Distance-regular Cayley graphs with least eigenvalue $-2$

We classify the distance-regular Cayley graphs with least eigenvalue $-2$ and diameter at most three. Besides sporadic examples, these comprise of the lattice graphs, certain triangular graphs, and line graphs of incidence graphs of certain projective planes. In addition, we classify the possible connection sets for the lattice graphs and obtain some results on the structure of distance-regular Cayley line graphs of incidence graphs of generalized polygons.

the recent survey by Van Dam, Koolen, and Tanaka [13]. Here we study the question which distance-regular graphs are Cayley graphs. This question has been well-studied for distance-regular graphs with diameter two, that is, for strongly regular graphs, see the survey paper on partial difference sets by Ma [25]. Miklavič and Potočnik [26,27] classified the distance-regular circulant graphs and distance-regular Cayley graphs on dihedral groups, whereas Miklavič andŠparl [28] studied a particular class of distance-regular Cayley graphs on abelian groups. See also the monograph by Konstantinova [20] for some basic facts and problems on Cayley graphs and distanceregular graphs.
It is well-known that graphs with least eigenvalue −2 have been classified by using root lattices, see [9, §3.12]. In particular, it follows that a distance-regular graph with least eigenvalue −2 is strongly regular or the line graph of a regular graph with girth at least five. The strongly regular graphs with least eigenvalue −2 have been classified by Seidel [34]. We will give an overview of which of these graphs is a Cayley graph and in particular, we will classify the possible connection sets for the lattice graphs, using some general results that we obtain for the distance-regular line graphs of incidence graphs of generalized polygons. We will also classify the Cayley graphs with diameter three among the distance-regular line graphs, in particular the line graphs of Moore graphs and the line graphs of incidence graphs of projective planes. What remains open is to classify which line graphs of incidence graphs of generalized quadrangles and hexagons are Cayley graphs.

Preliminaries
Let G be a finite group with identity element e and S ⊆ G \ {e} be a set such that S = S −1 (we call S inverse-closed). An (undirected) Cayley graph Cay(G, S) with connection set S is the graph whose vertex set is G and where two vertices a and b are adjacent (denoted by a ∼ b) whenever ab −1 ∈ S. The Cayley graph Cay(G, S) is connected if and only if the subgroup S generated by S is equal to G. In the literature, it is sometimes assumed explicitly that a Cayley graph is connected. In this case, the connection set is also called a generating set. Here we follow the terminology used by Alspach [4]. We denote the order of an element a ∈ G by O(a), the subgroup generated by a by a and the cyclic group of order n by Z n . Furthermore, the cycle graph of order m is denoted by C m and the line graph of a graph Γ by L(Γ ).

Distance-regular graphs
A strongly regular graph with parameters (v, k, λ , µ) is a k-regular graph with v vertices such that every pair of adjacent vertices has λ common neighbors and every pair of non-adjacent vertices has µ common neighbors. Here we exclude disjoint unions of complete graphs and edgeless graphs, and therefore strongly regular graphs are connected with diameter two.
A connected graph with diameter d is distance-regular whenever for all vertices x and y, and all integers i, j ≤ d, the number of vertices at distance i from x and distance j from y depends only on i, j, and the distance between x and y. A distance-regular graph with diameter two is the same as a strongly regular graph.
A generalized d-gon is a point-line incidence structure whose (bipartite) incidence graph has diameter d and girth 2d. It is of order (s,t) if every line contains s + 1 points, and every point is on t + 1 lines. For s = t, both the incidence graph and its line graph are distance-regular. This line graph can also be viewed as the point graph of a generalized 2d-gon of order (s, 1). For some basic background on generalized polygons, we refer to the monographs by Godsil and Royle [16, §5.6] and Brouwer, Cohen, and Neumaier [9, §6.5].
The (adjacency) spectrum of a graph is the multiset of eigenvalues of its adjacency matrix. As mentioned in the introduction, distance-regular graphs with least eigenvalue −2 can be classified. In particular, we have the following. -If d = 2, then Γ is a cocktail party graph, a triangular graph, a lattice graph, the Petersen graph, the Clebsch graph, the Shrikhande graph, the Schläfli graph, or one of the three Chang graphs, -If d = 3, then Γ is the line graph of the Petersen graph, the line graph of the Hoffman-Singleton graph, the line graph of a strongly regular graph with parameters (3250, 57, 0, 1), or the line graph of the incidence graph of a projective plane, -If d = 4, then Γ is the line graph of the incidence graph of a generalized quadrangle of order (q, q), -If d = 6, then Γ is the line graph of the incidence graph of a generalized hexagon of order (q, q).
Recall that the triangular graph T (n) is the line graph of the complete graph K n , the lattice graph L 2 (n) is the line graph of the complete bipartite graph K n,n (a generalized 2-gon), and the cocktail party graph CP(n) is the complete multipartite graph with n parts of size two. Note also that a projective plane is a generalized 3-gon.
We note that the distance-regular graphs with least eigenvalue larger than −2 are also known. Besides the complete graphs (with least eigenvalue −1), there are the cycles of odd length, and these are clearly Cayley graphs.

