Subject: Cutting-Edge Research in Signal Processing and Machine Learning for Wireless Communications

Hi Elman,

This newsletter covers recent preprints exploring the intersection of signal processing and machine learning in wireless communications, sensing, and biomedical domains. We'll delve into key themes, highlighting innovative approaches to optimization, robustness, and efficiency.

General Overview

This collection of preprints showcases the growing influence of machine learning and advanced signal processing techniques across various applications. Several papers focus on improving the efficiency and robustness of wireless systems. Larsson and Michelusi (2025) (Larsson & Michelusi, 2025) propose a novel contraction mapping framework for analyzing decentralized gradient descent, offering tight convergence bounds for various scenarios. Focusing on the physical layer, Hussain et al. (2025) introduce a DFT-based beam alignment method for ultra-massive MIMO systems in the near-field, leveraging far-field codebooks and correlation interferometry. Hague and Felton (2025) present a Chebyshev approximation method for synthesizing equiripple MIMO beampatterns, facilitating waveform set design that closely approximates desired correlation matrices. Addressing parameter-free SAR imaging challenges, Yang et al. (2025) introduce Principal Component Maximization (PCM), a novel technique exploiting the low-dimensional structure of SAR signals for image formation without prior system knowledge.

Another prominent theme is applying machine learning to enhance system performance and address complex optimization problems. Tosi et al. (2025) explore cross-environment transfer learning for location-aided beam prediction in mmWave networks, demonstrating significant improvements in prediction accuracy across different environments. Moryakova et al. (2025) employ a nonlinear least-squares algorithm to optimize cascaded power amplifiers, minimizing total nonlinearities. Kokkinis et al. (2025) utilize deep reinforcement learning for dynamic video-haptic radio resource slicing in the Tactile Internet, optimizing resource allocation based on latency, packet loss, and synchronization. For joint communications and sensing, Zakeri et al. (2025) propose a dynamic precoding design framework using Lyapunov optimization, maximizing sensing SNR while meeting communication requirements. Hua et al. (2025) investigate SWIPT in cell-free massive MIMO using stacked intelligent metasurfaces, developing efficient algorithms for joint phase-shift configuration and power allocation.

Several contributions focus on specific signal processing techniques and their applications. Dzwonkowski et al. (2025) propose a multitaper approach for post-processing compact antenna responses measured in non-anechoic conditions, enabling line-of-sight response extraction. Dong et al. (2025) investigate joint ADS-B and 5G for hierarchical UAV networks, proposing a MEC-based optimization algorithm to enhance trajectory surveillance. Uchimura et al. (2025) analyze and optimize near-field beams with uniform linear arrays, deriving closed-form expressions for phase distributions to generate Bessel and curving beams. Liu et al. (2025) propose a dynamic array partitioning framework for ISAC systems, minimizing Bayesian Cramér-Rao bounds for DOA estimation. Addressing the trade-off between pilot and data transmission, Chakraborty and Sun (2025) design an optimal pilot scheduling policy for throughput maximization based on the Age of Channel State Information.

Beyond wireless communications, the preprints cover diverse applications, including biomedical signal processing and energy efficiency. Luo et al. (2025) propose a single sparse graph enhanced expectation propagation algorithm for uplink MIMO-SCMA, improving error rate performance. Savoini et al. (2025) develop a CNN-based system for early lameness detection in horses using IMU sensor data. Güvengir et al. (2025) compare machine learning models for classifying power quality events in transmission grids. Xue et al. (2025) apply survival analysis and deep learning to predict Li-ion battery remaining useful life. Dutta et al. (2025) explore the opportunities and challenges of IoT-based smart precision farming in soilless agriculture. Elkeshawy et al. (2025) propose a federated learning approach for device activity detection in mMTC networks. Finally, Vasques et al. (2025) analyze sensor data for improving energy efficiency in buildings.

Overall, these preprints emphasize optimization, robustness, and efficiency in developing practical solutions for complex real-world problems. The exploration of novel architectures and algorithms, such as contraction mappings, PCM, and multitaper methods, highlights continuous innovation. The focus on integrated sensing and communication (ISAC) underscores its growing importance in future wireless networks. Finally, applying these techniques to diverse areas, from biomedical signal processing to agricultural monitoring, demonstrates their broad potential impact.

Paper Highlights

Unified Analysis of Decentralized Gradient Descent

Unified Analysis of Decentralized Gradient Descent: a Contraction Mapping Framework by Erik G. Larsson, Nicolo Michelusi https://arxiv.org/abs/2503.14353

This paper presents a significant advancement in the analysis of decentralized optimization algorithms. Decentralized Gradient Descent (DGD) and diffusion are crucial for distributed machine learning and multi-agent systems. However, existing analytical tools often fall short in providing tight convergence bounds and intuitive explanations. This work introduces a novel framework based on the Mean Hessian Theorem (MHT) and contraction mappings, offering a simpler, more powerful, and insightful approach.

