2. Preliminaries

In this section, we introduce the notation used throughout our paper, which is mostly conventional, and present some basic results.

Given a vector $x\in \mathbb{R}^n$ , we write $||x||=||x||_2$ for the Euclidean norm. If $A\in \mathbb{R}^{n\times n}$ is a symmetric matrix, let $\lambda_i(A)$ denote the i-th largest eigenvalue of A. The spectral radius of A is denoted by $||A||$ , and it can be defined in a number of equivalent ways:

\begin{align*}||A||=\max\{|\lambda_1(A)|,|\lambda_n(A)|\}=\max_{x\in \mathbb{R}^n, ||x||=1} ||Ax||.\end{align*}

The Frobenius norm of A is defined as

\begin{align*}||A||_F^2:=\langle A,A\rangle=\mathrm{tr}(A^2)=\sum_{i=1}^{n}\lambda_i(A)^2,\end{align*}

where $\langle A,B\rangle=\sum_{i,j}A(i,j)B(i,j)$ is the entrywise scalar product on the space of matrices.

2·1. Basics of cuts

In this section, we present some basic results and definitions about cuts.

It will be convenient to work with r-uniform multi-hypergraphs H (or r-multigraph for short). That is, we allow multiple edges on the same r vertices. A k-cut in H is partition of V(H) into k parts $V_1,\dots,V_k$ , together with all the edges that have at least one vertex in each of the parts. We denote by $e(V_1,\dots,V_k)$ the number of such edges.

Definition 2·1. The Max-k-Cut of H is the maximum size of a k-cut, and it is denoted by $\mathrm{mc}_k(H)$ .

In case we consider a random partition, i.e. when each vertex of H is assigned to one of $V_i$ independently with probability $1/k$ , the expected size of a cut is $({S(r,k)k!}/{k^r})e(H)$ , where S(r, k) is the number of unlabelled partitions of $\{1,\dots,r\}$ into k non-empty parts.

Definition 2·2. The k-surplus of H is

\begin{align*}\mathrm{sp}_k(H)=\mathrm{mc}_k(H)-\frac{S(r,k)k!}{k^r}e(H).\end{align*}

Clearly, $\mathrm{sp}_k(H)$ is always nonnegative. Next, we recall several basic facts from [Reference Conlon, Fox, Kwan and Sudakov6].

Lemma 2·3 (Corollary 3·2 in [Reference Conlon, Fox, Kwan and Sudakov6]). Let H be an r-multigraph on n vertices. Then for every $2\leq k\leq r$ ,

\begin{align*}\mathrm{sp}_k(H)=\Omega_r\left(\frac{e(H)}{n}\right).\end{align*}

Lemma 2·4 (Theorem 1·4 in [Reference Conlon, Fox, Kwan and Sudakov6]). Let H be an r-multigraph on n vertices with no isolated vertices. Then for every $2\leq k\leq r$ ,

\begin{align*}\mathrm{sp}_{k}(H)=\Omega_r(n).\end{align*}

In the case of multi-graphs G, we write simply $\mathrm{mc}(G)$ and $\mathrm{sp}(G)$ instead of $\mathrm{mc}_2(G)$ and $\mathrm{sp}_2(G)$ . We also need the following multi-graph analogue of a well-known observation, which together with its variants has been used extensively, see e.g. [Reference Alon, Krivelevich and Sudakov2, Reference Glock, Janzer and Sudakov11].

Lemma 2·5. Let G be a multi-graph and let $V_1,\dots,V_k\subset V(G)$ be disjoint sets. Then

\begin{align*}\mathrm{sp}(G)\geq \sum_{i=1}^k \mathrm{sp}(G[V_i]).\end{align*}

Proof. By adding singleton sets, we may assume that the sets $V_i$ partition V(G). Let $(A_i,B_i)$ be a partition of $V_i$ such that the number of edges between $A_i$ and $B_i$ is $e(G[V_i])/2+\mathrm{sp}(G[V_i])$ . Define the random partition (X, Y) of V(G) as follows: for $i\in [k]$ , let either $(X_i,Y_i)=(A_i,B_i)$ or $(X_i,Y_i)=(B_i,A_i)$ independently with probability $1/2$ , and let $X=\bigcup_{i=1}^kX_i$ and $Y=\bigcup_{i=1}^k Y_i$ . It is straightforward to show that the expected number of edges in the cut given by (X, Y) is exactly $e(G)/2+\sum_{i=1}^k\mathrm{sp}(G[V_i])$ , finishing the proof.

Given an r-multigraph H, for every $q\leq r$ , the underlying q-multigraph is the q-multigraph on vertex set V(H), in which each q-tuple is added as an edge as many times as it is contained in an edge of H. The final lemma in this section is used to deduce Theorem 1·4 from our results about 3-graphs.

