Balancing computational efficiency with reconstruction accuracy remains a core challenge in high-dimensional tensor recovery. We propose Adaptive Convex-Concave Tensor Robust Principal Component Analysis (AC2-TRPCA), which jointly leverages data-adaptive column-orthogonal subspace learning within a reduced-core representation and a convex-concave singular value separation strategy. A nonconvex regularizer applied to the reduced core tensor yields a tighter rank approximation and substantially lowers computational and memory complexity. On the theoretical side, we establish an approximate recovery guarantee under a modified tensor incoherence condition and prove that the Alternating Direction Method of Multipliers (ADMM)-based algorithm converges to a generalized Karush–Kuhn–Tucker (KKT) point. Experiments on synthetic data, hyperspectral image denoising, and video background modeling demonstrate that AC2-TRPCA improves reconstruction accuracy by 1.9–4.0 dB while substantially reducing runtime overhead, enabling the processing of large-scale tensors within a fraction of the time required by existing approaches. The source code is publicly available at https://github.com/chy0926/Provable-Tensor-Robust-Principal-Component-Analysis-via-Adaptive-Convex-Concave-Regularization.
In modern electronic warfare, dual-polarization radar systems encounter jamming with varying polarization states due to real-world perturbations. Conventional unidirectional adaptive polarization filters with slow transients exhibit poor robustness in non-stationary environments. To address these limitations, this paper proposes a signal modeling scheme for simultaneous dual-polarization radar under suppressive jamming. Based on this scheme, the Wiener optimal filter is derived, yielding a closed-form upper bound on the interference suppression ratio (ISR). Theoretical analysis reveals that the optimal ISR depends solely on the jamming-to-noise ratio (JNR) in the cross-polarization channel and is independent of the specific jamming polarization state, establishing a valuable performance benchmark for robust filter design. Furthermore, a forward-backward recursive least squares (FB-RLS) algorithm is proposed for adaptive polarization filtering. By adapting bidirectional range-sample processing within a pulse repetition time to polarization-domain adaptive filtering, FB-RLS substantially enhances transient tracking performance and convergence stability, enabling consistent deep interference suppression across the full range dimension even under jamming with unintended polarization variations. Comprehensive evaluations through simulations in single-polarization noise jamming scenario and jamming scenario with unintended polarization variations, along with validation on field-measured data, demonstrate that FB-RLS consistently outperforms the compared methods, confirming its superior practical robustness in complex electronic warfare environments.
To address the lack of interpretability in Convolutional Neural Networks (CNNs) for ship-radiated noise classification, this paper proposes a visual analysis framework integrating Guided Backpropagation (GBP) and Gaussian Mixture Models (GMMs). Firstly, a convolutional kernel visualization process based on GBP is constructed, combined with a multi-frame feature alignment algorithm to eliminate time-frequency interference from harmonic line components, enabling the precise reconstruction of deep convolutional kernel response features. Secondly, a GMM-based pattern decomposition method is proposed, which models activation locations to decompose a single convolutional kernel into multiple sub-patterns with distinct frequency distributions. Experiments were conducted on a CNN model with CQT spectrogram inputs, validated on the DeepShip dataset. The visualization results reveal that deep convolutional kernels contain two primary patterns: line spectra and background, detecting harmonic line components and continuous spectrum features in LOFAR spectrograms, respectively. After GMM decomposition, the response regions of the sub-patterns become more concentrated, feature expression becomes more consistent, and an association mechanism linking “convolutional kernel - sub-pattern - ship class” is established. This research provides a quantitative basis for understanding CNN decision logic, filling a gap in interpretability studies for ship-radiated noise classification models.
