Work

Voronovskaya-Type Analysis of Neural Network Operators via Fejér and de la Vallée Poussin Kernels

Year: 2026
Type: preprint
Abstract: We develop a rigorous connection between classical positive-operator approximation theory and neural network operators. Three results are established, each with explicit hypotheses. First, we prove a Voronovskaya-type asymptotic expansion for shallow neural operators built from the Fejér kernel, obtaining the rate \(n^{-1}\) with exact constant \(2\log 2\) for the leading term; the curvature of the target appears as the limiting factor, and the constant is shown to be optimal for the Fejér kernel. Second, under an explicit spectral hypothesis on the deep Neural Tangent Kernel (NTK) and an explicit spectral approximation inequality, we derive a depth-dependent bound in Sobolev norms of the form \(n^{-2(s-1)}(\log n)^{\gamma(d-1)}\), where the sign \(\gamma \in \{+1, -1\}\) encodes the direction of the logarithmic correction; the sharp rate \(n^{-2s}\) is stated as a conjecture under an additional non-degeneracy condition on the NTK eigenfunctions. Third, we characterize the rational symbol \(\mathbf{1}_{[-n,n]}(\xi)/(1+\mu|\xi|^m)\) as the unique minimizer of a fractional-Sobolev spectral risk, and we show separately that the modulated de la Vallée Poussin kernel \((1-|\xi/n|^m)_+\) is the minimizer of a shape-constrained problem with a monotonicity constraint. Structural remarks clarify the scope of each result, and we indicate open problems concerning sharp constants, unconditional spectral asymptotics, and high-dimensional extensions.
Language: English
Open Access status: green
OpenAlex ID: W7236760920
Cites:
Cited by:
Related to: 10