Source-linked AI summary
A Mixed Elliptic--Hyperbolic Monge--Amp\`ere Problem: Modeling and a Least-Squares Solver
G. V. Carvalho, M. J. H. Anthonissen, J. H. M. ten Thije Boonkkamp, L. Kusch, K. Mitra, W. L. IJzerman
TL;DR
Smooth periodic reflector arrays motivate surfaces with both convex/concave and saddle-shaped regions, beyond the purely elliptic or hyperbolic geometries described. The paper partitions the source domain into regions governed by coupled elliptic and hyperbolic Monge–Amp`ere equations and develops a least-squares approximation. Its examples demonstrate smooth reflector configurations, including periodic solutions.
Problem
Smooth periodic arrays require combining different curvature signatures, while classical elliptic or hyperbolic formulations lead to globally convex/concave or saddle-shaped geometries.
Source: §1 · p. 2Method
The paper partitions the source into elliptic and hyperbolic subregions, couples their target-density contributions, and develops a numerical scheme for the mixed system.
Source: §1 · pp. 2–3Results
The proposed approach generates smooth reflector configurations, including periodic solutions.
Source: §1 · p. 3Takeaways & Limitations
The mixed formulation provides a framework for smooth periodic reflectors combining locally convex or concave and saddle-shaped regions.
Source: §1 · pp. 2–3The system is underdetermined and may admit multiple solutions, while existence and regularity remain unknown and the current formulation is restricted to rectangular source and target domains.
Source: §1 · p. 3, §6 · p. 24
Abstract
from arXiv · showhide
We consider a transport problem motivated by freeform optical design of a parallel-to-far-field reflector. Classical formulations rely on either elliptic or hyperbolic Monge--Amp\`ere equations, leading to globally convex/concave or saddle-shape reflector geometries, respectively. However, the design of smooth reflector arrays may require optical surfaces combining both types of curvature. To address this, we introduce a mixed elliptic--hyperbolic Monge--Amp\`ere system obtained by partitioning the source domain into one elliptic and one hyperbolic subregion, and we develop a least-squares method for its numerical approximation. Numerical examples demonstrate the ability of the proposed framework to compute both isolated and periodic mixed-type transport solutions, yielding smooth reflector configurations.
1 Introduction
The paper addresses freeform parallel-to-far-field reflector design, where smooth periodic arrays require surfaces combining different curvature signatures. It introduces a mixed transport formulation and numerical method, with the stated aim of producing smooth reflector configurations.
1 Introduction
- The reflector design problem seeks a height function u whose mapping redirects source light onto target directions while inducing the desired target intensity.The mapping pushes the source density to the target density, and the associated potential allows recovery of the reflector function.Source: §1 · pp. 1–2
- Smooth periodic arrays motivate combining locally convex or concave regions with saddle-shaped regions, since replicated convex or concave elements can have slope discontinuities at their interfaces.The paper notes that such cusps may generate undesirable optical artifacts if not manufactured perfectly sharp.Source: §1 · p. 2
- The formulation splits the source into convex and saddle-shaped subregions, coupling elliptic and hyperbolic Monge–Amp`ere equations and their separate target-density contributions.The contributions are coupled to match the full target density.Source: §1 · pp. 2–3
- The mixed system is underdetermined and may admit multiple solutions; the paper focuses on its formulation and numerical approximation rather than existence or regularity analysis.The paper also presents smooth reflector configurations, including periodic solutions.Source: §1 · p. 3
- Mixed Monge–Amp`ere formulations for mass-transport problems had not previously been considered to the authors’ knowledge.The passage contrasts this gap with mixed Monge–Amp`ere studies for isometric embedding.Source: §1 · p. 3
2 Properties of mapping solutions
The section establishes mapping properties and boundary constraints used to characterize transport solutions on rectangular domains. It relates determinant sign to boundary orientation and segmentation, and identifies the admissible elliptic segmentations.
