Weak residuals drive the network update directly: NeRI beats PINN and WAN on nonsymmetric and low-regularity elliptic problems
Synopsis
NeRI turns the solution of elliptic operator equations into a single network update driven by the current weak residual: at each step it builds a state-dependent linear functional whose derivative with respect to the trial function recovers that residual, so no global energy functional and no extra test network are needed. It can be read as a neural-tangent-kernel-preconditioned version of classical residual iteration, and penalty and hard-constrained parameterizations extend it to elliptic variational inequalities. In numerical experiments on a nonsymmetric problem with a corner singularity, a discontinuous-diffusion problem and an obstacle-type variational inequality, NeRI's relative errors were smaller than those of the compared PINN and WAN.
Interpretation
When solving elliptic operator equations, network parameters can be updated purely by the current weak residual, without first writing a global energy functional as a fixed loss. The deep Ritz method relies on an energy functional tied to the equation, so it is not directly available when the operator is nonsymmetric or non-potential and no suitable global energy exists. On a nonsymmetric elliptic problem with a crack-tip corner singularity, NeRI's relative error was smaller than those of PINN and WAN, with PINN noticeably larger near the singular point and WAN in between.
Each step freezes the current approximation, builds a state-dependent linear functional, takes one gradient step on it, and then re-evaluates the weak residual at the new approximation. That functional only generates the current update direction and is not minimized to convergence as a surrogate variational problem, so there is no inner minimization loop; where an energy functional exists, the update direction coincides with the deep Ritz gradient direction. On a nonsymmetric problem with piecewise-constant diffusion and limited solution regularity, NeRI's relative error was smaller than those of PINN and WAN, with PINN noticeably larger near the interfaces, and the reference solution computed by finite elements on a uniform triangulated grid with centroid quadrature.
The same residual-driven mechanism covers elliptic variational inequalities through penalty and hard-constrained parameterizations, updating only the trial network. Weak adversarial networks parameterize both trial and test networks and run a minimax optimization, whereas NeRI gives an equivalent weak residual pairing through an implicitly defined test function without computing that function explicitly. On a nonsymmetric obstacle-type variational inequality with a biactive region and jumps in the solution's second derivatives across a circle, NeRI's relative error was the smallest of the three, below Prox-PINN and WAN.
Perspective
The numerical experiments cover a two-dimensional nonsymmetric problem with a crack singularity, a low-regularity problem with piecewise-constant diffusion, a high-dimensional p-Laplace problem, a high-dimensional biharmonic problem and a nonsymmetric obstacle-type variational inequality, compared against PINN, WAN, Prox-PINN and IDRM errors reported in the literature. The theory addresses strongly monotone, Lipschitz operator equations and gives finite-time estimates under gradient flow with wide two-layer networks.
The frozen-flow approximation in the error decomposition holds on a fixed time interval with constants that may depend on it, so these estimates do not by themselves establish convergence in a joint large-width and long-time limit. In the p-Laplace example the value of p puts the operator outside the strong-monotonicity assumption, so it lies outside the error analysis. Convergence analysis for the hard-constrained variational inequality formulation is not given, and stationarity of the parameter iteration does not in general imply satisfaction of the original variational inequality.