Vertex-transitivity and edge-transitivity
Recall that a graph Γ is vertex-transitive whenever the automorphism group of Γ acts transitivity on the vertex set of Γ , i.e. if x is a fixed vertex of Γ , then {x σ |σ ∈ Aut(Γ )} is equal to the set of vertices of Γ . It is clear that Cayley graphs are vertex-transitive. In fact, a graph Γ is a Cayley graph if and only if the automorphism group Aut(Γ ) of Γ contains a regular subgroup, see [4,Thm. 2.2].
A graph Γ is called edge-transitive whenever the automorphism group of Γ acts transitivity on the edge set of the graph. Because line graphs play an important role in this paper, also the concept of edge-transitivity is relevant. Indeed, the following result provides us with a connection between the vertex-transitivity of the line graph of a graph Γ and the edge-transitivity of Γ . [31,Thm. 5.3] Let Γ be a connected graph which is not isomorphic to the complete graphs K 2 , K 4 , a triangle with an extra edge attached, and two triangles sharing an edge. Then the automorphism group of Γ and its line graph are isomorphic, with the natural group isomorphism ϕ : Aut(Γ ) → Aut(L(Γ )), defined by ϕ(σ ) = σ for σ ∈ Aut(Γ ), where σ acts on the line graph of Γ such that σ ({v, w}) = {σ (v), σ (w)}, where v and w are adjacent in Γ .

Lemma 2.3 A connected regular graph is edge-transitive if and only if its line graph is vertex-transitive.
Proof Let Γ be connected and regular. If Γ is isomorphic to K 2 or K 4 , then Γ is edge-transitive and the line graph of Γ is vertex-transitive. On the other hand, if Γ is not isomorphic to K 2 or K 4 , then the automorphism group of Γ and its line graph are isomorphic with the natural group isomorphism by Theorem 2.2, which completes the proof. ⊓ ⊔

Groups and products
Two subgroups H and K in G are conjugate whenever there exists an element g ∈ G such that K = g −1 Hg. The semidirect product G of a group N by a group H is denoted by H ⋉ N or N ⋊ H. It has the property that it contains a normal subgroup N 1 isomorphic to N and a subgroup H 1 isomorphic to H such that G = N 1 H 1 and Let G be a finite group with subgroups H and K such that G = HK and the intersection of H and K is the identity of G. Then G is called a general product of H and K (see [11]).
Finally we mention a result that we will use in Section 5.2.

