Papers

Graph signal processing tasks that leverage spectral information typically assume access to the complete graph topology, which is often unavailable in practice. We propose a systematic framework for subgraph filter learning (SFL), where subgraph-supported operators approximate ambient graph filters under partial observations. We formulate SFL as a statistical learning problem in which optimal subgraph operators are inherently data-dependent. To address the difficulty of directly estimating such operators, we develop a subgraph filter algebra based on distance-aware Laplacian constructions, defining a structured and controllable class of filters for effective approximation. We further establish performance risk bounds under the least squares loss, quantifying how well the learned operator approximates the restricted ambient mapping. Experiments real-world datasets show that, for SFL tasks, the proposed algebraic models consistently outperform polynomial filters, distribution-agnostic operators, and direct numerical filter learning baselines that attempt to recover the underlying structure from data.

Reconfigurable intelligent surfaces (RISs) need measurement methodologies that quantify the three-dimensional scattering response of a programmed surface, not only the received power of one bistatic link. This paper presents an automated anechoic-chamber methodology for full-sphere characterization of a varactor-based RIS around 5 GHz. The developed platform combines a state-of-the-art spherical near-field scanner with a near-field-to-far-field transformation, which is suitable for antenna measurements, with a common, low-scattering mount for the RIS and the excitation antenna. It enables programming of several independent DC voltages for remote RIS beam steering configuration via a control unit integrated into the chamber, which features a multichannel digital-to-analog converter. The RIS states are synthesized from a pre-measured phase-voltage characteristic and evaluated against an unbiased-RIS reference using local angular power averages. Measurements at 5.3 GHz serve as a proof of operation: the workflow programs one-dimensional and two-dimensional anomalous-reflection states, localizes the resulting beams over the sphere, and extracts target-window gain, maximum-window gain, and pointing error from the same dataset. For five measured RIS beam-steering states, the prescribed target windows show relative beam-steering gains up to 8.8 dB, while the locally detected main-lobe windows show relative gains up to 14.0 dB with an average of 11.3 dB. The reported gain is not active amplification by the RIS; it is a reference-normalized local power ratio for a fixed measurement configuration.

Bilateral teleoperation systems that include joint flexibility better reflect real robotic systems used in surgery, space, and rehabilitation. However, joint flexibility together with time-varying communication delays makes it difficult to maintain stable and coordinated motion between the master and slave robots. To address this, we propose a hybrid control method that combines a stable Proportional-plus-Damping (P+d) controller with a model-free deep reinforcement learning agent based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. The P+d controller provides basic stability under bounded delays, while the learning agent adjusts and tunes the remote-side proportional and damping gains in real time to reduce vibrations and improve tracking. Stability is guaranteed for bounded time-varying delays using Lyapunov-Krasovskii analysis. The approach provides a practical solution for teleoperation systems facing both joint flexibility and uncertain network delays.

In this paper, we revisit the problem of an encrypted model predictive control (MPC) design, representing a significant challenge in the recent field of secure process control. Existing methods in secure optimization-based control are non-existent and even partial implementation fails to address the closed-loop system stability and recursive feasibility properties of the constrained MPC. To overcome these limitations, we propose a novel approach that utilizes a polynomial approximation of the optimal control law. This method evaluates the explicit control law within a fully homomorphic encryption framework, ensuring that the controller is securely deployed on any third-party or cloud-based platform, with both process data and controller coefficients protected. Experimental results from a laboratory-scale implementation and validation of the proposed privacy-aware control method demonstrate its advantages.

Objective: Brain-Computer Interfaces (BCIs) enable the control of external devices by decoding user intentions from electroencephalography (EEG). However, substantial EEG variability within and between users remains a major challenge. To better understand this variability, we propose interpretable metrics that independently quantify temporal, spatial, and frequency variability in BCI related brain activity within and between users. Methods: We propose a framework to quantify variability by extracting EEG features and defining variability as their dispersion around their centroid using appropriate distance functions. Using two motor imagery BCI datasets (N = 133 users), we investigated the relationship between BCI performance and the variability metrics through within-user and cross-user classification experiments. Results: Negative correlations of -0.2 to -0.4 were observed across most conditions, suggesting that lower variability is associated with higher BCI performance. Moreover, the metrics revealed differences in robustness to variability between the deep learning and Riemannian-based classifiers, with the former showing weaker correlations. Conclusion: The results demonstrate the effectiveness of the proposed variability metrics and suggest that reducing variability may improve BCI performance while revealing differences in the sensitivity of classification models to different types of variability. Significance: The framework quantifies temporal, spatial, and frequency variability at multiple hierarchical levels (within-trial, between-trial, and between-trial-group), providing interpretable measures to better understand EEG variability and support more robust BCIs. It could also be used to characterize dataset variability, evaluate classifier sensitivity, incorporate variability into objective functions, and provide variability-based user feedback.

