Rainbow matchings and rainbow connectedness

Aharoni and Berger conjectured that every bipartite graph which is the union of n matchings of size n + 1 contains a rainbow matching of size n. This conjecture is a generalization of several old conjectures of Ryser, Brualdi, and Stein about transversals in Latin squares. There have been many recent partial results about the Aharoni-Berger Conjecture. In the case when the matchings are much larger than n + 1, the best bound is currently due to Clemens and Ehrenm\"uller who proved the conjecture when the matchings are of size at least 3n/2 + o(n). When the matchings are all edge-disjoint and perfect, then the best result follows from a theorem of H\"aggkvist and Johansson which implies the conjecture when the matchings have size at least n + o(n). In this paper we show that the conjecture is true when the matchings have size n + o(n) and are all edge-disjoint (but not necessarily perfect). We also give an alternative argument to prove the conjecture when the matchings have size at least $\phi n + o(n)$ where $\phi\approx 1.618$ is the Golden Ratio. Our proofs involve studying connectedness in coloured, directed graphs. The notion of connectedness that we introduce is new, and perhaps of independent interest.


Introduction
A Latin square of order n is an n × n array filled with n different symbols, where no symbol appears in the same row or column more than once. Latin squares arise in different branches of mathematics such as algebra (where Latin squares are exactly the multiplication tables of quasigroups) and experimental design (where they give rise to designs called Latin square designs). They also occur in recreational mathematics-for example completed Sudoku puzzles are Latin squares.
In this paper we will look for transversals in Latin squares. A transversal in a Latin square of order n is a set of n entries such that no two entries are in the same row, same column, or have the same symbol. One reason transversals in Latin squares are interesting is that a Latin square has an orthogonal mate if, and only if, it has a decomposition into disjoint transversals. See [15] for a survey about transversals in Latin squares. It is easy to see that not every Latin square has a transversal (for example the unique 2 × 2 Latin square has no transversal). However, it is possible that every Latin square contains a large partial transversal. Here, a partial transversal of size m means a set of m entries such that no two entries are in the same row, same column, or have the same symbol.
There are several closely related, old, and difficult conjectures which say that Latin squares should have large partial transversals. The first of these is a conjecture of Ryser that every Latin square of odd order contains a transversal [13]. Brualdi and Stein conjectured that every Latin square contains a partial transversal of size n − 1.
There have been many partial results about this conjecture. It is known that every Latin square has a partial transversal of size n − o(n)-Woolbright [16] and independently Brower, de Vries and Wieringa [6] proved that every Latin square contains a partial transversal of size n − √ n. This has been improved by Hatami and Schor [11] to n − O(log 2 n). Häggkvist and Johansson proved a related result about Latin rectangles. For m n a m × n Latin rectangle is an m × n array of n symbols where no symbol appears in the same row or column more than once. A transversal in a Latin rectangle is a set of m entries no two of which are in the same row, column, or have the same symbol. Häggkvist and Johansson proved the following. Theorem 1.2 (Häggkvist and Johansson, [10]). For every , there is an m 0 = m 0 ( ) such that the following holds. For every n (1 + )m m 0 , every m × n Latin rectangle can be decomposed into disjoint transversals. This theorem is proved by a probabilistic argument, using a "random greedy process" to construct the transversals. The above theorem gives yet another proof that every sufficiently large n × n Latin square has a partial transversal of size n − o(n)-indeed if we remove n rows of a Latin square we obtain a Latin rectangle to which Theorem 1.2 can be applied.
In this paper we will look at a strengthening of Conjecture 1.1. The strengthening we'll look at is a conjecture due to Aharoni and Berger which takes place in a more general setting than Latin squares-namely edge coloured bipartite graphs. To see how the two settings are related, notice that there is a one-to-one correspondence between n × n Latin squares and proper edge colourings of K n,n with n colours-indeed to a Latin square S we associate the colouring of K n,n with vertex set {x 1 , . . . , x n , y 1 , . . . , y n } where the edge between x i and y j receives colour S i,j . Notice that this colouring is proper i.e. adjacent edges receive different colours. Recall that a matching in a graph is a set of disjoint edges. We call a matching rainbow if all of its edges have different colours. It is easy to see that partial transversals in the Latin square S correspond to rainbow matchings in the corresponding coloured K n,n . Thus Conjecture 1.1 is equivalent to the statement that "in any proper n-colouring of K n,n , there is a rainbow matching of size n − 1." One could ask whether a large rainbow matching exists in other settings. Recall that a simple graph is a graph which contains at most one edge between any pair of vertices. A multigraph is a graph which may contain multiple edges between vertices. Aharoni and Berger posed the following conjecture, which generalises Conjecture 1.1.

Conjecture 1.3 (Aharoni and
Berger, [1]). Let G be a multigraph, properly edge coloured by n colours, with at least n+1 edges of each colour. Then G contains a rainbow matching with n edges.
This conjecture was first posed in a different form in [1] as a conjecture about matchings in tripartite hypergraphs (Conjecture 2.4 in [1]). It was first stated as a conjecture about rainbow matchings in [2].
The above conjecture has attracted a lot of attention recently, and there are many partial results. Just like in Conjecture 1.1, one natural way of attacking Conjecture 1.3 is to prove approximate versions of it. As observed by Barat, Gyárfás, and Sárközy [4], the arguments that Woolbright, Brower, de Vries, and Wieringa used to find partial transversals of size size n − √ n in Latin squares actually generalise to bipartite graphs to give the following. Theorem 1.4 (Woolbright, [16]; Brower, de Vries, and Wieringa, [6]; Barat, Gyárfás, and Sárközy, [4]). Let G be a bipartite multigraph, properly edge coloured by n colours, with at least n edges of each colour. Then G contains a rainbow matching with n − √ n edges.
Barat, Gyárfás, and Sárközy actually proved something a bit more general in [4]-for every k, they gave an upper bound on the number of colours needed to find a rainbow matching of size n − k.
Another approximate version of Conjecture 1.3 comes from Theorem 1.2. It is easy to see that Theorem 1.2 is equivalent to the following "let G be a bipartite graph consisting of n edge-disjoint perfect matchings, each with at least n + o(n) edges. Then G can be decomposed into disjoint rainbow matchings of size n" (to see that this is equivalent to Theorem 1.2, associate an m-edge coloured bipartite graph with any m×n Latin rectangle by placing a colour k edge between i and j whenever (k, i) has symbol j in the rectangle).
