Direct/adaptive-mixture phase-gradient learning for neural-network quantum states with complex phase structure
Researchers have developed a new learning method called direct/adaptive-mixture phase-gradient learning for neural-network quantum states. This technique addresses challenges in optimizing these states when they possess complex phase structures, which are common in areas like gauge fields and fermionic statistics. By using a direct estimator for the phase gradient with lower variance than traditional methods, the new approach significantly reduces errors, as demonstrated on a 100-site flux ladder where it achieved a median error of 0.89%. The adaptive mixture of estimators further improves performance by minimizing failed runs, highlighting estimator design as a critical factor for complex-valued neural quantum states. AI
IMPACT Enhances optimization techniques for complex quantum systems, potentially accelerating research in condensed matter physics and quantum computing.