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

Aharoni’s rainbow cycle conjecture holds up to an additive constant

Patrick Hompe  and  Tony Huynh [email protected] Dipartimento di Informatica, Sapienza Università di Roma, Italy [email protected]
Abstract.

In 2017, Aharoni proposed the following generalization of the Caccetta-Häggkvist conjecture: if GG is a simple nn-vertex edge-colored graph with nn color classes of size at least rr, then GG contains a rainbow cycle of length at most n/r\lceil n/r\rceil.

In this paper, we prove that, for fixed rr, Aharoni’s conjecture holds up to an additive constant. Specifically, we show that for each fixed r1r\geqslant 1, there exists a constant αrO(r5log2r)\alpha_{r}\in O(r^{5}\log^{2}r) such that if GG is a simple nn-vertex edge-colored graph with nn color classes of size at least rr, then GG contains a rainbow cycle of length at most n/r+αrn/r+\alpha_{r}.

1. Introduction

A digraph DD is simple if for all distinct vertices u,vV(D)u,v\in V(D), there is at most one arc from uu to vv. Perhaps the most famous open problem concerning digraphs is the following conjecture of Caccetta and Häggkvist from 1978.

Conjecture 1.1 ([3]).

Let DD be a simple nn-vertex digraph with minimum out-degree at least rr. Then DD contains a directed cycle of length at most n/r\lceil n/r\rceil.

Although the Caccetta-Häggkvist conjecture has attracted considerable attention over the years, proving it still remains out of reach. We highlight here one partial result of Shen [8], which is in the spirit of our paper.

Theorem 1.2.

Let DD be a simple nn-vertex digraph with minimum out-degree at least rr. Then DD contains a directed cycle of length at most n/r+73\lceil n/r\rceil+73.

Let GG be an edge-colored graph. We say that GG is simple if no color class contains parallel edges. A subgraph of GG is rainbow if it does not contain two edges of the same color. In this paper, we focus on the following generalization of the Caccetta-Häggkvist conjecture, due to Aharoni [1].

Conjecture 1.3.

Let GG be a simple nn-vertex edge-colored graph with nn color classes of size at least rr. Then GG contains a rainbow cycle of length at most n/r\lceil n/r\rceil.

It is well-known that Aharoni’s conjecture implies the Caccetta-Häggkvist conjecture (see for example [5] or [4]). For completeness, we include the proof here.

Proof of Conjecture 1.1 assuming Conjecture 1.3.

Let DD be a simple nn-vertex digraph with minimum out-degree at least rr. We create an edge-coloured graph GG from DD with V(G)=V(D)V(G)=V(D) as follows. For each arc (u,v)A(D)(u,v)\in A(D) we create an edge uvE(G)uv\in E(G) with color uu. Note that GG has nn color classes of size at least rr (since DD has minimum out-degree at least rr). Moreover, since DD is a simple digraph it follows that GG is a simple edge-colored graph. By Conjecture 1.3, GG contains a rainbow cycle of length at most n/r\lceil n/r\rceil. This rainbow cycle necessarily corresponds to a directed cycle in DD. ∎

Note that Conjecture 1.3 holds when r=1r=1, since we can choose one edge from each color class and use the easy fact that every nn-vertex graph with nn edges contains a cycle. The r=2r=2 case of Conjecture 1.3 was proven by DeVos et al. [5].

Theorem 1.4 ([5]).

Let GG be a simple nn-vertex edge-colored graph with nn color classes of size at least 2. Then GG contains a rainbow cycle of length at most n/2\lceil n/2\rceil.

The r=3r=3 case is still open, but the following approximate result was obtained by Clinch et al. [4].

Theorem 1.5 ([4]).

Let GG be a simple nn-vertex edge-colored graph with nn color classes of size at least 33. Then GG contains a rainbow cycle of length at most 4n/9+74n/9+7.

For general rr, the best bound is due to Hompe and Spirkl [6], who proved that Aharoni’s conjecture holds up to a multiplicative constant.

Theorem 1.6 ([6]).

Let c=1011c=10^{11} and GG be a simple nn-vertex edge-colored graph with nn color classes of size at least crcr. Then GG contains a rainbow cycle of length at most n/r\lceil n/r\rceil.

In this paper, we strengthen Theorem 1.5 and Theorem 1.6, by proving that Aharoni’s conjecture holds for each fixed rr up to an additive constant. In fact, we prove a stronger ‘defect’ version which allows some color classes to have size less than rr. To state our main result precisely, for an edge-colored graph GG, let C(G)C(G) be the set of colors appearing in E(G)E(G), and for a color cC(G)c\in C(G), let cGc_{G} be the edges of E(G)E(G) with color cc. For each rr\in\mathbb{N}, we define

𝖽𝖾𝖿r(G)cC(G),|cG|r(r|cG|).\mathsf{def}_{r}(G)\coloneqq\sum_{c\in C(G),|c_{G}|\leqslant r}(r-|c_{G}|).

The following is our main theorem.

Theorem 1.7.

For all r2r\geqslant 2, there exists a constant αrO(r5log2r)\alpha_{r}\in O(r^{5}\log^{2}r) such that if GG is a simple nn-vertex edge-colored graph with nn color classes of size at least 22 and at most rr, then GG contains a rainbow cycle of length at most

n+𝖽𝖾𝖿r(G)r+αr.\frac{n+\mathsf{def}_{r}(G)}{r}+\alpha_{r}.

As an immediate corollary of Theorem 1.7, we obtain the result claimed in the abstract.

Corollary 1.8.

For all r1r\geqslant 1, there exists a constant αrO(r5log2r)\alpha_{r}\in O(r^{5}\log^{2}r) such that if GG is a simple nn-vertex edge-colored graph with nn color classes of size at least rr, then GG contains a rainbow cycle of length at most

nr+αr.\frac{n}{r}+\alpha_{r}.
Proof.

