Poised Corner Stencils: High-Order Finite Differences with Singular Enrichment at Reentrant Corners

High-order finite difference methods on Cartesian grids can now reach fourth- and sixth-order accuracy near curved boundaries. They do this by picking, for each grid point, a set of nearby nodes that determines a unique local polynomial. Such a node set is called "poised" or "unisolvent." This machinery stops working at a reentrant corner, meaning an inward-pointing corner such as the notch of an L-shaped domain. Near that corner the true PDE solution behaves like $r^{\pi/\omega}\sin(\pi\theta/\omega)$, where $\omega$ is the interior angle and $(r,\theta)$ are polar coordinates centered on the corner. That function is not a polynomial. No polynomial stencil, however well chosen, can recover it, and the accuracy collapses toward the rate the singularity allows. The usual fix is to spend more grid points near the corner through graded or adaptive meshes. The alternative we pursue is to add the known singular functions to the local approximation space and keep the grid uniform. It is not obvious how. Once the space mixes polynomials with fractional powers, the algebraic theory that tells us which node sets are poised no longer applies.

This proposal builds that theory and the stencil generator that goes with it. Its central observation is simple. Polynomials and the corner functions both scale cleanly under dilation about the corner. As a result, whether a corner-anchored stencil is poised, and how well conditioned it is, does not depend on the grid spacing $h$. The whole corner problem therefore reduces to a finite, offline catalogue of stencil shapes for each corner angle. The second observation is the price of this approach. Away from the corner, the singular function becomes locally indistinguishable from a polynomial, so the enriched system becomes nearly degenerate. This fixes a quantitative "enrichment shutoff" distance that we can derive and then test.

Core Claim

Let $V_n^\omega = P_n \oplus \mathrm{span}\{ r^{\alpha_k}\sin(\alpha_k\theta) : \alpha_k = k\pi/\omega \notin \mathbb{Z},\ \alpha_k < n+1\}$ be the degree-$n$ polynomials enriched with every non-polynomial corner mode below the target order. For $\omega = 3\pi/2$ these modes have exponents $2/3, 4/3, 8/3, 10/3,\dots$. The $k=3$ mode $r^2\sin 2\theta$ is already a polynomial and drops out. For a crack ($\omega=2\pi$) the exponents are $1/2, 3/2, 5/2,\dots$. Every basis function is homogeneous about the corner, so dilating a node set $X$ by $h$ gives

$$V(hX) = V(X)\,\mathrm{diag}(h^{d_j}),$$

where $V(X)$ is the interpolation matrix, meaning each basis function $j$ evaluated at each node, and $d_j$ is the degree or exponent of basis function $j$. Two consequences follow. A corner-anchored stencil is poised on every grid exactly when its unit-scale version is poised. After diagonal rescaling, its condition number is also independent of $h$. We claim this makes it possible to extend combinatorial poised-set search to enriched spaces. The corner-mode factor $\sin(\alpha_k\theta)$ vanishes on both corner edges, so boundary nodes carry no information about the singular coefficients. The search must therefore enforce angular coverage of the interior, a constraint that has no counterpart in the polynomial case.

For stencils of size $h$ centered at distance $d \gg h$ from the corner, the singular mode differs from its degree-$n$ Taylor polynomial by roughly $d^{\alpha}(h/d)^{n+1}$, which is the Taylor remainder. The smallest singular value of the rescaled enriched matrix should therefore shrink like $(h/d)^{n+1-\alpha}$ relative to the polynomial part. We predict a shutoff ratio $d^\ast/h$ at which the enriched stencils can be switched to plain polynomial ones. At that distance, the polynomial truncation error $h^{n+1}d^{\alpha-n-1}$ matches the error caused by conditioning. This ratio should grow only slowly with the target order.

Relation to Prior Work

The method comes from the PLG finite difference paper. It finds poised triangular lattices on truncated Cartesian grids by mapping them to "D-permutations" and pruning a depth-first search, and it reports 4th–6th order accuracy and about 5,350 times lower error than second-order embedded-boundary methods on Poisson problems. Its test geometries are ellipses, rotated boxes, and star-shaped smooth domains. It assumes a locally smooth solution and a purely polynomial space, and its group structure depends on polynomial degree counting, which does not hold for fractional exponents. We keep its search-and-prune architecture and its finding that 2 to 5 redundant least-squares points stabilize conditioning. We replace the poisedness test and add the angular-coverage constraint.

RBF-LOI motivates a different design choice. It fits the polynomial space to arbitrary node clusters so that unisolvence holds automatically. However, it reports that its fixed global tolerance fails on irregular, non-smooth clouds and needs manual tuning. That failure is the one our shutoff law is meant to replace with a derived, distance-dependent rule. Orthogonal polynomials on quadratic surfaces of revolution shows that a basis built for a singular geometry, in that case a cone apex, keeps spectral accuracy. It depends on rotational symmetry, however, and does not cover polygonal reentrant corners.

The singularity literature both motivates and constrains the proposal. The nonconforming primal hybrid method shows that conforming methods for the vector Laplacian converge to the wrong solution on non-convex domains, and that a consistent method still converges only at about $h^{1/3}$ in energy on the L-shape. Consistency alone does not buy accuracy. r-adaptive graded meshes recover the optimal $N^{-1}$ rate for linear discontinuous Galerkin, a finite element method whose elements need not match across edges, on L-shaped and crack domains. It does so through analytic grading, with $\gamma = 2/3 - \pi/(3\omega)$ for the $L^2$ norm. This result is the strongest competitor. Its grading pushes nodes into clustered, nearly degenerate configurations, which is exactly what harms high-order reconstruction. Boundary-concentrated meshes double normal-derivative rates, from 0.33 to 0.66 at reentrant corners, at a logarithmic cost in elements. We keep the grid uniform instead of grading it.

