The Pachner graph and the simplification of 3-sphere triangulations

# The Pachner graph and the simplification of 3-sphere triangulations

## Abstract

It is important to have fast and effective methods for simplifying 3-manifold triangulations without losing any topological information. In theory this is difficult: we might need to make a triangulation super-exponentially more complex before we can make it smaller than its original size. Here we present experimental work suggesting that for 3-sphere triangulations the reality is far different: we never need to add more than two tetrahedra, and we never need more than a handful of local modifications. If true in general, these extremely surprising results would have significant implications for decision algorithms and the study of triangulations in 3-manifold topology.

The algorithms behind these experiments are interesting in their own right. Key techniques include the isomorph-free generation of all 3-manifold triangulations of a given size, polynomial-time computable signatures that identify triangulations uniquely up to isomorphism, and parallel algorithms for studying finite level sets in the infinite Pachner graph.

ACM classes F.2.2; G.2.1; G.2.2; D.1.3

Keywords Triangulations, 3-manifolds, Pachner moves, 3-sphere recognition, isomorph-free enumeration

## 1Introduction

Triangulations of 3-manifolds are ubiquitous in computational knot theory and low-dimensional topology. They are easily obtained and offer a natural setting for many important algorithms.

Computational topologists typically allow triangulations in which the constituent tetrahedra may be “bent” or “twisted”, and where distinct edges or vertices of the same tetrahedron may even be joined together. Such triangulations (sometimes called semi-simplicial or pseudo-triangulations) can describe rich topological structures using remarkably few tetrahedra. For example, the 3-dimensional sphere can be built from just one tetrahedron, and more complex spaces such as non-trivial surface bundles can be built from as few as six [15].

An important class of triangulations is the one-vertex triangulations, in which all vertices of all tetrahedra are joined together as a single point. These are simple to obtain [11], and they are often easier to deal with both theoretically and computationally [6].

Keeping the number of tetrahedra small is crucial in computational topology, since many important algorithms are exponential (or even super-exponential) in the number of tetrahedra [5]. To this end, topologists have developed a rich suite of local simplification moves that allow us to reduce the number of tetrahedra without losing any topological information [2].

The most basic of these are the four Pachner moves (also known as bistellar moves). These include the 3-2 move (which reduces the number of tetrahedra but preserves the number of vertices), the 4-1 move (which reduces both numbers), and also their inverses, the 2-3 and 1-4 moves. It is known that any two triangulations of the same closed 3-manifold are related by a sequence of Pachner moves [21]. Moreover, if both are one-vertex triangulations then we can relate them using 2-3 and 3-2 moves alone [14].

However, little is known about how difficult it is to relate two triangulations by a sequence of Pachner moves. In a series of papers, Mijatović develops upper bounds on the number of moves required for various classes of 3-manifolds [17]. All of these bounds are super-exponential in the number of tetrahedra, and some even involve exponential towers of exponential functions. For relating one-vertex triangulations using only 2-3 and 3-2 moves, no explicit bounds are known at all.

In this paper we focus on one-vertex triangulations of the 3-sphere. Here simplification is tightly linked to the important problem of 3-sphere recognition, where we are given an input triangulation and asked whether represents the 3-sphere. This problem plays an key role in other important topological algorithms, such as connected sum decomposition [11] and unknot recognition [9], and it is now becoming important in computational 4-manifold topology. We can use Pachner moves for 3-sphere recognition in two ways:

• They give us a direct 3-sphere recognition algorithm: try all possible sequences of Pachner moves on up to Mijatović’s upper bound, and return “true” if and only if we reach one of the well-known “canonical” 3-sphere triangulations with one or two tetrahedra.

• They also allow a hybrid recognition algorithm: begin with a fast and/or greedy procedure to simplify as far as possible within a limited number of moves. If we reach a canonical 3-sphere triangulation then return “true”; otherwise run a more traditional 3-sphere recognition algorithm on our new (and hopefully simpler) triangulation.

