{
  "id": "delaunay-triangulation",
  "title": "Delaunay Triangulation",
  "lang": "en",
  "source_lang": "en",
  "summary": "A Delaunay triangulation connects points into simplices whose circumspheres contain no other input points, linking proximity geometry, efficient algorithms, and numerical mesh generation.",
  "category": "mathematics",
  "subcategory": "geometry",
  "type": "concept",
  "aliases": [
    "Delaunay tessellation",
    "Delone triangulation"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/delaunay-triangulation/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/delaunay-triangulation/",
      "json": "https://aiwiki.page/zh/delaunay-triangulation.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A Delaunay triangulation is a geometric subdivision that joins a set of points into triangles, or into higher-dimensional simplices, subject to an empty-circumsphere condition. In the plane, the open disk bounded by the circumcircle of every triangle contains no other input point. For points in general position, this condition determines a unique triangulation of their convex hull. Degenerate configurations can instead produce nonsimplicial Delaunay cells, which admit multiple valid triangulations.\n\nNamed after the mathematician Boris Delaunay, whose surname is also transliterated as Delone, the construction is fundamental to computational geometry. It connects nearest-neighbor relationships with convexity, provides efficient spatial data structures, and supplies meshes for scientific computation. Its planar form has an important angle-optimality property, but that property does not extend directly to three-dimensional tetrahedral meshes. The distinction between an unconstrained point-set triangulation and a boundary-conforming computational mesh is equally important in applications."
    },
    {
      "heading": "Definition, existence, and degeneracies",
      "text": "Let \\(P=\\{p_1,\\ldots,p_n\\}\\) be a finite set of distinct points in the Euclidean plane, not all collinear. A straight-line triangulation of \\(P\\) partitions its convex hull into nonoverlapping triangles whose vertices are precisely the input points. Triangles meet only along shared edges or vertices, every input point is used, and no edge passes through another input point without being subdivided there. A triangle’s circumcircle is the unique circle passing through its three vertices. The triangulation is Delaunay when every triangle has an empty open circumdisk: other input points may lie on its circumcircle, but not strictly inside it.\n\nThe word “open” is essential. Consider four points forming a square. They lie on one circle, and either diagonal divides the square into two triangles satisfying the empty-circle condition. Thus the Delaunay triangulation need not be unique. In the plane, a sufficient general-position assumption is that no three points are collinear and no four are cocircular. Under these assumptions, the Delaunay triangulation exists and is unique. The no-three-collinear condition simplifies the combinatorics; the exclusion of empty cocircular configurations is what removes ambiguity between alternative diagonals.\n\nWithout a general-position assumption, it is useful to distinguish a Delaunay subdivision from a chosen triangulation of that subdivision. An empty circle through four or more input points can support a convex polygonal cell rather than a single triangle. Adding noncrossing diagonals inside each such cell produces a Delaunay triangulation, but those diagonals are not uniquely determined by the geometry. Collinear point sets are intrinsically one-dimensional and do not admit a two-dimensional triangulation with positive-area faces. Duplicate coordinates likewise represent a single geometric site unless an application gives them additional, nongeometric meaning.\n\nFor a planar triangulation with \\(n\\) vertices and \\(h\\) vertices on the convex-hull boundary, including any points lying along boundary edges, the numbers of bounded triangular faces and edges are\n\n\\[t=2n-h-2,\\qquad e=3n-h-3.\\]\n\nThese identities follow from the planar Euler characteristic and from counting triangle-edge incidences. They show that planar Delaunay triangulations require only linear output space. Boundary edges belong to one triangle; interior edges belong to two.\n\nIn \\(\\mathbb{R}^{d}\\), the same definition uses \\(d\\)-dimensional simplices and empty circumspheres. A simplex has \\(d+1\\) affinely independent vertices: a triangle in two dimensions and a tetrahedron in three. Full-dimensional point sets admit Delaunay triangulations of their convex hull, while lower-dimensional sets are treated within their affine hull. Uniqueness generally requires excluding relevant configurations of \\(d+2\\) cospherical points. A planar triangulation is therefore the simplest nontrivial instance of a construction defined throughout Euclidean space."
