This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

On connected components with many edges

Sammy Luo
Abstract.

We prove that if HH is a subgraph of a complete multipartite graph GG, then HH contains a connected component HH^{\prime} satisfying |E(H)||E(G)||E(H)|2|E(H^{\prime})||E(G)|\geq|E(H)|^{2}. We use this to prove that every three-coloring of the edges of a complete graph contains a monochromatic connected subgraph with at least 1/61/6 of the edges. We further show that such a coloring has a monochromatic circuit with a fraction 1/6o(1)1/6-o(1) of the edges. This verifies a conjecture of Conlon and Tyomkyn. Moreover, for general kk, we show that every kk-coloring of the edges of KnK_{n} contains a monochromatic connected subgraph with at least 1k2k+54(n2)\frac{1}{k^{2}-k+\frac{5}{4}}\binom{n}{2} edges.

Department of Mathematics, Stanford University, Stanford, CA 94305, USA.
Supported by NSF GRFP Grant DGE-1656518. Email: [email protected].

1. Introduction

A classical observation of Erdős and Rado is that in any two-coloring of the edges of the complete graph KnK_{n}, one of the color classes forms a connected graph. In [3], Gyárfás proves the following generalization of this observation: For any k2k\geq 2, in every kk-coloring of the edges of KnK_{n}, there is a monochromatic connected component with at least n/(k1)n/(k-1) vertices. This observation has since been extended in numerous ways, such as by replacing KnK_{n} with a graph of high minimum degree [5] or with a nearly-complete bipartite graph [2], or adding a constraint on the diameter of the large monochromatic component [7]. See [4] for a survey of earlier work on the subject.

The arguments used in this subject tend to focus on sparse spanning structures like double stars. As such, there is a surprising lack of progress on the corresponding question about edges in a monochromatic connected component. That is, what is the largest value of M=M(n,k)M=M(n,k) such that every kk-edge-coloring of KnK_{n} has a monochromatic connected component with at least MM edges?

This question was raised by Conlon and Tyomkyn in [1], in the context of determining the multicolor Ramsey numbers for trails (see Section 3.2 for the relevant definitions). After showing that M(n,2)=29n2+o(n2)M(n,2)=\frac{2}{9}n^{2}+o(n^{2}), they sketch a simple argument that shows M(n,k)116k2n2+O(n)M(n,k)\geq\frac{1}{16k^{2}}n^{2}+O(n) for all kk; with a slightly more careful analysis, their argument in fact yields M(n,k)14k2n2+O(n)M(n,k)\geq\frac{1}{4k^{2}}n^{2}+O(n). In the other direction, they examine a construction of Gyárfás in [4] to show that M(n,k)12k(k1)n2+O(n)M(n,k)\leq\frac{1}{2k(k-1)}n^{2}+O(n) for infinitely many values of kk (specifically, when k1k-1 is a prime power), conjecturing that this upper bound is tight in the case k=3k=3.

In this note, we improve the general lower bound on M(n,k)M(n,k), as well as a corresponding lower bound for the trail Ramsey problem, bringing it to asymptotically within a factor 1O(1k2)1-O\left(\frac{1}{k^{2}}\right) of the upper bound for infinitely many values of kk.

Theorem 1.

For any k2k\geq 2, in every kk-coloring of the edges of KnK_{n}, there is a monochromatic connected component with at least M(n,k)1k2k+54(n2)M(n,k)\geq\frac{1}{k^{2}-k+\frac{5}{4}}\binom{n}{2} edges.

By building on the overarching ideas in [4] and introducing some key new insights, we manage to strengthen this lower bound in the case k=3k=3 to prove the tight lower bound conjectured by Conlon and Tyomkyn.

Theorem 2.

In every 33-coloring of the edges of KnK_{n}, there is a monochromatic component containing at least a sixth of the edges. That is, M(n,3)16(n2)M(n,3)\geq\lceil\frac{1}{6}\binom{n}{2}\rceil. Moreover, for nn sufficiently large, this bound is sharp.

While equality holds in this bound for sufficiently large nn, there are, as we will see from the proof, small values of nn for which M(n,3)>16(n2)M(n,3)>\lceil\frac{1}{6}\binom{n}{2}\rceil. In particular, equality holds for all n18n\geq 18, but not for n=17n=17.

The key result we use is a new inequality that may be of independent interest. Given a subgraph HH of a complete multipartite graph GG, it relates the edge counts of GG and HH to the largest edge count of a connected component HH^{\prime} of HH.

Theorem 3.

Let GG be a complete rr-partite graph for some r2r\geq 2, and let HH be a subgraph of GG. Then HH contains a connected component HH^{\prime} satisfying

|E(H)||E(H)|2|E(G)|.|E(H^{\prime})|\geq\frac{|E(H)|^{2}}{|E(G)|}.

In effect, this result settles the density analogue of the coloring question of determining M(n,k)M(n,k). Instead of partitioning the edges of a graph GG into color classes, we are fixing a subgraph HH of GG with a given fraction δ=|E(H)||E(G)|\delta=\frac{|E(H)|}{|E(G)|} of its edges, and asking about the component of HH with the most edges. Theorem 3 can then be restated as: if |E(H)|=δ|E(G)||E(H)|=\delta|E(G)|, then HH contains a connected component HH^{\prime} with |E(H)|δ2|E(G)||E(H^{\prime})|\geq\delta^{2}|E(G)|. Equality can be attained asymptotically when δ=1k\delta=\frac{1}{k} for any positive integer k2k\geq 2: Let V(G)=V1VrV(G)=V_{1}\cup\cdots\cup V_{r} be an rr-partition of V(G)V(G), split each ViV_{i} into kk (roughly) equally sized vertex sets {Vi,j}j=1k\{V_{i,j}\}_{j=1}^{k}, and for 1jk1\leq j\leq k, let HjH_{j} be the subgraph of GG induced on i=1rVi,j\bigcup_{i=1}^{r}V_{i,j}. Then indeed, H=j=1kHjH=\bigcup_{j=1}^{k}H_{j} is a subgraph of GG whose components have edge counts

|E(Hj)|=1k2|E(G)|+O(n)=|E(H)|2|E(G)|+O(n).|E(H_{j})|=\frac{1}{k^{2}}|E(G)|+O(n)=\frac{|E(H)|^{2}}{|E(G)|}+O(n).

The simplest case of Theorem 3, when G=KnG=K_{n}, already immediately implies an improvement over previously known lower bounds on M(n,k)M(n,k).

Corollary 4.

In every kk-coloring of the edges of KnK_{n}, there is a monochromatic connected component with at least 1k2(n2)\frac{1}{k^{2}}\binom{n}{2} edges.

Proof.

