A Stable Aggregation Method for Quantum Federated Learning

Fig. 2 illustrates the angular aggregation problem using a single-parameter slice of the QNN parameter space. Although a full QNN model contains many rotation parameters, the one-dimensional example captures the essential issue. The ordinary Euclidean mean maps seam-adjacent client angles to $14.35^{\circ}$, whereas the circular mean remains near $176.4^{\circ}$, which is geometrically consistent with the client cluster around the $-180^{\circ}/+180^{\circ}$ seam.

To this end, we introduce the following three big ideas. We develop Adaptive Aggregation with two Gains (A2G) ¹¹1A preliminary version of A2G is presented in the International Conference on Quantum Communications, Networking, and Computing (QCNC 2026) [20]. as a novel framework that uses client reliability and update-control gains to compute a stable global model update in quantum federated learning.

$\circ$ QoS-weighted aggregation: QoS-weighted aggregation gives more importance to clients with more reliable updates. The client weight depends on data size, quantum fidelity, latency, and update stability. Reliable clients contribute more to the global model, while noisy or delayed clients contribute less.

$\circ$ Circular A2G: Circular A2G is an adaptive aggregation method for quantum model parameters that are angles. It uses circular geometry instead of ordinary Euclidean averaging. This avoids wrong updates when angles are close to the $-\pi/\pi$ wrap-around boundary.

$\circ$ MP-A2G: MP-A2G means midpoint-projected A2G. It first computes a geometry-aware aggregation direction and then applies a one-shot midpoint-based correction before updating the server model. This aims to reduce overly aggressive movement toward an unstable aggregate. We discuss details in Sec III later.

Importantly, the A2G circular update, midpoint-projected A2G update, and SCM-A2G update do not collapse toward the misleading Euclidean aggregate. Instead, starting from the current global point $\theta_{t}=-170^{\circ}$, they move conservatively toward the geometry-consistent region of the angular parameter space. This behaviour is essential: it shows that the proposed updates respect the circular topology of the quantum parameter manifold and avoid artificial averaging artifacts caused by treating angles as ordinary Euclidean scalars.

It should be noted that our idea of angular discrepancy has a direct hardware-level interpretation, as shown in Table I. In our IBM quantum hardware validation, each aggregated angle is encoded into a single-qubit $R_{y}(\theta)$ circuit and evaluated using the Pauli-$Z$ expectation value. The Euclidean mean yields a positive hardware measurement, indicating that it realizes a physically different quantum state from the geometry-consistent solutions. In contrast, the circular, midpoint-projected, and SCM-based updates produce negative IBM hardware measurements that closely match the corresponding simulator values. This demonstrates that the choice of aggregation geometry is not merely a mathematical detail; it directly determines the quantum state implemented on real hardware.

Therefore, geometry-aware consistent aggregation is both theoretically necessary and experimentally meaningful. By preserving the intrinsic angular structure of the parameter space, the proposed A2G and SCM-A2G updates avoid spurious Euclidean averaging, maintain consistency with the intended quantum evolution, and produce hardware-realized states that align with the true circular geometry of the model.

Circular aggregation corrects the angular seam artefact, but it still computes the client-supported direction mainly from the current global model. In noisy and heterogeneous QFL, this first direction may be affected by non-IID client updates, stochastic local QNN training, finite-shot noise, and QoS variation. A direction that appears suitable at the current point may become less reliable along the induced movement.

To address this issue, as shown in the Figure 3 midpoint-projected aggregation introduces an intermediate check. Before accepting the final server update, the server evaluates a midpoint along the candidate movement and recomputes the client-supported direction from that midpoint. This midpoint decision acts as a geometry-aware stability check: it tests whether the proposed movement remains meaningful after the server has begun to move.

The proposed self-consistent midpoint aggregation further strengthens this idea. Instead of applying only a one-shot midpoint correction, SCM-A2G accepts a server movement only when the movement is supported by its own midpoint. Thus, the next global model is not obtained by direct Euclidean averaging, nor merely by a circular client mean. It is obtained as a QoS-weighted, torus-aware, self-consistent movement from the current global model.

Our main contributions are as follows.

1. 1.

   We identify a key aggregation challenge in QFL: client reliability varies, and QNN parameters lie on periodic angular spaces. We show that Euclidean averaging can fail near angular seams and can lead to different quantum observables on IBM hardware.

2. 2.

   We propose SCM-A2G-QFL, a QoS-aware and geometry-aware aggregation method with midpoint self-consistency.

We show how FedAvg, QoS-weighted aggregation, circular A2G, MP-A2G, and SCM-A2G fit into one aggregation framework. We evaluate SCM-A2G-QFL using accuracy, validation performance, update norm, volatility, manifold dispersion, and SCM residual.