Lemma 2·6. Let H be an r-multigraph, and let H’ be the underlying $(r-1)$ -multigraph. Then

\begin{align*}2r\mathrm{sp}_{r}(H)\geq \mathrm{sp}_{r-1}(H)\geq \frac{1}{2}\mathrm{sp}_{r-1}(H').\end{align*}

Proof. We start with the first inequality. Let $U_1,\dots,U_{r-1}$ be a partition of V(H) such that $e_{H}(U_1,\dots,U_{r-1})=\mathrm{mc}_{r-1}(H)$ . Let $V_{r}$ be a random subset of V(H), each element chosen independently with probability $1/r$ , and let $V_i=U_i\setminus V_r$ for $i=1,\dots,r-1$ . Note that if an edge e of H contains at least one element of each $U_1,\dots,U_{r-1}$ , then exactly one of these sets contains two vertices of e. Hence, the probability that e is cut by $V_1,\dots,V_{r}$ is $c_r=2(1-\frac{1}{r})^{r-1}({1}/{r})\geq ({1}/{2r})$ . Hence,

\begin{align*}\mathbb{E}(e_H(V_1,\dots,V_r))&= c_r\mathrm{mc}_{r-1}(H)=c_r\mathrm{sp}_{r-1}(H)+c_r\frac{r!/2}{(r-1)^{r-1}}e(H)\\&=c_r\mathrm{sp}_{r-1}(H)+\frac{r!}{r^r}e(H),\end{align*}

so $\mathrm{sp}_{r}(H)\geq c_r \mathrm{sp}_{r-1}(H)$ .

Now let us prove the second inequality. Let $V_1,\dots,V_{r-1}$ be a partition of V(H) such that $e_{H'}(V_1,\dots,V_{r-1})=\mathrm{mc}_{r-1}(H')$ . If e is an edge of H cut by $V_1,\dots,V_{r-1}$ , then among the $(r-1)$ -element subsets of e, exactly two are cut by $V_1,\dots,V_{r-1}$ . Thus, $e_H(V_1,\dots,V_{r-1})=({1}/{2})e_{H'}(V_1,\dots,V_{r-1})$ . Therefore,

\begin{align*}\mathrm{mc}_{r-1}(H)&\geq \frac{1}{2}\mathrm{mc}_{r-1}(H')=\frac{1}{2}\mathrm{sp}_{r-1}(H')+\frac{(r-1)!/2}{(r-1)^{r-1}}e(H')\\&=\frac{1}{2}\mathrm{sp}_{r-1}(H')+\frac{r!/2}{(r-1)^{r-1}}e(H).\end{align*}

Hence, $\mathrm{sp}_{r-1}(H)\geq ({1}/{2})\mathrm{sp}_{r-1}(H').$

2·2. Probabilistic constructions

In this section, we verify our claims about the surplus of random 3-graphs.

Lemma 2·7. For every ${10\log n}/{n^2}\leq p\leq {1}/{n}$ , there exists a 3-graph H on n vertices such that H has $\Theta(pn^3)$ edges, the maximum degree of H is $O(pn^2)$ , the maximum co-degree of H is $O(\log n)$ , and $\mathrm{sp}_3(H)=O(\sqrt{p}n^2)$ .

Proof. Let p be such that ${10\log n}/{n^2}\leq p\leq {1}/{n}$ , and let H be the 3-graph on n vertices, in which each of the $\binom{n}{3}$ potential edges are included independently with probability p. We assume that n is sufficiently large. By the Chernoff–Hoeffding theorem, for every $x\in (0,1/2)$ , we have

\begin{align*}\mathbb{P}\left(|e(H)/\binom{n}{3}-p|\geq x\right)\leq 2\exp\left(-\frac{x^2\binom{n}{3}}{4p}\right).\end{align*}

Hence, setting $x=20\sqrt{p}n^{-3/2}$ ,

\begin{align*}\mathbb{P}\left(|e(H)-p\binom{n}{3}|\geq 20\sqrt{p}n^{3/2}\right)\leq 0.1.\end{align*}

Also, by similar concentration arguments, the maximum degree of H is at most $2p\binom{n}{2}$ with probability $0.9$ by our assumption $pn^2\geq 10\log n$ . Moreover, the maximum co-degree is at most $10\log n$ with probability $0.9$ by our assumption that $p\leq 1/n$ .

Let X, Y, Z be a partition of the vertex set into three parts, and let $N=(n/3)^3\geq |X||Y||Z|=T$ . For every $x\gt 0$ ,

\begin{align*}\mathbb{P}\left(e(X,Y,Z)\geq N(p+x)\right)&\leq \mathbb{P}\left(\frac{e(X,Y,Z)}{T}\geq p+x\frac{N}{T}\right)\\&\leq \exp\left(-\frac{(xN/T)^2\cdot T}{4p}\right)\leq \exp\left(-\frac{x^2 N}{4p}\right),\end{align*}

where the second inequality is due to the Chernoff-Hoeffding theorem. Set $x={20\sqrt{p}}/{n}$ . The number of partitions of V(H) into three sets is at most $3^n$ , so

\begin{align*}\mathbb{P}(\exists X,Y,Z: e(X,Y,Z)\geq N(p+x))\leq 3^n\exp\left(-\frac{x^2 N}{4p}\right)\lt 0.1.\end{align*}

Hence, with probability 0.9, $\mathrm{mc}_3(H)\leq N(p+x)\leq pN+3\sqrt{p}n^2$ .

In conclusion, with positive probability, there exists a 3-graph H on n vertices such that $|e(H)-p\binom{n}{3}|\leq O(\sqrt{p}n^{3/2})$ , the maximum degree of H is at most $pn^2$ , the maximum co-degree is at most $10\log n$ , and $\mathrm{mc}_3(H)\leq p({n^3}/{27})+O(\sqrt{p}n^2)$ . For such a 3-graph H, we have

\begin{align*}\mathrm{sp}_3(H)=\mathrm{mc}_3(H)-\frac{2}{9}e(H)\leq p\frac{n^3}{27}+O(\sqrt{p}n^2)-\frac{2}{9}p\binom{n}{3}+O(\sqrt{p}n^{3/2})=O(\sqrt{p}n^2).\end{align*}

In the last equality, we used that $p\cdot \frac{n^3}{27}-\frac{2}{9}p\binom{n}{3}=O(pn^2)$ .