In a kk-coloring of the edges of KnK_{n}, one of the color classes has at least 1k(n2)\frac{1}{k}\binom{n}{2} edges. Taking HH to be this color class in Theorem 3 with G=KnG=K_{n} yields Corollary 4. ∎

The proof of Theorem 2 requires a more detailed argument, but similarly follows from this one case of Theorem 3, while the proof of Theorem 1 requires the use of Theorem 3 in full generality.

In the next section, we give an elementary proof of Theorem 3. We then study the coloring version of the problem, as well as the corresponding trail Ramsey problem, in Section 3.

2. Proof of Theorem 3

In the case G=KnG=K_{n}, the proof of Theorem 3 is very simple, but nevertheless includes a key insight: Instead of taking a component that maximizes the number of edges right away, we consider a component HH^{\prime} with maximum average degree d¯(H)=|E(H)|12|V(H)|\bar{d}(H^{\prime})=\frac{|E(H^{\prime})|}{\frac{1}{2}|V(H^{\prime})|}. Two observations are crucial here: First, by the so-called generalized mediant inequality, this highest average degree must be at least the average degree of the whole graph HH, since

|E(H)|12|V(H)|=|E(Hi)|12|V(Hi)|,\frac{|E(H)|}{\frac{1}{2}|V(H)|}=\frac{\sum|E(H_{i})|}{\frac{1}{2}\sum|V(H_{i})|},

where the sums are over connected components HiH_{i} of HH, and the right hand side is a generalized mediant of the average degrees of the individual components. Second, the number of vertices of HH^{\prime} is at least one more than its maximum degree, so |V(H)|d¯(H)+1|V(H^{\prime})|\geq\bar{d}(H^{\prime})+1. Letting δ=|E(H)||E(G)|\delta=\frac{|E(H)|}{|E(G)|}, we then obtain the bound

|E(H)|\displaystyle|E(H^{\prime})| 12(d¯(H))(d¯(H)+1)=(d¯(H)+12)\displaystyle\geq\frac{1}{2}(\bar{d}(H^{\prime}))(\bar{d}(H^{\prime})+1)=\binom{\bar{d}(H^{\prime})+1}{2}
(d¯(H)+12)=(2|E(H)|n+12)\displaystyle\geq\binom{\bar{d}(H)+1}{2}=\binom{\frac{2|E(H)|}{n}+1}{2}
=(δ(n1)+12)δ2(n2)\displaystyle=\binom{\delta(n-1)+1}{2}\geq\delta^{2}\binom{n}{2}
=δ2|E(G)|=|E(H)|2|E(G)|,\displaystyle=\delta^{2}|E(G)|=\frac{|E(H)|^{2}}{|E(G)|},

as desired.

The general case will use both of these observations in a modified setting. Instead of the average degree, we will work with a slightly different quantity whose denominator is a weighted vertex count; we will then, perhaps counterintuitively, lower bound this weighted vertex count by the modified analogue of the average degree, in order to obtain the bound we seek. The core of our proof is the following general inequality.

Lemma 5.

If a1,,ara_{1},\dots,a_{r} and b1,,brb_{1},\dots,b_{r} are nonnegative real numbers, then

((i=1rai)(i=1rbi)i=1raibi)2((i=1rai)2i=1rai2)((i=1rbi)2i=1rbi2).\left(\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sum_{i=1}^{r}a_{i}b_{i}\right)^{2}\geq\left(\left(\sum_{i=1}^{r}a_{i}\right)^{2}-\sum_{i=1}^{r}a_{i}^{2}\right)\left(\left(\sum_{i=1}^{r}b_{i}\right)^{2}-\sum_{i=1}^{r}b_{i}^{2}\right).
Proof.

The Cauchy-Schwarz inequality yields

(i=1rai)(i=1rbi)i=1raibi(i=1rai)(i=1rbi)(i=1rai2)(i=1rbi2).\displaystyle\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sum_{i=1}^{r}a_{i}b_{i}\geq\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sqrt{\left(\sum_{i=1}^{r}a_{i}^{2}\right)\left(\sum_{i=1}^{r}b_{i}^{2}\right)}.

This last quantity is clearly nonnegative, since (i=1rai)2i=1rai2\left(\sum_{i=1}^{r}a_{i}\right)^{2}\geq\sum_{i=1}^{r}a_{i}^{2} and (i=1rbi)2i=1rbi2\left(\sum_{i=1}^{r}b_{i}\right)^{2}\geq\sum_{i=1}^{r}b_{i}^{2}. So,

((i=1rai)(i=1rbi)i=1raibi)2((i=1rai)(i=1rbi)(i=1rai2)(i=1rbi2))2\displaystyle\left(\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sum_{i=1}^{r}a_{i}b_{i}\right)^{2}\geq\left(\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sqrt{\left(\sum_{i=1}^{r}a_{i}^{2}\right)\left(\sum_{i=1}^{r}b_{i}^{2}\right)}\right)^{2}
=(i=1rai)2(i=1rbi)2+(i=1rai2)(i=1rbi2)2(i=1rai)(i=1rbi)(i=1rai2)(i=1rbi2)\displaystyle=\left(\sum_{i=1}^{r}a_{i}\right)^{2}\left(\sum_{i=1}^{r}b_{i}\right)^{2}+\left(\sum_{i=1}^{r}a_{i}^{2}\right)\left(\sum_{i=1}^{r}b_{i}^{2}\right)-2\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)\sqrt{\left(\sum_{i=1}^{r}a_{i}^{2}\right)\left(\sum_{i=1}^{r}b_{i}^{2}\right)}
(i=1rai)2(i=1rbi)2+(i=1rai2)(i=1rbi2)(i=1rai)2(i=1rbi2)(i=1rbi)2(i=1rai2)\displaystyle\geq\left(\sum_{i=1}^{r}a_{i}\right)^{2}\left(\sum_{i=1}^{r}b_{i}\right)^{2}+\left(\sum_{i=1}^{r}a_{i}^{2}\right)\left(\sum_{i=1}^{r}b_{i}^{2}\right)-\left(\sum_{i=1}^{r}a_{i}\right)^{2}\left(\sum_{i=1}^{r}b_{i}^{2}\right)-\left(\sum_{i=1}^{r}b_{i}\right)^{2}\left(\sum_{i=1}^{r}a_{i}^{2}\right)
=((i=1rai)2i=1rai2)((i=1rbi)2i=1rbi2),\displaystyle=\left(\left(\sum_{i=1}^{r}a_{i}\right)^{2}-\sum_{i=1}^{r}a_{i}^{2}\right)\left(\left(\sum_{i=1}^{r}b_{i}\right)^{2}-\sum_{i=1}^{r}b_{i}^{2}\right),

where the last inequality is an application of the AM-GM inequality. ∎

Given a graph GG and vertex sets S,TV(G)S,T\subseteq V(G), let

eG(S,T)=#{(s,t)S×T:sGt}.e_{G}(S,T)=\#\{(s,t)\in S\times T:\>s\sim_{G}t\}.

