918–25%2.3–2.8
1.5list569–5890.9%78.4–78.6
1.5product6 241–7 56910–12%75–94
2.0list4 018–4 1276.3–6.5%122–128
2.0product42 025–46 22566–73%82–109

Table XIV. H14_{14}, DMET+SQD on the same embeddings (optimized LUCJ, 10410^4 shots, seeds 1–3): subspace size, fraction of the space and error (kcal mol−1^{-1} vs. embedding FCI).

Compactness, not correlation strength

The leading weight does not order these outcomes [Fig. 8(a)]: N₂ at 2.5 Å (w1=0.10w_{1} = 0.10) is handled, H₁₄ at 2.0 Å (w1=0.14w_{1} = 0.14) is not. What differs is how many determinants carry the state [Fig. 9(a)]: 42–49 for stretched N₂, where three bonds break between two atoms, but 13 376 for H₁₄ at 2.0 Å, where every embedded bond breaks (Table XI; all counts in the MO basis in which GENIE works). Against n99/Nn_{99}/N the picture is ordered [Fig. 8(b)]. Below n99/N≈1/3n_{99}/N \approx1/3, GENIE stays within a small factor of the best classical selector, apart from failed seeds. Between 1/3 and 1, the outcome depends on the circuit: N₂ at N=100N = 100 (n99/N=0.42n_{99}/N = 0.42–0.49) fails, while H₁₄ at 1.0 Å (N=200N = 200, 0.56) and 1.5 Å (N=104N = 10^{4}, 0.48) succeed. Above 1, every selected-CI method fails in absolute terms, because the budget is smaller than the state; there, H₁₄ at 2.0 Å needs 160 operators to keep up with CIPSI, and so does the H₁₂ ring at 2.0 Å (Sec. III G).

The determinant sets themselves show what goes wrong [Fig. 9(b,c)]. Where GENIE matches CIPSI in energy, it also captures the same exact weight, typically missing 10−4×10−410^{-4} \times10^{-4} of it. The failures are failures of reach: at N=100N = 100 the N₂ sets miss 9–32 % of the weight against 0.1–4 % for CIPSI, and for H₁₄ at 2.0 Å the 80-operator circuits miss 48 % (CCSD angles) and 85–88 % (free angles) against 4.7 % for CIPSI. Resolved by excitation level, the 80-operator set misses 13 % of the weight in doubles and 12 % in quadruples of the leading determinant, which an 80-operator product of excitations cannot generate in sufficient number. With 160 operators the missed weight drops to 3.9 %, below CIPSI’s 4.7 %; the remaining loss sits in odd excitation levels, which CIPSI covers better [Fig. 9(c)].

Compactness is basis dependent. In the raw Schmidt basis, which is localized, n99n_{99} of H₁₄ is 1.3–1.8 times smaller (87, 2 981 and 7 648 at 1.0, 1.5 and 2.0 Å), whereas for N₂ it is 1.3–2.2 times larger (64–91; Table XI). A stretched hydrogen chain is closer to a product of local bonds, which localized orbitals express compactly; this points to localized or natural-orbital pools for spread states. Since n99n_{99} of a spread state grows combinatorially with the number of strongly correlated orbitals, this last regime belongs to fragment methods such as LAS [10, 12], which divide the state into weakly coupled pieces, each compact.

Predicting it

Figure 10 compares nestn_{\mathrm{est}} with the exact n99n_{99}. Converged CCSD amplitudes track n99n_{99} over more than three orders of magnitude, from 6 (H₁₀, 1.0 Å) to 13 376 (H₁₄, 2.0 Å): within a factor of three for the ten embeddings used to build the router (and 0.4–3.3 for the six blind-test embeddings, Sec. III G), with H₁₀ at 1.0 Å overestimated four-fold (24 vs. 6) and H14_{14} at 1.5 Å underestimated four-fold (1 195 vs. 4 769). MP2 underestimates spread states by up to a factor of 40 (H14_{14} at 2.0 Å: 319), while unconverged CCSD overestimates stretched N2_2 four- to five-fold; using CCSD when it converges and MP2 otherwise gives the best of both. An earlier estimate based on fractionally occupied natural orbitals, which counts which orbitals are correlated, failed by more than an order of magnitude in both directions (App. § 1.3.6): what matters is how many determinants carry weight.

Figure 10. Compactness estimate nestn_{\mathrm{est}}, Eq. (5), from MP2 and CCSD amplitudes against the exact n99n_{99} (MO basis); triangles: blind-test systems. Open symbols: CCSD did not converge (stretched N2_2), and the router uses MP2. Shaded band: within a factor of three.

What the circuit selects

Figure 9. What the circuit selects. (a) Weight of the exact embedded state outside its kk largest determinants (MO basis); the crossing with 1% defines n99n_{99}. Vertical lines: budgets used. (b) Exact weight missed by the GENIE determinant set against that missed by CIPSI at the same NN, for every N2_2 and H14_{14} run (H10_{10} and runs with both below 10−610^{-6} omitted). (c) H14_{14} at 2.0 Å, N=104N = 10^4: exact weight per excitation level relative to the leading determinant (bars: weight missed; dashes: total weight at that level) for CIPSI and GENIE with 80 and 160 operators.

the cautious side: “deeper” for H14_{14} at 1.0 Å with budget 200, and “not GENIE” for the 36-determinant H10_{10} embeddings at 1.5 and 2.0 Å with N=20N = 20, where nest overestimates n99n_{99} and GENIE is exact. The diagnostics cost one CCSD or MP2 calculation of the embedding, negligible next to training.

