Subject: Cutting-Edge Advancements in Signal Processing and Communication

Hi Elman,

This newsletter covers recent preprints exploring cutting-edge advancements in signal processing and communication, with a strong emphasis on applying machine learning techniques.

General Overview

This collection of preprints explores cutting-edge advancements in signal processing and communication, with a strong emphasis on applying machine learning techniques to diverse applications. Several papers focus on improving the performance and robustness of wireless communication systems, particularly in challenging scenarios. Bian et al. (2024) introduce LISAC, a learned coded waveform design for integrated sensing and communication (ISAC) with OFDM, employing RNNs for encoding and decoding to optimize both sensing and communication performance. Bayat et al. (2024) propose novel time and frequency synchronization techniques for multiuser uplink OTFS in high-mobility scenarios, utilizing a spectrally efficient pilot pattern and Chebyshev polynomial basis expansion. Hedhly et al. (2024) analyze the impact of fractional Doppler on OTFS-NOMA systems, considering users with different mobility profiles and highlighting the performance impact of inter-Doppler interference. Addressing the limitations of traditional hybrid precoding in massive MIMO, Guo (2024) proposes a meta-learning-driven approach for maximizing spectral efficiency, demonstrating robustness and exceeding the performance of fully digital precoding in some cases.

Another prominent theme is the application of deep learning to signal analysis and interpretation. Han (2024) explores arrhythmia classification using graph neural networks based on correlation matrices derived from extracted ECG features. Merkofer et al. (2024) introduce WAND, a wavelet analysis-based neural decomposition method for MRS signals, effectively removing artifacts and improving metabolite quantification. Zhang et al. (2024) propose FSOS-AMC, a few-shot open-set learning framework for automatic modulation classification, leveraging a multi-scale attention network and meta-prototype training. For human activity recognition, Bi et al. (2024) provide a comprehensive overview of FMCW radar principles and deep learning techniques, including CNNs, LSTMs, and Transformers. Abbott et al. (2024) investigate data augmentation for cardiac auscultation signals using generative deep learning and signal processing, demonstrating improved model robustness through synthetic audio generation.

Several contributions focus on specific applications and challenges in sensing and communication. Tholeti et al. (2024) present online waveform selection algorithms for cognitive radar, utilizing reinforcement learning to adapt bandwidth based on feedback from ballistic missile trajectories. Afriat et al. (2024) develop a novel estimator for source localization of an unknown transmission in dense multipath environments, deriving the CRLB and proposing a practical optimization algorithm. Wang et al. (2024) propose SACGNN, a heterogeneous graph neural network for cooperative ISAC beamforming in cell-free MIMO systems, demonstrating performance gains over conventional schemes. Addressing synchronization challenges in ISAC, Wang et al. (2024) explore optimizing fingerprint-spectrum-based synchronization by deriving a near-optimal window function and proposing a practical selection criterion.

The application of near-field communications is also explored. Zhang et al. (2024) propose a meta-learning framework for adaptive codebook design in near-field communications, demonstrating improved performance and generalization. Zhao et al. (2024) investigate designing unimodular waveforms with good correlation properties for large-scale MIMO radar using manifold optimization. Qiu et al. (2024) address the near-field beam split issue in IRS-aided THz communications, proposing a double-layer delta-delay IRS beamforming scheme. Droulias et al. (2024) explore beam bending for 6G near-field communications, introducing near-field virtual routing and demonstrating its potential for dynamic routing and blockage avoidance. Jiang et al. (2024) propose a subarray decomposition scheme for efficient near-field channel characteristics analysis in large-scale MIMO.

Finally, several papers address broader applications of signal processing and machine learning. Numerous other preprints contribute to areas such as data-independent KLT approximations, crowd size estimation using mmWave radar, benchmarking LLMs in sensor processing, biointerfaces for separating neural drives, near-pilotless MIMO communications, V2X communication standards, multimodal human sensing, semantic communication, thermal management of base stations, PRACH jamming in 5G networks, real-time inference with two-way delay, social learning, computational wave imaging, optimizing RIS impairments, covert communications, and approximate computing in OFDM-based radar processors.

Paper Highlights

Bending Beams for 6G Near-Field Communications

Bending beams for 6G near-field communications by Sotiris Droulias, Giorgos Stratidakis, Angeliki Alexiou https://arxiv.org/abs/2410.08099

Caption: Bending Beam Propagation: (a) Ray trajectories and caustic formation. (b) Corresponding wave intensity distribution.

