Abstract 1 Introduction 2 Preliminaries 3 Our Main Result 4 Implications for Restricted Subclasses 5 Open Problems References

Approximating Barnette’s Conjecture

Michael A. Bekos ORCID Department of Mathematics, University of Ioannina, Greece Michael Kaufmann ORCID Department of Computer Science, University of Tübingen, Germany Maximilian Pfister ORCID Department of Computer Science, University of Tübingen, Germany
Abstract

A well-known conjecture, named after David W. Barnette, asserts that every 3-regular, 3-connected, bipartite, planar graph (for short, Barnette graph) is Hamiltonian. As another step towards addressing Barnette’s conjecture positively, we show that every n-vertex Barnette graph admits a subhamiltonian cycle containing 5⁢n6 edges, improving upon the previous bound of 2⁢n3. Equivalently, every Barnette graph admits a 2-page book embedding in which at least 5⁢n6 consecutive vertex pairs along the spine are connected by edges. As a byproduct, we present a simple proof for a known result that guarantees the existence of Hamiltonian cycles in a certain subclass of Barnette graphs.

Keywords and phrases:
Barnette’s Conjecture, Subhamiltonicity, Book embeddings
Funding:
Michael A. Bekos: Supported by the HFRI grant 26320.
Maximilian Pfister: Supported by the DFG grant SCHL 2331/1-1.
Copyright and License:
[Uncaptioned image] © Michael A. Bekos, Michael Kaufmann, and Maximilian Pfister; licensed under Creative Commons License CC-BY 4.0
2012 ACM Subject Classification:
Mathematics of computing → Combinatorics
; Mathematics of computing → Graph theory
Editors:
Vida Dujmović and Fabrizio Montecchiani

1 Introduction

Barnette’s conjecture [3] is an intriguing open problem in topological graph theory; it asserts that every 3-regular, 3-connected, bipartite, planar graph (for short, Barnette graph) is Hamiltonian, i.e., it contains a simple cycle visiting each vertex exactly once. The conjecture is inspired by two earlier conjectures by Tait [12] and Tutte [13], both of which were disproven.

While Barnette’s conjecture [3] remains open, several related results have been established. For instance, Goodey [7] showed that Barnette graphs with faces of degree 4 and 6 are Hamiltonian. Feder and Subi [5] extended this to Barnette graphs whose faces are 3-colored such that two colors contain only degree-4 and degree-6 faces. The authors in [2] proved Hamiltonicity under certain (improper) 2-colorings of the dual faces. Alternative forms of the conjecture have also been studied. Kelmans [11] showed that Barnette’s conjecture is equivalent to the existence of a Hamiltonian cycle containing exactly one of any pair of edges on a common face. Other forms of equivalency include the existence of a Hamiltonian cycle through (i) any edge [8], (ii) any path of length 3 with two consecutive edges on the boundary of a face [9], and (iii) the middle edge of any path of length 3 avoiding its outer edges [9].

Using a computer-assisted method, Brinkmann, Goedgebeur and McKay [4] verified that all Barnette graphs with up to 90 vertices are Hamiltonian. Their proof relies on the fact that all Barnette graphs can be derived from the cube graph via cube- and C4-expansions (see Figure 1) [10]. As the former operation preserves Hamiltonicity, the main challenge in Barnette’s conjecture lies in handling the latter; see a related open problem at Section 5.

(a)
(b)
(c)
Figure 1: Starting from (a) the cube graph all Barnette graphs can be generated with either (b) cube-expansions or (c) C4-expansions.

Our contribution.

Since Hamiltonicity in a planar graph is equivalent to admitting a 2-page book embedding with all consecutive vertices (including the first and the last) connected, studying subhamiltonian cycles (that is, linear vertex orders in such embeddings) offers a relaxed yet meaningful perspective. In this context, it is known that every n-vertex Barnette graph admits a subhamiltonian cycle containing at least 2⁢n3 edges [1]. Building on this, we revisit and refine the approach of [1], improve the bound to 5⁢n6, and discuss its potentials as a step toward resolving Barnette’s conjecture. Note that a bound of n settles the conjecture.