The main result of this paper is an approximate version of Conjecture 1.3 in the case when the matchings in G are disjoint, but not necessarily perfect. Theorem 1.5. For all > 0, there exists an N = N ( ) = 10 20 −16 −1 such that the following holds. Let G be a bipartite simple graph, properly edge coloured by n N colours, with at least (1 + )n edges of each colour. Then G contains a rainbow matching with n edges. Unlike the proof of Theorem 1.2 which can be used to give a randomised process to find a rainbow matching, the proof of Theorem 1.5 is algorithmic. In fact, it can be shown that the matching guaranteed by Theorem 1.5 can be found in polynomial time.
Another very natural approach to Conjecture 1.3 is to to prove it when the matchings have size much larger than n + 1. When the matchings have size 2n, the result becomes trivial. Lemma 1.6. Let G be a multigraph, properly edge coloured with n colours, each with at least 2n edges of each colour. Then G contains a rainbow matching with n edges. This lemma is proved by greedily choosing disjoint edges of different colours. We can always choose n edges this way, since each colour class has 2n edges (one of which must be disjoint from previously chosen edges).
There have been several improvements to the 2n bound in Lemma 1.6. Aharoni, Charbit, and Howard [2] proved that matchings of size 7n/4 are sufficient to guarantee a rainbow matching of size n. Kotlar and Ziv [12] improved this to 5n/3 . Clemens and Ehrenmüller [8] further improved the constant to 3n/2 + o(n). Finally, Aharoni, Kotlar, and Ziv improved the o(n) term to give the following. Theorem 1.7 (Aharoni, Kotlar, and Ziv, [3]). Let G be a multigraph, properly coloured with n colours, with at least 3n/2 + 1 edges of each colour. Then G contains a rainbow matching with n edges.
Though we won't improve on this theorem, we give an alternative proof which gives a weaker bound of φn + o(n) where φ ≈ 1.618 is the Golden Ratio. Theorem 1.8. Let G be a bipartite graph, properly coloured by n colours, with at least φn + 20n/ log n edges of each colour. Then G contains a rainbow matching with n edges. Theorems 1.5 and 1.8 are proved by studying paths in auxiliary directed graphs. This approach is new and the results we prove about directed graphs may be of independent interest. In particular, we introduce a new notion of connectivity, which we call "rainbow k-connectivity". In the next section we give an informal sketch of the proof of Theorem 1.5. In Section 2 we prove the results about a number of lemmas about directed graphs which we will need. In Section 3 we prove Theorem 1.5. In Section 4 we prove Theorem 1.8. In Section 5 we make some concluding remarks about the techniques used in this paper. For all standard notation we follow [5].

Sketch of proofs
In this section we informally present the main ideas in our proof of Theorem 1.5. Let G be a bipartite simple graph with bipartition classes X and Y properly coloured by n colours and with at least (1 + )n edges of each colour. Let M be a maximum rainbow matching in G. Let X 0 = X \ V (M ) and Y 0 = Y \ V (M ). If G doesn't satisfy the conclusion of Theorem 1.5, then there is some colour, c * , which is not present in M . Previous approaches to the Aharoni-Berger Conjecture [2,3,8,12] revolved around trying the electronic journal of combinatorics 24(1) (2017), #P1.13 to perform local manipulations on M until it can be extended to a larger matching. These local manipulations have a similar flavour to "alternating paths" which are often used to study matchings in uncoloured graphs. The following definition is a special case of the kinds of manipulations we will use in the full proof of Theorem 1.5. Definition 1.9. An X 0 -switching σ is a sequence of distinct edges σ = (e 0 , m 1 , e 1 , . . . , m −1 , e −1 , m ) such that m i ∈ M , e i goes between X 0 and m i+1 ∩ Y , and e i and m i have the same colour.
If e 0 has colour c 0 and m has colour c , then we say that σ is an X 0 -switching from c 0 to c . If σ is an X 0 -switching from the colour c * which is not present in M to some other colour c , then it is an easy exercise to show that the following is another rainbow matching in G.
In Section 3 we prove Lemma 3.3 which is a generalization of the statement of M σ being a rainbow matching. Furthermore M σ is a rainbow matching of the same size as M and missing c , the colour of m . Thus X 0 -switchings can be used to go between maximum rainbow matchings with different missing colours. At a very high level, our proof can be summarized as "we look for switchings in a graph until we find one which allows us to find a matching bigger than M ." The key idea of this paper is that switchings can be studied using auxiliary directed graphs. The following is a special case of the directed graphs which we will use to study switchings.
Definition 1.10. The directed graph D X 0 is defined as follows.
• The vertex set of D X 0 is the set of colours of edges in G.
• For two colours u and v ∈ V (D X 0 ), there is a directed edge from u to v in D X 0 whenever there is an x ∈ X 0 such that there is a colour u edge from x to the vertex m v ∩ Y in G, where m v is the colour v edge of M . In this case uv is coloured by "x".
What does the above definition have to do with X 0 -switchings? It is an easy exercise to show that X 0 -switchings are in one-to-one corrospondance with rainbow paths in the directed graph D X 0 . Specifically, given an X 0 -switching σ = (e 0 , m 1 , e 1 , m 2 , e 2 , . . . , m −1 , e −1 , m ) from c to c , then the set of edges of D X 0 corresponding to e 0 , e 1 , . . . , e −1 form a rainbow path from c to c . This correspondence is made precise in Lemma 3.6 in Section 3. Thus studying X 0 -switchings is exactly the same as studying rainbow paths in D X 0 . Since rainbow paths are more familiar objects that X 0 -switchings, this opens up more powerful techniques to study them. For two vertices u, v ∈ D X 0 , we use d R (u, v) to denote the length of the shortest rainbow path from c * to v The first observation one makes about D X 0 is that any vertex v ∈ D X 0 must have . This is a special case of Lemma 3.7 which we prove in Section 3.
the electronic journal of combinatorics 24(1) (2017), #P1.13 The idea behind it's proof is to consider an X 0 -switching σ corresponding to a minimal length path from c * to v. If d + (v) n − d R (c * , v) didn't hold, then it is possible to show that a colour v edge between X 0 and Y 0 can be added to M σ in order to get a larger matching, contradicting the maximality of the original matching M .
Thus we have that vertices close to c * in D X 0 have linear degree. A large part of the proof of Theorem 1.5 involves showing that such graphs have a large highly connected subgraph. But how do we define connectivity of coloured graphs? The following definition is new, and perhaps of independent interest. Definition 1.11. An edge coloured directed graph G is said to be rainbow k-edge-connected if for any set S of at most k −1 colours and any pair of vertices u and v, there is a rainbow u to v path whose edges have no colours from S.
