Complexity Framework for Forbidden Subgraphs II: Edge Subdivision and the "H"-graphs

Let G be a T-subgraph-free graph. We call vertices of degree 2 in G white and vertices of degree at least 3 black. The black graph is a subgraph of G induced by black vertices and a black component is a connected component in the black graph.

We first describe some helpful rules to solve the problem and a set of reductions simplifying the input graph, i.e. reductions transforming G into a graph G' that has fewer edges and/or vertices and that has a Hamiltonian cycle if and only if G has. We emphasize that by deleting an edge or a vertex from an H-subgraph-free graph, we obtain an H-subgraph-free graph again.

We start with some obvious rules:

• (R1) if the graph has vertices of degree 0 or 1, then stop: G has no Hamiltonian cycle.

• (R2) if the graph contains a vertex adjacent to more than two white vertices, then stop: G has no Hamiltonian cycle.

• (R3) if the graph is disconnected, then stop: G has no Hamiltonian cycle.

• (R4) if the graph contains a vertex v adjacent to exactly two white vertices, then delete the edges connecting v to all other its neighbours (if there are any).
  

Now we introduce a reduction applicable to a graph G containing an induced subgraph shown on the left in Figure ?, in which vertices a,b,c have degree 3 in G. The reduction depends on the degree of x. If the degree of x is also 3, the reduction consists in deleting the edges ab and xc. Otherwise, it transforms the graph as shown in Figure ?. We refer to this reduction as the diamond reduction and denote it by (R5).

One more reduction is illustrated in Figure ?. We will refer to this reduction as the butterfly reduction and will denote it by (R6).

In our algorithm to solve the problem we implement the above rules and reductions whenever they are applicable. We now develop more reductions allowing us to bound the number of vertices in black components. We assume that none of the above rules and reductions is applicable to G.

Application of Claim ? to vertices of large degree either shows that G has no Hamiltonian cycle or reduces the input graph to a graph of maximum degree 12. We will refer to this reduction as the large degree reduction and will denote it by (R7).

Since graphs of diameter D and maximum degree \Delta have fewer than \frac{\Delta}{\Delta-2}(\Delta -1)^D vertices, we conclude that after eliminating vertices of large degree, every black component has fewer than \frac{12}{10}11^7 vertices.

To develop more rules and reductions, assume G has a Hamiltonian cycle C. We can further assume that not all vertices of the graph are black, since otherwise the graph contains fewer than \frac{12}{10}11^7 vertices, in which case we can solve the problem by brute-force. A sequence of consecutive vertices of C surrounded by white vertices will be called a black interval. Observe that each black interval consists of at least 2 vertices (according to (R4)).

Let K be a black component of G. We will call the vertices of K that have white neighbours the contact vertices. Note that K may consists of one or more intervals. Each interval gives rise to exactly two contact vertices, and hence the number of contact vertices in K is even.

Let us show that for T-subgraph-free graphs, then number of intervals is at most 2.

By Claim ?, if G has a Hamiltonian cycle, then every black component has two or four contact vertices.

• (R8) If a black component K has exactly two contact vertices, check if K has a Hamiltonian path connecting the contact vertices. If such a path does not exist, then stop: the input graph has no Hamiltonian cycle. Otherwise, choose arbitrarily a Hamiltonian path connecting the contact vertices, include the edges of the path in the solution and delete all other edges from K.

Rule (R8) destroys black components with two contact vertices, i.e. after its implementation all vertices in such components become white.

Now we discuss the case where each black component has exactly four contact vertices. Let K be such a component with contact vertices v_1,v_2,v_3,v_4. If G has a Hamiltonian cycle, then the vertices of K can be partitioned into two parts each of which forms a path connecting a pair of contact vertices. We will call such a partition a pairing (of contact vertices) and will refer to a pairing as the set of edges in the two paths. Also, we will say that two pairings are of the same type, if they pair the contact vertices in the same way. Clearly, if all possible pairings in K have the same type, then it is irrelevant which one to choose, since non-contact vertices of K have no neighbours outside of K.

The above discussion justifies the following two rules.

• (R9) If a black component K with four contact vertices does not admit any pairing, then stop: the input graph has no Hamiltonian cycle.

• (R10) If in a black component K with four contact vertices all possible pairings have the same type, then choose arbitrarily any such pairing and delete all other edges from K. If this procedure disconnects the graph, then stop: the input graph has no Hamiltonian cycle.

Finally, we analyse the situation when each black component of G admits pairings of at least two different types.

We summarize the discussion in the following algorithm to solve the problem.

• 1. Apply rules and reductions (R1) – (R7) as long as they are applicable.

• 2. If the algorithm did not stop at Step 1 and the graph has fewer than \frac{12}{10}11^7 vertices, then solve the problem by brute-force. Otherwise, check the number of contact vertices in black components. If there is a black component with the number of contact vertices different from 2 or 4, then stop: G has no Hamiltonian cycle.

• 3. If the algorithm did not stop at Step 2, then apply (R8) to black components with two contact vertices, and (R9) and (R10) to black components with four contact vertices.

• 4. If the algorithm did not stop at Step 3, then find a Hamiltonian cycle according to Claim ?.

Reductions (R8), (R9), (R10) can be implemented in constant time, because the number of vertices in each black component is bounded by a constant. It is also obvious that all other rules, and hence all steps of the algorithm can be implemented in polynomial time. The correctness of the algorithm follows from the proofs of the claims.

First, branch on all O(2^k n^{3k}) options (so a polynomial number, as k is fixed) of solution paths that have at most three internal vertices. For each branch, we remove the guessed solution paths and the neighbours of the vertices on these paths. Let k still be the number of terminal pairs. We now only look for solution paths with at least four vertices. Branch on all O(n^{4k}) options of choosing the first two vertices a_z,b_z on the solution path from every terminal z\in \{s_i,t_i\} for i\in \{1,\ldots,k\}. In each branch, we remove all other neighbours of z,a_z from the graph, so every terminal z now has degree 1, while a_z has degree 2. We discard the branch if (\dagger) \{a_z,b_z\} \cap \{a_{z'},b_{z'}\} \not= \emptyset for some terminals z,z' or one of a_z,b_z is the same or neighbours one of a_{z'},b_{z'} for some terminals z,z' not from the same terminal pair.

We now start a recursive procedure. We first preprocess the input. If b_{s_i} and b_{t_i} are adjacent for some i \in \{1,\ldots,k\}, then we remove the solution path s_i,a_{s_i}, b_{s_i}, b_{t_i},a_{t_i},t_i and their neighbours from the graph. If b_{z} and b_z' are adjacent for some terminals z,z' that do not form a terminal pair, we discard the branch.

We run the polynomial-time algorithm for k-Disjoint Paths from [RS95] on the remaining terminal pairs. If this results in a no-answer, then we discard the branch. Else, we found a solution P_1,\ldots,P_k. We may assume that each path P_i is induced, or we may shortcut it. If P_1,\ldots,P_k is also a solution of k-Induced Disjoint Paths, then we return a yes-answer. Otherwise, there is (say) an edge (x_1,x_2) between paths x_1 \in P_1 and x_2 \in P_2. We pick x_1 such that it is closest to t_1 on P_1 and under that condition we pick x_2 such that it is closest to t_2 on P_2.

Let z_1,z_3 be the two neighbours of x_1 on P_1 and z_2,z_4 the two neighbours of x_2 on P_2. We let S=\{z_1,x_1,z_3,z_2,x_2,z_4\}. Observe that S does not contain any terminal by \dagger and the preprocessing. Suppose any z \in \{z_1,z_2,z_3,z_4\} has two neighbours outside of S. Then G has a \ensuremath{\mathbb{H}}_2 as a subgraph. Thus we may assume (\ddagger) that each z \in \{z_1,z_2,z_3,z_4\} has at most one neighbour outside of S.

By the choice of (x_1,x_2) and the fact that P_1 is induced, z_3 has no neighbours in S except x_1. Suppose the edge (z_1,z_2) exists and one of \{x_1,x_2\} has a neighbour outside of S. Then there is an \ensuremath{\mathbb{H}}_2 with middle path x_1,z_1,z_2 since s_2 \not\in S. Suppose the edge (z_1,z_4) exists and one of \{x_1,x_2\} has a neighbour outside of S. Then there is an \ensuremath{\mathbb{H}}_2 with middle path x_1,z_1,z_4 since s_2 \not\in S.