Since we have already noted that Conjecture 1.3 holds when r=1r=1, we may assume r2r\geqslant 2. Next, let GG’ be obtained from GG by deleting edges so that each color class has size exactly rr. By applying Theorem 1.7 to GG^{\prime} and noting that 𝖽𝖾𝖿r(G)=0\mathsf{def}_{r}(G^{\prime})=0, it follows that GG^{\prime} (and hence also GG) contains a rainbow cycle of length at most nr+αr\frac{n}{r}+\alpha_{r}. ∎

Our proof shows that we may take αr=3000000r5log2(r)\alpha_{r}=3000000r^{5}\log^{2}(r). An interesting open problem is to improve the dependence on rr. Given that the Caccetta-Häggkvist conjecture is still open, likely the best one could hope for is αr\alpha_{r}\leqslant\ell for some universal constant \ell. This would be a generalization of Theorem 1.2, up to an additive constant.

We would also like to point out that there have been many recent results on Aharoni’s conjecture. These are too numerous to list here, but we refer the interested reader to [4] for a survey of the state of the art.

Paper Outline

In Section 2 we state some preliminary lemmas needed in the proof of our main theorem. Our proof is split into two cases, depending on whether there are many ‘non-star’ vertices or few non-star vertices. We set-up our notation and assumptions in Section 3, where we also prove a key lemma used in the proof. Then, we handle the many non-star case in Section 4, and the few non-star case in Section 5.

2. Preliminaries

We begin with some results from the literature needed in the proof of our main theorem. We say that a graph is excess-kk if it contains at least kk more edges than vertices. Bollobás and Szemerédi [2] proved the following upper bound on the girth of excess-kk graphs.

Theorem 2.1.

For all n4n\geqslant 4 and k2k\geqslant 2, every nn-vertex, excess-kk graph has girth at most

2(n+k)3k(logk+loglogk+4).\frac{2(n+k)}{3k}(\log k+\log\log k+4).

In Theorem 2.1, and for this paper, all logs are in base 2. We will use the following immediate corollary of Theorem 2.1 (see [6]).

Corollary 2.2.

For all n4n\geqslant 4 and k2k\geqslant 2, every nn-vertex, excess-kk graph has girth at most

14(n+k)logk3k.\frac{14(n+k)\log k}{3k}.

Now, let DD be a digraph. A sink is a vertex with out-degree equal to zero. For each rr\in\mathbb{N}, we define

𝖽𝖾𝖿r(D)uU(rdeg+(u)),\mathsf{def}_{r}(D)\coloneqq\sum_{u\in U}(r-\deg^{+}(u)),

where UU is the set of vertices of DD of out-degree at most rr. We require the following theorem of Shen [7].

Theorem 2.3.

Let DD be a simple nn-vertex digraph with no sink, and let gg be the length of a shortest directed cycle of DD. If g2r1g\geq 2r-1, then nr(g1)+1𝖽𝖾𝖿r(D)n\geq r(g-1)+1-\mathsf{def}_{r}(D).

We will use the following immediate corollary of Theorem 2.3.

Corollary 2.4.

Let DD be a simple nn-vertex digraph with no sink, and let gg be the length of a shortest directed cycle of DD. Then

gn+𝖽𝖾𝖿r(D)r+2r1.g\leqslant\frac{n+\mathsf{def}_{r}(D)}{r}+2r-1.
Proof.

The statement is obviously true if g2r2g\leqslant 2r-2. Thus, we may assume that g2r1g\geqslant 2r-1. It follows by Theorem 2.3 that

gn+𝖽𝖾𝖿r(D)1r+1n+𝖽𝖾𝖿r(D)r+2r1,g\leqslant\frac{n+\mathsf{def}_{r}(D)-1}{r}+1\leqslant\frac{n+\mathsf{def}_{r}(D)}{r}+2r-1,

as desired. ∎

3. The Set-up

For the remainder of the paper, GG is a simple edge-colored graph with nn vertices, nn colors, and each color class of size at least 22 and at most rr, for r2r\geq 2. In other words, GG satisfies the hypotheses of Theorem 1.7. Let k:=78rlogrk:=\lceil 78r\log{r}\rceil and f(r)=80k2r2f(r)=80k^{2}r^{2}. We will show that, for fixed rr, we may take αr:=(4r+2)f(r)+2r\alpha_{r}:=(4r+2)f(r)+2r in Theorem 1.7.

To prove Theorem 1.7, we proceed by induction on nn. The base case is nαrn\leq\alpha_{r}, which is true since there is always a rainbow cycle of length at most nn. Thus, for the remainder of the paper, we also assume that n>αrn>\alpha_{r}.

For a subset of vertices XV(G)X\subseteq V(G), we let G[X]G[X] be the induced subgraph of GG on XX, and C(X)C(X) be the set of colors which appear in G[X]G[X]. By our earlier definition of C(G)C(G) for an edge-colored graph GG, an alternative way of writing this definition is C(X)=C(G[X])C(X)=C(G[X]).

Let UV(G)U\subseteq V(G). A galaxy rooted at UU is a collection M:={Mu:uU}M:=\{M_{u}:u\in U\} of subgraphs of GG such that

  • MuM_{u} is a star centred at uu, for all uUu\in U,

  • |E(Mu)|1|E(M_{u})|\geqslant 1 and all edges of MuM_{u} are the same color cuc_{u},

  • cucvc_{u}\neq c_{v}, for all uvu\neq v.