The above definition differs from usual notions of connectivity, since generally the avoided set S is a set of edges rather than colours.
The key intermediate result we prove is that every properly edge coloured directed graph D has a rainbow k-edge-connected subset C of size roughly δ + (D). Lemma 2.10 will make this precise. Since every vertex in , this can be used to get a k-edge-connected subset C of size roughly n − o(n) where the "o(n)" term depends on k. How does such a highly connected subset help in the original problem of finding a matching in G? Recall that the vertices of D X 0 correspond to colours in G.

In particular all vertices
Now X 1 is a subset of X which has some kind of "flexibility" property which comes from the rainbow k-connectedness of C. In particular, given any small set S of vertices in X 1 it is possible to find an X 0 -switching σ corresponding to some rainbow path in C, such that M σ is disjoint from S. The precise notion of flexibility which we obtain is given in Definition 3.2. Though far from obvious, it turns out that all the arguments in this section can be repeated with X 0 replaced by X 1 . This is made precise in Lemma 3.8. Repeating all the arguments with X 1 instead of X 0 gives a new set X 2 containing X 1 , which still has a degree of "flexibility". This process cannot go on forever since at each iteration X i is roughly n bigger than X i−1 . The only way the process can terminate is by finding a matching larger than M in the step during which we established that vertices in D X 0 have large out-degree. Thus, if the various parameters are chosen suitably, then we obtain that the original matching M must have used every colour i.e. Theorem 1.5.

Paths in directed and coloured graphs
In this section we prove results about paths in various types of directed graphs. All graphs in this section have no multiple edges, although we allow the same edge to appear twice in opposite directions. In directed graphs, "path" will always mean a sequence of vertices x 1 , x 2 , . . . , x k such that x i x i+1 is a directed edge for i = 1, . . . , k−1. The vertices x 1 and x k are called the endpoints of the path, and x 2 , . . . , x k−1 are called the internal vertices of the path. We say that a path P internally avoids a set of vertices S if {x 2 , . . . , x k−1 } ∩ S = ∅.
the electronic journal of combinatorics 24(1) (2017), #P1. 13 We will use additive notation for concatenating paths-for two paths P = p 1 . . . p i and Q = q 1 . . . q j , P + Q denotes the path with vertex sequence p 1 . . . p i q 1 . . . q j . Let N + G (v) denote the out-neighbourhood of a vertex v in a graph G i.e. the set of vertices x for which vx is an edge in G. When the identity of G is clear from the context, we will abbreviate this to N + (v). Let d G (u, v) denote the length of the shortest part from u to v in a graph G.
We will look at coloured graphs. An edge colouring of a graph is an assignment of colours to the edges of a graph. A total colouring is an assignment of colours to both the vertices and edges of a graph. For any coloured graph we denote by c(v) and c(uv) the colour assigned to a vertex or edge respectively.
An edge colouring is out-proper if for any vertex v, the outgoing edges from v all have different colours. Similarly an edge colouring is in-proper if for any vertex v, the ingoing edges from v all have different colours. We say that an edge colouring is proper if it is both in and out-proper (notice that by this definition it is possible to have two edges with the same colour at a vertex v-as long as one of the edges is oriented away from v and one is oriented towards v). A total colouring is proper if the underlying edge colouring and vertex colourings are proper and the colour of any vertex is different from the colour of any edge containing it. A totally coloured graph is rainbow if all its vertices and edges have different colours. For two vertices u and v in a coloured graph, d R (u, v) denotes the length of the shortest rainbow path from u to v. We say that a graph is vertex-rainbow if all its vertices have different colours.
This section will mostly be about finding highly connected subsets in directed graphs. The following is the notion of connectivity that we will use.
k − 1 and any vertices x, y ∈ A \ S, there is an x to y path of length d in D avoiding S.
Notice that a directed graph D is strongly k-connected if, and only if, V (D) is (k, ∞)connected in D. Also notice that it is possible for a subset A ⊆ V (D) to be highly connected without the induced subgraph D[A] being highly connected-indeed if D is a bipartite graph with classes X and Y where all edges between X and Y are present in both directions, then X is a (|Y |, 2)-connected subset of D, although the induced subgraph on X has no edges.
We will also need a generalization this notion of connectivity to coloured graphs Definition 2.2. Let A be a set of vertices in a coloured digraph D. We say that A is (k, d)-rainbow connected in D if, for any set of at most k − 1 colours S and any vertices x, y ∈ A, there is a rainbow x to y path of length at most d in D internally avoiding colours in S.
Notice that in the above definition, we did not specify whether the colouring was a edge colouring, vertex colouring, or total colouring. The definition makes sense in all three cases. For edge colourings a path P "internally avoiding colours in S" means P not having edges having colours in S. For vertex colourings a path P "internally avoiding the electronic journal of combinatorics 24(1) (2017), #P1.13 colours in S" means P not having vertices having colours in S (except possibly for the vertices x and y). For total colourings a path P "internally avoiding colours in S" means P having no edges or vertices with colours in S (except possibly for the vertices x and y).
Comparing the above definition to "rainbow k-edge-connectedness" defined in the previous section we see that an edge coloured graph is rainbow k-connected exactly when it is (k, ∞)-rainbow connected.
We'll need the following lemma which could be seen as a weak analogue of Menger's Theorem. It will allow us to find rainbow paths through prescribed vertices in a highly connected set. Lemma 2.3. Let D be a totally coloured digraph and A a (3kd, d)-rainbow connected subset of D. Let S be a set of colours with |S| k and a 1 , . . . , a k be vertices in A such that no a i has a colour from S and a 1 , . . . , a k all have different colours.
Then there is a rainbow path P from a 1 to a k of length at most kd which passes through each of a 1 , . . . , a k and avoids S Proof. Using the definition of (3kd, d)-rainbow connected, there is a rainbow path P 1 from a 1 to a 2 of length d avoiding colours in S. Similarly for i k, since the total number of colours in S together with P 1 + · · · + P i−1 is k + 2d(i − 1) + 1 3kd there is a rainbow path P i from a i to a i+1 of length d internally avoiding the k colours in S and the 2d(i − 1) + 1 colours in P 1 + · · · + P i−1 . Joining the paths P 1 , . . . , P k−1 gives the required path.
To every coloured directed graph we associate an uncoloured directed graph where two vertices are joined whenever they have a lot of short paths between them.
