Accelerated Flow for Probability distributions

  • 2019-01-10 18:42:38
  • Amirhossein Taghvaei, Prashant G. Mehta
  • 3


This paper presents a methodology and numerical algorithms for constructingaccelerated gradient flows on the space of probability distributions. Inparticular, we extend the recent variational formulation of acceleratedgradient methods in (wibisono, et. al. 2016) from vector valued variables toprobability distributions. The variational problem is modeled as a mean-fieldoptimal control problem. The maximum principle of optimal control theory isused to derive Hamilton's equations for the optimal gradient flow. TheHamilton's equation are shown to achieve the accelerated form of densitytransport from any initial probability distribution to a target probabilitydistribution. A quantitative estimate on the asymptotic convergence rate isprovided based on a Lyapunov function construction, when the objectivefunctional is displacement convex. Two numerical approximations are presentedto implement the Hamilton's equations as a system of $N$ interacting particles.The continuous limit of the Nesterov's algorithm is shown to be a special casewith $N=1$. The algorithm is illustrated with numerical examples.


Introduction (beta)



Conclusion (beta)