arxiv: 2605.14989 · v1 · submitted 2026-05-14 · 📡 eess.SP

Recognition: unknown

Map2APS: A Physically Grounded Benchmark for Direct Angle Power Spectrum Prediction from Urban Geometry

Junxi Huang , Xiucheng Wang , Nan Cheng , Kailong Wang , Ruijin Sun , Zhisheng Yin

Authors on Pith no claims yet

Pith reviewed 2026-05-15 03:16 UTC · model grok-4.3

classification 📡 eess.SP

keywords angle power spectrumurban geometryintelligent ray tracingbenchmarkbeam managementMIMOmap-to-APSgeneralization

0 comments

The pith

Map2APS benchmark enables direct prediction of angle power spectra from urban maps with 0.948 cosine similarity on unseen layouts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Map2APS, a benchmark dataset built from intelligent ray-tracing records to support direct prediction of angle power spectra from urban geometry inputs. It spans 51 equal-height maps and 2.55 million transmitter-receiver samples with a strict cross-map split that tests generalization to new city layouts. Representative models are evaluated, and MS-AReg is presented as a baseline that reaches 0.948 cosine similarity, 1.20 degree peak location error, and 0.101 ms per sample inference on the 249,993-sample held-out set. The work supports beam management and MIMO processing by supplying directional power information without repeated full simulations at runtime.

Core claim

Map2APS supplies a physically grounded collection of angle power spectra derived from intelligent ray-tracing path records over 51 urban maps and approximately 2.55 million Tx-Rx pairs. Under cross-map evaluation the MS-AReg model records a cosine similarity of 0.948, a peak location error of 1.20 degrees, and an inference latency of 0.101 ms per sample across the full test set of 249,993 samples, together with additional dominant-direction metrics that check preservation of decision-relevant arrival angles.

What carries the argument

The Map2APS benchmark dataset generated from intelligent ray-tracing propagation records, together with the MS-AReg multi-scale regression model that performs direct map-to-APS prediction.

If this is right

Wireless beam management systems can obtain directional power estimates directly from map geometry without executing full ray-tracing at every location.
The observed low latency supports integration into real-time MIMO precoding loops operating at millisecond timescales.
Dominant-peak recall metrics indicate that predictions retain the arrival directions most relevant for link adaptation decisions.
Cross-map generalization performance suggests the approach can scale to network planning across multiple cities without retraining per environment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Extending the benchmark to maps with varying building heights and moving scatterers would test robustness beyond the current equal-height assumption.
Coupling the fast inference model with periodic real-world calibration could reduce dependence on simulation-only ground truth.
The same map-to-spectrum formulation could be adapted for other propagation quantities such as delay power spectra in integrated sensing and communication systems.

Load-bearing premise

Intelligent ray-tracing simulations supply sufficiently accurate ground truth for real-world angle power spectra and the chosen 51 maps with cross-map splits adequately represent generalization to arbitrary unseen city layouts.

What would settle it

Direct comparison of the model's predicted angle power spectra against field measurements collected by real radio equipment in an urban environment absent from the benchmark maps.

read the original abstract

Angle power spectrum (APS) characterizes the directional distribution of received signal power and is directly relevant to beam management and MIMO processing. While environment-aware learning has been widely studied for radio maps and path loss, direct map-to-APS prediction still lacks a standardized large-scale benchmark. This paper presents Map2APS, a physically grounded benchmark constructed from intelligent ray-tracing (IRT) path-level propagation records. Map2APS covers 51 equal-height urban maps and approximately 2.55 million Tx--Rx samples, with a strict cross-map split for evaluating generalization to unseen urban layouts. We benchmark representative model families and introduce MS-AReg as a strong reference baseline. On the full held-out test set of 249{,}993 samples, MS-AReg achieves a cosine similarity of 0.948, a peak location error of 1.20$^\circ$, and an inference latency of 0.101 ms/sample. We further report dominant-direction metrics, including Top-1 dominant peak hit rate and dominant peak recall, to evaluate whether predicted spectra preserve decision-relevant arrival directions. The benchmark, code, and evaluation scripts are released at https://github.com/UNIC-Lab/aps-data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Map2APS supplies a new large-scale benchmark for map-to-APS prediction from ray-tracing data with proper cross-map splits, but the results only show how well models copy the simulator rather than real propagation.