Semantic communication has emerged as a promising paradigm for next-generation intelligent communications by prioritizing semantic fidelity over conventional bit-level accuracy. Although deep joint source-channel coding (JSCC) has demonstrated strong robustness under varying channel conditions, existing methods still face a trade-off between structural preservation and textural reconstruction. Most existing approaches rely on either convolutional neural networks (CNNs) or Transformers. CNN-based methods are effective at capturing local textures but struggle to model long-range structural dependencies, while Transformer-based methods capture global information yet often fail to preserve fine-grained textures under aggressive compression. To address these limitations, we propose Structural-Textural JSCC (ST-JSCC), a novel framework that jointly exploits structural and textural representations for robust image transmission. Specifically, a parallel architecture is designed, consisting of a linear attention branch for efficient global structure modeling and a CNN branch for local texture preservation. Furthermore, a Gated Fusion Module (GFM) is introduced to adaptively integrate complementary features from both branches. Experimental results on benchmark datasets demonstrate that ST-JSCC consistently outperforms state-of-the-art methods, especially under low-SNR conditions. The source code will be made publicly available upon publication.
Active reconfigurable intelligent surfaces (RISs) can mitigate severe multiplicative fading through signal amplification, but fully-connected architectures consume substantial hardware power because each reflecting element requires a dedicated power amplifier (PA). This paper investigates a sub-connected active RIS (SC-ARIS)-assisted integrated sensing and communication (ISAC) system, where multiple reflecting elements share one PA, reducing active-hardware power consumption at the cost of lower amplification-control flexibility. We formulate a highly non-convex energy efficiency (EE) maximization problem by jointly optimizing the base station transmit beamforming, radar receive filter, and SC-ARIS reflection coefficients, subject to a minimum radar signal-to-noise ratio (SNR) requirement, power budgets, and hardware constraints. To solve it, we develop an alternating optimization framework combining Dinkelbach’s method and fractional programming. By exploiting the phase freedom of the radar receive filter, the non-convex radar SNR constraint is transformed into a linear constraint on the transmit beamformer for a fixed receive filter, while projection-based recovery yields hardware-feasible SC-ARIS coefficients. Simulations show that the proposed design achieves higher EE than fully-connected and random-phase baselines, although the fully-connected benchmark attains a higher communication sum-rate. These results demonstrate a favorable trade-off among EE, ISAC performance, and active-hardware complexity.
Real-world underwater polarization images often suffer from severe scattering and spatially non-uniform illumination, making reliable restoration challenging in the absence of paired training data. In this paper, we propose a zero-shot underwater polarization image restoration framework that combines a physically grounded polarization imaging model with joint internal and external prior learning in a plug-and-play optimization scheme. The polarization model is first employed to decouple the observed signal into illumination, reflectance, and transmission components. To address spatially varying illumination, we introduce an internal prior by constructing a pixel-wise adaptive weighting matrix from local variance and gradient cues, enabling adaptive illumination smoothing while preserving structural transitions. A pre-trained deep denoising network is further incorporated as an external prior to refine the reflectance and transmission components, avoiding handcrafted regularization terms. In addition, an adaptive noise estimation strategy is adopted to automatically control the denoising strength during optimization. Experiments on real-world underwater polarization images show that the proposed method achieves visually consistent restoration and superior performance across multiple no-reference image quality metrics without relying on underwater paired datasets.
Estimating individualized treatment effects under continuous treatments (e.g., medication dosage) is challenging because treatment-response heterogeneity is often entangled with treatment-assignment bias in observational data. This paper proposes H-AdvCE, a Hellinger distance-based representation learning framework for cluster-aware continuous treatment effect estimation. The key idea is to learn a latent representation that preserves outcome-relevant heterogeneity while reducing dependence between the representation and treatment assignment. Theoretically, we derive a new counterfactual generalization error bound based on the Hellinger distance and show that it is tighter than commonly used Kullback–Leibler (KL) divergence-based alternatives. To optimize this objective, we introduce a variational Hellinger regularizer and formulate the learning problem as a min–max adversarial game. Empirically, H-AdvCE achieves strong performance on IHDP, News, and TCGA benchmarks for continuous dose-response estimation. Beyond prediction accuracy, the learned latent space reveals response-aware subgroups with distinct dose-response patterns, demonstrating its potential for clustering-driven interpretation in continuous treatment effect analysis. Our code is available at https://github.com/syzhao117/AdvCE.
