arXiv:2611.24355v1 Announce Type: new Abstract: Residual-based methods, including physics-informed neural networks (PINNs), physics-informed Kolmogorov--Arnold networks, and some physics-informed neural operator architectures, approximate PDE solutions by minimizing a discrete least-squares loss formed from the PDE and boundary residuals over a prescribed trial class. This structure is shared by classical least-squares Kansa collocation, whose kernel trial spaces admit an established convergence theory. For neural trial classes, approximation power alone does not ensure that discrete residuals control the continuous solution error. For linear elliptic Dirichlet problems, we extend the kernel-based collocation framework to neural trial classes through elliptic regularity, discrete stability, and consistency. Elliptic regularity supplies the PDE estimate relating solution error to continuous residuals. Discrete stability identifies admissible trial subclasses in which continuous residuals are controlled by a computable discrete loss. This step also quantifies the required number of collocation points. Furthermore, the stability analysis naturally demands a scale-dependent, physics-balanced boundary weight. For the constrained problem, consistency requires accurate comparison functions within the same admissible subclass, a requirement not supplied by neural approximation estimates alone. Specializing to multilayer perceptrons with tanh activations (tanh-MLPs), we construct admissible comparison networks and establish convergence for the constrained residual minimization problem. We further obtain conditional convergence of unconstrained physics-balanced PINNs based on tanh-MLPs under joint network and sampling refinement.