Multi-Colouring of Kneser Graphs: Notes on Stahl’s Conjecture

For any (n',k')-colouring of a graph G, each colour class (the set of vertices whose colour set contains a particular colour) is an independent set, hence contains at most \alpha(G) vertices. (Here \alpha(G) is the independence number of G, the size of a largest independent set in G.) Since each vertex appears in k' colour classes, we have k'\cdot|V(G)|\le n'\cdot\alpha(G), hence \chi_{k'}(G)\ge\dfrac{k'|V(G)|}{\alpha(G)}.

For Kneser graphs, by definition we have |V(K(n,k))|=\dbinom{n}{k}, while the celebrated Erdős-Ko-Rado Theorem [erdos_intersection_1961] means that \alpha(K(n,k))=\dbinom{n-1}{k-1} for all n\ge2k\ge2. Substituting those values in the lower bound for \chi_{k'}(G) above immediately gives \chi_{k'}(K(n,k))\ge \dfrac{k'{\binom{n}{k}}}{\binom{n-1}{k-1}}= \dfrac{k'n}{k}.

If 0\le r<\dfrac{k}{n-2k}, then \Bigl\lfloor\dfrac{r(n-2k)}{k}\Bigr\rfloor=0, and hence Theorem 1.5 immediately shows that \chi_{qk-r}(K(n,k))\ge qn-2r. We then have \chi_{qk-r}(K(n,k))=qn-2r by (?).

If r=\dfrac{k}{n-2k} is an integer and 0<r\le k-1, then \dfrac{k'n}{k}= \dfrac{(qk-r)n}{k}= qn-\dfrac{n}{n-2k}= qn-2r-1 is an integer. But as qk-r is not a multiple of k, since 1\le r\le k-1, by Theorem 2.4 we have that K(n,k) is not (qn-2r-1, qk-r)-colourable. We can conclude that \chi_{qk-r}(K(n,k))= qn-2r by (?).

Fix some k\ge 2, n\ge 2k+1 and 0\le r\le k-1. Take q\ge q_0(n,k)+1, where q_0(n,k) is a function that we will specify later. Fix an (x,qk-r)-colouring C of K(n,k) for some x\le qn-2r (which we know exists by (?)). We will show this means there also exists an (x-n, (q-1)k-r)-colouring of K(n,k).

Let y be the number of non-trivial colour classes in C (the colour classes that are not a subset of \Bigl\{F\in\dbinom{[n]}{k}\Bigm\vert i\in F\Bigr\} for some i\in[n]). Hence there are x-y trivial colour classes. By counting the appearance of each vertex in all colour classes, we have

Since (qk-r)\dbinom{n}{k}=\Bigl(qn-\dfrac{rn}{k}\Bigr)\dbinom{n-1}{k-1}, this gives

We claim that if q\ge q_0(n,k)+1, where

then for all i\in [n] there is a trivial colour class in C that is centred at i. For a contradiction, assume that there is no colour class centred at i^* for some i^*\in [n], i.e. none of the colour classes in C is a subset of \mathcal{F}= \Bigl\{F\in \dbinom{[n]}{k}\Bigm\vert i^*\in F\Bigr\}. Then each trivial colour class contains at most \dbinom{n-2}{k-2} vertices in \mathcal{F}, and each non-trivial colour class contains at most \dbinom{n-1}{k-1} - \dbinom{n-k-1}{k-1} vertices in \mathcal{F}. Counting the appearance of each vertex in \mathcal{F} in all colour classes, we find

Since (qn-r)|\mathcal{F}|=(qn-r)\dbinom{n-1}{k-1}=\dfrac{(qk-r)(n-1)}{k-1}\dbinom{n-2}{k-2} and \dbinom{n-1}{k-1}-\dbinom{n-2}{k-2}=\dbinom{n-2}{k-1}, this gives

Combining (?) and (?), we obtain

We can rearrange this to

Since r\le k-1, the right-hand side of this inequality is smaller than q_0(n,k)+1 as defined in (?), a contradiction.