Now either the edges (z_1,z_2) and (z_1,z_4) do not exist (see Figure ?) or at least one of them exists and x_1,x_2 have no neighbours outside S (see Figure ?). In the former case, we will apply Rule 1, while in the latter case, we will apply Rule 2 (see Figure ? and ? for the description of these two rules). Using \ddagger, we can prove that Rules 1 and 2 are safe and that they both preserve \dagger and H_2-subgraph-freeness (\spadesuit). We can recognize both rules and apply them in polynomial time. This makes the instance smaller by one vertex and we recurse. Hence, our algorithm will terminate in polynomial time.

As every graph that does not map to C_5 must contain a C_5-critical subgraph, we get from Theorem 17 that \ensuremath{\mathbb{H}}_3-subgraph-free C_5-colouring is solved by detecting the non-existence of the graphs K_3, E_1, E_2, E_3 and F_n (odd n\geq 3) in a \ensuremath{\mathbb{H}}_3-subgraph-free input graph G. We can detect the presence of the fixed graphs K_3, E_1, E_2 and E_3 in G in polynomial time by brute force. What we must show is that we can detect the presence of any odd C_5-flower F_n in polynomial time. To do so, we simply observe that for a fixed centre vertex, v_0 we can make an auxiliary graph on its neighbours putting an edge between two if there is a path on three edges between them in G. Now, G contains an odd C_5-flower with centre v_0 if and only if this auxiliary graph contains an odd cycle. We can check this in polynomial time for each v_0, so can find an odd C_5-flower in G polynomial time.

We observe that for every graphs G and every p\geq 1, G_p is a no-instance of Hamilton Cycle unless G was a cycle.

Under any kind of subdivision, k-Induced Disjoint Paths reduces to k-Disjoint Paths over the same graph, which is in {\sf P} for all k\geq 2, as shown in [Sh80] for k=2 and in [RS95] for every k\geq 3.

We first prove that for all p\geq 4, and for all x,y \in V(C_5), there is a walk of length p in C_5 from x to y. First let p=4. To walk a distance of zero: walk two forward then two back. To walk at distance one (without loss of generality) forward: walk four backward. To walk at distance two (without loss of generality) forward: walk one back, one forward, and two forward. Now let p=5. To walk a distance of zero: walk five forward. To walk at distance one (without loss of generality) forward: walk two forward, two back and one forward. To walk at distance two (without loss of generality) forward: walk one back, one forward, and three back. Finally, let p\geq 6. Keep moving one forward then one back until one of the two previous cases applies.

Now let G be a graph. We give each vertex in G a label from \{1,\ldots,5\}. From the above it follows that for every p\geq 3, we can extend c to a homomorphism from G_p to C_5; in other words, G_p is a yes-instance of C_5-Colouring.

Let G be a graph. We show that for all p\geq 3, G_p is a yes-instance of Star 3-Colouring. We do this by giving each vertex in G a label from \{1,2,3\}. The resulting labelling c might not be a 3-colouring, but this is not important: we will show that we can extend c to a star 3-colouring of G_p as follows.

Consider an edge e in G and let P be the corresponding path (of p+1 edges) in G_p. It suffices to give two star 3-colourings of this path, so that the first three vertices are distinct colours and the last three vertices are distinct colours: one in which the first and last vertices are the same colour and one in which they are a different colour. Let us identify a 3-colouring of P by a sequence of length p+1 over \{1,2,3\}. If p+1 is a multiple of three, then use (123)^{\frac{p+1}{3}} for the different colour and (123)^{\frac{p+1}{3}-1}231 for the same colour. If p+1 is 1 \bmod 3, then use (123)^{\frac{p}{3}-1}2132 for the different colour and (123)^{\frac{p}{3}}1 for the same colour. If p+1 is 2 \bmod 3, then use (123)^{\frac{p-1}{3}}12 for the different colour and (123)^{\frac{p-1}{3}}21 for the same colour.

It suffices to prove the statement for k=10. We do this by induction. For the base case, a graph with one vertex is star 10-colourable. Now take a vertex v in a graph G and assume G \setminus \{v\} has an injective 10-colouring. As G is subcubic, v has at most three neighbours, each of which have at most two more neighbours each. Thus there are at most nine vertices whose colour we wish to avoid. As we have ten colours in total, this means that we can safely colour v.

Let \ell\geq 1. As every (\ensuremath{\mathbb{H}}_{\ell},\ensuremath{\mathbb{H}}_{\ell+1},\ldots )-subgraph-free graph is H_\ell-minor-free, we apply Theorem 28.

We proceed by contraposition. If a class of graphs has unbounded treewidth, then every grid appears as a minor in some graph [RS86]. Let us explain the argument for n=2 first. We consider that the 3 \times 3-grid appears as a minor in some graph G in our class and let f be the minor map from G to the 3 \times 3-grid. Consider the three vertices in the 3 \times 3-grid that form the central row as u,v,w (in succession). Choose u' \in f^{-1}(u), v' \in f^{-1}(v), w' \in f^{-1}(w) so that u',v',w' have degree greater than 2, noting that such vertices must exist. If the distance in G between u' and v' is even, of length 2i, then there is an \ensuremath{\mathbb{H}}_{2i} subgraph in G with central path from u' to v'. If the distance in G between v' and w' is even, of length 2i, then there is an \ensuremath{\mathbb{H}}_{2i} subgraph in G with central path from v' to w'. Else, there is a path of even length 4i from u' to w' and then there is an \ensuremath{\mathbb{H}}_{4i} subgraph in G with central path from u' to w'.

For (\ensuremath{\mathbb{H}}_n,\ensuremath{\mathbb{H}}_{2n},\ldots)-subgraph-free graphs, consider the Abelian group (\mathbb{Z}/n\mathbb{Z}). The Davenport constant of an Abelian group G is the minimum d so that any sequence of elements of G contains a non-empty consecutive subsequence of zero-sum (that adds to the identity element 0). It is known that for (\mathbb{Z}/n\mathbb{Z}) the Davenport constant is n (see page 24 in [GR09]). Take an (n+1) \times (n+1)-grid and consider some row not at the top or bottom of the grid with vertices w_1,\ldots,w_{n+1} in succession. Consider some w'_1 \in f^{-1}(w_1), \ldots, w'_{n+1} \in f^{-1}(w_{n+1}) where f is the minor map as before, and the distances x_i between w'_{i+1} and w'_i. Using the Davenport constant, there is a subsequence x_j,\ldots,x_{j'} (j'>j) such that x_j+\ldots+x_{j'}=0 \bmod n. Now choose w'_j,\ldots,w'_{j'+1} as the central path in some \ensuremath{\mathbb{H}}_{in}.

Reduce from 3-Colourability which is known to be NP-complete even for planar graphs [Da80]. Let G be a planar graph. Replace each edge e by three new vertices a_e, b_e, c_e that are made adjacent only to the two end-vertices of e in G. Let G' be the resulting graph. Then every vertex of V(G')\setminus V(G) has degree 2 in G. Moreover, G' is planar and bipartite with partition classes V(G')\setminus V(G) and V(G). It remains to observe that G has a 3-colouring if and only if G' has a star 3-colouring.

It is well known [HN90] and easy to see that there is a reduction from K_5-Colouring, which is to decide if a graph has a K_5-colouring, that is, a homomorphism from G to the complete graph K_5 on five vertices. This problem is well known to be NP-complete [HN90]. Let G be a graph, and let G' be the 2-subdivision of G. We note that G' is (\ensuremath{\mathbb{H}}_1,\ensuremath{\mathbb{H}}_2,\ensuremath{\mathbb{H}}_4,\ensuremath{\mathbb{H}}_5,\ldots)-subgraph-free (but may contain many instances of \ensuremath{\mathbb{H}}_\ell where \ell=0\mod 3). Moreover, G has a K_5-colouring if and only if G' has a C_5-colouring.

The problem is ready seen to be in NP. We will now prove NP-hardness.

Let \ell\geq 1 be an integer. Let \phi be an instance of 3-Satisfiability, consisting of m clauses C_1,\ldots,C_m on n variables z_1,\ldots,z_n. For each clause C_j (j=1,\ldots,m), with C_j = y_{3j-2} \vee y_{3j-1} \vee y_{3j}, then y_i (i = 1,\ldots,3m) is a literal from \{z_1,\ldots,z_n, \overline{z}_1,\ldots,\overline{z}_n\}.