We write e(S,T)e(S,T) for eG(S,T)e_{G}(S,T) when the graph in question is unambiguous. Let G[S]G[S] denote the induced subgraph of GG on the vertex set SS. Lemma 5 immediately implies the following.

Corollary 6.

Let GG be a complete multipartite graph. For any S,TV(G)S,T\subseteq V(G), we have

e(S,T)24|E(G[S])||E(G[T])|.e(S,T)^{2}\geq 4|E(G[S])||E(G[T])|.
Proof.

Suppose that GG is rr-partite. Let V(G)=V1VrV(G)=V_{1}\cup\cdots\cup V_{r} be a partition of the vertices of GG into rr independent sets. Let ai=|ViS|a_{i}=|V_{i}\cap S| and bi=|ViT|b_{i}=|V_{i}\cap T|, so

e(S,T)=1i,jrijaibj=(i=1rai)(i=1rbi)i=1raibi,e(S,T)=\sum_{\begin{subarray}{c}1\leq i,j\leq r\\ i\neq j\end{subarray}}a_{i}b_{j}=\left(\sum_{i=1}^{r}a_{i}\right)\left(\sum_{i=1}^{r}b_{i}\right)-\sum_{i=1}^{r}a_{i}b_{i},

while

|E(G[S])|=12((i=1rai)2i=1rai2),|E(G[T])|=12((i=1rbi)2i=1rbi2).|E(G[S])|=\frac{1}{2}\left(\left(\sum_{i=1}^{r}a_{i}\right)^{2}-\sum_{i=1}^{r}a_{i}^{2}\right),\qquad|E(G[T])|=\frac{1}{2}\left(\left(\sum_{i=1}^{r}b_{i}\right)^{2}-\sum_{i=1}^{r}b_{i}^{2}\right).

Then Lemma 5 indeed yields e(S,T)24|E(G[S])||E(G[T])|e(S,T)^{2}\geq 4|E(G[S])||E(G[T])|, as desired. ∎

Proof of Theorem 3.

Let V=V(G)=V1VrV=V(G)=V_{1}\cup\cdots\cup V_{r} be a partition of the vertices of GG into rr independent sets, let H1,,HkH_{1},\dots,H_{k} be the connected components of HH, and let Vi,=ViHV_{i,\ell}=V_{i}\cap H_{\ell}. We have

|E(H)|=j=1k|E(H)|,|Vi|==1k|Vi,| for all i.|E(H)|=\sum_{j=1}^{k}|E(H_{\ell})|,\qquad|V_{i}|=\sum_{\ell=1}^{k}|V_{i,\ell}|\>\text{ for all }i.

For any subset SVS\subseteq V, define f(S)=12eG(S,V)f(S)=\frac{1}{2}e_{G}(S,V). Note that

f(S)=121i,jrij|ViS||Vj|,f(S)=\frac{1}{2}\sum_{\begin{subarray}{c}1\leq i,j\leq r\\ i\neq j\end{subarray}}|V_{i}\cap S||V_{j}|,

so f(S)f(S) can be viewed as a weighted vertex count for SS, with the property that

=1kf(V(H))=12=1keG(V(H),V)=12eG(V,V)=|E(G)|.\sum_{\ell=1}^{k}f(V(H_{\ell}))=\frac{1}{2}\sum_{\ell=1}^{k}e_{G}(V(H_{\ell}),V)=\frac{1}{2}e_{G}(V,V)=|E(G)|.

Then by the generalized mediant inequality, for some H=HH^{\prime}=H_{\ell} we have

|E(H)|f(V(H))=1k|E(H)|=1kf(V(H))=|E(H)||E(G)|.\frac{|E(H^{\prime})|}{f(V(H^{\prime}))}\geq\frac{\sum_{\ell=1}^{k}|E(H_{\ell})|}{\sum_{\ell=1}^{k}f(V(H_{\ell}))}=\frac{|E(H)|}{|E(G)|}.

Now, Corollary 6 applied with S=V(H)S=V(H^{\prime}), T=VT=V yields f(V(H))2|E(G[V(H)])||E(G)||E(H)||E(G)|f(V(H^{\prime}))^{2}\geq|E(G[V(H^{\prime})])||E(G)|\geq|E(H^{\prime})||E(G)|, so that

|E(H)||E(G)||E(H)|f(V(H))|E(H)||E(G)|,\frac{|E(H)|}{|E(G)|}\leq\frac{|E(H^{\prime})|}{f(V(H^{\prime}))}\leq\sqrt{\frac{|E(H^{\prime})|}{|E(G)|}},

which rearranges to the desired inequality. ∎

3. Coloring problems

We can now apply Theorem 3 to the corresponding coloring problems. We have already seen that applying Theorem 3 with G=KnG=K_{n} readily yields the simple lower bound on M(n,k)M(n,k) given in Corollary 4. We now discuss how to improve this lower bound, before turning to a closely related problem on monochromatic trails and circuits.

3.1. Lower bound on M(n,k)M(n,k) for general kk

Our strategy for improving the lower bound on M(n,k)M(n,k) is as follows: First, assuming that no monochromatic component has too many edges, we show that in the color with the highest density (say, red), we can upper bound the number of vertices covered by any set of components (or else we can finish with an average degree argument on the rest of the red components). Through a smoothing argument, this yields an upper bound on the sum of squares of the number of vertices in each red component, and thus a lower bound on the number of edges in the complete multipartite graph GG formed by deleting from KnK_{n} all edges within the vertex sets of the red components. We finish by applying Theorem 3 to GG to find a non-red component with many edges.

Proof of Theorem 1.

Fix a kk-coloring χ\chi of the edges of KnK_{n}, and suppose that the largest number of edges in a monochromatic connected component in this coloring is z(n2)z\binom{n}{2}. Without loss of generality, let red be the color with the most edges, so there are at least 1k(n2)\frac{1}{k}\binom{n}{2} red edges in our coloring. Let 𝒞1={R1,R2,,Rm}\mathcal{C}_{1}=\{R_{1},R_{2},\dots,R_{m}\} be the set of red components, with

|V(R1)||V(R2)||V(Rm)|,i=1m|V(Ri)|=n.|V(R_{1})|\geq|V(R_{2})|\geq\cdots\geq|V(R_{m})|,\qquad\sum_{i=1}^{m}|V(R_{i})|=n. (3.1)

Let x=1(n2)i=1m|E(Ri)|x=\frac{1}{\binom{n}{2}}\sum_{i=1}^{m}|E(R_{i})|, so x1kx\geq\frac{1}{k}. By assumption, |E(Ri)|z(n2)|E(R_{i})|\leq z\binom{n}{2} for 1im1\leq i\leq m. As in the proof of Corollary 4, applying Theorem 3 with KnK_{n} as GG and its red color class (i.e. the spanning subgraph formed by its red edges) as HH yields a red component with at least x2(n2)x^{2}\binom{n}{2} edges, so by assumption we have zx21k2z\geq x^{2}\geq\frac{1}{k^{2}}. Define δ=k1z0\delta=k-\frac{1}{\sqrt{z}}\geq 0, so z=1(kδ)2z=\frac{1}{(k-\delta)^{2}}.

