Corrádi and Hajnal’s theorem for sparse random graphs

Let \delta = \delta_{\ref{thm_decomposition}}(\gamma), \varepsilon = \varepsilon_{\ref{thm_triangleblowup}}(\delta), \xi = \xi_{\ref{thm_decomposition}}(\delta,\varepsilon) and C=\max\{C_{\ref{thm_decomposition}}(\varepsilon),C_{\ref{thm_triangleblowup}}(\varepsilon, \xi)\}, D=D_{\ref{thm_decomposition}}(\varepsilon). By Theorem 6.6, there exist set B, S, and (W_i)_{i=1}^{3k} which satisfies (i) - (iv) of Theorem 6.6. Furthermore for each t \in [k], by Theorem 1.4, each (\delta,\varepsilon,p)-strong-super-regular triple (W_{3t-2}, W_{3t-1}, W_{3t}) contains a perfect triangle packing. Therefore all the vertices except B can be covered by vertex-disjoint triangles. Since |B| \le Dp^{-2}, this completes the proof.

Let \eta=1/140, \varepsilon_1 = \varepsilon_{\ref{prop_findtriangle}}(3\eta/2, \delta/2), and \varepsilon=(\eta/4)\min\{\varepsilon_{\ref{lemma_main}}(\delta, \eta), \varepsilon_1\}. Let C = C_{\ref{lemma_main}}(\delta, \eta, \xi). Let (W_1, W_2, W_3) be a (\delta, \varepsilon, p)-strong-super-regular triple and m = |W_1| = |W_2| = |W_3|. By Lemma 7.1, there exists a matching M_{12}, M_{13}, M_{23} and sets Q_1, Q_2, Q_3 and X_1, X_2, X_3 satisfying (P1), (P2), (P3), and (P4). Let W_i' = W_i \setminus (Q_i \cup X_i) and note that

Let x = m - |M_{12}| - |M_{13}| - |M_{23}| and note that (1 - 6\eta)m < x < (1 - 3\eta/2)m. By applying Proposition 2.7, we can find x vertex-disjoint triangles inside the triple (W_1', W_2', W_3').

Note that the remaining |W_1'| - x = |M_{23}| - |X_1| vertices in W_1' together with the set X_1, can be matched with the set M_{23} to construct |M_{23}| vertex-disjoint triangles, by property (P4). Similarly, the remaining vertices in W_2' \cup X_2 and W_3' \cup X_3 can be matched with M_{13} and M_{12}, respectively. Therefore, we have found a perfect triangle packing of (W_1, W_2, W_3).

By regularity of the pair (V_1, V_2), for every pair of subsets V_i'' \subset V_i' such that |V_i''| \ge (\varepsilon_1/\varepsilon_2)|V_i'| \ge \varepsilon_1|V_i| for i \in \{1,2\}, we have

Since \max\{\varepsilon_1/\varepsilon_2, 2\varepsilon_1\} = \varepsilon_1/\varepsilon_2, the pair (V_1', V_2') is (\varepsilon_1/\varepsilon_2,p)-regular. The density condition immediately follows from the definition of regularity.

For i \in \{1, 2\}, let U_i be a subset of V_i of size at least \varepsilon|V_i|. Note that

The conclusion easily follows from the triangle inequality.

Note that without loss of generality, we may assume that \alpha \leq 1/2. Furthermore, let \varepsilon = \min\{\delta/4, \alpha/20\} and let \varepsilon' = 2\varepsilon/\alpha. Let us greedily remove triangles from (V_1, V_2, V_3) until we cannot do it anymore and denote the remaining triple by (V'_1, V'_2, V'_3). If |V'_i| \leq \alpha|V_i| for some i, then there is nothing left to prove, so we may assume that |V'_i| > \alpha|V_i| for all i. Let W_i be the set of all those vertices in V'_i that were \varepsilon-typical in the original triple and note that |W_i| \geq |V'_i| - \varepsilon|V_i| \geq (\alpha/2)|V_i|. By Proposition 2.1, the triple (W_1, W_2, W_3) is (\varepsilon', p)-regular and the density of each pair (W_i, W_j) is at least (d_{ij} - \varepsilon)p. Since \varepsilon' < 1/2, there is a vertex v \in W_1 with \deg(v,W_i) \geq (d_{ij}-\varepsilon-\varepsilon')p|W_i| for i \in \{2, 3\}. For i \in \{2, 3\}, let N_i and N'_i be the sets from the definition of an \varepsilon-typical vertex for v and let M_i = N_i \cap W_i and M'_i = N'_i \cap W_i. Since

and (N'_2, N'_3) was (\varepsilon,p)-regular with density at least (1-\varepsilon)\delta p and (1-\varepsilon)\delta p > \varepsilon p, the pair (M'_2, M'_3) has positive density. It follows that (W_1, W_2, W_3) contains a triangle, but this is impossible.

Let v \in V_1 be any such vertex and for j \in \{2, 3\}, let N_j = N(v) \cap V_j. Since v is \varepsilon-typical, (1-\varepsilon)d_{1j}p|V_j| \leq |N_j| \leq (1+\varepsilon)d_{1j}p|V_j|. Moreover, there exist subsets N'_j \subset N_j satisfying |N'_j| \geq (1-\varepsilon)|N_j| such that (N'_2,N'_3) is an (\varepsilon,p)-regular pair with density d_vp, where d_v \in [(1-\varepsilon)d_{23}, (1+\varepsilon)d_{23}]. Moreover, let M_j = N_j \setminus X_j and similarly let M'_j = N'_j \setminus X_j.