No-Reference Image Quality Assessment (NR-IQA) aims to assess the perceptual quality of images in accordance with human subjective perception. Recent methods based on multimodal large language models (MLLMs) use vision-language pretraining to improve semantic generalization, but visual encoders pretrained by Contrastive Language-Image Pretraining (CLIP) can underrepresent subtle, spatially distributed distortions in late visual tokens, which can make score regression rely excessively on semantically salient content. To address this representation mismatch, we propose the Distortion-Aware Residual Recalibration Adapter (DARRA), a lightweight adapter inserted into late Vision Transformer blocks. DARRA constructs local residual responses from patch-token neighborhood differences, separates residual evidence from coarse structural context, and applies bounded Inverted Energy Reweighting (IER) to retain nonzero residual responses with low distinctiveness before fusion. After structure-guided fusion, DARRA injects the fused residual evidence through a zero-initialized branch, preserving the pretrained visual pathway at initialization while learning distortion-sensitive corrections during fine-tuning. Experiments on nine IQA datasets show consistent gains on synthetic and model-processed distortion datasets while maintaining competitive performance on authentic datasets. Component ablations and controlled comparisons of reweighting mechanisms further identify Inverted Energy Reweighting as the main contributor to the observed improvement.
Absolute Pose Regression (APR) and relocalization estimate a cameras pose relative to a known scene from a single query image. CNN-based direct regression methods are fast and lightweight but rely solely on visual data, ignoring scene geometryreducing accuracy in unbounded scenes with photometric variations. To address this, this paper proposes a “mapping-before-positioning” pipeline. First, a Neural Radiance Field (NeRF) is trained for such scenes, integrating geometric information to improve relocalization performance. Then, NeRFs supervision signal builds a lightweight, high-precision single-shot relocalization model with strong photometric generalization. For effective unbounded scene reconstruction, an NeRF modeling method is proposed. By performing dense sampling on foreground regions and sparse sampling on background regions, this method ensures high-precision NeRF reconstruction for unbounded scenes. Second, a scene-invariant feature attention sampling strategy is designed to capture invariant features in the scene, thereby mitigating interference from photometric variations. Finally, a NeRF-supervised training pipeline is developed for the single-shot relocalization model, forming a passive relocalization scheme. Leveraging the scene-invariant features captured during NeRF sampling and the introduced geometric information, this method significantly enhances the photometric generalization capability of pose estimation. Experiments show centimeter-level accuracy in 0.2 s with a model size of just several megabytes.
The design of integrated sensing and jamming waveform (ISJW) aims to seamlessly unify radar sensing and electronic jamming functionalities within a single transmission framework. Although existing research has made progress in shared waveform construction and false-target generation, a unified framework remains elusive for achieving range–velocity (RV) deception while mitigating the sidelobe incoherence caused by velocity-induced pulse agility. To address these issues, this paper proposes an ISJW design method for RV deception via hybrid intelligent optimization. Specifically, a multi-dimensional modulation ISJW model is established, and the formation mechanism of two-dimensional false targets on the RV plane is analyzed. Furthermore, the pulse-agility period is derived from the slow-time phase evolution induced by relative velocities, and a joint mismatched filtering approach is introduced to recover sidelobe coherence. Subsequently, a constant-modulus multi-objective optimization problem is formulated to balance sidelobe suppression and inter-pulse similarity. To solve this, an integrated differential evolution and annealing (IDEA) algorithm is developed, incorporating a group-wise incremental synthesis mechanism to reduce computational overhead. Simulations demonstrate that the proposed method achieves RV deception for multiple false targets, effectively mitigates Doppler defocusing, and preserves sensing performance.