2 Preliminaries

Basic definitions.

We assume familiarity with basic concepts from Graph Theory and Graph Drawing. A k-page book embedding of a graph defines an order of its vertices and partitions its edges into k sets of non-crossing edges. A graph admitting a 2-page book embedding is called subhamiltonian, as it can be augmented to Hamiltonian by adding edges between consecutive vertices. The vertex order of such an embedding is called a subhamiltonian cycle.

Constructing Barnette graphs.

As mentioned, all Barnette graphs can be constructed by iteratively applying two operations, cube-expansions and C4-expansions, starting from the cube graph (see Figure 1). In a cube-expansion, a vertex x is replaced by seven new vertices x1,…,x7 and the following edges: (x1,y1), (x2,y2), (x3,y3), (x1,x4), (x1,x5), (x4,x6), (x5,x6), (x2,x4), (x2,x7), (x3,x7), (x6,x7), and (x3,x5), where y1, y2, and y3 are the neighbors of x. In a C4-expansion, two edges e1 and e2 with an odd distance along the boundary of a face are each subdivided twice, and the resulting new vertices are connected with two additional edges by preserving both planarity and bipartiteness.

The algorithm by Alam et al. [1]

Let G be an embedded Barnette graph with n vertices. By Lemma 2 of [1], the edges and the faces of G can be colored with three colors (red, blue and green) with the following properties (refer to Fig. 3(a) of [1] for an illustration): (i) every facial cycle of G is bichromatic, i.e., its edges alternate between two colors along its boundary, (ii) no two neighboring faces are of the same color, (iii) every face of G is colored differently from its bounding edges, (iv) the edge that bounds two faces of G colored x and y is of color z, where {x,y,z}={red,blue,green}, (v) the subgraph Gb⁢g∗ of the dual G∗ of G induced by the blue and green faces of G is connected. Let 𝒯b⁢g∗ be a spanning tree of Gb⁢g∗. This tree together with property (iv) yields a partition of the red edges of G into those that are crossed by 𝒯b⁢g∗ (corresponding to those belonging to 𝒯b⁢g∗) and the remaining ones that are not crossed by 𝒯b⁢g∗. Then, the algorithm by Alam et al. [1] computes a subhamiltonian cycle C of G by a traversal of 𝒯b⁢g∗, assuming that 𝒯b⁢g∗ is rooted at an arbitrary blue leaf ρ as follows. For an arbitrary edge (u,v) of G that is crossed by 𝒯b⁢g∗, let p and q be the faces that are bounded by (u,v), that is, (p,q) is an edge of 𝒯b⁢g∗. The removal of (p,q) from 𝒯b⁢g∗ results in two trees 𝒯p∗ and 𝒯q∗, such that w.l.o.g. ρ belongs to 𝒯p∗. For the recursive step, it is assumed that a simple and plane subhamiltonian cycle Cp for the subgraph Gp of G induced by the vertices of the faces of G in 𝒯p∗ has been computed, which does not necessarily span all vertices of Gp, but it satisfies the following invariant properties:

  1. I.1

    The edge (u,v) is on Cp.

  2. I.2

    Every red edge of Gp is in the interior of Cp or on Cp.

  3. I.3

    Every blue edge of Gp is in the exterior of Cp or on Cp.

  4. I.4

    Every green edge of Gp is on Cp.

  5. I.5

    For every red edge e of Gp that is not crossed by 𝒯b⁢g∗ and that is adjacent to two faces h and h′ of G, with h∈𝒯p∗ and h′∉𝒯p∗, one of the following holds:

    1. (a)

      if h is a blue face, then both endpoints of e are on Cp,

    2. (b)

      if h is a green face, then none of the endpoints of e is on Cp.

