Singular parameters and missing limits in neural PDE solvers

DOI10.48550/arXiv.2610.06770MaRDI QIDQ7367904arXivFDO

Authors Daniel Fernández

Publication date 5 October 2026

Published in arXiv

Tldr The paper analyzes the behavior of deep neural networks for solving partial differential equations, particularly focusing on the issue of missing limits when parameters grow without bound. The main contribution is to connect this phenomenon with unbounded hidden parameters or redundant neurons in tanh-networks, and propose a method to recover missing functions by incorporating kernel derivatives into the model.

Full work available at URL https://doi.org/10.48550/arXiv.2610.06770

Mathematics Subject Classification ID

Approximation by other special function classes (41A30) Multidimensional problems (41A63) Sobolev spaces and other spaces of ``smooth functions, embedding theorems, trace theorems (46E35) Numerical radial basis function approximation (65D12) Artificial neural networks and deep learning (68T07)

arXiv classification

cs.LGmath.NA

Cites work

• Affine symmetries and neural network identifiability
• Approximation by polynomials
• Convergence Guarantees for Gradient-Based Training of Neural PDE Solvers: From Linear to Nonlinear PDEs
• DGM: a deep learning algorithm for solving partial differential equations
• Error estimates and condition numbers for radial basis function interpolation
• Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating PDEs
• Fourier Series in Control Theory
• Geometry and singularities of the Prony mapping
• Limit problems for interpolation by analytic radial basis functions
• Networks and the best approximation property
• Nonclosedness of sets of neural networks in Sobolev spaces
• On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type PDEs
• On the convergence of PINNs
• Phase diagram for two-layer ReLU neural networks at infinite-width limit
• Physics-informed kernel learning
• Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
• Prevalence of neural collapse during the terminal phase of deep learning training
• Singularities Affect Dynamics of Learning in Neuromanifolds
• Solving Nonlinear PDEs with Sparse Radial Basis Function Networks
• Solving PDEs with radial basis functions
• Some trigonometrical inequalities with applications to the theory of series.
• Sparse RBF Networks for PDEs and nonlocal equations: function space theory, operator calculus, and training algorithms
• Stable Computations with Gaussian Radial Basis Functions
• Tensor Rank and the Ill-Posedness of the Best Low-Rank Approximation Problem
• The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems
• Topological properties of the set of functions generated by neural networks of fixed size
• Two-Layer Neural Networks for Partial Differential Equations: Optimization and Generalization Theory
• Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks
• WAN discretization of PDEs: best approximation, stabilization, and essential boundary conditions
• When and why PINNs fail to train: A neural tangent kernel perspective

This page was built for publication: Singular parameters and missing limits in neural PDE solvers

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7367904)