So we can assume that for all i\in [n] there is a trivial colour class in C that is centred at i. By removing one such trivial colour class centred at i for each i\in [n], we remove n colour classes in total and at most k colours for each vertex. This gives an (x-n,(q-1)k-r)-colouring of K(n,k), thus proving \chi_{(q-1)k-r}(K(n,k))\le \chi_{qk-r}(K(n,k))-n.

Take k\ge2 and n\ge n_0(k)+1, with n_0(k) as in Lemma 3.5 (hence definitely n\ge2k+1), and assume there exists an (x, qk-r)-colouring of K(n,k) for some x\le qn-2r. We will prove there is an (x-q,qk-r)-colouring of K(n-1,k), which shows \chi_{qk-r}(K(n-1,k))\le \chi_{qk-r}(K(n,k))-q.

We first claim there are at least (q-1)n+1 trivial colour classes in the (x,qk-r)-colouring of K(n,k). If this is not the case, then there are at most (q-1)n colour classes that appear on more than \alpha^*(K(n,k)) vertices. Hence counting the total number of appearance of each vertex in all colour classes, we have

Since \dbinom{n}{k}= \dfrac{n}{k}\dbinom{n-1}{k-1}= \dfrac{n}{k}\alpha(K(n,k)), we can rearrange this to

where we use that r\le k-1.

This contradicts Lemma 3.5. Hence there are at least (q-1)n+1 trivial colour classes in the (x, qk-r)-colouring, where each trivial colour class is a subset of \Bigl\{F\in \dbinom{[n]}{k}\Bigm\vert i\in F\Bigr\} for some i\in[n]. Therefore there is some i^*\in [n] such that at least q trivial colour classes are subsets of \Bigl\{F\in \dbinom{[n]}{k}\Bigm\vert i^*\in F\Bigr\}. Removing those q trivial colour classes, we obtain an (x-q,qk-r)-colouring of K(n-1,k), as required.

Fix k\ge2. Since Stahl’s conjecture holds for n=2k, we can assume n\ge2k+1. Let n_0(k) be the integer as in Lemma 3.4. So if n\ge n_0(k)+1, then for all 0\le r\le k-1 we have

For each n\ge2k+1, let q_0(n,k) be the integer as in Lemma 3.3. So if q\geq q_0(n,k)+1, then for all 0\le r\le k-1 we have

Next note that for all n\ge2k and q'\ge 2 we have

since combining a \bigl(\chi_{(q'-1)k-r}(K(n,k)),(q'-1)k-r\bigr)-colouring and a (n,k)-colouring of K(n,k) produces a (q'k-r)-multi-colouring.

Assume the conditions in Theorem 1.8 hold. I.e. for each 2k\le n\le n_0(k) there is a q_n\ge q_0(n,k) such that \chi_{q_nk-(k-1)}(K(n,k))= q_nn-2(k-1). Then we immediately have \chi_{q_nk-r}(K(n,k))= q_nn-2r for all 0\le r\le k-1, since for any non-empty graph G we have \chi_{k'+1}(G)\ge \chi_{k'}(G)+2 (by Lemma 2.2), and \chi_{q_nk}(K(n,k))= q_nn. Combining this with (?), (Proof), and (Proof), gives \chi_{qk-r}(K(n,k))= qn-2r for all 2k\le n\le n_0(k), q\ge1 and 0\le r\le k-1.

In particular we have that \chi_{qk-r}(K(n,k))= qn-2r for n=n_0(k) and all q\ge1 and 0\le r\le k-1. Combining this with (?) and (Proof) gives \chi_{qk-r}(K(n,k))= qn-2r for all n\ge n_0(k), q\ge1 and 0\le r\le k-1, completing the proof.

We can write

We first estimate -\prod\limits_{i=1}^{k-1}\dfrac{n-k-i}{n-i}= -\prod\limits_{i=1}^{k-1}\Bigl(1-\dfrac{k}{n-i}\Bigr)\le -\Bigl(1-\dfrac{k}{n-k+1}\Bigr)^{k-1}.