    },
    {
      "heading": "Voronoi duality and the lifting construction",
      "text": "The closest companion of a Delaunay triangulation is the Voronoi diagram. For an input site \\(p\\), its Voronoi cell is\n\n\\[V(p)=\\{x:\\|x-p\\|\\leq \\|x-q\\|\\text{ for every }q\\in P\\}.\\]\n\nIt consists of locations at least as close to \\(p\\) as to any other site, using ordinary Euclidean distance. Each comparison defines a half-space bounded by a perpendicular bisector, so every Voronoi cell is convex. The cells cover space and intersect along shared faces. Sites on the convex-hull boundary have unbounded Voronoi cells; sites strictly inside the hull have bounded cells.\n\nFor a nondegenerate planar configuration, two sites are connected by a Delaunay edge exactly when their Voronoi cells share an edge. Three sites form a Delaunay triangle exactly when their Voronoi cells meet at a Voronoi vertex. That vertex is the triangle’s circumcenter: it is equidistant from the three sites, and no other site is closer. The empty-circle condition follows immediately. Corresponding Delaunay and Voronoi edges are perpendicular, although their supporting lines need not intersect within both finite edges. An obtuse triangle’s circumcenter lies outside the triangle, so the duality is combinatorial rather than a claim that each dual feature lies inside its primal counterpart.\n\nDegeneracies require a more careful statement. At the center of a square, four Voronoi cells meet. The corresponding Delaunay cell is a quadrilateral, not a uniquely selected pair of triangles. Choosing a diagonal adds a simplicial refinement that does not correspond to a positive-length Voronoi edge. Thus the Voronoi diagram naturally determines the Delaunay subdivision, while a triangulation of a degenerate cell requires an additional choice.\n\nA second interpretation converts the empty-circle condition into convex geometry. Lift each planar point \\(p=(x,y)\\) onto a paraboloid in three-dimensional space:\n\n\\[\\widehat p=(x,y,x^2+y^2).\\]\n\nCompute the convex hull of the lifted points and project its lower faces vertically back to the plane. Their projections form the Delaunay subdivision. To see why, write a circle with center \\((a,b)\\) and radius \\(r\\) as\n\n\\[(x-a)^2+(y-b)^2=r^2.\\]\n\nOn the paraboloid, its points lie in the plane\n\n\\[z=2ax+2by+r^2-a^2-b^2.\\]\n\nA planar point inside the circle lifts below this plane. Consequently, a lower supporting face—one with all other lifted sites on or above its plane—projects to a cell with an empty circumcircle.\n\nThe construction extends by lifting \\(x\\in\\mathbb{R}^{d}\\) to \\((x,\\|x\\|^2)\\in\\mathbb{R}^{d+1}\\). It establishes a direct connection between Delaunay computation and convex-hull algorithms, explains existence, and turns cocircular or cospherical degeneracies into coplanar lower faces. It also provides the algebraic foundation for weighted Delaunay constructions."
    },
    {
      "heading": "Local criteria, angle optimality, and proximity graphs",
      "text": "Although the definition is global, a planar Delaunay triangulation can be recognized using a local test. Suppose two adjacent triangles share an interior edge \\(ab\\), with opposite vertices \\(c\\) and \\(d\\). The edge is locally Delaunay if \\(d\\) does not lie inside the circumcircle of \\(abc\\). When the two triangles form a convex quadrilateral, the equivalent angular condition is\n\n\\[\\angle acb+\\angle adb\\leq \\pi.\\]\n\nEquality occurs when the four vertices are cocircular. If the sum exceeds \\(\\pi\\), replacing diagonal \\(ab\\) by \\(cd\\) is an edge flip that corrects this local violation. A planar triangulation is Delaunay exactly when all its interior edges are locally Delaunay.\n\nThis criterion underlies the celebrated max–min angle property. Among all straight-line triangulations of a fixed planar point set, a Delaunay triangulation maximizes the smallest triangle angle. More strongly, if all triangle angles are sorted in increasing order, its angle sequence is lexicographically maximal. Flipping a non-Delaunay edge improves the local angle ordering; repeated improvements lead to a triangulation satisfying the empty-circle condition. Cocircular alternatives may have the same optimal angle sequence.\n\nThe property is relative to the available vertices, not an absolute guarantee of well-shaped triangles. If two points are extremely close and neighboring points are far away, any triangulation can contain narrow triangles. An acute angle forced by an input boundary likewise cannot disappear merely through diagonal changes. Delaunay triangulation therefore avoids unnecessarily small planar angles without ensuring that all angles exceed a fixed threshold. Nor does it generally minimize total edge length or make every triangle acute.\n\nIts edges also capture important proximity relationships. The Euclidean minimum spanning tree, which connects all sites with minimum total edge length, can be found within the Delaunay structure. Every edge of that tree has an empty closed disk with the edge as diameter, so it is necessarily a Delaunay edge. This permits an initially dense geometric graph problem to be reduced to a graph with only \\(O(n)\\) planar edges. The Gabriel graph uses a related diameter-disk emptiness rule and is another subgraph of the Delaunay graph, with precise inclusions depending on how boundary ties are defined.\n\nThe planar Delaunay graph is also a geometric spanner. There is a universal constant \\(C\\) such that any two input vertices \\(p\\) and \\(q\\) can be joined along Delaunay edges by a path of total length at most \\(C\\|p-q\\|\\). This is a bound on worst-case detour, not a statement that an edge joins every nearby pair or that graph shortest paths equal straight-line distances.\n\nThese relationships place Delaunay triangulations between geometry and graph theory. They preserve useful local connections in a sparse planar structure, while their triangles provide a continuous spatial subdivision. The angle theorem, however, is specifically planar: empty circumspheres in three dimensions do not yield an analogous guarantee against poorly shaped tetrahedra."
