Connectivity of Triangulation Flip Graphs in the Plane

Given a finite point set P in general position in the plane, a full triangulation of P is a maximal straight-line embedded plane graph on P. A partial triangulation of P is a full triangulation of some subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P'$$\end{document}P′ of P containing all extreme points in P. A bistellar flip on a partial triangulation either flips an edge (called edge flip), removes a non-extreme point of degree 3, or adds a point in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P \setminus P'$$\end{document}P\P′ as vertex of degree 3. The bistellar flip graph has all partial triangulations as vertices, and a pair of partial triangulations is adjacent if they can be obtained from one another by a bistellar flip. The edge flip graph is defined with full triangulations as vertices, and edge flips determining the adjacencies. Lawson showed in the early seventies that these graphs are connected. The goal of this paper is to investigate the structure of these graphs, with emphasis on their vertex connectivity. For sets P of n points in the plane in general position, we show that the edge flip graph is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lceil {n}/{2}-2\rceil $$\end{document}⌈n/2-2⌉-vertex connected, and the bistellar flip graph is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(n-3)$$\end{document}(n-3)-vertex connected; both results are tight. The latter bound matches the situation for the subfamily of regular triangulations (i.e., partial triangulations obtained by lifting the points to 3-space and projecting back the lower convex hull), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(n-3)$$\end{document}(n-3)-vertex connectivity has been known since the late eighties through the secondary polytope due to Gelfand, Kapranov, & Zelevinsky and Balinski’s Theorem. For the edge flip-graph, we additionally show that the vertex connectivity is at least as large as (and hence equal to) the minimum degree (i.e., the minimum number of flippable edges in any full triangulation), provided that n is large enough. Our methods also yield several other results: (i) The edge flip graph can be covered by graphs of polytopes of dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lceil {n}/{2} -2\rceil $$\end{document}⌈n/2-2⌉ (products of associahedra) and the bistellar flip graph can be covered by graphs of polytopes of dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n-3$$\end{document}n-3 (products of secondary polytopes). (ii) A partial triangulation is regular, if it has distance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n-3$$\end{document}n-3 in the Hasse diagram of the partial order of partial subdivisions from the trivial subdivision. (iii) All partial triangulations of a point set are regular iff the partial order of partial subdivisions has height \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n-3$$\end{document}n-3. (iv) There are arbitrarily large sets P with non-regular partial triangulations and such that every proper subset has only regular triangulations, i.e., there are no small certificates for the existence of non-regular triangulations.


Introduction
Triangulations of point sets play a role in many areas including mathematics, numerics, computer science, and processing of geographic data, [4,11,22,24]. A natural way to provide structure to the set of all triangulations is to consider a graph, called flip graph, with the triangulations as vertices and with pairs of triangulations adjacent if they can be obtained from each other by a minimal local change, called a flip (see below for the precise definition). One of the first and most prominent results on flip graphs of planar points sets (for edge flips), proved by Lawson [21] in 1972, is that they are connected. The corresponding question of connectedness of flip graphs in higher dimensions remained a mystery until Santos [33] showed, in 2000, that in dimension 5 and higher, there exist point sets for which the graph (for bistellar flips) is not connected. The question is still open in dimension 3 and 4.
Here, we concentrate on point sets in general position in the plane and investigate "how" connected the flip graphs are, i.e., determine the largest k (in terms of n, the size of the underlying point set) such that k-vertex connectivity holds. Moreover, we supply some structural results for the flip graph.
The preceding discussion swept under the rug that there are several types of triangulations, most prominently full triangulations (as mostly considered in computational geometry), partial triangulations (as primarily considered in discrete geometry), and regular triangulations (also called coherent triangulations or weighted Delaunay tri-angulations). We will address the vertex connectivity for full and partial triangulations, and we will discuss some implications for regular triangulations, for which the vertex connectivity of the flip graph has been known since the late eighties via the secondary polytope introduced by Gelfand et al. [15] (see [24,Corollary 5.3.2]).
Let us first supply the basic definitions and then discuss the results in more detail.
Notation We let N denote the set of positive integers and we let N 0 := N ∪ {0}. We write A ⊕ B for the symmetric difference of sets A and B. Definition 1.1 (point set). Throughout this paper we let P denote a finite planar point set in general position (i.e., no three points on a line) with n ≥ 3 points. The set of extreme points of P (i.e., the vertices of the convex hull of P) is denoted by xtr P, and we let P • := P \ xtr P denote the set of inner (i.e., non-extreme) points in P. We consistently use the notation h = h(P) := |xtr P| and n • = n • (P) := |P • | = n − h, and we let E hull = E hull (P) ⊆ P 2 denote the set of edges of the convex hull of P, |E hull | = h. Definition 1.2 (plane graph). For graphs G = (P , E), P ⊆ P, E ⊆ P 2 , on P we often identify edges { p, q} ∈ E with their corresponding straight line segments pq. We let VG := P and EG := E. A graph G on P is plane if no two straight line segments corresponding to edges in EG cross (i.e., they are disjoint except for possibly sharing an endpoint). For G plane, the bounded connected components of the complement of the union of the edges are called regions, the set of regions of G is denoted by RG.
Note that regions are open sets. Note also that isolated points in a plane graph are ignored in the definition of regions.

Definition 1.3 (full, partial, regular triangulation).
(a) A full triangulation of P is a maximal plane graph T = (P, E). (b) A partial triangulation of P is a full triangulation T = (P , E) with xtr P ⊆ P ⊆ P (hence E hull ⊆ ET ). (c) A regular triangulation of P is a triangulation obtained by projecting (back to R 2 ) the edges of the lower convex hull of a generic lifting of P to R 3 (i.e., we add third coordinates to the points in P so that no four lifted points are coplanar).
Points in V • T := P • ∩ VT are called inner points of T and points in P • \ V • T are called skipped in T (clearly, for a full triangulation V • T = P • and no points are skipped). Edges in E • T := ET \ E hull are called inner edges and edges in E hull are called boundary edges. T full (P), T part (P), and T reg (P) will denote the set of all full, partial, and regular triangulations of P, respectively.
Every full and every regular triangulation of P is also a partial triangulation of P. If P is in convex position (i.e., xtr P = P), all three notions coincide. It is well known, that there are point sets with non-regular triangulations, [24], see Sect. 10

Results-Full Triangulations
The edge flip graph is the subgraph of the bistellar flip graph induced by T full (P) (see Figs. 2 and 3). Lawson [21] showed that the edge flip graph is connected. A trivial upper bound for the vertex connectivity (see Definition 2.1) of any graph is the minimum vertex degree of the graph; for the edge flip graph this is the minimum number of flippable edges in any triangulation of P. This number has been investigated by Hurtado et al. in 1999 [20, Theorem 4.1].
Theorem 1.6 [20] Any T ∈ T full (P) has at least n/2 −2 flippable edges. This bound is tight for all 3 ≤ n ∈ N (i.e., for every n there is a set P, |P| = n, with a triangulation with exactly n/2 − 2 flippable edges).
We provide a lower bound on the vertex connectivity of the edge flip graph that matches this lower bound of n/2 − 2 for the minimum vertex degree, actually in the refined form max { n/2 − 2 , h − 3}, see Hoffmann et al. [18]. Note, however, that there are sets of points where all full triangulations allow more than max { n/2 − 2 , h − 3} edge flips. (Consider, e.g., the top left point set in Fig. 3: n = 6, h = 4, thus max { n/2 − 2 , h − 3} = 1, but the edge flip graph has minimum vertex degree 2.) We show that, for P large enough, the minimum vertex degree always determines the vertex connectivity of the edge flip graph. Obviously, for n ≥ n 0 , (i) implies (ii). In fact, we do not know whether the restriction "n large enough" is required in (i). Still, apart from covering the range to n 0 , we consider the proof for (ii) of independent interest, since it provides some extra insight to the structure of the edge flip graph via so-called subdivisions, and it is an introduction to the proof for the bistellar flip graph. Although some concrete value of n 0 could be easily extracted from the proof, getting a reasonable value is lengthy and tedious, and it is not clear where to stop. Moreover, we believe that the min-degree bound holds for all point sets. This is supported by a recent computer-assisted investigation for small n in [12] with the help of the Graz data base of planar order types [1], with the result that for n up to 10, the vertex connectivity is determined by the smallest degree of the edge flip graph. This endeavour was nontrivial, given the huge number of order types (14 309 547 for ten points), and for each order type the large number of triangulations (ranging from 250 to 4 550 for ten points).

Results-Partial Triangulations
The bistellar flip graph is connected, as it follows easily from the connectedness of the edge flip graph, see [24,Sect. 3.4.1]. Here is the counterpart of Theorem 1. Theorem 1.8 [25] Any T ∈ T part (P) allows at least n − 3 flips. This bound is tight for all P.
Again, we show that the vertex connectivity equals the minimum degree. This answers (for points in general position) a question mentioned by De Loera, Rambau, and Santos in 2010 [24,Exercise 3.23], and by Lee and Santos in 2017 [22, p. 442]. Before we mention further results, we provide some context. Along the way, we encounter some tools and provide intuition relevant later in the paper.

Context-Convex Position, Associahedron, and Balinski's Theorem
Suppose P is in convex position. Then T full (P) = T part (P) = T reg (P) is the set of triangulations of a convex n-gon whose study goes back to Euler, with one of the first appearances of the Catalan Numbers. There is an (n−3)-dimensional convex polytope, the associahedron, whose vertices correspond to the triangulations of a convex n-gon, and whose edges correspond to edge flips between these triangulations, see [9] for a historical account. That is, the 1-skeleton (graph) of this polytope is isomorphic to the flip graph of P, see Fig. 2 (left) for n = 5 and Fig. 4 for n = 6. Here Balinski's Theorem from 1961 comes into play. Theorem 1.10 (Balinski's Theorem [3]). The 1-skeleton of a convex d-dimensional polytope is at least d-vertex connected.
Thus we can conclude that the flip graph of n points in convex position is (n − 3)vertex connected. The face structure of the associahedron is easily explained via so-called subdivisions (also known as polyhedral subdivisions or convex decompositions), which will feature prominently in our arguments. Specialized to the setting of convex position, a subdivision S is a plane graph on P with all boundary edges present, and some diagonals. We identify every such subdivision with the set of all its possible completions to a triangulation (which we will call refinements). Figure 5 illustrates on two examples for n = 6, that if S misses i edges towards a triangulation, then its completions to triangulations correspond to the vertices of an i-face of the associahedron; in particular, no edge missing (a triangulation) corresponds to a vertex, and one edge missing corresponds to an edge (a flip). This correspondence will guide our intuition, even in more general settings in which there is no polytope in the background.

