![[2510.26722] Non-Convex Over-the-Air Heterogeneous Federated Learning: A Bias-Variance Trade-off](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2510.26722] Non-Convex Over-the-Air Heterogeneous Federated Learning: A Bias-Variance Trade-off
Summary
This paper explores the challenges of heterogeneous federated learning in wireless networks, focusing on the bias-variance trade-off in non-convex scenarios. It presents a novel algorithm that optimizes model updates to enhance convergence and generalization.
Why It Matters
As federated learning becomes increasingly important in distributed AI applications, understanding the impact of heterogeneous conditions on model performance is crucial. This research addresses a significant gap in existing methodologies, providing insights that could improve the efficiency and effectiveness of federated learning systems in real-world scenarios.
Key Takeaways
- The study addresses the limitations of existing OTA federated learning designs under heterogeneous wireless conditions.
- It introduces a structured bias in model updates to manage variance effectively.
- A novel successive convex approximation algorithm is proposed to optimize power control in federated learning.
- The research validates its approach through experiments on non-convex image classification tasks.
- Findings highlight the importance of balancing bias and variance for improved model convergence.
Computer Science > Machine Learning arXiv:2510.26722 (cs) [Submitted on 30 Oct 2025 (v1), last revised 13 Feb 2026 (this version, v4)] Title:Non-Convex Over-the-Air Heterogeneous Federated Learning: A Bias-Variance Trade-off Authors:Muhammad Faraz Ul Abrar, Nicolò Michelusi View a PDF of the paper titled Non-Convex Over-the-Air Heterogeneous Federated Learning: A Bias-Variance Trade-off, by Muhammad Faraz Ul Abrar and Nicol\`o Michelusi View PDF HTML (experimental) Abstract:Over-the-air (OTA) federated learning (FL) has been well recognized as a scalable paradigm that exploits the waveform superposition of the wireless multiple-access channel to aggregate model updates in a single use. Existing OTA-FL designs largely enforce zero-bias model updates by either assuming \emph{homogeneous} wireless conditions (equal path loss across devices) or forcing zero-bias updates to guarantee convergence. Under \emph{heterogeneous} wireless scenarios, however, such designs are constrained by the weakest device and inflate the update variance. Moreover, prior analyses of biased OTA-FL largely address convex objectives, while most modern AI models are highly non-convex. Motivated by these gaps, we study OTA-FL with stochastic gradient descent (SGD) for general smooth non-convex objectives under wireless heterogeneity. We develop novel OTA-FL SGD updates that allow a structured, time-invariant model bias while facilitating reduced variance updates. We derive a finite-time stationarity boun...
