Researchers Improve Quantum State Optimisation with New Phase-Gradient Learning Method

Yi-Ran Xue and colleagues at University of Massachusetts, in collaboration with Hefei National Laboratory, A. Alikhanyan National Science Laboratory, Nanjing University, Jiangsu Physical Science Research Centre, and 1 other institutions, have overcome a key limitation in applying neural-network quantum states to complex quantum systems. Optimising these states is often hindered by inaccuracies in estimating phase gradients, particularly in systems exhibiting complex phase structures common in gauge fields and fermionic statistics. Direct estimation of the phase gradient, rather than relying on stochastic methods, sharply reduces variance and improves accuracy, achieving a median error of $0.89\%$ on a 100-site flux ladder where conventional methods plateau at $1.8\%$. Further validation on chiral XXX chains confirms that estimator design is a key factor in developing effective complex-valued neural quantum states.

Direct energy differentiation unlocks efficient optimisation of quantum states

A direct estimator of the phase gradient, a property analogous to the slope of a hill indicating the steepest direction of ascent in a quantum system, was employed to overcome optimisation challenges in neural-network quantum states. These states represent a variational approach to approximating the ground state wavefunction of complex quantum many-body systems, utilising artificial neural networks as flexible parameterisations. The underlying principle involves training the network to minimise the energy of the system, effectively finding the lowest energy state. Instead of relying on a stochastic estimator, which approximates the phase gradient through Monte Carlo sampling and inherently suffers from statistical noise, the local energy of the system was differentiated with respect to the network parameters to calculate the phase gradient directly. This differentiation provides an analytical, deterministic estimate, circumventing the limitations of stochastic methods. The energy functional used is central to this approach, representing the Hamiltonian of the quantum system being simulated.

This new method for optimising neural-network quantum states resulted in improved performance on complex quantum systems, tackling challenges arising when modelling systems with specific properties, such as those found in materials with strong magnetic fields or those lacking time-reversal symmetry. These systems often exhibit non-trivial topological phases and complex correlations, making accurate simulation computationally demanding. Employing a 100-site flux ladder, a model system used to study frustrated magnetism and quantum phase transitions, the team reached a median error of 0.89 per cent, a sharp reduction from the 1.8 per cent plateau observed with conventional methods. The flux ladder’s geometry and magnetic frustration introduce significant challenges for traditional numerical methods. Further analysis of the training process indicated the improvement stems from eliminating unsuccessful optimisation runs, rather than uniformly enhancing already-converged solutions, and that the benefits increase with the complexity of the magnetic flux being modelled. This suggests the direct estimator is particularly effective in navigating complex energy landscapes with multiple local minima.

Phase gradient estimation limits optimisation of complex neural quantum states

Error rates dropped to 0.89% when using this direct estimator for optimising neural-network quantum states, a significant improvement over standard baselines which plateaued at 1.8%. This breakthrough overcomes a longstanding limitation in modelling complex quantum systems, particularly those with intricate phase structures found in materials exhibiting gauge fields or fermionic statistics. Gauge fields, arising from the presence of charged particles, and fermionic statistics, governing the behaviour of particles like electrons, introduce complex phases into the wavefunction, making accurate optimisation difficult. Researchers from University of Massachusetts and Nanjing University established that the fragility in optimising these states originates from inaccuracies in estimating phase gradients, not the network’s representational capacity. This is crucial because it indicates the bottleneck lies in the optimisation algorithm, not the ability of the neural network to represent the ground state. Validation on chiral XXX chains, a one-dimensional spin model exhibiting topological properties, establishes estimator design as a vital element in developing effective complex-valued neural quantum states; wider or deeper networks utilising the original approach degraded from 8.4% to 24.6% error. This demonstrates that simply increasing network size does not solve the problem and highlights the importance of accurate gradient estimation.

Improving quantum simulations through refined phase gradient calculations

Artificial intelligence techniques are being refined to better simulate the behaviour of complex quantum systems, unlocking potential advances in materials science and fundamental physics. Quantum simulation aims to understand the properties of materials and phenomena that are intractable for classical computers. The team at Nanjing University and the University of Massachusetts have demonstrated that a key stumbling block isn’t the neural network’s ability to learn, but the way it calculates essential quantum properties, specifically estimating the ‘phase gradient’, a measure of how a quantum state changes. The phase gradient dictates the direction and magnitude of adjustments needed to the network parameters during the optimisation process. Refining the calculation method, through direct measurement of the energy change instead of relying on statistical guesswork, dramatically improves accuracy in simulating complex quantum systems. This identifies the source of instability in optimising neural-network quantum states not as a limitation of the networks themselves, but as an inaccuracy in calculating this property. The conventional stochastic approach relies on estimating the phase gradient from a limited number of samples, leading to statistical fluctuations and hindering convergence. By switching from a statistical estimation of the phase gradient to a direct calculation based on local energy differences, a substantial reduction in error during simulations was achieved. This direct calculation provides a more precise and reliable signal for guiding the optimisation process, leading to more accurate and efficient simulations of quantum systems and potentially accelerating the discovery of new materials and physical phenomena.

The researchers successfully improved the accuracy of neural-network quantum state simulations by refining how the phase gradient is calculated. This matters because inaccurate gradient estimation was previously limiting the performance of these simulations, not the networks’ learning capacity itself. Using a direct calculation of energy change, rather than a statistical estimate, reduced the median error to 0.89% on a 100-site flux ladder, significantly outperforming standard methods which plateaued at 1.8% or even degraded with increased network size. The authors demonstrated this improved method on both flux ladders and chiral XXX chains, suggesting it offers a robust solution for complex-valued quantum states.