In [1], the authors recently proposed and solved a surveillance-evasion differential game in which an agile pursuer (the prying pedestrian) seeks to remain within a given surveillance range of a less agile evader for as long as possible while the evader seeks to escape as quickly as possible. In this paper, we provide initial results that extend this game from the 1 versus 1 (1v1) setting to a 2 versus 1 (2v1) setting with two pursuers and one evader. By deriving and exploiting semi-explicit or geometric reinterpretations of the existing 1v1 results, we derive partial solutions to the 2v1 game for the case of static pursuers and for the case of an evader that is at least twice as fast as the pursuers. While the 2v1 results of this paper build on the 1v1 results of [1], a different solution approach is developed to avoid a coordinate transformation that reduces the 1v1 game to two dimensions but which is ineffective at simplifying the 2v1 game. Beyond enabling progress on the 2v1 game, our new approach yields new geometric interpretations of the optimal pursuer and evader strategies in the 1v1 game, and opens further possible extensions.

We present CS-P2L, a framework coupling covariance steering (CS) with the Pick-to-Learn (P2L) meta-algorithm for certified controller synthesis over high-fidelity stochastic simulators. The method iteratively evaluates policies on simulator rollouts, tightens surrogate constraints using the worst-case violations, and provides compression-based probabilistic guarantees on the true violation probability given a confidence level. On a spacecraft powered-descent problem with uncertain gravity, CS-P2L certifies a violation bound of 4.9% with 600 rollouts, whereas standalone covariance steering underestimates the violation rate by roughly a factor of two.

Probabilistic Ising machines (PIMs) based on probabilistic bits offer a hardware-friendly route to solve combinatorial optimization problems, but most digital implementations achieve high throughput by exploiting sparse interactions. This limits their applicability to dense problems, for which memory bandwidth and data movement become the dominant bottlenecks. Here, we show a resource-efficient pipelined Field-Programmable Gate Array architecture enabling high-throughput execution of fully-connected PIMs while maintaining scalability and modularity. This architecture design combines a deeply pipelined (>20 stages) probabilistic bit update path, which overlaps spin evaluation and local-field updates, with a bandwidth-aware on-chip memory organization for the coupling and bias matrices. The architecture supports 512 p-bits with 16-bit fixed-point coefficients and 1024 and 2048 p-bits with 10-bit and 2-bit coefficients, respectively, and operates at up to 300 MHz. At fixed degree of parallelization, it delivers an order-of-magnitude higher update rate than an optimized non-pipelined baseline, while improving the time-area trade-off for dense workloads. Validation on portfolio optimization and low-density parity-check decoding shows close agreement with software references and substantial reductions in time-to-solution relative to the non-pipelined design, establishing pipelining as an effective route to scalable digital probabilistic computing for dense optimization problems.

We present VibeVoice-ASR-BitNet, a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs. We apply heterogeneous quantization tailored to the computational characteristics of each stage: the VAE acoustic tokenizer uses full-pipeline INT8 quantization (I8_S) with kernel fusion and SIMD optimization, while the autoregressive language model adopts BitNet-style ternary weights (I2_S). To preserve accuracy under aggressive compression, we employ a progressive quantization-aware training strategy. For inference, we implement custom SIMD kernels and fused operators within the ggml framework targeting both ARM and x86 platforms, achieving real-time recognition (RTF < 1) on low-thread-count CPUs. VibeVoice-ASR-BitNet is 1.6–2.3x faster than Whisper.cpp at comparable model sizes (~1.6 GB), with only modest accuracy degradation compared to the FP16 baseline.