After computing the QoS-aware weights, the server constructs a geometry-aware candidate movement. In Euclidean FL, the server can directly average client parameter vectors. However, QNN parameters are often rotation angles and therefore satisfy the periodic equivalence $\theta\equiv\theta+2\pi$. Consequently, ordinary Euclidean subtraction can be misleading near the $-\pi/\pi$ angular seam. To preserve the torus geometry of the QNN parameter space, we define the local client direction using the wrapped angular difference:

where $\boldsymbol{\theta}_{t}$ is the current global QNN parameter vector and $\boldsymbol{\theta}_{i,t}$ is the locally trained parameter vector returned by client $i$ at round $t$.

The QoS-weighted global direction is then computed as

where $w_{i,t}$ is the normalized QoS trust weight from Eq. (3). The initial geometry-aware candidate movement is

where $\beta_{t}\in(0,1]$ is the geometry gain controlling the size of the server movement.

A one-shot circular A2G update would therefore be

This update respects angular periodicity and avoids Euclidean seam artefacts. However, the direction $\mathbf{v}_{t}$ is computed only from the current global point $\boldsymbol{\theta}_{t}$. In noisy and heterogeneous QFL, this initial direction may no longer be sufficiently supported once the server begins to move. This motivates the midpoint and self-consistency refinement introduced next.

The midpoint refinement is introduced to avoid accepting a server movement that is supported only at the current global point but becomes unreliable along the movement path. The one-shot A2G movement $\mathbf{u}_{t}^{(0)}=\beta_{t}\mathbf{v}_{t}$ is used only as the initial candidate. SCM then treats the server movement as a variable $\mathbf{u}$ and refines it until the movement is supported by its own midpoint. Given any candidate movement $\mathbf{u}$, SCM defines the induced midpoint as

From this midpoint, the server recomputes the wrapped client directions:

and forms the QoS-weighted midpoint-supported direction

The map $\boldsymbol{\psi}_{t}(\mathbf{u})$ may differ from the initial direction $\mathbf{v}_{t}$, because the client directions are evaluated from the midpoint induced by $\mathbf{u}$, rather than directly from $\boldsymbol{\theta}_{t}$.

Here, $\mathbf{u}$ denotes a generic candidate server movement. The final SCM-accepted movement is denoted by $\mathbf{u}_{t}^{\star}$, which is the candidate movement satisfying the midpoint fixed-point condition.

Thus, the accepted server movement is not merely the first QoS-weighted direction computed at $\boldsymbol{\theta}_{t}$. Instead, it is a movement whose own midpoint continues to support the accepted movement. The next global model is then updated as

In practice, $\mathbf{u}_{t}^{\star}$ is approximated by an under-relaxed fixed-point iteration. Starting from the geometry-aware candidate movement $\mathbf{u}_{t}^{(0)}=\beta_{t}\mathbf{v}_{t}$, SCM iterates

where $\omega\in(0,1]$ is the solver relaxation factor. If the midpoint recomputed direction disagrees with the current candidate movement, the iteration adjusts the movement through $\boldsymbol{\psi}_{t}(\mathbf{u}_{t}^{(r)})$. Hence, SCM changes the accepted movement according to the QoS-weighted client evidence observed from the midpoint, rather than accepting the one-shot direction.

The iteration stops when

or when a maximum number of SCM iterations is reached. The final iterate is then used as the accepted movement $\mathbf{u}_{t}^{\star}$.

We record the SCM residual as

which measures how closely the accepted movement satisfies the midpoint self-consistency condition.

Let the global objective be

where $F_{i}$ is the local objective of client $i$, and $\boldsymbol{\theta}\in\mathbb{T}^{d}$ denotes the torus-valued QNN parameter vector. For analysis, we work in a local tangent chart induced by the wrapped angular difference $\operatorname{wrap}_{[-\pi,\pi)}(\cdot)$.

The ideal centralized descent direction at round $t$ is $-\nabla F(\boldsymbol{\theta}_{t})$. SCM-A2G instead applies the accepted server movement

where

In the convergence analysis, $R_{\mathrm{SCM},t}$ denotes the corresponding descent-scale residual contribution defined through Assumption IV.5. Thus, $R_{\mathrm{SCM},t}$ measures how much the remaining SCM fixed-point error contributes to the one-step descent inequality.ndition.

The midpoint-supported direction can be decomposed as

where $\boldsymbol{\xi}_{l,t}$ denotes stochastic local optimization error, $\boldsymbol{\xi}_{g,t}$ denotes client heterogeneity error, $\boldsymbol{\xi}_{q,t}$ denotes QoS reweighting bias, and $\boldsymbol{\xi}_{m,t}$ denotes midpoint displacement error.

