Numerical Approximation Capacity of Neural Networks with Bounded Parameters: Do Limits Exist, and How Can They Be Measured?

The Universal Approximation Theorem posits that neural networks can theoretically possess unlimited approximation capacity with a suitable activation function and a freely chosen or trained set of parameters. However, a more practical scenario arises when these neural parameters, especially the nonlinear weights and biases, are bounded. This leads us to question: $\mathbf{\text{Does}}\mathbf{\text{the}}\mathbf{\text{approximation}}\mathbf{\text{capacity}}\mathbf{\text{of}}\mathbf{\text{a}}\mathbf{\text{neural}}\mathbf{\text{network}}\mathbf{\text{remain}}\mathbf{\text{universal,}}\mathbf{\text{or}}\mathbf{\text{does}}\mathbf{\text{it}}\mathbf{\text{have}}\mathbf{\text{a}}\mathbf{\text{limit}}\mathbf{\text{when}}\mathbf{\text{the}}\mathbf{\text{parameters}}\mathbf{\text{are}}\mathbf{\text{practically}}\mathbf{\text{bounded?}}\mathbf{\text{And}}\mathbf{\text{if}}\mathbf{\text{it}}\mathbf{\text{has}}\mathbf{\text{a}}\mathbf{\text{limit,}}\mathbf{\text{how}}\mathbf{\text{can}}\mathbf{\text{it}}\mathbf{\text{be}}\mathbf{\text{measured?}}$ Our theoretical study indicates that while universal approximation is theoretically feasible, in practical numerical scenarios, Deep Neural Networks (DNNs) with any analytic activation functions (such as Tanh and Sigmoid) can only be approximated by a finite-dimensional vector space under a bounded nonlinear parameter space (NP space), whether in a continuous or discrete sense. Based on this study, we introduce the concepts of $\epsilon \mathit{\text{outer}}\mathit{\text{measure}}$ and $\mathit{\text{Numerical}}\mathit{\text{Span}}\mathit{\text{Dimension}}\mathit{\text{(NSdim)}}$ to quantify the approximation capacity limit of a family of networks both theoretically and practically. Furthermore, drawing on our new theoretical study and adopting a fresh perspective, we strive to understand the relationship between back-propagation neural networks and random parameter networks (such as the Extreme Learning Machine (ELM)) with both finite and infinite width. We also aim to provide fresh insights into regularization, the trade-off between width and depth, parameter space, width redundancy, condensation, and other related important issues.