1. Introduction
Clustering is one of the central problems in network analysis and machine learning [Reference Newman, Watts and Strogatz59, Reference Ng, Jordan and Weiss60, Reference Shi and Malik65]. Many clustering algorithms make use of graph models, which represent pairwise relationships among data. A well-studied probabilistic model is the stochastic block model (SBM), which was first introduced in [Reference Holland, Laskey and Leinhardt39] as a random graph model that generates community structure with given ground truth for clusters so that one can study algorithm accuracy. The past decades have brought many notable results in the analysis of different algorithms and fundamental limits for community detection in SBMs in different settings [Reference Coja-Oghlan20, Reference Guédon and Vershynin37, Reference Montanari and Sen54, Reference Van70]. A major breakthrough was the proof of phase transition behaviours of community detection algorithms in various connectivity regimes [Reference Abbe, Bandeira and Hall2, Reference Abbe and Sandon5, Reference Bordenave, Lelarge and Massoulié12, Reference Massoulié52, Reference Mossel, Neeman and Sly55, Reference Mossel, Neeman and Sly57, Reference Mossel, Neeman and Sly58]. See the survey [Reference Abbe1] for more references.
Hypergraphs can represent more complex relationships among data [Reference Battiston, Cencetti, Iacopini, Latora, Lucas, Patania, Young and Petri10, Reference Benson, Gleich and Leskovec11], including recommendation systems [Reference Bu, Tan, Chen, Wang, Wu, Zhang and He13, Reference Li and Li49], computer vision [Reference Govindu34, Reference Wen, Du, Li, Bian and Lyu73], and biological networks [Reference Michoel and Nachtergaele53, Reference Tian, Hwang and Kuang68], and they have been shown empirically to have advantages over graphs [Reference Zhou, Huang and Schölkopf79]. Besides community detection problems, sparse hypergraphs and their spectral theory have also found applications in data science [Reference Harris and Zhu38, Reference Jain and Oh40, Reference Zhou and Zhu80], combinatorics [Reference Dumitriu and Zhu26, Reference Friedman and Wigderson29, Reference Soma and Yoshida66], and statistical physics [Reference Cáceres, Misobuchi and Pimentel14, Reference Sen64].
With the motivation from a broad set of applications, many efforts have been made in recent years to study community detection on random hypergraphs. The hypergraph stochastic block model (HSBM), as a generalisation of graph SBM, was first introduced and studied in [Reference Ghoshdastidar and Dukkipati31]. In this model, we observe a random uniform hypergraph where each hyperedge appears independently with some given probability depending on the community structure of the vertices in the hyperedge.
Succinctly put, the HSBM recovery problem is to find the ground truth clusters either approximately or exactly, given a sample hypergraph and estimates of model parameters. We may ask the following questions about the quality of the solutions (see [Reference Abbe1] for further details in the graph case).
1. Exact recovery (strong consistency): With high probability, find all clusters exactly (up to permutation).
2. Almost exact recovery (weak consistency): With high probability, find a partition of the vertex set such that at most $o(n)$ vertices are misclassified.
3. Partial recovery: Given a fixed $\gamma \in (0.5,1)$, with high probability, find a partition of the vertex set such that at least a fraction $\gamma$ of the vertices are clustered correctly.
4. Weak recovery (detection): With high probability, find a partition correlated with the true partition.
For exact recovery of uniform HSBMs, it was shown that the phase transition occurs in the regime of logarithmic expected degrees in [16, Reference Eli Chien, Lin and Wang17, Reference Lin, Eli Chien and Wang50]. The thresholds are given for binary [Reference Gaudio and Joshi30, Reference Kim, Bandeira and Goemans43] and multiple [Reference Zhang and Tan77] community cases, by generalising the techniques in [Reference Abbe, Bandeira and Hall2–Reference Abbe and Sandon4]. After our work appeared on arXiv, thresholds for exact recovery on non-uniform HSBMs were given by [Reference Dumitriu and Wang25, Reference Wang71]. Strong consistency on the degree-corrected non-uniform HSBM was studied in [Reference Deng, Xu and Ying24]. Spectral methods were considered in [6, 16, Reference Cole and Zhu21, Reference Gaudio and Joshi30, Reference Yuan, Zhao and Zhao75, Reference Zhang and Tan77], while semidefinite programming methods were analysed in [Reference Alaluusua, Avrachenkov, Vinay Kumar and Leskelä8, Reference Kim, Bandeira and Goemans43, Reference Lee, Kim and Chung46]. Weak consistency for HSBMs was studied in [16, Reference Eli Chien, Lin and Wang17, Reference Ghoshdastidar and Dukkipati32, Reference Ghoshdastidar and Dukkipati33, Reference Ke, Shi and Xia42].
For detection of the HSBM, the authors of [Reference Angelini, Caltagirone, Krzakala and Zdeborová9] proposed a conjecture that the phase transition occurs in the regime of constant expected degrees. The positive part of the conjecture for the binary and multi-block case was solved in [Reference Pal and Zhu62] and [Reference Stephan and Zhu67], respectively. Their algorithms can output a partition better than a random guess when above the Kesten-Stigum threshold, but can not guarantee the correctness ratio. [Reference Gu and Pandey35, Reference Gu and Polyanskiy36] proved that detection is impossible and the Kesten-Stigum threshold is tight for $m$-uniform hypergraphs with binary communities when $m =3, 4$, while KS threshold is not tight when $m\geq 7$, and some regimes remain unknown.
1.1. Non-uniform hypergraph stochastic block model
The non-uniform HSBM was first studied in [Reference Ghoshdastidar and Dukkipati32], which removed the uniform hypergraph assumption in previous works, and it is a more realistic model to study higher-order interaction on networks [Reference Lung, Gaskó and Suciu51, Reference Wen, Du, Li, Bian and Lyu73]. It can be seen as a superposition of several uniform HSBMs with different model parameters. We first define the uniform HSBM in our setting and extend it to non-uniform hypergraphs.
Definition 1.1 (Uniform HSBM). Let $V=\{V_1,\dots V_k\}$ be a partition of the set $[n]$ into $k$ blocks of size $\frac{n}{k}$ (assuming $n$ is divisible by $k$). Let $m \in \mathbb{N}$ be some fixed integer. For any set of $m$ distinct vertices $i_1,\dots i_m$, a hyperedge $\{i_1,\dots i_m\}$ is generated with probability $a_m/\binom{n}{m-1}$ if vertices $i_1,\dots i_m$ are in the same block; otherwise with probability $b_m/ \binom{n}{m-1}$. We denote this distribution on the set of $m$-uniform hypergraphs as
Definition 1.2 (Non-uniform HSBM). Let $H = (V, E)$ be a non-uniform random hypergraph, which can be considered as a collection of $m$-uniform hypergraphs, i.e., $H = \bigcup _{m = 2}^{M} H_m$ with each $H_m$ sampled from (1).
Examples of $2$-uniform and $3$-uniform HSBM, and an example of non-uniform HSBM with ${\mathcal M} = \{2, 3\}$ and $k=3$ is displayed in Fig. 1a, b, c respectively.
1.2. Main results
To illustrate our main results, we first introduce the concepts $\gamma$-correctness and sigal-to-noise ratio to measure the accuracy of the obtained partitions.
Definition 1.3 ($\gamma$-correctness). Suppose we have $k$ disjoint blocks $V_1,\dots, V_k$. A collection of subsets $\widehat{V}_1,\dots, \widehat{V}_k$ of $V$ is $\gamma$-correct if $|V_i\cap \widehat{V}_i|\geq \gamma |V_i|$ for all $i\in [k]$.
Definition 1.4. For model 1.2 under Assumption 1.5, we define the signal-to-noise ratio ($\mathrm{SNR}$) as
Let ${\mathcal M}_{\max }$ denote the maximum element in the set $\mathcal M$. The following constant $C_{{\mathcal M}}(k)$ is used to characterise the accuracy of the clustering result,
Note that a non-uniform HSBM can be seen as a collection of noisy observations for the same underlying community structure through several uniform HSBMs of different orders. A possible issue is that some uniform hypergraphs with small SNR might not be informative (if we observe an $m$-uniform hypergraph with parameters $a_m=b_m$, including hyperedge information from it ultimately increases the noise). To improve our error rate guarantees, we start by adding a pre-processing step (Algorithm 3) for hyperedge selection according to SNR and then apply the algorithm on the sub-hypergraph with maximal SNR.
We state the following assumption that will be used in our analysis of Algorithms 1 ($k=2$) and 2 ($k\geq 3$).
Assumption 1.5. For each $m\in{\mathcal M}$, assume $a_m, b_m$ are constants independent of $n$, and $a_m \geq b_m$. Let ${\mathcal M}_{\max }$ denote the maximum element in the set $\mathcal M$. Given $\nu \in (1/k, 1)$, assume that there exists a universal constant $C$ and some $\nu$-dependent constant $C_{\nu } \gt 0$, such that
One does not have to take too large a $C$ for (4a); for example, $C = (2^{1/{\mathcal M}_{\max }} - 1)^{-1/3}$ should suffice, but even smaller $C$ may work. Both of the two inequalities above constant prevent the hypergraph from being too sparse, while (4b) also requires that the difference between in-block and across-blocks densities is large enough. The choices of $C, C_{\nu }$ and their relationship will be discussed in Remark 5.15.
1.2.1. The 2-block case
We start with Algorithm 1, which outputs a $\gamma$-correct partition when the non-uniform HSBM $H$ is sampled from model 1.2 with only $2$ communities. Inspired by the innovative graph algorithm in [Reference Chin, Rao and Van19], we generalise it to non-uniform hypergraphs while we provide a complete and detailed analysis at the same time.
Theorem 1.6 ($k=2$). Let $\nu \in (0.5, 1)$ and $\rho = 2\exp ({-}C_{{\mathcal M}}(2) \cdot \mathrm{SNR}_{{\mathcal M}}(2))$ with $\mathrm{SNR}_{{\mathcal M}}(k)$, $C_{{\mathcal M}}(k)$ defined in (2), (3), and let $\gamma = \max \{\nu,\, 1 - 2\rho \}$. Then under Assumption 1.5, Algorithm 1 outputs a $\gamma$-correct partition for sufficiently large $n$ with probability at least $1 - O(n^{-2})$.
1.2.2. The $k$-block case
For the multi-community case ($k \geq 3$), another algorithm with more subroutines is developed in Algorithm 2, which outputs a $\gamma$-correct partition with high probability. We state the result as follows.
Theorem 1.7 ($k\geq 3$). Let $\nu \in (1/k, 1)$ and $\rho = \exp ({-}C_{{\mathcal M}}(k) \cdot \mathrm{SNR}_{{\mathcal M}}(k))$ with $\mathrm{SNR}_{{\mathcal M}}(k), C_{{\mathcal M}}(k)$ defined in (2), (3), and let $\gamma = \max \{\nu,\, 1 - k\rho \}$. Then under Assumption 1.5, Algorithm 2 outputs a $\gamma$-correct partition for sufficiently large $n$ with probability at least $1 - O(n^{-2})$.
The time complexities of Algorithms 1 and 2 are $O(n^{3})$, with the bulk of time spent in Stage 1 by the spectral method.
To the best of our knowledge, Theorems 1.6 and 1.7 are the first results for partial recovery of non-uniform HSBMs. When the number of blocks is $2$, Algorithm 1 guarantees a better error rate for partial recovery as in Theorem 1.6. This happens because Algorithm 1 does not need the merging routine in Algorithm 2: if one of the communities is obtained, then the other one is also obtained via the complement.
Remark 1.8. Taking ${\mathcal M} = \{2 \}$, Theorem 1.7 can be reduced to [Reference Chin, Rao and Van19, Lemma 9] for the graph case. The failure probability $O(n^{-2})$ can be decreased to $O(n^{-p})$ for any $p\gt 0$, as long as one is willing to pay the price by increasing the constants $C$, $C_v$ in (4a), (4b).
Our Algorithms 1 and 2 can be summarised in 3 steps:
1. Hyperedge selection: select hyperedges of certain sizes to provide the maximal signal-to-noise ratio (SNR) for the induced sub-hypergraph.
2. Spectral partition: construct a regularised adjacency matrix and obtain an approximate partition based on singular vectors (first approximation).
3. Correction and merging: incorporate the hyperedge information from adjacency tensors to upgrade the error rate guarantee (second, better approximation).
The algorithm requires the input of model parameters $a_m,b_m$, which can be estimated by counting cycles in hypergraphs as shown in [Reference Mossel, Neeman and Sly55, Reference Yuan, Liu, Feng and Shang74]. Estimation of the number of blocks can be done by counting the outliers in the spectrum of the non-backtracking operator, e.g., as shown (for different regimes and different parameters) in [Reference Angelini, Caltagirone, Krzakala and Zdeborová9, Reference Le and Levina44, Reference Saade, Krzakala and Zdeborová63, Reference Stephan and Zhu67].
1.2.3. Weak consistency
Throughout the proofs for Theorems 1.6 and 1.7, we make only one assumption on the growth or finiteness of $d$ and $\mathrm{SNR}_{{\mathcal M}}(k)$, and it happens in estimating the failure probability as noted in Remark 1.10. Consequently, the corollary below follows, which covers the case when $d$ and $\mathrm{SNR}_{{\mathcal M}}(k)$ grow with $n$.
Corollary 1.9 (Weak consistency). For fixed $M$ and $k$, if $\mathrm{SNR}_{{\mathcal M}}(k)$ defined in (2) goes to infinity as $n\to \infty$ and $\mathrm{SNR}_{{\mathcal M}}(k) = o(\log n)$, then with probability $1- O(n^{-2})$, Algorithms 1 and 2 output a partition with only $o(n)$ misclassified vertices.
The paper [Reference Ghoshdastidar and Dukkipati32] also proves weak consistency for non-uniform HSBMs, but in a much denser regime than we do here ($d = \Omega (\log ^2(n))$, instead of $d=\omega (1)$, as in Corollary 1.9). In fact, we now know that strong consistency should be achievable in this denser regime, as [Reference Dumitriu and Wang25] shows. When restricting to the uniform HSBM case, Corollary 1.9 achieves weak consistency under the same sparsity condition as in [Reference Ahn, Lee and Suh7].
Remark 1.10. To be precise, Algorithms 1 and 2 work optimally in the $\mathrm{SNR}_{{\mathcal M}} = o (\log n)$ regime. When $\mathrm{SNR}_{{\mathcal M}}(k) = \Omega (\log n)$, it implies that $\rho = n^{-\Omega (1)}$, and one may have $e^{-n\rho } = \Omega (1)$ in (31), which may not decrease to $0$ as $n\to \infty$. Therefore the theoretical guarantees of Algorithms 5 and 6 may not remain valid. This, however, should not matter: in the regime when $\mathrm{SNR}_{{\mathcal M}}(k) = \Omega (\log n)$, strong (rather than weak) consistency is expected, as per [Reference Dumitriu and Wang25]. Therefore, the regime of interest for weak consistency is $\mathrm{SNR}_{{\mathcal M}} = o(\log n)$.
1.3. Comparison with existing results
Although many algorithms and theoretical results have been developed for hypergraph community detection, most of them are restricted to uniform hypergraphs, and few results are known for non-uniform ones. We will discuss the most relevant results.
In [Reference Ke, Shi and Xia42], the authors studied the degree-corrected HSBM with general connection probability parameters by using a tensor power iteration algorithm and Tucker decomposition. Their algorithm achieves weak consistency for uniform hypergraphs when the average degree is $\omega (\log ^2 n)$, which is the regime complementary to the regime we studied here. They discussed a way to generalise the algorithm to non-uniform hypergraphs, but the theoretical analysis remains open. The recent paper [Reference Zhen and Wang78] analysed non-uniform hypergraph community detection by using hypergraph embedding and optimisation algorithms and obtained weak consistency when the expected degrees are of $\omega (\log n)$, again a complementary regime to ours. Results on spectral norm concentration of sparse random tensors were obtained in [Reference Cooper23, Reference Jain and Oh40, Reference Lei, Chen and Lynch47, Reference Nguyen, Drineas and Tran61, Reference Zhou and Zhu80], but no provable tensor algorithm in the bounded expected degree case is known. Testing for the community structure for non-uniform hypergraphs was studied in [Reference Jin, Ke and Liang41, Reference Yuan, Liu, Feng and Shang74], which is a problem different from community detection.