We say that xV(G)x\in V(G) is an MM-neighbor of uUu\in U if xx is a neighbor of uu in MuM_{u}. We let C(M):={cu:uU}C(M):=\{c_{u}:u\in U\} be the set of MM-colors. Finally, for each uUu\in U we let |u|M:=|E(Mu)|\lvert u\rvert_{M}:=|E(M_{u})|. We will use the following key lemma multiple times in the proof.

Lemma 3.1.

Let R=X{u1,,um}V(G)R=X\cup\{u_{1},\cdots,u_{m}\}\subseteq V(G), M:={Mui:i[m]}M:=\{M_{u_{i}}:i\in[m]\} be a galaxy rooted at {u1,,um}\{u_{1},\cdots,u_{m}\}, CXC(X)C_{X}\subseteq C(X), and dd\in\mathbb{R} be such that

  1. (1)

    V(Mui){ui}X{u1,,ui1}V(M_{u_{i}})\setminus\{u_{i}\}\subseteq X\cup\{u_{1},\dots,u_{i-1}\} for all i[m]i\in[m],

  2. (2)

    CXC(M)=C_{X}\cap C(M)=\emptyset,

  3. (3)

    for every pair of vertices u,vXu,v\in X, there exists a rainbow path consisting of colors in CXC_{X} from uu to vv in G[X]G[X] of length at most dd.

Then for every pair of vertices u,vRu,v\in R, there exists a rainbow path consisting of colors in CXC(M)C_{X}\cup C(M) from uu to vv in G[R]G[R] of length at most

m+i=1m(r|ui|M)r+d+2.\displaystyle\frac{m+\sum_{i=1}^{m}(r-\lvert u_{i}\rvert_{M})}{r}+d+2.
Proof.

See Figure 1 for an example with m=7m=7. For each uRu\in R, we define a function p:RRp:R\to R as follows. If uXu\in X, then p(u)=up(u)=u. If uXu\notin X, but some MM-neighbor vv of uu is in XX, we arbitrarily choose one such vv, and define p(u)=vp(u)=v. Otherwise, define p(u)=uip(u)=u_{i}, where ii is the minimum index such that uiu_{i} is an MM-neighbor of uu. For example, in Figure 1, we have p(u4)=u2p(u_{4})=u_{2}.

Let u,vRu,v\in R. Let PuP^{u} be the path beginning at uu and repeatedly applying the map pp until we reach a vertex in XX. Define PvP^{v} analogously, but starting from vv.

First suppose there exists uaV(Pu)u_{a}\in V(P^{u}) and ubV(Pv)u_{b}\in V(P^{v}) such that uau_{a} and ubu_{b} have a common MM-neighbor in {u1,,um}\{u_{1},\dots,u_{m}\}. Choose uau_{a} and ubu_{b} such that a+ba+b is maximum. By symmetry, we may assume a>ba>b. Let PaP_{a} be the subpath of PuP^{u} from uu to uau_{a}, and let PbP_{b} be the subpath of PvP^{v} from vv to ubu_{b}. Let Qa=Pa{ua}Q_{a}=P_{a}\setminus\{u_{a}\} and Qb=Pb{ub}Q_{b}=P_{b}\setminus\{u_{b}\}. By the maximality of a+ba+b,

m\displaystyle m 1+wV(Qa)|w|M+wV(Qb)|w|M\displaystyle\geqslant 1+\sum_{w\in V(Q_{a})}|w|_{M}+\sum_{w\in V(Q_{b})}|w|_{M}
=1+wV(Qa)(r(r|w|M))+wV(Qb)(r(r|w|M)).\displaystyle=1+\sum_{w\in V(Q_{a})}(r-(r-|w|_{M}))+\sum_{w\in V(Q_{b})}(r-(r-|w|_{M})).

Rearranging, we have

|E(Pb)|+|E(Pa)|\displaystyle|E(P_{b})|+|E(P_{a})| m1+wV(Qa)(r|w|M)+wV(Qb)(r|w|M)r\displaystyle\leqslant\frac{m-1+\sum_{w\in V(Q_{a})}(r-|w|_{M})+\sum_{w\in V(Q_{b})}(r-|w|_{M})}{r}
m+i=1m(r|ui|M)r.\displaystyle\leqslant\frac{m+\sum_{i=1}^{m}(r-|u_{i}|_{M})}{r}.

Let uu_{\ell} be a common MM-neighbor of uau_{a} and ubu_{b} in {u1,,um}\{u_{1},\dots,u_{m}\}. Then PaPb{uau,ubu}P_{a}\cup P_{b}\cup\{u_{a}u_{\ell},u_{b}u_{\ell}\} is a rainbow path consisting of colors in CXC(M)C_{X}\cup C(M) from uu to vv in G[R]G[R] of length at most

m+i=1m(r|wi|M)r+d+2,\frac{m+\sum_{i=1}^{m}(r-|w_{i}|_{M})}{r}+d+2,

as required.

For the remaining case, let uXu^{\prime}\in X and vXv^{\prime}\in X be the other ends of PuP^{u} and PvP^{v}. Let Qu=PuuQ^{u}=P^{u}\setminus u^{\prime} and Qv=PvvQ^{v}=P^{v}\setminus v^{\prime}, and let SuS^{u} be QuQ^{u} minus its last vertex and SvS^{v} be QvQ^{v} minus its last vertex. If both uu and vv are in XX, then we are done by (3). Thus, by symmetry, we may assume that u=uau=u_{a} for some a[m]a\in[m] and that either vXv\in X or v=ubv=u_{b} for some b<ab<a. By the previous case, we may assume that for all xV(Qu)x\in V(Q^{u}) and yV(Qv)y\in V(Q^{v}), xx and yy do not have any common MM-neighbors in {u1,,um}\{u_{1},\dots,u_{m}\}. Therefore,

m\displaystyle m 1+wV(Sa)|w|M+wV(Sb)|w|M\displaystyle\geqslant 1+\sum_{w\in V(S_{a})}|w|_{M}+\sum_{w\in V(S_{b})}|w|_{M}
=1+wV(Sa)(r(r|w|M))+wV(Sb)(r(r|w|M)).\displaystyle=1+\sum_{w\in V(S_{a})}(r-(r-|w|_{M}))+\sum_{w\in V(S_{b})}(r-(r-|w|_{M})).