In this paper, we address supervisory control in discrete-event systems with event forcing. In particular, partial observation of the supervisor is considered. Unlike traditional supervisory control, which relies on event enablement and disablement, the forcing mechanism allows a supervisor to preempt unwanted transitions by actively triggering forcible events. Under partial observation, we face a new challenge where the supervisor's forcing decisions must be consistent across all indistinguishable strings meanwhile guarantee all violating transitions are excluded through preemption. We introduce the concept of forcing consistency as a necessary and sufficient condition for the existence of a supervisor that achieves a given specification. We prove that forcing consistency is strictly stronger than the notion of forcibility but, unlike the latter, is not closed under union. These results lay the theoretical foundation for supervisor synthesis in the presence of both forcing mechanisms and observation limitations.

As Integrated Sensing And Communication (ISAC) systems push toward higher sensing resolution, multi-band processing has emerged as a key enabler, with Frequency Range 3 (FR3) (7-24 GHz) standing out for its combination of wide bandwidth and favorable propagation. A common assumption underlying existing multi-band ISAC techniques is that targets behave as frequency-invariant point scatterers, enabling coherent combination of measurements across widely spaced subbands. However, this assumption does not hold over wide fractional bandwidths, since real objects exhibit frequency-dependent scattering mechanisms and migrating scattering centers. This paper provides the first systematic experimental characterization of the frequency anisotropy of everyday objects for ISAC, using channel measurements collected with a calibrated vector network analyzer over the 6-24 GHz band, across 10 objects and 120 viewpoints. We process bistatic channel impulse responses and synthetic aperture radar images to quantify multi-band coherence, extracting the real part of the cross-band correlation coefficient. Our results reveal complex, non-trivial coherence structures that vary substantially with object type and viewing angle, highlighting the need to account for frequency anisotropy in multi-band ISAC system design.

In this paper, we study the control co-design (CCD) synthesis problem for a class of systems with parabolic partial differential equation (PDE) dynamics. We first derive a sufficient stability condition for the PDE. By spatially discretizing the PDE and using the sufficient stability condition, we propose a computationally tractable approximate CCD problem. We solve the approximate CCD problem using a gradient-based method. Finally, we justify our proposed approach through an example.

Advances in AI-based voice conversion have enabled a wide range of media applications, including films, audiobooks, and games. However, most research and public benchmarks still focus on natural human speech, leaving designed vocalizations, such as monster growls and robotic voices, underexplored, partly due to the lack of publicly available resources. To address this gap, we introduce the Designed Vocalizations Dataset, constructed by curating diverse raw vocal sources, including speech and animal vocalizations, and applying professional vocal effects processing to produce corresponding effect modified variants. We further provide a standardized test set with explicit seen/unseen splits over source timbre groups and preset styles to assess generalization under controlled conditions. Finally, we report baseline benchmark results to support reproducible evaluation and future research. The dataset and demo samples are available at https://ncai-official.github.io/speech/publications/designed-vocalizations-dataset/.

Steerable catheters are the primary tool for cardiac electrophysiology (EP) procedures including radiofrequency ablation, where the tip must be positioned precisely at target tissue while maintaining controlled, stable contact. The central control problem is therefore not merely tip tracking and not merely force regulation; it is the regulation of catheter–tissue interaction dynamics. The interaction state must encode how the tip moves relative to tissue, how persistent friction and contact forces bias that motion, and how safety limits reshape what motion is physically allowable. Existing methods regulate these interaction dynamics through different mechanisms. Classical impedance control~\cite{hogan1985} shapes the tip port as a virtual mechanical impedance $Z(s) = M_d s^2 + D_d s + K_d$, providing passive compliance without an explicit contact model. Three complementary design requirements motivate the present formulation: (i)~an explicit force-related constraint, (ii)~compensation for steady error under persistent loading, and (iii)~a prediction model that can incorporate trajectory and actuator information.

