General Overview

This collection of preprints explores emerging trends in signal processing, communication systems, and machine learning, with a particular focus on novel applications and optimization techniques. Several papers delve into the potential of intelligent surfaces for enhancing wireless communication. Chondrogiannis et al. (2025) Chondrogiannis et al. (2025) investigate optical reconfigurable intelligent surfaces (ORISs) for free-space optical (FSO) multiple access systems, demonstrating improved performance of ORIS-enabled non-orthogonal multiple access (NOMA) compared to orthogonal methods. Similarly, Jia et al. (2025) Jia et al. (2025) propose stacked intelligent metasurfaces (SIM) for near-field multiuser beamfocusing, highlighting the superior spatial gain achieved in near-field communication and the SIM's ability to suppress inter-user interference. Zhuo et al. (2025) Zhuo et al. (2025) address channel estimation in RIS-aided mmWave MU-MIMO systems, proposing a correlation-based three-stage strategy with reduced pilot overhead. These works collectively showcase the potential of intelligent surfaces for shaping and controlling electromagnetic waves to improve communication performance.

Another prominent theme is the application of machine learning techniques to enhance various aspects of signal processing and communication. Deka et al. (2025) Deka et al. (2025) provide a comprehensive review of deep unfolding techniques for next-generation wireless communication, covering applications in signal detection, channel estimation, and beamforming design. Hou et al. (2025) Hou et al. (2025) introduce NOMANet, a graph neural network (GNN)-based power allocation scheme for NOMA networks, demonstrating significant improvements in energy efficiency. Xie et al. (2025) Xie et al. (2025) also leverage GNNs for optimizing antenna placement and power allocation in pinching-antenna systems. Furthermore, Dabush and Routtenberg (2025) Dabush & Routtenberg (2025) explore efficient sampling allocation strategies for graph-filter-based signal recovery, proposing cost functions based on the Cramér-Rao bound and Bayesian MSE. These studies underscore the growing role of machine learning in optimizing complex communication systems.

Beyond wireless communication, several papers address diverse applications of signal processing. Levi et al. (2025) Levi et al. (2025) present a novel approach for estimating food intake quantity using inertial signals from smartwatches, demonstrating the feasibility of dietary monitoring with readily available wearable devices. Goerttler et al. (2025) Goerttler et al. (2025) focus on explainable sleep stage classification, retrieving filter spectra in CNNs to understand spectral processing of EEG channels. Malinas et al. (2025) Malinas et al. (2025) tackle high-dimensional sequential change detection, introducing the Normalized High-Dimensional Kullback-Leibler divergence (NHDKL) to analyze the asymptotic performance of change detection procedures. Parsay et al. (2025) Parsay et al. (2025) present a comparative study of WiFi, video, and visual inspection for gait disorder assessment based on a large-scale clinical trial.

Several innovative methodologies and theoretical contributions are also presented. Geiger et al. (2025) Geiger et al. (2025) propose a novel phenomenological model of equalization-enhanced phase noise, representing it as a time-varying all-pass filter. Emenonye et al. (2025) Emenonye et al. (2025) analyze geolocation with large LEO constellations using Fisher information, providing conditions for 9D localization. Lv et al. (2025) Lv et al. (2025) investigate beam training for pinching-antenna systems, proposing scalable codebook designs and a three-stage beam training scheme. Kim et al. (2025) Kim et al. (2025) introduce weighted-sum energy efficiency maximization in user-centric uplink cell-free massive MIMO, developing optimization algorithms based on the Dinkelbach and quadratic transforms. These contributions advance the theoretical understanding and practical implementation of various signal processing and communication techniques.

