BSDE with quadratic growth and unbounded terminal value (Q2431750)

The authors consider a backward stochastic differential equation \[ Y_t=\xi+\int_t^Tf(s,Y_s,Z_s)ds- \int_t^TZ_s\cdot dB_s,\qquad 0\leq t\leq T,\tag{1} \] where \(B\) is a standard Brownian motion in \({\mathbb R}^d\) (the natural filtration generated by it is denoted by \(({\mathcal F}_t)_t\)), \(f\) is a measurable random predictable function such that \[ | f(t,y,z)| \leq\alpha +\beta | y| +(\gamma/2)| z| ^2,\qquad (t,y,z)\in[0,T]\times{\mathbb R}\times{\mathbb R}^d\tag{2} \] for some \(\alpha,\beta\geq 0\), \(\gamma>0\), and \(\xi\) is a real \({\mathcal F}_T\)-measurable random variable such that \[ E\exp \{\gamma e^{\beta T}| \xi| \}<\infty\tag{3} \] (so \(\xi\) need not be bounded). It is proved that under these conditions equation (1) has a solution \((Y,Z)\); moreover, if \(Ee^{\lambda| \xi| }<\infty\) for some \(\lambda>\gamma e^{\beta T}\), then \(E[\int_0^T| Z_s| ^2ds]^{1/2}<\infty\). Let \(H(p)=p(\alpha\gamma+\beta\ln p ){\mathbf1 }_{[1,\infty)}(p)+\gamma\alpha{\mathbf1 }_{(-\infty,1)}(p)\), \(p\in{\mathbb R}\), and \(\{\Phi_t(x)\}_{t\in[0,T]}\) be the solution of \(\Phi_t= e^{\gamma x}+\int_t^TH(\Phi_s)ds\). The authors prove that under (2), (3) the solution \(Y\) of (1) satisfies \[ -(1/\gamma)\ln E(\Phi_t(-\xi)| {\mathcal F}_t)\leq Y_t\leq (1/\gamma)\ln E(\Phi_t(\xi)| {\mathcal F}_t). \] This result is generalized to coefficients \(f\) with superlinear growth in \(y\). Some results of comparison type are also given.