(a) p is green, q blue
(b) p is blue, q is green
Figure 2: The solid (dotted) gray edges belong to 𝒯b⁢g∗ (Gb⁢g∗∖𝒯b⁢g∗). The bold (plain) red edges are (are not, respectively) crossed by 𝒯b⁢g∗. Cycle Cq is drawn dotted black. Figure reproduced from [1].

Depending on the colors of faces p and q, the traversal of 𝒯b⁢g∗ is continued by extending the cycle Cp to vertices of face q (thus obtaining a new cycle 𝒞q) and by maintaining Invariants I.1-I.5 as illustrated in Figure 2. Once the traversal of 𝒯b⁢g∗ has been completed, a simple and plane cycle 𝒞 has been computed, in which, by Invariants I.2-I.4, every green edge of G is on 𝒞, while every red (blue) edge of G is in the interior (exterior) of cycle 𝒞 or on 𝒞. Since 𝒯b⁢g∗ is a spanning tree of Gb⁢g∗, every green edge of G bounds a face that is in 𝒯b⁢g∗, and by Invariant I.4 both its endpoints are consecutive along 𝒞. As every vertex is incident to a green edge, it follow that cycle 𝒞 is indeed subhamiltonian. The following properties directly follow from Invariants I.4 and I.5, respectively (see [1]).

Property 1 ([1]).

Every green edge of G belongs to the subhamiltonian cycle 𝒞.

Property 2 ([1]).

Every red edge of G that is not crossed by 𝒯b⁢g∗ belongs to the subhamiltonian cycle 𝒞.

The authors in [1] used Properties 1 and 2 to show that every Barnette graph admits a subhamiltonian cycle with 2⁢n3 edges.

3 Our Main Result

To prove our main theorem, i.e., that every Barnette graph admits a subhamiltonian cycle containing at least 5⁢n6 edges, we first establish a few more properties of the algorithm by Alam et al. [1]. To this end, let G be a Barnette graph and let f be a face of it. By Property (i) of the edge coloring of G, the edges on the boundary of f alternate between two colors. We say that two edges of the same color c are consecutive along f if and only if they appear consecutively in a clockwise traversal of the edges of color c in f.

Property 3.

A blue edge appearing between two consecutive red edges of a green face of G that are crossed by 𝒯b⁢g∗ belongs to the subhamiltonian cycle 𝒞.

Proof.

Let fg be a green face of G which contains a blue edge eb appearing between two consecutive red edges er and er′ of fg that are crossed by 𝒯b⁢g∗. Since er and er′ are crossed by 𝒯b⁢g∗, after the algorithm processes face fg in the traversal of 𝒯b⁢g∗, by Invariant I.1 both edges er and er′ are on the subhamiltonian cycle constructed so far. By Invariant I.3, eb is either in the exterior or on the subhamiltonian cycle constructed so far. Since eb appears between er and er′ along fg, it follows that eb cannot be in its exterior. Since the subhamiltonian cycle is not extended along blue edges (as those are not crossed by 𝒯b⁢g∗), eb will be part of 𝒞. ◀

Property 4.

Every green face of G that corresponds to a leaf of 𝒯b⁢g∗ contributes a red edge (that is crossed by 𝒯b⁢g∗) to the subhamiltonian cycle 𝒞.

Proof.

Let fg be a green face of G that corresponds to a leaf of 𝒯b⁢g∗ and let fb be its parent in 𝒯b⁢g∗. Just before the algorithm processes face fg in the traversal of 𝒯b⁢g∗, the red edge of G that is crossed by the edge (fg,fb) of 𝒯b⁢g∗ necessarily belongs to the subhamiltonian cycle constructed so far by Invariant I.1. Since fg is a leaf of 𝒯b⁢g∗ and since fg is green, the subhamiltonian cycle will not be extended any further after the algorithm processes face fg by Invariant I.5b. Hence, the red edge of G that is crossed by the edge (fg,fb) of 𝒯b⁢g∗ will remain on the subhamiltonian cycle as desired. ◀

Property 5.

