Over-The-Air Computation in Correlated Channels

TL;DR

This paper introduces an analog over-the-air computation scheme for a broad class of functions, including linear and some nonlinear functions, with error bounds applicable to various fading and noise conditions, enhancing efficiency in distributed systems.

Contribution

It proposes a novel OTA computation method for nonlinear functions, providing nonasymptotic error guarantees under diverse channel conditions without assuming identical distributions.

Findings

Error bounds valid for fast-fading channels and sub-Gaussian noise.

Applicable to channels with correlations and bursty interference.

Potential to significantly reduce communication costs in distributed ML applications.

Abstract

Over-the-Air (OTA) computation is the problem of computing functions of distributed data without transmitting the entirety of the data to a central point. By avoiding such costly transmissions, OTA computation schemes can achieve a better-than-linear (depending on the function, often logarithmic or even constant) scaling of the communication cost as the number of transmitters grows. Among the most common functions computed OTA are linear functions such as weighted sums. In this work, we propose and analyze an analog OTA computation scheme for a class of functions that contains linear functions as well as some nonlinear functions such as $p$-norms of vectors. We prove error bound guarantees that are valid for fast-fading channels and all distributions of fading and noise contained in the class of sub-Gaussian distributions. This class includes Gaussian distributions, but also many other practically relevant cases such as Class A Middleton noise and fading with dominant line-of-sight components. In addition, there can be correlations in the fading and noise so that the presented results also apply to, for example, block fading channels and channels with bursty interference. We do not rely on any stochastic characterization of the distributed arguments of the OTA computed function; in particular, there is no assumption that these arguments are drawn from identical or independent probability distributions. Our analysis is nonasymptotic and therefore provides error bounds that are valid for a finite number of channel uses. OTA computation has a huge potential for reducing communication cost in applications such as Machine Learning (ML)-based distributed anomaly detection in large wireless sensor networks. We illustrate this potential through extensive numerical simulations.