Quantum Generative Adversarial Networks: Improving Dynamics Simulation with an Ancilla Qubit

Abstract

Simulating complex quantum systems remains a critical challenge, as conventional quantum techniques– such as those based on the Suzuki–Trotter decomposition—often result in deep circuits that demand substantial computational resources. Quantum Generative Adversarial Networks (QGANs) offer a promising alternative by learning the time evolution of target Hamiltonian using significantly fewer gates. However, standard QGAN architectures commonly suffer from unstable convergence and learning plateaus in the loss landscape, which hinder training and prevent the generator from achieving high-fidelity solutions. To address these limitations, we propose augmenting the generator with an ancilla qubit, expanding the learning space, and providing additional degrees of freedom that enable training to progress when the model becomes trapped in certain regions of the loss landscape. In this work, we investigate the effect of incorporating an ancilla under various connectivity topologies and at different stages of training, in order to perturb the optimization landscape and aid the generator overcome problematic training cases Simulation results demonstrate that ancilla-assisted QGANs successfully escape learning plateaus and other non-convergent behaviours, particularly when the ancilla’s connectivity links distant regions of the ansatz. Notably, the optimized fidelity overall improves when the ancilla is introduced mid-way through the training.