Context-Regular Triangulations and Secondary Polytope
All examples of bistellar flip graphs we have seen so far-Figs. 2 and 3, and sets in convex position-are graphs of polytopes, which is not true in general. (Up to six points, there is only one order type which has partial triangulations that are not .) The reason is simply that for the underlying point sets in these examples all partial triangulations are regular. Specifically, there is the following generalization of the associahedron, a highlight of the topic of triangulations from the late eighties due to Gelfand et al. [15] (see [24,Thm. 5.1.9]) (here specialized to points in the plane). Theorem 1.11 (secondary polytope [15]). For every set P of n points in general position in the plane, there is an (n − 3)-dimensional polytope -poly(P), the secondary polytope of P, whose 1-skeleton is isomorphic to the bistellar flip graph of regular triangulations of P.
Again, with the help of Balinski's Theorem (Theorem 1.10), we immediately get: The bistellar flip graph of regular triangulations of P is (n − 3)-vertex connected. Our Theorem 1.9 is the corresponding result for the bistellar flip graph of all partial triangulations. Note, however, that it is not a generalization, since it does not imply the result for regular triangulations.

Context-Simplicial Complex of Plane Graphs and Its Dual Flip Complex
A polytopal representation of the edge flip graph or the bistellar flip graph of all partial triangulations is not known. In fact, a clean result as with the secondary polytope is not possible (see [24] for details). However, there is a construction, due to Orden and Santos [31], of a simple high-dimensional polytope whose 1-skeleton represents all socalled pseudo-triangulations of a planar point set and flips between them, following an earlier construction by Rote et al. [32] of a corresponding polytope for pointed pseudo-triangulations.
Moreover, the edge flip graph of full triangulations of a planar point set does form the 1-skeleton of a closely related higher-dimensional structure, the flip complex, first described by Orden and Santos [31] and later rediscovered by Lubiw et al. [26]. The flip complex is not a polytope, but it is a polytopal complex (informally, a collection of convex polytopes that intersect only in common faces) with a particularly simple topology (it is homotopy equivalent to a ball). For the reader familiar with these notions, we briefly review how our findings fit into this context (for a quick reference for the relevant terminology from topology theory, see, e.g., [5]). This is not essential as a tool for our proofs and the rest of the paper can be followed without this context. Rather, our results shed some extra light on these structures considered in the literature.
Following [26], let us consider i.e., these are the edge sets of plane graphs on P. Since this family of sets is closed under taking subsets, it is a simplicial complex, and we call its elements faces of T. By definition, the dimension of a face F ∈ T is dim F := |F|−1. In [26], it is shown that T is shellable and homeomorphic to an (m −1)-dimensional ball (this also follows from the results in [31]). This fact was instrumental in the proof of the so-called orbit conjecture regarding flips in edge-labeled full triangulations formulated in [7] and proved in [26].
A face F of dimension m − 2 is a plane graph on P with exactly one edge missing towards a full triangulation. If F is obtained from a full triangulation by removal of a flippable edge, there are exactly two ways to complete it to a triangulation-we call F an interior (m − 2)-face; otherwise, there is exactly one way to do so-we call F a boundary (m − 2)-face. A face of any dimension is called a boundary face if it is contained in a boundary (m − 2)-face, and it is called an interior face, otherwise (hence, all facets are interior faces). We have the following. Theorem 1.12 [26,Proposition 3.7] F ∈ T is an interior face iff E hull ⊆ F, and all bounded regions of F are convex (in other words, iff (P, F) is what we will call a subdivision, see Definition 4.1 and also discussion in Sect. 1.3).
From the Coarsening Lemma 4.10 in Sect. 4, the following property of T is implied.

Theorem 1.13 Every interior face F of T contains an interior face of dimension
The flip graph can be seen as a structure dual to T, with the vertices of the flip graph corresponding to the facets of T, and two vertices adjacent in the flip graph, if the corresponding facets of T share (i.e., contain) a common (m −2)-face (which is clearly interior). This can be generalized to the flip complex X = T * , see [26], with each face of X (which is dual to an interior face of T) corresponding to a subdivision as we define it below (see Theorem 1.12 above). Each such face is a product of associahedra, and the vertices of such a face of the flip complex are the triangulations refining this subdivision (Definition 4.2). In terms of the flip complex, the Coarsening Lemma 4.10 and Theorem 1.13 can be restated as saying that every inclusion-maximal face of the flip complex is of dimension at least max {n/2 − 2, h − 3} (see also Theorem 6.1).

Results-Regular Triangulations
We study the (well-known, see [24]) partially ordered sets of full and partial subdivisions of P, respectively (see Definitions 4.1 and 7.1), in which triangulations are the minimal elements. We introduce the notions of slack of a subdivision (Definitions 4.5 and 7.2), perfect coarsening (Definition 8.1) and perfect coarsener (Definition 8.2), and we prove the so-called Coarsening Lemmas 4.10 and 8.6 (these can be considered extensions of Theorems 1.6 and 1.8). We consider these notions and lemmas our main contributions besides Theorems 1.7 and 1.9. Together with a sufficient condition for the regularity of partial triangulations and subdivisions (Theorem 10.1 and Regularity Preservation Lemma 10.10), these yield several other results on the structure of flip graphs. In particular, they allow us to settle, in the unexpected direction, a question by Francisco Santos (in personal communication, 2019) regarding the size of certificates for the existence of non-regular triangulations of a given point set in the plane. Theorem 1.14 For all n ∈ N there is a set of at least n points in general position in the plane, which has non-regular triangulations, and for which any proper subset has only regular triangulations (in other words, only full triangulations of the set can be non-regular).
This should be seen in contrast with the situation in higher dimensions. In large enough dimension, every point set with non-regular triangulations has a subset of bounded size (in the dimension) with non-regular triangulations. This holds since, (a) the vertices of any realization of the cyclic polytope with 12 vertices in dimension 8 has non-regular triangulations [23] (see [24,Sect. 5.5.2]), and (b) every large enough set of points in general position in R 8 has a subset of 12 points which are vertices of a cyclic polytope (this follows from Ramsey's Theorem [13,16,36]).

Approach
All our vertex connectivity bounds rely on a local variant of Menger's Theorem, the Local Menger Lemma 2.3. This lemma says that, assuming connectedness, in order to show k-vertex connectivity, it is enough to show k internally vertex-disjoint paths between any two vertices at distance 2. Then, in order to establish the min-degree bound of Theorem 1.7 (i), we explicitly construct the necessary paths between vertices at distance 2, i.e., triangulations T [e] and T [ f ].
For the bounds in Theorems 1.7 (ii) and 1.9, we look at the neighborhood of a triangulation T (which is in one-to-one correspondence with the flippable elements in T ), supplied with a compatibility relation between the flippable elements (for the edge flip graph, two flippable edges e and f are compatible, if e remains flippable after flipping f , Definition 3.6). We call this the link 1 of T , a structure motivated by the vertex figure of a vertex in a polytope, see [39, p. 54]: Recall that for a vertex v in a d-polytope P, its vertex figure is the (d − 1)-polytope P obtained by intersecting P υ Fig. 6 The vertex figure of a polytope, with the mapping of paths in the vertex figure to the 1-skeleton of the polytope indicated with a hyperplane that separates v from the remaining vertices of the polytope. Vertices of P correspond to edges of P incident to v, edges in the graph of P correspond to 2-faces of P incident to v. As indicated in Fig. 6, there is a natural way of mapping paths in the graph of P to paths in the graph of P. Using the Local Menger Lemma 2.3, this can be easily made an inductive proof of Balinski's Theorem (Theorem 1.10). We follow exactly this line of thought for flip graphs (where 4-and 5-cycles will play the role of 2-faces), except that we will not need induction: The link of a triangulation avoids 4-cycles in its complement, which turns out to directly yield sufficient vertex connectivity (Lemma 2.4).
That is, we borrow intuition from polytope theory, although we know that the edge flip graph and the bistellar flip graph are in general not graphs of polytopes. However, as we will see in Sects. 6 and 11.1, the edge flip graph and the bistellar flip graph can be covered by graphs of n/2 − 2 -and (n − 3)-polytopes, respectively.
It is perhaps worth emphasizing that the arguments we use to prove lower bounds for the vertex connectivity of the flip graphs (in particular, Lemma 2.3) assume the well-known fact that these graphs are always connected-we do not prove this. That is, our techniques will probably not help to address the open question whether flip graphs in dimensions 3 and 4 are connected. If at all, they might help to analyze the vertex connectivity of the connected components of flip graphs.

Paper Organization
In Sect. 2 we show the two lemmas on graph connectivity we mentioned already in Sect. 1.7: The Local Menger Lemma and a lemma about the vertex connectivity of graphs with no 4-cycle in their complement. These may be of independent interest, if only as exercises after teaching Menger's Theorem in class.
Then the paper splits in three parts, where the second and third part on partial triangulations are largely independent of the first part about full triangulations.
-Section 3 shows the min-degree bound of Theorem 1.7 (i). Section 4 prepares the proof of Theorem 1.7 (ii), which is presented in Sect. 5. We prove the Unoriented Edges Lemma 4.9, which captures the essence of and entails Theorems 1.6 (from [20]) and 1.8 (from [25]) above. It allows us to extend them further for our purposes. Moreover, along the way, we give a short proof of the bound for so-called simultaneously flippable edges by Souvaine et al. [35], and we indicate, how it may help getting insight on so-called k-holes in point sets [34] In Sect. 6 we briefly indicate polytopal substructures in the edge flip graph. -Sections 7-9 show the (n − 3)-bound in Theorem 1.9.
-Building on the tools developed in Sects. 7-9, in particular the Coarsening Lemma 8.6, Sect. 10 proves a nontrivial new sufficient condition for the regularity of partial triangulations (Theorem 10.1), which can be considered as a generalization of the regularity of stacked triangulations (i.e., triangulations obtained by successively adding points of degree 3 to a triangulation of xtr P). This will give us a number of implications to be presented in Sect. 11: Covering the bistellar flip graph by graphs of (n − 3)-polytopes, a characterization of point sets for which all partial triangulations are regular, and, finally, Theorem 1.14 on the size of certificates for the existence of non-regular triangulations in T part (P).
We conclude by discussing open problems in Sect. 12.

