arxiv: 2605.08035 · v1 · submitted 2026-05-08 · 📡 eess.SP · cs.LG

Recognition: 2 theorem links

· Lean Theorem

PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting

William Bjorndahl , Maninder Pal Singh , Farhad Nouri , Joseph Camp

Authors on Pith no claims yet

Pith reviewed 2026-05-11 02:09 UTC · model grok-4.3

classification 📡 eess.SP cs.LG

keywords RF propagation modelingmap-free reconstruction3D Gaussian splattingsite-specific wireless modelingsparse signal measurementspath loss predictionwireless radiance fields

0 comments

The pith

PropSplat reconstructs accurate site-specific RF fields from sparse signal measurements alone, without maps or geographic data.

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

The paper demonstrates that radio frequency propagation can be modeled for a specific location using only a modest number of direct signal strength readings taken along transmitter-receiver paths. It replaces reliance on detailed 3D maps, floor plans, or terrain databases with a collection of 3D anisotropic Gaussian shapes. Each Gaussian adjusts a simple baseline path-loss formula whose exponent is also learned from the data. If the approach holds, rapid RF environment modeling becomes feasible in areas where maps are unavailable or quickly outdated. The Gaussians are placed only along the measured paths and tuned end-to-end to match observed strengths, yielding lower prediction error than prior map-dependent methods on both outdoor and indoor test sets.

Core claim

PropSplat represents propagation effects through 3D anisotropic Gaussian primitives. Each primitive stores a scalar offset to an explicit baseline path-loss model whose exponent is learned during optimization. The primitives are initialized exclusively along observed transmitter-receiver trajectories and adjusted solely against sparse received-signal-strength values, with no external geographic, clutter, or floor-plan information supplied.

What carries the argument

3D anisotropic Gaussian primitives that add learned scalar offsets to a baseline path-loss model, initialized along observed paths and optimized end-to-end against sparse signal-strength data.

If this is right

Sparse outdoor drive-test measurements spaced 300 m apart produce 5.38 dB RMSE across multiple frequencies and terrain types.
Indoor Bluetooth measurements yield 0.19 m mean localization error, an order of magnitude lower than competing wireless radiance-field approaches.
No floor plans, terrain databases, or clutter maps are required at any stage of training or inference.
The same representation supports both field prediction and receiver localization tasks within one optimization.

Where Pith is reading between the lines

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

The method could be extended to update models in real time by adding new measurement paths as they become available.
Because only RF-native data are used, the approach may integrate directly with crowdsourced signal reports from mobile devices.
If the Gaussian primitives can be made differentiable with respect to frequency, the same framework might predict behavior across bands without retraining.

Load-bearing premise

A modest set of 3D anisotropic Gaussians placed only along measured paths and tuned to sparse signal-strength readings can capture the main propagation behavior of an unknown environment.

What would settle it

Apply the method to a new environment containing many small reflective surfaces or sharp terrain changes not aligned with the initial path lines; if the resulting RMSE on held-out measurements exceeds that of a conventional ray-tracing model that uses an accurate 3D map, the core claim is falsified.

Figures

Figures reproduced from arXiv: 2605.08035 by Farhad Nouri, Joseph Camp, Maninder Pal Singh, William Bjorndahl.

**Figure 1.** Figure 1: PropSplat represents radio propagation effects as learnable 3D Gaussian ellipsoids splatted throughout the environment. Each Gaussian encodes a spatially localized offset from a baseline path loss model. Red Gaussians (G1, G2, G4) represent increased path loss due to obstruction or shadowing and blue Gaussians (G3, G5) represent decreased path loss from constructive propagation effects such as reflection.… view at source ↗

**Figure 2.** Figure 2: PropSplat training process where Gaussians evolve from initialization in Iteration 0 through migration toward high [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: PropSplat applied to Ofcom London 5850 MHz drive-test measurements. Routes span up to 8 km from Tx. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: PropSplat performance, varying the measurement dis [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: PropSplat’s path loss prediction is spatially decompos [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