Blind channel estimation (BCE) is essential for recognizing non-cooperative bio-inspired communication signals. However, click-mimicking signals, a typical category of cetacean-inspired communication signals, exhibit short duration, wide bandwidth, and non-stationary characteristics, leading to limited performance of cross-relation model-based methods and source signal reconstruction methods. To address this issue, this paper proposes a scenario-adaptive blind channel estimation method based on orthogonal matching pursuit (OMP) for click-mimicking signals. Underwater acoustic channels are classified into short-delay and long-delay scenarios according to whether at least one delay interval between adjacent propagation paths exceeds the signal duration, and dedicated estimation methods are developed for each scenario. For short-delay channels, the BCE problem is formulated as a sparse vector recovery model by exploiting cross-relations and multipath sparsity in single-input multiple-output systems. The model is improved through measurement vector selection, dictionary matrix optimization, and an improved OMP algorithm to enhance estimation accuracy. For long-delay channels, a source signal estimation approach is proposed based on cross-correlation and autocorrelation properties of multichannel received signals. The BCE problem is transformed into a channel recovery model with a known signal, which is solved using OMP to reduce computational complexity. Simulation and experimental results demonstrate the robustness and accuracy of the proposed method.
The ℓp/ℓq (0 < p ≤ 1, q > 1) norm ratio, a nonconvex sparsity metric, satisfies scale invariance property that is desired for sparse data. In this paper, we focus on establishing theoretical guarantees for stable recovery of sparse signal with partially known support by the weighted ℓp/ℓq-minimization. To this end, sufficient conditions and error upper-bounds are provided and summarized as two theorems. Theorem 1 is proposed based on restricted isometry property and restricted orthogonality property, and Theorem 2 is proposed under the framework of cumulative coherence. From Theorem 1, we can obtain Corollary 1 which is for the stability of the solution to the original ℓp/ℓq-minimization. When more than 50% of the known support is accurate, Theorem 1 outperforms Corollary 1 due to a weaker sufficient condition and tighter error upper-bounds. In particular, we are interested in a representative parameter-free metric ℓ1/ℓ2. By setting p=1 and q=2 and considering the aforesaid assumption of having the accurate support over 50%, we reduce Theorem 1 to Corollary 2 which is for the stable recovery guarantee of the weighted ℓ1/ℓ2-minimization. Under some cases, in terms of the weakness of the sufficient condition and the tightness of the error upper-bound, Corollary 2 is better than a stable recovery guarantee in a previous work that focuses on the weighted ℓ1/ℓ2-minimization. Furthermore, we present numerical experiments to demonstrate the two comparisons above, and to visually depict the sufficient condition and the error upper-bounds in Theorem 1.
Monocular 3D object detection is critical for roadside perception but faces significant challenges due to high camera mounting positions and large pitch angles, which cause strong perspective distortion and scale-resolution variation, especially for distant objects. Under fixed roadside cameras on mostly planar roads, image coordinates offer approximate perspective cues for depth-related modeling, but are not universally valid geometric constraints. To address this, we propose MonoCAP, a Coordinate-Aware and Perspective-Aware Learning framework. Specifically, we design a Coordinate-Aware Hybrid Vision Encoder (CAHVE) that incorporates coordinate-aware positional cues through 2D positional embeddings and Coordinate Attention, enhancing spatially sensitive feature representation for distant objects. Furthermore, a Geo-Visual Synergistic Attention (GVSA) mechanism is introduced to learn a data-driven fusion ratio between ground-aware geometric features and visual semantic features, providing a more flexible geo-visual aggregation strategy than manually fixed fusion. Additionally, we construct a Perspective-Aware Deformable Attention (PADA) module, which introduces vertical-coordinate cues as soft perspective guidance for multi-scale feature selection and spatial sampling, improving feature adaptation to the near-large-far-small scale variation under fixed roadside viewpoints. Extensive experiments on the DAIR-V2X-I and Rope3D benchmarks demonstrate that MonoCAP achieves competitive performance and particularly improves the detection of distant objects under typical roadside perspective variations.