If one chooses $p=1/n$ , we have $m=e(H)=\Theta(n^2)$ , every co-degree in H is $O(\log n)$ , and $\mathrm{sp}_3(H)=O(n^{3/2})=O(m^{3/4})$ , showing the almost tightness of Theorem 1·2.

3. Surplus and Energy

In this section, we prove a lower bound on the surplus in terms of the energy of the graph. Let G be a multi-graph on vertex set V, and let $|V|=n$ . The adjacency matrix of G is the matrix $A\in \mathbb{R}^{V\times V}$ defined as $A(u,v)=k$ , where k is the number of edges between u and v.

Recall that $\mathrm{mc}(G)={e(G)}/{2}+\mathrm{sp}(G)$ . Next, we claim that

\begin{equation*} \mathrm{sp}(G)=\max_{x\in \{-1,1\}^{V}} -\frac{1}{2}\sum_{uv\in E(G)}x(u)x(v). \end{equation*}

Indeed, every $x\in \{-1,1\}^{V}$ corresponds to a partition X, Y of V(G), where $X=\{v\in V(G):x(v)=1\}$ , and then

\begin{align*}e(X,Y)-\frac{m}{2}=\frac{1}{2}(e(X,Y)-e(X)-e(Y))=-\frac{1}{2}\sum_{uv\in E(G)}x(u)x(v).\end{align*}

We define the surplus of arbitrary symmetric matrices $A\in\mathbb{R}^{n\times n}$ as well. We restrict our attention to symmetric matrices, whose every diagonal entry is 0. Let

(3·1) \begin{equation} \mathrm{sp}(A)=\max_{x\in \{-1,1\}^{n}} -\frac{1}{2}\sum_{i,j\in [n]}A(i,j)x(i)x(j). \end{equation}

Then $\mathrm{sp}(G)=\mathrm{sp}(A)$ if A is the adjacency matrix of G. Observe that by the condition that every diagonal entry of A is zero, $\sum_{i,j}A(i,j)x(i)x(j)$ is a multilinear function, so its minimum on $[{-}1,1]^n$ is attained by one of the extremal points. Therefore, (3·1) remains true if the maximum is taken over all $x\in [{-}1,1]^{n}$ . We introduce the semidefinite relaxation of $\mathrm{sp}(A)$ . Let

(3·2) \begin{equation} \mathrm{sp}^*(A)=\max-\frac{1}{2}\sum_{i,j}A(i,j)\langle z_i,z_j\rangle,\end{equation}

where the maximum is taken over all $z_1,\dots,z_n\in\mathbb{R}^{n}$ such that $||z_i||\leq 1$ for $i\in [n]$ . Clearly, $\mathrm{sp}^*(A)\geq \mathrm{sp}(A).$ On the other hand, we can use the following symmetric analogue of Grothendieck’s inequality [Reference Grothendieck13] to show that $\mathrm{sp}^{*}(A)$ also cannot be much larger than $\mathrm{sp}(A)$ .

Lemma 3·1 ([Reference Alon, Makarychev, Makarychev and Naor3, Reference Charikar and Wirth5]). There exists a universal constant $C\gt 0$ such that the following holds. Let $M\in \mathbb{R}^{n\times n}$ be symmetric. Let

\begin{align*}\beta=\sup_{x\in [{-}1,1]^{n}}\sum M_{i,j}x(i)x(j),\end{align*}

and let

\begin{align*}\beta^{*}=\sup_{v_1,\dots,v_n\in \mathbb{R}^{n},||v_i||\leq 1}\sum M_{i,j}\langle v_i,v_j\rangle.\end{align*}

Then $\beta\leq \beta^{*}\leq C\beta \log n$ .

Applying this lemma with the matrix $M=-({1}/{2})A$ , we get that

\begin{align*}\mathrm{sp}^*(A)\leq C\ \mathrm{sp}(A)\log n.\end{align*}

Finally, it is convenient to rewrite equation (3·2) as

(3·3) \begin{equation} \mathrm{sp}^*(A)=\max-\frac{1}{2}\langle A,X\rangle, \end{equation}

where the maximum is taken over all positive semidefinite symmetric matrices $X\in \mathbb{R}^{n\times n}$ which satisfy $X(i,i)\leq 1$ for every $i\in [n]$ . Indeed, (3·2) and (3·3) are equivalent, as if $z_1,\dots,z_n\in\mathbb{R}^n$ such that $||z_i||\leq 1$ , then the matrix X defined as $X(i,j)=\langle z_i,z_j\rangle$ is positive semidefinite and satisfies $X(i,i)=||z_i||^2\leq 1$ . Also, every positive semidefinite matrix X such that $X(i,i)\leq 1$ is the Gram matrix of some vectors $z_1,\dots,z_n\in \mathbb{R}^n$ satisfying $||z_i||\leq 1$ .

In the next lemma, we connect the spectral properties of A to the surplus. Let $\lambda_1\geq \dots\geq \lambda_n$ be the eigenvalues of A.

Definition 3·2. The energy of $A\in \mathbb{R}^{n\times n}$ is defined as

\begin{align*}\mathcal{E}(A)=\sum_{i=1}^{n}|\lambda_i|,\end{align*}

and if A is the adjacency matrix of a graph G, then the energy of G is $\mathcal{E}(G)=\mathcal{E}(A)$ .

