Corner-Aware Random Features: Spectrally Enriched Mesh-Free Solvers for Interfaces with Sharp Corners

Many physical problems involve two materials meeting along an interface, for example heat flowing through a composite or current through a cornered inclusion. Near a sharp corner or vertex of that interface, the solution is not smooth. It behaves like a power of the distance to the corner, $r^{\alpha}$, with a fractional exponent $\alpha$. Numerical methods that assume smoothness lose accuracy there. Mesh-free "random feature" solvers fit the solution as a linear combination of fixed, randomly initialized basis functions by solving one least-squares problem. On smooth interfaces they reach spectral accuracy, meaning the error falls exponentially as basis functions are added. That advantage disappears once a single corner appears.

The exponent $\alpha$ is not a mystery. Zoom in on a corner and the geometry looks like a cone. The exponent is set by the lowest eigenvalue of a small eigenvalue problem on the cone's cross-section, which is a slice of the unit sphere. This proposal computes that cross-sectional spectrum numerically for each corner, including cross-sections that themselves have corners and coefficient jumps. It then adds the resulting singular functions $r^{\alpha_k}\phi_k(\theta)$ to the random feature basis. The claim is that this restores exponential convergence on cornered interfaces at a cost of a handful of extra unknowns. It should also deliver the corner's stress-intensity-like coefficient as a by-product.

Core Claim

Three directions meet here, with unequal weight. Spectral domain characterization supplies the mechanism. Poisson type inequalities with respect to a cone and their applications shows that harmonic functions in a cone $C_n(\Gamma)$ grow or decay like $r^{\aleph^\pm}$, with $\aleph^\pm$ fixed by the first Laplace–Beltrami eigenvalue $\lambda_1$ of the spherical cross-section $\Gamma$. That paper proves this only for $C^2$-smooth $\partial\Gamma$, a single homogeneous material, and continuous analysis with no solver.

Support cone dependency provides the organizing principle. The singular behavior at a corner depends only on the local tangent cone of the interface and on the coefficient contrast, not on the global shape. Each corner can therefore be treated by a local spectral computation on its cone. This echoes the way The extremals of the Alexandrov–Fenchel inequality for convex polytopes reduces polytope behavior to shared normal cones. We borrow only that localization viewpoint, not its mixed-volume machinery.

Geometric singularity handling is the application regime. Finite volume formulation of the MIB method for elliptic interface problems explicitly states that its scheme "does not explicitly accommodate geometric singularities—such as sharp edges, corners, or cusps on the interface." The random feature method of The Random Feature Method for Solving Interface Problems is demonstrated only on smooth and nearly touching interfaces.

Concretely, consider $-\nabla\cdot(\beta\nabla u)=f$ with $\beta$ piecewise constant across the interface. At a corner, seek local solutions of the form $u=r^{\alpha}\phi(\theta)$. Substituting gives a transmission eigenvalue problem on the cross-section $\Gamma\subset S^{n-1}$:

$$-\nabla_S\cdot(\beta\nabla_S\phi)=\lambda\,\beta\,\phi \ \text{ on } \Gamma, \qquad [\phi]=0,\ [\beta\,\partial_\nu\phi]=0 \text{ across the interface arcs on } \Gamma .$$

Here $\nabla_S$ is the spherical gradient and $\partial_\nu$ is the derivative normal to the interface arc within the sphere. The admissible exponents follow from $\lambda=\alpha(\alpha+n-2)$, which is the relation behind Shu et al.'s $\aleph^+$.

In 2D the cross-section is a circle and the problem is one-dimensional. In 3D, the cross-section at a polyhedral vertex is a spherical polygon whose own corners violate the $C^2$ assumption. Interface edges contribute separate two-dimensional wedge exponents in the plane normal to the edge. Our contribution is three things together. We compute these spectra robustly for arbitrary polygonal and polyhedral interface corners with contrast. We embed the resulting singular functions in a random feature solver with a partition of unity, which splits the domain into overlapping patches blended by smooth weights. And we show that this, rather than local refinement, is what recovers exponential convergence.

Singular-function enrichment is a classical idea in finite elements generally, though not represented in this file. What is new is pairing it with a mesh-free spectral-accuracy solver and an automatic cone-spectrum computation for 3D vertices with contrast.

Research Program

The first component is a cone spectrum solver. In 2D, the eigenproblem on the circle reduces to a transcendental equation in $\alpha$ for given angles and contrast, which can be solved to machine precision. In 3D, we discretize the transmission problem on a spherical polygon using random features on the sphere with the two-set strategy of Chi et al.: separate feature sets on each material sector, coupled through jump conditions at collocation points. We take the lowest few eigenpairs. For a cross-section with its own corners, the eigenfunctions are themselves singular at the spherical vertices. We handle this recursively by enriching the spherical basis with the 2D wedge functions at those vertices. This recursion is the precise point where Shu et al.'s smoothness assumption fails, and we address it directly.

We will check the 3D solver against cases with closed-form answers. Circular cones have Legendre-function eigenvalues. The octant cross-section has known values for the homogeneous problem.

The second component is the enriched global solver. We start from the partition-of-unity random feature method. At each corner we add $K$ functions $\chi(r)\,r^{\alpha_k}\phi_k(\theta)$, where $\chi$ is a smooth cutoff. In 3D we add edge functions along each interface edge, multiplied by a smooth function of arclength. Their unknown coefficients join the same least-squares system. Singular and smooth features will be nearly linearly dependent, so the system will be badly conditioned. We therefore solve it with the count-sketch preconditioned LSQR of High-precision randomized iterative methods for the random feature method, which reports reducing condition numbers up to $10^{18}$ to between 3 and 7.

