GCNO: Gramian Chebyshev Neural Operator for Physics-Based Compression of Wireless Channels

Technical Summary: GCNO for Physics-Based Compression of Wireless Channels

Problem Statement
Modern wireless systems utilizing massive MIMO (Multiple-Input Multiple-Output) with large antenna arrays face a significant bottleneck in Channel State Information (CSI) feedback. In Frequency-Division Duplex (FDD) systems, user devices must report the downlink channel matrix to the base station (BS). As antenna counts ($N_r \times N_t$) increase toward 6G standards, the feedback overhead becomes prohibitive, requiring $2N_rN_t$ real numbers to describe a single channel matrix.

Existing deep learning (DL) approaches typically treat the channel matrix as an image, compressing it into a fixed-length latent code via a neural encoder and reconstructing it via a matched neural decoder. This paradigm suffers from two primary limitations:

1. Inflexible Payload: Fixed-length codes cannot adapt to the varying complexity of different channels, preventing feedback overhead from scaling with the actual information content.
2. Lack of Generalization: Encoder-decoder pairs are jointly trained for specific antenna configurations. Changing the number of antennas generally requires retraining both networks, hindering deployment in dynamic or evolving hardware environments.

Methodology: Gramian Chebyshev Neural Operator (GCNO)
The authors propose GCNO, a physics-based, variable-rate compressor that shifts the representation from a fixed latent code to a sparse set of dominant propagation paths. The core hypothesis is that outdoor and high-frequency channels are dominated by a small number of strong paths, which can be described by complex gains and spatial directions rather than a full matrix.

The GCNO pipeline operates as follows:

1. Physics-Based Representation: Instead of learning a latent space, the system aims to recover $K$ dominant paths, where each path $k$ is defined by a complex gain $g_k$ and transmit/receive spatial directions $(u_{t,k}, u_{r,k})$. The feedback payload consists of $4K$ real values (real/imaginary parts of gain and effective angles). The BS reconstructs the channel analytically using the known array response model, eliminating the need for a learned decoder.

2. Evidence Map Construction: To avoid the non-convexity of directly predicting continuous directions, GCNO first constructs a "direction-space evidence map." This is done by computing the inner product between the measured channel and a dictionary of precomputed array responses on a fixed grid. This map highlights candidate direction pairs but suffers from "off-grid" leakage, where a single physical path spreads energy across multiple grid cells.

3. Gramian Chebyshev Neural Operator (GCNO):

   • Adaptive Filtering: Standard neural operators use fixed filters (e.g., Fourier or graph Laplacians) that do not adapt to the specific structure of a given channel. GCNO introduces a novel mechanism where it constructs Gramian matrices from the receive and transmit feature correlations of the current channel sample.
   • Chebyshev Filtering: These sample-specific Gramians are used to define Chebyshev polynomial filters. The learned coefficients (polynomial weights) are shared across all antenna configurations, but the filtering operation itself adapts to the specific channel structure via the Gramians. This allows the model to process inputs of varying sizes (different $N_r, N_t$) without retraining.
   • Output: The network predicts a score map (identifying likely path locations) and continuous Taylor offsets to refine the grid-based directions.

4. Off-Grid Refinement and Selection:

   • Taylor Correction: Selected grid directions are refined using a first-order Taylor expansion to move them off the discrete grid, correcting the "off-grid" error.
   • Adaptive Path Selection: A greedy selection process iteratively adds paths that significantly reduce the channel reconstruction error. This allows $K$ to vary per channel, enabling variable-rate compression.
   • Least Squares (LS) Recovery: Once directions are selected, the complex gains are recovered using a standard least-squares solver. The neural network does not predict gains, only directions and selection priorities.

5. Training: The model is trained in a label-free manner using only the channel reconstruction error (NMSE). Ground-truth path directions, gains, or path counts are not used. During training, a differentiable approximation of the hard selection process is used; at inference, a deterministic hard selection is applied.

Key Contributions

1. Variable-Rate Path Compression: The paper formulates CSI feedback as a variable-length list of path tuples. This adapts the feedback size to the channel's complexity (number of dominant paths) and allows the BS to reconstruct the channel analytically without a matched neural decoder.
2. Gramian Chebyshev Neural Operator: The introduction of GCNO, which constructs channel-specific filters from receive-transmit correlations while sharing learned coefficients across antenna counts. This design improves path recovery and enables transfer to unseen antenna configurations.
3. Continuous Refinement and LS Recovery: The method refines directions within grid cells using Taylor offsets and recovers gains via LS, all without requiring path labels. Ablation studies confirm that removing the continuous refinement significantly degrades performance.

Experimental Results
The authors evaluated GCNO on ray-traced datasets from three distinct environments: Arizona State University (ASU), Dallas, and Seattle.

• Rate-Distortion Performance: Across all environments, GCNO achieved lower Normalized Mean Squared Error (NMSE) at the same payload compared to eight state-of-the-art neural feedback baselines (including CsiNet, CRNet, and SwinCFNet). Conversely, it achieved the same accuracy with significantly fewer transmitted bits.
• Generalization to Unseen Antenna Counts: A model trained on a $32 \times 32$ array configuration was applied directly to six other array sizes (e.g., $16 \times 16$, $32 \times 64$) without retraining. GCNO maintained high reconstruction accuracy, whereas neural baselines suffered significant performance drops when the antenna count changed.
• Ablation Studies:
  • Removing the Taylor refinement (off-grid correction) caused a massive increase in NMSE (e.g., from -20 dB to -3.8 dB in ASU), proving the necessity of continuous direction estimation.
  • Replacing the GCNO backbone with standard CNNs, FNOs, or polynomial filters degraded performance, indicating the benefit of the sample-specific Gramian filtering.
• Efficiency: GCNO has only ~95K trainable parameters, significantly fewer than the best-performing baseline (SwinCFNet, ~11M parameters).

Significance and Claims
The paper claims that representing wireless channels through their underlying multipath structure, rather than as fixed-size latent codes, offers a superior trade-off between feedback overhead and reconstruction accuracy. The primary significance lies in the physics-aligned design:

• By relying on the geometric channel model for reconstruction, the method decouples the compression algorithm from the specific antenna configuration, enabling robust transfer to unseen hardware.
• The variable-rate nature of the feedback allows the system to adapt to channel complexity naturally, avoiding the waste of transmitting fixed-length codes for simple channels.
• The approach is trained solely on reconstruction error, making it practical for deployment without the need for ground-truth path annotations.

The authors conclude that while their experiments focus on narrowband spatial channels, the shared structure of path directions and delays suggests that this approach could be extended to wideband OFDM systems, potentially offering even greater compressibility.