Graph Connectivity
We follow the graph theory books of Bollobás [6] and Diestel [10].
connected, has at least k + 1 vertices, and removing any set of at most k − 1 vertices (and their incident edges) leaves the graph connected.
Note that a graph is 1-vertex connected iff it is connected and has at least two vertices. Here is a classical result due to Menger from 1927 [29], see [6, Theorem III.5].
Theorem 2.2 (Menger's Theorem [29]). Let u and v be distinct nonadjacent vertices of a graph G. Then the minimal number of vertices separating u from v is equal to the maximal number of internally vertex-disjoint u-v paths.
Convention From now on, we will use vertex-disjoint short for "internally vertexdisjoint." We will need a local variant of Menger's Theorem.

Lemma 2.3 (Local Menger).
Let k ∈ N and let G be a simple undirected graph. Assume that G is connected. Then G is k-vertex connected iff G has at least k + 1 vertices and for any pair of vertices u and v at distance 2 there are k pairwise vertexdisjoint u-v paths.
Proof The direction (⇒) follows from Menger's Theorem 2.2. For (⇐), suppose that for any pair u and v of vertices at distance 2 there are k pairwise vertex-disjoint u-v paths. We show that no pair of vertices x and y can be separated by removal of a set V of at most k − 1 vertices (different from x and y). This is true for x and y adjacent. If x and y are not adjacent, let (x = x 0 , x 1 , . . . , x = y) be an x-y path which uses the minimal number of vertices in V , and among those, a shortest such path (hence . If no vertex on this path is in V we are done.
Otherwise, consider x i ∈ V , 0 < i < . We can replace the subpath ( by an x i−1 -x i+1 path using none of the vertices in V as internal vertices (x i−1 and x i+1 have distance 2 and hence such a path exists, since there are k vertex-disjoint x i−1 -x i+1 paths and |V | < k). We obtain an x-y walk with one less overlap with V , and we can turn this into an x-y path with less overlap with V ; contradiction. (A walk is a path with repetitions of vertices allowed.) For the proofs of Theorems 1.7(ii) and 1.9, here is a special property of a graph that guarantees that the minimum vertex degree determines exactly the vertex connectivity. As briefly discussed in Sect. 1.7, we will apply this lemma not directly to the flip graph, but rather to the link, a neighborhood structure of a triangulation.

Lemma 2.4
Let G be a graph with its complement having no cycle of length 4, i.e., for any sequence Proof It suffices to show that for any two distinct nonadjacent vertices u and v there are δ vertex-disjoint u-v paths. Let z 1 , z 2 , . . . , z be the set of vertices adjacent both to u and to v. This gives u-v paths (u, Otherwise, there are vertices x +1 , . . . , x δ adjacent to u and not to y, and there are vertices y +1 , . . . , y δ adjacent to v and not to u. For < j ≤ δ consider the sequence (u, y j , x j , v): None of {u, y j }, {x j , v}, and {u, v} are edges in G, hence {y j , x j } ∈ EG and (u, x j , y j , v) is a u-v path. That is, we have found another δ − paths connecting u to v. All paths constructed are easily seen to be vertex-disjoint.

Min-Degree Bound for Full Triangulations
Convention In Sects. 3-6 we will mostly use triangulation short for "full triangulation." In this section we prove Theorem 1.7 (i) (Sect. 3.5) via the following lemma (and the Local Menger Lemma 2.3). For the proof of Lemma 3.1 (see Sect. 3.4), we need a better understanding of how flippable edges interact. This will exhibit short cycles, more concretely, 4-or 5-cycles in the edge flip graph (called elementary cycles in [26]). Subpaths of these short cycles will be the building blocks for the T [e]-T [ f ] paths as claimed in Lemma 3.1.

Two Consecutive Flips
otherwise.

Interplay of Two Flippable Edges
Here comes an essential lemma about the interplay of two flippable edges in a triangulation.

Lemma 3.5 Let e and f be two distinct edges both of which are flippable in T ∈ T full (P). Then e is flippable in T
Proof We distinguish three cases, depending on the shape of terr e ∪ terr f . This can be composed of two disjoint quadrilaterals (recall that territories are open sets), or it is a pentagon, if e and f are incident to a common region of T . In short, the lemma states that, provided e and f are flippable in T , and T [ f , e] are defined, they may be equal or not.
With this in mind, we give the following definition (Fig. 8).

Definition 3.6
Let T ∈ T full (P) and let e and f be flippable edges in ET with e = f .

Proof of Lemma 3.1
Proof , with these paths required to be vertex-disjoint. For this we distinguish four cases, depending on g. , , There are at most eight such edges g, since for this to happen, g must be an edge of terr T e = terr T [e] e or of terr T f . For such an edge g choose an edge . Because of our degree condition, g * must exist. Note that g may equal g * , if g is flippable in T [ f ] but not in T . Now we choose edges x and y flippable in T with the three sets terr T x, terr T y, and terr T e ∪ terr T f ∪ terr T g ∪ terr T g * pairwise disjoint. These conditions allow for the following path: Every g in this final category is paired up with a different g * and a different pair {x, y} is chosen for building such a T [e]-T [ f ] path. P large enough will allow us to do so, by Theorem 1.6 and since we have to deal only with at most a constant (at most eight) such cases.
while previous cases gave signatures (again of internal vertices) (c) g compatible with e and f in T The vertex-disjointness of these paths can be readily concluded from these sequences and the proof is completed. As mentioned before, while "P large enough" is essential in Lemma 3.1, we do not know whether it can be removed in Theorem 1.7 (i).

Coarsening Full Subdivisions
In preparation of the proof of Theorem 1.7 (ii), which will be presented in Sect. 5, we introduce full subdivisions of a point set, which form a partially ordered set under a relation we call refinement, and slack, a parameter of subdivisions. As mentioned when discussing the flip complex in Sect. 1.5 above, subdivisions correspond to the faces of the flip-complex. The refinement partial order corresponds to inclusion of one face in another, and the slack of a subdivision is equal to the dimension of the corresponding face of the flip complex. This will allow us to prove a lower bound on how many edges are compatible with a given flippable edge in a triangulation.

Definition 4.1 (full subdivision).
A full subdivision S of P is a connected plane graph with VS = P and E hull ⊆ ES, and all regions of S convex.
Convention In Sect. 4 up to Sect. 6 we will use subdivision short for "full subdivision."  (In the literature, triangulations in T full S are sometimes also called S-constrained triangulations of P.) Obviously, is a partial order on subdivisions of P with the triangulations the minimal elements. Note that if S 1 S 2 , then every region of S 1 is contained in some region of S 2 , hence the name "refinement." Here is a notion that allows us to identify edges in a subdivision that can be individually removed (not necessarily simultaneously removed) in order to obtain a coarsening.

Definition 4.3 (locked)
. In a graph G on P, an edge e ∈ EG is locked at endpoint p if the angle obtained at p (between the edges adjacent to e in radial order around p) after removal of e is at least π (in particular, if p has degree 1 or 2 in G). See Fig. 11.
An edge in a triangulation is flippable iff it is not locked at any endpoint (in a triangulation, an edge can be locked at at most one of its endpoints). If S is a subdivision, then removal of an edge e in S gives a subdivision iff e is locked at none of its endpoints in S (here, for a subdivision, an edge can be locked at both of its endpoints). Hence, the -maximal subdvisions are those with all of their edges locked. Here are some simple fundamental properties of locking.

Observation 4.4 Let G be a graph on P and p ∈ VG.
(i) Any two edges locked at p must be consecutive in the radial order around p.
(ii) There are at most two edges locked at p, unless p has degree 3 in G.
(iii) If G is a subdivision and p ∈ P • is of degree 3, then the three edges incident to p are locked at p.

Definition 4.5 (slack, refined slack).
Let r be a region of a subdivision S, a convex set bounded by a k-gon, k ≥ 3. Then we define the slack, sl r , of r as k − 3, and the refined slack, sl * r , of r as sl r /2 . See Fig. 11. The slack, sl S, of subdivision S is the sum of the slacks of its regions. The refined slack, sl * S, is the sum of the refined slacks of its regions.
Note that k − 3 is the number of edges it takes to triangulate a k-gon. Hence, the slack of a subdivision is the number of edges which have to be added towards a triangulation. The refined slack (k − 3)/2 = k/2 − 1 of a k-gon cannot be easily motivated at this point. It will become apparent only in the proof of Lemma 4.9 (ii) where it is used as the excess (relative to a triangle) of the largest independent set of a k-cycle. We need the following well-known facts on the number of edges and regions of triangulations.

