According to my computations:

The function which minimizes $$\int_\Omega \|\operatorname{curl} f\|^2\,dx$$ should satisfy $$\operatorname{curl}(\operatorname{curl}f) = 0$$ everywhere on $\Omega$, provided $\operatorname{curl} f = 0$ on $\partial \Omega$.

I followed the same kind of computation that the one demonstrating the the argmin of $\int_\Omega \|\nabla f\|^2~dx$ should satisfy $\Delta f = 0$. However, I am not sure whether my computations are right... May anyone check that please ?

1) We first start with a functional $$G(f) = \int_\Omega \|\operatorname{curl} f\|^2\,dx.$$

2) We compute $$V(f,h) = \lim_{\epsilon\rightarrow 0} \frac{G(f+\epsilon h)-G(f)}{\epsilon} = 2\int_\Omega \operatorname{curl} f\cdot\operatorname{curl} h\,dx.$$

3) We use the identity : $$\operatorname{div}(A\times B) = -A\cdot\operatorname{curl} B + B\cdot\operatorname{curl} A,$$ with $A=\operatorname{curl} f$ and $B=h$.

4) We obtain $$\int_\Omega \operatorname{curl} f\cdot \operatorname{curl} h\:dx = -\int_\Omega \operatorname{div}(\operatorname{curl} f\times h)\,dx + \int_\Omega h\cdot \operatorname{curl}(\operatorname{curl}f)\,dx.$$

5) We use the divergence theorem to obtain : $$\int_\Omega \operatorname{curl} f\cdot \operatorname{curl} h\,dx = -\int_{\partial\Omega} \operatorname{curl} f\times h\,ds + \int_\Omega h\cdot\operatorname{curl}(\operatorname{curl}f)\,dx$$

6) We assumed $\operatorname{curl} f = 0$ on $\partial\Omega$, so the first term is zero.

7) $V(f,h)$ should equal zero for all $h$ for the function to be minimized, so $$\int_\Omega h\cdot\operatorname{curl}(\operatorname{curl}f) dx = 0\quad\forall h,$$ which implies $\operatorname{curl}(\operatorname{curl}f) = 0$ locally.

I guess this reasonning may be wrong in several places..... or is it right?? Thanks!

  • 1
    $\begingroup$ Your reasoning looks correct to me. I think you don't even need $ \nabla \times F = 0 $ on the boundary because $ h $ is supposed to vanish there anyway, thereby fulfilling (6) without need for the extra information. I may be a bit shaky on my vector calculus, but I think that given $ \nabla \times ( \nabla \times F) = 0 $ we can deduce that $ \nabla \times F = \nabla g $ for some $ g $, which is impossible unless $ g $ is constant and therefore $ \nabla \times F = 0 $ which finally means that $ F = \nabla \phi $ for some $ \phi $. I think. $\endgroup$ – anon Jun 8 '11 at 19:12
  • $\begingroup$ great, thanks :) I'm mainly worried about point 7 : I know that for scalar valued functions, if $\int f g=0~~\forall g$ then $f=0$ (almost everywhere!), but I don't know if this applies to vector valued functions, by replacing the product by a scalar product between them... $\endgroup$ – WhitAngl Jun 8 '11 at 20:07
  • 2
    $\begingroup$ Since it must be true for all $ h(x) = u(x) e_i $ for scalar functions $ u $ and $ \{ e_1, \dots e_n \} $ the vector space basis, we can still make the conclusion with sound logic. $\endgroup$ – anon Jun 8 '11 at 21:00
  • $\begingroup$ @anon: if $\Omega$ is a simply connected domain then $\nabla\times W = 0 \implies W = \nabla g$. Else it is better to just use the double curl identiy to get $\nabla\times(\nabla\times A) = \nabla (\nabla\cdot A) - \nabla^2 A$. $\endgroup$ – Willie Wong Apr 4 '12 at 13:03
  • $\begingroup$ Also note that the minimizers are in general not unique: if you replace $f \to f + \nabla h$ for some scalar function $h$, necessarily that $\nabla \times f = \nabla\times f + \nabla\times(\nabla h)$ so if $f$ is a minimizer and $h$ any function compactly supported in $\Omega$, $f + \nabla h$ is another minimizer. Generally this degree of freedom is gotten rid of by prescribing the divergence of $f$ (the constraint $\nabla\cdot f = k$ for some function $k$), in which case the double curl equation reduces to a Poisson equation. $\endgroup$ – Willie Wong Apr 4 '12 at 13:07

OP is mostly perfect until some later parts.

For the curl operator we have a similar integration by parts formula to the divergence theorem: $$ \int_{\Omega}\nabla \cdot (u\nabla v) = \int_{\Omega}(u\Delta v + \nabla u\cdot \nabla v) = \int_{\partial \Omega} u\nabla v\cdot \nu \,dS, $$ where $\nu$ is the outward unit normal to the boundary. The curl integration by parts formula is: $$ \int_{\Omega}\nabla \cdot (\phi\times \psi) = \int_{\Omega}(\psi\cdot\nabla \times \phi - \phi\cdot \nabla\times \psi) = \int_{\partial \Omega} \phi\times \psi\cdot \nu \,dS = \int_{\partial \Omega} \nu\times \phi\cdot \psi \,dS.\tag{1} $$

Say now you wanna minimize (here I use $u$ and $v$ instead of $f$ and $h$): $$ \mathcal{G}(v) = \int_{\Omega}|\nabla\times v|^2\,dx, $$ if the minimizer is $u$ and the variational limit is also no problem: $$ 0= \frac{d}{d\epsilon}\mathcal{G}(u+\epsilon v)\Bigg|_{\epsilon= 0} = 2\int_{\Omega} \nabla \times u\cdot \nabla \times v. $$

Now we use formula (1), letting $\phi = \nabla \times u$, and $\psi = v$: $$ \int_{\Omega}(v\cdot\nabla \times (\nabla \times u)- \nabla \times u\cdot \nabla\times v) =\left\{\begin{aligned} \int_{\partial \Omega} \big(\nu\times (\nabla \times u)\big)\cdot v\,dS \\ \\ \int_{\partial \Omega} (v\times \nu) \cdot (\nabla \times u)\,dS \end{aligned}\right. $$ There are two choices of way of writing boundary terms to be zero:

  1. You can set for the minimizer $\color{blue}{\nu\times (\nabla \times u)|_{\partial \Omega} = 0}$, and only tangential part suffices. $\nabla \times u|_{\partial \Omega = 0}$ is too strong.

  2. You can set for the test function spaces $\color{red}{\nu\times v = 0}$, then $\nu \times u$ can be prescribed as any permissible boundary data $g$ (those who have 0 surface divergence).

Choosing blue, you have to modular a gradient in a proper function space $H$ to set up a well-defined problem: $$ \left\{\begin{aligned} \nabla\times (\nabla \times u) &=0 \quad \text{ in } \Omega, \\ \nu \times (\nabla \times u) &= 0\quad \text{ on } \partial \Omega. \end{aligned}\right.\tag{2} $$ For $u \in H/\nabla p$ the quotient space, which means we don't tell the difference between $u $ and $u+\nabla p$, above problem has a unique solution. The only solution for (2) is zero, a.k.a. $u = 0$ or say $u = \nabla p$. Once we add a gauge condition $\nabla \cdot u = f$, then (2) is equivalent to the Neumann problem for $p$: $$ \left\{\begin{aligned} \Delta p &= f \quad \text{ in } \Omega, \\ \nabla p\cdot \nu &= 0\quad \text{ on } \partial \Omega. \end{aligned}\right. $$

Choosing red, the problem is: $$ \left\{\begin{aligned} \nabla\times (\nabla \times u) &=0 \quad \text{ in } \Omega, \\ \nu \times u &= g\quad \text{ on } \partial \Omega. \end{aligned}\right.\tag{3} $$ Again this problem has a unique solution in $H/\nabla p$, except this time $p$ is constant on the boundary which implies $\nabla p\times \nu =0$. If we set $g$ to be 0, then above problem again has a zero solution, which means $u = \nabla p$. Now a gauge condition $\nabla \cdot u = f$, then (3) can be transformed to a Dirichlet boundary problem for $p$: $$ \left\{\begin{aligned} \Delta p &= f \quad \text{ in } \Omega, \\ p &= \mathrm{Constant}\quad \text{ on } \partial \Omega. \end{aligned}\right. $$

Summary: the minimizer $u = \mathrm{arg}\min\mathcal{G}(v)$ is a gradient field $\nabla p$, the boundary condition of $p$ relies on the boundary condition of $u$. Both problem (2) and (3) don't require the normal component of $\nabla \times u$ or $u$ to be imposed. If you prescribe $\nabla\times u = 0$ on the boundary, which implies $\nabla\times u\cdot \nu = 0$, the problem may be over-determined.

  • $\begingroup$ thanks a lot for the insight! :) $\endgroup$ – WhitAngl Jun 25 '13 at 19:10
  • $\begingroup$ @WhitAngl Hey, while I was cooking up the answer I didn't realize it was a 2-year old question, have you got any trouble with this anymore recently? $\endgroup$ – Shuhao Cao Jun 25 '13 at 19:14
  • $\begingroup$ I was partly asking out of curiosity, and also due to a project I tried and didn't work out well... this could as well be to the way I formulated it (rather than your simpler Laplace equation), or a bug or anything. I might take another look at some point :) The answer was interesting anyway. Thanks! $\endgroup$ – WhitAngl Jun 25 '13 at 19:32

(This is supposed to be a shorter comment but I find myself powerless).

I haven't checked you reasoning, but I wanted to offer a simple (standard) proof. Some authors call this Dirichlet's principle.

Define $E(f) := \int_{\Omega} \| \nabla f \|^2 dx$. It's clear that $E(f) \geq 0$.

Assume $f, g$ are equal on $\partial \Omega$ Let's see that, if $f$ is harmonic then $E(f) \leq E(g)$.

For that, let $u := f - g$. Now calculate $E(g) = E(f - u)$ using Green's first identity (here you'll use that $u = 0$ on $\partial \Omega$ and that $\Delta f = 0$).

You'll get immediately that $E(g) = E(f) + E(u)$ , and using that $E(u) \geq 0$ that $E(f) \leq E(g)$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.