A study of backward stochastic differential equation on a Riemannian manifold (Q2042810)

Backward stochastic differential equations (BSDEs) have previously been introduced on Riemannian manifolds that can be covered by single coordinate patch. The present paper introduces a definition valid for any compact Riemannian manifold \(N\) by first embedding it into \(\mathbb{R}^L\) for sufficiently large \(L\) by the Nash embedding theorem. For some \(m\geq 1\), let \(f(p,u)\) be a \(C^1\) map from \(N\times T^mN\) to \(TN\) (where \(T^mN\) denotes the \(m\)-fold fiberwise tensor product of \(TN\) with itself) such that, for some \(C_0\), \[ |f(p,u)|_{T_pN}\leq C_0(1+|u|_{T^m_pN}) \text{ and } |\nabla_pf(p,u)|_{T_pN}+ |\nabla_pf(p,u)|_{T_pN}\leq C_0 ; \] \(f\) is the generator of the BSDE. We let \(\overline{f}\) and \(\overline{A}\) be extensions of \(f\) and the second fundamental form of the embedding (of \(N\) into \(\mathbb{R}^L\)), respectively, to a neighborhood of \(N\) in \(\mathbb{R}^L\). Fix a time horizon \(T>0\). Further, let \(B_t\) be a Brownian motion in \(\mathbb{R}^m\) with natural filtration \(\mathcal{F}_t\), and let \(\xi\) be an \(N\)-valued \(\mathcal{F}_T\)-measurable random variable. Then \((Y,Z)\) solves the BSDE \[ Y_t=\xi-\sum_{i=1}^m\int_t^T Z_s^i\, dB_s^i- \sum_{i=1}^m\frac{1}{2}\int_t^T \overline{A}(Y_s)(Z_s^i, Z_s^i) \, ds +\int_t^T \overline{f}(Y_s,Z_s)\, ds\tag{*} \] if \((Y,Z)\) solves \((*)\) viewed as an \(\mathbb{R}^L\times \mathbb{R}^{mL}\)-valued process and if, for some \(q\geq 2\), \(Y\) and \(Z\) are predictable with \[ \mathbb{E}\left[ \sup_{0\leq t\leq T} |Y_t|^q\right] \quad\text{and}\quad \mathbb{E}\left[ \left(\int_0^T |Z_t|^2dt\right)^{q/2}\right] \] both finite, and \(Y\) is a.s.\ pathwise continuous (on \([0,T]\)). In the case when \(N\) can be covered by a single coordinate patch, this agrees with the intrinsic definition in the literature. This formulation allows the authors to apply existing results about the existence and uniqueness of solutions to BSDEs in \(\mathbb{R}^L\) to show that \((*)\) has a unique solution under various conditions, such as a bound on the Malliavin derivative of \(\xi\) or the condition that \(m=1\) and \(\xi=h(B_T)\) for some \(h\in C^1(\mathbb{T};N)\) (which does not require any convexity condition). An analogous definition, and an existence result, are given for \(L^2(\mathbb{T}^m;N)\)-valued BSDEs. Concrete applications are given to the problem of finding a \(\nabla\)-martingale on \(N\) with prescribed terminal value.