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

Isfahan University of Technology, Isfahan, [email protected]://orcid.org/0000-0003-4401-2110Institute for Research in Fundamental Sciences (IPM), Tehran, [email protected] research was supported by a grant from IPM. \CopyrightRamin Javadi and Meysam Miralaei\hideLIPIcs

A Conjecture on Rainbow Hamiltonian Cycle Decomposition

Ramin Javadi    Meysam Miralaei
Abstract

Wu in 1999 conjectured that if HH is a subgraph of the complete graph K2n+1K_{2n+1} with nn edges, then there is a Hamiltonian cycle decomposition of K2n+1K_{2n+1} such that each edge of HH is in a separate Hamiltonian cycle. The conjecture was partially settled by Liu and Chen (2023) in cases that |V(H)|n+1|V(H)|\leqslant n+1, HH is a linear forest, or n5n\leqslant 5. In this paper, we settle the conjecture completely. This result can be viewed as a complete graph analogous of Evans conjecture and has some applications in linear arboricity conjecture and restricted size Ramsey numbers.

keywords:
Hamiltonian Cycle Decomposition, Evans Conjecture, Linear Arboricity, Size Ramsey Number
category:
\supplement

1 Introduction

An (edge) decomposition of a graph G=(V,E)G=(V,E) is a partition 𝒞=(C1,,Cn)\mathcal{C}=(C_{1},\ldots,C_{n}) of the edge set EE into disjoint subsets CiC_{i}’s. By |Ci||C_{i}| we mean the number of edges in CiC_{i}. Given a decomposition 𝒞=(C1,,Cn)\mathcal{C}=(C_{1},\ldots,C_{n}) and a subgraph HH of GG, we say that HH is rainbow in 𝒞\mathcal{C} if no two edges of HH are contained in the same set CiC_{i}, for some ii. A Hamiltonian cycle (resp. path) decomposition of a graph G=(V,E)G=(V,E) is a decomposition of GG into subsets C1,,CnEC_{1},\ldots,C_{n}\subseteq E such that each CiC_{i} induces a Hamiltonian cycle (resp. path) in GG.

It is well-known that the complete graph K2n+1K_{2n+1} has a decomposition into nn Hamiltonian cycles. In 1982, Hilton [8], in a seminal work, proposed a procedure for generating all Hamiltonian cycle decompositions of K2n+1K_{2n+1} and found necessary and sufficient conditions for extending a decomposition of KrK_{r} (for some r2nr\leqslant 2n) to a Hamiltonian cycle decomposition of K2n+1K_{2n+1}.

Wu [14], while studying linear arboricity conjecture, raised the question that for any subgraph HH of K2n+1K_{2n+1} with nn edges, if the complete graph K2n+1K_{2n+1} has a Hamiltonian cycle decomposition in which HH is rainbow?

Conjecture 1.

[14] Let HH be a subgraph of K2n+1K_{2n+1} with nn edges. Then, K2n+1K_{2n+1} admits a Hamiltonian cycle decomposition 𝒞\mathcal{C} such that HH is rainbow in 𝒞\mathcal{C}.

In [11], Liu and Chen partially proved Conjecture 1 when n5n\leqslant 5, or HH has at most n+1n+1 vertices, or HH is a linear forest, where a linear forest is a disjoint union of paths.

Theorem 1.1.

[11] Suppose that HH is a subgraph of K2n+1K_{2n+1} with nn edges such that either

  • HH has at most n+1n+1 vertices, or

  • HH is a linear forest, or

  • n5n\leqslant 5.

Then, there is a Hamiltonian cycle decomposition 𝒞\mathcal{C} of K2n+1K_{2n+1} such that HH is rainbow in 𝒞\mathcal{C}.

There are also some partial results regarding Conjecture 1 in [9]. In this paper, we settle Conjecture 1 completely. The conjecture has some applications in linear arboricity conjecture as well as Ramsey theory.

The linear arboricity of a graph GG denoted by la(G)\operatorname{la}(G) is the minimum number kk such that the edge-set of GG can be decomposed into kk linear forest. Linear arboricity conjecture [1] asserts that if GG has maximum degree Δ\Delta, then la(G)(Δ+1)/2\operatorname{la}(G)\leqslant\lceil(\Delta+1)/2\rceil (see e.g. [7, 10] for the literature). A graph GG is called critical if la(G)(Δ(G)+1)/2\operatorname{la}(G)\geqslant\lceil(\Delta(G)+1)/2\rceil and for any edge eE(G)e\in E(G), la(G)>la(Ge)\operatorname{la}(G)>\operatorname{la}(G\setminus e). One can check that if Conjecture 1 is true, then removing any n1n-1 edges from K2n+1K_{2n+1} yields a critical graph. It can also be seen that if the conjecture is true, then for every graph GG with nn vertices and at most (n1)2/2(n-1)^{2}/2 edges, la(G)(n1)/2\operatorname{la}(G)\leqslant(n-1)/2 (see [9]).

Conjecture 1 can also be viewed as a complete graph similar result of Evans conjecture [6]. In fact, Evans conjecture (proved by Smetianuk [13]) equivalently asserts that a proper edge coloring of any subgraph HH of Kn,nK_{n,n} with nn edges, can be extended to a proper edge coloring of Kn,nK_{n,n}. Andersen and Hilton [2] proved analogous results regarding proper edge coloring of complete graphs. In particular, they proved that

Theorem 1.2.

[2]

  • Let HH be a subgraph of K2nK_{2n} with n1n-1 edges. Any proper edge coloring of HH can be extended to a proper edge coloring of K2nK_{2n} with 2n12n-1 colors.

  • Let HH be a subgraph of K2n+1K_{2n+1} with n+1n+1 edges. Any proper edge coloring of HH can be extended to a proper edge coloring of K2n+1K_{2n+1} with 2n+12n+1 colors.

One can view Conjecture 1 as an analogous result of Theorem 1.2 where proper edge coloring is replaced with Hamiltonian cycle decomposition and HH is colored with nn different colors. Therefore, finding necessary and sufficient conditions for extending an arbitrary coloring of HH to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} is an interesting problem and can be viewed as a generalization of Conjecture 1 (see Concluding Remarks and Conjecture 3).

Another application of Conjecture 1 is in Ramsey theory. Given graphs G1,,GkG_{1},\ldots,G_{k}, the Ramsey number R(G1,,Gk)R(G_{1},\ldots,G_{k}) is the minimum number nn such that in every edge coloring of KnK_{n} with kk colors, there is a copy of GiG_{i} with color ii, for some i[k]i\in[k]. The restricted size Ramsey number denoted by r^(G1,,Gk)\hat{r}^{*}(G_{1},\ldots,G_{k}) is the minimum number mm such that there is a graph GG with R(G1,,Gk)R(G_{1},\ldots,G_{k}) vertices and mm edges, such that in every edge coloring of GG with kk colors, there is a copy of GiG_{i} with color ii for some i[k]i\in[k]. Miralaei and Shahsiah in [12] proved some upper bounds for the restricted size Ramsey numbers of stars versus cliques and conjectured that equality holds for these upper bounds. They mentioned that if Conjecture 1 is correct, then their conjecture is proved for the case that the number of stars of size even is positive and even. So, combining our result with their proof can yield the following result.

Theorem 1.3.

Let m,n,p1,,pm,q1,,qnm,n,p_{1},\ldots,p_{m},q_{1},\ldots,q_{n} be positive integers and ss be the number of even integers in the set {qi}i[n]\{q_{i}\}_{i\in[n]}. If ss is positive and even, then

r^(K1,q1,,K1,qn,Kp1,,Kpm)=(t2)s2+1,\hat{r}^{*}(K_{1,q_{1}},\ldots,K_{1,q_{n}},K_{p_{1}},\ldots,K_{p_{m}})=\binom{t}{2}-\dfrac{s}{2}+1,

where t=R(K1,q1,,K1,qn,Kp1,,Kpm)t=R(K_{1,q_{1}},\ldots,K_{1,q_{n}},K_{p_{1}},\ldots,K_{p_{m}}).

In order to prove Conjecture 1, we need the following result from [8] which deals with the extension of an edge decomposition of KrK_{r} to a Hamiltonian cycle decomposition of K2n+1K_{2n+1}.

