Let $\Omega\subset \mathbb R^N$ be open bounded, smooth boundary. Let $u_1$, $u_2\in H^{1}(\Omega)$ such that $T[u_1]=T[u_2]=T[\omega]$ where $T$ stands for the trace operator and $\omega\in H^1(\Omega)$ is a fixed function.

Define $$ F(u):=\inf_{v\in\mathcal V}\left\{\int_\Omega |\nabla u|^2v^2dx + \int_\Omega |\nabla v|^2+(1-v)^2dx \right\}, $$ where $\mathcal V:=\{v\in H^1(\Omega),\,0\leq v\leq 1\}$.

Question: does there exist a continuous path $a(t): [0,1]\to H^1(\Omega)$ between $u_1$ and $u_2$ satisfies the following conditions?

  1. $a(0)=u_1$, $a(1)=u_2$
  2. $T[a(t)]=T[\omega]$ for all $t\in (0,1)$
  3. $a(t)$ is continuous in $L^2$ sense, i.e., if $t\to t_0$, then $a(t)\to a(t_0)$ in $L^2$.
  4. $F(a(t))\leq \max\{F(u_1),F(u_2)\}$, for all $t\in (0,1)$.

Moreover, if $F(u_1)\leq F(a(t))\leq F(u_2)$ would be great, but it is not necessary... (here we assume $F(u_1)\leq F(u_2)$)

I tried the affine connection, i.e., $a(t)=tu_2+(1-t) u_2$, but it does not seems to work....

Any help, hint, or reference would be really welcome!

  • 1
    $\begingroup$ Why doesn't the affine path work? Seems to work to me... Moreover $\nabla\alpha_t=t\nabla u_2+(1-t)\nabla u_1$ and the convexity of the norm squared even gives you 4. and the additional condition you want. $\endgroup$
    – Teri
    Dec 4, 2015 at 18:09
  • $\begingroup$ @Teri: here the post is updated. $\endgroup$
    – JumpJump
    Dec 4, 2015 at 19:38
  • $\begingroup$ @Teri: I edit the problem a bit more. Now it is definitely a non-trivial problem... $\endgroup$
    – JumpJump
    Dec 5, 2015 at 18:28
  • $\begingroup$ Note that the compactness properties of $H^1(\Omega)$ imply that the infimum is attained. Using this one can still prove that $F$ is a convex functional, i.e. $F(tu_2+(1-t)u_1)\le tF(u_2)+(1-t)F(u_1)$. Property 4. and the additional condition always follow from this convexity. $\endgroup$
    – Teri
    Dec 6, 2015 at 20:50
  • $\begingroup$ @Teri Can you be more specific and write your comments as an answer? Thank you! $\endgroup$
    – JumpJump
    Dec 10, 2015 at 22:05

1 Answer 1


I've to admit that my original argument had a (stupid) mistake. However the argument below should work.

First let me relabel: call the two functions $u_1,u_2$ by $u_0,u_1$ so as to be compatible with $u_t:=tu_1+(1-t)u_0$.

Note that given any $u\in H^1(\Omega)$, the infimum in $$F(u)=\inf_{v\in \mathcal V}\{\int_\Omega|\nabla u|^2v^2 dx+\int_\Omega(|\nabla v|^2+(1-v)^2)dx\}$$ is obtained: take a minimizing sequence for $F(u)$ and, using the Rellich-Kondrachov theorem, find $v\in \mathcal V$ and a subsequence $v_k$ so that $v_k\to v$ in $L^2$, $\nabla v_k\rightharpoonup \nabla v$ weakly in $L^2$, and $v_k\to v$ pointwise almost everywhere. Then \begin{align*} F(u)\le &\int_\Omega|\nabla u|^2v^2 dx+\int_\Omega(|\nabla v|^2+(1-v)^2)dx\\ \le &\liminf_{k\to\infty}\left[\int_\Omega|\nabla u|^2v_k^2 dx+\int_\Omega(|\nabla v_k|^2+(1-v_k)^2)dx\right]=F(u). \end{align*}