Some results on generalized polygons
Let Γ be a distance-regular line graph of the incidence graph of a generalized d-gon of order (q, q). Then Γ can also be seen as the point graph of a generalized 2d-gon of order (q, 1). It follows that each vertex of Γ is contained in two maximal cliques, of size q + 1, and every edge of Γ is contained in a unique maximal clique. Thus, Γ does not have K 1,3 nor K 1,2,1 as an induced subgraph. Moreover, Γ has diameter d and every induced cycle in Γ is either a 3-cycle or a 2d-cycle. We will use these properties to derive some general results on the structure of the connection set in case Γ is a Cayley graph. We first claim that a ⊆ S ∪ {e} for all a ∈ S. In order to prove this, let a ∈ S and n = O(a) = 2d. If n = 2 or 3, then a ⊆ S ∪{e} since S = S −1 . If n ≥ 4, then it is clear that the induced subgraph Γ a of Γ on a contains a cycle e ∼ a ∼ a 2 ∼ · · · ∼ a n−1 ∼ e of length n (see also [3,Lemma 2.6]). Because n = 3 and n = 2d, it follows that this cycle is not an induced cycle. Thus, there must be an extra edge in Γ a , that is, an edge that is not generated by a or a −1 , and hence a i ∈ S for some i with 1 < i < n − 1. Now e is adjacent to a, a −1 , and a i , and because Γ does not contain an induced subgraph K 1,3 , it follows that a 2 ∈ S, or a i−1 ∈ S, or a i+1 ∈ S. Let us consider the case that a i−1 ∈ S, with i > 2. By considering the induced subgraph on {e, a, a i−1 , a i }, it follows that a i−2 ∈ S because Γ does not contain an induced subgraph K 1,2,1 . Similarly, by considering the induced subgraph on {e, a −1 , a i−1 , a i }, it follows that a i+1 ∈ S. By repeating this argument, it follows that a ⊆ S ∪ {e}. The other cases go similarly, which proves our claim.
Let H and K be the two cliques of size q + 1 that contain e. Then S = (H ∪K)\ {e} and H ∩ K = {e}. What remains to be shown is that H and K are subgroups of G.
Let a ∈ H \ {e}. Because the graph induced on a is a clique, and there are no edges between H \ {e} and K \ {e}, it follows that a ⊆ H, In particular, a −1 ∈ H. Now let a, b ∈ H, and let us show that ba −1 ∈ H, thus showing that H is a subgroup of G. If a = b, a = e, or b = e, then this clearly implies that ba −1 ∈ H. In the other cases, we have that b ∼ a, so ba −1 ∈ S. Because ba −1 ∼ a −1 , and there are no edges between H \ {e} and K \ {e}, it follows that ba −1 ∈ H. Thus, H -and similarly K -is a subgroup of G.
⊓ ⊔ The condition that a ⊆ S ∪ {e} for every element a of order 2d in S is not redundant. Indeed, the lattice graph L 2 (2), which is the line graph of K 2,2 (the incidence graph of a generalized 2-gon of order (1, 1)) is isomorphic to the Cayley graph Cay(Z 4 , {±1}). Both elements a in S = {±1} have order 4, but a 2 / ∈ S, and indeed S ∪ {e} does not contain a nontrivial subgroup of Z 4 .
The proof of Theorem 3.1 indicates that the condition a ⊆ S ∪ {e} for every element a of order 2d in S can be replaced by the condition that a 2 ∈ S for every element a of order 2d in S. We can in fact generalize this as follows. Proof Let a be of order 2d in S, and assume that there exists an element s ∈ S such that s = a, a −1 and sas −1 ∈ S. By Theorem 3.1, it suffices to prove that a ⊆ S ∪ {e}.
Because e is adjacent to a, a −1 , and s, and G has no induced subgraph K 1,3 , it follows that there is at least one edge within {a, a −1 , s}. If a and a −1 are adjacent, then a 2 ∈ S. Because Γ does not contain an induced subgraph K 1,2,1 , it then follows by induction and by considering the induced subgraph on {e, a, a i , a i+1 } (for i ≥ 2) that a ⊆ S ∪ {e}. So let us assume that a and a −1 are not adjacent. Without loss of generality, we may thus assume that s is adjacent a −1 , and hence that sa ∈ S. Now e is adjacent to sa, a and s. Furthermore, sa is adjacent to a and s since sas −1 ∈ S. It follows, again because Γ does not contain an induced subgraph K 1,2,1 , that a is adjacent to s. Using the same argument once more gives that a and a −1 are adjacent, which is a contradiction that finishes the proof. ⊓ ⊔