The direct algorithm lies well outside the realm of feasibility: Mijatović’s bound is super-exponential in the number of tetrahedra, and the running time is at least exponential in Mijatović’s bound. Current implementations [3] use the hybrid method, which is extremely effective in practice. Experience suggests that when is the 3-sphere, the greedy simplification almost always gives a canonical triangulation. If simplification fails, we revert to the traditional algorithm of Rubinstein [22]; although this runs in exponential time, recent improvements by several authors have made it extremely effective for moderate-sized problems [6].1

Our aims in this paper are:

• to measure how easy or difficult it is in practice to relate two triangulations of the 3-sphere using Pachner moves, or to simplify a 3-sphere triangulation to use fewer tetrahedra;

• to understand why greedy simplification techniques work so well in practice, despite the prohibitive theoretical bounds of Mijatović;

• to investigate the possibility that Pachner moves could be used as the basis for a direct 3-sphere recognition algorithm that runs in sub-exponential or even polynomial time.

Fundamentally this is an experimental paper (though the theoretical underpinnings are interesting in their own right). Based on an exhaustive study of million triangulations (including million one-vertex triangulations of the 3-sphere), the answers appear to be:

• we can relate and simplify one-vertex triangulations of the 3-sphere using remarkably few Pachner moves;

• both procedures require us to add at most two extra tetrahedra, which explains why greedy simplification works so well;

• the number of moves required to simplify such a triangulation could also be bounded by a constant, which means polynomial-time 3-sphere recognition may indeed be possible.

These observations are extremely surprising, especially in light of Mijatović’s bounds. If they can be proven in general—yielding a polynomial-time 3-sphere recognition algorithm—this would be a significant breakthrough in computational topology.

In Section 2 we outline preliminary concepts and introduce the Pachner graph, an infinite graph whose nodes represent triangulations and whose arcs represent Pachner moves. This graph is the framework on which we build the rest of the paper. We define simplification paths through the graph, as well as the key quantities of length and excess height that we seek to measure.

We follow in Section 3 with two key tools for studying the Pachner graph: an isomorph-free census of all closed 3-manifold triangulations with tetrahedra (which gives us the nodes of the graph), and isomorphism signatures of triangulations that can be computed in polynomial time (which allow us to construct the arcs of the graph).

Section 4 describes parallel algorithms for bounding both the length and excess height of simplification paths, and presents the highly unexpected experimental results outlined above. We finish in Section 5 with a discussion of the implications and consequences of these results.

## 2Triangulations and the Pachner graph

A 3-manifold triangulation of size is a collection of tetrahedra whose faces are affinely identified (or “glued together”) in pairs so that the resulting topological space is a closed 3-manifold.2 We are not interested in the shapes or sizes of tetrahedra (since these do not affect the topology), but merely the combinatorics of how the faces are glued together. Throughout this paper, all triangulations and 3-manifolds are assumed to be connected.

We do allow two faces of the same tetrahedron to be identified, and we also note that distinct edges or vertices of the same tetrahedron might become identified as a by-product of the face gluings. A set of tetrahedron vertices that are identified together is collectively referred to as a vertex of the triangulation; we define an edge or face of the triangulation in a similar fashion.

Figure 1 illustrates a 3-manifold triangulation of size . Here the back two faces of the first tetrahedron are identified with a twist, the front faces of the first tetrahedron are identified with the front faces of the second using more twists, and the back faces of the second tetrahedron are identified together by directly “folding” one onto the other. This is a one-vertex triangulation since all eight tetrahedron vertices become identified together. The triangulation has three distinct edges, indicated in the diagram by three distinct arrowheads.

Mijatović [17] describes a canonical triangulation of the 3-sphere of size , formed by a direct identification of the boundaries of two tetrahedra. In other words, given two tetrahedra and , we directly identify face with , with , and so on. The resulting triangulation has four faces, six edges, and four vertices.

The four Pachner moves describe local modifications to a triangulation. These include:

• the 2-3 move, where we replace two distinct tetrahedra joined along a common face with three distinct tetrahedra joined along a common edge;

• the 1-4 move, where we replace a single tetrahedron with four distinct tetrahedra meeting at a common internal vertex;

• the 3-2 and 4-1 moves, which are inverse to the 2-3 and 1-4 moves.

