Large deviations for multiscale diffusion via weak convergence methods (Q424511)

The authors derive a large deviation principle (LDP) for two-scale SDEs via a Laplace principle and weak convergence methods. The framework consists of SDEs with Lipschitz continuous coefficients that are locally periodic in the fast time scale but admits a generalization for special cases. Such a framework is motivated by rugged energy landscapes in chemical physics or biology and importance sampling. Results are stated under the three possible limit regimes for the ratio of the time scale \(\delta\) and the noise level \(\epsilon\), i.e \(\epsilon/\delta\to 0,\infty\) or \(\gamma\in(0,\infty)\). If a large deviation functional \(S\) has compact level sets (i.e. a good rate function) then the LDP for solutions \(X^{\epsilon}\) is equivalent to the Laplace principle \[ \lim_{\epsilon\downarrow 0}-\epsilon\ln\mathbb{E}\left[ \exp\left\{ -\frac{h(X^{\epsilon})}{\epsilon} \right\}\right]=\inf_{\phi\in\mathcal{C}([0,1])} \left[ S(\phi)-h(\phi) \right] \] for bounded continuous \(h\) and the infimum is taken over the sample space of \(X^{\epsilon}\). Their method consists of expressing the prelimit left hand side as control problem with \(\bar X^{\epsilon}\) solution to a controlled SDE and the control \(u\). Weak convergence of \(\bar X^{\epsilon}/\delta\) then gives that the functional \(S\) is given by a variational equation for pairs \((\phi,P)\), where the probability measures \(P\) is determined by \(\phi\) and the generator of this limit. The method allows to obtain a control that nearly achieves the LDP lower bound at prelimit level.