Theorem 1.4.

[8] Let 1r2n1\leqslant r\leqslant 2n be an integer. A decomposition 𝒞=(C1,,Cn)\mathcal{C}=(C_{1},\ldots,C_{n}) of KrK_{r} can be extended to a Hamiltonian cycle decomposition 𝒞=(C1,,Cn)\mathcal{C}^{\prime}=(C^{\prime}_{1},\ldots,C^{\prime}_{n}) of K2n+1K_{2n+1} where CiCiC_{i}\subseteq C^{\prime}_{i}, for each i[n]i\in[n], if and only if the induced subgraph of GG on each CiC_{i} is a linear forest with at most 2n+1r2n+1-r disjoint paths ((counting a vertex of KrK_{r} with no edge in CiC_{i} as a path of lengths 0)0).

Note that any linear forest on rr vertices with ee edges has exactly rer-e connected components. Therefore, Theorem 1.4 can be reformulated as follows.

Corollary 1.5.

[8] Let 1r2n1\leqslant r\leqslant 2n be an integer. An edge decomposition 𝒞=(C1,,Cn)\mathcal{C}=(C_{1},\ldots,C_{n}) of KrK_{r} can be extended to a Hamiltonian cycle decomposition 𝒞=(C1,,Cn)\mathcal{C}^{\prime}=(C^{\prime}_{1},\ldots,C^{\prime}_{n}) of K2n+1K_{2n+1} where CiCiC_{i}\subseteq C^{\prime}_{i}, for each i[n]i\in[n], if and only if

  1. 1.

    for each i[n]i\in[n], the induced subgraph of KrK_{r} on CiC_{i} is a linear forest, and

  2. 2.

    for each i[n]i\in[n], |Ci|2r2n1|C_{i}|\geqslant 2r-2n-1.

To prove Conjecture 1, we take a minimal counterexample HH to the conjecture with nn edges, in the sense that nn is the smallest integer for which the conjecture does not hold and among all counterexamples with nn edges, HH is the graph with the smallest number of vertices. We give rise to a contradiction in two phases. First, in Section 2, we embed the components of HH which are non-isomorphic to K2K_{2} into an edge decomposition of KrK_{r} for some rr with desired properties and then in Section 3, we extend this decomposition to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which HH is rainbow. This comes to a contradiction and proves Conjecture 1.

1.1 Notations and Terminologies

For positive integers m,nm,n, mnm\leqslant n, the sets {1,,n}\{1,\ldots,n\} and {m,,n}\{m,\ldots,n\} are respectively denoted by [n][n] and [m,n][m,n]. Given a graph G=(V(G),E(G))G=(V(G),E(G)), v(G)v(G) and e(G)e(G) respectively stand for |V(G)||V(G)| and |E(G)||E(G)|. For a subset FE(G)F\subseteq E(G), the induced subgraph of GG on FF is denoted by G[F]G[F]. For any vertex vV(G)v\in V(G), the set of neighbors of vv in GG and G[F]G[F] are respectively denoted by NG(v)N_{G}(v) and NF(v)N_{F}(v). Also, degG(v)\deg_{G}(v) and degF(v)\deg_{F}(v) stand for the number of edges incident with vv in E(G)E(G) and FF, respectively. For two disjoint subsets A,BV(G)A,B\subset V(G), eG(A,B)e_{G}(A,B) stands for the number of edges in GG with one endpoint in AA and one endpoint in BB and we define eF(A,B)=eG[F](A,B)e_{F}(A,B)=e_{G[F]}(A,B). Also, we define NF(A)=aANF(a)N_{F}(A)=\cup_{a\in A}N_{F}(a). If HH is a subgraph of GG, then GHG\setminus H is the graph obtained from GG by removing all edges of HH.

2 Embedding Components Non-isomorphic to K2K_{2}

Let HH be a minimal counterexample to Conjecture 1 with e(H)=ne(H)=n. Also, let HH^{\prime} be the subgraph of HH consisting of all connected components of HH which are non-isomorphic to K2K_{2}. In this section, we prove that there exists an edge decomposition of the complete graph on v(H)v(H^{\prime}) vertices with some desired properties in which HH^{\prime} is rainbow. Then, in the next section, we extend such decomposition to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which HH is rainbow.

Theorem 2.1.

Let HH be a minimal counterexample to Conjecture 1 with nn edges. Also let HH^{\prime} be the subgraph of HH containing all its connected components which are not isomorphic to K2K_{2} and suppose that HH^{\prime} has rr vertices and tt edges. Then, the complete graph KrK_{r} admits an edge decomposition (P1,,Pn)(P_{1},\ldots,P_{n}) such that

  • (1)

    For each i[n]i\in[n], PiP_{i} induces a linear forest in KrK_{r}.

  • (2)

    HH^{\prime} is rainbow in (P1,,Pt)(P_{1},\ldots,P_{t}).

  • (3)

    For each i[t]i\in[t], |Pi|2r2n1|P_{i}|\geqslant 2r-2n-1.

  • (4)

    For each i[t+1,n]i\in[t+1,n], |Pi|2r2n|P_{i}|\geqslant 2r-2n.

Proof 2.2.

Due to Theorem 1.1, HH is not a linear forest and n6n\geqslant 6. Also, every connected component of HH^{\prime} has at least two edges. Thus, HH^{\prime} has at most (t1)/2(t-1)/2 connected components and so rt+(t1)/2=(3t1)/2r\leqslant t+(t-1)/2=(3t-1)/2. Also, we can assume that rn+1r\geqslant n+1. To see this, suppose that rnr\leqslant n. Then, we can extend HH^{\prime} to a graph H^\widehat{H} such that v(H^)e(H^)=nv(\widehat{H})\leqslant e(\widehat{H})=n. Thus, by Theorem 1.1, K2n+1K_{2n+1} admits a Hamiltonian cycle decomposition  𝒞=(C1,,Cn)\mathcal{C}=(C_{1},\ldots,C_{n}) such that H^\widehat{H} is rainbow in 𝒞\mathcal{C}. Without loss of generality, we can assume that HH^{\prime} is rainbow in C1,,CtC_{1},\ldots,C_{t}. Also, since v(H)=rv(H^{\prime})=r, we can remove 2n+1r2n+1-r vertices from K2n+1K_{2n+1} to obtain an edge decomposition (P1,,Pn)(P_{1},\ldots,P_{n}) of KrK_{r} where HH^{\prime} is rainbow in P1,,PtP_{1},\ldots,P_{t}. Conditions (3) and (4) trivially hold as rnr\leqslant n. Therefore, we can assume that rn+1r\geqslant n+1.

Now, let s=r/2s=\lceil r/2\rceil and H′′H^{\prime\prime} be a subgraph of HH^{\prime} with e(H′′)=se(H^{\prime\prime})=s. Since HH is a minimal counterexample and H′′H^{\prime\prime} has fewer edges than HH, there is a Hamiltonian cycle decomposition 𝒞=(C1,,Cs)\mathcal{C}=(C_{1},\ldots,C_{s}) for K2s+1K_{2s+1} such that H′′H^{\prime\prime} is rainbow in 𝒞\mathcal{C}. Since v(H′′)v(H)=rv(H^{\prime\prime})\leqslant v(H^{\prime})=r, we can remove from K2s+1K_{2s+1} one vertex whenever rr is even and two vertices whenever rr is odd to obtain an edge decomposition 𝒫=(P1,,Ps)\mathcal{P}=(P_{1},\ldots,P_{s}) of KrK_{r} in which H′′H^{\prime\prime} is rainbow and each PiP_{i} contains r1r-1 edges when rr is even and at least r2r-2 edges when rr is odd. Since v(H)=rv(H^{\prime})=r, we can assume that HH^{\prime} is a subgraph of KrK_{r}. The subgraph HH′′H^{\prime}\setminus H^{\prime\prime} has exactly tst-s edges which are distributed in the subsets PiP_{i}’s. Now, define tst-s new empty subsets Ps+1,,PtP_{s+1},\ldots,P_{t} and move each edge of HH′′H^{\prime}\setminus H^{\prime\prime} from P1,,PsP_{1},\ldots,P_{s} to a separate set PiP_{i}, s+1its+1\leqslant i\leqslant t. Also, define Pt+1,,PnP_{t+1},\ldots,P_{n} as empty subsets. By abuse of notation, let 𝒫=(P1,,Pn)\mathcal{P}=(P_{1},\ldots,P_{n}) be the obtained decomposition of KrK_{r} where HH^{\prime} is rainbow in P1,,PtP_{1},\ldots,P_{t}. Also, without loss of generality, assume that |P1||P2||Ps||Ps+1|==|Pt|=1>|Pt+1|==|Pn|=0|P_{1}|\geqslant|P_{2}|\geqslant\cdots\geqslant|P_{s}|\geqslant|P_{s+1}|=\cdots=|P_{t}|=1>|P_{t+1}|=\cdots=|P_{n}|=0. We are going to move some edges from large PiP_{i}’s to small PiP_{i}’s to ensure that each PiP_{i} induces a linear forest such that

|Pi|{2r2n11it,2r2nt+1in.\displaystyle|P_{i}|\geqslant\begin{cases}2r-2n-1&1\leqslant i\leqslant t,\\ 2r-2n&t+1\leqslant i\leqslant n.\end{cases} (1)

Now, we consider the following two cases.

