Subject: Cutting-Edge Advances in Wireless Communication Systems

Hi Elman,

This newsletter covers recent groundbreaking research in wireless communication systems, spanning innovative signal processing techniques, machine learning applications, and novel hardware designs.

General Overview

This collection of papers explores innovative approaches to enhance wireless communication systems, leveraging advanced signal processing, machine learning, and novel hardware designs. Several papers focus on optimizing resource allocation and performance in complex communication scenarios. Merluzzi and Filippou (2025) introduce the concept of goal-oriented spectrum sharing, proposing a cross-layer optimization framework to balance the needs of goal-oriented and data-oriented systems (Merluzzi & Filippou, 2025). Similarly, Zhong et al. (2025) address resource allocation in integrated communication-radar systems, maximizing transmission rate while maintaining sensing performance (Zhong et al., 2025). Elkeshawy, Fares, and Nafkha (2025) investigate lightweight learning for activity detection in cell-free massive MIMO, demonstrating high accuracy with reduced complexity (Elkeshawy et al., 2025). Another contribution by the same authors (Elkeshawy et al., 2025) further explores the robustness of this approach to impairments and fixed-point representation.

Another prominent theme is the exploration of novel antenna designs and their impact on communication performance. Zhao et al. (2025) investigate beamforming for systems with movable and rotatable antennas, proposing an optimization approach using the sequential quadratic programming algorithm (Zhao et al., 2025). Yang et al. (2025) introduce flexible cylindrical arrays, demonstrating significant performance gains compared to uniform cylindrical arrays (Yang et al., 2025). Furthering this line of research, Yang et al. (2025) propose flexible intelligent metasurfaces (FIMs), which combine element movement and passive beamforming for enhanced signal strength (Yang et al., 2025). Zhou et al. (2025) explore the use of rotatable antennas in integrated sensing and communication systems, demonstrating performance gains in both communication and sensing capabilities (Zhou et al., 2025).

Several papers focus on specific signal processing techniques and their applications. Dizani and Olfat (2025) analyze decode-and-forward relaying for single-carrier frequency-domain equalization (SC-FDE) systems (Dizani & Olfat, 2025). Xiao et al. (2025) propose a multipath component power delay profile based joint range and Doppler estimation method for AFDM-ISAC systems (Xiao et al., 2025). Zhang et al. (2025) develop a deep learning-based OTFS channel estimation and symbol detection scheme using a plug-and-play framework (Zhang et al., 2025). Zhang (2025) derives standard Heisenberg's uncertainty principles for Cohen's class time-frequency distribution with specific kernels (Zhang, 2025). He and Zhang (2025) introduce the symplectic Wigner distribution in the linear canonical transform domain, demonstrating its superior time-frequency energy concentration for LFM signals (He & Zhang, 2025).

Applications of machine learning and signal processing to specific domains are also explored. De Rus et al. (2025) use explainable AI to investigate elevation cues in HRTFs across multiple datasets (De Rus et al., 2025). Wang et al. (2025) present MobiVital, a self-supervised system for contactless respiration monitoring using UWB radar (Wang et al., 2025). Azeraf et al. (2025) explore real-time pollutant identification using optical micro-sensors and machine learning (Azeraf et al., 2025). Khaked et al. (2025) assess the robustness of deep learning HAR models against real-world variabilities (Khaked et al., 2025). Grimaldi et al. (2025) introduce topological dictionary learning for sparse representation of signals on cell complexes (Grimaldi et al., 2025).

Finally, several papers address emerging challenges and propose novel solutions in diverse areas. A range of topics are covered, including DSP-free coherent interconnect architecture, privacy-preserving distributed median consensus algorithms, discrete-valued signal estimation, EEG data compression, RIS-assisted joint sensing and communications, uncertainty quantification for fast NFC tag identification, Farrow filter structures for digital resampling, stacked intelligent metasurfaces modeling, stochastic geometry for sensing and communications, low-rank matrix regression, autonomous robotic radio source localization, online conformal prediction, risky fault-chain search methods, phasor-pursuit directional modulation, off-grid target parameter estimation, pinching antenna-assisted NOMA systems, mid-band propagation measurements, time-interleaved ADCs calibration, beyond diagonal RIS in cognitive radio networks, PPG peak detection, MIMO VLC systems, and more. You can find the full list of papers and authors with their respective arXiv links in the references section of this newsletter.

Paper Highlights

BioSerenity-E1: A Self-Supervised EEG Model for Medical Applications

BioSerenity-E1: a self-supervised EEG model for medical applications by Ruggero G. Bettinardi, Mohamed Rahmouni, Ulysse Gimenez https://arxiv.org/abs/2503.10362

Caption: This figure illustrates the two-phase self-supervised pretraining framework of BioSerenity-E1. The first phase, pre-trained tokenizer, uses a transformer-based VQ-VAE to reconstruct spectral projections of EEG data, learning a compressed latent representation. The second phase, masked token predictor, trains a twin deep transformer network to predict masked EEG segments, forcing the model to learn spatiotemporal dependencies within the EEG signal.

