Atypical exit events near a repelling equilibrium (Q2039432)

Consider a stochastic diffusion process \(X\) with a small diffusion coefficient of order \(\varepsilon\). Assume that the associated dynamical system obtained in the limit \(\varepsilon=0\) has an unstable equilibrium point and that the stochastic system starts from this point. We know from the classical Freidlin-Wentzell theory that the probability of quitting a neighbourhood of this equilibrium before some fixed time is exponentially small, with a rate of order \(1/\varepsilon^2\). However, in longer times, the process will eventually quit this neighbourhood at a time \(\tau\). It was known that the escape time is of order \(\log(1/\varepsilon)\), and that the exit location \(X_\tau\) is obtained by following the most unstable direction(s). The aim of the paper under review is to consider more unlikely exit locations occurring in the case of an abnormally long stay in the neighbourhood. The main result states that under convenient assumptions, \(\mathbb{P}(X_\tau\in A)\sim\varepsilon^{\rho(A)}\mu(A)\) for \(\rho(A)\) taking values in a discrete set depending on the geometry of the dynamical system. Moreover, the conditional distribution of \(X_\tau\) given \(\{X_\tau\in A\}\) is studied. For the proof one notices that it is sufficient to consider arbitrarily small neighbourhoods, since after quitting such a neighbourhood the process is almost deterministic (due to the unstability of the equilibrium). On the other hand, on a small neighbourhood the process can be approximated by a Gaussian process. The approximation error was already studied by the same authors by applying Malliavin's calculus, so they have to study the behaviour of the Gaussian process.