• Start Date: November 7, 2018
  • Event Start Time: 11:00 AM
  • Event End Time: 12:00 PM
  • Seminar Series: Theoretical Computer Science Seminar
  • Presenter(s): Cole Franks - Rutgers University
  • Event Location: Conference Room 301 | Rutgers University | CoRE Building | 96 Frelinghuysen Road
  • Presentation Type: Stand Alone Presentation
  • Abstract:

    In many cases, a group action on a manifold can be associated with a map 
    called a moment map. Surprisingly, the image of the manifold under this 
    map is always a convex polytope, known as its moment polytope. Moment 
    polytope membership can often be expressed as an optimization problem 
    which is convex only in the Abelian case. Whether moment polytope 
    membership can be decided in polynomial time is a fundamental open 
    problem with deep consequences in algebraic complexity theory and 
    elsewhere.

    I'll discuss an approach to moment polytope membership through 
    optimization. For example, in a recent joint work we obtain an extremely 
    simple algorithm for tensor scaling, i.e. approximately scaling tensors 
    to arbitrary prescribed marginals (whenever possible). Improving the 
    run-time of our tensor scaling algorithm to polylog(eps) would result 
    in a polynomial time algorithm for moment polytope membership in this 
    special case, and would have applications in polynomial identity 
    testing, computing Brascamp-Lieb constants, and optimization problems 
    related to tensor rank, to name a few. If time permits I will discuss a 
    geodesically convex optimization algorithm that works in a much more 
    general setting.

     

    Joint work with with Peter Bürgisser, Ankit Garg, Rafael Oliveira, Michael Walter, 
    and Avi Wigderson