Fractional chromatic number vs. Hall ratio

Set C:=300. Let k=k(n):=\frac{1}{2}\log_3(n) and B_1,\ldots,B_k with |B_i|=\frac{n}{3^i} be defined as in Lemma 3.1. Note that for every sufficiently large n, we then have |B_1|\ge\cdots\ge|B_k|=\sqrt{n}\gg k. Hence, the preconditions of Lemma 2.2 are met, which implies that w.h.p. as n (and hence k=k(n)) tends to infinity, we have \chi_f(G_{\mathcal{B}})\ge \frac{k}{10\log k}=\frac{\log n}{20\log(3)\log(\log_3(n)/2)}\ge \frac{\log n}{20\log(3)\log \log n}\ge \frac{\log n}{C\log \log n}. On the other hand, Lemma 3.1 implies that \rho(G_{\mathcal{B}})\le 300(\log \log n)^2=C(\log \log n)^2 with probability at least 1/3 for all sufficiently large n. Altogether, for n sufficiently large we find that \chi_f(G_{\mathcal{B}})\ge \frac{\log n}{C\log\log n} and \rho(G_{\mathcal{B}})\le C(\log \log n)^2 hold simultaneously with positive probability.

Note further that for every n we have |V(G_{\mathcal{B}})|<n\sum_{i=1}^{\infty}3^{-i}=\frac{n}{2}. By the probabilistic method, it follows that for every sufficiently large n there exists a graph G satisfying \chi_f(G)\ge \frac{\log n}{C\log\log n} and \rho(G)\le C(\log \log n)^2 with at most n vertices. By adding isolated vertices as necessary, we can w.l.o.g. assume that G has in fact exactly n vertices, since the addition of isolated vertices to a graph leaves both the fractional chromatic number and the Hall ratio unchanged. This establishes the existence of graphs with the desired properties, concluding the proof of Theorem 1.3.

Fix \delta>0 arbitrarily and let G=G(\mathcal{B}) be the random graph as in Lemma 4.1. Then by the latter w.h.p. we have that \rho(G)\le 4+\delta and that for every subgraph H of G there exists an independent set in G touching at least \frac{|E(H)|}{4+\delta} edges of H. Since the number of edges of H touched by an independent set I in G equals the sum \sum_{v\in I}\mathrm{deg}_H(v), this is equivalent to saying \alpha_{\text{deg}_H}(G)\ge \frac{|E(H)|}{4+\delta} for every H\subseteq G.

On the other hand, we have

for every sufficiently large n. Thus, the preconditions of Lemma 2.2 are met, and we obtain that with high probability as n\rightarrow\infty (and hence k\rightarrow \infty) we have

Summarizing, w.h.p. as n\rightarrow \infty the graph G has at most

vertices and satisfies \rho(G)\le 4+\delta, \alpha_{\text{deg}_H}(G)\ge \frac{|E(H)|}{4+\delta} for every H\subseteq G as well as \chi_f(G)\ge \frac{\log \log n}{50\log \log \log n}. Since the latter three properties are maintained by adding to G a set of n-|V(G)| isolated vertices, this implies the assertion of Theorem 1.5.

Let \mathcal{B}=(B_i)_{i=1}^{k} be a collection of disjoint sets with |B_1|\ge \dots\ge |B_k| and let G:=G_{\mathcal{B}} denote the associated random graph as in Construction 2.1. Let us denote by w:V(G)\rightarrow \mathbb{R}_+ the positive weight function defined as w(v):=\frac{1}{|B_i|} for every v\in B_i. Then clearly \sum_{v\in V(G)}{w(v)}=k, and hence \chi_f(G)\le \frac{k}{10\log k} implies the existence of an independent set I\subseteq V(G) such that \sum_{v \in I}{w(v)}=\sum_{i=1}^{k}\frac{|I\cap B_i|}{|B_i|}\ge 10\log k. Let J\subseteq I be defined as the union of all the sets I\cap B_i such that \frac{|I\cap B_i|}{|B_i|}\ge \frac{\log k}{k}. We then have \sum_{i=1}^{k}{\frac{|J\cap B_i|}{|B_i|}}\ge 10\log k-\log k=9\log k.