Next, it is straightforward to check that if k=2, then we have \dfrac{1}{\binom{n-1}{k-1}}=\dfrac{1}{n-1}<\dfrac{1}{n-2}=\dfrac{2k-2}{k(n-2k+2)} for all n\ge3. For k\ge2 and n\ge2k+1, we can estimate by induction on k:

This shows that \dfrac{1}{\binom{n-1}{k-1}}<\dfrac{2k-2}{k(n-2k+2)} for all n\ge2k+1\ge5.

All in all, this means that it suffices to find n_0(k)\ge2k+1 such that for all n\ge n_0(k)+1 we have

This inequality is equivalent to \Bigl(1-\dfrac{k}{n-k+1}\Bigr)^{k-1}\ge 1-\dfrac{1}{k}, hence to

Now we use the standard inequalities 1-\dfrac{1}{x}\le \ln(x)\le x-1 to obtain that it is enough to guarantee (k-1)\cdot\dfrac{-k}{n-2k+1}\ge \dfrac{-1}{k}. This holds for n\ge k^3-k^2+2k-1, completing the proof.

We first estimate, using the definition of q_0(n,k) in (?),

In the second inequality we used that \dbinom{n-2}{k-1}\ge \dbinom{n-k-1}{k-1}\ge \dbinom{2k+1-k-1}{k-1}= k.

We bound \prod\limits_{i=1}^{k-1}\dfrac{n-i}{n-k-i} by estimating its logarithm for n\ge 2k+1:

We will estimate this by setting, for x\ge2k+1:

Then we have f'(x)= \bigl(\ln(x-1)-\ln(x-k-1)\bigr)-\bigl(\ln(x-k)-\ln(x-2k)\bigr). Now set g(y):=\ln y-\ln(y-k), for y\ge k+1. It’s straightforward to check that g'(y)<0 for y\ge k+1. Hence g(y) is strictly decreasing and so for x\ge2k+1 we have f'(x)= g(x-1)-g(x-k)<0. So also f(x) is strictly decreasing; which means that f(x)\le f(2k+1) for all x\ge2k+1. Using this we obtain

We can conclude that \prod\limits_{i=1}^{k-1}\dfrac{n-i}{n-k-i}< \mathrm{e}^{2k\ln{2}}= 4^k, which gives q_0(n,k)< 4^k(n-2k).

Translated into homomorphisms, we are asking for pairs (n,k) and (n',k') such that there is both a homomorphism K(n,k)\rightarrow K(n',k') and a homomorphism K(n',k')\rightarrow K(n,k)

If n=n' and k=k', then clearly there are homomorphisms K(n,k)\rightarrow K(n',k') and K(n',k')\rightarrow K(n,k).
If n=2k and n'=2k', then both K(n,k) and K(n',k') are bipartite. We can easily map any bipartite graph to any graph with at least one edge, by mapping all vertices in one part of the bipartition to one endvertex of an edge, and all vertices in the other part to the other endvertex of that same edge.

For the reverse implication, assume that we have homomorphisms K(n,k)\rightarrow K(n',k') and K(n',k')\rightarrow K(n,k). This means that \chi_{k'}(K(n,k))\le n' and \chi_k(K(n',k'))\le n. On the other hand, Observation 1.4 gives \chi_{k'}(K(n,k))\ge\dfrac{k'n}{k} and \chi_k(K(n',k'))\ge\dfrac{kn'}{k'}. So we have both \dfrac{k'n}{k}\le n' and \dfrac{kn'}{k'}\le n. This leads to the equality \dfrac{n'}{k'}=\dfrac{n}{k} and in particular gives that both \chi_{k'}(K(n,k))=\dfrac{k'n}{k} and \chi_k(K(n',k'))=\dfrac{kn'}{k'}. If conclusion (2) does not hold, we must have \dfrac{n'}{k'}=\dfrac{n}{k}>2. We obtain n'= \dfrac{k'n}{k}>2k' and n=\dfrac{kn'}{k'}>2k. Then Theorem 2.4 gives that k' is a multiple of k and k is a multiple of k'. That means k=k', hence also n=n', and so conclusion (1) holds.