Higher Order Bipartiteness vs Bi-Partitioning in Simplicial Complexes

Abstract

Bipartite graphs are a fundamental concept in graph theory and have significant applications in data analysis and modeling complex processes. A graph is bipartite if its vertices can be divided into two sets, with no connections within each set. Detecting bipartite graphs can be done by examining their cycles; a graph is bipartite if and only if there are no odd cycles present. This characteristic is essential in various problem domains, such as clustering and matching. Bipartite graphs can be identified through the spectrum of the Laplacian, a graph is bipartite if and only if the maximum eigenvalue of the normalized Laplacian equals two. However, graphs are limited to representing pairwise interactions, to model higher-order connections in complex systems, hypergraphs and simplicial complexes are required. This raises the question of what is bipartiteness for simplicial complexes of higher dimensions - whether obtaining the maximum eigenvalue of the normalized Laplacian is equivalent to partitioning the simplexes into two, similar to the graph case. We address this problem by fully characterizing those complexes for which the maximum eigenvalue of the Laplacian is obtained, known as disorientable simplicial complexes, based on the parity of the length of cycles in their down dual graph. We show that a $N$-dimensional simplicial complex is disorientable if and only if its down dual graph has no simple odd cycle of distinct edges and no twisted even cycle of distinct edges. We note that having no odd cycle is the key to answering the bi-partitioning problem of the simplexes.

1 introduction

Bipartite graphs are a useful tool in diverse domains, from matching problems and coding theory to social networks and biomedical applications such as cancer detection[1, 2, 3]. Exploring the properties of bipartite graphs, and their applications not only has enriched graph theory but also has opened the door to innovative solutions in challenging applications [1]. In the past few years, there has been a trend in understanding cycles in networks in terms of function, dynamics, and synchronizability [4]. Cycles and bipartite graphs are closely related. The vertices of a bipartite graph can be partitioned into two sets such that no two vertices inside the same set are joined by an edge. Equivalently a graph is bipartite if and only if it has no odd cycles. One can test for a graph being bipartite by observing the spectrum of the graph Laplacian, a matrix computed based on the incidence relations between vertices and edges of the graph. For the normalized Laplacian, the graph is bipartite if and only if two is the maximum eigenvalue of its Laplacian. Consider the spectrum of the normalized Laplacian of an unweighted graph in increasing order, $0=\lambda_{1}\leq...\leq\lambda_{n}\leq 2$; the graph is connected if and only if $\lambda_{2}$ is bigger than zero and is bipartite iff $\lambda_{n}$ is equal to two. While the down side of the spectrum tells the number of connected components (i.e. graph’s topology), the up side reveals bipartiteness. Both the top and bottom eigenvalues then can help us to partition the vertices of the graph into two sets in two different ways; if $\lambda_{2}$ is positive but small we can partition the vertices into two sets with very few edges connecting the sets while elements of each set are highly connected to each other and they form clusters. And when $\lambda_{n}$ is close to two we can partition the set of the vertices into two sets in a way that for each of these sets, there are few connections, and almost all of the edges, connect the vertices of one set to the other (Fig.1.).

Figure 1: Partitioning vertices based on the Laplacian spectrum. In (a) vertices are partitioned based on $\lambda_{n}$ and in (b) based on the $\lambda_{2}$.