Now let $w_j$ be a function in $\mathcal V$ for which the infimum $F(u_j)$ is attained, $j=0,1$, and let $w_t=tw_1+(1-t)w_0$. By the edit below $$t\mapsto w_t^2|\nabla u_t|^2$$ is convex so we may estimate \begin{align*} F(u_t)\le & \int_\Omega |\nabla u_t|^2w_t^2 dx+\int_\Omega(|\nabla w_t|^2+(1-w_t)^2)dx\\ \le & t\int_\Omega|\nabla u_1|^2w_1^2dx+(1-t)\int_\Omega |\nabla u_0|^2w_0^2dx\\ &\ \ + t\int_\Omega(|\nabla w_1|^2+(1-w_1)^2)dx+(1-t)\int_\Omega(|\nabla w_0|^2+(1-w_0)^2)dx\\ =&tF(u_1)+(1-t)F(u_0). \end{align*} Thus $F$ is convex and this gives you the properties you want.

EDIT: Denote $$f(t)=f_x(t)=|\nabla u_t(x)|^2, g(t)=g_x(t)=w_t(x)^2.$$ Both are convex functions, but this alone doesn't guarantee the convexity of their product. Instead we use the fact that they are essentially quadratic.

Let us show that $(fg)''(t)\ge 0$. Compute \begin{align*} f'(t)=&2\nabla(u_1-u_0)\cdot \nabla u_t\\ g'(t)=&2(w_1-w_0)w_t\\ f''(t)=& 2|\nabla(u_1-u_0)|^2\\ g''(t)=& 2(w_1-w_0)^2. \end{align*} Next check that \begin{align*} (fg)''(t)&=f''(t)g(t)+2f'(t)g'(t)+g''(t)f(t)\\ =&2[|\nabla(u_1-u_0)|^2w_ t^2+4w_t\nabla(u_1-u_0)\cdot (w_1-w_0)\nabla u_t+(w_1-w_0)^2|\nabla u_t|^2]\\ \ge& 2(|\nabla(u_1-u_0)|w_t-|w_1-w_0||\nabla u_t|)^2-4|w_1-w_0|w_t|\nabla(u_1-u_0)||\nabla u_t|\\ \ge & 0. \end{align*}

  • $\begingroup$ you mean that $w_t^2$ and $|\nabla u_t|^2$ is nonnegative and nondecreasing so that the product is convex? Since generally the product of nonnegative convex functions may not convex. However, we have $w_t^2$ is convex but $\nabla u_t$ may not be positive always. $\endgroup$
    – JumpJump
    Dec 11, 2015 at 20:30
  • $\begingroup$ I mean, why the function $|\nabla u_t|^2w_t^2$ is convex? We usually need $|\nabla u_t|^2$ and $w_t^2$ have nonnegative or nonpositive derivative at the same time. But I don't think we have in this case. $\endgroup$
    – JumpJump
    Dec 11, 2015 at 22:38
  • $\begingroup$ @tankonetoone You're right, the convexity of the product needs a separate argument, it doesn't follow automatically. I've made an edit to my answer to deal with this. $\endgroup$
    – Teri
    Dec 14, 2015 at 19:31
  • 3
    $\begingroup$ Sorry if this question is naive. But how you obtain the last inequality? You are basically saying that $(|a|-|b|)^2-2|ab|\geq 0$, which is not true in generally. For example, take $a=b$. $\endgroup$
    – JumpJump
    Dec 14, 2015 at 20:29
  • $\begingroup$ @tankonetoone You're right, again. I'm busy with a lot of stuff currently (hence the slow replies) but I'll try to see if anything can be done about the question. It seems that the convexity of $w_t^2|\nabla u_t|^2$ does not necessarily hold.. $\endgroup$
    – Teri
    Dec 19, 2015 at 14:52

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.