Replayed on all runs, the router of § 1 E never recommends GENIE where it then failed, apart from seed failures (Table XII): GENIE where it matched CIPSI, “deeper” for N2_2 at budget 100, and “not GENIE” for H14_{14} at 2.0 Å with 80 operators, where 160 operators then caught up with CIPSI. Its three deviations are all on

casenestn_{\mathrm{est}}NNrouteroutcome
N2_2 1.1591 100GENIEboth < 0.05 kcal mol−1^{-1}
N2_2 1.8–2.546–63200GENIE1.1–1.8c^{c}
N2_2 1.8–2.546–63100deeper1.4–490
H14_{14} 1.0138∼7 000GENIE≤ 0.025 kcal mol−1^{-1}; free angles exact
H14_{14} 1.0138200deeper1.02
H14_{14} 1.51 19510410^4GENIE1.2
H14_{14} 1.51 195200not GENIEall fail
H14_{14} 2.014 80010410^4not GENIE12 (1.3 at 160 ops)
H14_{14} 2.014 800200not GENIEall fail
H10_{10} 1.024113GENIEexact
H10_{10} 1.5, 2.03620not GENIEexactd^{d}

Table XII. Router recommendation versus outcome (GENIE/best classical at the same NN; 80 operators unless stated). c^{c}One failed seed at 2.5 Å (ratio 1 486); ratios on errors below 0.1 kcal mol−1^{-1} (2.1, 2.5 Å) reach 1.4–5.6. d^{d}36-determinant spaces; nest=36n_{\mathrm{est}} = 36 overestimates n99=14n_{99} = 14–17.

Blind test on new systems

To test whether the compactness criterion and the router generalize, we chose two systems after the router had been fixed and ran them unchanged (Table II): symmetric bond breaking in water, a polyatomic heteroatom molecule, and a ring of twelve hydrogen atoms. The ring is a 4n4n ring with a near-degenerate frontier pair, so its leading weight is only 0.19 already at 1.0 Å; CCSD nonetheless converges, and its estimate w1=0.92w_1 = 0.92 is far off. The diagnostics were computed before the classical baselines.

systemRRw1w_1nestn_{\mathrm{est}}n99n_{99}NNrouterGENIE (seeds)best classicalratio
H2_2O0.960.975320200GENIE0.320.281.2
H2_2O0.960.9753201 864–1 884GENIE0.004 / 0.0050.000135–52
H2_2O1.440.878129200deeper0.710.581.2
H2_2O1.440.8781292 433–2 438GENIE0.003 / 0.0030.0000470
H2_2O1.920.6211134200deeper1.000.861.2
H2_2O1.920.62111342 352–2 355GENIE0.004 / 0.0020.0000821–49
H12_{12} ring1.00.19124205200deeper29.329.21.0
H12_{12} ring1.00.1912420510⁴GENIE0.27 / 0.310.00551–59
H12_{12} ring1.50.118542 197200not GENIE44.141.71.1
H12_{12} ring1.50.118542 19710⁴GENIE1.12 / 1.130.323.4–3.5
160 ops1.510⁴0.590.321.8
H12_{12} ring2.00.1112 53113 258200not GENIE11544.12.6
H12_{12} ring2.00.1112 53113 25810⁴not GENIE71.5 / 9.35.114 / 1.8
160 ops2.010⁴8.855.11.7

Table II. Blind test of the router on systems it was not tuned on: H2_2O/6-31G (AVAS O 2s2p + H 1s, 18 qubits, symmetric stretch, ROH_{\mathrm{OH}} in Å) and a ring of twelve hydrogen atoms (STO-3G, six-atom impurity, 18–20 qubits). Predictions use the rule of § 1 E unchanged and depend only on the embedding. Errors in kcal mol−1^{-1} against the embedding FCI; best classical: CIPSI or ASCI in the MO basis at the same NN.

The estimate transfers: nestn_{\mathrm{est}} is within a factor of 0.4–3.3 of the exact n99n_{99} for all six new embeddings (Figure 10, triangles). So does the classification. Water is compact at every bond length (n99=20n_{99} = 20–34): GENIE is within 1.2 of CIPSI at 200 determinants, the router’s “deeper” at 1.44 and 1.92 Å was again cautious—and below 0.01 kcal mol−1^{-1} without cap. The ring at 2.0 Å (n99=13 258n_{99} = 13\,258) is spread like H14_{14} at 2.0 Å (13 376) and behaves like it: the router says “not GENIE”, one seed fails (14× behind ASCI, the set missing 45% of the exact weight) and the other reaches 1.8×, while every selector is 5–8 kcal mol−1^{-1} off. At 1.5 Å (n99/N=0.22n_{99}/N = 0.22) the router says GENIE, which ends 1.1 kcal mol−1^{-1} from the exact energy but 3.5 times behind ASCI, the largest gap in the compact regime apart from seed failures. As for H14_{14}, a 160-operator circuit closes most of these gaps: with the failing seed at 2.0 Å the error drops from 71.5 to 8.85 kcal mol−1^{-1} (1.7× ASCI), and at 1.5 Å from 1.12 to 0.59 kcal mol−1^{-1} (1.8×). In the ring at 1.0 Å, where the leading weight is only 0.19, the state is nevertheless compact (n99=205n_{99} = 205) and GENIE reaches 0.3 kcal mol−1^{-1}; with N=200<n99N = 200 < n_{99} every selector fails alike. Over the twelve blind cases the router never recommended GENIE where it then failed by more than a factor of 3.5 at the given budget, apart from the large-NN precision floor discussed next; its other misses are on the cautious side.