Many empirical networks and complex systems however incorporate higher-order relations between elements and therefore are modeled as hypergraphs and/or simplicial complexes, rather than graphs. Simplicial complexes are generalizations of graphs where there are not only vertices and edges but also possibly triangles, tetrahedrons, and so on. To develop a systematic tool for the structural analysis of simplicial complexes, different methods, and theories have been extended from graphs to these higher-order structures. An important question is that can the graph partitioning scheme that we described, be generalized to simplicial complexes? The main starting point to answer this question is the discrete ”Hodge” Laplacian which is a generalization of the graph Laplacian for simplicial complexes. There have been many advances in our understanding of the spectrum of discrete Hodge Laplacian on simplicial complexes in the past few years [5, 6, 7]. Moreover, Laplacian-based methods have become popular for detecting the structure and dynamics of complex networks modeled by simplicial complexes in the past few years[8]. It is well known, due to Eckmann [9], that for any fixed $d$ where $N\geq d\geq 0$ ($N$ is the dimension of the complex), the minimum eigenvalue of the $d$-Laplacian can tell us about the topology of the complex in dimension $d$. This theorem has led to many theoretical findings regarding the minimum eigenvalues of the Laplacian on simplicial complexes[6, 7] as well as clustering methods for $d$-simplexes in such a way that the clusters represent the topology of the complex in dimension $d$ [10, 11]. This is a generalization of what was described at the beginning based on the down side of the Laplacian spectrum for clustering the vertices of a graph. However, there is very little known about the maximum eigenvalue of the Laplacian and specifically the higher dimensional analogues for bipartiteness. A higher-order notion of bipartiteness called ”disorientability”, was introduced in [12] as a structure that the spectrum of the (up) Hodge Laplacian achieves its maximum possible value; the graph is the simplest setting and a one-dimensional simplicial complex is disoriantable if and only if it is bipartite. It is known that simple random walks on the vertices of a connected bipartite graph are periodic. It has been shown that the same is true when the random walks are on higher dimensional simplexes of a general simplicial complex and such random walks are periodic if and only if the complex is disorientable [12, 13, 14]. Random walks on graphs and simplicial complexes are the main tools in tackling diverse complex real-world problems from ranking web pages in Google page rank algorithm [15] to signal processing and flow network decomposition[11]. Considering the numerous theoretical implications and practical applications that bipartite graphs have, it is very desirable to develop our understanding of disorientability in simplicial complexes not just in terms of the spectrum of the Laplacian, but in terms of the combinatorial structure and shape of the complex and the parity of length of higher order cycles and see if we can have similar simplified characterization for recognizing whether a simplicial complex is disorientable and does disorientability give some kind of bi-portioning of the simplicies of higher dimensions in a similar manner like bipartite graphs. Furthermore, since more data sets and some complex problems are nowadays represented by simplicial complexes and not graphs it seems necessary to develop the applied partitioning methods for partitioning the simplexes of dimensions higher than zero and explore further the applications, and devise algorithms.

In this manuscript, we state conditions under which the simplicial complex is disorientable, in terms of the length of the higher dimensional cycles in the complex. Consequently, any simplicial complex can become disorientable by a finite number of splittings of some of the maximum dimensional simplexes; note that this is very similar to the graph case where by dividing some of the edges into two, and playing with the parity of the length of their cycles, we can always make the graph bipartite. And the fewer number of splittings needed to make the complex bipartite, the closer the maximum eigenvalue would be to the possible maximum eigenvalue of the Laplacian in the corresponding dimension. To the best of our knowledge, this is the first full characterization of disorientability from the combinatorial structure of the higher-dimensional cycles in simplicial complexes. We hope this characterization paves the way for the development of methods and effective algorithms as well as exploring new applications when dealing with beyond-binary interactions in a variety of complex systems and modeling problems.

First, we recall some preliminary notions.