These four moves are illustrated in Figure ?. Essentially, the 1-4 and 4-1 moves retriangulate the interior of a pyramid, and the 2-3 and 3-2 moves retriangulate the interior of a bipyramid. It is clear that Pachner moves do not change the topology of the triangulation (i.e., the underlying 3-manifold remains the same). Another important observation is that the 2-3 and 3-2 moves do not change the number of vertices in the triangulation.

Two triangulations are isomorphic if they are identical up to a relabelling of tetrahedra and a reordering of the four vertices of each tetrahedron (that is, isomorphic in the usual combinatorial sense). Up to isomorphism, there are finitely many distinct triangulations of any given size.

Pachner originally showed that any two triangulations of the same closed 3-manifold can be made isomorphic by performing a sequence of Pachner moves [21].3 Matveev later strengthened this result to show that any two one-vertex triangulations of the same closed 3-manifold with at least two tetrahedra can be made isomorphic through a sequence of 2-3 and/or 3-2 moves [14]. The two-tetrahedron condition is required because it is impossible to perform a 2-3 or 3-2 move upon a one-tetrahedron triangulation (each move requires two or three distinct tetrahedra).

In this paper we introduce the Pachner graph, which describes how distinct triangulations of a closed 3-manifold can be related via Pachner moves. We define this graph in terms of nodes and arcs, to avoid confusion with the vertices and edges that appear in 3-manifold triangulations.

It is clear that the arcs are well-defined (since Pachner moves are preserved under isomorphism), and that arcs do not need to be directed (since each 2-3 or 1-4 move has a corresponding inverse 3-2 or 4-1 move). In the full Pachner graph , each arc runs from some level to a nearby level or . In the restricted Pachner graph , each arc must describe a 2-3 or 3-2 move, and must run from some level to an adjacent level . Figure 2 shows the first few levels of the restricted Pachner graph of the 3-sphere.

We can now reformulate the results of Pachner and Matveev as follows:

To simplify a triangulation we essentially follow a path through or from a higher level to a lower level, which motivates the following definition.

Both the length and excess height measure how difficult it is to simplify a triangulation: the length measures the number of Pachner moves, and the excess height measures the number of extra tetrahedra required. For the 3-sphere, the only known bounds on these quantities are the following:

In the restricted Pachner graph, no explicit bounds on these quantities are known at all.

## 3Key tools

Experimental studies of the Pachner graph are difficult: the graph itself is infinite, and even the finite level sets grow super-exponentially in size. By working with isomorphism classes of triangulations, we keep the level sets considerably smaller than if we had used labelled triangulations instead. However, the trade-off is that both the nodes and the arcs of the graph are more difficult to construct.

In this section we outline two key algorithmic tools for studying the Pachner graph: a census of triangulations (which enumerates the nodes at each level), and polynomial-time computable isomorphism signatures (which allow us to construct the arcs).

### 3.1A census of triangulations

To enumerate the nodes of Pachner graphs, we build a census of all 3-manifold triangulations of size , with each triangulation included precisely once up to isomorphism. Because we are particularly interested in one-vertex triangulations as well as triangulations of the 3-sphere, we extract such triangulations into separate censuses with the help of the highly optimised 3-sphere recognition algorithm described in [7]. The final counts are summarised in Table 1.

The algorithms behind this census are sophisticated; see [4] for some of the techniques involved. The constraint that the triangulation must represent a 3-manifold is critical: if we just enumerate all pairwise identifications of faces up to isomorphism, there are at least

possibilities for . To enforce the 3-manifold constraint we use a modified union-find algorithm that tracks partially-constructed edge links and vertex links; see [4] for details.

Even with this constraint, we can prove that the census grows at a super-exponential rate:

The proof is detailed, and is given in the appendix.

For the largest case , the enumeration of all 3-manifold triangulations up to isomorphism required days of CPU time as measured on a single 1.7 GHz IBM Power5 processor (though in reality this was reduced to 2–3 days of wall time using 32 CPUs in parallel). The time required to extract all 3-sphere triangulations from this census was negligible in comparison.

### 3.2Isomorphism signatures

To construct arcs of the Pachner graph, we begin at a node—that is, a 3-manifold triangulation —and perform Pachner moves. Each Pachner move results in a new triangulation , and our main difficulty is in deciding which node of the Pachner graph represents .