The large-NN rows show a second, distinct shortfall. Where the budget far exceeds n99n_{99}, CIPSI converges to 10−510^{-5}–10−210^{-2} kcal mol−1^{-1} while GENIE stops at 10−310^{-3}–0.3 kcal mol−1^{-1}, 20–70 times higher (and 40–700 times for N2_2 and H14_{14} at 1.0 Å; Figure 8, open symbols). These errors are far below chemical accuracy, but they show that the learned circuit does not converge as fast as classical selection once the important determinants are found: the remaining small-weight determinants are reached by chance rather than by selection.

Figure 8. GENIE error relative to the best classical selector at the same size against (a) the exact leading weight w1w_1 and (b) the compactness n99n_{99} relative to the budget NN in the MO basis (80 operators; stars and arrows: the same runs with 160 operators, H14_{14} at 2.0 Å and the H12_{12} ring at 1.5 and 2.0 Å). Filled: budgeted runs of N2_2 and H14_{14} and of the blind test (H2_2O, H12_{12} ring, Sec. III G). Open grey: large-NN runs, where GENIE is below 0.35 kcal mol−1^{-1} and CIPSI nearly exact. Correlation strength does not order the failures; compactness does.

Scope and the LAS alternative

This is an impurity-level study: every error above is measured against the exact solution of the same embedding Hamiltonian, which isolates the solver. The total DMET energy of a single impurity is not a whole-molecule energy; for H10_{10} it lies 77–386 kcal mol−1^{-1} above whole-molecule FCI, increasingly so on stretching, independently of the solver (App. § 1.3.8). A whole-molecule comᵃIncluding the one-time LUCJ optimization (25–50 min).

GENIEDMET+SQDCIPSI / ASCI
hardwareA100 + 24 threads10–24 threads1 node
time, N₂ (18–20 qb)700–5100 s1600–2000 sᵃ≤ 1–12 s
time, H₁₄ (20 qb)0.3–3.9 h0.8–0.9 hᵃ≤ 17–44 s
time, N₂/cc-pVDZ (24–26 qb)0.8–2.3 h0.03–1.3 h≤ 0.1–0.6 hᵇ
circuits executed12 00010
shots1.2 × 10⁹10⁴–10⁵0
CX per circuit4.4–46 k3 499ᶜ–
classical diagonalizations∼12 000 + refinement6∼ log₂ N

Table 3. Resources per geometry. Times are wall-clock for one run (one seed); CX counts are for the best circuit, before routing for GENIE and after heavy-hex routing for LUCJ.

ᵇIncluding exact diagonalization in 2.9 × 10⁶ determinants.

ᶜAll-pairs LUCJ, two repetitions, after heavy-hex routing (N₂, 20 qubits).

parison requires all fragments and an energy partitioning, which we leave to future work, as do the embedding errors that impurity-level benchmarks cannot see [8, 9].

For spread states the router points to fragment methods. On H₁₀, where whole-molecule FCI is available, LASSQD agrees with LASSCF to within 1 kcal mol⁻¹, but the error against FCI depends on the fragmentation: 38.5, 41.0 and 19.5 kcal mol⁻¹ with H₂-unit fragments at 1.0, 1.5 and 2.0 Å, and 84.5, 32.5 and 9.5 kcal mol⁻¹ with two halves (App. H). For [Fe(H₂O)₄]²⁺bpym₄⁺, where a DMET impurity would need about 40 qubits, LAS needs two ten-qubit fragments; our LASSCF lies 1.05 mEhE_{\mathrm{h}} below the published value [12], and LASSQD warm-started from it reproduces it to 0.3 µEhE_{\mathrm{h}} for all four circuit/seed combinations (App. I).

Resources and hardware feasibility

Figure 11 and Table III compare the resources. At equal subspace size GENIE is the most expensive method on every count. It takes 200–600 times longer than CIPSI in wall-clock time for N₂ and H₁₄ (4–20 times for cc-pVDZ, where the recorded CIPSI time is dominated by the exact reference) while not being more accurate, executes 10⁴ times more circuits than SQD, and its circuits are 5–40 times deeper. Against SQD it buys accuracy: for H₁₄ at 1.5 Å SQD is 75–94 kcal mol⁻¹ off after 0.8 h, GENIE 3.0–3.2 kcal mol⁻¹ after 1.2–2.9 h. Against classical selection it buys nothing in these systems; the comparison with CIPSI is lost by two to three orders of magnitude in time.

Figure 11. Error against wall-clock time per geometry, every run shown. GENIE: one training run on an NVIDIA A100 with 24 CPU threads. DMET+SQD: one run on 10–24 CPU threads plus the one-time LUCJ optimization of that geometry (1 555–3 010 s). CIPSI: the recorded time also includes the exact diagonalization and the oracle, so it is an upper bound. At 2.5 Å (panel a), SQD’s product subspace with 10410^4 shots covers the whole space (error 0, drawn at the axis).

The circuits are not executable on present hardware. With an effective error pp per CX the circuit fidelity is roughly e−pnCXe^{-p n_{\mathrm{CX}}}. At the error rates of our noise model (p≈8.5×10−3p \approx8.5 \times10^{-3}, which reproduces the estimated fidelity 1.4×10−131.4 \times10^{-13} of the routed 3 499-CX LUCJ circuit), the 80-operator N₂ circuits (7400 CX) have a fidelity of 10−2710^{-27} and the H₁₄ circuits below 10−3910^{-39}. Under such noise the measured bitstrings are close to uniform, and, as shown for SQD (App. E), the subspace is then filled rather than selected; no learned structure survives. The fidelity reaches 0.35–0.5 for the 80-operator circuits only at p=10−4p = 10^{-4}, and 0.6–0.9 for all circuits at p=10−5p = 10^{-5}, i.e. with error-corrected logical gates. Training would also have to run on the device, adding 1.2×1091.2 \times10^{9} shots per geometry. GENIE in its present form therefore targets early fault-tolerant hardware. Shorter circuits are the obvious lever: gate-penalized rewards and gate-efficient pools reduced two-qubit counts by 83–98% in Ref. [24].

