⚛️ quantum physics

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

This paper introduces an autoregressive projective quantum Monte Carlo framework that utilizes recurrent neural networks to guide stochastic dynamics, demonstrating superior accuracy and scalability over standard methods for determining ground-state properties in both Hermitian and non-Hermitian quantum many-body systems.

Original authors: Lavoisier Wah, Remmy Zen, Flore K. Kunst

Published 2026-08-11

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Lavoisier Wah, Remmy Zen, Flore K. Kunst

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, cosmic game of chess, but instead of 64 squares and 32 pieces, the board has billions of invisible squares, and the pieces are subatomic particles like electrons and atoms. Physicists call this the "quantum many-body problem." The goal is to figure out how these particles arrange themselves when they are at their lowest energy state, known as the "ground state." Think of it like trying to find the absolute flattest, most relaxed pose a tangled ball of yarn can take. If you can solve this, you can predict how materials conduct electricity, how magnets work, or even how new super-materials might behave.

The trouble is, the number of possible arrangements for these particles grows so fast it's like trying to count every grain of sand on every beach on Earth, all at once. Traditional math tools often get stuck or give up. To tackle this, scientists use a clever trick called "Quantum Monte Carlo." Imagine you are trying to find the deepest valley in a foggy mountain range. You can't see the whole map, so you send out hundreds of tiny explorers (called "walkers") to wander around. They take random steps, but if they stumble into a steep cliff, they might fall back; if they find a gentle slope, they might split into more explorers to cover more ground. Over time, the crowd of explorers naturally gathers in the deepest valley, revealing the answer. However, if the fog is too thick or the terrain too weird, the explorers might get lost or wander into the wrong valleys.

Now, imagine giving those explorers a GPS. That is exactly what this paper does. The researchers, Lavoisier Wah, Remmy Zen, and Flore K. Kunst, have built a new system where a smart computer brain—a type of artificial intelligence called a Recurrent Neural Network (RNN)—acts as a guide for the explorers. Instead of wandering blindly, the explorers are nudged by the AI toward the most promising paths. The team tested this "Autoregressive Projective Quantum Monte Carlo" method on two types of quantum systems: the standard kind we know well (Hermitian) and a more exotic, tricky kind where energy can leak out or behave strangely (Non-Hermitian).

Here is what they found: When they let the AI guide the explorers, the system found the "deepest valley" (the ground state) much faster and with much higher precision than when the explorers wandered on their own. In fact, their AI-guided method was so good that it beat other popular computer methods that try to guess the answer directly. They showed that this works for both normal quantum systems and the weird, non-Hermitian ones, which are usually very hard to simulate because they are prone to a "sign problem"—a mathematical glitch where positive and negative numbers cancel each other out, making the signal disappear. By using a special mathematical trick (a gauge rotation) to fix the glitch, their AI-guided explorers could still find the right path.

The results were impressive. For a system with 20 spins (particles), their method was at least ten times more accurate than the old, unguided version. Even when they scaled up to 150 spins—a size where the old methods start to struggle—their new approach stayed accurate, with errors remaining tiny. They also found that the AI didn't need to be retrained from scratch every time; it could learn as it went, improving its "GPS" with every round of simulation. While they didn't solve every possible quantum puzzle, they proved that this specific combination of AI and Monte Carlo simulation is a powerful new tool. It suggests that by letting machine learning guide the random walks, we can simulate complex quantum materials that were previously too difficult to study, opening the door to understanding everything from new magnets to exotic states of matter.

Technical Summary: Autoregressive Projective Quantum Monte Carlo

Problem Statement
Accurately determining the ground-state properties of quantum many-body systems remains a central challenge in computational physics. Traditional approaches, such as Variational Monte Carlo (VMC), rely on optimizing a trial wavefunction but are upper-bounded by the variational principle, often failing to capture strong correlations efficiently. Projective Quantum Monte Carlo (PQMC), also known as Diffusion Monte Carlo (DMC), offers a more powerful alternative by projecting the ground state from a trial wavefunction via imaginary-time evolution. However, standard PQMC (specifically the Simple PQMC or SPQMC) stochastically explores the entire configuration space, leading to high variance and inefficiency. While importance sampling using a guiding wavefunction ( $\psi_T$ ) can mitigate this, existing methods often rely on Restricted Boltzmann Machines (RBMs) for $\psi_T$ , which suffer from sampling overhead, correlations, and the need for complex block-averaging techniques. Furthermore, extending these methods to non-Hermitian (NH) systems is particularly difficult due to the "sign problem" arising from complex off-diagonal Hamiltonian elements, a regime where neither simple PQMC nor PQMC guided by neural networks has been previously explored.