A naïve approach might be to search through nodes at the appropriate level of the Pachner graph and test each corresponding triangulation for isomorphism with . However, this approach is infeasible: although isomorphism testing is fast (as we prove below), the sheer number of nodes at level of the graph is too large (see Theorem ?).

What we need is a property of the triangulation that is easy to compute, and that uniquely defines the isomorphism class of . This property could be used as the key in a data structure with fast insertion and fast lookup (such as a hash table or a red-black tree), and by computing this property we could quickly jump to the relevant node of the Pachner graph.

Here we define such a property, which we call the isomorphism signature of a triangulation. In Theorem ? we show that isomorphism signatures do indeed uniquely define isomorphism classes, and in Theorem ? we show that they are small to store and fast to compute.

A labelling of a triangulation of size involves: (i) numbering its tetrahedra from 1 to inclusive, and (ii) numbering the four vertices of each tetrahedron from 1 to 4 inclusive. We also label the four faces of each tetrahedron from 1 to 4 inclusive so that face is opposite vertex . A key ingredient of isomorphism signatures is canonical labellings, which we define as follows.

As an example, consider the triangulation of size described by Table 2. This table lists the precise gluings of tetrahedron faces. For instance, the second cell in the bottom row indicates that face 2 of tetrahedron 3 is glued to tetrahedron 2, in such a way that vertices of tetrahedron 3 map to vertices of tetrahedron 2 respectively. This same gluing can be seen from the other direction by examining the first cell in the middle row.

It is simple to see that the labelling for this triangulation is canonical. The sequence is (reading tetrahedron numbers from left to right and then top to bottom in the table), and tetrahedron 2 first appears before tetrahedron 3 as required. Looking closer, the first appearance of tetrahedron 2 is in the second cell of the top row where vertices map to , and the first appearance of tetrahedron 3 is in the following cell where vertices map to . In both cases the gluings use the identity map.

In summary, we can choose any of the tetrahedra to label as tetrahedron 1, and we can choose any of the labellings of its four vertices. From here the remaining labels are forced, and can be deduced in time. The full proof is given in the appendix.

To encode the full set of face gluings for a canonical labelling, we could simply convert a table of gluing data (such as Table 2) into a sequence of bits. For practical implementations we use a more compact representation, which will be described in the full version of this paper.

It is clear that implies that and are isomorphic, since both signatures encode the same gluing data. Conversely, if and are isomorphic then their canonical labellings are the same (though they might be enumerated in a different order). In particular, the lexicographically smallest canonical labellings will be identical; that is, .

To encode a full set of face gluings, at worst we require a table of gluing data such as Table 2, with cells each containing four integers. Because some of these integers require bits (the tetrahedron labels), it follows that the total size of is .

The algorithm to generate is spelled out explicitly in its definition. The canonical labellings of can be enumerated in time (Lemma ?). Because a full set of face gluings has size , we can encode the bit sequences and select the lexicographically smallest in time, giving a time complexity of overall.

This space complexity of is the best we can hope for, since Theorem ? shows that the number of distinct isomorphism signatures for size triangulations grows like .

It follows from Theorems ? and ? that isomorphism signatures are ideal tools for constructing arcs in the Pachner graph, as explained at the beginning of this section. Moreover, the relevant definitions and results are easily extended to bounded and ideal triangulations (which are beyond the scope of this paper). We finish with a simple but important consequence of our results:

## 4Analysing the Pachner graph

As discussed in the introduction, our focus is on one-vertex triangulations of the 3-sphere. We therefore direct our attention to , the restricted Pachner graph of the 3-sphere.

In this section we develop algorithms to bound the shortest length and smallest excess height of any simplification path from a given node at level of . By running these algorithms over the full census of one-vertex triangulations of the 3-sphere (as described in Section 3.1), we obtain a computer proof of the following results:

The bound is required because there are no simplification paths in starting at level 2 or below (see Figure 2). For a computer proof becomes computationally infeasible.