Finally, several papers explore specific system designs and applications. Firouzjaei et al. (2025) contribute three papers on UAV-enabled IoT networks, focusing on energy harvesting, relay support, and disaster response (Firouzjaei et al., 2025, Firouzjaei et al., 2025, Firouzjaei et al., 2025). Wang et al. (2025) Wang et al. (2025) propose a pre-equalization aided grant-free massive access scheme for mMIMO systems. Zhang et al. (2025) Zhang et al. (2025) investigate positioning-aided channel estimation for multi-LEO satellite downlink communications. These works demonstrate the potential of emerging technologies for addressing real-world challenges in communication and sensing. Overall, this collection of preprints highlights the ongoing innovation in signal processing and communication, driven by the increasing demand for efficient, reliable, and intelligent systems.

Paper Highlights

WiFi Holds Promise for Diagnosing Gait Disorders

Gait Disorder Assessment Based on a Large-Scale Clinical Trial: WiFi vs. Video vs. Doctor's Visual Inspection by Alireza Parsay, Mert Torun, Philip R. Delio, Yasamin Mostofi https://arxiv.org/abs/2502.05328

Caption: This illustration depicts the setup of the clinical trial where a subject walks between two laptops acting as WiFi transmitter (TX) and receiver (RX). The WiFi signals reflected off the subject are analyzed to assess gait disorders, achieving surprising accuracy in distinguishing between healthy and unhealthy gaits. This approach allows for remote and non-invasive monitoring of gait, potentially revolutionizing healthcare delivery for gait disorders.

Neurological gait disorders significantly impact the quality of life for a large population, especially among older adults. Accurate and timely diagnosis is essential for effective intervention and management. This research presents a significant advancement in gait disorder assessment through a year-long clinical trial involving WiFi sensing technology. The study, conducted in collaboration with the Neurology Associates of Santa Barbara, encompassed 114 subjects with diverse neurological conditions such as Parkinson's disease, Neuropathy, Post Stroke effects, Dementia, and Arthritis, representing a broad spectrum of gait abnormalities.

A key aspect of this study was the focus on generalizability. The researchers leveraged publicly available online videos of both healthy and disordered gaits to train their WiFi-based gait classification system. A novel video-to-RF pipeline was developed to convert these videos into synthetic RF data, mimicking the signal reflections captured by WiFi transceivers during real-world walking scenarios. This innovative approach addressed the challenge of acquiring large, diverse datasets of real-world RF gait data.

The clinical trial involved a controlled environment within the neurology center, where subjects walked between two laptops functioning as WiFi transmitter and receiver. The received WiFi signals were then processed and analyzed by the system. For comparison, a parallel vision-based gait assessment system was also developed and tested under the same conditions, providing a direct comparison between the two modalities. Remarkably, the WiFi-based system achieved a per-subject accuracy of 85.61% and a per-class accuracy of 85.47% in differentiating between healthy and unhealthy gaits. The study also analyzed the impact of disease severity, age, height, and weight on classification accuracy, finding, as expected, that accuracy improved with increasing disease severity, reaching 100% for severe cases.

To benchmark the performance of both the WiFi and vision-based systems, the researchers conducted a comprehensive survey among 70 neurologists. The neurologists were presented with videos of the same subjects and asked to provide diagnoses based solely on visual observation. Surprisingly, both the WiFi and video-based systems outperformed the neurologists in diagnostic accuracy. The neurologists achieved an average per-class accuracy of 73.66% for binary classification (healthy vs. unhealthy) and 30.89% when classifying specific underlying disorders. This finding underscores the potential of these technologies to enhance clinical diagnosis and improve access to healthcare, particularly in regions with limited access to specialists.

While the current research primarily focuses on binary gait classification (healthy vs. unhealthy), future work will explore the potential of WiFi for classifying specific gait disorders. Furthermore, integrating additional medical data, such as patient history and blood test results, with the RF sensing data is expected to further improve diagnostic accuracy and pave the way for automated smart health systems. This integrated approach holds promise for revolutionizing healthcare delivery, particularly in underserved communities, by providing a cost-effective and accessible means of assessing and monitoring gait disorders.