Let us build a graph G^\ell_\phi with two specified vertices x and y of degree 2. There will be a hole containing x and y in G_\phi if and only if there exists a truth assignment satisfying \phi.

For each literal y_j (j = 1,\ldots,3m), prepare a graph G(y_j) on 20 named vertices

together with certain paths in between using unnamed vertices, as drawn in Figure ?. (We drop the subscript j in the labels of the vertices for clarity.)

For i = 1, 2, 3 add paths of length \ell between \alpha^{i+} and \alpha^{(i+1)+}; \alpha^{i-} and \alpha^{(i+1)-}; \beta^{i+} and \beta^{(i+1)+}; and \beta^{i-} and \beta^{(i+1)-}. Also add the edges \alpha^{1+}\beta^{1-}, \alpha^{1-}\beta^{1+}, \alpha^{4+}\beta^{4-}, \alpha^{4-}\beta^{4+}, \alpha \alpha^{1+}, \alpha \alpha^{1-}, \alpha^{4+} \alpha', \alpha^{4-} \alpha', \beta \beta^{1+}, \beta \beta^{1-}, \beta^{4+} \beta', \beta^{4-} \beta'.

For each clause C_j (j = 1,\ldots,m), prepare a graph G(C_j) with 10 named vertices

together with certain paths in between using unnamed vertices, as drawn in Figure ?. (We drop the subscript j in the labels of the vertices for clarity.) Add paths of length \ell between the following pairs of vertices: c^{12+} and c^{1+}; c^{12+} and c^{2+}; c^{12-} and c^{1-}; c^{12-} and c^{2-}; c^{0+} and c^{12+}; c^{0+} and c^{3+}; c^{0-} and c^{12-}; c^{0-} and c^{3-}.

For each variable z_i .(i = 1,\ldots,n), prepare a graph G(z_i) with 2z_i^- + 2z^+_ i vertices, where z^-_i is the number of times \overline{z}_i appears in clauses C_1, \ldots,C_m and z^+_i is the number of times z_i appears in clauses C_1, \ldots,C_m.

Let G(z_i) consist of two internally disjoint paths P^+_i and P^-_i with common endpoints d^+_i and d^-_i and lengths 1+ (2\ell)z^-_i and 1+ (2\ell)z^+_i, respectively. Label the vertices of P^+_i and P^-_i as in Figure ?.

The final graph G^\ell_\phi will be constructed from the disjoint union of all the graphs G(y_j), G(C_i), and G(z_i) with the following modifications:

• For j = 1,\ldots, 3m-1, add paths of length \ell between the pairs: \alpha'_j and \alpha_{j+1}; \beta'_j and \beta_{j+1}.

• For j = 1,\ldots, m-1, add a path of length \ell between c^{0-}_j and c^{0+}_{j+1}.

• For j = 1,\ldots, n-1, add a path of length \ell between d^{-}_i and d^{+}_{i+1}.

• For i = 1,\ldots,n-1, let y_{n_1},\ldots,y_{n_{z^-_i}} be the occurrences of \overline{z}_i over all literals. For j = 1,\ldots,z^-_i, delete the path between p^+_{i,2j-1} and p^+_{i,2j} and add the four edges p^+_{i,2j-1}\alpha^{2+}_{n_j}, p^+_{i,2j-1}\beta^{2+}_{n_j}, p^+_{i,2j}\alpha^{3+}_{n_j}, p^+_{i,2j}\beta^{3+}_{n_j}.

• For i = 1,\ldots,n-1, let y_{n_1},\ldots,y_{n_{z^+_i}} be the occurrences of z_i over all literals. For j = 1,\ldots,z^+_i, delete the path between p^-_{i,2j-1} and p^-_{i,2j} and add the four edges p^-_{i,2j-1}\alpha^{2+}_{n_j}, p^-_{i,2j-1}\beta^{2+}_{n_j}, p^-_{i,2j}\alpha^{3+}_{n_j}, p^-_{i,2j}\beta^{3+}_{n_j}.

• For i = 1,\ldots,m and j=1,2,3, add the edges \alpha^{2-}_{3(i-1)+j}c^{j+}_i, \alpha^{3-}_{3(i-1)+j}c^{j-}_i, \beta^{2-}_{3(i-1)+j}c^{j+}_i, \beta^{3-}_{3(i-1)+j}c^{j-}_i.

• Add a path of length \ell between the pairs of vertices: \alpha'_{3m}d^+_1 and d^+_1; \beta'_{3m}d^+_1 and c^{0+}_1.

• Add the vertex x and add paths of length \ell between the pairs of vertices: x and \alpha_1; x and \beta_1.

• Add the vertex y and add paths of length \ell between the pairs of vertices: y and c^{0-}_m; y and d^{-}_n.

It is easy to verify that the maximum degree of G^\ell_\phi is 3. Moreover, the size of G^\ell_\phi is polynomial (actually linear) in the size n+m of \phi, and x and y are non-adjacent, and both have degree 2.

We also claim that for \ell\geq 4, G^\ell_\phi is \ensuremath{\mathbb{H}}_4,\ldots,\ensuremath{\mathbb{H}}_{\ell}-subgraph-free, which can be seen as follows (note that G^\ell_\phi contains \ensuremath{\mathbb{H}}_1, \ensuremath{\mathbb{H}}_2 and \ensuremath{\mathbb{H}}_3 as subgraphs). Owing to the length of the \ell-paths that populate our construction, we need only verify the omission of these graphs on the connected components of the graph G^\ell_\phi after the removal of these paths (except a pendant edge from the corresponding connected component at the extremities of an instance of these paths). In this fashion, we only need to check for omission of the given graphs in the non-trivial cases drawn in Figure ?. Indeed, the two cases are isomorphic. Let i=4,5. Any two vertices of degree at least three that are separated by a path of length i must be in the subgraph C_6 at distance 6-i from one another. If i=4 then these vertices have a common neighbour so the \ensuremath{\mathbb{H}}_i can’t be completed. If i=5 then these two vertices are adjacent. For 6\leq i \leq \ell-i it is not possible to find two vertices of degree at least three that are separated by a path of length \ell.

We now show that \phi is satisfied by a truth assignment if and only if G^\ell_\phi contains a hole passing through x and y.

``\Rightarrow'' First assume that \phi is satisfied by a truth assignment \xi \in \{0,1\}^n. We will pick a set of vertices that induce a hole containing x and y.

1. Pick vertices x and y.

2. For i = 1,\ldots,3m, pick the vertices \alpha_i,\alpha'_i,\beta_i,\beta'_i.

3. For i = 1,\ldots,3m, if y_i is satisfied by \xi , then pick the vertices \alpha^{1+}_i,\alpha^{2+}_i,\alpha^{3+}_i,\alpha^{4+}_i,\beta^{1+}_i,\beta^{2+}_i, \beta^{3+}_i,\beta^{4+}_i. Otherwise, pick the vertices \alpha^{1-}_i,\alpha^{2-}_i,\alpha^{3-}_i,\alpha^{4-}_i,\beta^{1-}_i,\beta^{2-}_i,\beta^{3-}_i,\beta^{4-}_i.

4. For i = 1,\ldots,n, if \xi(i) = 1, then pick all the vertices of the path P^+_i and all the neighbours of the vertices in P^+_i of the form \alpha^{2+}_k or \alpha^{3+}_k for any k.

5. For i = 1,\ldots,n, if \xi(i) = 0, then pick all the vertices of the path P^-_i and all the neighbours of the vertices in P^-_i of the form \alpha^{2+}_k or \alpha^{3+}_k for any k.

6. For i = 1,\dots,m, pick the vertices c^{0+}_i and c^{0-}_i. Choose any j \in \{3i-2, 3i-1, 3i\} such that \xi satisfies y_j. Pick vertices \alpha^{2-}_j and \alpha^{3-}_j. If j = 3i - 2, then pick the vertices c^{12+}_j,c^{1+}_j,c^{12-}_j,c^{1-}_j. If j = 3i - 1, then pick the vertices c^{12+}_j,c^{2+}_j,c^{12-}_j,c^{2-}_j. If j = 3i, then pick the vertices c^{3+}_j,c^{3-}_j.