2 Background

Simplicial complexes- A simplicial complex $K$ on a vertex set $V=\{v_{1},...,v_{n}\}$ consists of a collection of simplices, that is, subsets of $V$ with the requirement that all these subsets are closed under inclusion. A subset that contains $d+1$-vertices, is called a $d$-simplex. A $0$-simplex is simply a vertex, a $1$-simplex is an edge, and a $2$-simplex is a triangle. The dimension of the complex is the maximum $d$, where we have at least one $d$-simplex in K. For computational purposes, we need to define an orientation for each d-simplex when $d\geq$ 1. An orientation of a $d$-simplex is an equivalence class of orderings of its vertices, where two orderings are equivalent if they differ by an even permutation. For simplicity, we can choose the reference orientation of the simplices induced by the ordering of the vertex labels. A $0$-simplex (a node) can have only one orientation. Hence, issues of orientation do not arise in graph-theoretic settings. Let $S_{d}$ be the set of all $d$-simplexes in $K$ with $0\leq d\leq N$ and $[S_{d}]$ be the set of all oriented $d$-simplexes. For any $d$-simplex $\sigma$ ($d\neq 0$) we have two opposite orientations, clockwise and counterclockwise (note that both are in $[S_{d}]$). A face of a $d$-simplex $\sigma$ is a subset of $\sigma$ with cardinality $d$, i.e., with one element of $\sigma$ omitted. If $\rho$ is a face of a $d$-simplex $\sigma_{d}$, then $\sigma_{d}$ is called a co-face of $\rho_{d-1}$. The degree of a $d$-simplex is the number of its co-faces (i.e., of dimension $d+1$). A $d$-simplex is called branching if its degree is bigger than $2$. A $d$-cycle of length $l$ ($0\leq d\leq N-1$) is a chain of $d$ simplexes, $\sigma^{0}_{d}...,\sigma^{l}_{d}$ such that for each $i$, $\sigma^{i}$ and $\sigma^{i+1}$ are upper adjacent (i. e. they share a co-face) and $\sigma^{0}=\sigma^{l}$. The cycle is non-twisted (or simple) if all of the vertices of $\sigma^{0}$ coincide exactly with themselves in $\sigma^{l}$ and is twisted if at least for one vertex such coincidence does not happen. Note that graphs can not have twisted cycles and this just can happen in higher dimensions. We call the simplicial complex non-twisting if it does not have a twisted cycle and otherwise, we call it twisting. A discrete cylinder and torus with respectively simple $0$-cycle(s) and $1$-cycle(s) are examples of non-twisting simplicial complexes and discrete Mobius strip and Klein bottle are twisting as they have respectively a $0$-twisting cycle and a $1$-twisting cycle. Twisting cycles do not naturally arise in geometric and topological data analysis (TDA) when constructing Simplicial complexes from a set of data points. However, we present it here for the sake of completeness of our characterization.
Boundary and co-boundary matrices and the Laplacian- We can extend face/co-face relations to the oriented $d$-simplexes with the help of the boundary/coboundary operators. The $d$-th chain group $C_{d}(K)$ of $K$ is a vector space with real coefficients with the basis $S_{d}$. The boundary map $\partial_{d}:C_{d}(K)\rightarrow C_{d-1}(K)$ is a linear operator defined by

$$\partial_{d}[i_{0},...i_{d}]=\sum_{j=0}^{d}(-1)^{j}[i_{0},.i_{j-1},i_{j+1}..i_{d}].$$ (2.1)

After choosing a basis for $C_{d}$, the boundary operator $\partial_{d}$ can be represented by a matrix $B_{d}$, which enables us to simply perform computations via matrix calculus. For graphs, the matrix $B_{1}$ is the node-to-edge incidence matrix. Likewise, the higher-order boundary maps can be interpreted as higher-order incidence matrices between simplices and their (co-)faces for each $B_{i}$. The transpose of the boundary matrix $B_{i}$ denoted by $B_{i}^{T}$ represents the co-boundary map $\partial_{d}^{T}:C_{d}\rightarrow C_{d+1}$. Subsequently, we can define the $d$-th discrete (combinatorial) Hodge Laplacian as

$$L_{d}=B_{d}^{T}B_{d}+B_{d+1}B_{d+1}^{T}.$$ (2.2)

For $d=0$, $L_{0}=B_{1}B_{1}^{T}$ and for $d=N$, $L_{N}=B_{N}^{T}B_{N}$. $B_{d}^{T}B_{d}$ and $B_{d+1}B_{d+1}^{T}$ are respectively called the down and up Laplacians.

