Exact hierarchical polar differential layer
Implementation & benchmark of arXiv:2609.03461 — Adaptively-refinable polar-spline discrete differential forms: hierarchical construction, exactness, and applications
Source paper: Adaptively-refinable polar-spline discrete differential forms: hierarchical construction, exactness, and applications arXiv:2609.03461 ⓘ · analyzed Sep 4, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace a Cartesian grid or singular polar convolution near the center of a disk with a hierarchy of polar-spline coefficient features. Use separate coefficient spaces for scalar fields, vector 1-forms, and 2-forms, together with sparse derivative matrices satisfying the discrete complex identity D1 D0 = 0. Refine only cells with large learned or physics residuals, while retaining the pole constraints that make the representation smooth despite the collapsed angular boundary.
Formulas
Mathematical statement
The polar map in equation (10) is F(theta,rho) = sum_{i=1}^{m_theta} sum_{j=1}^{m_rho} F_{ij} B^theta_i(theta) B^rho_j(rho), where theta and rho are parameter coordinates, F_{ij} are planar control points, and B^theta_i and B^rho_j are univariate B-spline basis functions; all control points on rho=0 coincide at the pole. The paper constructs scalar, 1-form, and 2-form spaces P^0, P^1, P^2 as pole-compatible subspaces of tensor-product spline spaces. Their basis counts are n^0 = n^{0,int} - m_theta + 2, n^1 = n^{1,int} - 2m_theta + 2, and n^2 = n^{2,int} - m_theta, with n^{0,int} = m_theta(m_rho-1)+1, n^{1,int} = 2m_theta(m_rho-1), and n^{2,int} = m_theta(m_rho-1). Here m_theta and m_rho are the numbers of angular and radial spline functions. The spaces form a discrete de Rham complex, so if D0 maps scalar coefficients to 1-form coefficients and D1 maps 1-form coefficients to 2-form coefficients, then D1 D0 = 0; this is the finite-dimensional statement that curl-grad and div-curl vanish. The cohomology result H^0 approximately R and H^1 = H^2 = 0 means constants are the only globally undifferentiated scalar modes and there are no spurious closed-but-not-exact interior modes on the disk. The hierarchical extension preserves the same cohomological structure under admissible local refinement.
Implementation notes
Integrate this as a spatial encoder and differential output head for a neural operator on a disk, annulus-with-center, or spherical patch. First build a polar hierarchy with angular coordinate theta and radial coordinate rho. At level l, store scalar coefficients x0 in P^0_l, vector or 1-form coefficients x1 in P^1_l, and scalar density or 2-form coefficients x2 in P^2_l. Evaluate the corresponding B-spline basis at query points using Cox-de Boor recursion, but tie all basis functions touching rho=0 according to the pole constraints; equivalently, use the reduced dimensions n^0, n^1, and n^2 above. Precompute sparse incidence or derivative matrices D0_l and D1_l from the univariate B-spline derivative relation and tensor-product Kronecker construction, then project them into the pole-compatible bases. Verify numerically that ||D1_l D0_l|| is at machine precision. A neural block can update all three fields using z0 = MLP0([x0, D1^T x1]), z1 = MLP1([x1, D0 x0, D1^T x2]), and z2 = MLP2([x2, D1 x1]); alternatively use D0 x0 and D1 x1 as fixed linear channels. For adaptive refinement, compute a cell score e_K = mean_{q in K} ||r(q)||^2, where r is the PDE residual or supervised prediction error, mark cells with e_K above a percentile threshold, split them radially or angularly, prolongate coefficients with the exact B-spline refinement matrix, and rebuild D0 and D1. Reject a refinement if ||D1D0|| exceeds 1e-8. The mathematical quantities are the pole tying, basis dimensions, prolongation, and exact derivative matrices; network weights and residual thresholds are estimated empirically. First test on a 2D manufactured Poisson problem and a divergence-free vector-field prediction task using a 6-layer CNN or FNO baseline on a 64x64 disk mask. Compare fixed Cartesian, uniform polar, and adaptive polar models at equal parameter count. Success means lower center-region error and fewer active coefficients at the same L2 error, with exact zero divergence for fields produced as D0 phi or exact zero div-curl residual for fields passed through D1.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.