BioSerenity has introduced BioSerenity-E1, a groundbreaking self-supervised foundation model poised to transform clinical EEG analysis. The model tackles the inherent difficulties of manual EEG interpretation, a process known for being time-consuming, demanding specialized expertise, and often lacking consistent accessibility. Traditional automated methods struggle with the complexity and noise inherent in EEG signals, while deep learning approaches are often hampered by the limited availability of labeled medical data. BioSerenity-E1 leverages the power of self-supervised learning to overcome these limitations, learning generalizable representations from vast amounts of unlabeled EEG data. This approach significantly reduces the reliance on labeled data for downstream tasks, opening doors for more efficient and effective EEG analysis across various medical applications.

BioSerenity-E1's innovative two-phase self-supervised pretraining framework begins with spectral tokenization. A transformer-based VQ-VAE architecture is trained to reconstruct log-multitaper spectral projections of the EEG signal, effectively learning a compressed, discrete latent representation. This tokenizer utilizes multitaper spectral estimation to mitigate the impact of noise and artifacts, while the logarithmic scaling of the power spectrum enhances frequency differentiation within the clinically relevant range (1-45 Hz). The second phase involves masked token prediction, where a twin deep transformer network is trained to predict the latent representations of masked EEG segments. This process compels the model to learn complex spatiotemporal dependencies within the EEG signal, further refining its understanding of the underlying brain activity. The model's training was conducted on a combined dataset comprising 4000 hours of EEG data from both proprietary and public sources (TUH EEG Corpus).

The efficacy of BioSerenity-E1 was rigorously evaluated across three critical clinical tasks: seizure detection, normal/abnormal EEG classification, and multiclass pathology differentiation. For seizure detection, using the TUH-Seizure dataset, BioSerenity-E1 achieved an impressive AUROC of 0.926 ± 0.002 and a Sensitivity of 0.909 ± 0.035, surpassing the performance of other models. In normal/abnormal classification, the model attained an AUPRC of 0.970 ± 0.001 on proprietary data and 0.910 ± 0.002 on TUH-Abnormal. Furthermore, on a multiclass pathology differentiation task using unbalanced data, BioSerenity-E1 achieved a Weighted F1 score of 0.730 ± 0.001. These results underscore the model's capability to effectively capture and interpret complex EEG patterns across a range of clinical scenarios.

The practical utility of BioSerenity-E1 is further highlighted by its performance in low-data regimes. When trained on less than 10% of the available data, the model demonstrated significant improvements in AUPRC, ranging from +2% to 17% compared to other approaches. This finding emphasizes the potential of BioSerenity-E1 to address the critical challenge of data scarcity in clinical AI development, particularly for rare diseases or conditions where acquiring large labeled datasets is difficult. Future research will focus on expanding the training dataset to further enhance model robustness and exploring entropy regularization techniques to optimize codebook utilization.

CoDiPhy: A General Framework for Applying Denoising Diffusion Models to the Physical Layer of Wireless Communication Systems

CoDiPhy: A General Framework for Applying Denoising Diffusion Models to the Physical Layer of Wireless Communication Systems by Peyman Neshaastegaran, Ming Jian https://arxiv.org/abs/2503.10297

CoDiPhy presents a novel framework that leverages conditional denoising diffusion models to tackle a wide range of wireless physical layer problems. While generative models, including denoising diffusion models (DMs), have shown promise in wireless applications due to their ability to learn complex data distributions, they often rely on the limiting assumption of Gaussian signal models and are typically designed for specific problems, hindering broader applicability. CoDiPhy addresses these limitations by incorporating a conditional encoder as a guidance mechanism. This encoder maps problem-specific observations, denoted as x_p, to a high-dimensional latent space, effectively removing the constraint of Gaussian signal models and allowing the diffusion model to operate with any signal model.

The CoDiPhy architecture combines a conditional encoder, a time embedding layer, and a U-Net-based main neural network for noise prediction. Unlike conventional DMs, which typically predict the original signal, CoDiPhy utilizes a noise prediction neural network. This key adaptation allows CoDiPhy to serve as a versatile solution for various tasks, including detection, estimation, and predistortion. The training objective is to minimize the difference between the predicted noise and the true noise, ε ~ N(0, I), by optimizing a loss function derived from the evidence lower bound: L_θ = E_t,xp,x0 [||ε_θ(x_t, x_p, t) – ε||²].

The paper demonstrates CoDiPhy's adaptability through two compelling case studies. The first focuses on signal detection in an OFDM uplink scenario, where CoDiPhy achieves near-optimal performance, with a gap of less than 0.5 dB to the linear minimum mean squared error (LMMSE) receiver with perfect channel state information (CSI) at an uncoded bit-error-rate (BER) of 10^-2. The second case study addresses phase noise (PN) estimation in single-carrier systems. Here, CoDiPhy significantly outperforms the conventional pilot symbol assisted modulation (PSAM) estimation, achieving up to 6 dB better mean squared error (MSE) performance across a wide signal-to-noise ratio (SNR) range.

