Point-Cloud-Assistant Localized Statistical Channel Prediction by Tangent Gaussian Splatting

Problem Statement

Accurate, site-specific channel information is essential for optimizing next-generation wireless networks. While Localized Statistical Channel Modeling (LSCM) has emerged as a state-of-the-art method for estimating the Angular Power Spectrum (APS) from low-overhead Reference Signal Received Power (RSRP) measurements, it suffers from a critical limitation: lack of spatial generalization. LSCM can only reconstruct APS at locations where RSRP measurements exist. It cannot interpolate or extrapolate APS to unmeasured regions without nearby measurements, severely restricting its utility for large-scale network planning and digital twin construction.

Existing alternatives fail to address this specific gap:

• Geometry-based stochastic models (GBSM) lack localized environmental detail.
• Deterministic ray tracing requires prohibitive computational costs and precise material parameters.
• Deep learning radio map predictors (e.g., RadioUNet) typically predict scalar signal power rather than the high-dimensional APS required for beamforming and antenna configuration.
• Neural scene representations (e.g., NeRF, 3DGS adapted for wireless) often require dense ground-truth APS labels for training, which are rarely available in outdoor deployments.

The paper defines the Localized Statistical Channel Prediction (LSCP) problem: predicting the APS at completely unmeasured outdoor grids by leveraging sparse, low-dimensional RSRP data and dense environmental priors (specifically LiDAR point clouds).

Methodology: PC-TGS Framework

The authors propose Point-Cloud-Assisted Tangent Gaussian Splatting (PC-TGS), a unified framework that repurposes 3D Gaussian Splatting (3DGS) for the radio domain. Unlike visual 3DGS which renders RGB colors, PC-TGS synthesizes radio signal propagation. The framework consists of four core modules:

1. Relaxed-Mean Reparameterization (RM):

   • To handle dense and noisy LiDAR point clouds, the framework does not use raw points directly. Instead, it employs a differentiable "soft selection" mechanism to extract $N$ representative virtual scatterers.
   • Each scatterer's position is a trainable, weighted aggregation of the raw point cloud, subject to simplex constraints and a "Maximum Element Constraint" (MEC) to ensure each scatterer anchors to a specific physical region. This filters noise while preserving key geometric features like building facades.

2. Attributes Construction:

   • Each virtual scatterer $n$ is parameterized by:
     • Position ($\mu_n$) and Covariance ($\Sigma_n$): Encodes the location and anisotropic spatial extent (orientation/spread) of the scatterer.
     • Complex Attenuation ($\alpha_n$): A learnable magnitude and phase shift, modeled via an MLP based on the distance from the base station to capture path loss and material interaction.
     • Directional Gain ($G_n$): Modeled using Complex Spherical Harmonics (SH) to capture angle-dependent scattering patterns from the Base Station (BS) to the scatterer and from the scatterer to the receiver.

3. Tangent-Plane Projection:

   • To map 3D scatterers to the 2D angular domain (AoD/AoA), the framework projects all Gaussians onto a local tangent plane at each discretized angular bin.
   • This projection aligns the environmental geometry with the angular domain used for APS computation, differing fundamentally from the camera-based projections used in visual 3DGS.

4. APS Synthesis via Electromagnetic Splatting:

   • The framework aggregates contributions from scatterers using a depth-aware accumulation rule. Scatterers are sorted by distance (front-to-back) to simulate occlusion and shadowing.
   • The final APS is computed by summing the coherent contributions of multipath components, where Gaussian opacities represent signal attenuation.
   • GWA-based Integration: To avoid computationally expensive numerical integration over spherical patches, the authors derive a closed-form Gaussian Weighted Average (GWA) approximation. This replaces the hard-edged integration region with a moment-matched Gaussian window, providing a tractable formula with a provable error bound (Theorem 1).

5. Recursive Fine-Tuning (RFT):

   • To enhance extrapolation, the model employs a boundary-propagation strategy. It initially trains on measured grids, then iteratively predicts APS at boundary grids, augments the training set with these predictions, and retrains. This allows the model to propagate learned environmental structure into unmeasured regions.

Key Contributions

1. LSCP Framework: The first formal definition and solution for extrapolating APS to unmeasured locations using sparse RSRP and dense LiDAR priors, moving beyond the interpolation limits of LSCM.
2. Physics-Inspired 3DGS: A rigorous reinterpretation of 3D Gaussian Splatting for radio propagation, where Gaussians act as virtual scatterers and the rendering process simulates multipath superposition and occlusion.
3. Efficient Scatterer Modeling: A relaxed-mean reparameterization algorithm that efficiently extracts robust virtual scatterers from noisy, dense point clouds, bridging raw sensing data with statistical channel modeling.
4. Tangent-Plane Projection & GWA: A scalable projection method and a closed-form integration formula (GWA) with a provable error bound, enabling fast and mathematically principled APS synthesis.
5. Unified Pipeline: An end-to-end differentiable framework that jointly integrates sparse measurements, dense geometry, and electromagnetic modeling, achieving robust performance in data-sparse environments.

Experimental Results

The framework was evaluated on a city-scale dataset (5 million LiDAR points, 6,310 RSRP samples) in a 1,000 $\times$ 1,200 $m^2$ area.

• Performance: PC-TGS significantly outperformed state-of-the-art baselines (including WNOMP, TMSBL, Ray Tracing, RadioUNet, and MM-LSCM) in extrapolation tasks (Sub-Tasks 2 and 3).
  • Sub-Task 2 (Unmeasured RSRP): PC-TGS achieved 7.45 dB MAE, compared to 9.12 dB (MM-LSCM) and 10.31 dB (RadioUNet).
  • Sub-Task 3 (Unmeasured Rotated RSRP): PC-TGS achieved 7.57 dB MAE, compared to 9.94 dB (MM-LSCM).
• Efficiency: PC-TGS offers fast inference (~74 ms per query), representing a ~5$\times$ speedup over MM-LSCM and significantly faster than online methods like Ray Tracing or Kriging-based extrapolation which take seconds per query.
• Ablation Studies:
  • Scatterer Count: Performance stabilized at $N=2000$ virtual scatterers; increasing further degraded accuracy and speed.
  • LiDAR Density: The model remained robust even with 50% random point cloud sparsification. Removing LiDAR entirely (SM-PC-TGS) caused a significant performance drop, confirming the value of geometric priors.
  • Noise Robustness: The framework tolerated geometric noise up to 1.0 m, supporting the use of image-derived point clouds (e.g., DUSt3R) as a LiDAR alternative.
  • RFT Impact: The Recursive Fine-Tuning scheme specifically improved extrapolation accuracy (Sub-Tasks 2 & 3) by ~0.8–1.0 dB without compromising performance on measured regions.

Significance and Claims

The paper claims that PC-TGS addresses a fundamental gap in wireless network optimization: the ability to predict channel statistics in areas where no signal measurements exist. By fusing sparse radio data with dense environmental geometry in a physically grounded, differentiable framework, PC-TGS enables:

• Geometry-Aware Prediction: Leveraging the physical structure of the environment (buildings, walls) to infer multipath behavior in unmeasured locations.
• Data Efficiency: Reducing the need for exhaustive drive tests by utilizing LiDAR priors to generalize from sparse RSRP samples.
• Scalability: Providing a computationally efficient alternative to ray tracing for large-scale digital twin construction.

The authors emphasize that PC-TGS is not merely an application of visual 3DGS but a fundamentally redesigned framework grounded in radio propagation physics, capable of true spatial extrapolation where previous methods fail.