Remark 3.3
In view of the above, if there exists an element a ∈ S of order 2d such that a S ∪ {e}, then a and a −1 are not adjacent. We may therefore assume that a ∈ H and a −1 ∈ K, where H and K are the two maximal cliques (but not subgroups) containing e. But clearly the set Ka is a maximal clique containing a and e. Because every edge is in a unique maximal clique, it follows that Ka = H. Therefore, in this case, S = (K ∪ Ka) \ {e}. In the case of the Cayley graph Cay(Z 4 , {±1}), we indeed have K = {−1, 0} and H = K + 1.
As a first application of the above, we obtain that the (distance-regular) line graph of the Tutte-Coxeter graph is not a Cayley graph.

Proposition 3.4 The line graph of the Tutte-Coxeter graph is not a Cayley graph.
Proof The Tutte-Coxeter graph is the incidence graph of a generalized quadrangle (4-gon) of order (2, 2). It has 30 vertices and 45 edges. If its line graph is a Cayley graph Cay(G, S), then |G| = 45 and |S| = 4. Because G has no element of order 8, it follows from Theorem 3.1 that there exist two subgroups H and K of G such that S = (H ∪ K) \ {e}, where |H| = |K| = 3 and H ∩ K = {e}. Furthermore, the group G is an abelian group isomorphic to Z 3 × Z 3 × Z 5 or Z 9 × Z 5 since G has only one subgroup of order 9 and one subgroup of order 5 by Sylow's theorems. By the structure of the connection set S, it now follows that G must be the abelian group isomorphic to Z 3 × Z 3 × Z 5 but in this case the Cayley graph Cay(G, S) is not connected, a contradiction. Therefore the line graph of the Tutte-Coxeter graph is not a Cayley graph.

⊓ ⊔
We finish this section with a result that shows that the obtained structure of S in the above fits naturally with line graphs of bipartite graphs. Proof From the structure of S, it follows that each vertex is in two maximal cliques, and every edge is in a unique maximal clique. By a result of Krausz [21] (see [37,Thm. 7.1.16]) it follows that Γ is a line graph of a graph, Γ ′ , say. The graph Γ ′ has the maximal cliques of Γ as vertices, and two such cliques are adjacent in Γ ′ if and only if they intersect; the corresponding edge in Γ ′ is the vertex in Γ that is contained in both cliques.
Because S = (H ∪ K) \ {e} and H ∩ K = {e}, we can distinguish between two kinds of maximal cliques. We call such a clique an H-clique if the edges in the clique are generated by an element in H, and the other cliques are similarly called K-cliques. Now it is clear that every edge in Γ ′ has one vertex in the set of H-cliques and the other vertex in the set of K-cliques. Thus Γ ′ is bipartite.

Strongly regular graphs
In this section, we will determine which strongly regular graphs with least eigenvalue −2 are Cayley graphs, using the case of diameter d = 2 in the classification given in Theorem 2.1.

The sporadic graphs
Besides the three infinite families of strongly regular graphs with least eigenvalue −2, we have to consider the Petersen graph, the Clebsch graph, the Shrikhande graph, the Schläfli graph, and the Chang graphs. The Petersen graph is the unique strongly regular graph with parameters (10, 3, 0, 1). It is the complement of the line graph of the complete graph K 5 , and therefore it is not a Cayley graph, by Corollary 4.6 below (see also [16,Lemma 3.1.3]).