Usually, the energy of G is denoted by E(G), but we use the notation $\mathcal{E}(G)$ to not confuse with the edge set of G, or with the expectation $\mathbb{E}$ . In a certain weak sense, the energy measures how random-like a graph is. Indeed, if G is a d-regular graph on n vertices, then the energy of G is at most $O(\sqrt{d} n)$ , and this bound is attained by the random d-regular graph. Next, we show that the energy is a lower bound (up to a constant factor) for $\mathrm{sp}^*(G)$ .

Proof. Let $v_1,\dots,v_n$ be an orthonormal set of eigenvectors of A such that $\lambda_i$ is the eigenvalue corresponding to $v_i$ . Let

\begin{align*}X=\sum_{i:\lambda_i\lt 0}v_i\cdot v_i^T.\end{align*}

Then X is positive semidefinite. Also, for $j=1,\dots,n$ ,

\begin{align*}X(j,j)=\sum_{i:\lambda_i\lt 0}v_i(j)^2\leq \sum_{i=1}^n v_i(j)^2=1.\end{align*}

Hence, by (3·3),

\begin{align*}\mathrm{sp}^{*}(A)\geq -\frac{1}{2}\langle X,A\rangle.\end{align*}

Observe that $A=\sum_{i=1}^{n}\lambda_i (v_i\cdot v_i^T)$ . Hence,

\begin{align*}-\frac{1}{2}\langle X,A\rangle&=-\frac{1}{2}\left\langle \sum_{i:\lambda_i\lt 0}v_i v_i^T,\sum_{i=1}^n\lambda_iv_i v_i^T\right\rangle\\&=-\frac{1}{2}\sum_{i:\lambda_i\lt 0}\sum_{j=1}^{n}\lambda_j\langle v_i,v_j\rangle^2=-\frac{1}{2}\sum_{i:\lambda_i\lt 0}\lambda_i.\end{align*}

As $\sum_{i=1}^{n}\lambda_i=\mbox{tr}(A)=0$ , we have $\sum_{i:\lambda_i\lt 0}\lambda_i=-\frac{1}{2}\mathcal{E}(A)$ . This finishes the proof.

Finally, we note that the energy is a norm on the space of $n\times n$ symmetric matrices. On this space, the energy coincides with the Schatten 1-norm, also known as trace norm, or nuclear norm. In particular, it follows that the energy is subadditive.

Lemma 3·5. Let $A,B\in \mathbb{R}^{n\times n}$ be symmetric matrices. Then

\begin{align*}\mathcal{E}(A+B)\leq \mathcal{E}(A)+\mathcal{E}(B).\end{align*}

4. Sampling coloured multi-graphs

This section is concerned with edge-coloured multi-graphs. Our goal is to show that if one randomly samples the colours in such a graph, then the resulting graph has large cuts. We use bold letters to distinguish random variables from deterministic ones. In the rest of this section, we work with the following setup.

Setup. Let G be a multi-graph with n vertices and m edges, and let $\phi:E(G)\rightarrow \mathbb{N}$ be a colouring of its edges. Let A be the adjacency matrix of G. Let $\Delta$ be the maximum degree of G, and assume that every colour appears at most D times at every vertex for some $D\leq \Delta$ . Let $p\in ({100}/{m},0.9)$ , and sample the colours appearing in G independently with probability p. Let $\textbf{H}$ be the sub-multi-graph of G of the sampled colours, and let $\textbf{B}$ be the adjacency matrix of $\textbf{H}$ .

In our first lemma, we show that pA is a good spectral approximation of $\textbf{B}$ . In order to do this, we use a result of Oliviera [Reference Oliveira16] on the concentration of matrix martingales. More precisely, we use the following direct consequence of Theorem 1·2 in [Reference Oliveira16].

Lemma 4·1 (Theorem 1·2 in [Reference Oliveira16]). Let $\textbf{M}_1,\dots,\textbf{M}_m$ be independent random $n\times n$ symmetric matrices, and let $\textbf{P}=\textbf{M}_1+\dots+\textbf{M}_m$ . Assume that $||\textbf{M}_i||\leq D$ for $i=1,\dots,m$ , and define

\begin{align*}W=\sum_{i=1}^m\mathbb{E}(\textbf{M}_i^2).\end{align*}

Then for all $t\gt 0$ ,

\begin{align*}\mathbb{P}(||\textbf{P}-\mathbb{E}(\textbf{P})||\geq t)\leq 2n\exp\left(-\frac{t^2}{8||W||+4Dt}\right).\end{align*}

Lemma 4·2. For every $t\geq 0$ ,

\begin{align*}\mathbb{P}\left(||pA-\textbf{B}||\geq t\right)\leq 2n\exp\left(-\frac{t^2}{8p\Delta D+4Dt}\right).\end{align*}

Proof. Let $1,\dots,m$ be the colours of G, let $A_i$ be the adjacency matrix of colour i, and let $\textbf{I}_i\in \{0,1\}$ be the indicator random variable that colour i is sampled. Then setting $\textbf{M}_i=\textbf{I}_i A_i$ , we have $\textbf{B}=\textbf{M}_1+\dots+\textbf{M}_n$ . Moreover, $||\textbf{M}_i||\leq ||A_i||\leq D$ , as the spectral radius is upper bounded by the maximum $\ell_1$ -norm of the row vectors, which in turn is the maximum degree of the graph of colour i. Now let us bound $||W||$ , where

\begin{align*}W=\sum_{i=1}^{m}\mathbb{E}(\textbf{M}_i^2)=p\sum_{i=1}^m A_i^2.\end{align*}

