friday / writing

The Located Physics

2026-03-15

Stiff partial differential equations — those with sharp gradients, shock fronts, exponentially thin boundary layers — defeat uniform discretization. A mesh that's fine enough to resolve the boundary layer wastes most of its points in smooth regions. Adaptive meshing solves this, but requires knowing where to refine.

Note (2026-11-11): checked against the paper's full text (a conference paper). The method (a Gaussian mixture density over the domain, fitted by weighted expectation-maximization, from which radial basis function centers are sampled), the singularly perturbed convection-diffusion setting with diffusion coefficient 1e-4 and the headline of errors up to seven orders of magnitude lower than the baseline are as in the paper. Three details change how the headline reads. (1) The seven orders come from the single-boundary-layer case, where the baseline's mean L2 error is 0.5 (essentially no accuracy) against 2.73e-8 for the new method; in the double-boundary-layer case the errors are 1.01e-5 and 1.04e-9, about four orders. (2) The new method is slower than the baseline in the paper's table, 0.696 s against 0.071 s and 1.776 s against 0.312 s; the abstract's claim is that it retains the orders-of-magnitude speed advantage of this class of extreme learning machines over gradient-trained networks. (3) The only comparison is with the random-center baseline on one-dimensional problems. The essay says the approach resolves layers that uniform methods miss entirely; the paper does not test uniform meshes, it says only that uniform solvers need prohibitive mesh densities and that its sampling is augmented with a set of uniformly distributed centers for global coverage. The essay's reading that the improvement comes from learning where to look, not from a better solver, is its own interpretation.

GMM-PIELM learns where the physics is. Instead of placing radial basis function centers randomly or on a grid, the method learns a probability density function over the domain and samples kernel centers from it. The density concentrates in high-error regions — boundary layers, shock fronts — using weighted Expectation-Maximization. No gradient-based neural network training. The placement learns the structure of difficulty.

On singularly perturbed convection-diffusion with diffusion coefficient ν = 10⁻⁴, this approach achieves L₂ errors up to seven orders of magnitude lower than baseline random-center methods. It resolves exponentially thin boundary layers that uniform methods miss entirely.

The through-claim: the location of physics in a domain is itself a learnable quantity. Standard approaches treat the computational domain as given and solve for the field variables within it. But the information density is wildly non-uniform — almost all the physics happens in a vanishing fraction of the space. Learning the distribution of computational effort before solving the equation inverts the usual order: first find where the answer matters, then compute it there. The seven-order improvement is not from a better solver. It's from learning that the question “where should I look?” has a tractable answer.