The averaging principle for stochastic differential equations driven by a Wiener process revisited (Q2122136)

Let \(X^\varepsilon\) be the solution of the following stochastic differential equation (SDE): \[ dX^\varepsilon(t)=\sigma(X^\varepsilon(t),m(t/\varepsilon)d\beta(t)), \] where \(\sigma\) is deterministic and smooth, \(m(\cdot)\) an ergodic Markov process and \(\beta\) a one dimensional Brownian motion. The averaging principle states that, when \(\varepsilon\searrow 0\), then \(X^\varepsilon\) converge in distribution to the solution of the SDE \[ d\bar X(t)=\Sigma(\bar X(t))d\beta(t), \] with \(\Sigma^2(\cdot)=\overline{\sigma^2(\cdot)}=\int\sigma(\cdot,m)^2d\mu(m)\), where \(\mu\) is the invariant measure of the process \(m\).\\ In the present paper, the author gives a new proof of this result, based on the decomposition of \(X^\varepsilon\) into the sum of two processes, representing respectively the average evolution and the fluctuations around this average. More precisely, he writes \(X^\varepsilon=Y^\varepsilon+Z^\varepsilon\), where \((Y^\varepsilon ,Z^\varepsilon )\) solves \[ \left\{\begin{array}{l} dY^\varepsilon(t)=\bar\sigma(Y^\varepsilon(t)+Z^\varepsilon(t))d\beta(t)\\ dZ^\varepsilon(t)=(\sigma(Y^\varepsilon(t)+Z^\varepsilon(t),m(t/\varepsilon))-\bar\sigma(Y^\varepsilon(t)+Z^\varepsilon(t)))d\beta(t) \end{array}\right. \] and converges in distribution to the solution of \[ \left\{\begin{array}{l} dY(t)=\bar\sigma(Y(t)+Z(t))d\beta^1(t)\\ dZ(t)=\langle \sigma\rangle (Y(t)+Z(t))d\beta^2(t), \end{array}\right. \] where \(\langle\sigma\rangle^2=(\sigma-\bar\sigma)^2\) and \(\beta^1\) and \(\beta^2\) are two independent Brownian motions. This decomposition permits to explain the presence of \(\Sigma\) instead of \(\bar\sigma\) in the limit and also why the convergence only holds in distribution.