read the original abstract

Building a site-specific propagation model typically requires either ray-tracing over detailed 3D maps or dense measurement campaigns. Both approaches are expensive and often infeasible for rapid deployments where geographic data is unavailable or outdated. We present PropSplat, a map-free propagation modeling method that reconstructs radio frequency (RF) fields using 3D anisotropic Gaussian primitives. Each Gaussian encodes a scalar path loss offset relative to an explicit baseline path loss model with a learnable path loss exponent. Gaussians are initialized along observed transmitter--receiver paths and optimized end-to-end to learn the propagation environment without external information like floor plans, terrain databases, or clutter data. We evaluate PropSplat against wireless radiance field methods NeRF$^2$, GSRF, and WRF-GS+ on two real-world datasets. On large-scale outdoor drive-tests spanning multiple topographical regions at six sub-6 GHz frequencies, PropSplat achieves 5.38 dB RMSE when training measurements are spaced 300m apart and outperforms WRF-GS+ (5.87 dB), GSRF (7.46 dB), and NeRF$^2$ (14.76 dB). On indoor Bluetooth Low Energy measurements, PropSplat achieves 0.19m mean localization error, an order of magnitude better than NeRF$^2$ (1.84m), while achieving near-identical received signal strength prediction accuracy. These results show that accurate site-specific propagation reconstruction is achievable from sparse RF-native measurements. The need for geographic data as a prerequisite for scalable RF environment modeling is reduced.

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

PropSplat gets measurable gains over prior RF radiance-field baselines on real outdoor and indoor data by initializing anisotropic Gaussians along measurement paths and fitting scalar offsets to a learnable path-loss baseline.

read the letter

PropSplat adapts 3D Gaussian splatting to RF field reconstruction without maps. It places anisotropic Gaussians only along observed transmitter-receiver paths and lets each one carry a scalar offset on top of an explicit path-loss model whose exponent is learned during optimization. That combination produces the reported numbers: 5.38 dB RMSE on the large outdoor drive-test set at 300 m spacing, beating WRF-GS+ at 5.87 dB and the other two baselines by larger margins, plus a big drop in indoor localization error to 0.19 m on BLE data. The concrete improvement on real measurements is the clearest takeaway so far. The path-initialized setup plus the explicit baseline appears to be the part that is not already in the cited NeRF2, GSRF, or WRF-GS+ papers, so the method is new in that specific form. The outdoor results across multiple frequencies and terrains give some evidence that the approach can work with sparse RF-native data alone. The main soft spots are the lack of error bars, ablation tables, or stability checks in the abstract, which makes it hard to tell how much of the gain comes from the Gaussians versus simply tuning the single exponent. The initialization constraint also leaves open the question of how much site-specific structure is actually recovered off the paths; regions between the lines depend on the splatting kernels extrapolating from on-path data, and without targeted checks on unsampled volumes the claim that RF measurements alone suffice rests on the assumption that the kernels capture the dominant effects. This work is aimed at people building propagation models for rapid wireless deployments or RF sensing applications who want to skip geographic databases. A reader already following radiance-field methods applied to signals would get the most out of it. The real-data results and the clear technical distinction are enough to justify sending it to peer review rather than desk rejection, though the referees will need to see the missing ablations and off-path validation to judge how far the central claim holds.

Referee Report

2 major / 3 minor

Summary. The paper introduces PropSplat, a map-free RF propagation modeling technique that represents the environment using 3D anisotropic Gaussian primitives. Each Gaussian encodes a scalar path-loss offset relative to an explicit baseline model whose exponent is learned from data. Primitives are initialized exclusively along observed Tx-Rx paths and optimized end-to-end against sparse received-signal-strength measurements. On large-scale outdoor drive-test data at six sub-6 GHz frequencies with 300 m measurement spacing, the method reports 5.38 dB RMSE, outperforming WRF-GS+ (5.87 dB), GSRF (7.46 dB), and NeRF² (14.76 dB). On indoor BLE data it achieves 0.19 m mean localization error while matching RSS prediction accuracy of prior methods.

Significance. If substantiated, the approach would meaningfully reduce reliance on geographic data for site-specific RF modeling, enabling faster deployment in environments where maps are unavailable or stale. The quantitative gains on real-world outdoor and indoor datasets indicate practical potential for scalable, measurement-driven propagation reconstruction.

major comments (2)

[§3.2] §3.2 (Gaussian initialization and optimization): Gaussians are placed only along observed Tx-Rx paths. The manuscript must demonstrate, via explicit off-path metrics or visualizations, that the anisotropic splatting kernels produce site-specific deviations from the baseline model in unsampled volumes; otherwise the reconstruction reduces to the learnable-exponent baseline and the map-free claim is not fully supported.
[Table 1, §4.1] Table 1 and §4.1 (quantitative results): The reported 5.38 dB RMSE and comparisons to baselines lack error bars, standard deviations across runs, or ablation studies on Gaussian count, initialization density, and exponent learning. These omissions make it impossible to judge whether the 0.49 dB improvement over WRF-GS+ is statistically reliable or sensitive to optimization details.

minor comments (3)