Conclusion and discussion

Within DMET, GENIE is, to our knowledge, the first quantum impurity solver whose sampling circuit is learned. Measured at equal subspace size against classical selection rather than against the embedded FCI, it closes the gap of one to three orders of magnitude between fixed-form sampling and classical selection. It reaches CIPSI quality at small budgets but not ASCI quality, and in a 26-qubit embedding stays an order of magnitude behind CIPSI at its own uncapped size, at two to three orders of magnitude more wall-clock time and with circuits of 10⁴ CX gates that require fault-tolerant hardware. Standard DMET+SQD reaches good energies by filling the space, which noise makes complete; where filling is not possible, as for H₁₄, it fails by 75–130 kcal mol⁻¹. Both observations are consistent with critical assessments of QSCI [19, 20] and with the strength of modern classical SCI [21].

The compactness criterion held in a blind test on two new systems, where it predicted the failure for the stretched hydrogen ring and success for water. It is not specific to GENIE. It applies to any quantum selection scheme, including Krylov variants [3, 4] and DMET with learned classical post-processing [15]: a sampler can pay off only where the state is compact in some representation and yet hard to find classically, a regime none of our systems reaches, and the one in which claimed advantages have so far been shown [22]. For spread states, localized orbitals, which stretched H₁₄ is 1.3–1.8 times more compact, paired or generalized excitation pools and multireference initial states are the natural next steps; on noisy hardware, configuration recovery [2] becomes relevant to GENIE as well.

Methodologically we recommend that quantum subspace methods report the subspace size and compare against CIPSI and ASCI at that size, in the orbital basis in which those methods perform best; that seed statistics be reported; and that embedding studies separate the solver error, as here, from the embedding error against whole-molecule references. A cheap compactness estimate should accompany any claim of where a method is expected to work.

Code availability

The QuEnAIs package (DMET, LAS, GENIE interface, classical baselines, router) and all bench mark grids used here are archived as version 0.3.0 at doi:10.5281/zenodo.23736693 [31] (tag v0.3.0 of the GitHub repository); docs/REPRODUCE_GENIE_PAPER.md maps every result to its benchmark grid. Benchmark results, exported analysis data, embedding Hamiltonians and the figure sources are archived at doi:10.5281/zenodo.23737678 [33]. Third-party components (the GQE-for-QSCI trainer, PyCI, block2) are not redistributed and are installed from their upstream sources. Reproduction details are given in App. J.

Acknowledgments

This work was carried out at the Fraunhofer Institute for Industrial Mathematics ITWM in the QuEnAIS project. We thank the ITWM department of High Performance Computing for computing resources and support. Calculations were performed on the ITWM computing cluster (NVIDIA A100 GPU nodes and CPU nodes).

References

Systems and DMET implementation

The impurity is selected with AVAS [26] on atom-indexed labels. The impurity orbitals and their Schmidt bath, obtained from the reference one-particle density matrix, span the embedding space. In detail: (i) Mean field. The whole molecule is described by UHF. For N2_2, H10_{10}, H14_{14}, H2_2O and the H12_{12} ring at 2.0 Å it coincides with RHF to <10−5< 10^{-5} kcal mol−1^{-1}. For the H12_{12} ring at 1.0 and 1.5 Å it breaks spin symmetry and lies 53.5 and 5.2 kcal mol−1^{-1} below RHF, as expected for a 4n4n ring with a near-degenerate frontier pair. Only the environment and EcoreE_\mathrm{core} depend on this choice; all solver errors are measured against the exact solution of the same embedding Hamiltonian, and the embedded reference for the circuits is the closed-shell embedded RHF. (ii) Impurity. The impurity orbitals are the AVAS active orbitals (molecular orbitals of mixed atomic character, not localized atomic orbitals). (iii) Reference density. Instead of the mean-field density, the Schmidt decomposition acts on a CASCI density in the AVAS space, with the active one-electron Hamiltonian screened by the mean-field potential of the non-active electrons, combined with the mean-field density outside the active space. This places the bath where the correlated density couples impurity and environment. (iv) Bath. The impurity–environment block of the reference density (Löwdin basis) is decomposed by singular value decomposition; singular values below 10−810^{-8} are discarded, so a bath of zero orbitals is allowed. The number of bath orbitals is the larger of the position of the largest gap in the singular values and the number needed to cover 99.9% of ∑σ2\sum\sigma^2, capped at the number of impurity orbitals and at 18 embedding orbitals in total. (v) Electron count. The number of embedded electrons is the rounded trace of the reference density in the embedding space, not the active-space count. (vi) Environment. The environment density is the reference density projected onto the complement of the embedding space; it enters hpqh_{pq} through Coulomb and exchange terms. EcoreE_\mathrm{core} is fixed so that the projected UHF density reproduces the whole-molecule UHF energy within the embedding Hamiltonian. (vii) Chemical potential. A one-shot grand-canonical shift h→h−μ1h \to h-\mu\mathbb{1}, Ecore→Ecore+μNE_\mathrm{core} \to E_\mathrm{core}+\mu N is applied; it cancels exactly for the fixed-particle-number solvers used here. (viii) One-shot. No correlation potential is fitted and no self-consistency loop is run; the energy reported is that of the embedding Hamiltonian (Sec. III H). Table IV lists every system, its embedding and the resulting qubit count (Jordan–Wigner, two qubits per spatial orbital).