Case 1. n+1r(4n1)/3n+1\leqslant r\leqslant(4n-1)/3.

First, we claim that |Pns|4r4n1|P_{n-s}|\geqslant 4r-4n-1. To prove the claim, by the contrary, suppose that |Pns|4r4n2|P_{n-s}|\leqslant 4r-4n-2. Therefore, |Pi|4r4n2|P_{i}|\leqslant 4r-4n-2, for all nsisn-s\leqslant i\leqslant s. Since before moving edges of HH′′H^{\prime}\setminus H^{\prime\prime}, each PiP_{i} has at least r2r-2 edges, we have

2nr2tr2e(HH′′)\displaystyle 2n-r\geqslant{2t-r}\geqslant 2\,e(H^{\prime}\setminus H^{\prime\prime}) 2i=nss(r2|Pi|)\displaystyle\geqslant 2\sum_{i=n-s}^{s}\left(r-2-|P_{i}|\right)
2i=nss(r24r+4n+2)\displaystyle\geqslant 2\sum_{i=n-s}^{s}\left(r-2-4r+4n+2\right)
2(rn+1)(4n3r).\displaystyle\geqslant 2(r-n+1)(4n-3r).

Thus, fn(r)=2(rn+1)(4n3r)2n+r0f_{n}(r)=2(r-n+1)(4n-3r)-2n+r\leqslant 0. On the other hand, when n+1r(4n1)/3n+1\leqslant r\leqslant(4n-1)/3, the function fn(r)f_{n}(r) attains its minimum on r=n+1r=n+1 or r=(4n1)/3r=(4n-1)/3. However, fn(n+1)=3n11>0f_{n}(n+1)=3n-11>0 as n6n\geqslant 6 and fn((4n1)/3)=1>0f_{n}((4n-1)/3)=1>0. This is a contradiction which proves the claim. Therefore, |P1||Pns|4r4n1|P_{1}|\geqslant\cdots\geqslant|P_{n-s}|\geqslant 4r-4n-1.

Now, for each ii, s+1its+1\leqslant i\leqslant t, do the following. Let abab be the edge of HH′′H^{\prime}\setminus H^{\prime\prime} in PiP_{i}. Consider the set Pj=PisP_{j}=P_{i-s} and let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PjP_{j}. Also, let e1e_{1} and e2e_{2} be two edges in PjP_{j} incident with aa and bb, respectively (if exist) such that in G[Pj{e1,e2}]G[P_{j}\setminus\{e_{1},e_{2}\}], aa and bb are in different connected components. Now, move 2r2n22r-2n-2 edges from Pj{e,e1,e2}P_{j}\setminus\{e^{\prime},e_{1},e_{2}\} to PiP_{i}. Since |Pj|4r4n12r2n2+3|P_{j}|\geqslant 4r-4n-1\geqslant 2r-2n-2+3, this can be done. Since PiP_{i} has already contained the edge abab and we excluded the edges e1e_{1} and e2e_{2}, so after moving these 2r2n22r-2n-2 edges, G[Pi]G[P_{i}] is a linear forest. Also, the remaining edges in PjP_{j} are at least 4r4n1(2r2n2)=2r2n+14r-4n-1-(2r-2n-2)=2r-2n+1.

Now, for each ii, t+1int+1\leqslant i\leqslant n, do the following. Consider the set Pj=PisP_{j}=P_{i-s} and let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PjP_{j}. Now, move 2r2n2r-2n arbitrary edges from Pj{e}P_{j}\setminus\{e^{\prime}\} to PiP_{i}. Since |Pj|4r4n12r2n+1|P_{j}|\geqslant 4r-4n-1\geqslant 2r-2n+1, this can be done. Also, the remaining edges in PjP_{j} are at least 4r4n1(2r2n)=2r2n14r-4n-1-(2r-2n)=2r-2n-1. On the other hand, for each ii, ns+1isn-s+1\leqslant i\leqslant s, |Pi|(r2)e(HH′′)=r2t+s2r2n1|P_{i}|\geqslant(r-2)-e(H^{\prime}\setminus H^{\prime\prime})=r-2-t+s\geqslant 2r-2n-1, as tnt\leqslant n.

Hence, every PiP_{i} satisfies Condition (1), G[Pi]G[P_{i}] is a linear forest and we are done.

Case 2. 4n/3r(3t1)/24n/3\leqslant r\leqslant(3t-1)/2.

It is clear that in this case (8n+3)/9tn(8n+3)/9\leqslant t\leqslant n. First, we claim that |P2n2s|3r3nϵ|P_{2n-2s}|\geqslant 3r-3n-\epsilon, where ϵ=max{tn+1,0}\epsilon=\max\{t-n+1,0\}. To prove the claim, by the contrary, suppose that |P2n2s|3r3nϵ1|P_{2n-2s}|\leqslant 3r-3n-\epsilon-1. Therefore, |Pi|3r3nϵ1|P_{i}|\leqslant 3r-3n-\epsilon-1, for all 2n2sis2n-2s\leqslant i\leqslant s. Now, define δ=0\delta=0 when rr is even and δ=1\delta=1 when rr is odd. Since before moving edges of HH′′H^{\prime}\setminus H^{\prime\prime}, each PiP_{i} has at least r1δr-1-\delta edges, so

2trδ=2t2s=2e(HH′′)\displaystyle 2t-r-\delta=2t-2s=2\,e(H^{\prime}\setminus H^{\prime\prime}) 2i=2n2ss(r1δ|Pi|)\displaystyle\geqslant 2\sum_{i=2n-2s}^{s}\left(r-1-\delta-|P_{i}|\right)
2i=2n2ss(r1δ(3r3nϵ1))\displaystyle\geqslant 2\sum_{i=2n-2s}^{s}(r-1-\delta-(3r-3n-\epsilon-1))
(3r4n+2)(3n2rδ+ϵ).\displaystyle\geqslant(3r-4n+2)(3n-2r-\delta+\epsilon).