This research envisions future wireless systems that go beyond simply faster data transmission. The authors explore the innovative concept of "bending beams" for near-field communication in 6G, offering solutions for dynamic blockage avoidance, interference management, and connecting users on curved paths. This approach departs from traditional beamforming, which relies on straight-line propagation, by tailoring beams to follow virtually any trajectory.

The underlying principle is wavefront engineering, manipulating the phase and amplitude of radiated waves to achieve the desired curvature. The beam's trajectory is determined by the phase profile at the input plane, governed by the equation: $\frac{d\phi(x)}{dx} = k \frac{df(z_c)/dz_c}{\sqrt{1+(df(z_c)/dz_c)^2}}$, where $\phi(x)$ is the input phase, k is the wavenumber, and $f(z_c)$ describes the desired trajectory. Airy beams, a type of self-accelerating wave, provide a practical solution, propagating along parabolic paths and even demonstrating autofocusing capabilities. The paper also outlines designing beams for arbitrary trajectories using ray optics, where the beam's path is defined by the caustic of its equivalent rays.

The authors address practical implementation with large antenna arrays, considering footprint size, shape, inter-element spacing, available phase levels, and operating frequency. The footprint size, controlled by the aperture size, determines the maximum propagation distance along the curved trajectory ($z_{max} = \frac{L_x + \beta z_0^2 + x_0}{\beta}$). The footprint shape allows tailoring the power distribution along the beam path. Importantly, bending beams are resilient to blockage, often outperforming conventional beamforming, and are immune to beam split effects in wideband applications.

The concept of near-field virtual routing (NFVR) is introduced, where a single beam serves a user's curved trajectory, ensuring uninterrupted connectivity regardless of position and speed. This eliminates complex tracking and multiple beams required by traditional methods. Bending beams also create interference-free regions with curved boundaries, enhancing interference management. Applications in wireless power transfer and energy-efficient connectivity leverage autofocusing to concentrate power.

The results demonstrate promise in various applications. In dynamic blockage scenarios, multiple beams with different trajectories can reach the same receiver, enabling seamless switching based on blocker positions. The impact of practical limitations, like limited active elements and quantized phase levels, is quantified, showing high efficiency even with constraints. The robustness of bending beams across a wide frequency range highlights their suitability for wideband applications. This work opens a new paradigm in wireless communications, dynamically shaping the environment to optimize connectivity and enable new functionalities.

Exploiting Moving Arrays for Near-Field Sensing

Exploiting Moving Arrays for Near-Field Sensing by Yilong Chen, Zixiang Ren, Xianghao Yu, Lei Liu, Jie Xu https://arxiv.org/abs/2410.09358

Caption: Comparison of CRB for moving and fixed arrays under isotropic and SEM transmissions, demonstrating the superior performance of moving arrays in near-field target localization.

This paper presents a novel approach to near-field MIMO sensing using moving arrays, providing a cost-effective alternative to extremely large antenna arrays (ELAAs). The system consists of a base station (BS) with a uniform linear array (ULA) on a moving platform. By using the near-field channel characteristics created by the array's movement, the system locates a point target in 2D space with a limited number of antenna elements, contrasting with traditional fixed arrays relying on far-field assumptions and often needing massive antenna deployments for near-field sensing.

The analysis centers on deriving the Cramér-Rao bound (CRB) for estimating the target's 2D coordinates. A key finding is the CRB's dependence on the transmit signal waveform over time for moving arrays, unlike fixed arrays where it depends solely on the sample covariance matrix. This emphasizes waveform design's importance in moving array-based sensing. The derived CRB for the moving array is proportional to that of an equivalent ELAA matching the platform's size, revealing that movement effectively expands the aperture, enabling near-field sensing of distant targets infeasible for conventional fixed arrays under far-field assumptions. The CRB for the moving array is:

CRB({sₗ}) = σ²/2|b|² (Gᵧᵧ + Gₓₓ)/(GₓₓGᵧᵧ - ℜ{Gₓᵧ}²)

where Gpq = Σₗ∊L tr(AᴴpₗAᵩₗsₗsᴴₗ) - (Σₗ∊L tr(AᴴₗAₚₗsₗsᴴₗ) Σₗ∊L tr(AᴴₗAᵩₗsₗsᴴₗ))/Σₗ∊L tr(AᴴₗAₗsₗsᴴₗ), for p, q ∈ {x,y}.