Orbital-basis consistency

The GENIE operator pool is built from CCSD amplitudes of the embedded RHF and therefore refers to the embedded MO basis, and the circuit starts from the lowest orbitals occupied. In the raw Schmidt basis that determinant is not Hartree–Fock: for N2_2 the overlap between its occupied space and the RHF occupied space can be as small as 10−310^{-3}, and its energy lies 20 000–40 000 kcal mol−1^{-1} above HF. The circuit then starts in the wrong state, and the pool’s excitation indices address the wrong orbitals. After rotation, energies are unchanged; determinant indices and weights (w1w_1, oracle ordering) refer to the MO basis. Table V shows the effect.

Classical baselines and full N2_2 data

CIPSI. Starting from the lowest-diagonal determinant, the set is grown by adding the candidates with the largest Epstein–Nesbet contributions ∣⟨Da∣H∣Ψ⟩∣2/(E−Haa)|\langle D_a|H|\Psi\rangle|^2/(E-H_{aa}), doubling the size per iteration and re-diagonalizing, until exactly NN determinants are reached.

ASCI. Seeded with the CIPSI set; each iteration re-ranks members by ∣c∣|c| and candidates by their first-order amplitude and keeps the best NN, so determinants can leave the set. ASCI can beat the oracle because the largest-∣c∣|c| set is not energy-optimal.

Fixed-size HCI. Heat-bath CI adds a determinant aa if max⁡i∣Haici∣>ϵ\max_i |H_{ai}c_i|>\epsilon. To compare at fixed size, log⁡10ϵ\log_{10}\epsilon is bisected (30 steps between −8-8 and 1) for the largest converged HCI set with at most NN determinants, which is topped up to exactly NN by heat-bath score in one step and re-diagonalized. The one-step top-up and the absence of an energy denominator make this variant weaker than CIPSI at small NN; it is reported for completeness and is not the reference baseline; production semistochastic HCI with an energy-based size control would be closer to CIPSI. For spaces above ∼105\sim10^5 determinants HCI is skipped.

Training variance. In a plain CAS(10e,80) N2_2 problem at 2.1 Å (w1=0.19w_1=0.19, budget 200, 120 epochs), five seeds gave 47.37, 5.28, 5.11, 5.27 and 0.155 mEhE_\mathrm{h} (CIPSI 1.64, oracle 0.110 mEhE_\mathrm{h}). An earlier “GENIE beats CIPSI” claim based on a single favorable run was withdrawn on this basis.

DMET with SQD

Circuit. Heavy-hex pair patterns lose most of the correlation even after optimization (Table VII); we therefore use all-pairs LUCJ with two repetitions, CCSD-initialized and refined with ten linear-method iterations. One batch containing all samples and six configuration-recovery iterations are used.

LUCJ variantoff-HF weightE−EFCIE - E_{\mathrm{FCI}}
heavy-hex, 1 rep, not optimized0.000429.5
all pairs, 4 reps, not optimized0.02670.004
any variant, optimized0.02620.000

Table VII. H4_4 (one two-orbital fragment, first LAS cycle; exact off-HF weight 0.0262): LUCJ variants. Error of the circuit energy (mEhE_{\mathrm{h}}).

Exploration shots. Replacing 25–50% of the shots by samples of a random shallow circuit makes every tested system exact (H8_8, N2_2/STO-3G and N2_2/cc-pVDZ), with subspaces of 99.9–100% of the space and 30 min to 3 h of classical diagonalization per run.

Noise model

Samples are drawn with Qiskit Aer [29] from a noise model built on the FakeSherbrooke backend: circuits are routed onto the heavy-hex coupling map, depolarizing errors with the device-average one- and two-qubit rates multiplied by a scale s∈{0,0.1,1}s \in\{0,0.1,1\} are attached to every gate, and readout errors are included. The estimated circuit fidelity is ∏g(1−spg)\prod_g(1-s p_g); s=0s=0 keeps the routing without errors and must reproduce the noiseless result.

Compactness diagnostics and router

HF weight from CCSD. With closed-shell amplitudes in intermediate normalization and τijab=tijab+tjiab\tau_{ij}^{ab}=t_{ij}^{ab}+t_{ji}^{ab}, the spin-adapted norms of the singly and doubly excited parts are

∥C1∥2=2∑ia(tia)2,(5)\lVert C_1\rVert^2=2\sum_{ia}(t_i^a)^2, \tag*{(5)}
∥C2∥2=∑ijabτijab(2τijab−τijba),(6)\lVert C_2\rVert^2=\sum_{ijab}\tau_{ij}^{ab}(2\tau_{ij}^{ab}-\tau_{ij}^{ba}), \tag*{(6)}

and w1est=1/(1+∥C1∥2+∥C2∥2)w_1^{\mathrm{est}}=1/(1+\lVert C_1\rVert^2+\lVert C_2\rVert^2). For two electrons in two orbitals this gives 1/(1+t2)1/(1+t^2), the exact intermediate-normalization weight.

Determinant-level coefficients. Each tiat_i^a represents an α\alpha and a β\beta single; each τijab\tau_{ij}^{ab} an αβ\alpha\beta double; for i<ji<j, a<ba<b the antisymmetrized τijab−τijba\tau_{ij}^{ab}-\tau_{ij}^{ba} represents an αα\alpha\alpha and a ββ\beta\beta double. Overlapping pairs in the quadruple term of Eq. (5) are not excluded, so this term is an upper estimate. The count is evaluated in O(nlog⁡n)O(n\log n) by sorting ∣cD∣|c_D| and counting partners above τ/∣cD∣\tau/|c_D| by binary search, and is capped at the size of the space.