The results of Theorem ? are astonishing, especially in light of Mijatović’s super-exponential bounds. Furthermore, whilst it can be shown that the excess height bound of is tight, the length estimate of is extremely rough: the precise figures could be much smaller still. These results have important implications, which we discuss later in Section 5.

In this section we describe the algorithms behind Theorem ?, and we present the experimental results in more detail. Our algorithms are constrained by the following factors:

• Their time and space complexities must be close to linear in the number of nodes that they examine, due to the sheer size of the census.

• They cannot loop through all nodes in , since the graph is infinite. They cannot even loop through all nodes at level , since there are too many to enumerate.

• They cannot follow arbitrary breadth-first or depth-first searches through , since the graph is infinite and can branch heavily in the upward direction.4

Because of these limiting factors, we cannot run through the census and directly measure the shortest length or smallest excess height of any simplification path from each node. Instead we develop fast, localised algorithms that allow us to bound these quantities from above. To our delight, these bounds turn out to be extremely effective in practice. The details are as follows.

### 4.1Bounding excess heights

In this section we compute bounds so that, from every node at level of the graph , there is some simplification path of excess height . As in Theorem ?, we compute these bounds for each in the range .

In step 2(a) we construct arcs by performing 2-3 moves. We only construct arcs from nodes already in , which means we only work with a small portion of level for each . In step 2(b) we use union-find to update the number of components in small time complexity.

It is clear that Algorithm ? is correct for any : once we have the subgraph is connected, which means there is a path from any node at level to any other node at level . By Theorem ? at least one such node allows a 3-2 move, and so any node at level has a simplification path of excess height .

However, it is not clear that Algorithm ? terminates: it might be that every simplification path from some node at level passes through nodes that we never construct at higher levels . Happily it does terminate for all , giving an output of each time. Table ? shows how the number of components changes throughout the algorithm in each case.

It is straightforward to show that the space and time complexities of Algorithm ? are linear and log-linear respectively in the number of nodes in (other small polynomial factors in and also appear). Nevertheless, the memory requirements for were found to be extremely large in practice (29 GB), and for they were too large for the algorithm to run (estimated at 400–500 GB). In the case of a two-phase approach was necessary:

1. Use Algorithm ? for the transition from level to level , and terminate if .

2. From each node at level , try all possible combinations of a 2-3 move followed by a 3-2 move. Let be the endpoint of such a combination (so is also a node at level ). If then merge the components and decrement if necessary. Otherwise do nothing (since would never have been constructed in the original algorithm).

3. If after this procedure then output ; otherwise terminate with no result.

It is important to note that, if this two-phase approach does output a result, it will always be the same result as Algorithm ?. Essentially Step 2 simulates the transition from level to in the original algorithm, with the advantage of a much smaller memory footprint (since it does not store any nodes at level ), but with the disadvantage that it cannot move on to level if required (and so it cannot output any result if ).

Of course by the time we reach there are reasons to suspect that (following the pattern for ), and so this two-phase method seems a reasonable (and ultimately successful) approach. For the memory consumption was 50 GB, which was (just) within the capabilities of the host machine.

### 4.2Bounding path lengths

Our next task is to compute bounds so that, from every node at level of , there is some simplification path of length . Once again we compute for .

Because it is infeasible to perform arbitrary breadth-first searches through , we only consider paths that can be expressed as a series of jumps, where each jump involves a pair of 2-3 moves followed by a pair of 3-2 moves. This keeps the search space and memory usage small: we always stay within levels , and , and we never need to explicitly store any nodes above level . On the other hand, it means that our bounds are very rough—there could be much shorter simplification paths that we do not detect.

To identify the initial set we simply attempt to perform 3-2 moves. When we process each node , we must enumerate all jumps out from ; that is, all combinations of two 2-3 moves followed by two 3-2 moves. The number of such combinations is in general.

This time we can guarantee both correctness and termination if . Because the initial set is non-empty (Theorem ?), and from our height experiments in Section 4.1 we know that our search will eventually reach all of level . It follows that every node at level of has a path of length to some , and therefore a simplification path of length . Table ? shows how the search progresses for each .

This time the space and time complexities are linear and log-linear respectively in the number of nodes at level (again with further polynomial factors in ). This is considerably smaller than the number of nodes processed in Algorithm ?, and so for Algorithm ? memory is not a problem: the case runs in under 4 GB.