For each i, let W_i = V_i \setminus X_i and recall that (1-\varepsilon)|V_i| \leq |W_i| \leq |V_i|. For every i and j, let d'_{ij}p be the density of the pair (W_i, W_j). Since (V_i,V_j) is (\varepsilon, p)-regular, d'_{ij} \in [d_{ij} - \varepsilon, d_{ij} + \varepsilon]. It follows that

And by the given condition \deg(v, X_j) \le \varepsilon p |V_j|, we have

Moreover, since |N_j| \geq (1-\varepsilon)\delta p|V_j|, we have

By Proposition 2.1, the pair (M'_2, M'_3) is (2\varepsilon,p)-regular with density d'_vp satisfying

Therefore, if \varepsilon is sufficiently small, then \deg(v,W_j) = |M_j| \in [(1-\varepsilon')pd'_{1j}|W_j|, (1+\varepsilon')pd'_{1j}|W_j|], |M'_j| \geq (1-\varepsilon')|M_j|, and (M'_2, M'_3) is (\varepsilon',p)-regular with density d'_vp, where d'_v \in [(1-\varepsilon')d'_{23}, (1+\varepsilon')d'_{23}]. It follows that v is \varepsilon'-typical in (W_1, W_2, W_3).

For an \varepsilon-typical vertex v \in V_2, let N_1 = N(v) \cap V_1 and N_3 = N(v) \cap V_3. Recall from Definition 2.4 that |N_i| \geq (1-\varepsilon)d_{2i}p|V_i| and that there exist N_i' \subset N_i with |N_i'| \ge (1-\varepsilon)|N_i| for i \in \{1,3\} such that (N_1', N_3') is (\varepsilon,p)-regular and has density at least (1-\varepsilon)d_{13}p. It follows that at least (1-\varepsilon)|N_3'| vertices w \in N_3' have at least ((1-\varepsilon)d_{13} - \varepsilon)p|N_1'| common neighbours with v in N_1'. Since

and 3\varepsilon + \varepsilon/\delta \le \varepsilon', each such edge \{v,w\} is \varepsilon'-good. Since there are at least (1-\varepsilon)|V_2| typical vertices in V_2 and

the total number of \varepsilon'-good edges between V_2 and V_3 is at least (1-4\varepsilon)d_{23}p|V_2||V_3|. Finally, since the number of edges between V_2 and V_3 is exactly d_{23}p|V_2||V_3|, the total number of non-\varepsilon'-good edges is at most 4\varepsilon d_{23}p|V_2||V_3|.

Let \varepsilon = \min\{\varepsilon'/4, \varepsilon'\delta/4, (\varepsilon'\delta)^2/32\}. By Proposition 2.12, at most 4\varepsilon d_{23}p|V_2||V_3| edges in (V_2,V_3) are not \varepsilon'-good. Let \alpha|V_1| be the number of \varepsilon-typical vertices in V_1 that are not \varepsilon'-good. By definition, the neighbourhood of every such vertex contains at least \varepsilon'd_{12}d_{13}d_{23}p^3|V_2||V_3| edges that are not \varepsilon'-good. Therefore, our assumption on the maximum number of common neighbours of the endpoints of edges in (V_2, V_3) implies that

and hence \alpha \leq 16\varepsilon/(\varepsilon'd_{12}d_{13}) \leq \varepsilon'/2. Finally, since at most \varepsilon|V_1| vertices in V_1 are not \varepsilon-typical and \varepsilon \leq \varepsilon'/2, the number of vertices in V_1 that are not \varepsilon'-good is at most \varepsilon'|V_1|.

Fix an \varepsilon-typical vertex v \in V_1. For each j \in \{2,3\}, let N_j = N(v) \cap V_j. Since v is \varepsilon-typical, there are N'_j \subset N_j with |N'_j| \geq (1-\varepsilon)|N_j| \geq (1-\varepsilon)^2pd_{1j}|V_j| such that (N'_2,N'_3) is (\varepsilon,p)-regular with density d'_{23}p, where (1-\varepsilon)d_{23} \leq d'_{23} \leq (1+\varepsilon)d_{23}. Let N''_j be the set of vertices in N'_j that have at most (d'_{23}+\varepsilon)p|N'_k| neighbours in N'_k and note that |N''_j| \geq (1-\varepsilon)|N'_j| by (\varepsilon, p)-regularity of (N'_2, N'_3). Since 1/d'_{23} \leq 1/((1-\varepsilon)d_{23}) \leq 2/\delta, we have

and then (?) follows. Note that |N''_j| \geq (1-\varepsilon)|N'_j| \geq (1-\varepsilon)^2d_{1j}p|V_j| and

Moreover, if v is \varepsilon-good, then at most \varepsilon d_{12}d_{13}d_{23}p^3|V_2||V_3| edges in (N''_2,N''_3) are not \varepsilon-good. Now (?) follows.

To prove (i), as a first step, we fix a set W \subset V. We may assume that Dp^{-1}n/|W| \le n as otherwise, the claim is vacuously true. Suppose that there are Dp^{-1}n/|W| vertices v \in V \setminus W such that \deg(v,W) \neq (1 \pm \rho)|W|p. Then there exists a set B \subset V \setminus W of size Dp^{-1}n/(2|W|) such that either \deg(v,W) > (1+\rho)|W|p for all v \in B or \deg(v,W) < (1-\rho)|W|p for all v \in B. This clearly implies that e(B,W) \neq (1 \pm \rho)|B||W|p for some B as above. Since e(B, W) is a sum of independent binomial random variables and \mathbb{E}[e(B,W)] = |B||W|p = Dn/2, by Chernoff’s inequality,

By the union bound, the probability that such a set B exists is at most 2^ne^{-\rho^2Dn/12}.

