Donsker theorems for diffusions: necessary and sufficient conditions (Q2569224)

Let \(\{X_t,{\mathbf P}_x\}\) denote a regular diffusion on an open interval \(I\) with finite speed measure \(m\) and local time \(l_t(x)\), let \(\Lambda\) be a collection of signed measures on \(I\), and consider the random maps \({\mathbf H}_t\) on \(\Lambda\) defined by \[ {\mathbf H}_t\lambda:= t^{1/2} \int_I(l_t(x)/t- 1/m(x))\,\lambda(dx). \] The finite-dimensional distributions of \({\mathbf H}_t\) converge weakly to those of a Gaussian, zero-mean random mapping \({\mathbf H}\) on \(\Lambda\) with covariance function \[ {\mathbf E}\,{\mathbf H}\lambda\,{\mathbf H}\nu= (4/m(I)) \int_I(\lambda(l, x]- F(x)\lambda(I))(\nu(l,x]- F(x) \nu(I))\,ds(x), \] where \(F\) is the distribution function corresponding to \(m/m(I)\) and \(s\) is the scale function. The random map \({\mathbf H}\) induces a natural pseudometric \(d^2_{{\mathbf H}}(\lambda, \nu)={\mathbf E}({\mathbf H}\lambda-{\mathbf H}\nu)^2\) on \(\Lambda\). Suppose \(\sup_{\lambda\in\Lambda}\|\lambda\|< \infty\). Then, for every fixed \(t\), \({\mathbf H}_t\) is a random map into the space \(\ell^\infty(\Lambda)\) of uniformly bounded real-valued functions on \(\Lambda\) equipped with the supremum-norm, and the question is, whether the weak convergence of \({\mathbf H}_t\) to \({\mathbf H}\) takes place in \(\ell^\infty(\Lambda)\) with a tight, Borel measurable limiting process. Using majorizing techniques for continuous martingales, the authors show that the answer is positive if and only if \({\mathbf H}\) admits a bounded \(d_{{\mathbf H}}\)-uniformly continuous version. In that case, for every \(x\in I\), \({\mathbf H}_t\overset{{\mathbf P}_x} \Rightarrow{\mathbf H}\text{ in } C_b(\Lambda, d_{{\mathbf H}}),\text{ as }{t}\infty.\) The applications discussed include the convergence of the local times density estimator and of the empirical distribution function on the full state space.