Autoregressive Projective Quantum Monte Carlo: From a Hermitian to a Non-Hermitian Perspective

AI summary

Overview

• Research area: Computational quantum many-body physics — projective quantum Monte Carlo (PQMC) combined with neural-network quantum states (NQS), extended from Hermitian to non-Hermitian lattice models.
• Technical level: Advanced. It assumes familiarity with imaginary-time projection, Green's functions, importance sampling, Trotter error, autoregressive neural networks, and parity-time (PT) symmetry.
• Scope (one sentence): The paper introduces an autoregressive, recurrent-neural-network-guided PQMC algorithm and benchmarks it against unguided PQMC and variational RNN ansätze on a staggered-field 1D Ising chain, both Hermitian and non-Hermitian.

What This Paper Is About

PQMC projects the ground state out of a trial wavefunction by evolving walkers in imaginary time, and its accuracy hinges on how good that guiding wavefunction is. The authors turn the usual relationship around: instead of using Monte Carlo merely to train a neural network, they let a recurrent neural network (RNN) learn from the walkers and then use it to guide the Monte Carlo. The goal is a method that stays polynomial in cost while producing more accurate ground-state energies and smaller fluctuations than standard PQMC, and that also works for non-Hermitian Hamiltonians — a regime the authors state had not previously been studied with PQMC of any kind.

Key Contributions

1. Autoregressive PQMC framework. An RNN supplies the guiding (trial) wavefunction inside a self-learning PQMC loop. Because the autoregressive structure factorizes the wavefunction by the chain rule, configurations can be sampled independently with tractable likelihoods, unlike restricted Boltzmann machines, which require Gibbs sampling over hidden units.
2. Extension to non-Hermitian systems. The authors apply PQMC — simple or neural-guided — to a PT-symmetric non-Hermitian Ising chain with a complex staggered field $g=\eta+i\xi$, and show a unitary (gauge) rotation can mitigate the sign problem for this model, while explicitly noting the trick is not general.
3. Systematic benchmarking against exact and analytic references. Results are checked against exact diagonalization (ED), the Jordan–Wigner–Bogoliubov (JWB) analytic solution, and, for the non-Hermitian case, high-order series expansions and previously published variational RNN results.
4. Practical characterization of cost and convergence. The paper reports wall-clock times, walker counts, convergence in $N_s = 5$ to $10$ stints, and a comparison of the finite-time-step bias of continuous-time, symmetrized, and non-symmetrized PQMC.