In our approach, we relied on knowing the tensors for each uniform hypergraph. However, in computations, we only ran the spectral algorithm on the adjacency matrix of the entire hypergraph since the stability of tensor algorithms does not yet come with guarantees due to the lack of concentration, and for non-uniform hypergraphs, $M-1$ adjacency tensors would be needed. This approach presented the challenge that, unlike for graphs, the adjacency matrix of a random non-uniform hypergraph has dependent entries, and the concentration properties of such a random matrix were previously unknown. We overcame this issue and proved concentration bounds from scratch down to the bounded degree regime. Similar to [Reference Feige and Ofek28, Reference Le, Levina and Vershynin45], we provided here a regularisation analysis by removing rows in the adjacency matrix with large row sums (suggestive of large degree vertices) and proving a concentration result for the regularised matrix down to the bounded expected degree regime (see Theorem 3.3).
In terms of partial recovery for hypergraphs, our results are new, even in the uniform case. In [Reference Ahn, Lee and Suh7, Theorem 1], for uniform hypergraphs, the authors showed detection (not partial recovery) is possible when the average degree is $\Omega (1)$; in addition, the error rate is not exponential in the model parameters, but only polynomial. Here, we mention two more results for the graph case. In the arbitrarily slowly growing degrees regime, it was shown in [Reference Fei and Chen27, Reference Zhang and Zhou76] that the error rate in (2) is optimal up to a constant in the exponent. In the bounded expected degrees regime, the authors in [Reference Chin and Sly18, Reference Mossel, Neeman and Sly56] provided algorithms that can asymptotically recover the optimal fraction of vertices, when the signal-to-noise ratio is large enough. It’s an open problem to extend their analysis to obtain a minimax error rate for hypergraphs.
In [Reference Ghoshdastidar and Dukkipati32], the authors considered weak consistency in a non-uniform HSBM model with a spectral algorithm based on the hypergraph Laplacian matrix, and showed that weak consistency is achievable if the expected degree is of $\Omega (\log ^2 n)$ with high probability [Reference Ghoshdastidar and Dukkipati31, Theorem 4.2]. Their algorithm can’t be applied to sparse regimes straightforwardly since the normalised Laplacian is not well-defined due to the existence of isolated vertices in the bounded degree case. In addition, our weak consistency results obtained here are valid as long as the expected degree is $\omega (1)$ and $o(\log n)$, which is the entire set of problems on which weak consistency is expected. By contrast, in [Reference Ghoshdastidar and Dukkipati32], weak consistency is shown only when the minimum expected degree is $\Omega (\log ^2(n))$, which is a regime complementary to ours and where exact recovery should (in principle) be possible: for example, this is known to be an exact recovery regime in the uniform case [Reference Eli Chien, Lin and Wang17, Reference Kim, Bandeira and Goemans43, Reference Lee, Kim and Chung46, Reference Zhang and Tan77].
In subsequent works [Reference Dumitriu and Wang25, Reference Wang71] we proposed algorithms to achieve weak consistency. However, their methods can not cover the regime when the expected degree is $\Omega (1)$ due to the lack of concentration. Additionally, [Reference Wang, Pun, Wang, Wang and So72] proposed Projected Tensor Power Method as the refinement stage to achieve strong consistency, as long as the first stage partition is partially correct, as ours.
1.4. Organization of the paper
In Section 2, we include the definitions of adjacency matrices of hypergraphs. The concentration results for the adjacency matrices are provided in Section 3. The algorithms for partial recovery are presented in Section 4. The proof for the correctness of our algorithms for Theorem 1.7 and Corollary 1.9 are given in Section 5. The proof of Theorem 1.6, as well as the proofs of many auxiliary lemmas and useful lemmas in the literature, are provided in the supplemental materials.
2. Preliminaries
Definition 2.1 (Adjacency tensor). Given an $m$-uniform hypergraph $H_m=([n], E_m)$, we can associate to it an order- $m$ adjacency tensor $\boldsymbol{\mathcal{A}}^{(m)}$. For any $m$-hyperedge $e = \{ i_1, \dots, i_m \}$, let $\boldsymbol{\mathcal{A}}^{(m)}_e$ denote the corresponding entry $\boldsymbol{\mathcal{A}}_{[i_{1},\dots,i_{m}]}^{(m)}$, such that
Definition 2.2 (Adjacency matrix). For the non-uniform hypergraph $H$ sampled from model 1.2, let $\boldsymbol{\mathcal{A}}^{(m)}$ be the order- $m$ adjacency tensor corresponding to the underlying $m$-uniform hypergraph for each $m\in{\mathcal M}$. The adjacency matrix ${\mathbf{A}} \,:\!=\, [{\mathbf{A}}_{ij}]_{n \times n}$ of the non-uniform hypergraph $H$ is defined by
We compute the expectation of ${\mathbf{A}}$ first. In each $m$-uniform hypergraph $H_m$, two distinct vertices $i, j\in V$ with $i \neq j$ are picked arbitrarily since our model does not allow for loops. Assume for a moment $\frac{n}{k} \in \mathbb{N}$, then the expected number of $m$-hyperedge containing $i$ and $j$ can be computed as follows.
• If $i$ and $j$ are from the same block, the $m$-hyperedge is sampled with probability $a_m/\binom{n}{m-1}$ when the other $m-2$ vertices are from the same block as $i$, $j$, otherwise with probability $b_m/\binom{n}{m-1}$. Then
\begin{align*} \alpha _m\,:\!=\,{\mathbb{E}}{\mathbf{A}}_{ij} = \binom{\frac{n}{k} -2}{m-2} \frac{a_m}{\binom{n}{m-1}} + \left [ \binom{n-2}{m-2} - \binom{\frac{n}{k} - 2}{m-2} \right ]\frac{b_m}{\binom{n}{m-1}} \,. \end{align*}• If $i$ and $j$ are not from the same block, we sample the $m$-hyperedge with probability $b_m/\binom{n}{m-1}$, and
\begin{align*} \beta _m \,:\!=\,{\mathbb{E}}{\mathbf{A}}_{ij} = \binom{n-2}{m-2} \frac{b_m}{\binom{n}{m-1}}\,. \end{align*}
By assumption $a_m \geq b_m$, then $\alpha _m \geq \beta _m$ for each $m\in{\mathcal M}$. Summing over $m$, the expected adjacency matrix under the $k$-block non-uniform $\mathrm{HSBM}$ can be written as
where ${\mathbf{J}}_{\frac{n}{k}}\in{\mathbb{R}}^{\frac{n}{k} \times \frac{n}{k}}$ denotes the all-one matrix and
Lemma 2.3. The eigenvalues of ${\mathbb{E}}{\mathbf{A}}$ are given below:
Lemma 2.3 can be verified via direct computation. Lemma 2.4 is used for approximately equi-partitions, meaning that eigenvalues of $\widetilde{{\mathbb{E}}{\mathbf{A}}}$ can be approximated by eigenvalues of ${\mathbb{E}}{\mathbf{A}}$ when $n$ is sufficiently large.
Lemma 2.4. For any partition $(V_1, \dots, V_k)$ of $V$ where $n_i \,:\!=\, |V_i|$, consider the following matrix
Assume that $n_i = \frac{n}{k} + O(\sqrt{n} \log n)$ for all $i\in [k]$. Then, for all $1\leq i \leq k$,
Note that both $(\,\widetilde{{\mathbb{E}}{\mathbf{A}}} + \alpha{\mathbf{I}}_{n})$ and $(\,{\mathbb{E}}{\mathbf{A}} + \alpha{\mathbf{I}}_{n})$ are rank $k$ matrices, then $\lambda _i(\widetilde{{\mathbb{E}}{\mathbf{A}}}) = \lambda _i({\mathbb{E}}{\mathbf{A}}) = -\alpha$ for all $(k+1) \leq i \leq n$. At the same time, $\mathrm{SNR}$ in (2) is related to the following quantity
When ${\mathcal M} = \{2\}$ and $k$ is fixed, $\mathrm{SNR}$ in (2) is equal to $\frac{(a - b)^2}{k[a + (k-1)b]}$, which corresponds to the $\mathrm{SNR}$ for the undirected graph in [Reference Chin, Rao and Van19], see also [Reference Abbe1, Section 6].
3. Spectral norm concentration
The correctness of Algorithms 2 and 1 relies on the concentration of the adjacency matrix of $H$. The following two concentration results for general random hypergraphs are included, which are independent of HSBM model. The proofs are deferred to Section A.
Theorem 3.1. Let $H=\bigcup _{m =2}^{M} H_m$, where $H_m = ([n], E_m)$ is an Erdős-Rényi inhomogeneous hypergraph of order $m$ for each $m\in \{2, \cdots, M\}$. Let $\boldsymbol{\mathcal{T}}^{\,(m)}$ denote the probability tensor such that $\boldsymbol{\mathcal{T}}^{(m)} ={\mathbb{E}} \boldsymbol{\mathcal{A}}^{(m)}$ and $\boldsymbol{\mathcal{T}}^{(m)}_{[i_1,\dots, i_m]} = d_{[i_1,\dots,i_m]}/ \binom{n}{m-1}$, denoting $d_m=\max d_{[i_1,\dots, i_m]}$. Suppose for some constant $c\gt 0$,
Then for any $K\gt 0$, there exists a constant $ C= 512M(M-1)(K+6)\left [ 2 + (M-1)(1+K)/c \right ]$ such that with probability at least $1-2n^{-K}-2e^{-n}$, the adjacency matrix ${\mathbf{A}}$ of $H$ satisfies
The inequality (10) can be reduced to the result for graph case obtained in [Reference Feige and Ofek28, Reference Lei and Rinaldo48] by taking ${\mathcal M} = \{2\}$. The result for a uniform hypergraph is obtained in [Reference Lee, Kim and Chung46]. Note that $d$ is a fixed constant in our community detection problem, thus the Assumption 3.1 does not hold and the inequality (9) cannot be directly applied. However, we can still prove a concentration bound for a regularised version of ${\mathbf{A}}$, following the same strategy of the proof for Theorem 3.1.
Definition 3.2 (Regularized matrix). Given any $n\times n$ matrix ${\mathbf{A}}$ and an index set $\mathcal{I}$, let ${\mathbf{A}}_{\mathcal{I}}$ be the $n\times n$ matrix obtained from ${\mathbf{A}}$ by zeroing out the rows and columns not indexed by $\mathcal{I}$. Namely,
Since every hyperedge of size $m$ containing $i$ is counted $(m-1)$ times in the $i$-th row sum of ${\mathbf{A}}$, the $i$-th row sum of ${\mathbf{A}}$ is given by
Theorem 3.3 is the concentration result for the regularised ${\mathbf{A}}_{\mathcal{I}}$, by zeroing out rows and columns corresponding to vertices with high row sums.
Theorem 3.3. Following all the notations in Theorem 3.1, for any constant $\tau \gt 1$, define
Let ${\mathbf{A}}_{\mathcal{I}}$ be the regularised version of ${\mathbf{A}}$, as in Definition 3.2. Then for any $K\gt 0$, there exists a constant $C_{\tau }=2( (5M+1)(M-1)+\alpha _0\sqrt{\tau })$ with $\alpha _0=16+\frac{32}{\tau }(1+e^2)+128M(M-1)(K+4)\left (1+\frac{1}{e^2}\right )$, such that $ \| ({\mathbf{A}}-{\mathbb{E}}{\mathbf{A}})_{\mathcal{I}}\|\leq C_{\tau }\sqrt{d}$ with probability at least $1-2(e/2)^{-n} -n^{-K}$.
4. Algorithms blocks
In this section, we are going to present the algorithmic blocks constructing our main partition method (Algorithm 2): pre-processing (Algorithm 3), initial result by spectral method (Algorithm 4), correction of blemishes via majority rule (Algorithm 5), and merging (Algorithm 6).
4.1. Three or more blocks ($k\geq 3$)
The proof of Theorem 1.7 is structured as follows.
Lemma 4.1. Under the assumptions of Theorem 1.7, Algorithm 4 outputs a $\nu$-correct partition $U_1^{\prime }, \cdots, U_k^{\prime }$ of $Z = (Z\cap V_{1}) \cup \cdots \cup (Z\cap V_{k})$ with probability at least $1 - O(n^{-2})$.
Lines $4$ and $6$ contribute most complexity in Algorithm 4, requiring $O(n^{3})$ and $O(n^{2}\log ^2(n))$ each (technically, one should be able to get away with $O(n^2 \log (1/\varepsilon ))$ in line 4, for some desired accuracy $\varepsilon$ to get the singular vectors). We will conservatively estimate the time complexity of Algorithm 4 as $O(n^3)$.
Lemma 4.2. Under the assumptions of Theorem 1.7, for any $\nu$-correct partition $U_1^{\prime }, \cdots, U_k^{\prime }$ of $Z = (Z\cap V_{1}) \cup \cdots \cup (Z\cap V_{k})$ and the red hypergraph over $Z$, Algorithm 5 computes a $\gamma _{\mathrm{C}}$-correct partition $\widehat{U}_1, \cdots, \widehat{U}_k$ with probability $1 -O(e^{-n\rho })$, while $\gamma _{\mathrm{C}} = \max \{\nu,\, 1 - k\rho \}$ with $ \rho \,:\!=\, k\exp\!\left ({-}C_{{\mathcal M}}(k) \cdot \mathrm{SNR}_{{\mathcal M}}(k) \right )$ where $\mathcal M$ is obtained from Algorithm 3, and $\mathrm{SNR}_{{\mathcal M}}(k)$ and $C_{{\mathcal M}}(k)$ are defined in (2), (3).
Lemma 4.3. Given any $\nu$-correct partition $\widehat{U}_1, \cdots, \widehat{U}_k$ of $Z = (Z\cap V_{1}) \cup \cdots \cup (Z\cap V_{k})$ and the blue hypergraph between $Y$ and $Z$, with probability $1 -O(e^{-n\rho })$, Algorithm 6 outputs a $\gamma$-correct partition $\widehat{V}_1, \cdots, \widehat{V}_{k}$ of $V_1 \cup V_2 \cup \cdots \cup V_k$, while $\gamma =\max \{\nu,\, 1 - k\rho \}$.
The time complexities of Algorithms 5 and 6 are $O(n)$, since each vertex is adjacent to only constant many hyperedges.
4.2. The binary case ($k = 2$)
The spectral partition step is given in Algorithm 7, and the correction step is given in Algorithm 8.
Lemma 4.4. Under the conditions of Theorem 1.6, the Algorithm 7 outputs a $\nu$-correct partition $V_1^{\prime }, V_2^{\prime }$ of $V = V_1 \cup V_2$ with probability at least $1 - O(n^{-2})$.
Lemma 4.5. Given any $\nu$-correct partition $V_1^{\prime }, V_2^{\prime }$ of $V = V_1 \cup V_2$, with probability at least $1 -O(e^{-n\rho })$, the Algorithm 8 computes a $\gamma$-correct partition $\widehat{V}_1, \widehat{V}_2$ with $\gamma = \{\nu,\, 1 - 2\rho \}$ and $\rho = 2\exp ({-}C_{{\mathcal M}}(2) \cdot \mathrm{SNR}_{{\mathcal M}}(2))$, where $\mathrm{SNR}_{{\mathcal M}}(2)$ and $C_{{\mathcal M}}(2)$ are defined in (2), (3).
5. Algorithm’s correctness
We are going to present the correctness of Algorithm 2 in this section. The correctness of Algorithm 1 is deferred to Section C. We first introduce some definitions.
Vertex set splitting and adjacency matrix
In Algorithm 2, we first randomly partition the vertex set $V$ into two disjoint subsets $Z$ and $Y$ by assigning $+1$ and $-1$ to each vertex independently with equal probability. Let ${\mathbf{B}} \in{\mathbb{R}}^{|Z|\times |Y|}$ denote the submatrix of ${\mathbf{A}}$, while ${\mathbf{A}}$ was defined in (6), where rows and columns of ${\mathbf{B}}$ correspond to vertices in $Z$ and $Y$ respectively. Let $n_i$ denote the number of vertices in $Z\cap V_i$, where $V_i$ denotes the true partition with $|V_i| = \frac{n}{k}$ for all $i\in [k]$, then $n_i$ can be written as a sum of independent Bernoulli random variables, i.e.,
and $|Y\cap V_i| = |V_i| - |Z\cap V_i| = \frac{n}{k} - n_i$ for each $i \in [k]$.