[Eq. 3] Notation for the path-loss offset inside each Gaussian (Eq. 3) is introduced without a clear link to the final field reconstruction formula; a single consolidated equation would improve readability.
[§4.2] The indoor localization experiment reports mean error but does not specify the number of test points or the exact train/test split used for the 0.19 m figure.
[Figure 4] Figure 4 caption should explicitly state the measurement spacing and frequency band shown, rather than referring only to 'drive-test data'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below, providing our response and indicating the revisions we will make to the manuscript.

read point-by-point responses

Referee: [§3.2] §3.2 (Gaussian initialization and optimization): Gaussians are placed only along observed Tx-Rx paths. The manuscript must demonstrate, via explicit off-path metrics or visualizations, that the anisotropic splatting kernels produce site-specific deviations from the baseline model in unsampled volumes; otherwise the reconstruction reduces to the learnable-exponent baseline and the map-free claim is not fully supported.

Authors: We agree that explicit evidence of off-path generalization is necessary to fully support the map-free claim. While the end-to-end optimization of anisotropic Gaussians against sparse RSS measurements inherently allows the model to learn volumetric deviations (as evidenced by the 0.49 dB RMSE improvement over WRF-GS+ and larger gains over other baselines on held-out data), we will strengthen the manuscript by adding visualizations of the reconstructed RF field in unsampled volumes away from Tx-Rx paths. We will also include quantitative off-path metrics, such as RMSE computed on a subset of test measurements not used for initialization, comparing the full PropSplat model against the learnable-exponent baseline alone. These additions will directly illustrate the site-specific effects captured by the splatting kernels. revision: yes
Referee: [Table 1, §4.1] Table 1 and §4.1 (quantitative results): The reported 5.38 dB RMSE and comparisons to baselines lack error bars, standard deviations across runs, or ablation studies on Gaussian count, initialization density, and exponent learning. These omissions make it impossible to judge whether the 0.49 dB improvement over WRF-GS+ is statistically reliable or sensitive to optimization details.

Authors: We acknowledge that the absence of statistical measures and ablations limits the ability to assess robustness. In the revised manuscript, we will augment Table 1 with error bars showing the standard deviation of RMSE over multiple optimization runs (e.g., 5 random seeds). We will also expand §4.1 with ablation studies examining the sensitivity to Gaussian count, initialization density along observed paths, and the contribution of learning the path-loss exponent versus fixing it. These results will confirm that the reported improvements are statistically reliable and not artifacts of specific hyperparameter choices. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained data-driven optimization

full rationale

The paper presents PropSplat as an optimization procedure that initializes 3D anisotropic Gaussians along observed Tx-Rx paths and fits both the Gaussians' offsets and a single learnable path-loss exponent directly to sparse RF measurements. This is a standard supervised fitting process whose outputs on held-out test locations are not forced by construction to equal the training inputs; the model must generalize via the splatting kernels. No self-citation is invoked as a load-bearing uniqueness theorem, no ansatz is smuggled from prior author work, and no known empirical pattern is merely renamed. The central claim therefore rests on empirical performance numbers rather than definitional equivalence.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the modeling choice that environmental effects can be represented as additive scalar offsets from a baseline path-loss model using a modest number of 3D Gaussians, plus the assumption that gradient-based optimization on sparse measurements will converge to a useful representation.

free parameters (1)

path loss exponent
Learnable scalar in the explicit baseline path-loss model that is jointly optimized with the Gaussian offsets.

axioms (1)

domain assumption Environmental propagation effects can be captured by additive scalar offsets encoded in 3D anisotropic Gaussians placed along observed paths.
This is the central representational assumption stated in the abstract.

pith-pipeline@v0.9.0 · 5600 in / 1345 out tokens · 36282 ms · 2026-05-11T02:09:09.135098+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Each Gaussian encodes a scalar path loss offset relative to an explicit baseline path loss model with a learnable path loss exponent. Gaussians are initialized along observed transmitter–receiver paths and optimized end-to-end
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

3D anisotropic Gaussian primitives... anisotropic shapes capture directional propagation effects

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

https://doi.org/10.1145/3592433 Xiaonan Kong and Riley G

B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., vol. 42, no. 4, Jul. 2023. [Online]. Available: https://doi.org/10.1145/3592433

work page doi:10.1145/3592433 2023
[2]

Mildenhall, B

B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: representing scenes as neural radiance fields for view synthesis,”Commun. ACM, vol. 65, no. 1, p. 99–106, Dec. 2021. [Online]. Available: https://doi.org/10.1145/3503250

work page doi:10.1145/3503250 2021
[3]

Nerf2: Neural radio-frequency radiance fields,

X. Zhao, Z. An, Q. Pan, and L. Yang, “Nerf2: Neural radio-frequency radiance fields,” inProceedings of the 29th Annual International Conference on Mobile Computing and Networking, ser. ACM MobiCom ’23. New York, NY , USA: Association for Computing Machinery,

work page
[4]