Thus, fn,t(r)=(3r4n+2)(3n2rδ+ϵ)2t+r+δ0f_{n,t}(r)=(3r-4n+2)(3n-2r-\delta+\epsilon)-2t+r+\delta\leqslant 0. On the other hand, when 4n/3r(3t1)/24n/3\leqslant r\leqslant(3t-1)/2, the function fn,t(r)f_{n,t}(r) attains its minimum on r=4n/3r=4n/3 or r=(3t1)/2r=(3t-1)/2. However, fn,t(4n/3)=2n2tδ+2ϵ>0f_{n,t}(4n/3)=2n-2t-\delta+2\epsilon>0, as ϵ=1\epsilon=1 whenever t=nt=n and ϵ=0\epsilon=0 whenever t<nt<n. Also, 2fn,t((3t1)/2)=(9t8n+1)(3n3t+1δ+ϵ)t+2δ12f_{n,t}((3t-1)/2)=(9t-8n+1)(3n-3t+1-\delta+\epsilon)-t+2\delta-1. Note that if t=nt=n, then ϵ=1\epsilon=1 and 2fn,t((3t1)/2)=(n+1)(1δ)+2δ>02f_{n,t}((3t-1)/2)=(n+1)(1-\delta)+2\delta>0. So, (8n+3)/9tn1(8n+3)/9\leqslant t\leqslant n-1, then n12n\geqslant 12, ϵ=0\epsilon=0 and the function gn(t)=2fn,t((3t1)/2)g_{n}(t)=2f_{n,t}((3t-1)/2) attains its minimum on t=(8n+3)/9t=(8n+3)/9 or t=n1t=n-1. However, gn((8n+3)/9)=4n/92δ4/3>0g_{n}((8n+3)/9)=4n/9-2\delta-4/3>0 as n12n\geqslant 12 and gn(n1)=(n8)(4δ)n+2δ=n(3δ)+10δ32>0g_{n}(n-1)=(n-8)(4-\delta)-n+2\delta=n(3-\delta)+10\delta-32>0 as n12n\geqslant 12. This leads to a contradiction and proves the claim. Therefore, |P1||P2n2s|3r3nϵ|P_{1}|\geqslant\cdots\geqslant|P_{2n-2s}|\geqslant 3r-3n-\epsilon. Now, for each ii, s+1its+1\leqslant i\leqslant t, do the following process.

Suppose that Pi={ab}P_{i}=\{ab\}, where abab is an edge of HH′′H^{\prime}\setminus H^{\prime\prime}. We are going to move 2r2n22r-2n-2 edges from some sets PjP_{j} and PP_{\ell} to PiP_{i} to guarantee Condition (1). First, consider Pj=PisP_{j}=P_{i-s} which has at least 3r3nϵ3r-3n-\epsilon edges. Also, let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PjP_{j}. Also, we can choose at most two edges e1,e2e_{1},e_{2} in PjP_{j} incident with a,ba,b, respectively, such that in G[Pj{e1,e2}]G[P_{j}\setminus\{e_{1},e_{2}\}], vertices aa and bb are in different connected components. Now, let PjP_{j}^{\prime} be a set of rn2r-n-2 arbitrary edges in Pj{e,e1,e2}P_{j}\setminus\{e^{\prime},e_{1},e_{2}\} (it exists since 3r3nϵ3rn23r-3n-\epsilon-3\geqslant r-n-2). Move all edges in PjP_{j}^{\prime} from PjP_{j} to PiP_{i}. Then, PiP_{i} now induces a linear forest in GG with rn1r-n-1 edges. Let IiI_{i} and JiJ_{i} be respectively the set of all inner vertices and endpoints of the paths in G[Pi]G[P_{i}].

Now, consider P=Pi2s+tP_{\ell}=P_{i-2s+t} and let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PP_{\ell}. Now, let PP_{\ell}^{\prime} be a subset of PP_{\ell} containing all edges incident with IiI_{i} along with at most one edge incident with each vertex in JiJ_{i} such that if z,wz,w are two endpoints of a path in G[Pi]G[P_{i}] then zz and ww are in different connected components of G[PP]G[P_{\ell}\setminus P_{\ell}^{\prime}]. Since PP_{\ell} induces a linear forest, we have |P|2|Ii|+|Ji|=2|Pi|=2(rn1)|P_{\ell}^{\prime}|\leqslant 2|I_{i}|+|J_{i}|=2|P_{i}|=2(r-n-1). Therefore, since |P|3r3n1|P_{\ell}|\geqslant 3r-3n-1, we have |P(P{e})|3r3n1(2r2n2)1=rn|P_{\ell}\setminus(P^{\prime}_{\ell}\cup\{e^{\prime}\})|\geqslant 3r-3n-1-(2r-2n-2)-1=r-n. Finally, move rnr-n arbitrary edges from P(P{e})P_{\ell}\setminus(P^{\prime}_{\ell}\cup\{e^{\prime}\}) to PiP_{i}. Choosing such edges guarantees that PiP_{i} induces a linear forest with 2r2n12r-2n-1 edges.

Finally, for each ii, t+1int+1\leqslant i\leqslant n, do the following process. First, note that since tn1t\leqslant n-1, we have ϵ=0\epsilon=0. First, consider Pj=Pi+t2sP_{j}=P_{i+t-2s}. Also, let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PjP_{j}. We choose an arbitrary subset PjPj{e}P^{\prime}_{j}\subseteq P_{j}\setminus\{e^{\prime}\} such that |Pj|=rn1|P^{\prime}_{j}|=r-n-1. This can be done since rn13r3n1r-n-1\leqslant 3r-3n-1. Now, move all rn1r-n-1 edges in PjP^{\prime}_{j} to PiP_{i}. Then, PiP_{i} is now a linear forest in GG with rn1r-n-1 edges. Let IiI_{i} and JiJ_{i} be respectively the set of all inner vertices and endpoints of the paths in G[Pi]G[P_{i}].

Now, consider P=Pi+n2sP_{\ell}=P_{i+n-2s} and let ee^{\prime} be the unique edge of H′′H^{\prime\prime} in PP_{\ell}. Now, let PP_{\ell}^{\prime} be a subset of PP_{\ell} containing all edges incident with IiI_{i} along with at most one edge incident with each vertex in JiJ_{i} such that if z,wz,w are two endpoints of a path in G[Pi]G[P_{i}] then zz and ww are in different connected components of G[PP]G[P_{\ell}\setminus P_{\ell}^{\prime}]. Since PP_{\ell} induces a linear forest, we have |P|2|Ii|+|Ji|=2|Pi|=2(rn1)|P_{\ell}^{\prime}|\leqslant 2|I_{i}|+|J_{i}|=2|P_{i}|=2(r-n-1). Therefore, since |P|3r3n|P_{\ell}|\geqslant 3r-3n, we have |P(P{e})|3r3n(2r2n2)1=rn+1|P_{\ell}\setminus(P^{\prime}_{\ell}\cup\{e^{\prime}\})|\geqslant 3r-3n-(2r-2n-2)-1=r-n+1. Finally, move rn+1r-n+1 arbitrary edges from P(P{e})P_{\ell}\setminus(P^{\prime}_{\ell}\cup\{e^{\prime}\}) to PiP_{i}. Choosing such edges guarantees that PiP_{i} induces a linear forest with 2r2n2r-2n edges.

At the end of this process, for each i[s+1,t]i\in[s+1,t], |Pi|=2r2n1|P_{i}|=2r-2n-1 and for each i[t+1,n]i\in[t+1,n], |Pi|=2r2n|P_{i}|=2r-2n. Also, for each j[1,2n2s]j\in[1,2n-2s], we have moved from PjP_{j} at most rnr-n edges whenever t=nt=n and at most rn+1r-n+1 edges whenever t<nt<n. Therefore, at most rnϵ+1r-n-\epsilon+1 edges are moved from PjP_{j}. Thus, |Pj|3r3nϵ(rnϵ+1)=2r2n1|P_{j}|\geqslant 3r-3n-\epsilon-(r-n-\epsilon+1)=2r-2n-1. Finally, for each i[2n2s+1,s]i\in[2n-2s+1,s], |Pi|(r2)e(HH′′)=r2t+s2r2n1|P_{i}|\geqslant(r-2)-e(H^{\prime}\setminus H^{\prime\prime})=r-2-t+s\geqslant 2r-2n-1 as ntn\geqslant t and sr/2s\geqslant r/2. Hence, 𝒫=(P1,,Pn)\mathcal{P}=(P_{1},\ldots,P_{n}) satisfies the conditions of the theorem. This completes the proof.

3 Embedding Components Isomorphic to K2K_{2}

Let HH be a minimal counterexample to Conjecture 1 with nn edges. Suppose that H=H(nt)K2H=H^{\prime}\cup(n-t)K_{2}, where HH^{\prime} is the union of all connected components of HH non-isomorphic to K2K_{2} and let tt and rr be respectively the number of edges and vertices of HH^{\prime}. In Theorem 2.1, we proved that HH^{\prime} can be embedded into an edge decomposition of KrK_{r} with certain properties. In this section, we are going to prove that such decomposition can be extended to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which HH is rainbow and so Conjecture 1 is settled completely. In fact, we prove the following theorem.

Theorem 3.1.

