A probabilistic representation of the solution of some quasi-linear PDE with a divergence form operator. Application to existence of weak solutions of FBSDE. (Q2574637)

From MaRDI portal
scientific article
Language Label Description Also known as
English
A probabilistic representation of the solution of some quasi-linear PDE with a divergence form operator. Application to existence of weak solutions of FBSDE.
scientific article

    Statements

    A probabilistic representation of the solution of some quasi-linear PDE with a divergence form operator. Application to existence of weak solutions of FBSDE. (English)
    0 references
    0 references
    29 November 2005
    0 references
    The first part of the paper deals with Markov processes generated by time-inhomogeneous uniformly elliptic differential operators on \(\mathbb R^N\) in a divergence form, i.e.\ \(L=\partial _i(a_{ij}(t,x)\partial _j)+b_i(t,x)\partial _i\) where the coefficients \(a\) and \(b\) are measurable and uniformly bounded. It is shown that, indeed, there exists a continuous strong Markov process \((\Omega ,\mathcal F,X_t,\mathbb P_{s,y},\mathcal F_{s,t})\) generated by \(L\), and, for every time \(s\) and \(y\in \mathbb R^N\), the process \(X\) admits a decomposition \(X_t=y+M_t+V_t\), \(t\geq s\), where \(M\) is a continuous square-integrable martingale and \(V\) is a continuous process with zero-quadratic variation. Moreover, an explicit form of the process \(V\) is given using the fundamental solution \(\Gamma \) of the equation \(\partial _t+L=0\) and some \(\mathcal G_t\)-martingale \(\overline M_t\) where \(\mathcal G_t=\sigma (X_r:t\leq r\leq T)\) is the filtration relative to the future of \(X\). Although \(X\) is not (in general) a semimartingale, an Itô-type formula characterizing the process \(u(t,X_t)\) where \(u\) is a one-times differentiable function is proven. As an application, a linear Feynman-Kac formula is proven for the Markov semigroup corresponding to \(L\), and the author further points out that the decomposition \(X=y+M+V\) may be used to construct stochastic integrals of type \(\int f(X_t)\,dX_t\) as in the case of time-homogeneous generators \(L\), cf.\ e.g.\ \textit{A.\ Rozkosz} [Stochastic Processes Appl. 63, 11--33 (1996; Zbl 0870.60073)] or \textit{T.\ J.\ Lyons} and \textit{L.\ Stoica} [Ann.\ Probab.\ 27, 1--49 (1999; Zbl 0969.60078)]. The second part deals with a link between a semilinear partial differential equation \[ \partial _t u+Lu+h(t,x,u,\nabla u)=0\tag{1} \] on a cylinder \([0,T]\times \mathcal O\) endowed with a terminal condition \(u(T)=g\), and a backward stochastic differential equation \[ dY=h(t,X_t,Y_t,Z_t)\,dt-Z_t\,dM\tag{2} \] with the terminal condition \(Y_T=g(X_T)\) and unknowns \(Y\) and \(Z\). It is proven that for a certain set \(\mathcal N\subseteq [0,T]\times \mathbb R^N\) of points \((s,y)\), the connection between a solution \(u\) of the equation (1) and a \(\mathbb P_{s,y}\)-solution \((Y,Z)\) of the equation (2) is that \(Y_t=u(t,X_t)\) and \(Z_t=\nabla u(t,X_t)\). Furthermore, if the infinitesimal generator has a particular form \[ L=\partial _i(a_{ij}(t,x,u(t,x))\partial _j)+b_i(t,x,u(t,x),\nabla u(t,x))\partial _i \] for a function \(u\), then the following holds: (A) if \((s,y)\in \mathcal N\), \(u\) solves the quasilinear PDE \[ \partial _t u+Lu+h(t,x,u,\nabla u)=0\tag{3} \] and \((Y,Z)\) is the unique \(\mathbb P_{s,y}\)-solution of the equation (2), then \(Y_t=u(t,X_t)\) and \(Z_t=\nabla u(t,X_t)\); (B) if \[ du(t,X_t)=h(t,X_t,u(t,X_t),\nabla u(t,X_t))\,dt-\nabla u(t,X_t)\,dM,\qquad u(T,X_T)=g(X_T) \] holds \(\mathbb P_{s,y}\)-almost surely for every \((s,y)\in \mathcal N\), then \(u\) is a solution of the equation (3). In the third part, additional smoothness assumptions are imposed upon the matrix function \((a_{ij})\). This results into the fact that if the function \(u\) is a solution of the equation (3), then \(\nabla _xu\) is bounded and \(\partial _t u\), \(\partial _{x_ix_j}u\) are \(L^2\)-integrable. Finally, this regular solution is used to prove that the system of a forward stochastic differential equation \[ dX=\sigma (t,X_t,Y_t)\, dB+\widehat b(t,X_t,Y_t,Z_t)\, dt, \quad X_s=x, \tag{4a} \] and a backward stochastic differential equation \[ dY=h(t,X_t,Y_t,Z_t)\, dt-Z_t\sigma (t,X_t,Y_t)\,\text dB, \quad Y_T=g(X_T), tag4b \] where \((Y,Z)\) are unknowns, \(B\) is a Brownian motion, \(\sigma \) satisfies \(\sigma \sigma ^T=a\) and \(\hat b_i=\partial _ka_{ki}+\partial _ya_{ki}z_k+b_i\), has a weak solution whose uniqueness is left as an open problem even in case the solutions of the equation (3) are known to be unique. The last part of the paper contains a weak existence theorem for the FBSDE (4a), (4b) with \(\widehat b=b\) under the assumption of existence of a solution of the equation (3) where the second order differential operator \(L\) is slightly modified, and an illustrative example of weak existence of a solution of a one-dimensional FBSDE and its corresponding quasilinear PDE is given.
    0 references
    0 references
    forward-backward stochastic differential equation
    0 references
    time reversal
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references