Main Findings

• Better accuracy than unguided PQMC at large size. For the Hermitian staggered-field Ising chain, Table 1 compares SPQMC and autoregressive PQMC against JWB. At $N=50$ spins: relative error $1.145\times10^{-4}$ (SPQMC) versus $2.315\times10^{-6}$ (autoregressive), with energies $-88.035485019236$ versus $-88.025202331421$ against a JWB value of $-88.025406110236$. At $N=100$: $3.337\times10^{-4}$ versus $5.861\times10^{-6}$ (energies $-176.010647660133$ and $-176.049780386649$ against JWB $-176.050812220459$). At $N=150$: $1.141\times10^{-3}$ versus $1.158\times10^{-5}$ (energies $-264.378585600678$ and $-264.073160328081$ against JWB $-264.076218330689$).
• Lower energy variance. Energy variance per spin in the same table: at $N=50$, $2.213\times10^{-4}$ (SPQMC) versus $2.523\times10^{-5}$ (autoregressive); at $N=100$, $2.284\times10^{-4}$ versus $2.883\times10^{-5}$; at $N=150$, $5.111\times10^{-4}$ versus $8.221\times10^{-5}$.
• Modest extra runtime. At $N=50$: 01:37:11 (SPQMC) versus 01:50:37; at $N=100$: 04:53:47 versus 07:22:23; at $N=150$: 07:14:59 versus 10:00:01. Walker counts were $5\times10^{4}$, $10^{5}$, and $1.5\times10^{5}$ respectively, with the same total projection time in all runs.
• Fewer stints than RBM guiding. Convergence was observed within $N_s = 5$ to $N_s = 10$ stints, described as approximately four times fewer than the number of stints required for RBMs in Ref. [41], and this window held regardless of system size.
• Reduced fluctuations around the exact energy. In Fig. 5(a) for $N=20$, $g=1.6$, $\Delta\tau=10^{-2}$, $N_w=2\times10^{4}$, $MCs=2\times10^{3}$ and projection time $\Delta\tau_p=20$, the guided curve tracks ED and the analytic JWB result much more tightly than SPQMC or PQMC with a random $\psi_T$; each peak marks the start of a new stint. Fig. 5(b) shows correspondingly lower energy variance for the guided method.
• Non-variational character made explicit. Because PQMC uses a mixed estimator (Eq. 8), energies are not bounded from below by the true ground-state energy; instantaneous local energies can dip below it. The authors argue that a more expressive ansatz shrinks these excursions, making the method behave "variational-like."
• Error peaks at criticality. In Fig. 6(a), for $N=20$, the relative error with respect to ED grows in the ferromagnetic phase ($g<1$), decreases in the paramagnetic phase ($g>1$), and is largest near $g_c\approx 1$. At $N=20$ the autoregressive method is reported as at least one order of magnitude more accurate than SPQMC. Fig. 6(b) shows both methods improve with more walkers, at higher cost.
• Finite-time-step bias quantified. In Fig. 2 ($N_w=10^{3}$, $N=20$), the symmetrized SPQMC curve fits $ax^2$ with $a=0.285(7)$ and the non-symmetrized curve fits $ax$ with $a=0.263(5)$, while continuous-time PQMC shows no meaningful fit and mitigates the bias; the paper adopts continuous-time PQMC.
• Non-Hermitian ground state captured. For the PT-unbroken regime with open boundary conditions and $g=1.6$, the real part of the energy coincides with the negative absolute value of the energy (sign added for convenience), all local energies are real, and the algorithm converges to the true ground state, defined as the state with the smallest real part of the energy (Fig. 9(a), $N=10$ versus ED). For $N=100$ (Fig. 9(b)), autoregressive PQMC is compared with variational RNN results from Ref. [52] and with high-order series expansions; the authors report that adding non-Hermiticity shrinks the ferromagnetic phase relative to the Hermitian transition line. The excerpt ends mid-sentence while stating the relative error, so the numerical value of that comparison is not available here.
• Outperforms variational RNNs. The abstract states the approach consistently achieves lower energies and higher fidelity than conventional variational RNN ansätze regardless of system size or Hermiticity; the supporting comparison is deferred to Appendix F, which is not included in the provided text.

Methodology in Plain English

The starting point is the standard projective (diffusion) Monte Carlo recipe. A population of "walkers" — each a full spin configuration — is evolved in imaginary time. Walkers that land in low-energy regions get duplicated (branching), and low-weight ones die out, so the population drifts toward the ground state. A reference energy $E_r = E + \mu\log(N_{\text{avg}}/N_{\text{cur}})$ keeps the population near a target size.

Unguided PQMC wanders blindly through configuration space, which is inefficient. The fix is importance sampling with a guiding function $\psi_T$: walkers are pushed along the gradient of $\psi_T$ (drift) while still diffusing, and the branching weight is set by the local energy. The energy is then read off as a mixed estimator between $\psi_T$ and the projected state.

The new ingredient is where $\psi_T$ comes from. Rather than optimizing it separately beforehand, the authors alternate: run PQMC for a fixed number of steps using the current RNN as guide, then train the RNN by maximum likelihood (equivalently, minimizing KL divergence) on the resulting walker configurations, then repeat. The RNN's autoregressive form gives exact conditional probabilities and independent samples, which makes both the training and the proposal moves cleaner than RBM-based guiding. Continuous-time PQMC is used to remove the finite-$\Delta\tau$ bias that the first-order Trotter splitting would otherwise introduce.

