Turán-type and tiling problems in oriented graphs

Let a,b,c\in \mathbb{N} with 2\leqslant a \leqslant b \leqslant c and \gcd(a+b+c,c^2-ab)=1. Set h=a+b+c. Choose

Let G be an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta)n.

We claim that G contains a D_{a,b,c}-factor. By Lemma 4.1, it suffices to consider the case where V(G) is not (D_{a,b,c},\beta,t)-closed for any t\in\mathbb{N}. Then Lemma 4.2 implies that G is \gamma-extremal. By Lemma 4.3, there exist a set S\subseteq V(G) with |S|\leqslant h-1 and a constant-sized D_{a,b,c}-tiling \mathcal{H} in G-S such that V(G)\setminus (S\cup V(\mathcal{H})) admits a \gamma^{1/3}-superextremal partition (V_1,V_2,V_3) with |V_i|\equiv 0\pmod h for all i\in[3]. Since n\in h\mathbb{N}, it follows that |S|\equiv 0\pmod h, and hence |S|=0. Let G^\prime=G-V(\mathcal{H}). Since \mathcal{H} is a constant-sized tiling, we have

Applying Lemma 4.5 to G^\prime with parameters 2\eta and \gamma^{1/3}, we conclude that G^\prime contains a D_{a,b,c}-factor \mathcal{H}^\prime. Then \mathcal{H}\cup\mathcal{H}^\prime is a D_{a,b,c}-factor of G, as desired.

Let b,c\in \mathbb{N} with 1\leqslant b\leqslant c, and set h=1+b+c. Choose

Let G be an n-vertex semi-regular tournament. Assume that either \gcd(1+b+c,c^2-b)=1 or 3\mid (1+b+c). If the former holds, then G contains a D_{1,b,c}-factor by the same argument as in the proof of Theorem 1.4. Thus we may assume that 3\mid (1+b+c). Moreover, by Lemma 4.1, it suffices to consider the case where V(G) is not (D_{1,b,c},\beta,t)-closed for any t\in \mathbb{N}.

By Lemma 4.2, G is \gamma-extremal. Then Lemma 4.4 implies that there exists a constant-sized D_{1,b,c}-tiling \mathcal{H} such that V(G)\setminus V(\mathcal{H}) admits a \gamma^{1/3}-superextremal partition (V_1,V_2,V_3) with |V_i|\equiv 0\pmod{1+b+c} for each i\in[3]. Let G^\prime=G-V(\mathcal{H}). Since \mathcal{H} is a constant-sized tiling, we have

Applying Lemma 4.5 to G^\prime with parameters \eta and \gamma^{1/3}, we conclude that G^\prime contains a D_{1,b,c}-factor \mathcal{H}^\prime. Then \mathcal{H}\cup\mathcal{H}^\prime is a D_{1,b,c}-factor of G, as desired.

Let a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b\leqslant c. Set h=a+b+c. Choose

Suppose that G is an n-vertex oriented graph with \delta^0(G)\geqslant(1/2-\eta)n. Choose U\subseteq V(G) with |U|\leqslant h-1 such that |V(G)\backslash U|\equiv 0\pmod h. Set G^{\ast}=G-U.

If V(G^{\ast}) is (D_{a,b,c},\beta,t)-closed in G^{\ast} for some t\in\mathbb{N}, then, by Lemma 4.1, G^{\ast} has a D_{a,b,c}-factor, and hence G has the desired tiling since |U|\leqslant h-1. Thus we may assume that V(G^{\ast}) is not (D_{a,b,c},\beta,t)-closed in G^{\ast} for any t\in \mathbb{N}. Then G^{\ast} is \gamma-extremal by Lemma 4.2. Since G^{\ast}=G-U and U is a constant-sized set, G is clearly \gamma-extremal as well. By Lemma 4.3, there exist a set S\subseteq V(G) with |S|\leqslant 2h-3 (in particular, |S|\leqslant h-1 when \gcd(a+b+c,c^2-ab)=1) and a constant-sized D_{a,b,c}-tiling \mathcal{H} in G-S such that V(G)\backslash (S\cup V(\mathcal{H})) admits a \gamma^{1/3}-superextremal partition (V_1,V_2,V_3) with |V_i|\equiv 0\pmod h for all i\in[3]. Let G^{\prime}=G-(S\cup V(\mathcal{H})). Clearly, \delta^{0}(G^{\prime})\geqslant (1/2-\eta)n-|S\cup V(\mathcal{H})|\geqslant (1/2-2\eta)|G^\prime|. Applying Lemma 4.5 to G^\prime with parameters 2\eta and \gamma^{1/3}, we obtain a D_{a,b,c}-factor \mathcal{H}^{\prime} in G^{\prime}, which together with \mathcal{H} forms a D_{a,b,c}-factor of G-S. This proves the first statement.

Suppose now that G has no D_{a,b,c}-factor and that n\in h\mathbb{N}. It follows from Lemma 4.1 that V(G) is not (D_{a,b,c},\beta,t)-closed, and then Lemma 4.2 implies that G is \gamma-extremal. The “moreover” part follows immediately from Definition 2 ?.

Given a,b,c,t\in \mathbb{N} with 1\leqslant a\leqslant b \leqslant c, and set h=a+b+c. Choose

where 3\eta plays the role of \eta_2 in Lemma 5.1 with parameters \eta_1 and \delta.

Suppose that G is an n-vertex oriented graph with \delta^{0}(G)\geqslant (1/2-\eta)n. Then every vertex of G has at most 2\eta n missing edges. Since V(G) is (D_{a,b,c}, \beta, t)-closed, Lemma 5.3 implies that every S\in \binom{V(G)}{h} has at least \beta^2 n disjoint (D_{a,b,c}, ht)-absorbers in G. Applying Lemma 5.2 with parameters \beta^2, ht, and h, we obtain a (D_{a,b,c}, \delta)-absorbing set A \subseteq V(G) with |A| \leqslant \beta^2 n.

Let G^{\prime}=G-A. Then \delta^0(G^{\prime}) \geqslant(1/2-\eta)n-|A|\geqslant(1/2-\eta_1)|G^\prime| and \delta(G^{\prime})\geqslant |G^{\prime}|-2\eta n\geqslant (1-3\eta)|G^{\prime}|, since \eta\ll\beta\ll\eta_1\ll 1. Applying Lemma 5.1 to G^{\prime} with parameters \eta_1 and \eta_2=3\eta, we obtain a D_{a,b,c}-tiling \mathcal{H} in G^{\prime} covering all but at most \delta n vertices. Let R=V(G^{\prime}) \backslash V(\mathcal{H}) be the set of uncovered vertices. Clearly, |R|\leqslant \delta n. Since A is a (D_{a,b,c}, \delta)-absorbing set and h \mid (|A|+|R|), G[A \cup R] contains a D_{a,b,c}-factor \mathcal{H}^\prime, which together with \mathcal{H} forms a D_{a,b,c}-factor of G.

Given a,b,c\in\mathbb{N} with 1\leqslant a\leqslant b\leqslant c, let

Suppose that G is an n-vertex oriented graph with \delta^0(G)\geqslant(1/2-\eta)n. By Lemma 6.1, there exists an integer t^\prime>0 such that V(G) admits a (D_{a,b,c},\beta^\prime,t^\prime)-closed partition \mathcal{P}^\prime=(U_1,U_2,\ldots,U_p) with p\leqslant \lceil 1/\alpha\rceil and |U_i|\geqslant \alpha n/2 for each i\in [p].

If L^{\beta^\prime}_{\mathcal{P}^\prime}(D_{a,b,c}) contains a 2-transferral \mathbf{v}=\mathbf{e}_i-\mathbf{e}_j for some distinct i,j\in [p], then by Lemma 6.2 we may merge the corresponding parts U_i and U_j, that is, replace \mathcal{P}^\prime with the partition

Repeating this procedure for all 2-transferrals, we eventually obtain a (D_{a,b,c},\beta,t)-closed partition \mathcal{P} with t\geqslant t^\prime such that L^\beta_{\mathcal{P}}(D_{a,b,c}) is 2-transferral-free. Since V(G) is not (D_{a,b,c},\beta,t)-closed, the partition \mathcal{P} must contain at least two parts. The lemma now follows immediately from Lemma 6.3, applied with \alpha/2 in place of \alpha.