Proposition 4.1 (Folklore). The Petersen graph is not a Cayley graph.
It is well-known that the complement of the Clebsch graph is the folded 5-cube, which is strongly regular with parameters (16, 5, 0, 2) (see [10, p. 119]). The ddimensional cube Q d is the distance-regular graph whose vertex set can be labeled with the 2 d binary d-tuples such that two vertices are adjacent whenever their labels differ in exactly one position (clearly this is a Cayley graph). The folded d-cube is the distance-regular graph that can be obtained from the cube Q d−1 by adding a perfect matching that connects vertices at distance d − 1 (see [7]). It is evident that the folded Thus, the Clebsch graph is a Cayley graph.
The Shrikhande graph is a strongly regular graph with the same parameters as the lattice graph L 2 (4) and can be constructed as a Cayley graph This construction 'on the torus' is accredited to Biggs [6] by Gol'fand, Ivanov, and Klin [17, p. 182].
The Schläfli graph is the unique strongly regular graphs with parameters (27,16,10,8). It follows from the work by Liebeck, Praeger, and Saxl [23] (see also [22,Lemma 2.6]) that it is a Cayley graph over the semidirect product Z 9 ⋊ Z 3 . Using GAP [36], we checked that with G = Z 9 ⋊ Z 3 = a, b|a 9 = b 3 = 1, b −1 ab = a 7 and S = {a, a 8 , a 3 , a 6 , b, b 2 , a 7 b, a 5 b 2 , a 2 b, a 4 b 2 }, the Cayley graph Γ = Cay(G, S) indeed is the complement of the Schläfli graph. Note that Γ is also the point graph of the unique generalized quadrangle of order (2, 4), with lines thus being the triangles in Γ . Therefore, these lines can be obtained as the right 'cosets' of the five triangles {e, a, a 2 b}, {e, a 8 , a 7 b},{e, a 3 , a 6 },{e, b, b 2 },{e, a 5 b 2 , a 4 The Schläfli graph can also be constructed as a Cayley graph over (Z 3 × Z 3 )⋊ Z 3 , the other nonabelian group of order 27. Indeed, we again checked with GAP [36] that Γ = Cay(G ′ , S ′ ) for G ′ = a, b, c|a 3 = b 3 = c 3 = e, abc = ba, ac = ca, bc = cb and , a 2 , b, b 2 , c, c 2 , cba, a 2 b 2 c 2 , aba, bab}. In this case all nonidentity elements of the group have order 3, and hence the triangles through e are subgroups H 1 , . . . , H 5 of G ′ , with trivial intersection and S ′ = (H 1 ∪ · · · ∪ H 5 ) \ {e} (cf. Theorem 3.1). Again, the cosets of these subgroups give the lines of the generalized quadrangle of order (2, 4). From the above, we conclude the following. The Chang graphs are strongly regular graphs with the same parameters as the line graph of the complete graph K 8 . These three graphs can be obtained by Seidel switching in L(K 8 ). According to [8], the orders of the automorphism groups of these graphs are 384, 360, and 96, respectively.

Proposition 4.3 The three Chang graphs are not Cayley graphs.
Proof Let Γ be one of the Chang graphs, and suppose on the contrary that it is a Cayley graph, and hence that it is vertex-transitive. Let x be a fixed vertex in Γ . Then the order of {x σ |σ ∈ Aut(Γ )} is equal to 28 since Γ is vertex-transitive. It follows that the index of Aut(Γ ) over the stabilizer of x is 28. Therefore 28 must divide the order of Aut(Γ ), which is a contradiction.

The infinite families
A cocktail party graph is a complete multipartite graph with parts of size two, and clearly such a graph is a Cayley graph. By [1,Prop. 2.6], we obtain the following result.

