This paper studies finite-blocklength performance analysis and near-limit receiver design for short-blocklength coded fractional orthogonal time-frequency space (OTFS) systems. To characterize the finite-blocklength performance limits, we first develop a direct achievability benchmark based on the s-tilted random-coding union (RCU) bound evaluated directly on the physical fractional OTFS channel. To obtain more tractable benchmarks, we further introduce two equivalent-SNR proxies: a channel-based proxy determined by the effective channel strength, and a detector-based proxy inferred from the detector output. These proxies enable approximate RCU, meta-converse (MC), and normal-approximation (NA) evaluations through a binary-input additive white Gaussian noise (BIAWGN) surrogate. On the receiver side, we propose a practical near-limit receiver that combines partial Gaussian approximation (PGA) soft-output detection, used to cope with the dense coupling induced by fractional delay/Doppler leakage, with locally constrained ordered-statistics decoding (LC-OSD) for efficient near-ML decoding in the short-blocklength regime. We also develop a decoder-oriented frame-error-rate (FER) predictor via saddlepoint analysis by approximating the list-tail probability under a prescribed search budget. Numerical results show that the proposed finite-blocklength bounds for coded fractional OTFS follow trends consistent with the well-known finite-blocklength behavior of the BIAWGN channel. Moreover, under different code rates and fractional OTFS configurations, the proposed receiver based on PGA detection and LC-OSD consistently remains close to the proposed finite-blocklength bounds, and achieves clear gains over coded fractional OTFS receivers employing 5G LDPC codes with BP-based recovery as well as 5G polar codes with CA-SCL-based recovery. The proposed saddlepoint approximation also effectively predicts the receiver performance.
Terahertz (THz) communication systems are highly vulnerable to phase noise and multipath distortion caused by hardware imperfections and complex propagation environments. To address these challenges, this paper proposes a pilot-free jointly optimized UNet (PF-JO-UNet) transceiver employing chirped discrete Fourier transform spread orthogonal frequency division multiplexing (DFT-s-OFDM). To the best of the authors’ knowledge, this is the first work to deploy UNets at both the transmitter and receiver to address THz phase noise and multipath distortion via end-to-end learning and optimization. Unlike conventional schemes, the proposed framework performs joint phase noise mitigation and multipath equalization without relying on pilots, thereby improving spectral efficiency. Numerical results demonstrate that, by combining PF-JO-UNet with chirped DFT-s-OFDM, the proposed framework not only eliminates the error floors observed in single-carrier and DFT-s-OFDM systems, but also achieves BER performance close to the ideal phase-noise-free case and higher spectral efficiency than existing phase-noise mitigation schemes. The results further reveal that linear chirping with a DFT spreading factor of at least 4, together with odd-indexed chirp rates, is the preferred configuration for robustness enhancement.
Orthogonal Time Frequency Space (OTFS) modulation is promising for high-mobility communications and sensing, yet practical fractional delay-Doppler (DD) shifts can severely degrade its performance. Such off-grid effects induce two-dimensional energy leakage in the DD domain, weakening the sparsity structure exploited by conventional estimators and thereby impairing both communication reliability and sensing performance. To address this challenge, we propose a hybrid model-data driven Bayesian inference framework for dynamic off-grid channel estimation. Specifically, we develop a spatiotemporal Markov hyperprior prediction-aided variational inference-based sparse Bayesian learning (SMHP-VI-SBL) algorithm, where the off-grid VI-SBL backend iteratively optimizes continuous-valued delay and Doppler parameters, while the SMHP module leverages a spatiotemporal network (ST-Net) to learn support evolution from historical frames and maps these predictions into a hyperprior via a Markov model to warm-start VI-SBL, thereby reducing the required iterations and improving estimation accuracy. QuaDRiGa-based simulations demonstrate that, compared with the Off-Grid VI-SBL baseline, the proposed SMHP-VI-SBL requires approximately 2.4× fewer iterations to converge and achieves an average normalized mean squared error (NMSE) improvement of 1.12 dB, demonstrating its efficiency for high-mobility OTFS systems. In addition, we develop a USRP-based prototype system and conduct field trials, confirming the feasibility of the proposed algorithm in real-world environments.
Sixth-generation (6G) is expected to support diverse applications, where waveform design plays an important role. In this paper, a universal linear algebraic modulation (ULAM) is proposed for 6G as a general waveform design, which subsumes most of the existing fourth- (4G) and fifth-generation (5G) waveforms as special cases. ULAM is empowered by the novel discrete canonical transform (DCaT), which is governed by a parametric quintuplet and covers most well-known transforms adopted and proposed for 4G/5G. Furthermore, algebraic-theoretic tools are employed to systematically design the quintuplet and enable maximum diversity transmissions over various channels, such as time-or frequency-selective channels, and multiple-input multiple-output (MIMO) fading channels. Moreover, grouping (a.k.a. vector) technique is incorporated in the design of ULAM, called GULAM to offer tradeoffs among performance, complexity, and peak-to-average power ratio. Simulation results corroborate all theoretical findings and propositions, and demonstrate promising performance and flexibility of the proposed ULAM and GULAM schemes.
Traditional digitally implemented iterative inference tasks, such as sparse channel estimation, typically rely on frequent large-scale binary discrete fixed-point operations and data movement, and therefore suffer from high computation power consumption and a pronounced memory-wall bottleneck. To address this issue, this paper proposes a hardware-native variational message passing architecture for memristor-based in-memory computing, termed VMP-IMC. The proposed method decomposes the iterative probabilistic inference into a linear projection stage and a nonlinear shrinkage stage, which are implemented by quasi-static differential memristor arrays and lightweight digital proximal operators, respectively. This hardware-oriented decomposition reduces repeated high-dimensional digital arithmetic and data movement, thereby better matching the computing characteristics of crossbar-based in-memory computing. Meanwhile, the unavoidable nonidealities of the IMC substrate, such as quantization truncation and conductance drift, are explicitly modeled as state-dependent heteroscedastic hardware uncertainty and incorporated into the inference operator design, so as to suppress their impact on the iterative trajectory and convergence behavior. Simulation results show that, under the same low-precision quantization and significant device-drift constraints, the proposed VMP-IMC achieves more than 1.5 dB performance gain and approximately 6:3× reduction in energy per channel estimation compared with conventional fully digital algorithms. These results demonstrate that the proposed architecture effectively improves the co-adaptation between complex sparse inference and memristor-based IMC hardware, and provides a viable path toward low-power and high-efficiency wireless baseband processing.