Definition 5.1. The splitting $V = Z \cup Y$ is perfect if $|Z\cap V_i| = |Y\cap V_i| = n/(2k)$ for all $i\in [k]$. And the splitting $Y = Y_1 \cup Y_2$ is perfect if $|Y_1\cap V_i| = |Y_2\cap V_i| = n/(4k)$ for all $i\in [k]$.
However, the splitting is imperfect in most cases since the size of $Z$ and $Y$ would not be exactly the same under the independence assumption. The random matrix ${\mathbf{B}}$ is parameterised by $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m \in{\mathcal M}}$ and $\{n_i\}_{i=1}^{k}$. If we take expectation over $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m \in{\mathcal M}}$ given the block size information $\{n_i\}_{i=1}^{k}$, then it gives rise to the expectation of the imperfect splitting, denoted by $\widetilde{{\mathbf{B}}}$,
where $\alpha$, $\beta$ are defined in (8). In the perfect splitting case, the dimension of each block is $n/(2k)\times n/(2k)$ since ${\mathbb{E}} n_i = n/(2k)$ for all $i \in [k]$, and the expectation matrix $\overline{{\mathbf{B}}}$ can be written as
In Algorithm 4, $Y_1$ is a random subset of $Y$ obtained by selecting each element with probability $1/2$ independently, and $Y_2 = Y\setminus Y_1$. Let $n^{\prime }_{i}$ denote the number of vertices in $Y_1\cap V_i$, then $n^{\prime }_{i}$ can be written as a sum of independent Bernoulli random variables,
and $|Y_2\cap V_i| = |V_i| - |Z\cap V_i| - |Y_1\cap V_i| = n/k - n_i - n^{\prime }_{i}$ for all $i \in [k]$.
Induced sub-hypergraph
Definition 5.2 (Induced Sub-hypergraph). Let $H = (V, E)$ be a non-uniform random hypergraph and $S\subset V$ be any subset of the vertices of $H$. Then the induced sub-hypergraph $H[S]$ is the hypergraph whose vertex set is $S$ and whose hyperedge set $E[S]$ consists of all of the edges in $E$ that have all endpoints located in $S$.
Let $H[Y_1 \cup Z]$(resp. $H[Y_2 \cup Z]$) denote the induced sub-hypergraph on vertex set $Y_1\cup Z$ (resp. $Y_2 \cup Z$), and ${\mathbf{B}}_1 \in{\mathbb{R}}^{|Z|\times |Y_1|}$ (resp. ${\mathbf{B}}_2 \in{\mathbb{R}}^{|Z|\times |Y_2|}$) denote the adjacency matrices corresponding to the sub-hypergraphs, where rows and columns of ${\mathbf{B}}_1$ (resp. ${\mathbf{B}}_2$) are corresponding to elements in $Z$ and $Y_1$ (resp., $Z$ and $Y_2$). Therefore, ${\mathbf{B}}_1$ and ${\mathbf{B}}_2$ are parameterised by $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m \in{\mathcal M}}$, $\{n_i\}_{i=1}^{k}$ and $\{n^{\prime }_{i}\}_{i=1}^{k}$, and the entries in ${\mathbf{B}}_1$ are independent of the entries in ${\mathbf{B}}_2$, due to the independence of hyperedges. If we take expectation over $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m \in{\mathcal M}}$ conditioning on $\{n_i\}_{i=1}^{k}$ and $\{n^{\prime }_{i}\}_{i=1}^{k}$, then it gives rise to the expectation of the imperfect splitting, denoted by $\widetilde{{\mathbf{B}}}_1$,
where
The expectation of the perfect splitting, denoted by $\overline{{\mathbf{B}}}_1$, can be written as
where
The matrices $\widetilde{{\mathbf{B}}}_2, \overline{{\mathbf{B}}}_2$ can be defined similarly, since dimensions of $|Y_2\cap V_i|$ are also determined by $n_i$ and $n^{\prime }_{i}$. Obviously, $\overline{{\mathbf{B}}}_2 = \overline{{\mathbf{B}}}_1$ since ${\mathbb{E}} n^{\prime }_{i} ={\mathbb{E}} (n/k- n_{i} - n_i^{\prime }) = n/(4k)$ for all $i\in [k]$.
Fixing Dimensions
The dimensions of $\widetilde{{\mathbf{B}}}_1$ and $\widetilde{{\mathbf{B}}}_2$, as well as blocks they consist of, are not deterministic – since $n_i$ and $n^{\prime }_i$, defined in (12) and (13) respectively, are sums of independent random variables. As such, we cannot directly compare them. In order to overcome this difficulty, we embed ${\mathbf{B}}_1$ and ${\mathbf{B}}_2$ into the following $n\times n$ matrices:
Note that ${\mathbf{A}}_1$ and ${\mathbf{A}}_2$ have the same size. Also by definition, the entries in ${\mathbf{A}}_1$ are independent of the entries in ${\mathbf{A}}_2$. If we take expectation over $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m\in{\mathcal M}}$ conditioning on $\{n_i\}_{i=1}^{k}$ and $\{n^{\prime }_{i}\}_{i=1}^{k}$, then we obtain the expectation matrices of the imperfect splitting, denoted by $\widetilde{{\mathbf{A}}}_1$(resp. $\widetilde{{\mathbf{A}}}_2$), written as
The expectation matrix of the perfect splitting, denoted by $\overline{{\mathbf{A}}}_1$(resp. $\overline{{\mathbf{A}}}_2$), can be written as
Obviously, $\widetilde{{\mathbf{A}}}_i$ and $\widetilde{{\mathbf{B}}}_i$(resp. $\overline{{\mathbf{A}}}_i$ and $\overline{{\mathbf{B}}}_i$) have the same non-zero singular values for $i = 1, 2$. In the remaining of this section, we will deal with $\widetilde{{\mathbf{A}}}_i$ and $\overline{{\mathbf{A}}}_i$ instead of $\widetilde{{\mathbf{B}}}_i$ and $\overline{{\mathbf{B}}}_i$ for $i = 1, 2$.
5.1. Spectral partition: Proof of Lemma 4.1
5.1.1. Proof outline
Recall that ${\mathbf{A}}_1$ is defined as the adjacency matrix of the induced sub-hypergraph $H[Y_1\cup Z]$ in equation (5.2). Consequently, the index set should contain information only from $H[Y_1\cup Z]$. Define the index sets
where $d = \sum _{m \in{\mathcal M}}(m-1)a_m$, and $\mathrm{row}(i)\big |_{Y_1 \cup Z}$ is the row sum of $i$ on $H[Y_1\cup Z]$. We say $\mathrm{row}(i)\big |_{Y_1 \cup Z}=0$ if $i\not \in Y_1\cup Z$, and for vertex $i\in Y_1\cup Z$,
As a result, the matrix $({\mathbf{A}}_1)_{\mathcal{I}_1}$ is obtained by restricting ${\mathbf{A}}_1$ on index set $\mathcal{I}_1$. The next $4$ steps guarantee that Algorithm 4 outputs a $\nu$-correct partition.
(i) Find the singular subspace ${\mathbf{U}}$ spanned by the first $k$ left singular vectors of $({\mathbf{A}}_1)_{\mathcal{I}_1}$.
(ii) Randomly pick $s = 2k \log ^2 n$ vertices from $Y_2$ and denote the corresponding columns in ${\mathbf{A}}_2$ by $\boldsymbol{a}_{i_1},\dots,\boldsymbol{a}_{i_s}$. Project each vector $\boldsymbol{a}_i - \overline{\boldsymbol{a}}$ onto the singular subspace ${\mathbf{U}}$, with $\overline{\boldsymbol{a}}\in{\mathbb{R}}^{n}$ defined by $\overline{\boldsymbol{a}}(j) = \unicode{x1D7D9}_{j\in Z} \cdot (\overline{\alpha } + \overline{\beta })/2$, where $\overline{\alpha }$, $\overline{\beta }$ were defined in (17).
(iii) For each projected vector $P_{{\mathbf{U}}}(\boldsymbol{a}_i - \overline{\boldsymbol{a}})$, identify the top $n/(2k)$ coordinates in value and place the corresponding vertices into a set $U^{\prime }_i$. Discard half of the obtained $s$ subsets, those with the lowest blue edge densities.
(iv) Sort the remaining sets according to blue hyperedge density and identify $k$ distinct subsets $U^{\prime }_1, \cdots, U^{\prime }_k$ such that $|U^{\prime }_i \cap U^{\prime }_j| \lt \lceil (1 - \nu )n/k\rceil$ if $i\neq j$.
Based on the $4$ steps above in Algorithm 4, the proof of Lemma 4.1 is structured in $4$ parts.
(i) Let $\widetilde{{\mathbf{U}}}$ denote the subspace spanned by first $k$ left singular vectors of $\widetilde{{\mathbf{A}}}_1$ defined in (19). Subsection 5.1.2 shows that the subspace angle between ${\mathbf{U}}$ and $\widetilde{{\mathbf{U}}}$ is smaller than any $c\in (0, 1)$ as long as $a_m, b_m$ satisfy certain conditions depending on $c$.
(ii) The vector $\widetilde{{\boldsymbol \delta }}_i$, defined in (22), reflects the underlying true partition $Z\cap V_{k(i)}$ for each $i\in [s]$, where $k(i)$ denotes the membership of vertex $i$. Subsection 5.1.3 shows that $\overline{{\boldsymbol \delta }}_i$, an approximation of $\widetilde{{\boldsymbol \delta }}_i$ defined in (23), can be recovered by the projected vector $P_{{\mathbf{U}}}(\boldsymbol{a}_{i} - \overline{\boldsymbol{a}})$, since projection error $\|P_{{\mathbf{U}}}(\boldsymbol{a}_{i} - \overline{\boldsymbol{a}}) - \overline{{\boldsymbol \delta }}_{i}\|_2 \lt c\|\overline{{\boldsymbol \delta }}_{i}\|_2$ for any $c\in (0, 1)$ if $a_m, b_m$ satisfy the desired property in part (i).
(iii) Subsection 5.1.4 indicates that the coincidence ratio between the remaining sets and the true partition is at least $\nu$, after discarding half of the sets with the lowest blue edge densities.
(iv) Lemma 5.13 proves that we can find $k$ distinct subsets $U^{\prime }_i$ within $k\log ^2 n$ trials with high probability.
5.1.2. Bounding the angle between ${\mathbf{U}}$ and $\widetilde{{\mathbf{U}}}$
The angle between subspaces ${\mathbf{U}}$ and $\widetilde{{\mathbf{U}}}$ is defined as
A natural idea is to apply Wedin’s $\sin \Theta$ Theorem (Lemma D.7). Lemma 5.3 indicates that the difference of $\sigma _i(\widetilde{{\mathbf{A}}}_1)$ and $\sigma _i(\overline{{\mathbf{A}}}_1)$ is relatively small, compared to $\sigma _i(\overline{{\mathbf{A}}}_1)$.
Lemma 5.3. Let $\sigma _i(\overline{{\mathbf{A}}}_1)$(resp. $\sigma _i(\widetilde{{\mathbf{A}}}_1)$) denote the singular values of $\overline{{\mathbf{A}}}_1$ (resp. $\widetilde{{\mathbf{A}}}_1$) for all $i\in [k]$, where the matrices $\overline{{\mathbf{A}}}_1$ and $\widetilde{{\mathbf{A}}}_1$ are defined in (20) and (19) respectively. Then
with $\overline{\alpha }$, $\overline{\beta }$ defined in (17). Moreover, with probability at least $1 - 2k\exp ({-}k\log ^2(n))$,
Therefore, with Lemma 5.3, we can write $ \sigma _i(\widetilde{{\mathbf{A}}}_1) = \sigma _i(\overline{{\mathbf{A}}}_1)(1 + o(1)).$ Define ${\mathbf{E}}_1 \,:\!=\,{\mathbf{A}}_1 - \widetilde{{\mathbf{A}}}_1$ and its restriction on $\mathcal{I}_1$ as
as well as ${\boldsymbol \Delta }_{1}\,:\!=\, (\widetilde{{\mathbf{A}}}_1)_{\mathcal{I}_1} - \widetilde{{\mathbf{A}}}_1$. Then $({\mathbf{A}}_1)_{\mathcal{I}_1} - \widetilde{{\mathbf{A}}}_1$ is decomposed as
Lemma 5.4. Let $d = \sum _{m\in{\mathcal M}} (m-1)a_m$, where $\mathcal M$ is obtained from Algorithm 3. There exists a constant $C_1\geq (2^{1/{\mathcal M}_{\max }} - 1)^{-1/3}$ such that if $d \geq C_1$, then with probability at least $1 - \exp\!\left ({-} d^{-2}n/{\mathcal M}_{\max } \right )$, no more than $d^{-3}n$ vertices have row sums greater than $20{\mathcal M}_{\max }d$.
Lemma 5.4 shows that the number of high-degree vertices is relatively small. Consequently, Corollary 5.5 indicates $\|{\boldsymbol \Delta }_1\| \leq \sqrt{d}$ with high probability.
Corollary 5.5. Assume $d \geq \max \{C_1, \sqrt{2}\}$, where $C_1$ is the constant in Lemma 5.4, then $\|{\boldsymbol \Delta }_1\| \leq \sqrt{d}$ with probability at least $1 - \exp\!\left ({-} d^{-2}n/{\mathcal M}_{\max } \right )$.
Proof of Corollary 5.5. Note that $n - |\mathcal{I}| \leq d^{-3}n$ and $\mathcal{I} \subset \mathcal{I}_1$, then $n - |\mathcal{I}_1| \leq d^{-3}n$. From Lemma 5.4, there are at most $d^{-3} n$ vertices with row sum greater than $20{\mathcal M}_{\max }d$ in the adjacency matrix ${\mathbf{A}}_1$, then the matrix ${\boldsymbol \Delta }_1 = (\widetilde{{\mathbf{A}}}_1)_{\mathcal{I}_1} - \widetilde{{\mathbf{A}}}_1$ has at most $2d^{-3} n^{2}$ non-zero entries. Every entry of $\widetilde{{\mathbf{A}}}_1$ in (19) is bounded by $\alpha$, then,
Moreover, taking $\tau =20{\mathcal M}_{\max }, K=3$ in Theorem 3.3, with probability at least $1 - n^{-2}$
where constant $C_3$ depends on ${\mathcal M}_{\max }$. Together with upper bounds for $\|({\mathbf{E}}_1)_{\mathcal{I}_1}\|$ and $\|{\boldsymbol \Delta }_1\|$, Lemma 5.6 shows that the angle between ${\mathbf{U}}$ and $\widetilde{{\mathbf{U}}}$ is relatively small with high probability.
Lemma 5.6. For any $c\in (0,1)$, there exists some constant $C_2$ such that, if
then $\sin \angle ({\mathbf{U}}, \widetilde{{\mathbf{U}}}) \leq c$ with probability $1 - n^{-2}$. Here $\angle ({\mathbf{U}}, \widetilde{{\mathbf{U}}})$ is the angle between ${\mathbf{U}}$ and $\widetilde{{\mathbf{U}}}$.