Definition 2.4. Let D be a totally coloured digraph and m ∈ N. We define an uncoloured directed graph D m as follows. The vertex set of D m is V (D), and xy is an edge of D m whenever there are m internally vertex disjoint paths P 1 , . . . , P m , each of length 2 and going from x to y such that P 1 ∪ · · · ∪ P m is rainbow.
It turns out that for properly coloured directed graphs D, the uncoloured graph D m has almost the same minimum degree as D. The following lemma will allow us to study short rainbow paths in coloured graphs by first proving a result about short paths in uncoloured graphs.
Lemma 2.5. For all > 0 and m ∈ N, there is an N = N ( , m) = (5m + 4)/ 2 such that the following holds. Let D be a properly totally coloured vertex-rainbow directed graph on at least N vertices. Then we have For w ∈ V (D), we define r v (w) = #rainbow paths of length 2 from v to w.
The following claim shows that W is large.
the electronic journal of combinatorics 24(1) (2017), #P1.13 Proof. For any u ∈ N + D (v) we let ux or x has the same colour as v or vu}.
Since D is properly coloured, and all the vertices in D have different colours, we have that |{x ∈ N + (u) : ux or x has the same colour as v or vu}| 4. This implies that Notice that for a vertex x, we have x ∈ N (u) if, and only if, the path vux is rainbow. Indeed vu has a different colour from v and u since the colouring is proper. Similarly ux has a different colour from u and x. Finally ux and x have different colours from v and vu by the definition of N (u). Therefore .
The following claim shows that W is contained in N + Dm (v). Claim 2.7. If w ∈ W , then we have vw ∈ E(D m ).
Proof. From the definition of W , we have 5m distinct rainbow paths P 1 , . . . , P 5m from v to w of length 2. Consider an auxiliary graph G with V (G) = {P 1 , . . . , P 5m } and P i P j ∈ E(G) whenever P i ∪ P j is rainbow.
We claim that δ(G) 5m − 4. Indeed if for i = j we have P i = vxw and P j = vyw, then, using the fact that the colouring on D is proper and the vertex-rainbow, it is easy to see that the only way P i ∪ P j could not be rainbow is if one of the following holds: Thus if P i = vxw had five non-neighbours vy 1 w, . . . , vy 5 w in G, then by the Pigeonhole Principle for two distinct j and k we would have one of c(y j w) = c(y k w), c(y j ) = c(y k ), the electronic journal of combinatorics 24(1) (2017), #P1.13 or c(vy j ) = c(vy k ). But none of these can occur for distinct paths vy j w and vy k w since the colouring on D is proper and the vertex-rainbow. Therefore δ(G) 5m − 4 holds. It is easy to see that G has a clique of size at least |V (G)|/5 = m (for example by Turán's Theorem, or Brooks' Theorem.) The union of the paths in this clique is rainbow, showing that vw ∈ E(D m ).
Since v was arbitrary, this implies the lemma.
The following lemma shows that every directed graph with high minimum degree contains a large, highly connected subset.
Proof. We start with the following claim.
• Otherwise we stop with M = i.
We will show that thatÃ = A M satisfies the conditions of the claim. Notice that by the construction of A M , it certainly satisfies the first condition. Thus we just need to show This follows from the fact that the sets B 0 , . . . , LetÃ be the set given by the above claim.
We claim that A satisfies the conditions of the lemma.
To show that |A| δ + (D) − |D|, notice that we have the electronic journal of combinatorics 24(1) (2017), #P1.13 The first inequality come from bounding |N − (v) ∩Ã| by 2 |D| for v ∈ A and by |Ã| for v ∈ A. The second inequality follows from the second property ofÃ in Claim 2.9.
Rearranging we obtain Now, we show that A is (k, d)-connected in D. As in Definition 2.1, let S be a subset of V (D) with |S| k − 1 and let x, y be two vertices in A \ S. We will find a path of length d from x to y inÃ \ S. Notice that since |D| 32k −2 , we have |S| 2 |D|/32.
The following is a generalization of the previous lemma to coloured graphs. This is the main intermediate lemma we need in the proof of Theorem 1.5.
As in Definition 2.2, let S be a set of k colours and x, y ∈ A. Let S V be the vertices of D with colours from S. Since D is vertex-rainbow, Using the property of D m , for each edge uv ∈ P , there are at least m choices for a triple of three distinct colours (c 1 , c 2 , c 3 ) and a vertex y(uv) such that there is a path the electronic journal of combinatorics 24(1) (2017), #P1.13 uy(uv)v with c(uy(uv)) = c 1 , c(y(uv)) = c 2 , and c(y(uv)v) = c 3 . Since m 9d + 3k 6|E(P )| + 3|V (P )| + 3|S|, we can choose such a triple for every edge uv ∈ P such that for two distinct edges in P , the triples assigned to them are disjoint, and also distinct from the colours in S and colours of vertices of P .
Let the vertex sequence of P be u, x 1 , x 2 , . . . , x p , v. The following sequence of vertices is a rainbow path from u to v of length 2|P | d internally avoiding colours in S P = u, y(ux 1 ), To show that P is a rainbow path we must show that all its vertices and edges have different colours. The vertices all have different colours since the vertices in D all had different colours. The edges of P all have different colours from each other and the vertices of P by our choice of the vertices y(x i x i+1 ) and the triples of colours associated with them.
We'll need the following simple lemma which says that for any vertex v there is a set of vertices N t 0 close to v with few edges going outside N t 0 .
Lemma 2.11. Suppose we have > 0 and D a totally coloured directed graph. Let v be a vertex in D and for t ∈ N, A corollary of the above lemma is that for any vertex v in a properly coloured directed graph, there is a subgraph of D close to v which has reasonably large minimum out-degree. Proof. Apply Lemma 2.11 to D in order to obtain a number t 0 −1 with |N t 0 +1 (v)| |N t 0 (v)| + |D|. We claim that the set N = N t 0 (v) satisfies the conditions of the lemma.
Suppose, for the sake of contradiction that there is a vertex Notice that since the colouring on D is proper and all vertices in D have different colours, the path P + y is rainbow for all except at most 2|P | of the vertices y ∈ N + (x). Therefore we have |N The goal of this section is to prove an approximate version of Conjecture 1.3 in the case when all the matchings in G are disjoint. The proof will involve considering auxiliary directed graphs to which Lemmas 2.10 and 2.12 will be applied. We begin this section by proving a series of lemmas (Lemmas 3.3 -3.8) about bipartite graphs consisting of a union of n 0 matchings. The set-up for these lemmas will always be the same, and so we state it in the next paragraph to avoid rewriting it in the statement of every lemma.
