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

Abstract: Large antenna arrays allow wireless systems to serve more users and achieve higher data rates, but they also make channel feedback expensive: the receiving device must repeatedly report a large complex-valued channel matrix to the base station. Most neural compressors treat this matrix like an image and replace it with a fixed-length code that only a matched neural decoder can interpret. The message therefore does not adapt to channel complexity, and changing the antenna count typically requires retraining. We ask whether a device can instead report only the few dominant propagation paths underlying each channel. We introduce the Gramian Chebyshev Neural Operator (GCNO), a physics-based, variable-rate compressor that identifies a sample-dependent set of path directions. GCNO uses receive-transmit channel structure to locate paths, a first-order Taylor correction to refine directions that fall between grid points, and least squares to recover their complex strengths. It is trained without path labels, and the base station reconstructs the channel analytically from the transmitted path tuples rather than through a learned decoder. Across three ray-traced environments, GCNO achieves better reconstruction accuracy at the same payload - or lower payload at the same accuracy - than neural feedback baselines, and transfers to unseen antenna counts without retraining.

• The paper introduces GCNO, a label-free neural-operator codec that represents wireless channels with variable-length path tuples and reconstructs them analytically, achieving up to -19.23 dB median NMSE at 16 floats with roughly 100× fewer parameters than SwinCFNet.
• The method combines input-adaptive Gramian-Chebyshev filtering, Taylor-based off-grid direction refinement, adaptive path selection, and ridge least-squares gain recovery to avoid fixed grids and jointly trained decoders.
• The paper demonstrates transfer across antenna sizes and propagation scenes, with beamforming performance within approximately 0.004 bit/s/Hz of perfect CSI, while identifying wideband OFDM, richer multipath, and noisy estimation as important open challenges.

Motivation and problem statement

In FDD massive MIMO, the user equipment (UE) must report downlink channel state information (CSI) to the base station (BS), and for an $N_r \times N_t$ array pair this is a complex matrix requiring $2N_rN_t$ real numbers. The dominant learned approach—initiated by CsiNet and extended through convolutional, recurrent, attention, and transformer variants—compresses the matrix into a fixed-length latent code that only a jointly trained neural decoder can interpret. This design has two structural drawbacks: feedback length does not adapt to per-channel complexity, and changing the antenna count generally requires retraining both encoder and decoder.

The paper under review, "GCNO: Gramian Chebyshev Neural Operator for Physics-Based Compression of Wireless Channels" (2608.18522), proposes instead to feed back a variable-length list of dominant propagation paths. Under the geometric channel model, each path contributes $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$, so a channel dominated by $K$ paths needs only $4K$ real values (complex gain plus two effective spatial angles). The BS reconstructs analytically from these tuples using known array responses; no matched decoder exists at all. Because path geometry is a property of the environment rather than of the array, the same representation applies across antenna counts, and the payload naturally tracks channel richness.

The central technical obstacle is estimating continuous path directions. Grid-based estimators spread off-grid path energy across neighboring cells, so several grid coefficients may be needed per physical path; directly regressing continuous directions with a neural network yields a highly non-convex objective due to rapidly oscillating steering phases. The paper resolves this with a hybrid pipeline in which the grid serves only as an intermediate search structure.

Method

GCNO operates on a matched-filter evidence map $C_{ij}=\langle \mathbf D_{ij},\mathbf H\rangle_F$ computed against a fixed dictionary of candidate array-response atoms $$\mathbf D_{ij}=\mathbf a_{N_r}(u_i^r)\mathbf a_{N_t}(u_j^t)^{\mathsf H}$$ over a $28\times28$ direction grid. Both the raw normalized channel and the evidence map are encoded by lightweight stems, fused on the direction grid, and processed by three GCNO layers.

The GCNO backbone is the paper's architectural contribution. At each layer it forms sample-specific receive- and transmit-side Gramians $$\mathbf G_r=\sum_c \mathbf X_c\mathbf X_c^{\mathsf H}$$ and $$\mathbf G_t=\sum_c \mathbf X_c^{\mathsf H}\mathbf X_c$$, normalizes them into operators $2N_rN_t$0 with spectra in $2N_rN_t$1, and filters features bilaterally:

$2N_rN_t$2

Learned Chebyshev coefficients are indexed only by feature channels and polynomial orders—not by antenna or grid position—which is what confers size flexibility. Unlike FNO's fixed Fourier basis or static graph-Chebyshev filters, the Gramians are recomputed per input, so filtering adapts to each channel's correlation structure while remaining a neural operator across discretizations.

Three downstream components complete the pipeline. Off-grid refinement: each candidate direction receives bounded Taylor corrections ($2N_rN_t$3), and the corrected atom is approximated by a first-order expansion $2N_rN_t$4 using fixed analytical derivative dictionaries, so no gradient flows through newly generated steering vectors. Adaptive selection: candidates are scanned by score order, admitted only if they reduce reconstruction error beyond a threshold, yielding a data-dependent $2N_rN_t$5. Least-squares gains: complex gains are recovered by joint ridge LS rather than predicted by the network. Training replaces hard selection with a smooth soft-attention surrogate over at most eight candidates and optimizes a log-NMSE rate-distortion objective with duplicate, offset, and score regularizers—entirely label-free; ground-truth paths never enter training, validation, or checkpoint selection.

