the backward stochastic differential equation

A backward stochastic differential equation (BSDE) is an SDE solved BACKWARD in time, from a terminal condition rather than an initial one — but with a twist that makes it deeply different from just running an ODE backward. The catch is adaptedness: the solution must, at every time t, depend only on information available up to t (it must be adapted to the filtration), yet its value at the FINAL time T is prescribed. You cannot simply integrate backward, because that would make today's value depend on the future. The resolution forces the solution to be a PAIR of processes, and BSDEs have become a unifying language for stochastic control, nonlinear pricing, and nonlinear PDE.

The standard BSDE is dY_t = - f(t, Y_t, Z_t) dt + Z_t dB_t with terminal condition Y_T = xi, where xi is a given F_T-measurable random variable (the terminal payoff) and f is the driver (generator). A solution is a pair (Y_t, Z_t) of adapted processes: Y_t is the value, and the EXTRA process Z_t is the martingale-representation integrand that is precisely what allows Y to be adapted despite its terminal pinning — Z is the 'control' that keeps Y from peeking into the future, and it also has a financial reading as the hedging portfolio. The Pardoux-Peng theorem gives existence and uniqueness of (Y,Z) when the driver f is Lipschitz in (y,z) and xi is square-integrable; the simplest case f = 0 just recovers the martingale representation theorem, Y_t = E[ xi given F_t ]. The nonlinear Feynman-Kac formula then links BSDEs to PDEs: when xi = g(X_T) and f depends on the state through a forward SDE, Y_t = V(t, X_t) where V solves a semilinear parabolic PDE, and Z_t = sigma^T grad_x V — the BSDE is the probabilistic representation of a nonlinear PDE, generalising the linear Feynman-Kac formula.

BSDEs matter because they unify several threads: the adjoint equation of the stochastic maximum principle IS a (linear) BSDE; the value process of a nonlinear pricing problem (with transaction costs, different borrowing/lending rates, or risk measures) is a BSDE whose driver encodes the nonlinearity; and dynamic risk measures and g-expectations are built from BSDEs. An honest caveat: the standard existence-uniqueness theorem needs the driver Lipschitz and the data square-integrable; QUADRATIC-growth drivers (which arise from exponential utility and from robust/risk-sensitive control) require their own, harder theory (Kobylanski), and reflected BSDEs (constraining Y >= an obstacle) are the BSDE form of optimal stopping and American options. Dropping the Lipschitz/integrability hypotheses is not safe — existence can fail or solutions can blow up.