Research Program

The first part is theoretical. We will prove the dilation-invariance statement for corner-anchored stencils. We will then characterize poisedness failure for $V_n^\omega$. Failure means the nodes lie on the zero set of some $p + \sum_k c_k r^{\alpha_k}\sin(\alpha_k\theta)$, which is a transcendental curve rather than an algebraic one. We will derive the conditioning law in the transition zone, $d/h \to \infty$, including the constants for $\omega \in \{3\pi/2, 2\pi\}$ and a sweep of angles between $\pi$ and $2\pi$.

The second part is algorithmic. We will extend the prune-and-search generator so that it enumerates corner-anchored stencils on the reference grid for each angle and each corner position relative to the grid. The corner may sit on a grid node or at a generic offset, which gives a finite family. Each candidate will be accepted by a scale-free singular-value threshold rather than an exact determinant test. The generator also enforces angular coverage. The resulting catalogue is computed once and reused at every resolution. Stencils in the transition annulus are chosen by the shutoff rule. Stencils beyond it fall back to standard polynomial stencil generation.

The third part assembles a finite-difference discretization of Poisson and heat problems. Because the corner modes are harmonic, the discrete Laplacian built from enriched reconstructions is exact on them. For Poisson problems with smooth forcing, whose particular solutions can contain logarithmic corner terms when exponents collide with integers, we will add those terms where the corner-expansion theory requires them. We will report when these terms matter.

Baselines

The naive reference point is PLG-FD with polynomial stencils only, run on the same uniform L-shaped and crack grids. It uses the method and stencil selection of the PLG paper, which reports no corner results. We expect its global error to fall to the singularity-limited rate. We will measure that rate rather than assume it.

The strongest existing method is analytically graded r-adaptive discontinuous Galerkin from the graded-mesh paper. It reports the optimal $N^{-1}$ rate on L-shaped and crack Laplace problems, with the best $L^2$ grading at $\gamma = 4/9$ for $\omega = 3\pi/2$. We will reimplement it using the paper's analytic $\gamma$, not a tuned value. We will compare against it at equal degrees of freedom and equal wall-clock time on identical exact-solution benchmarks, and also on the Poisson L-shape problem, which has no closed form and requires a fine reference solution. To give it a fair fight, we will also run higher-order dG on the same graded meshes.

The proposal beats this baseline if two conditions hold. First, our sixth-order configuration shows error decaying clearly faster than $N^{-1}$ in $L^2$ and $L^\infty$. Our target is a rate near $N^{-2}$ or better, since $h^{4}$ or higher with $N \sim h^{-2}$ gives $N^{-2}$. Second, it reaches lower absolute error at matched cost.

For the vector Laplacian, the reference point is the roughly $h^{1/3}$ energy rate reported by the hybrid method. That comparison is secondary, because our first scope is scalar problems.

Two ablations isolate the key ingredients. The first keeps the enriched space $V_n^\omega$ but chooses stencils by nearest-neighbour oversampled least squares instead of catalogue selection. It tests whether the poisedness machinery matters or whether enrichment alone is enough. The second replaces the derived shutoff with fixed radii, which tests the conditioning law.

Why not just…?

Why not just grade the mesh toward the corner and keep polynomial stencils? Grading is proven to restore $N^{-1}$ for linear dG (graded meshes). That is a strong result, and at low order it may be the right tool. For high-order reconstruction, grading produces the clustered, nearly degenerate node sets that break poisedness and conditioning, the tension raised by PLG and by anisotropic virtual elements. Grading also leaves the uniform grid, which removes the main advantage of Cartesian finite differences. The matched-cost comparison above will show which approach wins at each order. If grading wins at fourth order, we will say so.

Why not just add the singular functions and oversample with least squares, skipping poisedness entirely? This is the most serious threat, and the first ablation exists to settle it. Our reasons for doubting it are specific. Boundary nodes carry no singular information, so nearest-neighbour clouds near the corner can be close to rank-deficient in exactly the singular directions. In the transition zone the enriched system degenerates, and extra points do not repair that. The PLG paper found that only 2 to 5 redundant points were needed once the base set was poised. If unconstrained oversampling matches our accuracy and conditioning at comparable stencil size, the catalogue is unnecessary, and the contribution shrinks to the shutoff law.

Why not subtract the singular part with a global coefficient, or map the corner away? Subtracting a global singular function adds a non-local unknown and a coupling row, which breaks the local stencil structure. It also still needs several modes for high order. Mappings move the difficulty into the coefficients, as Olver and Xu note for non-standard geometries, and RBF-LOI shows that local flattening introduces distortion. We will implement global single-coefficient subtraction on the L-shape as an extra comparison. We expect it to be competitive when one mode dominates and to lose at sixth order.

Proposed Experiments

The benchmarks are the Laplace L-shape and crack problems with exact singular solutions and the Poisson L-shape problem from the graded-mesh paper. We add a heat equation on the L-shape and the smooth PLG test geometries, where enrichment is unused, as a regression check. We will report $L^2$ and $L^\infty$ errors against $N$ and wall-clock time, observed convergence orders, stencil condition numbers against $h$, and search time for building the catalogue.

Three checks decide the outcome. The first is whether rescaled corner-stencil condition numbers stay flat across at least four grid refinements, as the dilation argument predicts. The second is whether measured transition-zone singular values follow the predicted $(h/d)^{n+1-\alpha}$ slope, and whether the derived shutoff beats every fixed radius. The third is whether the full method outperforms both graded dG and the enriched-least-squares ablation at matched cost for orders four and six on the L-shape and crack domains. If the flatness check fails, the offline-catalogue premise is wrong. If the ablation matches the full method, we will narrow the contribution to the conditioning law.