Theoretical analysis

Two appendix results give the framework formal grounding. First, a chain-rule argument shows that any differentiable encoder–decoder that exactly reconstructs locally identifiable $2N_rN_t$6-path channels requires a latent of dimension at least $2N_rN_t$7; the path tuple attains this bound exactly, and a fixed-width autoencoder must pay $2N_rN_t$8 even when most channels need fewer paths. This is deliberately narrow—it covers unquantized, locally exact codecs and says nothing about lossy or entropy-coded schemes—but it establishes that variable-rate path feedback is not merely competitive but information-theoretically minimal within its scope. Second, Lemma B.1 factorizes the evidence map as $2N_rN_t$9 with rank at most $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$0, so the Gramian subspaces spanned by GCNO's filters lie inside the physical footprint spans; when footprints are separated, the linear core reduces to applying a learned bivariate polynomial response $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$1 to each path. Normalization guarantees spectral norms of all Chebyshev factors are at most one, giving a stable, size-independent bound on layer outputs.

Empirical results

Evaluation uses DeepMIMO ray-traced channels from ASU, Dallas, and Seattle (10,000/2,000/1,500 splits), against eight retrained encoder–decoder baselines including TransNet, StarCANet, and SwinCFNet. The headline findings are strong:

• Rate–distortion dominance. GCNO sits on the best quantized RD frontier across all three environments. At 16 transmitted floats it reaches median NMSE of $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$2 dB (Seattle) and $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$3 dB (Dallas), versus $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$4 and $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$5 dB for SwinCFNet—the strongest baseline—at roughly 100× fewer trainable parameters (95K vs. 10.9M).
• Taylor refinement is essential. Retraining without Taylor offsets collapses median NMSE to around $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$6 to $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$7 dB and degrades angular error from ~0.23° to ~2°, consistently across environments and payloads (improvements of at least 6.6 dB in nine additional comparisons). Off-grid correction, not architecture alone, makes one-tuple-per-path compression viable.
• Backbone ablation. Replacing only the three GCNO layers with CNN, dilated-CNN, FNO, static-Chebyshev, or learned-polynomial backbones—keeping inputs, selection, and LS identical—worsens the RD trade-off, so the gain cannot be attributed solely to the analytical decoder.
• Array-size transfer. A model trained at $$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$8 transfers without any parameter update to six other configurations ($$g_k\,\mathbf{a}_{N_r}(u_{r,k})\mathbf{a}_{N_t}(u_{t,k})^{\mathsf H}$$9 through $K$0), retaining strong accuracy where frozen baselines adapted by a deterministic interpolation rule lose most of theirs.
• Cross-scene transfer. Frozen models evaluated across scene pairs show GCNO best in all six directions, improving over SwinCFNet by 0.17–8.44 dB.
• Downstream utility. Beamforming with reconstructed CSI achieves spectral efficiency within ~0.004 bit/s/Hz of perfect CSI at 20 dB SNR (single stream), with mean squared alignment of the dominant right singular vector of 0.9973.
• Robustness checks. A baseline-favorable seed study on Seattle (the closest comparison) retains positive margins for every GCNO seed and payload target, with a minimum margin of 0.65 dB over SwinCFNet. A fixed-$K$1 control also beats Grid-OMP by nearly 12 dB at equal payload, indicating the fixed dictionary plus LS alone does not explain the result.

Physical fidelity diagnostics confirm the representation is meaningful: strong-path F1 of 0.832 and median joint angle error of 0.225°, exceeding the closest baselines, though recall drops on richer channels—an expected consequence of compactness.

Limitations and open questions

The authors are explicit about scope. All experiments concern narrowband spatial channels; extension to wideband OFDM requires adding a delay coordinate per path, which the authors argue may improve compressibility further but do not demonstrate. The identifiability theorem excludes coincident paths, diffuse energy, quantization, and entropy coding. Cross-array transfer is established only for seven uniform-linear-array shapes under a common adapter protocol, not for arbitrary geometries or calibration changes. Recall limitations on multipath-rich channels remain, and the Grid-OMP comparison is acknowledged as a single lightweight classical control rather than a survey of sparse-recovery methods. Whether the Gramian construction retains its advantage at substantially larger grids or under realistic estimation noise is left open.

Conclusion

This paper reframes CSI compression as label-free extraction of a variable-length set of physical path parameters, supported by a channel-adaptive neural-operator backbone whose Gramian–Chebyshev filters share coefficients across array sizes. The combination of a provably minimal $K$2 representation, Taylor-corrected off-grid directions, LS gain recovery, and analytic BS-side reconstruction yields consistent rate–distortion improvements over eight neural baselines at a fraction of their parameter count, along with robust transfer across antenna counts and propagation scenes. The main open question is whether the same physics-aligned formulation extends to wideband, delay-aware channel models without sacrificing its current simplicity.