arXiv:2609.36702
Quantum Fidelity Landscape-Guided Prior Calibration for Single-Circuit QGAN Image Generation
Xue Yang, Rigui Zhou, Dax Enshan Koh, Siong Thye Goh, Yitao Tang, ShiZheng Jia, Young-Wook Cho, Hongyu Chen
Abstract
Quantum Generative Adversarial Networks (QGANs) have emerged as representative generative models in the Noisy Intermediate-Scale Quantum (NISQ) era and have attracted increasing attention in quantum machine learning. However, most existing QGAN methods rely on patch-based decomposition strategies, which weaken the global consistency of generated images and increase quantum resource overhead. In this work, we investigate a simpler approach: pixel-level, end-to-end image generation using a single-quantum-circuit QGAN. By analyzing the structural matching relationship between the quantum prior and the target data distribution in Hilbert space, we provide a new theoretical perspective for understanding the training behavior of naive end-to-end QGANs. Specifically, we introduce the Quantum Fidelity Landscape (QFL), defined as the pairwise-fidelity structure induced by an ensemble of quantum states and preserved under shared unitary transformations of the quantum generation process. We show that, under a fixed Lipschitz readout, this invariant imposes a one-sided bound on decoded sample separation, motivating calibration of the prior-induced QFL before adversarial training. To validate this theoretical insight, we propose BasicQGAN, a QGAN framework incorporating quantum prior calibration. Before adversarial optimization, BasicQGAN aligns the prior-induced QFL with the data-induced QFL. Experimental results on small-scale grayscale image datasets show that BasicQGAN achieves stable and effective end-to-end pixel-level image generation while requiring fewer qubits and trainable parameters than representative patch-based quantum generators. Furthermore, experiments with different initial quantum-state ensembles show that QFL-calibrated ensembles achieve better generative performance.
AI Summary
This paper asks how the quantum latent prior affects training of single-circuit QGANs that generate complete pixel images rather than independent patches. It defines the quantum fidelity landscape (QFL) as an ensemble’s pairwise-fidelity matrix and uses its invariance under a shared unitary to motivate prior calibration before adversarial training. The analysis combines unitary invariance, measurement contractivity and the Fuchs–van de Graaf inequality. For a fixed readout whose classical post-processing is $L$-Lipschitz in total variation, decoded pairs satisfy $d(x_i,x_j)\leq L\sqrt{1-F(\rho_i,\rho_j)}$. For amplitude-magnitude vectors, it obtains $\|a_i-a_j\|_2^2\leq2(1-\sqrt{F(\rho_i,\rho_j)})$. These are one-sided separation constraints, not guarantees of distribution recovery. BasicQGAN calibrates product-state rotation-angle ranges by matching the prior and amplitude-encoded data’s empirical fidelity distributions using KDE, mean and variance discrepancies, then trains a shared circuit with WGAN-GP. Noise-free simulations cover MNIST at $16\times16$ and $28\times28$, plus small letter and shape datasets. On $16\times16$ digit 0, calibrated-prior FID is 20.38, versus 73.93 and 72.64 for two uncalibrated broad priors and 33.97 for an extremely narrow prior. The $16\times16$ generator uses eight qubits and 320 trainable parameters. Reported comparisons show lower FID than patch-based PQWGAN at $16\times16$, but higher FID than the baselines at $28\times28$. Stated limitations include ideal simulation without hardware or noise validation, small grayscale benchmarks, and degraded higher-resolution performance. Histogram calibration does not establish full geometric matching or successful generation.
Error found: The paper says BasicQGAN and PQWGAN(Global) use identical generator circuits to isolate prior calibration, but its resource table assigns them different qubit and parameter counts at both image resolutions. The training pseudocode minimizes the generated samples' critic score, reversing the stated WGAN-GP generator objective, although it is unclear whether this error occurs in the experimental implementation.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.