Dual graphs- We can create dual graphs of simplicial complexes based on the upper adjacency and/or lower adjacency connections between the simplexes. For this purpose, we consider a N-simplicial complex $K$ and we fix a number $d$ ( $0\leq d\leq N)$. The up-dual graph $G$ of $K$ in dimension $d$ is constructed as follows: each $d$-simplex in $K$ becomes a vertex in the dual graph $G$, and there is an edge between two vertices in $G$ if their corresponding d-simplexes in $K$ share a co-face. Similarly, the down-dual graph $G$ of $K$ in dimension $d$ is a graph $G$ that its vertices are the $d$-simplexes of $K$ but there is an edge between two vertices in $G$ if their corresponding d-simplexes in $K$ share a face. Particularly, the down-dual graph of a graph is called a line graph. We call a $N$-complex $d$-connected if the up dual graph in dimension $d$ is connected.

Higher order bipartiteness- There are two types of generalizations for bipartiteness for a $N$-simplicial complex; a popular notion is called disorientability. A disorientation of a $N$-complex $K$ is a choice of the orientation of its $N$-simplexes, that whenever two arbitrary $N$-simplexes intersect in a $(N-1)$-simplex, they induce the same orientation on it (see Fig.2.). If $K$ has a disorientation it is said to be disorientable. A graph is disorientable iff it is bipartite.

Figure 2: Example of a 2-d disorientable simplicial complex.

Another natural but less popular analogue is “$(d+1)$-partiteness”: having some partition $A_{0},...,A_{d}$ of $V$ so that every $d$-simplex contains one vertex from each $A_{i}$. A $(d+1)$-partite complex is seen to be disorientable, but the opposite does not necessarily hold for $d\geq 2$ [12]. Since this second case is not related to the spectrum of the Laplacian and does not have the theoretical implications that disorientability has we will omit it here.

3 Higher Order Bipartiteness

To explore disorientable $N$-dimensional simplicial complexes we use their down-dual graphs and we start with graphs as 1-d complexes. As already mentioned in [12] bipartite graphs are disorientable and in fact, these two are equivalent in graphs. We elaborate on this from a signs perspective and the line graph of a graph.
If a graph is disorientable, there is a choice of orientations on its edges that adjacent edges induce the same orientation on their common vertex. Equivalently, there is an assignment of $\pm$ to vertices such that no two adjacent vertex have the same sign, namely we have a bipartite graph where each partition is labeled with one of these signs representing the head/tails of the oriented edges. Now let’s look at the line graph of a bipartite graph. We have the following simple observation:

Lemma 3.1.

The line graph of a non-branching bipartite graph is bipartite.

Proof.

We note that when going to line graphs of a general graph, its cycles and the parity of their length are preserved. But we might get some more cycles in the line graph that did not exist before. A moment of pose clarifies that such cycles are obtained due to the existence of branching vertices, i.e. those vertices that have deg bigger or equal than 3. For any such vertex $v$ with degree $\alpha\geq 3$ we would have a sequence of cycles with lengths $\alpha$, $\alpha-1$,.., three. So if there exists no branching vertex, there is a one-to-one correspondence between the cycles in the graphs and its dual (line graph). Therefore if the graph is bipartite and non-branching its line graph is also bipartite. ∎

As the result of the above Lemma, the disoientability of the graph is equivalent to assigning $+$ and $-$ to the edges of its line graph such that no two adjacent edges have the same sign. But what if the bipartite graph has some branching vertices? In this case, the line graph is not bipartite and has (fundamental) odd cycles that for every such cycle, all of its edges relate to the same corresponding branching vertex. We note that if we consider general (not necessarily fundamental) cycles the existing odd cycles include more than one edge corresponding to that branching vertex.
Therefore, the graph is disorientable iff it is bipartite or equivalently iff its line graph has no odd cycle of distinct edges, namely they correspond to different vertices in the original graph. In the language of signs, a graph is disorientable iff in its line graph, we can assign + and - to the edges such that no two adjacent edges that correspond to different vertices in the original graph, get the same sign.

