REM Confidence Interval

The confidence interval is fundamentally derived from the posterior variance of a Bayesian model, most commonly a Gaussian Process (GP). After observing sparse sensor data, the GP computes a predictive distribution at each unobserved location. The variance of this distribution quantifies the model's epistemic uncertainty—how much the prediction might deviate from the true RF power due to limited nearby measurements. A wide interval signals high uncertainty in areas far from sensors or in complex propagation environments.

The width of the confidence interval is directly governed by the variogram or kernel function parameters. As the distance from the nearest sensor increases, spatial correlation decays, causing the predictive variance to grow. The interval captures the nugget effect (measurement noise at zero distance), the sill (maximum variance), and the range (distance where correlation becomes negligible). This geostatistical property ensures the interval reflects the physical reality of radio wave propagation and sensor density.

A well-calibrated 95% confidence interval means that if the REM estimation process were repeated many times, the true spectrum value would fall within the computed interval 95% of the time. Calibration is critical for operational safety. An overconfident, narrow interval risks harmful interference to incumbents, while an underconfident, wide interval leads to inefficient spectrum underutilization. Proper calibration requires rigorous empirical validation against held-out sensor data.

Unlike a simple global error bar, REM confidence intervals are heteroscedastic, meaning the uncertainty varies spatially. Areas with dense sensor coverage or simple line-of-sight propagation exhibit tight, low-variance intervals. Conversely, deep urban canyons, shadowed regions behind terrain, or locations near interference sources show inflated intervals. This spatial granularity allows a cognitive radio to make a localized risk assessment, such as transmitting at higher power only where confidence is high.

The total confidence interval aggregates uncertainty from multiple sources. Beyond sensor measurement noise and spatial interpolation error, a significant component is propagation model mismatch. When a REM uses a deterministic model like Longley-Rice or ray tracing, the interval must account for errors in the digital elevation model, building geometry inaccuracies, and atmospheric assumptions. Bayesian frameworks can integrate this prior model uncertainty into the final predictive variance.

In a Spectrum Access System (SAS), the confidence interval is not merely a visualization but a hard decision boundary. A secondary user is permitted to transmit only if the upper bound of the confidence interval for predicted interference at an incumbent receiver remains below a regulatory threshold. This conservative policy, known as worst-case uncertainty bounding, mathematically guarantees incumbent protection even under estimation error, transforming statistical uncertainty into a binary spectrum availability decision.

The primary Bayesian non-parametric method for generating REM confidence intervals. A Gaussian Process defines a distribution over functions where any finite set of points follows a multivariate normal distribution. Key properties:

• Mean function: The predicted spectrum value at any coordinate
• Covariance kernel: Encodes spatial correlation assumptions (e.g., Matérn, squared exponential)
• Predictive variance: Directly provides the confidence interval—high variance in unsensed regions, low variance near sensor nodes
• Naturally handles noisy observations without overfitting

A geostatistical best linear unbiased predictor that is mathematically equivalent to Gaussian Process regression under specific conditions. Ordinary Kriging assumes an unknown constant mean, while Universal Kriging incorporates a deterministic trend. The kriging variance quantifies estimation uncertainty:

• Depends solely on the variogram model and spatial configuration of sensors
• Zero at measurement locations, increasing with distance
• Forms the basis for confidence ellipsoids around interpolated values
• Widely used in spectrum cartography for its computational efficiency

The foundational geostatistical function that quantifies spatial autocorrelation of RF measurements as a function of separation distance. The empirical variogram is computed as:

• γ(h) = ½E[(Z(x) - Z(x+h))²] where h is the lag distance
• Nugget: Micro-scale variation or measurement error at zero distance
• Sill: The variance where spatial correlation flattens
• Range: Distance beyond which measurements are uncorrelated
• Poor variogram fitting directly inflates confidence interval width and degrades REM reliability

The statistical framework underlying REM uncertainty quantification. Unlike frequentist methods, Bayesian approaches treat unknown spectrum values as random variables with prior distributions:

• Prior: Initial belief about spectrum occupancy (e.g., propagation model output)
• Likelihood: Probability of sensor observations given the true field
• Posterior: Updated belief after incorporating measurements—this is the REM estimate
• Credible interval: The Bayesian analog of a confidence interval, directly interpretable as 'there is a 95% probability the true value lies in this range'

A cognitive map architecture that extends confidence intervals into the temporal domain. By integrating recurrent neural networks or temporal Gaussian Processes, predictive REMs forecast future spectrum states with associated uncertainty bounds:

• Epistemic uncertainty: Model uncertainty due to limited training data, reducible with more samples
• Aleatoric uncertainty: Inherent noise in spectrum dynamics, irreducible
• Confidence intervals widen as the prediction horizon extends
• Enables proactive spectrum handoff before congestion or primary user arrival

A sensing failure mode that directly impacts REM confidence. When a secondary user is obstructed from detecting a primary transmitter by terrain or buildings, the REM may show falsely low occupancy probability with misleadingly narrow confidence intervals:

• Creates 'confidence holes' where the map is overconfident in vacant spectrum
• Mitigated by cooperative sensing and shadow fading margin incorporation
• Confidence intervals must account for sensor geometry, not just measurement noise
• Drives the need for physics-informed priors in Bayesian REM construction