Rearranging, we have

|E(Pu)|+|E(Pv)|\displaystyle|E(P_{u})|+|E(P_{v})| |E(Qu)|+|E(Qv)|+2\displaystyle\leqslant|E(Q_{u})|+|E(Q_{v})|+2
m1+wV(Su)(r|w|M)+wV(Sv)(r|w|M)r+2\displaystyle\leqslant\frac{m-1+\sum_{w\in V(S_{u})}(r-|w|_{M})+\sum_{w\in V(S_{v})}(r-|w|_{M})}{r}+2
m+i=1m(r|ui|M)r+2.\displaystyle\leqslant\frac{m+\sum_{i=1}^{m}(r-|u_{i}|_{M})}{r}+2.

By (3), there is a rainbow path PP consisting of colors in CXC_{X} from uu^{\prime} to vv^{\prime} in G[X]G[X] of length at most dd. Since CXC(M)=C_{X}\cap C(M)=\emptyset, PuPvPP^{u}\cup P^{v}\cup P is a rainbow path consisting of colors in CXC(M)C_{X}\cup C(M) from uu to vv in G[R]G[R] of length at most

m+i=1m(r|ui|M)r+d+2,\frac{m+\sum_{i=1}^{m}(r-|u_{i}|_{M})}{r}+d+2,

as required. This completes the proof of the lemma. ∎

XXu1u_{1}u2u_{2}u3u_{3}u4u_{4}u5u_{5}u6u_{6}u7u_{7}
Figure 1. An example of an edge-colored graph satisfying the hypotheses of Lemma 3.1. Unlabelled vertices are the neighbours of {u1,,u7}\{u_{1},\dots,u_{7}\} in XX. Other vertices of XX and edges with both ends in XX are not depicted.

4. Many Non-Stars

A color class of GG is a star class if its edges form a star. A vertex of GG is a star vertex if it is the centre of a star class, and is otherwise a non-star vertex. Let NN be the set of non-star vertices of GG.

Recall that k=78rlogrk=\lceil 78r\log{r}\rceil, f(r)=80k2r2f(r)=80k^{2}r^{2}, and αr=(4r+2)f(r)+2r\alpha_{r}=(4r+2)f(r)+2r. In this section we consider the case where |N|>f(r)|N|>f(r). We say that a subset NNN^{\prime}\subseteq N covers a color class AA if every edge in AA is incident to at least one vertex in NN^{\prime}.

Claim 4.1.

Each color class is covered by at most 4(|N|k2)4\binom{|N|}{k-2} sets NNN^{\prime}\subseteq N of size kk.

Proof.

First suppose AA is a star class. Then AA is covered by at most

(|N||A|k|A|)(|N|2k2)4(|N|k2)\binom{|N|-|A|}{k-|A|}\leqslant\binom{|N|-2}{k-2}\leqslant 4\binom{|N|}{k-2}

sets NNN^{\prime}\subseteq N of size kk (since |A|2|A|\geqslant 2).

So, suppose instead that AA is a non-star class. If there exists a matching M={u1v1,u2v2}M=\{u_{1}v_{1},u_{2}v_{2}\} of size 22 in AA, then we note that every set of vertices which covers AA must contain either u1u_{1} or v1v_{1}, and either u2u_{2} or v2v_{2}. It follows that AA is covered by at most

4(|N|k2)4\binom{|N|}{k-2}

subsets of size kk in NN, as required.

Finally, suppose that there does not exist a matching of size 2 in AA. Since AA is not a star, it follows that AA is a triangle. Then the number of subsets of NN of size kk that cover AA is at most

3(|N|2k2)4(|N|k2).\displaystyle 3\binom{|N|-2}{k-2}\leq 4\binom{|N|}{k-2}.

This completes the proof. ∎

Lemma 4.2.

If |N|>f(r)|N|>f(r), then GG contains a rainbow cycle of length at most

n+cC(G)(r|cG|)r+αr.\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r}.
Proof.

First suppose that fewer than (|N|k)24k2\frac{(|N|-k)^{2}}{4k^{2}} color classes contain an edge with an end in NN. By Claim 4.1, at most

(|N|k)24k24(|N|k2)<(|N|k)\displaystyle\frac{(|N|-k)^{2}}{4k^{2}}\cdot 4\binom{|N|}{k-2}<\binom{|N|}{k}

subsets of NN of size kk cover some color class. Therefore, there exists a subset NNN^{\prime}\subseteq N of size kk such that NN^{\prime} does not cover any color class. By choosing an edge of each color with neither end in NN^{\prime} we obtain a rainbow subgraph HH of GG with |E(H)|=n|E(H)|=n and |V(H)||V(G)N|=nk|V(H)|\leqslant|V(G)\setminus N^{\prime}|=n-k. In particular, HH is excess-kk. Since r2r\geq 2 and nkn\geq k (otherwise we have a rainbow cycle of length at most k<αrk<\alpha_{r}),  Corollary 2.2 gives that HH has a rainbow cycle of length at most

14(n+k)logk3k\displaystyle\frac{14(n+k)\log k}{3k} 28nlogk3k\displaystyle\leqslant\frac{28n\log{k}}{3k}
28n(log78+logr+loglogr)234rlogr\displaystyle\leqslant\frac{28n(\log{78}+\log{r}+\log{\log{r}})}{234r\log{r}}
28n(log78+2)logr234rlogr\displaystyle\leqslant\frac{28n(\log{78}+2)\log{r}}{234r\log{r}}
nr,\displaystyle\leqslant\frac{n}{r},

as desired.