Unoriented Edges Lemma
Definition 4.8 (partially oriented graph, well-oriented graph). Let G be a graph G on P with some (not necessarily all) of its edges oriented to one (not both) of its endpoints and some edges unoriented; we call this a partially oriented graph. G is well oriented if (a) no edge is directed to a point in xtr P, and (b) if edges e and f are directed to a common point p, then e and f have to be consecutive in radial order around p.
Note right away that condition (b) for a well-oriented graph shows that no point in such a graph can have indegree more than 3, and indegree 3 is possible only for points of degree 3. Moreover, given a subdivision, suppose we orient each locked inner edge to one locking endpoint (and we leave all other edges, in particular the boundary edges, unoriented), then Observation 4.4 shows that this is a well-oriented graph (hence, S in Fig. 11, center, serves as an example of a well-oriented graph). This is how we will later employ the following lemma.
(ii) Suppose S is well oriented. Then the number of unoriented inner edges is at least . Since S has 3C 3 + 2C 2 + C 1 oriented edges, the number of unoriented inner edges is (ii) Let every inner point charge 1 to each region incident to it that lies between two incoming edges. The overall charge made is exactly 3C 3 +C 2 (if the indegree is 3, the degree is 3; if the indegree is 2, the two incoming edges are consecutive). While each triangular region can be charged at most once, other regions can be multiply charged: A region with slack d, a (d + 3)-gon, is charged at most times (vertices charging a region r cannot be consecutive along the boundary of r ). Hence, if is the number of regions charged at all, the overall charge made to these regions is at most . That is, We plug this bound on C 3 into the number of unoriented edges derived in (4) above: The left bound in (ii) on the number of unoriented follows readily. Moreover, d/2 ≤ d for d ∈ N 0 , hence D * ≤ D, i.e., (D + D * )/2 ≤ D and the right bound in (ii) is implied. Proof Orient each edge in S to a locking endpoint (ties broken arbitrarily). This gives a well-oriented graph S without unoriented inner edges. Since Lemma 4.9, (i) and Here is the essential implication on the number of compatible edges. We see that Corollary 4.11 (i) is Theorem 1.6 by Hurtado et al. [20]. Actually, the set ET \ ES in the argument for Corollary 4.11(i) is what is called ps-flippable (pseudo-simultaneously flippable) by Hoffmann et al. [18], where a lower bound of max {n/2 − 2, h − 3} is shown for such ps-flippable edges. The proof (not the statement) of [18, Lemma 1.3] implies Lemma 4.10 above. We want to claim that the proof given above, including the proof of the Unoriented Edges Lemma 4.9 employed, is more concise. We emphasize, though, that our proof of the Unoriented Edges Lemma and its set-up is clearly inspired by the proof of the n/2 − 2 -bound for flippable edges in [20] (see also [25] for Lemma 4.9 (i)).
We conclude this section by mentioning a few further applications of Lemma 4.9 (which will play no role in the rest of the paper). A set of pairwise independently flippable edges is often called simultaneously flippable in the literature. Here is a streamlined proof of the known tight lower bound on their number by Souvaine et al. from 2011 [35], based on the Unoriented Edges Lemma 4.9.
Theorem 4.12 [35] Every triangulation T of P has a set of at least (n − 4)/5 edges that are pairwise independently flippable (simultaneously flippable).
Proof Let S be a -maximal coarsening of T . Orient edges in S (all locked) to a locking endpoint (ties broken arbitrarily). Lemma 4.9(ii) guarantees at least (n − 4 − (sl S + sl * S))/2 unoriented edges in S. Since all inner edges are oriented, sl S + sl * S ≥ n − 4 holds.
For each region r of S of slack d, T induces a triangulation of this (d + 3)-gon. Its d diagonals can be 3-colored with no triangle incident to two edges of the same color (start with an edge and spread the coloring along the dual tree of the triangulation of r ). Each color class offers a set of simultaneously flippable edges, one of size at least d/3 . It is easy to verify that d/3 We combine the simultaneously flippable edges collected for each region of S, which gives an overall set of at least (sl S + sl * S)/5 ≥ (n − 4)/5 simultaneously flippable edges.
While the lower bounds of (n − 4)/2 for the number of flippable edges, and of (n − 4)/5 for the size of a set of simultaneously flippable edges are tight (see [20] and [14], respectively), it is interesting to observe from the proofs that both bounds cannot be simultaneously attained for the same point set. That is, if the number of flippable edges is close to its lower bound then this forces the existence of a set of more than (n − 4)/5 simultaneously flippable edges, and vice versa. This can be quantified as follows.

Theorem 4.13
If the number of flippable edges in a triangulation T of P is α, and if the largest set of simultaneously flippable edges in T has size β, then α +β ≥ 4(n − 4)/5.
(The claim is still true if α is the largest size of a set of pseudo-simultaneously flippable edges, see [18].) Proof Let S be a -maximal coarsening of T , with d 1 , d 2 , . . . , d |RS| the slacks of the regions of S. Then α ≥ i d i and, by the proof of Theorem 4.12, Let us conclude this section with an observation about interior-disjoint k-holes, as it emerged from a discussion with Manfred Scheucher (see also [34]). A k-hole of P is a subset of k points in convex position whose convex hull is disjoint from all other points in P. A k-hole and an -hole are called interior-disjoint, if their respective convex hulls are interior-disjoint (they can share up to two points). Harborth showed in 1978 [17], that every set of at least ten points has a 5-hole. This was recently strengthened to the existence of another interior-disjoint 4-hole [19,Theorem 2]. We show, how our framework allows an easy proof of this fact.
Theorem 4.14 [19] Every set P of at least ten points has a 5-hole and a 4-hole which are interior-disjoint.
Proof Since P has a 5-hole [17], there is a subdivision of P with a region of slack 2. Consider a -maximal coarsening S of such a subdivision with r a region of largest slack. We know that sl r ≥ 2. If sl r ≥ 4, i.e., it is a k-gon with k ≥ 7, then we can add a diagonal to r which divides r into a 5-gon and a (k − 3)-gon which readily gives the claimed interior-disjoint holes. Otherwise, if k ≤ 6, we have sl r + sl * r ≤ 5, while sl S + sl * S ≥ n − 4 ≥ 6 (Lemma 4.9 (ii)). Hence, there must be another region r ∈ RS with positive slack, i.e., a k -gon with k ≥ 4.

Link of a Full Triangulation
The link of a triangulation is the graph representing the compatibility relation among its flippable edges. In Sect. 1.7, the intuition for links as counterparts of vertex figures in polytopes was briefly explained. We will see that for proving Theorem 1.7 (ii) (in Sect. 5.2 below) it is enough to prove n/2 − 3 -vertex connectivity of links. Here is the special property of links that will immediately show that the vertex connectivity is determined by the minimum vertex degree (via Lemma 2.4). Proof For e and f flippable edges in T we show that there is at most one flippable edge g that is compatible with neither e nor f ; this implies the lemma. Such a g has to be an edge of both terr e and terr f . If terr e and terr f are disjoint, they share at most one edge, since terr e and terr f are convex, see Fig. 13, left. Otherwise, terr e and terr f overlap in a triangle , of which e and f are edges; the third edge g of this triangle is a common edge of terr e and terr f , see Fig. 13, right. No other edge of terr f can appear on the boundary of terr e: Consider the line λ through edge f . All edges of terr f other than e and g lie on the side of λ opposite to , and terr e is on the same side of λ as , since it contains and f is an edge of terr e; again, convexity of terr e and of terr f is essential here.

Proof of Theorem 1.7 (ii)
Proof We want to show that for n ≥ 5, the edge flip graph is max { n/2 − 2 , h − 3}vertex connected. We employ the Local Menger Lemma 2.3. We know that the edge flip graph is connected [21]. What is left to show is that for any triangulation T ∈ T full (P) and

Covering the Edge Flip Graph with Polytopes
Recall from Sect. 1.3 that for a convex k-gon there is a (k − 3)-polytope whose 1-skeleton is isomorphic to the edge flip graph of triangulations of the k-gon, an associahedron, which we denote by A k−3 (the index reflecting its dimension; to be precise, A k−3 is some representative realization of the associadedron); A 1 is an edge, A 2 is a convex pentagon, etc. If we consider all triangulations of the k-gon with a given diagonal present, we get a facet of A k−3 ; all facets of A k−3 can be obtained in Then the set, T full S , triangulations refining S, with its edge flip graph is represented by the product (see [39]) It is now easy to see that every edge of the edge flip graph finds itself in the 1-skeleton of such a max { n/2 − 2 , h − 3}-dimensional polytope contained in the edge flip graph (see also the discussion in Sect. 1.5). One can strengthen this and show that any pair of incident edges in the edge flip graph is part of a subgraph isomorphic to the 1-skeleton of some n/2−2 -dimensional polytope (a gluing of products of associahedra). This has been discussed in [37], but we decided to skip that part in this version.

Partial Subdivisions-Slack and Order
In Sects. 7-9 we move on to proving Theorem 1.9, the (n − 3)-vertex connectivity for the bistellar flip graph of partial triangulations. As indicated in the introduction, the proof will follow a similar line as for the edge flip graph, using subdivisions and links, but several new aspects and challenges will appear. Convention From now on, in Sects. 7-11, we will mostly use triangulation short for "partial triangulation," but we return to using "full triangulation" and "full subdivision." We define partial subdivisions, which form a poset in which the triangulations of P are the minimal elements-our definition is a specialization, to the plane and general position, of the established notion of a polyhedral subdivision [24]. These partial subdivisions are plane graphs, possibly with isolated points. Hence, it may be useful to point out a subtlety in the definition of regions of a plane graph (Definition 1.2): We defined them as the bounded connected components in the complement of the union the edges (as line segments in R 2 ), not taking the isolated points into account. That is, regions can contain isolated points of the graph, and isolated points will not keep them from being convex. Observe that a partial subdivision S is a full subdivision of P iff VS = P and V by S = ∅. Also, a partial subdivision S is a full subdivision of VS \ V by S.
Convention From now on, in Sects. 7-11, we will mostly use subdivision for "partial subdivision." As it should have become clear by now, VS is essential in the definition of a subdivision S, it is not simply the set of endpoints of edges in S, there are also bystanders. For example, for T ∈ T part (P), all graphs (P , ET ) with VT ⊆ P ⊆ P are subdivisions of P, all different. VS partitions into boundary points, involved points, and bystanders, i.e., VS = xtr P∪ V inv S∪ V by S. Moreover, there are the skipped points, A first important example of a subdivision is obtained from a triangulation T and an element x flippable in T , i.e., {T , T [x]} is an edge of the bistellar flip graph: If x = e is a flippable edge, then T ±e has one convex quadrilateral region Q; all other regions are triangular. We obtain T and T [e] from T ±e by adding one or the other of the two diagonals of Q to T ±e . If x = p is a flippable point, then T ± p is almost a triangulation, all regions are triangular, except that p ∈ VT ± p is a bystander. We obtain T and T [ p] by either removing this point from T ± p or by adding the three edges from p to the points of the triangular region in which p lies. The subdivision T ±x is close to a triangulation and, in a sense, represents the flip between T and T [x].
To formalize and generalize this we generalize the notion of slack from full to partial subdivisions. Definition 7.2 (slack of partial subdivision). Given a subdivision S of P, we call a region of S active if it is not triangular or if it contains at least one point in VS (necessarily a bystander) in its interior. For r ∈ RS, the slack, sl r , of r is |Vr | − 3. The slack of S, sl S, is the sum of slacks of its regions.
Note that a region is active iff it has nonzero slack.

Observation 7.3 For a subdivision S with s bystanders we have
Proof The slack of a region r equals the number of edges it takes to triangulate r (ignoring bystanders) plus the number of bystanders in r . Thus, sl S is the number of edges it takes to triangulate (VS \ V by S, ES) (a full subdivision of VS \ V by S) plus |V by S|. Now the claim follows from Lemma 4.6 (or Observation 4.7).

Observation 7.4 Let S be a subdivision.
(i) sl S = 0 iff S is a triangulation iff S has no active region.
(ii) sl S = 1 iff S has exactly one active region of slack 1; this region is either a convex quadrilateral, or a triangular region with one bystander in its interior. (iii) sl S = 2 iff S has either (a) exactly two active regions, both of slack 1, or (b) exactly one active region of slack 2, where this region is either a convex pentagon, or a convex quadrilateral with one bystander in its interior, or a triangular region with two bystanders in its interior.
The example in Fig. 17 hides some of the intricacies of the partial order ; e.g., in general, it is not true that all paths from a triangulation to S triv have the same length n − 3. S triv is the unique coarsest ( -maximal) element (quite contrary to the poset of full subdivisions, where there were several -maximal full subdivisions). The triangulations (i.e., subdivisions of slack 0) are the minimal elements.   1 each), then T part S has cardinality 4, spanning a 4-cycle in the bistellar flip graph of P (Fig. 18). If there is exactly one active region in S (of slack 2), then T part S has cardinality 5, spanning a 5-cycle (see Fig. 2).

