Andrásfai and Vega graphs in Ramsey-Turán theory

Define

Consider a triangle-free graph G on n vertices with \alpha(G)\le s and e(G)=\mathrm{ex}(n,s), where

Since \gamma is sufficiently small, we have s<(k-1)n/(3k-4) and Fact 5.2 implies e(G)\le g_\ell(n, s) for every \ell\ne k. On the other hand, Lemma 3.5 yields

and it remains to prove that this holds with equality. It will be convenient to rewrite (?) as

Since

and

we have

and thus

Therefore, the set

satisfies

and, consequently,

Now our argument will proceed as follows. We shall verify that the graph G'=G-A satisfies the assumptions of Theorem 1.3 and, hence, e(G') is bounded from above by g_k(n-|A|,s). Then, using the fact that all vertices in A are of small degree, we derive an upper bound of g_k(n,s) on the number of edges in G.

For the minimum degree of G' we obtain

As the graph G' is triangle-free and satisfies \alpha(G')\le s, this shows that Theorem 1.3 applies to G' and we are led to

Now our choice of \zeta and \gamma yields

and for this reason (?) can be continued to

Every vertex in A has degree at most (1-\zeta)s in G, so we arrive at

which together with (?) concludes the proof of Theorem 1.2.

Observe that adding any edges to H can neither increase its independence number nor decrease its minimum degree. Thus, we may and shall assume that H is a maximal triangle-free graph.

Due to Fact 5.2, it suffices to consider the case

Since \delta(H) > (k+1)n/(3k+2) > n/3, Theorem 5.1 tells us that H is a proper blow-up of some graph J, which is either an Andrásfai graph \Gamma_{\ell}, or a Vega graph {\Upsilon}_{i}^{\mu\nu}. In the latter case we set \ell=9i-(6+\mu+\nu). Now in both cases J has a proper \ell-regular blow-up on 3\ell-1 vertices and Lemma 2.2 yields k = \ell. If J=\Gamma_k is an Andrásfai graph the assertion follows from Lemma 3.3 and if J is a Vega graph we use Lemma 4.1 instead.

The double counting argument

establishes part ?. If H is a proper blow-up, then (?) yields

and part ? follows.

Notice that every balanced blow-up H(p) of H is a proper blow-up of J as well. By applying this observation to a sufficiently large integer p we learn that we may assume, without loss of generality, that H is a proper blow-up of F. Now Fact 2.1? reveals

which together with (?) implies

Taking reciprocals we obtain

i.e., d-1<k<d+1.

As usual, we write the vertex set of F in the form V(F)=\{v_1, \ldots, v_r\}. Setting x_i=|V_i| for i\in [r] we have

which proves the desired upper bound on e(H).

Suppose from now on that this estimate holds with equality and that rs\ne kn. This means that (x_i-x)|N(V_i)|=(x_i-x)s holds for every i\in [r], or in other words that x_i>x implies |N(V_i)|=s. This proves the second bullet. Now, if the first bullet fails, we have |N(V_i)|=s for every i\in [r] and (?) yields the contradiction rs=kn.

In order to prove the desired upper bound on e(H) we remark that Lemma 2.3 applied to r=3k-1 and x=(k-1)n-(3k-4)s yields

Let us now study the case that equality holds in this estimate. If clause ? failed, then the trivial upper bound e(H)\le ns/2 would contradict e(H)=g_k(n, s).

The remaining two clauses follow from the moreover-part in Lemma 2.3 provided that its assumption (3k-1)s\ne kn holds. So it remains to deal with the case s=kn/(3k-1). Now (k-1)n - (3k-4)s = n/(3k-1) is at the same time the average size of the vertex classes of H and a lower bound on the sizes of these vertex classes. In other words, H is the balanced blow-up F\bigl(n/(3k-1)\bigr) and ?, ? hold trivially.

The first part of ? follows from \Gamma_{k-1}\cong \Gamma_k - \{v_0, v_k, v_{2k}\}. Furthermore, the graph \Gamma_k^-=\Gamma_k-\{v_{k}\} satisfies N(v_0, \Gamma_k^-)\subseteq N(v_1, \Gamma_k^-) and N(v_{2k}, \Gamma_k^-)\subseteq N(v_{2k-1}, \Gamma_k^-), which proves \Gamma_k^- to be a subgraph of the blow-up of \Gamma_k - \{v_0, v_k, v_{2k}\} obtained by doubling the vertices v_1 and v_{2k-1}. This establishes the second part of ? and the last part follows inductively.

Proceeding with ? we observe that, due to ?, \Gamma_k contains a pentagon \Gamma_2 as a subgraph, whence \chi(\Gamma_k)\ge \chi(\Gamma_2)=3. To verify the reverse inequality, we partition V(\Gamma_k) into the three independent sets \{v_0, v_1, \ldots, v_{k-1}\}, \{v_k, v_{k+1},\ldots, v_{2k-1}\}, and \{v_{2k}, v_{2k+1}, \ldots, v_{3k-2}\}.

