Abstract
In the emerging high mobility vehicle-to-everything (V2X) communications using millimeter wave (mmWave) and sub-THz, multiple-input multiple-output (MIMO) channel estimation is an extremely challenging task. At mmWaves/sub-THz frequencies, MIMO channels exhibit few leading paths in the space-time (ST) domain (i.e., directions or arrival/departure and delays). Algebraic low-rank (LR) channel estimation exploits ST channel sparsity through the computation of position-dependent MIMO channel eigenmodes leveraging recurrent training vehicle passages in the coverage cell. LR requires vehicles’ geographical positions and tens to hundreds of training vehicles’ passages for each position, leading to significant complexity and control signaling overhead. Here, we design a deep-learning (DL)-based LR channel estimation method to infer MIMO channel eigenmodes in V2X urban settings, starting from a single least squares (LS) channel estimate and without needing vehicle’s position information. Numerical results show that the proposed method attains comparable mean squared error (mse) performance as the position-based LR. Moreover, we show that the proposed model can be trained on a reference scenario and be effectively transferred to urban contexts with different ST channel features, providing comparable mse performance without an explicit transfer learning procedure. This result eases the deployment in arbitrary dense urban scenarios.
What Problem Does This Paper Address?
As we move toward 6G, Vehicle-to-Everything (V2X) communications are shifting to millimeter-wave (mmWave) and sub-THz frequencies to find more bandwidth. However, these high frequencies bring a major challenge: channel estimation.
In high-mobility environments, the wireless channel changes rapidly. Traditional Least Squares (LS) estimation is often too noisy. A more advanced technique, Algebraic Low-Rank (LR) estimation, exploits the fact that these channels are "sparse," meaning signals only travel along a few dominant paths (as visualized in Figure 4). While LR is highly accurate, it traditionally requires precise geographical coordinates of the vehicle and hundreds of previous vehicle passages at that exact spot to "learn" the local environment. In a dense city, managing this position-dependent data creates a massive signaling and computational burden. The authors ask: Can we get the accuracy of Low-Rank estimation without needing to know exactly where the car is?
Methods: DL-Based Low-Rank Estimation
The authors propose a DL-based LR channel estimation method. Instead of looking up a table based on GPS, a Deep Neural Network (DNN) "sees" a single, noisy LS estimate and predicts the underlying space-time (ST) eigenmodes of the channel.
Key Technical Features: * Architecture: The DNN uses convolutional layers to extract spatial features from the input channel matrix, followed by fully connected layers. * The "Thin SVD" Layer: To ensure the output matrices are mathematically valid (unitary), the authors incorporate a "thin" Singular Value Decomposition (thSVD) operator at the end of the network. This projects the network's output onto the Stiefel manifold. * Modal Filtering: The inferred eigenmodes are used to create a projection matrix ($\Pi_{DL}$) that filters out noise from the initial LS estimate, effectively "denoising" the channel.
Datasets and Simulation Environment
The study uses high-fidelity simulated data rather than generic stochastic models to capture the nuances of urban propagation.
- Scenarios: Five distinct urban road crossings in Milan, Italy, ranging from standard intersections to complex roundabouts (see Figure 3).
- Traffic Simulation: Realistic vehicle trajectories generated via SUMO (Simulation of Urban MObility).
- Channel Modeling: Ray-tracing performed with Altair WinProp at 28 GHz, capturing specific Direction of Arrival (DoA), Direction of Departure (DoD), and delays for each geographical point.
- Scale: A dataset of $2.5 \times 10^5$ channel samples per scenario.
Key Insights and Results
- Roads as Constraints: The core insight is that urban geometry (buildings and roads) constrains wireless paths. Because vehicles follow predictable paths, the "modes" of the channel are highly structured. A DNN can learn these patterns globally for a cell, removing the need for per-meter positioning.
- Massive Accuracy Gains: The DL-based method outperforms standard LS estimation by approximately 15 dB in Mean Squared Error (MSE), matching the performance of position-based LR.
- Implicit Transferability: Surprisingly, a model trained on one intersection performs remarkably well on completely different intersections without any retraining for frequency-flat channels. The network learns a generalized mapping of urban sparsity.
- Fine-Tuning for Complexity: In frequency-selective cases (50 MHz bandwidth), retraining just the last two layers of the DNN on the new scenario improved performance by about 2 dB, showing that the bulk of the learned features are universal.
Conclusion
By moving from a position-based lookup to a data-driven inference, the proposed system reduces control signaling overhead and eliminates the need for precise GPS data. This makes high-performance MIMO channel estimation much more feasible for real-world 6G deployment in dense urban environments.
Summary updated