Adaptive posterior contraction rates for empirical Bayesian drift estimation of a diffusion
share
Gaussian process (GP) priors are attractive for estimating the drift of an SDE with continuous time observations, because of their conjugate posteriors. However, their performance strongly depends on the choice of the hyper-parameters. In this article we employ the marginal maximum likelihood estimator to estimate the scaling and / or smoothness parameter(s) of the prior and we show that the corresponding posterior has optimal rates of convergence. General theorems do not apply directly to this model, as the usual test functions are with respect to a random Hellinger-type metric. We allow for continuous and discrete, one- and two-dimensional sets of hyper-parameters, where optimising over the two-dimensional set of smoothness and scaling hyper-parameters is shown to be beneficial in terms of the adaptivity range.
READ FULL TEXT