Hyper Compressed Fine-Tuning of Large Foundation Models

Hyper Compressed Fine-Tuning of Large Foundation Models with Quantum Inspired Adapters by Snehal Raj, Brian Coyle https://arxiv.org/abs/2502.06916

Caption: This image visualizes the construction of Quantum-Inspired Adapters for Parameter-Efficient Fine-Tuning (PEFT). The green blocks represent the adapter matrix ΔW_Q, built from compound matrices, which is multiplied with the original pre-trained weight matrix W (blue) to produce the adapted weights W_adapt. The increasing complexity of the green blocks from left to right demonstrates the use of higher-order compound matrices and multiple blocks, enabling greater expressiveness while maintaining parameter efficiency.*

Fine-tuning large pre-trained language and vision models, such as BERT and GPT-3, for specific tasks has become increasingly challenging due to the computational and storage demands associated with updating their vast number of parameters. Parameter-Efficient Fine-Tuning (PEFT) methods address this challenge by updating only a small subset of the model's parameters, typically through the use of adapter modules. This paper introduces a novel PEFT approach called Quantum-Inspired Adapters, inspired by Hamming-weight preserving quantum circuits from the field of quantum machine learning. These adapters offer a powerful combination of expressiveness and parameter efficiency by operating in a combinatorially large space while simultaneously preserving orthogonality in weight parameters.

The core idea behind Quantum-Inspired Adapters lies in constructing orthogonal adapters using compound matrices up to a certain order k. Given a base matrix A, the compound matrix C_k = A^(k) is formed by taking determinants of all k x k submatrices of A. These compound matrices, known for their orthogonality properties, are used to build a block-diagonal adapter matrix ΔW_Q. The adapter is applied multiplicatively: W_adapt = ΔW_QW_, where W** is the original pre-trained weight matrix. This multiplicative approach, coupled with the orthogonality of the compound matrices, helps preserve the spectral properties of W**, mitigating the risk of catastrophic forgetting, a common issue in fine-tuning where the model loses its pre-trained knowledge. The block-diagonal structure and the limited order k of the compound matrices ensure parameter efficiency.

The authors evaluated the performance of Quantum-Inspired Adapters on the GLUE and VTAB benchmarks, comparing them to established PEFT methods like LoRA (Low-Rank Adaptation), OFT (Orthogonal Fine-Tuning), and BOFT (Block-Orthogonal Fine-Tuning). On the GLUE benchmark, using a specific configuration of the adapter, the method achieved 99.2% of the performance of existing fine-tuning methods while achieving a remarkable 44x parameter compression compared to LoRA. This result demonstrates the ability of Quantum-Inspired Adapters to achieve near-optimal performance with significantly fewer trainable parameters. Compared to other orthogonal fine-tuning methods like OFT and BOFT, the adapters achieved 98% relative performance with a 25x parameter reduction. On a subset of the VTAB benchmark, the adapters also demonstrated competitive performance with a 13.6x parameter reduction compared to LoRA.

Ablation studies, designed to understand the contribution of individual components of the method, revealed that both orthogonality and the inclusion of the first-order compound matrix (equivalent to OFT) are crucial for achieving high performance. Replacing the determinant operation in compound matrix construction with other operations like 'max' or 'avg' on minors led to significantly worse results. The authors also explored increasing the parameter count by multiplying multiple compound adapters, which led to performance improvements. These findings suggest that Quantum-Inspired Adapters offer a promising avenue for efficient fine-tuning of large models, particularly in resource-constrained settings. The approach combines expressiveness with parameter efficiency, providing a potential alternative to existing PEFT techniques. Future work will focus on extending the method to more complex architectures and investigating potential quantum speedups for compound matrix operations.

Revolutionizing Optical Wireless Communication

Intelligent Reconfigurable Optical Wireless Ether by Hongwei Cui, Soung Chang Liew https://arxiv.org/abs/2502.06128

Caption: This diagram illustrates the Optical Wireless Ether (OWE) concept, showcasing a network of Ether Amplifiers (EAs) that extend and guide light propagation between a station and an Access Point (AP). The EAs amplify and retransmit signals, creating a virtual medium for optical wireless communication, with both Line-of-Sight (LOS) and reflection channels contributing to signal propagation. The dashed lines represent the wireless optical links, while the solid lines represent wired connections.

