Universality of Shallow and Deep Neural Networks on Non-Euclidean Spaces

Mathematician Vugar Ismailov has published a groundbreaking theoretical paper establishing that neural networks maintain their universal approximation capabilities far beyond traditional Euclidean spaces. The work, titled 'Universality of Shallow and Deep Neural Networks on Non-Euclidean Spaces,' proves that both shallow and deep networks can approximate any continuous vector-valued function on arbitrary topological spaces, including locally convex spaces. This extends classical approximation theorems that were previously limited to Euclidean domains (ℝⁿ), providing rigorous mathematical foundations for applying neural networks to complex data structures like graphs, manifolds, and other non-Euclidean geometries that increasingly appear in modern AI applications.

A key technical achievement is the paper's treatment of width-constrained deep networks, where Ismailov identifies conditions under which networks with uniformly bounded hidden layer widths retain universal approximation power as depth increases. Using Ostrand's extension of the Kolmogorov superposition theorem, the paper derives explicit width bounds expressed in terms of topological dimension for products of compact metric spaces. This 23-page mathematical framework bridges general topology with neural network theory, offering new tools for analyzing AI systems on diverse data types while connecting to 35 references across topology, functional analysis, and machine learning. The work provides theoretical justification for the empirical success of neural networks on non-Euclidean data and opens avenues for more principled architecture design.