Computational Statistics and Molecular Simulation: A Practical Cross-Fertilization (18w5023)
Christian Robert (Universite Paris Dauphine)
Luke Bornn (Simon Fraser University and Harvard)
Gabriel Stoltz (Ecole des Ponts)
In fact, we want to emphasize practical cross-fertilization that is not only based on attending lectures from researchers from other communities, but also operated through an active participation in "live'' hands-on sessions, where a researcher from one domain presents a numerical test case of interest to everyone, runs simulations and comments on their results, and emphasizes the actual challenges at hand in this problem. This presentation will be followed, and in fact hopefully interrupted, by all sorts of questions, comments, remarks from more theoretically oriented researchers, or more numerically active researchers. Let us list a few topics of interest, originally motivated by one scientific field but applicable to the other one. A transfert from molecular simulation to computational statistics is marked as [MD-$>$CS], the reverse transfert being denoted by [CS-$>$MD]:
[MD-$>$CS] methods for accelerated molecular dynamics, such as the ones developed by Arthur Voter in Los Alamos, which aim at sampling correct dynamics for metastable systems; [MD-$>$CS] automatic importance sampling techniques, based on the computation of the free energy in molecular simulation, and whose adaptation to cases relevent in computational statistics still requires further work; [CS-$>$MD] massively parallel MCMC, where the size of the data or the complexity of the statistical likelihood prohibits algorithms that operate on the whole of it at once, requiring techniques that split and merge partial calculations in the most efficient manner. nstances are pre-fetching, collaborative and median mixing, all proposed in the recent past. The design of numerical methods based on a sparsification of the data would be most welcome; [CS-$>$MD] intractable constants in both statistics and physics, where parts of the likelihood are results of intractable integrals, with an ever increasing array of solutions tackling special cases without a global, all-encompassing perspective; [MD-$>$CS] the design of variance reduction techniques for non-reversible dynamics, for which the invariant measure may in fact be unknown.
In order to strengthen the interactions generated by the live numerical sessions, and in particular to allow large computations during the workshop week and the weeks immediately after the workshop, we plan to grant access to the computing cluster of CERMICS (at Ecole des Ponts, France) for the participants expressing their interest in that option. If local computing alternatives are possible, we would obviously appreciate getting access to those.
An effort will also be made to promote the presence of younger researchers at this workshop. It is indeed paramount that young researchers in our research areas are exposed as early as possible to researchers from other research areas in order to start looking for possible numerical methods and analysis outside their immediate scientific surrounding. We believe that changing the habits of the younger researchers is one of the most performant ways of cross-fertilizing fields. While few names are included in the list below, for the obvious reason that the youngest researchers of 2018 are not yet at a position we can detect them, we believe that the various doctoral schools where the more mature researchers are involved will be very open to our request of involving their youngest members at the time.