Shortest path embeddings of graphs on surfaces

The classical theorem of F\'{a}ry states that every planar graph can be represented by an embedding in which every edge is represented by a straight line segment. We consider generalizations of F\'{a}ry's theorem to surfaces equipped with Riemannian metrics. In this setting, we require that every edge is drawn as a shortest path between its two endpoints and we call an embedding with this property a shortest path embedding. The main question addressed in this paper is whether given a closed surface S, there exists a Riemannian metric for which every topologically embeddable graph admits a shortest path embedding. This question is also motivated by various problems regarding crossing numbers on surfaces. We observe that the round metrics on the sphere and the projective plane have this property. We provide flat metrics on the torus and the Klein bottle which also have this property. Then we show that for the unit square flat metric on the Klein bottle there exists a graph without shortest path embeddings. We show, moreover, that for large g, there exist graphs G embeddable into the orientable surface of genus g, such that with large probability a random hyperbolic metric does not admit a shortest path embedding of G, where the probability measure is proportional to the Weil-Petersson volume on moduli space. Finally, we construct a hyperbolic metric on every orientable surface S of genus g, such that every graph embeddable into S can be embedded so that every edge is a concatenation of at most O(g) shortest paths.


Introduction
Fáry's theorem and joint crossing numbers. A famous theorem of Fáry [11] states that any simple planar graph can be embedded so that edges are represented by straight line segments. In this article we investigate analogues of this theorem in the context of graphs embedded into surfaces. We focus on the following problem: Given a surface S, is there a metric on S such that every graph embeddable into S can be embedded so that edges are represented by shortest paths ? We call such an embedding a shortest path embedding, and such a metric a universal shortest path metric. 1 Before being enticed by this question, we were motivated to consider it by a number of problems involving joint embeddings of curves or graphs on surfaces arising from seemingly disparate settings. The literature on the subject goes back at least 15 years with Negami's work related to diagonal flips in triangulations [25]. He conjectured that there exists a universal constant c such that for any pair of graphs G 1 and G 2 embedded in a surface S, there exists a homeomorphism h : S → S such that h(G 1 ) and G 2 intersect transversely at their edges and the number of edge crossings satisfies cr(h(G 1 ), G 2 ) ≤ c|E(G 1 )| · |E(G 2 )|.
Recently, on one hand, Matoušek, Sedgwick, Tancer, and U. Wagner [20,21], working on decidability of embeddability of 2-complexes into R 3 and on the other hand, Geelen, Huynh, and Richter [13], in a quest for explicit bounds for graph minors, were faced with a similar question and provided bounds for related problems. Joint crossing number type problems are dually equivalent to problems of finding a graph with a specific pattern within an embedded graph while bounding the multiplicity of the edges used. This is a fundamental concern of computational topology of surfaces where one is interested in finding objects with a fixed topology and minimal combinatorial complexity, e.g., short canonical systems of loops [19], short pants decompositions [6] or short octagonal decompositions [4]; see also [5].
Negami provided the upper bound cr(h(G 1 ), G 2 ) ≤ cg|E(G 1 )| · |E(G 2 )|, and despite subsequent discoveries [1,27], his conjecture is still open. In a paper that refines Negami's work [27], Richter and Salazar wrote "this [conjecture] seems eminently reasonable: why should two edges be forced to cross more than once?". The connection with our work is that if two graphs are embedded transversally by shortest path embeddings, then indeed no two edges cross more than once, since otherwise one of them could be shortcut. In particular, a proof that every surface admits a universal shortest path metric would imply Negami's conjecture, actually even if we allowed to subdivide each edge of the embedded graph constantly many times.
We note that prior to our work, Schaefer [28, paragraph on Geodesic crossing numbers] had considered similar questions, mainly for drawing edges of a graph by geodesics. We provide the details below including answers to some of Schaefer's questions. We also note that our methods easily yield a new proof of Negami's theorem for orientable surfaces; see Corollary 20. Beyond crossing numbers, the existence or non-existence of shortest path universal metrics might be relevant in curvature free and extremal Riemannian geometry.
Related work. Various results in graph drawing [33] revolve around generalizing Fáry's theorem to find drawings of graphs with additional constraints, for instance drawing the edges with polylines with few bends. On the other hand, only few extensions to graphs embedded in surfaces are known. Two classical avatars of Fáry's theorem in the plane are of relevance to our work: Tutte's barycentric embedding theorem [35] and the Koebe-Andreev-Thurston circle packing theorem (see, for example, the book of Stephenson [29]). Both have been generalized to surfaces, providing positive answers to the following questions: 1. Given a surface S, a metric m, and a graph G embeddable into S, can we embed the graph G so that every edge is represented by a geodesic with respect to m?
2. Given a graph G embeddable into S, does there exist a metric m on S so that G embeds into S with shortest paths? a.
b. c. d. Figure 1: a. and b. Two geodesics crossing many times. c. A grid embedded in a torus with geodesics. d. A reembedding of this grid with shortest paths.
The first question was considered by Schaefer [28]; a positive answer for many metrics had been previously given by Y. Colin de Verdière [7] who generalized Tutte's barycentric embedding approach using a variational principle. The idea behind this approach is to start with a topological embedding of the graph, replace the edges by springs, and let the system reach an equilibrium. Y. Colin de Verdière proved that for any metric of non-positive curvature, the edges become geodesics with disjoint interiors when the system reaches stability; moreover, this embedding is essentially unique within its homotopy class. However, geodesics need not be shortest paths, and two geodesics can intersect an arbitrarily large number of times, see Figure 1. Yet, these examples do not provide a negative answer to the second question, or to our main question, since we could change the embedding by a homeomorphism of the torus (thus even preserving the combinatorial map) to obtain a shortest path embedding.
The second question also has a positive answer, which can be proved via a generalization of the circle packing theorem to closed surfaces [29]. Namely, for every triangulation T of a surface, there exists a metric of constant curvature so that T can be represented as the contact graph of a family of circles. The representation of the triangulation that places a vertex at the center of its corresponding circle is an embedding with shortest paths. Such a representation can be computed efficiently and can be used as a tool for representing graphs on surfaces [23]. However, the metric is determined by the triangulation, which makes this approach ill-suited for our purpose.
Our results. Our objective here is a mix of these last two results. On the one hand, we require shortest paths and not geodesics, on the other hand, we want a single metric for each surface and not one which depends on the triangulation. We will also consider the relaxation of our problem where we are allowed to use concatenations of shortest paths: we say that a metric is a k-universal shortest path metric if every topologically embeddable graph can be represented by an embedding in which edges are drawn as concatenations of k shortest paths. This is akin to various problems in graph drawing where graphs are embedded with polylines with a bounded number of bends instead of straight lines [9,32].
