CGAL 5.2 - 2D Periodic Hyperbolic Triangulations
User Manual

Author
Iordan Iordanov and Monique Teillaud

This package enables the computation of Delaunay triangulations of the Bolza surface, which is the most symmetric surface of genus 2. The Bolza surface is a hyperbolic closed compact orientable surface.

A triangulation of the Bolza surface can be seen as a periodic triangulation of the hyperbolic plane.

The Bolza Surface

Let \(\mathbb{H}^2\) denote the hyperbolic plane, represented in the Poincaré disk model (see Chapter 2D Hyperbolic Delaunay Triangulations). The Bolza surface \(\mathcal{M}\) is defined as the quotient of \(\mathbb H^2\) under the action of a group \(\mathcal G\) that we will introduce now.

Consider the regular hyperbolic octagon \( \mathcal D_O \) centered at the origin with all angles equal to \( \pi/4\), as shown in Figure 43.1. Note that \(\mathcal D_O\) is unique up to rotation and cannot be scaled since this operation would change its angles. Consider the four hyperbolic translations \( a,b,c,d\) with their respective inverses \(\overline{a}, \overline{b}, \overline{c}, \overline{d}\) that identify the opposite sides of \( \mathcal D_O \). The axes of these translations, shown in Figure 43.1 - Left, are diameters of the Poincaré disk. The four translations \(a, b, c, d\) generate a (non-commutative) discrete group of orientation-preserving isometries, with finite presentation

\[ \mathcal{G} = \left< a,b,c,d \; \bigg| \; abcd\overline{a}\overline{b}\overline{c}\overline{d} \right>. \]

Note that equivalent translations of the group \(\mathcal G\) can be reduced to a unique minimal representative.

The group \(\mathcal G\) is acting on \(\mathbb H^2\): two points \(p\) and \(q\) in \(\mathbb H^2\) are in the same orbit under the action of \(\mathcal G\) if there exists an element \(g \in \mathcal G\) such that \(g(p) = q\), as shown in Figure 43.1 - Center. The Bolza surface \(\mathcal{M}\) is defined as the quotient of \(\mathbb H^2\) under the action of \(\mathcal G\), named the fundamental group of \(\mathcal M\):

\[ \mathcal{M} = \mathbb H^2 / \mathcal{G}. \]

The natural projection map from \(\mathbb H^2\) onto \(\mathcal M\) is denoted with \(\pi\). By definition, all points of \(\mathbb H^2\) that belong to the same orbit under the action of \(\mathcal G\) project by \(\pi\) onto the same point of the surface \(\mathcal M\).

Figure 43.1 Left: The hyperbolic translations \(a,b,c,d\) and their inverses identify opposite sides of the octagon \(\mathcal D_O\). Center: Illustration of periodicity in the hyperbolic plane. The figure shows a few periodic copies of the points in the central octagon. Right: The half-open octagon \(\mathcal D\) is an original domain for \(\mathcal{M}\). Note that only one vertex of the octagon is contained in the original domain.

The half-open octagon \(\mathcal D\) shown in Figure 43.1 - Right contains exactly one representative of each point of \(\mathcal{M}\); \(\mathcal D\) is called the original domain of \(\mathcal{M}\). A point outside \(\mathcal D\) is the image of a point in \(\mathcal D\), its representative, under the action of a uniquely defined translation in the group \(\mathcal{G}\).

Figure 43.2 illustrates how a genus-2 surface can be obtained by identifying opposite sides of \(\mathcal D\) under the action of \(\mathcal G\).

Figure 43.2 Topological construction of a genus-2 surface from the original domain \(\mathcal D\) of the Bolza surface. Each open side is paired with the opposite closed side. Note that all vertices of the octagon project to the same point on the surface, which is represented uniquely by the only vertex of \(\mathcal D\).

Representation

In the following, and when there is no risk of confusion, the same notation will be used for a point on the surface \(\mathcal M\) and its representative in \(\mathcal D\). Similarly, \(\mathcal{P}\) will denote both a set of points on the surface and the set of their representatives in \(\mathcal D\).

We require that all input points lie inside \(\mathcal D\).

Data Structure

The Delaunay triangulation of \(\mathcal{M}\) defined by a point set \(\mathcal{P}\) is the projection by \(\pi\) of the Delaunay triangulation in the plane \(\mathbb H^2\) of the (infinite) set of points \(\mathcal{G P}\) onto \(\mathcal{M}\), provided that some condition (detailed in Section Simplicial Embedding Condition below) holds. More details can be found in [1].