Now that (i) is proved, we easily get (ii) by applying the union bound.

Fix x, W, and E as in the statement of this proposition. For a vertex w \in W and an edge \{u,v\} \in E, let B(u,v,w) denote the event that w is adjacent to both u and v. Let X be the random variable denoting the number of events B(u,v,w) that occur in G(n,p). Note that each of the “bad” events described in the statement of this lemma implies that X \geq \xi np^2x \geq 2x^2p^2. Moreover, observe that X \leq x^2, so we can restrict our attention to the case x \geq \xi np^2. Since all B(u,v,w) are mutually independent, X has binomial distribution with parameters x^2 and p^2, and hence by Chernoff’s inequality,

for some absolute positive constant c. Since for each x, there are at most {n \choose x}n^{2x} pairs (W,E) with |W|=|E|=x, the probability that some “bad” event occurs is at most

Finally, note that

provided that np^2 \geq (4/c\xi)\log n.

The constant D(\varepsilon, \rho) will be chosen later. Let X be a fixed set of size at least Dp^{-2}. Without loss of generality we may assume that \rho \leq 1/2.

First expose the edges between X and V \setminus X and call a pair of vertices \{v,w\} \in V \setminus X bad if v and w do not have (1 \pm \rho)|X|p^2 common neighbours in X. By Proposition 4.2 (i), with probability 1 - e^{(1 - \rho^2D^{1/2}/50)n}, there are at most D^{1/2}p^{-1}n/|X| vertices that do not satisfy \deg(v, X) = (1 \pm \rho/2)|X|p. Even if each of these vertices forms bad pairs with all n vertices, there are at most D^{1/2}p^{-1}n^2/|X| such bad pairs. For each vertex that satisfies \deg(v, X) = (1 \pm \rho/2)|X|p, again by Proposition 4.2 (i), with probability 1 - e^{(1 - \rho^2D^{1/2}/50)n}, there are at most 2D^{1/2}p^{-1}n/(|X|p) other vertices w which do not have (1 \pm \rho)|X|p^2 common neighbours with w in X. Therefore, if D is sufficiently large, then with probability at least 1 - e^{-2n}, the total number of bad pairs is at most

Finally, expose the edges within V \setminus X. By Chernoff’s inequality, with probability 1 - e^{-\varepsilon n^2p/20}, at most \varepsilon n^2p bad pairs will form an edge.

Since n^2p \gg n, if we fix the set X, both of the above events happen with probability at least 1 - e^{-2n}. Since there are at most 2^n choices for X, we can take the union bound over all choices of X to derive the conclusion.

For the sake of brevity, denote G(n,p) by G and let V = V(G). Let \varepsilon be a small positive constant (we will fix it later), let C = C_{\ref{prop_randomgraphproperties1}}(\varepsilon), and let D_0(\varepsilon) be a constant satisfying D_0 \ge \max \{D_{\ref{prop_nontypicalvertices} (ii)}(\varepsilon, \varepsilon), D_{\ref{prop_randomgraphproperties2}}(\varepsilon, \varepsilon), 1\}.

Without loss of generality we may assume that |T| = \varepsilon n. Let X_0 \subset V \setminus T be an arbitrary set of size 2D_0p^{-2}. By assuming that the events from Propositions 4.1, 4.2 (ii), and 4.4 hold, we will show that there exists a triangle in G' which intersects X_0 but not T. This will prove that there are at most 2D_0p^{-2} vertices that are not contained in triangles that do not hit T. Let T' be the collection of all the vertices v that satisfy \deg(v, X_0 \cup T) \ge (1 + \varepsilon)|X_0 \cup T|p and note that |T'| \le D_0p^{-1} by Proposition 4.2 (ii). Let T'' = T \cup T' and X = X_0 \setminus T'. Note that |T''| \le 2\varepsilon n and |X| \ge |X_0| - |T'| \ge D_0 p^{-2}. Let D be the constant defined by |X| = Dp^{-2} and note that D \ge D_0. It suffices to show that there exists a triangle in G' which contains a vertex from X but not from T''.

Let Y = V \setminus (X \cup T'') and fix a vertex x \in X. Note that

where the last two inequalities follow from the fact that x \not\in T' and our assumption on p. Finally, let N_x = N_G(x) \cap Y and fix an arbitrary subset N_x' \subset N_{G'}(x) \cap Y of size (2/3 - 3\varepsilon)np.

It suffices to show that the number of triangles xy_1y_2 in G' such that x \in X and y_1, y_2 \in N_x' is nonzero. Let this number be M. To bound M from below, first bound the number of triangles xy_1y_2 in G such that x \in X, y_1, y_2 \in N_x, and y_1y_2 is an edge of the graph G' (we will later subtract the number triangles whose y_1 or y_2 is not in N_x'). Let this number be M_0. Since |X \cup T''|, |Y| \geq \varepsilon n, by Proposition 4.1 (iii), we have

By Proposition 4.4, the number of edges \{v, w\} in G'[Y] that form a triangle in G with fewer than (1 - \varepsilon)D vertices in X is at most \varepsilon n^2p given that D_0 is large enough. Thus,

To obtain a bound on M from M_0, we can subtract the number of triangles xy_1y_2 as above such that either y_1 or y_2 is not in N_x'. Since |N_x| \leq (1 + \varepsilon)np by Proposition 4.1 (i), we have

Thus, if \varepsilon is small enough, by Proposition 4.1 (iii) we have,

Therefore there exists a triangle as claimed, provided that \varepsilon is sufficiently small.