Let $\mu_1\geq \dots\geq \mu_n\geq 0$ be the eigenvalues of W. Then for every k,

\begin{align*}\sum_{i=1}^n \mu_i^k=\mathrm{tr}(W^k)=\sum_{v_1,\dots,v_k\in V(G)}W(v_1,v_2)\dots W(v_{k-1},v_k)W(v_k,v_1).\end{align*}

Observe that $A_i^2(a,b)$ is the number of walks of length 2 between a and b using edges of colour i, so $W(a,b)/p=\sum_{i=1}^m A_i^2(a,b)$ is the number of monochromatic walks of length 2 between a and b in G. Hence, $\mathrm{tr}(W^k/p^k)$ is the number of closed walks $(v_i,e_i,w_i,f_i)_{i=1,\dots,k}$ , where $v_i,w_i$ are vertices, $e_i,f_i$ are edges, $v_i,w_i$ are endpoints of $e_i$ , $w_i,v_{i+1}$ are endpoints of $f_i$ (where the indices are meant modulo k), and $e_i$ and $f_i$ has the same colour. The total number of such closed walks is upper bounded by $n\Delta^k D^k$ , as there are n choices for $v_1$ , if $v_i$ is already known there are at most $\Delta$ choices for $e_i$ and $w_i$ , and as the colour of $f_i$ is the same as of $e_i$ , there are at most D choices for $f_i$ and $v_{i+1}$ . Hence, we get

\begin{align*}\mu_1^k/p^k\leq n\Delta^k D^k.\end{align*}

As this holds for every k, we deduce that $||W||=\mu_1\leq p\Delta D$ . But then the inequality

\begin{align*}\mathbb{P}\left(||pA-\textbf{B}||\geq t\right)\leq 2n\exp\left(-\frac{t^2}{8p\Delta D+4Dt}\right)\end{align*}

is a straightforward consequence of Lemma 4·1.

Next, we use the previous lemma to show that the expected energy of $pA-\textbf{B}$ is large. This is somewhat counter-intuitive, as the previous lemma implies that none of the eigenvalues of $pA-\textbf{B}$ is large. However, we show that the Frobenius norm of $pA-\textbf{B}$ must be large in expectation, which tells us that the sum of the squares of the eigenvalues is large. Given none of the eigenvalues is large, this is only possible if the energy of $pA-\textbf{B}$ is large as well.

Lemma 4·3.

\begin{align*}\mathbb{E}(\mathcal{E}(pA-\textbf{B}))=\Omega\left(\frac{pm}{\sqrt{\Delta D}\log m}\right).\end{align*}

Proof. Let $\lambda_1\geq\dots\geq\lambda_n$ be the eigenvalues of $pA-\textbf{B}$ , let $\lambda=||pA-\textbf{B}||$ , and $t=20\log m \sqrt{\Delta D}$ . Without loss of generality, we may assume that G has no isolated vertices, so $m\geq 2n$ . But then, by the previous lemma, $|\lambda_i|\leq \lambda\leq t$ with probability at least

\begin{align*}1-2n\exp\left(-\frac{t^2}{8p\Delta D+4Dt}\right).\end{align*}

Here,

\begin{align*}\frac{t^2}{8p\Delta D+4Dt}=\frac{400(\log m)^2 \Delta D}{8p\Delta D+80 D (\log m)\sqrt{\Delta D}}\geq \frac{400(\log m)^2 \Delta D}{8p\Delta D+80(\log m) \Delta D}\gt 4\log m.\end{align*}

Thus, we conclude

\begin{align*}\mathbb{P}(\lambda\gt t)\leq 2nm^{-4}\lt m^{-2}.\end{align*}

Consider the Frobenius norm of $pA-\textbf{B}$ . We have

(4·4) \begin{equation} ||pA-\textbf{B}||_F^2=\sum_{i=1}^{n}\lambda_i^2\leq \lambda\sum_{i=1}^{n}|\lambda_i|=\lambda\cdot \mathcal{E}(pA-\textbf{B}).\end{equation}

Let $\textbf{Y}=||pA-\textbf{B}||_F^2$ . If $u,v\in V(G)$ , then $\textbf{B}(u,v)=z_1\textbf{I}_1+\dots+z_r\textbf{I}_r$ , where $\textbf{I}_1,\dots,\textbf{I}_r$ are the independent indicator random variables of colours appearing between u and v in G, and $z_1,\dots,z_r$ are the multiplicities of colours. Note that $z_i\geq 1$ and $z_1+\dots+z_r=A(u,v)$ . Hence,

\begin{align*}\mathbb{E}((pA(u,v)-\textbf{B}(u,v))^2)&=\mbox{Var}(\textbf{B}(u,v))=\sum_{i=1}^{r}\mbox{Var}(z_i\textbf{I}_r)\\&=\sum_{i=1}^{r}(p-p^2)z_i^2\geq (p-p^2)A(u,v).\end{align*}

Therefore,

\begin{align*}\mathbb{E}(\textbf{Y})\geq \sum_{u,v\in V(G)}(p-p^2) A(u,v)=2(p-p^2)m=\Omega(pm).\end{align*}

By (4·4), we can write

\begin{align*}\mathbb{E}(\mathcal{E}(pA-\textbf{B}))\geq \mathbb{E}\left(\frac{\textbf{Y}}{\lambda}\right).\end{align*}

