An ENO-based method for second-order equations and application to the control of dike levels (Q421325): Difference between revisions

The present work uses a model in which the stochastic behaviour of economic growth is modelled in continuous time. The resulting optimization problem leads to a so-called Hamilton-Jacobi-Bellman (HJB) equation. It is a system of second-order partial differential equations that needs to be solved backwards in time. This is achieved by numerical approximation. Formulating an optimal control problem requires an expression for the total future expected discounted costs and input variables which govern optimal values. The authors consider an impulsive control formulation, where it is assumed that the input variable increases instantaneously (both the optimal input variable and the intervention times are typically not known in such optimal stopping problems and need to be determined). The numerical treatment of impulse control model equations falls into two parts: the uncontrolled problem, (i.e., between intervention times \(t_K\) and \(t_{K+1}\)) and the impulse control. The uncontrolled part of the problem (typically of convection-reaction type) has the following general form: \[ \frac{\partial u}{\partial t}+ \nabla . f(u) = \nabla . K (u,t) \nabla u + s(u,t), \tag{1} \] complemented with appropriate boundary and initial conditions. Here the convection part \(\nabla.f\) relates to the deterministic part of the system dynamics, while the diffusion part relates to the stochastic behaviour. The source term, \(s\) in (1), accounts for the running costs and discounting (is brought into conservative form whereas the original operator in a control problem is not). The HJB equations are solved numerically with an essentially non-oscillatory (ENO) method, which combines high-order accuracy with convergence to the relevant (viscosity, or entropy) solution. The model equations are of purely convective type when the system dynamics are fully deterministic (\(K(u,t) = 0\) in (1)) and fully diffusive when the drift in a stochastic state system vanishes, i.e., \(f(u)=0\). Convection (a fully deterministic problem without the source terms), here the model equation (1), reduces to \[ \frac{\partial u}{\partial t}+ \nabla . f(u) = 0, \tag{2} \] which is a hyperbolic conservation law. Main result: The ENO method, as originally intended for convection-type problems (with the purpose of extending it to the discretization of the diffusion operator), is presented. By integrating (2) over a control volume a discretization (inherently conservative) can be obtained. There is no central difference scheme \(L\) of order of accuracy higher than two for which the numerical method \(E\) (a forward-Euler-type numerical method, \(u_j^{n+1}=E(u^n,j)\)), is monotone.