A dynamical system comprises a collection of states, and a transition law transforms states from one to another. Almost every natural dynamic phenomenon, engineering systems as well as computer algorithms can be presented as a dynamical system. When a dynamical system is measured it generates data.
The art of data-driven discovery of dynamics is to uncover the underlying dynamics from a stream of data. One can try to simply approximate the transition law from data, but as shown in [3] this is inadequate for preserving important features of the dynamics such as asymptotic behavior and topology of invariant sets. An alternative was proposed in [1] where it was proposed to add a small randomness to the dynamial law. The result is a Markov transition process with low noise, close approximation of the invariant set as well as the dynamics law.
The technique is based on partitining the phase space. It was pointed out in [1] that if the partitioning is chosen to be a Rokhlin tower [4, 5] then the Markov approximation actually recreates long orbits of the deterministic dynamics.
The leftmost panel shows dataset generated by the Lorenz 63 attractor, and the middle panel shows the orbits created by a low-noise Markov process created from that data. The autocorrelations of the timeseries generated by the two datasets show a very close match.
Future directions.
1) Implementing codes in Julia language.
2) Finding efficient ways to partition a chaotic attractor
3) Devising a numerical method for constructing Rokhlin towers.
4) Creating a numerical procedure for converting the Markov approximation into a switched flow [2].
5) Finding sufficient conditions for stochastic stability.
Prerequisites.
1) Programming : familiarity with Matlab / Julia; familiarity with programming techniques in Linear Algebra
2) Linear Algebra : matrix theory, SVD, eigen-decomposition, least squares methods
3) Analysis : continuity and differntiability, limits and convergence
4) Functional analysis : Banach spaces, Hilbert spaces, kernel integrals
5) Measure theory : probability measures, absolute continuity, Markov transitions, Borel sets, abstract Measure theory
Sources
[1] Reconstructing dynamical systems as zero-noise limits - S Das (original paper presenting the Markov approximation technique)
[2] Discrete-time dynamics, step-skew products, and pipe-flows - S Das (equivalence of step-skew products, discrete-time dynamics and piper flows)
[3] Learning Theory for Dynamical Systems - T Berry, S Das (common generalization of ESNs and delay-coordinates)
[4] Multiple Rokhlin tower theorem - a simple proof - S J Eigen, V S Prasad (foundational paper 1 on Rohli towers)
[5] Coding a stationary process to one with prescribed marginals - S Alpern, V S Prasad (foundational paper 2 on Rohli towers)