Lu and Tsai propose MLSG, a multilevel stochastic-gradient neural solver that represents boundary densities with a network and scales to million-level discretizations for second-kind boundary integral equations
Synopsis
Lu and Tsai propose a multilevel stochastic-gradient neural solver (MLSG) for second-kind boundary integral equations that represents the unknown boundary density with a neural network optimized by stochastic residual minimization over a hierarchy of successively refined Nyström discretizations, warm-starting each level with the previous level's network parameters; the algorithm avoids grid-transfer operators and hierarchical fast-summation machinery, relying instead on batched kernel evaluations and standard network forward and backward passes, and is demonstrated on Laplace/Poisson and Helmholtz problems in two and three dimensions plus an exterior Robin problem on a hypersurface in R^4, under both parametric and signed-distance surface representations at up to million-scale discretizati
Figure 5 : Reconstructed total field | u + u inc | |u+u^{\mathrm{inc}}| for the four-sphere problem. Left: n = 256 n=256 , κ = 4 \kappa=4 . Right: n = 512 n=512 , κ = 8 \kappa=8 .
arXivInterpretation
Introduces MLSG: a neural network represents the unknown boundary density, trained by stochastic residual minimization across a hierarchy of successively refined Nyström discretizations, with each level warm-started from the previous level's parameters. Compared with conventional boundary-element/integral-equation workflows, the scheme retains a continuous, grid-independent density representation and explicitly avoids grid-transfer operators and hierarchical fast-summation machinery. The abstract states the algorithmic structure and design goal, and notes reliance on batched kernel evaluations and standard network forward and backward passes that map onto modern GPU architectures; specific convergence constants and complexity bounds require the full text.
Provides an analysis of where convergence rates come from: for uniformly stable second-kind discretizations, strongly nonuniform contraction rates originate in the empirical neural tangent kernel (NTK) rather than in the discretized operator. Brings within-level parameter updates that can reshape the NTK, and refinement that re-samples the tangent kernel on a richer discrete space and reveals directions not adequately resolved on coarser levels, into the explanatory framework. The abstract states this conclusion and its mechanism; the proof conditions and scope of applicability require the full text.
Gives a cross-level estimate bounding the warm-start loss in terms of the preceding training tolerance and quadrature error, motivating a tolerance schedule that balances optimization and discretization errors. Offers a quantitative basis for when to stop a coarse level and when to refine, rather than relying on empirical settings alone. The abstract states the existence and purpose of this estimate without giving constants or numbers; it is a theoretical result to be verified in the full text.
Demonstrates the method on Laplace/Poisson and Helmholtz problems in two and three dimensions and an exterior Robin problem on a hypersurface in R^4, covering both parametric and signed-distance surface representations at up to million-scale discretizations. Extends the setting beyond common two- and three-dimensional cases to an exterior Robin problem on a higher-dimensional hypersurface, while covering two surface representations. The abstract lists problem types, dimensions, surface representations, and the discretization scale range, but gives no error values or comparisons against baseline methods.
Perspective
The work targets second-kind boundary integral equations, particularly settings with uniformly stable discretizations; it starts from a continuous, grid-independent density representation and suits problems that admit Nyström discretization and batched kernel evaluations on GPUs. The experiments described in the abstract cover Laplace/Poisson and Helmholtz problems in two and three dimensions plus an exterior Robin problem on a hypersurface in R^4, with parametric and signed-distance surface representations at up to million-scale discretizations. For researchers and implementers who want to control residual tolerance at high resolution without introducing grid-transfer or hierarchical fast-summation machinery, this route offers an alternative to try.
The abstract gives no specific error values, convergence constants, runtimes, or comparisons with baseline methods, so the practical computational advantage of the approach relative to conventional multilevel boundary elements or fast algorithms at a given residual tolerance still needs the full text's data. The NTK analysis addresses uniformly stable second-kind discretizations, and whether the conclusions hold for other discretizations or operator classes is an open question. How the tolerance schedule in the cross-level estimate should be chosen in practice, and memory and accuracy behavior at million-scale discretizations, are not elaborated in the abstract. In addition, the available text is the arXiv abstract and metadata page without body figures or numerical tables, so the experimental details summarized here are limited to what the abstract states.