Parkinson's disease (PD) manifests in motor impairments that are detectable through digitized spiral drawings. This study introduces an explainable framework for PD screening using a novel radial-sampling feature fusion approach. We transform 2D spiral images into 1D revolution signals via a systematic ray-sampling technique to extract three distinct revolutions. We integrate spatial metrics, such as inter-revolution spacing variability and RMS radial derivatives, with spectral descriptors derived from Fast Fourier Transform (FFT) analysis across low, mid, and high harmonic bands. A total of 20 features were utilized to train state-of-the-art machine learning models, including Support Vector Machines (SVMs), Random Forests (RFs), and Light Gradient Boosting Machines (LightGBMs). Among these, the RF classifier demonstrated superior performance. Subsequent 5-fold cross-validation stability analysis along with feature importance analysis identified RMS radial derivative of the outer revolution as the most critical biomarker. Stratified Cross-Validation demonstrates that combining spatial and frequency features significantly enhances detection accuracy compared to single-domain methods, facilitating effective clinical deployment even in data-scarce environments. This interpretable pipeline provides a robust, low-cost white-box screening tool, offering a practical alternative to opaque deep-learning models for early clinical intervention.

Radio map (RM) estimation aims to reconstruct the spatial distribution of wireless signal characteristics, such as received signal strength (RSS), from sparse measurements, a task that is critical for spectrum management, interference mitigation, and localization in modern wireless networks. Traditional approaches, including interpolation and deep learning, either struggle to capture complex propagation effects or require large-scale retraining for each new sampling pattern, which limits their generalization. More recently, prior-based methods have combined pre-trained generative models with measurements to reduce the need for deployment-time model fine-tuning, but they typically treat the prior as a simple regularizer and lack explicit transmitter-aware integration. In this paper, we propose RadioTrace, a novel RM estimation framework without deployment-time fine-tuning that tightly integrates sparse RSS measurements with a frozen pre-trained diffusion prior. RadioTrace incorporates transmitter (Tx) location estimation directly into the denoising loop, iteratively refining Tx coordinates based on reconstruction quality to guide the generative process. To further enhance robustness, we introduce a propagation-guided K-means initialization that mitigates poor local minima in the Tx update and provides a geometry-consistent starting point. Moreover, we provide a stochastic stability analysis for the Tx-coordinate refinement component, showing that the Tx update remains stable under perturbations induced by diffusion sampling and Tx-map relaxation. Extensive experiments demonstrate that RadioTrace achieves competitive performance with state-of-the-art learning-based methods under random sampling, and maintains strong reconstruction quality under restricted-area sampling, highlighting its adaptability, robustness, and practical relevance.

We propose a notion of robust adaptive backup control barrier functions for nonlinear control affine systems with parametric uncertainty in both the drift dynamics and actuation matrix. Backup control barrier functions guarantee safety by predicting the system's trajectory under a pre-certified safe controller. However, these predictions rely on the model and can be inaccurate when the system contains unknown parameters. To address this issue, we estimate the unknown parameters using element-wise certified adaptive estimators that provide a parameter adaptation law and component-wise estimation error bounds. We compute the backup flow using the estimated model and tighten the safety conditions using these certified bounds. The resulting safety conditions account for the sensitivity of the predicted flow to parameter estimation errors. Moreover, to handle uncertainty in the actuation matrix, we use a duality-based reformulation that enables the use of a computationally efficient quadratic-program-based safety filter. We prove that controllers satisfying the proposed robust adaptive backup control barrier function constraints guarantee safety under parametric uncertainty and input constraints.

Motivated by reinforcement learning in harsh environments, we consider the problem of learning an optimal policy subject to adversarially corrupted feedback. Specifically, at each time-step, an adversary can perturb both the reward and state observations of the learner following the Huber contamination model. To defend against such data corruption, we propose BR-Async-Q: a novel, epoch-based, robust Q-learning algorithm built upon two key ideas: (i) partitioning the online data stream into batches to reduce variance, and (ii) constructing robust estimates of the Bellman optimality operator using such batched data. We prove a high-probability $\ell_\infty$ error bound for BR-Async-Q that matches that for vanilla Q-learning, up to a small additive term that scales with the fraction of corrupted samples. To our knowledge, this provides the first robustness guarantee for asynchronous Q-learning subject to both reward and state corruption. Furthermore, when only rewards are corrupted, the dependence of our algorithm's bound on the corruption fraction is minimax optimal.