The given vertices do not yet induce a connected component, because we need to add the vertices of \ell-paths in between. Thus, if p and q are vertices which we selected that have an \ell-path between them (drawn as a dashed edge in the associated gadget), then we need to add the interior vertices of this path also.

It suffices to show that the chosen vertices induce a hole containing x and y. The only potential problem is that for some k, one of the vertices \alpha^{2+}_k,\alpha^{3+}_k,\alpha^{2-}_k,\alpha^{3-}_k was chosen more than once. If \alpha^{2+}_k and \alpha^{3+}_k were picked in Step 3, then y_k is satisfied by \xi. Therefore, \alpha^{2+}_k and \alpha^{3+}_k were not chosen in Step 4 or Step 5. Similarly, if \alpha^{2-}_k and \alpha^{3-}_k were picked in Step 6, then y_k is satisfied by \xi and \alpha^{2-}_k and \alpha^{3-}_k were not picked in Step 3. Thus, the chosen vertices induce a hole in G^\ell_\phi containing vertices x and y.

``\Leftarrow'' To show the reverse implication, assume G^\ell_\phi contains a hole H passing through x and y. The hole H must contain \alpha_1 and \beta_1, and the paths leading to them, since they are the only two path neighbours of x. Next, either both \alpha^{1+}_1 and \beta^{1+}_1 are in H or both \alpha^{1-}_1 and \beta^{1-}_1 are in H.

Without loss of generality, let \alpha^{1+}_1 and \beta^{1+} be in H (the same reasoning that follows will hold true for the other case). Since \alpha^{1-}_1 and \beta^{1-} are both neighbours of two members of H, they cannot be in H. Thus, \alpha^{2+}_1 and \beta^{2+}_1, and the paths to them, must be in H. Since \alpha^{2+}_1 and \beta^{2+}_1 have the same neighbours outside G(y_1), it follows that H must contain \alpha^{3+}_1 and \beta^{3+}_1, and the paths that lead to them. Also, H must contain \alpha^{4+}_1 and \beta^{4+}_1, and the paths that lead to them. Suppose that \alpha^{4-}_1 and \beta^{4-}_1 are in H. Because \alpha^{i-}_1 has the same neighbour as \beta^{i-}_1 outside G(y_1) for i = 2, 3, it follows that H must contain \alpha^{3-}_1, \alpha^{2-}_1, \alpha^{1-}_1. But then H is not a hole containing x, a contradiction. Therefore, \alpha^{4-}_1 and \beta^{4-}_1 cannot both be in H, so H must contain \alpha'_1, \beta'_1, \alpha_2, \beta_2, and the paths to them.

By induction, we see for i = 1, 2, \ldots,3m that H must contain \alpha_i,\alpha'_i,\beta_i,\beta'_i. Also, for each i, either H contains \alpha^{1+}_i,\alpha^{2+}_i,\alpha^{3+}_i,\alpha^{4+}_i,\beta^{1+}_i,\beta^{2+}_i,\beta^{3+}_i,\beta^{4+}_i or H contains \alpha^{1-}_i,\alpha^{2-}_i,\alpha^{3-}_i,\alpha^{4-}_i, \beta^{1-}_i,\beta^{2-}_i,\beta^{3-}_i,\beta^{4-}_i.

As a result, H^\ell_\phi must also contain d^+_1 and c^{0+}_1 and the paths to them. By symmetry, we may assume H^\ell_\phi contains p^+_{1,1} and \alpha^{2+}_k, for some k. Since \alpha^{1+}_k is adjacent to two vertices in H, H must contain \alpha^{3+}_k and the path of length \ell toward it. Similarly, H cannot contain \alpha^{4+}_k, so H contains p^+_{1,2} and p^+_{1,3}, as well as the paths through these. By induction, we see that H contains p^+_{1,i} for i = 1, 2, \ldots, z^+_i and d^-_1 and the \ell-paths in between. If H contains p^-_{1,z^-_i}, then H must contain p^-_{1,i} for i = z^-_i, \ldots, 1, a contradiction. Thus, H must contain d^+_2 and the \ell-path to it. By induction, for i = 1, 2, \ldots,n, we see that H contains all the vertices of the path P^+_i or P^-_i and by symmetry, we may assume H contains all the neighbours of the vertices in P^+_i or P^-_i of the form \alpha^{2+}_k or \alpha^{3+}_k, for any k.

Similarly, for i = 1, 2, \ldots,m, it follows that H must contain c^{0+}_i and c^{0-}_i. Also, H contains one of the following:

• c^{12+}_i,c^{1+}_i,c^{12-}_i,c^{1-}_i and either \alpha^{2-}_j and \alpha^{3-}_j or \beta^{2-}_j and \beta^{3-}_j (where \alpha^{2-}_j is adjacent to c^{1+}_i).

• c^{12+}_i,c^{2+}_i,c^{12-}_i,c^{2-}_i and either \alpha^{2-}_j and \alpha^{3-}_j or \beta^{2-}_j and \beta^{3-}_j (where \alpha^{2-}_j is adjacent to c^{2+}_i).

• c^{3+}_i,c^{3-}_i and either \alpha^{2-}_j and \alpha^{3-}_j or \beta^{2-}_j and \beta^{3-}_j (where \alpha^{2-}_j is adjacent to c^{3+}_i).

We can recover the satisfying assignment \xi as follows. For i = 1, 2, \ldots,n, set \xi(i) = 1 if the vertices of P^+_i are in H and set \xi(i) = 0 if the vertices of P^-_i are in H. By construction, it is easy to verify that at least one literal in every clause is satisfied, so \xi is indeed a satisfying assignment.

Let x and y have neighbours x_1, x_2 and y_1, y_2 respectively. We replace x and its incident edges by the following. Create vertices p_1, q_1, r_1, p_2, q_2, r_2, s_1, s_2. Add edges s_1p_1, p_1q_1,q_1x_1, q_1r_1, r_1s_2 and s_2r_2, r_2q_2, q_2x_2, p_2s_1, p_2q_2. In the same way, we replace y and its incident edges by vertices a_1,b_1,c_1,a_2,b_2,c_2,t_1,t_2 and edges a_1t_1, a_1b_1, b_1y_1, b_1c_1, c_1t_2 and a_2t_1, a_2b_2, b_2y_2, b_2c_2, c_2t_2. Note that any vertex that is introduced has degree at most 3. For an integer d\geq 0, we now subdivide every incident edge of every newly introduced vertex d times. We denote the new graph by G^d and say that G^d is the d-replacement for G. Note that G^d has maximum degree 3 as well. See also Fig. ?, where we display the replacement of x for d=1. It is readily seen that for every integer d\geq 0, G^d has maximum degree  3.

To prove the remainder of the lemma, first assume that d=0. We observe that the paths s_1,p_1,q_1,x_1 and s_2,r_2,q_2,x_2 are mutually induced. The paths s_1,p_2,q_2,x_2 and s_2,r_1,q_1,x_1 are also mutually induced. Moreover, these are the only two options that can co-exist, in the following sense. A path from s_1 to x_1 that uses only edges of this gadget and that does not have s_2 as an internal vertex has to pass through q_1. Similarly, a path from s_2 to x_2 that uses only edges of this gadget and that does not have s_1 as an internal vertex has to pass through q_2. The same observations hold with respect to the gadget replacing vertex y. Hence, G has a hole containing x and y if and only if G^d has mutually induced paths between s_1 and t_1 and between s_2 and t_2.

Note that any incident edges of every vertex of G^d that does not belong to G can be subdivided an arbitrary number of times without affecting the correctness of the reduction (q_1,q_2,b_1,b_2 remain bottlenecks). So, the claim also holds for d\geq 1, and by taking d sufficiently large, we obtain the desired subcubic (\ensuremath{\mathbb{H}}_4,\ldots,\ensuremath{\mathbb{H}}_\ell)-subgraph-free graph G'.

That the first two items hold are clear. To see that C_5-critical G cannot have a removable thread, remove the thread and C_5-colour the remaining graph G'. As the thread has length at least k, we can extend the C_5-colouring of G' to G; contradicting the fact that G was C_5-critical. To see that G cannot have a redundant thread, remove it, and get a C_5-colouring of the remaining graph G'. We can extend this to a C_5-colouring of G be letting the colouring on the redundant thread be the same as on the path of G' of the same length between the endpoints of the thread. That redundant vertices cannot occur in a critical graph is clear– they can always get the same colour as the vertex they are a subclone of, so removing them does not change the possible colourings. That any vertex in a C_4 is a 3^+ vertex is because a 2-vertex in a C_4 is redundant. That

Let G be \ensuremath{\mathbb{H}}_3-subgraph-free and C_5-critical. If it is a K_3 the lemma is vacuously true, so we may assume by criticality that it is K_3-free.