Optical Wireless Communication (OWC) holds immense potential for high-bandwidth, license-free communication. However, the inherent limitations of light propagation, such as its directionality and inability to penetrate obstacles, severely restrict its coverage area. This paper introduces a novel solution: Optical Wireless Ether (OWE), a reconfigurable fabric of interconnected ether amplifiers (EAs) designed to extend and guide light propagation, effectively creating a virtual medium for OWC. Unlike traditional relay networks, OWE operates at the analog level, utilizing an "instantaneous-propagate" (IP) mechanism that bypasses the latency associated with buffering and digital processing. The key innovation lies in the strategic placement and gain control of EAs, which amplify and retransmit signals without decoding, thereby shaping the propagation characteristics of the OWE. A critical challenge in OWE design is mitigating self-interference and preventing amplifier saturation due to feedback loops created by signal reflections and scattering within the environment.

The paper presents a rigorous theoretical analysis of OWE, establishing stability constraints to prevent amplifier saturation. A crucial condition for stability is that the spectral radius of the matrix HᵀG must be less than 1, where H represents the channel response matrix between EAs and G is the diagonal matrix of EA gains. This condition ensures the convergence of the iterative channel response, preventing runaway amplification and signal distortion. Based on this stability analysis, the paper formulates optimization problems for various application scenarios. In a single-basic-service-set (single-BSS) scenario, the objective is to maximize the signal-to-noise ratio (SNR) at the access point (AP) by optimizing EA gains while respecting the stability constraint. For multiple-BSS scenarios, the optimization problem extends to maximizing the sum of normalized SNRs across all BSSs while minimizing mutual interference.

Simulations and experiments were conducted to validate the theoretical analysis and demonstrate the feasibility of OWE. In single-BSS simulations, optimizing EA gains resulted in substantial SNR improvements compared to both disabling EAs and using maximum equal gain settings. For example, in certain scenarios, the SNR improvement reached an impressive 81.66 dB compared to the equal gain setting. In multiple-BSS simulations, the OWE successfully served multiple BSSs simultaneously with minimal mutual interference, demonstrating the effectiveness of the optimization framework in promoting fairness and efficient resource allocation. A prototype OWE consisting of four EAs was also developed and tested experimentally. The results confirmed that the optimal gain settings significantly improved both SNR (over 7 dB improvement compared to disabled EAs) and packet loss rate (up to tenfold reduction).

This research pioneers the concept and implementation of OWE, offering a transformative approach to extend the reach and enhance the performance of OWC. The theoretical framework and optimization strategies presented provide a solid foundation for future research and development in this promising area. The experimental validation confirms the practical viability of OWE, paving the way for its application in various scenarios demanding high-bandwidth, low-latency wireless communication. Future work will focus on extending the experimental validation to multiple-BSS scenarios and exploring more sophisticated EA designs and control algorithms to further enhance OWE performance and adaptability.

Conclusion

This newsletter highlights the exciting advancements happening at the intersection of signal processing, communication systems, and machine learning. From leveraging WiFi signals for gait disorder diagnosis to developing quantum-inspired algorithms for compressing large language models and creating a reconfigurable "ether" for optical wireless communication, the research covered in this newsletter showcases the innovative spirit driving the field forward. The Parsay et al. paper demonstrates the potential of readily available technologies like WiFi to revolutionize healthcare diagnostics, offering a non-invasive and cost-effective approach to gait assessment. The Raj and Coyle paper tackles the computational challenges of working with massive language models, drawing inspiration from quantum computing to develop highly efficient fine-tuning methods. Finally, the Cui and Liew paper introduces a paradigm shift in optical wireless communication with the concept of Optical Wireless Ether, offering a novel solution to extend coverage and enhance performance. These diverse research directions collectively point towards a future where intelligent systems seamlessly integrate with our lives, improving healthcare, enhancing communication, and pushing the boundaries of what's possible.