Coresets have become an invaluable tool for solving $k$-means and kernel $k$-means clustering problems on large datasets with small numbers of clusters. On the other hand, spectral clustering works well on sparse graphs and has recently been extended to scale efficiently to large numbers of clusters. We exploit the connection between kernel $k$-means and the normalised cut problem to combine the benefits of both. Our main result is a coreset spectral clustering algorithm for graphs that clusters a coreset graph to infer a good labelling of the original graph. We prove that an $\alpha$-approximation for the normalised cut problem on the coreset graph is an $O(\alpha)$-approximation on the original. We also improve the running time of the state-of-the-art coreset algorithm for kernel $k$-means on sparse kernels, from $\tilde{O}(nk)$ to $\tilde{O}(n\cdot \min (k, d_{avg}))$, where $d_{avg}$ is the average number of non-zero entries in each row of the $n\times n$ kernel matrix. Our experiments confirm our coreset algorithm is asymptotically faster on large real-world graphs with many clusters, and show that our clustering algorithm overcomes the main challenge faced by coreset kernel $k$-means on sparse kernels which is getting stuck in local optima.

We study the following $\mathsf{KS}_2(c)$ problem: let $c \in\mathbb{R}^+$ be some constant, and $v_1,\ldots, v_m\in\mathbb{R}^d$ be vectors such that $\|v_i\|^2\leq α$ for any $i\in[m]$ and $\sum_{i=1}^m \langle v_i, x\rangle^2 =1$ for any $x\in\mathbb{R}^d$ with $\|x\|=1$. The $\mathsf{KS}_2(c)$ problem asks to find some $S\subset [m]$, such that it holds for all $x \in \mathbb{R}^d$ with $\|x\| = 1$ that \[ \left|\sum_{i \in S} \langle v_i, x\rangle^2 - \frac{1}{2}\right| \leq c\cdot\sqrtα,\] or report no if such $S$ doesn't exist. Based on the work of Marcus et al. and Weaver, the $\mathsf{KS}_2(c)$ problem can be seen as the algorithmic Kadison-Singer problem with parameter $c\in\mathbb{R}^+$. Our first result is a randomised algorithm with one-sided error for the $\mathsf{KS}_2(c)$ problem such that (1) our algorithm finds a valid set $S \subset [m]$ with probability at least $1-2/d$, if such $S$ exists, or (2) reports no with probability $1$, if no valid sets exist. The algorithm has running time \[ O\left(\binom{m}{n}\cdot \mathrm{poly}(m, d)\right)~\mbox{ for }~n = O\left(\frac{d}{ε^2} \log(d) \log\left(\frac{1}{c\sqrtα}\right)\right), \] where $ε$ is a parameter which controls the error of the algorithm. This presents the first algorithm for the Kadison-Singer problem whose running time is quasi-polynomial in $m$, although having exponential dependency on $d$. Moreover, it shows that the algorithmic Kadison-Singer problem is easier to solve in low dimensions. Our second result is on the computational complexity of the $\mathsf{KS}_2(c)$ problem. We show that the $\mathsf{KS}_2(1/(4\sqrt{2}))$ problem is $\mathsf{FNP}$-hard for general values of $d$, and solving the $\mathsf{KS}_2(1/(4\sqrt{2}))$ problem is as hard as solving the $\mathsf{NAE\mbox{-}3SAT}$ problem.