For $\phi\in C^1(\mathbb{R}^N)$ with $$\omega_{\phi}=\{x\in\mathbb{R}^N\ |\ \phi(x)>0\}$$ being a bounded set with $\nabla\phi (x)\neq 0,\ \forall\ x\in\phi^{-1}(0)=\partial\omega_{\phi}\neq \emptyset$ consider the signed distance function:

$$ d_{\phi}:\mathbb{R}^N\to\mathbb{R},\ d_{\phi}(x)=\begin{cases} -\text{dist}(x,\partial\omega_{\phi}),\ x\in\omega_{\phi} \\ 0,\ x\in\partial\omega_{\phi} \\ \text{dist}(x,\partial\omega_{\phi}),\ x\in\mathbb{R}^N\setminus\overline{\omega}_{\phi}\end{cases}$$

I have some technical questions that seem to be true about $d_{\phi}$.

Let's take a point $x_0\in\mathbb{R}^N$ with a unique $y(0)\in\partial\omega_{\phi}$ such that: $|x_0-y|=\text{dist}(x_0,\partial\omega_{\phi})$ and any $\psi\in C^1(\mathbb{R}^N)$ with $|\psi(x)|<M,\forall\ x\in\mathbb{R}^N$.

For $|\theta|$ sufficiently small we associate to $x_0$ a set denoted $Y(\theta)$ such that $\forall y\in Y(\theta)\subseteq\mathbb{R}^N$ we have $\phi(y)+\theta\psi(y)=0$ and:


How can we prove that:

(1) There is a selection for the multifunction $Y:[0;\theta_0)\rightrightarrows \phi^{-1}(0)$ denoted $y=y(\theta)$ which is continuous and differentiable for $\theta\in [0,\theta_0)$.

Note that we cannot apply here Michael's Selection Theorem because in general $Y(\theta)$ is not a convex set. (https://en.wikipedia.org/wiki/Michael_selection_theorem)

(2) $\lim\limits_{\theta\to 0} d_{\phi+\theta\psi}(x_0)=d_{\phi}(x_0)$

(3) Exists $\lim\limits_{\theta\to 0}\dfrac{d_{\phi+\theta\psi}(x_0)-d_{\phi}(x_0)}{\theta}$

I tried a lot to prove the continuity but I did not succeed.

It is well known (from KKT conditions) that $y(0)$ and $y(\theta)$ satisfy:

$$\begin{cases} \phi(y(0))=0,\ \phi(y(\theta))+\theta\psi(y(\theta))=0 \\ \lambda\nabla \phi (y(0))=y(0)-x_0,\ \lambda_{\theta}\Big (\nabla\phi(y(\theta))+\theta\psi(y(\theta))\Big )=y(\theta)-x_0\end{cases} $$

I proved easily that for $\theta$ sufficiently small $\nabla\phi+\theta\nabla\psi\neq 0$ on $(\phi+\theta\psi)^{-1}(0)\neq \emptyset$. Also note that if we prove $(1)$ then the answer to $(2)$ follows immediately and $(3)$ can be computed from KKT conditions given the result: $-\dfrac{\psi(y(0))}{|\nabla\phi(y(0))|}$.

Maybe the following sources will help:

Set of points with a unique closest point in a compact set

Concavity near the boundary of the distance function

Finally, are there some references that treats the signed distance function with the level set method (not with a shape derivative approach, but a functional approach)?


Your assertions are wrong. Assume that in a neighboorhood of $(0,0)$ we have $$ \phi(x)=(x_1-1)^2-1+x_2^2-x_2^4\\ \psi(x)=x_2^2 $$ Then for $x_0=(1,0)$ the unique minimum is $y_0=(0,0)$. However for $\theta>0$ there are multiple minima regardless of how small $\theta$ is.

  • 1
    $\begingroup$ Nice example! However we can take a selection of that multifunction that associates to $x_0$ the points on $\phi^{-1}(0)$ which realizes the minimum distance between $x_0$ and $\phi^{-1}(0)$ that is both continuous and differentiable w.r.t. $\theta$ (in your example we could take $y>0$). I don't think that all my assertions are wrong. Somehow I could calculate the limit in $(4)$ being $-\dfrac{\psi(y(0))}{|\nabla\phi(y(0))|}$, but the details are a bit messy (even on the continuity condition) which is true in general. $\endgroup$
    – Bogdan
    Dec 31 '20 at 17:20
  • $\begingroup$ What assumptions are wrong? $\endgroup$ Jan 4 at 0:40
  • $\begingroup$ I have edited my post. Now I think there are none. $\endgroup$
    – Bogdan
    Jan 4 at 6:02

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.