Let C be a constant satisfying C \geq 2 and e^{C^2/15} \geq 8e/\varepsilon. If p \geq (\log n/n)^{1/2}, then by Proposition 4.1, a.a.s. \delta(G(n,p)) \geq (1-\varepsilon/4)np and each pair of vertices of G(n,p) has at most 2np^2 common neighbours. If Cn^{-1/2} \leq p < (\log n/n)^{1/2}, then still a.a.s. \delta(G(n,p)) \geq (1-\varepsilon/4)np, but G(n,p) may contain some edges whose endpoints have more than 2np^2 common neighbours. Let H be the subgraph consisting of all such edges, and let v be an arbitrary vertex. By Chernoff’s inequality, the probability that v and some other vertex have more than 2np^2 common neighbours is at most e^{-np^2/15}. Therefore,

so a.a.s. \Delta(H) < (\varepsilon/4)np. Finally, let G = G(n,p) - H. Clearly, the endpoints of every edge of G have at most 2np^2 common neighbours. Moreover, by Proposition 4.1 (i), we may assume that \delta(G) > (1-\varepsilon/2)np.

Let X be an arbitrary fixed set of (\varepsilon/4)p^{-2} vertices of G and let W = \{v \notin X \colon \deg(v,X) \geq 2|X|p\}. By Chernoff’s inequality, the probability that a vertex v belongs to W is e^{-\Omega(p^{-1})} and these events are independent for different vertices. Since p \gg e^{-\Omega(p^{-1})}, Chernoff’s inequality implies that a.a.s. |W| \leq (\varepsilon/4)np. Moreover, since our assumption on p implies that |X| \leq (\varepsilon/8) n, we can apply Chernoff’s inequality and deduce that a.a.s. \deg(u,X) \leq (\varepsilon/4) np for every vertex u.

Let G' be the subgraph of G obtained by deleting all edges within X, all edges between X and W, and deleting edges incident to any y \notin X \cup W according to the following rule – for every triangle xyz in G with x \in X and z \notin X \cup W, remove the edge yz. It is quite easy to see that no vertex of X is contained in a triangle in G'. Let us now estimate \delta(G'). Since a vertex u \in X \cup W lost only edges connecting it to X and W, we have

Since a vertex y \notin X \cup W is incident to at most (\varepsilon/2)p^{-1} vertices x \in X and it has at most 2np^2 common neighbours with each such x, we then have

Thus G' has the required properties.

Let \varepsilon_1 = \varepsilon_{\ref{prop_findtriangle}}(\frac{1}{2},\delta) and \varepsilon_0 = \min \{ \frac{\varepsilon'}{4(3/\gamma + 1)}, \varepsilon_{\ref{prop_randomgraphtypical}}(\varepsilon_1,\frac{\delta}{2}), \frac{\delta}{2}, \varepsilon_1 \}. Assume that \varepsilon \in (0, \varepsilon_0) is given.

Let C_t = \{3t-2, 3t-1, 3t\} for t \in [k] be triangles of the graph R. For each vertex i \in [3k], call an index t \in [k] i-rich or rich with respect to i if i is adjacent to all three vertices of C_t, and assume that there are g_i i-rich indices. Then by the minimum degree condition on R, we have

which is equivalent to g_i \ge \gamma k. Thus for each vertex i \in [3k] of R, we can assign an i-rich index t \in [k] to it so that every index in [k] is chosen by at most (3k)/(\gamma k) = 3/\gamma vertices.

Consider the following process that adjusts the parts one by one. Throughout the process, we will maintain sets B \subset V(G) and Z_i \subset V_i for each i \in [3k]; they are empty at the beginning. Call a triangle C_t balanced if the sets V_{3t-2} \setminus Z_{3t-2}, V_{3t-1} \setminus Z_{3t-1}, and V_{3t} \setminus Z_{3t} have equal cardinalities. Assume that the triangles C_1, \ldots, C_{t-1} are already balanced and we are trying to balance the triangle C_t. Without loss of generality, we may assume that |V_{3t-i} \setminus Z_{3t-i}| - |V_{3t} \setminus Z_{3t}| = x_i m and 0 \le x_i \le 2\varepsilon for i \in \{1,2\}. Thus we have to remove x_i m vertices from V_{3t-i} for i \in \{1,2\} in order to make C_t balanced. By moving at most 2 arbitrary vertices from each set V_{3t-1} and V_{3t-2} to B and also to Z_{3t-1} and Z_{3t-2}, respectively, we may assume that both x_1 m and x_2 m are divisible by 3. First consider the set V_{3t-1} and let s be the rich index with respect to 3t-1 which we have chosen above. As we will later establish, for every i \in [3k], |V_{i} \setminus Z_i| \ge m/2 throughout the process. Therefore by Proposition 2.1, the triple (V_{3t-1} \setminus Z_{3t-1}, V_{3s-j} \setminus Z_{3s-j}, V_{3s-k} \setminus Z_{3s-k}) inherits the regularity of (V_{3t-1}, V_{3s-j}, V_{3s-k}) and is always (2\varepsilon, p)-regular of density at least \delta/2 for every pair \{j, k\} \subset \{0,1,2\}. By Proposition 4.8, a.a.s. it must also be \varepsilon_1-typical. Thus by Proposition 2.7 we can find x_1 n /3 triangles across this triple. Do this for each pair \{j, k\} and update the sets Z_{3t-1}, Z_{3s-2}, Z_{3s-1}, and Z_{3s} by placing all the vertices of these triangles into corresponding parts. Note that even though the sizes of the sets in C_s have decreased, the number by which they decreased is the same for all three of them and thus after performing the same procedure for V_{3t-2}, the triangles C_1, \ldots, C_t will be balanced.