Let GG be a graph with tt edges and rr vertices, where G≇tK2G\not\cong tK_{2} and also let ntn\geqslant t be an integer. Suppose that KrK_{r} admits an edge decomposition 𝒫=(P1,,Pn)\mathcal{P}=(P_{1},\ldots,P_{n}) such that

  • for each i[n]i\in[n], PiP_{i} induces a linear forest in KrK_{r},

  • GG is rainbow in P1,,PtP_{1},\ldots,P_{t} and

  • for each i[t]i\in[t], |Pi|2r2n1|P_{i}|\geqslant 2r-2n-1 and for each i[t+1,n]i\in[t+1,n], |Pi|2r2n|P_{i}|\geqslant 2r-2n.

Then, 𝒫\mathcal{P} can be extended to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which the graph G(nt)K2G\cup(n-t)K_{2} is rainbow.

In order to prove Theorem 3.1, we begin with the given decomposition 𝒫\mathcal{P} of KrK_{r} and apply an idea by Hilton [8] to add vertices of K2n+1KrK_{2n+1}\setminus K_{r} one by one and we try to construct a rainbow matching in the remaining classes PiP_{i}, t+1int+1\leqslant i\leqslant n. For this, we need a couple of lemmas. First, let us review some definitions.

For an edge coloring of a multigraph GG with colors 1,,k1,\ldots,k, let Ci(v)C_{i}(v) be the set of edges of color ii incident with vv and let Ci(u,v)C_{i}(u,v) be the set of edges of color ii with endpoints u,vu,v. An edge coloring of GG is called a balanced edge coloring if for every vertex vV(G)v\in V(G),

max1i<jk||Ci(v)||Cj(v)||1,\max_{1\leqslant i<j\leqslant k}\big{|}|C_{i}(v)|-|C_{j}(v)|\big{|}\leqslant 1,

and for every pair of vertices u,vV(G)u,v\in V(G),

max1i<jk||Ci(u,v)||Cj(u,v)||1.\max_{1\leqslant i<j\leqslant k}\big{|}|C_{i}(u,v)|-|C_{j}(u,v)|\big{|}\leqslant 1.

We need the following result due to de Werra.

Lemma 3.2.

[3, 4, 5] For every bipartite multigraph GG and every positive integer kk, GG admits a balanced kk-edge coloring.

Let GG be a bipartite multigraph with bipartition (X,Y)(X,Y). By a pairing of GG on XX, we mean a collection {x:xX}\{\mathcal{E}_{x}:x\in X\} such that each x\mathcal{E}_{x} is a family of disjoint two-subsets of the edges incident with xx. If {e1,e2}x\{e_{1},e_{2}\}\in\mathcal{E}_{x}, then we say e1e_{1} and e2e_{2} form an xx-pair. We have also an additional condition that if there are more than one edge between two vertices xx and yy, then at most one of these parallel edges is paired with an edge xyxy^{\prime}, yyy^{\prime}\neq y. We also need the following lemma which can be derived from the proof of Theorem 1 in [8]. For the completeness, we provide a proof of the following lemma in Appendix A.

Lemma 3.3.

[8] Let FF be a bipartite multigraph with bipartition (X,Y)(X,Y) such that the degree of each vertex in YY is even. Also, consider a pairing of FF on XX. Then, FF has a balanced edge-coloring with two colors such that if e1,e2E(F)e_{1},e_{2}\in E(F) form an xx-pair for some xXx\in X, then e1e_{1} and e2e_{2} have different colors.

Finally, we need the following lemma.

Lemma 3.4.

Suppose that GG is a bipartite multigraph with bipartition (X,Y)(X,Y) whose edge set is partitioned into two sets A,BA,B such that for every yYy\in Y, degB(y)degA(y)\deg_{B}(y)\geqslant\deg_{A}(y) and for every xXx\in X, degA(x),degB(x)η\deg_{A}(x),\deg_{B}(x)\leqslant\eta for some integer η\eta. Also, suppose that for some vertex x0Xx_{0}\in X, we have either

  • degA(x0)>degB(x0)\deg_{A}(x_{0})>\deg_{B}(x_{0}), or

  • degA(x0)=degB(x0)\deg_{A}(x_{0})=\deg_{B}(x_{0}) is odd, η\eta is even and for every yYy\in Y, degB(y)\deg_{B}(y) is even.

Then, there is a subset CE(G)C\subset E(G) such that

  • degC(y)=degA(y)\deg_{C}(y)=\deg_{A}(y), for all yYy\in Y,

  • degC(x0)=degA(x0)1\deg_{C}(x_{0})=\deg_{A}(x_{0})-1, and

  • degA(x)degC(x)η\deg_{A}(x)\leqslant\deg_{C}(x)\leqslant\eta, for all xX{x0}x\in X\setminus\{x_{0}\}.

Proof 3.5.

Let X~\widetilde{X} be the set of all vertices xXx\in X such that there is an alternating path PxP_{x} from x0x_{0} to xx with edges in AA and BB alternately starting from an edge in AA (note that x0X~x_{0}\in\widetilde{X}).

If for some xX~{x0}x\in\widetilde{X}\setminus\{x_{0}\}, degA(x)<η\deg_{A}(x)<\eta, then we set C=(APx)(PxA)C=(A\setminus P_{x})\cup(P_{x}\setminus A) which satisfies all the conditions. Now, suppose that for all xX~{x0}x\in\widetilde{X}\setminus\{x_{0}\}, degA(x)=ηdegB(x)\deg_{A}(x)=\eta\geqslant\deg_{B}(x). Let Y~=NA(X~)\widetilde{Y}=N_{A}(\widetilde{X}) and so eA(X~,Y~)eB(X~,Y~)e_{A}(\widetilde{X},\widetilde{Y})\geqslant e_{B}(\widetilde{X},\widetilde{Y}). On the other hand, it is clear that NB(Y~)X~N_{B}(\widetilde{Y})\subseteq\widetilde{X} and since for every yYy\in Y, degB(y)degA(y)\deg_{B}(y)\geqslant\deg_{A}(y), we have eB(X~,Y~)eA(X~,Y~)e_{B}(\widetilde{X},\widetilde{Y})\geqslant e_{A}(\widetilde{X},\widetilde{Y}). If degA(x0)>degB(x0)\deg_{A}(x_{0})>\deg_{B}(x_{0}), then we have eA(X~,Y~)>eB(X~,Y~)e_{A}(\widetilde{X},\widetilde{Y})>e_{B}(\widetilde{X},\widetilde{Y}) which is a contradiction. Now, suppose that degA(x0)=degB(x0)\deg_{A}(x_{0})=\deg_{B}(x_{0}) is odd, η\eta is even and for every yYy\in Y, degB(y)\deg_{B}(y) is even. Then, eB(X~,Y~)=eA(X~,Y~)=η|X~|η+degA(x0)e_{B}(\widetilde{X},\widetilde{Y})=e_{A}(\widetilde{X},\widetilde{Y})=\eta|\widetilde{X}|-\eta+\deg_{A}(x_{0}) is odd and eB(X~,Y~)=yY~degB(y)e_{B}(\widetilde{X},\widetilde{Y})=\sum_{y\in\widetilde{Y}}\deg_{B}(y) is even which is again a contradiction. This contradiction completes the proof.

Now, we are ready to prove Theorem 3.1.

Proof 3.6 (Proof of Theorem 3.1.).

We start with the edge decomposition 𝒫\mathcal{P} of KrK_{r} with the described properties and for each s[0,nt]s\in[0,n-t], we are going to prove that 𝒫\mathcal{P} can be extended to a decomposition 𝒬=(Q1,,Qn)\mathcal{Q}=(Q_{1},\ldots,Q_{n}) of K2s+rK_{2s+r} such that GsK2G\cup sK_{2} is rainbow in Q1,,Qt+sQ_{1},\ldots,Q_{t+s} and

|Qi|{4s+2r2n1i[t+s],4s+2r2ni[t+s+1,n].|Q_{i}|\geqslant\begin{cases}4s+2r-2n-1&i\in[t+s],\\ 4s+2r-2n&i\in[t+s+1,n].\end{cases} (2)

For s=0s=0 we have 𝒬=𝒫\mathcal{Q}=\mathcal{P}. Suppose that it has been done until some s<nts<n-t and we do it for s+1s+1.