2 Properties of mapping solutions
- Under the stated hypotheses, a mapping satisfying the system’s Jacobian and transport conditions is a diffeomorphism from X to Y.The section assumes axis-aligned rectangular domains and densities bounded away from zero and infinity with equal mass.Source: §2 · pp. 3–4
- For mappings satisfying the specified local condition, the global transport constraint can be imposed by requiring m(∂X) = ∂Y.The authors identify this boundary characterization as useful for the later numerical treatment.Source: §2 · p. 4
- The sign of det Dm determines whether a positively oriented boundary retains or reverses its orientation under the mapping.The paper associates positive orientation with the elliptic case and reversed orientation with the hyperbolic case.Source: §2 · pp. 3, 5
- Boundary segment assignments are cyclical and determined by the image of one segment together with the sign of det Dm.For the hyperbolic system, clockwise travel along ∂X becomes counter-clockwise along ∂Y.Source: §2 · p. 6
- Elliptic solutions exclude segmentations that map horizontal boundary segments to vertical ones, leaving the listed admissible choices and yielding convex potentials.The excluded mapping would contradict the positive Jacobian condition; the resulting Jacobian matrix is symmetric positive definite in the domain.Source: §2 · pp. 6–7
3 Mixed elliptic-hyperbolic Monge–Amp`ere system
The transport formulation splits the source into a convex left region and a saddle-like right region, coupling elliptic and hyperbolic Monge–Ampère systems while preserving differentiability. A periodicity constraint extends the mapping and potential smoothly across repeated cells.
3 Mixed elliptic-hyperbolic Monge–Amp`ere system
- The rectangular-domain transport formulation admits solutions m = grad u with u convex on the left subdomain and saddle-like on the remainder, while preserving differentiability throughout the domain.The subdomains are X L = (−a, 0) × (−b, b) and X R = (0, a) × (−b, b).Source: §3 · p. 7
- The coupling enforces continuity across the interface and requires the transported densities from both subdomains to sum to the target density.Each subdomain maps to the target, and gL + gR = g.Source: §3 · p. 9
- The backward formulation removes the unknown densities gL and gR, but leaves four scalar inverse-map fields governed by only three first-order scalar PDEs.The paper identifies this imbalance as indicating possible non-uniqueness.Source: §3 · p. 9
- For the stated uniform-density example, each admissible choice of α generates a different solution of the mixed system.The construction sets β(y1) = α(y1) − y1 + 1, with α satisfying the stated endpoint and derivative constraints.Source: §3 · p. 10
3.1 Additional constraint: periodicity of the mapping
- Imposing periodicity in x1 lets the mapping and potential extend smoothly across repeated copies of the source domain.The resulting reflector is C1 across adjacent-cell interfaces, and the target-density distribution is preserved while its density scales by N + 1.Source: §3.1 · pp. 10–11
- The periodic construction repeats the source density across N additional cells, and the extended map transports it to a scalar multiple of the target density.The repeated cells carry the same density; the resulting map preserves the relative spatial distribution while scaling total mass.Source: §3.1 · pp. 10–11
4 Least-squares method for the mixed MA system
This section presents an iterative least-squares scheme for the mixed Monge–Ampère system and recovery of a potential from the computed mappings. It also describes residual-based stopping criteria motivated by non-uniqueness, discretization, and the numerical procedures used in the experiments.
4.1 Method description
- Each iteration updates densities, symmetric matrix fields, and mappings through three least-squares projections that enforce density, determinant, transport-boundary, and interface-continuity constraints.The matrix projection has a characterization covering both elliptic and hyperbolic cases.Source: §4 · p. 11, §4.1 · p. 12
- The parameter ε bounds the determinants away from zero and reduced degenerate mappings in experiments, while the matrix projection is uniquely defined for almost every eigenvalue pair.The matrix minimization is reduced to a one-dimensional problem whose stationary points can be computed analytically.Source: §4.1 · pp. 12, 15
4.2 Recovery of the potential function
- Potential recovery assumes the computed mappings are invertible; ε reduced noninvertibility, but global invertibility is not theoretically guaranteed.The method inverts the two mappings, forms a piecewise forward map, and projects it onto conservative vector fields to approximate the transport potential.Source: §4 · p. 11, §4.2 · p. 18
- Stationarity of the potential-recovery functional yields a boundary condition, while interface terms cancel using continuity of m across ΓX.For the periodic case, a similar interface condition can be added to allow smooth extension.Source: §4.2 · p. 18
4.3 Stopping criteria
- Because non-uniqueness makes convergence to a particular discretized or exact solution unreasonable to expect, the method monitors residuals and stops when their relative changes stabilize.The residuals measure the mixed-system equations, while a third function measures pointwise change in the mappings; their L1 norms are used for global estimates.Source: §4.3 · pp. 18–19
4.4 Numerical implementation
- The target domain is discretized on a uniform N1 × N2 grid, with affine initial mappings satisfying the Dirichlet conditions; other initial mappings can yield different approximations.The authors relate this dependence on initialization to the underdetermined system.Source: §4.4 · p. 19
- The implementation uses numerical projection and recovery steps, finite differences, interpolation, and trapezoidal estimates.It computes the gL-projection zero by bisection after trapezoidal integration, solves the quartic P*-projection pointwise via Ferrari’s method [13, p. 32], and uses second-order central finite differences for recovery; sparse LU decompositions solve the resulting linear systems.Source: §4.4 · p. 19
5 Numerical results
The numerical examples show second-order residual convergence for constant densities and first-order convergence when the target density is discontinuous. The computed potentials have the intended mixed curvature, and the periodic example extends smoothly across replicated domains.
5.1 Example 1: Constant source density
- τµ can keep evolving after τR and τ× stabilize, so the stopping rule uses τR and τ× rather than requiring all three indicators to converge.In the 50 × 50 run, τµ reverses its trend and converges toward zero only after iteration 21,277, coinciding with ε-clamping and the end of drift along approximate solutions.Source: §5.1 · pp. 20–21
- The results show second-order convergence of both residuals in both cases: without the periodicity condition and with it enforced in the latter case, as discussed in Remark 3.1.The convergence study used uniform N × N grids with N ∈ {32, 64, 128, 256, 512} and relative tolerances δR, δ× = 10−5.Source: §5.1 · p. 21
- The potential is convex on X_L and saddle-shaped on X_R, while its periodic extension is smooth across replicated copies of X.The potential was computed on a 512 × 512 grid; the smooth extension is shown for the periodicity-condition run.Source: §5.1 · p. 21
- Monte Carlo sampling shows a small overall discrepancy from the bin-averaged target density, with localized error near ΓY linked to the interface derivative mismatch.The comparison uses 100 × 100 bins and Ns = 10^9 samples.Source: §5.1 · pp. 21, 23
5.2 Example 2: Discontinuous target density
- For the discontinuous target, mappings remain continuous but lose regularity near the density jump; the periodic condition aligns the second components along the specified boundary segment.The mappings are shown for a 512 × 512 discretization, with and without the periodicity condition.Source: §5.2 · p. 23
- First-order convergence of τR and τ× occurs for the discontinuous target density, consistent with mapping nonsmoothness near the density jump.The finite-difference estimates are second order, but errors near the jump contribute O(h) to the global residual.Source: §5.2 · p. 23
- In the discontinuous-target example, the periodic extension is slanted because δu is nonzero, and the authors could not control this parameter.Source: §5.2 · pp. 23–24
6 Conclusion
The paper introduces a mixed elliptic–hyperbolic Monge–Ampère transport formulation, with Dirichlet boundary conditions and a least-squares solver. Numerical examples compute isolated and periodic mixed-type solutions for smooth reflector configurations, while the current formulation remains limited to rectangular domains and a two-region partition.
6 Conclusion
- The least-squares method computes isolated and periodic mixed-type solutions that induce smooth reflector configurations relevant to optical array design.The formulation couples elliptic and hyperbolic Monge–Ampère equations through a decomposition of the source domain.Source: §6 · p. 24
- Dirichlet boundary conditions derived from boundary orientation and segmentation help enforce mapping of the target boundary to the boundaries of both source subdomains.The system couples elliptic and hyperbolic Monge–Ampère equations through the source-domain decomposition.Source: §6 · p. 24
- The current formulation is restricted to rectangular source and target domains and partitions the source into two regions, while existence and regularity remain unknown.The authors identify more general geometries and multiple subdomains as directions for further work.Source: §6 · p. 24