PAS-QFL: Personalized Ansatz Selection for Quantum Federated Learning under Client Data Heterogeneity

PAS-QFL (Personalized Ansatz Selection for Quantum Federated Learning) is a framework designed to address client data heterogeneity by decomposing each client’s Quantum Neural Network (QNN) into a globally shared ansatz and a client-specific private ansatz. Unlike traditional QFL methods that force all clients to use a fixed, monolithic ansatz structure, PAS-QFL personalizes the structure of the private layer to adapt to local data characteristics, rather than just its parameters.

The framework operates by selecting the shared ansatz using a stability-aware criterion based on Macro-F1 to ensure reliable cross-client aggregation, while each client independently selects its private ansatz based on local validation performance. During training, clients update both shared and private parameters locally but upload only the shared parameters to the central server, preserving the structural consistency required for federated aggregation while allowing for personalized decision heads.

Experiments demonstrate that PAS-QFL improves average Macro-F1 over fixed-ansatz baselines, particularly in scenarios with class-imbalanced non-IID data. By allowing different clients to favor different ansatz structures, the method mitigates the instability and unfairness associated with forcing a single ansatz on heterogeneous clients, providing stronger local adaptation without compromising global model coherence.

PAS-QFL addresses a practical limitation in quantum federated learning (QFL): most existing approaches assume that all participating clients can use a single, globally shared quantum neural network (QNN) ansatz, even though real federated settings often involve highly heterogeneous and non-IID local data. The paper argues that this assumption is especially problematic under class-imbalanced distributions, where different clients may have different local label frequencies, feature scales, or task biases. In such regimes, a one-size-fits-all ansatz can be poorly matched to some clients’ data, leading to slower convergence, higher local loss, or degraded global generalization.

The key contribution is a personalized ansatz selection framework that treats the QNN architecture itself as a client-specific design variable rather than a fixed global hyperparameter. PAS-QFL enables each quantum client to select or adapt an ansatz structure that better matches its local data distribution, while still participating in a federated training process that preserves data privacy. The central insight is that ansatz suitability is data-dependent: under non-IID, class-imbalanced conditions, clients may benefit from different circuit structures, parameterizations, or inductive biases. By aligning the ansatz with local data characteristics, the method aims to reduce negative transfer, improve local training efficiency, and produce a more robust federated QNN.

This work matters because it moves QFL closer to realistic deployment scenarios, where quantum clients may be distributed edge devices, sensors, or heterogeneous hardware with non-identical data distributions. Rather than focusing only on parameter aggregation, PAS-QFL highlights architectural personalization as a critical dimension of quantum federated optimization. It is relevant not only to QFL, but also to broader questions in quantum machine learning about how to choose expressive, efficient, and data-appropriate ansatzes in distributed and privacy-preserving settings.