Set k=2n2sr+1k=2n-2s-r+1 and suppose that the vertex set of K2s+rK_{2s+r} is {u1,,u2s+r}\{u_{1},\ldots,u_{2s+r}\}. We define an auxiliary bipartite multigraph G~\widetilde{G} with a bipartition X={u1,,u2s+r}X=\{u_{1},\ldots,u_{2s+r}\} and Y={c1,,cn}Y=\{c_{1},\ldots,c_{n}\} such that cic_{i} is adjacent to uju_{j} with an edge if uju_{j} is an endpoint of a path in QiQ_{i} and cic_{i} is adjacent to uju_{j} with two parallel edges if uju_{j} is incident with no edge in QiQ_{i}. Then, we have

degG~(uj)\displaystyle\deg_{\widetilde{G}}(u_{j}) =k,\displaystyle=k, j[2s+r],\displaystyle\forall\,j\in[2s+r], (3)
degG~(ci)\displaystyle\deg_{\widetilde{G}}(c_{i}) =4s+2r2|Qi|,\displaystyle=4s+2r-2|Q_{i}|, i[n].\displaystyle\forall\,i\in[n]. (4)

To see (3), for each {0,1,2}\ell\in\{0,1,2\}, let nn_{\ell} be the number of sets QiQ_{i} where degQi(uj)=\deg_{Q_{i}}(u_{j})=\ell. Therefore, degG~(uj)=2n0+n1=2(nn1n2)+n1=2nn12n2\deg_{\widetilde{G}}(u_{j})=2n_{0}+n_{1}=2(n-n_{1}-n_{2})+n_{1}=2n-n_{1}-2n_{2}. On the other hand, uju_{j} is incident with 2s+r12s+r-1 edges in K2s+rK_{2s+r}. So, n1+2n2=2s+r1n_{1}+2n_{2}=2s+r-1 and we have degG~(uj)=2n2sr+1=k\deg_{\widetilde{G}}(u_{j})=2n-2s-r+1=k.

To see (4), note that if QiQ_{i} is empty, then degG~(ci)=2(2s+r)\deg_{\widetilde{G}}(c_{i})=2(2s+r). Now, adding any edge to QiQ_{i} reduces the degree of cic_{i} in G~\widetilde{G} by two (because of its endpoints). Thus, degG~(ci)=4s+2r2|Qi|\deg_{\widetilde{G}}(c_{i})=4s+2r-2|Q_{i}|.

Therefore, by (2), for each i[s+t]i\in[s+t], we have degG~(ci)2k\deg_{\widetilde{G}}(c_{i})\leqslant 2k and for each i[s+t+1,n]i\in[s+t+1,n], degG~(ci)2k2\deg_{\widetilde{G}}(c_{i})\leqslant 2k-2. Now, we prove the following claim.

Claim 2.

There exist two disjoint subgraphs G1,G2G_{1},G_{2} of G~\widetilde{G} such that for =1,2\ell=1,2,

  1. i.

    for each j[2s+r]j\in[2s+r], degG(uj)=1\deg_{G_{\ell}}(u_{j})=1,

  2. ii.

    for each i[n]i\in[n], degG(ci)2\deg_{G_{\ell}}(c_{i})\leqslant 2 and degG(cs+t+1)1\deg_{G_{\ell}}(c_{s+t+1})\leqslant 1,

  3. iii.

    for each i[n]i\in[n], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x, for some integer xx, then

    degG1(ci)+degG2(ci){4x1is+t,3xi=s+t+1,5xs+t+2in.\deg_{G_{1}}(c_{i})+\deg_{G_{2}}(c_{i})\geqslant\begin{cases}4-x&1\leqslant i\leqslant s+t,\\ 3-x&i=s+t+1,\\ 5-x&s+t+2\leqslant i\leqslant n.\end{cases}
  4. iv.

    for each i[n]i\in[n], if cic_{i} is adjacent to some vertices uj1u_{j_{{}_{1}}} and uj2u_{j_{{}_{2}}} in GG_{\ell}, then uj1u_{j_{{}_{1}}} and uj2u_{j_{{}_{2}}} are not the endpoints of a path in QiQ_{i},

  5. v.

    if cs+t+1c_{s+t+1} is adjacent to some vertex uj1u_{j_{{}_{1}}} in G1G_{1} and uj2u_{j_{{}_{2}}} in G2G_{2}, then uj1u_{j_{{}_{1}}} and uj2u_{j_{{}_{2}}} are not the endpoints of a path in Qs+t+1Q_{s+t+1}. Also, G1G2G_{1}\cup G_{2} has no parallel edges.

Proof 3.7 (Proof of Claim 2).

First, let G^\widehat{G} be obtained from G~\widetilde{G} by duplicating each edge of G~\widetilde{G}. Also, we add a new vertex uu^{\prime} and join it to all vertices cs+t+1,,cnc_{s+t+1},\ldots,c_{n} each one with 4 parallel edges. Therefore, for all j[2s+r]j\in[2s+r], degG^(uj)=2k\deg_{\widehat{G}}(u_{j})=2k and degG^(u)2k2\deg_{\widehat{G}}(u^{\prime})\leqslant 2k-2. Also, for each i[n]i\in[n], we have degG^(ci)4k\deg_{\widehat{G}}(c_{i})\leqslant 4k.

By Lemma 3.2, there is a balanced kk-edge coloring {L1,,Lk}\{L_{1},\ldots,L_{k}\} of G^\widehat{G} with k=2n2sr+1k=2n-2s-r+1 colors. Fix a color class LqL_{q}. Since degG^(uj)=2k\deg_{\widehat{G}}(u_{j})=2k and degG^(ci)4k\deg_{\widehat{G}}(c_{i})\leqslant 4k, we have degLq(uj)=2\deg_{{L}_{q}}(u_{j})=2 and degLq(ci)4\deg_{{L}_{q}}(c_{i})\leqslant 4. Also, for each i[s+t]i\in[s+t], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x, then degLq(ci)(4k4x)/k4x\deg_{{L}_{q}}(c_{i})\geqslant\lfloor(4k-4x)/k\rfloor\geqslant 4-x (since k4k\geqslant 4 as r2t1r\leqslant 2t-1). Moreover, for each i[s+t+1,n]i\in[s+t+1,n], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x, then degLq(ci)(4k4x+4)/k5x\deg_{{L}_{q}}(c_{i})\geqslant\lfloor(4k-4x+4)/k\rfloor\geqslant 5-x (as k4k\geqslant 4 and x1x\geqslant 1).

Since degG^(u)2k2\deg_{\widehat{G}}(u^{\prime})\leqslant 2k-2, there is a color class say L1L_{1}, such that uu^{\prime} is incident with exactly one edge in L1L_{1}. Without loss of generality, suppose that four parallel edges between uu^{\prime} and cs+t+1c_{s+t+1} are in the color classes L1,L2,L3L_{1},L_{2},L_{3} and L4L_{4}. Now, for each q[4]q\in[4], let L^q\widehat{L}_{q} be the edge-set obtained from LqL_{q} by removing all edges incident with uu^{\prime}.

