This paper proposes a Hadamard row-permutation based time-modulation matrix design for direction-of-arrival (DOA) estimation with time-modulated linear arrays (TMLAs). Fixed time-modulation schemes may lead to nonuniform harmonic observation energy, angle-dependent performance loss, and limited resolution for closely spaced sources. To address these issues, the proposed method optimizes the assignment between Hadamard binary switching sequences and physical array elements through row permutation, while preserving binary implementation and the underlying Hadamard row orthogonality. A whitened harmonic-domain MUSIC framework is adopted to handle the colored noise induced by time modulation. Based on the whitened effective array manifold, a Cramér–Rao bound (CRB)-inspired projected-derivative metric and a steering-vector correlation penalty are combined to guide the row-permutation optimization, which is solved by a greedy row-exchange algorithm. Simulation results verify that the proposed design provides a more uniform nonzero-harmonic power distribution and achieves improved single-source accuracy, multi-source root-mean-square error (RMSE), and closely spaced source resolution compared with conventional and unoptimized Hadamard-based time-modulation schemes.
Face super-resolution (FSR) under extreme degradation must reconstruct frequency bands with different recoverability: coarse facial geometry survives downsampling better than oriented fine detail. Existing feed-forward methods do not explicitly allocate successive refinement steps from the current spectral profile, while supervised multi-step objectives can at tenuate long-range credit or encourage locally optimal updates. We propose PolicySR, a finite-horizon policy that observes a compact multi-level, multi-orientation wavelet profile of decoder features and emits a low-dimensional action that modulates a shared decoder. During training, a focal subband reward weights each band's improvement by its pre-action normalized error; trajectory-level proximal policy optimization yields an empirically observed coarse-to-fine schedule. On the CelebA and Helen datasets, PolicySR improves all reported PSNR, SSIM, and LPIPS values over state-of-the-art methods from 4× to 16×. The full model reaches 22.71dB on Helen at 16×; early exit retains 22.68dB with 10.1M parameters and 23ms latency.
Outdoor-to-indoor sub-terahertz links leak energy through building partitions, enabling penetration-path eavesdropping. This letter studies when polarization-dependent wall penetration can be a physical-layer secrecy resource rather than a favorable-placement effect. Near the pseudo-Brewster angle, TM internal reflection is strongly suppressed, changing Fabry–Pérot fading for the terminal closest to that incidence condition. We formulate a polarimetric stochastic penetration model with co-polar TE/TM slab response, bounded depolarization, roughness loss, and random multipass phase. The closed-form Rician parameters expose a fading-asymmetry term that remains after mean path gain is normalized. We then propose GAPS, a geometry-aware polarization rule using the legitimate incidence angle and an admissible eavesdropper sector, with a minimax fallback when the passive eavesdropper is not localized. At 140 GHz, GAPS achieves 0.741 bps/Hz versus 0.371 bps/Hz for an agnostic baseline in the studied glass case; after gain normalization, it retains 0.0697 bps/Hz versus 0.0446 bps/Hz for a no-Brewster control.
Conventional beamforming (CBF) suffers from broad mainlobes, sidelobe leakage, and background fluctuations in low-frequency acoustic detection, which limits the resolution of closely spaced sources. This letter proposes a structure-guided weighted deconvolution method for high-resolution direction-of-arrival (DOA) estimation. A positive direction-dependent structural weight is constructed from the normalized CBF spectrum by exploiting amplitude concentration and symmetric-gradient information, thereby emphasizing candidate source-related angular structures and improving the peak-to-valley contrast before deconvolution. The weighting parameters are determined offline through Bayesian optimization and then fixed for independent testing. The optimized weight generates a structurally refined CBF spectrum while keeping the steering-manifold-based point spread function (PSF) matrix unchanged, allowing the weighted spectrum to be used with different PSF-based deconvolution solvers. Weighted DAMAS and weighted Richardson–Lucy (RL) deconvolution are developed under this framework. Simulation results show that the proposed methods achieve higher resolution probability while maintaining accurate DOA estimation for closely spaced sources under low-SNR conditions.
This letter proposes a frequency-optimized multi-frequency joint sparse recovery method for grating-lobe suppression in distributed radar. Single-frequency sparse recovery in distributed large-baseline arrays may suffer from highly correlated grating-lobe atoms, resulting in unstable angular support estimation. To address this problem, a grating-lobe consistency metric is constructed from frequency-dependent observation matrices, and an optimized frequency set is selected to reduce the cross-frequency consistency of ambiguous lobes. A multi-frequency joint sparse reconstruction model is then developed to recover the common angular support of true targets. Simulation results demonstrate that the proposed method effectively suppresses residual grating lobes and improves angular support stability.
