Connectivity Bounds for GHZ Graphs

Let d_{v}(u) denote the number of neighbours of a vertex u \in G. In G_c^w, let d_{\mathbf{i}}(u) denote the number of edges of colour i incident on u. Let \Delta_{\mathbf{i}}(G_c^w) = \max\limits_{v \in V(G)} d_{\mathbf{i}}(v) and \delta_{\mathbf{i}}(G_c^w) = \min\limits_{v \in V(G)} d_{\mathbf{i}}(v). We prove the following three lemmas first.

Since \Delta_v(G) \leq |V(G)|-1, Lemmas ? and ? imply Theorem 6.

Towards a contradiction, suppose that there is a loopless connected multigraph G'' such that

Choose a colouring c and a weight function w witnessing this dimension. Delete all redundant edges and then all redundant colour classes. The dimension is unchanged. Let the resulting graph be G. Then

Moreover, G is a subgraph of G''. Suppose that G were disconnected, with component vertex sets C_1,\ldots,C_s, where s\geq 2. For each feasible monochromatic colour i, Observation 12, with its factors grouped by components of G, gives 1=w(\mathbf{i}_{V(G)})=\prod_{t=1}^{s}w(\mathbf{i}_{C_t}). If two such colours i,j existed, assigning colour i to C_1 and colour j to all other components would, by Observation 11 and the same component factorisation, give the non-zero weight w(\mathbf{i}_{C_1})\prod_{t=2}^{s}w(\mathbf{j}_{C_t}). This contradicts the GHZ condition. Thus a disconnected G would satisfy \mu(G,c,w)\leq 1, contradicting \lambda(G'')\geq 1 and \mu(G,c,w)\geq\lambda(G'')+1. Hence G is connected. In particular,

Let (S,S') be a minimum edge cut of G, so S'=V(G)\setminus S and |(S,S')|=\lambda(G). Since G has a perfect matching, |V(G)| is even, and therefore |S| and |S'| have the same parity.

First suppose that |S| and |S'| are odd. Every monochromatic perfect matching must contain at least one edge of the cut (S,S'). Therefore every feasible monochromatic colour appears on at least one edge of the cut. Hence

contradicting \mu(G,c,w)\ge \lambda(G'')+1.

Now suppose that |S| and |S'| are even. Let A be the set of feasible colours for which no monochromatic perfect matching crosses the cut (S,S'), and put r=|A|. Let B be the remaining feasible colours and put r'=|B|. Thus

For every colour i\in B, some colour-i monochromatic perfect matching crosses the cut. Since both sides of the cut have even cardinality, such a matching uses at least two cut edges of colour i. Therefore

We claim that r\le 1. Suppose not, and choose two colours 1,2\in A. Since no colour-i monochromatic perfect matching crosses the cut for i=1,2, all such matchings split between the two sides of the cut. Hence

In particular, w(1_S) and w(2_{S'}) are non-zero.

Consider the inherited vertex colouring vc in which every vertex of S gets colour 1 and every vertex of S' gets colour 2. Any perfect matching inducing vc contains no cut edge: a cut edge in an ordinary edge-colouring would give the same colour to its two endpoints. Therefore, by Observation 11,

This is impossible because vc is non-monochromatic. Hence r\le 1.

If r'\ge 1, then

again contradicting \mu(G,c,w)\ge \lambda(G'')+1. If r'=0, then \mu(G,c,w)=r\le 1, also a contradiction. Therefore no such G'' exists.

If G is a complete graph, from Observation 14, \mu(G) \leq |V(G)|-1. As \kappa(G)=|V(G)|-1, it is easy to see that the given statement holds. We now consider simple connected graphs not isomorphic to complete graphs.

Since T is a minimum vertex separator, |T|=\kappa(G), and since G is simple we have |E(T)|\leq \binom{\kappa(G)}{2}. Hence Theorem 9 follows.

Towards a contradiction, let \mu(G) \geq 3. Therefore, there exists a colouring c and weight function w, such that \mu(G,c,w)\geq 3. Let three of its non-redundant colours be 1, 2 and 3. Consider the union of a colour 1 perfect matching and a colour 2 perfect matching. Their union is a disjoint union of even cycles whose edges alternate between colours 1,2. Since G is simple and the two matchings are edge-disjoint, every such cycle has length at least 4. With only 6 vertices, their union must therefore be a single 6-cycle. Let the vertices of such a 6 cycle be labelled abcdef in clockwise direction. Without loss of generality, let c(ab)=c(cd)=c(ef)=1 and c(bc)=c(de)=c(fa)=2.