read the letter

The paper introduces Map2APS, a dataset of roughly 2.55 million Tx-Rx pairs drawn from 51 equal-height urban maps, with labels generated by intelligent ray-tracing. It enforces a strict cross-map split so models are tested on entirely unseen city layouts, and it supplies MS-AReg as a reference baseline that reaches 0.948 cosine similarity and 1.20° peak error on the 250k-sample held-out set. Dominant-direction metrics are also reported, which is the right focus for beam-management use cases. Releasing the data, code, and evaluation scripts is helpful and makes the benchmark immediately usable by others. That is the concrete advance over earlier radio-map or path-loss work. The numbers are computed directly against the simulated labels without obvious circularity. The soft spot is exactly the one the stress-test flags: there is no quantitative comparison of the IRT-generated APS against field measurements. The benchmark therefore measures how faithfully a model can reproduce the simulator on new maps, not how well it would match physical reality. The equal-height maps add another simplification that may not carry over to real cities with varied building heights. This work is aimed at researchers building learning models for environment-aware MIMO and beamforming who need a standardized testbed. It deserves peer review because the dataset scale, split protocol, and baseline results are substantial enough to warrant referee scrutiny, especially on the simulation-to-reality gap. A revised version that adds even preliminary measurement validation would be stronger, but the current version is already worth sending out.

Referee Report

2 major / 2 minor

Summary. The paper introduces Map2APS, a benchmark for direct map-to-APS prediction constructed from intelligent ray-tracing (IRT) path-level records. It covers 51 equal-height urban maps and ~2.55 million Tx-Rx samples with a strict cross-map split, benchmarks model families, and presents MS-AReg as a reference baseline. On the held-out test set of 249,993 samples, MS-AReg reports cosine similarity 0.948, peak location error 1.20°, inference latency 0.101 ms/sample, plus dominant-direction metrics such as Top-1 hit rate and peak recall. The benchmark, code, and scripts are released publicly.

Significance. If the IRT simulations provide faithful APS labels, the work supplies a large-scale, reproducible benchmark for environment-aware learning in wireless communications, directly relevant to beam management. Strengths include the dataset scale, cross-map generalization protocol, multiple decision-oriented metrics, and full public release of data and evaluation code.

major comments (2)

[Section 3] Section 3 (Dataset Construction): The manuscript provides no quantitative comparison of IRT-generated APS against any field measurements. All reported metrics (cosine similarity 0.948, 1.20° peak error) are therefore computed exclusively against simulated labels, so the claim that Map2APS constitutes a 'physically grounded' benchmark usable for real beam management rests on an untested assumption about IRT fidelity.
[Section 4] Section 4 (Experimental Setup): All 51 maps use equal-height buildings. This design choice simplifies the geometry but leaves open whether the reported generalization performance (cross-map split) would hold for realistic urban layouts that include height variation; the current results therefore test generalization only within a restricted class of environments.

minor comments (2)

[Abstract] Abstract and §5: The exact test-set size is given as 249,993; confirm this figure is consistent with the total 2.55 million samples after the cross-map split and state the precise train/validation/test counts in a single table.
[Figure 2] Figure 2 and §4.3: The visualization of predicted vs. ground-truth APS would benefit from an explicit color-scale bar and a note on whether the plotted spectra are normalized per sample.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and outline targeted revisions to improve clarity and transparency while preserving the core contributions of the Map2APS benchmark.

read point-by-point responses

Referee: [Section 3] Section 3 (Dataset Construction): The manuscript provides no quantitative comparison of IRT-generated APS against any field measurements. All reported metrics (cosine similarity 0.948, 1.20° peak error) are therefore computed exclusively against simulated labels, so the claim that Map2APS constitutes a 'physically grounded' benchmark usable for real beam management rests on an untested assumption about IRT fidelity.

Authors: We acknowledge that the manuscript does not present new field measurements comparing IRT-generated APS to real-world data. The term 'physically grounded' is intended to indicate that labels are produced by deterministic ray-tracing based on electromagnetic propagation principles rather than purely empirical or stochastic models. IRT has been extensively validated against measurements in the wireless literature; we will add a new subsection in Section 3 that cites representative validation studies and explicitly discusses the simulation-to-reality gap as a limitation. This revision clarifies scope without altering the reported results or requiring new experiments. revision: partial
Referee: [Section 4] Section 4 (Experimental Setup): All 51 maps use equal-height buildings. This design choice simplifies the geometry but leaves open whether the reported generalization performance (cross-map split) would hold for realistic urban layouts that include height variation; the current results therefore test generalization only within a restricted class of environments.