For item (1), assume that u is a 3^+ vertex in a C_4 with its 4^+-neighbour x, in G. As G is K_3-free, we may assume that the C_4 is u,v,w,x, and is induced. As x is a 4^+ vertex it has neighbours y and y' not in the C_4. If u is a 4^+-vertex then it has neighbours t and t' not in the C_4, and as G is K_3-free, they are distinct from y and y'. But then \{t,t'\},u,v,w,x,\{y,y'\} is an \ensuremath{\mathbb{H}}_3, which is impossible.

For item (2), assume that C is the 5-cycle a,b,c,d,e,a, that a is a 3^+ vertex, and that d is a 4^+ vertex. As G is K_3-free, C is induced, and so a has neighbours x,x' not in \{b,c,d\}. If d has degree 5, then it has two neighbours y,y' not in \{a,b,c,x,x'\} and so \{x,x'\},a,b,c,d,\{y,y'\} is an \ensuremath{\mathbb{H}}_3.

If C_5-critical G contained a proper subgraph G' that was saturated, then that subgraph is disconnected from G by the removal of its one unsaturated vertex, but this contradicts the fact that G, being C_5-critical, is 2-connected by Fact 1.

Towards contradiction, let G be a \ensuremath{\mathbb{H}}_3-subgraph-free graph of girth at least 6 that is C_5-critical. As G has no C_5-colouring it must contain an odd cycle, and so by assumption, it has odd girth 2k+1 \geq 7. Let C be a C_{2k+1} in G with vertex set \{-k,-(k-1), \dots, 0, \dots, k\} for which two vertices are adjacent if they differ by 1 modulo 2k+1. Let X be the set of 3^+-vertices of C.

As C is a shortest odd cycle in G it is induced, and there is no vertex x \not\in C adjacent to two vertices of C at distance greater than 2. So because G is \ensuremath{\mathbb{H}}_3-subgraph-free, X contains no two vertices that are distance 3 apart. On the other hand, as G is C_5-critical and so contains no 3-threads X contains at least one of i-1, i, and i+1 for every vertex i of C. It follows that if X contains i and i+1, then it does not contain i + 3 and i+4, and so it must contain i+2 and i+5. But these are distance 3 apart; so we conclude that X contains no two consecutive vertices.

We may assume, w.l.o.g, that X contains the vertex 0. So it contains neither -1 or 1, or -3 or 3. It must therefore contain -2 and 2. Shifting this argument cyclically by 2, we see that X must contain every second vertex of C. But this is impossible as C is an odd cycle.

Let G be an \ensuremath{\mathbb{H}}_3-subgraph-free C_5-critical graph of girth 5. We show that G is either E_1, E_2 or an odd C_5-flower F_n.

We start with two claims that deal with the more difficult cases. Let B_1 and B_2 be the graphs made from two copies of C_5 shown in Figure ?.

Recall that B_2, shown in Figure ?, consists two C_5s whose intersection is a path of length two.

Now, we have assumed that G has girth 5. Let C be a 5-cycle in G with vertices 0,1,2,3,4. By part (4) of Fact 3, there are at least three 3^+-vertices in C, so we may assume that 1 and 4 are 3^+-vertices. By girth, they have distinct neighbours, a and z respectively, that are not in C. Further, a and z are non-adjacent as an edge between them would make a B_2, which contradicts Claim ?.

By girth, neither of a or z have a second edge to C, and they both have 3-paths to 3-vertices in C, so they are both 2-vertices. They have new neighbours b and y respectively, not necessarily distinct.

As C has at least one more 3-vertex, we may assume that b has a 3-path to this vertex i. If b is a 3-vertex, then it must have an edge to a neighbour of i, other than 1 and 4. Whatever this edge is, would make a B_1 or a B_2, so we would be done by one of the claims. Thus b is a 2-vertex.

If b=y we have a removable thread, so let c be the new neighbour of b. As c has a 3-path to 1, if it is a 3-vertex then it has an edge to 0 or 2, making a B_1. So we may assume that c is a 2 vertex, and so we have a removable thread. Thus G is not critical, which is a contradiction.

Item (1) is straightforward from the fact that G has no induced \ensuremath{\mathbb{H}}_3 or C_4, and (2) comes from Fact 2 and item (1).

We show item (3). Any 5-cycle C in G is induced and has no 3^+-threads, so (by an easy version of the argument in the proof of Proposition 37) must contain two 3^+-vertices, call them 1 and 4, at distance 2 in C. By (1), these are, in fact, 3-vertices. Let a and b, be their respective neighbours not in C.

If there is no other 3^+-vertex in C not, then remove C, and C_5-colour G' = G \setminus C. The only edges from C to G' are the edges 1a and 4b, and whatever colours a and b get, we can extend this colouring to a C_5-colouring of G, contradicting the fact that G was C_5-critical. Thus C has at least three 3^+-vertices, as needed.

To see that S_1 is saturated, note that all 3-vertices have 3-paths to other 3-vertices, so are saturated. For all 2-vertices but the one in the middle of the right cycle, there is a 3-path to a 3-vertex not in a C_5 with it. Adding an edge to such a 2-vertex, the edge would then have to go to a neighbour of the 3-vertex, but checking each case, this would make a C_3 or C_4. The middle 2-vertex is not saturated, but one non-saturated vertex is okay.

In an even C_5-flower, the 3-vertices all have 3-paths to other 3-vertices, so are saturated. The 2-vertices all have 3-paths to the middle 3^+-vertex, so if they have another edge, it must be to one of its neighbours. If this neighbour is a new vertex, we have a C_4, and if it is a neighbour in F_n, then it is a 3-vertex, which is already saturated. So the 2-vertices are saturated. Only the middle vertex is not. But again, one unsaturated vertex is allowed.

Let G be a \ensuremath{\mathbb{H}}_{3}-subgraph-free C_5-critical graph of girth 4. We start with a pair of claims showing that G can contain no biclique other than a C_4.

With Claim ? proved, we are ready to finish our proof of the proposition. As G has girth 4 we may assume it contains a 4-cycle C on the vertices 0,1,2,3. Every vertex i of C has degree at least 3, so has at least one more neighbour i'. Let N_i be the set of neighbours of i in G \setminus C for vertex i of C. No two vertices of C share a common neighbour not in C or we would have a C_3 or a K_{2,3}, so the sets N_i are disjoint; and so there are no edges from N_i to j \in V(C) for i \neq j.

Notice the vertices of C are saturated in C \cup \bigcup N_i, as by definition of the N_i they can have no other neighbours in G.

We may therefore assume that 1' \in N_1 has degree at least 3. It must have an edge to some 3' in N_3 or we have an \ensuremath{\mathbb{H}}_3. For i = 1,3 let N'_i = N(i') \setminus (V(C) \cup N). So our graph looks like this.

Now, the following hold, reducing the figure to what you get by ignoring the dashed bits.

1. N_1 and N_3 are disjoint, as otherwise we would have a K_{2,3}. If there is some 1^* \in N_1\setminus \{1'\} and some 3^* \in N_3\setminus \{3'\}, then we have a \ensuremath{\mathbb{H}}_3 on \{2,1^* \},1',3',3, \{0,3^* \}. So we may assume that N_3 = \{3'\}.

2. If N_2 has two vertices, then we can find an \ensuremath{\mathbb{H}}_3 unless N_1 = \{1'\} and N'_1 and N'_3 are empty. But if all these conditions hold, then taking B = \{0,1,2,3,1',3'\} and B' = \{0',2'\} in Lemma 44 we get that 0',2' is an edge in G. But this makes a E_3. So we may assume that N_2 = \{2'\} and similarly that N_0 = \{0'\}.

3. By Lemma 43, if 2' were a 3^+ vertex, it would have an edge to 0 or 0', making a K_{2,3} or a copy of E_3 respectively. Thus 2', and similarly 0', is a 2-vertex.