Continuously recorded high-resolution waveform measurements provide rich information about fast power system dynamics. However, they require automated methods to identify events. This problem is addressed by developing a spectrogram-based framework to jointly detect, localize, and classify events in real-world continuously recorded waveforms at the terminal of an Inverter-Based Resource. We recast this problem as a temporal object detection problem on spectrogram images, as they capture the transient and harmonic signatures more explicitly than in raw waveform data. Each time-series waveform is transformed using the short-time Fourier transform, and the resulting per-channel spectrograms are stacked as a tensor for event detection. We benchmark this method against a detector operating directly on raw time-series measurements. Experiments on single-phase disturbances and three-phase faults demonstrate that the proposed spectrogram method consistently improves event detection, localization, and classification over the raw waveform baseline.

Emerging immersive applications are expected to support heterogeneous groups of users with fundamentally different communication and computation requirements. Using the Musical Metaverse (MM) as a representative example, we show how such applications expose limitations of current 5G network slicing and orchestration frameworks, which remain largely service-centric and operate through decoupled application and network management mechanisms. While existing Quality of Service (QoS) and slicing techniques provide traffic differentiation, they cannot jointly account for application semantics, shared resources, and differentiated end-to-end requirements across multiple user classes within the same service instance. To address these limitations, we introduce application-integrated slicing, a framework that unifies application and network orchestration within a common end-to-end service model and enables differentiated KPI targets for multiple user classes within a single logical slice. Using a MM reference scenario deployed over cloud-edge network, we show that application-integrated slicing reduces both resource provisioning cost and QoS violation rates compared with conventional decoupled approaches.

The State-of-Charge (SoC) is an important parameter of a battery energy storage system (BESS), and its balance problem is also an issue worth studying in a multi-BESS network. Recently, some researchers have proposed a power allocation method, claiming that as long as the power sharing state and SoC balance state can be obtained in real-time, it can not only maintain supply and demand balance, but also ensure that any BESS will not exit early due to insufficient energy storage. Considering this, we are attempting to design {a distributed dynamic average tracking algorithm} based on multi-agent systems (MASs) with prescribed transient and steady state performance to estimate the power sharing and SoC balance states \textcolor{blue}{of} a BESSs network with dynamic load in a distributed manner. Two versions of the solution are designed here, where one is event triggered and the other is self triggered. These two schemes ensure the performance of the estimators under intermittent communication to varying degrees, \textcolor{blue}{respectively}. In each constructed estimation scheme, prescribed performance control (PPC) method is implemented to ensure the expected steady-state and dynamic performance, which is encapsulated in a performance function. Thus, it achieves almost zero error estimation of the fast time-varying {power} sharing and the SoC balance states. In addition, consensus and average tracking performance are decoupled, which provides convenience for \textcolor{blue}{parameters} tuning. Finally, to verify the results of \textcolor{blue}{the} theoretical analysis, some cases are studied and relevant simulations are performed on a 4 bus system.

Battery energy storage systems (BESSs) are often integrated into the smart grid as the key equipment for valley filling and peak suppression. However, the internal power consumption and capacity degradation of battery cells can not be ignored. In addition, there are certain electric trading between the owner of the grid-connected BESSs and the electric company (EC). Therefore, an expenditure function for grid-connected BESSs is constructed in this paper, in which internal power consumption, capacity degradation and power trading are covered while meeting the balance of power supply and demand. On this basis, the Karush-Kuhn-Tucker (KKT) condition of the function is summarized as the consensus problem of marginal cost (MC) converging to time-phased electricity price. Thus, we plan to design a distributed MC consensus scheme for economic dispatch (ED) in BESSs based on multi-agent systems (MASs) on a small-time scale. In view of the low control accuracy and response speed of the existing distributed proportional protocol, this paper proposes a distributed ED (DED) scheme with reset mechanism based on a proportional integral (PI) control. When the proportional term encounters zero crossing, the integral term of the control scheme is reset to 0, which ensures that the signs of the two are aligned, thus accelerating MCs convergence and restraining overshoot. Parameter conditions for consensus, regularity and Zeno-free behavior are given through the relevant theoretical analysis. Several simulation cases are designed to verify the designed DED algorithm.

