Federated learning facilitates collaborative model training across multiple clients while preserving data privacy. However, its performance is often constrained by limited communication resources, particularly in systems supporting a large number of clients. To address this challenge, integrating over-the-air computations into the training process has emerged as a promising solution to alleviate communication bottlenecks. The system significantly increases the number of clients it can support in each communication round by transmitting intermediate parameters via analog signals rather than digital ones. This improvement, however, comes at the cost of channel-induced distortions, such as fading and noise, which affect the aggregated global parameters. To elucidate these effects, this paper develops a theoretical framework to analyze the performance of over-the-air federated learning in large-scale client scenarios. Our analysis reveals three key advantages of scaling up the number of participating clients: (1) Enhanced Privacy: The mutual information between a client’s local gradient and the server’s aggregated gradient diminishes, effectively reducing privacy leakage. (2) Mitigation of Channel Fading: The channel hardening effect eliminates the impact of small-scale fading in the noisy global gradient. (3) Improved Convergence: Reduced thermal noise and gradient estimation errors benefit the convergence rate. These findings solidify over-the-air model training as a viable approach for federated learning in networks with a large number of clients. The theoretical insights are further substantiated through extensive experimental evaluations.

Leveraging over-the-air computations for model aggregation is an effective approach to cope with the communication bottleneck in federated edge learning. By exploiting the superposition properties of multi-access channels, this approach facilitates an integrated design of communication and computation, thereby enhancing system privacy while reducing implementation costs. However, the inherent electromagnetic interference in radio channels often exhibits heavy-tailed distributions, giving rise to exceptionally strong noise in globally aggregated gradients that can significantly deteriorate the training performance. To address this issue, we propose a novel gradient clipping method, termed Median Anchored Clipping (MAC), to combat the detrimental effects of heavy-tailed noise. We also derive analytical expressions for the convergence rate of model training with analog over-the-air federated learning under MAC, which quantitatively demonstrates the effect of MAC on training performance. Extensive experimental results show that the proposed MAC algorithm effectively mitigates the impact of heavy-tailed noise, hence substantially enhancing system robustness.

Federated learning enables multiple clients to collaboratively train a common model while concurrently preserving data privacy. However, its performance is often constrained by limited communication resources, especially when the system encounters a large number of clients. Under those circumstances, integrating over-the-air computations into the model training procedure is considered an effective approach to coping with the communication bottleneck. Specifically, by uploading each client's intermediate parameters via analog transmissions instead of digital ones, the system can dramatically extend the number of clients it simultaneously supports in each communication round. However, that is achieved at the expense of introducing channel distortions, particularly fading and noise, in the aggregated global parameter. To demystify these effects, the present paper develops a theoretical framework to analyze the performance of the over-the-air federated model training process. Our analysis unveils a three-fold benefit from system scaling up, i.e., as the number of participating clients increases: (i) the privacy leakage, quantified by the mutual information between each client's locally possessed gradient and the edge's globally aggregated one, substantially decreases, (ii) the impairment of small-scale fading disappears due to the channel hardening effect, and (iii) the convergence rate is enhanced as thermal noise and gradient estimation error can be reduced. To that end, it establishes over-the-air model training as a viable approach for implementing federated learning in scenarios with a large number of clients. We corroborate the theoretical findings with extensive experiments.

Computing and minimizing the worst-case bound on the expected shortfall risk of a portfolio given partial information on the distribution of the asset returns is an important problem in risk management. One such bound that been proposed is for the worst-case distribution that is “close” to a reference distribution where closeness in distance among distributions is measured using [Formula: see text]-divergence. In this paper, we advocate the use of such ambiguity sets with a tree structure on the univariate and bivariate marginal distributions. Such an approach has attractive modeling and computational properties. From a modeling perspective, this provides flexibility for risk management applications where there are many more choices for bivariate copulas in comparison with multivariate copulas. Bivariate copulas form the basis of the nested tree structure that is found in vine copulas. Because estimating a vine copula is fairly challenging, our approach provides robust bounds that are valid for the tree structure that is obtained by truncating the vine copula at the top level. The model also provides flexibility in tackling instances when the lower dimensional marginal information is inconsistent that might arise when multiple experts provide information. From a computational perspective, under the assumption of a tree structure on the bivariate marginals, we show that the worst-case expected shortfall is computable in polynomial time in the input size when the distributions are discrete. The corresponding distributionally robust portfolio optimization problem is also solvable in polynomial time. In contrast, under the assumption of independence, the expected shortfall is shown to be #P-hard to compute for discrete distributions. We provide numerical examples with simulated and real data to illustrate the quality of the worst-case bounds in risk management and portfolio optimization and compare it with alternate probabilistic models such as vine copulas and Markov tree distributions.