Discrete adjoint computations for relaxation Runge-Kutta methods (Q6158981): Difference between revisions

The authors present novel adjoint computations and properties of the relaxation Runge-Kutta method. They start with the first-order initial value system: \(y^{\prime }(t)=f(y(t),t)\), \(0<t\leq T\), completed with the initial condition \(y(0)=y_{init}\). A generic \(s\)-stage Runge-Kutta method is specified by a \(s\times s\) matrix \(A_{s}\) and two vectors \(b_{s}\) and \(c_{s}\) in \(\mathbb{R}^{s}\), and is written as \(y_{k}=y_{k-1}+\Delta t\sum_{i=1}^{s}b_{i}F_{k,i}\), \(Y_{k,i}=y_{k-1}+\Delta t\sum_{i=1}^{s}a_{ij}F_{k,i}\), \(i=1,\ldots ,s\), where \( F_{k,i}=f(Y_{k,i},t_{k-1}+c_{i}\Delta t)\). The authors define the smooth and convex entropy function \(\eta :\mathbb{R}^{N}\rightarrow \mathbb{R}\) associated to this problem as the solution to the problem: \(\frac{d\eta }{dt} (y(t))=\nabla \eta (y(t))^{\intercal }f(y(t),t)\). The relaxation Runge-Kutta method considers the modified equation: \(y_{k}=y_{k-1}+\gamma _{k}\Delta t\sum_{i=1}^{s}b_{i}F_{k,i}\), where the relaxation parameter \(\gamma _{k}\) is the non-zero root (closest to one) of the nonlinear scalar function: \( r_{k}(\gamma ;y)=\eta (y_{k-1}+\gamma \Delta t\sum_{i=1}^{s}b_{i}F_{k,i})-\eta (y_{k-1})-\gamma \Delta t\sum_{i=1}^{s}b_{i}\nabla \eta (Y_{k,i})^{\intercal }F_{k,i}\). The authors then move to the optimal control problem: \(min_{u}c(y,u)\) such that \(E(y,u)=0 \), with \(E(y,u)=y^{\prime }(t)-f(y(t),t)\), \(0<t\leq T\), \(y(0)-u\). \(c\) is the cost functional, \(y\) the vector of state variable and \(u\) the vector of control variables. The authors introduce the time-stepping scheme: \(\delta _{k}=\delta _{k-1}+\Delta t\sum_{i=1}^{s}b_{i}J_{k,i}\Delta _{k,i}+w_{k}\), \( \Delta _{k,i}=\delta _{k-1}+\Delta t\sum_{i=1}^{s}a_{ij}J_{k,j}\Delta _{k,j}\) , \(j=1,\ldots ,s\), where \(J_{k,i}=\frac{\partial f}{\partial y} (Y_{k,i},t_{k-1}+c_{i}\Delta t)\).\ \(\delta _{k}\) is the linearized Runge-Kutta approximation and \(\Delta _{k,i}\) the linearized Runge-Kutta internal stages. The adjoint Runge-Kutta formulas follow from the solution to \(E(y)\lambda =w\), a block upper-triangular system, which can be solved by back substitution to yield: \(\lambda _{k-1}=\lambda _{k}+\sum_{i=1}^{s}\Lambda _{k,i}+w_{k-1}\), \(\Lambda _{k,i}=\Delta tJ_{k,i}b_{i}\lambda _{k}+\sum_{i=1}^{s}a_{ij}\Lambda _{k,j}\), \(i=1,\ldots ,s \). \(\lambda _{k}\) is the adjoint Runge-Kutta approximation and \(\Lambda _{k}\) the adjoint Runge-Kutta internal stages. The authors then compute linearization time-stepping formulas for the incremental direction technique and relaxation methods. They prove some time-symmetry properties of these two methods. In the last part of their paper, they apply these methods to examples, choosing 2-, 3- and 4-stages Runge-Kutta schemes. They specify the linearizations, the consistency of the discrete relaxation Runge-Kutta adjoints, and the preservation of the time-symmetry property, in these cases.