Adaptive Aggregation with Two Gains in QFL

1. Problem Statement

Federated Learning (FL) deployed over quantum-enabled and heterogeneous classical networks faces significant performance degradation due to several unique challenges that traditional aggregation methods (like FedAvg) fail to address:

• Communication Heterogeneity: Quantum networks suffer from stochastic variations in teleportation fidelity, communication latency, and device instability.
• Geometric Mismatch: Many quantum models (e.g., Variational Quantum Circuits) utilize parameters that exist on non-Euclidean manifolds (e.g., periodic angles on tori or circles). Traditional Euclidean averaging assumes a flat parameter space, leading to "geometry-induced drift" and divergence when aggregating local models.
• Limitations of Prior Work: Existing methods like FedAvg ignore QoS and geometry; Riemannian FL ignores network latency/fidelity; and current Quantum FL research focuses on model development rather than robust aggregation strategies.

2. Methodology: A2G Framework

The authors propose A2G (Adaptive Aggregation with Two Gains), a dual-gain framework that jointly regulates the aggregation process using two distinct mechanisms:

A. QoS Gain ($\alpha$): Trust Weighting

This component modulates client importance based on physical-layer communication metrics. It calculates a trust score for each client $i$ at round $t$ using:

• Teleportation Fidelity ($F_{i,t}$): Quality of the quantum link.
• Latency ($\tau_{i,t}$): Communication delay.
• Instability ($\sigma^2_{i,t}$): Variance in channel performance.

The QoS factor is computed as:
$$q_{i,t} = \frac{F_{i,t}^\alpha}{(\tau_{i,t} + \epsilon)^\gamma (\sigma^2_{i,t} + \epsilon)^\delta}$$
This factor is combined with data size proportions to generate normalized trust weights $w^{(\alpha)}_{i,t}$. Clients with high fidelity and low latency/instability receive higher weights, effectively filtering out unreliable nodes.

B. Geometry Gain ($\beta$): Manifold Correction

This component addresses the non-Euclidean nature of quantum parameters. Instead of simple averaging, A2G uses Riemannian geometry (Logarithmic and Exponential maps) to aggregate models:

1. Tangent Vector: Local models are mapped to the tangent space of the global model using the Riemannian logarithm map ($\text{Log}_{\theta_t}$).
2. Weighted Sum: A QoS-weighted tangent vector $v_t$ is computed.
3. Correction Term: The manifold correction $\Psi$ is defined as $\text{Exp}_{\theta_t}(v_t) - \theta_t$.
4. Interpolation: The geometry gain $\beta \in [0, 1]$ interpolates between purely Euclidean aggregation ($\beta=0$) and full Riemannian averaging ($\beta=1$).

C. The Update Rule

The global model update rule combines these elements:
$$\theta_{t+1} = \theta_t - \eta \sum w^{(\alpha)}_{i,t} g_{i,t} + \beta \Psi(\theta_t, \{\theta_{i,t}\})$$
Note: In the paper's experiments, the server-side gradient step ($\eta$) is set to 0, relying entirely on the geometry-controlled interpolation of client models.

3. Key Contributions

1. Universal Dual-Gain Framework: A2G is the first FL mechanism to jointly incorporate QoS (fidelity/latency) and Model Geometry (curvature) into a single adaptive rule.
2. Unified Formulation: It recovers FedAvg, QoS-weighted averaging, and Riemannian aggregation as special cases, making it applicable to classical, quantum, and hybrid systems.
3. Convergence Guarantees: The authors provide a theoretical convergence analysis under smoothness, bounded-variance, and finite-QoS assumptions, showing that QoS gains accelerate convergence while geometry gains correct curvature bias.
4. Teleportation-Aware Trust: Formalizes a reliability mechanism specifically for noisy quantum channels, weighting clients by their actual transmission quality.
5. Stabilization of Noisy QNNs: Demonstrates that small geometry gains ($\beta \ll 1$) significantly suppress client drift and improve stability compared to traditional averaging.

4. Experimental Results

Experiments were conducted on a quantum-classical hybrid testbed using two medical imaging datasets (Breast Cancer Wisconsin and Breast-Lesions-USG) with non-IID data partitions.

• Geometry Gain Tuning:
  • Optimal $\beta$: A small geometry gain ($\beta = 0.05$) yielded the best results, achieving 68.25% accuracy on the Breast-Lesions-USG dataset.
  • Comparison: This was a 13.65% (approx. 25% relative) improvement over the standard FedAvg baseline ($\beta = 1.0$).
  • Trade-off: Aggressive geometry corrections ($\beta \geq 0.3$) led to volatile learning dynamics and lower accuracy, suggesting that gentle curvature correction is superior to full manifold averaging in noisy settings.
• Robustness to Noise:
  • A2G maintained high stability under varying teleportation noise levels (bit-flip probabilities $p \in \{0.01, 0.06, 0.12\}$).
  • Small $\beta$ values resulted in lower instability and latency spikes compared to the baseline.
• QoS Impact: The QoS weighting successfully down-weighted clients with poor fidelity or high latency, preventing them from degrading the global model.

5. Significance and Conclusion

The paper presents a critical advancement for Quantum Federated Learning (QFL) by addressing the dual challenges of network unreliability and geometric complexity.

• Practical Impact: A2G offers a communication-efficient, gradient-free aggregation strategy that avoids the high cost of server-side gradient computation in quantum circuits (which suffer from barren plateaus and shot noise).
• Theoretical Value: It provides the first convergence analysis for a unified framework that handles both QoS heterogeneity and Riemannian geometry.
• Future Outlook: A2G serves as a foundational mechanism for next-generation distributed quantum intelligence, enabling scalable and noise-resilient learning across emerging quantum-classical networks.

In summary, A2G-QFL proves that adapting aggregation rules to both the physical quality of the network and the mathematical geometry of the model is essential for the success of federated learning in quantum environments.