Overview

Universal approximation asks when neural networks and related models can approximate target functions, operators, or probability transformations, a prerequisite for learning what those models can express. Common approaches prove density under conditions on activation functions and depth, derive error rates for smoothness classes, and study architectures such as residual networks, invariant models, Kolmogorov–Arnold networks, and signature-based models. Results now extend beyond compact finite-dimensional domains to derivatives, path-dependent processes, variable-size sets, partial differential equations, and measure-valued maps. A central tension is that universality may fail or require substantial resources under restrictions such as narrow architectures, small embeddings, atomic inputs, finite precision, or physical noise, making quantitative and practical guarantees important open concerns.

Latest papers

CardsList

32 papers3 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.

  1. Re-uploading quantum data: a universal function approximator for quantum inputs

    Sep 23, 2025Hyunho Cha, Daniel K. Park, Jungwoo LeeQuantum Machine LearningUniversal Approximation

  2. On shallow feedforward neural networks with inputs from a topological space

    Apr 3, 2025Vugar IsmailovFeedforward Neural NetworksNeural Network Approximation Theory