Map-Guided Spectrum Sensing That Keeps Its Promise in Unseen Cities

Radio devices that borrow unused spectrum, known as secondary users, must avoid interfering with the licensed owner, the primary user. Much of the theory behind this kind of access assumes perfect sensing: a device can always tell whether the owner is transmitting. Real detectors make two kinds of mistakes. A false alarm wastes a free channel. A missed detection causes a collision, and regulators cap how often that may happen. A natural way to help a lone device decide is to give it a predicted radio map, meaning an estimate of how strongly the primary transmitter's signal reaches each street location, computed from building footprints by a fast neural network. Once the device trusts that map, however, the perfect-sensing assumption has simply moved from the detector into the map. Map predictors are known to degrade in cities they were not trained on, so their errors become missed detections. Those missed detections are not spread evenly. They cluster in the shadowed spots where the map is most wrong.

This note proposes to borrow the central move of spatial normalization: expressing every subject's data in a shared reference coordinate system, so that a position means the same thing across individuals. Here the "subjects" are cities. The goal is to calibrate map errors in that shared frame, so that a calibration learned on six training cities still holds in a seventh. The deliverable is a sensing rule for an isolated secondary user. It combines its own energy detector with a map prior, and it holds the collision rate below the regulatory cap at each kind of location, not only on average, in a held-out city.

Core Claim

The claim is that map-prediction error has stable statistics once each location is described in transmitter-relative, geometry-anchored coordinates. Those coordinates are distance to the transmitter, whether the line of sight is blocked, the number of buildings the direct ray crosses, and the depth of the location inside a building's shadow. In raw pixel coordinates the statistics are unstable. If the claim holds, the operating-point rule from Joint Design and Separation Principle for Opportunistic Spectrum Access in the Presence of Sensing Errors can be applied separately at each location. That paper shows the optimal detector sets its missed-detection probability $\delta$ equal to the allowed collision probability $\zeta$, and that the user can then simply obey the detector. This works only if the location's effective $\delta$ is known, and the canonical frame is what makes $\delta$ estimable in a new city.

There is a reason not to use a full anatomical-style template warp. Cities share no anatomy, so a canonical "average city" analogous to the Allen brain atlas has no meaning. What carries over from the atlas work is landmark anchoring: aligning data to consistent physical landmarks rather than raw coordinates. Examples include the blood-vessel landmarks in Differential Spatial Representation of Taste Modalities in the Rat Gustatory Cortex, the body-axis anchors of Kobitski et al., and the anatomy-relative frames for deforming tissue in Yuhang Liu et al.. In our setting the transmitter and building edges are the landmarks. WiCAT supplies the motivating evidence: registering cortices to a shared atlas with fixed coordinate embeddings enabled zero-shot transfer to unseen animals. We test whether the analogous move transfers error calibration, not representations, across cities.

Research Program

The ground truth is the ray-traced pathloss in RadioMap3DSeer and the Istanbul blind set from The First Pathloss Radio Map Prediction Challenge. The prior is a PMNet-style predictor trained on the six training cities.

We define the primary user's protected region as the set of locations where the true received primary power exceeds a protection level. A secondary user at such a location collides if it transmits while the primary is active. At each location the secondary user observes an energy-detector statistic. That statistic is simulated from the true pathloss plus small-scale fading and noise, with a short sensing window so that the single-sensor error floor identified in Spectrum Sensing: To Cooperate or Not to Cooperate? is active.

The decision rule is a likelihood-ratio test whose threshold depends on the map's predicted primary power at that location. Given an error model for the map, we choose the threshold at each location to meet $\delta = \zeta$.

The error model is the object of study. We map each location to its canonical coordinate $c$ and fit a conditional distribution of map residuals given $c$. The goal is that for every stratum $g$ of the canonical frame (for example, non-line-of-sight locations behind two or more buildings at 200–400 m),

$$\Pr\big(\text{collision} \mid \text{primary active},\ c \in g\big) \le \zeta,$$

which holds on held-out cities. We use group-conditional split conformal calibration on residuals, which gives a distribution-free bound under exchangeability within each stratum. This is the source of the guarantee, and the experiments test whether "exchangeable within canonical strata across cities" is a reasonable assumption.