    },
    {
      "heading": "Construction algorithms and complexity",
      "text": "Several algorithmic families construct the same geometric object through different intermediate representations. Their efficiency depends both on geometric operations and on locating the region affected by each operation. For \\(n\\) planar sites, optimal general-purpose algorithms run in \\(O(n\\log n)\\) time in standard comparison-based settings, with \\(O(n)\\) output storage. Here asymptotic notation suppresses constant factors and lower-order terms. The sorting lower bound can be related to Delaunay construction by placing input values along a suitable curve and recovering their order from the resulting adjacency structure.\n\nOne approach begins with any valid triangulation and repeatedly flips non-Delaunay edges. Charles Lawson developed influential local-transformation methods in the 1970s. Each flip changes only two adjacent triangles, making the method conceptually simple and useful for mesh repair. A naive flipping sequence can take quadratically many operations in the plane, however, so local simplicity alone does not establish an optimal running time. Maintaining a queue of potentially illegal edges avoids repeatedly testing unaffected regions.\n\nIncremental algorithms insert points one at a time. The Bowyer–Watson method, published independently by Adrian Bowyer and David Watson in 1981, removes triangles whose circumcircles contain the new point. Their union forms a cavity; connecting its boundary edges to the new point restores a Delaunay triangulation. Another incremental formulation first splits the containing triangle, or the adjacent triangles when insertion lies on an edge, and then legalizes nearby edges by flipping. Both require special handling of hull changes, boundary cases, and exact degeneracies.\n\nRandom insertion order gives strong expected bounds. With an appropriate point-location structure, randomized incremental construction achieves expected \\(O(n\\log n)\\) time and linear final space in two dimensions. The expectation concerns the randomized algorithm, not a requirement that the input points themselves follow a random distribution. Conflict graphs or a history of replaced triangles can help locate subsequent insertion regions. A simple walk from a nearby previously located triangle is often effective in practice, but its performance depends on the input and query order.\n\nDivide-and-conquer algorithms split the sites, triangulate the subsets, and merge them while enforcing the empty-circle condition. Michael Shamos and Dan Hoey presented an optimal planar approach in 1975. Leonidas Guibas and Jorge Stolfi’s 1985 quad-edge framework supplied a compact way to represent and manipulate both a planar subdivision and its dual. Other methods compute the Voronoi diagram first and then obtain the triangulation by duality; Steven Fortune’s 1987 sweep-line algorithm constructs a planar Voronoi diagram in \\(O(n\\log n)\\) time.\n\nLifting provides another route: compute a convex hull one dimension higher and retain its lower faces. Dynamic structures support insertion, deletion, and moving vertices, while parallel implementations distribute point processing or partition the domain. The best practical choice depends on dimension, input ordering, degeneracy frequency, memory access, and whether later operations require the full Voronoi dual. In higher dimensions, the output itself may be superlinear, so planar complexity bounds cannot simply be carried over."