Lemma 7.8 Any proper refinement S of a subdivision S of slack 2 has slack at most 1.
Proof For a refinement of S we add m edges, thereby involving s bystanders, and we remove s bystanders (some of these parameters may be 0, but not all, if the refinement is proper). We have sl S = sl S −(m −2s +s ) (easy consequence of Observation 7.3) and we want to show m − 2s + s > 0.
Since sl S = 2, S has at most two bystanders and thus s ≤ 2. If s = 0, then m − 2s + s > 0 holds, since some of the three parameters have to be positive. If s = 1, we observe that we need at least three edges to involve a bystander and m − 2s ≥ 3 − 2 · 1 = 1. If s = 2, we need at least five edges to involve two bystanders and m − 2s ≥ 5 − 2 · 2 = 1.
For D ≥ 3, a proper refinement of a subdivision of slack D can have slack D or even higher (Fig. 19). The proof fails, since we can involve three bystanders with six edges. Intuitively, as briefly alluded to at the end of Sect. 1.3 (for the special case of convex position), one can think of the subdivisions as the faces of a higherdimensional geometric structure behind the bistellar flip graph, with slack playing the role of dimension, analogous to the secondary polytope for regular triangulations. The following lemma shows that-for slack at most 2-we have the property corresponding to the fact that faces of dimension d are either equal, or intersect in a common face of smaller dimension (possibly empty). This correspondence fails for slack exceeding 2. Proof If T part S 1 ∩ T part S 2 contains some triangulation, then we easily see that S 1 ∧ S 2 := (VS 1 ∩ VS 2 , ES 1 ∪ ES 2 ) is a subdivision, and T part S 1 ∧ S 2 = T part S 1 ∩ T part S 2 . Two edges incident to a vertex of a polytope may span a 2-face, or not; same here, which gives rise to the following definition:  This needs some time to digest. In particular, if two flippable edges e and f share a common endpoint of degree 4, then they are compatible (Fig. 20, bottom left), quite contrary to the situation for full triangulations as treated in Sect. 3.3 (see Definition 3.6). The configurations of two flippable but incompatible elements are shown in Fig. 20 (two rightmost): (a) Two flippable edges e and f whose removal creates a nonconvex pentagon and whose common endpoint q has degree at least 5. (b) A flippable edge e and a flippable point p of degree 3 whose removal creates a nonconvex quadrilateral region whose reflex point q has degree at least 5 in the triangulation. What is essential for us is that whenever x and y are compatible in a triangulation T , then there is a cycle of length 4

Observation 7.11 Let T ∈ T part (P).
(i) A skipped point p ∈ P • \ V • T is compatible with every flippable element of T .
(ii) Any two flippable points p, q ∈ P • are compatible.

Coarsening Partial Subdivisions
As in Sects. 4 and 5 for full triangulations, the existence of many coarsenings is essential for the vertex connectivity of the bistellar flip graph. However, note right away that going via -maximal subdivisions-as for full subdivisions-will not work: Here, for partial subdivisions, there is a unique -maximal element, the trivial subdivision. Moreover, note that for full subdivisions (as employed in Sect. 4), if S 1 S 2 , then (S 1 , S 2 ) is an edge in the Hasse-diagram of the partial order iff sl S 2 = sl S 1 + 1. For partial subdivisions, this is not the case (Fig. 21). (i) We call S 2 a direct coarsening of S 1 (and S 1 a direct refinement of S 2 ), in symbols S 1 ≺ dir S 2 , if S 1 S 2 and any subdivision S with S 1 S S 2 satisfies S ∈ {S 1 , S 2 } (equivalently, if (S 1 , S 2 ) is an edge in the Hasse diagram of ). (ii) We call S 2 a perfect coarsening of S 1 (S 1 a perfect refinement of S 2 ), in symbols S 1 ≺ 1 S 2 , if S 1 ≺ dir S 2 and sl S 2 = sl S 1 + 1.
The reflexive transitive closure of ≺ dir is exactly , while ≺ * 1 ⊆ and, in general, the inclusion is proper.
To motivate the upcoming definitions, let us discuss a few possibilities of coarsenings, direct coarsenings and perfect coarsenings. There are the simple operations of removing an unlocked edge, and of adding a skipped point p ∈ P \ VS as a bystander. For a triangulation, we can isolate a point of degree 3. How does this generalize to subdivisions? Removing the edges incident to a point of degree 3 does not work if some incident edge is locked at its other endpoint (e.g., p 0 in Fig. 22). If, however, no edge incident to a given point p (of any degree) is locked at the respective other endpoint, then we can isolate this point for a coarsening S . Unless p has degree 3, S is not a direct coarsening of S, though. If p has degree at least 4, some incident edge, say e, is not locked at p, thus not locked at all, and therefore, S S S for S := (VS, ES \ {e}). Finally, suppose we want to isolate all points in a set U of points for obtaining a coarsening S . For this to work, it is necessary that no edge e connecting U with the outside is locked at the endpoint of e not in U . However, this is not a sufficient condition, because several edges connecting U with a point not in U can collectively create a reflex vertex by their removal (e.g., U = {p 0 , p 1 , p 2 } in Fig. 22). Moreover, for S ≺ dir S to hold, U cannot be incident to unlocked edges, and no nonempty subset of U can be suitable for such an isolation operation.  Examples of prime coarseners are given in Fig. 23.

Remark 1 Why is it necessary to put the extra condition "
where U is exactly the set of new bystanders, i.e., V by S \ V by S = U " in Definition 8.2 (i)? U is supposed to be the set of points that get isolated in the coarsening step, this is essential for the definition of increment. However, there might be a subset of U that covers already all edges incident to U (see Fig. 24). We thank Valentin Stoppiello for pointing this out, it was overlooked by us in the earlier version [38].
The following observation, a simple consequence of Observation 7.3, explains the term "increment". Let us note right away, that the increment can be negative.  The following observation lists all ways of obtaining direct and perfect coarsenings.

Isolating a prime coarsener: For U a prime coarsener in S, S is obtained from S by removal of the set, E U , of all edges incident to points in
(ii) S is a perfect coarsening of S iff it is obtained from S by adding a single point, removing a single unlocked edge, or by isolating a perfect coarsener.
We are prepared for the right formulation and proof of the Coarsening Lemma. We let S be a subdivision of slack D ≥ 1 assuming the assertion holds for slack less than D.

Case 2.
There is no bystander in S. Again we employ a partial orientation of S. The choice of the orientation is somewhat more intricate and we will proceed in three phases (Fig. 25). We keep the invariant that the unoriented inner edges are exactly the unlocked inner edges.
In a first phase, we orient all locked inner edges to all of their locking endpoints, i.e., we temporarily allow edges to be directed to both ends (to be corrected in the third phase); edges directed to both endpoints are called mutual edges. We can give the following interpretation to an edge directed from p to q: If we decide to isolate p (i.e., remove all incident edges of p) for a coarsening of S, then q becomes a reflex point of some region and we have to isolate q as well (i.e., every coarsener containing p must contain q as well). In particular, if { p, q} is a mutual edge, then either both or none of the points p and q will be isolated. In fact, if we consider the graph G with VG := V • S and EG the mutual edges in the current orientation, then in any coarsening of S either all points in a connected component of G are isolated, or none.
A connected component K of G is called a candidate component, (a) if all edges connecting K with points outside are directed towards K , (b) no point in K is incident to an unoriented edge, (c) all points in K have indegree 3, and (d) the mutual edges in K do not form any cycle (i.e., they have to form a spanning tree of K ). It follows that if K has k points then the number of edges is 3k − (k − 1) = 2k + 1. The term "candidate" refers to the fact that removing all edges incident to K seems like a direct coarsening step with incrementing the slack by 1 (Observation 7.3); however, while individual edges connecting K to the rest of the graph are not locked at their endpoints outside K , some of these edges collectively may actually create a reflex vertex in this way (see K and q in Fig. 25, left). So K is only a candidate for a perfect coarsener.
We start the second phase of orienting edges further. In the spirit of our remarks about candidate components of G, suppose q is an inner point outside a candidate K of G (thus all edges connecting q to K are directed from q to K ), such that removing the edges connecting q to K creates a reflex angle at q. Then we orient one (and only one) of the edges connecting q to K , say { p, q}, also to q, thereby making this edge mutual. (The reader might be worried that q now joins the candidate component while possibly not having indegree 3 as required in a candidate component. However, this just means that the enlarged component is not a candidate component, i.e., we have lost a candidate component.) We call all the edges connecting K to q, except for { p, q}, the witnesses of the extra new orientation of { p, q} from p to q. We successively proceed orienting edges, with the graph G of mutual edges evolving in this way and candidate components growing or disappearing. (The reader will correctly observe that our approach is very conservative towards prime coarseners, but by what we observed and by what will follow, since we are interested only in perfect coarseners, we can afford to leave alone connected components other than the candidate components.) The process will clearly stop at some point when the second phase is completed. We freeze G and denote it by G * .
Before we start the third phase, let us make a few crucial observations: K q Fig. 25 Orientation after phase 1, with candidate components shaded (left); after phase 2 (middle), with the connected components of G * ; after phase 3 (right), with unoriented edges bold (each of these can be removed for a coarsening of slack 1 larger), and with the candidate components with a leader shaded (perfect coarseners) (i) If p, q are inner points in the same connected component of G * , then any coarsener contains both or none (i.e., if a connected component is a coarsener, then it is prime). This holds after phase 1, and whenever we expand a connected component, it is maintained. (ii) During the second phase, an edge can be witness only once, and it is and will never be directed to the endpoint where it witnesses. Why? (a) Before it becomes a witness, it connects different connected components of G, after that it is and stays in a connected component of G. (b) Before it becomes a witness, it is not directed to the endpoint to which it witnesses an orientation, after that it is and stays in a connected component of G and can therefore not get an extra direction.
(An unoriented edge can never get an orientation and it can never be a witness.) (iii) If we remove, conceptually, for each incoming edge of a point q the witnesses (which direct away from q) for the orientation of this edge to q, then among remaining incident edges, all the incoming edges are locked at q (an incoming edge that was oriented already in the first phase to q has no witness). In particular, the indegree of q cannot exceed 3, and if q is incident to some not ingoing edge which is not a witness for any edge incoming at q, then the indegree of q is at most 2. (We might generate incoming edges to a point q that are not consecutive around q.) (iv) If an unoriented edge e connects two points of the same connected component of G * , then both endpoints have indegree at most 2 (recall that this edge e cannot be a witness at its endpoints). If an edge e is directed from a connected component K of G * to a point outside K , then the tail of this edge e has indegree at most 2 (recall that e cannot be a witness at all, since its endpoints are in different connected components of G * ). (v) A candidate component K of G * is a perfect coarsener. It is a coarsener (otherwise, we would have expanded it further), it is a prime coarsener (see (i) above) and inc K = 1 (we have argued before that a candidate component increases the slack by exactly 1).
The third phase will make sure that each mutual edge loses exactly one direction. Our goal is to have in every connected component K of G * at most one point with indegree 3. To be more precise, only candidate components have exactly one point with indegree 3, others don't. Consider a connected component K .
(a) If the mutual edges form cycles in K , choose such a cycle c and keep for each edge on c one orientation so that we have a directed cycle, counterclockwise, say. All other mutual edges in K keep the direction in decreasing distance in G * to c, ties broken arbitrarily. This completed, no point in K has indegree 3, since there is always a mutual edge incident that decreases the distance to c and the incoming direction of this edge will be removed. (b) If K has points of indegree at most 2, choose one such point p with indegree at most 2, orient all mutual edges in K in decreasing distance in G * to p, ties broken arbitrarily. Again, this completed, no point in K will have indegree 3. (c) If none of the above applies, the mutual edges of K form a spanning tree and all points in K have indegree 3. Moreover, all edges connecting K with points outside are directed towards K and no edge within K is unoriented (violation of these properties force a point of indegree at most 2). So this is a candidate component. We choose an arbitrary point p in K , call it the leader of K , and for all mutual edges keep the orientation of decreasing distance in G * to p (ties cannot occur, mutual edges form a tree). Now the leader p is the only point of K with indegree 3, all other points in K have indegree exactly 2.
Phase 3 is completed. Let us denote the obtained partial orientation of S as S * . It has identified certain connected components of G * which have a leader of indegree 3. In fact, every point of indegree 3 after phase 3 is part of a perfect coarsener (probably of size 1).
We can now describe a sufficient supply of perfect coarsenings of S. Let N := |VS| and let C 3 be the number of points of indegree 3 in S * . We know that there are at least N − 3 − D − C 3 unoriented inner edges (Lemma 4.9). In this way we have identified at least n − 3 − D perfect coarsenings.
Here are two immediate implications which we will need later: The first in the vertex connectivity proof in Sect. 9 and the second for the result about covering of the bistellar flip graph by (n − 3)-polytopes in Sect. 11.