The State of Charge (SoC) balance, power sharing, and frequency restoration are common control objectives of battery energy storage systems (BESSs). However, the SoC balance scheme induced by the power allocation through existing droop controllers can cause the capacity parameters of battery cells to be unequal to the droop coefficient, which is the result of battery capacity degradation. Under this limitation, previous capacity based droop controllers and the secondary controllers no longer suitable to address the imprecise power sharing and frequency restoration caused by this problem. Therefore, a power allocation scheme based on the current SoC level ratio is designed to induce a new droop controller and ensure that the SoC simultaneously drops to 0. In order to restore frequency in a distributed manner and obtain SoC level ratio, a distributed nominal frequency controller, SoC average estimator, and power sharing controller are designed based on multi-agent systems (MASs) in both asymptotic and finite time manners. In the asymptotic scheme, the steady-state performance of the SoC estimator is adjustable, and power sharing and frequency restoration are zero errors. In the finite time scheme, the influence of parameters on convergence time is well analyzed, and some conservative calculation methods are provided. Several time-domain simulation examples are designed on an improved IEEE57 bus system to verify the distribution of asymptotic and finite time schemes.

Reliable and continuous human tracking is essential for natural human-robot interaction, particularly for companion robots. However, many existing approaches rely on wearable tags or fixed cameras with limited fields of view, which reduces system flexibility and often causes tracking failures when the target moves outside the sensing range. In this paper, we present a human tracking approach based on a gimbal-mounted camera integrated into a mobile robot. By actively controlling the gimbal mechanism, the camera can dynamically adjust its viewing direction to maintain the target within the field of view, even under substantial relative motion between the robot and the human. Furthermore, a continual learning strategy is applied to the person re-identification (ReID) task to adapt to changes in appearance and environmental conditions during long-term tracking. Experimental results demonstrate that the proposed system significantly improves the stability and continuity of human tracking, enables real-time re-identification, and provides responsive feedback for reliable tracking of human motion from walking to running. User studies further indicate that the proposed approach enhances user comfort by eliminating the need for wearable tags.

Service robots are increasingly deployed in elderly-care facilities to alleviate caregiver workload and enhance the quality of daily care. However, most existing studies focus on isolated service functions and lack integrated capabilities for continuous companionship, natural interaction, and safety monitoring. In this paper, we present an intelligent companion robot system that unifies active visual human-following, real-time LLM-driven speech interaction for intent understanding and task execution, and VLM-based safety monitoring for fall detection and abnormal posture assessment. The perception layer ensures robust human tracking and uses an active gimbal to maintain the user in view during occlusions or abrupt movements. At the interaction layer, a Large Language Model interprets spoken requests and maps them to robot actions, enabling escorting and semantic navigation. Simultaneously, a VLM-based safety agent continuously analyzes visual observations to detect fall-related or abnormal postures and triggers emergency responses when necessary. Experimental results demonstrate the system's ability to reliably follow and interact with humans, while effectively detecting potential falls to ensure user safety.

The sixth-generation (6G) of mobile networks will be shaped not only by artificial intelligence (AI)-enabled network automation and optimization, but also by the need to serve AI as a 6G-native workload. Emerging AI services introduce traffic and compute demands that differ from conventional mobile broadband. Their user experience depends on how quickly useful information is delivered, how bursty and asymmetric multimodal flows are handled, and where inference, retrieval, caching, and content processing are executed. This article presents a joint connectivity-compute view of AI-native 6G. We first characterize representative AI service traffic in terms of uplink/downlink throughput skew, burstiness, and token latency. Next, we discuss how fifth-generation extended reality awareness mechanisms can evolve toward AI traffic characteristics awareness in 6G. Finally, we introduce AI Grid as a distributed AI infrastructure platform for placing workloads according to latency, cost, policy, and service-level constraints. Together, AI-aware connectivity and AI Grid enable 6G as a distributed intelligence platform.