In order to bound the right-hand side, we condition on the event $(\lambda\leq t)$ :

\begin{align*}\mathbb{E}\left(\frac{\textbf{Y}}{\lambda}\right)\geq \mathbb{P}(\lambda\leq t)\cdot \mathbb{E}\left(\frac{\textbf{Y}}{\lambda}| \lambda\leq t\right)\geq \frac{1}{t}\mathbb{P}(\lambda\leq t)\cdot \mathbb{E}(\textbf{Y}|\lambda\leq t).\end{align*}

Here, using that $\mathbb{P}(\lambda\gt t)\leq m^{-2}$ and $\max(\textbf{Y})\leq m^2$ , we can further write

\begin{align*}\mathbb{P}(\lambda\leq t)\cdot \mathbb{E}(\textbf{Y}|\lambda\leq t)&=\mathbb{E}(\textbf{Y})-\mathbb{P}(\lambda\gt t)\cdot\mathbb{E}(\textbf{Y}|\lambda\gt t)\\&\geq \mathbb{E}(\textbf{Y})-\mathbb{P}(\lambda\gt t)\cdot\max(\textbf{Y})\geq \Omega(pm).\end{align*}

Thus, $\mathbb{E}(\mathcal{E}(pA-\textbf{B}))\geq \Omega({pm}/{t})$ , finishing the proof.

Our final technical lemma shows that $M+\textbf{B}$ has large expected surplus for any fixed matrix M.

Lemma 4·4. Let $M\in \mathbb{R}^{V(G)\times V(G)}$ be symmetric with only zeros in the diagonal. Then

\begin{align*}\mathbb{E}(\mathrm{sp}(M+\textbf{B}))=\Omega\left(\frac{pm}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Proof. Again, we may assume that G does not contain isolated vertices. Let

\begin{align*}t=\frac{\varepsilon pm}{\sqrt{\Delta D}\log m},\end{align*}

where $\varepsilon\gt 0$ is the constant hidden by the $\Omega(.)$ notation in Lemma 4·3. Thus,

\begin{align*}\mathbb{E}(\mathcal{E}(\textbf{B}-pA))\geq t.\end{align*}

First, assume that $\mathcal{E}(M+pA)\gt {t}/{8}$ . Then by Corollary 3·4, we have $\mathrm{sp}(M+pA)=\Omega({t}/{\log n})$ . This means that there exists a vector $x\in \{-1,1\}^{V(G)}$ such that

\begin{align*}\mathrm{sp}(M+pA)=-\frac{1}{2}\sum_{u,v\in V(G)}(M(u,v)+pA(u,v))x(u)x(v)=\Omega\left(\frac{t}{\log n}\right).\end{align*}

But then

\begin{align*} \mathbb{E}(\mathrm{sp}(M+\textbf{B}))&\geq \mathbb{E}\left[{-}\frac{1}{2}\sum_{u,v\in V(G)}(M(u,v)+\textbf{B}(u,v))x(u)x(v)\right] \\[5pt] &= \,-\frac{1}{2}\sum_{u,v\in V(G)}(M(u,v)+pA(u,v))x(u)x(v)=\mathrm{sp}(M+pA).\end{align*}

so we are done.

Hence, we may assume that $\mathcal{E}(M+pA)\leq {t}/{8}$ . Writing $(\textbf{B}-pA)=(M+\textbf{B})+(-pA-M)$ , we can use the additivity of the energy (Lemma 3·3) to deduce that $\mathcal{E}(\textbf{B}-pA)\leq \mathcal{E}(M+\textbf{B})+\mathcal{E}(M+pA)$ . From this,

\begin{align*}\mathbb{E}(\mathcal{E}(M+\textbf{B}))\geq \mathbb{E}\left(\mathcal{E}(\textbf{B}-pA)-\mathcal{E}(M+pA)\right)\geq \frac{t}{2}.\end{align*}

Therefore, by Corollary 3·4,

\begin{align*}\mathbb{E}(\mathrm{sp}(M+\textbf{B}))\geq \Omega\left(\mathbb{E}\left(\frac{\mathcal{E}(M+\textbf{B})}{\log n}\right)\right)=\Omega\left(\frac{pm}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

In the last step, we are also using the fact that G does not contain isolated vertices, whence $m = \Omega(n)$ .

5. MaxCut in 3-graphs

In this section, we prove our main theorems. We start with the proof of Theorem 1·3, which we restate here for the reader’s convenience.

Theorem 5·1. Let H be a 3-multigraph. Assume that H contains an induced subghypergraph with m edges, maximum degree $\Delta$ and maximum co-degree D. Then

\begin{align*}\mathrm{sp}_3(H)\geq \Omega\left(\frac{m}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Sample the vertices of H independently with probability $p=1/3$ , and let $\textbf{X}$ be the set of sampled vertices. First, we condition on $\textbf{X}\setminus Q$ , that is, we reveal $\textbf{X}$ outside of Q, and treat $\textbf{X}\cap Q$ as the source of randomness. We use bold letters to denote those random variables that depend on $\textbf{X}\cap Q$ . Let us introduce some notation, see also Figure 1 for an illustration:

1. (1) $V=S\setminus \textbf{X}$ ;

2. (2) $G_0$ is the graph on vertex set V which contains those edges of G, whose colour is in Q,

3. $A\in \mathbb{R}^{V\times V}$ is the adjacency matrix of $G_0$ ;

4. (3) $\textbf{G}_0^*$ is the subgraph of $G_0$ in which we keep those edges, whose colour is in $ \textbf{X}\cap Q$ ,

5. $\textbf{B}\in \mathbb{R}^{V\times V}$ is the adjacency matrix of $G_0^*$ ;

6. (4) $G_1^*$ is the graph on vertex set V which contains those edges of G, whose colour is in $\textbf{X}\setminus Q$ ;

7. $M\in \mathbb{R}^{V\times V}$ is the adjacency matrix of $G_1^*$ ,

8. (5) $\textbf{G}^*$ is the subgraph of $G[\textbf{X}^c]$ , whose edges have colour in $\textbf{X}$ . (Here, $\textbf{X}^c=V(H)\setminus \textbf{X}$ .)

