Condition Numbers and Eigenvalue Spectra of Shallow Networks on Spheres

We present an estimation of the condition numbers of the mass and stiffness matrices arising from shallow ReLUᵏ neural networks defined on the unit sphere~Sᵈ. In particular, when \θⱼ^*\ⱼ₌₁ⁿ ⊂ Sᵈ is antipodally quasi-uniform, the condition number is sharp. Indeed, in this case, we obtain sharp asymptotic estimates for the full spectrum of eigenvalues and characterize the structure of the corresponding eigenspaces, showing that the smallest eigenvalues are associated with an eigenbasis of low-degree polynomials while the largest eigenvalues are linked to high-degree polynomials. This spectral analysis establishes a precise correspondence between the approximation power of the network and its numerical stability.