How does federated quantum learning work when every device runs a different circuit?

Quantum federated learning promises to train compact quantum neural networks on private data spread across many devices. The problem is that these devices rarely share the same hardware capabilities.

A single shared architecture cannot fit every client. Some processors support more qubits, others allow deeper circuits, and noise profiles differ wildly. Earlier personalized approaches let each client search its own architecture through quantum neural architecture search. That solved the hardware mismatch but created a new one: averaging parameters across structurally different circuits mixes operations that have no semantic correspondence.

Why parameter averaging breaks across different circuits

In classical federated learning, FedAvg works because every client trains an identical model architecture. The weights at layer three mean the same thing on every device. Quantum circuits do not guarantee that correspondence. A rotation gate on qubit two in a five-qubit circuit performs a different logical operation than a rotation gate on qubit two in a seven-qubit circuit. Even when qubit counts match, differing entanglement patterns change the meaning of each parameter.

Personalized quantum neural architecture search lets each client pick a circuit that fits its device. One client might use a hardware-efficient ansatz with linear entanglement. Another might choose a problem-inspired ansatz with all-to-all connectivity. The search space includes gate types and placement. Measurement strategies add another dimension. The resulting architectures share no layer-to-layer mapping. Averaging their parameters produces a meaningless hybrid.

The paper states this mixing combines semantically inconsistent circuit operations. No alignment procedure exists for parameters across heterogeneous quantum architectures. The authors argue that prototype sharing bypasses the need for any such alignment. This is the core claim. The method never aggregates model parameters across clients.

Prototype sharing as a semantic bridge

A prototype is the mean latent representation of training samples for a class. The client computes these vectors from its trained network. Since the latent space is defined by the client's architecture, prototypes live in a space only that client understands. Class labels are shared across the federation.

The server receives prototypes per class from each client. It averages same-class prototypes across clients. This yields a global prototype per class. Global prototypes carry no architectural information; they are high-dimensional vectors. The server broadcasts them to all clients.

Each client refines local prototypes by pulling them toward the global ones. The paper calls this using global prototypes as federated semantic anchors. The local network trains with a loss term encouraging latent representations to match refined prototypes. The client's architecture remains unchanged. No parameters leave the device. Only class-wise summary vectors traverse the network.

What the experiments measure and what they do not

The authors report a 3.70% accuracy improvement over FedAvg on their benchmark. They also claim enhanced class-consistent representation alignment. The comparison uses FedAvg as a baseline, which assumes identical architectures across clients. The experimental setup does not disclose how many clients participated, what quantum hardware was simulated, or whether real device noise models were used. The dataset names do not appear in the abstract. The number of communication rounds is not stated. The search space size for the architecture search is not given.

The method assumes that class-wise prototypes capture enough semantic information to guide training across architectural boundaries. It assumes the server can be trusted with prototype vectors. It assumes clients have sufficient local data to compute meaningful prototypes for every class. The paper does not discuss what happens when a client lacks samples for a particular class. It does not analyze communication cost relative to parameter-based federated learning. It does not address whether prototype vectors could leak information about the training data.

The submission date is 1 Oct 2026. The version is v1. No peer-reviewed publication is cited. The work remains a preprint.