Proof of Lemma 5.6. From (22) and Corollary 5.5, with probability at least $1 - n^{-2}$,
Since $\sigma _{k+1}(\widetilde{{\mathbf{A}}}_1)=0$, using Lemma 5.3 to approximate $\sigma _k(\widetilde{{\mathbf{A}}}_1)$, we obtain
Then for any $c\in (0, 1)$, we can find $C_2 = [2^{{\mathcal M}_{\max } + 2}(C_3 + 1)/c]$ such that $\|({\mathbf{A}}_1)_{\mathcal{I}_1} - \widetilde{{\mathbf{A}}}_1\|\leq (1-1/\sqrt{2}) \sigma _k(\widetilde{{\mathbf{A}}}_1)$. By Wedin’s Theorem (Lemma D.7), the angle $\angle ({\mathbf{U}}, \widetilde{{\mathbf{U}}})$ is bounded by
5.1.3. Bound the projection error
Randomly pick $s=2k\log ^2 n$ vertices from $Y_2$. Let $\boldsymbol{a}_{i_1},\dots,\boldsymbol{a}_{i_s}$, $\widetilde{\boldsymbol{a}}_{i_1}$,$\dots$, $\widetilde{\boldsymbol{a}}_{i_s}$, $\overline{\boldsymbol{a}}_{i_1},\dots, \overline{\boldsymbol{a}}_{i_s}$ and $\boldsymbol{e}_{i_1}, \dots,\boldsymbol{e}_{i_s}$ be the corresponding columns of ${\mathbf{A}}_2$, $\widetilde{{\mathbf{A}}}_2$, $\overline{{\mathbf{A}}}_2$ and ${\mathbf{E}}_2\,:\!=\,{\mathbf{A}}_2 - \widetilde{{\mathbf{A}}}_2$ respectively, where ${\mathbf{A}}_2$, $\widetilde{{\mathbf{A}}}_2$ and $\overline{{\mathbf{A}}}_2$ were defined in (18), (19) and (20). Let $k(i)$ denote the membership of vertex $i$. Note that entries of vector $\widetilde{\boldsymbol{a}}_{i}$ are $\widetilde{\alpha }_{ii}$, $\widetilde{\beta }_{ij}$ or $0$, according to the membership of vertices in $Z$, where $\widetilde{\alpha }_{ii}$, $\widetilde{\beta }_{ij}$ were defined in (15a), (15b). Then the corresponding vector $\widetilde{{\boldsymbol \delta }}_i \in{\mathbb{R}}^{n}$ with the entries given by
can be used to recover the vertex set $Z\cap V_{k(i)}$ based on the sign of elements in $\widetilde{{\boldsymbol \delta }}_i$. However, it is hard to handle with $\widetilde{{\boldsymbol \delta }}_i$ due to the randomness of $\widetilde{\alpha }_{ii}$, $\widetilde{\beta }_{ij}$ originated from $n_i$ and $n^{\prime }_i$. Note that $n_i$ and $n^{\prime }_i$ concentrate around $n/(2k)$ and $n/(4k)$ respectively as shown in Lemma 5.3. Thus a good approximation of $\widetilde{{\boldsymbol \delta }}_i$, which rules out randomness of $n_i$ and $n^{\prime }_i$, was given by $\overline{{\boldsymbol \delta }}_i \,:\!=\, \overline{\boldsymbol{a}}_{i} - \overline{\boldsymbol{a}}$, with entries given by $\overline{\boldsymbol{a}}(j)\,:\!=\, \unicode{x1D7D9}_{j\in Z}\cdot (\overline{\alpha } + \overline{\beta })/2$, where $\overline{\alpha }$ and $\overline{\beta }$ were defined in (17), and
By construction, $\overline{{\boldsymbol \delta }}_i$ identifies vertex set $Z \cap V_{k(i)}$ in the case of perfect splitting for any $i\in \{i_1, \cdots, i_s\}\cap Y_2\cap V_{k(i)}$. However, the access to $\overline{{\boldsymbol \delta }}_i$ is limited in practice, thus the projection $P_{{\mathbf{U}}}(\boldsymbol{a}_i - \overline{\boldsymbol{a}})$ is used instead as an approximation of $\overline{{\boldsymbol \delta }}_i$. Lemma 5.7 proves that at least half of the projected vectors have small projection errors.
Lemma 5.7. For any $c\in (0, 1)$, there exist constants $C_1$ and $C_2$ such that if $d\gt C_1$ and
then among all projected vectors $P_{{\mathbf{U}}}(\boldsymbol{a}_{i} - \overline{\boldsymbol{a}})$ for $i\in \{i_1, \cdots, i_s\}\cap Y_2$, with probability $1 - O(n^{-k})$, at least half of them satisfy
Proof Lemma 5.7. Note that $\overline{{\boldsymbol \delta }}_i = P_{\overline{{\mathbf{U}}}}\overline{{\boldsymbol \delta }}_i$, where $\overline{{\mathbf{U}}}$ is spanned by the first $k$ left singular vectors of $\overline{{\mathbf{A}}}_1$ with $\mathrm{rank}(\overline{{\mathbf{A}}}_1) = k$, and $\overline{{\mathbf{A}}}_1$, $\overline{{\mathbf{A}}}_2$ preserve the same eigen-subspace. The approximation error between $ P_{{\mathbf{U}}}(\boldsymbol{a}_i - \overline{\boldsymbol{a}})$ and $\overline{{\boldsymbol \delta }}_i$ can be decomposed as
Then by triangle inequality,
Note that $\|\overline{{\boldsymbol \delta }}_i\| = O(n^{-\frac{1}{2} })$ and $n^{\prime }_{i}$ concentrates around $n/(4k)$ for each $i\in [k]$ with deviation at most $\sqrt{n}\log (n)$, then by definitions of $\overline{\alpha }$ and $\overline{\beta }$ in (17),
Meanwhile, by an argument similar to Lemma 5.6, it can be proved that $\sin \angle ({\mathbf{U}}, \overline{{\mathbf{U}}}) \lt c/2$ for any $c\in (0, 1)$, if constants $C_1, C_2$ are chosen properly, hence $\|P_{{\mathbf{U}}} - P_{\overline{{\mathbf{U}}}}\|\cdot \|\overline{{\boldsymbol \delta }}_i\|_{2}\lt \frac{c}{2} \|\overline{{\boldsymbol \delta }}_i\|_{2}$. Lemma 5.8 shows that at least half of the indices from $\{i_1, \cdots, i_s\}\cap Y_2$ satisfy $\|P_{{\mathbf{U}}}\boldsymbol{e}_i \|_2 \lt \frac{c}{2}\|\overline{{\boldsymbol \delta }}_{i}\|_2$, which completes the proof.
Lemma 5.8. Let $d = \sum _{m\in{\mathcal M}_{\max }} (m-1)a_m$. For any $c\in (0, 1)$, with probability $1 - O(n^{-k\log n})$, at least $\frac{s}{2}$ of the vectors $\boldsymbol{e}_{i_1}, \dots,\boldsymbol{e}_{i_s}$ satisfy
Definition 5.9. The vector $\boldsymbol{a}_i$ satisfying (25) is referred as good vector. The index of the good vector is hence referred to as a good vertex.
To avoid introducing extra notations, let $\boldsymbol{a}_{i_1}, \dots,\boldsymbol{a}_{i_{s_1}}$ denote good vectors with $i_1, \cdots, i_{s_1}$ denoting good indices. Lemma 5.8 indicates that the number of good vectors is at least $s_1\geq \frac{s}{2} = k\log ^2 n$.
5.1.4. Accuracy
We are going to prove the accuracy of the initial partition obtained from Algorithm 4. Lemmas 5.10, 5.11 and 5.13 are crucial in proving our results. We present the proof logic first and defer the Lemma statements later.
For each projected vector $P_{{\mathbf{U}}}(\boldsymbol{a}_{i} - \overline{\boldsymbol{a}})$, let $U^{\prime }_{i}$ denote the set of its largest $\frac{n}{2k}$ coordinates, where $i\in \{i_1, \cdots, i_s\}$ and $s = 2k \log ^2 n$. Note that vector $\overline{{\boldsymbol \delta }}_{i_j}$ in (23) only identifies blocks $V_{k(i_j)}$ and $V\setminus V_{k(i_j)}$, which can be regarded as clustering two blocks with different sizes. By Lemma 5.7, good vectors satisfy $\|P_{{\mathbf{U}}}(\boldsymbol{a}_{i_j} - \overline{\boldsymbol{a}}) - \overline{{\boldsymbol \delta }}_{i_j}\|_2 \lt c \, \|\overline{{\boldsymbol \delta }}_{i_j}\|_2$ for any $c\in (0, 1)$. Then by Lemma 5.10 (after proper normalisation), for a good index $i_j$, the number of vertices in $U^{\prime }_{i_j}$ clustered correctly is at least $(1 - \frac{4}{3}kc^{2})\frac{n}{k}$. By choosing $c = \sqrt{3(1-\nu )/(8k)}$, the condition $|U^{\prime }_{i_j} \cap V_i| \gt (1+\nu )/2 |U^{\prime }_{i_j}|$ in part (ii) of Lemma 5.11 is satisfied. In Lemma 5.6, we choose
where $C_3$ defined in (22). Hence, with high probability, all good vectors have at least $\mu _{\mathrm{T}}$ blue hyperedges (we call this “high blue hyperedge density”). From Lemma 5.8, at least half of the selected vectors are good. Then, in Algorithm 4, throwing out half of the obtained sets $U^{\prime }_{i}$ (those with the lowest blue hyperedge density) guarantees that the remaining sets are good.
Recall that, by choosing constant appropriately, we can make the subspace angle $\sin \angle ({\mathbf{U}}, \overline{{\mathbf{U}}}) \lt c$ for any $c\in (0, 1)$ ($\overline{{\mathbf{U}}}$ is spanned by the first $k$ left singular vectors of $\overline{{\mathbf{A}}}_1$). Then for each vector $\overline{{\boldsymbol \delta }}_{i_1}, \cdots, \overline{{\boldsymbol \delta }}_{i_k}$ with each $i_j$ selected from different vertex set $V_j$, there is a vector $P_{{\mathbf{U}}}(\boldsymbol{a}_{i_j} - \overline{\boldsymbol{a}})$ in ${\mathbf{U}}$ arbitrarily close to $\overline{{\boldsymbol \delta }}_{i_j}$, which was proved by Lemma 5.7. From (i) of Lemma 5.11, so obtained $U_{i_j}^{\prime }$ must satisfy $|U_{i_j}^{\prime }\cap V_j|\geq \nu |U_{i_j}^{\prime }|$ for each $j\in [k]$. The remaining thing is to select $k$ different $U_{i_j}^{\prime }$ with each of them concentrating around distinct $V_j$ for each $j\in [k]$. This problem is equivalent to finding $k$ vertices in $Y_2$, each from a different partition class, which can be done with $k\log ^2(n)$ samplings as shown in Lemma 5.13.
To summarise, this section is a more precise and quantitative version of the following argument: with high probability,
Lemma 5.10 (Adapted from Lemma 23 in [Reference Chin, Rao and Van19]). Suppose $n, k$ are such that $\frac{n}{k} \in \mathbb{N}$. Let $\boldsymbol{v},\bar{\boldsymbol{v}} \in{\mathbb{R}}^n$ be two unit vectors, and let $\bar{\boldsymbol{v}}$ be such that $\frac{n}{k}$ of its entries are equal to $\frac{1}{\sqrt{n}}$ and the rest are equal to $-\frac{1}{\sqrt{n}}$. If $\sin \angle (\bar{\boldsymbol{v}},\boldsymbol{v}) \lt c \leq 0.5$, then $\boldsymbol{v}$ contains at least $(1 - \frac{4}{3}kc^{2}) \frac{n}{k}$ positive entries $\boldsymbol{v}_i$ such that $\bar{\boldsymbol{v}}_i$ is also positive.
Lemma 5.11. Suppose that we are given a set $X \subset Z$ with size $|X| = n/(2k)$. Define
and $\mu _{\mathrm{T}} \,:\!=\, (\mu _1 + \mu _2)/2 \in [\mu _1, \mu _2]$. There is a constant $c\gt 0$ depending on $k, a_m, \nu$ such that for sufficiently large $n$,
(i) If $|X\cap V_i| \leq \nu |X|$ for each $i\in [k]$, then with probability $1 - e^{-cn}$, the number of blue hyperedges in the hypergraph induced by $X$ is at most $\mu _{\mathrm{T}}$.
(ii) Conversely, if $|X\cap V_i| \geq \frac{1+ \nu }{2}|X|$ for some $i\in \{1, \dots, k\}$, then with probability $1 - e^{-cn}$, the number of blue hyperedges in the hypergraph induced by $X$ is at least $\mu _{\mathrm{T}}$.
Remark 5.12. Lemma 5.11 is reduced to [Reference Chin, Rao and Van19, Lemma 31] when ${\mathcal M} = \{2\}$.
Lemma 5.13. Through random sampling without replacement in Step $6$ of Algorithm 4, we can find at least $k$ indices $i_1,\dots, i_k$ in $Y_2$ among $k\log ^2 n$ samples such that with probability $1 -n^{-\Omega (\log n)}$,
5.2. Local correction: Proof of Lemma 4.2
For notation convenience, let $U_i\,:\!=\, Z \cap V_i$ denote the intersection of $Z$ and true partition $V_i$ for all $i\in [k]$. In Algorithm 2, we first colour the hyperedges with red and blue with equal probability. By running Algorithm 4 on the red hypergraph, we obtain a $\nu$-correct partition $U_1^{\prime }, \dots, U_k^{\prime }$, i.e.,
In the rest of this subsection, we condition on the event that (27) holds true.
Consider a hyperedge $e = \{i_{1}, \cdots, i_{m}\}$ in the underlying $m$-uniform hypergraph. If vertices $i_{1}, \cdots, i_{m}$ are from the same block, then $e$ is a red hyperedge with probability $a_{m}/2\binom{n}{m-1}$; if vertices $i_{1}, \cdots, i_{m}$ are not from the same block, then $e$ is a red hyperedge with probability $b_{m}/2\binom{n}{m-1}$. The presence of those two types of hyperedges can be denoted by
respectively. For any finite set $S$, let $[S]^{l}$ denote the family of $l$-subsets of $S$, i.e., $[S]^{l} = \{Z| Z\subseteq S, |Z| = l \}$. Consider a vertex $u\in U_1\,:\!=\, Z \cap V_1$. The weighted number of red hyperedges, which contains $u\in U_1$ with the remaining vertices in $U_j^{\prime }$, can be written as
where $\mathcal{E}^{(a_m)}_{1, j}\,:\!=\, E_m([U_1]^{1}, [ U_1 \cap U_j^{\prime }]^{m-1} )$ denotes the set of $m$-hyperedges with one vertex from $[U_1]^{1}$ and the other $m-1$ from $[U_1 \cap U^{\prime }_{j}]^{m-1}$, while $ \mathcal{E}^{(b_m)}_{1, j} \,:\!=\, E_m \Big ([U_1]^{1}, \,\, [U_j^{\prime }]^{m-1} \setminus [U_1\cap U_j^{\prime }]^{m-1} \Big )$ denotes the set of $m$-hyperedges with one vertex in $[U_1]^{1}$ while the remaining $m-1$ vertices in $[U_j^{\prime }]^{m-1}\setminus [U_1 \cap U_j^{\prime }]^{m-1}$(not all $m$ vertices are from $V_1$) with their cardinalities
We multiply $(m-1)$ in (28) as weight since the rest $m-1$ vertices are all located in $U^{\prime }_{j}$, which can be regarded as $u$’s neighbours in $U^{\prime }_{j}$. According to the fact $|U^{\prime }_j \cap U_j| \geq (\nu n/2k)$ in (27) and $|U_{j}^{\prime }| = n/(2k)$ for $j\in [k]$,
To simplify the calculation, we take the lower and upper bound of $\big|\mathcal{E}^{(a_m)}_{1, 1}\big|$ and $\big|\mathcal{E}^{(a_m)}_{1, j}\big|\,(j\neq 1)$ respectively. By taking expectation with respect to $T_{e}^{(a_m)}$ and $T_{e}^{(b_m)}$, then for any $u \in U_1$, we have
By assumptions in Theorem 1.7, ${\mathbb{E}} S_{11}^{\prime }(u) -{\mathbb{E}} S_{1j}^{\prime }(u) = \Omega (1)$. Define
In Algorithm 5, vertex $u$ is assigned to $\widehat{U}_{i}$ if it has the maximal number of neighbours in $U_i^{\prime }$. If $u\in U_1$ is mislabelled, then one of the following events must happen:
• $S_{11}^{\prime }(u) \leq \mu _{\mathrm{C}}$, meaning that $u$ was mislabelled by Algorithm 5.
• $S_{1j}^{\prime }(u) \geq \mu _{\mathrm{C}}$ for some $j\neq 1$, meaning that $u$ survived Algorithm 5 without being corrected.
Lemma 5.14 shows that the probabilities of those two events can be bounded in terms of the SNR.
Lemma 5.14. For sufficiently large $n$ and any $u\in U_1 = Z \cap V_1$, we have
where $\rho \,:\!=\, \exp\!\left ({-}C_{{\mathcal M}}\cdot \mathrm{SNR}_{{\mathcal M}} \right )$ with $\mathrm{SNR}_{{\mathcal M}}$ and $C_{{\mathcal M}}$ defined in (2).
