Imputing Radio Maps Many Times to Set Safe Power Limits for Shared Spectrum

When an unlicensed radio shares a band with a licensed one, it may only transmit as loudly as the licensed receiver can tolerate. That limit depends on the radio channel between the two devices, and in practice that channel is rarely measured where it matters. Systems therefore fill the gaps in a sparse map of signal measurements, called a radio map, and read off a single "best guess" value. A best guess is the wrong input for a safety limit. If it is too low, the licensed user suffers illegal interference. If it is padded with a worst-case margin, the unlicensed user's throughput collapses. We propose to fill the map many times instead. Each fill is a plausible version of the unmeasured radio environment that is consistent with the measurements, and with the fact that some sensors went silent because the signal was too weak to read. We then choose transmit powers so that the interference ceiling holds across these versions, and tune the safety level on held-out data so the realized violation rate matches a stated target.

In two sentences: treat radio-map completion as multiple imputation (drawing several random completions of the missing data, not one), and keep the draws joint so that the summed interference from several unlicensed transmitters can be evaluated. Then calibrate a single scalar with conformal risk control (a method that gives finite-sample guarantees on an error rate) so that the power ceilings derived from those draws violate the licensed user's limit no more often than a chosen rate $\epsilon$. The test is whether this recovers throughput lost to conservative margins without paying for it in violations.

Core Claim

Underlay spectrum sharing caps each secondary transmitter at $\min(P_{\max}, I_p/g_{sp})$, where $P_{\max}$ is the hardware limit, $I_p$ the interference ceiling at the primary receiver, and $g_{sp}$ the gain from secondary transmitter to primary receiver. Makki et al. analyze this dual ceiling. Huang et al. show that once CSI becomes stale (correlation dropping from 1.0 to 0.7), the conservative margins needed to stay safe make the achievable bit rate fall sharply. Their uncertainty is a fixed analytical perturbation, not something learned from correlated observations. The radio-map literature does estimate unmeasured gains from data. It does so with single point estimates scored by average error: Rahman et al. use nuclear-norm matrix completion and Kriging on UAV measurements, Deep Spectrum Cartography uses learned generative priors, and Zou et al. use inverse-distance weighting and Kriging. Pham et al. warn that low average error can coexist with compressed variance and underestimated extremes. For interference protection, those underestimated extremes are exactly the violations.

Our claim is that three ingredients, used together, give a better throughput–safety trade-off than point-estimate maps with either fixed or conformal margins. The first is joint multiple imputation of the gain field. The second is explicit handling of censoring: a missing reading from a sensor below its sensitivity threshold carries the information "the value was low." The third is a conformally calibrated chance constraint on aggregate interference. The advantage should appear mainly in two regimes: when several secondary transmitters contribute correlated interference to the same protected location, and when missingness is informative. Both regimes defeat marginal, per-link error bars.

None of the connecting work in the file does this. The cartography papers produce one map and never set powers. The CSI-robust allocation papers never learn the uncertainty from data. Lee et al. apply conformal risk control to spectrum sensing, bounding false-negative rates in binary occupancy detection, but not to power ceilings or imputed fields.

Research Program

Imputation model. The imputation model treats each grid cell as a row whose features are spatial coordinates together with the other observed variables: other frequency bands in the Zou data, other altitudes in AERPAW. It fills missing gains with chained equations. In this method (MICE, cycling through variables as in Liew et al.), each incomplete variable is regressed on all the others, and imputed values are drawn at random rather than set to the conditional mean. Cross-band and cross-altitude correlations are what make chained equations more than spatial interpolation.

Censoring. Censored cells are drawn from the conditional distribution truncated above at the sensor's sensitivity threshold, instead of being treated as missing at random. As a higher-capacity alternative imputer, we substitute RefiDiff (Ahamed et al.). RefiDiff warms up with a tree model and then refines with a diffusion model, a generator that turns noise into samples by repeated denoising. It reports strong results under all three standard missingness regimes on tabular data, but it has never been run on radio fields. Each imputation yields a complete joint field, so the $M$ draws $g^{(m)}_k(x)$ give, for every secondary transmitter $k$ and protected location $x$, a sample of the joint gains.

Power allocation. Powers $P_k$ are chosen to maximize secondary sum rate subject to a sample-based chance constraint:

$$\frac{1}{M}\sum_{m=1}^{M} \mathbf{1}\Big[\textstyle\sum_k P_k\, g^{(m)}_k(x) > I_p\Big] \le \delta \quad \text{for all protected } x,$$

where $\delta$ is an internal tolerance. We do not pool the draws with Rubin's rules, the standard multiple-imputation recipe for combining estimates. Those rules target means and standard errors under approximate normality, while this decision depends on a tail.

Calibration. The internal tolerance $\delta$ is set by conformal risk control on calibration maps. The realized violation indicator is monotone in $\delta$, which is the condition the method of Lee et al. needs. The result is a finite-sample guarantee that expected realized violation stays at or below $\epsilon$. That guarantee requires the calibration scenes to be exchangeable with deployment scenes, which in practice means a dense reference survey of comparable terrain. We state this assumption openly and test what happens when it breaks.

