The problem begins with a practical constraint. Conventional material characterization at these frequencies relies on terahertz time-domain spectroscopy or vector network analyzers, both of which require phase-coherent measurements that capture both the amplitude and the phase of the transmitted wave. In real deployment scenarios, such as profiling walls for site-specific ray tracing or assessing materials in the field for integrated sensing and communication systems, only the received power—the intensity envelope—is reliably measurable. At sub-terahertz wavelengths, surface roughness on the scale of typical construction tolerances randomizes the optical phase, washing out the interference fringes that coherent methods depend on. The researchers therefore worked with the incoherent penetration-loss model recommended by the ITU-R P.2040-4 standard, in which total loss splits into a frequency-independent interface reflection term governed by permittivity and a bulk absorption term governed by both permittivity and thickness.

Under this intensity-only constraint, the authors uncovered a severe information bottleneck. The Fisher information matrix, the mathematical object that quantifies how much information a set of measurements carries about unknown parameters, becomes nearly rank-one when measurements are taken at a single incidence angle. Its condition numbers, which measure how badly conditioned the estimation problem is, range from ten million to a billion across common building materials. The physical origin is elegant: the absorption gradient at every frequency points along essentially the same direction in the two-dimensional parameter space of permittivity and thickness, so sweeping across the 100-to-300-gigahertz band adds little independent information. The two Jacobian columns—the sensitivity profiles of the loss with respect to each parameter—have a cosine similarity exceeding 0.996, meaning they are almost perfectly collinear. In plain terms, permittivity and thickness are jointly unidentifiable from a single-angle intensity measurement.

To make this abstract diagnosis concrete, the researchers deployed neural networks as diagnostic probes rather than as miracle solvers. They trained a fully connected deep network with roughly 930,000 parameters to estimate permittivity and thickness from noisy 201-point penetration-loss spectra, then compared its mean squared error against the Cramér–Rao lower bound, the statistical floor that limits any unbiased estimator. The ratio of the bound to the actual error, which they call the efficiency eta, was mapped as a heatmap across the parameter space, complemented by a second metric called Gain that measures improvement over a naive constant predictor. Together, these two indicators distinguish regions where the data themselves are information-poor from regions where an estimator simply fails to exploit available information.

The results split into two strikingly different regimes depending on how the loss tangent, the material’s dissipative property, is treated. When the wall material class is known—drywall, glass, concrete—so that the loss tangent is bounded to a tabulated class-representative value, the problem is genuinely learnable. The network achieved a root-mean-square error of 1.57 on permittivity against a naive baseline of 2.72, a 66.6 percent gain, and reduced absolute errors by 41 to 89 percent on four of five ITU-R reference materials. But when the loss tangent follows a continuous, two-parameter conductivity law that varies sample to sample—the arguably more realistic open-world case where the material class is unknown—the same architecture degraded to within 1.5 percent of the naive predictor. The extra nuisance entropy washes out the absorption-slope signature the network needs. Notably, the authors explicitly corrected an earlier claim from their own work: the continuous frequency-dependent model is the harder case, not the easier one.

Crucially, the team demonstrated that this precision limit is not an artifact of neural network design. They benchmarked against a maximum-likelihood estimator that is essentially Cramér–Rao-efficient, with efficiency ratios between 0.61 and 1.09, and found it cannot beat the same one-to-two RMSE floor on permittivity. Monte-Carlo dropout and deep ensemble estimators reproduced both the accuracy floor and the spatial efficiency pattern, with Spearman rank correlations between 0.88 and 1.00 across architectures differing by a factor of twenty in parameter count. The bottleneck, the authors conclude, is information-geometric rather than architectural. Even uncertainty-aware methods that report their own confidence offer no escape: the ensemble spread grew only mildly toward the degenerate region and correlated weakly with true error, which is precisely why a Cramér–Rao-based diagnostic, not an estimator’s self-reported variance, is the right instrument for certifying identifiability.

The framework survived an extensive battery of robustness checks. Substituting independently published terahertz time-domain spectroscopy material parameters into the same forward model reproduced the near-degeneracy, with condition numbers of the same order and Jacobian collinearity above 0.995—a sanity check on measured parameters, though the authors are careful to note it is not a validation against measured spectra, which remains future work. Dynamic-range censoring, alternative noise models, incidence-angle uncertainty, and mismatched forward models all confirmed the same model-agnostic bottleneck. A two-layer wall analysis showed the problem only worsens, with condition numbers climbing to ten billion or beyond.

The security payoff of the framework is a worked example that translates estimation uncertainty into certifiable protection-zone boundaries. For a learnable plasterboard wall, the propagation of estimation error into loss uncertainty yields a margin of about 16 decibels at three meters, an interception probability of roughly four in a hundred thousand, and a confident certificate. A fifteen-centimeter concrete wall presents a subtler verdict: it is link-budget-safe, since its roughly 256-decibel penetration loss buries any eavesdropper’s signal far below threshold, but it is not sensing-certifiable, because its entire spectrum exceeds the receiver’s 60-decibel dynamic range and its parameters cannot be identified at all. The genuinely dangerous case is a moderate-loss wall whose eavesdropper signal sits near the secrecy threshold, where a few decibels of estimation uncertainty can flip the verdict.

The study also points to a concrete remedy. Adding even a single oblique measurement angle rotates the absorption gradient and introduces an information direction orthogonal to the normal-incidence one, cutting the Cramér–Rao bound by factors of two for thin low-loss slabs and more than twenty-five for thick absorptive materials. Two-angle diversity reduced the concrete condition number from 580 million to 8.5 million and the best-case loss uncertainty from 68 to 5.8 decibels, restoring certifiability, and the gain remains robust to one-to-two degrees of pointing error. For 6G designers, the message is twofold: neural networks can approach the theoretical limit of what intensity-only wall sensing can deliver, but that limit itself is set by physics, and overcoming it will require measuring smarter, not training longer.

Subject of Research: Information-theoretic efficiency diagnosis of neural-network estimators for sub-THz intensity-only wall sensing

Article Title: Information-theoretic efficiency diagnosis of neural-network estimators for sub-THz intensity-only wall sensing

Article References: Wu, Q., & Huang, W. (2026). Information-theoretic efficiency diagnosis of neural-network estimators for sub-THz intensity-only wall sensing. Cybersecurity, 9(1), Article 233. https://doi.org/10.1186/s42400-026-00659-3

Image Credits: AI Generated

DOI: 10.1186/s42400-026-00659-3

Keywords: Information-theoretic, efficiency, diagnosis, neural-network, estimators, sub-THz, intensity-only, wall, sensing, scientific research