Optimization and Learning for Large-Scale MIMO-OFDM Wireless Systems
Theory, Algorithms, and Applications
Time: Fri 2022-12-02 14.00
Location: F3, Lindstedtsvägen 26 & 28, Stockholm
Subject area: Telecommunication Optimization and Systems Theory Electrical Engineering
Doctoral student: Shashi Kant , Nätverk och systemteknik, Ericsson AB
Opponent: Prof. Dr. Christoph Studer, ETH Zürich, Department of Information Technology Electrical Engineering
Supervisor: Professor Carlo Fischione, Nätverk och systemteknik, ACCESS Linnaeus Centre; Prof. Mats Bengtsson, Teknisk informationsvetenskap; Bo Göransson, Skolan för elektroteknik och datavetenskap (EECS); Gabor Fodor, Reglerteknik
The requirements for next-generation wireless communications networks, particularly fifth-generation (5G) and beyond, are driven by at least three broad use cases. These include enhanced mobile broadband services to support extremely high data rates in terms of network or per user in both uplink and downlink, massive machine-type communications to accommodate massive internet-of-things applications, and critical machine-type communications to handle mission-critical applications that require ultra-high reliability and low latency.
These new-generation wireless communication systems adopt orthogonal frequency division multiplexing (OFDM) with cyclic prefix and multiple antennas at the transmitter and receiver (MIMO). There are many attractive characteristics of OFDM, namely robustness to the adverse effects of time dispersion due to multipath fading, simplicity in equalization, and flexibility in supporting both low and high symbol rates---thereby supporting a variety of various quality-of-service requirements.
It has been known for a long time that OFDM has problems with high out-of-band emissions (OOBE) and high peak-to-average-power ratio (PAPR). The OOBE must be adequately suppressed since high OOBE causes significant interference in the adjacent channels. Furthermore, high PAPR typically requires expensive linear radio frequency (RF) transmitter components and consequently costly digital predistortion to manage and mitigate OOBE resulting from the distortion caused by RF components, e.g., power amplifiers. Additionally, there are practical 5G standard constraints, which necessitate using only data-carrying subcarriers for OOBE and PAPR reduction. Hence, it is of utmost importance to reduce OOBE and PAPR for MIMO-OFDM-based systems and mitigate/minimize the signal distortion at the receiver(s) to meet the new generation systems’ requirements encompassing various use cases.
In this thesis, we seek principled approaches to tackle the distortion-based OOBE and PAPR reduction problems. More specifically, we present optimization formulations for these well-known issues in large-scale MIMO-OFDM-based systems, such as 5G New Radio (NR), and future extensions thereof. Unfortunately, these problems cannot be solved via a general-purpose optimization solver since these off-the-shelf solvers typically employ interior-point-based methods, which have prohibitive complexity for state-of-the-art radio hardware systems. Hence, we propose large-scale optimization techniques to tackle these problems resulting in implementation-friendly algorithms. More concretely, we develop (near) optimal and computationally-efficient data-dependent solutions by proposing a type of three-operator alternating direction method of multipliers (ADMM) method that essentially employs a divide-and-conquer approach to solve the huge and cumbersome OOBE and PAPR reduction problems in large-scale MIMO-OFDM-based systems. Moreover, in the last part of the thesis, we also investigate the application of our proposed three-operator ADMM (TOP-ADMM) for federated learning (FL) over networks that capitalize on the potentially rich datasets generated at the physical layer and/or RF hardware of a base station located near an edge server.
In summary, this thesis develops principled, implementation-friendly, and standards-agnostic algorithms for distortion-based OOBE and PAPR reduction algorithms using first-order optimization algorithms, which provide insights into the trade-off between computational complexities and in-band and out-of-band performance. Furthermore, we develop a novel yet simple TOP-ADMM first-order algorithm suitable for tackling centralized and distributed optimization problems. Additionally, this thesis studies the feasibility of the TOP-ADMM algorithm for edge FL exploiting rich datasets available at the base station(s) besides (private) datasets at the users. Finally, this thesis may provide input to the systemization and implementation of large-scale MIMO-OFDM-based wireless communication systems.