For any j[1,m1]j\in[1,m-1], let GjG_{j} be the complete graph on i=j+1mV(Ri)\bigcup_{i=j+1}^{m}V(R_{i}), with its coloring induced by χ\chi. Applying Theorem 3 with GjG_{j} as GG and the red color class of GjG_{j} as HH yields a red connected component H=RiH^{\prime}=R_{i} with |E(H)||E(H)|2|E(Gj)||E(H^{\prime})|\geq\frac{|E(H)|^{2}}{|E(G_{j})|}. Since

|E(H)|=i=j+1m|E(Ri)|=x(n2)i=1j|E(Ri)|x(n2)jz(n2),|E(H)|=\sum_{i=j+1}^{m}|E(R_{i})|=x\binom{n}{2}-\sum_{i=1}^{j}|E(R_{i})|\geq x\binom{n}{2}-jz\binom{n}{2},

while |E(Gj)|=(|V(Gj)|2)=(ni=1j|V(Ri)|2)|E(G_{j})|=\binom{|V(G_{j})|}{2}=\binom{n-\sum_{i=1}^{j}|V(R_{i})|}{2}, we have

z(n2)|E(H)|max(xjz,0)2(n2)2(ni=1j|V(Ri)|2).z\binom{n}{2}\geq|E(H^{\prime})|\geq\frac{\max(x-jz,0)^{2}\binom{n}{2}^{2}}{\binom{n-\sum_{i=1}^{j}|V(R_{i})|}{2}}.

Since max(xjz,0)zxz1\frac{\max(x-jz,0)}{\sqrt{z}}\leq\frac{x}{\sqrt{z}}\leq 1, this implies

(ni=1j|V(Ri)|2)max(xjz,0)2z(n2)(max(xjz,0)zn2).\binom{n-\sum_{i=1}^{j}|V(R_{i})|}{2}\geq\frac{\max(x-jz,0)^{2}}{z}\binom{n}{2}\geq\binom{\frac{\max(x-jz,0)}{\sqrt{z}}n}{2}.

Since ni=1j|V(Ri)|1n-\sum_{i=1}^{j}|V(R_{i})|\geq 1, and the function f(X)=(X2)f(X)=\binom{X}{2} is increasing for X1X\geq 1, this then implies

i=1j|V(Ri)|(1x/z+jz)nfor all j[1,m1].\sum_{i=1}^{j}|V(R_{i})|\leq(1-x/\sqrt{z}+j\sqrt{z})n\qquad\text{for all }j\in[1,m-1]. (3.2)

We can now give an upper bound on i=1m(|V(Ri)|2)\sum_{i=1}^{m}\binom{|V(R_{i})|}{2} by solving the corresponding convex optimization problem. The proof of the following technical lemma will be deferred until the end of the section.

Lemma 7.

Let x,z>0x,z>0. Subject to the constraints v1vm0v_{1}\geq\cdots\geq v_{m}\geq 0, i=1mvi=1\sum_{i=1}^{m}v_{i}=1, and

i=1jvi1x/z+jzfor all j[1,m1],\sum_{i=1}^{j}v_{i}\leq 1-x/\sqrt{z}+j\sqrt{z}\qquad\text{for all }j\in[1,m-1],

the quantity i=1mvi2\sum_{i=1}^{m}v_{i}^{2} is maximized when v1=1x/z+zv_{1}=1-x/\sqrt{z}+\sqrt{z}, vi=zv_{i}=\sqrt{z} for 2ixz2\leq i\leq\lfloor\frac{x}{z}\rfloor, vxz+1=x/zxzzv_{\lfloor\frac{x}{z}\rfloor+1}=x/\sqrt{z}-\lfloor\frac{x}{z}\rfloor\sqrt{z}, and vi=0v_{i}=0 for i>xz+1i>\lfloor\frac{x}{z}\rfloor+1.

Let vi=|V(Ri)|nv_{i}=\frac{|V(R_{i})|}{n}. Since (3.1) and (3.2) hold, we can apply Lemma 7 to obtain

i=1mvi2(1x/z+z)2+(xz1)z+((x/zxz)z)2(1x/z+z)2+(x/z1)z,\sum_{i=1}^{m}v_{i}^{2}\leq(1-x/\sqrt{z}+\sqrt{z})^{2}+\left(\left\lfloor\frac{x}{z}\right\rfloor-1\right)z+\left(\left(x/z-\left\lfloor\frac{x}{z}\right\rfloor\right)\sqrt{z}\right)^{2}\leq(1-x/\sqrt{z}+\sqrt{z})^{2}+(x/z-1)z,

so that we have

i=1m(|V(Ri)|2)\displaystyle\sum_{i=1}^{m}\binom{|V(R_{i})|}{2} =n2i=1mvi2n2n2((1x/z+z)2+(x/z1)z)n2.\displaystyle=\frac{n^{2}\sum_{i=1}^{m}v_{i}^{2}-n}{2}\leq\frac{n^{2}((1-x/\sqrt{z}+\sqrt{z})^{2}+(x/z-1)z)-n}{2}.

Finally, let GG be the complete mm-partite graph obtained from KnK_{n} by removing all edges within each V(Ri)V(R_{i}), with its coloring induced by χ\chi. There are no red edges in GG, so by the pigeonhole principle, one of the k1k-1 remaining colors has at least 1k1|E(G)|\frac{1}{k-1}|E(G)| edges in GG. Let HH be the spanning subgraph of GG induced by the edges in that color. Applying Theorem 3 then yields a monochromatic connected component HH^{\prime} with at least 1(k1)2|E(G)|\frac{1}{(k-1)^{2}}|E(G)| edges. Then by assumption we have

z\displaystyle z 1(k1)2|E(G)|(n2)=1(k1)2(1i=1m(|V(Ri)|2)(n2))\displaystyle\geq\frac{1}{(k-1)^{2}}\frac{|E(G)|}{\binom{n}{2}}=\frac{1}{(k-1)^{2}}\left(1-\frac{\sum_{i=1}^{m}\binom{|V(R_{i})|}{2}}{\binom{n}{2}}\right)
1(k1)2(1n2((1x/z+z)2+(x/z1)z)nn2n)\displaystyle\geq\frac{1}{(k-1)^{2}}\left(1-\frac{n^{2}((1-x/\sqrt{z}+\sqrt{z})^{2}+(x/z-1)z)-n}{n^{2}-n}\right)
1(k1)2(1(1x/z+z)2(x/z1)z),\displaystyle\geq\frac{1}{(k-1)^{2}}\left(1-(1-x/\sqrt{z}+\sqrt{z})^{2}-(x/z-1)z\right),

which rearranges to give

1(k1)2z(1x/z+z)2+(x/z1)z=1zx2(1+2/z)x+(1+2z).1-(k-1)^{2}z\leq(1-x/\sqrt{z}+\sqrt{z})^{2}+(x/z-1)z=\frac{1}{z}x^{2}-(1+2/\sqrt{z})x+(1+2\sqrt{z}).