H14: full data

Table XIII lists every GENIE run on H₁₄ together with the classical selectors at the same size, and Table XIV the DMET+SQD runs on the same embeddings.

Whole-molecule energies and LAS

Single-impurity DMET energy. For H₁₀ (impurity: 1s of atoms 4 and 5) the total DMET energy with an exact impurity solver lies 77, 198 and 386 kcal mol−1^{-1} above whole-molecule FCI at 1.0, 1.5 and 2.0 Å: one impurity in a mean-field environment does not describe a stretched chain. This error is the same for every solver, so it does not affect the impurity-level comparisons of the main text, but it rules out a whole-molecule ranking of DMET against LAS with a single impurity.

LAS on hydrogen chains. LASSCF/LASSQD [10, 12] with AVAS on all H 1s and two fragmentations: H₂ units (2 orbitals, 1α1β each) and two halves (5 or 7 orbitals, opposite-MSM_S doublets); LASSQD with optimized LUCJ, 10410^4 shots, seeds 1–3. Table XV lists errors against whole-molecule FCI for H₁₀ and total energies for H₁₄ (no FCI reference). LASSQD reproduces LASSCF to better than 10−310^{-3} kcal mol−1^{-1} with H₂-unit fragments (4 qubits) in every case. With halves it does so in most runs, but lies 1.7–2.7 kcal mol−1^{-1} above LASSCF for the seven-orbital (14-qubit) fragments of H₁₄ at 1.0 Å, and one seed each lies 0.9 (H₁₀, 1.0 Å) and 0.34 kcal mol−1^{-1} (H₁₄, 1.5 Å) above. For H₁₄ halves at 2.0 Å, LASSQD warm-started from the converged LASSCF ends 8.0, 32 and 41 kcal mol−1^{-1} above it for the three seeds: with seven strongly correlated orᵃNumerical noise of the reference (<2×10−7Eh< 2 \times10^{-7} E_{\mathrm{h}}). Free-angle GENIE runs reach the embedding FCI with 25–51 % of the space at every geometry.

bitals per fragment, the SQD fragment solver no longer reproduces the exact fragment solution.

Halves are better than H2_2 units for stretched chains and worse near equilibrium; the fragmentation decides the accuracy.

N2_2 with atomic fragments. One fragment per N (4 orbitals, 5 electrons). Quartet fragments (opposite MSM_S) lie 1 320 kcal mol−1^{-1} above FCI at short range and 6.9 at 2.5 Å; doublet fragments 34–270. LASSQD (optimized LUCJ, 104^{4} shots) reproduces LASSCF to ≲0.1\lesssim0.1 kcal mol−1^{-1} except where seeds land on different LASSCF stationary points (1.6, 2.0, 2.25 Å). Non-parallelity errors (kcal mol−1^{-1}): CASSCF 0.09, CASCI 0.21, DMET+SQD 29.7, LASSCF (doublet) 237, CCSD(T) 313, HF 435.

Determinant carryover (ScF/STO-3G, (8e,9o), fragments Sc(6o) + F(3o)). LASSQD error vs. LASSCF (kcal mol−1^{-1}, five seeds) with/without carryover [12]: 103^{3} shots 5.5±10.8 / 10.4±9.2; 3 × 103^{3}: 0.37±0.34 / 5.0±4.6; 104^{4}: 0.14 ± 0.32 / 2.3 ± 3.3; 3 × 104^{4}: 0.000 / 0.60 ± 1.33. The threshold (10−4^{-4}–10−6^{-6}) does not matter.

[Fe(H2_2O)4_4]2_2bpym4+^{4+}

6-31G, charge +4, singlet coupling of two high-spin Fe(II); AVAS Fe 3d (12e,10o); fragments Fe0_0 (5o; 4α2β) and Fe1_1 (5o; 2α4β); geometry from Ref. [11]. Reference LASSCF −3655.496377 EhE_{\mathrm{h}} [12].

LASSCF. With exact integrals our two-step optimizer (orbital step at fixed RDMs, then fragment resolve) reaches −3655.4974221 EhE_{\mathrm{h}} after the 50-cycle limit (213 620 s) with ΔE≈−2.4×10−7Eh\Delta E \approx-2.4 \times10^{-7} E_{\mathrm{h}} per cycle and orbital gradient 6.6 × 10−4^{-4}: the energy is converged to ∼5 μEhE_{\mathrm{h}}, the gradient is not, because the line search stalls along a soft orbital–CI coupled mode. It lies 1.05 mEhE_{\mathrm{h}} below the reference.

LASSQD. Four runs warm-started from this LASSCF at fixed orbitals (heavy-hex and all-pairs LUCJ, seeds 1 and 2), which isolates the fragment-solver error, give −3655.4974223 EhE_{\mathrm{h}} in every case, 0.24 μEhE_{\mathrm{h}} below LASSCF (1.5 × 10−4^{-4} kcal mol−1^{-1}): the ten-qubit fragment SQD solves are exact. A self-consistent run with density fitting was not completed.