Every green 4-face of G that does not correspond to a leaf of 𝒯b⁢g∗ contributes two blue edges to the subhamiltonian cycle 𝒞.

Proof.

Let fg be a green 4-face (i.e., of degree 4) of G and assume that fg does not correspond to a leaf of 𝒯b⁢g∗. Since fg is a 4-face of G, it consists of two red and two blue edges that alternate along its boundary. Since fg does not correspond to a leaf of 𝒯b⁢g∗, both red edges are crossed by 𝒯b⁢g∗. Now the statement directly follows from Property 3. ◀

Property 6.

Every green 6-face of G that does not correspond to a leaf of 𝒯b⁢g∗ contributes at least one blue edge to the subhamiltonian cycle 𝒞.

Proof.

Let fg be a green 6-face of G that does not correspond to a leaf of 𝒯b⁢g∗. Hence, its boundary contains at least two red edges that are crossed by 𝒯b⁢g∗. These two red edges are consecutive on the boundary of fg and thus the result follows by Property 3. ◀ So far, 𝒯b⁢g∗ has only been assumed to be a spanning tree of Gb⁢g∗. In the following, we show that it can be chosen in such a way that the preceding property generalizes to all green faces.

Property 7.

There exists a choice of the tree 𝒯b⁢g∗ such that every green face of G that does not correspond to a leaf of 𝒯b⁢g∗ contributes at least one blue edge to the subhamiltonian cycle 𝒞.

Proof.

Let fg be a green face of G that is not a leaf in 𝒯b⁢g∗. Since fg is not a leaf, its boundary contains at least two red edges that are crossed by 𝒯b⁢g∗. Our goal is to locally modify 𝒯b⁢g∗ so that these two red edges appear consecutively along the boundary of fg; Once this goal is achieved, we can apply Property 3 to obtain the desired result. Denote by r1,…,rk the red edges on the boundary of fg in clockwise order. Let also e1,…,ek be the corresponding dual edges of Gb⁢g∗. Finally, denote by vg the vertex of Gb⁢g∗ corresponding to fg. W.l.o.g. assume that r1 is crossed by 𝒯b⁢g∗, that is, e1 belongs to 𝒯b⁢g∗. Let rj be the edge on the boundary of f that is crossed by 𝒯b⁢g∗, such that j>1 and j is minimum. Note that rj exists, because fg is not a leaf in 𝒯b⁢g∗. If j=2 or j=k holds, then r1 and r2 are consecutive and we are done. Thus, assume 2<j<k. Consider the graph 𝒯b⁢g∗+e2. Since 𝒯b⁢g∗ is spanning, 𝒯b⁢g∗+e2 contains a cycle. Let e be the edge of this cycle (distinct from e2) that is also incident to vg. If e≠e1, then replacing e with e2 in 𝒯b⁢g∗ yields a new spanning tree in which e1 and e2 are consecutive, as desired. Hence, suppose e=e1. In this case, we obtain a new spanning tree by replacing e1 with e2 in 𝒯b⁢g∗ and we observe that the distance between the relevant pair of red edges crossed by this tree has been decreased. Repeating this argument a finite number of times will eventually yield a configuration in which the two red edges are consecutive, completing the proof. ◀ We are now ready to prove the main theorem of this section.

Theorem 1.

Every n-vertex Barnette graph admits a subhamiltonian cycle with at least 5⁢n6 edges.

Proof.