As a result, the probability that either of those events happened is bounded by $\rho$. The number of mislabelled vertices in $U_1$ after Algorithm 5 is at most
where $\Gamma _{t}$ (resp. $\Lambda _{t}$) are i.i.d indicator random variables with mean $\rho _1^{\prime }$ (resp. $\rho _j^{\prime }$, $j\neq 1$). Then
Let $t_1\,:\!=\, n \rho/2$, where $\nu$ denotes the correctness after Algorithm 4, then by Chernoff bound (Lemma D.1),
Then with probability $1 - O(e^{-n\rho })$, the fraction of mislabelled vertices in $U_1$ is smaller than $k\rho$, i.e., the correctness of $U_1$ is at least $\gamma _{\mathrm{C}} \,:\!=\, \max \{\nu \,, 1 - k\rho \}$. Therefore, Algorithm 5 outputs a $\gamma _{\mathrm{C}}$-correct partition $\widehat{U}_1, \cdots, \widehat{U}_{k}$ with probability $1- O(e^{-n\rho })$.
5.3. Merging: Proof of Lemma 4.3
By running Algorithm 5 on the red hypergraph, we obtain a $\gamma _{\mathrm{C}}$-correct partition $\widehat{U}_1, \cdots, \widehat{U}_{k}$ where $\gamma _{\mathrm{C}}\,:\!=\,\max \{\nu \,, 1 - k\rho \} \geq \nu$, i.e.,
In the rest of this subsection, we shall condition on this event and abbreviate $Y\cap V_l$ by $W_l\,:\!=\, Y\cap V_l$. The failure probability of Algorithm 6 is estimated by the presence of hyperedges between vertex sets $Y$ and $Z$.
Consider a hyperedge $e = \{i_{1}, \cdots, i_{m}\}$ in the underlying $m$-uniform hypergraph. If vertices $i_{1}, \cdots, i_{m}$ are all from the same cluster $V_{l}$, then the probability that $e$ is an existing blue edge conditioning on the event that $e$ is not a red edge is
and the presence of $e$ can be represented by an indicator random variable $\zeta _e^{(a_m)} \sim \mathrm{Bernoulli}\left (\psi _m\right )$. Similarly, if vertices $i_{1}, \cdots, i_{m}$ are not all from the same cluster $V_l$, the probability that $e$ is an existing blue edge conditioning on the event that $e$ is not red
and the presence of $e$ can be represented by an indicator variable $\xi _e^{(b_m)} \sim \mathrm{Bernoulli}\left (\phi _m\right )$.
For any vertex $w\in W_l \,:\!=\, Y\cap V_l$ with fixed $l \in [k]$, we want to compute the number of hyperedges containing $w$ with all remaining vertices located in vertex set $\widehat{U}_{j}$ for some fixed $j\in [k]$. Following a similar argument given in Subsection 5.2, this number can be written as
where $\widehat{\mathcal{E}}^{(a_m)}_{l, j}\,:\!=\, E_m([W_l]^{1}, [U_l\cap \widehat{U}_j]^{m-1})$ denotes the set of $m$-hyperedges with $1$ vertex from $[W_l]^{1}$ and the other $m-1$ vertices from $[U_l\cap \widehat{U}_j]^{m-1}$, while $ \widehat{\mathcal{E}}^{(b_m)}_{l, j} \,:\!=\, E_m ([W_l]^{1}, \,\, [\widehat{U}_j]^{m-1} \setminus [U_l\cap \widehat{U}_j]^{m-1})$ denotes the set of $m$-hyperedges with $1$ vertex in $[W_l]^{1}$ while the remaining $m-1$ vertices are in $[\widehat{U}_j]^{m-1}\setminus [U_l\cap \widehat{U}_j]^{m-1}$, with their cardinalities
Similarly, we multiply $(m-1)$ in (35) as weight since the rest $m-1$ vertices can be regarded as $u$’s neighbours in $\widehat{U}_{j}$. By accuracy of Algorithm 5 in (32), $|\widehat{U}_{j} \cap U_j| \geq \nu n/(2k)$, then
Taking expectation with respect to $\zeta _e^{(a_m)}$ and $\xi _e^{(b_m)}$, for any $w \in W_l$, we have
By assumptions in Theorem 1.7, ${\mathbb{E}} \widehat{S}_{ll}(w) -{\mathbb{E}} \widehat{S}_{lj}(w) = \Omega (1)$. We define
After Algorithm 6, if a vertex $w\in W_l$ is mislabelled, one of the following events must happen
• $\widehat{S}_{ll}(w) \leq \mu _{\mathrm{M}}$, which implies that $u$ was mislabelled by Algorithm 6.
• $\widehat{S}_{lj}(w) \geq \mu _{\mathrm{M}}$ for some $j\neq l$, which implies that $u$ survived Algorithm 6 without being corrected.
By an argument similar to Lemma 5.14, we can prove that for any $w\in W_l$,
where $\rho \,:\!=\, \exp\!\left ({-}C_{{\mathcal M}}\cdot \mathrm{SNR}_{{\mathcal M}} \right )$. The misclassified probability for $w\in W_l$ is upper bounded by $\sum _{j=1}^{k}\hat{\rho }_{j} \leq k\rho$. The number of mislabelled vertices in $W_l$ is at most $ R_l = \sum _{t=1}^{|W_l|}\Gamma _{t} \,,$ where $\Gamma _{t}$ are i.i.d indicator random variables with mean $k \rho$ and ${\mathbb{E}} R_l \leq n/(2k)\cdot k\rho = n\rho/2$. Let $t_l\,:\!=\, n \rho/2$, by Chernoff bound (Lemma D.1),
Hence with probability $1 - O(e^{-n\rho })$, the fraction of mislabelled vertices in $W_l$ is smaller than $k\rho$, i.e., the correctness in $W_l$ is at least $\gamma _{\mathrm{M}} \,:\!=\,\max \{ \nu,\, 1 - k\rho \}$.
5.4. Proof of Theorem 1.7
Now we are ready to prove Theorem 1.7. The correctness of Algorithms 5 and 6 are denoted by $\gamma _{\mathrm{C}}$ and $\gamma _{\mathrm{M}}$ respectively, then with probability at least $1 - O(e^{-n\rho })$, the correctness $\gamma$ of Algorithm 2 is $ \gamma \,:\!=\, \min \{\gamma _{\mathrm{C}}, \gamma _{\mathrm{M}}\} = \max \{\nu,\, 1 - k\rho \}.$ We will have $\gamma = 1 - k\rho$ if $\nu \leq 1 - k\rho$, equivalently,
otherwise $\gamma = \nu$. The inequality (37) holds since
where the first two inequalities hold since $d \,:\!=\, \sum _{m\in{\mathcal M}}(m-1)a_m$ and Condition (4b), while the last inequality holds by taking $C_{\nu }\geq \max \{ \sqrt{({\mathcal M}_{\max }-1)/C_{{\mathcal M}} }, C_2 \}$ with $C_2$ defined in (26).
Remark 5.15. The lower bound $C$ in (4a) comes from the requirement in Lemma 5.4 that only a few high-degree vertices be deleted. The constant $C_{\nu }$ in (4b) comes from the requirement in Lemma 5.6 that the subspace angle is small. When $C$ is not so large (or the hypergraph is too sparse), one could still achieve good accuracy $\gamma$ if $C_{\nu }$ is large enough (the difference between $a_m$ and $b_m$ is large enough).
Remark 5.16. Condition (37) indicates that the improvement of accuracy from local refinement (Algorithms 5 and 6) will be guaranteed when $\text{SNR}_{{\mathcal M}}(k)$ is large enough. If $\text{SNR}_{{\mathcal M}}(k)$ is small, we use correctness of Algorithm 4 instead, i.e., $\gamma = \nu$, to represent the correctness of Algorithm 2.
5.5. Proof of Corollary 1.9
For any fixed $\nu \in (1/k, 1)$, $\mathrm{SNR}_{{\mathcal M}}(k) \to \infty$ implies $\rho \to 0$ and
Since
Condition (4b) is satisfied. Applying Theorem 1.7, we find $\gamma = 1 - o(1)$, which implies weak consistency. The constraint of $\mathrm{SNR}_{{\mathcal M}}(k) = o(\log n)$ is used in the proof of Lemma 4.2, see Remark 1.10.
Acknowledgements
I.D. and H.W. are partially supported by NSF DMS-2154099. I.D. and Y.Z. acknowledge support from NSF DMS-1928930 during their participation in the programme Universality and Integrability in Random Matrix Theory and Interacting Particle Systems hosted by the Mathematical Sciences Research Institute in Berkeley, California during the Fall semester of 2021. Y.Z. is partially supported by NSF-Simons Research Collaborations on the Mathematical and Scientific Foundations of Deep Learning. Y.Z. thanks Zhixin Zhou for his helpful comments. The authors thank the anonymous reviewers for detailed comments and suggestions which greatly improve the presentation of this work.
Appendix A. Proof of Theorems 3.1 and 3.3
A.1. Discretization
To prove Theorem 3.1, we start with a standard $\varepsilon$-net argument.
Lemma A.1 (Lemma 4.4.1 in [Reference Vershynin69]). Let ${\mathbf{W}}$ be any Hermitian $n\times n$ matrix and let $\mathcal{N}_{\varepsilon }$ be an $\varepsilon$-net on the unit sphere $\mathbb{S}^{n-1}$ with $\varepsilon \in (0,1)$, then $ \|{\mathbf{W}}\| \leq \frac{1}{1-\varepsilon }\sup _{\boldsymbol{x}\in \mathcal{N}_{\varepsilon }} |\langle{\mathbf{W}} \boldsymbol{x}, \boldsymbol{x}\rangle |.$
By [Reference Vershynin69, Corollary 4.2.13], the size of $\mathcal{N}_{\varepsilon }$ is bounded by $|\mathcal{N}_{\varepsilon }|\leq (1+2/\varepsilon )^n$. We would have $\log |\mathcal{N}| \leq n \log (5)$ when $\mathcal{N}$ is taken as an $(1/2)$-net of $\mathbb S^n$. Define ${\mathbf{W}} \,:\!=\,{\mathbf{A}} -{\mathbb{E}}{\mathbf{A}}$, then ${\mathbf{W}}_{ii}=0$ for each $i\in [n]$ by the definition of adjacency matrix in equation (7), and we obtain
For any fixed $\boldsymbol{x}\in \mathbb{S}^{n-1}$, consider the light and heavy pairs as follows.
where $d = \sum _{m=2}^{M}(m-1)d_m$. Thus by the triangle inequality,
and by equation (A.1),
A.2. Contribution from light pairs
For each $m$-hyperedge $e\in E_m$, we define $\boldsymbol{\mathcal{W}}^{(m)}_e \,:\!=\, \boldsymbol{\mathcal{A}}^{(m)}_e -{\mathbb{E}}\boldsymbol{\mathcal{A}}^{(m)}_e.$ Then for any fixed $\boldsymbol{x}\in \mathbb{S}^{n-1}$, the contribution from light couples can be written as
where the constraint $i\neq j$ comes from the fact ${\mathbf{W}}_{ii} =0$ and we denote
Note that ${\mathbb{E}} \boldsymbol{\mathcal{Y}}_e^{(m)} = 0$, and by the definition of light pair equation (A.2),
Moreover, equation (A.4) is a sum of independent, mean-zero random variables, and
when $n\geq 2m-2$, where $d_m=\max d_{[i_1,\dots, i_m]}$ and $d = \sum _{m=2}^{M}(m-1)d_m$. Then Bernstein’s inequality (Lemma D.3) implies that for any $\alpha \gt 0$,
Therefore by taking a union bound,
where we choose $\alpha =5M(M-1)$ in the last line.
A.3. Contribution from heavy pairs
Note that for any $i\not =j$,
and
Therefore it suffices to show that, with high probability,
Here we use the discrepancy analysis from [Reference Cook, Goldstein and Johnson22, Reference Feige and Ofek28]. We consider the weighted graph associated with the adjacency matrix ${\mathbf{A}}$.
Definition A.2 (Uniform upper tail property, UUTP). Let ${\mathbf{M}}$ be an $n\times n$ random symmetric matrix with non-negative entries and ${\mathbf{Q}}$ be an $n\times n$ symmetric matrix with entries ${\mathbf{Q}}_{ij}\in [0,a]$ for all $i,j\in [n]$. Define
We say that ${\mathbf{M}}$ satisfies the uniform upper tail property $\mathbf{UUTP}(c_0,\gamma _0)$ with $c_0\gt 0,\gamma _0\geq 0$, if for any $a,t\gt 0$,
where function $f_{{\mathbf{Q}}}({\mathbf{M}})\,:\,{\mathbb{R}}^{n \times n} \mapsto{\mathbb{R}}$ is defined by $f_{{\mathbf{Q}}}({\mathbf{M}})\,:\!=\, \sum _{i,j=1}^n{\mathbf{Q}}_{ij}{\mathbf{M}}_{ij}$ for ${\mathbf{M}} \in {\mathbb{R}}^{n \times n}$, and function $h(x) \,:\!=\, (1 + x) \mathrm{log}(1 + x) - x$ for all $x\gt -1$.
Lemma A.3. Let ${\mathbf{A}}$ be the adjacency matrix of non-uniform hypergraph $H = \bigcup _{m=2}^{M}H_m$, then ${\mathbf{A}}$ satisfies $\mathbf{UUTP}(c_0,\gamma _0)$ with $c_0 = [M(M-1)]^{-1}, \gamma _0 = 0$.
Proof of Lemma A.3. Note that
where $\boldsymbol{\mathcal{Z}}^{(m)}_e = \boldsymbol{\mathcal{W}}^{(m)}_e \big ( \sum _{\{i,j\}\subset e, i\neq j}{\mathbf{Q}}_{ij}\big )$ are independent centred random variables upper bounded by $ |\boldsymbol{\mathcal{Z}}^{(m)}_e| \leq \sum _{ \{i,j\} \subset e,\, i\neq j}{\mathbf{Q}}_{ij}\leq M(M-1)a$ for each $ m\in \{2, \dots, M\}$ since ${\mathbf{Q}}_{ij} \in [0, a]$. Moreover, the variance of the sum can be written as
where the last inequality holds since by definition ${\mathbb{E}}{\mathbf{A}}_{ij} = \sum _{m=2}^M \,\,\sum _{\substack{e\in E_m\\ \{i,j\}\subset e}}{\mathbb{E}} [\boldsymbol{\mathcal{A}}^{(m)}_{e}]$. Then by Bennett’s inequality Lemma D.4, we obtain
where the inequality holds since the function $x\cdot h(1/x) = (1+x)\log (1 + 1/x) - 1$ is decreasing with respect to $x$.
Definition A.4 (Discrepancy property, DP). Let ${\mathbf{M}}$ be an $n\times n$ matrix with non-negative entries. For $S,T\subset [n]$, define $ e_{{\mathbf{M}}}(S,T)=\sum _{i\in S, j\in T}{\mathbf{M}}_{ij}.$ We say ${\mathbf{M}}$ has the discrepancy property with parameter $\delta \gt 0$, $\kappa _1\gt 1, \kappa _2\geq 0$, denoted by $\mathbf{DP}(\delta,\kappa _1,\kappa _2)$, if for all non-empty $S,T\subset [n]$, at least one of the following hold:
(1) $e_{{\mathbf{M}}}(S,T)\leq \kappa _1 \delta |S| |T|$;
(2) $e_{{\mathbf{M}}}(S,T) \cdot \log \left(\frac{e_{{\mathbf{M}}}(S,T)}{\delta |S|\cdot |T|}\right)\leq \kappa _2 (|S|\vee |T|)\cdot \log \left(\frac{en}{|S|\vee |T|}\right)$.
Lemma A.5 shows that if a symmetric random matrix ${\mathbf{A}}$ satisfies the upper tail property $\mathbf{UUTP}(c_0,\gamma _0)$ with parameter $c_0\gt 0,\gamma _0\geq 0$, then the discrepancy property holds with high probability.