    },
    {
      "heading": "Geometric predicates and numerical robustness",
      "text": "A Delaunay implementation makes discrete decisions from numerical coordinates: whether three points turn left or right, whether a fourth point lies inside a circumcircle, and which faces should be adjacent. A single incorrect sign can produce a crossed edge, an inverted triangle, or a corrupted adjacency structure. Robustness therefore depends less on computing visually plausible circumcenters than on evaluating a small set of geometric predicates consistently.\n\nFor planar points \\(a=(a_x,a_y)\\), \\(b=(b_x,b_y)\\), and \\(c=(c_x,c_y)\\), orientation is determined by\n\n\\[\\operatorname{orient}(a,b,c)= \\det \\begin{pmatrix} a_x&a_y&1\\\\ b_x&b_y&1\\\\ c_x&c_y&1 \\end{pmatrix}.\\]\n\nA positive value indicates counterclockwise order, a negative value clockwise order, and zero collinearity. The use of a determinant makes the test polynomial in the coordinates and avoids dividing by a potentially small quantity. It also supplies a signed area, apart from a factor of two.\n\nFor counterclockwise \\(a,b,c\\), the in-circle predicate is the sign of\n\n\\[\\operatorname{incircle}(a,b,c,d)= \\det \\begin{pmatrix} a_x&a_y&a_x^2+a_y^2&1\\\\ b_x&b_y&b_x^2+b_y^2&1\\\\ c_x&c_y&c_x^2+c_y^2&1\\\\ d_x&d_y&d_x^2+d_y^2&1 \\end{pmatrix}.\\]\n\nA positive value places \\(d\\) inside the circumcircle, a negative value outside, and zero on the circle. Reversing the orientation of \\(a,b,c\\) reverses this sign convention. The determinant is closely related to the orientation of lifted points, linking practical predicates to the paraboloid construction.\n\nOrdinary floating-point arithmetic can misclassify nearly collinear or nearly cocircular configurations because cancellation leaves a small result with an unreliable sign. Comparing every result with a fixed numerical tolerance is not a general solution: the relevant scale varies, and separately applied tolerances can produce logically inconsistent decisions. Explicitly calculating a circumcenter and comparing distances can be especially problematic for nearly collinear triangles, whose circumcenters may lie very far away.\n\nAdaptive exact predicates first evaluate a fast approximation and an error bound. If the sign is certified, computation stops; otherwise, additional precision is used until the exact sign for the represented coordinates is established. Jonathan Shewchuk’s work in the 1990s made floating-point expansion arithmetic and adaptive orientation and in-circle tests widely used. Exact arithmetic on rational or integer coordinates provides another approach, although intermediate values and computational costs can grow.\n\nExact zeros still need explicit treatment. Symbolic perturbation, including the simulation-of-simplicity framework introduced by Herbert Edelsbrunner and Ernst Mücke in 1990, interprets degeneracies through a consistent infinitesimal ordering. Alternatively, an implementation may preserve nonsimplicial cells or select diagonals deterministically. These strategies resolve combinatorial ambiguity; they do not remove uncertainty in measured coordinates. Exact predicates certify decisions about stored input values, not about unknown physical locations. Geometric constructions such as new circumcenters may require separate accuracy guarantees even when all predicates are exact."
    },
    {
      "heading": "Boundary constraints and Delaunay mesh refinement",
      "text": "An unconstrained Delaunay triangulation fills the convex hull of its sites. It does not automatically preserve coastlines, polygonal boundaries, holes, material interfaces, or internal fractures. A segment representing such a feature may cross unconstrained Delaunay edges, and concave parts of a domain may be filled by unwanted triangles. Computational meshes therefore often require a constrained Delaunay triangulation built from a planar straight-line graph: vertices and noncrossing segments that encode required geometry.\n\nConstraint edges must appear in the triangulation, even when they would violate the ordinary empty-circle rule. The remaining triangles satisfy a visibility-qualified condition. Informally, a triangle’s circumdisk may contain an input vertex if intervening constraint segments block that vertex’s visibility to the triangle. Equivalently, in a valid constrained triangulation, each flippable unconstrained interior edge is locally Delaunay. Constraint edges are exempt from flipping, because replacing them could destroy the prescribed domain geometry.\n\nA conforming Delaunay triangulation is different. It adds extra vertices, commonly called Steiner points, so that the required segments become chains of edges in an ordinary Delaunay triangulation. A constrained triangulation can often use only the original vertices; a conforming construction may need many additional ones. The distinction matters when an application requires the strict Euclidean Voronoi–Delaunay relationship rather than only boundary preservation.