arXiv:2610.18681
Explicitly correlated ground states with a feedback-free variational hybrid algorithm
Benjamin Corbett, Akimasa Miyake
Abstract
Explicitly correlated methods, which represent dominant short-range correlations by a diagonal Jastrow-type correlator, are designed to accurately describe systems of strongly correlated electrons, such as the fully periodic 2D Fermi--Hubbard model. However, as the reference state is compact in momentum space whereas the correlator acts naturally in real space, there is a fundamental mismatch of bases that classical algorithms resolve only by restricting the reference or the correlator. Here we propose a hybrid algorithm in which a quantum processor prepares a low-entanglement momentum-space matrix product state (MPS), precomputed classically by the transcorrelated method at the cost of a nonvariational energy, rotates it to real space, and measures in bases fixed by the Hamiltonian. Because the measurement bases are independent of the correlator, the correlator can be variationally optimized entirely in classical post-processing with no quantum-classical feedback loop. We demonstrate our algorithm on 32 qubits of Quantinuum's H2-2 with a bond-dimension-4 MPS for the half-filled $4 \times 4$ lattice and achieve a relative ground-state energy error of $16 \pm 4\%$ after postselection, reduced to $3.5 \pm 0.5\%$ by calibration using low-order moments from the MPS. Compared with other feedback-free hybrid algorithms, even under generous assumptions, our method requires fewer samples than quantum-selected configuration interaction and fewer two-qubit gates than quantum Krylov methods.
AI Summary
The paper addresses the basis mismatch between compact momentum-space MPS references and real-space diagonal correlators for correlated-electron ground states. Classical variational Monte Carlo requires real-space reference amplitudes that are generally hard to compute, while transcorrelated DMRG avoids this requirement but yields nonvariational energies. The proposed hybrid algorithm prepares a classically computed transcorrelated MPS, Fourier-rotates it to real space, and samples occupation and hopping measurement bases. These bases are independent of the diagonal correlator $C_\theta$, allowing its nonlinear optimization entirely in classical post-processing using the same data. Sample probabilities are calibrated against low-order moments computed from the reference MPS. For the half-filled periodic $4\times4$ Hubbard model at $U/t=4$, classical benchmarks give $E=-13.609\pm0.005$ with bond dimension 16 and a 103-parameter correlator, versus the exact $-13.621855$. A 32-qubit Quantinuum H2-2 demonstration uses a bond-dimension-4 reference, 555 two-qubit gates per occupation-measurement circuit, and approximately 36,000 samples. Relative energy error decreases from $16\pm4\%$ after particle-number postselection to $3.5\pm0.5\%$ after calibration. Benchmark comparisons find lower sample requirements than raw sampled QSCI and lower circuit gate counts than the examined Krylov implementations, under explicitly stated resource models. The variational bound holds for exact expectations, not generally for finite-sample or calibrated noisy estimates. Flexible correlators can exploit sampling inconsistencies, including divergent objectives. Sampling studies on $4\times4$ and $6\times6$ lattices motivate—but do not establish—half-filled cost scaling $\exp(cN)\operatorname{poly}(N)$ with $c<0.1$. Larger doped lattices, richer-correlator sampling costs, and quantum advantage remain untested.
Their features were selected by an LLM-assisted heuristic search using the available samples, without reference to the exact ground-state energy.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.