We know that degL^q(cs+t+1)3\deg_{\widehat{L}_{q}}(c_{s+t+1})\leqslant 3. If degL^1(cs+t+1)=3\deg_{\widehat{L}_{1}}(c_{s+t+1})=3, then we apply Lemma 3.4 by setting G=G^uG=\widehat{G}\setminus u^{\prime}, A=L^1A=\widehat{L}_{1}, B=L^2B=\widehat{L}_{2}, x0=cs+t+1x_{0}=c_{s+t+1} and η=4\eta=4 to obtain a set CC and define L1=CL^{\prime}_{1}=C. Also, if degL^1(cs+t+1)2\deg_{\widehat{L}_{1}}(c_{s+t+1})\leqslant 2, then define L1=L^1L^{\prime}_{1}=\widehat{L}_{1}. It is clear that degL1(cs+t+1)2\deg_{L^{\prime}_{1}}(c_{s+t+1})\leqslant 2 and degL^1(ci)degL1(ci)4\deg_{\widehat{L}_{1}}(c_{i})\leqslant\deg_{L^{\prime}_{1}}(c_{i})\leqslant 4, for all is+t+1i\neq s+t+1 and degL1(uj)=degL^1(uj)\deg_{L^{\prime}_{1}}(u_{j})=\deg_{\widehat{L}_{1}}(u_{j}), for all jj. With a similar argument, using the color classes L^3\widehat{L}_{3} and L^4\widehat{L}_{4}, we can find a color class L3L^{\prime}_{3} such that degL3(cs+t+1)2\deg_{L^{\prime}_{3}}(c_{s+t+1})\leqslant 2 and degL^3(ci)degL3(ci)4\deg_{\widehat{L}_{3}}(c_{i})\leqslant\deg_{L^{\prime}_{3}}(c_{i})\leqslant 4, for all is+t+1i\neq s+t+1 and degL3(uj)=degL^3(uj)\deg_{L^{\prime}_{3}}(u_{j})=\deg_{\widehat{L}_{3}}(u_{j}), for all jj. Let L=L1L3L^{\prime}=L^{\prime}_{1}\cup L^{\prime}_{3}. Then, we have

  • for all j[2s+r]j\in[2s+r], degL(uj)=4\deg_{L^{\prime}}(u_{j})=4,

  • for each i[s+t]i\in[s+t], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x, for some x0x\geqslant 0, then 82xdegL(ci)88-2x\leqslant\deg_{L^{\prime}}(c_{i})\leqslant 8,

  • if degG~(cs+t+1)=2k2x\deg_{\widetilde{G}}(c_{s+t+1})=2k-2x, for some x1x\geqslant 1, then 62xdegL(cs+t+1)46-2x\leqslant\deg_{L^{\prime}}(c_{s+t+1})\leqslant 4,

  • for each i[s+t+2,n]i\in[s+t+2,n], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x, for some x1x\geqslant 1, then 102xdegL(ci)810-2x\leqslant\deg_{L^{\prime}}(c_{i})\leqslant 8 except for possibly one cic_{i}, say cs+t+2c_{s+t+2}, for which 92xdegL(cs+t+2)89-2x\leqslant\deg_{L^{\prime}}(c_{s+t+2})\leqslant 8.

Now, we apply Lemma 3.3 for the graph F=G^[L]F=\widehat{G}[L^{\prime}] and the following pairing: if there are two parallel edges e1,e2e_{1},e_{2} joining cic_{i} to uju_{j}, then e1e_{1} and e2e_{2} form an ii-pair. If cic_{i} is adjacent to uju_{j} and uju_{j^{\prime}}, where uju_{j} and uju_{j^{\prime}} are the endpoints of one path in QiQ_{i}, then ciujc_{i}u_{j} and ciujc_{i}u_{j^{\prime}} form an ii-pair. Also, if there is more than one edge incident with cic_{i} still not paired off, then such edges are paired arbitrarily (one edge may possibly be remained unpaired). Therefore, there is a balanced two-coloring for FF, called E1,E2E_{1},E_{2}, such that if cic_{i} is adjacent to some vertices uj1u_{j_{{}_{1}}} and uj2u_{j_{{}_{2}}} in EiE_{i}, i=1,2i=1,2, then uj1u_{j_{{}_{1}}} and uj2u_{j_{{}_{2}}} are not the endpoints of a path in QiQ_{i}. Now, since degL(cs+t+2)92x\deg_{L^{\prime}}(c_{s+t+2})\geqslant 9-2x, one of the sets E1,E2E_{1},E_{2}, say E1E_{1}, satisfies degE1(cs+t+2)5x\deg_{E_{1}}(c_{s+t+2})\geqslant 5-x. Finally, applying Lemma 3.2 to the graph G^[E1]\widehat{G}[E_{1}], we find a balanced two-edge-coloring E11,E12E_{11},E_{12}. Setting G1=G~[E11]G_{1}=\widetilde{G}[E_{11}] and G2=G~[E12]G_{2}=\widetilde{G}[E_{12}], we obtain the desired subgraphs.

G2:G_{2}:cs+t+1c_{s+t+1}uj6u_{j_{{}_{6}}}uj3u_{j_{{}_{3}}}uj4u_{j_{{}_{4}}}cic_{i}G1:G_{1}:cs+t+1c_{s+t+1}uj5u_{j_{{}_{5}}}uj1u_{j_{{}_{1}}}uj2u_{j_{{}_{2}}}cic_{i}uj5u_{j_{{}_{5}}}uj6u_{j_{{}_{6}}}u2s+r+1u_{2s+r+1}u2s+r+2u_{2s+r+2}Qs+t+1Q_{s+t+1}uj2u_{j_{{}_{2}}}uj1u_{j_{{}_{1}}}uj3u_{j_{{}_{3}}}uj4u_{j_{{}_{4}}}u2s+r+1u_{2s+r+1}u2s+r+2u_{2s+r+2}QiQ_{i}
Figure 1: Adding edges to the sets Q1,,QnQ_{1},\ldots,Q_{n} using auxiliary graphs G1,G2G_{1},G_{2}.

Now, we are ready to extend the decomposition 𝒬\mathcal{Q} of K2s+rK_{2s+r} to a decomposition of K2s+r+2K_{2s+r+2} by adding two new vertices u2s+r+1u_{2s+r+1} and u2s+r+2u_{2s+r+2} as follows (see Figure 1). For each =1,2\ell=1,2, if uju_{j} is adjacent to cic_{i} in GG_{\ell}, then we add the edge u2s+r+uju_{2s+r+\ell}u_{j} to QiQ_{i}. Also, we add the edge u2s+r+1u2s+r+2u_{2s+r+1}u_{2s+r+2} to Qs+t+1Q_{s+t+1}. So, by abuse of notation, we obtain a decomposition 𝒬=(Q1,,Qn)\mathcal{Q}=(Q_{1},\ldots,Q_{n}) for K2s+r+2K_{2s+r+2}. It is clear that G(s+1)K2G\cup(s+1)K_{2} is rainbow in Q1,,Qs+t+1Q_{1},\ldots,Q_{s+t+1}. Moreover, because of Conditions (ii), (iv) and (v) of Claim 2, each QiQ_{i} induces a linear forest in K2s+r+2K_{2s+r+2}. Now, we check Condition (2). First, for each i[n]i\in[n], if degG~(ci)=2k2x\deg_{\widetilde{G}}(c_{i})=2k-2x for integer xx, then, by (4) and Condition (iii) of Claim 2, we have

|Qi|{4s+2r2n1+x+4x=4(s+1)+2r2n11is+t,4s+2r2n1+x+3x+1=4(s+1)+2r2n1i=s+t+1,4s+2r2n1+x+5x=4(s+1)+2r2ns+t+2in.|Q_{i}|\geqslant\begin{cases}4s+2r-2n-1+x+4-x=4(s+1)+2r-2n-1&1\leqslant i\leqslant s+t,\\ 4s+2r-2n-1+x+3-x+1=4(s+1)+2r-2n-1&i=s+t+1,\\ 4s+2r-2n-1+x+5-x=4(s+1)+2r-2n&s+t+2\leqslant i\leqslant n.\end{cases}

Note that when i=s+t+1i=s+t+1, we have included the new edge us+t+1us+t+2u_{s+t+1}u_{s+t+2} into Qs+t+1Q_{s+t+1} and so we have at least 4x4-x new edges therein. Hence, inductively we can construct an edge decomposition 𝒬=(Q1,,Qn)\mathcal{Q}=(Q_{1},\ldots,Q_{n}) for K2n2t+rK_{2n-2t+r} such that

  • for each i[n]i\in[n], QiQ_{i} induces a linear forest in K2n2t+rK_{2n-2t+r},

  • G(nt)K2G\cup(n-t)K_{2} is rainbow in Q1,,QnQ_{1},\ldots,Q_{n} and

  • for each i[n]i\in[n], |Qi|2n4t+2r1|Q_{i}|\geqslant 2n-4t+2r-1.

Now, we apply Corollary 1.5 to extend 𝒬\mathcal{Q} to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which G(nt)K2G\cup(n-t)K_{2} is rainbow. This completes the proof.