The right hand side is a quadratic in xx that is decreasing for xz2+zx\leq\frac{z}{2}+\sqrt{z}. Since by assumption zx1k\sqrt{z}\geq x\geq\frac{1}{k}, we then have

1(k1)2z1zk2(1+2/z)1k+(1+2z).1-(k-1)^{2}z\leq\frac{1}{zk^{2}}-(1+2/\sqrt{z})\frac{1}{k}+(1+2\sqrt{z}).

Substituting in z=1(kδ)2z=\frac{1}{(k-\delta)^{2}} yields

1(k1)2(kδ)2(kδ)2k21+2(kδ)k+1+2kδ,1-\frac{(k-1)^{2}}{(k-\delta)^{2}}\leq\frac{(k-\delta)^{2}}{k^{2}}-\frac{1+2(k-\delta)}{k}+1+\frac{2}{k-\delta},

which upon rearrangement becomes

0\displaystyle 0 (k1)2k2+(kδ)4k(kδ)22(kδ)3k+2(kδ)k2\displaystyle\leq(k-1)^{2}k^{2}+(k-\delta)^{4}-k(k-\delta)^{2}-2(k-\delta)^{3}k+2(k-\delta)k^{2}
=k3+k2kδ2+2k3δ2kδ3+δ4\displaystyle=-k^{3}+k^{2}-k\delta^{2}+2k^{3}\delta-2k\delta^{3}+\delta^{4}
=(kδ2)2k(k2δ2)(12δ).\displaystyle=(k-\delta^{2})^{2}-k(k^{2}-\delta^{2})(1-2\delta).

This implies

12δ(kδ2)2k(k2δ2)<1k,1-2\delta\leq\frac{(k-\delta^{2})^{2}}{k(k^{2}-\delta^{2})}<\frac{1}{k},

so that δ>k12k\delta>\frac{k-1}{2k}. Thus, the coloring χ\chi contains a monochromatic connected component with at least z(n2)z\binom{n}{2} edges, where

z=1(kδ)2>1(k12+12k)21k2k+14+114k+14k21k2k+54,z=\frac{1}{(k-\delta)^{2}}>\frac{1}{(k-\frac{1}{2}+\frac{1}{2k})^{2}}\geq\frac{1}{k^{2}-k+\frac{1}{4}+1-\frac{1}{4k}+\frac{1}{4k^{2}}}\geq\frac{1}{k^{2}-k+\frac{5}{4}},

as desired. ∎

Proof of Lemma 7.

Since the feasible region is compact, there exists a choice of the viv_{i} such that the desired maximum is attained. Fix such a maximizing choice of the viv_{i}. For j[m]j\in[m], let AjA_{j} denote the given constraint on i=1jvi\sum_{i=1}^{j}v_{i}, and let SS be the set of jj for which equality holds in AjA_{j}. Let m=xzm^{\prime}=\lfloor\frac{x}{z}\rfloor. Since i=1mvi=1\sum_{i=1}^{m}v_{i}=1, the constraints AjA_{j} for j>mj>m^{\prime} cannot be tight, so S[m]S\subseteq[m^{\prime}]. For any i<ji<j and any ε(0,vj]\varepsilon\in(0,v_{j}], replacing (vi,vj)(v_{i},v_{j}) with (vi+ε,vjε)(v_{i}+\varepsilon,v_{j}-\varepsilon) increases the value of i=1mvi2\sum_{i=1}^{m}v_{i}^{2}, so by maximality, no such “smoothing” operation is possible. That is, we can assume the following equality condition for all (i,j)(i,j) with i<ji<j: Either vi=vi1v_{i}=v_{i-1}, or vj=vj+1v_{j}=v_{j+1}, or S[i,j1]S\cap[i,j-1]\neq\emptyset. If S=[m]S=[m^{\prime}], then we are in exactly the maximizing case described (since by the equality conditions for (m+1,i)(m^{\prime}+1,i) we have vi=0v_{i}=0 for all im+2i\geq m^{\prime}+2), so assume otherwise.

First, suppose there is some i0[m]i_{0}\in[m^{\prime}] such that vi0<zv_{i_{0}}<\sqrt{z}, and pick the smallest such index i0i_{0}. Let i1[m]i_{1}\in[m] be the largest index such that vi1>0v_{i_{1}}>0, and note that i1>i0i_{1}>i_{0} since i=1i0vi<1\sum_{i=1}^{i_{0}}v_{i}<1. Then we have vi0<vi01v_{i_{0}}<v_{i_{0}-1} and vi1>vi1+1v_{i_{1}}>v_{i_{1}+1}. But as i=1jvi(1x/z+(j1)z)+vj<(1x/z+(j1)z)+z\sum_{i=1}^{j}v_{i}\leq(1-x/\sqrt{z}+(j-1)\sqrt{z})+v_{j}<(1-x/\sqrt{z}+(j-1)\sqrt{z})+\sqrt{z} for any ji0j\geq i_{0}, we have S[i0,i1]=S\cap[i_{0},i_{1}]=\emptyset, contradicting the equality condition for (i0,i1)(i_{0},i_{1}). So, we can assume vizv_{i}\geq\sqrt{z} for all imi\leq m^{\prime}. In particular, if jSj\in S for some j[m]j\in[m^{\prime}], then we recursively obtain vi=zv_{i}=\sqrt{z} for all i[j+1,m]i\in[j+1,m^{\prime}], so [j,m]S[j,m^{\prime}]\subseteq S.

Thus, we can assume 1S1\notin S, i.e. v1<1x/z+zv_{1}<1-x/\sqrt{z}+\sqrt{z}. By the equality condition for (1,2)(1,2), we must then have v2=v3v_{2}=v_{3}. Let j03j_{0}\geq 3 be the smallest index such that vj0+1vj0v_{j_{0}+1}\neq v_{j_{0}}. By the equality condition for (1,j0)(1,j_{0}), there is some j1S[2,j01]j_{1}\in S\cap[2,j_{0}-1]. Then

1x/z+j1z=i=1j1vi=v1+(j11)v2<1x/z+z+(j11)v2,1-x/\sqrt{z}+j_{1}\sqrt{z}=\sum_{i=1}^{j_{1}}v_{i}=v_{1}+(j_{1}-1)v_{2}<1-x/\sqrt{z}+\sqrt{z}+(j_{1}-1)v_{2},

which implies v2>zv_{2}>\sqrt{z}. But then i=1j1+1vi=1x/z+z+j1v2>1x/z+(j1+1)z\sum_{i=1}^{j_{1}+1}v_{i}=1-x/\sqrt{z}+\sqrt{z}+j_{1}v_{2}>1-x/\sqrt{z}+(j_{1}+1)\sqrt{z}, violating Aj1+1A_{j_{1}+1}. This is a contradiction, so in fact S=[m]S=[m^{\prime}], and we are in the desired maximizing case. ∎