Computational details and reproducibility

Software and hardware. QuEnAIs [31] (DMET, LAS, baselines, router), PySCF [30], the GQE-for-QSCI trainer on CUDA-Q [27] (adapted to accept embedding Hamiltonians), Qiskit/Qiskit Aer [29] and ffsim [28] (versions: PySCF 2.14.0, CUDA-Q 0.15.1, PyTorch 2.13.0 with PyTorch Lightning 2.6.5, Qiskit 2.5.1, Qiskit Aer 0.17.2, qiskit-addon-sqd 0.12.1, ffsim 0.0.83, NumPy 2.4.6, SciPy 1.17.1). GENIE runs on an NVIDIA A100 (40 GB) with 24 CPU threads; SQD, LAS and baselines on 10–24-thread CPU nodes.

Reproducibility. Every number in this paper is a row of a benchmark grid run with tools/run_benchmark.py and collected into results.csv; classical baselines come from tools/selection_baselines.py and diagnostics from tools/regime_diagnostics.py --exact. Grids: n2_gqe_mo, n2_gqe_budget, n2_sqd_list, n2_noise, n2_dz-gqe-mo, h_chain-gqe, h14-gqe_depth, h_chain_compare (+ h14sqd, _h14halves_warm), fe2bpym_las, fe2bpym_lassqd_warm. Each run is identified by a hash of its parameters; its directory holds the input, logs, pickled intermediate results and a one-line row.json.

1.0 Å1.5 Å2.0 Å
H10_{10}, H2_2 units38.541.019.5
H10_{10}, halves84.532.59.5
H14_{14}, H2_2 units−7.43226−6.89705−6.59858
H14_{14}, halves−7.41492−6.95024−6.63159

Table XV. LAS on hydrogen chains (STO-3G). H10_{10}: error against whole-molecule FCI (kcal mol−1^{-1}); H14_{14}: LASSCF total energy (EhE_\mathrm{h}). LASSQD equals LASSCF within the stated tolerance.

taskcost
GENIE, N2_2/STO-3G12–85 min
GENIE, H14_{14} (80 operators)0.3–3.0 h
GENIE, N2_2/cc-pVDZ0.8–2.3 h
GENIE circuit, H14_{14} 2.0 Å, 160 / 320 ops∼\sim2.1 / 4.6×104\times10^4 CXf^f
LUCJ optimization, per geometry (18–20 qubits)1 555–3 010 s once
SQD run after optimization70–200 s
LUCJ after heavy-hex routing, N2_21 204–3 499 CX, depth 488–3 852
noisy SQD (s=1s = 1), N2_24.3 h
LASSCF Fe2_2, exact integrals59 h

Table XVII. Wall times and circuit sizes.

f^f As reported by the trainer, before routing.