Note that in the end, |B| \le 4k. Moreover, throughout the process, by the restriction that every index is the chosen rich index for at most 3/\gamma other indices, we always have, |Z_i| \le 2\varepsilon m \cdot (3/\gamma + 1) \le \min\{\varepsilon' m, m/2\} as claimed. Define S = \big( \cup_{i=1}^{3k} Z_i \big) \setminus B and we have the sets B and S as claimed.

Given a \gamma, let \delta = \delta_{\ref{prop_reducedmindegree}}(\gamma) and \varepsilon_0 = \varepsilon_{\ref{prop_reducedmindegree}}(\gamma). Moreover, for a given \varepsilon \in (0,\varepsilon_0), let \varepsilon_1 \le \min \{ \varepsilon/2, \varepsilon_{\ref{prop_typicalstability}}(\varepsilon, \delta), \delta/2\}, \varepsilon_2 \le \min \{ \varepsilon_{\ref{lemma_leftovertriangles}}^2/36^2, \varepsilon_1^2/400, \varepsilon_{\ref{lemma_balancingpartition}}(\delta, \varepsilon_1/3, \gamma/2) \}, \varepsilon_3 \le (1/27)\min\{ \varepsilon_{\ref{prop_randomgraphtypical}}(\varepsilon_2, \delta)^3, \varepsilon_2\}.

Let K = K_{\ref{thm_regularity}}(\varepsilon_3, 2, 1/\varepsilon_3), \eta = \min \{\eta_{\ref{thm_regularity}}(\varepsilon_3, 2, 1/\varepsilon_3), 1\}, \xi = \varepsilon_3/(4K), C = \max\{C_{\ref{prop_randomgraphproperties1}}(\eta), C_{\ref{lemma_leftovertriangles}} \}, and D = \max\{3D_{\ref{lemma_leftovertriangles}}, 18KD_{\ref{prop_nontypicalvertices} (ii)}(1/3,\xi), 24K \}.

Proposition 4.1 (iii) implies that G(n,p) is a.a.s. (\eta, 2, p)-upper-uniform. Thus we can apply the regularity lemma, Theorem 2.3, to obtain an (\varepsilon_3,p)-regular partition V_0, V_1, \ldots, V_{3k} of the graph G', where each part forms a regular pair with at least (3-3\varepsilon_3)k other parts. Let m = |V_i|, note that m \ge n/(2K), and let R be the reduced graph with parameter \delta. Since G' has minimum degree at least (2/3 + \gamma)np, by Proposition 4.7, a.a.s. the reduced graph has minimum degree at least (2 + \gamma/2)k. Thus by Theorem 2.16, we may assume that (V_{3t-2}, V_{3t-1}, V_{3t}) forms an (\varepsilon_3,p)-regular triple of density at least \delta for all t \in [k]. By Proposition 4.4, a.a.s. there are at most (\varepsilon_3/2) pm^2 edges in (V_{3t-1}, V_{3t-2}) whose endpoints have more than 2|V_{3t}|p^2 common neighbours in V_{3t}. Similar estimate holds for the edges in (V_{3t-2}, V_{3t}) and (V_{3t-1}, V_{3t}). Delete all such edges for all t \in [k] to obtain the subgraph G''. Then in the graph G'', each triple (V_{3t-2}, V_{3t-1}, V_{3t}) is (3\varepsilon_3^{1/3}, p)-regular by Proposition 2.2.

By Proposition 4.8, we may assume that every (3\varepsilon_3^{1/3},p)-regular triple of density at least \delta is \varepsilon_2-typical. Thus for each index i, if we let X_i \subset V_i be the collection of non \varepsilon_2-typical vertices, then |X_i| \le \varepsilon_2 |V_i|. Furthermore, for each t \in [k], add to X_{3t} the collection of those vertices v \in V_{3t} such that \deg(v, V_{3t-j}) \neq (1 \pm \varepsilon_2) d_{3t,3t-j}p|V_{3t-j}| for some j \in \{1,2\} and define X_{3t-1} and X_{3t-2} accordingly (there are at most 4\varepsilon_2 n such vertices by regularity). By adding arbitrary vertices to X_i if necessary, we may assume that |X_i| = 5\varepsilon_2 |V_i|. Move all the vertices in X_i from V_i to V_0 and denote the resulting partition by (V_i')_{i=0}^{3k}. We then have |V_0'| \le |V_0| + \sum_{i}|X_i| \le 6\varepsilon_2 n.

Consider the following process of finding triangles that absorbs the vertices in V_0'. Let T be the empty set; we will update it throughout the process. Apply Lemma 4.5 to find a triangle which hits V_0' but not T and move all the vertices of this triangle into T. If |T \cap V_i'| \ge \sqrt{\varepsilon_2} |V_i'|- 3 for some index i \in [3k], then move all the vertices of V_i' into T. This way, we will have

throughout the process. Terminate the process when we cannot find such triangles anymore. Then, a.a.s. we must have |V_0'| \le (D/3)p^{-2}. Let B_0 be the collection of all the remaining vertices of V_0', and S_0 be the set of vertices in the copies of the triangles that we found. Let V_i'' = V_i' \setminus S_0 = V_i \setminus (B_0 \cup S_0) and note that for all i, since |V_i \cap (B_0 \cup S_0)| \le |X_i| + \sqrt{\varepsilon_2}|V_i| \le (\varepsilon_1/6)|V_i|, then |V_i''| \ge (1 - \varepsilon_1/6)|V_i|.

