Dynamics and geometry from high dimensional data

Markov state models for data assimilation and interpretation

Eric Vanden-Eijnden

New York University

Markov State Models (MSMs) have emerged has a popular way to interpret and analyze time-series data generated e.g. in the context of molecular dynamics simulations. The basic idea of these models is to represent the dynamics of the system as memoryless jumps between predefined sets in the systems state space, i.e. as a Markov jump process (MJP). Performing this mapping typically involves two nontrivial questions: first how to define these sets and second how to learn the rate matrix of the MJP once the sets have been identified? Both these questions will be discussed in the talk, with emphasis put on the problem of model specification error. Some of the outputs of MSMs, in particular in terms of free energy, reaction rate, etc. will also be discussed.