Auxiliary SDEs for homogenization of quasilinear PDEs with periodic coefficients. (Q1889783)

The author considers the system of \(Q\) quasilinear equations (here \((\theta_{\epsilon})_l\) which depends on \((t,x) \in [0,T]\times {\mathbb R}^P\) and \(l=1, \dots Q\)) \( {\mathcal E}(\varepsilon)\): \[ \begin{multlined}{{\partial (\theta_{\varepsilon})_l}\over{\partial t}} +{{1}\over{2}} \sum_{i,j=1}^P a_{i,j}(\varepsilon^{-1}x, \theta_{\epsilon}) {{\partial^2 (\theta_{\varepsilon})_l}\over{\partial x_i \partial x_j}} + \sum_{i=1}^P \big[\varepsilon^{-1} b_{i}(\varepsilon^{-1}x, \theta_{\varepsilon}) + c_{i}(\varepsilon^{-1}x, \theta_{\varepsilon}, \nabla_x \theta_{\varepsilon})\big] {{\partial (\theta_{\varepsilon})_l}\over{\partial x_i}} \\+ \varepsilon^{-1} e_l(\varepsilon^{-1}x, \theta_{\epsilon}) + f_l(\varepsilon^{-1}x, \theta_{\epsilon}, \nabla_x \theta_{\epsilon}) = 0,\end{multlined} \] \[ \theta_{\varepsilon}(T,x) = H(x), \] where the coefficients \(a, b, c, e, f\) are \([0,1]^P\)-periodic functions and \(a\) is assumed to be symmetric. The aim is to study the behaviour of the system when \(\varepsilon\) goes to zero. To treat (from a probabilistic point of view) this system, the author represents the system by the forward-backward stochastic differential equations (here \(X_r, Y_r, Z_r\) depend on \((\varepsilon,t,x)\)) \(E(\varepsilon, t, x)\): \[ \begin{aligned} X_s &= x + \int_t^s (\varepsilon^{-1} b + c) (\varepsilon^{-1}X_r, Y_r, Z_r) \,dr + \int_t^s \sigma (\varepsilon^{-1}X_r, Y_r) \,d B_r ,\\ Y_s &= H(X_T) + \int_s^T \varepsilon^{-1} e + f) (\varepsilon^{-1}X_r, Y_r, Z_r) \,dr - \int_s^T Z_r\sigma (\varepsilon^{-1}X_r, Y_r) \,d B_r ,\\ {\mathbf E}&\int_t^T(| X_s| ^2+| Y_s| ^2+| Z_s| ^2) \,ds < + \infty , \end{aligned} \] where \((\varepsilon^{-1} b + c)(x,y,z) = \varepsilon^{-1} b(x,y) + c(x,y,z)\) and \((\varepsilon^{-1} e + f)(x,y,z) = \varepsilon^{-1} e(x,y) + f(x,y,z)\), \((B_t)_{t\geqslant 0}\) is a Brownian motion and \(\sigma\) is such that \(\sigma \sigma^{\ast} = a\). The connection between \({\mathcal E}(\varepsilon)\) and \(E(\varepsilon, t, x)\) can be roughly summarized as follows: for every \(s \in [t,T]\) \(Y_s(\varepsilon,t,x) = \theta_{\varepsilon}(s, X_s(\varepsilon,t,x))\). To study problems \(E(\varepsilon, t, x)\) the strategy is the following: pass to some modified processes (denoted by \(\widehat{X}\) and \(\widehat{Y}\)) in order to get rid of the terms \(\varepsilon^{-1} b\) and \(\varepsilon^{-1} e\), which leads to the study of the ``auxiliary problems''; estimate the distance between \(\theta (\cdot , \widehat{X}(\varepsilon, t,x))\) and \(\widehat{Y}(\varepsilon, t,x)\), where \(\theta\) is the solution of the presumed limit system; to do this, since \(\theta\) is not regular enough, the author needs to pass through a regularization of \(\theta\), an appropriate sequence \((\zeta_n)_n\), and introduce some ``auxiliary SDEs'', a new approach needed to by-pass a problem due to \((\zeta_n)_n\) (a control, uniform in \(n\), of \(D_{xx}\zeta_n\)). Finally the author establishes that \[ \begin{multlined} \lim_{\varepsilon \to 0} \biggl( {\mathbf E} \sup_{t \leqslant s \leqslant T} | Y_s(\varepsilon, t,x) - \theta (s, X_s(\varepsilon, t,x))| ^2 + {\mathbf E} \int_t^T | Z_s(\varepsilon, t,x) - \nabla_x \theta(s,X_s(\varepsilon, t,x)) \\ \chi_1(\varepsilon^{-1}X_s(\varepsilon, t,x),Y_s(\varepsilon, t,x)) - \chi_2(\varepsilon^{-1}X_s(\varepsilon, t,x),Y_s(\varepsilon, t,x)) | ^2\,ds\biggr) = 0 \end{multlined} \] where \(\chi_1\) and \(\chi_2\) are appropriate corrector terms.