University of California, Berkeley
This project contains three loosely related themes: computational applications of Markov chains, approximate counting, and algorithmic apects of finite metric spaces.
During the past decade, the so-called "Markov chain Monte Carlo method" has emerged as a powerful algorithmic paradigm, with applications in such areas as approximate counting, statistical physics and combinatorial optimization. The PI is pushing ahead with these applications, in the following three directions:
Development of new techniques based on multicommodity flow to analyze geometric Markov chains (where the state space is the set of integer points inside a convex body).
Investigation of techniques such as Chebyshev acceleration, which can potentially modify a Markov chain so as to substantially speed up its convergence.
Improved understanding of newly emerging connections between the rapid mixing property of certain Markov chains on the configurations of physical systems and the thermodynamic properties of those systems.
In parallel with recent activity aimed at classifying NP-hard optimization problems according to their degree of approximability, the PI is contributing to a similar classification of #P-hard counting problems. Specific aims here include:
Investigation of novel algorithms for the central problem of approximating the permanent.
Development of approximation-preserving reductions between counting problems so that, for example, a class of problems that are equivalent to the permanent can be constructed.
Finite metric spaces provide a clean and elegant paradigm that captures (and often streamlines) several previously ad hoc algorithmic ideas. The third part of the project involves a systematic study of finite metrics, and specifically:
The design of new techniques for l1 embedding, both in the general case (to achieve a distortion within o(log n) of optimal) and for special metrics (such as those supported on planar graphs). This has algorithmic applications to, among others, the central problem of approximating the sparsest cut in a network.
Development of improved techniques for embeddings into low-dimensional spaces. These have applications to algorithms for ubiquitous problems such as nearest-neighbor searching and clustering.
Approximation of complex metrics by simpler ones. This would allow existing algorithms that work well for simple metrics (such as trees) to be applied more widely.