Our results focus on Riemannian metrics of constant curvature, and our techniques are organized by the sign of the curvature. We first observe that for the sphere and the projective plane, since there is a unique Riemannian metric of curvature 1, the circle packing approach applies to all graphs. Then, with the aid of irreducible triangulations, we provide flat metrics (i.e., of zero curvature) on the torus and the Klein bottle for which every graph admits a shortest path embedding. Theorem 1. The sphere S 2 , the projective plane RP 2 , the torus T 2 , and the Klein bottle K can be endowed with a universal shortest path metric.
This result could lead to the idea that shortest path embeddings can be achieved for any metric, i.e., that every metric is a universal shortest path metric. We prove that this is not the case already for the unit square flat metric on the Klein bottle (arguably the first example to consider). Theorem 2. Let K denote the Klein bottle endowed with the unit square flat metric on the polygonal scheme aba −1 b. Then there exists a graph embeddable into K which cannot be embedded into K so that the edges are shortest paths.
In higher genus, the number of irreducible triangulations is too large to check all cases by hand. Hyperbolic surfaces of large genus are hard to comprehend, but the probabilistic point of view allows us to show that if there exist universal shortest path metrics of constant curvature −1 at all, their fraction tends to 0 as the genus tends to infinity.
Theorem 3. For any ε > 0, with probability tending to 1 as g goes to infinity, a random hyperbolic metric is not a O(g 1/3−ε )-universal shortest path metric. In particular, with probability tending to 1 as g goes to infinity, a random hyperbolic metric is not a universal shortest path metric.
Here the probability measure on the space hyperbolic surfaces is proportional to the Weil-Petersson volume, see Section 5. Our proof is an application of deep results on this volume by Mirzakhani [22] and Guth, Parlier, and Young [15].
For a given graph G and a metric m on S, Schaefer [28] defines the geodesic crossing number of G as the minimal number of crossings of any drawing of G in S in which edges are represented by geodesics. Schaefer asks if this definition is equivalent to the analogous definition with shortest paths instead of geodesics. Notice that the examples in Theorems 2 and 3 have nonpositive curvature, hence, combined with the aforementioned result of Y. Colin de Verdière imply that some graphs have geodesic crossing number zero but shortest path crossing number nonzero, answering Schaefer's question.
For genus g > 1 we do not know if there exist shortest path universal metrics. But relaxing the question to concatenations of shortest paths and combining ideas from hyperbolic geometry and computational topology, we provide for every orientable surface of genus g an O(g)-universal shortest paths metric. The proof relies on the octagonal decompositions ofÉ. Colin de Verdière and Erickson [4] and a variant of the aforementioned theorem of Y. Colin de Verdière [7]. In this article we focused on Riemannian metrics of constant curvature, but we remark that both of our last results also hold in some setting of piecewise-Euclidean metrics as well. For the upper bound, it suffices to replace hyperbolic hexagons with Euclidean ones, and the rest of the proof works similarly. The lower bound can be derived following the heuristic strong parallels between the Weil-Petersson volume form on moduli space and the counting measure on the space of N = 4g Euclidean triangles randomly glued together. In particular the results that we use have analogs in this latter space: see Brooks and Makover [2] and the second half of the article of Guth, Parlier, and Young [15].
We have stated our results for graphs in this introduction. We note that one could consider the problem of shortest path embeddings for a graph with a fixed embedding up to a homeomorphism of the surface (i.e., for a combinatorial map), which is more in the spirit of Negami's conjecture. Our positive results can be stated in this stronger version; i.e., in our proofs the map is preserved. Our negative results would be weaker if the map had to be preserved, and in fact the proofs deal firstly with the statements for maps and then we derive the analog for graphs with some extra work.

Open questions.
The main open question is the existence of universal shortest path metrics, or O(1)-universal shortest path metrics. Natural candidates for these are given by certain celebrated extremal metrics like the ones occurring as lower bounds for Gromov's systolic inequality [3,14].
Lists of irreducible triangulations exist for the double torus and the non-orientable surface of genus up to four [30]. While the numbers are too big to be investigated by hand as we did for the torus and the Klein bottle, it may be possible to investigate some computerized approach to test their shortest path embeddability for some well chosen hyperbolic metric.
Our Theorem 4 only deals with orientable surfaces. A similar approach might work for non-orientable surfaces as well, the key issue being to generalize the octagonal decompositions ofÉ. Colin de Verdière and Erickson [4] to the non-orientable setting. We leave this as an open problem.

Preliminaries
In this article we only deal with compact surfaces without boundaries. By the classification theorem, these are characterized by their orientability and their genus, generally denoted by g. Orientable surfaces of genus 0 and 1 are respectively the sphere S 2 and the torus T 2 , while non-orientable surfaces of genus 1 and 2 are the projective plane RP 2 and the Klein bottle K. The orientable surface of genus g is denoted by S g . The Euler genus is equal to the genus for non-orientable surfaces and equals twice the genus for orientable surfaces.
By a path on a surface S we mean a continuous map p : [0, 1] → S, and a closed curve denotes a continuous map γ : S 1 → S. These are simple if they are injective. We will be using occasionally the notions of homotopy, homology, and universal cover, we refer to Hatcher [16] for an introduction to these concepts. All the graphs that we consider in this paper are simple graphs unless specified otherwise, i.e., loops and multiple edges are disallowed. An embedding of a graph G into a surface S is, informally, a crossing-free drawing of G on S. We refer to Mohar and Thomassen [24] for a thorough reference on graphs on surfaces, and only recall the main definitions. A graph embedding is cellular if its faces are homeomorphic to open disks. Euler's formula states that v − e + f = 2 − g for any graph with v vertices, e edges, and f faces cellularly embedded in a surface S of Euler genus g. When the graph is not cellularly embedded, this becomes an inequality: v − e + f ≥ 2 − g. A triangulation of a surface is a cellular graph embedding such that all the faces are adjacent to three edges. An isomorphism between two triangulations is a bijection between the vertices, edges and faces that respects incidences. By a slight abuse of language, we will sometimes refer to an embedding of a triangulation, by which we mean an embedding of its underlying graph which is homeomorphic to the given triangulation. A pants decomposition of an orientable surface S is a family of disjoint curves Γ such that cutting S along all of the curves of Γ gives a disjoint union of pairs of pants, i.e., spheres with three boundaries. Every orientable surface except the sphere and the torus admits a pants decomposition with 3g − 3 closed curves and 2g − 2 pairs of pants. Note that all the pants decompositions are not topologically the same, i.e., are not related by a self-homeomorphism of the surface. A class of pants decompositions equivalent under such homeomorphisms will be called the (topological) type of the pants decomposition. We say that an embedding f : G → S contains a pants decomposition if there exists a subgraph H ⊆ G such that f : H → S is a pants decomposition of S.