Satellite-derived Land Surface Temperature (LST) provides spatially comprehensive data that ground stations cannot match. However, its utility is frequently limited by severe data gaps due to the presence of clouds. As LST is essential for understanding land-atmosphere interactions, numerous methods have been proposed to address this challenge. Yet, the development of a scalable and adaptable pipeline for generating gap-free LST datasets and reconstructing cloud-contaminated pixels remains challenging. Moreover, the reconstruction of extensive missing regions in fine-spatial-resolution observations is particularly difficult. To address this challenge, we propose a Multimodal Fast Fourier Convolutional GAN for reconstructing cloud-contaminated pixels in fine-resolution (30 m) Landsat imagery to generate gap-free clear-sky LST products. The method leverages Fast Fourier Convolution to enable a global receptive field across the image, and is guided by a stack of data consisting of satellite observations and Synthetic Aperture Radar (SAR) data. Across all LST quantiles, the interquartile range of scene-averaged RMSE (computed over reconstructed pixels) is consistently between 0.8 K and 1.8 K. The proposed approach enables the recovery of extensive missing regions, including scenes with more than 70% cloud-induced gaps, while relying on auxiliary data that are readily available at a near-global scale.

In this study, a novel encoder-only Transformer-based receiver architecture is presented for BPSK signals transmitted over Faster-than-Nyquist (FTN) signaling channels that introduce intentional inter-symbol interference (ISI) with a compression factor of $\tau=0.8$. A complete end-to-end communication chain encompassing BPSK modulation, RRC pulse shaping, and the ISI coefficients arising from matched filtering was constructed and evaluated. The proposed Transformer receiver was benchmarked against the optimal BCJR detector over an $E_b/N_0$ range of 0-8 dB. To systematically close the BER gap to the BCJR, a two-stage training strategy combining multi-SNR pretraining and per-SNR curriculum fine-tuning was developed. The computational complexity and inference latency of the Transformer receiver were analyzed in comparison with a GRU based receiver. Attention map visualizations revealed that the Transformer autonomously identifies the FTN-induced ISI memory structure without requiring any prior channel knowledge; as the SNR increases, the attention weights become significantly concentrated around the center token and its nearest neighbors.

Near-field sampling seeks to represent electromagnetic fields between transmitting and receiving regions using a minimal number of measurement points while preserving the dominant spatial modes. This paper develops a geometry-aware sampling framework based on spatial degrees of freedom (DoF). A view-length formulation is used to derive closed-form expressions for the propagating-mode DoF density for simple line-source geometries, providing both the total DoF and its local distribution. One-DoF sampling points are obtained from equal increments of the cumulative DoF density, yielding an adaptive nonuniform sampling strategy up to the knee of the singular-value spectrum. To improve the representation of the remaining modes beyond the knee, a reactive-mode density is introduced to guide the placement of additional edge samples. An operator-based sampling error functional is formulated and shown to be lower-bounded by the neglected singular values of the continuous channel operator. Numerical results demonstrate that the proposed sampling strategy closely approaches the optimal performance obtained from singular-value decomposition and significantly outperforms sampling based solely on the propagating-mode DoF density.

Machine-learning-based channel predictors must operate under stringent latency, memory, and computational constraints while remaining robust to noisy and time-varying observations. This paper develops a causal channel-prediction framework based on three single-layer gated recurrent unit variants: an unconstrained lightweight GRU (L-GRU), a stability-aware GRU (SA-GRU) with a spectral bound on the candidate-state recurrent matrix, and a doubly constrained lightweight GRU (DCL-GRU) with additional control of the reset-gate recurrent matrix. A sufficient condition is derived for contraction of the complete candidate-state mapping, while preserving the parameter count and inference-time structure of the baseline architecture. These guarantees apply to the candidate-state mapping and do not directly imply contraction of the complete GRU hidden-state transition. The models are trained on 2x2 MIMO channels generated using the 3GPP CDL-A model, and their hyperparameters are selected through Bayesian optimisation with Optuna's Tree-structured Parzen Estimator. Across the considered SNR range, the constrained variants retain competitive prediction accuracy and achieve optimisation runtimes close to L-GRU, with speedups of 1.72x and 1.76x relative to a five-layer GRU for SA-GRU and DCL-GRU, respectively. All audited constrained runs satisfy the prescribed spectral bounds. Under temporary observation corruption followed by recursive prediction, SA-GRU reduces the mean and peak hidden-state deviations by approximately 15.3% and 13.0%, respectively, relative to L-GRU, whereas L-GRU achieves the lowest rollout NMSE. These results highlight an explicit trade-off between prediction accuracy, empirical rollout robustness, and candidate-state stability guarantees.