By Proposition 2.1, (V_{3t-2}'', V_{3t-1}'', V_{3t}'') forms a (2\varepsilon_3,p)-regular triple of density at least \delta - \varepsilon_3 \ge \delta/2 for all t \in [k]. Apply Lemma 5.1 to (V_i'')_{i=1}^{3k} to obtain sets B_1 and S_1. Observe that |B_0 \cup B_1| \le (D/3)p^{-2} + 4K \le (2D/3)p^{-2} and G[S_0 \cup S_1] contains a perfect triangle packing. Also, most crucially, if we let BS = B_0 \cup S_0 \cup B_1 \cup S_1 and W_i = V_i \setminus BS = V_i'' \setminus (B_1 \cup S_1), then all W_i have equal sizes and moreover, |W_i| \geq (1 - \varepsilon_1/6 - \varepsilon_1/3)m \ge (1 - \varepsilon_1/2)m.

We will remove some vertices from each set W_i to make the triples (W_{3t-2}, W_{3t-1}, W_{3t}) super-regular for all t \in [k]. Since \xi n \le |X_i| \le |BS \cap V_i| \le (\varepsilon_1/2) m for all i, by Proposition 4.2 (ii), there are at most (D/(18K))p^{-1} vertices which have more than (2\varepsilon_1/3) p m neighbours in BS \cap V_i for each fixed i \in [3k]. Let Y_1 be the collection of such vertices for the set BS \cap V_2 and BS \cap V_3 which lie in V_1 and similarly define Y_2, Y_3. By placing arbitrary vertices into Y_1, Y_2, or Y_3 as necessary, we may assume that |Y_1| = |Y_2| = |Y_3| \le (D/(9k))p^{-1}. Consider the set W_i' = W_i \setminus Y_i for i \in \{1,2,3\}. Then since |Y_i| + |BS \cap V_i| \le (D/(9K))p^{-1} + (\varepsilon_1/2) m \le \varepsilon_1 m, in total we removed at most \varepsilon_1 m vertices from each part of (V_1,V_2,V_3) to obtain (W_1',W_2',W_3'). By the definition of the sets X_i at the beginning, all the vertices in W_1' were \varepsilon_1-typical in the triple (V_1, V_2, V_3), and by the choice of Y_1, they have at most (2\varepsilon_1/3) p m + |Y_2| \le \varepsilon_1 pm neighbours in the deleted portion in V_2 (similar for V_3). Thus by Proposition 2.10, all the vertices in W_1' are \varepsilon-typical in the triple (W_1',W_2',W_3'). Also, since all the vertices of V_1 not in X_1 had (1 \pm \varepsilon_1)d_{12}p|V_2| neighbours in V_2, they will still have (1 \pm 2\varepsilon_1)d_{12}p|V_2| neighbours in W_2' and similar for other choices of indices. Moreover, the triple (W_1', W_2', W_3') inherits the regularity of (V_1, V_2, V_3) and is (2\varepsilon_3, p)-regular of density at least \delta - \varepsilon_3 > \delta/2, see Proposition 2.1. Thus by the fact 2\varepsilon_1 < \varepsilon and 2\varepsilon_3 < \varepsilon, (W_1',W_2',W_3') is (\delta/2, \varepsilon,p)-super-regular. Repeat the above process for all other triples. Let B be the union of B_0, B_1, and Y_i for all i \in [3k] as above so that |B| \le (2D/3)p^{-2} + (D/3)p^{-1} = Dp^{-1} and S = S_0 \cup S_1. We also have the bound |W_i'| \ge |V_i|/2 \ge \xi n for all i. Moreover, (v) will hold since in G'' all the edges between W_2' and W_3' have at most 2|V_i|p^2 common neighbours in V_1, and therefore at most 4|W_1'|p^2 in W_1' (similar for other indices).

Let \varepsilon = \min\{1/100, \delta/50\}. Fix i, j, and k with \{i,j,k\} = \{1,2,3\}. For each vertex v \in V_i, let N_j = N(v) \cap V_j and N_k = N(v) \cap V_k. By construction, M_{12} \cup M_{13} \cup M_{23} is a matching.

Finally, note that Claims Proof–Proof imply that (M1)–(M4) are satisfied with probability 1 - o(1) (one needs to apply the union bound over all choices of vertices in order to deduce (M2) and (M3) from Proof and Proof).

Let \delta = \min\{\delta_{\ref{thm_cleanstage1}}(\gamma)/2, 3/4\}. Without loss of generality, we may assume that \varepsilon \leq 2/\delta. Furthermore, let \varepsilon_3 = \varepsilon_{\ref{prop_typicalstability}}(\varepsilon, 2\delta), \varepsilon_1 = \min\{\varepsilon\delta^2/1180, \varepsilon_3\delta^2/40\}, \varepsilon_2 = \min\{\varepsilon_{\ref{prop_manygoodvertices}}(\varepsilon_1, 2\delta), \varepsilon_{\ref{thm_cleanstage1}}(\gamma), \varepsilon_{\ref{lemma_randommatching}}(\delta), \varepsilon/2\}, and \xi = \xi_{\ref{thm_cleanstage1}}(\varepsilon_2)/2. Let \eta = 12\varepsilon_1/\delta^2. Moreover, let

and C = \max\{ C_{\ref{thm_cleanstage1}}(\varepsilon_2), C_{\ref{prop_smallexpansion}}(3\varepsilon_1\xi) \}.