The MHT provides a linear approximation of the gradient difference between two points, simplifying the analysis of gradient-based methods. The framework leverages three contraction mapping theorems: the Contraction Convergence Theorem (C2T), the Concatenated C2T (C³T), and a new Noisy C³T (NC3T). The NC3T is particularly relevant for decentralized optimization, as it handles noise whose variance scales with the input.

This framework allows for the derivation of tight convergence bounds for DGD and diffusion under strongly convex and smooth objective functions with arbitrary undirected topologies. The bound takes the form ||x_t - x|| ≤ (1 – ημ)^t||x₀ - x*||*, where η is the step size, μ is the strong convexity parameter, and x** is the optimal solution. This holds for η ≤ 2/(L + μ), where L is the smoothness parameter. The analysis extends to scenarios with multiple local gradient updates, time-varying step sizes η_t, and various noise sources like stochastic gradients, communication noise, and random network topologies. For instance, with time-varying step sizes η_t ~ 1/t, the error vanishes as O(1/t). Importantly, the framework avoids the unrealistic assumption of uniformly bounded gradient noise variance.

This MHT+contraction framework offers tighter bounds, simpler proofs, and handles more realistic noise models compared to existing methods. Its modular approach, separating convergence speed and asymptotic error analysis, provides valuable insights into the algorithms' behavior. The analysis of DGD with random link failures demonstrates the framework's versatility in handling practical scenarios. While focusing on strongly convex, smooth objectives, and symmetric weight matrices, this work sets a strong foundation for future research in decentralized optimization, potentially enabling the analysis of more complex scenarios and algorithm variants.

Multi-Modal Self-Supervised Semantic Communication

Multi-Modal Self-Supervised Semantic Communication by Hang Zhao, Hongru Li, Dongfang Xu, Shenghui Song, Khaled B. Letaief https://arxiv.org/abs/2503.13940

Caption: This figure illustrates the two-stage training process for a multi-modal semantic communication system. Stage 1 depicts joint pre-training of semantic encoders using self-supervised learning with augmented data from different modalities (e.g., RGB and depth images). Stage 2 shows the supervised fine-tuning phase, where feature vectors from the pre-trained encoders are fed into a channel model, followed by a semantic decoder and an inference target, minimizing inference loss.

Semantic communication, a key enabler for 6G, aims to transmit meaning rather than raw data. While demonstrating significant potential, existing methods often overlook the substantial communication costs incurred during training, particularly in multi-modal scenarios. This paper introduces a novel multi-modal self-supervised learning approach to address this challenge, significantly reducing training overhead while maintaining or exceeding the performance of supervised methods.

The proposed two-stage approach begins with multi-modal self-supervised learning for task-agnostic feature extraction. Inspired by DeCUR, this stage employs a novel loss function, L_pre-train = L_cross + L_intra, to learn shared and unique information across modalities by decoupling cross-modal and intra-modal representations. This pre-training eliminates the need for labels, drastically reducing communication overhead. The second stage involves supervised fine-tuning of the transmitter and receiver using labeled data and a cross-entropy loss, L_fine. This dual-phase strategy effectively captures both modality-invariant and modality-specific features, crucial for accurate and efficient semantic communication.

Evaluated on the NYU Depth V2 dataset for a classification task using RGB and depth data, the proposed method demonstrates a substantial reduction in training-related communication overhead compared to supervised learning. It achieves faster convergence during fine-tuning and consistently higher test accuracy across various SNR levels in AWGN channels. Remarkably, even with only 50% of the label information, the self-supervised approach significantly outperforms the fully supervised baseline, showcasing its robustness and efficiency in data-scarce scenarios. Unlike some other self-supervised methods, the proposed approach maintains stable training dynamics, avoiding exploding gradient issues.

This research highlights the advantages of incorporating multi-modal self-supervised learning in semantic communication systems. By reducing the communication burden during training, this approach paves the way for more efficient and scalable edge inference, especially in dynamic wireless environments. The ability to learn effectively from limited labeled data further enhances the practicality and adaptability of this framework for real-world applications.

Conclusion

This newsletter highlighted two impactful papers that advance the state-of-the-art in wireless communication and signal processing. The first paper introduces a powerful and insightful framework for analyzing decentralized optimization algorithms, offering tighter convergence bounds and handling realistic noise models. This contribution has significant implications for distributed machine learning and multi-agent systems. The second paper tackles the challenge of efficient training in multi-modal semantic communication, leveraging self-supervised learning to drastically reduce communication overhead while maintaining or exceeding the performance of supervised methods. This approach is particularly relevant for resource-constrained edge devices and dynamic wireless environments, paving the way for more efficient and scalable semantic communication systems. Both papers showcase the innovative application of advanced techniques to address critical challenges in next-generation wireless systems, promising significant improvements in performance, efficiency, and robustness.