### 4.3Parallelisation and performance

For , both Algorithms ? and ? have lengthy running times: Algorithm ? requires a very large number of nodes to be processed at levels 9, 10 and 11 of the Pachner graph, and Algorithm ? spends significant time enumerating the available jumps from each node.

We can parallelise both algorithms by processing nodes simultaneously (in step 2 of Algorithm ?, and during each stage of the breadth-first search in Algorithm ?). We must be careful however to serialise any updates to the graph.

The experiments described here used an 8-core 2.93 GHz Intel Xeon X5570 CPU with 72 GB of RAM (using all cores in parallel). With the serialisation bottlenecks, Algorithms ? and ? achieved roughly and CPU utilisation for the largest case , and ran for approximately 6 and 15 days respectively. All code was written using the topological software package Regina [1].

## 5Discussion

As we have already noted, the bounds obtained in Section 4 are astonishingly small. Although we only consider , this is not a small sample: the census includes million triangulations including million one-vertex 3-spheres; moreover, nine tetrahedra are enough to build complex and interesting topological structures [4]. Our results lead us to the following conjectures:

If true, this result (combined with Theorem ?) would reduce Mijatović’s bound in Theorem ? from to for one-vertex triangulations of the 3-sphere. Furthermore, it would help explain why 3-sphere triangulations are so easy to simplify in practice.

There are reasons to believe that a proof might be possible. As a starting point, a simple Euler characteristic argument shows that every closed 3-manifold triangulation has an edge of degree ; using at most two “nearby” 2-3 moves, this edge can be made degree three (the setting for a possible 3-2 simplification). The details will appear in the full version of this paper.

This is a bolder conjecture, since the length experiments are less consistent in their results. However, the fact remains that every 3-sphere triangulation of size can be simplified after just three jumps, and this number does not rise between and .

If true, this second conjecture would yield an immediate polynomial-time 3-sphere recognition algorithm: for any triangulation of size we can enumerate all combinations of three jumps, and test each resulting triangulation for a 3-2 move down to tetrahedra. By repeating this process times, we will achieve either a recognisable 2-tetrahedron triangulation of the 3-sphere, or else a proof that our input is not a 3-sphere triangulation.

Even if Conjecture ? is false and the length bounds do grow with , this growth rate appears to be extremely slow. A growth rate of or even would still yield the first known sub-exponential 3-sphere recognition algorithm (using the same procedure as above), which would be a significant theoretical breakthrough in algorithmic 3-manifold topology.

Looking forward, it is natural to ask whether this behaviour extends beyond the 3-sphere to triangulations of arbitrary 3-manifolds. Initial experiments suggest “partially”: the Pachner graphs of other 3-manifolds also appear to be remarkably well-connected, though not enough to support results as strong as Conjectures ? and ? above. We explore these issues further in the full version of this paper.

## Acknowledgements

The author is grateful to the Australian Research Council for their support under the Discovery Projects funding scheme (project DP1094516). Computational resources used in this work were provided by the Queensland Cyber Infrastructure Foundation and the Victorian Partnership for Advanced Computing.

Benjamin A. Burton
School of Mathematics and Physics, The University of Queensland
Brisbane QLD 4072, Australia
(bab@maths.uq.edu.au)

Here we offer full proofs for Theorem ? and Lemma ?, which were omitted from the main text to simplify the exposition.

An upper bound of is easy to obtain. If we count all possible gluings of tetrahedron faces, without regard for isomorphism classes or other constraints (such as the need for the triangulation to represent a closed 3-manifold), we obtain an upper bound of

Proving a lower bound of is more difficult—the main complication is that most pairwise identifications of tetrahedron faces do not yield a 3-manifold at all [8]. We work around this by first counting 2-manifold triangulations (which are much easier to obtain), and then giving a construction that “fattens” these into 3-manifold triangulations without introducing any unwanted isomorphisms.

To create a 2-manifold triangulation of size (the size must always be even), we identify the edges of distinct triangles in pairs. Any such identification will always yield a closed 2-manifold (that is, nothing can “go wrong”, in contrast to the three-dimensional case).