So, we may assume that at least (|N|k)24k2\frac{(|N|-k)^{2}}{4k^{2}} color classes have at least one edge with an end in NN. Note that |N|k|N|/2|N|-k\geq|N|/2 since |N|>f(r)2k|N|>f(r)\geq 2k. Therefore, by averaging, there exists a vertex vNv\in N with at least |N|16k2>f(r)16k2=5r2\frac{|N|}{16k^{2}}>\frac{f(r)}{16k^{2}}=5r^{2} colors having at least one edge incident with vv. Therefore, there exists a rainbow star TT centred at vv with |V(T)|>5r2.|V(T)|>5r^{2}.

Now, let XX be a maximal set of vertices containing V(T)V(T) such that

  1. (1)

    exactly |X|1|X|-1 colors of GG appear in G[X]G[X],

  2. (2)

    for all u,vXu,v\in X, there exists a rainbow path from uu to vv in G[X]G[X] with length at most

    |X|+cC(X)(r|cG|)r5r+2,\frac{|X|+\sum_{c\in C(X)}(r-|c_{G}|)}{r}-5r+2,

where |cG||c_{G}| is the size of the color class in GG, and C(X)C(X) is the set of colors which appear in G[X]G[X].

Note that V(T)V(T) satisfies the first condition since otherwise G[V(T)]G[V(T)] contains a rainbow triangle and we have obtained a short rainbow cycle. Moreover, V(T)V(T) satisfies the second condition since |V(T)|>5r2|V(T)|>5r^{2} and every color class has size at most rr. Since V(T)V(T) satisfies both conditions and clearly V(T)V(T)V(T)\subseteq V(T), it follows that V(T)V(T) is a valid candidate and therefore XX is a well-defined set.

Now, let u1,,umu_{1},\cdots,u_{m} be a maximal sequence of vertices in GXG\setminus X such that there is a star color class AiA_{i} centred at uiu_{i} with V(Ai){ui}X{u1,,ui1}V(A_{i})\setminus\{u_{i}\}\subseteq X\cup\{u_{1},\dots,u_{i-1}\}. Note that u1,,umu_{1},\dots,u_{m} exist, since we allow m=0m=0. Let T=X{u1,,um}T^{\prime}=X\cup\{u_{1},\cdots,u_{m}\}, cic_{i} denote the color of AiA_{i}, and C(T)=C(X){c1,,cm}C^{\prime}(T^{\prime})=C(X)\cup\{c_{1},\cdots,c_{m}\}. Let d:=|X|+cC(X)(r|cG|)r5r+2d:=\frac{|X|+\sum_{c\in C(X)}(r-|c_{G}|)}{r}-5r+2. By Lemma 3.1, for all u,vTu,v\in T^{\prime} there exists a rainbow path in G[T]G[T^{\prime}] from uu to vv consisting of colors in C(X){c1,,cm}C(X)\cup\{c_{1},\dots,c_{m}\} of length at most

m+i=1m(r|ui|M)r+d+2\displaystyle\frac{m+\sum_{i=1}^{m}(r-\lvert u_{i}\rvert_{M})}{r}+d+2 =m+i=1m(r|ui|M)r+(|X|+cC(X)(r|cG|)r5r+2)+2\displaystyle=\frac{m+\sum_{i=1}^{m}(r-\lvert u_{i}\rvert_{M})}{r}+\left(\frac{|X|+\sum_{c\in C(X)}(r-|c_{G}|)}{r}-5r+2\right)+2
=|T|+cC(T)(r|cG|)r5r+4.\displaystyle=\frac{|T^{\prime}|+\sum_{c\in C^{\prime}(T^{\prime})}(r-|c_{G}|)}{r}-5r+4.

It follows that exactly |X|1+m=|T|1|X|-1+m=|T^{\prime}|-1 colors appear in G[T]G[T^{\prime}]; otherwise, G[T]G[T^{\prime}] contains a rainbow cycle of length at most

|T|+cC(T)(r|cG|)r5r+5n+𝖽𝖾𝖿r(G)r5r+5n+𝖽𝖾𝖿r(G)r.\frac{|T^{\prime}|+\sum_{c\in C^{\prime}(T^{\prime})}(r-|c_{G}|)}{r}-5r+5\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}-5r+5\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}.

Therefore, C(T)=C(T)C^{\prime}(T^{\prime})=C(T^{\prime}). Now, let 𝒜\mathcal{A} be the set of color classes AA such that no edge of AA is in G[T]G[T^{\prime}] and some edge in AA has exactly one end in TT^{\prime}. For each A𝒜A\in\mathcal{A}, choose an edge eAAe_{A}\in A with exactly one end in TT^{\prime}, and let YV(G)TY\subseteq V(G)\setminus T^{\prime} be the set of ends of these edges not in TT^{\prime}. Note that |Y|=|𝒜||Y|=|\mathcal{A}|; otherwise G[TY]G[T^{\prime}\cup Y] contains a rainbow cycle of length at most

|T|+cC(T)(r|cG|)r5r+6n+𝖽𝖾𝖿r(G)r5r+6n+𝖽𝖾𝖿r(G)r.\frac{|T^{\prime}|+\sum_{c\in C(T^{\prime})}(r-|c_{G}|)}{r}-5r+6\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}-5r+6\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}.