Given a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b\leqslant c, set h=a+b+c and choose

Suppose that G is an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta)n and that G is \gamma-extremal. By Lemma 7.2, there exists a \gamma^{1/3}-superextremal partition of V(G), say (U_1,U_2,U_3). Moreover, by Lemma 7.3, G contains linearly many disjoint copies of D_{a,b,c} of type-(i,i+1,i+2) for each i\in[3].

Note that R_h(|U_1|)\leqslant h-1, and each copy of D_{a,b,c} of type-(1,2,3) uses exactly a vertices of U_1. Thus, after removing a constant-sized family \mathcal{H}_1 of disjoint copies of D_{a,b,c} of type-(1,2,3), we have R_h(|U_1\backslash V(\mathcal{H}_1)|)\leqslant a-1. For each i\in[3], let S_i be any subset of U_i\backslash V(\mathcal{H}_1) such that

Clearly, |S_1|\leqslant a-1 and |S_2|,|S_3|\leqslant h-1.

If at least one of |S_2| and |S_3| is at most b+c-1, then set \mathcal{H}=\mathcal{H}_1, let S=S_1\cup S_2\cup S_3, and define V_i=U_i\backslash (S\cup V(\mathcal{H})) for each i\in[3]. We claim that \mathcal{H} and S are the desired tiling and set, respectively. Indeed, |S|\leqslant (a-1)+(b+c-1)+(h-1)=2h-3. Since S\cup V(\mathcal{H}) has constant size, the partition (V_1,V_2,V_3) of V(G)\backslash (S\cup V(\mathcal{H})) is clearly \gamma^{1/3}-superextremal. Moreover, |V_i|\equiv 0\pmod h for each i\in[3]. This completes the proof in this case.

Thus we may assume that |S_2|,|S_3|\geqslant b+c. Applying Lemma 7.3 again, we can find two disjoint copies of D_{a,b,c} in G-V(\mathcal{H}_1), one of type-(1,2,3) and the other of type-(2,3,1). Let \mathcal{H}_2 consist of these two copies, and set \mathcal{H}=\mathcal{H}_1\cup \mathcal{H}_2. For each i\in[3], let S_i^{\prime} be any subset of U_i\backslash V(\mathcal{H}) such that |S_i^{\prime}|=R_h(|U_i\backslash V(\mathcal{H})|). Since h=a+b+c and a\leqslant b\leqslant c, we have |S_1^{\prime}|=h+|S_1|-(a+c)\leqslant a+b-1,|S_2^{\prime}|=|S_2|-(a+b),|S_3^{\prime}|=|S_3|-(b+c). Moreover, as |S_2|,|S_3|\leqslant h-1, it follows that |S_2^{\prime}|+|S_3^{\prime}|\leqslant a+c-2, and hence

It is straightforward to check that \mathcal{H} and S_1^{\prime}\cup S_2^{\prime}\cup S_3^{\prime} are the desired tiling and set, respectively.

We now prove the “in particular” part. Let S be any subset of V(G) with |S|\leqslant h-1 such that |V(G)\backslash S| is divisible by h, and set r_i=R_h(|U_i\backslash S|) for each i\in[3]. Since \gcd(h,c^2-ab)=1, the system of congruences (7.4) in Lemma 7.4 has an integer solution, say (x_0,y_0,z_0). Note that (R_h(x_0),R_h(y_0),R_h(z_0)) is also an integer solution to (7.4), so we may assume w.l.o.g. that 0\leqslant x_0,y_0,z_0\leqslant h-1. In particular, x_0,y_0, and z_0 are all bounded by a constant. By Lemma 7.3, we can find a D_{a,b,c}-tiling \mathcal{H} consisting of x_0 copies of D_{a,b,c} of type-(1,2,3), y_0 copies of type-(2,3,1), and z_0 copies of type-(3,1,2). Let V_i=U_i\backslash (S\cup V(\mathcal{H})) for each i\in[3]. It is not difficult to check that |U_i\cap V(\mathcal{H})|\equiv r_i\pmod{h} for each i\in[3], and hence |V_i|\equiv 0\pmod h. Therefore, \mathcal{H} and S are the desired D_{a,b,c}-tiling and set, which completes the proof.

Given b,c\in\mathbb{N} with 1\leqslant b\leqslant c, set h=1+b+c. Let G be a semi-regular tournament on n\in h\mathbb{N} vertices that is \gamma-extremal. Since 3\mid h and n\in h\mathbb{N}, it follows that 3\mid n. By Lemma 7.2, V(G) has a \gamma^{1/3}-superextremal partition, say (V_1,V_2,V_3). Relabeling the parts if necessary, we may assume w.l.o.g. that V_1 is the largest part. Moreover, by reversing all edges of G and swapping the labels of V_2 and V_3 if necessary, we may further assume that |V_2|\geqslant |V_3|. Set |V_1|=n/3+g, |V_2|=n/3+f, and |V_3|=n/3-g-f, where g and f are integers since 3\mid n. It follows from |V_1|\geqslant |V_2|\geqslant |V_3| that g\geqslant 0 and g\geqslant f\geqslant -g-f.

We denote by M_{21} and M_{13} maximum matchings in E(V_2,V_1) and E(V_1,V_3), respectively. Next we claim that |M_{21}|\geqslant 2g+f-1 and |M_{13}|\geqslant g-f-1. Considering the out-degrees of vertices in V_2, we have

Let H be the underlying bipartite graph corresponding to the set of edges from V_2 to V_1. Clearly, e(H)\geqslant|V_2|(g+f/2-1/2). On the other hand, since (V_1,V_2,V_3) is \gamma^{1/3}-superextremal, 2 implies that every vertex of V_2 has degree at most n/6 + O(\gamma^{1/3}n) in H. Since H is bipartite, by König Theorem, the size of a maximum matching equals the size of a minimum vertex cover in H. Thus, we obtain that

It follows from \gamma\ll 1 that |M_{21}|>2g+f-2, and hence |M_{21}|\geqslant 2g+f-1. Similarly, by considering the in-degrees of vertices in V_3, we obtain |M_{13}|\geqslant g-f-1.

Next we claim that there exists a D_{1,b,c}-tiling \mathcal{H}_1 such that

If g=0, then |V_1|=|V_2|=|V_3|, and there is nothing to prove. Thus we may assume that g\geqslant 1.

When f\geqslant 0, since

we may first choose R_h(g) edges from M_{21} and then choose another R_h(f) edges from the remaining edges of M_{21}. By Lemma 8.1, we obtain that the former R_h(g) edges have vertex sets that are \gamma-transfer for (V_1,V_3), while the latter R_h(f) edges have vertex sets that are \gamma-transfer for (V_2,V_3).

When f<0, we have

By Lemma 8.1, we may choose R_h(-f) edges from M_{13} whose vertex sets are \gamma-transfer for (V_1,V_2). Moreover,

so there exist R_h(g+f) edges in M_{21} that are disjoint from the previously chosen R_h(-f) edges in M_{13}. The vertex sets of these edges are \gamma-transfer for (V_1,V_3).

By Definition 4, in either case the edges chosen above can be extended to a constant-sized D_{1,b,c}-tiling \mathcal H_1 such that

for some k\in\{0,1,2\}. Moreover, (V_1\setminus V(\mathcal H_1),\,V_2\setminus V(\mathcal H_1),\,V_3\setminus V(\mathcal H_1)) is still a \gamma^{1/3}-superextremal partition of V(G)\setminus V(\mathcal H_1).

Recall that h is divisible by 3. Hence kh/3 is an integer. Let

Note that x,y, and z are small integers. By Lemma 7.3, there is a collection \mathcal{H}_2 of disjoint copies of D_{1,b,c} in G- V(\mathcal{H}_1), consisting of x copies of D_{1,b,c} of type-(1,2,3), y copies of D_{1,b,c} of type-(2,3,1), and z copies of D_{1,b,c} of type-(3,1,2), respectively. Next we claim that

Here we only consider the case b\equiv0\pmod{3}; the other two cases can be verified similarly.