Let \mathcal{T} denote the following set of k-tuples of integers:

For any (t_1,\ldots,t_k) \in \mathcal{T}, let us denote by \mathcal{E}(t_1,\ldots,t_k) the random event that there exists an independent set in G_{\mathcal{B}} containing exactly t_i vertices from B_i for every i\in [k]. Above we argued that \chi_f(G)\le \frac{k}{10\log k} implies that \mathcal{E}(t_1,\ldots,t_k) occurs for some (t_1,\ldots,t_k)\in \mathcal{T}. Thus, our goal will be to bound the probabilities of the individual events \mathcal{E}(t_1,\ldots,t_k) for all (t_1,\ldots,t_k)\in \mathcal{T} and then apply a union bound. So let us consider t=(t_1,\ldots,t_k)\in \mathcal{T} arbitrary, and let P=\mathrm{supp}(t):=\{i\in [k]|t_i>0\}. We then have:

Using |B_1|\ge\dots\ge |B_k| and (t_1,\ldots,t_k)\in \mathcal{T}, we can estimate as follows:

Plugging into the above, we obtain that

using k\ge 3 in the second to last inequality. Summing this over all (t_1,\ldots,t_k)\in \mathcal{T}, clustering the elements of \mathcal{T} according to their support, we obtain:

Since |B_i|\ge k for all i\in [k] by assumption and since the term \log(x)-2\frac{(\log k)^2}{k}x is monotonically decreasing for x\ge k, we obtain the bound

as desired. This concludes the proof of the lemma.

For simplicity, let us denote G:=G_{\mathcal{B}} in the following. The following claim is at the heart of the idea of the proof. Roughly speaking it states that, with decent probability, any sequence of consecutive sets B_i that is not too long does not contain relatively small subgraphs with too many edges. We then later use this sparsity result to give lower bounds on the independence number of subgraphs of G. For convenience, in the following we set B_r:=\emptyset for all integers r\in \mathbb{Z}\setminus [k].

In the rest of the proof, let us refer to the above property/random event as \mathcal{A}.

In the rest of the proof, we will argue that if G satisfies property \mathcal{A}, then \rho(G)\le 300 (\log \log n)^2 for n large enough. Combined with Claim Proof, this will then directly prove that \mathbb{P}[\rho(G)\le 300(\log\log n)^2]\ge \mathbb{P}[\mathcal{A}]>1/3.

So in the following, let us assume that \mathcal{A} holds. We have to show that every non-empty subgraph H of G satisfies \alpha(H)\ge \frac{|V(H)|}{300 (\log \log n)^2}. So let H\subseteq G be given arbitrarily and let s:=|V(H)|>0. Then 0<s\le |V(G)|<n\sum_{i=1}^{\infty}3^{-i}=\frac{n}{2}<2|B_1|. Hence, there exists a (unique) index i\in [k] such that 2|B_{i+1}|<s\le 2|B_i|. Let X:=V(H)\setminus \bigcup_{\ell=i+1}^{k}B_\ell, and note that

For each integer 0\le j<\log_2(i), let us define a subset X_j of X as

and define X_{-1}:=X\cap B_i. Then the sets X_{-1},\ldots,X_{\lceil\log_2(i)\rceil-1} partition X.

We next claim that for every j=-1,0,\ldots,\lceil\log_2(i)\rceil-1, we have \alpha(H[X_j])\ge \frac{|X_j|^2}{7s}. This is easy to see when j=-1, since then by definition X_{-1}\subseteq B_i is an independent set in G, and hence we have \alpha(H[X_{-1}])=|X_{-1}|\ge \frac{|X_{-1}|^2}{7s}. Now, consider any integer 0\le j<\log_2(i). By definition of s and X_j we have 2|B_{i+1}|<s\le 2|B_i|, |X_j|\le s and X_j\subseteq \bigcup_{\ell=i-2^{j+1}+1}^{i-2^j}B_\ell. Hence, the induced subgraph H[X_j] meets all preconditions of the statement \mathcal{A}. In particular, our assumption that \mathcal{A} holds now implies that we cannot have |E(H[X_j])|\ge 3s, in other words, we must have |E(H[X_j])|< 3s.

