arXiv:2611.04496
Broken-Symmetry Reference States for Sample-Based Quantum Diagonalization with Near-DMRG Accuracy
Kunal Kumar, Tomonori Shirakawa, Boseong Kim, Ryo Wakizaka, Ikko Hamamura, Mitsuhisa Sato, Seiji Yunoki, Mario Motta, Hiroshi Horii, Yuri Kobayashi
Abstract
Sample-based quantum diagonalization (SQD) accuracy remains limited for strongly correlated systems like iron-sulfur clusters. This accuracy heavily depends on the mean-field reference state, which determines the sampling ansatz and configuration-recovery target. We extend the qcsc-prefect workflow to support a broken-symmetry unrestricted Hartree-Fock (BS-UHF) starting point, derived from atom-localized antiferromagnetic initial guesses. For the $[Fe_4S_4(SCH_3)_4]^{2-}$ cubane (54 electrons, 36 orbitals, 72 data qubits; 88 total qubits), this lowers the mean-field reference by $\approx 534$ mHa relative to the Restricted Hartree-Fock (RHF) solution (-327.0809 vs. -326.5474 Ha). Combined with deep multi-step configuration recovery on ROQUO GPU nodes at subspace dimension $d = 3 \times 10^{10}$, we reach $E = -327.235076$ Ha. This is 4.3 mHa above the finite-$D = 4000$ DMRG value (-327.2393952 Ha) and $\sim 512-517$ mHa below a previous same-active-space RHF-reference SQD calculation. Extrapolating to the $d \to \infty$ limit via a Gaussian-in-log $d$ fitting ansatz gives an asymptotic estimate $E_\infty = -327.244468$ Ha $\pm 1.05$ mHa, just 0.33 mHa above our discarded-weight-extrapolated DMRG estimate ($E_{D\to\infty} = -327.244797$ Ha $\pm 0.46$ mHa), well within their combined $1σ$ uncertainty. We characterize the determinant-subspace composition driving convergence, show matching validation on $[Fe_2S_2(SCH_3)_4]^{2-}$, and report multi-HPC execution on Fugaku, Miyabi-G, and ROQUO. We conclude that reference-state preparation, rather than subspace size alone, is a key driver of SQD energy accuracy under strong electronic correlation.
AI Summary
The paper asks how the mean-field reference affects sample-based quantum diagonalization (SQD) for strongly correlated iron–sulfur clusters. It extends `qcsc-prefect` to prepare atom-localized antiferromagnetic broken-symmetry UHF references, initialize LUCJ sampling circuits from unrestricted coupled-cluster amplitudes, and perform checkpointed configuration recovery and classical subspace diagonalization. A candidate-generation booster recombines sampled excitation patterns. For $[\mathrm{Fe}_4\mathrm{S}_4(\mathrm{SCH}_3)_4]^{2-}$ in CAS(54e,36o), the selected BS-UHF reference is $534$ mHa below RHF. A 72-data-qubit circuit, using 88 physical qubits, supplies five million hardware shots; the resulting pool is reused throughout recovery. At subspace dimension $d=3\times10^{10}$, the directly computed energy is $-327.235076$ Ha: $4.3$ mHa above finite-$D=4000$ DMRG and $9.7$ mHa above extrapolated DMRG. An empirical Gaussian-in-$\log d$ fit to five SQD points estimates $-327.244468\pm1.05$ mHa, versus DMRG’s $-327.244797\pm0.46$ mHa. A smaller CAS(30e,20o) cluster provides an additional SQD–DMRG comparison. Rank-resolved wavefunction weights document increasing triple-and-higher excitation content with subspace size. The extrapolation uncertainties are fit-conditional; the larger-cluster calculation samples only $3.4\times10^{-4}\%$ of the complete space. The paper also reports iteration-capped Davidson solves with residual convergence errors, unmeasured final-state spin contamination, and unmatched SQD/DMRG reference preparation. Comparisons with earlier RHF-based SQD additionally differ in dimension, recovery depth, samples, and hardware. No quantum advantage or accuracy for chemical energy differences is established.
Central claim may not hold: The paper's strong attribution of the historical energy improvement almost entirely to the broken-symmetry reference is not isolated from a 300-fold increase in subspace dimension and other workflow changes. The quoted SQD extrapolation uncertainty is conditional on an empirical three-parameter fit to five points and does not quantify functional-form uncertainty or unresolved diagonalization error. The largest projected diagonalizations terminate with several-mHa last-iteration energy changes, so small recovery-step changes do not demonstrate eigensolver convergence at the accuracy claimed for the extrapolation. The dimer's fitted infinite-dimension asymptote differs from its fitted complete-space endpoint, so its agreement with DMRG does not validate that asymptote as the exact-diagonalization limit.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 22 pages, 10 figures, 7 tables. Data: https://doi.org/10.5281/zenodo.23064186 (FCIDUMPs and raw quantum-hardware sample pools)