The third component is analysis. We will state and test an error bound of the form $\|u-u_N\|\lesssim e^{-cN^{\gamma}}+\varepsilon_\alpha N^{p}$. The first term is the smooth-remainder approximation error. The second captures exponent misspecification, where $\varepsilon_\alpha=|\alpha-\hat\alpha|$. The prediction is that error saturates at a floor proportional to $\varepsilon_\alpha$. The reason is that subtracting $r^{\hat\alpha}\phi$ leaves a residual $r^{\alpha}-r^{\hat\alpha}\approx(\alpha-\hat\alpha)\,r^{\alpha}\log r$. That residual is still singular and limits the smooth part's convergence. This makes the accuracy of the cone spectrum directly measurable.

Baselines

The naive reference point is the unenriched partition-of-unity random feature method of The Random Feature Method for Solving Interface Problems, run on the same cornered problems. We expect it to converge only algebraically, at a rate tied to $\alpha$. This defines the gap the proposal must close.

The strongest competitor on the question of whether enrichment is needed at all is the same random feature method with geometrically graded patches. Patch sizes shrink by a fixed ratio toward each corner, mimicking hp-refinement. This is the one alternative that could in principle also recover exponential convergence. We run it with the same total degrees of freedom, the same preconditioned solver, and a grading ratio and number of layers tuned on a held-out geometry with the same budget used to tune our cutoff radius and $K$. Beating it means lower relative $L^2$ and $L_\infty$ error at every matched degree-of-freedom count above a small threshold, and a measurably steeper log-error versus $\sqrt{N}$ slope.

The strongest grid-based competitor is MIB-FVM. Its reported orders of 1.85 to 2.34 hold on low-regularity but smooth-interface benchmarks. No reported result found on Lacuna for MIB-FVM, random features, or PINNs on interfaces with sharp corners. We will therefore run MIB-FVM ourselves on the cornered benchmarks, compare at matched unknown counts, and report its observed order.

For neural solvers we use the domain-decomposition PINN with gradient-enhanced jump conditions of On convergence of neural network methods for solving elliptic interface problems. We also use the weak random feature method of Weak Random Feature Method for Solving Partial Differential Equations. On its 2D static heat problem with non-smooth data it reports an $L_\infty$ error of 0.785, against 1.189 for a PINN. That is the only quantitative low-regularity comparison in the file, and we reproduce it as a sanity check.

Two ablations isolate the key ingredient. The first enriches with the wrong exponent: the homogeneous-material exponent that ignores contrast, or, in 3D, a smooth-cross-section approximation that replaces the spherical polygon by a circular cap of equal area. The second fits the exponent by nonlinear least squares instead of computing it from the spectrum. If the computed spectrum matters, the first ablation should show the predicted error floor proportional to $\varepsilon_\alpha$. The second should be slower and less stable without being more accurate.

Why not just…?

Why not just refine the basis near the corner? Algebraic refinement only buys algebraic rates. Geometric grading is a real threat, because hp-style refinement is known to recover exponential convergence in finite elements. That is exactly why graded patches are our main baseline at matched degrees of freedom. If grading matches enrichment, the finding is still useful, but the proposal's claim narrows to the coefficient-extraction and conditioning advantages. We will report this honestly.

Why not just use the weak formulation, which tolerates low regularity? Moving derivatives onto test functions lets WRFM converge for non-smooth solutions. It does not change what the basis can represent: smooth features still approximate $r^{\alpha}$ slowly. WRFM also scales badly in dimension $d\ge 10$, though that is irrelevant here. We include WRFM on the 2D cornered problems and expect algebraic convergence. If it instead shows exponential decay, our premise is wrong.

Why not just use a domain-decomposition PINN, which has convergence proofs? The Wu et al. proof relies on trace theorems and interface embeddings that assume a regular interface, and it breaks at re-entrant corners. Its reported results are on circular and flower-shaped interfaces. It also offers no mechanism to target the corner exponent. We run it at matched wall-clock time and expect slow, noisy convergence near corners.

Why not just fit $\alpha$ from data instead of solving an eigenproblem? For a single 2D corner this may work. In 3D vertices, however, several exponents and non-trivial angular profiles $\phi_k$ must be fitted jointly, which is a non-convex problem. The second ablation tests this directly.

Proposed Experiments

The file contains no benchmark with cornered interfaces; PDEBench is explicitly "limited to rectangular spatial domains". We will therefore build a small suite with exact or highly resolved reference solutions.

In 2D, we use wedge and checkerboard-type inclusions with contrasts from 1 to 1000. The upper end matches the stress level of Wu et al. Exact solutions are built as $r^{\alpha}\phi(\theta)$ plus a smooth part. We also include a polygonalized version of the Cao et al. jigsaw geometry at $\beta^+=1$, $\beta^-=10$.

In 3D, we use a cube and an L-prism inclusion, with contrast 10 and 100. For these, reference solutions come from the enriched solver at very high resolution, cross-checked against graded-mesh finite elements.

The metrics are relative $L^2$ error, $L_\infty$ error, $H^1$ seminorm error, error in the extracted singular coefficient, condition number before and after preconditioning, and wall-clock time.

The decisive check is a plot of log error against $\sqrt{N}$ (or $N^{1/3}$ in 3D) for all methods. Enriched features should be straight-line exponential. Unenriched features should be algebraic. The wrong-exponent ablation should plateau at a floor proportional to $|\alpha-\hat\alpha|$. A second decisive check is that the 3D cone spectrum solver reproduces known circular-cone and octant eigenvalues to at least eight digits. If enrichment beats graded patches by less than an order of magnitude at matched cost, the method's practical case reduces to coefficient extraction, and the write-up will say so.