Since 3\mid(1+b+c) and b\equiv0\pmod{3}, it follows that c\equiv2\pmod3. Recall that in this case \mathcal{H}_2 consists of kh/3 copies of D_{1,b,c} of type-(1,2,3) and 2kh/3 copies of D_{1,b,c} of type-(3,1,2). As b\equiv0\pmod{3}, we have 2bk/3\in\mathbb{Z}, and hence 2bkh/3\equiv0\pmod h. This implies that |V_1\cap V(\mathcal{H}_2)|=kh/3+2bkh/3\equiv kh/3\pmod h. Similarly, |V_2\cap V(\mathcal{H}_2)|=(b+2c)kh/3\equiv kh/3\pmod{h} and |V_3\cap V(\mathcal{H}_2)|=(c+2)kh/3\equiv kh/3\pmod{h}, since b+2c\equiv c+2\equiv1\pmod3.

Therefore, \mathcal{H}=\mathcal{H}_1\cup\mathcal{H}_2 is the desired D_{1,b,c}-tiling. Let V_i^\prime=V_i\backslash V(\mathcal{H}). Clearly, |V_i^\prime|\equiv0 \pmod h. Moreover, since |V(\mathcal{H})| is constant, it is not difficult to check that (V_1^\prime,V_2^\prime,V_3^\prime) is a \gamma^{1/3}-superextremal partition of V(G)\backslash V(\mathcal{H}).

Set h=a+b+c. Choose \varepsilon,d and \eta such that

Let B be the set of \gamma^{1/2}-bad vertices in G. By Lemma 7.1, we have |B|\leqslant \gamma^{1/3}n. We divide the proof into the following three steps.

Step 1. Cover bad vertices by a D_{a,b,c}-tiling.

Let S be any subset of V(G)\backslash B with |S|\leqslant \gamma^{1/4}n, and let v be any vertex in B. In this step, we show that G has a D_{a,b,c}-tiling \mathcal{H}_v satisfying the following properties.

(i) v\in V(\mathcal{H}_v) and V(\mathcal{H}_v)\backslash\{v\}\subseteq V(G)\backslash (S\cup B);

(ii) \mathcal{H}_v consists of three disjoint copies of D_{a,b,c};

(iii) |V(\mathcal{H}_v)\cap V_1|=|V(\mathcal{H}_v)\cap V_2|=|V(\mathcal{H}_v)\cap V_3|.

Set \widetilde{V}_i= V_i\backslash (S\cup B) for each i\in [3]. It follows from ? and |S\cup B|\leqslant 2 \gamma^{1/4}n that |\widetilde{V}_i|=n/3\pm O(\gamma^{1/4}n). Suppose that v belongs to V_j for some j\in [3]. Since (V_1,V_2,V_3) is \gamma-superextremal, by 2 we have

Set V^\ast_{j+1}=N^+(v,\widetilde{V}_{j+1}) and V^\ast_{j-1}=N^-(v,\widetilde{V}_{j-1}). By the degree condition, v misses at most 2\eta n edges in G. Then one of |N^+(v,\widetilde{V}_{j})| and |N^-(v,\widetilde{V}_{j})| is at least (|\widetilde{V}_j|-2\eta n)/2\geqslant n/7 by ?. Let V^\ast_j be the larger of these two neighborhoods.

Note that every vertex in V^\ast_1\cup V^\ast_2\cup V^\ast_3 is \gamma^{1/2}-good. Thus, for each u\in V^\ast_i and i\in[3], we have

It is easy to verify that (V_i^\ast,V_{i+1}^\ast) is (\gamma,1/2)-regular for i\in[3]. Applying Lemma 5.5 to (V_i^\ast,V_{i+1}^\ast,V_{i+2}^\ast) for each i, there exists \beta>0 such that it contains at least \beta n^{a+b+c} copies of D_{a,b,c} of type-(i,i+1,i+2). Since each copy of D_{a,b,c} intersects at most (a+b+c)n^{a+b+c-1} others, we can choose three pairwise disjoint copies H_1,H_2,H_3 in G[V^\ast_1\cup V^\ast_2\cup V^\ast_3] such that H_i is of type-(i,i+1,i+2) for each i\in[3].

Recall that v\in V_j\cap B for some j\in[3]. Thus, v\notin V_j^{\ast}, and hence v does not belong to H_1\cup H_2\cup H_3. By construction, v either dominates all vertices in V_j^{\ast} or is dominated by all of them. Hence, for any x\in V(H_1)\cap V_j^{\ast}, there is a copy H_1^\ast of D_{a,b,c} in G[(H_1-x)\cup \{v\}]. Then \{H^\ast_1,H_{2},H_{3}\} is the desired D_{a,b,c}-tiling \mathcal{H}_v satisfying (i)-(iii).

We carry out the above process sequentially for each v\in B, defining S to be the set of vertices in V(G)\backslash B that are used in the previously constructed D_{a,b,c}-tilings. Clearly, |S|=O(\gamma^{1/3}n)\ll\gamma^{1/4}n, since |B|\leqslant \gamma^{1/3}n and |V(\mathcal{H}_v)|=3h for each v\in B. Therefore, we can find a D_{a,b,c}-tiling \mathcal{H}_0 of size 3h|B| consisting of the three tiles in \mathcal{H}_v for each v\in B. Note that each \mathcal{H}_v uses exactly h vertices from each V_i. Thus |V_i\backslash V(\mathcal{H}_0)|\equiv 0 \pmod {h} for each i\in [3].

Set V_i^0=V_i\backslash V(\mathcal{H}_0) for each i\in [3]. It follows from ? and |V(\mathcal{H}_0)|=O(\gamma^{1/3}n) that |V_i^0|=n/3\pm O(\gamma^{1/3}n). Moreover, every vertex in V_i^0 is \gamma^{1/2}-good for each i\in[3].

Step 2. Balance |V_1^0|,|V_2^0|, and |V_3^0| by a D_{a,b,c}-tiling.

By relabeling the sets if necessary, we may assume w.l.o.g. that |V_3^0|\leqslant |V_1^0|,|V_2^0|. Set k_i=(|V_i^0|-|V_3^0|)/h for i\in [2]. Since |V^0_{i}|=n/3\pm O(\gamma^{1/3} n) and |V^0_{i}|\equiv 0\pmod h for each i\in[3], we have k_1,k_2 = O(\gamma^{1/3} n), and both k_1 and k_2 are integers. Next we embed a D_{a,b,c}-tiling \mathcal{H}_i of size k_i in G[V_i^0] for each i\in[2] such that

For each v\in V_1^0, since it is \gamma^{1/2}-good, we have d^-(v,V_2), d^+(v,V_3)=O(\gamma^{1/2} n). Combining ? with the degree condition, we obtain

Thus d^\pm(v,V^0_1)\geqslant d^\pm(v,V_1)-|V(\mathcal{H}_0)| \geqslant |V^0_1|/2-O(\gamma ^{1/3}n) as |V(\mathcal{H}_0)|=O(\gamma^{1/3}n) and |V_1^0|=n/3\pm O(\gamma^{1/3}n). We apply the Diregularity Lemma to G[V^0_1] with \varepsilon, M^\prime = 1/\varepsilon, M^{\prime\prime}=1, and d. By Lemma 3.4, there exists a reduced oriented graph R with vertex set [k] and \delta^0(R)\geqslant (1/2-O(\gamma^{1/3})-d-3\varepsilon)k. Applying Lemma 6.6 (i) to R, there is an edge from N^+_R(1) to N^-_R(1), which together with 1 forms a directed triangle in R. Note that Lemma 6.6 holds for all sufficiently large oriented graphs with \delta^0(G)\geqslant (1/2-1/100)|G|. Thus we can greedily find at least \gamma^{1/4}k disjoint triangles in R, since \eta \ll \varepsilon \ll d\ll \gamma \ll 1/100.

