Semi-Gridless Variational Bayes Channel Estimation in XL-MIMO: Near-Field Modeling and Inference

1. Problem Statement

The transition to sixth-generation (6G) communications involves the deployment of Extremely Large Antenna Arrays (ELAAs) and high-frequency bands (mmWave/THz). These technologies fundamentally alter electromagnetic propagation:

• Near-Field Dominance: As array apertures and frequencies increase, the Rayleigh distance (the boundary between near-field and far-field) extends to hundreds of meters. Consequently, many communication links operate in the near-field region, where wavefronts are spherical rather than planar.
• Modeling Challenge: Conventional channel estimation (CE) methods rely on the far-field assumption (planar wavefronts characterized solely by Angle of Arrival/AoA). In the near-field, the channel depends on both the Direction of Arrival (DoA) and the distance (range) to the scatterer/user.
• Limitations of Existing Methods:
  • Subspace methods (e.g., MUSIC): High computational complexity due to exhaustive searches and require large sample sizes to overcome rank deficiency.
  • Compressive Sensing (e.g., P-SOMP, P-SIGW): Rely on discrete codebooks (grids) for angles and distances. This introduces "grid mismatch" errors when true parameters fall between grid points and incurs storage/latency overhead.
  • Deep Learning: Requires massive training data and suffers from generalization issues in environments different from the training set.
  • Sparse Bayesian Learning (SBL): Often requires prior knowledge of sparsity patterns and struggles with general array geometries (e.g., extending from Uniform Linear Arrays to Uniform Planar Arrays).

2. Methodology

The authors propose a Semi-Gridless Variational Bayesian (SG-VB) algorithm. The core innovation lies in reformulating the near-field channel model to separate frequency (angle) and distance components, allowing for distinct estimation strategies for each.

A. New Channel Representations

The paper derives new mathematical models for both Uniform Linear Arrays (ULA) and Uniform Planar Arrays (UPA) using the Fresnel approximation.

• ULA Model: The steering vector is decomposed into a standard array response $a(\omega)$ (dependent on DoA) and a distance-dependent vector $c(s)$.
• UPA Model: The steering vector is decomposed into three components: $a(\omega)$ (azimuth), $c(\psi)$ (elevation), and $d(s)$ (distance).
• Key Insight: This separation allows the algorithm to treat DoA estimation as a continuous frequency estimation problem and distance estimation as a separate parameter search.

B. Variational Bayesian Inference Framework

The SG-VB algorithm approximates the intractable posterior distribution of channel parameters using a mean-field variational distribution.

• Latent Variables: The algorithm infers distributions for path gains ($\nu$ or $\alpha$), DoAs ($\omega, \psi$), distances ($s$), and noise precision ($\gamma$).
• Sparsity Enforcement: A Gamma-Gaussian prior is used for path gains to enforce channel sparsity, enabling closed-form Bayesian updates.

C. The "Semi-Gridless" Strategy

The algorithm employs a hybrid approach to balance accuracy and complexity:

1. Gridless DoA Estimation:
   • Instead of a discrete grid, the DoA parameters are modeled using a von Mises distribution (the circular analog of the Gaussian distribution).
   • This allows for closed-form updates of the variational parameters, eliminating spectral discretization artifacts (grid mismatch) and enabling continuous DoA estimation.
2. Coarse-to-Fine Distance Estimation:
   • Distance estimation utilizes a coarse-to-fine grid search.
   • Step 1: A coarse grid search identifies the likely distance range.
   • Step 2: A Newton-Raphson iterative refinement (using first and second derivatives of the log-likelihood) fine-tunes the distance estimate to sub-grid accuracy.

D. Algorithm Execution

The algorithm uses Coordinate Ascent Variational Inference (CAVI) to iteratively update the variational distributions of all latent variables until convergence. Closed-form update rules are derived for all variables, ensuring computational efficiency.

3. Key Contributions

1. Novel SG-VB Algorithm: A low-complexity, scalable channel estimation method applicable to both ULAs and UPAs without requiring prior knowledge of subarray sparsity.
2. Hybrid Estimation Strategy:
   • Gridless DoA: Uses von Mises distributions to estimate continuous angles, avoiding grid mismatch.
   • Refined Distance: Uses a coarse-to-fine search with Newton-Raphson updates for high-precision range estimation.
3. Closed-Form Derivations: The authors derive closed-form variational distributions for all latent variables, enabling efficient updates and the direct estimation of scatterer positions (crucial for advanced transceiver design).
4. Generalizability: The framework is explicitly designed to handle both 1D (ULA) and 2D (UPA) array geometries, a significant improvement over many existing methods limited to ULAs.

4. Simulation Results

Extensive simulations were conducted comparing SG-VB against benchmarks: MUSIC, P-SOMP, P-SIGW, SBL, Least Squares (LS), and Oracle LS.

• Channel Reconstruction (NMSE):
  • SG-VB consistently outperforms all baselines across the entire Signal-to-Noise Ratio (SNR) range.
  • At high SNRs, SG-VB approaches the performance of the Oracle LS (the theoretical lower bound), significantly outperforming grid-based methods (MUSIC, P-SOMP) which suffer from performance saturation due to grid mismatch.
• DoA Estimation (MSE):
  • SG-VB achieves superior angle estimation accuracy, particularly at high SNRs (>10 dB), offering up to 20 dB gain over grid-based methods.
  • Grid-based methods saturate quickly as SNR increases because they cannot resolve angles outside their discrete grid.
• Distance Estimation (NMSE):
  • SG-VB provides robust distance estimation across various user distances (5m to 135m).
  • While P-SIGW performs well at very short distances, SG-VB maintains superior performance across the full range and is significantly more computationally efficient.
• Array Geometry:
  • The method demonstrates effectiveness for both ULAs and UPAs. For UPAs, it outperforms 3D-MUSIC, which suffers from error propagation when estimating distance based on pre-estimated DoAs.
• Complexity:
  • The computational complexity is $O(L(N^2 + KN))$, which is higher than P-SOMP but offers significantly better accuracy. It is more scalable than MUSIC ($O(N^3)$) for large arrays.

5. Significance

This work addresses a critical bottleneck in 6G XL-MIMO systems: the need for accurate, low-complexity channel estimation in the near-field region.

• Enabling Near-Field Technologies: By accurately estimating both angle and distance, the proposed method unlocks the potential of near-field beamforming and multiplexing gains.
• Scalability: The ability to handle UPAs without a massive increase in complexity makes it suitable for future massive MIMO deployments.
• Practicality: The "semi-gridless" approach offers a practical compromise, avoiding the high complexity of fully gridless methods while eliminating the accuracy limitations of fixed-grid compressive sensing.
• Advanced Applications: The ability to estimate scatterer locations directly supports advanced transceiver designs, such as near-field beam focusing and integrated sensing and communication (ISAC).