Efficient Bayesian estimation of a non-Markovian Langevin model driven by correlated noise

Abstract: Data-driven modeling of non-Markovian dynamics is a recent topic of research with applications in many fields such as climate research, molecular dynamics, biophysics, or wind power modeling. In the frequently used standard Langevin equation, memory effects can be implemented through an additional hidden component which functions as correlated noise, thus resulting in a non-Markovian model. It can be seen as part of the model class of partially observed diffusions which are usually adapted to observed data via Bayesian estimation, whereby the difficulty of the unknown noise values is solved through a Gibbs sampler. However, when regarding large data sets with a length of $10^6$ or $10^7$ data points, sampling the distribution of the same amount of latent variables is unfeasible. For the model discussed in this work, we solve this issue through a direct derivation of the posterior distribution of the Euler-Maruyama approximation of the model via analytical marginalization of the latent variables. Yet, in the case of a nonlinear noise process, the inverse problem of model estimation proves to be ill-posed and still numerically expensive. We handle these complications by restricting the noise to an Ornstein-Uhlenbeck process, which considerably reduces the ambiguity of the estimation. Further, in this case, the estimation can be performed very efficiently if the drift and diffusion functions of the observed component are approximated in a piecewise constant manner. We illustrate the resulting procedure of efficient Bayesian estimation of the considered non-Markovian Langevin model by an example from turbulence.