By Theorem 5.2, there exists a further subgraph G' of G whose vertex set can be partitioned into sets B_0, S_0, and (W_i)_{i=1}^{3k} such that |B_0| \leq (D/2)p^{-2}, G[S_0] contains a perfect triangle packing, and for all t \in \{1, \ldots, k\}, the triple (W_{3t-2}, W_{3t-1}, W_{3t}) is (2\delta, \varepsilon_2, p)-super-regular in G' and satisfies |W_{3t-2}| = |W_{3t-1}| = |W_{3t}|. Moreover, the endpoints of no edge in (W_{3t-2}, W_{3t-1}) have more than 4|W_{3t}|p^2 common neighbours in W_{3t} (and a similar statement holds for other choices of indices). We will show that each such triple contains a slightly smaller (\delta, \varepsilon, p)-strong-super-regular triple in such a way that all but at most O(p^{-1}) leftover vertices can be covered by vertex-disjoint triangles. Obviously, this will imply the assertion of the theorem.

Without loss of generality, we will only consider the triple (W_1, W_2, W_3). For the sake of brevity, let m = |W_1| = |W_2| = |W_3| and note that m \geq 2\xi n. Without loss of generality, we can condition on the event that G(n,p) satisfies the assertions of

• Proposition 4.2 (ii) with \rho = 1/4 and \xi = \varepsilon_1m/n,

• Proposition 4.4 with \varepsilon = \varepsilon_1\varepsilon_2\delta^3m^2/(4n^2), and \rho = 1/4, and

• Proposition 4.3 with \xi = 6\varepsilon_1m/(2n).

For each i, let X_i \subset W_i be the collection of vertices that are not \varepsilon_1-good. By Proposition 2.13, |X_i| \leq \varepsilon_1 m and we may assume that |X_i| = \varepsilon_1 m. We perform the following cleaning-up procedure. While constantly updating the sets X_1, X_2, and X_3, repeat the following. If there exists an i and a vertex v \in W_i \setminus X_i such that either

1. (A) |N(v) \cap X_j| \geq 4\varepsilon_1pm for some j or

2. (B) the neighbourhood of v contains more than \varepsilon_2d_{12}d_{13}d_{23}p^3m^2 edges whose endpoints have more than 5\varepsilon_1p^2m common neighbours in X_i,

then move v to X_i.

It is not hard to check that (W_1 \setminus X_1, W_2 \setminus X_2, W_3 \setminus X_3) is (\delta/2, \varepsilon', p)-strong-super-regular. Unfortunately, this conclusion does not help us at the moment as we first need to absorb X_1 \cup X_2 \cup X_3 into vertex-disjoint triangles and in the process of absorbing those vertices, we may use some vertices from the triple (W_1 \setminus X_1, W_2 \setminus X_2, W_3 \setminus X_3).

For every i, let Y_i \subset X_i be the set of vertices in X_i that have more than 4\varepsilon_1pm neighbours in X_j for some j with j \neq i. Since |X_j| \leq 3\varepsilon_1m and we assumed that G(n,p) satisfies the assertion of Proposition 4.2 (ii) with \xi = \varepsilon_1m/n, then |Y_i| \leq Dp^{-1}/(6k). By adding arbitrary vertices of X_i to Y_i, we can guarantee that |Y_1| = |Y_2| = |Y_3|. For every i and j, let E_{ij} = (X_i, W_j) \cup (W_i, X_j). Since (W_i, W_j) is (\varepsilon_2, p)-regular and |X_i| \ge \varepsilon_1 |W_i| \ge \varepsilon_2 |W_i|, we have

Fix a vertex v \in W_i \setminus Y_i. We check that (A2) in Lemma 6.1 is satisfied. Let (N''_j, N''_k) be as in (A2) in Lemma 6.1. Since |N''_j \cap X_j| \leq |N(v) \cap X_j| \leq 4\varepsilon_1pm and similarly, |N''_k \cap X_k| \leq 4\varepsilon_1pm, we have

Lemma 6.1 implies that a.a.s. for each i and j, there exists an M_{ij} \subset (W_i \setminus X_i, W_j \setminus X_j) such that (M1)–(M4) in Lemma 6.1 are satisfied with V_i = W_i and V'_i = W_i \setminus Y_i for each i. Let Q_i be defined as in Lemma 6.1.

Fix any such triangle packing and for each i, let X'_i = X_i \cup T_i, where T_i \subset Q_i is the set of vertices in W_i \setminus X_i that are covered by the triangle packing. Note that |T_i| = |X_j \setminus Y_j| + |X_k \setminus Y_k|. Let W'_i = W_i \setminus X'_i. Since for each i, |X'_i| = |X_i| + |T_i| = |X_1 \setminus Y_1| + |X_2 \setminus Y_2| + |X_3 \setminus Y_3| + |Y_i| and |Y_1| = |Y_2| = |Y_3|, the sets W'_1, W'_2, and W'_3 have the same number of elements. Denote this number by m' and note that m' \geq m - 9\varepsilon_1m \geq m/2 \geq \xi n.

Finally, let B = B_0 \cup \bigcup_{i=1}^{3k}Y_i and let S = S_0 \bigcup_{i=1}^{3k}(T_i \cup (X_i \setminus Y_i)). Clearly, the sets B, S, and (W'_i)_{i=1}^{3k} partition the vertex set of G,

and G[S] contains a perfect triangle packing. Finally, by Claim Proof, for each t \in [k], the triple (W'_{3t-2}, W'_{3t-1}, W'_{3t}) is (\varepsilon, \delta, p)-strong-super-regular and satisfies |W'_{3t-2}| = |W'_{3t-1}| = |W'_{3t}| \geq \xi n.