Available: https://doi.org/10.1145/3570361.3592527

[Online]. Available: https://doi.org/10.1145/3570361.3592527

work page doi:10.1145/3570361.3592527
[5]

GSRF: Complex-valued 3d gaussian splatting for efficient radio-frequency data synthesis,

K. Yang, G. Dong, S. JI, W. Du, and M. Srivastava, “GSRF: Complex-valued 3d gaussian splatting for efficient radio-frequency data synthesis,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id=E3oNDQ8e9r

work page 2025
[6]

Active sampling and gaus- sian reconstruction for radio frequency radiance field,

C.-S. Gau, X. Chen, T. Javidi, and X. Zhang, “Active sampling and gaus- sian reconstruction for radio frequency radiance field,”arXiv preprint arXiv:2412.08003, 2024

work page arXiv 2024
[7]

Rfcanvas: Modeling rf channel by fusing visual priors and few-shot rf measurements,

X. Chen, Z. Feng, K. Sun, K. Qian, and X. Zhang, “Rfcanvas: Modeling rf channel by fusing visual priors and few-shot rf measurements,” in Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems, ser. SenSys ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 464–477. [Online]. Available: https://doi.org/10.1145/36660...

work page doi:10.1145/3666025.3699351 2024
[8]

Wrf-gs: Wireless radiation field reconstruction with 3d gaussian splatting,

C. Wen, J. Tong, Y . Hu, Z. Lin, and J. Zhang, “Wrf-gs: Wireless radiation field reconstruction with 3d gaussian splatting,” inIEEE INFOCOM 2025 - IEEE Conference on Computer Communications, 2025, pp. 1– 10

work page 2025
[9]

Neural representation for wireless radiation field reconstruction: A 3d gaussian splatting approach,

——, “Neural representation for wireless radiation field reconstruction: A 3d gaussian splatting approach,”IEEE Transactions on Wireless Communications, vol. 25, pp. 7490–7504, 2026

work page 2026
[10]

Laik: Location-specific analysis to infer key performance indicators,

R. Enami, S. Gupta, D. Rajan, and J. Camp, “Laik: Location-specific analysis to infer key performance indicators,”IEEE Transactions on Vehicular Technology, vol. 70, no. 5, pp. 4406–4418, 2021

work page 2021
[11]

Machine learning-based path loss model- ing with simplified features,

J. Ethier and M. Ch ˆateauvert, “Machine learning-based path loss model- ing with simplified features,”IEEE Antennas and Wireless Propagation Letters, 2024

work page 2024
[12]

Map-based path loss prediction in multiple cities using convolutional neural networks,

R. G. Dempsey, J. Ethier, and H. Yanikomeroglu, “Map-based path loss prediction in multiple cities using convolutional neural networks,”IEEE Antennas and Wireless Propagation Letters, 2025

work page 2025
[13]

Application of machine learning for radiowave propagation modelling below 6 ghz,

M. Z. Bocus and A. Lodhi, “Application of machine learning for radiowave propagation modelling below 6 ghz,”IEEE Access, 2025

work page 2025
[14]

Path loss prediction based on machine learning techniques: Principal component analysis, artificial neural network, and gaussian process,

H.-S. Jo, C. Park, E. Lee, H. K. Choi, and J. Park, “Path loss prediction based on machine learning techniques: Principal component analysis, artificial neural network, and gaussian process,”Sensors, vol. 20, no. 7,

work page
[15]

Available: https://www.mdpi.com/1424-8220/20/7/1927

[Online]. Available: https://www.mdpi.com/1424-8220/20/7/1927

work page 1927
[16]

Recursive estimation of dynamic rss fields based on crowdsourcing and gaussian processes,

I. Santos, J. J. Murillo-Fuentes, and P. M. Djuri ´c, “Recursive estimation of dynamic rss fields based on crowdsourcing and gaussian processes,” IEEE Transactions on Signal Processing, vol. 67, no. 5, pp. 1152–1162, 2019

work page 2019
[17]

Uk radiowave propagation measurement data for frequencies below 6 ghz,

Ofcom, “Uk radiowave propagation measurement data for frequencies below 6 ghz,” Office of Communications (Ofcom), Tech. Rep., Aug. 2019, accessed: 2025-05-06. [Online]. Available: https://www.ofcom.org.uk/siteassets/resources/documents/spectrum/uk- radiowave-propagation-measurement/sub-6ghz-propagation- measurement-data.pdf?v=324106

work page 2019