Figure 3: A bipartite graph and its line graph. (a) is bipartite and (b) is the line graph of (a) and has two odd cycles of length 3 corresponding to the branching red and green vertices in (a).

This simple but fundamental change of perspective from graphs to their line graphs helps us to explore bipartiteness in general simplicial complexes. We start with the simplest cases: non-branching and non-twisting simplicial complexes and we gradually develop the idea for the general case.

1. If the simplicial complex is non-branching and non-twisting:

   Lemma 3.2.

   A non-branching, non-twisting simplicial complex is dissorientable if and only if in its down dual graph, we can assign + and - to the edges in such a way that no two adjacent edges have the same sign. We note that this is equivalent to the condition that the down-dual graph has no odd cycle which is the same as bipartiteness.

   Proof.

   If the simplicial complex is disorientable, by definition there is an assignment of orientations on $N$-simplexes such that they induce the same orientation on their common face. Therefore, in its corresponding down dual graph, we would be able to have a single assignment of +/- to each edge representing the induced orientations on the common $(N-1)$-faces (in the simplicial complex) such that no two adjacent edges get the same sign. This means that the down-dual graph has no odd cycle. Therefore it is bipartite. For the reverse, if there is such assignment of +/- to the edges of the down-dual graph, by definition the simplicial complex is disorientable as each such sign for each edge can be interpreted as the unique induced sign on the common N-1 simplex (from its oriented N-simplexes cofaces). ∎

   Remark 3.3.

   The above lemma can be considered as a higher-dimensional version of the two-coloring problem on graphs. We note that in this case, we can color the max-dimensional simplexes with two colors in such a way that no two adjacent simplexes that share a face have the same color. Recall that a graph can be colored by two colors if and only if it is bipartite. Also planer graphs (graphs that can be drawn without any of their edges crossing) can be colored using at most four colors, such that no two adjacent nodes have the same color.

2. The simplicial complex is branching and non-twisting:

   Lemma 3.4.

   A branching non-twisting simplicial complex is disorientable if and only in its down dual graph, either all of the edges of an odd cycle correspond to a branching $(N-1)$-simple and/or that odd cycle includes more than one edge corresponding to that branching simplex.

   Proof.

   Similar to the branching bipartite graphs that are already described, for every branching $(N-1)$-simple$\alpha$, in the down dual graph we would have a cycle of length= degree$\alpha$, as well as all cycles with length between three and degree of $\alpha$, such that all of the edges of these cycles correspond to $\alpha$ or its subsets. Therefore, all of the edges corresponding to all of these cycles in the down dual graph are adjacent and get the same sign and when moving to another adjacent edge (that corresponds to another $(N-1)$-simple), the sign is changed. Therefore, we might have odd cycles, but they can just be for the branching $(N-1)$-simplexes, and no other fundamental odd cycle will exist. Also, when considering general cycles (not necessarily fundamental ones) odd cycle will include more than one edge corresponding to that branching simplex. ∎

3. The simplicial complex is non-branching:

   Lemma 3.5.

   A non-branching simplicial complex is dissorientable if and only if in its down dual graph, all the twisting cycles have odd lengths and all the non-twisting cycles have even lengths.

   Proof.

   It is straightforward to see that for the twisting cycles, the twisted part acts as an extra (hidden) $N$-dimensional simplex where the twisted connection corresponds to two faces of this simplex, with opposite orientations. Therefore, every twisted cycle as opposed to simple cycles should have odd length to induce the dissorientability condition along the cycle. The rest of the proof will be as before and we omit it here. ∎

As the direct result of the above lemmas, we have the following theorem that enables us to check the dissorientability of a general simplicial complex based on the length of the cycles of its down dual graph.

Theorem 3.6.

A simplicial complex is dissorientable if and only if in its down dual graph, the (possible) odd cycles only correspond to the branching of some of $(N-1)$-simplexes and/or twisted cycles and there is no twisted cycle of even length.