Lemma A.5 (Lemma 6.4 in [Reference Cook, Goldstein and Johnson22]). Let ${\mathbf{M}}$ be an $n\times n$ symmetric random matrix with non-negative entries. Assume that for some $\delta \gt 0$, $\delta \gt 0$, ${\mathbb{E}}{\mathbf{M}}_{ij}\leq \delta$ for all $i,j\in [n]$ and ${\mathbf{M}}$ has $\mathbf{UUTP}(c_0,\gamma _0)$ with parameter $c_0,\gamma _0 \gt 0$. Then for any $K\gt 0$, the discrepancy property $\mathbf{DP}(\delta,\kappa _1,\kappa _2)$ holds for ${\mathbf{M}}$ with probability at least $1 - n^{-K}$ $ \kappa _1 = e^2(1+\gamma _0)^2, \kappa _2 = \frac{2}{c_0}(1+\gamma _0)(K+4).$
When the discrepancy property holds, then deterministically the contribution from heavy pairs is $O(\sqrt{d})$, as shown in the following lemma.
Lemma A.6 (Lemma 6.6 in [Reference Cook, Goldstein and Johnson22]). Let ${\mathbf{M}}$ be a non-negative symmetric $n\times n$ matrix with all row sums bounded by $d$. Suppose ${\mathbf{M}}$ has ${\mathbf{M}}$ has $\mathbf{DP}(\delta,\kappa _1,\kappa _2)$ with $\delta =Cd/n$ for some $C\gt 0,\kappa _1\gt 1$, $\kappa _2\geq 0$. Then for any $x\in \mathbb{S}^{n-1}$,
where $ \alpha _0=16+32C(1+\kappa _1)+64\kappa _2 (1+\frac{2}{\kappa _1\log \kappa _1}).$
Lemma A.7 proves that ${\mathbf{A}}$ has bounded row and column sums with high probability.
Lemma A.7. For any $K\gt 0$, there is a constant $\alpha _1\gt 0$ such that with probability at least $1-n^{-K}$,
with $\alpha _1= 4 + \frac{2(M-1)(1+K)}{3c}$ and $d\geq c\log n$.
Proof. For a fixed $i\in [n]$,
Then for $\alpha _1= 4 + \frac{2(M-1)(1+K)}{3c}$, by Bernstein’s inequality, with the assumption that $d\geq c\log n$,
Taking a union bound over $i\in [n]$, then equation (A.9) holds with probability $1-n^{-K}$.
Now we are ready to obtain equation (A.8).
Lemma A.8. For any $K\gt 0$, there is a constant $\beta$ depending on $K, c, M$ such that with probability at least $1-2n^{-K}$,
Proof. By Lemma A.3, ${\mathbf{A}}$ satisfies $\mathbf{UUTP}(\frac{1}{M(M-1)},0)$. From equation (A.6) and Lemma A.5, the property $\mathbf{DP}(\delta, \kappa _1,\kappa _2)$ holds for ${\mathbf{A}}$ with probability at least $1-n^{-K}$ with
Let $\mathcal{E}_1$ be the event that $\mathbf{DP}(\delta, \kappa _1,\kappa _2)$ holds for ${\mathbf{A}}$. Let $\mathcal{E}_2$ be the event that all row sums of ${\mathbf{A}}$ are bounded by $\alpha _1 d$. Then ${\mathbb{P}} (\mathcal{E}_1 \cap \mathcal{E}_2)\geq 1-2n^{-K}.$ On the event $\mathcal{E}_1 \cap \mathcal{E}_2$, by Lemma A.6, equation (A.11) holds with $ \beta = \alpha _0\alpha _1,$ where
A.4. Proof of Theorem 3.1
Proof. From equation (A.5), with probability at least $1-2e^{-n}$, the contribution from light pairs in equation (A.3) is bounded by $2\alpha \sqrt{d}$ with $\alpha =5M(M-1)$. From equations (A.7) and (A.11), with probability at least $1-2n^{-K}$, the contribution from heavy pairs in equation (A.3) is bounded by $2\sqrt{d}+2\beta \sqrt{d}$. Therefore with probability at least $1-2e^{-n}-2n^{-K}$,
where $C_M$ is a constant depending only on $c, K, M$ such that $ C_M= 2(\alpha +1+\beta ).$ In particular, we can take $\alpha =5M(M-1),$ $\beta =512M(M-1)(K+5)\left ( 2+\frac{(M-1)(1+K)}{c}\right )$, and $ C_M=512M(M-1)(K+6)\left ( 2+\frac{(M-1)(1+K)}{c}\right ).$ This finishes the proof of Theorem 3.1.
A.5. Proof of Theorem 3.3
Let ${\mathcal{S}}\subset [n]$ be any given subset. From equation (A.5), with probability at least $1-2e^{-n}$,
Since there are at most $2^n$ many choices for $\mathcal{S}$, by taking a union bound, with probability at least $1-2(e/2)^{-n}$, we have for all ${\mathcal{S}}\subset [n]$, equation (A.12) holds. In particular, by taking ${\mathcal{S}} = \mathcal{I} = \{i\in [n]\,:\, \mathrm{row}(i)\leq \tau d\}$, with probability at least $1-2(e/2)^{-n}$, we have
Similar to equation (A.7), deterministically,
Next we show the contribution from heavy pairs for ${\mathbf{A}}_{\mathcal{I}}$ is bounded.
Lemma A.9. For any $K\gt 0$, there is a constant $\beta _{\tau }$ depending on $K, c, M,\tau$ such that with probability at least $1-n^{-K}$,
Proof. Note that ${\mathbf{A}}$ satisfies $\mathbf{UUTP}\left (\frac{1}{M(M-1)},0 \right )$ from Lemma A.3. According to Lemma A.5, with probability at least $1-n^{-K}$, $\mathbf{DP}(\delta,\kappa _1,\kappa _2)$ holds for ${\mathbf{A}}$ with
The $\mathbf{DP}(\delta,\kappa _1,\kappa _2)$ property holds for ${\mathbf{A}}_{\mathcal{I}}$ as well, since ${\mathbf{A}}_{\mathcal{I}}$ is obtained from ${\mathbf{A}}$ by restricting to $\mathcal{I}$. Note that all row sums in ${\mathbf{A}}_{\mathcal{I}}$ are bounded by $\tau d$. By Lemma A.6,
where we can take $ \alpha _0=16+\frac{32}{\tau }(1+e^2)+128M(M-1)(K+4)\left (1+\frac{1}{e^2}\right ).$
We can then take $\beta _{\tau }=\alpha _0\sqrt{\tau }$ in equation (A.15). Therefore, combining equations (A.13), (A.14), (A.16), with probability at least $1-2(e/2)^{-n}-n^{-K}$, there exists a constant $C_{\tau }$ depending only on $\tau, M, K$ such that $ \|({\mathbf{A}}-{\mathbb{E}}{\mathbf{A}})_{\mathcal{I}}\|\leq C_{\tau } \sqrt{d},$ where $C_{\tau }=2( (5M+1)(M-1)+\alpha _0\sqrt{\tau })$. This finishes the proof of Theorem 3.3.
Appendix B. Technical Lemmas
B.1. Proof of Lemma 2.4
Proof. By Weyl’s inequality (Lemma D.5), the difference between eigenvalues of $\widetilde{{\mathbb{E}}{\mathbf{A}}}$ and ${\mathbb{E}}{\mathbf{A}}$ can be upper bounded by
The lemma follows, as $\lambda _i({\mathbb{E}}{\mathbf{A}}) = \Omega \left ( n (\alpha - \beta ) \right )$ for all $1 \leq i \leq k$.
B.2. Proof of Lemma 5.3
Proof. We first compute the singular values of $\overline{{\mathbf{B}}}_1$. From equation (16), the rank of matrix $\overline{{\mathbf{B}}}_1$ is $k$, and the least non-trivial singular value of $\overline{{\mathbf{B}}}_1$ is
where $\mathcal{M}$ is obtained from Algorithm 3. By the definition of $\overline{{\mathbf{A}}}_1$ in equation (20), the least non-trivial singular value of $\overline{{\mathbf{A}}}_1$ is
Recall that $n_i$, defined in equation (12), denotes the number of vertices in $Z\cap V_i$, which can be written as $n_i = \sum _{v\in V_i} \mathbf{1}_{\{ v \in Z\} }.$ By Hoeffding’s Lemma D.2,
Similarly, $n^{\prime }_{i}$, defined in equation (13), satisfies
As defined in equations (14) and (16), both $\widetilde{{\mathbf{B}}}_1$ and $\overline{{\mathbf{B}}}_1$ are deterministic block matrices. Then with probability at least $1 - 2k\exp\!\left ({-} k\log ^2 (n)\right )$, the dimensions of each block inside $\widetilde{{\mathbf{B}}}_1$ and $\overline{{\mathbf{B}}}_1$ are approximately the same, with deviations up to $\sqrt{n}\log (n)$. Consequently, the matrix $\widetilde{{\mathbf{A}}}_1$, which was defined in equation (19), can be treated as a perturbed version of $\overline{{\mathbf{A}}}_1$. By Weyl’s inequality (Lemma D.5), for any $i\in [k]$,
As a result, with probability at least $1 - 2k\exp ({-}k \log ^2(n))$, we have
B.3. Proof of Lemma 5.4
Proof. Without loss of generality, we can assume $\mathcal M=\{2,\dots, M\}$. If $\mathcal M$ is a subset of $\{2,\dots, M\}$, we can take $a_m=b_m=0$ for $m\not \in \mathcal M$. Note that in fact, if the best SNR is obtained when $\mathcal{M}$ is a strict subset, we can substitute $\mathcal{M}_{\max }$ for $M$.
Let $X\subset V$ be a subset of vertices in hypergraph $H = (V, E)$ with size $|X| = cn$ for some $c\in (0,1)$ to be decided later. Suppose $X$ is a set of vertices with high degrees that we want to zero out. We first count the $m$-uniform hyperedges on $X$ separately, then weight them by $(m-1)$, and finally sum over $m$ to compute the row sums in ${\mathbf{A}}$ corresponding to each vertex in $X$. Let $E_m(X)$ denote the set of $m$-uniform hyperedges with all vertices located in $X$, and $E_m (X^{c})$ denote the set of $m$-uniform hyperedges with all vertices in $X^{c}=V\setminus X$, respectively. Let $E_m(X, X^{c})$ denote the set of $m$-uniform hyperedges with at least $1$ endpoint in $X$ and $1$ endpoint in $X^{c}$. The relationship between total row sums and the number of non-uniform hyperedges in the vertex set $X$ can be expressed as
If the row sum of each vertex $v \in X$ is at least $20Md$, where $d = \sum _{m=2}^{M} (m-1)a_m$, it follows
Then either
B.3.1. Concentration of $\sum _{m=2}^{M} m(m-1)|E_m(X)|$
Recall that $\big |E_m(X)\big |$ denotes the number of $m$-uniform hyperedges with all vertices located in $X$, which can be viewed as the sum of independent Bernoulli random variables $T_{e}^{(a_m)}$ and $T_{e}^{(b_m)}$ given by
Let {$V_{1}, \dots, V_{k}$} be the true partition of $V$. Suppose that there are $\eta _{i}cn$ vertices in block $V_{i}\cap X$ for each $i\in [k]$ with restriction $\sum _{i=1}^{k}\eta _{i} = 1$, then $\big |E_m(X)\big |$ can be written as
where $E_{m}(X, a_m) \,:\!=\, \cup _{i=1}^{k}E_{m}(V_i\cap X)$ denotes the union for sets of hyperedges with all vertices in the same block $V_i\cap X$ for some $i\in [k]$, and
denotes the set of hyperedges with vertices crossing different $V_i\cap X$. We can compute the expectation of $\big |E_m(X)\big |$ as
Then
As $\sum \limits _{i=1}^k \eta _i=1$, it follows that $ \sum _{i=1}^k \binom{\eta _i c n}{m} \leq \binom{cn}{m}$ by induction, thus
where both terms on the right are positive numbers. Using this and taking $b_m = a_m$, we obtain the following upper bound for all $n$,
Note that $\sum _{m = 2}^{M} m(m-1) |E_m(X)|$ is a weighted sum of independent Bernoulli random variables (corresponding to hyperedges), each upper bounded by $M^2$. Also, its variance is bounded by
We can apply Bernstein’s Lemma D.3 and obtain
B.3.2. Concentration of $\sum _{m=2}^{M} (m-1)^2|E_m(X, X^{c})|$
For any finite set $S$, let $[S]^{j}$ denote the family of $j$-subsets of $S$, i.e., $[S]^{j} = \{Z| Z\subseteq S, |Z| = j \}$. Let $E_m([Y]^{j}, [Z]^{m-j})$ denote the set of $m$-hyperedges, where $j$ vertices are from $Y$ and $m-j$ vertices are from $Z$ within each $m$-hyperedge. We want to count the number of $m$-hyperedges between $X$ and $X^c$, according to the number of vertices located in $X^c$ within each $m$-hyperedge. Suppose that there are $j$ vertices from $X^c$ within each $m$-hyperedge for some $1\leq j \leq m-1$.
(i) Assume that all those $j$ vertices are in the same $[V_{i}\setminus X]^{j}$. If the remaining $m-j$ vertices are from $[V_{i}\cap X]^{m-j}$, then this $m$-hyperedge is connected with probability $a_m/ \binom{n}{m-1}$, otherwise $b_m/ \binom{n}{m-1}$. The number of this type $m$-hyperedges can be written as
\begin{align*} \sum _{i=1}^{k} \left [ \sum _{e\in \mathcal{E}^{(a_m)}_{j, i}} T_{e}^{(a_m)} + \sum _{e\in \mathcal{E}^{(b_m)}_{j, i}} T_{e}^{(b_m)} \right ]\,, \end{align*}where $\mathcal{E}^{(a_m)}_{j, i}\,:\!=\, E_m([V_i \cap X^c]^{j}, [V_i\cap X]^{m-j})$, and\begin{align*} \mathcal{E}^{(b_m)}_{j, i} \,:\!=\, E_m \Big ([V_i\cap X^c]^{j}, \,\, [X]^{m-j} \setminus [V_i\cap X]^{m-j} \Big ) \end{align*}denotes the set $m$-hyperedges with $j$ vertices in $[V_i\cap X^c]^{j}$ and the remaining $m-j$ vertices in $[X]^{j}\setminus [V_i\cap X]^{j}$. We compute all possible choices and upper bound the cardinality of $\mathcal{E}^{(a_m)}_{j, i}$ and $\mathcal{E}^{(b_m)}_{j, i}$ by\begin{align*} \big |\mathcal{E}^{(a_m)}_{j, i} \big | \leq &\, \binom{ (\frac{1}{k} - \eta _{i}c)n}{j} \binom{\eta _{i}cn}{m - j}\,,\quad \big |\mathcal{E}^{(b_m)}_{j, i} \big | \leq \binom{ (\frac{1}{k} - \eta _{i}c)n}{j} \left [ \binom{cn}{m - j} - \binom{\eta _{i}cn}{m - j} \right ]\,. \end{align*}(ii) If those $j$ vertices in $[V\setminus X]^{j}$ are not in the same $[V_i \cap X]^{j}$ (which only happens $j \geq 2$), then the number of this type hyperedges can be written as $\sum _{e\in \mathcal{E}^{(b_m)}_{j}} T_{e}^{(b_m)}$, where
\begin{align*} \mathcal{E}^{(b_m)}_{j} \,:\!=\,&\, E_m \Big ([V\setminus X]^{j}\setminus \big ( \cup _{i=1}^{k}[V_i\setminus X]^{j} \big ), \,\,\, [X]^{m-j}\Big )\,,\\ \big | \mathcal{E}^{(b_m)}_{j} \big | \leq &\, \left [ \binom{(1-c)n}{j} -\sum _{i=1}^{k} \binom{( \frac{1}{k} - \eta _{i}c) n}{j} \right ] \binom{cn}{m-j}\,. \end{align*}
Therefore, $|E_m(X, X^c)|$ can be written as a sum of independent Bernoulli random variables,
Then the expectation can be rewritten as
where we used the fact $\binom{(1-c)n}{1} = \sum _{i=1}^{k}\binom{(1/k - \eta _i c)n}{1}$ in the first equality and Vandermonde’s identity $\left({n_{1} + n_{2} \atop m}\right) = \sum _{j=0}^{m}\left({n_{1} \atop j}\right)\left({n_{2} \atop m-j}\right)$ in last equality. Note that
counts the number of subsets of $V$ with $m$ elements such that at least one element belongs to $X$ and at least one element belongs to $X^c$. On the other hand,
counts the number of subsets of $V$ with $m$ elements such that all elements belong to a single $V_i$, and given such an $i$, that at least one element belongs to $X \cap V_i$ and at least one belongs to $X^c \cap V_i$.
