A new preprint on arXiv investigates a fundamental question in deep learning theory: how wide must a neural network be to approximate all continuous functions from R^n to R^m? The authors analyze this in the topology of uniform convergence on compact subsets, a standard setting for universal approximation results.

The abstract indicates that the answer is tied to a seemingly unrelated area of topology: immersions of real projective space. By linking the minimal width problem to this geometric question, the paper offers a new perspective on the limits of neural network expressivity. The full proof and precise statement are not available in the abstract, but the announced connection suggests a deeper mathematical structure underlying universal approximation.