As for orbits of points, all faces of the Delaunay triangulation of \(\mathcal{G P}\) that are in the same orbit under the action of \(\mathcal{G}\) project by \(\pi\) onto the same face on the surface \(\mathcal{M}\). We can define a unique canonical representative for each orbit, which has at least one vertex in \(\mathcal D\). Some canonical faces have vertices both inside and outside \(\mathcal D\); such faces can be uniquely specified by three pairs of points in \(\mathcal D\) and reduced translations of \(\mathcal{G}\) (points in the original domain are paired with the identity translation \(\mathbb 1)\). The underlying combinatorial triangulation is a 2D Triangulation Data Structure enriched in each face by the three translations that are paired with the point in each vertex of the canonical representative (see Figure 43.3).

Figure 43.3 Representation of a face \(f\) stored in the triangulation data structure. Each vertex \(v_i\) stores a point \(p_i\) paired with a translation \(\tau_i\).

More precisely, the translations are elements of the subset \(\mathcal N\) of \(\mathcal G\) for which the image of \(\mathcal D_O\) has at least one vertex in common with \(\mathcal D_O\). These images of \(\mathcal D_O\) by translations in \(\mathcal N\) are shaded in Figure 43.4; we consider them to be ordered counterclockwise around \(\mathcal D_O\), arbitrarily starting with the one corresponding to translation \(abcd\). The canonical representative in \(\mathbb H^2\) of a face on \(\mathcal M\) is such that

  • either all vertices of the representative lie in \(\mathcal D\), or
  • at least one of the vertices of the representative lies in \(\mathcal D\) and is as close as possible to \(abcd\) in the ordering defined above.

Figure 43.4 Among the three faces in the orbit that have at least one vertex in \(\mathcal D\), the canonical representative is the green one: it is closest, in the counterclockwise order, to the region labeled \(abcd\). The translations \(\mathbb 1, a,\) and \(\overline{b}\), corresponding to the points \(p, q\), and \(r\) are stored in the face in the data structure.

Simplicial Embedding Condition

Let us now give details on the condition mentioned above. The projection by \(\pi\) of the Delaunay triangulation in \(\mathbb H^2\) of \(\mathcal{G P}\) is a triangulation of \(\mathcal{M}\) if and only if its combinatorial graph does not contain loops (i.e., edges having two identical vertices) or double edges (i.e., two edges sharing the same two vertices), or, equivalently, if the projection is a simplicial complex:

  • any face of a simplex is a simplex, and
  • two simplices either do not intersect or share one common face.

Some point sets do not define a triangulation of \(\mathcal M\). For instance, a single point does not define a triangulation of \(\mathcal M\), as all edges of the induced subdivision would be loops. Another example is shown in Figure 43.5.

Figure 43.5 The three points in the central octagon define a non-simplicial subdivision of the Bolza surface. The pink edge between the two blue vertices corresponds to a loop on the surface. The two blue edges correspond to a double edge on the surface; the two green edges, too.

We initialize a triangulation of \(\mathcal M\) with a predetermined set \(\mathcal Q\) of 14 points, called dummy points (see Figure 43.6), which ensures that the projection by \(\pi\) of the Delaunay triangulation of \(\mathcal{G} (\mathcal{P}\cup\mathcal{Q})\) onto \(\mathcal M\) is a simplicial complex for any set of input points \(\mathcal{P}\) [1].

Figure 43.6 Delaunay triangulation of \(\mathcal M\) defined by the 14 dummy points. The set \(\mathcal Q\) of dummy points contains the origin, the vertex of \(\mathcal D\), the midpoints of the closed sides of \(\mathcal D\), and the (hyperbolic) midpoints of the segments between the origin and the vertices of the octagon.

If sufficiently many well-distributed points are inserted, the dummy points become unnecessary, i.e., the subdivision remains a simplicial complex even if we remove them. Otherwise some dummy points are left to ensure that the final subdivision is a simplicial complex.

Software Design

The main class of this package is Periodic_4_hyperbolic_Delaunay_triangulation_2, which implements Delaunay triangulations of the Bolza surface \(\mathcal M\). The prefix "Periodic_4" emphasizes that the triangulation in the universal covering \(\mathbb H^2\) is periodic in the four directions defined by the hyperbolic translations \( a,b,c\), and \(d\).

The implementation is fully dynamic, supporting both point insertion and vertex removal. However, a vertex can be removed only if the subdivision remains a simplicial complex.

The class Periodic_4_hyperbolic_Delaunay_triangulation_2 provides high-level geometric functionality specific to Delaunay triangulations, such as point insertion and vertex removal, the side-of-circle test, finding the conflicting region of a given point, and dual functions. It inherits from a base class, Periodic_4_hyperbolic_triangulation_2, which provides functionality common to triangulations in general, such as point location [2], and access functions, but supports neither point insertion nor vertex removal.