We remark that in Gyárfás’s construction, after removing all edges contained in the vertex set of each red component, each non-red component is left with approximately 1(k1)2(11k1)(n2)=k2(k1)3(n2)\frac{1}{(k-1)^{2}}(1-\frac{1}{k-1})\binom{n}{2}=\frac{k-2}{(k-1)^{3}}\binom{n}{2} edges. Since k2(k1)3=1k2k+1+1k2\frac{k-2}{(k-1)^{3}}=\frac{1}{k^{2}-k+1+\frac{1}{k-2}}, this suggests that the method used to prove Theorem 1 cannot show a lower bound on M(n,k)M(n,k) better than 1k2k+1(n2)\frac{1}{k^{2}-k+1}\binom{n}{2} without introducing additional ideas.

3.2. Multicolor Ramsey numbers of trails and circuits

A trail is a walk without repeated edges, and a circuit is a trail with the same first and last vertex. The (kk-color) Ramsey problem for trails is the question of finding the largest mm such that every kk-coloring of the edges of KnK_{n} contains a monochromatic trail of length mm.

Answering a question of Osumi [6], Conlon and Tyomkyn [1] show that every 22-coloring of the edges of KnK_{n} contains a monochromatic circuit with at least 29n2+O(n3/2)\frac{2}{9}n^{2}+O(n^{3/2}) edges, and this is asymptotically tight. For the case of general kk, they observe that by deleting a forest in each color class of a kk-coloring of KnK_{n} to make each color class Eulerian (i.e. ensuring every vertex has even degree in each color), one can reduce this Ramsey problem to a variant of the problem of determining M(n,k)M(n,k). Where previously we colored the edges of KnK_{n} and found a large monochromatic component, we now apply the same procedure to the graph obtained by deleting at most knkn edges from KnK_{n}. We now prove a lower bound for the general case of this problem, analogous to the bound on M(n,k)M(n,k) given in Theorem 1.

Corollary 8.

Every kk-coloring of the edges of KnK_{n} contains a monochromatic circuit (and hence a monochromatic trail) of length at least 1k2k+54(n2)+Ok(n)\frac{1}{k^{2}-k+\frac{5}{4}}\binom{n}{2}+O_{k}(n).

Proof.

Fix a kk-coloring of E(Kn)E(K_{n}). As in [1], we can remove from each color class a forest that meets all odd degree vertices, leaving a coloring where every color class, and hence every monochromatic connected component, is Eulerian. Let χ\chi be the resulting partial kk-coloring of E(Kn)E(K_{n}), where the (at most knkn) removed edges are left uncolored, and let z(n2)z\binom{n}{2} be the largest number of edges in a monochromatic component in this coloring. It suffices to show that z1k2k+54+Ok(n1)z\geq\frac{1}{k^{2}-k+\frac{5}{4}}+O_{k}(n^{-1}), since every monochromatic component is Eulerian, and thus contains an Eulerian circuit. Fix a color (say, red) with x(n2)x\binom{n}{2} edges, where x1k(n2)nk(n2)=1k2n1x\geq\frac{1}{k}\frac{\binom{n}{2}-nk}{\binom{n}{2}}=\frac{1}{k}-\frac{2}{n-1}. As before, applying Theorem 3 with G=KnG=K_{n} immediately yields zx21k2O(n1)z\geq x^{2}\geq\frac{1}{k^{2}}-O(n^{-1}). Note that this means there is a monochromatic circuit with at least (1k2O(n1))(n2)=1k2(n2)+O(n)\left(\frac{1}{k^{2}}-O(n^{-1})\right)\binom{n}{2}=\frac{1}{k^{2}}\binom{n}{2}+O(n) edges.

To improve this lower bound further, we proceed as in the proof of Theorem 1. Letting R1,,RmR_{1},\dots,R_{m} be the red components, such that (3.1) holds, we obtain (3.2) in the same manner as before, so we can once again apply Lemma 7 to obtain the same upper bound on i=1m(|V(Ri)|2)\sum_{i=1}^{m}\binom{|V(R_{i})|}{2} in terms of xx and zz. Let GG be the complete mm-partite graph with parts V(R1),,V(Rm)V(R_{1}),\dots,V(R_{m}), with its partial coloring induced by χ\chi. Since GG has no red edges, and at most knkn edges are uncolored, one of its color classes HH has at least 1k1(|E(G)|kn)=(1k1Ok(n1))|E(G)|\frac{1}{k-1}(|E(G)|-kn)=\left(\frac{1}{k-1}-O_{k}(n^{-1})\right)|E(G)| edges. Applying Theorem 3 as before yields

z\displaystyle z (1k1Ok(n1))2|E(G)|(n2)(1(k1)2Ok(n1))(1i=1m(|V(Ri)|2)(n2))\displaystyle\geq\left(\frac{1}{k-1}-O_{k}(n^{-1})\right)^{2}\frac{|E(G)|}{\binom{n}{2}}\geq\left(\frac{1}{(k-1)^{2}}-O_{k}(n^{-1})\right)\left(1-\frac{\sum_{i=1}^{m}\binom{|V(R_{i})|}{2}}{\binom{n}{2}}\right)
(1(k1)2+Ok(n1))(1(1x/z+z)2(x/z1)z).\displaystyle\geq\left(\frac{1}{(k-1)^{2}+O_{k}(n^{-1})}\right)\left(1-(1-x/\sqrt{z}+\sqrt{z})^{2}-(x/z-1)z\right).

Letting z=1(kδ)2z=\frac{1}{(k-\delta)^{2}} and noting that x1kO(n1)x\geq\frac{1}{k}-O(n^{-1}), we can perform the same rearrangements and substitutions as in the proof of Theorem 1, simply separating out the Ok(n1)O_{k}(n^{-1}) terms at each step, to derive the inequality 12δ<1k+Ok(n1)1-2\delta<\frac{1}{k}+O_{k}(n^{-1}), and thus

z1k2k+54Ok(n1)=1k2k+54+Ok(n1).z\geq\frac{1}{k^{2}-k+\frac{5}{4}-O_{k}(n^{-1})}=\frac{1}{k^{2}-k+\frac{5}{4}}+O_{k}(n^{-1}).

Then the coloring χ\chi contains a monochromatic connected component, and hence a monochromatic Eulerian circuit, with at least z(n2)1k2k+54(n2)+Ok(n)z\binom{n}{2}\geq\frac{1}{k^{2}-k+\frac{5}{4}}\binom{n}{2}+O_{k}(n) edges, as claimed. ∎

3.3. Three colors

In this section, we prove the lower and upper bounds on M(n,3)M(n,3) in separate lemmas in order to establish Theorem 2. We then discuss the behavior of M(n,3)M(n,3) for small values of nn, and conclude by describing how to adapt our proofs to obtain asymptotically tight bounds for the Ramsey numbers of trails and circuits in three colors.