Authors: We agree that restricting all maps to equal-height buildings limits the tested environment class. This design decision was made to control for building height while focusing on footprint and layout diversity, enabling a large-scale, reproducible benchmark. The cross-map split still evaluates generalization across distinct urban geometries within this class. We will revise the experimental setup description and add an explicit limitations paragraph in the discussion section stating that extension to height-varying buildings remains future work. The reported metrics and conclusions will be qualified accordingly. revision: partial

Circularity Check

0 steps flagged

No circularity: benchmark metrics are direct comparisons to independently generated IRT labels on held-out splits

full rationale

The paper generates APS ground truth via IRT simulations on 51 urban maps, then evaluates models (including the introduced MS-AReg baseline) on a strict cross-map held-out test set of 249993 samples. Reported metrics (cosine similarity 0.948, peak error 1.20°) are computed directly against these simulated labels without any parameter fitting on the test set or reduction of predictions to inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked to justify the core benchmark construction or performance claims. The derivation chain is self-contained empirical evaluation against an external simulation process.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the domain assumption that ray-tracing simulations serve as reliable ground truth and that the chosen urban maps support meaningful generalization testing; no free parameters or invented entities are introduced in the benchmark construction itself.

axioms (1)

domain assumption Intelligent ray-tracing simulations accurately represent urban radio propagation for angle power spectra
The entire benchmark is constructed from IRT path-level propagation records treated as ground truth.

pith-pipeline@v0.9.0 · 5531 in / 1206 out tokens · 66016 ms · 2026-05-15T03:16:48.973836+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · 1 internal anchor

[1]

Millimeter wave mobile communications for 5g cellular: It will work!

T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y . Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5g cellular: It will work!”IEEE Access, vol. 1, pp. 335–349, 2013. [Online]. Available: https://doi.org/10.1109/ACCESS.2013.2260813

work page doi:10.1109/access.2013.2260813 2013
[2]

An overview of signal processing techniques for millimeter wave MIMO systems,

R. W. Heath Jr., N. Gonz ´alez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques for millimeter wave MIMO systems,”IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 436–453, 2016

work page 2016
[3]

TR 38.901: Study on channel model for frequencies from 0.5 to 100 ghz,

3GPP, “TR 38.901: Study on channel model for frequencies from 0.5 to 100 ghz,” 3rd Generation Partnership Project (3GPP), 2022, accessed: 2026-04-07. [Online]. Available: https://www.3gpp.org/ dynareport/38901.htm

work page 2022
[4]

DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” inProc. Information Theory and Applications Workshop (ITA), San Diego, CA, USA, Feb. 2019, pp. 1–8

work page 2019
[5]

Radiounet: Fast radio map estimation with convolutional neural networks,

R. Levie, C ¸ . Yapar, G. Kutyniok, and G. Caire, “Radiounet: Fast radio map estimation with convolutional neural networks,”IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 4001–4015, 2021. [Online]. Available: https://doi.org/10.1109/TWC.2021.3054977

work page doi:10.1109/twc.2021.3054977 2021
[6]

Rme-gan: A learning framework for radio map estimation based on conditional generative adversarial network,

S. Zhang, A. Wijesinghe, and Z. Ding, “Rme-gan: A learning framework for radio map estimation based on conditional generative adversarial network,”IEEE Internet of Things Journal, vol. 10, no. 20, pp. 18 016–18 027, 2023. [Online]. Available: https://doi.org/10.1109/ JIOT.2023.3278235

work page arXiv 2023
[7]

Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,

X. Wang, K. Tao, N. Cheng, Z. Yin, Z. Li, Y . Zhang, and X. Shen, “Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 2, pp. 738–750, 2025

work page 2025
[8]

Sionna rt: Differentiable ray tracing for radio propagation modeling,

J. Hoydis, F. A. Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna rt: Differentiable ray tracing for radio propagation modeling,” in2023 IEEE Globecom Workshops (GC Wkshps), 2023, pp. 317–321. [Online]. Available: https: //doi.org/10.1109/GCWkshps58843.2023.10465179

work page doi:10.1109/gcwkshps58843.2023.10465179 2023
[9]

A Tutorial on Learning-Based Radio Map Construction: Data, Paradigms, and Physics-Awareness

X. Wang, Y . Pan, and N. Cheng, “A tutorial on learning-based radio map construction: Data, paradigms, and physics-awareness,” arXiv preprint arXiv:2603.17499, 2026. [Online]. Available: https: //arxiv.org/abs/2603.17499

work page internal anchor Pith review Pith/arXiv arXiv 2026
[10]

Environment-aware hybrid beamforming by leveraging channel knowledge map,

D. Wu, Y . Zeng, S. Jin, and R. Zhang, “Environment-aware hybrid beamforming by leveraging channel knowledge map,”IEEE Transac- tions on Wireless Communications, vol. 23, no. 5, pp. 4990–5005, 2024

work page 2024
[11]

[Online]

Altair Engineering Inc.,Intelligent Ray Tracing (IRT), Altair Engineering Inc., 2026, altair WinProp Documentation, accessed: 2026-04-13. [Online]. Available: https://help.altair.com/winprop/topics/ winprop/user guide/propagation methods/propgation models indoor irt.htm

work page 2026
[12]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?...

work page 2021
[13]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778. [Online]. Available: https://openaccess.thecvf.com/content cvpr 2016/ html/He Deep Residual Learning CVPR 2016 paper.html

work page 2016