In this article, we will also be dealing with notions coming from Riemannian geometry, we refer to the book of do Carmo for more background [8]. By a metric we always mean a Riemannian metric, which associates to every point of a surface the curvature at this point. The Gauss-Bonnet theorem ties geometry and topology; it implies that the sign of a metric of constant curvature that a topological surface accepts is determined solely by its Euler genus.
A Riemannian metric induces a length functional on paths and closed curves. A path or a closed curve is a geodesic if the functional is locally minimal. Shortest paths between two points are global minima of the length functional. Unlike in the plane, geodesics are not, in general, shortest paths; in addition, neither geodesics nor shortest paths are unique in general. If we have a shortest path embedding of a graph where every edge is drawn as the unique shortest path between its endpoints, we speak of shortest paths embedding with uniqueness.
3 Shortest path embeddings for low genus surfaces Theorem 1. The sphere S 2 , the projective plane RP 2 , the torus T 2 , and the Klein bottle K can be endowed with a universal shortest path metric.
In the theorem above, for S 2 and RP 2 we use the round metric of positive constant curvature scaled to 1. In the case of torus we use the flat metric obtained by the identification of the opposite edges of the square. In the case of the Klein bottle we can show that an analogous result fails with the flat square metric on the polygonal scheme aba −1 b, as we will see in Section 4. But we can get the result for the metric obtained by the identification of the edges of a rectangle of dimensions 1 × b where b = 4/3 + ε for some small ε > 0. (The edges of length 1 are identified coherently, whereas the edges of length b are identified in opposite directions.) In all cases we can get shortest path embeddings with uniqueness. Actually, for the torus and the Klein bottle, uniqueness will be a convenient assumption for inductive proofs.
The sphere and the projective plane. By the circle packing theorem any planar graph can be represented as the contact graph of a circle packing on the sphere (endowed with the standard round metric) [29,Theorem 4.3]. On the sphere each circle is the boundary of a cap (a metric ball), and by the center of the circle we mean the center of the corresponding cap. It is easy to see that drawing each edge (u, v) of a contact graph, by the shortest path between the centers the circles corresponding to u and v is an embedding. Since these are shortest paths, this proves Theorem 1 for S 2 .
For the projective plane, a similar circle packing theorem follows from the spherical case. Since we could not find a reference in the literature we include a proof here.
Henceforth, the sphere and the projective plane are always endowed with their usual spherical metrics (of constant curvature). For any circle packing P ⊂ S 2 , consider its contact graph, which we denote by C(P ), together with the embedding of C(P ) to S 2 in which the edge corresponding to touching circles C v and C w is drawn by the geodesic between the centers of the circles C v and C u in S 2 . We will only consider packings P for which this embedding of C(P ) is a triangulation of S 2 , and we call the corresponding triangulation the carrier of P . A map S 2 → S 2 mapping a circle packing P to itself induces an automorphism of its carrier (as defined in the preliminaries), which by a slight abuse of language we call an automorphism of the circle packing P .
Proposition 5. Every triangulation of RP 2 is the carrier of a circle packing in RP 2 .
In particular, for any triangulation T of RP 2 , this provides a shortest path embedding of T . Since any simple graph embedded on RP 2 can be extended to a triangulation, this proves Theorem 1 for RP 2 .
Proof. Let T be a triangulation of RP 2 . Let π : S 2 → RP 2 be the projection map sending each pair of antipodal points in S 2 to a point in RP 2 . LetT be the double cover of T , which is a triangulation of S 2 induced by π −1 (T ), and let i be the automorphism i :T →T induced by the antipodal map. By the Koebe-Andreev-Thurston theorem there exists a circle packingP ⊂ S 2 whose carrier is isomorphic toT . Furthermore, this circle packing is unique up to Möbius transformations [34,Chapter 13], in particular, any automorphism ofP is induced by a Möbius transformation of S 2 . Thus the map i is induced by a Möbius transformation φ : S 2 → S 2 . Furthermore, since i is fixed-point free, so is φ: if φ(x) = x for x within a circle corresponding to a vertex v, then i(v) = v which is a contradiction, and similarly, a fixed point on the intersection of two circles, or in the region between three adjacent circles would correspond to a fixed edge or face for i, which is also a contradiction. Now, any fixed point free Möbius transformation of the sphere is the antipodal map upto a Möbius transformation [36]. Specifically, there exists another Möbius transformation τ : S 2 → S 2 such that τ −1 • φ • τ is the antipodal map. We can conclude that the circle packingQ = τ −1 (P ) is centrally symmetric and therefore it projects to a circle packing Q = π(Q) in RP 2 . Since the carrier ofQ is isomorphic toT , the carrier of Q is isomorphic to T .
Minimal triangulations. Let S be a surface and T be a triangulation of it. The triangulation T is called reducible, if it contains an edge e such that the contraction of e yields again a triangulation, which we denote by T /e. We refer to e as a contractible edge (we do not mean contractibility in a topological sense). On the other hand, a triangulation is minimal (or irreducible), if no edge can be contracted this way. For every surface there is a finite list of minimal triangulations. In particular, for the torus T 2 this list consists of 21 triangulations found by Lawrencenko [17] and for the Klein bottle K there are 29 minimal triangulations found by Sulanke [31].
The strategy of the proof of Theorem 1 for T 2 and K is to show that it is sufficient to check Theorem 1 for minimal triangulations with appropriate fixed metric; see Lemma 6. Then, since every embedded graph can be extended to a triangulation (possibly with adding new vertices), we finish the proof by providing the list of shortest path embeddings of the minimal triangulations.
Lemma 6. Let S be a surface equipped with a flat metric. Let T be a reducible triangulation with contractible edge e. Let us assume that T /e admits a shortest path embedding with uniqueness into S. Then T admits a shortest path embedding with uniqueness into S as well.
The restriction on flat metrics in the lemma above does not seem essential, but this is all we need and this way the proof is quite simple.