(n − 3)-Connectivity for Partial Triangulations
To complete the proof of Theorem 1.9, the (n − 3)-vertex connectivity of the bistellar flip graph, we need again links, now for partial triangulations, which are graphs that represent the compatibility relation among flippable elements (Definition 7.10).

Link of a Partial Triangulation
Recall that if x is a flippable element in a triangulation T then T ±x denotes the sub- (See Sect. 1.7 for some intuition for this definition and Fig. 26 for an example.) We will see that it is enough to prove (n −4)-vertex connectivity of all links. The following lemma implies, via Lemma 2.4, that the vertex connectivity of links is determined by the minimum vertex degree. Proof Recall that all p ∈ P • \ V • T are flippable and compatible with every flippable element (Observation 7.11 (i)), so we can assume {x 0 , Moreover, if p, q ∈ V • T are two distinct points flippable in T , then p q (Observation 7.11 (ii)). Hence, we assume that no two consecutive elements in the cyclic sequence (x 0 , x 1 , x 2 , x 3 ) are points; w.l.o.g. let x 0 = e and x 2 = f be edges.
Recall from Definition 3.2, that for an inner edge e in a triangulation T , its territory terr e, is defined as the interior of the closure of the union of the two regions in T incident to e. Obviously, e is flippable in T iff the quadrilateral terr T e is convex. Note that for an element x to be incompatible with edge e, x must appear on the boundary  of terr e, and analogously elements incompatible with f must appear on the boundary of terr f . We show that there is at most one flippable element in the intersection of the boundaries of terr e and terr f (Fig. 27). This is obvious, if terr e ∩ terr f is empty or a single point (recall that A denotes the closure of A ⊆ R 2 ). If this intersection is an edge and its two endpoints, we observe that among any edge and its two incident points, at most one element can be flippable (inner degree 3 points cannot be adjacent and cannot be incident to a flippable edge). This covers already all possibilities if terr e and terr f are disjoint (since they are convex). Finally, terr e ∩ terr f can be a triangle (see argument in the proof of Lemma 5.2), in which case the common boundary consists of the common endpoint of e and f , clearly not flippable, and an edge with its two endpoints; again, at most one of these three can be flippable.  Proof Let x be a vertex of Lk part T . T ±x , a subdivision of slack 1, has at least n − 4 perfect coarsenings of slack 2 (Lemma 8.6). Each such coarsening equals T ±x,y for some y x, i.e., y is a neighbor of x in Lk part T . Distinct coarsenings yield distinct compatible elements y (since {T [x], T , T [y]} ⊆ T part T ±x,y and T part T ±x,y spans a cycle, T [y] is determined as the other neighbor of T on this cycle). That is, the minimum vertex degree in Lk full T is at least n − 4. Lk part T has no cycle of length 4 in its complement (Lemma 9.2). The lemma follows by Lemma 2.4.

Proof of Theorem 1.9
Proof We know that the bistellar flip graph is connected [24,Sect. 3.4.1], and it has at least n − 2 vertices, since it is nonempty and every vertex has degree at least n − 3 (Corollary 8.7 (i)). Hence, for (n −

Regular Subdivisions by Successive Perfect Refinements
Suppose h = 3 and consider stacked triangulations of P, i.e., we start with the triangulation (xtr P, E hull ), and then we successively add points in P • by connecting a new point to the three vertices of the triangle where it lands in (Fig. 29). It is easily seen that this yields regular triangulations. The result of this section is the following sufficient condition for the regularity of a subdivision (Definition 10.2 below), which can be seen as a generalization of the regularity of stacked triangulations (Fig. 30). The condition is not necessary, see Sect. 10.2.
Theorem 10.1 If S ≺ * 1 S triv for a subdivision S, then S is a regular subdivision.
In other words, all subdivisions, in particular, all triangulations in the ≺ 1 -lower closure of S triv are regular. This condition will eventually allow us to show the covering of the bistellar flip graph by graphs of (n − 3)-polytopes. The proof of Theorem 10.1 stretches out over several definitions and lemmas with a conclusion in Sect. 10.6. Before we give a brief outline of this proof shortly in Sect. 10.3, we first introduce some notions.

Height Functions, Liftings, and Regular Subdivisions
We see that if sl S < |VS| − 3, there are always compliant height functions not in (S). In order to extract among those a realizing height function we consider mountains and valleys in the lifting given by a compliant height function.  (i) ω is linear on S iff the ω-labeling of S is constant 0. (ii) ω realizes S iff the ω-labeling of S is constant .

Mother of Examples
In order to understand the subtleties of whether a subdivision is regular or not, we should briefly discuss the mother of examples, see [24]. For this consider the configuration in Fig. 32. Whether or not the displayed subdivisions are regular or not depends on how exactly the three dashed lines (as indicated in S) meet.
(a) If the three dashed lines meet in a common point, then S is a regular subdivision, but none of T and T is regular. (b) If the three dashed lines do not meet in a common point, then S is not a regular subdivision, but one of T and T is regular, the other one not.
The example allows us to clarify a few points.
-The condition in Theorem 10.1 for regularity is not necessary (consider case (b) with T regular, and note T ≺ * 1 S triv ). This is inherently so, since the condition in Theorem 10.1 depends only on the order type (or oriented matroid) of the point set. In fact, note that a perturbation of the point set does not change the order type of the set, but it affects how the dashed lines meet and, therefore, whether subdivisions are regular or not.
-The condition in Theorem 10.1 cannot be generalized to: If S ≺ * 1 S and S is regular, then S is regular. In fact, adding a single edge in a subdivision may switch from regular to non-regular (case (a)). The right generalization will be given in the Regularity Preservation Lemma 10.10 below.

Outline of Proof of Theorem 10.1
It is easy to see that if p is an inner point of degree 3 in a triangulation T , then for any height function ω (which, as we observed, is T -compliant), the ω-labeling assigns the same value to the three edges incident to p. We will generalize this observation for an S-compliant height function ω in two ways: (A) If p ∈ V • S and not all incident edges are 0-labeled, then the ⊕-labeled andlabeled edges incident to p cannot be separated by a line through p (Lemma 10.7).
(In particular, this forces the ω-labeling to be constant on the edges incident to an inner point of degree 3 in any subdivision.) (B) If K is a perfect coarsener of S, then the ω-labeling assigns the same label to all the edges E K incident to a perfect coarsener K (Lemma 10.9).
Here is another simple observation about an inner point p of degree 3 in a triangulation T . Removing the three edges incident to p (while keeping p as a bystander) yields a subdivision S with T ≺ 1 S. Obviously, if S is a regular subdivision, then T is a regular triangulation (this was behind our observation about stacked triangulations at the beginning of this section): Given a height function ω realizing S, we can always perturb p downwards (decrease ω p by a sufficiently small value ε), obtaining a height function that realizes T . Again, this allows an appropriate generalization: (C) Suppose S is a regular subdivision with dim (S) = |VS|−sl S. Then every perfect refinement S of S, i.e., S ≺ 1 S, is regular, and, moreover, dim (S ) = |VS |−sl S (Regularity Preservation Lemma 10.10).
Since S triv is regular and dim (S triv ) = 3 = |VS triv | − sl S triv , this immediately yields an inductive argument for Theorem 10.1. For the proof of (C), we consider the perfect coarsener K of S whose isolation leads to S, and a height function ω 1 that realizes S. First, we show that dim (S ) > dim (S), and, therefore, a height function ω ∈ (S ) \ (S) exists. According to (B), the ω -labeling assigns the same label to all edges incident to K , and since ω / ∈ (S), this label cannot be 0. Hence, either ω or −ω assigns constant , and it can be used for a controlled perturbation ω 1 + εω which realizes S . We will now carefully work out these steps.