Proposition 4.4 A Cayley graph Cay(G, S) is a cocktail party graph if and only if G
has an element a of order 2 and S = G \ a .
Consider a set X of size n and let V be the collection of all subsets of size m in X, with m ≥ 2 and n ≥ 2m + 1. The Kneser graph K(n, m) is the graph with vertex set V such that two vertices A and B in V are adjacent whenever |A ∩ B| = 0. The Kneser graph K(n, 2) is the complement of the triangular graph T (n). Godsil [14] characterized the Cayley graphs among the Kneser graphs.
Theorem 4.5 [14] Except in the following cases, the Kneser graph K(n, m) is not a Cayley graph.
As a corollary, we obtain a result first obtained by Sabidussi [32]. Note that the triangular graphs T (2) and T (3) are complete graphs, and that T (4) is isomorphic to the cocktail party graph CP(3). Corollary 4.6 [32] The triangular graph T (n) is a Cayley graph if and only if n = 2, 3, 4 or n ≡ 3 (mod 4) and n is a prime power.
Godsil [15] gave the following construction of the triangular graph T (n) as a Cayley graph Cay(G, S) for prime powers n ≡ 3 (mod 4). Let F be the field of order n. For a, b ∈ F, let the map T a,b : F → F be defined by T a,b (x) = ax+b. Let G be the group of maps T a,b , with a a non-zero square and b arbitrary. It is not hard to see that G acts regularity on the edges of the complete graph K n (with vertex set F), using that −1 is a non-square (whence the assumption that n ≡ 3 (mod 4)). As connection set S one can take the set of maps T a,b ∈ G such that either T a,b (0) ∈ {0, 1} or T a,b (1) ∈ {0, 1} (thus mapping the vertex {0, 1} of the triangular graph to an adjacent vertex).
As a final family of strongly regular graphs with least eigenvalue −2, we consider the lattice graphs. Let n ≥ 2. The lattice graph L 2 (n) is the line graph of the complete bipartite graph K n,n . It is isomorphic to the Cartesian product of two complete graphs K n , and hence to the Cayley graph Cay(Z n × Z n , {(0, 1), . . . , (0, n − 1), (1, 0), . . . , (n − 1, 0)}). Because K n,n is the incidence graph of a generalized 2-gon, we can apply the results of Section 3. We will use these to give a characterization of the lattice graphs as Cayley graphs. Theorem 4.7 Let n ≥ 2, let G be a finite group, S be an inverse-closed subset of G, and let Γ = Cay(G, S). Then the following hold: -If G is a general product of two of its subgroups H and K of order n and S = (H ∪ K) \ {e}, then Γ is isomorphic to the lattice graph L 2 (n), -If Γ is isomorphic to the lattice graph L 2 (n) and a ⊆ S ∪ {e} for every element a of order 4 in S, then G is a general product of two of its subgroups H and K of order n and S = (H ∪ K) \ {e}.
Proof Let G be a general product of two of its subgroups H and K of order n and let S = (H ∪ K) \ {e}. Then |G| = n 2 . By using the results in Section 3, we know that every vertex in Γ is in two maximal cliques of size n. A simple counting argument shows that there are 2|G|/n = 2n maximal cliques in Γ . By Lemma 3.5 and its proof, it follows that Γ is the line graph of a bipartite graph Γ ′ on the 2n maximal cliques of Γ , and that each clique is on n edges in Γ ′ . This implies that Γ ′ is the complete bipartite graph K n,n , and hence Γ is the lattice graph L 2 (n).
To prove the second item, suppose that Γ is isomorphic to the lattice graph L 2 (n) and a ⊆ S ∪ {e} for every element a of order 4 in S. It follows by Theorem 3.1 that there are subgroups H and K of order n in G such that H ∩ K = {e} and S = (H ∪ K) \ {e}. Now K is a maximal clique in Γ . Let g be a vertex not in K. Then the structure of the lattice graph implies that g is adjacent to precisely one vertex k ∈ K. Thus gk −1 ∈ S, and hence it follows that gk −1 ∈ H (because if it were in K, then so would g), so g = hk for some h ∈ H. Therefore G is the general product of H and K, which completes the proof.