Proof. Let v be the vertex of T /e obtained by the contraction of e. We first consider the shortest path drawing of T /e. Then we perform the appropriate vertex splitting of v (the inverse operation of the contraction) in a close neighborhood of v so that we get a shortest path embedding of T . In order to see that this is indeed possible, let us consider the subgraph G formed by the edges incident to v. It is a simply connected set, which lifts isometrically to the universal cover so that the edges are realized by straight segments (since they are shortest paths). Thus we may choose ε > 0 small enough such that the ε-neighborhood N ε G of G is simply connected. Moreover, by compactness, for each edge uv of T /e, there exists ε ≤ ε such that for every v in the ε -neighborhood of v, the geodesic segment connecting u and v inside N ε G is the unique shortest path between u and v in S; see Figure 2 and footnote 2 . Therefore, it is sufficient to perform the vertex splitting of v in a sufficiently small neighborhood of v so that we do not introduce new intersections.
The minimal triangulations of T 2 and K. In Figure 3 we provide a list of shortest path embeddings with uniqueness of minimal triangulations of the torus with a flat metric obtained by identifying the opposite edges of the unit square. They are in the same order as in 2 Indeed, let us consider two functions d, dOUT : The function dOUT is well defined as the function g(y) := dist(u, y) + dist(y, x) is continuous and attains its minimum on the compact set S \ N ε G . By the triangle inequality |dOUT (x) − dOUT (x )| ≤ dist(x, x ) for x, x ∈ S which implies that dOUT is continuous. Finally, we observe that d(v) < dOUT(v) as the shortest path connecting u and v is unique. Therefore there is an open Nv. This is the required ε needed for the edge uv. We just skip drawings of the triangulations 7 to 17, because they are all analogous to the triangulation 6, they only have different patterns of diagonals. It is clear that every edge is a geodesic. In order to check that each of them is drawn as a shortest path, it is sufficient to verify that each edge projects vertically and horizontally to a segment of length less than 1 2 .
For the Klein bottle K, we also provide a metric such that all the minimal triangulations admit shortest path embeddings with uniqueness. We obtain this metric as the identification of the edges of the rectangle R = [0, a] × [0, b], where a = 1 and b = 4/3 + ε for sufficiently small ε. The edges of length 1 are identified in coherent directions. The edges of length b are identified in the opposite directions. The value b = 4/3 + ε is set up in such a way that if we consider the points p = (0, 3 4 then the shortest path between p and q is the horizontal path of height 1 4 b. However, when we shift p along the boundary of R a little bit closer to the center, say by 1 1000 , then the shortest path becomes the diagonal edge connecting the left copy of p and q, see Figure 5.
There are 29 minimal triangulations of the Klein bottle. A list of 25 of them was first found by Lawrencenko and Negami [18]. Later on, Sulanke [31] found a gap in the claimed completeness of this list and provided a complete list containing 4 additional triangulations. These triangulations split into two classes. The 25 triangulations of the first class are named Kh1-Kh25 and the four triangulations of the second class are named Kc1-Kc4. The triangulations from the second class are those that contain a cycle of length 3 which splits the Klein bottle into two Möbius bands.
We begin examining the triangulations of the first class. We present shortest path embeddings Mb2 Mb3 Only for the drawings of Kh3, Kh12, Kh13, and Kh25 we did more significant movements. It is routine (but tedious) to check that all the edges are indeed drawn as shortest paths. For many edges this can be checked easily. For few not so obvious cases the general recipe is to use the universal cover approach and Lemma 8. Now let us focus on the triangulations in the second class. All of them are obtained by gluing two triangulations of the Möbius bands along their boundaries. In our case, we split K into two bands by a cycle depicted on the bottom left picture of Figure 4. There is an isometric homeomorphism which maps one band to another and which preserves the common boundary pointwise. Therefore, it is sufficient to present the shortest paths embeddings with uniqueness into the bands, as on the middle three pictures. Then we get drawings of Kc1-Kc4 using this homeomorphism. For example, Kc1 is obtained by gluing two copies of Mb1 together, as depicted on the bottom right picture of Figure 4. The vertices on the pictures are the lattice points of the same lattice as above with exception of two points of Mb3. The points on the 'central' cycle of Mb3 have coordinates ( 1 6 , b 2 ), ( 4 9 , b 2 ), and ( 8 9 , b 2 ). Note that we have significantly redrawn the original drawings of Lawrencenko and Negami [18], but it is easy to check that we get the same triangulations, because each triangulation of the Möbius band Mb1-Mb3 is quite small.

Square flat metric on the Klein bottle
The task of this section is to prove the following theorem. Theorem 2. Let K denote the Klein bottle endowed with the unit square flat metric on the polygonal scheme aba −1 b. Then there exists a graph embeddable into K which cannot be embedded into K so that the edges are shortest paths.
We consider the minimal triangulation Kc1 (see Figure 4, bottom, right) and we denote by G the underlying graph for this triangulation. We will prove that G does not admit a shortest path embedding into K with the square metric. First, we observe that the triangulation Kc1 is the only embedding of G into K. Proposition 7. G has a unique embedding into the Klein bottle.
Proof of Proposition 7. G has 9 vertices and their degrees are (8,8,8,5,5,5,5,5,5). Since no other irreducible triangulation of the Klein bottle has this degree sequence, any other hypothetical embedding of G into the Klein bottle is either non-cellular or has a reducible edge. In the first case, it means that G is cellularly embeddable into the sphere or the projective plane, which is not the case. Indeed, it is obtained as the gluing of two copies of K 6 along a triangle, and therefore contains K 5 ⊕ K 5 (two copies of K 5 identified along an edge minus that edge) as a minor, which does not embed into the projective plane [24, Figure 6.4]. In the latter case, we observe that an edge contraction cannot decrease the degree of all three degree 8 vertices, and thus we reach a contradiction since a triangulation on 8 vertices cannot have a degree 8 vertex.
For contradiction, let us assume that G admits a shortest paths embedding into K. We know that Kc1 is obtained by gluing two triangulations of a Möbius band along a cycle of length 3 (the triangle corresponding to this cycle is not part of the triangulation). Let abc be this cycle. With a slight abuse of notation we identify this cycle with its image in the (hypothetical) shortest path embedding into K. Our strategy is to show that already abc cannot be embedded into K with shortest path edges, which will give the required contradiction. By Proposition 7, we know that abc splits K to two Möbius bands. Let X = R 2 be the universal cover of K (with standard Euclidean metric). Let π : X → K be the isometric projection corresponding to the cover. We will represent the Klein bottle with the flat-square metric as the unit square [0, 1] 2 with suitable identification of the edges (aba −1 b, as in the previous section). We will use the convention that π((0, 1) 2 ) = (0, 1) 2 ; that is, the projection is the identity on the interior of this square. See Figure 6.
Given a point p ∈ K we set X p := π −1 (p). Finally, let V p be the Voronoi diagram in X corresponding to the set X p . Lemma 8. Let p and q be two points in K and γ be an arc (edge) connecting them, considered as a subset of K. Then γ is the unique shortest path between p and q if and only if there are p ∈ X p , q ∈ X q such that γ = π(p q ) where p q denotes the straight edge connecting p and q in X and q belongs to the open Voronoi cell for p in V p .
