Spanning trees in sparse expanders

Take k=800, and choose \frac{1}{n}\ll c\ll\frac{1}{C}\ll\frac{1}{k}. Let G be an (n,d)-expander with d\ge C\Delta\sqrt{n}, where \Delta\le c\sqrt{n}. By Corollary 2.2, every T\in \mathcal{T}(n, \Delta) has either at least \frac{n}{4k\Delta} pendant stars or \frac{n}{4k\Delta} vertex-disjoint caterpillars each of length k. For the former case, by applying Lemma 5.1 with \gamma=\frac{1}{4k}, one can find a copy of T in G. For the latter case, as c\ll \frac{1}{C}, Lemma 5.2 immediately gives a copy of T in G.

To build a desired star-matching, we shall show a generalized Hall’s condition (see, e.g., Bollobás [Bo98])), that is,

From the assumption that every set X of at most m vertices in A has at least d|X| neighbors in B, we know that (Proof) holds when |X|\le m. It remains to consider the following two cases depending on the size of |X|:

• m<|X|\le |A|-m: By the assumption (2) and |B|=\sum_{v\in A}f(v), we have

  |N(X)\cap B|> |B|-m\ge \sum_{v\in A}f(v)-\sum_{v\in A\setminus X}f(v)=\sum_{v\in X}f(v),

  where the second inequality follows as m\le |A|-|X|\le\sum_{v\in A\setminus X}f(v).

• |A|-m<|X|\le |A|: Let T:=B\setminus N(X). We may assume that T\neq\emptyset for otherwise we are done. As there is no edge between X and T, we obtain that |T|<m and N(T)\cap A\<A\setminus X. Then it follows by assumption that d |T|\le|N(T)\cap A|\le|A\setminus X|\le\sum_{v\in A\setminus X}f(v). Thus we have

  |N(X)\cap B|=|B|-|T|=\sum_{v\in X}f(v)+\sum_{v\in A\setminus X}f(v)-|T|>\sum_{v\in X}f(v).

In all cases, condition (Proof) holds and this yields a desired star-matching covering A\cup B.

Given integers m,d_1,\Delta\in \mathbb{N} as in the statement, we fix G to be an m-joined graph and W\<V(G) such that G d_1-expands into W with d_1\ge \Delta+2. Now we build the desired tree array in the following two steps. At the first step, we shall use Lemma 4.2 to build pairwise vertex-disjoint paths with internal vertices in W, each connecting a prescribed pair of vertices in \mathcal{I}. Let S be an empty graph on vertex set V(\mathcal{I}).

Therefore, by repeatedly applying Lemma 4.2 to G' with (d_1,s,m) in place of (d,\ell, m), we can connect the pairs of vertices from \mathcal{I} by vertex-disjoint paths P_1, \ldots,P_t each of length s in G'. This can be done as in this process the resulting subgraph S_i:=S\cup P_1\cup\cdots\cup P_{i}, i\le t, would still be a (d_1, m)-extendable graph of order at most

For the second step, let R:=V(S_t)\setminus V(\mathcal{I}). Then we shall repeatedly use Lemma 4.3 to build for all vertices v\in R, pairwise vertex-disjoint \Delta-ary trees T_v of height s, where T_v has root v and all other vertices in W\setminus R. In fact, in the initial step, we have that

For an arbitrary vertex v\in R, we apply Lemma 4.3 to obtain a desired copy T_v of \Delta-ary tree of height s with a root v and in particular we have that S_t\cup T_v is also (d_1, m)-extendable in G[W\cup V(\mathcal{I})]. We repeat this step for every v\in R iteratively such that the corresponding \Delta-ary trees T_v are pairwise vertex disjoint. Note that this can be done as at the end of this process the resulting subgraph \bigcup_{v\in R}T_v\cup S_t would still be (d_1, m)-extendable with maximum degree \Delta+2\le d_1 and order at most

The family \{T_v: v\in R\} of trees together with the family \{P_i: i\in[t]\} of paths as above form a desired (W,\mathcal{I},s,\Delta)-tree array.

