Häberle & Bölcskei derive optimal scattering network design for low-dim data

A new theoretical paper by Konstantin Häberle and Helmut Bölcskei tackles a foundational question in representation learning: how to design scattering networks that maximally separate points on low-dimensional datasets. Scattering networks are wavelet-based architectures that compute invariant features by cascading filters and nonlinearities. The authors consider a stripped-down version — fixed monomial nonlinearities, no pooling — so the only free design variable is the frame generated by the network's filters. They model data as rectifiable sets (curves, surfaces, or manifolds embedded in high dimensions) and derive bounds on the separation capacity, i.e., the ability to assign distinct feature vectors to distinct input points.

The key contribution is a two-part design criterion: (i) the filters must sample the data across sufficiently many frequency bands, and (ii) the matrices that couple the frame to the local geometry of the dataset must be well-conditioned. In plain terms, the network succeeds when its filters are dense enough in frequency and when the filter responses are not overly correlated with each other. These results provide formal guidance for practitioners building scattering networks for tasks like point-cloud classification, manifold learning, and invariant recognition. The paper (arXiv:2607.06048) is 19 pages and sits at the intersection of machine learning, information theory, and harmonic analysis.