Proof of Lemma 8. Any path κ with endpoints p and q lifts to some path κ with endpoints p ∈ X p and q ∈ X q (κ , p , and q are not determined uniquely). This lift preserves the length of the path. Vice versa, any path connecting a point in X p with a point in X q projects to a path connecting p and q (not necessarily simple), again preserving the length. Therefore, γ is the shortest path in K connecting p and q if and only if it lifts to a straight edge realizing the distance between X p and X q in X. Such an edge connects p ∈ X p and q ∈ X q . By symmetry, we can fix q arbitrarily and we look for the closest p . Then, a point p is the unique point of X p closest to q if and only if q belongs to the open Voronoi cell for p in V p . This is what we need. Now let us lift the cycle abc to a path a b c a in X; see Figure 6. Given a curve in X, we call the length of its projection to the x-axis, the "horizontal length" of the curve; similarly we speak about the horizontal distance and the vertical distance of two points in X.
Lemma 9. The horizontal distance between a and a is at least 2.
Proof of Lemma 9. If we consider the point a fixed, then the position of a in X a determines the homotopy class of the cycle abc in π 1 (K). Therefore, it also determines the homology classes of this cycle in H 1 (K; Z 2 ) and in H 1 (K; Z). We note that the cycle abc must be homologically trivial in H 1 (K; Z 2 ) because it bounds a Möbius band; however, it is homologically nontrivial in H 1 (K; Z) because it bounds a Möbius band (which is non-orientable) on both sides. In addition the cycle abc is two sided, that is, its (regular) neighborhood is an annulus and not a Möbius band.
The horizontal distance between a and a must be a non-negative integer. We will rule out the cases when this distance is 0 or 1.
If this distance is 1, then the cycle abc is not two-sided (this can be read on the lift), a contradiction.
If the horizontal distance is 0 and the vertical distance is odd, then abc is homologically nontrivial in H 1 (K; Z 2 ). (It is sufficient to consider the segment connecting a and a and project it to a cycle z in K. Then z is homotopy equivalent to abc.) A contradiction.
Similarly, if the horizontal distance is 0 and the vertical distance is even, then abc is homologically trivial in H 1 (K; Z). (Again we project the segment connecting a and a .) A contradiction.
Lemma 10. Let γ be a unique shortest path in K connecting points p and q. Let γ be a lift of γ with endpoints p and q . Then the horizontal distance in X between p and q is less than 5 8 . Proof. Let C be the open Voronoi cell for p in V p . By Lemma 8, q belongs to C. Therefore, it is sufficient to check that every point c of C has horizontal distance less than 5 8 from p . Without loss of generality, we may assume that the x-coordinate of p equals 0 since shifting p in horizontal direction only shifts X p and V p (note that this is not true for the vertical direction). For contradiction, there is a c in C at distance at least 5 8 and without loss of generality the x-coordinate of c is positive. Let p be the point of X p with x-coordinate equal 1 which is vertically closest to c (pick any suitable point in case of draw); see the picture on the left. The vertical distance between c and p is at most 1 2 . A simple calculation, using the Pythagoras theorem, gives that p is at most as far from c as p . A contradiction.
Finally, we summarize how the previous lemmas yield a contradiction. By Lemma 9, the horizontal distance between a and a is at least 2. On the other hand, Lemma 10 gives that the horizontal length of each of the edges a b , b c , and c a is at most 5 8 , altogether at most 15 8 . This gives the required contradiction, which finishes the proof of Theorem 2.

Asymptotically almost all hyperbolic metrics are not universal
Before stating the main theorem of this section, we will give some very quick background on the geometry of surfaces, we refer to Farb and Margalit [10] for a proper introduction. The Teichmüller space T g of a surface S of genus g denotes the set of hyperbolic metrics on S, such that two metrics are equivalent if they are related by an isometry isotopic to the identity. In some contexts, like ours, one might also want to identify metrics related by an isometry (not necessarily isotopic to the identity). The corresponding space is called the moduli space M g of the surface, and is obtained by quotienting T g by the mapping class group of S, i.e., its group of homeomorphisms. This moduli space can be endowed with multiple structures, here we will be interested in a particular one, called the Weil-Petersson metric. This metric provides M g with a Riemannian structure of finite volume, and therefore by renormalizing, we obtain a probability space, allowing to choose a random metric. We can now state the main theorem of this section.
Theorem 3. For any ε > 0, with probability tending to 1 as g goes to infinity, a random hyperbolic metric is not a O(g 1/3−ε )-universal shortest path metric. In particular, with probability tending to 1 as g goes to infinity, a random hyperbolic metric is not a universal shortest path metric.
The proof is a consequence of two important results on random hyperbolic metrics. The first is a small variant of a theorem of Guth, Parlier, and Young [15,Theorem 1] that relies on the work of Wolpert [37]. Before stating it, we need some definitions.
Given a hyperbolic metric m on a surface S, we say that m has total pants length at least if in any pants decomposition Γ of S, the lengths of the closed curves of Γ sum up to at least . We say that m has total pants length of type ξ at least if in any pants decomposition Γ of S of type ξ, the lengths of the closed curves of Γ sum up to at least . Theorem 11. For any ε > 0 and any family of types of pants decomposition (ξ g ), a random metric on M g has total pants length of type ξ g at least g 4/3−ε with probability tending to 1 as g → ∞.
Proof of Theorem 11. This bound is obtained with a similar technique as the proof of Theorem 1 of Guth, Parlier, and Young [15]. We refer to their article for more details, and as in their proof, we will discard non super-exponential terms, e.g., n! ≈ n n . For every a, b, c ∈ R + there exists a unique hyperbolic metric on a pair of pants with boundary lengths a, b and c (see for example Ratcliffe [26,Theorem 9.7.3]). For a pants decomposition of fixed type ξ g , the Weil-Petersson volume form on moduli space is the push forward of the form on Teichmüller space which is identified with R 6g−6 and the i denote the lengths of the (geodesic) boundaries of the pants decomposition, while the τ i quantify how much the metric twists around each geodesic. Since every full twist gives a homeomorphic metric, the subset of Teichmüller space {( i , τ i ) | i i ≤ L, 0 ≤ τ i ≤ i } projects surjectively onto the region of moduli space corresponding to surfaces with total pants length of type ξ g at most L. The volume of this set is bounded by ( L g ) 6g , which is to be compared with the total volume of moduli space ≈ g 2g . For L smaller than g 4/3−ε , the ratio tends to zero, which proves the theorem.