⊓ ⊔
We recall from Section 3 that the lattice graph L 2 (2) is isomorphic to the Cayley graph Cay(Z 4 , {±1}), which is an example such that G = Z 4 cannot be written as a general product HK with inverse-closed sets H and K of size 2. We now conclude this section by giving the classification of all strongly regular Cayley graphs with least eigenvalue at least −2 (which follows from the above). Recall that the only strongly regular graph with least eigenvalue larger than −2 is the 5-cycle. Theorem 4.8 A graph Γ is a strongly regular Cayley graph with least eigenvalue at least −2 if and only if Γ is isomorphic to one of the following graphs.

Distance regular graphs with diameter three
In this section, we will determine which distance-regular graphs with least eigenvalue −2 and diameter three are Cayley graphs. By the classification given in Theorem 2.1, we again have to consider a few sporadic examples and an infinite family. Proof Let Γ be a strongly regular graph with parameters (3250, 57, 0, 1) and suppose that the line graph of Γ is a Cayley graph. Then L(Γ ) is vertex-transitive and therefore Γ is edge-transitive by Lemma 2.3. On the other hand, it is known that Γ is not vertextransitive, see [10,Prop. 11.2], and therefore Γ must be bipartite by [16,Lemma 3.2.1], which is a contradiction.

The line graphs of Moore graphs
⊓ ⊔

The line graphs of the incidence graphs of projective planes
Recall that a projective plane of order q is a point-line incidence structure such that each line has q + 1 points, each point is on q + 1 lines, and every pair of points in on a unique line. It is the same as a generalized 3-gon of order (q, q) and a 2-(q 2 + q + 1, q + 1, 1) design. Currently, projective planes of order q are only known to exist for prime powers q, and for q = 1. For q > 1, the classical construction of a projective plane of order q uses the finite field GF(q) and gives the so-called Desarguesian plane of order q. We note that Loz, Mačaj, Miller,Šiagiová,Širáň, and Tomanová [24] showed that the (distance-regular) incidence graph of a Desarguesian plane is a Cayley graph. Here we will consider the line graph, however. For q = 1, the line graph of the incidence graph is a 6-cycle, which is a Cayley graph. We therefore assume from now on that q > 1. We note that the dual incidence structure of a projective plane is also a projective plane; if a projective plane is isomorphic to its dual, then we say it is self-dual. Consider now a projective plane π of order q, and let Γ π be the incidence graph of π. Recall from Theorem 2.2 that the automorphism group of Γ π and its line graph L(Γ π ) are isomorphic. A collineation (automorphism) of π is a permutation of the points and lines that maps points to points, lines to lines, and that preserves incidence. If π is not self-dual, then an automorphism of the incidence graph Γ π must be a collineation. Additionally, if the projective plane is self-dual, then the automorphism group of Γ π has index 2 over the automorphism group of π; in this case the plane has so-called correlations (isomorphisms between the plane and its dual; see also [29]) on top of collineations.
By construction, a vertex in L(Γ π ) corresponds to an incident point-line pairalso called flag -of π. If L(Γ π ) is a Cayley graph (or more generally, is vertextransitive), then we have a group of collineations and correlations of π that is transitive on flags. In particular, we have the following lemma. Lemma 5.4 Let π be a projective plane of order q, with q even. If L(Γ π ) is a Cayley graph, then π has a collineation group acting regularly on its flags.
Proof If L(Γ π ) is a Cayley graph, then there must be a group G of automorphisms of Γ π acting regularly on the edges of Γ π . The group G therefore has order (q + 1)(q 2 + q + 1). Moreover, G is (isomorphic to) a group of collineations and correlations of π that acts regularly on its flags. If this group contains correlations, then it has an index 2 subgroup of collineations, but this is impossible because the order of G is odd. Hence π has a collineation group acting regularly on its flags.