Moreover, exactly |TY|1|T^{\prime}\cup Y|-1 colors appear in G[TY]G[T^{\prime}\cup Y]; otherwise G[TY]G[T^{\prime}\cup Y] contains a rainbow cycle of length at most

|T|+cC(T)(r|cG|)r5r+7n+𝖽𝖾𝖿r(G)r5r+7n+𝖽𝖾𝖿r(G)r.\frac{|T^{\prime}|+\sum_{c\in C(T^{\prime})}(r-|c_{G}|)}{r}-5r+7\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}-5r+7\leqslant\frac{n+\mathsf{def}_{r}(G)}{r}.

Note that when adding YY to TT^{\prime} the ‘rainbow diameter’ increases by at most 22, regardless of the size of YY. Therefore, for all u,vTYu,v\in T^{\prime}\cup Y, there exists a rainbow path from uu to vv in G[TY]G[T^{\prime}\cup Y] of length at most

(|T|+cC(T)(r|cG|)r5r+4)+2|TY|+cC(TY)(r|cG|)r5r+6|Y|r.\left(\frac{|T^{\prime}|+\sum_{c\in C(T^{\prime})}(r-|c_{G}|)}{r}-5r+4\right)+2\leqslant\frac{|T^{\prime}\cup Y|+\sum_{c\in C(T^{\prime}\cup Y)}(r-|c_{G}|)}{r}-5r+6-\frac{|Y|}{r}.

It follows that |Y|4r1|Y|\leqslant 4r-1; otherwise TYT^{\prime}\cup Y contradicts the maximality of XX. Now, we contract TT^{\prime} to a single vertex, delete all edges with colors in TT^{\prime}, and remove parallel edges of the same color. Let the resulting edge-colored graph be GG^{\prime}. Note that by the construction of TT^{\prime}, GG^{\prime} contains |V(G)||V(G^{\prime})| color classes of size at least 22, and at most 4r14r-1 color classes have less edges in GG^{\prime} than they do in GG (since |𝒜|4r1|\mathcal{A}|\leqslant 4r-1). By induction on nn, GG^{\prime} has a rainbow cycle DD of length at most

|V(G)|+cC(G)(r|cG|)r+αr|V(G)|+(4r1)(r2)+cC(G)(r|cG|)r+αr.\frac{|V(G^{\prime})|+\sum_{c\in C(G^{\prime})}(r-|c_{G^{\prime}}|)}{r}+\alpha_{r}\leqslant\frac{|V(G^{\prime})|+(4r-1)(r-2)+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r}.

Observe that DD is either a rainbow cycle in GG, or a rainbow path between uu and vv, for some u,vTu,v\in T^{\prime}. If the former holds, then

|V(D)|n+cC(G)(r|cG|)r+αr,|V(D)|\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r},

since |V(G)|+(4r1)(r2)|V(G)|+5r21|V(G)|+|T|1=n|V(G^{\prime})|+(4r-1)(r-2)\leqslant|V(G^{\prime})|+5r^{2}-1\leqslant|V(G^{\prime})|+|T^{\prime}|-1=n.

In the latter case, from our earlier application of Lemma 3.1, we have that there exists a rainbow path PP in G[T]G[T^{\prime}] from uu to vv of length at most

|T|+cC(T)(r|cG|)r5r+4.\frac{|T^{\prime}|+\sum_{c\in C(T^{\prime})}(r-|c_{G}|)}{r}-5r+4.

By the construction of GG^{\prime}, the colors used in PP are disjoint from the colors used in DD. Therefore, DPD\cup P is a rainbow cycle in GG of length at most

|E(D)|+|E(P)|\displaystyle|E(D)|+|E(P)| |V(G)|+|T|+(4r1)(r2)+cC(G)(r|cG|)r5r+4+αr\displaystyle\leqslant\frac{|V(G^{\prime})|+|T^{\prime}|+(4r-1)(r-2)+\sum_{c\in C(G)}(r-|c_{G}|)}{r}-5r+4+\alpha_{r}
=n+1+(4r1)(r2)+cC(G)(r|cG|)r5r+4+αr\displaystyle=\frac{n+1+(4r-1)(r-2)+\sum_{c\in C(G)}(r-|c_{G}|)}{r}-5r+4+\alpha_{r}
n+cC(G)(r|cG|)r+αr,\displaystyle\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r},

since 1+(4r1)(r2)r(5r4)1+(4r-1)(r-2)\leqslant r(5r-4) simplifies to r2+5r3r^{2}+5r\geqslant 3 which is true for r2r\geq 2. This completes the proof of the lemma. ∎

5. Few Non-Stars

Recall from the previous section that NN is the set of non-star vertices of GG. In this section, we complete the proof by handling the case |N|f(r)|N|\leq f(r).

Lemma 5.1.

If |N|f(r)|N|\leq f(r), then GG contains a rainbow cycle of length at most

n+cC(G)(r|cG|)r+αr.\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r}.
Proof.

Let N={v1,v2,,vt}N=\{v_{1},v_{2},\cdots,v_{t}\} be an arbitrary fixed ordering of the vertices of NN. Let U=V(G)NU=V(G)\setminus N and choose a galaxy M:=(Mu:uU)M:=(M_{u}:u\in U) rooted at UU by arbitrarily choosing a star class MuM_{u} of color cuc_{u} centred at uu for each uUu\in U. An edge eE(G)e\in E(G) is an MM-edge if eE(Mu)e\in E(M_{u}) for some uUu\in U. For YV(G)Y\subseteq V(G) we define γ(Y)\gamma(Y) to be the set of star vertices uV(G)Yu\in V(G)\setminus Y such that at least one MM-neighbor of uu is in YY.