There is, however, the issue of connectedness to deal with (recall from the beginning of Section 2 that all triangulations in this paper are assumed to be connected). To ensure that a labelled 2-manifold triangulation is connected, we insist that for each , the first edge of the triangle labelled is identified with some edge from one of the triangles labelled . Of course many connected labelled 2-manifold triangulations do not have this property, but since we are proving a lower bound this does not matter.

We can now place a lower bound on the number of labelled 2-manifold triangulations. First we choose which edges to pair with the first edges from triangles ; from the property above we have choices. We then pair off the remaining edges, with possibilities overall. Finally we note that each of the pairs of edges can be identified using one of two possible orientations. The total number of labelled 2-manifold triangulations is therefore at least

Each isomorphism class can contain at most labelled triangulations, and so the number of distinct isomorphism classes of 2-manifold triangulations is bounded below by

We fatten each 2-manifold triangulation into a 3-manifold triangulation as follows. Let denote the closed 2-manifold described by the original triangulation.

1. Replace each triangle with a prism and glue the vertical faces of adjacent prisms together, as illustrated in Figure ?. This represents a bounded 3-manifold, which is the product space .

2. Cap each prism at both ends with a triangular pillow, as illustrated in Figure ?. The two faces of each pillow are glued to the top and bottom of the corresponding prism, effectively converting each prism into a solid torus. This produces the closed 3-manifold , and the complete construction is illustrated in Figure ?.

3. Triangulate each pillow using two tetrahedra, which are joined along three internal faces surrounding an internal vertex. Triangulate each prism using tetrahedra, which again all meet at an internal vertex. Both triangulations are illustrated in Figure ?.

If the original 2-manifold triangulation uses triangles, the resulting 3-manifold triangulation uses tetrahedra. Moreover, if two 3-manifold triangulations obtained using this construction are isomorphic, the original 2-manifold triangulations must also be isomorphic. The reason for this is as follows:

• Any isomorphism between two such 3-manifold triangulations must map triangular pillows to triangular pillows. This is because the internal vertex of each triangular pillow meets only two tetrahedra, and no other vertices under our construction have this property.

• By “flattening” the triangular pillows into 2-dimensional triangles, we thereby obtain an isomorphism between the underlying 2-manifold triangulations.

It follows that, for , we obtain a family of pairwise non-isomorphic 3-manifold triangulations.

This result is easily extended to . Let denote the number of distinct isomorphism classes of 3-manifold triangulations of size .

• Each triangulation of size has at least distinct 2-3 moves available (since any face joining two distinct tetrahedra defines a 2-3 move, and there are at least such faces).

• On the other hand, each triangulation of size has at most distinct 3-2 moves available (since each 3-2 move is defined by an edge that meets three distinct tetrahedra, and the triangulation has at most edges in total).

It follows that for any . This gives for sufficiently large and all , and so we obtain with no restrictions on .

For the result is trivial, since all possible labellings are canonical. For we observe that, if we choose (i) any one of the tetrahedra to label as tetrahedron 1, and (ii) any one of the possible labellings of its four vertices, then there is one and only one way to extend these choices to a canonical labelling of .

To see this, we can walk through the list of faces , where represents face of tetrahedron . The first face amongst that is joined to an unlabelled tetrahedron must in fact be joined to tetrahedron 2 using the identity map; this allows us to deduce tetrahedron 2 as well as the labels of its four vertices.

We inductively extend the labelling in this manner: once we have labelled tetrahedra and their corresponding vertices, the first face amongst that is joined to an unlabelled tetrahedron must give us tetrahedron and the labels for its four vertices (again using the identity map). The resulting labelling is canonical, and all of the labels can be deduced in time using a single pass through the list . The factor is required for manipulating tetrahedron labels, each of which requires bits.

It follows that there are precisely canonical labellings of , and that these can be enumerated in time using iterations of the procedure described above.