Finally, let S\subseteq V(\Gamma_k) be a nonempty independent set we want to cover by the neighbourhood of an appropriate vertex. By symmetry we may suppose that v_k\in S, which due to (?) entails S\subseteq \{v_1, v_2, \ldots, v_{2k-1}\}. Let i, j\in [2k-1] be the smallest and largest index with v_i, v_j\in S. Now S\subseteq \{v_i, \ldots, v_j\}, a further application of (?) shows j-i \le k-1, and altogether we have S\subseteq N(v_{j+k}).

The upper bound holds because of

By applying this estimate to V_{i+k+1}, \ldots, V_{i+2k-2} instead of V_i we infer

As usual we denote the vertex classes of H corresponding to the vertices of \Gamma_k by V_0, \ldots, V_{3k-2}. Recall that by Fact 3.2 we have

In particular, H has the properties enumerated in the moreover-part of Corollary 2.4 and clause ? corresponds to the estimates on s stated in the lemma. Next, ? allows us to assume, without loss of generality, that

In the special case

we have

which yields |V_i|=3s-n for every i\ne 0, k, 2k-1, meaning that a=(k-1)n - (3k-4)s exemplifies H\in \mathscr{G}^n_k(s). By symmetry we may therefore suppose |V_{2k-1}|>(k-1)n - (3k-4)s from now on, which in view of Corollary 2.4? entails

i.e.,

Because of |V_1|+\ldots+|V_k|=|N(V_{2k})|\le s and (Proof) this yields

which together with (?) establishes the first bullet in Definition 3.4.

Next we observe that, since by (?) the lower bound on |V_0| provided by Fact 3.2 holds with equality, an easy inspection of the proof of Fact 3.2 discloses

and for the same reason (?) leads to |V_{2k+1}|=\ldots=|V_{3k-2}|=3s-n. Combined with (?) these equations confirm the third bullet in Definition 3.4 and, finally, the second bullet follows easily in the light of v(H)=n and (Proof).

The upper bound on |X| +|Y| follows from

Similarly,

yields the desired upper bound on |V_0| + |V_{2i-1}|. It remains to prove |Z|\le \omega_{\mu\nu}(z)(3s-n) for every z\in V(\Upsilon_i)\smallsetminus\{v_0, v_{2i-1}, x, y\}. Following the same strategy we have just used, this task reduces to exhibiting a list of 3\omega_{\mu\nu}(z) vertices of \Upsilon_i whose neighbourhoods cover the entire vertex set \omega_{\mu\nu}(z) many times and the set Z itself even once more. As there are several cases and plenty of vertices to consider, it seems useful to devise the following notation. For a set Q\subseteq V(\Upsilon_i) we denote its characteristic function by {\mathds 1}(Q), so

for all q\in V(\Upsilon_i). If Q=\{q_1, \ldots, q_r\} comes with an explicit enumeration of its members, it will be convenient to omit a pair of curly braces, thus writing {\mathds 1}(q_1, \ldots, q_r) for this function. For instance, the functions f and g considered earlier can now be represented as

Instead of {\mathds 1}\bigl(V(\Upsilon_i)\bigr) we shall just write {\mathds 1}. Next, given a function h\colon V(\Upsilon_i)\longrightarrow{\mathds Z} we let \Sigma(h)=h(V(\Upsilon_i))=\sum_{t\in V(\Upsilon_i)}h(t) be the sum of the values it attains and by \mathscr{N}(h)\colon V(\Upsilon_i)\longrightarrow{\mathds Z} we mean the function \sum_{t\in V(\Upsilon_i)}h(t)\,{\mathds 1}\bigl(N(t)\bigr). In other words, this function satisfies

So we used earlier that

the upper bound on |X|+|Y| relies on the fact that the hexagon \mathscr{C}_6=\{a,v,c,u,b,w\} has the properties

and soon we are going to need that the inner Andrásfai graph \Gamma=\{v_0,v_1,\dots, v_{3i-2}\} satisfies

Now it suffices to exhibit for every vertex z\in V(\Upsilon_i)\smallsetminus\{v_0, v_{2i-1}, x, y\} a function

such that

1. \mathscr{N}(h^z) agrees on V(\Upsilon^{\mu\nu}_i) with \omega_{\mu\nu}(z)\cdot {\mathds 1}+{\mathds 1}(z)

2. and \Sigma(h^z)=3\omega_{\mu\nu}(z).

Indeed, once we have such a function h^z at our disposal, we can imitate the above reasoning and argue that

which proves the desired inequality |Z|\le \omega_{\mu\nu}(z)(3s-n).