These results highlight CoDiPhy's potential as a powerful and versatile tool for various physical layer tasks. Its ability to handle non-Gaussian signal models, thanks to the conditional encoder, significantly broadens the application scope of diffusion models in wireless communications. Future research directions include evaluating CoDiPhy's robustness in out-of-distribution scenarios, comparing its performance with discriminative models, and exploring its application in non-traditional areas like beamforming with reconfigurable intelligent surfaces and signal detection in free-space optical communications.

MobiVital: Self-supervised Time-series Quality Estimation for Contactless Respiration Monitoring Using UWB Radar

MobiVital: Self-supervised Time-series Quality Estimation for Contactless Respiration Monitoring Using UWB Radar by Ziqi Wang, Derek Hua, Wenjun Jiang, Tianwei Xing, Xun Chen, Mani Srivastava https://arxiv.org/abs/2503.11064

Caption: This flowchart illustrates the MobiVital pipeline for extracting high-quality respiration waveforms from UWB radar data. The process involves amplitude and phase calculation, waveform inversion detection using a biology-informed algorithm, and signal quality scoring using a self-supervised autoregressive model. This ultimately allows for the selection of the best breath waveform from the initial UWB data matrix.

Respiration waveforms, providing insights beyond simple respiration rates, are increasingly recognized as essential biomarkers for diagnosing diseases, guiding rehabilitation, and monitoring athletic performance. While wearable chest bands offer accurate measurements, their intrusive nature necessitates the development of contactless alternatives. Ultra-wideband (UWB) radar has emerged as a promising contactless solution, but existing methods primarily focus on respiration rate estimation, often overlooking critical waveform quality issues like distortion and inversion. These issues stem from the challenge of extracting the optimal waveform from the UWB data matrix, where selecting the best candidate time series among multiple possibilities is difficult without ground truth data. MobiVital addresses this challenge by estimating signal quality and detecting inversions without relying on ground truth information.

MobiVital introduces a novel approach to enhance the quality of respiration waveforms derived from UWB radar data. For signal quality estimation, MobiVital employs a self-supervised autoregressive model trained on high-quality respiration data. Leveraging the generalization limitations of machine learning models, the model's prediction accuracy on candidate time series during deployment serves as a surrogate for signal quality, with higher accuracy indicating better quality. For inversion detection, MobiVital utilizes a biology-informed algorithm based on the natural "duty cycle" of respiration, calculating the ratio of average peak widths of the original and flipped signals to identify and correct inversions. This innovative approach addresses the challenge of waveform inversion, a common issue in UWB radar-based respiration monitoring.

To foster reproducible research, the authors have released a valuable 24-hour UWB radar vital signal dataset collected from 12 subjects, along with time-synchronized ground truth data from wearable sensors. This dataset encompasses both tripod-mounted and handheld sensor scenarios, offering diversity in respiration patterns and motion challenges. The performance of MobiVital was evaluated against three baseline methods: Variance, Signal-to-Noise Ratio (SNR) estimation, and Constant False Alarm Rate (CFAR). The correlation coefficient (r) between the extracted waveform and the ground truth served as the evaluation metric, calculated as:

r = Σ( x_i – x̄ )( y_i – ȳ ) / √[Σ( x_i – x̄ )² · Σ( y_i – ȳ )²]

where x represents the measured signal and y represents the ground truth signal. Results demonstrate that MobiVital substantially outperforms the baseline methods, achieving an average correlation coefficient of 0.819. This represents a 7% improvement over the closest baseline (SNR) and a remarkable 34% improvement over SNR without pre-inversion. The inversion detection algorithm proved crucial, doubling the performance of CFAR, Variance, and SNR. Furthermore, MobiVital's performance in a downstream task of respiration rate estimation also surpassed the baselines, achieving a lower respiration rate error using both signal processing and learning-based methods. The study highlights inter-user variability, suggesting the need for larger and more diverse datasets in future research. The authors envision MobiVital as a stepping stone towards real-time breath coaching systems on mobile platforms, offering personalized feedback and guidance for respiration training.

Conclusion

This newsletter highlights a convergence of advanced techniques across signal processing, machine learning, and hardware design to address critical challenges in wireless communication. BioSerenity-E1 demonstrates the potential of self-supervised learning to revolutionize medical applications, specifically EEG analysis, by overcoming the limitations of manual interpretation and data scarcity. CoDiPhy introduces a versatile framework for applying denoising diffusion models to various physical layer tasks, offering significant performance gains in signal detection and estimation. MobiVital tackles the complexities of contactless respiration monitoring using UWB radar, emphasizing the importance of waveform quality and providing a valuable open dataset for further research. These advancements collectively pave the way for more robust, efficient, and intelligent wireless communication systems across diverse domains.