Let G be an n-vertex Barnette graph and let Fg, Fr and Fb be the set of green, red and blue faces of G, respectively. Since G is 3-regular, Euler’s formula directly yields that G has (i) 3⁢n2 edges out of which n2 are green, n2 are red and n2 are blue, and (ii) n2+2 faces. Hence, |Fg|+|Fr|+|Fb|=n2+2. Assume w.l.o.g. that |Fg|≥|Fr|≥|Fb|. By this assumption, Fg contains at least 13⁢(n2+2) faces. Thus, for some x>0, we may write:

|Fg|=n6+x (1)

Since |Fg|≥|Fr|≥|Fb|, it follows:

|Fr|≥n6−x2+1 (2)

We next claim that the average degree of the faces in Fg is at most 6. To see this, observe that edges on the boundary of the faces of Fg are in total n, since by Properties (iii) and (iv) of the edge coloring of G, they correspond to the red and the blue edges of G. Thus, by Eq.(1) the average degree of the faces in Fg is nn6+x, which is at most 6 as claimed. It follows that for every face in Fg of degree greater than 6, there is at least one face in Fg which has degree 4. With these observations in mind, we are now ready to count the number of edges contained in the subhamiltonian cycle 𝒞.

  • ■

    By Property 1, subhamiltonian cycle 𝒞 contains all green edges of G, which are n2 in total.

  • ■

    By Property 2, subhamiltonian cycle 𝒞 contains all red edges of G not crossed by 𝒯b⁢g∗. We claim that these are at least n6−x2. To see this, recall that all red edges of G are n2 in total. Since 𝒯b⁢g∗ is a spanning tree of 𝒢b⁢g∗, the number of red edges crossed by 𝒯b⁢g∗ is:

    |Fb|+|Fg|−1=n2+2−|Fr|−1=n2−|Fr|+1.

    Therefore, the number of red edges of G which are not crossed by 𝒯b⁢g∗ is:

    n2−(n2−|Fr|+1)=|Fr|−1⁢≥(2)⁢n6−x2
  • ■

    Each green face of G contributes an additional edge to 𝒞. To see this, observe that a green face corresponding to leaf in 𝒯b⁢g∗ contributes at least one (red) edge to 𝒞 by Property 4. Each of the remaining faces contributes at least one (blue) edge by Properties 5 and 7.

From the discussion above, it follows that the number of edges belonging to the subhamiltonian cycle 𝒞 is at least

n2+n6−x2+n6+x=5⁢n6+x2

which concludes the proof. ◀

4 Implications for Restricted Subclasses

In this section, we focus on certain restricted subclasses of Barnette graphs for which the algorithm yields a Hamiltonian cycle.

Observation 2.

Let G be a Barnette graph. If 𝒯b⁢g∗ is such that any green face f of G is (i) a leaf in 𝒯b⁢g∗ or (ii) its degree in 𝒯b⁢g∗ is exactly half of the degree of f, then G is Hamiltonian.

Proof.

A green leaf f contains exactly 12⁢|f| red edges, one of which is crossed by 𝒯b⁢g∗. The crossed edge will be part of 𝒞 due to Property 4, while the others are part of 𝒞 by construction. A green non-leaf f has, by assumption, maximum degree in 𝒯b⁢g∗, i.e., all its red edges are crossed by 𝒯b⁢g∗. But then Property 3 directly implies that 12⁢|f| blue edges of f are part of 𝒞. Hence, for any green face, half of its boundary edges are part of 𝒞. Since the faces in Fg are disjoint and cover all red and blue edges, they include exactly n edges. Summing over all green faces, this contributes a total of n/2 edges to 𝒞. Together with all green edges (which are n/2 many), this means that |𝒞|=n and thus G is Hamiltonian. ◀ The following known result from the literature due to Florek [6] can hence be obtained as a direct corollary of Observation 2.

Corollary 3.

A Barnette graph is Hamiltonian if its faces are 3-colored, with adjacent faces having different color, such that one of the three colors contain only faces of degree 4.

Proof.