Under Assumptions 1–5, these terms are bounded in expectation by

Theorem IV.7 shows that SCM-A2G-QFL retains the standard non-convex FL convergence structure. The average squared gradient norm decreases at an $\mathcal{O}(1/T)$ rate up to a neighbourhood determined by stochastic local training, client heterogeneity, QoS-induced reweighting, geometry-induced movement error, and the SCM fixed-point residual. The residual term $\frac{1}{\beta T}\sum_{t=0}^{T-1}R_{\mathrm{SCM},t}^{2}$ is specific to the proposed self-consistent midpoint mechanism. It shows that the accuracy of the midpoint fixed-point solve directly affects the final convergence neighbourhood.

The result also explains the empirical stability observed in the experiments. The geometry gain $\beta$ controls how aggressively the server moves from the current global parameter $\boldsymbol{\theta}_{t}$ toward the QoS-weighted geometry-aware direction. Smaller accepted movements can reduce the midpoint displacement term $L^{2}\overline{U^{2}}$, but excessively small $\beta$ can increase the relative effect of the residual contribution $\varepsilon_{\mathrm{SCM}}^{2}/\beta$. Thus, $\beta$ represents a stability–progress trade-off rather than a purely stabilizing parameter.

Overall, SCM-A2G does not merely change the aggregation target. It controls the path by which the global QNN model moves from $\boldsymbol{\theta}_{t}$ to $\boldsymbol{\theta}_{t+1}$, requiring the accepted movement to remain consistent with the QoS-weighted direction evaluated at its own midpoint. This provides a theoretical explanation for the reduced late-round volatility observed in the experiments.

We use two binary classification datasets from substantially different domains. The Breast-Lesions-USG represents a medical ultrasound diagnosis task, where patient age and ultrasound descriptors are used to classify lesions as benign or malignant [23]. Invalid entries are excluded, categorical features are label-encoded, age is converted to numeric form, and principal component analysis (PCA) apply after standardization to obtain $d=2$ or $d=5$ QNN-compatible features.

The Bank Account Fraud (BAF) dataset serves as a benchmark for financial fraud detection. BAF was introduced at NeurIPS 2022 as a privacy-preserving, large-scale resource. It provides a realistic tabular benchmark for detecting fraud in bank account openings. [11]. The suite comprises six synthetic datasets generated from an anonymized real-world fraud detection environment and addresses practical challenges, including class imbalance, temporal dynamics, and controlled bias variants. Each dataset includes 32 input features and up to one million records. The binary target variable is fraud_bool. In the conducted experiments, missing values are removed, categorical variables are ordinal encoded, and numerical features are standardized. PCA is applied to obtain $d=4$ QNN-compatible features. For both datasets, train, validation, and test splits are performed prior to encoding, scaling, and PCA to prevent data leakage.

All experiments are implemented in Python using Qiskit, Qiskit Aer, and Qiskit Machine Learning. Qiskit Aer is used for simulator-based QNN execution. The main supporting libraries include NumPy, pandas, scikit-learn, and matplotlib. Fixed random seeds are used for dataset splitting, client partitioning, QNN parameter initialization, and stochastic optimizer behaviour wherever supported. The main experimental configuration is summarized in Table V. Dataset-specific adjustments are made when required by sample size, circuit dimension, and computational budget.

Prior to evaluating comprehensive federated QNN training, the angular component is first isolated. This is achieved by examining aggregation behavior through controlled one-dimensional stress tests. These scenarios represent coordinate-wise QNN rotation parameters on the torus and enable direct assessment of whether the server update remains geometrically consistent near angular seams i.e., the wrap-around boundary where $+\pi$ and $-\pi$ denote adjacent directions..

Table VI reports four controlled angular stress tests. Case A is a non-wrap control case, where the client angles are clustered away from the $-180^{\circ}/+180^{\circ}$ seam. Euclidean and circular aggregation therefore produce nearly identical targets, showing that the geometry-aware formulation does not distort benign angular configurations.

Cases B and D test wrap-around behaviour, where the client angles are geometrically concentrated near the seam. In these cases, Euclidean averaging produces seam-inconsistent targets, such as $0^{\circ}$ and $58.07^{\circ}$, while circular aggregation recovers the correct seam-consistent target. In contrast, MP-A2G and SCM-A2G do not directly jump to this target; they compute bounded server movements from the current global angle. Thus, values near $-171^{\circ}$ in Cases B and D reflect controlled relaxation rather than aggregation error.

This distinction is central: circular aggregation identifies the geometry-aware client target, whereas SCM-A2G determines the accepted next server state. The SCM update is constrained by the geometry gain and midpoint self-consistency, requiring the accepted movement to remain aligned with the QoS-weighted direction at its own midpoint. Fig. 5 confirms this fixed-point behaviour, as the SCM residual rapidly decreases to near zero in all cases. These tests therefore show that SCM-A2G preserves benign-case behaviour, avoids seam-induced Euclidean failures, and stabilizes the server update through midpoint fixed-point refinement.