As Fig. B.1 shows, $g_{c}$ only counts the blue pairs while $f_c$ counts red pairs in addition. By virtue of the fact that there are fewer conditions imposed on the sets included in the count for $f_c$, we must have $f_c \geq g_c$. Thus, rewriting equation (B.8), we obtain
Since both terms in the above sum are positive, we can upper bound by taking $a_m=b_m$ to obtain
By summing over $m$, the expectation of $\sum _{m=2}^{M}(m-1)^2|E_m(X, X^{c})|$ satisfies
where the last upper inequality holds when $c\in (0, 2^{1/M} - 1]$, since
where the last inequality holds by the following Claim.
Claim B.1. Let $m \geq 2$ be some finite integer. Then for $0 \lt c \lt = 2^{1/m} - 1$, it follows that $(1 + c)^{m} - 1 \leq 2mc$.
Proof of the Claim We finish the proof by induction. First, the argument $(1 + c)^{j} - 1 \leq 2jc$ holds true for the base cases $j = 1, 2$. Suppose that the argument holds for the case $j \geq 2$. For the case $j + 1 \leq m$, it follows that
where the last inequality holds true if $c(1 + c)^{j} \leq 2c$, and it holds since $c \leq 2^{1/m} - 1 \leq 2^{1/j} - 1$ for all $j\leq m$.
Similarly, we apply Bernstein Lemma D.3 again with $K=M^2$, $\sigma ^2\leq 8M^3cnd$ and obtain
By the binomial coefficient upper bound $\binom{n}{k}\leq (\frac{en}{k})^k$ for $1\leq k\leq n$, there are at most
many subsets $X$ of size $|X| = cn$. Let $d$ be sufficiently large so that $d^{-3} \leq c_0$. Substituting $c = d^{-3}$ in equation (B.11), we have
Taking $c=d^{-3}$ in equations (B.6) and (B.10), we obtain
Taking a union bound over all possible $X$ with $|X|=d^{-3}n$, we obtain with probability at least $1-2\exp (3d^{-3}\log d n-2d^{-2}n/M)\leq 1-2\exp ({-}d^{-2}n/M)$, no more than $d^{-3}n$ many vertices have total row sum greater than $20Md$. Note that we have imposed the condition that $c = d^{-3} \in (0, 2^{1/M} - 1]$ in (B.9), thus $d\geq (2^{1/M} - 1)^{-1/3}$, producing the lower bound in Assumption 1.5.
B.4. Proof of Lemma 5.8
Proof. Note that ${\mathbf{U}}$ is spanned by first $k$ singular vectors of $({\mathbf{A}}_1)_{\mathcal{I}_1}$. Let $\{\boldsymbol{u}_i\}_{i=1}^{k}$ be an orthonormal basis of the ${\mathbf{U}}$, then the projection $P_{{\mathbf{U}}} \,:\!=\, \sum _{l=1}^{k}\langle\boldsymbol{u}_{l},\cdot \, \rangle\boldsymbol{u}_{l}$. Let $k(i)$ index the membership of vertex $i$. For each fixed $i\in V_{k(i)} \cap Y_2\cap \{i_1, \cdots, i_s\}$,
As a consequence of independence between entries in ${\mathbf{A}}_1$ and entries in ${\mathbf{A}}_2$, defined in equation (18), it is known that $\{\boldsymbol{u}_l\}_{l=1}^{k}$ and $\boldsymbol{e}_i$ are independent of each other, since $\boldsymbol{e}_i$ are columns of ${\mathbf{E}}_2 \,:\!=\,{\mathbf{A}}_2 - \widetilde{{\mathbf{A}}}_2$. If the expectation is taken over $\{\boldsymbol{\mathcal{A}}^{(m)}\}_{m \in \mathcal{M}}$ conditioning on $\{\boldsymbol{u}_l\}_{l=1}^{k}$, then
where $\mathcal{M}$ is obtained from Algorithm 3. Expand each $\langle\boldsymbol{u}_{l},\boldsymbol{e}_i\rangle ^2$ and rewrite it into $2$ parts,
Part $(a)$ is the contribution from graph, i.e., $2$-uniform hypergraph, while part $(b)$ is the contribution from $m$-uniform hypergraph with $m\geq 3$, which only occurs in hypergraph clustering. The expectation of part $(a)$ in is upper bounded by $\alpha$ as defined in equation (8), since
where $\|\boldsymbol{u}_l\|_2^2 = \sum _{j=1}^{n}[\boldsymbol{u}_{l}(j)]^2 = 1$. For part $(b)$,
According to Definition 2.1 of the adjacency tensor, $\boldsymbol{\mathcal{A}}^{(m)}_{e_1}$ and $\boldsymbol{\mathcal{A}}^{(m)}_{e_2}$ are independent if hyperedge $e_1 \neq e_2$, then only the terms with hyperedge $e\supset \{i, j_1, j_2\}$ have non-zero contribution. Then the expectation of part $(b)$ can be rewritten as
Note that $|Y_2 \cup Z| \leq n$, then the number of possible hyperedges $e$, while $e\in E_m[Y_2\cup Z]$ and $e \supset \{i,j_1, j_2\}$, is at most $\binom{n}{m-3}$. Thus equation (B.13) is upper bounded by
where $\|\boldsymbol{u}_l\|_2 = 1$, $d = \sum _{m\in \mathcal{M}}(m-1)a_m$. With the upper bounds for part $(a)$ and $(b)$ in equation (B.12), the conditional expectation of $\|P_{{\mathbf{U}}}\boldsymbol{e}_i\|_2^2$ is bounded by
Let $X_i$ be the Bernoulli random variable defined by
By Markov’s inequality,
Let $\delta \,:\!=\, \frac{s}{2 \sum _{j=1}^{s}{\mathbb{E}} X_{i_j}} - 1$ where $s = 2k\log ^2(n)$. By Hoeffding Lemma D.2,
Therefore, with probability $1 - O(n^{-k\log (n)})$, at least $s/2$ of the vectors $\boldsymbol{e}_{i_1}, \dots,\boldsymbol{e}_{i_s}$ satisfy
Meanwhile, for any $c\in (0, 2)$, there exists some large enough constant $C_2 \geq 2^{\mathcal{M}_{\max } + 1}\sqrt{(\mathcal{M}_{\max } + 2)/k}/c$ such that if $\sum _{m \in \mathcal{M}}(m-1)(a_m -b_m) \gt C_2 k^{\mathcal{M}_{\max }-1}\sqrt{d}$, then
B.5. Proof of Lemma 5.10
Proof. Split $[n]$ into $V^{\prime}_1$ and $V^{\prime}_2$ such that $V^{\prime}_{1} = \{i|\boldsymbol{v}(i) \gt 0\}$ and $V^{\prime}_{2} = \{i|\boldsymbol{v}(i) \leq 0\}$. Without loss of generality, assume that the first $\frac{n}{k}$ entries of $\bar{\boldsymbol{v}}$ are positive. We can write $\boldsymbol{v}$ in terms of its orthogonal projection onto $\bar{\boldsymbol{v}}$ as
where $\bar{\boldsymbol{v}} \perp{\boldsymbol \varepsilon }$ with $\|{\boldsymbol \varepsilon }\|_2 \lt c$ and $c_{1} \geq \sqrt{1 - c^{2}}$. The number of entries of $\boldsymbol \varepsilon$ smaller than $-\frac{\sqrt{1 - c^{2}}}{\sqrt{n}}$ is at most $\frac{c^{2}}{1 - c^{2}}n$. Note that $c_{1} \geq \sqrt{1 - c^{2}}$, so at least $\frac{n}{k} - \frac{c^{2}}{1 - c^{2}}n$ indices $i$ with $\bar{\boldsymbol{v}}_i = \frac{1}{\sqrt{n}}$ will have $\boldsymbol{v}_i \gt 0$, thus the ratio we are seeking is at least
B.6. Proof of Lemma 5.11
Proof. We start with the following simple claim: for any $m \geq 2$ and any $\nu \in [1/2, 1)$,
Indeed, one quick way to see this is by induction on $m$; we will induct from $m$ to $m+2$. Assume the inequality is true for $m$; then
where we have used the induction hypothesis together with $1-2\nu \leq 0$ and $(1-\nu )^2\gt 0$. After easily checking that the inequality works for $m=2,3$, the induction is complete. We shall now check that the quantities defined in Lemma 5.11 obey the relationship $\mu _2 \geq \mu _1$ and $\mu _2 - \mu _1 = \Omega (n)$, for $n$ large enough. First, note that the only thing we need to check is that, for sufficiently large $n$,
in fact, we will show the stronger statement that for any $m \geq 2$ and $n$ large enough,
and this will suffice to see that the second part of the assertion, $\mu _2-\mu _1 = \Omega (n)$, is also true. Asymptotically, $\binom{\frac{\nu n}{2k}}{m} \sim \frac{\nu ^m}{m!} \left ( \frac{n}{2k} \right )^m$, $\binom{\frac{(1-\nu )n}{2k}}{m} \sim \frac{(1-\nu )^m}{m!} \left ( \frac{n}{2k} \right )^m$, and $\binom{\frac{(1+\nu )n}{4k}}{m} \sim \frac{\left ( \frac{1+\nu }{2} \right )^m}{m!} \left ( \frac{n}{2k} \right )^m$. Note then that equation (B.16) follows from equation (B.15).
Let {$V_{1}, \dots, V_{k}$} be the true partition of $V$. Recall that hyperedges in $H = \cup _{m\in \mathcal{M}}H_m$ are coloured red and blue with equal probability in Algorithm 2. Let $E_m(X)$ denote the set of blue $m$-uniform hyperedges with all vertices located in the vertex set $X$. Assume $|X \cap V_i| = \eta _i |X|$ with $\sum _{i=1}^{k}\eta _{i} = 1$. For each $m\in \mathcal{M}$, the presence of hyperedge $e\in E_{m}(X)$ can be represented by independent Bernoulli random variables
depending on whether $e$ is a hyperedge with all vertices in the same block. Denote by
the union of all $m$-uniform sets of hyperedges with all vertices in the same $V_i\cap X$ for some $i\in [k]$, and by
the set of $m$-uniform hyperedges with vertices across different blocks $V_i\cap X$. Then the cardinality $|E_m(X)|$ can be written as the
and by summing over $m$, the weighted cardinality $|E(X)|$ is written as
with its expectation
since
Next, we prove the two Statements in Lemma 5.11 separately. First, assume that $|X\cap V_i| \leq \nu |X|$ ( i.e., $\eta _i \leq \nu$) for each $i\in [k]$. Then
To justify the above inequality, note that since $\sum \limits _{i=1}^k \eta _i = 1$, the sum $\sum _{i=1}^k \binom{\eta _i \frac{n}{2k}}{m}$ is maximised when all but $2$ of the $\eta _i$ are $0$, and since all $\eta _i \leq \nu$, this means that
Note that $m(m-1)(T_{e}^{(a_m)} -{\mathbb{E}} T_{e}^{(a_m)})$ and $m(m-1)(T_{e}^{(b_m)} -{\mathbb{E}} T_{e}^{(b_m)})$ are independent mean-zero random variables bounded by $M(M-1)$ for all $m\in \mathcal{M}$, and ${\mathbb{V}\mathrm{ar}}(|E(X)|) \leq M^2(M - 1)^2{\mathbb{E}} |E(X)| = \Omega (n)$. Recall that $\mu _{\mathrm{T}}\,:\!=\, (\mu _1 + \mu _2)/2$. Define $t = \mu _{\mathrm{T}} -{\mathbb{E}}|E(X)|$, then $ 0 \lt (\mu _2 - \mu _1)/2 \leq t \leq \mu _{\mathrm{T}}$, hence $t = \Omega (n)$. By Bernstein’s Lemma D.3, we have
where $c\gt 0$ is some constant. On the other hand, if $|X\cap V_i| \geq \frac{1+\nu }{2}|X|$ for some $i\in [k]$, then
The above can be justified by noting that at least one $|X\cap V_i| \geq \frac{1+\nu }{2} |X|$, and that the rest of the vertices will yield a minimal binomial sum when they are evenly split between the remaining $V_j$. Similarly, define $t = \mu _{\mathrm{T}} -{\mathbb{E}}|E(X)|$, then $0 \lt (\mu _2 - \mu _1)/2 \leq -t = \Omega (n)$, and Bernstein’s Lemma D.3 gives
where $c^{\prime }\gt 0$ is some other constant.
B.7. Proof of Lemma 5.13
Proof. If vertex $i$ is uniformly chosen from $Y_2$, the probability that $i \notin V_{l}$ for some $l \in [k]$ is
where $n_t$ and $n_t^{\prime }$, defined in equations (12) and (13), denote the cardinality of $Z\cap V_t$ and $Y_1\cap V_t$ respectively. As proved in Appendix B.2, with probability at least $1 - 2\exp ({-}k \log ^2(n))$, we have
then ${\mathbb{P}}(i\notin V_{l} | i\in Y_2) = 1-\frac{1}{k} \Big (1 + o(1) \Big )\,.$ After $k\log ^2 n$ samples from $Y_2$, the probability that there exists at least one node which belongs to $V_l$ is at least
The proof is completed by a union bound over $l\in [k]$.
B.8. Proof of Lemma 5.14
Proof. We calculate ${\mathbb{P}}(S_{11}^{\prime }(u) \leq \mu _{\mathrm{C}})$ first. Define $t_{\mathrm{1C}} \,:\!=\, \mu _{\mathrm{C}} -{\mathbb{E}} S_{11}^{\prime }(u)$, then by Bernstein’s inequality (Lemma D.3) and taking $K = \mathcal{M}_{\max } - 1$,
where $\mathcal{M}$ is obtained from Algorithm 3 with $\mathcal{M}_{\max }$ denoting the maximum value in $\mathcal{M}$, and the last two inequalities hold since ${\mathbb{V}\mathrm{ar}}[S_{11}^{\prime }(u)] \leq (\mathcal{M}_{\max } - 1)^2{\mathbb{E}} S_{11}^{\prime }(u)$, and for sufficiently large $n$,
Similarly, for ${\mathbb{P}}(S_{1j}^{\prime }(u) \geq \mu _{\mathrm{C}})$, define $t_{\mathrm{jC}} \,:\!=\, \mu _{\mathrm{C}} -{\mathbb{E}} S_{1j}^{\prime }(u)$ for $j\neq 1$, by Bernstein’s Lemma D.3,
The last two inequalities holds since ${\mathbb{V}\mathrm{ar}}[S_{1j}^{\prime }(u)] \leq (\mathcal{M}_{\max } - 1)^2{\mathbb{E}} S_{1j}^{\prime }(u)$, and for sufficiently large $n$,
Appendix C. Algorithm correctness for the binary case
We will show the correctness of Algorithm 1 and prove Theorem 1.6 in this section. The analysis will mainly follow from the analysis in Section 5. We only detail the differences.
Without loss of generality, we assume $n$ is even to guarantee the existence of a binary partition of size $n/2$. The method to deal with the odd $n$ case was discussed in Lemma 2.4. Then, let the index set be $\mathcal{I} = \{i\in [n]\,:\, \mathrm{row}(i) \leq 20\mathcal{M}_{\max } d\}$, as shown in equation (11). Let $\boldsymbol{u}_i$ (resp. $\bar{\boldsymbol{u}}_i$) denote the eigenvector associated to $\lambda _i({\mathbf{A}}_{\mathcal{I}})$ (resp. $\lambda _i(\overline{{\mathbf{A}}})$) for $i =1, 2$. Define two linear subspaces ${\mathbf{U}}\,:\!=\, \mathrm{Span}\{\boldsymbol{u}_{1},\boldsymbol{u}_{2}\}$ and $\overline{{\mathbf{U}}}\,:\!=\, \mathrm{Span}\{\bar{\boldsymbol{u}}_{1},\bar{\boldsymbol{u}}_{2}\}$, then the angle between ${\mathbf{U}}$ and $\overline{{\mathbf{U}}}$ is defined as $ \sin \angle ({\mathbf{U}}, \overline{{\mathbf{U}}}) \,:\!=\, \|P_{{\mathbf{U}}} - P_{\overline{{\mathbf{U}}}}\|$, where $P_{{\mathbf{U}}}$ and $P_{\overline{{\mathbf{U}}}}$ are the orthogonal projections onto ${\mathbf{U}}$ and $\overline{{\mathbf{U}}}$, respectively.
