Density estimation for a class of continuous time processes (Q1368861)

A strictly stationary stochastic process \((X_t\), \(t\in\mathbb{R}^+)\) with a marginal density \(f\) is considered. A nonparametric linear wavelet estimator \(\widehat{f}_T\) based on the observed sample path \((X_t\), \(0\leq t\leq T)\) is constructed. Suppose that \(f\) has exponential tails and lies in a ball of a Besov space \(B_{p,q}^s\) (note that Sobolev spaces \(W_2^s\) and Hölder spaces \(C^s\) are particular cases of Besov spaces). The process \((X_t)\) is strongly mixing and satisfies the following ``local'' assumption: \[ \forall x\in\mathbb{R}, \quad \sup_{y\in\mathbb{R}} \int_0^\infty|f_t(x,y)- f(x)f(y)|dt\leq g(x), \tag{1} \] where \(f_t\), \(t>0\), is the joint density of the law of \((X_0,X_t)\) and \(g\) is a positive, integrable and bounded function. For the estimator \(\widehat{f}_T\) the risk associated with the loss function defined by the squared \(L_{p'}\)-norm on the whole line is studied, where \(p'\) is either integer or infinite, and \(p'\geq\max (2,p)\). It is shown that the risk converges to zero at the rate \(1/T\) if \(p'<\infty\), and \(\log T/T\) if \(p'=\infty\). This would be a standard result in parametric estimation but it is somewhat surprising in a nonparametric setup. Condition (1) controls the explosion of the joint density \(f_t\) when \(t\) goes to zero. It follows the idea of \textit{J. V. Castellana} and \textit{M. R. Leadbetter} [Stochastic Processes Appl. 21, 179-193 (1986; Zbl 0588.62156)] where the kernel estimators of the marginal density had the quadratic error of the rate \(1/T\). It is proved that the assumptions on \((X_t)\) hold for a wide class of diffusion process. The definitions of Besov spaces and wavelets are given in the Appendix.