Methodology
The authors introduce an Autoregressive Projective Quantum Monte Carlo (PQMC) framework that integrates Recurrent Neural Networks (RNNs) as guiding wavefunctions for stochastic dynamics. The methodology proceeds as follows:

Self-Learning Loop: The algorithm operates in a cycle of "stints." In each stint, a PQMC simulation is run using the current RNN parameters ( $\theta_s$ ) to guide the walkers (system replicas) via importance sampling. This generates a distribution of walkers approximating the ground state $|\psi_0|^2$ .
Autoregressive Sampling: Unlike RBMs, the RNN provides an explicit, normalized probability amplitude for any spin configuration via autoregressive factorization (chain rule). This allows for the exact evaluation of conditional probabilities and the generation of independent samples without the need for Gibbs sampling or marginalization over hidden units.
Parameter Update: The walker configurations generated during the stint are used to update the RNN parameters by maximizing the likelihood (minimizing the Kullback–Leibler divergence) between the RNN distribution and the PQMC walker distribution.
Continuous-Time Approximation: To mitigate biases associated with finite time steps ( $\Delta\tau$ ), the authors employ a continuous-time approximation, slicing the MC time step into infinitesimal intervals to reduce discretization errors.
Non-Hermitian Extension: For non-Hermitian systems, the authors address the sign problem (caused by complex off-diagonal elements) by applying a unitary gauge rotation (basis change). This transformation renders the Hamiltonian effectively stoquastic (sign-problem-free) within the $\mathcal{PT}$ -unbroken regime, allowing the standard PQMC formalism to be applied. The local energy is computed using a mixed estimator adapted for biorthogonal systems, utilizing only the right ground-state wavefunction.

Key Contributions

Framework Integration: The paper establishes the first framework combining autoregressive neural sampling (RNNs) with PQMC, moving beyond the traditional use of VMC with neural networks.
Algorithmic Advantages: By utilizing RNNs, the method achieves independent sampling, avoids the sampling overhead that necessitates block averaging, and enables direct, stable maximum-likelihood updates. This results in faster convergence (requiring fewer stints, $N_s \approx 5$ , compared to RBMs) and reduced variance.
Non-Hermitian Application: The work represents the first application of PQMC guided by an NQS to non-Hermitian lattice models, as neither simple PQMC nor neural-guided PQMC had been used to study such systems previously.

Results
The authors benchmarked their approach against standard SPQMC, exact diagonalization (ED), and analytical solutions (Jordan-Wigner-Bogoliubov) for both Hermitian and non-Hermitian 1D transverse-field Ising models.

Hermitian Systems: For the 1D Ising chain with an alternating magnetic field, the autoregressive PQMC consistently achieved lower energies and higher fidelity than SPQMC and PQMC with random trial wavefunctions. The method showed significantly reduced fluctuations in local energy and variance. It remained robust for large system sizes ( $N=150$ ), maintaining low relative errors ( $\epsilon_{rel} \sim 10^{-5}$ ) compared to exact solutions.
Non-Hermitian Systems: Applied to a $\mathcal{PT}$ -symmetric non-Hermitian Ising chain, the method successfully captured the ground-state properties in the $\mathcal{PT}$ -unbroken regime. The autoregressive PQMC outperformed variational RNN approaches (VMC), exhibiting a smaller relative error with respect to high-order series expansions and a slower scaling of error with increasing system size ( $N$ ).
Efficiency: The autoregressive approach required fewer training stints to converge compared to previous RBM-guided methods, demonstrating superior expressivity and efficiency in learning the ground-state wavefunction.

Significance
The paper claims that this work establishes two "firsts": the integration of autoregressive neural sampling into PQMC and its application to non-Hermitian discrete lattice models. The authors posit that their results highlight the versatility of neural-guided PQMC, offering a scalable route to simulate low-energy states in complex quantum many-body systems where traditional methods (perturbation theory, mean-field) or standard VMC fail. By providing a mechanism to mitigate the sign problem in specific non-Hermitian regimes, the framework opens the door for future explorations in higher dimensions and across arbitrary non-Hermitian Hamiltonians, paving the way for more accurate and efficient quantum simulations.

Technical Summary: Autoregressive Projective Quantum Monte Carlo

More like this