Valid {⊕, , 0}-Edge Labelings
Definition 10.6 Let S be a subdivision. Given a labeling α : E • S → {⊕, , 0}, we call an inner point p ∈ V • S α-pointed, if α is not constant 0 on the edges incident to p, and if there is a line through p that has all ⊕-labeled edges incident to p strictly on one side and all -labeled edges incident to p strictly on the other side of this line. (We do not require that both ⊕and -labeled edges incident to p exist.) We call α a valid labeling of E • S if no point in V • S is α-pointed (Fig. 33).
For example, for an inner point of degree 3 in a subdivision, a valid labeling must assign the same label to its three incident edges. We can now prove (A) above. Proof Let p ∈ V • S and suppose there is a line through p that has all ⊕-labeled edges incident to p on one side, and all -labeled edges incident to p on the other side. Sweep a vertical plane h parallel to in R 3 over p and observe its intersection with the ω-lifting f . On the side of the ⊕-labeled edges, this intersection must be a locally concave function, on the side of the -labeled edges a locally convex function. Consequently, it has to be locally linear at the point when h contains p and . Now it follows that f must be locally linear around p and all edges incident to p must be flat.

Lemma 10.8 Let K be a perfect coarsener in a subdivision S. In every valid {⊕, , 0}labeling of E • S, the edges E K incident to K all get the same label.
Proof We plan to prove the following.
Claim. With reference to the orientation process in the proof of Lemma 8.6, after the second phase, in any valid labeling, the edges incident to a candidate component of G * get the same label.
It is not obvious from the proof of Lemma 8.6 that every perfect coarsener is identified by the three phase process. In order to close this gap (from the claim to the assertion of the lemma), isolate K in S obtaining a subdivision S with a region r containing the points in K . Let V r be the vertices of this region, i.e., V r = Vr \ V by S . Now consider the subgraph S r of S induced by V r ∪ K . K is a perfect coarsener of S r whose isolation yields the trivial subdivision of V r ∪ K . It is the only coarsener of S r and sl S r = |VS r | − 4. Therefore, the procedure in the proof of Lemma 8.6 must identify K as a candidate component after the second phase. We establish the claim by showing the following invariant in the process during the second phase: (a) For every candidate component K , the edges E K incident to K obtain the same label in any valid labeling. (b) An edge gaining a new orientation in the second phase and its witnesses obtain the same label in any valid labeling.
At the end of the first phase, a point with indegree 3 in S has actually degree 3 in S, and therefore any valid labeling must give the same label to all incident edges. Since in a candidate component K , all points have indegree 3 in S (i.e., at this point, have degree 3 in S) and since a connected component is connected [sic!], it easily follows that all edges incident to a candidate component must have the same label in any valid labeling.
During the second phase, a newly oriented edge and its witnesses are part of the edges incident to a candidate component. Hence, given (a), invariant (b) is maintained after an orientation step of phase 2. We are left to show that (a) is preserved. Consider a point p of a candidate component. It must have indegree 3, all incident edges are either ingoing or witnesses for an ingoing edge (otherwise, indegree 3 is impossible); we know that each bundle of an ingoing edge and its witnesses have the same label, and such a bundle can be separated from the other incident edges by a line through the given point p (this is why the edge was oriented in phase 2). Hence, due to a simple consideration, any valid labeling must assign the same label to all incident edges. (The simple consideration: Suppose a single bundle is labeled ⊕, then this bundle can be separated from the other two bundles by a line, contradiction. Suppose exactly two bundles are labeled ⊕, then the remaining bundle can be separated from these two ⊕-labeled bundles by a line, which is a contradiction. Hence, if any incident edge is labeled ⊕, then all incident edges must be labeled ⊕. Similarly, for .) This completes the proof of the claim, and thus of the lemma. Now, with Lemma 10.7 we immediately get property (B).

Lemma 10.9
If ω is an S-compliant height function, then the ω-labeling is constant on any set of edges incident to a perfect coarsener of S.
Proof Let ω 1 ∈ R VS 1 be a height function that realizes S 1 . Case 1. S 1 is obtained from S 0 by adding a single point p ∈ P • \ V • S 0 . Then ω 1| VS 0 realizes S 0 and S 0 is regular. We have (S 0 ) = {ω| VS 0 | ω ∈ (S 1 )}. For ω ∈ (S 1 ), the value of ω p ( p the added point) is determined by ω| VS 0 , i.e., dim (S 1 ) = dim (S 0 ). Therefore, Case 2. S 1 is obtained from S 0 by removing a single unlocked edge or by isolating a perfect coarsener in S 0 . We have VS 1 = VS 0 . The set E * := ES 0 \ ES 1 is either a single unlocked edge or the set E K of edges incident to a perfect coarsener K in S 0 . Let r * be the region in S 1 generated by the removal of the edges in E * .
We have (S 0 ) ⊇ (S 1 ) and (with Lemma 10.3 (iii) and sl S 0 = sl S 1 − 1) Therefore, there must exist ω ∈ (S 0 ) \ (S 1 ), a height function that is not flat on r * and there is an edge in E * that is not flat in the ω -lifting of S 0 . In that case, all edges in E * are mountains or all are valleys (trivially true if |E * | = 1, otherwise by Lemma 10.9). Let us suppose that the ω -labeling is constant on E * (if not switch to −ω ). Now, for any sufficiently small positive ε ∈ R, the height function ω 0 := ω 1 + εω is compliant with S 0 and all inner edges in S 0 are valleys in the ω 0 -lifting of S 0 (ε has to be small enough such that all valleys in the ω 1 -lifting remain valleys in ω 0 ; this is a familiar operation, see [24,Lemma 2.3.16]). This establishes that S 0 is regular.
We are left to show that dim (S 0 ) = |VS 0 | − sl S 0 , or, equivalently, dim (S 0 ) = dim (S 1 ) + 1. This holds, if for any two ω , ω in (S 0 ) \ (S 1 ) there exists a ∈ R and ω ∈ (S 1 ) such that ω = aω + ω. Suppose that all edges in E * are valleys in ω , and all edges in E * are mountains in ω (switch signs, if necessary). Now consider ω t := (1 − t)ω + tω , t ∈ [0, 1]. There must be a value t in (0, 1) where some edge in E * is flat in the ω t -lifting, but then all edges have to be flat and ω t ∈ (S 1 ) (Lemma 10.9). We have shown

Proof of Theorem 10.1
Proof If S ≺ * 1 S triv then there is a sequence S triv is regular, sl S triv = n − 3 and (S triv ) = (S triv ), of dimension 3 = n − sl S triv = |VS triv |−sl S triv . Along with Lemma 10.10 we have an inductive argument that S 0 = S is regular.
We immediately get that successive perfect refinements of a subdivision S * fill the regions of S * with locally regular subdivisions. Definition 10.11 (restriction of subdivision). Let S S be subdivisions, and let r ∈ RS . Then the restriction of S to r , S| r , is the subgraph of S induced by VS ∩ r (r the closure of r ).

Corollary 10.12
Let S ≺ * 1 S . For r ∈ RS , we have S| r ≺ * 1 S triv (Vr ) and S| r is a regular subdivision of Vr.
Let us conclude this section with the remark, that successive perfect coarsening starting from a subdivision is a non-deterministic process, that may-even for the same subdivision-lead to S triv or not (see Fig. 21 at the beginning of Sect. 8).

Covering the Bistellar Flip Graph with Polytopes
Theorem 11.1 The edge set of the bistellar flip graph of P can be covered by subgraphs isomorphic to 1-skeletons of (n − 3)-polytopes (which are products of secondary polytopes). where -poly(A) denotes the secondary polytope of A ⊆ P [24], see also Sect. 1.4, Theorem 1.11. The dimension of this product is r (|Vr |−3) = sl S = n −3. Its faces correspond to the refinements S of S such that for each region r of S, S | r is regular, i.e., this includes T and T [x] (as vertices), and T ±x (as edge) (Corollary 10.12). This completes the argument.

Sets with All Triangulations Regular
We give characterizations of point sets for which all triangulations are regular (as, e.g., it is the case for point sets in convex position). In particular, we show that this can be easily read off the height of the partial order . In a first step we prove that property to be equivalent to requiring that all subdivisions are regular.

Lemma 11.2 All subdivisions are regular iff all triangulations are regular.
Proof The direction (⇒) is obvious. For (⇐) it suffices to show that every non-regular subdivision S with sl S > 0 has a direct refinement which is not regular. Case 1. S has a bystander p ∈ V by S. Clearly, the direct refinement (VS \ {p}, ES) is not regular iff S is not regular.
Case 2. S has no bystander. Since sl S > 0, S must have an active region r * which is a k-gon for k ≥ 4. Choose two crossing diagonals e 0 and e 1 in r * and consider the subdivisions S i := (VS, ES ∪ {e i }), i = 0, 1. We want to show that if S is not regular, then at least one of S 0 and S 1 is not regular. So let us suppose that, for i = 0, 1, ω i ∈ R VS is a height function realizing S i (as a regular subdivision) and, for t ∈ [0, 1], consider the convex combination ω t := (1 − t)ω 0 + tω 1 .
We say that a height function ω respects region r in subdivision S, if ω is linear on Vr and all points in VS (ω) \ Vr (ω) lie strictly above the plane spanned by Vr (ω) . Clearly, intersect (in the lifting in R 3 ). For that value of t, ω t is linear on Vr * . Moreover, all edges in S| r are valleys in the ω t -lifting, since these are valleys both in the ω 0 -lifting and the ω 1 -lifting (and that property is preserved for all convex combinations of ω 0 and ω 1 ). Hence, ω t realizes S and we have a contradiction.
We recall the definition of the height of an element in a partial order, and of the height of the partial order. Definition 11.3 (height). The height, hS, of a subdivision S (in the partial order ) is recursively defined: (a) If S is a triangulation, then hS := 0, and (b) if S is not a triangulation, then hS := 1 + max S ≺ dir S hS . (Equivalently, hS is the size of the longest -chain ending in S minus 1.) We let h max = h max (P) be the maximum height of any subdivision of P (i.e., h max = hS triv ).