We will always have bipartite graph called "G" with bipartition classes X and Y consisting of n+1 edge-disjoint matchings M 1 , . . . , M n+1 . These matchings will be referred to as colours, and the colour of an edge e means the matching e belongs to. There will always be a rainbow matching called M of size n in G. We set X 0 = X \ V (M ) and Y 0 = Y \ V (M ). The colour missing from M will denoted by c * .
Notice that for any edge e, there is a special colour (the colour c e of the edge e) as well as a special vertex in X (i.e. e ∩ X) and in Y (i.e. e ∩ Y ). In what follows we will often want to refer to the edge e, the colour c e , and the vertices e∩X and e∩Y interchangeably. To this end we make a number of useful definitions: • For an edge e, we let (e) C be the colour of e, (e) X = e ∩ X, and (e) Y = e ∩ Y . We will now introduce two important and slightly complicated definitions. Both Definition 3.1 and 3.2 will apply in the setting of a bipartite graph G with bipartition X ∪ Y consisting of n + 1 edge-disjoint matchings, and a rainbow matching M of size n missing colour c * . The first definition is that of a switching-informally this should be thought of as a sequence of edges of G \ M which might be exchanged with a sequence of edges of M in order to produce a new rainbow matching of size n. See Figure 1 for an illustration of a switching.  (iv) For all i = j, we have e i ∩ e j = e i−1 ∩ m j = ∅ and also c i = c j .
(v) For all i, (e i ) X ∈ X .
The next definition is that of a free subset of X-informally a subset X ⊂ X is free if there are matchings M which "look like" M , but avoid small subsets of X . Definition 3.2. Let X , T ⊆ X, k ∈ N, and c be a colour. We say that X is (k, T, c)-free if T ∩ X = ∅, c ∈ (X ∪ T ) C , and the following holds: Let A be any set of at most k edges in M \ ((T ) M ∪ (c) M ), B ⊆ X any set of at most k vertices such that (A) X ∩ B = ∅. Then there is a rainbow matching M of size n satisfying the following: • M misses the colour c.
In other words, M replaces all edges touching B by edges not touching B, whilst maintaining the edges of A and avoiding the colour c. It is worth noticing that X 0 is (n, ∅, c * )-free (always taking the matching M to be M in the definition). The following lemma is crucial-it combines the preceding two definitions together and says that if we have an X -switching σ for a free set X , then there is a new rainbow matching of size n which avoids (m(σ)) X . Lemma 3.3. Suppose that X is (2k, T, c)-free and σ = (e 0 , m 1 , e 1 , . . . , e −1 , m ) is an X -switching from c to (m ) C of length k. Let A be any set of at most k edges in M − (c) M and let B be any subset of X of order at most k. Suppose that the following disjointness conditions hold.
Then there is a rainbow matchingM of size n in G which misses colour (m ) C , agrees with M on A, and has (M ) X ∩ (m(σ)) X = (M ) X ∩ B = ∅.
Proof. We let A = m(σ) ∪ A and B = (e(σ)) X ∪ B. Notice that we have |A |, |B | 2k. Also from the definition of "switching", we have that for any i and j, the edges e i and m j never intersect in X which together with ( Finally, σ being an X -switching and B ⊆ X imply that B ⊆ X . Therefore we can invoke the definition of X being (2k, T, c)-free in order to obtain a rainbow matching M of size n avoiding B , agreeing with M on A , and missing colour c = (e 0 ) C . We let M = (M \ m(σ)) ∪ e(σ) = M + e 0 − m 1 + e 1 − m 2 + e 2 − · · · + e −1 − m .
We claim thatM is a matching which satisfies all the conditions of the lemma. Recall that B ⊇ (e(σ)) X , A ⊇ m(σ), and (A ) X ∩ B = ∅. Since M agreed with M on A and was disjoint from B , we get m(σ) ⊆ M and e(σ) ∩ (M \ m(σ)) = ∅. This implies thatM is a set of n edges and also that (M ) X = (M ) X \ (m(σ)) X ∪ (e(σ)) X is a set of n vertices. Finally notice that since ( ThusM is a set of n edges with n vertices in each of X and Y i.e. a matching. The matchingM is clearly rainbow, missing the colour (m ) C since m i and e i always have the same colour.
To see thatM agrees with M on edges in A, notice that M agreed with M on these edges since we had A ⊆ A . Since (σ) X ∩ (A) X = ∅ implies that σ contains no edges of A, we obtain thatM agrees with M on A.
We study X -switchings by looking at an auxiliary directed graph. For any X ⊆ X, we will define a directed, totally labelled graph D X . We call D X a "labelled" graph rather than a "coloured" graph just to avoid confusion with the coloured graph G. Of the electronic journal of combinatorics 24(1) (2017), #P1.13 course the concepts of "coloured" and "labelled" graphs are equivalent, and we will freely apply results from Section 2 to labelled graphs. The vertices and edges of D X will be labelled by elements of the set X ∪ { * }. Definition 3.4. Let X be a subset of X. The directed graph D X is defined as follows: • The vertex set of D X is the set of colours of edges in G. For any colour v ∈ V (D X ) present in M , v is labelled by "(v) X ". The colour c * is labelled by " * ".
• For two colours u and v ∈ V (D X ), there is a directed edge from u to v in D X whenever there is an x ∈ X such that there is a colour u edge from x to the vertex (v) Y in G. In this case uv is labelled by "x".
Notice that in the second part of this definition the labelling is well-defined since there cannot be colour u edges from two distinct vertices x and x to (v) Y (since the colour u edges form a matching in G).
Recall that a total labelling is proper if outgoing edges at a vertex always have different labels, ingoing edges at a vertex always have different labels, adjacent vertices have different labels, and an edge always has different labels from its endpoints. Using the fact that the matchings in G are disjoint we can show that D X is always properly labelled.
Lemma 3.5. For any X ⊆ X the total labelling on D X is always proper. In addition D X is vertex-rainbow.
Proof. Suppose that uv and u v are two distinct edges of D X with the same label x ∈ X . By definition of D X they correspond to two edges x(v) Y and x(v ) Y of G having colours u and u respectively. This implies that u and u are different since otherwise we would have two edges of the same colour leaving x in G (which cannot happen since colour classes in G are matchings). We also get that v and v are distinct since otherwise we would have edges of colours both u and u between x and (v) Y in G (contradicting the matchings forming G being disjoint).