The following is an immediate corollary of this theorem. Corollary 12. Let T g be a family of triangulations of S g , such that every member of T g contains a pants decomposition of fixed type ξ g . For any ε > 0, with probability tending to 1 as g → ∞, a shortest embedding of T g into a random hyperbolic surface of genus g has length at least Ω(g 4/3−ε ).
Theorem 13. With probability tending to 1, the diameter of a random hyperbolic surface of genus g is O(log g). Theorem 3 is proved by providing an explicit family of graphs G g which will embed badly. It is defined in the following way for g ≥ 2. Let ξ g be a type of pants decompositions for every value of g.
• We start with a pants decomposition of type ξ g of a surface S g .
• We place four vertices on every boundary curve.
• We triangulate each pair of pants with a bounded size triangulation so that each cycle of length 3 bounds a triangle in the triangulation, and any path connecting two boundary components of the pair of pants has length at least 4 (in particular G g is a simple graph and each cycle of length 3 in the graph G g bounds a triangle in the triangulation).
The following proposition controls the issues related to the flexibility of embeddings of graphs into surfaces.

Proposition 14.
There is a unique embedding of G g into S g , up to a homeomorphism; in particular every embedding contains a pants decomposition of type ξ g .
Proof of Proposition 14. Let v be the number of vertices, e be the number of edges and t be the number of triangles of the triangulation in the definition of G g (triangles in the graph-theoretical sense). By Euler's formula and by the construction we get v − e + t = χ where χ is the Euler characteristic of S g . Let us consider an embedding Ψ of G g into S g . Let f be the number of faces of this embedding and F be the set of faces. Euler's formula for this embedding gives v − e + f ≥ χ (we get an inequality because some of the faces need not be embedded cellularly). In particular, we get f ≥ t. On the other hand, we get 2e = 3t and 2e = σ∈F deg σ ≥ 3f since each edge is in exactly two faces. This gives 3t ≥ 3f . Therefore, both of the aforementioned inequalities have to be equalities. In particular, each σ ∈ F is a triangle bounded by a cycle of length 3 in G g . Since the number of cycles of length 3 in G g equals t = f , we deduce that Ψ coincides with the embedding from the definition of G g up to a homeomorphism.
Remark: We preferred to use a hands-on construction of the graphs G g , but another approach could be to rely on the theory of LEW-embeddings and use one of its results on uniqueness of embeddings, see for example Mohar and Thomassen [24,Corollary 5.2.3].
With these three results at hand we are ready to provide a proof of the theorem.
Proof of Theorem 3. We use the family of graphs G g previously defined. Since there are O(g) curves in a pants decomposition, it contains O(g) edges, and every embedding of G g into S g contains a pants decomposition of type ξ g by Proposition 14. Now, by Corollary 12, for every ε > 0, and for g large enough, the probability that the shortest possible embedding of G g into a random metric has length at least O(g 4/3−ε ) is at least 1 − ε/2. In particular, since there are O(g) edges in G g , some edge e g in this embedding must have length at least Ω(g 1/3−ε ). By Theorem 13, we can choose g large enough so that with probability at least 1 − ε 2 , the random hyperbolic metric has diameter O(log g). Hence, by the union bound, with probability 1 − ε both properties hold. Therefore, for every ε > 0, there exists some value g 0 such that for any g ≥ g 0 , in any embedding of G g , there exists an edge e g = (x, y) such that m (e g ) = Ω(g 1/3−ε ), but d m (x, y) ≤ diam(m) ≤ O(log g). This implies that e is not drawn by a shortest path. Similarly, subdividing each edge O(g 1/3−ε ) times will run into the same issue. This concludes the proof.  Our approach to prove Theorem 4 is to cut the surface S g with a hexagonal decomposition ∆, so that every edge of G is cut O(g) times by this decomposition ∆. The construction to do this is a slight modification of the octagonal decompositions provided byÉ. Colin de Verdière and Erickson [4, Theorem 3.1]. Each of the hexagons is then endowed with a specific hyperbolic metric m H , and pasting these together yields the hyperbolic metric m on S g . The hyperbolic metric m H is chosen so that the hexagons are convex, i.e., the shortest paths between points of a hexagon stay within this hexagon. Therefore, there only remains to embed the graph G cut along ∆, separately in every hexagon with shortest paths. To do this, we use a variant of a theorem of Y. Colin de Verdière [7] which generalizes Tutte's barycentric method to metrics of nonpositive curvature. b. a. c. Figure 7: a. An octagonal decomposition b. A hexagonal decomposition c. How to add one closed curve to upgrade an octagonal decomposition to a hexagonal decomposition Hexagonal decompositions. A hexagonal decomposition, respectively an octagonal decomposition of S g is an arrangement of closed curves on S g that is homeomorphic to the one pictured in Figure 7.b., respectively Figure 7.a. In particular, every vertex has degree four and every face has six sides, respectively eight sides.
Octagonal decompositions were introduced byÉ. Colin de Verdière and Erickson [4] where they showed how to compute one that does not cross the edges of an embedded graph too many times. We restate their theorem in our language.
Theorem 15 ([4, Theorem 3.1]). Let G be a graph embedded in a surface S g for g ≥ 2. There exists an octagonal decomposition Γ of S g such that each edge of G crosses each closed curve of Γ a constant number of times.
We observe that this octagonal decomposition can be upgraded to a hexagonal decomposition that still does not cross G too much: Corollary 16. Let G be a graph embedded in a surface S g . There exists a hexagonal decomposition ∆ of S g such that each edge of G crosses each closed curve of ∆ a constant number of times, except for maybe one closed curve which is allowed to cross each edge of G at most O(g) times. In particular, the number of crossing between every edge of G and ∆ is O(g).
Proof. The decomposition ∆ is simply obtained by taking the decomposition Γ and adding a single curve that follows closely a concatenation of O(g) subpaths of curves of Γ, see Figure 7c. The resulting arrangement of curves has the topology of a hexagonal decomposition, and the bounds on the number of crossings results directly from the construction.
Remark: We remark thatÉ. Colin de Verdière and Erickson [4] actually build a decomposition with hexagonal faces before discarding cycles to get an octagonal decomposition. However, their decomposition is not homeomorphic to our notion of hexagonal decomposition, and would be less adapted for the proof of Theorem 4 since some of their hexagonal faces are glued to themselves (in particular Proposition 17 would not hold).