Let 123 be one such fixed triangle in R, and let X_1,X_2,X_3 be the corresponding sets in G[V_1^0]. Then for each i\in[3], (X_i, X_{i+1}) is (\varepsilon, d)-regular and |X_i|=m for some integer m. Next we claim that there is a D_{a,b,c}-tiling \mathcal{D} in G[X_1\cup X_2\cup X_3] covering at least 2m vertices. Moreover, \mathcal{D} uses an equal number of vertices from X_1,X_2, and X_3. Indeed, let \mathcal{D} be initially empty and set X_i^{\prime}=X_i\backslash V(\mathcal{D}). By the same argument as in Step 1, as long as |X_i^{\prime}|> m/3, Lemma 5.5 implies that there are three disjoint copies of D_{a, b, c} in G[X^\prime_1\cup X^\prime_2\cup X^\prime_3] such that they use exactly h vertices from each X_i^{\prime}, where i\in[3]. Note that this is possible because (X^\prime_i, X^\prime_{i+1}) is (3\varepsilon,d/2)-regular by |X_i^{\prime}|> m/3 and Lemma 3.2.

Therefore, G[V_1^0] has a D_{a,b,c}-tiling covering at least \gamma^{1/4}k\cdot 2m vertices. Recall that km\geqslant |V_1^0|-\varepsilon |V_1^0|\geqslant n/4 and that k_1 = O(\gamma^{1/3} n). Then one can find the desired tiling \mathcal{H}_1 of size k_1 in G[V_1^0]. Similarly, there is a D_{a,b,c}-tiling \mathcal{H}_2 of size k_2 in G[V_2^0].

Set \mathcal{H}=\mathcal{H}_0\cup \mathcal{H}_1\cup \mathcal{H}_2. Let V_i^{\prime}=V_i\backslash V(\mathcal{H}) for each i\in[3]. By the construction above, we have |V_1^{\prime}|=|V_2^{\prime}|=|V_3^{\prime}|\equiv0\pmod h and |V(\mathcal{H})|=O(\gamma^{1/3}n).

Step 3. Find a D_{a,b,c}-factor in the resulting graph G[V_1^{\prime}\cup V_2^{\prime}\cup V_3^{\prime}].

Since |V^\prime _1|=|V^\prime _2|=|V^\prime _3|\equiv0\pmod h, there exists an integer t such that |V^\prime _i|=ht for each i\in[3]. Moreover, it follows from |V(\mathcal{H})|=O(\gamma^{1/3}n) that

For each i\in[3], let \{V^a_i,V^b_i,V^c_i\} be any partition of V^\prime _i with |V^j_i|=jt for each j\in\{a,b,c\}. Set W_i=V^a_i\cup V_{i+1}^b\cup V^c_{i+2}, where the indices are taken modulo 3. Then (W_1,W_2,W_3) is a partition of V(G)\backslash V(\mathcal{H}). By symmetry, it suffices to show how to find a D_{a,b,c}-factor in G[W_1].

Recall that Yuster [yusterorder20] showed that every jt-vertex oriented graph G with \delta(G)\geqslant(1-4^{-j})jt+4^j contains a TT_j-factor. Since \delta^0(G)\geqslant (1/2-\eta)n and |V^a_1|=at\geqslant an/(4h), we have \delta(G[V_1^a])\geqslant |V_1^a|-2\eta n\geqslant (1-8\eta h/a )|V_1^a|. It follows from \eta \ll 1 that G[V_1^a] contains a TT_a-factor \mathcal{T}_a. By a similar argument, there is a TT_b-factor \mathcal{T}_b in G[V_2^b] and a TT_c-factor \mathcal{T}_c in G[V_3^c]. Now we construct an auxiliary graph H as follows. Let V(H)=X_a\cup X_b\cup X_c, where each vertex of X_i corresponds to a copy of TT_i in \mathcal{T}_i for each i\in\{a,b,c\}. Clearly, |X_a|=|X_b|=|X_c|=t. For simplicity, for each u\in X_i with i\in\{a,b,c\}, let T_u be the copy of TT_i to which u corresponds. For each u\in X_i and v\in X_j with (i,j)\in\{(a,b),(b,c),(c,a)\}, we put uv\in E(H) if and only if every vertex of T_u dominates every vertex of T_v in G.

Recall that every vertex in V(G)\backslash V(\mathcal{H}) is \gamma^{1/2}-good. Then every copy of TT_i with i\in\{a,b,c\} has at most O(\gamma^{1/2} n) missing edges in G. This shows that for each u\in X_i, both its in-degree and out-degree in H are at least t-O(\gamma^{1/2} n)\geqslant (1-\gamma^{1/3})t, since t\geqslant n/(4h). Applying Lemma 9.1 to the underlying graph of H, we obtain a triangle-factor in H, which corresponds to a D_{a,b,c}-factor in G[W_1] by the construction of H. By the same argument, each of G[W_2] and G[W_3] contains a D_{a,b,c}-factor. Together with the former tiling \mathcal{H}, this yields a D_{a,b,c}-factor of G. This completes the proof.

Given a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b \leqslant c, let

where \eta is the constant from Theorem 5.4, and M is the constant in Lemma 3.3 with \varepsilon=\delta^3, d=\eta_1, M^\prime =1/\delta^3, and M^{\prime\prime}=1.

Suppose that G is an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta_1)n and \delta(G)\geqslant(1-\eta_2)n. Applying Lemma 3.3 to G with parameters \delta^3,\eta_1, M^\prime =1/\delta^3,M^{\prime\prime}=1, we obtain a partition V_0,V_1,\ldots,V_k of V(G) satisfying 1/\delta^3\leqslant k\leqslant M, |V_0|\leqslant \delta ^3 n and |V_i|=m for each i\in[k], where n/(2k) \leqslant m\leqslant n/k.

Let R be the reduced oriented graph. Then |R|=k and \delta^0(R)\geqslant (1/2-2\eta_1-3\delta^3)k\geqslant (1/2-\eta)k by Lemma 3.4. Applying Theorem 5.4 to R, we obtain a C_3-tiling \mathcal{T} in R covering all but at most three vertices. Let 123 be any fixed triangle in R. We next claim that G[V_1\cup V_2\cup V_3] contains a D_{a,b,c}-tiling covering all but at most 3\delta^2m vertices.

To see this, it suffices to show that whenever V_i^{\ast}\subseteq V_i satisfies |V_i^{\ast}|\geqslant \delta^2m for i\in[3], the graph G[V_1^{\ast}\cup V_2^{\ast}\cup V_3^{\ast}] contains a D_{a,b,c}-tiling \mathcal{F} such that |V(\mathcal{F})\cap V_i^\ast|=a+b+c for each i. Let (V_i^A,V_i^B,V_i^C) be an almost balanced partition of V_i^{\ast}, that is, the sizes of V_i^A, V_i^B, and V_i^C differ by at most one. Clearly, |V_i^A|, |V_i^B|, |V_i^C|\geqslant |V_i^{\ast}|/4\geqslant \delta^2 m/4. Since 123 is a triangle in R, each of (V_1,V_2), (V_2,V_3), and (V_3,V_1) is (\delta^3,\eta_1)-regular. Then Lemma 3.2 implies that each of (V_i^A,V_{i+1}^B), (V_{i+1}^B,V_{i+2}^C), and (V_{i+2}^C,V_i^A) is (4\delta,\eta_1/2)-regular for each i\in[3]. By Lemma 5.5, for each i\in[3] there exists a copy H_i of D_{a,b,c} of type-(V_i^A, V_{i+1}^B,V_{i+2}^C) in G[V_i^A\cup V_{i+1}^B\cup V_{i+2}^C]. Then \{H_1,H_2,H_3\} is the desired D_{a,b,c}-tiling of G[V_1^{\ast}\cup V_2^{\ast}\cup V_3^{\ast}]. Therefore, for every triangle 123 of R, the graph G[V_1\cup V_2\cup V_3] contains a D_{a,b,c}-tiling covering all but at most 3\delta^2m vertices.

Recall that \mathcal{T} is a C_3-tiling of R covering all but at most three vertices of R. Hence, by combining the tilings corresponding to the triangles in \mathcal{T}, we obtain the desired tiling of G, covering all but at most |V_0|+k\delta^2m + 3m \leqslant \delta^3n + k\delta^2 n/k + 3n/k\leqslant \delta n vertices, since |V_0|\leqslant \delta^3 n and m\leqslant n/k \leqslant \delta^3 n. This completes the proof.