A secondary study removes the perfect-sensing assumption from Distributed Spectrum Access with Spatial Reuse. That paper's convergence results assume perfect sensing and name sensing errors as open. We feed the per-location calibrated error rates into its spatial game and measure whether distributed learning still converges and how far the result falls from the centralized optimum.

Baselines

The naive reference point is a map-free detector that applies the $\delta = \zeta$ rule with a single global threshold, set for the worst-case received power over the protected region. It meets the cap trivially but wastes most opportunities. Any proposed method must beat its secondary throughput while matching its collision safety.

The second reference is a "perfect-map" detector. It takes the PMNet prediction as truth, which is the perfect-sensing assumption moved into the map, and sets location thresholds with no error model. We expect it to have high throughput and to violate the cap in shadowed strata on Istanbul. For context on how strong this prior is, the challenge reports PMNet as the winner with RMSE 0.0383 on the Istanbul blind test. Agile (KL, LoS) reached 0.0451 and PPNet 0.0507 (challenge paper). No reported result was found on Lacuna for collision rate or secondary throughput of any map-guided sensing scheme on these data.

The strongest competitor is pooled conformal calibration, the same procedure with a single residual distribution for all locations. It is the cheapest principled fix and gives a valid marginal guarantee. A second strong competitor is group-conditional calibration with groups defined by raw pixel position or raw distance alone. This variant isolates our key ingredient: the geometry-anchored canonical frame, as opposed to stratification in general.

All variants share the same trained PMNet, the same calibration split drawn from training cities, and an equal number of strata. They are tuned only on training-city validation folds and never on Istanbul. Beating them means three things on Istanbul. First, worst-stratum collision rate at or below $\zeta$ for $\zeta \in \{0.05, 0.1\}$. Second, a pooled method that exceeds $\zeta$ in some stratum. Third, secondary throughput within a few points of pooled calibration and well above the global-threshold detector.

Why not just…?

Why not just use pooled conformal calibration on the map residuals? This is the real threat. If map error in a new city is roughly uniform across geometry, pooled calibration will meet the cap everywhere and normalization adds nothing. We expect it to fail locally because the challenge reports that models degrade on Istanbul's unfamiliar morphology, and degradation is plausibly concentrated behind buildings. The worst-stratum collision comparison settles this. If pooled calibration passes, we report a negative result for the normalization hypothesis.

Why not just augment the map predictor with rotations and crops so that it generalizes better? Better maps shrink residuals but provide no collision guarantee, and the regulatory constraint concerns the tail of the error distribution, not its RMSE. We will include an augmented PMNet as the prior in every variant. If augmentation alone makes the perfect-map detector safe, that result would be decisive against our proposal.

Why not just use cooperative sensing? Hard one-bit fusion breaks the single-sensor floor (Duan et al.), but it needs active neighbors and a reporting channel. We target isolated users and report cooperative fusion with $N$ neighbors as an upper reference, to show how many neighbors the map prior is worth.

Why not just learn a deep generative prior for the map, as in deep spectrum cartography? Deep Spectrum Cartography completes maps from sparse power measurements on fixed floorplans and reaches about 0.08 reconstruction error at 5% coverage on the Mannheim indoor data. It does not address cross-environment transfer or detection error trade-offs. Its approach is complementary, since it could replace PMNet as the prior, but it does not supply the calibration.

Decisive Checks

The main experiment trains on the six RadioMap3DSeer cities and evaluates on Istanbul. It reports worst-stratum and fraction-of-strata collision violations, together with throughput, at two values of $\zeta$. Leave-one-city-out runs over the training cities test whether the result is specific to Istanbul.

The key diagnostic measures the gap between residual distributions across cities, stratum by stratum, in the canonical frame and in pixel coordinates. The hypothesis predicts smaller cross-city divergence in the canonical frame. If divergence is no smaller, the core claim is false regardless of the downstream numbers.

Finally, the spatial-game extension reports convergence and the gap to the centralized optimum under calibrated imperfect sensing. These are compared with the within-8% figure that Chen and Huang obtained under perfect sensing.