\n\nMesh generation usually imposes size and shape requirements as well. Delaunay refinement repeatedly inserts points associated with poor triangles or inadequately resolved boundaries, then updates the triangulation. A common triangle-quality measure is the circumradius-to-shortest-edge ratio,\n\n\\[\\rho=\\frac{R}{\\ell_{\\min}}.\\]\n\nBy the extended sine rule,\n\n\\[\\ell_{\\min}=2R\\sin\\theta_{\\min}, \\qquad \\rho=\\frac{1}{2\\sin\\theta_{\\min}},\\]\n\nwhere \\(\\theta_{\\min}\\) is the smallest angle. Bounding \\(\\rho\\) therefore bounds the minimum angle away from zero. A size field can separately prescribe smaller elements near interfaces, curved boundaries, or regions of rapid solution variation.\n\nAlgorithms associated with L. Paul Chew and Jim Ruppert established influential termination and quality guarantees. A typical refinement process splits an encroached boundary segment before inserting the circumcenter of a bad triangle; this priority helps preserve boundary conformity and controls interactions between refinement steps. For appropriate planar inputs without small input angles, classical analyses guarantee termination for minimum-angle targets around \\(20.7^\\circ\\), with precise thresholds depending on the algorithm and hypotheses.\n\nSmall boundary angles complicate such guarantees because the geometry itself forces small mesh angles. Curved boundaries also require geometric approximation or curved elements rather than merely more triangles. In three dimensions, circumsphere-based refinement controls some size measures but does not eliminate sliver tetrahedra. Practical mesh generators therefore combine Delaunay operations with segment protection, geometric approximation, smoothing, edge or face transformations, and explicit quality improvement. These are additional algorithms, not consequences of the empty-circle definition alone."
    },
    {
      "heading": "Higher dimensions, weights, and geometric generalizations",
      "text": "In three dimensions, a Delaunay triangulation partitions a point set’s convex hull into tetrahedra with empty open circumspheres. A generic tetrahedron corresponds to a vertex of the three-dimensional Voronoi diagram; a triangular face corresponds to a Voronoi edge; and a Delaunay edge corresponds to a Voronoi polygonal face. The familiar planar duality thus persists, but the output size and element-quality behavior change substantially.\n\nThe worst-case complexity in fixed dimension \\(d\\) is\n\n\\[O\\!\\left(n^{\\lceil d/2\\rceil}\\right),\\]\n\nand this order is attained by suitable point configurations. In particular, three-dimensional Delaunay triangulations can have quadratic size. Real engineering datasets often produce much smaller meshes, but no linear worst-case bound follows from spatial dimension three. Algorithms must account for both input size and the number of generated simplices.\n\nA distinctive three-dimensional obstacle is the sliver: a tetrahedron whose vertices lie nearly in one plane, often near a circle, while its edge lengths remain comparable. Such a tetrahedron can have extremely small volume and very small or very large dihedral angles without a large circumradius-to-shortest-edge ratio. Consequently, good edge-length and circumradius measures do not ensure a well-conditioned tetrahedron. This limits direct transfer of planar angle arguments to finite-element meshing.\n\nWeighted sites lead to regular triangulations, also called weighted Delaunay triangulations. Assign a scalar weight \\(w_p\\) to each point and replace squared distance by power distance,\n\n\\[\\pi_p(x)=\\|x-p\\|^2-w_p.\\]\n\nThe regions where one site minimizes this expression form a power diagram. Lifting a site to\n\n\\[\\widehat p=(p,\\|p\\|^2-w_p)\\]\n\nand projecting the lower convex hull gives the corresponding regular subdivision. Unlike the ordinary unweighted case, some weighted sites can be hidden: they have empty power cells and do not appear as vertices in the subdivision. Weight adjustment is useful in geometric modeling and in methods such as sliver exudation, which improves tetrahedral meshes by changing weights.\n\nOther generalizations modify the domain or the distance rule. Periodic Delaunay triangulations identify opposite boundaries of a fundamental cell and effectively triangulate a flat torus, avoiding artificial exterior boundaries in periodic simulations. Intrinsic Delaunay triangulations on piecewise-flat surfaces use surface geometry and local angle conditions rather than unrestricted ambient three-dimensional distances. Their edges may represent geodesic paths across an underlying surface mesh.\n\nA restricted Delaunay construction instead relates the ambient Voronoi diagram to a target surface or domain: simplices are selected through intersections of dual Voronoi features with that target. Such constructions support surface reconstruction and manifold meshing under appropriate sampling and regularity assumptions. Anisotropic variants incorporate direction-dependent measures, often expressed through a metric tensor, but require separate geometric and algorithmic analysis. The ordinary Euclidean empty-circle theorem does not automatically apply after replacing distance by an arbitrary spatially varying metric."