Let h=\sqrt{\log n} and d_1=\Delta^{2h}. Then G d_1-expands into W and |W|\ge \frac{n}{\Delta^{5h/2}}. Note that s\in [h-1,2h-1] and

Since n and also h are sufficiently large, we have that

We can apply Lemma 3.8 with t=m to obtain a desired (W,\mathcal{I},s,\Delta)-tree array.

Choose \frac{1}{n}\ll\frac{1}{C}\ll\frac{1}{k} and throughout the proof we have

Assume that an n-vertex tree T contains a collection of \frac{n}{4k\Delta} vertex-disjoint caterpillars each containing a central path P_i, 1\le i\le\frac{n}{4k\Delta}, of length k\ge 800, where x_i,y_i are the ends of P_i. In the forthcoming proofs, we shall write r=\frac{n}{8k\Delta} and consider two subcases as follows:

1. (1) at least r paths P_i are bare in T;

2. (2) at least r paths P_i are internally attached with leaves in T;

For the first case, we may assume that P_1,P_2,\ldots,P_{r} are bare paths in T and let T'=T-\cup_{i\in[r]}V(P_i-\{x_i,y_i\}). Then our embedding proceeds by first randomly partitioning V(G) into two parts V_1,V_2 such that

We can easily check that |V_1|\ge \frac{n}{2},|V_2|\ge\frac{n}{16\Delta}-21\Delta m\ge\frac{n}{32\Delta} as d\ge C\Delta^2. Note that G is an (n,d)-expander and

By Lemma 3.1, there exists a partition V_1\cup V_2=V such that G d_i-expands into V_i for all i\in[2]. Based on this, we can first apply Theorem 1.9 to G[V_1] to find an embedding of T' inside V_1. Indeed, as d_1\ge 2 \Delta, it suffices to ensure that the order of T' is no larger than |V_1| - 4\Delta \lceil\frac{|V_1|}{2d_1}\rceil. This easily follows since |V_1| - |T'|=21\Delta m >4\Delta\left\lceil\frac{|V_1|}{2d_1}\right\rceil. Let \phi be the resulting embedding of T' and X,Y be the set of images of ends x_i,y_i, respectively. Now it remains to embed the bare paths P_i,i\in[r] using V_2. By the choice of \frac{1}{C}\ll\frac{1}{k} and the fact that d_2\ge\frac{C}{160}\sqrt{n}\ge C_{\ref{pathcover2}}k\sqrt{n}, such a collection of vertex-disjoint paths can be obtained by Lemma 3.6 applied to G[X\cup Y\cup V_2] with (d_2,k+1,(k+1)r,V_2) in place of (d,\ell,n,W), which together with \phi completes the embedding of T.

For the second case, the proof strategy is similar to that in Section ?. Without loss of generality, we may assume that P_1,P_2,\ldots, P_{r} are central paths from vertex-disjoint caterpillars in T, each with at least one leg. Following the technique in Section ?, we take subpaths Q_i\< P_i, 1\le i\le r such that each Q_i has length k'=\lfloor k/2\rfloor and one can write Q_i=a^i_0a^i_1\ldots a^i_{k'} so that a^i_1 is attached with at least one leaf. For every j\in\{0,1,\ldots,k'\}, denote by A_j the set of the vertices a^i_j taken over all paths Q_i, 1\le i\le r. Moreover we write L for the set of all leaves in T that are attached to the internal vertices of the paths Q_i,1\le i\le r. Write L^+=N_T(L). Then A_1\<L^+ and r=|A_j|\le|L| for every j\in\{0,1,\ldots,k'\}.