Both classes Periodic_4_hyperbolic_triangulation_2 and Periodic_4_hyperbolic_Delaunay_triangulation_2 have two template parameters that reflect the separation between the geometric and combinatorial structures of the triangulation:

  • a geometric traits class, which provides all objects, predicates, and constructions that are necessary for the computation of Delaunay triangulations of the Bolza surface (see Section The Geometric Traits Parameter).
  • a triangulation data structure class, which encodes the combinatorial structure of the triangulation (see Section The Triangulation Data Structure Parameter).

The Geometric Traits Parameter

The geometric traits class must fulfill the requirements described in the concept Periodic_4HyperbolicDelaunayTriangulationTraits_2. Moreover, the traits class must represent hyperbolic translations of the group \(\mathcal G\) via the class Hyperbolic_octagon_translation.

A model for the concept Periodic_4HyperbolicDelaunayTriangulationTraits_2 offered by this package is the class Periodic_4_hyperbolic_Delaunay_triangulation_traits_2. The class has one template parameter:

The Triangulation Data Structure Parameter

The triangulation data structure is a container for the faces and vertices and maintains incidence and adjacency relations. This parameter must meet the requirements described in the concept TriangulationDataStructure_2, for which CGAL offers the model Triangulation_data_structure_2. This model is itself parameterized by a vertex base class and a face base class, which gives the possibility to customize the vertices and faces used by the triangulation data structure. To represent periodic hyperbolic triangulations, the face base and vertex base classes must be models of the concepts Periodic_4HyperbolicTriangulationFaceBase_2 and Periodic_4HyperbolicTriangulationVertexBase_2, respectively.

The default value for the triangulation data structure parameter is Triangulation_data_structure_2< Periodic_4_hyperbolic_triangulation_vertex_base_2<Traits>, Periodic_4_hyperbolic_triangulation_face_base_2<Traits> >, where Traits is a model of Periodic_4HyperbolicDelaunayTriangulationTraits_2.

Example

Most functionalities of periodic hyperbolic triangulations are similar to those of Euclidean 2D triangulations. We refer the reader to the following two examples of the 2D Triangulation package:

The example below shows the initialization of a periodic hyperbolic 2D Delaunay triangulation, the insertion of random points uniformly distributed in the unit disk for the Euclidean metric, and the properties of the triangulation after the insertion. It uses the default triangulation data structure.


File Periodic_4_hyperbolic_triangulation_2/p4ht2_example_insertion.cpp

#include <CGAL/Hyperbolic_octagon_translation.h>
#include <CGAL/Periodic_4_hyperbolic_Delaunay_triangulation_2.h>
#include <CGAL/Periodic_4_hyperbolic_Delaunay_triangulation_traits_2.h>
#include <CGAL/point_generators_2.h>
typedef Traits::FT NT;
typedef Triangulation::Vertex_handle Vertex_handle;
typedef Triangulation::Point Point;
typedef Triangulation::size_type size_type;
int main(int argc, char** argv)
{
int N;
if(argc < 2) {
std::cout << "usage: " << argv[0] << " [number_of_points_to_insert]" << std::endl;
std::cout << "Defaulting to 100k points..." << std::endl;
N = 100000;
} else {
N = atoi(argv[1]);
}
int N1 = N/2;
int N2 = N - N1;
// Generate N random points uniformly distributed with respect to the Euclidean
// metric in the disk with radius 0.85, which contains the fundamental domain.
// Some of the points will be outside the octagon, so they will not be inserted.
std::vector<Point> pts;
CGAL::Random_points_in_disc_2<Point,Creator> g(0.85);
std::copy_n(g, N1, std::back_inserter(pts));
// The triangulation is automatically initialized with the dummy points.
Triangulation tr;
// Batch-insert new points in the triangulation. Note that by default, after
// the insertion of a set of points, an attempt is made to remove all dummy points
// from the triangulation. Note the third boolean parameter in the call of the
// `insert()` function which suppresses the automatic removal of dummy points.
std::cout << "Batch-inserting " << N1 << " random points in the triangulation... "; std::cout.flush();
size_type N_batch_inserted = tr.insert(pts.begin(), pts.end(), false);
std::cout << "DONE! " << std::endl;
// Insert new points in the triangulation one by one. When points are inserted
// one by one, dummy points are not automatically removed.
std::cout << "Single-inserting " << N2 << " random points in the triangulation... "; std::cout.flush();
int N_single_inserted = 0;
for(int i=0; i<N2; ++i)
{
Vertex_handle vh = tr.insert(*(++g));
if(vh != Vertex_handle())
N_single_inserted++;
}
std::cout << "DONE! " << std::endl;
// Total number of inserted points.
std::size_t N_inserted = N_batch_inserted + N_single_inserted;
// Finally, we try to manually remove all dummy points from the triangulation.
std::cout << "Cleaning dummy points from the triangulation... "; std::cout.flush();
tr.try_to_remove_dummy_vertices();
std::cout << "DONE! " << std::endl;
// Make sure that the triangulation is valid.
assert(tr.is_valid());
std::size_t NV = tr.number_of_vertices();
std::size_t NF = tr.number_of_faces();
std::size_t NE = tr.number_of_edges();
// This function `tr.number_of_dummy_points()` returns the number of dummy points that
// are currently in the triangulation.
int NDP = tr.number_of_dummy_points();
std::cout << std::endl;
std::cout << "-------------- STATS --------------" << std::endl;
std::cout << "Random points generated: " << N << std::endl;
std::cout << "Vertices in the triangulation: " << NV << std::endl;
std::cout << "Dummy points in the triangulation: " << NDP << std::endl;
std::cout << "Random points inserted: " << N_inserted << std::endl;
std::cout << "Random points outside/duplicates: " << (N - N_inserted) << std::endl;
std::cout << std::endl;
std::cout << "---------- COMBINATORICS ----------" << std::endl;
std::cout << "Number of vertices NV: " << NV << std::endl;
std::cout << "Number of faces NF: " << NF << std::endl;
std::cout << "Number of edges NE: " << NE << std::endl;
// The number of vertices in the triangulation must equal the number of points
// inserted, plus the dummy points in the triangulation.
assert(N_inserted + NDP == NV);
std::cout << "Number of vertices is correct! " << std::endl;
// Note that the Euler relation is already verified by the function `is_valid()`.
assert((2 + NE - NV - NF) / 2 == 2);
std::cout << "Euler relation verified! " << std::endl << std::endl;
return EXIT_SUCCESS;
}