Lemma 9.

Every 33-coloring of the edges of KnK_{n} contains a monochromatic component with at least 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges.

Proof.

Let G=KnG=K_{n}, and call the three colors red, green, and blue. First, suppose one of the color classes (say, red) is connected. If there are at least 16|E(G)|\frac{1}{6}|E(G)| red edges, we are done. Otherwise, the other two colors together have at least 56|E(G)|\frac{5}{6}|E(G)| edges, so one of them has at least 512|E(G)|\frac{5}{12}|E(G)| edges. Applying Theorem 3 with G=KnG=K_{n} to the graph in that color then gives a monochromatic component with at least (512)2|E(G)|>16|E(G)|(\frac{5}{12})^{2}|E(G)|>\frac{1}{6}|E(G)| edges, so we are again done.

Thus, we can assume every color has at least two components. Without loss of generality, let red be the color with the most edges, so the red graph HH has at least 13|E(G)|\frac{1}{3}|E(G)| edges. Let HH^{\prime} be the component of HH with the highest average degree, so |V(H)|d¯(H)+113n|V(H^{\prime})|\geq\bar{d}(H^{\prime})+1\geq\frac{1}{3}n. We can assume |V(H)|12n|V(H^{\prime})|\leq\frac{1}{2}n; otherwise, we would have |E(H)|12|V(H)|d¯(H)>16|E(G)||E(H^{\prime})|\geq\frac{1}{2}|V(H^{\prime})|\bar{d}(H^{\prime})>\frac{1}{6}|E(G)|. Let V1=V(H)V_{1}=V(H^{\prime}) and V2=V(G)V1V_{2}=V(G)\setminus V_{1}, and let GG^{\prime} be the bipartite graph induced by GG between V1V_{1} and V2V_{2}, so GG^{\prime} has at least |V1||V2|29n2>13|E(G)||V_{1}||V_{2}|\geq\frac{2}{9}n^{2}>\frac{1}{3}|E(G)| edges, all of which must be green or blue.

Fix an edge in GG^{\prime} and consider the monochromatic component C1C_{1} of GG^{\prime} containing this edge. Without loss of generality, assume C1C_{1} is green. Suppose C1C_{1} covers all of V1V_{1}. Since every green edge in GG^{\prime} intersects V1V_{1}, this means there is exactly one green component of GG^{\prime} with a nonzero number of edges. Since there are at least two green components in GG, there is a vertex vV2v\in V_{2} not in C1C_{1}. Then all edges between vv and V1V_{1} must be blue, so all vertices of V1V_{1} are in the same blue component in GG^{\prime}. This implies that there is also exactly one blue component of GG^{\prime} with nonzeroly many edges. Thus, all edges of GG^{\prime} are in one of two monochromatic components, one of which then has at least 12|E(G)|>16|E(G)|\frac{1}{2}|E(G^{\prime})|>\frac{1}{6}|E(G)| edges, as desired.

Refer to caption
Figure 1. Lower bound: Exactly two components of each color

We are likewise done if C1C_{1} covers all of V2V_{2}, so we can assume V1C1V_{1}\setminus C_{1} and V2C1V_{2}\setminus C_{1} are both nonempty. Let A1=V1C1A_{1}=V_{1}\cap C_{1}, B1=V2C1B_{1}=V_{2}\cap C_{1}, A2=V1A1A_{2}=V_{1}\setminus A_{1}, B2=V2B1B_{2}=V_{2}\setminus B_{1}. Then all edges between A1A_{1} and B2B_{2}, or between A2A_{2} and B1B_{1}, can only be blue. Then there are at most two blue components, and thus by assumption exactly two. This in turn implies that all edges between A1A_{1} and B1B_{1}, or between A2A_{2} and B2B_{2}, can only be green, so there are exactly two green components C1C_{1} and C2C_{2}. Finally, all edges between B1B_{1} and B2B_{2} can only be red, so we conclude that there are exactly two red components, V1V_{1} and V2V_{2}; see Figure 1.

Since each of the three colors has exactly two components, one of the components has at least 16|E(G)|\frac{1}{6}|E(G)| edges, as desired. ∎

Lemma 10.

For sufficiently large nn, there exists a 33-coloring of the edges of KnK_{n} such that every monochromatic component contains at most 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges.

Proof.
Refer to caption
Figure 2. Initial construction for the M(n,3)M(n,3) upper bound

We consider the following modification of a construction by Gyárfás: Let V=V1V2V3V4V=V_{1}\cup V_{2}\cup V_{3}\cup V_{4} be a partition of the vertices of G=KnG=K_{n} into four parts, with

n4=|V1||V4|=n4.\left\lceil\frac{n}{4}\right\rceil=|V_{1}|\geq\cdots\geq|V_{4}|=\left\lfloor\frac{n}{4}\right\rfloor.

Letting E(U,W)E(U,W) denote the set of edges between vertex sets UU and WW, color E(V1,V2)E(V3,V4)E(V_{1},V_{2})\cup E(V_{3},V_{4}) red, E(V1,V3)E(V2,V4)E(V_{1},V_{3})\cup E(V_{2},V_{4}) green, and E(V1,V4)E(V2,V3)E(V_{1},V_{4})\cup E(V_{2},V_{3}) blue. For large enough nn, e(Vi,Vj)16(n2)e(V_{i},V_{j})\leq\lceil\frac{1}{6}\binom{n}{2}\rceil for 1i<j41\leq i<j\leq 4, so it remains to extend this partial coloring by coloring the edges within each of the ViV_{i} such that each monochromatic component has at most 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges at the end. It is natural to attempt the simple approach of distributing the edges within each ViV_{i} as evenly as possible among the three colors. However, the possibility of a slight difference in size among the ViV_{i} can yield a Θ(n)\Theta(n) difference among the numbers of edges in each component if we are not sufficiently careful; a different coloring strategy will turn out to be simpler to analyze.

We call an extension of the above partial coloring nice if each green or blue component contains exactly 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges. At least one nice coloring exists: we can color exactly 16(n2)e(V1,V3)\lceil\frac{1}{6}\binom{n}{2}\rceil-e(V_{1},V_{3}) of the edges within V1V_{1} and 16(n2)e(V2,V4)\lceil\frac{1}{6}\binom{n}{2}\rceil-e(V_{2},V_{4}) of the edges within V2V_{2} green, and exactly 16(n2)e(V2,V3)\lceil\frac{1}{6}\binom{n}{2}\rceil-e(V_{2},V_{3}) of the edges within V3V_{3} and 16(n2)e(V1,V4)\lceil\frac{1}{6}\binom{n}{2}\rceil-e(V_{1},V_{4}) of the edges within V4V_{4} blue (there are enough edges within each ViV_{i} to do this when nn is sufficiently large), and color all remaining edges red. See Figure 2 for a diagram of this coloring. Fix a nice coloring where the larger of the two red components contains as few edges as possible.