Let T':=T-(\bigcup_{j=1}^{k'-1} A_j\cup L). Now our embedding proceeds as follows.
Phase 0. Partition V(G).
We first randomly partition V(G) into four parts V_1,V_2,V_3,V_4 such that

Observe that for every i\in[4], |V_i|\ge r-21\Delta m\ge\frac{n}{16k\Delta} as d\ge C\Delta^2. Since G is an (n,d)-expander and

by Lemma 3.1, we can obtain such a partition with the following property:

1. G d_i-expands into V_i for all i\in[4].

Phase 1. Embed T' in G[V_1].
By property ?, we can apply Theorem 1.9 to G[V_1] to find a copy of T' inside V_1, denoted as \phi|_{T'}. Indeed, this easily follows since d_1\ge 2 \Delta in (?) and |V_1| -|T'|= 21\Delta m> 4\Delta\left\lceil\frac{|V_1|}{2d_1}\right\rceil.

Phase 2. Embed A_1 into V_2\cup V_3.
We first build a maximal matching M_1 between \phi(A_0) and V_2. Denote by A'_0, A_1' the sets of vertices in \phi(A_0) and V_2, respectively, that are not covered by M. Since G is m-joined, we have |A'_0|=|A'_1|< m. Moreover, by property ? and (?), we have that d(v,V_3)\ge d_3\ge 2m. Thus one can easily obtain a matching M_2 from A_0' to a subset B_1\<V_3 with |B_1|=|A_0'|. Let V_2'=(V_2\setminus A_1')\cup B_1 and V_3'=(V_3\setminus B_1)\cup A_0'. Then |V_3'\setminus V_3|=|V_2'\setminus V_2|\le m and note that V_2'=\phi(A_1). Furthermore, it is easy to verify that G \frac{d_2}{2}-expands into V_2' and \frac{d_3}{2}-expands into V_3' as d_i\ge 2m.

Phase 3. Embed the subpaths Q_i-a_0^i using V_3'
By the choice of \frac{1}{C}\ll\frac{1}{k} and the fact that \frac{d_3}{2}\ge\frac{C}{160k}\sqrt{n}\ge C_{\ref{pathcover2}}k\sqrt{n}, such a collection of vertex-disjoint paths can be obtained by Lemma 3.6 applied to G[V_2'\cup \phi(A_{k'})\cup V_3'] with (\frac{d_3}{2},k'+1,(k'+1)r,V_3') in place of (d,\ell,n,W). This yields an embedding \phi|_{T-L}.

Hence it remains to embed L into the remaining set R of vertices with R=V(G)-\phi(T-L).

Phase 4. Embed L into the leftovers via a star-matching.
Recall that L^+ is the neighborhood of L in T and A_1\<L^+. Also observe that V_4\<R. We shall build a star-matching between \phi(L^+) and R. To finish the embedding of T, we define an auxiliary function f:\phi(L^+)\rightarrow [\Delta] by taking f(v)=d_T(\phi^{-1}(v))-2, that is, the number of leaves we need to attach to every v\in \phi(L^+). Then as G \frac{d_2}{2}-expands into V_2', we observe that for every v\in R

Moreover, as V_4\<R and G d_4-expands into V_4 with d_4\overset{\eqref{eq56}}{>}\Delta, we obtain that

Applying Lemma 3.3 with (\phi(L^+),R,\Delta) in place of (A,B,d), we have a desired f-matching, which together with \phi|_{T-L} forms an embedding of T.

\blacksquare\blacksquare

We shall first create pairwise disjoint 40-fans F_v with root v for every v\in W_1\cup W_2 and the remaining vertices inside W_3. To achieve this, we choose a subset T\<W_3 such that |T|=\frac{|W_3|}{2} and d(v,T)\ge \frac{1}{4}d(v,W_3) for every v\in V(G). Such a set T can be found by choosing \frac{|W_3|}{2} vertices in W_3 uniformly at random and by applying Chernoff’s inequality and a union bound. Let T'=W_3\setminus T. We claim that for every set L of m vertices in W_1\cup W_2 and every set S\<W_3 of \frac{n}{100} vertices, there exists a 40-fan with a root in L and the remaining vertices in S. In fact, as G is m-joined, by averaging, there is a vertex, say v\in L, that has a set S_v of neighbors inside S with