The hyperbolic metric. We first endow each hexagon of the hexagonal decomposition with the hyperbolic metric m H of an equilateral right-angled hyperbolic hexagon. Since the hexagons have right angles and the vertices of a hexagonal decomposition have degree 4, this metric can be safely pasted between hexagons to endow S g with a hyperbolic metric m. The main property of this metric that we will use is the following one: Figure 8: The intersection graph of the hexagonal decomposition and the two involutions: σ 1 is the reflection across the horizontal plane pictured in gray, and σ 2 swaps every pair of triply-linked adjacent vertices (pictured by disks and stars).
x y x y Figure 9: Top: The surface is cut into four quadrants, and Q 1 is pictured here with a darker color. Bottom: If γ enters another hexagon than H, it can be shortcut.
Proposition 17. Every hexagon H, viewed as a subset of S g endowed with m, is convex, i.e., every path between x, y ∈ H that is a shortest path in H is also a shortest path in S g .
Proof. We will prove that for any two points x, y ∈ H, there exists a shortest path (in S g ) that is entirely contained in H. The proof relies on an exchange argument based on the symmetries of the hexagonal decomposition. We rely on two involutions which we first define on the intersection graph 3 of the hexagonal decomposition. This graph should be thought of as embedded on the surface of a g × 1 × 1 rectangular block (see Figure 8), and • The map σ 1 is the reflection across the mid-plane of the top and bottom facets, sending vertices above that plane to the corresponding ones below.
• The map σ 2 is the reflection accross the mid-plane of the front and back facets. Equivalently, it swaps each hexagon with its neighbor in the original octagonal decomposition.
Since all the hexagons are isometric, these two maps induce isometric involutions of the surface S g endowed with m, which we also denote by σ 1 and σ 2 . They allow us to cut S g into four quadrants, each of them being a linear concatenation of hyperbolic hexagons: the first one, which we denote by Q 1 is pictured in Figure 9, and we obtain Q 2 , Q 3 and Q 4 by applying respectively σ 1 , σ 2 and σ 1 σ 2 to Q 1 . Now, let x and y be two points in a hexagon H, which we assume without loss of generality (by applying σ 1 and/or σ 2 ) to be in Q 1 , and let γ be a shortest path between x and y. This path γ may wander out of H, but in this case we show that there is another shortest path between x and y that is entirely contained within H. For every maximal subpath α of γ in the interior of Q 3 ∪ Q 4 , we reflect α into Q 1 ∪ Q 2 using σ 2 . This results in a new path γ between x and y, with the same length as γ (since σ 1 is an isometry) that is entirely contained in Q 1 ∪ Q 2 . Then, for every maximal subpath α of γ in the interior of Q 2 , we reflect α into Q 1 by applying σ 1 , which yields a path γ between x and y entirely contained within Q 1 . Since γ has the same length as γ, it is also a shortest path.
At this stage, we claim that γ is actually contained in H. Indeed, if it were not, since Q 1 is a linear concatenation of hexagons, there would be another hexagon H that contains a subpath α of γ such that the two endpoints e 1 and e 2 of α lie on the same side s of H , and α ⊂ s. But then γ cannot be a shortest path, since this subpath α could be shortcut by following the side s between e 1 and e 2 instead of entering H , see Figure 9.
Thus we have found a shortest path between x and y that is entirely contained within H, which proves the proposition.
Finishing the proof. We prove in this paragraph how to reembed a graph embedded in a hexagon so that its edges are shortest paths. This allows us to finish the proof. Let us postpone the proof of this theorem for now, and show how to conclude the proof of Theorem 4. We first show how to upgrade a graph embedded in a disk to a triangulation.
Lemma 19. For any graph G embedded in a disk without dividing edges, there exists a triangulation G of the disk that contains G as a subgraph and that does not contain any dividing edges.
Proof. For every face F of G, we start by adding a vertex in F and edges connecting it to the vertices adjacent to the face. This does not add loops or dividing edges, but may add multiple edges if one vertex occurs multiple times on the boundary of a face. These are taken care of by subdividing them once again and triangulating.
All the pieces are now in place for the proof of Theorem 4.
Proof of Theorem 4. By Corollary 16, one can embed G into S g such that every edge of G is cut O(g) times by the hexagonal decomposition ∆. This defines a graph G = i G i such that each of the graphs G i is embedded in a single hexagon and G is obtained from G by subdividing every edge O(g) times. If there are dividing edges in G i , they can be removed by subdividing the edge once. By Lemma 19, one can upgrade all the G i to triangulations. We can then apply Theorem 18 in each of the hexagons separately, yielding embeddings with shortest paths. Since the vertices on the boundary did not move during the reembedding, this defines an embedding of G into S g . Since H is simply connected and m H is hyperbolic, there is a unique geodesic connecting any two points, and this geodesic is a shortest path. Therefore the edges of G are shortest paths in H. By Proposition 17, each edge of G is also a shortest path in S g . Therefore each edge of G is embedded as a concatenation of O(g) shortest paths.
We note that by subdividing each edge once more, the shortest paths we obtain are unique.
The proof of Theorem 18 is obtained in a spirit similar to the proof of the one of the celebrated spring theorem of Tutte [35]. However, there are two main differences which prevent us from directly appealing to the literature: on the one hand the metric is not Euclidean but hyperbolic, and on the other hand the boundary of the input polygon is not strictly convex, since there may be multiple vertices of G on a geodesic boundary of H. The hypothesis on dividing edges is tailored to circumvent the second issue, and in a Euclidean setting it was proved by Floater [12] that the corresponding embedding theorem holds. Regarding the first issue, Y. Colin de Verdière stated a Tutte embedding theorem [7,Theorem 3] for the hyperbolic setting with strictly convex boundary, yet he actually did not provide a proof for it. In Appendix A we show how to prove Theorem 18 in the generality that we need following the ideas laid out by Y. Colin de Verdière in the rest of his article [7]. This concludes the proof of Theorem 4.
Finally, we remark that this proof technique provides an alternative proof of Negami's Theorem [25] for orientable surfaces. If G 1 and G 2 are two graphs embedded on the orientable surface of genus g, a crude application of Theorem 4 shows that one can reembed both graphs with a homeomorphism such that each edge is realized as a concatenation of O(g) shortest paths for our hyperbolic metric. Since hyperbolic shortest paths in general position cross at most once, this gives embeddings of G 1 and G 2 such that there are O(g 2 ) crossings between each edge of G 1 and each edge of G 2 . Negami proved that O(g) crossings are actually enough, and a deeper look at our construction also achieves this better bound: it is easy to see that in our reembeddings, each edge is actually cut into O(1) subedges realized as shortest paths in each hexagon. Since there are O(g) hexagons, there are in total O(g) crossings between each edge of G 1 and G 2 , which yields the following: Corollary 20. There exists an absolute constant c > 0 such that if S g is an orientable surface of genus g, for any two embedded graphs G 1 , G 2 → S g there exists a homeomorphism h : S g → S g such that cr(h(G 1 ), G 2 ) ≤ cg|E(G 1 )| · |E(G 2 )| A Tutte's embedding theorem in a hyperbolic setting In this section, we explain the proof of the subsequent theorem, following the arguments of Y. Colin de Verdière [7].