Figure 1. An illustration of the different subsets of the vertex set used in the proof of Theorem 5·1.

Note that $G_0$ has maximum degree at most $2\Delta$ , as the degree of every vertex in $G_0$ is at most two times its degree in $H[S_0]$ . Also, every colour in $G_0$ appears at most D times at every vertex. Indeed, if some colour $q\in Q$ appears more than D times at some vertex $v\in V$ , then the co-degree of qv in $H[S_0]$ is more than D, contradicting our choice of $S_0$ . Thus, the matrices A and $\textbf{B}$ fit the Setup of Section 4.

Also, $\textbf{G}^*[V]$ is the edge-disjoint union of $\textbf{G}_0^*$ and $G_1^*$ , so the adjacency matrix of $\textbf{G}^*[V]$ is $M+\textbf{B}$ . Hence, $\mathrm{sp}(\textbf{G}^*)\geq \mathrm{sp}(\textbf{G}^*[V])=\mathrm{sp}(M+\textbf{B})$ , where the first inequality follows from Lemma 2·5. Thus, writing $f=e(G_0)$ , we can apply Lemma 4·4 to get

\begin{align*}\mathbb{E}(\mathrm{sp}(\textbf{G}^*)|\textbf{X}\setminus Q)\geq \mathbb{E}(\mathrm{sp}(M+\textbf{B})|\textbf{X}\setminus Q)=\Omega\left(\frac{f}{\sqrt{\Delta D}(\log f)^2}\right). \end{align*}

As we have $f\leq m$ ,

\begin{align*}\mathbb{E}(\mathrm{sp}(\textbf{G}^*)|\textbf{X}\setminus Q)=\Omega\left(\frac{f}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Furthermore,

\begin{align*}\mathbb{E}(\mathrm{sp}(\textbf{G}^*))=\mathbb{E}(\mathbb{E}(\mathrm{sp}(\textbf{G}^*)|\textbf{X}\setminus Q))\geq \Omega\left(\frac{\mathbb{E}(f)}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Note that f is the number of those edges in $G[S\setminus \textbf{X}]$ , whose colour is in Q. For every edge of G[S], whose colour is in Q, the probability that it survives is $(1-p)^2p=\Omega(1)$ . Hence, $\mathbb{E}(f)=\Omega(m)$ . In conclusion,

\begin{align*}\mathbb{E}(\mathrm{sp}(\textbf{G}^*))=\mathbb{E}(\mathbb{E}(\mathrm{sp}(\textbf{G}^*)|\textbf{X}\setminus Q))\geq \Omega\left(\frac{m}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Let $\textbf{Y},\textbf{Z}$ be a partition of $\textbf{X}^c$ such that $e_{\textbf{G}^*}(\textbf{Y},\textbf{Z})= ({1}/{2})e(\textbf{G}^*)+\mathrm{sp}(\textbf{G}^*)$ . Then $e_H(\textbf{X},\textbf{Y},\textbf{Z})=e_{\textbf{G}^*}(\textbf{Y},\textbf{Z})$ , so

\begin{align*}\mathbb{E}(e_H(\textbf{X},\textbf{Y},\textbf{Z}))&=\mathbb{E}(e_{\textbf{G}^*}(\textbf{Y},\textbf{Z}))=\frac{1}{2}\mathbb{E}(e(\textbf{G}^*))+\mathbb{E}(\mathrm{sp}(\textbf{G}^*))\\&=\frac{2}{9}e(H)+\Omega\left(\frac{m}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

Therefore, choosing a partition $(\textbf{X},\textbf{Y},\textbf{Z})$ achieving at least the expectation, we find that

\begin{align*}\mathrm{sp}_3(H)=\Omega\left(\frac{m}{\sqrt{\Delta D}(\log m)^2}\right).\end{align*}

From this, Theorem 1·2 follows almost immediately. In particular, we prove the following slightly more general result.

Theorem 5·2. Let H be a 3-graph with m edges and maximum co-degree D. Then

\begin{align*}\mathrm{sp}_3(H)=\Omega\left(\frac{m^{3/4}}{D^{3/4}(\log m)^2}\right).\end{align*}

If $e_G(S,T)\geq 1.6m$ , then $\mathrm{sp}(G)\geq 0.1 m$ , so we are done as Lemma 2·4 implies $\mathrm{sp}_3(H)\geq ({1}/{12})\mathrm{sp}(G)$ .

Hence, we may assume that $e_G(S,T)\lt 1.6m$ . But then we must have $e(G[S])\geq 1.3m$ , as $e(G[T])+e_G(S,T)+e(G[S])=e(G)=3m$ . Moreover,

\begin{align*}e(G[S])\leq 3e(H[S])+e_H(S,T) = 3e(H[S])+\frac{1}{2}e_G(S,T),\end{align*}

from which we conclude that $e(H[S])\geq 0.1 m$ . The maximum degree of H[S] is less than the maximum degree of G[S], which is less than $\Delta=100m^{1/2}D^{1/2}$ . Also, the maximum co-degree of H[S] is at most D, so applying Theorem 5·1 to the subhypergraph induced by S finishes the proof.