4. So 0' and 2' are not in any C_4, so have no edges to \{1',3'\} \cup N_1^*, and so in particular, are not in N'_1 \cup N'_3, (meaning N'_1 \cup N'_3 has no edges to \{0,2\}).

5. No vertex in N_1 has degree greater than 3. Indeed, the only possible neighbours of a vertex v of N_1 are in N'_1 \cup N'_3 \cup \{3'\} or are vertices not shown in the figure. So using the two of them that are not 1 or 3' as leaves, we have an \ensuremath{\mathbb{H}}_3 with the vertices v, 1,2,3,\{0,3'\}. The same holds for 3'– it has degree at most 3.

6. A vertex 1'' in N'_1 can only have as neighbours 0',1' or 2', or vertices not shown in the figure. If 1'' has degree at least 3, then using the two neighbours that are not 1' as leaves, there is an \ensuremath{\mathbb{H}}_3 on 1',3',3,\{0,2\}. Thus vertices in N'_1 and N'_3 are 2-vertices.

7. If there are distinct 1^* \in N_1 \setminus \{1'\} and 3'' \in N'_3, then we have a \ensuremath{\mathbb{H}}_3 on the vertices \{1',3''\},3',3,2,1,\{0,1^*\}. So either N_3' = N_1 \setminus \{1'\} = \{1^{*}\}, or at least one of N_3' and N_1 \setminus \{1'\} are empty. Assume the former. This vertex 1^* has the same two neighbours as 1' does, so they must both have distinct third neighbours 1^{**} and 1'' respectively. These are 2-vertices by (6), and the vertices 1^* and 1' are 3-vertices by (5). Applying Lemma 44 with B = \{0,1,2,3,1',1^*,3'\} and B' = \{0',2'\} gives that 2',0' is an edge of G, but this is not true. Thus either N'_3 is empty, or N_1 = \{1'\}.

In either case we apply Lemma 44 with B = V(C) \cup \{3'\} \cup N_1 and B' = N'_1 \cup \{0'\} \cup N'_3 \cup \{2'\}. (By definition, all neighbours of elements of B are in B \cup B', so elements of B are saturated; elements of B' have been argued to be 2-vertices.) Edges from 0' to 2' give a proper subgraph E_3 which is impossible. Edges from N'_1 to N'_3, or from \{0',2'\} to N'_1 \cup N'_3 leave the graph C_5-colourable, so adding them until there is at most one unsaturated vertex left gives a saturated C_5-colourable subgraph, contradicting the criticality of G.

Fix a 4-cycle C in G on the vertices 0,1,2,3. There can be no 3^+ vertex at distance exactly 3 from C as it would make an H_3 with the vertex in C closest to it.

Now, there are no 3^+-vertices at distance 2 or 3 from C. Assuming there are 3^+-vertices at distance greater than 3 from C, and let v be a closest such vertex to C. If v has distance 5 or more from C, then since the closest 3^+-vertex on its shortest path to C is at distance 1 from C, it contains a removable 3^+-thread. So v must have distance 4 from C. The vertex 1' at distance 1 from C on a shortest path to C must then be a 2-vertex, or we have a H_3– v and 1' would be 3^+ vertices at distance 3 and from the fact that v has distance 4 to C they cannot be adjacent or share a neighbour. So there is a removable 3-thread between v and C, a contradiction.

For the ’Moreover’ statement, assume that a 3^+-vertex v with an edge to C, has an edge to 0. As there is a 3-path to 2, then v and 2 are either adjacent or share a neighbour, and as G is C_3-free the shared neighbour would have to be a new vertex.

Assume the setup of the lemma. As G is C_5-critical it is 2-connected; as the subgraph B is 2-connected, Lemma 35 says that we can construct G from B by adding successive ears. Let P be a shortest ear of B in G; say it is between vertices u and v of B. As vertices in B are saturated, P must start and end with edges u,u' and v',v to 2-vertices u' and v' in B'. As vertices of B' have degree 1 in G', u' and v' are distinct. If there are any 3^+-vertices on P, then they must be adjacent to the C_4 in B; this contradicts the choice of P as the shortest ear. If there are any other 2-vertices in P, then we have a 3-thread, contradicting the criticality of G. So P is a 3-path through B'. Thus one gets P by adding an edge between vertices of B', as needed.

When we apply this lemma to a graph to extend it, it adds an edge between vertices of B'. If the graph remains C_5-colourable, then the resulting graph still satisfies the conditions of the lemma (with two vertices of B' moved into B), and we can apply it again. If the graph becomes non-C_5-colourable, then it is G.

Suppose we have a C_4 in which three vertices have degree greater than 2. Two of these must be adjacent in the C_4, let us call them u and v. If they both have private neighbours outside of the vertices of the C_4, then G contains \mathbb{A} as a subgraph, and is a no-instance. If they have a common neighbour outside of the vertices of the C_4, then G contains a graph built from C_4 and K_3 by identifying an edge of the former with an edge of the latter as a subgraph (which contains a C_5), and is a no-instance. If they both have their third neighbour in the vertices of the C_4, then the vertices of the C_4 induce a K_4 in G, and this is a no-instance. Finally, is the possibility that one has a new neighbour outside of the C_4, while the other has a new neighbour inside of the C_4. Then G contains as a subgraph a diamond with a pendant vertex attached to one of the vertices of degree 2, and is a no-instance.

By Lemma 48 we find that {\mathbb A} is not star 3-colourable.

We now prove that \overline{2 P_4 + 3 P_6} is not star 3-colourable. Assume otherwise. Let x_1 be the neighbour of u on one copy of P_4 and x_2 the neighbour of v on this path. Similarly, let y_1 be the neighbour of u on the second copy of P_4 and y_2 the neighbour of v. Label the internal vertices of the first copy of P_6 z_1 \dots z_4 where z_1 is the neighbour of u and z_4 is the neighbour of v. Let w_1 \dots w_4 be the internal vertices of the second copy of P_6 such that w_1 is adjacent to u and w_4 is adjacent to v.

First consider a proper star 3-colouring c in which u and v are coloured differently, say u is coloured 1 and v is coloured 2. It must be that either c(x_1)=2 or c(x_2)=1. Without loss of generality, c(x_1)=2. Then c(x_2)=3. Every remaining neighbour a of v then has colour 1, else we have a bichromatic P_4 with vertex set \{a,v,x_2, x_1 \}. To avoid a bichromatic P_4, c(y_1)=3. Each further neighbour b of u then has colour 2, else we have a bichromatic P_4 with vertex set \{b,u, y_1,y_2 \}. We now consider the first copy of P_6. Since c(z_1)=2, c(z_2)=3, else there is a bichromatic P_4 with vertex set \{x_1,u,z_1,z_2\}. However, c(z_3)=3, else we have a bichromatic P_4 with vertex set \{z_3, z_4, v, y_2\}. This leaves two adjacent vertices coloured 3, a contradiction.

Therefore, u and v must be coloured alike in any proper star 3-colouring c. Without loss of generality, they are both assigned colour 1. Without loss of generality, c(x_1)=2 and c(x_2)=3. Now consider the first copy of P_6. First assume that c(z_1)=2. Then, to avoid a bichromatic P_4 with vertex set \{z_2, z_1, u, x_1\}, c(z_2)=3. Now, to avoid a bichromatic P_4 along this path, at least one of z_3 and z_4 must be coloured 2. If c(z_3)=2 then c(z_4)=3 and we have a bichromatic P_4. Therefore, c(z_4)=2 and c(z_3)=1. Each remaining neighbour of v then has colour 3 to avoid a bichromatic P_4. Since c(w_4)=3, c(w_3)=2, else we have a bichromatic P_4. If c(w_2)=3, then c(w_1)=2 and we have a bichromatic P_4 on \{w_1,w_2,w_3,w_4\}. Thus, c(w_2)=1, whereupon c(w_1)=3 to avoid a bichromatic P_4 on \{x_1,u,w_1,w_2\}. The same arguments apply to the third copy of P_6 but now we have a bichromatic P_4 on 1313 with vertex set \{u, w_1, w_2 \} plus the first vertex on the third copy of P_6, a contradiction.