### Footnotes

1. See [7] for explicit measurements of running time.
2. It is sometimes useful to consider bounded triangulations where some faces are left unidentified, or ideal triangulations where the overall space only becomes a 3-manifold when we delete the vertices of each tetrahedron. Such triangulations do not concern us here.
3. As Mijatović notes, Pachner’s original result was proven only for true simplicial complexes, but it is easily extended to the more flexible definition of a triangulation that we use here [17]. The key step is to remove irregularities by performing a second barycentric subdivision using Pachner moves.
4. In general, a node at level can have up to distinct neighbours at level .

### References

1. Benjamin A. Burton, Regina: Normal surface and 3-manifold topology software, http://regina. sourceforge.net/, 1999–2010.
2. to3em, Face pairing graphs and 3-manifold enumeration, J. Knot Theory Ramifications 13 (2004), no. 8, 1057–1101.
3. to3em, Introducing Regina, the 3-manifold topology software, Experiment. Math. 13 (2004), no. 3, 267–272.
4. to3em, Enumeration of non-orientable 3-manifolds using face-pairing graphs and union-find, Discrete Comput. Geom. 38 (2007), no. 3, 527–571.
5. to3em, The complexity of the normal surface solution space, SCG ’10: Proceedings of the Twenty-Sixth Annual Symposium on Computational Geometry, ACM, 2010, pp. 201–209.
6. to3em, Optimizing the double description method for normal surface enumeration, Math. Comp. 79 (2010), no. 269, 453–484.
7. to3em, Quadrilateral-octagon coordinates for almost normal surfaces, Experiment. Math. 19 (2010), no. 3, 285–315.
8. Nathan M. Dunfield and William P. Thurston, Finite covers of random 3-manifolds, Invent. Math. 166 (2006), no. 3, 457–521.
9. Masao Hara, Seiichi Tani, and Makoto Yamamoto, is in , SODA ’05: Proceedings of the Sixteenth Annual ACM-SIAM Symposium on Discrete Algorithms, Society for Industrial and Applied Mathematics, 2005, pp. 359–364.
10. William Jaco, David Letscher, and J. Hyam Rubinstein, Algorithms for essential surfaces in 3-manifolds, Topology and Geometry: Commemorating SISTAG, Contemporary Mathematics, no. 314, Amer. Math. Soc., Providence, RI, 2002, pp. 107–124.
11. William Jaco and J. Hyam Rubinstein, 0-efficient triangulations of 3-manifolds, J. Differential Geom. 65 (2003), no. 1, 61–168.
12. William Jaco and Jeffrey L. Tollefson, Algorithms for the complete decomposition of a closed -manifold, Illinois J. Math. 39 (1995), no. 3, 358–406.
13. Bruno Martelli and Carlo Petronio, Three-manifolds having complexity at most 9, Experiment. Math. 10 (2001), no. 2, 207–236.
14. Sergei Matveev, Algorithmic topology and classification of 3-manifolds, Algorithms and Computation in Mathematics, no. 9, Springer, Berlin, 2003.
15. Sergei V. Matveev, Complexity theory of three-dimensional manifolds, Acta Appl. Math. 19 (1990), no. 2, 101–130.
16. to3em, Computer recognition of three-manifolds, Experiment. Math. 7 (1998), no. 2, 153–161.
17. Aleksandar Mijatović, Simplifying triangulations of , Pacific J. Math. 208 (2003), no. 2, 291–324.
18. to3em, Triangulations of Seifert fibred manifolds, Math. Ann. 330 (2004), no. 2, 235–273.
19. to3em, Simplical structures of knot complements, Math. Res. Lett. 12 (2005), no. 5-6, 843–856.
20. to3em, Triangulations of fibre-free Haken 3-manifolds, Pacific J. Math. 219 (2005), no. 1, 139–186.
21. Udo Pachner, P.L. homeomorphic manifolds are equivalent by elementary shellings, European J. Combin. 12 (1991), no. 2, 129–145.
22. J. Hyam Rubinstein, An algorithm to recognize the -sphere, Proceedings of the International Congress of Mathematicians (Zürich, 1994), vol. 1, Birkhäuser, 1995, pp. 601–611.
23. Abigail Thompson, Thin position and the recognition problem for , Math. Res. Lett. 1 (1994), no. 5, 613–630.
You are adding the first comment!
How to quickly get a good reply:
• Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
• Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
• Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters