How to compute the functional derivative of the following functional encountered in electrical impedance tomography?

45 Views Asked by At

In the context of electrical impedance tomography, one encounters the following partial differential equation.

Let $ \Omega $ be a bounded Lipschitz domain in $ \mathbb{R}^2 $ with $C^1$ boundary $ \partial \Omega $ with outward normal $ n $. Assuming $ \Omega $ contains material with electrical conductivity $ q(x) $ satisfying $ q(x) \geq q_0 > 0 $. Then the electrical potential $ u(x) $ inside $ \Omega $ satisfies \begin{gather} \label{eqn:eitpde} - \nabla \cdot (q \nabla u) = 0 \text{ in } \Omega \tag{1} \\ q \nabla u \cdot n = f \text{ on }\partial \Omega \tag{2} \label{eqn:eitbc} \end{gather} where $ f(x) $ is the applied current density on $ \partial \Omega $ satisfying $ \int_{\partial \Omega} f(x) \operatorname{d}\! x = 0 $.

In solving the inverse problem, one may define the following mismatching functional \begin{equation} \tag{3} \label{eqn:mf} F(q) = \frac{1}{2} \int_{\partial \Omega} (u - m)^2 \end{equation} (The simplified version is stated here for easier discussion, for the complete mismatching functional, please refer to the above reference.) Then it remains to compute the gradient of $F$, now according to my understanding, the gradient of $F$ is the functional derivative, i.e. Gateaux derivative of $F$.

From the above reference, the gradient of $F$ is computed as \begin{equation} \tag{4} \label{eqn:gradf} \frac{\operatorname{d}\!F}{\operatorname{d}\!q} = - \nabla u \cdot \nabla z \end{equation} where $ z $ satisfies \begin{gather} \label{eq:avpde} - \nabla \cdot (q \nabla z) = 0 \text{ in }\Omega \tag{5} \\ q \nabla z \cdot n = u - m \tag{6} \label{eq:avbc} \end{gather} with the constraint \begin{equation} \tag{7} \label{eqn:zcons} \int_{\partial \Omega} z = 0. \end{equation} My question is

How do one obtain \eqref{eqn:gradf}?

In attempt to compute the functional derivative, I have first assumed a perturbation in $ q $ by $ \epsilon \tilde{q} $ which induces a corresponding perturbation in $ u $ by $ \epsilon \tilde{u} $. The change in $F$, denoted by $ \delta F $ is then \begin{equation} \tag{8} \label{eqn:Fturb} \delta F(q) = \epsilon \int_{\partial \Omega} \tilde{u}(u - m) + O(\epsilon^2). \end{equation} But then, the perturbations $ \tilde{u} $ and $\tilde{q}$ are related by \begin{equation} \tag{9} \label{eqn:uqturb} - \nabla \cdot [(q + \epsilon \tilde{q}) \nabla (u + \epsilon \tilde{u})] = 0 \end{equation} which transforms to (with ignoring higher order terms) \begin{equation} \tag{10} \label{eqn:uqturbfinal} - \epsilon [ \nabla \cdot (\tilde{q} \nabla u) + \nabla \cdot (q \nabla \tilde{u})] = 0 \end{equation} I tried to multiply \eqref{eqn:uqturbfinal} by $ z $ and integrate over $ \Omega$, with applying Green's first identity to yield \begin{equation} \tag{11} \label{eqn:asq} \int_{\partial \Omega} q z \nabla \tilde{u} \cdot n - \int_\Omega q \nabla z \cdot \nabla \tilde{u} - \int_\Omega \tilde{q} \nabla z \cdot \nabla u = 0. \end{equation} But this is the place where I got stuck, it seems like there is no evident choice of $ z $ such that when \eqref{eqn:asq} is added to \eqref{eqn:Fturb}, the perturbations are decoupled. I would appreciate very much if anyone can give a guide on how should I proceed or point out what did I miss or what went wrong.

Update (2020-02-10)

So basically I reworked and apply appropriate Green's identities instead, Instead of \eqref{eqn:asq}, I've got \begin{equation} \tag{12} \label{eqn:asqnew} \int_{\partial \Omega} z \tilde{q} \nabla u \cdot n - \int_\Omega \tilde q \nabla z \cdot \nabla u + \int_{\partial \Omega} q z \nabla \tilde u \cdot n - \int_{\partial \Omega} q \tilde u \nabla z \cdot n - \int_\Omega \tilde{u} \nabla \cdot ( q \nabla z ) = 0. \end{equation} So by adding \eqref{eqn:asqnew} to \eqref{eqn:Fturb}, one obtain \begin{multline} \label{Fturbnew} \tag{13} \delta F(q)/\epsilon = \int_{\partial \Omega} \tilde{u}(u - m) + \int_{\partial \Omega} z \tilde{q} \nabla u \cdot n - \int_\Omega \tilde q \nabla z \cdot \nabla u \\ + \int_{\partial \Omega} q z \nabla \tilde u \cdot n - \int_{\partial \Omega} q \tilde u \nabla z \cdot n - \int_\Omega \tilde{u} \nabla \cdot ( q \nabla z ) + O(\epsilon^2). \end{multline} Now it is evident that if one assumes that $ \nabla u \cdot n $ is given, then $ \tilde q = 0 $ on $ \partial \Omega$ by \eqref{eqn:eitbc}. From the given gradient one also knows that the term $ \int_{\partial \Omega} q z \nabla \tilde u \cdot n $ should be zero, So that, if $ z $ indeed satisfies \eqref{eq:avpde} and \eqref{eq:avbc}, then \eqref{Fturbnew} becomes \begin{equation} \delta F(q)/\epsilon = - \int_\Omega \tilde q \nabla z \cdot \nabla u + O(\epsilon^2). \end{equation} So, one picks $ \tilde q = \nabla z \cdot \nabla u $ to ensure the functional decreases along the gradient direction. My new question would be:

How do you know the term $ \int_{\partial \Omega} q z \nabla \tilde u \cdot n $ is zero?