Channel-Selective Information Regulation for Low-SNR Automatic Modulation Recognition

Abstract

Wireless receivers operating in dense and contested spectral environments must often infer modulation type from short, corrupted, and weak observations. In such settings, the classification problem is shaped not only by additive noise but also by oscillator mismatch, fading, interference, finite observation windows, and the uneven distribution of informative signal structure across time and feature channels. The practical consequence is that many deep classifiers fail not because they lack nominal capacity, but because they allocate representational effort inefficiently: they preserve redundant activations, over-amplify nuisance responses, and allow fragile class evidence to be buried beneath feature congestion. This paper develops a technical research framework for low-SNR automatic modulation recognition centered on the concept of information regulation in hierarchical feature systems. The analysis treats deep modulation classifiers as nonlinear operators that compress, route, and reweight class evidence under noise and derives a representation-theoretic view of why channel-selective mechanisms become especially important when the latent signal manifold is weakly expressed. A mathematical formulation is given for the received waveform, multistage feature mapping, channel-gating dynamics, class margin evolution, and error concentration under adverse conditions. The paper then examines how channel-selective compression interacts with residual transport, multiscale memory, and global context summaries to improve latent separability without uncontrolled growth in computation. Rather than presenting architecture as an isolated engineering recipe, the study interprets low-SNR recognition through the geometry of preserved versus overloaded feature subspaces, the stability of decision boundaries under nuisance perturbation, and the statistical efficiency of selective feature amplification. The resulting treatment offers a principled account of why robust modulation recognition depends on controlling information flow as much as on increasing network depth or width.