Given a,b,c\in \mathbb{N} with a\leqslant b \leqslant c, let

where \sqrt \eta is the minimum of the constants given by Lemma 5.6 for s=a,b, and c with \mu. Set r_x=(d-\varepsilon)^x/\sqrt[x]{2} for each x\in\{a,b,c\}. If b=0, then there is nothing to prove by applying Lemma 5.6 to G[C]. Thus we may assume that 1\leqslant b\leqslant c.

Since (C,A) is (\varepsilon,d)-regular, we have e(C,A) \geqslant d|A||C|. Applying Lemma 3.5 to the underlying graph of G[C,A] with parameters d and \varepsilon=\varepsilon_0, there exists C^{\prime}\subseteq C with |C^{\prime}| \geqslant r_c |C| such that all but at most (\varepsilon_0 |C^{\prime}|)^c of the c-tuples in C^{\prime} have at least \varepsilon_0 r_c |A| common out-neighbors in A. By Lemma 3.2 and the fact that (B,C) is (\varepsilon,d)-regular, we have e(B,C^{\prime})\geqslant (d-\varepsilon)|B||C^{\prime}|. Applying Lemma 3.5 in the same way to the underlying graph of G[B,C^\prime], there exists C^{\prime\prime}\subseteq C^{\prime} with |C^{\prime\prime}| \geqslant r_c |C^{\prime}| such that all but at most (\varepsilon_0 |C^{\prime\prime}|)^c of the c-tuples in C^{\prime\prime} have at least \varepsilon_0 r_c |B| common in-neighbors in B.

Since \delta(G)\geqslant (1-\eta)n, every vertex has at most \eta n non-neighbors in G. As |C|\geqslant \xi n and |C^{\prime\prime}|\geqslant r_c^2|C|, we have \delta(G[C^{\prime\prime}]) \geqslant |C^{\prime\prime}|-\eta n \geqslant |C^{\prime\prime}|-\eta|C|/\xi \geqslant (1-\eta/(\xi r_c^2))|C^{\prime\prime}| \geqslant (1-\sqrt \eta)|C^{\prime\prime}|. Then Lemma 5.6 shows that G[C^{\prime\prime}] contains at least \mu|C^{\prime\prime}|^c copies of TT_c. Moreover, there are at least

“nice” copies of TT_c in G[C^{\prime\prime}] with the following property: all vertices of each such TT_c have many common out-neighbors in A and many common in-neighbors in B.

For any fixed “nice” copy of TT_c in G[C^{\prime\prime}], let A^\prime\subseteq A be the set of its common out-neighbors and let B^\prime \subseteq B be the set of its common in-neighbors. Since (A,B) is (\varepsilon,d)-regular and |A^\prime| \geqslant \varepsilon_0 r_c |A|\geqslant \varepsilon |A| and |B^\prime| \geqslant \varepsilon_0 r_c |B|\geqslant \varepsilon |B|, the definition of regular pair implies that e(A^\prime,B^\prime) \geqslant (d-\varepsilon) |A^\prime||B^\prime|. Applying Lemma 3.5 to the underlying graph of G[A^\prime,B^\prime] with \varepsilon =\varepsilon_0, there exists B^{\prime\prime}\subseteq B^\prime with |B^{\prime\prime}| \geqslant r_b |B^\prime| such that all but at most (\varepsilon_0 |B^{\prime\prime}|)^b of the b-tuples in B^{\prime\prime} have at least \varepsilon_0 r_b|A^\prime| common in-neighbors in A^\prime. Since \delta(G)\geqslant (1-\eta)n and |B|\geqslant\xi n, we have

By Lemma 5.6, G[B^{\prime\prime}] contains at least \mu |B^{\prime\prime}|^b- (\varepsilon_0|B^{\prime\prime}|)^b \geqslant \mu |B^{\prime\prime}|^b/2 “nice” copies of TT_b such that all vertices of each such copy have at least \varepsilon_0 r_b|A^\prime| common in-neighbors in A^\prime.

For any fixed “nice” copy of TT_b, let A^{\prime\prime}\subseteq A^\prime be the set of common in-neighbors of its vertices. In the same way, we have \delta(G[A^{\prime\prime}])\geqslant |A^{\prime\prime}|-\eta n \geqslant (1-\eta /(\xi \varepsilon_0^2 r_br_c))|A^{\prime\prime}|\geqslant (1-\sqrt \eta)|A^{\prime\prime}|. By Lemma 5.6 again, there are at least \mu |A^{\prime\prime}|^a copies of TT_a in G[A^{\prime\prime}]. Recall that there are at least \mu |C^{\prime\prime}|^c/2 choices for TT_c and \mu |B^{\prime\prime}|^b/2 choices for TT_b. After a simple calculation, G contains at least

copies of D_{a,b,c}, which completes the proof.

Observe that we may assume that s \geqslant 2. Let

where M is the constant obtained from Lemma 3.3 with parameters \varepsilon,d,M^\prime=1/\varepsilon,M^{\prime\prime}=1.

Suppose that G is an n-vertex oriented graph with \delta(G)\geqslant (1-\eta)n. Applying Lemma 3.3 with parameters \varepsilon,d,M^\prime=1/\varepsilon,M^{\prime\prime}=1, there is a partition V_0,V_1,\ldots,V_{k} of V(G) with k\geqslant 1/\varepsilon and |V_i|=m>n/(2M) for each i\in [k]. Let R be the reduced oriented graph of this partition. It follows from Lemma 3.4 that \delta(R)\geqslant (1-\eta-2d-3\varepsilon)k\geqslant (1-3d)k. Hence, we can greedily find a tournament on 2^{s-1} vertices as k\geqslant 1/\varepsilon\gg 2^{s-1}. Since 2^{s/2-1}\leqslant \overrightarrow{r}(s)\leqslant 2^{s-1}, it follows that R contains a copy of TT_s, say H. We may assume w.l.o.g. that V(H)=\{1,2,\ldots, s\} and that ij\in E(H) for all i<j. Applying Lemma 5.7 to the underlying graph of G[\cup_{i\in[s]}V_i] with parameters d and \delta, we obtain at least d^{s(s-1)/2}m^s/2 copies of TT_s in G. Since m\geqslant n/(2M), it follows that G contains at least \mu n^s copies of TT_s, completing the proof.

Given a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b\leqslant c, let h=a+b+c. Choose

Let \mathcal{P}=(U_1,U_2,\ldots,U_p) be a partition of V(G) with p\geqslant 2 and |U_i|\geqslant \alpha n for each i\in[p]. Applying Lemma 3.3 to G with parameters \varepsilon, d, M^\prime=1/\varepsilon, and M^{\prime\prime}=p, we obtain a refined partition X_0,X_1,\ldots,X_k of V(G), where 1/\varepsilon \leqslant k\leqslant M and |X_0|\leqslant \varepsilon n. Moreover, |X_i|=m for each i\in [k], and hence n=km+|X_0| implies that k\leqslant n/m\leqslant 2k\leqslant 2M. Let R^{\prime} be the reduced digraph of this partition, with vertex set [k]. For each i\in [p], set R_i=\{j\in [k]: X_j \subseteq U_i\}. Observe that \mathcal{P}_R=(R_1,R_2,\ldots,R_p) is a partition of [k]. Furthermore, we have

By Lemma 3.4, R^{\prime} contains a spanning oriented graph R with \delta ^0(R)\geqslant (1/2-\eta-d-3\varepsilon )k\geqslant (1/2-2d)k.

