Much of the licensed radio spectrum below 6 GHz sits idle. A review of cognitive radio networks puts idle licensed spectrum at over 60% at any given time and place. One reason is that unlicensed "secondary" users are kept out by conservative rules, since no one can cheaply check that their combined interference at a licensed receiver stays below a legal limit. That check is an integral: sum the transmit power of every secondary user, weighted by the signal loss (pathloss) from its location to the protected receiver, over wherever those users are. The loss field is known only at a few places where a sensor has actually measured it. Averaging such measurements taken at random places converges at the Monte Carlo rate $N^{-1/2}$, so each extra digit of accuracy costs 100 times more measurements.
We propose to treat the admission check as a Bayesian quadrature problem. Bayesian quadrature estimates an integral by fitting a Gaussian process, a probabilistic regression model that returns both a prediction and an error bar, to the integrand, and then integrating the model. In our design, a pretrained neural radio-map predictor supplies a cheap first guess of the pathloss field. A scalable sparse Gaussian process models only the correction that sensors reveal. The admission decision is then made from the posterior distribution of the aggregate interference itself, not from a pixel-accurate map. New sensor readings are requested only where they are most likely to change that decision. The claim is that this reaches correct and calibrated admission decisions with far fewer sensor measurements than averaging or map-then-sum pipelines, and that it stays correct in cities the neural predictor has never seen.
Core Claim
The quantity that governs spectrum reuse is a single number with an error bar: the aggregate interference $I$ at a primary receiver located at $x_p$,
$$I = \int_{\mathcal{S}} P\,\rho(x)\,10^{-L(x)/10}\,dx,$$
where $P$ is secondary transmit power, $\rho(x)$ is the density of secondary users over region $\mathcal{S}$, and $L(x)$ is the pathloss in dB from location $x$ to $x_p$. A secondary configuration is admitted if $\Pr(I > I_{\max} \mid \text{data}) \le \delta$, where $I_{\max}$ is the interference limit and $\delta$ is the tolerated violation risk.
Our model writes $L(x) = L_{\text{nn}}(x) + r(x)$. Here $L_{\text{nn}}$ is a neural radio-map prediction and $r$ is a sparse Gaussian process fitted to sensor residuals. We make three specific claims. First, integrating the posterior over $r$ gives a faster-than-$N^{-1/2}$ decrease in integral error as a function of sensor count $N$. Second, the joint posterior covariance, which the integral's variance needs, is exactly what cheap local Gaussian process methods throw away. Third, choosing sensor locations to resolve the threshold decision beats choosing them to reduce map error.
Three lines of prior work combine here, each in a different role. The pathloss challenge The First Pathloss Radio Map Prediction Challenge motivates the proposal. It shows neural predictors such as PMNet reach RMSE 0.0383 in about 10 ms, yet all submissions "performed significantly worse" on unseen Istanbul maps, and none produce uncertainty. Multilevel Bayesian Quadrature supplies the method. It places Gaussian processes on the differences between a cheap and an expensive fidelity level, and reports errors 10 to 100 times smaller than multilevel Monte Carlo on a tsunami simulator. That work uses exact Gaussian processes with cubic cost, assumes a hierarchy of simulators rather than simulator-plus-sensors, and never faces a threshold decision. Approximation-Aware Bayesian Optimization supplies the principle that a sparse Gaussian process should spend its limited capacity where the downstream decision is sensitive, not uniformly. We carry that principle from optimization to quadrature-with-threshold. We know of no work in the file that places a Bayesian-quadrature posterior over aggregate interference in spectrum sharing.
Research Program
The first component is the two-level surrogate. The coarse level is a PMNet-style predictor trained on the RadioMap3DSeer training cities. The fine level is a sparse variational Gaussian process with $m$ inducing points. The sparse process summarizes $n$ observations through $m \ll n$ learned "virtual" points, at $O(nm^2)$ cost, and is fitted to the residual between the sensor readings and $L_{\text{nn}}$. Its inputs are location together with the network's own prediction, so the residual kernel can depend on building context. This follows the multilevel logic of MLBQ, but the fine level is sparse measurement rather than a finer simulator. We will test directly whether the residual is actually smoother than $L$, which is the premise that makes the multilevel scheme pay off. The sharp shadow edges in urban maps may violate that premise.
The second component is the integral posterior. Interference is linear in received power, so it is the exponential of a Gaussian field and $I$ is not Gaussian. We handle this in two ways. The first is an exact moment computation: the pointwise mean of $10^{-L/10}$ is closed-form given the posterior mean and variance, and the second moment is a double integral of pairwise posterior covariances. The second is drawing $S$ joint posterior sample fields through the inducing-point representation, at $O(Gm)$ per sample on a grid of $G$ cells. We then read $\Pr(I > I_{\max})$ directly from the sampled integrals.
Monte Carlo is still used inside this step, but only over the cheap posterior and never over expensive measurements. The convergence gain we claim is in $N$, the number of sensor readings. A useful side effect is that a change in the user density $\rho$, which happens every scheduling interval, requires only re-integrating the fitted posterior, not refitting it. This is what makes a 60-second CBRS-style decision slot (F-CBRS) realistic.
The third component is decision-aware sensing. Given current data, the next sensor location is chosen to maximize the expected reduction in the probability of a wrong admission decision. Inducing-point locations are refit under a matching objective that weights residual error by each location's contribution to the variance of $I$. Two constraints shape this design. On the Sampling Problem for Kernel Quadrature shows that node placement can change quadrature error by orders of magnitude, but its SMC-KQ search is costly when evaluations are cheap. Kernel quadrature with DPPs gets optimal rates from repulsive node sets, but requires analytic kernel eigenfunctions, which do not exist for kernels conditioned on building layouts. We therefore use greedy sequential acquisition rather than either scheme.
