Restricted Delaunay Triangulation

• Restricted Delaunay Triangulation is a variant of the Delaunay complex that limits simplices to those whose associated Voronoi cells intersect a given manifold or subspace.
• It employs adaptive sampling and refinement strategies to ensure mesh quality and topological integrity during manifold reconstruction.
• The method's effectiveness varies with dimensions and sampling density, making it crucial for applications in mesh generation and computational geometry.

A restricted Delaunay triangulation (RDT) is a subcomplex of the Delaunay complex formed by restricting constructions—typically Voronoi cells or Delaunay simplices—to a subset such as a manifold, surface, or lower-dimensional feature embedded in a Euclidean or more general metric space. RDTs underpin both theoretical studies of manifold reconstruction and practical algorithms for mesh generation on submanifolds, piecewise smooth domains, and geometric complexes. The theoretical properties, algorithmic frameworks, and limitations of RDTs vary sharply depending on ambient and intrinsic dimension, metric background, and the topological and geometric complexity of the underlying domain.

1. Core Definitions and Mathematical Foundations

Let $M \subset \mathbb{R}^N$ be a compact, smooth $m$-dimensional submanifold of positive reach $\tau$. For a finite sample $L \subset M$, the usual Delaunay triangulation $\mathrm{Del}(L)$ in the ambient space is defined as the nerve of the Voronoi diagram: a simplex $\sigma \subset L$ belongs to $\mathrm{Del}(L)$ if the corresponding Voronoi cells have nonempty intersection. The restricted Delaunay complex $\mathrm{Del}|_M(L)$ is formed by those simplices of $\mathrm{Del}(L)$ whose dual Voronoi cells intersect $M$:

$$\mathrm{Del}|_M(L) = \{\, \sigma \in \mathrm{Del}(L) : \mathrm{Vor}(\sigma) \cap M \neq \emptyset \,\}$$

where $$\mathrm{Vor}(\sigma) = \bigcap_{p \in \sigma} \mathrm{Vor}_L(p)$$.

Under this construction, $\mathrm{Del}|_M(L)$ is the nerve of the cover $\{\, \mathrm{Vor}_L(p) \cap M \,\}_{p \in L}$. This definition generalizes naturally to restrictions on curve networks, surface patches, and volumetric regions in piecewise smooth complexes (Engwirda, 2016).

2. Sampling Theory and Topological Guarantees

For $M \subset \mathbb{R}^n$, let $L \subset M$ be an $\varepsilon$-sample, i.e., $$\forall x \in M,\, \exists p \in L : \|x - p\| \leq \varepsilon$$, and $L$ is $\varepsilon$-sparse if $\forall p \neq q \in L,\, \|p-q\| \geq \varepsilon$. Classical results (Amenta–Bern, Attali–Edelsbrunner, Guibas–Oudot) establish that if $M$ is a smooth curve in $\mathbb{R}^2$ or a smooth surface in $\mathbb{R}^3$ and $\varepsilon < 0.1\,\tau(M)$, then $\mathrm{Del}|_M(L)$ is homeomorphic to $M$. Moreover, for any sufficiently dense witness set $W \subset M$, the witness complex $\mathrm{Wit}(L, W)$ satisfies

$$\mathrm{Del}|_M(L) \subseteq \mathrm{Wit}(L, W), \quad \mathrm{Wit}(L, M) = \mathrm{Del}|_M(L)$$

For these dimensionalities, a uniform sample with sufficiently small $\varepsilon$ (relative to reach) ensures that the RDT captures the correct manifold topology (0803.1296).

3. Pathologies in Higher Dimensions and Metric Spaces

In ambient dimension $n \geq 4$ or for manifolds of intrinsic dimension $m \geq 3$, the analogy with surfaces breaks down. Explicit counterexamples demonstrate that for every $\mu < 1/3$, there exist compact closed hypersurfaces $M \subset \mathbb{R}^4$ and uniform $\varepsilon$-samples $L$ with $\varepsilon = \mu\tau$, such that:

• $\mathrm{Del}|_M(L)$ is not homeomorphic to $M$ (Theorem 4.1, (0803.1296)).
• Moreover, $\mathrm{Del}|_M(L)$ may fail even to be homotopy equivalent to $M$ (Theorem 4.2).
• The inclusion relations between RDT and witness complexes can fail: $\mathrm{Wit}(L,W) \not\subseteq \mathrm{Del}|_M(L)$, even for arbitrarily dense samples (Theorem 4.3).

