Solve $\mathcal{H}v = F$ for $v$, where $\mathcal{H}$ is a nonlinear operator, $v$ is an input parameter, and $F$ is a predetermined forcing term

30 Views Asked by At

I am recreating results for an algorithm published in this paper, Shutyaev et al (2018). Essentially, I take optimal solutions I have found for a PDE constrained minimisation problem, and use them to find the Hessian of my original cost function $J$ with respect to a parameter control variable $\beta(x)$. If I consider optimal solutions $(\eta(x,t), u(x,t))$, and some appropriate adjoint solution $(\eta^*(x,t), u^*(x,t)$ given the optimal control variable $\beta$ minimising $J$, then the hessian $\mathcal{H}$ of the cost function $J$ acting on some $v:= v(x)$ is defined as the successive solutions of

Tangent Model with initial time conditions \begin{align} \frac{ \partial\hat{\eta}} { \partial t} + \frac { \partial \hat{u}}{ \partial x} + \frac{ \partial \big(u\hat{\eta}\big)}{ \partial x} + \frac {\partial \big(\hat{u} \eta\big)} {\partial x} - \frac {\partial \big(v u\big)} { \partial x} - \frac {\partial \big(\beta \hat{u}\big)} {\partial x} = 0, & \\ \frac {\partial\hat{u}} {\partial t} + \frac {\partial\big(\hat{u} u\big)} { \partial x} + \frac {\partial \hat{\eta}} {\partial x} = 0, & \\ \hat{\eta}(x,0) = 0, \ \hat{u}(x,0) = 0.& \end{align}

Second Order Adjoint system with final time conditions \begin{align} \frac {\partial \bar{\eta}} {\partial t} + \hat{u} \frac {\partial \eta^*} {\partial x} + u \frac { \partial\bar{\eta}} {\partial x} + \frac {\partial \bar{u}}{\partial x} = - \hat{\eta}(x_i,t ;v), & \\ \frac {\partial \bar{u}} {\partial t} + ( 1 + \eta - \beta) \frac {\partial \bar{\eta}} {\partial x} + (\hat{\eta} - v ) \frac {\partial \eta^*} {\partial x} + u \frac {\partial \bar{u}} { \partial x} + \hat{u} \frac {\partial u^*} {\partial x} = 0, & \\ \bar{\eta}(x,T) = 0, \ \bar{u}(x,T) = 0.& \end{align} where $\hat{\eta}(x_i,t ;v)$ is the solution $\hat{\eta}$ given $v$ at a small set of known spatial locations $x_i$. Then the Hessian acting on $v$ is defined as \begin{equation} \mathcal{H}v = \int_0^T \big( \hat{u} \frac{\partial {\eta^*}} {\partial x} + u \frac{\partial \bar{\eta}}{\partial x} \big) dt . \end{equation}

Now, at a later stage of the paper, I am required to solve $\mathcal{H}v = F$ for $v$, where $\mathcal{H}$ is defined as above. $F:= F(x,t)$ is a forcing term I have already calculated, but I am struggling to find the appropriate way to solve $\mathcal{H}v = F$. If I were to use numerical linear algebra methods, I would need to discretize the operator $\mathcal{H}$ such that I could write it as a matrix $H$ multiplied by a vector $v$, and then use it to solve the matrix equation $Hv = F$. I'm struggling with how to discretise this expression in such a way. Alternatively, I tried to formulate this as an optimization problem where I look for the minimiser $v$ of $\parallel \mathcal{H}v - F\parallel_{\infty}$, but given the nonlinearity and nonconvexity of the function I keep hitting local minimums and fail to converge.

Any insights on how to approach this would be extremely helpful!

1

There are 1 best solutions below

1
On BEST ANSWER

Assuming you can solve the above PDEs relatively, you don't actually have to compute the matrix $H$, just the action of $H$ on the discretization of $v$, then you can use a Krylov linear solver, which is typical for these times of problems. All you have to do is formulate a function that takes in a discretization of $v$ and returns a discretization of $\mathcal{H}v$. You can then use something like GMRES to solve the linear system using this construction