Baselines

The naive reference sets powers from a single point-estimate map (inverse-distance weighting, which beat Kriging on the Zou data, and nuclear-norm completion) with no margin. It shows how often average-error-optimized maps violate the ceiling.

The conventional engineering answer is a worst-case margin: a fixed dB backoff chosen on the calibration maps so that violation is at most $\epsilon$. This mirrors the conservative-margin logic in Huang et al.. It defines the throughput floor we must beat.

The strongest competitor on the safety question is split conformal on a point estimator. We take the best point map, nuclear-norm completion (reported by Rahman et al. to beat the best Ordinary Kriging setup by about 0.2 dB RMSE on AERPAW), and add a per-link residual quantile calibrated with the same conformal procedure and the same calibration maps. Aggregate constraints are then handled by the union bound.

The strongest competitor on the modelling question is Gaussian-process conditional simulation, which draws joint Kriging realizations, with and without truncated (Tobit-style) handling of censored cells. Simple Kriging is included because Rahman et al. report it beating Ordinary Kriging by 0.8 dB with only 50 points.

The learned generative baseline is the cartography line of Deep Spectrum Cartography. DowJons reports a squared reconstruction error of about 0.08 at 5% sensor coverage on its indoor data, against about 0.35 for tensor completion. Where it is reproducible, we run it as a point map under the same conformal margin.

For violation rate or throughput under an interference ceiling, no reported result was found on Lacuna for any of these methods. The comparison numbers will therefore be generated here.

All methods see identical masks, identical calibration and test splits, the same $M$ wherever draws are used, and an equal hyperparameter search budget on a validation split. Beating them means matching the target violation rate $\epsilon$ (within conformal finite-sample slack) while achieving higher secondary sum rate than the conformal point-map baseline and the Gaussian-process simulation baseline.

Two ablations isolate the ingredients. The first keeps per-link marginal draws and breaks the cross-link joint structure by independently permuting draws across links. The second treats censored cells as missing at random.

Why not just…?

Why not just add a conformal margin to the best point map? This is the most serious threat. With a single secondary transmitter and random missingness, it may match our method, because a calibrated marginal quantile is all that problem needs. Its weakness is aggregation. Summing per-link quantiles through a union bound ignores the shared shadowing that makes gains from nearby transmitters move together, so it should over-reserve as the number of transmitters grows. Our sweep over one to several secondary transmitters shows directly whether the gap opens. If it does not, we report that joint imputation buys nothing over conformalized point maps.

Why not just use Kriging's predictive variance, or Gaussian-process conditional simulation? Conditional simulation is itself a form of joint multiple imputation, so we treat it as a co-equal baseline rather than a straw man. The bet is that chained equations exploit cross-band and cross-altitude structure that a purely spatial covariance does not, and that real shadowing is skewed. Deep Spectrum Cartography reports that smoothness priors degrade badly under heavy shadowing. If censored Gaussian-process simulation matches our method, the contribution narrows to the calibrated aggregate decision layer, and we will say so.

Why not just optimize against a worst-case uncertainty set? Worst-case sets must be sized for the adversarial error, which reproduces the throughput loss documented by Huang et al.. Varshney shows that overestimated cross-gains drive transmitters near the primary receiver to near-zero power. The fixed-backoff baseline measures exactly this cost at matched violation.

Why not just estimate the covariance directly, without imputing values? Direct parameter estimation, which Pham et al. favour for preserving correlations, returns global parameters. A power decision needs location-specific joint realizations, so it can at most serve as an input to the Gaussian-process baseline.

Proposed Experiments

Datasets and masks. We use the Zou et al. urban map, which has 1 m resolution, three bands (30 MHz, 115 MHz, 2 GHz) and three transmitters, and the AERPAW five-altitude LTE measurements from Rahman et al.. The dense measurements serve as ground truth, and we impose synthetic masks under three mechanisms. The first is random. The second is route-dependent, mimicking UAV or rover paths. The third is censoring below a sensitivity threshold, plus a variant in which readings are dropped when local power is high, the case where ignoring missingness is anti-conservative. Coverage runs from 50 points (the regime where Rahman et al. found Kriging fragile) up to dense.

Multi-transmitter scenarios. Where the Zou data separate the per-transmitter contributions, the three transmitters serve as secondary transmitters. Otherwise we build multi-transmitter scenarios by displacing and summing measured fields. This is a stated limitation, because it imposes correlation structure artificially. Protected locations are held-out cells.

Metrics. Our metrics are realized aggregate-violation rate against $\epsilon \in \{0.01, 0.05, 0.1\}$, secondary sum rate at matched violation, the full trade-off curve, and RMSE, the last only to expose cases where low average error still yields poor tails.

Decisive checks. Three results decide the outcome. First, at matched violation, does joint multiple imputation beat the conformal point map, and does that margin grow with the number of transmitters? Second, under the censoring masks, does the ablation that treats censored cells as random either violate the target (under high-power dropout) or forfeit throughput (under sensitivity censoring)? Third, when calibration comes from one terrain type and testing from another, how far does realized violation drift from $\epsilon$? This last check bounds how much the guarantee depends on a representative reference survey.