References

  1. [1]K. Kanno, M. Kohda, R. Imai, S. Koh, K. Mitarai, W. Mizukami, and Y. O. Nakagawa, Quantum-selected configuration interaction: classical diagonalization of Hamiltonians in subspaces selected by quantum computers (2023), arXiv:2302.11320.DOI
  2. [2]J. Robledo-Moreno, M. Motta, H. Haas, A. Javadi-Abhari, P. Jurcevic, W. Kirby, et al., Chemistry beyond the scale of exact diagonalization on a quantum-centric supercomputer, Sci. Adv. 10.1126/sciadv.adu9991 (2025), arXiv:2405.05068.
  3. [3]J. Yu, J. Robledo Moreno, J. T. Isouе, L. Bertels, D. Claudino, et al., Quantum-centric algorithm for sample-based Krylov diagonalization (2025), arXiv:2501.09702.arxiv.org/abs/2501.09702
  4. [4]S. Piccinelli, A. Baiardi, S. Barison, M. Rossmannek, A. Carrera Vazquez, et al., Quantum chemistry with provable convergence via randomized sample-based Krylov quantum diagonalization (2025), arXiv:2508.02578.arxiv.org/abs/2508.02578
  5. [5]G. Knizia and G. K.-L. Chan, Density matrix embedding: A simple alternative to dynamical mean-field theory, Phys. Rev. Lett. 109, 186404 (2012).pubmed.ncbi.nlm.nih.gov/23215304
  6. [6]G. Knizia and G. K.-L. Chan, Density matrix embedding: A strong-coupling quantum embedding theory, J. Chem. Theory Comput. 9, 1428 (2013).DOI
  7. [7]S. Wouters, C. A. Jiménez-Hoyos, Q. Sun, and G. K.-L. Chan, A practical guide to density matrix embedding theory in quantum chemistry, J. Chem. Theory Comput. 12, 2706 (2016), arXiv:1603.08443.DOI
  8. [8]A. Shajan, D. Kaliakin, A. Mitra, J. Robledo Moreno, Z. Li, M. Motta, C. Johnson, A. A. Saki, S. Das, I. Sitdikov, A. Mezzacapo, and J. Merz, Kenneth M., Toward quantum-centric simulations of extended molecules: Sample-based quantum diagonalization enhanced with density matrix embedding theory, J. Chem. Theory Comput. 10.1021/acs.jctc.5c00114 (2025).
  9. [9]A. K. Patra, A. K. S. V., S. S. P., R. Bhat, R. V., R. Maitra, and J. G., Quantum simulation of ligand-like molecules through sample-based quantum diagonalization in density matrix embedding framework (2025), arXiv:2511.22158.DOI
  10. [10]M. R. Hermes and L. Gagliardi, Multiconfigurational self-consistent field theory with density matrix embedding: The localized active space self-consistent field method, J. Chem. Theory Comput. 15, 972 (2019), arXiv:1810.03266.DOI
  11. [11]M. R. Hermes, R. Pandharkar, and L. Gagliardi, Variational localized active space self-consistent field method, J. Chem. Theory Comput. 16, 4923 (2020), arXiv:2003.02995.
  12. [12]Q. Wang, M. Motta, R. D’Cunha, K. J. Sung, M. R. Hermes, T. Gujarati, Y. Kawashima, Y.-y. Ohnishi, G. O. Jones, and L. Gagliardi, Sample-based quantum diagonalization as parallel fragment solver for the localized active space self-consistent field method (2025), arXiv:2512.14936.arxiv.org/abs/2512.14936
  13. [13]T. Shiota, T. M. Do, W. Mizukami, and Y. Yoshida, Quantum computing for accurate large-scale electronic-structure calculations: DFT-embedded, post-processed quantum-selected configuration interaction (2026), arXiv:2606.06015.arxiv.org/abs/2606.06015
  14. [14]M. Motta, K. J. Sung, K. B. Whaley, M. Head-Gordon, and J. Shee, Bridging physical intuition and hardware efficiency for correlated electronic states: the local unitary cluster Jastrow ansatz for electronic structure, Chem. Sci. 14, 11213 (2023).DOI
  15. [15]A. K. Patra, A. K. S. V., S. S. P., R. Bhat, R. Maitra, and J. G., Machine-learned compact subspace generation for quantum selected configuration interaction within density matrix embedding framework (2026), arXiv:2607.20585.arxiv.org/abs/2607.20585
  16. [16]B. Huron, J. P. Malrieu, and P. Rancurel, Iterative perturbation calculations of ground and excited state energies from multiconfigurational zeroth-order wavefunctions, J. Chem. Phys. 58, 5745 (1973).DOI
  17. [17]N. M. Tubman, J. Lee, T. Y. Takeshita, M. Head-Gordon, and K. B. Whaley, A deterministic alternative to the full configuration interaction quantum Monte Carlo method, J. Chem. Phys. 145, 044112 (2016), arXiv:1603.02686.DOI
  18. [18]A. A. Holmes, N. M. Tubman, and C. J. Umrigar, Heat-bath configuration interaction: An efficient selected configuration interaction algorithm inspired by heat-bath sampling, J. Chem. Theory Comput. 12, 3674 (2016), arXiv:1606.07453.DOI
  19. [19]P. Reinholdt, K. M. Ziems, E. R. Kjellgren, S. Coriani, S. P. A. Sauer, and J. Kongsted, Critical limitations in quantum-selected configuration interaction methods, J. Chem. Theory Comput. 21, 6811 (2025).DOI
  20. [20]N. Bonilla Vargas, Machine learning for sample-based quantum diagonalization: a review of generative configuration recovery and the classical-simulatbility frontier (2026), arXiv:2608.05314.
  21. [21]H. Zhang and M. Otten, From random determinants to the ground state (2025), arXiv:2511.14734.
  22. [22]W. Kirby, B. Pokharel, J. Robledo Moreno, et al., Observation of improved accuracy over classical sparse ground-state solvers using a quantum computer (2026), arXiv:2603.03496.arxiv.org/abs/2603.03496
  23. [23]K. Nakaji, L. B. Kristensen, R. Kemmoku, J. A. Campos-Gonzalez-Angulo, M. G. Vakili, H. Huang, M. Bagherimehrab, C. Gorgulla, F. Wong, A. McCaskey, J.-S. Kim, T. Nguyen, P. Rao, Q. Gao, M. Sugawara, N. Yamamoto, and A. Aspru-Guzik, The generative quantum eigensolver (GQE) and its application for ground state search (2024), arXiv:2401.09253.arxiv.org/abs/2401.09253
  24. [24]R. Kemmoku, Q. Gao, S. Kanno, K. Keithley, I. Hamamura, N. Yamamoto, and K. Nakaji, Generative circuit design for quantum-selected configuration interaction (2026), arXiv:2604.09756.arxiv.org/abs/2604.09756
  25. [25]K. Elgammal and M. Maußner, Scaling a CUDA-Q GQE + QSCI pipeline to 40 qubits for EUV photoresist chemistry (2026) arXiv:2607.23988arxiv.org/abs/2607.23988
  26. [26]E. R. Sayfutyarova, Q. Sun, G. K.-L. Chan, and G. Knizia, Automated construction of molecular active spaces from atomic valence orbitals, J. Chem. Theory Comput. 13, 4063 (2017).
  27. [27]NVIDIA, CUDA-Q: a platform for hybrid quantum–classical computing, https://github.com/NVIDIA/cuda-quantum (2026), version 0.15.1.
  28. [28]The ffsim developers, ffsim: faster simulation of fermionic quantum circuits, https://github.com/qiskit-community/ffsim (2026), version 0.0.83.
  29. [29]A. Javadi-Abhari et al., Quantum computing with Qiskit (2024), arXiv:2405.08810.
  30. [30]Q. Sun et al., Recent developments in the PySCF program package, J. Chem. Phys. 153, 024109 (2020).DOI
  31. [31]A. V. Loharkar, QuEnAIS: quantum embedding with active-space selection and quantum impurity solvers, version 3.0.3, Zenodo (2026), source: https://github.com/AwadhootVijayLoharkar/Quenais-Project, tag v0.3.0.
  32. [32]T. Weaving, A. Mingare, A. Ralli, and P. V. Coveney, Towards compact wavefunctions from quantum-selected configuration interaction, J. Chem. Theory Comput. 10.1021/acs.jctc.5c01994 (2025), arXiv:2509.02525.DOI
  33. [33]A. V. Loharkar, Data for the GENIE paper (QuEnAIS 0.3.0 benchmarks), Zenodo (2026).DOI

Paper details

vlnc:W_d5ye8n9a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

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