Theorem 3.7.

Every simplicial complex can become dissorientable by a finite number of splittings of some of the $N$-dimensional simplexes into two $N$-dimensional simplexes.

Proof.

We orient all of $N$-simplexes one by one in a compatible manner, namely in such a way that the adjacent $N$-simplexes induce the same orientation on their common face. If we can achieve this goal globally and for all of the lower-adjacent $N$-simplexes, we are done; i.e., the simplicial complex is disorientable. If not, there are at least two $N$-simplexes, $A$ and $B$, such that they induce opposite orientations on their common face. We note that this means $A$ and $B$ are adjacent vertices in the down dual graph and based on the previous theorem both are included in a cycle of odd (even for the twisting cycles) length. We chose one of them arbitrarily (A) and we split it into two $N$-simplexes $A^{\prime}$ and $A^{\prime\prime}$, where $A^{\prime\prime}$ and $B$ have non-empty intersections. We orient $A^{\prime}$ based on the orientation of A in such a manner that $A$ and $A^{\prime}$ have exactly the same orientation on the shared face. We then orient $A^{\prime\prime}$ in a compatible manner with $A^{\prime}$ where they induce the same orientation on their common face. Then trivially $A^{\prime\prime}$ will have a compatible orientation with $B$, meaning they induce the same orientation on their common face. If we repeat this process for any two non-compatible N-simplex we get our desired result. We note that by splitting $A$, we make the length of the simple cycle in the down-dual graph including $A^{\prime},A^{\prime\prime},B$ even, and the length of the twisted cycles will become odd. Also if the incompatibility is happening in a branching $(N-1)$-simplex, we might need to split more than one of the cofaces of such simplex; this splitting of course will not change the number of branching and consequently will not affect the length of its corresponding cycle in the down-dual graph. ∎

Remark 3.8.

Note that if splitting one/some of the $N$-simplex(es) divides its (their) free $(N-1)$-face(s) into two, then this will not have an effect on the other $N$-simplexes. But if we split the $N$-simplex in such a way that it divides its non-free face into two, due to the simplicial structure, this will also affect all other $N$-dimensional cofaces of the divided $(N-1)$ simplex, namely those which are lower adjacent to the original $N$-simplex (as shown in the next example).

Example.

Figure 4: Making simplicial complexes disorintable and the effect on the cycles of their corresponding down-dual graphs for Mobius strip and tetrahedron. The down-dual graph of each complex is drawn in its bottom. (a) is a non-disorientable Mobius strip as (b) has a twisting even cycle. (c) is obtained by splittings of two simplexes of (a), presented by red arrows, which is disorientable as in (d) twisting cycle(s) are even and simple cycle(s) are odd. Similarly, (e) and (g) are non-disorientable as based on respectively (f) and (h) they have odd simple cycles. (i) is obtained from (e) by splitting of all its four 2-simplexes and is disorientable as (j) has no odd simple simple.

In the example shown in Figure 4, we show how to make simplicial complexes disorintable and the effect on the cycles of their corresponding down-dual graphs.

We have two main rows of simplicial complexes (with blue) and the down-dual graph of each example is drawn at the bottom of each complex. The top row is the discrete Mobius strip and the bottom row corresponds to a tetrahedron.

Conclusions- We have fully characterized disorientability of simplicial complexes of any dimension in terms of the parity of the length of cycles in their down dual graphs. As a direct result, every simplicial complex can become dissorientable by a finite number of splittings of some of the $N$-dimensional simplexes into two $N$-dimensional simplexes (see Fig.4.) in a similar manner that any graph can become bipartite by a finite number of splitting of some of its edges into two. Such splittings do not change the topology of the complex.

This innovative perspective is new and is simple but fundamental as it allows to extension of a range of theoretical and applied cycle-based methods from bipartite graphs to higher dimensional disorientable simplicial complexes.