The pathologies arise from the flexibility of Voronoi duals and the creation or destruction of "sliver" simplices through infinitesimal geometric perturbations—these phenomena cannot be eliminated merely by densifying the sample (0803.1296). In general Riemannian manifolds of $m > 2$, even generic, arbitrarily dense samples do not guarantee that the intrinsic Delaunay complex is a triangulation of $M$ (Boissonnat et al., 2016).

4. Algorithmic Construction and Refinement Strategies

Constructing an RDT for a smooth submanifold can be achieved using ambient Euclidean distances: the restricted Voronoi diagram is computed by intersecting Voronoi cells with $M$, and their nerve forms $\mathrm{Del}|_M$. The intrinsic Delaunay triangulation uses geodesic distances but, under certain conditions (specifically, $\delta$-genericity, sparsity, and thickness constraints), the restricted Delaunay complex coincides with both intrinsic and tangential complexes. The tangential complex $\mathrm{Del}^T(P)$ (Boissonnat–Ghosh) projects local neighborhoods into tangent spaces and applies a weighted Delaunay construction there (Boissonnat et al., 2013).

A classical Delaunay-refinement paradigm, adapted to manifold or piecewise-smooth input, iteratively augments the sample to eliminate "bad" simplices (those violating quality or topological criteria) via either circumcenter insertion or "off-centre" (size-optimal) Steiner placements. In piecewise smooth complexes, RDT subcomplexes are defined for curves, surfaces, and volumes; topological constraints (e.g., manifoldhood) are maintained through dynamic predicates during refinement (Engwirda, 2016).

5. Mesh Quality, Topology Enforcement, and Practical Performance

Quality metrics for RDT-based meshes include radius–edge ratio, area–length and volume–length ratios, and dihedral angle bounds. The Frontal–Delaunay refinement scheme (Engwirda, 2016):

• Ensures shape bounds (e.g., $\rho(f) \leq \bar{\rho}_f$), mesh-size conformance $h(e) \leq (4/3) \bar{h}(x_e)$, and surface discretization error $$\varepsilon_1(e), \varepsilon_2(f) \leq \bar{\varepsilon}(x_e)$$.
• Enforces topological integrity: curve and surface restricted complexes are required to be 1-manifold and 2-manifold, respectively, including at non-manifold junctions.
• Employs sliver suppression based on volume–length criteria, off-centre refinement for high-quality placement along features, and protection strategies for sharp-angle curve segments.

Empirical comparisons with classical (weighted) Delaunay-refinement implementations demonstrate that Frontal–Delaunay strategies produce higher mean element quality, tighter edge-length distributions, and fewer mesh elements, while preserving or improving topological correctness (Engwirda, 2016).

Metric	JGSW-FD (Frontal–Delaunay)	CGAL-DR (Weighted)
Mean area–length	≈0.97–0.98	≈0.93–0.94
Min angle (deg)	≃23.5°	<20°
Std edge–length	0.05–0.07	0.15–0.20

6. Limitations, Open Problems, and Remedial Constructions

In codimension >1, higher intrinsic dimension ($m>2$), or arbitrary Riemannian metric, density-based sampling alone cannot guarantee that the restricted Delaunay or intrinsic Delaunay complex triangulates the manifold (0803.1296, Boissonnat et al., 2016). The breakdown is intrinsic and cannot be circumvented by naive refinement or generic perturbation.

Possible remedies include:

• Enforcing thickness and "power-protection" (i.e., lower bounds on the distance from simplex circumsphere centers to non-vertex sample points), as in the tangential Delaunay complex approach (Boissonnat et al., 2013).
• Augmenting sampling with sliver exudation (applying weights or removing nearly degenerate simplices).
• Explicit protection strategies and dynamic topological predicates for non-manifold features (Engwirda, 2016).

Outstanding challenges include characterizing the class of (Riemannian) metrics for which density plus separation conditions are sufficient, developing local feature size invariants stronger than the classical reach, and extending current algorithms to more general implicit or higher-order geometric domains.

7. Applications and Broader Impacts

Restricted Delaunay triangulation is fundamental in geometric modeling, meshing for numerical simulation (finite element, finite volume), and topological data analysis. RDT-based algorithms support high-quality mesh generation for piecewise-smooth domains and complex CAD objects, facilitating provable guarantees on mesh topology and element regularity (Engwirda, 2016). Their theoretical limitations in higher dimensions and on general metric backgrounds remain an active area of research, with significant implications for computational geometry, numerical PDEs, and geometric inference.