Wireless channel maps now follow receiver resolution: SpecFlowAR reaches 2.03 dB BPL@5 under a unified budget

The learning target for wireless channels changes from variable-length multipath component lists to fixed-grid angle power spectra and power delay profiles defined at a practical receiver's finite aperture and finite bandwidth. Most ray-tracing benchmarks previously supplied supervision as variable-size multipath component lists, while a practical receiver observes spectra only after finite-aperture beamforming and finite-bandwidth filtering, so converting lists into spectra requires explicit response kernels and discretization rules. UrbanADPSMap fixes the receiver-domain kernels, discretization and validity mask, so different methods train and are evaluated on the same target support, removing reconstruction-induced target ambiguity.

SpecFlowAR splits joint spectrum generation into a coarse-to-fine multi-scale process, first producing a low-resolution dispersion envelope and then refining beam-relevant peaks and low-power tails. Single-scale flow matching keeps the same flow-matching objective and receiver-spectrum interface but removes the coarse-to-fine autoregressive pyramid, so the difference between the two isolates progressive refinement. On the test set under a unified budget, SpecFlowAR gives BPL@5 of 2.03 dB, delay-spread error of 29.12 ns and azimuth-spread-of-arrival error of 10.82 degrees, all better than single-scale flow matching at 2.42 dB, 29.76 ns and 11.87 degrees.

The sampling budget is controlled explicitly by the number of function evaluations, so latency and quality for dense spectrum queries can be adjusted. The default setting uses four scales with 20 solver steps each, for 80 function evaluations in total, and iterative generators are compared at matched budgets from 12 to 80. At 80 function evaluations SpecFlowAR reports the lowest values on all three main metrics, but the budget sweep shows that increasing function evaluations does not yield a universal monotonic gain, and the preferred operating point depends on the metric.

The demonstrated setting is ray-traced urban scenes with carrier frequency, bandwidth, a 16-element half-wavelength uniform linear array and angle-delay grids fixed as stated, with training, validation and test splits separated by scene and test scenes unseen during training. The authors list calibration of these choices against measured channels and extension to variable hardware settings as future work.

The budget sweep shows that increasing function evaluations does not yield a universal monotonic gain, and the preferred operating point varies with the metric; spectral fidelity and receiver-facing metrics rank methods differently, with single-scale flow matching lower on APS RMSE while SpecFlowAR is lower on APS JSD, PDP RMSE and PDP JSD.