Using the well-known bound \alpha(F)\ge \frac{|V(F)|^2}{2|E(F)|+|V(F)|} which holds for every graph F, we now obtain that

as desired. Since X_{-1}, \ldots,X_{\lceil\log_2(i)\rceil-1} partition X, there exists some j such that |X_j|\ge \frac{|X|}{\lceil \log_2(i)\rceil+1}\ge \frac{s}{4(\lceil\log_2(i)\rceil+1)}. We now obtain

All in all, this shows that \mathcal{A} implies

for n large enough. As explained before, this concludes the proof of the lemma.

In the rest of the proof, let us denote G:=G_{\mathcal{B}}. The statements claimed in the lemma will be easy to deduce once we have established that G w.h.p. satisfies two sparsity properties that are isolated in the following claim. As in the proof of Lemma 3.1, for convenience let us set B_r:=\emptyset for all r\in\mathbb{Z}\setminus [k].

To conclude the proof of the lemma, we will now show that for every large enough n, if the statements (1) and (2) of Claim Proof hold, then we also must have that \rho(G)\le 4+\delta (i.e. \alpha(H)\ge \frac{|V(H)|}{4+\delta} for every H\subseteq G) and that for every subgraph H of G there exists an independent set in G touching at least \frac{|E(H)|}{4+\delta} edges of H. Note that w.l.o.g. it suffices to prove these two statements for all subgraphs H of G without isolated vertices(This is because we can always add all the isolated vertices of H to a suitable independent set in the subgraph of H induced by the vertices of degree at least one.). Once this is achieved, the statements of the lemma will be directly implied by Claim Proof.

So suppose in the remainder of the argument that statements (1) and (2) of Claim Proof hold, and consider any non-empty subgraph H of G without isolated vertices. We have to show that \alpha(H)\ge \frac{|V(H)|}{4+\delta} and that G contains an independent set touching at least \frac{|E(H)|}{4+\delta} edges of H. To do so, we start by observing

Hence, there exists a (unique) index i\in [k] such that k^5|B_{i+1}|<|V(H)|\le k^5|B_i|. Let X:=V(H)\setminus\bigcup_{\ell=i}^{k} B_\ell. Then X\subseteq \bigcup_{\ell=1}^{i-1}B_\ell, and since G[X] and all of its subgraphs have at most k^5|B_i| vertices, it follows from item (1) of Claim Proof that every subgraph of G[X] has average degree strictly less than 3. In particular, every subgraph of G[X] contains a vertex of degree at most 2, which means that G[X] is a 2-degenerate graph. This implies that \chi(G[X])\le 3. Let X':=V(H)\setminus \bigcup_{\ell=i+1}^{k} B_\ell. Since B_i is an independent set in G and X'\subseteq X \cup B_i, we conclude \chi(G[X'])\le 4. Consider a fixed proper 4-coloring of G[X']. Let us denote by I a color class of this coloring of maximum size and by J a color class touching the most edges of H. Then clearly |I|\ge \frac{|X'|}{4}. Further, denoting F:=E\left(G\left[\bigcup_{\ell=i+1}^{k}B_\ell\right]\right), the definition of X' implies that every edge in E(H)\setminus F has at least one endpoint in X'. Hence, J must touch at least \frac{1}{4}|E(H)\setminus F| edges of H. To conclude, it remains to compare the quantities |X'| and |V(H)| as well as |E(H)\setminus F| and |E(H)| to each other:

By definition of X' we have

and since k^5|B_{i+1}|<|V(H)| (by choice of i) and k=k(n)\rightarrow \infty for n\rightarrow \infty, this implies that |X'|\ge \frac{|V(H)|}{1+\delta/4} for n sufficiently large. Similarly, by item (2) of Claim Proof we have

where we used that H has no isolated vertices by assumption. Hence, for n sufficiently large we also must have |E(H)\setminus F|\ge \frac{|E(H)|}{1+\delta/4}. All in all, this implies that \alpha(H)\ge |I|\ge \frac{|X'|}{4}\ge \frac{|V(H)|}{4+\delta} and that J is an independent set in G touching at least \frac{|E(H)\setminus F|}{4}\ge \frac{|E(H)|}{4+\delta} edges of H. This is what had to be shown. As explained previously, this concludes the proof of the lemma.