Thus as G is m-joined, we can greedily pick a matching of 40 edges in S_v, which together with v form a 40-fan F_v as desired. Based on this claim we can greedily choose pairwise disjoint 40-fans F_v for all but at most m vertices v\in W_1\cup W_2, where all the remaining vertices of such fans F_v come from T'. This is possible as |T'|=\frac{|W_3|}{2}>\frac{n}{6}\ge 80|W_1\cup W_2|+\frac{n}{100}. Now it remains to find fans for a set L' of at most m vertices from W_1\cup W_2, and here we shall use T (reserved to cover the leftover vertices). This can be trivially done as for every v\in V(G), d(v,T)\ge \frac{1}{4}d(v,W_3)\ge\frac{|W_3|}{20|W|}d>\frac{d}{60}\ge 160m\ge 80|L'|+80 m. Denote by A_v the set of absorbers for v in F_v, v\in W_1\cup W_2 and let M=\cup_{v\in W_1\cup W_2}A_v. Then M is indeed a matching of size 160r in W_3. Let W_3'=W_3\setminus V(M) and thus |W_3'|=|W_3|-320r.

To describe how to route our paths through these triangles in the fans, we refer to an auxiliary bipartite graph H obtained by Lemma ? applied with n=3r. The bipartite graph H has maximum degree 40 and vertex classes X=[3r] and Y\cup Z such that |Y|=|Z|=2r and for any subset Z'\<Z with |Z'|=r there is a perfect matching between X and Z'\cup Y in H. Fix an arbitrary bijection \tau:Y\cup Z\to W_1\cup W_2 with \tau(Z)=W_1. Now we first pick pairwise disjoint matchings M_i\<M,i\in[3r] such that

1. for every i\in[3r] and every v\in \tau(N_H(i)), M_i consists of exactly one edge in A_v.

Recall that X_1=\{x_1,\ldots,x_{3r}\}, Y_1=\{y_1,\ldots,y_{3r}\}. Next, we shall build pairwise disjoint (x_i,y_i)-paths P_i of length \ell-2 for every i\in[3r] such that M_i\<E(P_i) and the other internal vertices come from W_3'\cup W_4. As |M_i|\le 40, we may take |M_i|=40 for instance and set M_i=\{u_jv_j: j\in[40]\} to ease the notation. Observe that \ell\ge 200> 3(|M_i|+1)+|M_i|, every such P_i ought to be constructed by connecting the corresponding 41 pairs of vertices

using pairwise disjoint paths P_{i,j} of length k_{i,j}\ge 3,j\in[41] satisfying

To do this, in total we have a set of \sum_{i\in[3r]}(|M_i|+1)\le 123r such pairs to connect as above, say \mathcal{P}, and the number of vertices used in all these connections is at most 3r\ell. Note that by the choice of \frac{1}{C}\ll\frac{1}{\ell} and r=\frac{n}{10^4\ell}, we have

By repeatedly applying Lemma 6.1 with d_1=m and U playing the role of the set of vertices in W_3' uncovered by previous connections, one can greedily connect all but a subfamily \mathcal{P}'\<\mathcal{P} of at most 2m pairs via vertex-disjoint paths P_{i,j} with all the internal vertices in W_3'. Now we shall connect the remaining pairs in \mathcal{P}' using W_4. Note that G d_4-expands into W_4 and thus for every v\in V(G),

Thus one can greedily complete the connection for \mathcal{P}' as the total number of vertices used in connections is at most 2m\ell.

Let A=\cup_{i\in[3r]}(P_i\setminus\{x_i,y_i\})\cup W_2 and thus |A|=3r(\ell-2)-r. We claim this is such a set as required by the claim. Indeed, let U\< W_1 be any set of size r. From the property of the graph H, we can find a perfect matching between \tau^{-1}(U)\cup Y and [3r]. For each i\in [3r], take the vertex v\in U\cup W_2 such that \tau^{-1}(v)\in \tau^{-1}(U)\cup Y is matched to i and also take an (x_i,y_i)-path of length \ell-1 on V(P_i)\cup\{v\}, i.e., in P_i replacing the unique edge uw\in P_i\cap A_v by the path uvw (see ?). These paths altogether cover U\cup A, as required.