We now complete the proof of part (i) of Lemma 6.3. Applying Lemma 6.6 (ii) to R, there exists a directed triangle T in R with V(T)\cap R_1\neq\emptyset and V(T)\cap \overline{R_1}\neq\emptyset. We may assume w.l.o.g. that T=123, with 1\in R_1 and 2\in \overline{R_1}. Clearly, (X_1,X_2), (X_2,X_3), and (X_3,X_1) are (\varepsilon,d)-regular. Applying Lemma 5.5 three times with parameters \eta,\beta,\xi=1/(2M),\varepsilon,d, we see that, for each j\in[3], there are at least \beta n^{h} copies of D_{a,b,c} of type-(j,j+1,j+2) in G[X_1\cup X_2\cup X_3]. Let \mathcal{H}_j denote the family of copies of D_{a,b,c} of type-(j,j+1,j+2). Note that if a+1=b=c, then \mathcal{H}_1 and \mathcal{H}_2 yield two vectors \mathbf{v}_1,\mathbf{v}_2\in I^{\beta}_\mathcal{P}(D_{a,b,c}) such that \mathbf{v}_2-\mathbf{v}_1=\mathbf{e}_1-\mathbf{e}_2. If a+1=b+1=c, then \mathcal{H}_2 and \mathcal{H}_3 yield two vectors \mathbf{v}_2,\mathbf{v}_3\in I^{\beta}_\mathcal{P}(D_{a,b,c}) such that \mathbf{v}_2-\mathbf{v}_3=\mathbf{e}_1-\mathbf{e}_2. In either case, there exists a 2-transferral in L^{\beta}_{\mathcal{P}}(D_{a,b,c}), as required.

We now prove part (ii) of Lemma 6.3. Suppose that L_\mathcal{P}^\beta(D_{a,b,c}) is 2-transferral-free. We first claim that R contains no “cross” K_4^-.

For convenience, let A=R_1 and B=V(R)\setminus R_1. Then |A|,|B|\geqslant |R_i|\geqslant \alpha k/2. We first show that either \text{cyc}(A,A,B)\leqslant \sqrt d\, k^3 or \text{cyc}(B,B,A)\leqslant \sqrt d\, k^3 in R. Using this fact, we then prove that G is \gamma-extremal. We may assume w.l.o.g. that \text{cyc}(A,A,B)\leqslant \sqrt d\, k^3, since the argument in the case \text{cyc}(B,B,A)\leqslant \sqrt d\, k^3 is identical. Combining the fact that there is no K^-_4 of type-(A,V(R),V(R),\overline{A}) with Lemma 6.9 applied with \varepsilon=\sqrt d and \alpha=\alpha/2, we conclude that R admits a \gamma-extremal partition (V_1,V_2,V_3). Let V^\ast_i=\bigcup_{j\in V_i}X_j for each i\in[3], and put all vertices of X_0 into V^\ast_1. Clearly, (V^\ast_1,V^\ast_2,V^\ast_3) is a partition of V(G). A simple calculation shows that |V^\ast_i|=(k/3\pm O(\gamma k))m\pm |X_0|=n/3\pm O(\gamma n), since |V_i|=k/3\pm O(\gamma k). Combining e(V_{i+1},V_i)=O(\gamma k^2) with (iii) of Lemma 3.4, we obtain e(V^\ast_{i+1},V^\ast_i)=O(\gamma n^2), using the fact that |X_0|\leqslant \varepsilon n\ll \gamma n. Hence (V^\ast_1,V^\ast_2,V^\ast_3) is a \gamma-extremal partition of V(G), proving part (ii) of Lemma 6.3.

Thus, it remains to show that either \text{cyc}(A,A,B)\leqslant \sqrt d\, k^3 or \text{cyc}(B,B,A)\leqslant \sqrt d\, k^3 holds in R. If \text{cyc}(A,A,B)\leqslant \sqrt d\, k^3, then there is nothing to prove. Thus, we may assume w.l.o.g. that \text{cyc}(A,A,B)> \sqrt d\, k^3. Then there exists an edge uv\in E_R(A,B) such that |N_R^{-,+}(uv,A)|\geqslant \sqrt d\, k^3/(|A||B|)\geqslant \sqrt d\, k. Moreover, we have N_R^{-,+}(uv,B)=\emptyset, since otherwise there would exist a cross K_4^-, contradicting Claim ?. For each S\in\{A,B\}, let S_1=N_R^{-,+}(uv,S), S_2=N_R^{+,-}(uv,S), S_3=N_R^{+,+}(uv,S), and S_4=N_R^{-,-}(uv,S), and set S^r=S\setminus(S_1\cup S_2\cup S_3\cup S_4). Note that u\in A^r and v\in B^r. Clearly, |A_1|=|N_R^{-,+}(uv,A)|\geqslant \sqrt d\, k, |B_1|=|N_R^{-,+}(uv,B)|=0, and |A^r|+|B^r|\leqslant 8dk+2 since \delta^0(R)\geqslant(1/2-2d)k.

For disjoint sets X,Y\subseteq V(R), we write X \Rightarrow Y to mean that all edges between X and Y are directed from X to Y. We now determine the adjacency relations among several of these sets. It is easy to check that A_4\cup B_4\Rightarrow A_3\cup B_3, since otherwise R would contain a cross K_4^- of type-(u,A_3\cup B_3,A_4\cup B_4,v). Similarly, we have A_1\Rightarrow A_2\cup A_3 and B_2\cup B_4\Rightarrow A_1, since otherwise R would contain a cross K_4^- of type-(A_2\cup A_3,A_1,u,v) or type-(B_2\cup B_4,v,A_1,u). Note that for any pair of vertices x,y\in V(R) (possibly x=y), all but at most 8dk vertices of R are common neighbors of x and y. Thus x and y have common neighbors in A_1, since |A_1|\geqslant \sqrt d\, k>8dk. This implies that A_2\Rightarrow B_3. Indeed, suppose that b_3a_2 is an edge from B_3 to A_2. Then there exists a_1\in N_R^{-}(a_2,A_1), since d(a_2,A)>0 and A_1 \Rightarrow A_2. It follows that R[\{a_1,a_2,v,b_3\}] is a cross K_4^-, contradicting Claim ?. By the same argument, we also have B_2\Rightarrow A_3 and B_4\Rightarrow B_2, since otherwise there would exist a cross K_4^- of type-(A_1,A_3,B_2,v) or type-(A_1,u,B_2,B_4). In summary, we have

Next we claim that \text{cyc}(B,B,A)\leqslant \sqrt d\, k^3 in R. To prove this, we need the following result.

Since |A^r|+|B^r|\leqslant 8dk+2, there are at most 9dk^3 directed triangles containing at least one vertex from A^r\cup B^r. Then Claim ? implies that

It remains to show that \text{cyc}(B_3,B_3,A_3),\text{cyc}(B_4,B_4,A_4)\leqslant 17dk^3. Recall that A_1\cup A_4\cup B_4\Rightarrow A_3 and N_R^{-,+}(uv)\cup N_R^{-,-}(uv)=A_1\cup A_4\cup B_4\subseteq \overline{B_3}. Hence, the degree condition \delta^0(R)\geqslant(1/2-2d)k implies that d^-(x,\overline{B_3})\geqslant d_R^{-,+}(uv)+d_R^{-,-}(uv)-2dk for each x\in A_3. By Lemma 6.8 and the degree condition \delta^0(R)\geqslant(1/2-2d)k, we have d_R^{-,+}(uv)=d_R^{+,-}(uv)\pm 8dk and d_R^{-,-}(uv)=d_R^{+,+}(uv)\pm 8dk. Therefore,

By the degree condition again, we obtain d^-(x)\leqslant (1/2+2d)k, and hence d^-(x,B_3)\leqslant 17dk. It follows that e(B_3,A_3)\leqslant 17dk^2, which in turn implies that \text{cyc}(B_3,B_3,A_3)\leqslant 17dk^3. By symmetry, considering the out-degree of any vertex y\in B_4, we similarly obtain \text{cyc}(B_4,B_4,A_4)\leqslant 17dk^3, since B_4\Rightarrow A_1\cup A_3\cup B_3. In summary,

since d\ll 1. This completes the proof.

Given a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b\leqslant c, choose

where M is given by Lemma 3.3 with parameters \varepsilon,d, M^\prime=1/\varepsilon, and M^{\prime\prime}=2. Set \alpha=1/(4M).