W.l.o.g. let all green faces have degree 4. Hence, a green face is either a leaf in 𝒯b⁢g∗ or its degree is 2 in 𝒯b⁢g∗, i.e., maximum. Thus, Observation 2 yields the desired result. ◀

5 Open Problems

In this work, we approached Barnette’s conjecture from the lens of 2-page book embeddings. Our work, while not resolving Barnette’s conjecture at this time, makes another step forward and raises new open problems and possible directions for Barnette’s conjecture:

  1. 1.

    Improve the lower bound of 5⁢n6 on the size of the subhamiltonian cycle guaranteed by Theorem 1; a more strategic choice of the underlying spanning tree may lead to a better result in this direction (although our preliminary attempts were not successful).

  2. 2.

    It is also worth asking whether, for a given Barnette graph with a given Hamiltonian cycle, there is always a choice of the spanning tree 𝒯b⁢g∗ such that our algorithm yields the given Hamiltonian cycle.

  3. 3.

    Is it true that all Barnette graphs generated solely by C4-expansions are Hamiltonian?

  4. 4.

    The previous question is interesting also when the faces have a certain maximum degree.

  5. 5.

    Extend the set of restricted subclasses for which Observation 2 can be shown. Potential candidates are known subclasses from the literature which admit a Hamiltonian cycle.

References

  • [1] Muhammad Jawaherul Alam, Michael A. Bekos, Vida Dujmović, Martin Gronemann, Michael Kaufmann, and Sergey Pupyrev. On dispersable book embeddings. Theor. Comput. Sci., 861:1–22, 2021. doi:10.1016/J.TCS.2021.01.035.
  • [2] Helmut Alt, Michael S. Payne, Jens M. Schmidt, and David R. Wood. Thoughts on Barnette’s conjecture. Australas. J Comb., 64:354–365, 2016. URL: http://ajc.maths.uq.edu.au/pdf/64/ajc_v64_p354.pdf.
  • [3] David W. Barnette. Conjecture 5. In W. T. Tutte, editor, Recent Progress in Combinatorics. Academic Press, New York, 1969. Proc. of the Third Waterloo Conference on Combinatorics.
  • [4] Gunnar Brinkmann, Jan Goedgebeur, and Brendan D. McKay. The minimality of the Georges-Kelmans graph. Math. Comput., 91(335):1483–1500, 2022. doi:10.1090/MCOM/3701.
  • [5] Tomas Feder and Carlos Subi. On Barnette’s conjecture. Technical report, Electronic Colloquium on Computational Complexity (ECCC), 2006.
  • [6] Jan Florek. Remarks on Barnette’s conjecture. J. Comb. Optim., 39(1):149–155, 2020. doi:10.1007/S10878-019-00460-8.
  • [7] P. R. Goodey. Hamiltonian circuits in polytopes with even sided faces. Israel Journal of Mathematics, 22(1):52–56, March 1975. doi:10.1007/bf02757273.
  • [8] A. Hertel. Hamiltonian cycles in sparse graphs. Master’s thesis, University of Toronto, 2004.
  • [9] A. Hertel. A survey and strengthening of Barnette’s conjecture, 2005. Undergraduate thesis.
  • [10] Derek A. Holton, Bennet Manvel, and Brendan D. McKay. Hamiltonian cycles in cubic 3-connected bipartite planar graphs. J. Comb. Theory B, 38(3):279–297, 1985. doi:10.1016/0095-8956(85)90072-3.
  • [11] A. K. Kelmans. Konstruktsii kubicheskih dvudolnyh 3-svyaznyh bez gamiltonovyh tsiklov. Sbornik Trudov Vserossiiskogo NII Sistemnykh Issledovanii, 10:64–72, 1986.
  • [12] P. G. Tait. Listing’s topologie. Philosophical Magazine, 17:30–46, 1884. Reprinted in Scientific Papers, Vol. II, pp. 85–98.
  • [13] W. T. Tutte. On hamiltonian circuits. Journal of the London Mathematical Society, 21(2):98–101, 1946. doi:10.1112/jlms/s1-21.2.98.