Let uv be an edge of D X labelled by x and x(v) Y the corresponding colour u edge of G. Then u cannot be labelled by "x" (since that would imply that the colour u edge at x would end at (u) Y rather than (v) Y ), and v cannot be labelled by "x" (since then there would be edges from x to (v) Y in G of both colours u and v).
The fact that D X is vertex-rainbow holds because M being a matching implies that (c) X is distinct for any colour c.
Recall that a path in a totally labelled graph is defined to be rainbow whenever all its vertices and edges have different colours. The reason we defined the directed graph D X is that rainbow paths in D X correspond exactly to X -switchings in G. Let P = v 0 , . . . , v be a path in D X for some X . For each i = 0, . . . , − 1 let e i be the colour v i edge of G corresponding to the edge v i v i+1 in D X . We define σ P to be the sequence of edges (e 0 , (v 1 ) M , e 1 , (v 2 ) M , e 2 , . . . , (v −1 ) M , e −1 , (v ) M ). Notice that (e(σ P )) X is the set of labels of edges in P , and (m(σ P )) X is the set of labels of vertices in P − v 0 .
The following lemma shows that if P is rainbow then σ P is a switching.
the electronic journal of combinatorics 24(1) (2017), #P1.13 Lemma 3.6. Let P = v 0 , . . . , v be a rainbow path in D X for some X ⊆ X. Then σ P is an X -switching from v 0 to v of length .
Proof. As in the definition of σ P , let e i be the colour v i edge of G corresponding to the edge v i v i+1 in D X . We need to check all the parts of the definition of "X -switching". For part (i), notice that (v 1 ) M , . . . , (v ) M are edges of M by definition of (.) M , whereas e i cannot be the colour Part (iv) follows from the fact that P is a rainbow path. Indeed to see that for i = j we have e i ∩ e j = ∅, notice that e i ∩ e j ∩ X = ∅ since v i v i+1 and v j v j+1 have different labels in D X , and that Although it will not be used in our proof, it is worth noticing that the converse of Lemma 3.6 holds i.e. to every X -switching σ there corresponds a unique rainbow path P in D X such that σ = σ P .
So far all our lemmas were true regardless whether the rainbow matching M was maximum or not. Subsequent lemmas will assume that M is maximum. The following lemma shows that for a free set X , vertices in D X have large out-degree. Proof. Notice that since P is contained in D X \ (T ) C and since X being (2k, T, c)-free implies X ∩ T = ∅, we can conclude that (σ P ) X ∩ T = ∅. Therefore, Lemma 3.3 applied with A = ∅ implies that for any B ⊆ X with |B| k and B ∩ (σ P ) X = ∅, there is a rainbow matching M of size n which is disjoint from B and misses colour v. Since there are no rainbow matchings of size n + 1 in G this means that there are no colour v edges from X \ (σ P ) X to Y 0 . Indeed if such an edge xy existed, then we could apply Lemma 3.3 with B = {x} in order to obtain a rainbow matching M missing colour v and vertex x which can be extended to a rainbow n + 1 matching by adding the edge xy.
We claim that there are at least (1 + )n + |X | − |X| − 2|P | colour v edges from X \ (σ P ) X . Indeed out of the (1 + )n colour v edges in G at most |X| − |X | of them can avoid X , and at most 2|P | of them can pass through (σ P ) X , leaving at least (1 + )n − (|X| − |X |) − 2|P | colour v edges to pass through X \ (σ P ) X . Since none of these edges can touch Y 0 , each of them must give rise to an out-neighbour of v in D X . This shows that |N + D X (v)| (1 + )n + |X | − |X| − 2|P | which implies the result.
the electronic journal of combinatorics 24(1) (2017), #P1.13 The following lemma is the essence of the proof of Theorem 1.5. It roughly says that given a free set X 1 containing X 0 , there is another free set X 2 containing X 0 such that X 2 is much bigger than X 1 , but has worse parameter k. The proof of this lemma combines everything in this section with Lemmas 2.3, 2.10 and 2.12 from Section 2.
• Suppose that we have X 1 , T 1 ⊆ X and a colour c 1 such that X 1 is (k 1 , T 1 , c 1 )-free and we also have X 0 ⊆ X 1 ∪ T 1 and |T 1 | k 1 − 30 −1 .
Since W ⊆ D , there is a path, Q, of length 10 −1 from c 1 to some q ∈ W . Let c 2 be any vertex in W with (c 2 ) X ∈ (σ Q ) X . Let T 2 = T 1 ∪ (σ Q ) X ∪ (c 1 ) X . Let X 2 = ((W ) X ∪ X 0 ) \ (T 2 ∪ (c 2 ) X ). We claim that X 2 , T 2 , and c 2 satisfy the conclusion of the lemma.
First we show that X 2 is (k 2 , T 2 , c 2 )-free. The facts that T 2 ∩ X 2 = ∅ and c 2 ∈ X 2 ∪ T 2 follow from the construction of X 2 , T 2 , and c 2 . Let A be any set of k 2 edges of M \ ((T 2 ) M ∪ (c 2 ) M ), and B ⊆ X 2 any set of k 2 vertices such that (A) X ∩ B = ∅. Let there is a rainbow path P in D of length 3k 2 d from q to c 2 which is disjoint from V (Q − q) and (A) C , passes through every colour of (B W ) C , and whose edges and vertices don't have labels in (A) X ∪ (σ Q ) X ∪ B X 0 \ (q) X . Notice that this means that Q + P is a rainbow path from c 1 to c 2 .
We apply Lemma 3.3 with X = X 1 , T = T 1 , c = c 1 , σ = σ Q+P , A = A, B = B X 0 . For this application notice that σ Q+P is an X 1 -switching of length k 1 /2, which holds because of Lemma 3.6 and because 2|Q| + 2|P | 20 −1 + 2k 2 d k 1 /2. We also need to check the various disjointness conditions-(A) X ∩T 1 = (A) X ∩(σ Q+P ) X = (A) X ∩B X 0 = ∅ the electronic journal of combinatorics 24(1) (2017), #P1.13 (which hold because (A) X was disjoint from T 2 , P , and B), (σ Q+P ) X ∩ T 1 = ∅ (which holds since vertices and edges in D have no labels from T 1 ), and (σ Q+P ) X ∩ B X 0 = ∅ (which holds since B was disjoint from T 2 and P had no labels from B X 0 ). Therefore Lemma 3.3 produces a rainbow matching M of size n which agrees with M on A, avoids (m(σ Q+P )) X ∪ B X 0 , and misses colour c 2 . Since P passes through every colour in (B W ) C , we have B W ⊆ (m(σ Q+P )) X and so M avoids all of B. Since A and B were arbitrary, we have shown that X 2 is (k 2 , T 2 , c 2 )-free.