Theorem 11.4
The following five conditions are equivalent.
(i) All triangulations are regular.
(ii) All subdivisions are regular.
Proof For (i) ⇔ (ii) see Lemma 11.2. For the rest we show the implication cycles all subdivisions are regular ⇒ all subdivisions are regular, (Note here: There is a maximum chain of size 1 + h max in the -partial order. If we remove the trivial subdivision (not corresponding to a proper face), we get still a chain of length h max .) (b) h max = n −3 ⇒ ≺ dir = ≺ 1 . Consider a maximal chain S 0 ≺ dir S 1 ≺ dir . . . ≺ dir S m ; because of maximality, S 0 is a triangulation (of slack 0), and S m = S triv (of slack n − 3). We know that sl S i ≤ sl S i−1 + 1 (Lemma 11.5 below) with equality iff S i−1 ≺ 1 S i . It follows that m ≥ n − 3. Moreover, if m = n − 3 (which is given if h max = n − 3), then S i−1 ≺ 1 S i for all i = 1, 2, . . . , n − 3. Since every pair S ≺ dir S is part of a maximal chain, the claim follows. (c) ≺ dir = ≺ 1 ⇒ all subdivisions are regular. Every subdivision S has a chain of direct coarsenings to S triv . If every direct coarsening is a perfect coarsening, this shows S ≺ * 1 S triv and therefore S is regular (Theorem 10.1). (d) ≺ dir = ≺ 1 ⇒ h = sl. For proving hS = sl S, we can proceed by induction on the height of S, where the induction basis holds without assumptions. With the assumption of ≺ dir =≺ 1 and with the induction hypothesis For the last equality, note that for a subdivision that is not a triangulation, a perfect refinement can always be obtained by simply adding an edge, or by involving a bystander as a point of degree 3. If there is no edge to add and if there is no bystander, we have a triangulation. (e) h = sl ⇒ h max = n − 3. If h = sl, then h max = hS triv = sl S triv = n − 3.

More Properties of Coarseners
We derive two more properties of prime and perfect coarseners, Lemma 11.5, which we used in the proof of Theorem 11.4 above, and Lemma 11.6, which we will need in Sect. 11.4 below.
Lemma 11.5 inc U ≤ 1 for every prime coarsener U in a subdivision S.
Proof Let r be the region in S obtained by removing from S the edges E U incident to U . The subgraph of S induced by U is connected (Observation 8.4 (iv)), that is, all points of U have to lie in the same region of S . We consider the restriction S| r (Definition 10.11), a subdivision of VS| r = VS ∩ r . Isolating U in S| r yields S | r , the trivial subdivision of VS| r . Therefore, that is, inc U = VS| r − 3 − sl S| r . On the one hand, U is the only coarsener of S| r (since U is prime and it exhausts all inner points in VS| r ). On the other hand, the Coarsening Lemma 8.6 guarantees VS| r − 3 − sl S| r , i.e., inc U (perfect) coarsenings of S| r . Therefore, inc U ≤ 1.

Large Minimal Sets with Non-Regular Triangulations-Proof of Theorem 1.14
Observation 11.7 If P ⊆ P and P has non-regular triangulations, then P has nonregular triangulations.
Given a set P with non-regular triangulations, is there always a small subset P of P that witnesses this fact? This is a question asked by Francisco Santos 3 (see also Sect. 12.4 for a computational complexity related motivation for this question). An 36 A twisted double-gon of ten points with subdivision S (left). Subset P * (case p * 0 = q 0 ) of a twisted double-gon (right) equivalent reformulation is: How large can minimal sets P with non-regular triangulations be? (Here, "minimal" means that every proper subset of P has only regular triangulations.) Santos describes such a minimal set of eight points and conjectures this to be the largest example of such a minimal set. We will show that there exist such minimal sets of arbitrarily large (even) size. Definition 11.8 (twisted double-gon). A set P of n points in general position, n = 2k even, is called a twisted double-gon if the following holds.
Let p 0 , p 1 , . . . , p k−1 be a counter-clockwise numbering of xtr P along the boundary of the convex hull of P.
(II) The set Q := P • of inner points is in convex position.
There is a numbering q 0 , q 1 , . . . , q k−1 of Q following the order along the boundary of the convex hull of Q such that for all i, 0 ≤ i ≤ k − 1, it holds: (III) q i is extreme in P \ {p i }.
(IV) q i is extreme in P \ {p i−1 , q i−1 }. (V) q i lies in the triangle q i−1 p i q i+1 . Figure 36 indicates that such twisted double-gons exist for all even n ≥ 6 (these are not double-circles, see e.g. [2]; actually, double-circles have only regular triangulations). For n = 6, this is the mother-of-examples configuration (Sect. 10.2). Here are a few simple observations. Observation 11.9 Let P be a twisted double-gon, with notation as in Definition 11.8.
(i) The graph S := (P, E hull ∪ {{q i , p i }, {q i , q i+1 mod k } | i = 0, 1, . . . , k − 1}) is a subdivision (uses (II) and (V)). Its slack is n − 3. (ii) If q i is involved in a subdivision of a subset of P where it is not extreme, it is connected to p i (by (III)) and to at least one of { p i−1 , q i−1 } (by (IV)). fact, the min-degree bound was our initial result, and the other results have emerged in the quest to get rid of the "P large enough" condition. As indicated right after Theorem 1.7 in Sect. 1.1, the min-degree bound holds for all point sets with up to ten points [12]. Since very small point sets initially seemed the most likely candidates for counterexamples (for instance, because Lemma 3.1 fails for a set of six points, Fig. 10), we conjecture that the "P large enough" condition can indeed be removed.

General vs. Special Position
When abandoning the general position assumption, i.e., allowing collinearities of three or more points, the situation changes significantly (actually, even repeated points have to be considered [24]). The edge flip graph may degenerate to a single vertex even for large point sets (Fig. 37). For bistellar flips, new flips as shown in Fig. 38 come into play. Theorem 1.8 ([25], see also [24,Thm. 3.4.9]) still holds in this situation and n − 3 bistellar flips are guaranteed. We are currently investigating with Grelier to what extent our methods generalize from general to special position towards (n − 3)-vertex connectivity of the bistellar flip graph, which is by no means obvious.

Stably Regular Triangulations
It might be interesting to understand, which regular triangulations are captured by Theorem 10.1, i.e., partial triangulations in the ≺ * 1 -lower closure of S triv . From the discussion in Sect. 10.2 (and Fig. 32) we recall one reason why this condition for regularity cannot possibly be necessary: It depends only on the order type (or oriented matroid) of P, while regularity depends on the coordinates of the point set. So let us define a partial triangulation T of P stably regular, if it is regular on any realization of the order type of P. (A careful formulation of "regular on any realization of the order type" is: Whenever there is a bijective mapping p → p from P to a set P which preserves orientation of triples, then the partial triangulation T of P obtained by VT = {p | p ∈ VT } and ET = {{p , q } | {p, q} ∈ ET } is also regular.) Note that if we can move the points in P freely while preserving T as a triangulation, then we can turn any triangulation into a regular one. This follows basically from Steinitz's Theorem (see [39,Theorem 4.1]).
We believe that Theorem 10.1 goes as far as one can go with an order type based condition.
Conjecture 12.1 A triangulation T ∈ T part (P) is stably regular iff T ≺ * 1 S triv .
One might ask whether the subdivisions in the ≺ * 1 -down set of S triv correspond to the face lattice of a polytope. In general, this is not the case, it is not even an abstract polytope, (see [27,Chapter 2A] for abstract polytopes): We can choose sl S as a rank function, have S triv as maximal/greatest element, and add an empty face as minimal/least element of rank −1. All flags (i.e., maximal chains) have the same size. But we fail on the so-called diamond axiom: "If S is a subface of S with the ranks differing by 2, then there are exactly two faces strictly between S and S ." Figure 21 shows an example where this axiom fails: S and S triv in this figure differ by 2 in slack, but there is only one subdivision strictly in between the two (in the ≺ 1 -order).

Minimal Sets with Non-Regular Triangulations
In Sect. 11.4 we have identified arbitrarily large minimal sets with non-regular triangulations. Given that such sets of arbitrary size exist, it is not clear how to check efficiently whether a set allows only regular triangulations. A first step would be to understand all minimal sets with non-regular triangulations. It is possible that twisted double-gons as defined in Definition 11.8 are the only such sets.
Let us mention here that the problem of deciding whether a given triangulation is regular can be answered in polynomial time via linear programming.

The Fundamental Group
Consider the set of closed walks in an edge or bistellar flip graph, starting in a fixed triangulation T * . Consider a cycle c = (T 0 , T 1 , . . . , T k−1 ), k ≥ 2, in the flip graph, i.e., T i and T i+1 mod k are adjacent triangulations for 0 ≤ i < k. An insertion of c adds c before an occurrence of T 0 in a closed walk, and a deletion of c removes c if it occurs before T 0 in a closed walk. Let us consider two closed walks equivalent, if they can be obtained from each other by insertion and deletion of cycles of length 2, or of cycles spanned by the refinements of a subdivision of slack 2 (these are 4-or 5-cycles, one can show that these are exactly the 4-and 5-cycles, both for the edge and the bistellar flip graph.). Lubiw et al. [26] show that in the edge flip graph all closed walks containing a triangulation T * are equivalent to the trivial walk (T * ); in other words, the fundamental group of the complex spanned by slack 2 subdivisions (which correspond to the 2-faces of the flip complex) is trivial. This is the decisive step in proving the orbit conjecture of [7] (see discussion in Sect. 1.5). It is natural to ask, whether the analogous result holds for the bistellar flip graph, or, more generally, whether there is a bistellar analogue of the flip complex.

Higher Dimensions
Flip graphs can be defined on triangulations in higher dimensions, i.e., tetrahedralizations in R 3 , etc., see [24], but even for sets of points in convex and general position in 3-space it is open, whether these graphs are always connected. As mentioned before, because of the Local Menger Lemma 2.3, we were able to show high vertex connectivity without ever providing any evidence that the graph is connected. Basically, we showed high vertex connectivity of the connected components of the flip graphs, and Lawson gave us the connectedness on top of it. Still, it might be interesting to see whether one can show, e.g., (n − 4)-vertex connectivity for bistellar flip graphs of sets in convex and general position in R 3 . A first step is supplied: There are always at least n − 4 flips for such sets [24, Proposition 3.6.19].

Mixing Rate of the Random Flip Process
Last, but not least, a natural next question is to show expansion properties of the flip graphs, ideally yielding rapid mixing of the process of flipping random edges in the flip graph. This is known for points in convex positions, still with a big gap between the upper bound of O(n 5 ) and the lower bound of (n 3/2 ) (McShine and Tetali 1998 [28], and Molloy et al. 2001 [30]). There are related results for lattice point sets by Caputo et al. 2013 [8].
Funding Open access funding provided by the Swiss Federal Institute of Technology Zürich.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.