For each XV(G)X\subseteq V(G), let CM(X)C_{M}(X) be the set of MM-colors which have an edge with both ends in XX. Now, we iteratively construct pairwise-disjoint sets of vertices T1,T2,,TtT_{1},T_{2},\cdots,T_{t} of GG and associated constants d1,d2,,dtd_{1},d_{2},\cdots,d_{t} such that for all j[t]j\in[t], we have the following four properties:

  1. (1)

    TjN={vj}T_{j}\cap N=\{v_{j}\}.

  2. (2)

    For each star vertex vi=1jTjv\notin\bigcup_{i=1}^{j}T_{j}, vv has at least one MM-neighbor which is not in i=1jTj\bigcup_{i=1}^{j}T_{j}.

  3. (3)

    For each pair of vertices u,vTju,v\in T_{j}, there exists a rainbow path of MM-edges between uu and vv in G[Tj]G[T_{j}] of length at most

    |Tj|+cCM(Tj)(r|cG|)r+dj.\frac{|T_{j}|+\sum_{c\in C_{M}(T_{j})}(r-|c_{G}|)}{r}+d_{j}.

  4. (4)

    i=1jdi+|γ(i=1jTi)|(4r+1)j\sum_{i=1}^{j}d_{i}+|\gamma(\bigcup_{i=1}^{j}T_{i})|\leqslant(4r+1)j.

Fix j[t]j\in[t], and suppose we have already constructed T1,,Tj1T_{1},\cdots,T_{j-1} with the associated constants d1,,dj1d_{1},\cdots,d_{j-1} such that they satisfy the above four properties. Let Gj:=Gi=1j1TiG_{j}:=G\setminus\bigcup_{i=1}^{j-1}T_{i}. For XV(Gj)X\subseteq V(G_{j}), let s(X)s(X) be the set of star vertices xXx\in X such that at least one MM-neighbor of xx is in i=1j1Ti\bigcup_{i=1}^{j-1}T_{i}. Let XX be a maximal set of vertices of GjG_{j} such that XN={vj}X\cap N=\{v_{j}\}, and for each pair of vertices u,vXu,v\in X, there exists a rainbow path of MM-edges between uu and vv in G[X]G[X] of length at most

d:=|X|+cCM(X)(r|cG|)r+|s(X)|.d:=\frac{|X|+\sum_{c\in C_{M}(X)}(r-|c_{G}|)}{r}+|s(X)|.

Note that XX exists since {vj}\{v_{j}\} is a candidate for XX. Let u1,,umu_{1},\dots,u_{m} be a maximal sequence of vertices in Gj(XN)G_{j}\setminus(X\cup N) such that for all i[m]i\in[m], all the MM-neighbors of uiu_{i} are in i=1j1TiX{u1,,ui1}\bigcup_{i=1}^{j-1}T_{i}\cup X\cup\{u_{1},\dots,u_{i-1}\}. Define Tj:=X{u1,,um}T_{j}:=X\cup\{u_{1},\dots,u_{m}\}, and dj:=|s(Tj)|+2d_{j}:=|s(T_{j})|+2.

Clearly, TjT_{j} satisfies Item 1 and Item 2 (by the maximality of u1,,umu_{1},\dots,u_{m}). Let C:={cui:i[m]}C^{\prime}:=\{c_{u_{i}}:i\in[m]\}. By Lemma 3.1, for any pair of vertices u,vTju,v\in T_{j}, there is a rainbow path of MM-edges from uu to vv in G[Tj]G[T_{j}] of length at most

m+cC(r|cG|)r+d+2\displaystyle\frac{m+\sum_{c\in C^{\prime}}(r-|c_{G}|)}{r}+d+2 =m+cC(r|cG|)r+(|X|+cCM(X)(r|cG|)r+|s(X)|)+2\displaystyle=\frac{m+\sum_{c\in C^{\prime}}(r-|c_{G}|)}{r}+\left(\frac{|X|+\sum_{c\in C_{M}(X)}(r-|c_{G}|)}{r}+|s(X)|\right)+2
|Tj|+cCM(Tj)(r|cG|)r+|s(Tj)|+2\displaystyle\leq\frac{|T_{j}|+\sum_{c\in C_{M}(T_{j})}(r-|c_{G}|)}{r}+|s(T_{j})|+2
=|Tj|+cCM(Tj)(r|cG|)r+dj,\displaystyle=\frac{|T_{j}|+\sum_{c\in C_{M}(T_{j})}(r-|c_{G}|)}{r}+d_{j},

since |s(X)||s(Tj)||s(X)|\leq|s(T_{j})|. It follows that Item 3 is satisfied. Finally, we show that Item 4 holds.

Claim 5.2.

i=1jdi+|γ(i=1jTi)|(4r+1)j\sum_{i=1}^{j}d_{i}+|\gamma(\bigcup_{i=1}^{j}T_{i})|\leqslant(4r+1)j.

Proof.

Let

Y:={ui=1jTi:at least one M-neighbor of u is in Tj}.Y:=\{u\notin\bigcup_{i=1}^{j}T_{i}:\text{at least one $M$-neighbor of $u$ is in $T_{j}$}\}.