Finally, |X 2 | > |X 1 | + n/2 holds because since (W ) X was disjoint from X 0 we have We are finally ready to prove Theorem 1.5. The proof consists of starting with X 0 and applying Lemma 3.8 repeatedly, at each step finding a free set X i which is n/2 bigger than X i−1 . This clearly cannot be performed more than 2 −1 times (since otherwise it would contradict |X i | |X| = |X 0 | + n), and hence the "there is no rainbow matching in G of size n + 1" clause of Lemma 3.8 could not be true.
Proof of Theorem 1.5. Let G be a bipartite graph which is the union of n 0 N disjoint matchings each of size at least (1 + )n 0 . Let M be the largest rainbow matching in G and c * the colour of any matching not used in M . Let n be the number of edges of M . Since M is maximum, Lemma 1.6 tells us that n N/2. Let X 0 = X \ M and Y 0 = Y \ M . Suppose for the sake of contradiction that n < n 0 .
For i = 0, . . . , 2 −1 we repeatedly apply Lemma 3.8 to X i , k i , T i , c i in order to obtain sets X i+1 , T i+1 ⊆ X and a colour c i+1 such that X i+1 is (k i+1 , T i+1 , c i+1 )-free, To see that we can repeatedly apply Lemma 3.8 this way we only need to observe that there are no rainbow n + 1 matchings in G, and that for i 2 −1 we always have n 10 20 −8 k i , k i 10 −1 , and |T i | 30 −1 i k i − 30 −1 .
But now we obtain that |X 2 −1 | > |X 0 | + n = |X| which is a contradiction since X i is a subset of X.

Golden Ratio Theorem
In this section we prove Theorem 1.8. The proof uses Theorem 1.4 as well as Lemma 2.11.
Proof of Theorem 1.8. The proof is by induction on n. The case "n = 1" is trivial since here G is simply a matching. Suppose that the theorem holds for all G which are unions of < n matchings. Let G be a graph which is the union of n matchings each of size φn+20n/ log n. Suppose that G has no rainbow matching of size n. Let M be a maximum the electronic journal of combinatorics 24(1) (2017), #P1.13 rainbow matching in G. By induction we can suppose that |M | = n − 1. Let c * be the missing colour in M .
Let X 0 = X \ V (M ) and Y 0 = Y \ V (M ). Notice that for any colour c there are at least (φ − 1)n + 20n/ log n colour c edges from X 0 to Y and at least (φ − 1)n + 20n/ log n colour c edges from Y 0 to X. If n < 10 6 , then this would give more than n colour c * edges from X 0 to Y , one of which could be added to M to produce a larger matching. Therefore, we have that n 10 6 .
We define an edge-labelled directed graph D whose vertices are the colours in G, and whose edges are labelled by vertices from X 0 ∪ Y 0 . We set cd an edge in D with label v ∈ X 0 ∪ Y 0 whenever there is a colour c edge from v to the colour d edge of M . Notice that D is out-proper-indeed if edges ux and uy ∈ E(D) had the same label v ∈ X 0 ∪ Y 0 , then they would correspond to two colour u edges touching v in G (which cannot happen since the colour classes of G are matchings).
Recall that d R (x, y) denotes the length of the shortest rainbow x to y path in D.
We'll need the following two claims. The graph M is clearly a rainbow graph with n edges. We claim that it is a matching.
Distinct edges e i and e j satisfy e i ∩ e j = ∅ since P is a rainbow path. The edge e i intersects V (M ) only in one of the vertices of m i , which are not present in M . This means that M is a rainbow matching of size n contradicting our assumption that M was maximum.
Claim 4.2. There is a set A ⊆ V (D) containing c * such that for all a ∈ A we have |N + (a) \ A| n/ log n and d R (c, a) log n.
Proof. This follows by applying Lemma 2.11 to D with = (log n) −1 .
Let A be the set of colours given by the above claim. Let M be the submatching of M consisting of the edges with colours not in A. Since c * ∈ A, we have |M | + |A| = n.
Let A X be the subset of X spanned by edges of M with colours from A, and A Y be the subset of Y spanned by edges of M with colours from A. Claim 4.1 shows that for any a ∈ A there are at most log n colour a edges between X 0 and Y 0 . Therefore there are at least (φ − 1)n + 20n/ log n − log n colour a edges from X 0 to Y ∩ (M ) Y . Using the property of A from Claim 4.2 we obtain that there are at least (φ − 1)n + 19n/ log n − log n colour a the electronic journal of combinatorics 24(1) (2017), #P1.13 edges from X 0 to A Y . Similarly, for any a ∈ A we obtain at least (φ−1)n+19n/ log n−log n colour a edges from Y 0 to A X .
By applying Theorem 1.4 to the subgraph of G consisting of the colour A edges between X 0 and A Y we can find a subset A 0 ⊆ A and a rainbow matching M 0 between X 0 and A Y using exactly the colours in A 0 such that we have Recall that for each a ∈ A 1 there is a colour a matching between Y 0 and A X of size at least (φ−1)n+19n/ log n−log n. Notice that the following holds The first inequality follows from φ 2 − φ − 1 = 0 as well as some simple bounds on √ n and log n for n 10 6 . The second inequality holds since x/ log x is increasing.
By induction there is a rainbow matching M 1 between Y 0 and A X using exactly the colours in A 1 . Now M ∪ M 0 ∪ M 1 is a rainbow matching in G of size n.

Concluding remarks
Here we make some concluding remarks about the techniques used in this paper.

Analogues of Menger's Theorem for rainbow k-edge-connectedness
One would like to have a version of Menger's Theorem for rainbow k-edge-connected graphs as defined in the introduction. In this section we explain why the most natural analogue fails to hold.
Consider the following two properties in an edge coloured directed graph D and a pair of vertices u, v ∈ D.
(i) For any set of k − 1 colours S, there is a rainbow u to v path P avoiding colours in S.
(ii) There are k edge-disjoint u to v paths P 1 , . . . , P k such that P 1 ∪ · · · ∪ P k is rainbow.