Numerical results confirm the theoretical analysis, showing the moving array significantly outperforms the conventional fixed array in near-field target localization, achieving CRB improvements of over four orders of magnitude. Under both isotropic and strongest eigen-mode (SEM) transmissions, the moving array excels, with SEM further enhancing accuracy. Increasing the number of symbols (L) or platform size effectively reduces the CRB for both moving and extended fixed arrays, with minimal impact on the conventional fixed array. Practical localization performance using a maximum likelihood estimator demonstrates the moving array's accurate target location with a clear peak in the log-likelihood function, unlike the ambiguous performance of the conventional fixed array.

This work introduces a promising direction for near-field sensing using moving arrays. Achieving near-field performance with limited antennas offers advantages in cost and complexity compared to ELAA-based solutions. The findings highlight waveform design's importance in optimizing moving array systems and pave the way for future research in this area.

Meta-Learning for Hybrid Precoding in Millimeter Wave MIMO System

Meta-Learning for Hybrid Precoding in Millimeter Wave MIMO System by Yifan Guo https://arxiv.org/abs/2410.09427

Caption: The Gradient-Guided Meta-Learning (GGML) framework iteratively optimizes analog (VRF) and digital (VD) precoders using gradient information as network input. Within each iteration, distinct Machine Learning (ML) blocks refine VRF and VD through circulation involving neural networks (NN), gradient updates (ΔV), and constraint enforcement (Retraction/Power Constraint). This process efficiently calculates and minimizes the loss function, defined as the negative spectral efficiency (R), to maximize overall system performance.

Hybrid precoding in mmWave MIMO systems addresses the high energy consumption of fully digital precoding by using fewer RF chains. However, the coupled analog and digital precoders and the constant modulus constraint of phase shifters create design challenges. Traditional optimization algorithms often have high computational complexity or suboptimal performance, while deep learning solutions lack scalability and robustness. This paper introduces a novel gradient-guided meta-learning (GGML) framework for hybrid precoding, a plug-and-play solution without pre-training.

Unlike conventional deep learning, GGML uses gradient information as network input, facilitating gradient information flow sharing and enabling system adaptation to changes. Instead of directly outputting the precoding matrix, GGML outputs gradient information for iteratively optimizing both precoders. The digital precoding network (DPN) optimizes VD, while the analog precoding network (APN) optimizes VRF. The loss function is the negative spectral efficiency (SE): L = -R(VRF, VD). GGML integrates manifold optimization (MO) for the constant modulus constraint but removes the computationally expensive gradient orthogonal projection, using a retraction function to map updated points back to the manifold.

Simulations using a Saleh-Valenzuela channel model show GGML's rapid convergence and superior performance over existing methods like WMMSE-MO, MM-AltMin, Element-AltMax, and CNN-based approaches. At 10dB SNR, with 64 transmit antennas, 4 RF chains, and 4 users, GGML improves performance by 2.1% over WMMSE-MO. It also demonstrates robustness to variations in system parameters like antenna number and RF chains. Remarkably, with increasing RF chains, GGML surpasses fully digital WMMSE precoding.

GGML's superior performance stems from effectively capturing correlations between coupled variables through meta-learning and efficiently handling the constant modulus constraint via the modified MO approach. By learning from gradient information and adapting to the environment, GGML offers a promising solution for hybrid precoding in mmWave MIMO systems, paving the way for more efficient and robust future wireless networks.

Conclusion

This newsletter highlights significant advancements in signal processing and communication, particularly focusing on the innovative application of machine learning techniques. From revolutionizing near-field communications with bending beams and moving arrays to enhancing the efficiency of hybrid precoding in mmWave MIMO systems using meta-learning, these papers present impactful contributions. The themes of enhanced performance, robustness, and adaptability to dynamic environments resonate throughout the selected works. The exploration of near-field applications, including beam bending and moving arrays, points towards a future where wireless systems can overcome traditional limitations and offer new functionalities. The development of sophisticated learning algorithms, such as the gradient-guided meta-learning framework for hybrid precoding, demonstrates the potential for machine learning to optimize complex systems and achieve unprecedented performance gains. These advancements collectively contribute to the ongoing evolution of communication and sensing technologies, promising more efficient, robust, and versatile systems for the future.