Georgia Institute of Technology College of Sciences

There has been substantial work on approximation algorithms for clustering data under distance-based objective functions such as k-median, k-means, and min-sum objectives. This work is fueled in part by the hope that approximating these objectives well will indeed yield more accurate solutions. That is, for problems such as clustering proteins by function, or clustering images by subject, there is some unknown correct "target" clustering and the implicit assumption is that clusterings that are approximately optimal in terms of these distance-based measures are also approximately correct in terms of error with respect to the target. In this work we show that if we make this implicit assumption explicit -- that is, if we assume that any c-approximation to the given clustering objective Phi is epsilon-close to the target -- then we can produce clusterings that are O(epsilon)-close to the target, even for values c for which obtaining a c-approximation is NP-hard. In particular, for the k-median, k-means, and min-sum objectives, we show that we can achieve this guarantee for any constant c > 1. Our results show how by explicitly considering the alignment between the objective function used and the true underlying clustering goals, one can bypass computational barriers and perform as if these objectives were computationally substantially easier. This talk is based on joint work with Avrim Blum and Anupam Gupta (SODA 2009), Mark Braverman (COLT 2009), and Heiko Roeglin and Shang-Hua Teng (ALT 2009).

It is known that, relative to any fixed vertex q of a finite graph, there exists a unique q-reduced divisor (G-Parking function based at q) in each linear equivalence class of divisors. In this talk, I will give an efficient algorithm for finding such reduced divisors. Using this, I will give an explicit and efficient bijection between the Jacobian group and the set of spanning trees of the graph. Then I will outline some applications of the main results, including a new approach to the Random Spanning Tree problem, efficient computation of the group law in the critical and sandpile group, efficient algorithm for the chip-firing game of Baker and Norine, and the relation to the Riemann-Roch theory on finite graphs.

We consider the #P complete problem of counting the number of independent sets in a given graph. Our interest is in understanding the effectiveness of the popular Belief Propagation (BP) heuristic. BP is a simple and iterative algorithm that is known to have at least one fixed point. Each fixed point corresponds to a stationary point of the Bethe free energy (introduced by Yedidia, Freeman and Weiss (2004) in recognition of Hans Bethe's earlier work (1935)). The evaluation of the Bethe Free Energy at such a stationary point (or BP fixed point) leads to the Bethe approximation to the number of independent sets of the given graph. In general BP is not known to converge nor is an efficient, convergent procedure for finding stationary points of the Bethe free energy known. Further, effectiveness of Bethe approximation is not well understood. As the first result of this paper, we propose a BP-like algorithm that always converges to a BP fixed point for any graph. Further, it finds an \epsilon approximate fixed point in poly(n, 2^d, 1/\epsilon) iterations for a graph of n nodes with max-degree d. As the next step, we study the quality of this approximation. Using the recently developed 'loop series' approach by Chertkov and Chernyak, we establish that for any graph of n nodes with max-degree d and girth larger than 8d log n, the multiplicative error decays as 1 + O(n^-\gamma) for some \gamma > 0. This provides a deterministic counting algorithm that leads to strictly different results compared to a recent result of Weitz (2006). Finally as a consequence of our results, we prove that the Bethe approximation is exceedingly good for a random 3-regular graph conditioned on the Shortest Cycle Cover Conjecture of Alon and Tarsi (1985) being true. (Joint work with Venkat Chandrasekaran, Michael Chertkov, David Gamarnik and Devavrat Shah)