Sparse identification of nonlinear dynamical systems provides a powerful framework for signal processing when governing equations remain unknown. Feature-selection-based methods must balance parsimony with accurate representation. However, existing methods typically perform selection in a single pass and permanently exclude candidates without reconsideration, which is problematic when measurement noise or weak signal strength causes true governing terms to be incorrectly pruned. To address this issue, we introduce an adaptive three-stage framework with an integrated recovery mechanism. First, an adaptive thresholding stage reduces the candidate feature library while preserving potential governing terms. Second, a forward-backward stepwise selection guided by the Bayesian information criterion provides a theoretically grounded model selection mechanism without requiring specification of regularization parameters. Third, a residual-based post-audit examines residual correlations to identify potentially missed features, and bootstrap resampling removes terms with unreliable coefficient estimates. Extensive experiments on synthetic systems, partial differential equations and real-world applications are conducted to demonstrate the effectiveness of the proposed method.
To improve the performance of direction of arrival and range estimation for mixed far-field and near-field sources, we propose a localization scheme tailored for space-constrained platforms, which incorporates a symmetric sparse nested array (SSNA) with scalable aperture and multiple coprime rational frequencies. The physical aperture of the SSNA, formulated as a function of antenna number of subarrays, can scale with uniform step, significantly improving flexibility to maximize the utilization of the available geometric space on platforms. Benefiting from its extended unit inter-antenna spacing, the SSNA cooperated with multi-frequency can give enhanced aperture and degrees of freedom (DOF) in space-frequency domain, both of which outperform those provided by the existing arrays, thereby enabling unambiguous and high-precision parameter estimation. The closed-form expressions of coarray DOFs with the scaled apertures, are derived for both space- and space-frequency-based difference coarrays. A closed-form solution for the optimal array configuration which maximizes DOF is presented, and the maximum DOFs for the array with two and multiple frequencies are analyzed. The simulation results validate the superior performance of the proposed scheme over the existing techniques.
We address the Normalized Signal to Noise Ratio (NSNR) metric defined in the seminal paper by Reed, Mallett, and Brennan on adaptive detection. NSNR is the ratio between the SNR of a linear detector which uses an estimated noise covariance and the SNR of a clairvoyant detector based on the exact unknown covariance. It is not obvious how to evaluate NSNR since it is a function of the target vector. To close this gap, we consider the NSNR associated with the worst target. Using the Kantorovich Inequality, we provide a closed-form solution for the worst-case NSNR. Then, we prove that the classical Gaussian Kullback Leibler (KL) divergence bounds it. Motivated by these results, we derive a simple variant of a classic norm based estimator by incorporating KL in a leave-one-out cross-validation (LOOCV) framework. Numerical experiments with different true covariances and various estimates suggest that the KL metric is more correlated with the NSNR metric than competing norm-based metrics and simply changing the metric in the LOOCV estimator improves KL and NSNR performance.
We describe two Hilbert transform-based methods for augmenting covariance matrices from multivariate signals. The two methods are shown to be isometric under the Riemannian affine-invariant metric on the manifold of symmetric/Hermitian positive definite matrices. The augmented representations distinguish cases that standard covariances clearly cannot and, when paired with a Riemannian minimum distance to mean classifier, improve classification of both synthetic data and real electroencephalography (EEG) data without introducing any extra hyperparameters. This novel combination of methods also significantly improves tangent space classifier accuracy on the same dataset, outperforming state-of-the-art classifiers with similarly sized parameter grids. We also examine how the augmented covariances interact with the minimum distance to mean classifier and show how multivariate cross-covariance functions behave under the Hilbert transform.
This article proposes an auxiliary model based completely multidirection conjugate gradient iterative (AMUL-D-CGI) identification algorithm for output-error (OE) systems. The approach extends the multidirection gradient iterative method, which improves convergence by exploring multiple gradient directions per iteration within a Krylov subspace. However, this existing multidirection gradient method fails to ensure complete conjugacy among directions and suffers from a limited generative strategy. To overcome these issues, the proposed AMUL-D-CGI algorithm constructs a mutually conjugate direction set via a modified Gram-Schmidt process, integrating historical gradient information and eliminating error components in multiple directions. Theoretical analysis shows that the algorithm achieves an lth-order acceleration over conventional conjugate gradient methods under ideal conditions. Simulations validate significant improvements in convergence rate and parameter estimation accuracy.
