Radiomap blind prediction infers radio maps from environment and base station representations without field measurements. Because those representations are incomplete in practice, the same observable input can correspond to multiple physical states and therefore multiple radiomaps. The authors develop a population-level theory showing that the optimal deterministic target under incomplete observation is the conditional mean of the radiomap given the observables, and that prediction error separates into reducible predictor approximation error and irreducible uncertainty from missing physical information.
They also characterize the train-test risk gap and how enriching observations reduces uncertainty. A key finding is that propagation priors play a dual role: they provide physically meaningful guidance, but their implementation can bias the attainable predictor. This motivates RadioDecomp, which treats a prior-guided predictor as a correctable base and learns its remaining predictable discrepancy through residual refinement. Two instantiations, RadioFR and RadioLSR, are tested across random, cross-configuration, and cross-environment settings, and both improve on their respective bases.
The result matters because it gives radiomap prediction a firmer statistical foundation under the realistic condition of incomplete inputs, and it offers a concrete way to combine physics-based priors with learned residual correction rather than choosing between them.