Note that if we add YY to TjT_{j}, then the ‘rainbow diameter’ increases by at most 22, regardless of the size of YY. Thus, |Y|4r1|Y|\leqslant 4r-1; otherwise, YTjY\cup T_{j} contradicts the maximality of XX (see Lemma 4.2 for an analogous argument with more detail). Moreover, by definition, s(Tj)γ(i=1j1Ti)s(T_{j})\subseteq\gamma(\bigcup_{i=1}^{j-1}T_{i}) and γ(i=1jTi)Y(γ(i=1j1Ti)s(Tj))\gamma(\bigcup_{i=1}^{j}T_{i})\subseteq Y\cup(\gamma(\bigcup_{i=1}^{j-1}T_{i})\setminus s(T_{j})). Thus, |γ(i=1jTi)||Y|+|γ(i=1j1Ti)||s(Tj)||\gamma(\bigcup_{i=1}^{j}T_{i})|\leqslant|Y|+|\gamma(\bigcup_{i=1}^{j-1}T_{i})|-|s(T_{j})|. It follows that,

i=1jdi+|γ(i=1jTi)|\displaystyle\sum_{i=1}^{j}d_{i}+|\gamma(\bigcup_{i=1}^{j}T_{i})| dj+|Y||s(Tj)|+i=1j1di+|γ(i=1j1Ti)|\displaystyle\leqslant d_{j}+|Y|-|s(T_{j})|+\sum_{i=1}^{j-1}d_{i}+|\gamma(\bigcup_{i=1}^{j-1}T_{i})|
=2+|Y|+i=1j1di+|γ(i=1j1Ti)|\displaystyle=2+|Y|+\sum_{i=1}^{j-1}d_{i}+|\gamma(\bigcup_{i=1}^{j-1}T_{i})|
(4r+1)+i=1j1di+|γ(i=1j1Ti)|\displaystyle\leqslant(4r+1)+\sum_{i=1}^{j-1}d_{i}+|\gamma(\bigcup_{i=1}^{j-1}T_{i})|
(4r+1)+(4r+1)(j1)\displaystyle\leqslant(4r+1)+(4r+1)(j-1)
=(4r+1)j.\displaystyle=(4r+1)j.\qed

To finish, we have two cases. First, suppose that V(G)=i=1tTiV(G)=\bigcup_{i=1}^{t}T_{i}. Let AA be the set of edges of GG which are not MM-edges. Note that exactly |N||N| colors appear in AA. If some eAe\in A has both ends in some TiT_{i}, then we obtain a rainbow cycle of length at most

1+|Ti|+cCM(Ti)(r|cG|)r+di\displaystyle 1+\frac{|T_{i}|+\sum_{c\in C_{M}(T_{i})}(r-|c_{G}|)}{r}+d_{i} 1+n+cC(G)(r|cG|)r+(4r+1)f(r)\displaystyle\leqslant 1+\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+(4r+1)f(r)
n+cC(G)(r|cG|)r+(4r+2)f(r)\displaystyle\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+(4r+2)f(r)
n+cC(G)(r|cG|)r+αr.\displaystyle\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r}.

So we may assume that every edge in AA has ends in different TiT_{i}. Now, for each of the tt colors that appear in AA, keep one edge of that color, delete all other edges in GG, and contract each TiT_{i} to a single vertex to obtain GG^{\prime}. Since GG^{\prime} has tt vertices with tt edges, there is a cycle CC in GG^{\prime} of length at most tt. Now, consider the edges of CC in the original graph GG. Let the edges be xiyi+1x_{i}y_{i+1} for i/ti\in\mathbb{Z}/t\mathbb{Z}, such that xix_{i} and yiy_{i} are contained in the same set TjiT_{j_{i}} in {T1,,Tt}\{T_{1},\cdots,T_{t}\}.

By Item 3, there is a rainbow path PiP_{i} of MM-colors between xix_{i} and yiy_{i} contained in TjiT_{j_{i}} of length at most |Tji|+dji|T_{j_{i}}|+d_{j_{i}}. Joining the edges xiyi+1x_{i}y_{i+1} with the paths {Pi}\{P_{i}\} yields a rainbow cycle of length at most

n+cCM(G)(r|cG|)r+t+i=1tdi\displaystyle\frac{n+\sum_{c\in C_{M}(G)}(r-|c_{G}|)}{r}+t+\sum_{i=1}^{t}d_{i} n+cC(G)(r|cG|)r+f(r)+(4r+1)f(r)\displaystyle\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+f(r)+(4r+1)f(r)
n+cC(G)(r|cG|)r+αr,\displaystyle\leqslant\frac{n+\sum_{c\in C(G)}(r-|c_{G}|)}{r}+\alpha_{r},

as desired.

The remaining case is V(G)i=1tTiV(G)\neq\bigcup_{i=1}^{t}T_{i}. Let W=V(G)(i=1tTi)W=V(G)\setminus\left(\bigcup_{i=1}^{t}T_{i}\right). Let DD be the digraph with vertex set WW and for each vWv\in W and each MM-neighbor uu of vv in WW, put an arc from vv to uu in DD. By Item 2, every vertex vWv\in W has at least one of its MM-neighbors in WW. Therefore, DD has no sink. By Item 4, at most (4r+1)f(r)(4r+1)f(r) vertices in WW have an MM-neighbor outside of WW. Therefore, 𝖽𝖾𝖿r(D)(r1)(4r+1)f(r)\mathsf{def}_{r}(D)\leqslant(r-1)(4r+1)f(r). We now apply Corollary 2.4 in DD to obtain a directed cycle CC in DD of length at most

|W|+(r1)(4r+1)f(r)r+2r1\displaystyle\frac{|W|+(r-1)(4r+1)f(r)}{r}+2r-1 nr+αr.\displaystyle\leqslant\frac{n}{r}+\alpha_{r}.

The corresponding edges of CC in GG form a rainbow cycle, which completes the proof of the lemma. ∎

Lemma 4.2 and Lemma 5.1 together imply Theorem 1.7 and Corollary 1.8, as desired.

Acknowledgements

The second author is grateful to the Structural Graph Theory Downunder II workshop at the Mathematical Research Institute MATRIX (March 2022), for providing an ideal environment to work on this problem. The second author also thanks Katie Clinch, Jackson Goerner, and Freddie Illingworth for very stimulating discussions.

References

Index