Exact Algorithms and Lower Bounds for Stable Instances of Euclidean k-Means

We investigate the complexity of solving stable or perturbation-resilient instances of k-Means and k-Median clustering in fixed dimension Euclidean metrics (more generally doubling metrics). The notion of stable (perturbation resilient) instances was introduced by Bilu and Linial [2010] and Awasthi et al. [2012]. In our context we say a k-Means instance is α-stable if there is a unique OPT which remains optimum if distances are (non-uniformly) stretched by a factor of at most α. Stable clustering instances have been studied to explain why heuristics such as Lloyd's algorithm perform well in practice. In this work we show that for any fixed ϵ>0, (1+ϵ)-stable instances of k-Means in doubling metrics can be solved in polynomial time. More precisely we show a natural multiswap local search algorithm finds OPT for (1+ϵ)-stable instances of k-Means and k-Median in a polynomial number of iterations. We complement this result by showing that under a new PCP theorem, this is essentially tight: that when the dimension d is part of the input, there is a fixed ϵ_0>0 s.t. there is not even a PTAS for (1+ϵ_0)-stable k-Means in R^d unless NP=RP. To do this, we consider a robust property of CSPs; call an instance stable if there is a unique optimum solution x^* and for any other solution x', the number of unsatisfied clauses is proportional to the Hamming distance between x^* and x'. Dinur et al. have already shown stable QSAT is hard to approximate for some constant Q, our hypothesis is simply that stable QSAT with bounded variable occurrence is also hard. Given this hypothesis we consider "stability-preserving" reductions to prove our hardness for stable k-Means. Such reductions seem to be more fragile than standard L-reductions and may be of further use to demonstrate other stable optimization problems are hard.