Starting with the vertices in \Gamma we observe that for every k\in {\mathds Z}_{3i-1}\smallsetminus\{0, 2i-1\} we have

By adding (?) we infer

which shows that for k\ne 0, i-1, 2i-1 the function h^{v_k}={\mathds 1}(\mathscr{C}_6)+{\mathds 1}(v_{k-i}, v_k, v_{k+i}) has the required properties. Moreover, for k=i-1 we deduce from (?) with the help of (?) that

meaning that we can take h^{v_{i-1}}={\mathds 1}(\mathscr{C}_6)+{\mathds 1}(v_{i-1}, v_{2i-1}, v_{3i-2}) -\nu g. It remains to deal with the vertices on the hexagon. Starting with a, we observe

which in view of (?) and (?) entails that the function

satisfies

Together with \Sigma(h^a)=6(i-1)+(3i-1)+2-1=3(3i-2) this proves that h^a has all required properties. By symmetry, for \widetilde{h}^b=(i-1)\cdot{\mathds 1}(\mathscr{C}_6) +{\mathds 1}(\Gamma)+{\mathds 1}(b, x)-{\mathds 1}(v_{0}) we obtain \mathscr{N}(\widetilde{h}^b)=(3i-2){\mathds 1}+{\mathds 1}(b) and \Sigma(\widetilde{h}^b)=3(3i-2), so by (?) we may set h^b=\widetilde{h}^b-\nu g. Next,

reveals that the function h^c=(i-1)\cdot{\mathds 1}(\mathscr{C}_6)+{\mathds 1}(\Gamma)+{\mathds 1}(x)-{\mathds 1}(v_0, v_{2i-1}, w) has the property

and because of \Sigma(h^c)=6(i-1)+(3i-1)+1-3=3(3i-3) this establishes the desired bound |C|\le (3i-3)(3s-n). Utilising that similar to (?), (?) we have

one confirms easily that the functions

take care of the three remaining vertex classes.

As we shall prove by means of a complete case analysis, there exist two adjacent non-neighbours z', z'' of z in \Upsilon^{\mu\nu}_i such that each of the sets \{x, y\} and \{v_0, v_{2i-1}\} is either contained in N(z')\cup N(z'')\cup \{z\} or in its complement.

Once we have two such vertices z' and z'', the argument proceeds as follows. Due to z'z''\in E(\Upsilon_i) and zz', zz''\not\in E(\Upsilon_i) the union R=\{z\} \mathbin{\ifx\mathbin\mathop\vphantom{\cup}\fi \mathpalette\mov@rlay{\cup\cr\cdot} } \ifx\mathbin\mathop\expandafter\displaylimits\fi N(z') \mathbin{\ifx\mathbin\mathop\vphantom{\cup}\fi \mathpalette\mov@rlay{\cup\cr\cdot} } \ifx\mathbin\mathop\expandafter\displaylimits\fi N(z'') is disjoint. Let S=V(\Upsilon_i)\smallsetminus R be the complement of R and write

for the sets corresponding to R and S in H. Fact 4.5 implies

and, consequently, we have

i.e., |Z|\ge \omega_{\mu\nu}(z)(3s-n)-\lambda. So it remains to check that the auxiliary vertices z' and z'' do indeed exist and we list some possible choices in the table that follows.

This concludes the proof of Fact 4.6.

Recall that in the proof of Lemma 4.1 we viewed H as a blow-up of the k-regular graph G_i^{\mu\nu}. By Corollary 2.4 at least one of the particles occurring in this construction has the size (k-1)n-(3k-4)s = (3s-n) - \lambda. Let z\in V({\Upsilon}_{i}^{\mu\nu}) be a vertex one of whose \omega_{\mu\nu}(z) particles has this size. Now (?) needs to hold with equality and we have \lceil \lambda/\omega_{\mu\nu}(z)\rceil=\lambda. For this reason at least one of the equations \omega_{\mu\nu}(z)=1 or \lambda=1 is true. In the former case, z\in \{x, y, v_0, v_{2i-1}\} is immediate and we are done.

Now suppose for the sake of contradiction that \lambda=1 and that |Z|\ne 3s-n-\lambda holds for every z\in V({\Upsilon}_{i}^{\mu\nu})\cap\{x, y, v_0, v_{2i-1}\} with \omega_{\mu\nu}(z)=1. Together with the Facts 4.5 and 4.6 this yields |Z|\le \omega_{\mu\nu}(z)(3s-n) for all vertices z\in V({\Upsilon}_{i}^{\mu\nu}).

Concerning the set M=\{z\in V({\Upsilon}_{i}^{\mu\nu})\colon |Z| < \omega_{\mu\nu}(z)(3s-n)\} of vertices for which this estimate fails to be sharp we can deduce from

and (?) that |M|\le 3. Owing to \chi({\Upsilon}_{i}^{\mu\nu})=4 this shows that the neighbourhoods of the vertices in M cannot cover the entire vertex sets of {\Upsilon}_{i}^{\mu\nu} or, in other words, that there is a vertex z_\star\in V({\Upsilon}_{i}^{\mu\nu}) whose neighbourhood is disjoint to M. But now

contradicts (?). This completes the proof of Claim 4.7.

Following the proof of Theorem 1.2, we see that (?) holds with equality, for which reason the set A defined in (?) has to be empty. In other words, G'=G and the proof of Theorem 1.3 discloses that G' is a blow-up of either the Andrásfai graph \Gamma_k, or of a Vega graph \Upsilon^{\mu\nu}_i with k=9i-(6+\mu+\nu). In the former case, Lemma 3.5 shows that G is isomorphic to a graph in \mathscr{G}^n_k(s) and in the latter case we apply Lemma 4.4.