4
$\begingroup$

Let $K$ be a compact subset of $\mathbb{R}^n$ and $\mathcal{U}$ be a finite collection of open subsets covering $K$ satisfying the minimality property: for every $U\in \mathcal{U}$, the sub-collection $\mathcal{U}-\{U\}$ does not cover $K$.

Notation: Given any subset $A\subseteq K$ denote the relative complement $K-A:=\{x\in K:\, x\not\in A\}$. Given a point $x\in K$ and a subset $A\subseteq K$ let $\|x-A\|:=\inf_{a\in A} \|x-a\|$.

Is there a condition on the open cover $\mathcal{U}$ such that the partition of unity $\{\psi_U\}_{U \in \mathcal{U}}$ is such that each $$ \psi_U(x):= \frac{\|x-(K-U)\|^2}{\sum_{V \in \mathcal{U}}\,\|x-(K-V)\|^2} $$ is Lipschitz?

If so, what are explicit bounds on the Lipschitz constant of the functions $\{\psi_U\}_{U\in \mathcal{U}}$?

$\endgroup$
9
  • 1
    $\begingroup$ Note that if an open set $U$ can't be removed from the covering $\mathcal U$, i.e. $K\cap U\setminus \bigcup_{U\neq V\in\mathcal U}V\neq\emptyset$, then the $\psi_U$ supported in $U$ in any partition of unity subordinated to $\mathcal U$ must have Lipschitz constant at least 1/diam(U). $\endgroup$ Jun 13, 2022 at 11:21
  • $\begingroup$ @PietroMajer This I am having a lot of trouble seeing; I was also suspecting that $L=1/r$ where $r>0$ is the Lebesgue number of $\mathscr{U}$. (Ideally, do you have a reference to this or your claim?) $\endgroup$ Jun 13, 2022 at 11:30
  • $\begingroup$ Well, what I wrote is just this: if, for a given $x\in K$, $U$ is the only open set in $\mathcal U$ such that $x\in U$, then $\psi_V(x)=0$ for all $V\neq U$, so $\psi_U(x)=1$; since $\psi_U(y)=0$ on $\partial U$ we need the Lipschitz constant of $\psi_U$ at least 1/diam(U). $\endgroup$ Jun 13, 2022 at 13:49
  • $\begingroup$ @PietroMajer Ah true, but then can we upper-bound the Lipschitz constant in a similar fashion? $\endgroup$ Jun 13, 2022 at 13:57
  • 1
    $\begingroup$ @LSpice Thanks, its all fixed now; sorry about the typos. $\endgroup$ Jun 13, 2022 at 22:56

1 Answer 1

4
+200
$\begingroup$

For convenience write

$$ \phi_U(x) = \| x - (K-U)\|. $$

Let $N = |\mathcal{U}|$ the number of open sets in the cover. Define $$ f_1(x) = \frac{1}{N} \sum \phi_U(x), \qquad f_2(x) = \left( \frac1N \sum \phi_U^2(x) \right)^{1/2} $$ Note that since $\mathcal{U}$ is a cover we have $$ 0 < f_1(x) \leq f_2(x) \leq \sqrt{N} f_1(x) $$ Let $$ \delta_1 = \inf f_1, \quad \delta_2 = \inf f_2 $$ and we also have $0 < \delta_1 \leq \delta_2 \leq \sqrt{N} \delta_1 $. Note also that $\delta_1$ is a Lebesgue number of the covering $\mathcal{U}$.

Now, you've defined $$ \psi_U(x) = \frac{\phi_U^2(x)}{N f_2^2(x)} $$ so $$ N(\psi_U(x) - \psi_U(y)) = \frac{\phi_U^2(x) f_2^2(y) - \phi_U^2(y) f_2^2(x) }{f_2(x) f_2(y)} $$ $$ = \frac{(\phi_U(x) - \phi_U(y))(\phi_U(x) + \phi_U(y))}{f_2^2(x)} + \frac{\phi_U^2(y) (f_2^2(y) - f_2^2(x))}{f_2(x)^2 f_2(y)^2}. $$ So we have the following estimate $$ N |\psi_U(x) - \psi_U(y)| \leq \left|\frac{(\phi_U(x) - \phi_U(y))(\phi_U(x) + \phi_U(y))}{f_2^2(x)}\right| + N \left| \frac{f_2^2(y) - f_2^2(x)}{f_2(x)^2} \right| $$

Now, it suffices to consider only those $x,y$ with $\|x-y\| \leq \delta \leq \delta_1$ for a fixed $\delta$. This is because

  1. We know that $0 \leq \psi_U(x) \leq 1$ and hence if $\|x-y\| > \delta$, then the corresponding difference quotient is $\leq \frac{1}{\delta}$.
  2. On the other hand, if we define $$ \epsilon = \inf_{U\in \mathcal{U}} d( U \setminus \cup (\mathcal{U} \setminus \{U\}), K \setminus U ) $$ then essentially the same argument of Pietro Majer shows that the Lipschitz constant of the family is at least $1/\epsilon$. (Note: minimality of the covering is used to get that $U\setminus \cup(\mathcal{U}\setminus \{U\}) \neq\emptyset$.) Note that $\epsilon \geq \delta_1$. So we don't expect a much better upper bound estimate of the Lipschitz constant than $1/\delta$.

Now, $\phi_U(x)$ is well-known to be $1$-Lipschitz. So we have that $$\phi_U(x) + \phi_U(y) \leq 2 \phi_U(x) + \delta \leq (2N+1) f_1(x)$$ So $$ \left|\frac{(\phi_U(x) - \phi_U(y))(\phi_U(x) + \phi_U(y))}{f_2^2(x)}\right| \leq \frac{2N+1}{\delta_2} \|x - y\| $$ Similarly, we have $$ |f^2_2(x) - f^2_2(y)| \leq \frac1N \sum |\phi_U(x) - \phi_U(y)| |2\phi_U(x) + \delta| \leq \|x-y\| \cdot (\delta + 2 f_1(x)) $$ so $$ N \left| \frac{f_2^2(y) - f_2^2(x)}{f_2(x)^2} \right| \leq \frac{3N}{\delta_2} \|x-y\|$$

So putting everything together we find that the Lipschitz constant of the family $\psi_U$ are uniformly bounded by $6 / \delta_2$.

Summary

  1. For the finite collection $\mathcal{U}$, each function in the corresponding partition of unity is necessarily Lipschitz.
  2. The uniform Lipschitz constant can be controlled by a quantity similar to the Lebesgue number of the covering.
$\endgroup$
3
  • $\begingroup$ This has kept me busy for a day. The Lipschitz constant can also be written $6 \sqrt{N} / \delta_1$, which looks like a much better upper bound than what the published literature. Then it finally dawned on me that $\delta_1$ is quite a conservative lower bound for the Lebesgue number. It may generally only be the $N$-th part of the best possible Lebesgue number $\delta_\ast$. The best known bounds actually scale like $N / \delta_\ast$. $\endgroup$
    – shuhalo
    Aug 24, 2023 at 22:50
  • $\begingroup$ Is there a specific reason why one would use the squares of the distance functions rather than the distance functions themselves? $\endgroup$
    – shuhalo
    Aug 29, 2023 at 15:38
  • $\begingroup$ I used the square because that's what is used in the original question. @shuhalo Your question is probably better directed to the OP. $\endgroup$ Aug 30, 2023 at 1:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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