For the non-Hermitian model, the complex field makes the Hamiltonian non-stoquastic, so a unitary basis rotation is applied to strip the phase from the field while leaving the spectrum untouched — a mitigation the authors stress is model-specific, not a general cure for the sign problem.

All runs were performed on an HPC cluster with NVIDIA Quadro RTX 6000 GPUs (24 GB VRAM, CUDA 12.2); full RNN and PQMC hyperparameters are collected in Appendix D.

Why This Matters

• Impact on research: The paper argues for a reversal of the usual division of labor, in which NQSs serve QMC (guiding and variance reduction) rather than QMC serving NQSs (sampling and gradient estimation). It also opens PQMC — previously restricted to stoquastic, sign-problem-free Hamiltonians — to PT-symmetric non-Hermitian lattice models, where only variational NQS approaches had been used.
• Real-world applications (as motivations for the underlying problem):
  • Strongly correlated materials, where mean-field and perturbation theory fail.
  • Quantum chemistry and electronic-structure energetics of many-body systems.
  • Open and dissipative quantum systems, which non-Hermitian Hamiltonians model.
  • Benchmarking and validating quantum simulators and near-term quantum hardware on classically hard ground states.
• Industry relevance: Ground-state energy estimation underpins materials discovery, catalyst and drug-design pipelines that rely on electronic structure, and the calibration of quantum computing hardware. A polynomial-cost method that improves accuracy per unit of compute, and that reports explicit GPU wall-clock times, is directly relevant to groups deciding how to spend simulation budgets.

Future Directions

• Generalizing the sign-problem mitigation. The gauge-rotation trick works for the specific non-Hermitian Ising chain studied; the authors flag generalization to other non-Hermitian Hamiltonians as an open avenue. The similarity-transformation route for PT-symmetric Hamiltonians is noted to introduce nonlocality (Appendix C).
• Scaling to larger and higher-dimensional lattices. The excerpt motivates the work partly as an extension to lattice models; whether the method retains its advantage in two dimensions and beyond, and how walker counts must grow (already increased from $5\times10^4$ to $1.5\times10^5$ between $N=50$ and $N=150$), remains to be established.
• Fermionic and frustrated systems. The paper discusses the sign problem as a general obstacle, particularly for fermions, and does not demonstrate the method there.
• Tighter control of the mixed-estimator bias. Since PQMC is not variational and individual local energies can fall below the true ground-state energy, understanding and systematically reducing the downward excursions (from finite walker populations, estimator variance, guide quality, and propagator discretization) is a natural follow-up.

Target Audience

Researchers and graduate students in computational condensed-matter and quantum many-body physics who already work with quantum Monte Carlo or neural-network quantum states. It will also interest practitioners of non-Hermitian and PT-symmetric physics looking for a numerical alternative to variational ansätze, and machine-learning researchers working on autoregressive generative models applied to physical systems. Readers without a background in imaginary-time projection, importance sampling, and variational Monte Carlo will find the technical sections demanding.

Authors’ abstract

Accurately determining the ground-state properties of quantum many-body systems remains a central challenge. In this work, we introduce an autoregressive projective quantum Monte Carlo (PQMC) framework that leverages recurrent neural networks (RNNs) to guide the stochastic dynamics. By incorporating autoregressive sampling into PQMC, we demonstrate substantial improvements in accuracy compared to standard unguided PQMC, while retaining polynomial computational cost. We benchmark our approach against conventional variational RNN ansätze and find that the autoregressive PQMC consistently achieves lower energies and higher fidelity, regardless of system size or whether the Hamiltonian is Hermitian or non-Hermitian. Our results highlight the versatility and power of neural-guided PQMC methods, paving the way for promising scalable simulations of low-energy states in complex quantum many-body systems.