We call (a,d), (b,e), and (c,f) the antipodal pairs. We claim that any perfect matching M formed by non-cyclic edges must have at least one edge containing an antipodal pair. Suppose not. Observe that any edge connecting a vertex from \{a,c,e\} to \{b,d,f\} is either a cyclic edge or contains an antipodal pair. Thus M would have to match the vertices of \{a,c,e\} among themselves and the vertices of \{b,d,f\} among themselves, which is impossible because both sets have odd cardinality.

Therefore, for the monochromatic matching of colour 3, there is an edge containing an antipodal pair, say ad, without loss of generality. Consider the non-monochromatic inherited vertex colouring in which a,d are coloured 3, b,c are coloured 2 and e,f are coloured 1. By Observation 11, the weight of such a vertex colouring is w(ad)w(bc)w(ef). Since the vertex colouring is non-monochromatic, w(vc) = 0. Therefore, one of these three edges must have weight zero. Contradiction.

Towards a contradiction, let \mu(G) \geq 4. Therefore, there exists a colouring c and weight function w, such that \mu(G,c,w)\geq 4. Let four of its non-redundant colours be 1 (red), 2 (green), 3 (blue) and 4 (brown). Consider the union of a perfect matching of two different colours. Their union is a disjoint union of even cycles with alternating colours. Since G is simple and the matchings are edge-disjoint, every such cycle has length at least 4. With 8 vertices, their union is therefore either an 8-cycle or two 4-cycles.

Case 1. We first consider the case where the union of any two monochromatic perfect matchings of different colours is always two disjoint 4-cycles. Consider a red and a green perfect matching. Their union forms two disjoint cycles 1234 and 5678, where

are red and

are green.

Consider the non-monochromatic vertex colouring in which 1234 gets colour red and 5678 gets colour green. By Observation 11, this colouring has weight

and this weight must be zero. Hence one of the two factors is zero. Without loss of generality, assume

Since the red edges (1,2) and (3,4) are present, this forces the red edges (1,3) and (2,4) to be present and

Now consider the vertex colouring in which 1234 gets colour green and 5678 gets colour red. Again by Observation 11, one of w(2_{\{1,2,3,4\}}) and w(1_{\{5,6,7,8\}}) must be zero. But

Therefore

This forces the red edges (5,7) and (6,8) to be present and

We know that there should be a blue perfect matching. Note that all possible edges within \{1,2,3,4\} and within \{5,6,7,8\} are red or green. Therefore, all blue edges must have one vertex from 1,2,3,4 and the other from 5,6,7,8. Without loss of generality, let (2,5) be an edge from the blue perfect matching.

• Case 1.1 The blue perfect matching contains (3,8). Then the blue perfect matching can be completed in two ways as shown in ? and ?.

  • Case 1.1.1 (?.) There is a blue-red alternating Hamiltonian cycle 16834752 which is a union of two perfect matchings of different colours. By assumption such perfect matchings are not present in this case. Contradiction.

  • Case 1.1.2 (?.) There is a blue-red alternating Hamiltonian cycle 17834652 which is a union of two perfect matchings of different colours. By assumption such perfect matchings are not present in this case. Contradiction.

• Case 1.2 Vertex 3 can have an edge with 6 or 7. Note that both are identical red neighbours and hence without loss of generality let the edge be with 7. Then the blue perfect matching can be completed in two ways.

  • Case 1.2.1 (?.) There is a blue-red alternating Hamiltonian cycle 16843752. It is a union of two perfect matchings of different colours. Contradiction.

  • Case 1.2.2 (?.) There is a blue-red alternating Hamiltonian cycle 12564378. It is a union of two perfect matchings of different colours. Contradiction.

Case 2. We classify the remaining cases with respect to the fixed alternating red-green Hamiltonian cycle. Label its vertices 1,\ldots,8 in cyclic order. For a monochromatic perfect matching M of a colour i\notin\{\text{red},\text{green}\}, call an edge of M crossing if removing its endpoints from the Hamiltonian cycle leaves two paths of even order. Say that i is crossing if some such M contains a crossing edge. There are two cases.

1. Some remaining non-redundant colour, say blue, is not crossing. Then every edge of every blue monochromatic perfect matching joins vertices of the same parity in the cyclic ordering.

2. Every non-redundant colour other than red and green is crossing.

Case 2.1. Fix a blue monochromatic perfect matching M_b. Since blue is not crossing, M_b pairs vertices of equal parity: the four odd vertices among themselves and the four even vertices among themselves. Up to dihedral symmetry of the Hamiltonian cycle and interchange of red and green, its three possible types are:

These are precisely the three cases shown in ?.

