Mainz 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
Q: Fachverband Quantenoptik und Photonik
Q 74: Quantum Information – Concepts and Methods
Q 74.5: Vortrag
Freitag, 6. März 2026, 12:00–12:15, P 10
On the Generalization Limits of Quantum Generative Adversarial Networks with Pure State Generators — Jasmin Frkatovic1, •Akash Malemath1, Ivan Kankeu1, Yannick Werner1,2, Matthias Tschöpe1, Vitor Fortes Rey1,2, Sungho Suh3, Paul Lukowicz1,2, Nikolaos Palaiodimopoulos1,2, and Maximilian Kiefer Emmanouilidis1,2 — 1RPTU Kaiserslautern-Landau — 2DFKI, Kaiserslautern — 3Korea University, Seoul
Quantum Generative Adversarial Networks (QGANs) have emerged as promising candidates for quantum-enhanced generative modelling, yet their practical capabilities remain insufficiently understood. In this work, we investigate the generalization performance of two state-of-the-art fully quantum GAN architectures, QuGAN and IQGAN, in image generation tasks. Using extensive numerical experiments on MNIST and CIFAR-10, we systematically show that both models fail to learn the underlying data distribution and instead converge to reproducing only the dominant average features of each class, even under multi-class training and increased circuit expressivity. To explain these empirical failures, we derive an analytic lower bound on the achievable fidelity of pure-state quantum generators. Using the Helstrom bound, we prove that any QGAN whose generator outputs a single pure quantum state cannot approximate high-rank data distributions beyond the fidelity associated with the dataset’s leading eigenvector. Our results highlight intrinsic expressivity bottlenecks in current QGAN designs and motivate the development of quantum generators capable of producing mixed-state outputs or incorporating non-linear mechanisms.
Keywords: Quantum Generative Models; Quantum Machine Learning; Quantum Generative Adversarial Networks; Image Generation