Fix a vertex v\in V(G), and set A=N^{+}(v), B=N^{-}(v), and C=V(G)\setminus (A\cup B). Since \delta^{0}(G)\geqslant (1/2-\eta)n, we have |C|\leqslant 2\eta n. Hence

Applying Lemma 3.3 to G[A\cup B] with parameters \varepsilon,d, M^\prime=1/\varepsilon, and M^{\prime\prime}=2, we obtain a partition V_0,V_1,\ldots,V_k of A\cup B, where 1/\varepsilon \leqslant k\leqslant M, |V_j|=m\geqslant n/(2M) for each j\in[k], V_j\subseteq A for each j\in[i], and V_j\subseteq B for each i+1\leqslant j\leqslant k. Let R be the reduced oriented graph. By Lemma 3.4, we have

We may assume w.l.o.g. that R_A=\{1,2,\ldots,i\} and R_B=\{i+1,i+2,\ldots,k\}. Then (R_A,R_B) is a partition of V(R). Applying Lemma 6.6 (ii) to this partition, we find a triangle T in R such that V(T)\cap R_A\neq \emptyset and V(T)\cap R_B\neq \emptyset. We may assume that |V(T)\cap R_A|=2 and write T=12k. In particular, 1,2\in R_A and k\in R_B. Let S be the set of vertices u\in V_1 such that d^{+}(u,V_2)\geqslant dm/2 \text{ and } d^{-}(u,V_k)\geqslant dm/2. We claim that v is (D_{a,b,c},\beta,1)-reachable to every vertex in S.

Fix any u\in S, and set N_2=N^{+}(u,V_2) and N_k=N^{-}(u,V_k). By the definition of S, we have |N_i|\geqslant dm/2\geqslant dn/(4M) for each i\in\{2,k\}. Since \delta^0(G)\geqslant (1/2-\eta)n, every vertex of G misses at most 2\eta n edges. Hence, for some \sigma\in\{+,-\}, we have |N^\sigma(u,V_1)|\geqslant (m-2\eta n)/2\geqslant m/3, where we used m\geqslant n/(2M) and \eta\ll 1/M. Set N_1=N^\sigma(u,V_1).

Since 12k is a triangle in R, each of (V_1,V_2), (V_2,V_k), and (V_k,V_1) is (\varepsilon,d)-regular. By Lemma 3.2 and the fact that |N_1|,|N_2|,|N_k|\geqslant dm/2 with \varepsilon\ll d, it follows that (N_1,N_2), (N_2,N_k), and (N_k,N_1) are (\varepsilon^{1/2},d/2)-regular. Applying Lemma 5.5 with \xi=d/(4M), we conclude that G[N_1\cup N_2\cup N_k] contains at least \beta n^{a+b+c-1} copies of D_{a-1,b,c} of type (N_1,N_2,N_k) (where possibly a-1=0). Note that N_1\subseteq V_1\subseteq N^+(v), N_2\subseteq V_2\subseteq N^+(v), N_k\subseteq V_k\subseteq N^-(v). Therefore, every such copy of D_{a-1,b,c} together with v and together with u, respectively, forms a copy of D_{a,b,c}. This proves the claim.

Finally, since both (V_1,V_2) and (V_k,V_1) are (\varepsilon,d)-regular, Lemma 3.1 together with m\geqslant n/(2M) and \alpha=1/(4M) implies that |S|\geqslant (1-2\varepsilon)m\geqslant \alpha n. Hence, v is (D_{a,b,c},\beta,1)-reachable to every vertex in S, and this completes the proof.

Let 0<\eta\leqslant 1/100, and let G be an n-vertex oriented graph with \delta^{0}(G)\geqslant (1/2-\eta)n. Suppose that A and B are two disjoint subsets of V(G).

To prove (i), we may assume that |A|=(1/4+2\eta)n and |B|=(1/2-\eta)n. Since G[B] is oriented, there exists a vertex b\in B with d^-(b,B)\leqslant |B|/2. It follows from \delta^0(G)\geqslant (1/2-\eta)n that d^-(b,A)\geqslant (1/2-\eta)n-(n-|A|-|B|)-|B|/2>0. Thus, there is an edge from A to b, and hence from A to B. Similarly, there exists b^\prime\in B with d^+(b^\prime,B)\leqslant |B|/2, and therefore G contains an edge from B to A.

Now we prove (ii). Let A,B be a partition of V(G), and assume w.l.o.g. that |A|\leqslant |B|. If there exists a\in A with d^+(a,B)\geqslant (1/4+2\eta)n, then Lemma 6.6 (i) implies that there is an edge from N^+(a,B) to N^-(a), which together with a forms the desired triangle. Thus, we may assume that d^+(a,B)<(1/4+2\eta)n for every a\in A. By symmetry, d^-(a,B)<(1/4+2\eta)n for every a\in A. Since V(G)=A\cup B and \delta^0(G)\geqslant (1/2-\eta)n, it follows that d^\pm(a,A)>(1/4-3\eta)n for every a\in A, and hence

Fix any u\in A, and let B_1=N^+(u,B), B_2=N^-(u,B), B_3=B\setminus(B_1\cup B_2). Since d^\pm(u,B)<(1/4+2\eta)n, we have |B_1|,|B_2|<(1/4+2\eta)n. Moreover, \delta^0(G)\geqslant (1/2-\eta)n implies that |B_3|\leqslant 2\eta n. We may assume that e(B_1,B_2)=0, since otherwise any edge in E(B_1,B_2) together with u forms the desired C_3. Now B_2 contains a vertex, say w, with d^{-}(w,B_2)\leqslant |B_2|/2. Therefore,

By Lemma 6.6 (i), there is an edge from N^+(w) to N^-(w,A), which together with w forms the desired triangle.

Observe that d^{\sigma,\tau}(xy)=d^\sigma(x)-d^{\sigma,-\tau}(xy)-|N^\sigma(x)\setminus N(y)| and d^{-\sigma,-\tau}(xy)=d^{-\tau}(y)-d^{\sigma,-\tau}(xy)-|N^{-\tau}(y)\setminus N(x)|. The claim therefore follows immediately from the fact that G is an oriented graph with \delta^0(G)\geqslant(1/2-\eta)n.

We choose \beta such that \varepsilon \ll \beta \ll \gamma. For simplicity, we write \text{SN}^\sigma_\beta(A) as S^\sigma for \sigma\in\{+,-\}. Let S=V(G) \backslash(S^+\cup S^-\cup A). Applying Lemma 6.7 with \gamma=\beta and using the fact that \text{cyc}(A,A,\overline{A})\leqslant \varepsilon n^3, we obtain |S^-|=|S^+|\pm 2\beta n, |A|\leqslant(1/3+\beta)n, and |S|\leqslant2\beta n.

Next we claim that |A\cup S|, |S^+|, and |S^-| are all approximately equal. Observe that \text{cyc}(S^+, S^+, A)\leqslant \beta n \cdot \binom{|S^+|}{2} \leqslant \beta n^3 and \text{cyc}(S^+, S^+, S)\leqslant 2\beta n^3 since |S|\leqslant 2\beta n. For any vertices x\in S^+ and y\in S^-, we have d^-(x,A),d^+(y,A)\geqslant |A|-\beta n by the definition. Thus, at least |A|-2\beta n vertices in A form a directed triangle of type-(A,S^+,S^-) together with x and y. Therefore,

which implies that \text{cyc}(S^+,S^+,S^-)=\text{cyc}(S^+,S^-,S^+)\leqslant\varepsilon n^3 since |A|\geqslant \alpha n. Thus,

Applying Lemma 6.7 with (A,\varepsilon,\gamma)=(S^+,4\beta,\gamma), we obtain |S^+|\leqslant (1/3+\gamma)n. Moreover, since |S^-|=|S^+|\pm 2\beta n, we also have |S^-|\leqslant (1/3+2\gamma)n. Hence, (1/3-\gamma)n\leqslant |A\cup S|\leqslant (1/3+3\beta)n. It is then easy to check that |S^+|,|S^-|,|A\cup S|=n/3\pm O(\gamma n).

We now claim that G is \gamma-extremal by showing that e(S^+,A\cup S),\ e(A\cup S,S^-),\ e(S^-, S^+) = O(\gamma n^2). Recall that for each x\in S^+ and y\in S^-, we have d^-(x,A),d^+(y,A)\geqslant |A|-\beta n. Thus, e(S^+,A\cup S)\leqslant (\beta n+|S|)\cdot |S^+|\leqslant 3\beta n|S^+|=O(\gamma n^2). Similarly, we have e(A\cup S,S^-)=O(\gamma n^2). The degree condition \delta^0(G)\geqslant (1/2-\eta)n implies that every vertex has at most 2 \eta n non-neighbors in G. Hence,

which completes the proof.