C.1. Proof of Lemma 4.4
The strategy to bound the angle is similar to Subsection 5.1.2, except that we apply Davis-Kahan Theorem (Lemma D.6) here.
Define ${\mathbf{E}} \,:\!=\,{\mathbf{A}} - \overline{{\mathbf{A}}}$ and its restriction on $\mathcal{I}$, namely ${\mathbf{E}}_{\mathcal{I}} \,:\!=\, ({\mathbf{A}} - \overline{{\mathbf{A}}})_{\mathcal{I}} ={\mathbf{A}}_{\mathcal{I}} - \overline{{\mathbf{A}}}_{\mathcal{I}}$, as well as ${\boldsymbol \Delta }\,:\!=\, \overline{{\mathbf{A}}}_{\mathcal{I}} - \overline{{\mathbf{A}}}$. Then the deviation ${\mathbf{A}}_{\mathcal{I}} - \overline{{\mathbf{A}}}$ is decomposed as
Theorem 3.3 indicates $\|{\mathbf{E}}_{\mathcal{I}}\|\leq C_3\sqrt{d}$ with probability at least $1 - n^{-2}$ when taking $\tau =20\mathcal{M}_{\max }, K=3$, where $C_3$ is a constant depending only on $\mathcal{M}_{\max }$. Moreover, Lemma 5.4 shows that the number of vertices with high degrees is relatively small. Consequently, an argument similar to Corollary 5.5 leads to the conclusion $\|{\boldsymbol \Delta }\| \leq \sqrt{d}$ w.h.p. Together with upper bounds for $\|{\mathbf{E}}_{\mathcal{I}}\|$ and $\|{\boldsymbol \Delta }\|$, Lemma C.1 shows that the angle between ${\mathbf{U}}$ and $\overline{{\mathbf{U}}}$ is relatively small with high probability.
Lemma C.1. For any $c\in (0,1)$, there exists a constant $C_2$ depending on $\mathcal M_{\max }$ and $c$ such that if
then $\sin \angle ({\mathbf{U}}, \overline{{\mathbf{U}}}) \leq c$ with probability $1 - n^{-2}$.
Proof. First, with probability $1 - n^{-2}$, we have
According to the definitions in equation (8), $\alpha \geq \beta$ and $\alpha = O(1/n)$, $\beta = O(1/n)$. Meanwhile, Lemma 2.3 shows that $|\lambda _2(\overline{{\mathbf{A}}})| = [{-} \alpha + (\alpha - \beta )n/2]$ and $|\lambda _{3}(\overline{{\mathbf{A}}})| = \alpha$. The
Then for some large enough $C_2$, the following condition for Davis-Kahan Theorem (Lemma D.6) is satisfied
Then for any $c\in (0, 1)$, we can choose $C_2 = (C_3 + 1)/c$ such that
Now, we focus on the accuracy of Algorithm 7, once the conditions in Lemma C.1 are satisfied.
Lemma C.2 (Lemma 23 in [Reference Chin, Rao and Van19]). If $\sin \angle (\overline{{\mathbf{U}}},{\mathbf{U}}) \leq c \leq \frac{1}{4}$, there exists a unit vector $\boldsymbol{v}\in{\mathbf{U}}$ such that the angle between $\bar{\boldsymbol{u}}_{2}$ and $\boldsymbol{v}$ satisfies $\sin \angle (\bar{\boldsymbol{u}}_{2},\boldsymbol{v}) \leq 2 \sqrt{c}$.
The desired vector $\boldsymbol{v}$, as constructed in Algorithm 7, is the unit vector perpendicular to $P_{{\mathbf{U}}}\mathbf{1}_n$, where $P_{{\mathbf{U}}}\mathbf{1}_n$ is the projection of all-ones vector onto ${\mathbf{U}}$. Lemmas C.1 and C.2 together give the following corollary.
Corollary C.3. For any $c \in (0, 1)$, there exists a unit vector $\boldsymbol{v}\in{\mathbf{U}}$ such that the angle between $\bar{\boldsymbol{u}}_{2}$ and $\boldsymbol{v}$ satisfies $\sin \angle (\bar{\boldsymbol{u}}_{2},\boldsymbol{v}) \leq c \lt 1$ with probability $1 - O(e^{-n})$.
Proof. For any $c \in (0, 1)$, we could choose constants $C_2$, $C_3$ in Lemma C.1 such that $\sin \angle (\overline{{\mathbf{U}}},{\mathbf{U}}) \leq \frac{c^{2}}{4} \lt 1$. Then by Lemma C.2, we construct $\boldsymbol{v}$ such that $\sin \angle (\bar{\boldsymbol{u}}_{2},\boldsymbol{v}) \leq c$.
Lemma C.4 (Lemma 23 in [Reference Chin, Rao and Van19]). If $\sin \angle (\bar{\boldsymbol{u}}_{2},\boldsymbol{v}) \lt c \leq 0.5$, then we can identify at least $(1 - 8c^{2}/3)n$ vertices from each block correctly.
The proof of Lemma 4.4 is completed when choosing $C_2$, $C_3$ in Lemma C.1 s.t. $c \leq \frac{1}{4}$.
C.2 Proof of Lemma 4.5
The proof strategy is similar to Subsections 5.2 and 5.3. In Algorithm 1, we first colour the hyperedges with red and blue with equal probability. By running Algorithm 8 on the red graph, we obtain a $\nu$-correct partition $V_1^{\prime }, V_2^{\prime }$ of $V = V_1 \cup V_2$, i.e., $|V_{l}\cap V_{l}^{\prime }| \geq \nu n/2$ for $l = 1, 2$. In the rest of the proof, we condition on this event and the event that the maximum red degree of a vertex is at most $\log ^{2}(n)$ with probability at least $1 - o(1)$. This can be proved by Bernstein’s inequality (Lemma D.3).
Similarly, we consider the probability of a hyperedge $e = \{i_{1}, \cdots, i_{m}\}$ being blue conditioning on the event that $e$ is not a red hyperedge in each underlying $m$-uniform hypergraph separately. If vertices $i_{1}, \cdots, i_{m}$ are all from the same true cluster, then the probability is $\psi _m$, otherwise $\phi _m$, where $\psi _m$ and $\phi _m$ are defined in equations (29) and (30), and the presence of those hyperedges are represented by random variables $\zeta _e^{(a_m)} \sim \mathrm{Bernoulli}\left (\psi _m\right )$, $\xi _e^{(b_m)} \sim \mathrm{Bernoulli}\left (\phi _m\right )$, respectively.
Following a similar argument in Subsection 5.2, the row sum of $u$ can be written as
where $\mathcal{E}^{(a_m)}_{l, j}\,:\!=\, E_m([V_l]^{1}, [V_l\cap V^{\prime }_j]^{m-1})$ denotes the set of $m$-hyperedges with $1$ vertex from $[V_l]^{1}$ and the other $m-1$ vertices from $[V_l\cap V^{\prime }_j]^{m-1}$, while $ \mathcal{E}^{(b_m)}_{l, j} \,:\!=\, E_m \Big ([V_l]^{1}, \,\, [V^{\prime }_j]^{m-1} \setminus [V_l\cap V^{\prime }_j]^{m-1} \Big )$ denotes the set of $m$-hyperedges with $1$ vertex in $[V_l]^{1}$ while the remaining $m-1$ vertices in $[V^{\prime }_j]^{m-1}\setminus [V_l\cap V^{\prime }_j]^{m-1}$, with their cardinalities
According to the fact $|V_{l}\cap V_{l}^{\prime }| \geq \nu n/2$, $|V_l| = n/2$, $|V^{\prime }_l| = n/2$ for $l = 1, 2$, we have
To simplify the calculation, we take the lower and upper bound of $|\mathcal{E}^{(a_m)}_{l, l}|$ and $|\mathcal{E}^{(a_m)}_{l, j}|(j\neq l)$ respectively. Taking expectation with respect to $\zeta _e^{(a_m)}$ and $\xi _e^{(b_m)}$, for any $u \in V_l$, we have
By assumptions in Theorem 1.7, ${\mathbb{E}} S^{\prime }_{ll}(u) -{\mathbb{E}} S^{\prime }_{lj}(u) = \Omega (1)$. We define
After Algorithm 6, if a vertex $u\in V_l$ is mislabelled, one of the following events must happen
• $S^{\prime }_{ll}(u) \leq \mu _{\mathrm{C}}$,
• $S^{\prime }_{lj}(u) \geq \mu _{\mathrm{C}}$, for some $j\neq l$.
By an argument similar to Lemma 5.14, we can prove that
where $\rho = \exp\!\left ({-}C_{\mathcal M}(2)\cdot \mathrm{SNR}_{\mathcal{M}}(2) \right )$ and
As a result, the probability that either of those events happened is bounded by $\rho$. The number of mislabelled vertices in $V_1$ after Algorithm 5 is at most
where $\Gamma _{i}$ (resp. $\Lambda _{i}$) are i.i.d indicator random variables with mean $\rho _1^{\prime }$ (resp. $\rho _2^{\prime }$). Then
where $\nu$ is the correctness after Algorithm 4. Let $t_l\,:\!=\, (1 + \nu/2)n\rho$, then by Chernoff Lemma D.1,
which means that with probability $1- O(e^{- n\rho })$, the fraction of mislabelled vertices in $V_l$ is smaller than $2\rho$, i.e., the correctness of $V_l$ is at least $\gamma \,:\!=\, \max \{\nu, 1 - 2\rho \}$.
Appendix D. Useful Lemmas
Lemma D.1 (Chernoff’s inequality, Theorem 2.3.6 in [Reference Vershynin69]). Let $X_i$ be independent Bernoulli random variables with parameters $p_i$. Consider their sum $S_N = \sum _{i=1}^{N}X_i$ and denote its mean by $\mu ={\mathbb{E}} S_N$. Then for any $\delta \in (0, 1]$,
Lemma D.2 (Hoeffding’s inequality, Theorem 2.2.6 in [Reference Vershynin69]). Let $X_1,\dots, X_N$ be independent random variables with $X_i \in [a_i, b_i]$ for each $i\in \{1, \dots, N\}$. Then for any, $t \geq 0$, we have
Lemma D.3 (Bernstein’s inequality, Theorem 2.8.4 in [Reference Vershynin69]). Let $X_1,\dots,X_N$ be independent mean-zero random variables such that $|X_i|\leq K$ for all $i$. Let $\sigma ^2 = \sum _{i=1}^{N}{\mathbb{E}} X_i^2$. Then for every $t \geq 0$, we have
Lemma D.4 (Bennett’s inequality, Theorem 2.9.2 in [Reference Vershynin69]). Let $X_1,\dots, X_N$ be independent random variables. Assume that $|X_i -{\mathbb{E}} X_i| \leq K$ almost surely for every $i$. Then for any $t\gt 0$, we have
where $\sigma ^2 = \sum _{i=1}^{N}\mathrm{Var}(X_i)$, and $h(u) \,:\!=\, (1 + u)\log (1 + u) - u$.
Lemma D.5 (Weyl’s inequality). Let ${\mathbf{A}},{\mathbf{E}} \in{\mathbb{R}}^{m \times n}$ be two real $m\times n$ matrices, then $|\sigma _i({\mathbf{A}} +{\mathbf{E}}) - \sigma _i({\mathbf{A}})| \leq \|{\mathbf{E}}\|$ for every $1 \leq i \leq \min \{ m, n\}$. Furthermore, if $m = n$ and ${\mathbf{A}},{\mathbf{E}} \in{\mathbb{R}}^{n \times n}$ are real symmetric, then $|\lambda _i({\mathbf{A}} +{\mathbf{E}}) - \lambda _i({\mathbf{A}})| \leq \|{\mathbf{E}}\|$ for all $1 \leq i \leq n$.
Lemma D.6 (Davis-Kahan’s $\sin{\boldsymbol \Theta }$ Theorem, Theorem 2.2.1 in [Reference Chen, Chi, Fan and Ma15]). Let $\overline{{\mathbf{M}}}$ and ${\mathbf{M}} = \overline{{\mathbf{M}}} +{\mathbf{E}}$ be two real symmetric $n \times n$ matrices, $n \times n$ matrices, with eigenvalue decompositions given respectively by
Here, $\{ \overline{\lambda }_{i}\}_{i=1}^{n}$(resp. $\{\lambda _{i}\}_{i=1}^{n}$) stand for the eigenvalues of $\overline{{\mathbf{M}}}$(resp. ${\mathbf{M}}$), and $\overline{\boldsymbol{u}}_i$(resp. $\boldsymbol{u}_i$) denotes the eigenvector associated $\overline{\lambda }_{i}$(resp. $\lambda _i$). Additionally, for some fixed integer $r\in [n]$, we denote
The matrices $\boldsymbol \Lambda$, ${\boldsymbol \Lambda }_{\perp }$, ${\mathbf{U}}$, ${\mathbf{U}}_{\perp }$ are defined analogously. Assume that
and the projection matrices are given by $P_{\mathbf{U}} \,:\!=\,{\mathbf{U}}{\mathbf{U}}^{{\mathsf T}}$, $P_{\overline{{\mathbf{U}}}} \,:\!=\, \overline{{\mathbf{U}}} \,\overline{{\mathbf{U}}}^{{\mathsf T}}$, then one has $\|P_{\mathbf{U}} - P_{\overline{{\mathbf{U}}}}\| \leq (2\|{\mathbf{E}}\|/\Delta )$. In particular, suppose that $|\overline{\lambda }_1| \geq |\overline{\lambda }_2| \geq \cdots \geq |\overline{\lambda }_r| \geq |\overline{\lambda }_{r+1}| \geq \cdots |\overline{\lambda }_{n}|$ (resp. $|\lambda _1| \geq \cdots \geq |\lambda _{n}|$). If $\|{\mathbf{E}}\|\leq (1 - 1/\sqrt{2})(|\overline{\lambda }|_{r} - |\overline{\lambda }|_{r+1})$, then one has
Lemma D.7 (Wedin’s $\sin{\boldsymbol \Theta }$ Theorem, Theorem 2.3.1 in [Reference Chen, Chi, Fan and Ma15]). Let $\overline{{\mathbf{M}}}$ and ${\mathbf{M}} = \overline{{\mathbf{M}}} +{\mathbf{E}}$ be two $n_1 \times n_2$ real matrices and $n_2$ real matrices and $n_1\geq n_2$, with SVDs given respectively by
Here, $\overline{\sigma }_1 \geq \dots \geq \overline{\sigma }_{n_1}$ (resp. $\sigma _1 \geq \dots \geq \sigma _{n_1}$) stand for the singular values of $\overline{{\mathbf{M}}}$(resp. ${\mathbf{M}}$), $\overline{\boldsymbol{u}}_i$(resp. $\boldsymbol{u}_i$) denotes the left singular vector associated with the singular value $\overline{\sigma }_i$(resp. $\sigma _i$), and $\overline{\boldsymbol{v}}_i$(resp. $\boldsymbol{v}_i$) denotes the right singular vector associated with the singular value $\overline{\sigma }_i$(resp. $\sigma _i$). In addition, for any fixed integer $r\in [n]$, we denote
The matrices $\overline{{\boldsymbol \Sigma }}$, $\overline{{\boldsymbol \Sigma }}_{\perp }$, $\overline{{\mathbf{V}}}$, $\overline{{\mathbf{V}}}_{\perp }$ are defined analogously. If ${\mathbf{E}} ={\mathbf{M}} -{\mathbf{M}} - \overline{{\mathbf{M}}}$ satisfies $\|{\mathbf{E}}\|\leq \overline{\sigma }_r$ - with the projection matrices $P_{\mathbf{U}} \,:\!=\,{\mathbf{U}}{\mathbf{U}}^{{\mathsf T}}$, one has
In particular, if $\|{\mathbf{E}}\|\leq (1 - 1/\sqrt{2})(\overline{\sigma }_{r} - \overline{\sigma }_{r+1})$, then one has