Combining Theorems 2.1 and 3.1 settles Conjecture 1 as follows. Let HH be a minimal counterexample to Conjecture 1, where H=H(nt)K2H=H^{\prime}\cup(n-t)K_{2} and HH^{\prime} has no component isomorphic to K2K_{2}. Then, by Theorem 2.1, we can construct an edge decomposition 𝒫\mathcal{P} with the desired properties and by Theorem 3.1, we can extend it to a Hamiltonian cycle decomposition of K2n+1K_{2n+1} in which HH is rainbow. This contradicts the fact that HH is a counterexample. So, Conjecture 1 is settled.

4 Concluding Remarks

In this paper, we proved a conjecture by Wu asserting that for any subgraph HH of K2n+1K_{2n+1} with nn edges, there is a Hamiltonian cycle decomposition for K2n+1K_{2n+1} in which HH is rainbow. As it was mentioned in the introduction, this result is, in some point of view, a complete graph counterpart of Evans conjecture. In other words, the conjecture asserts that if the edges of HH are colored by nn different colors, then we can extend such coloring to a coloring of K2n+1K_{2n+1} such that each color class induces a Hamiltonian cycle. So, one may ask if its generalization to an arbitrary coloring of HH is correct. We state this assertion as the following conjecture.

Conjecture 3.

Let HH be a subgraph of K2n+1K_{2n+1} with nn edges which are colored such that each color class induces a linear forest. Then, the coloring can be extended to a coloring of K2n+1K_{2n+1} with nn colors such that each color class induces a Hamiltonian cycle.

Note that Conjecture 1 is a special case of Conjecture 3 when the edges of HH are colored with nn different colors. Also, it is noteworthy that in Conjecture 3, the number nn for the number of edges of HH is optimal, because if we consider H=P2P3H=P_{2}\cup P_{3} as a subgraph of K5K_{5} and color the edges of P3P_{3} with color 11 and the edge of P2P_{2} with color 22, then such a coloring cannot be extended to a Hamiltonian cycle decomposition of K5K_{5}.

References

  • [1] J. Akiyama, G. Exoo, and F. Harary. Covering and packing in graphs IV: Linear arboricity. Networks, 11(1):69–72, 1981.
  • [2] L. D. Andersen and A. J. W. Hilton. Symmetric Latin square and complete graph analogues of the evans conjecture. Journal of Combinatorial Designs, 2(4):197–252, 1994.
  • [3] D. de Werra. Balanced schedules. INFOR: Information Systems and Operational Research, 9(3):230–237, 1971.
  • [4] D. de Werra. A few remarks on chromatic scheduling. In Combinatorial Programming: Methods and Applications: Proceedings of the NATO Advanced Study Institute held at the Palais des Congrès, Versailles, France, 2–13 September, 1974, pages 337–342. Springer, 1975.
  • [5] D. de Werra. On a particular conference scheduling problem. INFOR: Information Systems and Operational Research, 13(3):308–315, 1975.
  • [6] T. Evans. Embedding incomplete Latin squares. The American Mathematical Monthly, 67(10):958–961, 1960.
  • [7] A. Ferber, J. Fox, and V. Jain. Towards the linear arboricity conjecture. Journal of Combinatorial Theory, Series B, 142:56–79, 2020.
  • [8] A. J. W. Hilton. Hamiltonian decompositions of complete graphs. Journal of Combinatorial Theory, Series B, 36(2):125–134, 1984.
  • [9] T. Khoplyklang. Hamiltonian decompositions and linear arboricity of graphs. PhD thesis, Queen Mary University of London, 2023.
  • [10] R. Lang and L. Postle. An improved bound for the linear arboricity conjecture. Combinatorica, pages 1–23, 2023.
  • [11] Y. Liu and Y. Chen. Rainbow subgraphs in Hamiltonian cycle decompositions of complete graphs. Discrete Mathematics, 346(8):113479, 2023.
  • [12] M. Miralaei and M. Shahsiah. On the multicolor size Ramsey number of stars and cliques. Discrete Mathematics, 343(8):111899, 2020.
  • [13] B. Smetianuk. A new construction on Latin squares I. A proof of the Evans conjecture, Ars Combinatorica, 11:155–172, 1981.
  • [14] J. Wu. The linear arboricity theory of graphs. PhD thesis, Shandong University, 1999.

Appendix A Proof of Lemma 3.3

Here we give a proof of Lemma 3.3. Let FF be a bipartite multigraph with bipartition (X,Y)(X,Y) such that e(F)=ee(F)=e and the degree of each vertex in YY is even. Also, consider a pairing of FF on XX and extend it as follows. If there are two edges incident with a vertex xXx\in X which are unpaired, then form a xx-pair and do this until there is at most one unpaired edge on each vertex of XX.

We proceed by induction on ee. There is nothing to prove for e=2e=2. So, assume that e>2e>2. Now, we claim that there is an even cycle or an even path (a path with even edges) in FF, say QQ, satisfying the following properties:

  1. 1.

    if QQ is a path, its endpoints are in XX,

  2. 2.

    for every vertex xV(Q)Xx\in V(Q)\cap X, if degQ(x)=2\deg_{Q}(x)=2, then the edges of QQ incident with xx form an xx-pair,

  3. 3.

    there is a balanced 2-edge coloring on QQ with the color classes E1E_{1} and E2E_{2} such that the edges of QQ are alternately in E1E_{1} and E2E_{2}.

In order to prove the claim, select an arbitrary edge e0=xi0yj0e_{0}=x_{i_{0}}y_{j_{0}} and place it in E1E_{1}. If e0e_{0} has an xi0x_{i_{0}}-pair, say e1=xi0yj1e_{1}=x_{i_{0}}y_{j_{1}}, place e1e_{1} in E2E_{2}. Since the degree of yj1y_{j_{1}} is even, it has still at least one unassigned edge, let it be e2=xi1yj1e_{2}=x_{i_{1}}y_{j_{1}} and place e2e_{2} in E1E_{1}. Continue in this way until one of the following occurs.

  • An edge e2s+1=xisyjs+1e_{2s+1}=x_{i_{s}}y_{j_{s+1}} is found (and placed in E2E_{2}), where yjs+1=yjpy_{j_{s+1}}=y_{j_{p}} for some 0ps0\leqslant p\leqslant s. In this case we have an even cycle and the claim is proved.

  • An edge e2s=xisyjse_{2s}=x_{i_{s}}y_{j_{s}} is found (and placed in E1E_{1}) and e2se_{2s} has no xisx_{i_{s}}-pair; then let xisx_{i_{s}} be the endpoint of this path. Now, there is at least one unassigned edge on yj0y_{j_{0}}. Let this edge be e1=xi1yj0e_{-1}=x_{i_{-1}}y_{j_{0}} and place it in E2E_{2}. If e1e_{-1} has no xi1x_{i_{-1}}-pair, let xi1x_{i_{-1}} be the other endpoint of the path QQ. If e1e_{-1} has an xi1x_{i_{-1}}-pair, let it be e2=xi1yj1e_{-2}=x_{i_{-1}}y_{j_{-1}} and place it in E1E_{1}. The vertex yj1y_{j_{-1}} still has an unassigned edge on it, say e3=xi2yj1e_{-3}=x_{i_{-2}}y_{j_{-1}}. Place e3e_{-3} in E2E_{2}. Continue in this way until an edge e(2r+1)=xir1yjre_{-(2r+1)}=x_{i_{-r-1}}y_{j_{-r}} is found (and placed in E2E_{2}) such that the edge e(2r+1)e_{-(2r+1)} has no xir1x_{i_{-r-1}}-pair. Then, let xir1x_{i_{-r-1}} be the other endpoint of the path constructed.

Let FF^{\prime} be the spanning subgraph of FF obtained by deleting the edges of QQ. By the induction hypothesis, FF^{\prime} has a balanced 22-edge coloring with the color classes E1E_{1}^{\prime} and E2E_{2}^{\prime} satisfying the assertion. Clearly, the edge coloring on FF with the color classes E1′′E_{1}^{\prime\prime} and E2′′E_{2}^{\prime\prime} with E1′′=E1E1E_{1}^{\prime\prime}=E_{1}\cup E_{1}^{\prime} and E2′′=E2E2E_{2}^{\prime\prime}=E_{2}\cup E_{2}^{\prime} is the desired balanced 22-edge coloring on FF.