For the sake of brevity, let m = |W_1| = |W_2| = |W_3|. Without loss of generality, we may assume that \delta \leq 1. Let \alpha = \delta^2/24 and let \beta be a positive constant satisfying \beta \log(e/\beta) < \alpha\eta/30. Moreover, let \varepsilon = \min\{\delta/2, \eta/4, \varepsilon_{\ref{lemma_randommatching}}(\delta), \beta \cdot (\varepsilon_0)_{\ref{prop_randomgraphtypical}}(\alpha\delta/16,\delta/2), \alpha\beta\delta/16\} and let \varepsilon' = 4\varepsilon/\delta. Finally, let C be sufficiently large so that C(\log n/n)^{1/2} \geq C_{\ref{lemma_randommatching}}(\log m/m)^{1/2} and without loss of generality we may assume that G(n,p) satisfies the assertion of Proposition 4.3 with \xi = \eta\delta^2m/(2n) and \xi = \eta\delta^2m/(12n), and Proposition 4.8 with \varepsilon'_{\ref{prop_randomgraphtypical}}=\alpha\delta/16 and \delta_{\ref{prop_randomgraphtypical}} = \delta/2.

For all i and j, let q_{ij} = \eta/(d_{ij}mp) and select each \varepsilon'-good edge of (W_i,W_j) independently with probability q_{ij}. Let M'_{ij} be the set of all selected edges in (W_i,W_j) and let M_{ij} \subset M'_{ij} be the set of all those edges that are not incident to any other selected edge. By Proposition 2.12, (W_i, W_j) contains at most \eta d_{ij}pm^2 edges that are not \varepsilon'-good. Since each v \in W_i is \varepsilon-good, its neighbourhood contains at most \eta d_{12}d_{13}d_{23}p^3m^2 edges that are not \varepsilon'-good. Therefore, Lemma 6.1 applies with E_{ij} being the set of non-\varepsilon'-good edges in (W_i, W_j) and V_i = V'_i = W_i.

Lemma 6.1 and Claim Proof imply that there exist M_{12}, M_{13}, and M_{23} such that M_{12} \cup M_{13} \cup M_{23} is a matching and for all i, j, and k (properties 1, 2, and 3 follow from (M1), (M3), and (M4) of Lemma 6.1, respectively, whereas property 4 follows from Claim Proof):

1. (\eta/2)m \leq |M_{jk}| \le 2\eta m,

2. the neighbourhood of every vertex in V_i contains at least (\eta/2)\delta^2p^2m edges of M_{jk},

3. the endpoints of each edge of M_{jk} have at least (1/2)\delta^2p^2m common neighbours in W_i \setminus Q_i,

4. for every set Y_i \subset W_i of size \beta m, all but at most \alpha\eta m edges of M_{jk} belong to the neighbourhood of some vertex of Y_i.

Fix any such M_{12}, M_{13}, and M_{23}. Next, for each i, let X_i be a random binomial subset of W_i \setminus Q_i, where each element is included with probability \eta/5. A simple application of Chernoff’s inequality combined with Property 3 above shows that if C is sufficiently large, then with probability 1 - o(1), for all i, j, and k:

1. |X_i| \leq (\eta/4)m,

2. the endpoints of each edge of M_{jk} have at least (\eta/12)\delta^2p^2m common neighbours in X_i.

Let X_1, X_2, and X_3 be arbitrary sets satisfying Proof and Proof and note that properties (P1)–(P3) are satisfied. It remains to show that (P4) is also satisfied.

Fix a Z_i \subset W_i of size |M_{jk}| such that X_i \subset Z_i. Let H be the bipartite graph on the vertex set Z_i \cup M_{jk}, where a vertex w \in Z_i is adjacent to an edge \{u,v\} \in M_{jk} if and only if \{u,v,w\} is a triangle in (W_1,W_2,W_3). Clearly, it suffices to prove that H contains a perfect matching. We check that H satisfies the assumptions of Proposition 2.15 with A = Z_i, B = M_{jk}, and L = \alpha\eta m.

Fix an S \subset Z_i. If 0 < |S| \leq (\eta/4)\delta^2m, then |N_H(S)| > |S| or otherwise there would be an x \in [1, (\eta/4)\delta^2m] such that G(n,p) contains some x independent edges and x vertices, each of which is adjacent to both ends of at least (\eta/2)\delta^2p^2m of those edges, see ?. This would contradict our assumption that G(n,p) satisfies the assertion of Proposition 4.3 with \xi = \eta\delta^2m/(2n). On the other hand, if |S| \geq (\eta/4)\delta^2m \geq \beta m, then by ?, |M_{jk} \setminus N_H(S)| \leq \alpha\eta m. Hence, |N_H(S)| \geq |S| as long as |Z_i \setminus S| \geq \alpha\eta m

Finally, fix a T \subset M_{jk} with 0 < |T| \leq \alpha\eta m = (\eta/24)\delta^2m. If |N_H(T)| < |T|, then there would be an x \in [1,(\eta/24)\delta^2m] such that G(n,p) contains x vertices and x independent edges whose endpoints have at least (\eta/12)\delta^2p^2m common neighbours among those x vertices (recall that X_i \subset Z_i), see Proof. This would contradict our assumption that G(n,p) satisfies the assertion of Proposition 4.3 with \xi = \eta\delta^2m/(12n).