Neural network approximation over finite dimensional Banach spaces (Q6957826)

The paper develops quantitative neural network approximation results for functions defined on finite-dimensional Banach spaces with values in \(\mathbb{R}^N\). The central idea is to exploit the linear homeomorphism between any finite-dimensional Banach space \(E\) and \(\mathbb{R}^N\), thereby reducing the approximation problem to a multivariate setting in Euclidean space.\N\NUsing a perturbed hyperbolic tangent activation function and a symmetrized density construction, the author introduces several multivariate neural network operators: Normalized operators, Quasi-interpolation operators, Kantorovich-type operators and Quadrature-type operators. For each operator, multivariate Jackson-type inequalities are established in terms of the first modulus of continuity. The paper proves both pointwise and uniform convergence and provides explicit quantitative error bounds, including exponentially decaying remainder terms. This work represents a solid and technically rigorous contribution to approximation theory, particularly within the author's established research program on neural network operators.