Performance

We have measured the insertion execution time of our implementation against other CGAL triangulations. Random points, uniformly distributed in the unit disk with respect to the Euclidean metric, are generated. From these generated points, we select 1 million points that lie in the original octagon \(\mathcal D_O\). We insert the same set of points in three triangulations:

  • a periodic hyperbolic Delaunay triangulation with CORE::Expr as number type;
  • a (non-periodic) Euclidean Delaunay triangulation with CORE::Expr as number type;
  • a (non-periodic) Euclidean Delaunay triangulation with double as number type.

We report results averaged over 10 executions of this experiment in Table 1 below. The experiments have been executed on two machines:

  • Machine 1: MacBook Pro (2015), CPU: Intel Core i5 @ 2.9 GHz, RAM: 16 GB @ 1867 MHz, OS: Mac OS X (10.10.5), Compiler: clang-700.1.81;
  • Machine 2: Dell Vostro 5471 (2018), CPU: Intel Core i5 @ 1.6 GHZ (up to 3.4 GHz in TurboMode), RAM: 8GB @ 2400 MHz, OS: Linux Mint 19 (kernel 4.15.0), Compiler: gcc version 7.3.0.
Table 1: Comparison of insertion times of 1 million random points
Triangulation type Machine 1 Machine 2
Periodic hyperbolic (CORE::Expr)
55 sec. 40 sec.
Non-periodic Euclidean (CORE::Expr) 24 sec. 20 sec.
Non-periodic Euclidean (double) 1 sec. 1 sec.

Another experiment shows that, on average, all dummy points are removed from the triangulation with the insertion of fewer than 200 random points uniformly distributed in the unit disk with respect to the Euclidean metric. We start with an empty triangulation of the Bolza surface (i.e., initialized with only the dummy points), and we insert random points one by one. After each insertion, the unnecessary dummy points (if any) are removed from the triangulation. As soon as the last dummy point has been removed, we stop the process and record the number of random points inserted. Results are shown in Figure 43.7.

Figure 43.7 Histogram of the number of random input points needed to remove all dummy points in a triangulation of the Bolza surface. The histogram shows results acquired from 1000 executions of the experiment described above. In these 1000 executions, the minimum number of random points needed to remove all dummy points was 23, the maximum 188, and on average 64 random points were needed.

Design and Implementation History

This package implements the algorithm for the computation of Delaunay triangulation of the Bolza surface proposed by Mikhail Bogdanov, Monique Teillaud, and Gert Vegter [1]. The implementation itself is described by Iordan Iordanov and Monique Teillaud [3].

In 2016, Iordanov started working on the 2D Periodic Hyperbolic Triangulations package.

The authors would like to thank their partners from the Astonishing and SoS projects for very helpful discussions.