Nonparametric Bayesian posterior contraction rates for discretely observed scalar diffusions

Abstract

We consider nonparametric Bayesian inference in a reflected diffusion model dXt = b(Xt)dt+σ(Xt)dWt, with discretely sampled observations X0, X∆, . . . , Xn∆. We analyse the nonlinear inverse problem corresponding to the 'low frequency sampling' regime where ∆ > 0 is fixed and n → ∞. A general theorem is proved that gives conditions for prior distributions Π on the diffusion coefficient σ and the drift function b that ensure minimax optimal contraction rates of the posterior distribution over Hölder-Sobolev smoothness classes. These conditions are verified for natural examples of nonparametric random wavelet series priors. For the proofs we derive new concentration inequalities for empirical processes arising from discretely observed diffusions that are of independent interest. Primary 62G05; secondary 60J60 62F15 62G20

Nonparametric Bayesian Volatility Estimation

“…Literature on nonparametric Bayesian volatility estimation in SDE models is scarce. We can list theoretical contributions (Gugushvili and Spreij, 2014a), (Gugushvili and Spreij, 2016), (Nickl and Söhl, 2017), and the practically oriented paper (Batz et al, 2017feb). The model in the former two papers is close to the one considered in the present work, but from the methodological point of view different Bayesian priors are used and practical usefulness of the corresponding Bayesian approaches is limited.…”

Nonparametric Bayesian inference for Gamma-type Lévy subordinators

“…As we already remarked elsewhere, the parameter β plays a role similar to the dispersion coefficient σ of a stochastic differential equation driven by a Wiener process. Derivation of nonparametric Bayesian asymptotics for the latter class of processes (all of which is a recent work) historically proceeded from the assumption of a known σ to the one where σ is unknown and has to be estimated; see van der Meulen and van Zanten (2013), Gugushvili and Spreij (2014) and Nickl and Söhl (2017b). In that sense the fact that at this stage we assume β is known does not appear unexpected or unnatural.…”

On frequentist coverage errors of Bayesian credible sets in moderately high dimensions

“…where R 1 and R 2 are random variables such that there exist positive constants c 1 and c 2 depending only on q appearing in Condition 2.5 (a) for which we have where the first two inequalities follows from Lemma C.2 and the last two inequalities follows from Lemma C.3. From inequalities (43) and (45) and from Lemma C.5, we have…”

“…Note that [13] also discusses non-Gaussian error distributions. See also [8,12,26,28,42,43,48] for related results. We refer the reader to [3,18,24,38] on the BvM theorem for quasi-posterior distributions.…”