The most natural analogue of Menger's Theorem for rainbow k-edge-connected graphs would say that for any graph we have (i) ⇐⇒ (ii). One reason this would be a natural analogue of Menger's Theorem is that there is fractional analogue of the statement (i) ⇐⇒ (ii). We say that a rainbow path P contains a colour c if P has a colour c edge.
the electronic journal of combinatorics 24(1) (2017), #P1.13 Proposition 5.1. Let D be a edge coloured directed graph, u and v two vertices in D, and k a real number. The following are equivalent.
(a) For any assignment of non-negative real numbers y c to every colour c, satisfying c a colour y c < k, there is a rainbow u to v path P with c contained in P y c < 1.
(b) We can assign a non-negative real number x P to every rainbow u to v path P , such that for any colour c we have P contains c x P 1 and also P a rainbow u to v path x P k.
Proof. Let k a be the minimum of c a colour y c over all choices of non-negative real numbers y c satisfying c contained in P y c 1 for all u to v paths P . Similarly, we let k b be the maximum of P a rainbow u to v path x P over all choices of non-negative real numbers x P satisfying P contains c x P 1 for all colours c.
It is easy to see that k a and k b are solutions of two linear programs which are dual to each other. Therefore, by the strong duality theorem (see [9]) we have that k a = k b which implies the proposition.
The reason we say that Proposition 5.1 is an analogue of the statement "(i) ⇐⇒ (ii)" is that if the real numbers y c and x P were all in {0, 1} then (a) would be equivalent to (i) and (b) would be equivalent to (ii) (this is seen by letting S = {c : y c = 1} and {P 1 , . . . , P k } = {P : x P = 1}).
Unfortunately (i) does not imply (ii) in a very strong sense. In fact even if (ii) was replaced by the weaker statement "there are k edge-disjoint rainbow u to v paths", then (i) would still not imply (ii).
Proposition 5.2. For any k there is a coloured directed graph D k with two vertices u and v such that the following hold.
(I) For any set of k colours S, there is a rainbow u to v path P avoiding colours in S.
(II) Any pair P 1 , P 2 of rainbow u to v paths have a common edge.
Proof. We will construct a multigraph having the above property. It is easy to modify the construction to obtain a simple graph. Fix m > 2k + 1. The vertex set of D is {x 0 , . . . , x m } with u = x 0 and v = x m . For each i = 0, . . . , m − 1, D has k + 1 copies of the edge x i x i+1 appearing with colours i, m + 1, m + 2, . . . , m + k. In other words G is the union of k + 1 copies of the path x 0 x 1 . . . x m one of which is rainbow, and the rest monochromatic.
Notice that D satisfies (II). Indeed if P 1 and P 2 are u to v paths, then they must have vertex sequence x 0 x 1 . . . x m . Since there are only m + k colours in D both P 1 and P 2 must have at least m − k edges with colours from {0, . . . , m − 1}. By the Pigeonhole Principle, since 2(m − k) > m, there is some colour i ∈ {0, . . . , m} such that both P 1 and P 2 have a colour i edge. But the only colour i edge in D is x i x i+1 which must therefore be present in both P 1 and P 2 . There is another, more subtle, reason why (i) does not imply (ii). Indeed if we had "(i) =⇒ (ii)" then this would imply that every bipartite graph consisting of n matchings of size n contains a rainbow matching of size n.
Indeed given a bipartite graph G with bipartition X ∪ Y consisting of n matchings of size n construct an auxiliary graph G by adding two vertices u and v to G with all edges from u to X and from Y to v present. These new edges all receive different colours which were not in G. It is easy to see for any set S of n − 1 colours, there is a rainbow u to v path in G i.e. (i) holds for this graph with k = n. In addition, for a set of paths P 1 , . . . , P t with P 1 ∪ · · · ∪ P t rainbow, it is easy to see that {P 1 ∩ E(G), . . . , P t ∩ E(G)} is a rainbow matching in G of size t.
Therefore if "(i) =⇒ (ii)" was true then we would have a rainbow matching in G of size n. However, as noted in the introduction, there exist Latin squares without transversals, and hence bipartite graphs consisting of n matchings of size n containing no rainbow matching of size n.
The above discussion has hopefully convinced the reader that the natural analogue of Menger's Theorem for rainbow k-edge-connectedness is not true. Nevertheless, it would be interesting to see if any statements about connectedness carry over to rainbow k-edgeconnected graphs.

Improving Theorem 1.5
One natural open problem is to improve the dependency of N on in Theorem 1.5. Throughout our proof we made no real attempt to do this. However there is one interesting modification which one can make in order to significantly improve the bound on N which we mention here.
Notice that the directed graphs D X in Section 1.5 and the directed graph D in Section 1.8 had one big difference in their definition-to define the graphs D X we only considered edges starting in X, whereas to define the graph D, we considered edges starting from both X 0 and Y 0 . It is possible to modify the proof of Theorem 1.5 in order to deal with directed graphs closer to those we used in the proof of Theorem 1.8. There are many nontrivial modifications which need to be made for this to work. However, the end result seems to be that the analogue of Lemma 3.8 only needs to be iterated O(log −1 ) many times (rather than O( −1 ) as in the proof of Theorem 1.5). This would lead to an improved bound on N in Theorem 1.5 N = O C log for some constant C. In the grand scheme of things, this is still a very small improvement to bound in Theorem 1.5, and so we do not include any further details here. It is likely that completely new ideas would be needed for a major improvement in the bound in Theorem 1.5.
Another desirable improvement to Theorem 1.5 would be to remove the condition that G is simple i.e. prove an approximate version of the Aharoni-Berger Conjecture. The assumption that G is simple appears only once in the proof of Theorem 1.5-it appears in the proof of Lemma 3.5. If G was a multigraph, then the labeling of the corresponding directed graph D X would not necessarily be proper. To see an example of this, let G be a graph with vertex set {x 1 , y 1 , . . . , x n , y n }, such that the edge x i y i appears n times, once with each of the colours 1, . . . n. For X = {x 1 , . . . , x n } and M a rainbow matching the electronic journal of combinatorics 24(1) (2017), #P1.13 of size n, the corresponding directad graph D X has vertex set {1, . . . , n} and ij an edge labeled by x j for all i = j. Thile this directed graph D X is out-proper, it is not in-proper. Moreover, it can be checked that this D X doesn't have any rainbow k-edge-connected subsets A with |A| 2 and k 1. This is a serious barrier to our proof strategy since it stops most of the machinery from Section 2 from working. This barrier didn't occur in the proof of Theorem 1.8, since in that theorem we only used the fact that D X contains a subgraph with high minimum degree, rather than a highly connected one.