Theorem 18. Let G be a graph embedded as a triangulation in a hyperbolic hexagon H endowed with the metric m H . If there are no dividing edges in G, i.e., edges between two non-adjacent vertices on the boundary of H, then G can be embedded with geodesics, with the vertices on the boundary of H in the same positions as in the initial embedding.
As announced, the proof follows from a spring-like construction, i.e. we think of the edges of the graph G as springs with some arbitrary stiffness, the vertices which are not on the boundary are allowed to move and we prove that the equilibrium state for this physical system is an embedding of the graph.
For an embedding ϕ : G → H, denote by e ij the map [0, 1] → H representing the edge (i, j). Starting with an embedding ϕ 0 : G → H and given assignments c i,j : E(G) → R + , we are interested in the map ϕ : G → H minimizing the energy functional with fixed vertices on the boundary of H. This is the equilibrium state of the spring system with the c i,j coefficients specifying the stiffness of the springs. We claim that ϕ is an embedding such that the edges are geodesics.
Step 1: Existence. The existence of ϕ follows from classical compactness considerations, since an Arzelà-Ascoli argument proves the compactness of sets with bounded energy. Then an extremum of E ϕ corresponds to a ϕ where all the arcs e i,j are geodesics. Furthermore, every vertex ϕ(x) which is not on the boundary lies in the strict hyperbolic convex hull of its neighbors which are not mapped to the same point.
Step 2: Curvature considerations. Since ϕ 0 provides an embedding of G into H, G can be seen as a topological subspace of H. The corresponding simplicial complex will be denoted by X (it is of course homeomorphic to H) and its set of vertices, edges and triangles by V , E, and T . By extending ϕ separately with a local homeomorphism in the interior of each non-degenerate triangle, we can extend it into a map Φ : X → H agreeing with ϕ on G. Now, the map Φ : X → H provides values for the angles of the non-degenerate triangles in X. For degenerate triangles, values of the angles are taken arbitrarily so that they sum to π (therefore morally their hyperbolic area is zero). For an interior vertex v, let us define the curvature K(v) = 2π − i α i v , where α i v are the angles adjacent to v. For a vertex v on the interior of a geodesic boundary, we define it by K(v) = π − i α i v , and on the six vertices of H, we take it to be K(v) = π/2 − i α i v . The area of a geodesic hyperbolic triangle is π minus the sum of its angles. Summing over all the triangles of Φ(X), we obtain |T |π − v i α i v = t∈T Area(t). With Euler's formula and double counting, this gives t∈T Area(t) = π + v K(v). Since the boundary is fixed, Φ has degree one and is thus surjective, therefore the sum of the areas of the triangles is at least the area of the hexagon, which is π since it is right-angled. Therefore v K(v) ≥ 0.
Step 3: Punctual degeneracies. In this step we investigate which subcomplexes of X can be mapped to a single point. We show that no triangle can be mapped to a single point, and that a set of edges mapped to a single point forms a path subgraph in G.
Let X 1 be a maximal connected subcomplex of X which is mapped to a point x by Φ. This subcomplex has to be simply connected, otherwise the region inside could be mapped to x as well which would reduce the value of E ϕ . Since the boundary edges are fixed by ϕ, X 1 does not contain any edge on the boundary or triangle adjacent to the boundary.
For every vertex v in Φ −1 (x), Φ(v) = ϕ(v) lies in the strict convex hull of its neighbors which are not mapped to x, as was observed in Step 1. Therefore the angles of the non-degenerate triangles adjacent to v sum up to at least 2π. Indeed the angular opening at ϕ(v) has to be at least π by the convexity hypothesis, but if a map S 1 → S 1 is not surjective then every point in the image has at least two pre-images, in which case this angular opening of at least π amounts to at least 2π in the sum of angles around v. This shows that K(x) := v∈Φ −1 (x) K(v) is nonpositive. Since the boundary edges are fixed, we also have K(v) ≤ 0 for the vertices on the boundary.
Summing over all the values of x, we obtain that v K(v) ≤ 0, and thus this sum is zero by the previous paragraph, and each of the K(x) is also zero.
From that we infer that X 1 contains no triangle: if it did, there would be at least 3 preimages of x for which the angles of the adjacent non-degenerate triangles would sum up to at least 2π. Summing them into K(x) we would obtain a nonzero value. Similarly, X 1 can only be a linear subgraph of G, and every triangle adjacent to a X 1 not reduced to a point is degenerate.
Step 4: Linear degeneracies. Now that we showed that triangles can not be mapped to points, we show that triangles are not mapped to lines either, or equivalently that edges are not mapped to points.
Let X 2 be a maximal connected subcomplex of X such that the image of the triangles of X 2 by ϕ are degenerate. Let us assume that X 2 is non-empty. Then the image Φ(X 2 ) is an arc of a geodesic of H: indeed if there was a broken line in Φ(X 2 ), around the breaking points there would be non-degenerate triangles adjacent to a X 1 not reduced to a point, which is absurd by the previous paragraph.
If this geodesic is not a boundary geodesic of H, two of the points on the boundary of X 2 are mapped to the endpoints of the arc of geodesic, and all the other vertices have their adjacent edges within X 2 because of the convexity condition. Therefore, there must be two arcs connecting the two boundary points, as in the top of Figure 10, which is impossible in the ∂X ∂H ϕ ϕ Figure 10: Any triangulation inducing a linear degeneracy would require either multiple edges (top) or a dividing edge (bottom).
simplicial complex X.
If this geodesic is on the boundary of H, then by the same convexity argument, two vertices of ∂X must map to the endpoints of this arc of geodesic, and the other vertices have all their edges within X 2 . Therefore there is a dividing edge connecting these two vertices, as in the bottom of Figure 10, which is a contradiction.
Step 5: Conclusion. Since X 2 is empty, no triangle in the image of Φ is degenerate. Furthermore, all the X 1 are reduced to a single point and thus K(v) is zero for all the vertices v. The only remaining possible pathology is if all the triangles adjacent to a non-boundary vertex v are mapped to a half-plane around Φ(v). By the convexity constraint, this can only happen if the edges adjacent to v are aligned, but this would yield degenerate triangles. Therefore Φ is a local homeomorphism of degree 1, hence it is a global homeomorphism and ϕ is an embedding.