Baselines
The naive reference is plain Monte Carlo averaging. We take $N$ sensor readings at locations drawn from $\rho$ and average $P\,10^{-L/10}$, with a CLT-based error bar used for the admission test. We also run a randomized quasi-Monte Carlo variant, which places the $N$ readings on a scrambled low-discrepancy set pushed through $\rho$. Both use the same measurement noise and the same $N$. We beat them if our method reaches a given relative integral error with several-fold fewer readings, and shows a steeper error-versus-$N$ slope on log-log axes.
The practitioner's reference is a static exclusion zone. Secondary users are admitted only outside a distance radius tuned so that no violation occurs on training cities. This is the rule that a Stackelberg game study and the cognitive-radio review blame for idle spectrum. We beat it if we admit more secondary power at the same or lower violation rate.
The strongest existing method is the neural map summed over the grid, with a safety margin. PMNet's prediction is integrated against $\rho$, and a margin is calibrated on held-out training-city maps to meet $\delta$. Its reported figure is map-level RMSE 0.0383 on RadioMap3DSeer, from The First Pathloss Radio Map Prediction Challenge. For the Istanbul split the paper gives no absolute number, and we found no reported integral-level or admission-level result on Lacuna for any radio-map method. We will therefore compute this baseline ourselves on identical maps. For the sparse-sensing regime, RadioDiff-3D reports RMSE 0.0481 at 10% coverage on UrbanRadio3D (RadioDiff-3D). We will run that model conditioned on the same sensor readings, sum its output, and calibrate a margin the same way. Every method gets the same maps, sensor budget, noise, and hyperparameter search budget. Beating these baselines means matching their in-distribution admitted power while holding the violation rate at $\delta$ on Istanbul, where fixed margins should fail.
Three ablations isolate the key ingredients. The first removes the network and uses a Gaussian process on raw pathloss. The second replaces the sparse joint posterior with nearest-neighbour local Gaussian processes (GPnn); these give accurate pointwise predictions but no cross-location covariance, so the integral's variance must be assembled as if cells were independent. The third keeps everything but chooses sensors to reduce map variance rather than decision error. An exact-Gaussian-process quadrature at small $n$ serves as an accuracy ceiling for the sparse approximation.
Why not just…?
Why not just sum a neural radio map and add a margin? In-distribution this is a serious competitor, and it may win on speed and even accuracy. The challenge paper's Istanbul result is the evidence against it: the error distribution shifts in unseen cities, so a margin tuned elsewhere either wastes spectrum or is violated. The decisive experiment is the out-of-distribution violation rate at fixed $\delta$. If a margin calibrated on training cities also holds on Istanbul, our main motivation weakens, and the contribution shrinks to the reduction in sensor count.
Why not just use quasi-Monte Carlo on the sensor locations? Quasi-Monte Carlo avoids any model and any matrix inversion. However, its rate advantage depends on the integrand having bounded variation in a favourable sense, and pathloss has sharp shadow discontinuities along building edges. It also cannot use the neural predictor's information. Our error-versus-$N$ curves will show where quasi-Monte Carlo is competitive. We expect it to be competitive when $\rho$ is spread thinly over open terrain, and we will report that regime honestly.
Why not just use a fast local Gaussian process and add up its pointwise variances? GPnn trains on $1.6 \times 10^6$ points in about 30 seconds on a laptop, which is attractive. But the variance of an integral is dominated by correlated errors, such as a whole street being mis-shadowed, and summing independent marginal variances underestimates it. Its authors also note that local models achieve only weak calibration. The ablation measures this directly as credible-interval coverage for $I$. If local models with a scalar variance inflation reach nominal coverage, we will adopt them, since they are cheaper.
Proposed Experiments
We use RadioMap3DSeer with its training cities for fitting the network, and held-out training-city maps plus the 84 blind Istanbul maps for evaluation. UrbanRadio3D (RadioDiff-3D) adds a 3D test with receivers at varying heights. Ground-truth $I$ is computed from the full ray-traced map. Sensors reveal individual pixels with added noise. The density $\rho$ comes in three types: uniform, street-concentrated, and hotspot mixtures.
We report four measurements. The first is relative integral error versus $N$ on log-log axes. The second is coverage of 90% credible intervals for $I$. The third is the admission violation rate against $\delta$, together with admitted secondary power as the measure of spectrum reuse. The fourth is wall-clock time per decision and per change of $\rho$.
The first decisive check is whether our error curve has a steeper slope than $-1/2$ in $N$ and whether it keeps that slope on Istanbul. The second is whether the joint sparse posterior achieves nominal coverage where the local-Gaussian-process ablation does not. The third is whether decision-aware sensing reduces wrong admissions at equal $N$ relative to map-variance sensing. If the third check fails, the proposal reduces to a well-calibrated two-level quadrature, which is still useful but a smaller claim.
References
- Intelligent and efficient development of wireless networks: A review of cognitive radio networks
- The First Pathloss Radio Map Prediction Challenge
- Multilevel Bayesian Quadrature
- Approximation-Aware Bayesian Optimization
- Interference management for unlicensed users in shared CBRS spectrum
- On the Sampling Problem for Kernel Quadrature
- Kernel quadrature with DPPs
- A stackelberg game for power control and channel allocation in cognitive radio networks
- RadioDiff-3D: A 3D× 3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication
- Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression