A probabilistic approach to Dirichlet problems of semilinear elliptic PDEs with singular coefficients (Q717887)

The following Dirichlet boundary value problem for the a class of semi-linear second-order elliptic PDE is treated: \[ \left\{\begin{aligned} &\mathcal A u(x)=-f(x,u(x),\nabla u(x))\\ &u|\partial D=\phi,\end{aligned}\right. \tag{1} \] where \(u(x)\) is defined in a bounded domain \(D\) (with boundary \(\partial D\)) of \(\mathbb R^d\). The p.d.o. \(\mathcal A\) is of the form \[ \mathcal A=\frac12 \sum_{i,j=1}^d\frac{\partial}{\partial x_i} \left(a_{ij}\frac{\partial u}{\partial x_j}\right)+\sum_{i=1}^d b_i\frac{\partial u}{\partial x_j}- ``\text{div}(\hat bu)" +q(x)u, \tag{2} \] in which the diffusion part of \(\mathcal A\), i.e., \[ L=\frac12 \sum_{i,j=1}^d \frac{\partial}{\partial x_i}\left(a_{ij}\frac{\partial u}{\partial x_j}\right) \tag{3} \] is supposed to be elliptic. The classical case of the problem where \(f\) (resp., \(\hat b\)) is identically null can be solved by probabilistic methods via the Feynman-Kac formula. The main difficulties to solve (1) lies, on the one hand, in the non-linearity of the measurable function \(f :\mathbb R^d \times\mathbb R \times\mathbb R^d \rightarrow\mathbb R\), and in the interpretation of the term \(\text{div}(\hat bu)\) in the distributional sense on the other hand. Note that all the terms \(a = ( a_{ij} ), b = (b_i )\) and \(\hat{b} = (\hat {b}_i )\) mentioned above are supposed to be real measurable functions on \(D\) that are eventually locally \(L^p\)-integrable. To be more precise, the author expresses the p.d.o. \(\mathcal A\) as an operator acting on the Sobolev space \(W^{1,2} ( D)\) as follows (see Def. 2.1): \[ \begin{multlined} (-\mathcal A u,v)_{L^2}=\frac12\sum_{i,j=1}^d\int_D a_{ij}\frac{\partial u}{\partial x_i}\frac{\partial v}{\partial x_j}dx+\sum_{i=1}^d\int_D b_i\frac{\partial u}{\partial x_i}v_i dx\\ -\sum_{i=1}^d\int_D\hat{b}_i u\frac{\partial v}{\partial x_i}dx- \sum_{i=1}^d\int_D quv dx. \end{multlined} \tag{4a} \] The boundary valued problem (1) is now written in the weak form as follows: \[ (-\mathcal A u,\phi)_{L^2}=(f(x,u,\nabla u),\phi)_{L^2},\;\forall \phi\in C^{1,2}(D)\text{ and }\forall u\in C(\overline{D})\text{ such that }u|\partial D=\phi.\tag{4b} \] Such a function \(u \in C (\overline{D})\), if it exists, is called a weak solution of the boundary value problem (1). The aim of the author is to construct a backward stochastic differential equation (BSDE) associated with the Dirichlet process given by the quadratic form (4a) and then to use the solution of the above mentioned SDE to give a stochastic representation the solution of (4b). The author mentions in the introduction that the term \(\hat{b} = (\hat{b}_i )\) can be tackled here by the time reversal of the Girsanov transform of the diffusion process \(X^0\) (associated with the elliptic p.d.o. \(L\), see (3)) from its exit-time \(\tau_D\) for \(D\). For a particular case (where \(f = 0\)), a weak solution of (4b) is given by \[ \begin{multlined} u(x)=\text{E}_x^0\left[\phi(X^0(\tau_D))\exp\left(\int_0^{\tau_D}\langle a^{-1}b\cdot X_s^0,dM_s^0\rangle\right)\right]\dots \\ +\text{E}_x^0\left[\int_0^{\tau_D}\langle a^{-1}\hat{b}\cdot X_s^0,dM_s^0\circ\tau_D\rangle\right]\dots \\ -\frac12 \text{E}_x^0\left[\int_0^{\tau_D}((b-\hat{b})a^{-1}(b-\hat{b})^*\cdot X_s^0+q(X_s^0))ds\right],\end{multlined} \tag{5} \] where \(M^0 = (M^0_s)\) is the martingale part in Fukushima's decomposition of \(X^0 = (X^0_s)\), \(r_t\) the reverse operator, and \(\langle .,. \rangle\) the inner product of \(\mathbb R^d\). The general case where \(f\) is non-linear is treated in Section 3. Since the diffusion process \(X = (X_s)\) associated with the p.d.o. \(L_2 = L + \sum_{i,j=1}^d b_i\frac{\partial}{\partial x_i}\) is not a semi-martingale, the author first shows the existence of a solution \((X,Z)\) for the BSDE \[ Y_s=\xi+\int_t^T f(s,Y_s,Z_s)ds-\int_t^T\langle Z_s,dM_s\rangle, \tag{6} \] in which \(M=(M_s)\) is the martingale part of \(X\). Note that the result is obtained under the hypotheses that the stochastic integrand \(f(s, Y_s, Z_s)\) must satisfy some monotone hypotheses and the terminal \(T\) is supposed to be deterministic. The case of random terminal \(T\) is studied in the next Section 4. In the last Section 5, the author comes back to the boundary value problem (4b) and shows in Theorem 5.1 the existence of a weak solution of (4b). The stochastic representation of the value \(u(x)\) (see (5)) is too complicated in the semi-linear case to be written down here. Note that the author treated only the case where \(f:\mathbb R^d \times\mathbb R^d \rightarrow\mathbb R\) satisfies {\((D_1)\)} \((y_1-y_2)(f(x,y_1)-f(x,y_2))\leq -J_1(x)|y_1-y_2|^2, \) {\((D_2)\)} \(|f(y_1,y_2)|\leq C,\) where \(J_1(x)\) is a measurable function and \(C\) is a constant. Reviewer's remark: It seems to the reviewer that the author investigates the semi-linear Dirichlet boundary value problem for a subclass of the following class of elliptic p.d.o: \[ \mathcal G=\frac12\sum_{i,j=1}^d \frac{\partial^2}{\partial x_i\partial x_j}(a_{ij}u)+\sum_{i,j=1}^d\frac{\partial}{\partial x_i}(b_iu)+cu, \tag{7a} \] in which the coefficients are supposed only measurable. In the particular case where \(\mathcal G\) can be written in the form \[ \mathcal G_0=\frac12\sum_{i,j=1}^d\frac{\partial}{\partial x_i}(a_{ij}\frac{\partial u}{\partial x_j})+q(x)u, \tag{7b} \] it is known that, under very mild conditions on the coefficients, we are actually in the case of a Brelot potential axiomatic. Therefore, the (linear and semi-linear) Dirichlet boundary value problem here can eventually be solved by Perron-Brelot methods. On the other hand, the Markov process associated with the above mentioned Brelot axiomatic is a Hunt process; therefore, we could used the exit-boundary (Martin boundary) of the above mentioned Hunt process for the domain \(D\) to give a stochastic representation of the weak solutions of the Dirichlet problem given by Perron-Brelot methods. Finally, it is interesting to remark that, if we add the distributional drift ``\(\text{div}(\hat{b}u)\)'' to the the p.d.o. \(\mathcal G_0\), then the adjoint of \(\mathcal G_0 u + \text{div}(\hat{b}u )\) turns out to be an elliptic p.d.o. of the same type as \(\mathcal G_0\) (see (4a)).