Case 2.1.1 (?). Consider the non-monochromatic vertex colouring in which 1,2,3,8 get colour blue and 4,5,6,7 get colour green. By Observation 11, the weight of this colouring is

This contradicts the GHZ condition, since every non-monochromatic inherited vertex colouring must have weight 0.

Case 2.1.2 (?). Consider the non-monochromatic vertex colouring in which 1,2 get the colour red, 4,5 get colour green and 3,6,7,8 get the colour blue. From Observation 11, the weight of such a colouring is w((1,2))w((4,5))w((3,7))w((6,8)) \neq 0. But a non-monochromatic vertex colouring should have weight 0. Contradiction.

We now record the structure forced by a crossing edge in a selected monochromatic perfect matching M_i, for a colour i other than red and green. Up to a dihedral symmetry and interchange of red and green, let (8,3) be a crossing edge of M_i. Consider the non-monochromatic vertex colouring assigning colour i to 3,8, red to 1,2, and green to 4,5,6,7. By Observation 11, its weight is w((8,3))w((1,2))w(\mathbf{2}_{\{4,5,6,7\}})=0. The first two factors are non-zero, so w(\mathbf{2}_{\{4,5,6,7\}})=0. The green matching (4,5),(6,7) has non-zero weight; its only possible alternative is (4,6),(5,7), since (5,6) is red. Thus (4,6) and (5,7) must both be green. After (8,3) is removed from M_i, four unmatched vertices lie in \{4,5,6,7\} and only two lie outside it. Therefore M_i contains an edge within \{4,5,6,7\}. The only available pair is (4,7): all other pairs already have red or green edges, and G is simple. The remaining vertices 1,2,5,6 can be matched in exactly two ways, (1,5),(2,6) or (1,6),(2,5), because (1,2),(5,6) are red. These are the two crossing-matching structures shown in ?; the derivation applies to any colour i, including brown.

Case 2.1.3. This is the case shown in ?. Now consider the fourth colour, which is brown.

Case 2.1.3.1 If brown is not crossing, no brown monochromatic perfect matching contains a crossing edge. The union of any selected brown monochromatic perfect matching with the fixed red-green Hamiltonian cycle must then be one of the three structures from ?. Further, as all antipodal pairs have a blue edge between them, the union of a brown perfect matching with a Hamiltonian cycle with alternating red and green colours must be like that of ?, which was already proven to be not possible.

Case 2.1.3.2 If brown is crossing, select a brown monochromatic perfect matching containing a crossing edge. By the crossing-edge derivation above, its union with the fixed red-green Hamiltonian cycle must be one of the two structures from ?. Further, as all antipodal pairs have a blue edge between them, the union of a brown perfect matching with a Hamiltonian cycle with alternating red and green colours must be like that of ?. Now consider the non-monochromatic vertex colouring in which 1,2,5,6 get the colour brown, and 3,4,7,8 get the colour blue. From Observation 11 and Observation 12, we know that the weight of such a vertex colouring is w((1,6))w((2,5))w((3,7))w((4,8))\ne 0. Contradiction.

Case 2.2. In this case, every non-redundant colour other than red and green is crossing. Select a blue monochromatic perfect matching and a brown monochromatic perfect matching, each containing a crossing edge. The crossing-edge derivation above applies separately to both selected matchings, so each has one of the two structures in ?, up to a symmetry of the fixed red-green Hamiltonian cycle. Their combined configurations are shown in ?.

The selected blue and brown matchings are edge-disjoint because the graph is simple. Up to the symmetries of the alternating Hamiltonian cycle and interchange of the names blue and brown, the possible pairs of blue and brown matchings are exactly the three configurations shown in ?.

Case 2.2.1 and Case 2.2.2 (? and ?) In both configurations, the following argument applies. Consider the non-monochromatic vertex colouring in which the vertices 5,6,7,8 are red, 1,4 are brown, and 2,3 are green. From Observation 11, this is possible only if the edges (5,7) and (6,8) are red. Now consider the non-monochromatic vertex colouring in which vertices 4,5,6,7 are green, vertices 1,2 are red, and vertices 8,3 are blue. From Observation 11, this is possible only if the edges (5,7) and (6,4) are green. But (5,7) is already red. Contradiction.

Case 2.2.3 (?) Consider the vertex colouring in which 1,2,5,6 get the colour red and 3,4,7,8 get the colour blue. From Observation 11 and Observation 12, the weight of the vertex colouring is w((1,2))w((5,6))w((3,7))w((4,8))\neq 0. Contradiction.

Therefore, \mu(G) \leq 3.