    },
    {
      "heading": "Interpolation, numerical models, and shape reconstruction",
      "text": "Delaunay triangulations provide a local framework for interpolation of scattered measurements. Suppose scalar values \\(f_i\\) are known at sites \\(p_i\\). Inside a triangle with vertices \\(p_1,p_2,p_3\\), barycentric coordinates express a point as\n\n\\[x=\\lambda_1p_1+\\lambda_2p_2+\\lambda_3p_3, \\qquad \\lambda_1+\\lambda_2+\\lambda_3=1, \\qquad \\lambda_i\\geq 0.\\]\n\nThe piecewise-linear interpolant is\n\n\\[f(x)=\\lambda_1f_1+\\lambda_2f_2+\\lambda_3f_3.\\]\n\nIt is continuous across shared edges because adjacent triangles agree on the edge endpoints and vary linearly along the edge. Its gradient is generally discontinuous across those edges. The triangulation supplies a geometrically meaningful partition, but interpolation accuracy still depends on sampling density, data noise, triangle shape, and the behavior of the underlying function.\n\nIn terrain modeling, a triangulated irregular network commonly applies planar Delaunay triangulation to horizontal coordinates and attaches measured elevations to the vertices. The resulting triangles represent a piecewise-planar surface. This is a two-dimensional triangulation supporting a height function, not the three-dimensional Delaunay tetrahedralization of the measured points. Ridges, stream channels, and other breaklines may need constrained edges; otherwise, the interpolated surface can smooth across important discontinuities in slope.\n\nFor numerical solutions of partial differential equations, triangular and tetrahedral meshes support finite-element basis functions and local integration. Delaunay geometry is particularly relevant to the cotangent discretization of the Laplace operator. For an interior edge whose opposite angles are \\(\\alpha\\) and \\(\\beta\\), a common weight is\n\n\\[w=\\frac12(\\cot\\alpha+\\cot\\beta).\\]\n\nFor nondegenerate incident triangles, the local Delaunay condition \\(\\alpha+\\beta\\leq\\pi\\) makes this weight nonnegative. Nonnegative weights are important in some discrete maximum-principle and diffusion formulations, although boundary treatment, material coefficients, and the complete discretization determine the final numerical properties. Boundary-edge weights require separate consideration.\n\nThe Voronoi dual also supports finite-volume methods. Orthogonality between a Delaunay edge and its dual Voronoi face simplifies certain two-point flux approximations for isotropic diffusion. Anisotropic coefficients, obtuse configurations, and boundary clipping can require more elaborate schemes. Neither orthogonality nor planar angle optimality alone guarantees accuracy or stability for every equation.\n\nFor point-cloud analysis, the Delaunay structure is a substrate for alpha complexes and alpha shapes. An alpha complex selects Delaunay simplices through radius-controlled intersections of balls with their Voronoi cells; it is not simply obtained by keeping every triangle below an arbitrary circumradius threshold without further rules. Varying the scale reveals connected components, tunnels, and cavities. This connects geometric reconstruction with algebraic topology and topological data analysis.\n\nScientific software reflects these varied uses. Triangle implements robust planar triangulation and refinement, CGAL provides extensive triangulation data structures and geometric algorithms, and Qhull computes Delaunay structures through convex-hull methods. Their conventions differ for degeneracies, omitted or merged points, boundary constraints, and numerical precision. A triangulation used for interpolation, topology, or simulation must therefore be interpreted together with the geometric assumptions and construction rules that produced it."
    }
  ],
  "links": [
    "convex-hull",
    "computational-geometry",
    "triangulation",
    "circle",
    "euler-characteristic",
    "simplex",
    "affine-space",
    "euclidean-space",
    "voronoi-diagram",
    "euclidean-distance",
    "geometry",
    "graph-theory",
    "big-o-notation",
    "randomization",
    "determinant",
    "floating-point-arithmetic",
    "finite-element-method",
    "metric-tensor",
    "interpolation",
    "partial-differential-equation",
    "laplace-operator",
    "algebraic-topology"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}