⊓ ⊔
For q even, we can therefore use the following characterization by Kantor [19].
Theorem 5.5 [19,Thm. A] Let q ≥ 2, let π be a projective plane of order q, and let F be a collineation group of π that is transitive on flags. Then either -PSL(3, q) is contained in F and π is Desarguesian, or -F is a Frobenius group of odd order (q + 1)(q 2 + q + 1), and q 2 + q + 1 is prime.
Recall that PSL (3, q) is the projective special linear group, which has order If L(Γ π ) is a Cayley graph Cay(G, S), then |G| = (q 2 + q + 1)(q + 1), and the action of G on the flags of π must be regular. Because the order of PSL(3, q) is larger than |G|, it follows that G is a Frobenius group of odd order (q 2 + q + 1)(q + 1), and that q 2 + q + 1 is prime. Recall that a Frobenius group is a group F which has a non- i.e. F is the semidirect product N ⋊ H (see [30]).

Proposition 5.6
If the line graph of the incidence graph of a projective plane π of order q is a Cayley graph Cay(G, S), where G corresponds to a group of collineations of π, then G is N ⋊ H in which N is a normal subgroup of prime order q 2 + q + 1 and H is a subgroup of odd order q + 1.
Proof It follows from the above that G is a Frobenius group of odd order (q 2 + q + 1)(q+1), and q 2 +q+1 is a prime number. It follows that G has a normal (q 2 +q+1)-Sylow subgroup N of order q 2 + q + 1 by Sylow's theorems. On the other hand, there exists a subgroup H of order q + 1 in G by Theorem 2.4, and the intersection of N and H is the identity element of G. Therefore G is N ⋊ H.

⊓ ⊔
It is widely believed that there is no non-Desarguesian plane admitting a collineation group acting transitively on flags. Thas and Zagier [35] showed that if such a plane exists, then its order is at least 2 × 10 11 .
On the other hand, Higman and McLaughlin [18] showed that the only Desarguesian planes admitting a collineation group acting regularly on flags are those of order 2 and 8. Indeed, the line graphs of the incidence graphs of these projective planes can be constructed as Cayley graphs as follows: Example 5. 7 The Heawood graph is the incidence graph of the Fano plane; its line graph is the unique graph with spectrum {4 1 , (1 + √ 2) 6 , (1 − √ 2) 6 , −2 8 } (see [12]). Let G = Z 7 ⋊ Z 3 = a, b|a 7 = b 3 = e, b −1 ab = a 2 . Let H = b , K = a −1 ba and S = (H ∪ K) \ {e} (cf. Theorem 3.1). By using GAP [36] and similar codes as in [2, p. 4], it is checked that the Cayley graph Cay(G, S) is indeed the line graph of the Heawood graph.
Similarly the line graph of the incidence graph of the (unique) projective plane of order 8 is obtained by taking G = Z 73 ⋊ Z 9 = a, b|a 73 = b 9 = e, b −1 ab = a 2 , H = b , K = a −1 ba , and S = (H ∪ K) \ {e}.
We may thus conclude the following. Theorem 5.8 Let Γ be a distance-regular Cayley graph with diameter three and least eigenvalue at least −2. Then Γ is isomorphic to one of the following graphs.
-The cycle C 6 or C 7 , -The line graph of the incidence graph of the Desarguesian projective plane of order 2 or 8, -The line graph of the incidence graph of a non-Desarguesian projective plane of order q, where q 2 + q + 1 is prime and q is even and at least 2 × 10 11 , -The line graph of the incidence graph of a projective plane of odd order with a group of collineations and correlations acting regularly on its flags.
It would be interesting to find out whether any of the results on collineations of projective planes can be extended to groups of collineations and correlations, and thus rule out the final case of Theorem 5.8. We could not find any such results in the literature.
Besides the line graph of the Tutte-Coxeter graph (see Proposition 3.4) we leave the case of the line graphs of incidence graphs of generalized quadrangles and hexagons open (cf. Theorem 2.1). For some results on flag-transitive generalized quadrangles, we refer to Bamberg, Giudici, Morris, Royle, and Spiga [5]; for flag-transitive generalized hexagons, we refer to Schneider and Van Maldeghem [33].