It remains to consider the case where c(z_1)=3. If c(z_2)=1 then c(z_3)=2 and c(z_4)=3. Additionally, every remaining neighbour c of u is coloured 2 to avoid a bichromatic P_4 with vertex set \{z_2, z_1, u, c \}. It must then be that c(w_1)=2, c(w_2)=3. In fact, this case is covered in the previous paragraph by an automorphism of \overline{2 P_4 + 3 P_6} that swaps u and v, as well as colours 2 and 3. Therefore c(z_2)=2. To avoid a bichromatic P_4 along this copy of P_6, c(z_3)=1. Now c(z_4)=2 since otherwise we have a bichromatic P_4 with vertex set \{z_3, z_4, v, x_2\}. However, this induces a bichromatic P_4 with vertex set \{z_2,z_3,z_4,v\}, a contradiction.

We may assume without loss of generality that G is connected. Suppose G is \mathbb{A}-subgraph-free. Note that G cannot contain a K_3 since it is bipartite. By assumption, G has more than two 3^+-vertices. Let us choose three in a row u,v,w on a path in that order, so as to minimise the total distance from u to w. Note it is possible that there is a path from u to w that is shorter than the path from u to w via v.We are not aware of an example like this that is both omits all \ensuremath{\mathbb{H}}_{2i} as a subgraph and has such u,v,w minimally chosen.

Suppose u,v,w are consecutive. Then u and w must have another common neighbour to avoid an \ensuremath{\mathbb{H}}_2. Now we violate Lemma 48. Thus, u,v,w may not be consecutive.

Suppose now that u and w have two common neighbours p and q. By the minimality of the length of the path from u to v to w we may now assume that the subpaths from u to v and v to w have length at most 2, and so as u,v,w are not consecutive one of them has length 2. Without loss of generality, let us assume this is u to v . According to Lemma 48, both p and q have degree 2 and neither of these may have an edge to v, therefore v has a distinct neighbour z. Since there is a path of length 2 from u to v we have \ensuremath{\mathbb{H}}_2 as a subgraph and finish this case.

Suppose now that u and w have one common neighbour p and two distinct neighbours u' and w', respectively. Since u,v,w are not consecutive, there is an \ensuremath{\mathbb{H}}_2 using u,p,w as the central path.

Suppose now that u and w have neighbours u',u'' and w',w'', respectively, so that these vertices are pairwise distinct. If the total length of the path on u,v,w is even then we have some \ensuremath{\mathbb{H}}_{2i}. W.l.o.g., assume the path from u to v is even, whereupon the path from v to w must be odd. There must be an edge from v to one of u',u'' (w.l.o.g. u') to avoid an \ensuremath{\mathbb{H}}_{2i}. By minimality of the path u,v,w, the distance from u to v along the path we chose must be 2 with middle vertex s. According to Lemma 48, both s and u' have degree 2 in G. If u has degree greater than 3 there is an \ensuremath{\mathbb{H}}_2 with middle path u,s,v (note that there cannot be an edge from u to the neighbour d of v along the path to w, by our original minimal choice of the path along u,v,w). Suppose u'' has no neighbours outside of \{u,u',s,v\} (noting that there might be an edge from u'' to v). Then v is a cut vertex. Note that none of the paths emanating from w and not going through v can have vertices of degree greater than 2, unless they are part of a squash racket whose distinguished vertex is identified with w, as we would introduce an \ensuremath{\mathbb{H}}_{2i}. The case of the squash racket can be dealt with just like the case of an even cycle so we do not concern ourselves with it specially. Note however that there can be only one squash racket attached at w and if there is a squash racket there can be no additional even cycles or paths attached at w. If the edge u'' to v exists then there can be no new neighbours of u'' without introducing an \mathbb{A} as a subgraph. This places us in the case shown in Figure ?, and the figure shows a star 3-colouring in this case.

Suppose now that u'' has a neighbour outside \{u,u',s,v\} which we call t (noting that in this case there can be no edge from u'' to v since then there would be three consecutive 3^+ vertices). Let us continue expanding from t along some path away from u'' which we enumerate t=t_1,t_2,\ldots. Suppose no such path like this reaches w (or v) avoiding u. Note that such a path cannot now arrive at either v or s since we have uncovered all of their neighbours (if v had more neighbours then there would be an \ensuremath{\mathbb{H}}_2 with central path u,s,v). If one of the vertices t_j has degree greater than 2 then we have a \ensuremath{\mathbb{H}}_{2i} involving a central path either from t_j to v via u or from t_j to w via u. Thus we have nothing more than a path here which must terminate at some t_{k}. This case will be superseded by a more general case which will come later (see Figure ?).

Suppose we have a path t_1,t_2,\ldots that eventually joins the path between v and w exclusive of v and w. Then we contradict minimality of the path from u to w via v. Thus we have a path t_1,t_2,\ldots that eventually joins w. (Possibly it did so via, w.l.o.g., w''.) This path must be of odd length since otherwise it would produce an \ensuremath{\mathbb{H}}_{2i}. If any vertex on this path has degree greater than 2 then it induces some \ensuremath{\mathbb{H}}_{2i} on a path to either u or w. Except in the case that it is distance 2 from w, an odd distance from u, and shares two neighbours with w (and these are w’s only neighbours). This case is drawn in Figure ?. Therefore, there is nothing but this path, except that we did not isolate the degree of w. Indeed, w may have arbitrarily high degree. Any path emanating from w and not eventually reaching v (we have accounted for all such paths) may not have a vertex c of degree greater than 2, since otherwise this vertex would produce an \ensuremath{\mathbb{H}}_{2i} with middle path to either w or v. Except in the case that c is at distance 2 from w and shares two neighbours the two neighbours w' and w'' with w, in which case c may have a path emanating from it in the form of c=c_1, c_2, etc. This case, in which there is a squash racket with distinguished vertex identified with w, will be superseded by the next. Note however that there can be only one squash racket attached at w and if there is a squash racket there can be no additional even cycles or paths. It follows then that some path may leave w and terminate in a vertex of degree 1, or return to w producing a cycle of even length. We claim the graph G is a yes-instance of star 3-colouring. Let us refer to Figure ?.

We may assume without loss of generality assume that G is connected. Suppose G is (\mathbb{A},\overline{2 P_4 + 3 P_6})-subgraph-free. Let the two 3^+ vertices be u and v. There may be many paths from u to v. Suppose one of these paths is of even length. Then, as G is bipartite, they must all be of even length. There may further be even cycles attached to u and even cycles attached to v. In order to avoid \ensuremath{\mathbb{H}}_{2i}, the number of these paths between u and v must be bounded, in particular, there can be no more than one path of length greater than or equal to 4. See Figure ? for the 3-star colouring in this case.

Suppose, one of the paths between u and v is of odd length. Then they must all be of odd length. (There may further be even cycles attached to u and even cycles attached to v.) Suppose there are paths between u and v of length both 1 and 3. Now, u and v must have third neighbours u' and v', respectively, that are not involved in these paths (e.g., by bipartiteness). Since u'\neq v', again by bipartiteness, we have an \mathbb{A} as a subgraph and this is a no-instance.

Now we consider three cases. Firstly, that there is a path of length 1 and all other paths are of length at least 5 (see Figure ?). Secondly, that all paths are of length at least 3 and there are at most two paths of length 5 (see Figure ?). Thirdly, that all paths are of length at least 3 and there is at most one path of length 3 (see Figure ?).

We will essentially follow the proof of Lemma 50 to show that, if we assume there are three 3^+-vertices, we can never uncover more than one more, unless G contains some H_{2i} as a subgraph or contains \mathbb{A} as a subgraph or contains some C_4 with (at least) three 3^+ vertices. Of course, as we follow the proof of Lemma 50, we can no longer use bipartiteness. This means we have an important new case at the start (later on we only used it to assume that certain cycles were even).

Suppose there are three consecutive 3^+-vertices u,v,w in a clique (the “triangle”). Suppose any two of these have a common neighbour, then the graph is not star 3-colourable by Lemma 48. Thus, they have private neighbours u',v',w', respectively, that form with them the “net”. Since G is connected, there is a path from the net, w.l.o.g. through u', to another 3^+ vertex t. Choose such a t closest to the triangle in the net. If t has two neighbours outside of the net, then there is an H_{2i} as a subgraph with central path either t,\ldots,u (if distance from t to u even) or either of t,\ldots,u,v and t,\ldots,u,w (if distance from t to u odd). Thus, t has a neighbour in the net and we have discovered a path from u to (w.l.o.g.) v, avoiding w, with a 3^+ vertex at distance either 1 or 2 from v. Let us choose a minimal such path (not permitting v'=u' as we already ruled this our by assumption). This minimal path induces a cycle in G. If it is C_4 or C_5 we are in a no-instance (in both cases there is a – potentially non-induced – C_5). W.l.o.g., assume it is of length at least 6.