Next, we prove Theorem 1·1.

Proof of Theorem 1·1. Let G be the underlying graph of H, and consider G as an edge weighted graph, where the weight of each edge is its multiplicity. We may assume that $\mathrm{sp}(G)\leq {m^{3/5}}/{(\log m)^2}$ , otherwise we are done by Lemma 2·4. Let $D=m^{1/5}$ . Say that an edge e of G is D-heavy if its weight is at least D.

Let S be the vertex set of a maximal matching of D-heavy edges.

Proof. Assume that $|S|\geq m^{2/5}$ , and let $e_1,\dots,e_s$ be a matching of D-heavy edges, where $s=m^{2/5}/2$ . By Lemma 2·5, we have

\begin{align*}\mathrm{sp}(G)\geq \sum_{i=1}^{s}\mathrm{sp}(G[e_i])\geq s\cdot \frac{D}{2}=\frac{m^{3/5}}{4},\end{align*}

a contradiction.

Let $\Delta=m^{3/5}$ , and let T be the set of vertices of G of degree more than $\Delta$ . Then $3m=e(G)\geq ({1}/{2})|T|\Delta$ , so $|T|\leq 6m^{2/5}$ . Let $U=S\cup T$ , then $|U|\leq 7m^{2/5}$ .

Proof. If $e(G[U])\gt 0.1m$ , then by Lemmas 2·5 and 2·3, we get

\begin{align*}\mathrm{sp}(G)\geq \mathrm{sp}(G[U])= \Omega\left(\frac{e(G[U])}{|U|}\right)=\Omega(m^{3/5}),\end{align*}

a contradiction.

Let $W=U^c$ . Then G[W] contains no D-heavy edges, and the maximum degree of G[W] is at most $\Delta$ . Also, $e_G(U,W)\leq 1.6m$ , as otherwise $\mathrm{sp}(G)\geq 0.1m$ . This implies that $e_H(U,W)\leq 0.8m$ and that $e(G[W])=e(G)-e(G[U])-e_G(U,W)\geq 1.3m$ . But then

\begin{align*}e(H[W])\geq \frac{1}{3}(e(G[W])-e_H(U,W))\gt 0.1m.\end{align*}

Thus, H[W] is an induced subhypergraph of H with at least $0.1m$ edges, maximum degree at most $\Delta$ , and maximum co-degree at most D. Applying Theorem 5·1, we conclude that

\begin{align*}\mathrm{sp}_3(H)\geq \Omega\left(\frac{m}{\sqrt{\Delta D}(\log m)^2}\right)=\Omega\left(\frac{m^{3/5}}{(\log m)^2}\right).\end{align*}

Finally, we prove Theorem 1·4.

Proof of Theorem 1·4. Let $H_r=H$ , and for $i=r-1,\dots,3$ , let $H_i$ be the underlying i-graph of $H_{i+1}$ . Then by Lemma 2·6, $\mathrm{sp}_{i+1}(H_{i+1})\geq ({1}/{4(i+1))}\mathrm{sp}_{i}(H_i)$ . From this, $\mathrm{sp}_r(H)\geq \Omega_r(\mathrm{sp}_3(H_{3}))$ . But $H_3$ has $r(r-1)\dots 4\cdot m$ edges, so $\mathrm{sp}_3(H_3)=\Omega_r ({m^{3/5}}/{(\log m)^2})$ by Theorem 1·1. Thus, $\mathrm{sp}_r(H)=\Omega_r ({m^{3/5}}/{(\log m)^2})$ as well. Note that we also have $\mathrm{sp}_{r-1}(H_{r-1})=\Omega_r ({m^{3/5}}/{(\log m)^2})$ . Combining this with $\mathrm{sp}_{r-1}(H) \geq ({1}/{2})\mathrm{sp}_{r-1}(H_{r-1})$ guaranteed by Lemma 2·6, we conclude that the statement also holds for $k = r-1$ .

In case H is linear, the maximum co-degree of $H_3$ is $O_r(1)$ , so applying Theorem 5·2 gives the desired result.

6. Concluding remarks

In this paper, we proved that every r-graph with m edges contains an r-partite subhypergraph with at least $({r!}/{r^r})m+m^{3/5-o(1)}$ edges. This is still somewhat smaller than the conjectured bound $({r!}/{r^r})m+\Omega_r(m^{2/3})$ of Conlon, Fox, Kwan and Sudakov [Reference Conlon, Fox, Kwan and Sudakov6]. In order to prove this conjecture, it is enough to prove the following strengthening of Theorem 1·3.

Indeed, our Theorem 1·3 is sharp (up to logarithmic terms) as long as the maximum co-degree is O(1). However, we believe the dependence on the co-degree in our lower bound is not necessary. On another note, we are unable to extend our methods to attack the problem of k-cuts in r-graphs in case $k\leq r-2$ . The following remains an intriguing open problem.

In the case of linear hypergraphs, we proved that $\mathrm{sp}_r(H)\gt m^{3/4-o(1)}$ , which is optimal up to the o(1) term. It would be interesting to extend this result for k-cuts as well in case $k\leq r-2$ , or to improve the o(1) term.