Suppose one of the red components in this coloring, without loss of generality the one on V1V2V_{1}\cup V_{2}, has more than 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges. Then V3V4V_{3}\cup V_{4} must have less than 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil red edges. Without loss of generality let V1V_{1} contain a red edge ee. If either V3V_{3} contains a green edge, or V4V_{4} contains a blue edge, we can switch the color of that edge with edge ee, preserving the sizes of the green and blue components while decreasing the size of the larger red component by one. Otherwise, V3V_{3} is entirely blue and red, and V4V_{4} is entirely green and red. Since the red component on V3V4V_{3}\cup V_{4} has less than 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges, neither V3V_{3} nor V4V_{4} can be entirely red (for sufficiently large nn, n42+(n42)>16(n2)\lfloor\frac{n}{4}\rfloor^{2}+\binom{\lfloor\frac{n}{4}\rfloor}{2}>\lceil\frac{1}{6}\binom{n}{2}\rceil). Then if V2V_{2} contains a red edge, we can similarly switch two edges to reduce the size of the larger red component by one, so we can assume V2V_{2} is entirely green and blue. But then there is one component of each color (including the larger red component) that does not have any edges within V2V3V4V_{2}\cup V_{3}\cup V_{4}, which means there are at least 316(n2)3\lceil\frac{1}{6}\binom{n}{2}\rceil edges incident to V1V_{1}, a contradiction since

e(V1,V)(n42)+n4(nn4)<316(n2),e(V_{1},V)\leq\binom{\lceil\frac{n}{4}\rceil}{2}+\left\lceil\frac{n}{4}\right\rceil\left(n-\left\lceil\frac{n}{4}\right\rceil\right)<3\left\lceil\frac{1}{6}\binom{n}{2}\right\rceil,

for sufficiently large nn. Thus indeed there is a construction of this form where every monochromatic component has at most 16(n2)\lceil\frac{1}{6}\binom{n}{2}\rceil edges. ∎

Combining Lemmas 9 and 10 yields Theorem 2 as desired.

The construction in the proof of Lemma 10 is well-defined for all n46n\geq 46. A more careful analysis, splitting into cases based on the value of nn modulo 44 and then using a more explicit construction in each case, shows that in fact M(n,3)=16(n2)M(n,3)=\left\lceil\frac{1}{6}\binom{n}{2}\right\rceil for all n11n\geq 11 except n=13,17n=13,17. However, the lower bound from Lemma 9 is not sharp for some of these small values of nn. Closely inspecting each step of our proof for n=17n=17, for example, we can deduce that M(17,3)=24M(17,3)=24, instead of the expected 2323; indeed, all of the bounds before the final step in the proof of Lemma 9 are loose enough to yield a component with at least 2424 edges unless we are in the case depicted in Figure 1, where one of the vertex sets AiA_{i} or BiB_{i} has size 55, and the other three sets have size 44. The set of size 55 then contains 1010 internal edges, some 44 of which are in the same color, yielding a component with 4+20=244+20=24 edges as claimed. This shows a genuine difference in the behavior of M(n,3)M(n,3) for these small values of nn due to integer-related constraints in the extremal configurations.

We can adjust the proof of Lemma 9 to give a lower bound on the size of the largest monochromatic circuit in a 33-coloring of the edges of KnK_{n}, as follows. First, as before, we can remove a forest in each color and leave each color class Eulerian. The resulting graph GG has at least (n2)3n\binom{n}{2}-3n edges, so all but at most 3n3\sqrt{n} of the vertices have degree at least n12nn-1-2\sqrt{n}. We then pass to the induced subgraph GG^{\prime} on these nn3nn^{\prime}\geq n-3\sqrt{n} vertices; the minimum degree of GG^{\prime} is at least n12nn3nn^{\prime}-1-2\sqrt{n}\geq n^{\prime}-3\sqrt{n^{\prime}} when nn is sufficiently large. The argument then proceeds largely as in the proof of Lemma 9, except that the condition of every green or blue component intersecting both V1V_{1} and V2V_{2} is strengthened by requiring every green or blue component to intersect each of V1V_{1} and V2V_{2} in more than 6n6\sqrt{n^{\prime}} vertices. The proof then concludes as before, reducing to the case where there are exactly two components in each color. When the edges between V(G)V(G^{\prime}) and V(G)V(G)V(G)\setminus V(G^{\prime}) are added back in, it remains true that there are at most two components in each color with a positive number of edges. So, at least one monochromatic component in GG has at least 16(|E(G)|(3n)2)112n2O(n)\frac{1}{6}(|E(G)|-(3\sqrt{n})^{2})\geq\frac{1}{12}n^{2}-O(n) edges. Since this component of GG is Eulerian, we have a circuit, and hence a trail, of the desired length, for all sufficiently large nn. The upper bound from Lemma 10 likewise applies to the size of the longest circuit, showing that the lower bound is asymptotically tight.

Acknowledgments. I would like to thank my advisor Jacob Fox for introducing me to this problem and for helpful conversations along the way, as well as David Conlon and Mykhaylo Tyomkyn for the insights from our later joint work that served as inspiration for some of the strengthened arguments in the revised version of this paper. In addition, I would like to thank the anonymous referees for their careful reading and helpful comments, including a specific suggestion that led to a significant strengthening in the bound for general kk in Theorem 1.

References

  • [1] Conlon, D., and Tyomkyn, M. Ramsey numbers of trails and circuits. arXiv preprint arXiv:2109.02633 (2021).
  • [2] DeBiasio, L., Krueger, R. A., and Sárközy, G. N. Large monochromatic components in multicolored bipartite graphs. J. Graph Theory 94, 1 (2020), 117–130.
  • [3] Gyárfás, A. Partition coverings and blocking sets in hypergraphs. Communications of the Computer and Automation Institute of the Hungarian Academy of Sciences 71 (1977), 62.
  • [4] Gyárfás, A. Large monochromatic components in edge colorings of graphs: a survey. In Ramsey theory, vol. 285 of Progr. Math. Birkhäuser/Springer, New York, 2011, pp. 77–96.
  • [5] Gyárfás, A., and Sárközy, G. N. Large monochromatic components in edge colored graphs with a minimum degree condition. Electron. J. Combin. 24, 3 (2017), Paper No. 3.54, 14.
  • [6] Osumi, M. Ramsey numbers of trails. arXiv preprint arXiv:2109.00734 (2021).
  • [7] Ruszinkó, M. Large components in rr-edge-colorings of KnK_{n} have diameter at most five. J. Graph Theory 69, 3 (2012), 337–340.