Let 0<1/n\ll\eta\ll\gamma\ll 1, and let G be an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta)n. Fix a \gamma-extremal partition (V_1,V_2,V_3) of V(G). Since \delta^0(G)\geqslant (1/2-\eta)n, every vertex v\in V(G) misses at most 2\eta n edges in G. Therefore, for each i\in[3], we have

We claim that for each i\in[3], there are at most \gamma^{1/3}n/6 vertices v\in V_i with d^+(v,V_{i+1})\leqslant |V_{i+1}|-\gamma^{1/2}n. Indeed, this follows immediately from the fact that every such v contributes at least \gamma^{1/2}n missing edges to V_{i+1}, together with \gamma\ll 1. By symmetry, for each i\in[3], there are at most \gamma^{1/3}n/6 vertices in V_{i+1} whose in-degree from V_i is at most |V_i|-\gamma^{1/2}n. This completes the proof.

Let 0<1/n\ll\eta\ll\gamma\ll 1, and let G be an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta)n. Fix a \gamma-extremal partition (V_1,V_2,V_3) of V(G). By Lemma 7.1, the number of \gamma^{1/2}-bad vertices with respect to this partition is at most \gamma^{1/3}n. Let v be any such bad vertex.

It follows from the degree condition that d^+(v,V_i),d^-(v,V_j)\geqslant n/6-O(\gamma n) for some i,j\in[3]. Moreover, since v has at most 2\eta n missing edges in G, ? implies that for each i\in[3], either d^+(v,V_i) or d^-(v,V_i) is at least n/6-O(\gamma n). Hence there exists i\in[3] such that d^-(v,V_{i-1})\geqslant n/6-O(\gamma n) and d^+(v,V_{i+1})\geqslant n/6-O(\gamma n). We then reassign v to the set V_i.

After reassigning all \gamma^{1/3}n bad vertices, we obtain a \gamma^{1/3}-superextremal partition of V(G), which proves the lemma.

Given a,b,c\in \mathbb{N} with 1\leqslant a\leqslant b\leqslant c, let 0<1/n\ll\eta\ll \beta\ll\gamma\ll 1. Suppose that G is an n-vertex oriented graph with \delta^0(G)\geqslant (1/2-\eta)n and that (V_1,V_2,V_3) is a \gamma-extremal partition of V(G). By ?, we have |V_i|=n/3\pm O(\gamma n) for each i\in[3].

By Lemma 7.1, all but at most \gamma^{1/3}n vertices of V(G) are \gamma^{1/2}-good with respect to the partition (V_1,V_2,V_3). It is not difficult to check that (V_i,V_{i+1}) is (\gamma^{1/4},1/2)-regular for each i\in[3]. Applying Lemma 5.5 to (V_i,V_{i+1},V_{i+2}) with parameters \gamma^{1/4},1/2 and (a+b+c)\beta, there are at least (a+b+c)\beta n^{a+b+c} copies of D_{a,b,c} of type-(i,{i+1},{i+2}). Since each copy of D_{a,b,c} intersects at most (a+b+c)n^{a+b+c-1} other copies of D_{a,b,c}, there are at least \beta n disjoint copies of D_{a,b,c} in G. This completes the proof.

We may assume w.l.o.g. that i=1. Let B be the set of \gamma^{1/2}-bad vertices in G. Since (V_1, V_2, V_3) is a \gamma-superextremal partition of V(G), we have |B|\leqslant \gamma^{1/3}n by Lemma 7.1. Moreover, ? shows that |V_i|=n/3\pm O(\gamma n) for each i\in[3]. We first show that the vertex set of every edge in E(V_2,V_1) is \gamma-transfer for (V_1,V_3). To see this, we need the following two claims.

Consider a fixed edge xy in E(V_2,V_1). Now we claim that \{x,y\} is \gamma-transfer for (V_1,V_3). We may assume that d^{-,+}(xy,V_1)<\gamma^{1/4}n since otherwise there is nothing to prove by Claim ?. Furthermore, if d^+(y,V_{3})\geqslant2\gamma^{1/4}n, then Claim ?, ? and 2 show that \{x,y\} is \gamma-transfer for (V_1,V_3). Therefore, we may further assume that d^+(y,V_{3})<2\gamma^{1/4}n. Clearly, d^+(y,V_1)\geqslant \lfloor(n-1)/2\rfloor-|V_2|-2\gamma^{1/4}n and d^{-,+}(xy,V_1)\geqslant d^-(x,V_1)+d^+(y,V_1)-|V_1|. Combining ? and d^{-,+}(xy,V_1)<\gamma^{1/4}n, we get that d^-(x,V_1)<n/6+O(\gamma^{1/4}n). Thus by ? and 2 we have

For any subset A of V(G)\backslash \{x,y\} with |A|=O(\gamma n), set J=N^+(x,V_1)\backslash (A\cup B). It follows from |A\cup B|\leqslant 2\gamma^{1/3}n that |J|= n/6\pm O(\gamma^{1/4}n). As G[J] is oriented, there exists y^{\ast}\in J such that d^+(y^{\ast},J)\leqslant |J|/2\leqslant n/12+O(\gamma^{1/4}n). Moreover, since y^{\ast}\notin B, we have d^+(y^{\ast},V_1)\geqslant \lfloor (n-1)/2\rfloor-|V_2|-O(\gamma n)\geqslant n/6-O(\gamma n). This implies that

Applying Claim ? again, we obtain that \{x,y^\ast\} is \gamma-transfer for (V_1,V_3). From Definition 4 and the fact that x, y^\ast \notin A, there exists a set S\subseteq V(G)\backslash A with the required property. Consequently, \{x,y\} is \gamma-transfer for (V_1,V_3).

Similarly, by computing d^\pm(x,V_3) and d^{-,+}(xy,V_2), one can prove that \{x,y\} is \gamma-transfer for (V_2,V_3), which completes the proof.

Note that (c^2-ab)-(a^2-bc)=(c-a)(a+b+c)=(c-a)h. Hence h\mid\bigl((c^2-ab)-(a^2-bc)\bigr), and therefore c^2-ab\equiv a^2-bc \pmod h. Similarly, c^2-ab\equiv b^2-ac\pmod h.

First assume that \gcd(h,\Delta)=1. Then there exists a unique integer \Delta^{-1}\in [h-1] such that \Delta\cdot \Delta^{-1}\equiv 1\pmod h. We claim that (x_0,y_0,z_0)=\bigl((a+c)\Delta^{-1},-(a+b)\Delta^{-1},0\bigr) is an integer solution to (A.1). Indeed, ax_0+cy_0+bz_0=\Delta^{-1}(a^2-bc)\equiv \Delta^{-1}\Delta\equiv 1\pmod h, so the first congruence in (A.1) holds. The second and third congruences can be verified similarly.

Now assume that (A.1) has an integer solution (x_1,y_1,z_1). Suppose for a contradiction that g=\gcd(h,\Delta)>1. Since the congruences in (A.1) hold modulo h and g\mid h, reducing them modulo g yields

Using the first two congruences, we obtain (a^2-bc)x_1+(ab-c^2)z_1 = a(ax_1+cy_1+bz_1)-c(bx_1+ay_1+cz_1) \equiv a+c\pmod g. Since c^2-ab\equiv a^2-bc\pmod h and g=\gcd(h,c^2-ab), it follows that c^2-ab\equiv a^2-bc \equiv 0\pmod g. Hence a+c\equiv(c^2-ab)(x_1-z_1)\equiv 0\pmod g.

Similarly, c(ax_1+cy_1+bz_1)-b(bx_1+ay_1+cz_1)\equiv b+c\pmod g, and together with c^2-ab\equiv b^2-ac\pmod h g=\gcd(h,c^2-ab), this implies that b+c\equiv 0\pmod g. Note that we also have h=a+b+c\equiv 0\pmod g. Therefore, a\equiv b\equiv c\equiv 0\pmod g. Substituting these congruences into the first congruence above yields 0\equiv 1\pmod g, a contradiction. Thus g=\gcd(h,\Delta)=1, completing the proof.