Suppose t is at distance 1 from v and we may w.l.o.g. assume it is v' itself. If v' has an edge to anywhere in the net, then we are in a no-instance (either we have a C_5 or diamond-with-pendant). Thus, some third neighbour v'' of v' is outside the net. Now there is an H_2 with central path v',v,u. Suppose t is at distance 2 from v. We have found two neighbours of t so let t' be a third neighbour, If t' is in the net then we are in a no-instance with a C_5; unless t'=v, in which case we were not in a minimal cycle (essentially v' could be t). Thus, t' is outside the net. Now there is an H_2 with central path t,v',v.

Now let us follow the proof as in the bipartite case (Lemma 50) but no longer using bipartiteness. Suppose u,v,w are consecutive. Then u and w must have another common neighbour to avoid an H_2. Now we violate Lemma 48. Thus, u,v,w may not be consecutive.

Suppose now that u and w have two common neighbours p and q. According to Lemma 48, both of these have degree 2 and neither of these may have an edge to v, therefore v has a distinct neighbour z. Suppose either u or v has degree greater than 3, w.l.o.g. u. Then there is an H_{2} with u,p,w forming the central path. Thus, u and w are of degree 3 and v is a cut vertex. If either of the paths from u to v or v to w is even then we have an H_{2i}. Thus we may assume both are odd. If z has degree greater than 2 then the paths from v to u and v to w must be even to avoid an H_{2i} with central path (e.g.) z,v,\ldots,u. But then there is an H_{2i} with central path v,\ldots,u. Thus, z has degree 1 or 2. Indeed, we may follow a path z=z_1, z_2 etc. and never find a vertex of degree greater than 2, unless it returns to v, as it would imply either an H_{2i} with central path z_j,\ldots,v (if j even) or (e.g.) z_j,\ldots,v,\ldots,u (if j is odd). Thus, either the path ends or it returns to v as a cycle. And there may be more than one such path or cycle. In this case, G has three 3^+-vertices.

Suppose now that u and w have one common neighbour p and two distinct neighbours u' and w', respectively. Since u,v,w are not consecutive, there is an H_2 using u,p,w as the central path.

Suppose now that u and w have neighbours u',u'' and w',w'', respectively, so that these vertices are pairwise distinct. If the total length of the path on u,v,w is even then we have some H_{2i}. W.l.o.g., assume the path from u to v is even, whereupon the path from v to w must be odd. There must be an edge from v to one of u',u'' (w.l.o.g. u') to avoid an H_{2i}. By minimality of the path u,v,w, the distance from u to v along the path we chose must be 2 with middle vertex s. According to Lemma 48, both s and u' have degree 2 in G. If u has degree greater than 3 there is an H_2 with middle path u,s,v (note that there cannot be an edge from u to the neighbour d of v along the path to w, by our original minimal choice of the path along u,v,w). Suppose u'' has no neighbours outside of \{u,u',s,v\} (noting that there might be an edge from u'' to v). Then v is a cut vertex. Note that none of the paths emanating from w and not going through v can have vertices of degree greater than 2, unless they are part of a squash racket whose distinguished vertex is identified with w, as we would introduce an H_{2i}. The squash racket case can be dealt with just like a cycle so we do not concern ourselves with it specially. Note however that there can be only one squash racket attached at w and if there is a squash racket there can be no additional cycles or paths. If the edge u'' to v exists then there can be no new neighbours of u'' without introducing an \mathbb{A} as a subgraph. In this case, G has three 3^+-vertices.

Suppose now that u'' has a neighbour outside \{u,u',s,v\} which we call t (noting that in this case there can be no edge from u'' to v since then there would be three consecutive 3^+-vertices). Let us continue expanding from t along some path away from u'' which we enumerate t=t_1,t_2,\ldots. Suppose no such path like this reaches w (or v) avoiding u. Note that such a path cannot now arrive at either v or s since we have uncovered all of their neighbours (if v had more neighbours then there would be an H_2 with central path u,s,v). If one of the vertices t_j has degree greater than 2 then we have a H_{2i} involving a central path either from t_j to v via u or from t_j to w via u. Thus we have nothing more than a path here which must terminate at some t_{k}. In this case, G has three 3^+-vertices.

Suppose we have a path t_1,t_2,\ldots that eventually joins the path between v and w exclusive of v and w. Then we contradict minimality of the path from u to w via v. Thus we have a path t_1,t_2,\ldots that eventually joins w. (Possibly it did so via, w.l.o.g., w''.) This path must be of odd length since otherwise it would produce an H_{2i}. If any vertex on this path has degree greater than 2 then it induces some H_{2i} on a path to either u or w. Except in the case that it is distance 2 from w, an odd distance from u, and shares two neighbours with w (and these are w’s only neighbours). In this case, G has four 3^+-vertices.

Therefore, there is nothing but this path, except that we did not isolate the degree of w. Indeed, w may have arbitrarily high degree. Any path emanating from w and not eventually reaching v (we have accounted for all such paths) may not have a vertex c of degree greater than 2, since otherwise this vertex would produce an H_{2i} with middle path to either w or v. Except in the case that c is at distance 2 from w and shares two neighbours the two neighbours w' and w'' with w, in which case c may have a path emanating from it in the form of c=c_1, c_2, etc. This case, in which there is a squash racket with distinguished vertex identified with w, will be superseded by the next. Note however that there can be only one squash racket attached at w and if there is a squash racket there can be no additional cycles or paths. It follows then that some path may leave w and terminate in a vertex of degree 1, or return to w producing a cycle. In this case, G has three 3^+-vertices.

Let us give the proof for d=2 where we consider a connected graph G from {\cal G}_2 that has exactly two 3^+-vertices before we make the obvious generalisation. Let u and v be the 3^+-vertices of G. Let P_x be the set of paths from x \in \{u,v\} that terminate with a vertex of degree 1. Let C_x be the set of paths from x \in \{u,v\} that return to x making a cycle. Let P_{u,v} be the set of paths from u to v. These account for all vertices of G. Recall that star colouring has the property that it can see up to distance 3 from a certain vertex in terms of the effect of colours. Consider the paths in P_{u,v} that are of length \geq 3+4+3=10. In any star 3-colouring of a path of length \geq 10, the middle four vertices must contain a consecutive run of three distinct colours. It follows that a path of length 10 covers all longer paths of length 1 \bmod 3. Then, paths of length 11 and 12 cover all longer paths of length 2 and 0 \bmod 3, respectively. Thus, we may w.l.o.g. consider that the paths in P_{u,v} are of length at most 12. Now, suppose there are more than N=3^6 of some particular length m of path. Then in any star 3-colouring of these, some sequence of three colours at the beginning and at the end must appear more than once. Since it appears more than once it could be repeated if there were additional paths of length m over and above the N without violating the condition of star 3-colouring. In this fashion, we may assume that there are at most 3^6 copies of a path of any length. But now we reduced our graph to one of bounded size \leq 5 \cdot 12 \cdot 12 \cdot 3^6 (this is: #sets, #path-lengths, #vertices, #end-types, respectively) and we can check using brute force whether it is star 3-colourable. If it is not star 3-colourable then may add it to our list of forbidden subgraphs.

Now, when d is larger, we consider P_x, C_x, P_{x,y} for all vertices x and y giving a quadratic number in d of sets to consider. The same bound on the length of paths applies as in the case d=2. The argument again reduces any graph to a graph of size at most f(d) for some function f which is O(d^2) and the result follows as in the case d=2.

Let G and G' be the graphs as defined in the proof of Theorem 31. We claim that G has a 3-colouring if and only if G' has an acyclic 3-colouring. First suppose G has a 3-colouring. From the proof of Theorem 31 we find that G' has a star 3-colouring, which is an acyclic 3-colouring by definition. Now suppose G' has an acyclic 3-colouring c. Then no two adjacent vertices u and v in G can be coloured alike by c. Hence, the restriction of c to G is a 3-colouring.

The polynomial results follow from combining the above observation that the problem satisfies C1 with Propositions 29 and 30, respectively. The NP-completeness result follows from Proposition 54.